An Introduction to Computer Networks

Release 1.0

Peter L Dordal

February 04, 2015

CONTENTS

0 Preface
0.1 Classroom Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
0.2 Progress Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
0.3 Technical considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1 An Overview of Networks
1.1 Layers . . . . . . . . . . . . .
1.2 Bandwidth and Throughput . .
1.3 Packets . . . . . . . . . . . . .
1.4 Datagram Forwarding . . . . .
1.5 Topology . . . . . . . . . . . .
1.6 Routing Loops . . . . . . . . .
1.7 Congestion . . . . . . . . . . .
1.8 Packets Again . . . . . . . . .
1.9 LANs and Ethernet . . . . . .
1.10 IP - Internet Protocol . . . . .
1.11 DNS . . . . . . . . . . . . .
1.12 Transport . . . . . . . . . . .
1.13 Firewalls . . . . . . . . . . .
1.14 Network Address Translation
1.15 IETF and OSI . . . . . . . . .
1.16 Berkeley Unix . . . . . . . .
1.17 Epilog . . . . . . . . . . . . .
1.18 Exercises . . . . . . . . . . .
2 Ethernet
2.1 10-Mbps classic Ethernet
2.2 100 Mbps (Fast) Ethernet
2.3 Gigabit Ethernet . . . . .
2.4 Ethernet Switches . . . .
2.5 Spanning Tree Algorithm
2.6 Virtual LAN (VLAN) . .
2.7 Epilog . . . . . . . . . .
2.8 Exercises . . . . . . . . .
3 Other LANs

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

3
3
5
5

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

7
7
7
8
9
11
12
12
13
14
16
21
22
24
25
26
29
29
29

.
.
.
.
.
.
.
.

33
33
43
44
45
47
51
52
52
55
i

3.1 Virtual Private Network . . . . . . .

3.2 Carrier Ethernet . . . . . . . . . . .
3.3 Wi-Fi . . . . . . . . . . . . . . . . .
3.4 WiMAX . . . . . . . . . . . . . . .
3.5 Fixed Wireless . . . . . . . . . . . .
3.6 Token Ring . . . . . . . . . . . . . .
3.7 Virtual Circuits . . . . . . . . . . . .
3.8 Asynchronous Transfer Mode: ATM
3.9 Epilog . . . . . . . . . . . . . . . .
3.10 Exercises . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

55
56
57
68
70
73
75
78
80
80

4 Links
4.1 Encoding and Framing . . .
4.2 Time-Division Multiplexing
4.3 Epilog . . . . . . . . . . .
4.4 Exercises . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

85
85
89
93
93

5 Packets
5.1 Packet Delay . . . . . .
5.2 Packet Delay Variability
5.3 Packet Size . . . . . . .
5.4 Error Detection . . . . .
5.5 Epilog . . . . . . . . .
5.6 Exercises . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

95
95
98
99
101
105
106

6 Abstract Sliding Windows

6.1 Building Reliable Transport: Stop-and-Wait
6.2 Sliding Windows . . . . . . . . . . . . . . .
6.3 Linear Bottlenecks . . . . . . . . . . . . . .
6.4 Epilog . . . . . . . . . . . . . . . . . . . .
6.5 Exercises . . . . . . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

109
109
113
117
123
123

7 IP version 4
7.1 The IPv4 Header . . . . . . . . . . . . . . . .
7.2 Interfaces . . . . . . . . . . . . . . . . . . . .
7.3 Special Addresses . . . . . . . . . . . . . . .
7.4 Fragmentation . . . . . . . . . . . . . . . . .
7.5 The Classless IP Delivery Algorithm . . . . .
7.6 IP Subnets . . . . . . . . . . . . . . . . . . .
7.7 Address Resolution Protocol: ARP . . . . . .
7.8 Dynamic Host Configuration Protocol (DHCP)
7.9 Internet Control Message Protocol . . . . . .
7.10 Unnumbered Interfaces . . . . . . . . . . . .
7.11 Mobile IP . . . . . . . . . . . . . . . . . . .
7.12 Epilog . . . . . . . . . . . . . . . . . . . . .
7.13 Exercises . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

127
127
129
130
131
133
135
140
143
145
148
149
150
150

8 IP version 6

ii

.
.
.
.
.
.

.
.
.
.
.
.

153

8.1 The IPv6 Header . . . . . . . . . .

8.2 IPv6 addresses . . . . . . . . . . .
8.3 IPv6 Multicast . . . . . . . . . . .
8.4 IPv6 Extension Headers . . . . . .
8.5 Network Prefixes . . . . . . . . . .
8.6 Neighbor Discovery . . . . . . . .
8.7 IPv6 Host Address Assignment . .
8.8 Globally Exposed Addresses . . .
8.9 ICMPv6 . . . . . . . . . . . . . .
8.10 Routerless Connection Examples
8.11 IPv6-to-IPv4 connectivity . . . .
8.12 Epilog . . . . . . . . . . . . . . .
8.13 Exercises . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

153
154
156
156
158
159
160
164
164
165
167
167
167

9 Routing-Update Algorithms
9.1 Distance-Vector Routing-Update Algorithm
9.2 Distance-Vector Slow-Convergence Problem
9.3 Observations on Minimizing Route Cost . .
9.4 Loop-Free Distance Vector Algorithms . . .
9.5 Link-State Routing-Update Algorithm . . .
9.6 Routing on Other Attributes . . . . . . . . .
9.7 Epilog . . . . . . . . . . . . . . . . . . . .
9.8 Exercises . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

169
169
174
176
178
180
183
184
184

.
.
.
.
.
.
.
.

189
189
191
192
192
196
197
209
210

10 Large-Scale IP Routing
10.1 Classless Internet Domain Routing: CIDR
10.2 Hierarchical Routing . . . . . . . . . . .
10.3 Legacy Routing . . . . . . . . . . . . . .
10.4 Provider-Based Routing . . . . . . . . .
10.5 Geographical Routing . . . . . . . . . .
10.6 Border Gateway Protocol, BGP . . . . .
10.7 Epilog . . . . . . . . . . . . . . . . . . .
10.8 Exercises . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

11 UDP Transport
11.1 User Datagram Protocol UDP . . .
11.2 Fundamental Transport Issues . . . .
11.3 Trivial File Transport Protocol, TFTP
11.4 Remote Procedure Call (RPC) . . . .
11.5 Epilog . . . . . . . . . . . . . . . . .
11.6 Exercises . . . . . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

215
215
224
225
230
234
234

12 TCP Transport
12.1 The End-to-End Principle . . .
12.2 TCP Header . . . . . . . . . . .
12.3 TCP Connection Establishment
12.4 TCP and WireShark . . . . . .
12.5 TCP simplex-talk . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

237
238
238
239
243
245

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

iii

12.6 TCP state diagram . . . . . . . . . .

12.7 TCP Old Duplicates . . . . . . . . .
12.8 TIMEWAIT . . . . . . . . . . . . . .
12.9 The Three-Way Handshake Revisited
12.10 Anomalous TCP scenarios . . . . .
12.11 TCP Faster Opening . . . . . . . . .
12.12 Path MTU Discovery . . . . . . . .
12.13 TCP Sliding Windows . . . . . . . .
12.14 TCP Delayed ACKs . . . . . . . . .
12.15 Nagle Algorithm . . . . . . . . . .
12.16 TCP Flow Control . . . . . . . . . .
12.17 TCP Timeout and Retransmission .
12.18 KeepAlive . . . . . . . . . . . . . .
12.19 TCP timers . . . . . . . . . . . . .
12.20 Epilog . . . . . . . . . . . . . . . .
12.21 Exercises . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

249
251
252
253
255
256
256
257
257
258
258
259
259
260
260
260

13 TCP Reno and Congestion Management

13.1 Basics of TCP Congestion Management
13.2 Slow Start . . . . . . . . . . . . . . . .
13.3 TCP Tahoe and Fast Retransmit . . . .
13.4 TCP Reno and Fast Recovery . . . . .
13.5 TCP NewReno . . . . . . . . . . . . .
13.6 SACK TCP . . . . . . . . . . . . . . .
13.7 TCP and Bottleneck Link Utilization .
13.8 Single Packet Losses . . . . . . . . . .
13.9 TCP Assumptions and Scalability . . .
13.10 TCP Parameters . . . . . . . . . . . .
13.11 Epilog . . . . . . . . . . . . . . . . .
13.12 Exercises . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

263
264
268
272
274
276
279
280
282
283
283
284
284

14 Dynamics of TCP Reno

14.1 A First Look At Queuing . . . . . . . .
14.2 Bottleneck Links with Competition . .
14.3 TCP Fairness with Synchronized Losses
14.4 Notions of Fairness . . . . . . . . . . .
14.5 TCP Reno loss rate versus cwnd . . . .
14.6 TCP Friendliness . . . . . . . . . . . .
14.7 AIMD Revisited . . . . . . . . . . . .
14.8 Active Queue Management . . . . . . .
14.9 The High-Bandwidth TCP Problem . .
14.10 The Lossy-Link TCP Problem . . . .
14.11 The Satellite-Link TCP Problem . . .
14.12 Epilog . . . . . . . . . . . . . . . . .
14.13 Exercises . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

289
289
290
297
304
305
307
309
311
313
315
315
315
315

15 Newer TCP Implementations

323
15.1 High-Bandwidth Desiderata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324

iv

15.2 RTTs . . . . .
15.3 Highspeed TCP
15.4 TCP Vegas . .
15.5 FAST TCP . .
15.6 TCP Westwood
15.7 TCP Veno . . .
15.8 TCP Hybla . .
15.9 TCP Illinois . .
15.10 H-TCP . . . .
15.11 TCP CUBIC .
15.12 Epilog . . . .
15.13 Exercises . .

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

324
325
326
328
330
332
333
334
335
336
340
341

16 Network Simulations: ns-2

16.1 The ns-2 simulator . . . . . . . . . . . . .
16.2 A Single TCP Sender . . . . . . . . . . . .
16.3 Two TCP Senders Competing . . . . . . .
16.4 TCP Loss Events and Synchronized Losses
16.5 TCP Reno versus TCP Vegas . . . . . . . .
16.6 Wireless Simulation . . . . . . . . . . . .
16.7 Epilog . . . . . . . . . . . . . . . . . . . .
16.8 Exercises . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

347
347
349
361
376
387
389
395
395

17 The ns-3 Network Simulator

17.1 Installing and Running ns-3
17.2 A Single TCP Sender . . . .
17.3 Wireless . . . . . . . . . . .
17.4 Exercises . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

399
399
400
408
414

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

415
415
416
416
417
418
430
431
434
440
444
446
448
450
452
452

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.

18 Queuing and Scheduling

18.1 Queuing and Real-Time Traffic . . . . . . .
18.2 Traffic Management . . . . . . . . . . . . .
18.3 Priority Queuing . . . . . . . . . . . . . . .
18.4 Queuing Disciplines . . . . . . . . . . . . .
18.5 Fair Queuing . . . . . . . . . . . . . . . . .
18.6 Applications of Fair Queuing . . . . . . . . .
18.7 Hierarchical Queuing . . . . . . . . . . . . .
18.8 Hierarchical Weighted Fair Queuing . . . . .
18.9 Token Bucket Filters . . . . . . . . . . . . .
18.10 Applications of Token Bucket . . . . . . . .
18.11 Token Bucket Queue Utilization . . . . . .
18.12 Hierarchical Token Bucket . . . . . . . . .
18.13 Fair Queuing / Token Bucket combinations .
18.14 Epilog . . . . . . . . . . . . . . . . . . . .
18.15 Exercises . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

19 Quality of Service
457
19.1 Net Neutrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458

19.2 Where the Wild Queues Are . . . . . . . .

19.3 Real-time Traffic . . . . . . . . . . . . . .
19.4 Integrated Services / RSVP . . . . . . . . .
19.5 Global IP Multicast . . . . . . . . . . . . .
19.6 RSVP . . . . . . . . . . . . . . . . . . . .
19.7 Differentiated Services . . . . . . . . . . .
19.8 RED with In and Out . . . . . . . . . . . .
19.9 NSIS . . . . . . . . . . . . . . . . . . . .
19.10 Comcast Congestion-Management System
19.11 Real-time Transport Protocol (RTP) . . .
19.12 Multi-Protocol Label Switching (MPLS) .
19.13 Epilog . . . . . . . . . . . . . . . . . . .
19.14 Exercises . . . . . . . . . . . . . . . . .
20 Network Management and SNMP
20.1 Network Architecture . . . . . . . . .
20.2 SNMP Basics . . . . . . . . . . . . .
20.3 SNMP Naming and OIDs . . . . . .
20.4 MIBs . . . . . . . . . . . . . . . . .
20.5 SNMPv1 Data Types . . . . . . . . .
20.6 ASN.1 Syntax and SNMP . . . . . .
20.7 SNMP Tables . . . . . . . . . . . . .
20.8 SNMP Operations . . . . . . . . . .
20.9 MIB Browsing . . . . . . . . . . . .
20.10 MIB-2 . . . . . . . . . . . . . . . .
20.11 SNMPv1 communities and security
20.12 SNMP and ASN.1 Encoding . . . .
20.13 SNMPv2 . . . . . . . . . . . . . . .
20.14 Table Row Creation . . . . . . . . .
20.15 SNMPv3 . . . . . . . . . . . . . . .
20.16 Exercises . . . . . . . . . . . . . .
21 Security
21.1 Code-Execution Intrusion . . . .
21.2 Stack Buffer Overflow . . . . . .
21.3 Heap Buffer Overflow . . . . . .
21.4 Network Intrusion Detection . . .
21.5 Secure Hashes . . . . . . . . . .
21.6 Shared-Key Encryption . . . . . .
21.7 Diffie-Hellman-Merkle Exchange
21.8 Public-Key Encryption . . . . . .
21.9 SSH and TLS . . . . . . . . . . .
21.10 Exercises . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

458
459
461
462
466
468
472
472
473
474
478
480
481

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

483
485
485
486
488
489
490
491
495
500
501
509
511
513
524
532
541

.
.
.
.
.
.
.
.
.
.

545
546
547
555
561
562
564
573
575
578
587

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

Bibliography

591

Indices and tables

593

vi

Bibliography

595

Index

601

vii

viii

An Introduction to Computer Networks, Release 1.0

Peter L Dordal
Department of Computer Science
Loyola University Chicago

Contents:

CONTENTS

An Introduction to Computer Networks, Release 1.0

CONTENTS

0 PREFACE

No man but a blockhead ever wrote, except for money. - Samuel Johnson
The textbook world is changing. On the one hand, open source software and creative-commons licensing
have been great successes; on the other hand, unauthorized PDFs of popular textbooks are widely available,
and it is time to consider flowing with rather than fighting the tide. Hence this open textbook, released for
free under the Creative Commons license described below. Mene, mene, tekel pharsin.
Perhaps the last straw, for me, was patent 8195571 for a roundabout method to force students to purchase
textbooks. (A simpler strategy might be to include the price of the book in the course.) At some point,
faculty have to be advocates for their students rather than, well, Hirudinea.
This is not to say that I have anything against for-profit publishing. It is just that this particular book does not
and will not belong to that category. In this it is in good company: there is Wikipedia, there is Gnu/Linux,
and there is an increasing number of other free online textbooks out there. The market inefficiencies of
traditional publishing are sobering: the return to authors of advanced textbooks is at best modest, and costs
to users are quite high.
This text is released under the Creative Commons license Attribution-NonCommercial-NoDerivs; this text
is like a conventional book, in other words, except that it is free. You may copy the work and distribute it
to others, but reuse requires attribution. Creation of derivative works eg modifying chapters or creating
additional chapters and distributing them as part of this work also requires permission.
The work may not be used for commercial purposes without permission. Permission is likely to be granted
for use and distribution of all or part of the work in for-profit and commercial training programs, provided
there is no direct charge to recipients for the work and provided the free nature of the work is made clear to
recipients (eg by including this preface). However, such permission must always be requested. Alternatively,
participants in commercial programs may be instructed to download the work individually.
The official book website (potentially subject to change) is [Link]. The book is available
there as online html, as a zipped archive of html files, in .pdf format, and in other formats as may prove
useful.

0.1 Classroom Use

This book is meant as a serious and more-or-less thorough text for an introductory college or graduate course
in computer networks, carefully researched, with consistent notation and style, and complete with diagrams
and exercises. My intent is to create a text that covers to a reasonable extent why the Internet is the way it is,
to avoid the endless dreary focus on TLAs (Three-Letter Acronyms), and to remain not too mathematical.
For the last, I have avoided calculus, linear algebra, and, for that matter, quadratic terms (though some
inequalities do sneak in at times). That said, the book includes a large number of back-of-the-envelope
calculations in settings as concrete as I could make them illustrating various networking concepts.
Overall, I tried to find a happy medium between practical matters and underlying principles. My goal has
been to create a book that is useful to a broad audience, including those interested in network management,
in high-performance networking, in software development, or just in how the Internet is put together.

An Introduction to Computer Networks, Release 1.0

The book can also be used as a networks supplement or companion to other resources for a variety of
other courses that overlap to some greater or lesser degree with networking. At Loyola, earlier versions of
this material have been used coupled with a second textbook in courses in computer security, network
management, telecommunications, and even introduction-to-computing courses for non-majors. Another
possibility is an alternative or nontraditional presentation of networking itself. It is when used in concert
with other works, in particular, that this books being free is of marked advantage.
Finally, I hope the book may also be useful as a reference work. To this end, I have attempted to ensure that
the indexing and cross-referencing is sufficient to support the drop-in reader. Similarly, obscure notation is
kept to a minimum.
Much is sometimes made, in the world of networking textbooks, about top-down versus bottom-up sequencing. This book is not really either, although the chapters are mostly numbered in bottom-up fashion.
Instead, the first chapter provides a relatively complete overview of the LAN, IP and transport network layers
(along with a few other things), allowing subsequent chapters to refer to all network layers without forward
reference, and, more importantly, allowing the chapters to be covered in a variety of different orders. As a
practical matter, when I use this text to teach Loyolas Introduction to Computer Networks course, I cover
the IP and TCP material more or less in parallel.
A distinctive feature of the book is the extensive coverage of TCP: TCP dynamics, newer versions of TCP
such as TCP Cubic, and a chapter on using the ns-2 simulator to explore actual TCP behavior. This has
its roots in a longstanding goal to find better ways to present competition and congestion in the classroom.
Another feature is the detailed chapter on queuing disciplines.
One thing this book makes little attempt to cover in detail is the application layer; the token example included is SNMP. While SNMP actually makes a pretty good example of a self-contained application, my
recommendation to instructors who wish to cover more familiar examples is to combine this text with the
appropriate application documentation.
For those interested in using the book for a traditional networks course, I with some trepidation offer the
following set of core material. In solidarity with those who prefer alternatives to a bottom-up ordering, I
emphasize that this represents a set and not a sequence.
1 An Overview of Networks
Selected sections from 2 Ethernet, particularly switched Ethernet
Selected sections from 3.3 Wi-Fi
Selected sections from 5 Packets
6 Abstract Sliding Windows
7 IP version 4 and/or 8 IP version 6
Selected sections from 9 Routing-Update Algorithms and 10 Large-Scale IP Routing
11 UDP Transport
12 TCP Transport
13 TCP Reno and Congestion Management
With some care in the topic-selection details, the above can be covered in one semester along with a survey
of selected important network applications, or the basics of network programming, or the introductory con-

0 Preface

An Introduction to Computer Networks, Release 1.0

figuration of switches and routers, or coverage of additional material from this book, or some other set of
additional topics. Of course, non-traditional networks courses may focus on a quite different sets of topics.

Peter Dordal
Shabbona, Illinois

0.2 Progress Notes

Edition 1.0 was declared complete as of March 2014.
The [Link] website carries both edition 1.0 and also the current 1.x edition, x>0. My plan
is to make sure, as far as is possible, that the two editions are classroom-compatible: existing exercises
will not be renumbered (though new exercises may be introduced with fractional numbers), and section
renumbering will be avoided to the extent practical.
As of November 2014 chapters on network management and security have been added; this completes the
set of chapters originally envisioned.
At this point I am actively seeking reviewers either for style or for technical accuracy.

0.3 Technical considerations

The book was prepared in reStructuredText using the linux Sphinx package, which can produce multiple
formats from the same source. That said, the primary format is html.
In order to read this book online in html format you will need the appropriate unicode character set installed.
At a minimum, the following characters must display properly:

=
The characters above should look roughly as they do in the following image:

If no available browser displays these properly, I recommend the pdf or epub formats. Generally Firefox
handles the necessary characters out of the box, as does Internet Explorer, but Chrome does not.
The diagrams in the body of the text are now all in bitmap .png format, although a few diagrams rendered
with line-drawing characters still appear in the exercises. I would prefer to use the vector-graphics .svg
format, but as of January 2014 most browsers do not appear to support zooming in on .svg images, which is
really the whole point.

0.2 Progress Notes

An Introduction to Computer Networks, Release 1.0

0 Preface

1 AN OVERVIEW OF NETWORKS

Somewhere there might be a field of interest in which the order of presentation of topics is well agreed upon.
Computer networking is not it.
There are many interconnections in the field of networking, as in most technical fields, and it is difficult
to find an order of presentation that does not involve endless forward references to future chapters; this
is true even if as is done here a largely bottom-up ordering is followed. I have therefore taken here a
different approach: this first chapter is a summary of the essentials LANs, IP and TCP across the board,
and later chapters expand on the material here.
Local Area Networks, or LANs, are the physical networks that provide the connection between machines
within, say, a home, school or corporation. LANs are, as the name says, local; it is the IP, or Internet
Protocol, layer that provides an abstraction for connecting multiple LANs into, well, the Internet. Finally,
TCP deals with transport and connections and actually sending user data.
This chapter also contains some important other material. The section on datagram forwarding, central
to packet-based switching and routing, is essential. This chapter also discusses packets generally, congestion, and sliding windows, but those topics are revisited in later chapters. Firewalls and network address
translation are also covered here and not elsewhere.

1.1 Layers
These three topics LANs, IP and TCP are often called layers; they constitute the Link layer, the Internetwork layer, and the Transport layer respectively. Together with the Application layer (the software you use),
these form the four-layer model for networks. A layer, in this context, corresponds strongly to the idea
of a programming interface or library (though some of the layers are not accessible to ordinary users): an
application hands off a chunk of data to the TCP library, which in turn makes calls to the IP library, which
in turn calls the LAN layer for actual delivery.
The LAN layer is in charge of actual delivery of packets, using LAN-layer-supplied addresses. It is often
conceptually subdivided into the physical layer dealing with, eg, the analog electrical, optical or radio
signaling mechanisms involved, and above that an abstracted logical LAN layer that describes all the
digital that is, non-analog operations on packets; see 2.1.2 The LAN Layer. The physical layer is
generally of direct concern only to those designing LAN hardware; the kernel software interface to the LAN
corresponds to the logical LAN layer. This LAN physical/logical division gives us the Internet five-layer
model. This is less a formal hierarchy as an ad hoc classification method. We will return to this below in
1.15 IETF and OSI.

1.2 Bandwidth and Throughput

Any one network connection eg at the LAN layer has a data rate: the rate at which bits are transmitted.
In some LANs (eg Wi-Fi) the data rate can vary with time. Throughput refers to the overall effective

An Introduction to Computer Networks, Release 1.0

transmission rate, taking into account things like transmission overhead, protocol inefficiencies and perhaps
even competing traffic. It is generally measured at a higher network layer than the data rate.
The term bandwidth can be used to refer to either of these, though we here try to use it mostly as a synonym
for data rate. The term comes from radio transmission, where the width of the frequency band available is
proportional, all else being equal, to the data rate that can be achieved.
In discussions about TCP, the term goodput is sometimes used to refer to what might also be called
application-layer throughput: the amount of usable data delivered to the receiving application. Specifically, retransmitted data is counted only once when calculating goodput but might be counted twice under
some interpretations of throughput.
Data rates are generally measured in kilobits per second (Kbps) or megabits per second (Mbps); in the
context of data rates, a kilobit is 103 bits (not 210 ) and a megabit is 106 bits. The use of the lowercase b
means bits; data rates expressed in terms of bytes often use an upper-case B.

1.3 Packets
Packets are modest-sized buffers of data, transmitted as a unit through some shared set of links. Of necessity,
packets need to be prefixed with a header containing delivery information. In the common case known as
datagram forwarding, the header contains a destination address; headers in networks using so-called
virtual-circuit forwarding contain instead an identifier for the connection. Almost all networking today
(and for the past 50 years) is packet-based, although we will later look briefly at some circuit-switched
options for voice telephony.

At the LAN layer, packets can be viewed as the imposition of a buffer (and addressing) structure on top
of low-level serial lines; additional layers then impose additional structure. Informally, packets are often
referred to as frames at the LAN layer, and as segments at the Transport layer.
The maximum packet size supported by a given LAN (eg Ethernet, Token Ring or ATM) is an intrinsic
attribute of that LAN. Ethernet allows a maximum of 1500 bytes of data. By comparison, TCP/IP packets
originally often held only 512 bytes of data, while early Token Ring packets could contain up to 4KB of
data. While there are proponents of very large packet sizes, larger even than 64KB, at the other extreme the
ATM (Asynchronous Transfer Mode) protocol uses 48 bytes of data per packet, and there are good reasons
for believing in modest packet sizes.
One potential issue is how to forward packets from a large-packet LAN to (or through) a small-packet LAN;
in later chapters we will look at how the IP (or Internet Protocol) layer addresses this.

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

Generally each layer adds its own header. Ethernet headers are typically 14 bytes, IP headers 20 bytes, and
TCP headers 20 bytes. If a TCP connection sends 512 bytes of data per packet, then the headers amount to
10% of the total, a not-unreasonable overhead. For one common Voice-over-IP option, packets contain 160
bytes of data and 54 bytes of headers, making the header about 25% of the total. Compressing the 160 bytes
of audio, however, may bring the data portion down to 20 bytes, meaning that the headers are now 73% of
the total; see 19.11.4 RTP and VoIP.
In datagram-forwarding networks the appropriate header will contain the address of the destination and
perhaps other delivery information. Internal nodes of the network called routers or switches will then make
sure that the packet is delivered to the requested destination.
The concept of packets and packet switching was first introduced by Paul Baran in 1962 ([PB62]). Barans
primary concern was with network survivability in the event of node failure; existing centrally switched
protocols were vulnerable to central failure. In 1964, Donald Davies independently developed many of the
same concepts; it was Davies who coined the term packet.
It is perhaps worth noting that packets are buffers built of 8-bit bytes, and all hardware today agrees what
a byte is (hardware agrees by convention on the order in which the bits of a byte are to be transmitted).
8-bit bytes are universal now, but it was not always so. Perhaps the last great non-byte-oriented hardware
platform, which did indeed overlap with the Internet era broadly construed, was the DEC-10, which had a
36-bit word size; a word could hold five 7-bit ASCII characters. The early Internet specifications introduced
the term octet (an 8-bit byte) and required that packets be sequences of octets; non-octet-oriented hosts had
to be able to convert. Thus was chaos averted. Note that there are still byte-oriented data issues; as one
example, binary integers can be represented as a sequence of bytes in either big-endian or little-endian byte
order. RFC 1700 specifies that Internet protocols use big-endian byte order, therefore sometimes called
network byte order.

1.4 Datagram Forwarding

In the datagram-forwarding model of packet delivery, packet headers contain a destination address. It is up
to the intervening switches or routers to look at this address and get the packet to the correct destination.
In the diagram below, switch S1 has interfaces 0, 1 and 2, and S2 has interfaces 0,1,2,3. If A is to send a
packet P to B, S1 must know that P must be forwarded out interface 2 and S2 must know P must be forwarded
out interface 3. In datagram forwarding this is achieved by providing each switch with a forwarding table of
destination,next_hop pairs. When a packet arrives, the switch looks up the destination address (presumed
globally unique) in this table, and finds the next_hop information: the address to which or interface by
which the packet should be forwarded in order to bring it one step closer to its final destination. In the
network below, a complete forwarding table for S1 (using interface numbers as next_hop values) would be:
S1
destination
A
C
B
D
E

next_hop
0
1
2
2
2

1.4 Datagram Forwarding

An Introduction to Computer Networks, Release 1.0

The table for S2 might be as follows, where we have consolidated destinations A and C for visual simplicity.
S2
destination
A,C
D
E
B

next_hop
0
1
2
3

Alternatively, we could replace the interface information with next-node, or neighbor, information, as all
the links above are point-to-point and so each interface connects to a unique neighbor. In that case, S1s
table might be written as follows (with consolidation of the entries for B, D and E):
S1
destination
A
C
B,D,E

next_hop
A
C
S2

A central feature of datagram forwarding is that each packet is forwarded in isolation; the switches involved do not have any awareness of any higher-layer logical connections established between endpoints.
This is also called stateless forwarding, in that the forwarding tables have no per-connection state. RFC
1122 put it this way (in the context of IP-layer datagram forwarding):
To improve robustness of the communication system, gateways are designed to be stateless,
forwarding each IP datagram independently of other datagrams. As a result, redundant paths can
be exploited to provide robust service in spite of failures of intervening gateways and networks.
Datagram forwarding is sometimes allowed to use other information beyond the destination address. In
theory, IP routing can be done based on the destination address and some quality-of-service information,
allowing, for example, different routing to the same destination for high-bandwidth bulk traffic and for lowlatency real-time traffic. In practice, many ISPs ignore quality-of-service information in the IP header, and
route only based on the destination.
By convention, switching devices acting at the LAN layer and forwarding packets based on the LAN address
are called switches (or, in earlier days, bridges), while such devices acting at the IP layer and forwarding
on the IP address are called routers. Datagram forwarding is used both by Ethernet switches and by IP
routers, though the destinations in Ethernet forwarding tables are individual nodes while the destinations in
IP routers are entire networks (that is, sets of nodes).

10

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

In IP routers within end-user sites it is common for a forwarding table to include a catchall default entry,
matching any IP address that is nonlocal and so needs to be routed out into the Internet at large. Unlike the
consolidated entries for B, D and E in the table above for S1, which likely would have to be implemented as
actual separate entries, a default entry is a single record representing where to forward the packet if no other
destination match is found. Here is a forwarding table for S1, above, with a default entry replacing the last
three entries:
S1
destination
A
C
default

next_hop
0
1
2

Default entries make sense only when we can tell by looking at an address that it does not represent a
nearby node. This is common in IP networks because an IP address encodes the destination network, and
routers generally know all the local networks. It is however rare in Ethernets, because there is generally
no correlation between Ethernet addresses and locality. If S1 above were an Ethernet switch, and it had
some means of knowing that interfaces 0 and 1 connected directly to individual hosts, not switches and S1
knew the addresses of these hosts then making interface 2 a default route would make sense. In practice,
however, Ethernet switches do not know what kind of device connects to a given interface.

1.5 Topology
In the network diagrammed in the previous section, there are no loops; graph theorists might describe this
by saying the network graph is acyclic, or is a tree. In a loop-free network there is a unique path between
any pair of nodes. The forwarding-table algorithm has only to make sure that every destination appears in
the forwarding tables; the issue of choosing between alternative paths does not arise.
However, if there are no loops then there is no redundancy: any broken link will result in partitioning the
network into two pieces that cannot communicate. All else being equal (which it is not, but never mind for
now), redundancy is a good thing. However, once we start including redundancy, we have to make decisions
among the multiple paths to a destination. Consider, for a moment, the following network:

Should S1 list S2 or S3 as the next_hop to B? Both paths A S1 S2 S4 B and A S1 S3 S4 B get there.

There is no right answer. Even if one path is faster than the other, taking the slower path is not exactly
wrong (especially if the slower path is, say, less expensive). Some sort of protocol must exist to provide a
mechanism by which S1 can make the choice (though this mechanism might be as simple as choosing to
route via the first path discovered to the given destination). We also want protocols to make sure that, if S1
reaches B via S2 and the S2 S4 link fails, then S1 will switch over to the still-working S1 S3 S4 B route.
As we shall see, many LANs (in particular Ethernet) prefer tree networks with no redundancy, while IP
has complex protocols in support of redundancy.

1.5 Topology

11

An Introduction to Computer Networks, Release 1.0

1.6 Routing Loops

A potential drawback to datagram forwarding is the possibility of a routing loop: a set of entries in the
forwarding tables that cause some packets to circulate endlessly. For example, in the previous picture we
would have a routing loop if, for (nonexistent) destination C, S1 forwarded to S2, S2 forwarded to S4,
S4 forwarded to S3, and S3 forwarded to S1. A packet sent to C would not only not be delivered, but
in circling endlessly it might easily consume a large majority of the bandwidth. Routing loops typically
arise because the creation of the forwarding tables is often distributed, and there is no global authority to
detect inconsistencies. Even when there is such an authority, temporary routing loops can be created due to
notification delays.
Routing loops can also occur in networks where the underlying link topology is loop-free; for example, in
the previous diagram we could, again for destination C, have S1 forward to S2 and S2 forward back to S1.
We will refer to such a case as a linear routing loop.
All datagram-forwarding protocols need some way of detecting and avoiding routing loops. Ethernet, for
example, avoids nonlinear routing loops by disallowing loops in the underlying network topology, and avoids
linear routing loops by not having switches forward a packet back out the interface by which it arrived. IP
provides for a one-byte Time to Live (TTL) field in the IP header; it is set by the sender and decremented
by 1 at each router; a packet is discarded if its TTL reaches 0. This limits the number of times a wayward
packet can be forwarded to the initial TTL value, typically 64.
In datagram routing, a switch is responsible only for the next hop to the ultimate destination; if a switch
has a complete path in mind, there is no guarantee that the next_hop switch or any other downstream switch
to agree to forward along that path. Misunderstandings can potentially lead to routing loops. Consider this
network:

D might feel that the best path to B is DECB (perhaps because it believes the AD link is to be avoided).
If E similarly decides the best path to B is EDAB, and if D and E both choose their next_hop for B
based on these best paths, then a linear routing loop is formed: D routes to B via E and E routes to B via D.
Although each of D and E have identified a usable path, that path is not in fact followed. Moral: successful
datagram routing requires cooperation and a consistent view of the network.

1.7 Congestion
Switches introduce the possibility of congestion: packets arriving faster than they can be sent out. This can
happen with just two interfaces, if the inbound interface has a higher bandwidth than the outbound interface;
another common source of congestion is traffic arriving on multiple inputs and all destined for the same
output.

12

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

Whatever the reason, if packets are arriving for a given outbound interface faster than they can be sent, a
queue will form for that interface. Once that queue is full, packets will be dropped. The most common
strategy (though not the only one) is to drop any packets that arrive when the queue is full.
The term congestion may refer either to the point where the queue is just beginning to build up, or to the
point where the queue is full and packets are lost. In their paper [CJ89], Chiu and Jain refer to the first point
as the knee; this is where the slope of the load v throughput graph flattens. They refer to the second point as
the cliff; this is where packet losses may lead to a precipitous decline in throughput. Other authors use the
term contention for knee-congestion.
In the Internet, most packet losses are due to congestion. This is not because congestion is especially bad
(though it can be, at times), but rather that other types of losses (eg due to packet corruption) are insignificant
by comparison.
When to Upgrade?
Deciding when a network really does have insufficient bandwidth is not a technical issue but an economic one. The number of customers may increase, the cost of bandwidth may decrease or customers
may simply be willing to pay more to have data transfers complete in less time; customers here can
be external or in-house. Monitoring of links and routers for congestion can, however, help determine
exactly what parts of the network would most benefit from upgrade.
We emphasize that the presence of congestion does not mean that a network has a shortage of bandwidth.
Bulk-traffic senders (though not real-time senders) attempt to send as fast as possible, and congestion is
simply the networks feedback that the maximum transmission rate has been reached.
Congestion is a sign of a problem in real-time networks, which we will consider in 19 Quality of Service.
In these networks losses due to congestion must generally be kept to an absolute minimum; one way to
achieve this is to limit the acceptance of new connections unless sufficient resources are available.

1.8 Packets Again

Perhaps the core justification for packets, Barans concerns about node failure notwithstanding, is that the
same link can carry, at different times, different packets representing traffic to different destinations and from
different senders. Thus, packets are the key to supporting shared transmission lines; that is, they support
the multiplexing of multiple communications channels over a single cable. The alternative of a separate
physical line between every pair of machines grows prohibitively complex very quickly (though virtual
circuits between every pair of machines in a datacenter are not uncommon; see 3.7 Virtual Circuits).
From this shared-medium perspective, an important packet feature is the maximum packet size, as this represents the maximum time a sender can send before other senders get a chance. The alternative of unbounded
packet sizes would lead to prolonged network unavailability for everyone else if someone downloaded a
large file in a single 1 Gigabit packet. Another drawback to large packets is that, if the packet is corrupted,
the entire packet must be retransmitted; see 5.3.1 Error Rates and Packet Size.
When a router or switch receives a packet, it (generally) reads in the entire packet before looking at the
header to decide to what next node to forward it. This is known as store-and-forward, and introduces
a forwarding delay equal to the time needed to read in the entire packet. For individual packets this

1.8 Packets Again

13

An Introduction to Computer Networks, Release 1.0

forwarding delay is hard to avoid (though some switches do implement cut-through switching to begin
forwarding a packet before it has fully arrived), but if one is sending a long train of packets then by keeping
multiple packets en route at the same time one can essentially eliminate the significance of the forwarding
delay; see 5.3 Packet Size.
Total packet delay from sender to receiver is the sum of the following:
Bandwidth delay, ie sending 1000 Bytes at 20 Bytes/millisecond will take 50 ms. This is a per-link
delay.
Propagation delay due to the speed of light. For example, if you start sending a packet right now
on a 5000-km cable across the US with a propagation speed of 200 m/sec (= 200 km/ms, about 2/3
the speed of light in vacuum), the first bit will not arrive at the destination until 25 ms later. The
bandwidth delay then determines how much after that the entire packet will take to arrive.
Store-and-forward delay, equal to the sum of the bandwidth delays out of each router along the path
Queuing delay, or waiting in line at busy routers. At bad moments this can exceed 1 sec, though that
is rare. Generally it is less than 10 ms and often is less than 1 ms. Queuing delay is the only delay
component amenable to reduction through careful engineering.
See 5.1 Packet Delay for more details.

1.9 LANs and Ethernet

A local-area network, or LAN, is a system consisting of
physical links that are, ultimately, serial lines
common interfacing hardware connecting the hosts to the links
protocols to make everything work together
We will explicitly assume that every LAN node is able to communicate with every other LAN node. Sometimes this will require the cooperation of intermediate nodes acting as switches.
Far and away the most common type of LAN is Ethernet, originally described in a 1976 paper by Metcalfe
and Boggs [MB76]. Ethernets popularity is due to low cost more than anything else, though the primary
reason Ethernet cost is low is that high demand has led to manufacturing economies of scale.
The original Ethernet had a bandwidth of 10 Mbps (megabits per second; we will use lower-case b for
bits and upper-case B for bytes), though nowadays most Ethernet operates at 100 Mbps and gigabit (1000
Mbps) Ethernet (and faster) is widely used in server rooms. (By comparison, as of this writing the data transfer rate to a typical faster hard disk is about 1000 Mbps.) Wireless (wi-fi) LANs are gaining popularity,
and in some settings have supplanted wired Ethernet to end-users.
Many early Ethernet installations were unswitched; each host simply tapped in to one long primary cable
that wound through the building (or floor). In principle, two stations could then transmit at the same time,
rendering the data unintelligible; this was called a collision. Ethernet has several design features intended
to minimize the bandwidth wasted on collisions: stations, before transmitting, check to be sure the line is
idle, they monitor the line while transmitting to detect collisions during the transmission, and, if a collision
is detected, they execute a random backoff strategy to avoid an immediate recollision. See 2.1.3 The Slot

14

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

Time and Collisions. While Ethernet collisions definitely reduce throughput, in the larger view they should
perhaps be thought of as a part of a remarkably inexpensive shared-access mediation protocol.
In unswitched Ethernets every packet is received by every host and it is up to the network card in each host
to determine if the arriving packet is addressed to that host. It is almost always possible to configure the
card to forward all arriving packets to the attached host; this poses a security threat and password sniffers
that surreptitiously collected passwords via such eavesdropping used to be common.
Password Sniffing
In the fall of 1994 at Loyola University I remotely changed the root password on several CS-department
unix machines at the other end of campus, using telnet. I told no one. Within two hours, someone else
logged into one of these machines, using the new password, from a host in Europe. Password sniffing
was the likely culprit.
Two months later was the so-called Christmas Day Attack (12.9.1 ISNs and spoofing). One of the
hosts used to launch this attack was Loyolas hacked [Link]. It is unclear the degree to which
password sniffing played a role in that exploit.
Due to both privacy and efficiency concerns, almost all Ethernets today are fully switched; this ensures that
each packet is delivered only to the host to which it is addressed. One advantage of switching is that it
effectively eliminates most Ethernet collisions; while in principle it replaces them with a queuing issue, in
practice Ethernet switch queues so seldom fill up that they are almost invisible even to network managers
(unlike IP router queues). Switching also prevents host-based eavesdropping, though arguably a better
solution to this problem is encryption. Perhaps the more significant tradeoff with switches, historically, was
that Once Upon A Time they were expensive and unreliable; tapping directly into a common cable was dirt
cheap.
Ethernet addresses are six bytes long. Each Ethernet card (or network interface) is assigned a (supposedly)
unique address at the time of manufacture; this address is burned into the cards ROM and is called the cards
physical address or hardware address or MAC (Media Access Control) address. The first three bytes of
the physical address have been assigned to the manufacturer; the subsequent three bytes are a serial number
assigned by that manufacturer.
By comparison, IP addresses are assigned administratively by the local site. The basic advantage of having
addresses in hardware is that hosts automatically know their own addresses on startup; no manual configuration or server query is necessary. It is not unusual for a site to have a large number of identically configured
workstations, for which all network differences derive ultimately from each workstations unique Ethernet
address.
The network interface continually monitors all arriving packets; if it sees any packet containing a destination
address that matches its own physical address, it grabs the packet and forwards it to the attached CPU (via a
CPU interrupt).
Ethernet also has a designated broadcast address. A host sending to the broadcast address has its packet
received by every other host on the network; if a switch receives a broadcast packet on one port, it forwards
the packet out every other port. This broadcast mechanism allows host A to contact host B when A does
not yet know Bs physical address; typical broadcast queries have forms such as Will the designated server
please answer or (from the ARP protocol) will the host with the given IP address please tell me your
physical address.

1.9 LANs and Ethernet

15

An Introduction to Computer Networks, Release 1.0

Traffic addressed to a particular host that is, not broadcast is said to be unicast.
Because Ethernet addresses are assigned by the hardware, knowing an address does not provide any direct
indication of where that address is located on the network. In switched Ethernet, the switches must thus have
a forwarding-table record for each individual Ethernet address on the network; for extremely large networks
this ultimately becomes unwieldy. Consider the analogous situation with postal addresses: Ethernet is
somewhat like attempting to deliver mail using social-security numbers as addresses, where each postal
worker is provided with a large catalog listing each persons SSN together with their physical location. Real
postal mail is, of course, addressed hierarchically using ever-more-precise specifiers: state, city, zipcode,
street address, and name / room#. Ethernet, in other words, does not scale well to large sizes.
Switched Ethernet works quite well, however, for networks with up to 10,000-100,000 nodes. Forwarding
tables with size in that range are straightforward to manage.
To forward packets correctly, switches must know where all active destination addresses in the LAN are
located; Ethernet switches do this by a passive learning algorithm. (IP routers, by comparison, use active
protocols.) Typically a host physical address is entered into a switchs forwarding table when a packet from
that host is first received; the switch notes the packets arrival interface and source address and assumes
that the same interface is to be used to deliver packets back to that sender. If a given destination address
has not yet been seen, and thus is not in the forwarding table, Ethernet switches still have the backup
delivery option of forwarding to everyone, by treating the destination address like the broadcast address,
and allowing the host Ethernet cards to sort it out. Since this broadcast-like process is not generally used
for more than one packet (after that, the switches will have learned the correct forwarding-table entries), the
risk of eavesdropping is minimal.
The host,interface forwarding table is often easier to think of as host,next_hop, where the next_hop node
is whatever switch or host is at the immediate other end of the link connecting to the given interface. In a
fully switched network where each link connects only two interfaces, the two perspectives are equivalent.

1.10 IP - Internet Protocol

To solve the scaling problem with Ethernet, and to allow support for other types of LANs and point-to-point
links as well, the Internet Protocol was developed. Perhaps the central issue in the design of IP was to
support universal connectivity (everyone can connect to everyone else) in such a way as to allow scaling to
enormous size (in 2013 there appear to be around ~109 nodes, although IP should work to 1010 nodes or
more), without resulting in unmanageably large forwarding tables (currently the largest tables have about
300,000 entries.)
In the early days, IP networks were considered to be internetworks of basic networks (LANs); nowadays
users generally ignore LANs and think of the Internet as one large (virtual) network.
To support universal connectivity, IP provides a global mechanism for addressing and routing, so that
packets can actually be delivered from any host to any other host. IP addresses (for the most-common
version 4, which we denote IPv4) are 4 bytes (32 bits), and are part of the IP header that generally follows
the Ethernet header. The Ethernet header only stays with a packet for one hop; the IP header stays with the
packet for its entire journey across the Internet.
An essential feature of IPv4 (and IPv6) addresses is that they can be divided into a network part (a prefix)
and a host part (the remainder). The legacy mechanism for designating the IPv4 network and host
address portions was to make the division according to the first few bits:
16

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

first few bits

0
10
110

first byte
0-127
128-191
192-223

network bits
8
16
24

host bits
24
16
8

name
class A
class B
class C

application
a few very large networks
institution-sized networks
sized for smaller entities

For example, the original IP address allocation for Loyola University Chicago was [Link], a class B.
In binary, 147 is 10010011. The network/host division point is not carried within the IP header; in fact,
nowadays the division into network and host is dynamic, and can be made at different positions in the
address at different levels of the network.
IP addresses, unlike Ethernet addresses, are administratively assigned. Once upon a time, you would get
your Class B network prefix from the Internet Assigned Numbers Authority, or IANA (they now delegate
this task), and then you would in turn assign the host portion in a way that was appropriate for your local
site. As a result of this administrative assignment, an IP address usually serves not just as an endpoint
identifier but also as a locator, containing embedded location information.
The Class A/B/C definition above was spelled out in 1981 in RFC 791, which introduced IP. Class D was
added in 1986 by RFC 988; class D addresses must begin with the bits 1110. These addresses are for
multicast, that is, sending an IP packet to every member of a set of recipients (ideally without actually
transmitting it more than once on any one link).
The network portion of an IP address is sometimes called the network number or network address or
network prefix; as we shall see below, most forwarding decisions are made using only the network portion.
It is commonly denoted by setting the host bits to zero and ending the resultant address with a slash followed
by the number of network bits in the address: eg [Link]/8 or [Link]/16. Note that [Link]/8 and
[Link]/9 represent different things; in the latter, the second byte of any host address extending the network
address is constrained to begin with a 0-bit. An anonymous block of IP addresses might be referred to only
by the slash and following digit, eg we need a /22 block to accommodate all our customers.
All hosts with the same network address (same network bits) must be located together on the same LAN; as
we shall see below, if two hosts share the same network address then they will assume they can reach each
other directly via the underlying LAN, and if they cannot then connectivity fails. A consequence of this rule
is that outside of the site only the network bits need to be looked at to route a packet to the site.
Each individual LAN technology has a maximum packet size it supports; for example, Ethernet has a maximum packet size of about 1500 bytes but the once-competing Token Ring had a maximum of 4 KB. Today
the world has largely standardized on Ethernet and almost entirely standardized on Ethernet packet-size limits, but this was not the case when IP was introduced and there was real concern that two hosts on separate
large-packet networks might try to exchange packets too large for some small-packet intermediate network
to carry.
Therefore, in addition to routing and addressing, the decision was made that IP must also support fragmentation: the division of large packets into multiple smaller ones (in other contexts this may also be called
segmentation). The IP approach is not very efficient, and IP hosts go to considerable lengths to avoid fragmentation. IP does require that packets of up to 576 bytes be supported, and so a common legacy strategy
was for a host to limit a packet to at most 512 user-data bytes whenever the packet was to be sent via a
router; packets addressed to another host on the same LAN could of course use a larger packet size. Despite
its limited use, however, fragmentation is essential conceptually, in order for IP to be able to support large
packets without knowing anything about the intervening networks.
IP is a best effort system; there are no IP-layer acknowledgments or retransmissions. We ship the packet
off, and hope it gets there. Most of the time, it does.
1.10 IP - Internet Protocol

17

An Introduction to Computer Networks, Release 1.0

Architecturally, this best-effort model represents what is known as connectionless networking: the IP layer
does not maintain information about endpoint-to-endpoint connections, and simply forwards packets like a
giant LAN. Responsibility for creating and maintaining connections is left for the next layer up, the TCP
layer. Connectionless networking is not the only way to do things: the alternative could have been some
form connection-oriented internetworking, in which routers do maintain state information about individual
connections. Later, in 3.7 Virtual Circuits, we will examine how virtual-circuit networking can be used to
implement a connection-oriented approach; virtual-circuit switching is the primary alternative to datagram
switching.
Connectionless (IP-style) and connection-oriented networking each have advantages. Connectionless networking is conceptually more reliable: if routers do not hold connection state, then they cannot lose connection state. The path taken by the packets in some higher-level connection can easily be dynamically rerouted.
Finally, connectionless networking makes it hard for providers to bill by the connection; once upon a time
(in the era of dollar-a-minute phone calls) this was a source of mild astonishment to many new users. The
primary advantage of connection-oriented networking, however, is that the routers are then much better positioned to accept reservations and to make quality-of-service guarantees. This remains something of a
sore point in the current Internet: if you want to use Voice-over-IP, or VoIP, telephones, or if you want to
engage in video conferencing, your packets will be treated by the Internet core just the same as if they were
low-priority file transfers. There is no priority service option.
Perhaps the most common form of IP packet loss is router queue overflows, representing network congestion.
Packet losses due to packet corruption are rare (eg less than one in 104 ; perhaps much less). But in a
connectionless world a large number of hosts can simultaneously decide to send traffic through one router,
in which case queue overflows are hard to avoid.

1.10.1 IP Forwarding
IP routers use datagram forwarding, described in 1.4 Datagram Forwarding above, to deliver packets, but
the destination values listed in the forwarding tables are network prefixes representing entire LANs
instead of individual hosts. The goal of IP forwarding, then, becomes delivery to the correct LAN; a separate
process is used to deliver to the final host once the final LAN has been reached.
The entire point, in fact, of having a network/host division within IP addresses is so that routers need to list
only the network prefixes of the destination addresses in their IP forwarding tables. This strategy is the key
to IP scalability: it saves large amounts of forwarding-table space, it saves time as smaller tables allow faster
lookup, and it saves the bandwidth that would be needed for routers to keep track of individual addresses.
To get an idea of the forwarding-table space savings, there are currently (2013) around a billion hosts on the
Internet, but only 300,000 or so networks listed in top-level forwarding tables. When network prefixes are
used as forwarding-table destinations, matching an actual packet address to a forwarding-table entry is no
longer a matter of simple equality comparison; routers must compare appropriate prefixes.
IP forwarding tables are sometimes also referred to as routing tables; in this book, however, we make
at least a token effort to use forwarding to refer to the packet forwarding process, and routing to refer
to mechanisms by which the forwarding tables are maintained and updated. (If we were to be completely
consistent here, we would use the term forwarding loop rather than routing loop.)
Now let us look at a simple example of how IP forwarding (or routing) works. We will assume that all
network nodes are either hosts user machines, with a single network connection or routers, which do
packet-forwarding only. Routers are not directly visible to users, and always have at least two different

18

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

network interfaces representing different networks that the router is connecting. (Machines can be both
hosts and routers, but this introduces complications.)
Suppose A is the sending host, sending a packet to a destination host D. The IP header of the packet will
contain Ds IP address in the destination address field (it will also contain As own address as the source
address). The first step is for A to determine whether D is on the same LAN as itself or not; that is, whether
D is local. This is done by looking at the network part of the destination address, which we will denote by
Dnet . If this net address is the same as As (that is, if it is equal numerically to Anet ), then A figures D is on
the same LAN as itself, and can use direct LAN delivery. It looks up the appropriate physical address for D
(probably with the ARP protocol, 7.7 Address Resolution Protocol: ARP), attaches a LAN header to the
packet in front of the IP header, and sends the packet straight to D via the LAN.
If, however, Anet and Dnet do not match D is non-local then A looks up a router to use. Most ordinary
hosts use only one router for all non-local packet deliveries, making this choice very simple. A then forwards
the packet to the router, again using direct delivery over the LAN. The IP destination address in the packet
remains D in this case, although the LAN destination address will be that of the router.
When the router receives the packet, it strips off the LAN header but leaves the IP header with the IP
destination address. It extracts the destination D, and then looks at Dnet . The router first checks to see
if any of its network interfaces are on the same LAN as D; recall that the router connects to at least one
additional network besides the one for A. If the answer is yes, then the router uses direct LAN delivery to the
destination, as above. If, on the other hand, Dnet is not a LAN to which the router is connected directly, then
the router consults its internal forwarding table. This consists of a list of networks each with an associated
next_hop address. These net,next_hop tables compare with switched-Ethernets host,next_hop tables;
the former type will be smaller because there are many fewer nets than hosts. The next_hop addresses in the
table are chosen so that the router can always reach them via direct LAN delivery via one of its interfaces;
generally they are other routers. The router looks up Dnet in the table, finds the next_hop address, and uses
direct LAN delivery to get the packet to that next_hop machine. The packets IP header remains essentially
unchanged, although the router most likely attaches an entirely new LAN header.
The packet continues being forwarded like this, from router to router, until it finally arrives at a router that
is connected to Dnet ; it is then delivered by that final router directly to D, using the LAN.
To make this concrete, consider the following diagram:

With Ethernet-style forwarding, R2 would have to maintain entries for each of A,B,C,D,E,F. With IP forwarding, R2 has just two entries to maintain in its forwarding table: 200.0.0/24 and 200.0.1/24. If A sends
to D, at [Link], it puts this address into the IP header, notes that 200.0.0 = 200.0.1, and thus concludes
D is not a local delivery. A therefore sends the packet to its router R1, using LAN delivery. R1 looks up the
destination network 200.0.1 in its forwarding table and forwards the packet to R2, which in turn forwards it
to R3. R3 now sees that it is connected directly to the destination network 200.0.1, and delivers the packet
via the LAN to D, by looking up Ds physical address.
In this diagram, IP addresses for the ends of the R1R2 and R2R3 links are not shown. They could be
assigned global IP addresses, but they could also use private IP addresses. Assuming these links are

1.10 IP - Internet Protocol

19

An Introduction to Computer Networks, Release 1.0

point-to-point links, they might not actually need IP addresses at all; we return to this in 7.10 Unnumbered
Interfaces.
One can think of the network-prefix bits as analogous to the zip code on postal mail, and the host bits
as analogous to the street address. The internal parts of the post office get a letter to the right zip code,
and then an individual letter carrier gets it to the right address. Alternatively, one can think of the network
bits as like the area code of a phone number, and the host bits as like the rest of the digits. Newer protocols that support different net/host division points at different places in the network sometimes called
hierarchical routing allow support for addressing schemes that correspond to, say, zip/street/user, or
areacode/exchange/subscriber.
The Invertebrate Internet
Once upon a time, each leaf node connected through its provider to the backbone, and traffic between
any two nodes (or at least any two nodes not sharing a provider) passed through the backbone. The
backbone still carries a lot of traffic, but it is now also common for large providers such as Google
to connect (or peer) directly with large residential ISPs such as Comcast. See, for example, [Link].
We will refer to the Internet backbone as those IP routers that specialize in large-scale routing on the
commercial Internet, and which generally have forwarding-table entries covering all public IP addresses;
note that this is essentially a business definition rather than a technical one. We can revise the table-size
claim of the previous paragraph to state that, while there are many private IP networks, there are about
300,000 visible to the backbone. A forwarding table of 300,000 entries is quite feasible; a table a hundred
times larger is not, let alone a thousand times larger.
IP routers at non-backbone sites generally know all locally assigned network prefixes, eg 200.0.0/24 and
200.0.1/24 above. If a destination does not match any locally assigned network prefix, the packet needs
to be routed out into the Internet at large; for typical non-backbone sites this almost always this means
the packet is sent to the ISP that provides Internet connectivity. Generally the local routers will contain a
catchall default entry covering all nonlocal networks; this means that the router needs an explicit entry only
for locally assigned networks. This greatly reduces the forwarding-table size. The Internet backbone can be
approximately described, in fact, as those routers that do not have a default entry.
For most purposes, the Internet can be seen as a combination of end-user LANs together with point-to-point
links joining these LANs to the backbone, point-to-point links also tie the backbone together. Both LANs
and point-to-point links appear in the diagram above.
Just how routers build their destnet,next_hop forwarding tables is a major topic itself, which we cover in
9 Routing-Update Algorithms. Unlike Ethernet, IP routers do not have a broadcast delivery mechanism
as a fallback, so the tables must be constructed in advance. (There is a limited form of IP broadcast, but it
is basically intended for reaching the local LAN only, and does not help at all with delivery in the event that
the network is unknown.)
Most forwarding-table-construction algorithms used on a set of routers under common management fall into
either the distance-vector or the link-state category. In the distance-vector approach, often used at smaller
sites, routers exchange information with their immediately neighboring routers; tables are built up this
way through a sequence of such periodic exchanges. In the link-state approach, routers rapidly propagate
information about the state of each link; all routers in the organization receive this link-state information and
each one uses it to build and maintain a map of the entire network. The forwarding table is then constructed
20

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

(sometimes on demand) from this map.

Unrelated organizations exchange information through the Border Gateway Protocol, BGP (10 Large-Scale
IP Routing), which allows for a fusion of technical information (which sites are reachable at all, and through
where) with policy information representing legal or commercial agreements: which outside routers are
preferred, whose traffic we will carry even if it isnt to one of our customers, etc.
Internet 2 is a consortium of research sites with very-high-bandwidth internal interconnections. Before
Internet 2, Loyolas largest router table consisted of maybe a dozen internal routes, plus one default route
to the outside Internet. After Internet 2, the default route still pointed to the commercial Internet, but the
master forwarding table now had to have an entry for every Internet-2 site so this traffic would take the
Internet-2 path. See 7 IP version 4, exercise 5.

1.10.2 The Future of IPv4

As mentioned earlier, allocation of blocks of IP addresses is the responsibility of the Internet Assigned
Numbers Authority. IANA long ago delegated the job of allocating network prefixes to individual sites; they
limited themselves to handing out /8 blocks (class A blocks) to the five regional registries, which are
ARIN North America
RIPE Europe, the Middle East and parts of Asia
APNIC East Asia and the Pacific
AfriNIC most of Africa
LACNIC Central and South America
As of the end of January 2011, the IANA finally ran out of /8 blocks. There is a table at
[Link] of all IANA assignments of /8
blocks; examination of the table shows all have now been allocated.
A few months after the IANA pool ran out, Microsoft purchased 666,624 IP addresses (2604 Class-C blocks)
in a Nortel bankruptcy auction for $7.5 million. By a year later, IP-address prices appeared to have retreated
only slightly. It is possible that the market for IPv4 address blocks will continue to develop; alternatively,
this turn of events may accelerate implementation of IPv6, which has 128-bit addresses.
An IPv4 address price in the range of $10 is unlikely to have much impact in residential Internet access,
where annual connection fees are often $600. Large organizations already use NAT extensively, leading to
the need for only a small number of globally visible addresses. The IPv4 address shortage does not even
seem to have affected wireless networking. It does, however, lead to inefficient routing tables, as a site that
once had a single /20 address block and thus a single backbone forwarding-table entry might now be
spread over more than a hundred /27 blocks and concomitant forwarding entries.

1.11 DNS
IP addresses are hard to remember (nearly impossible in IPv6). The domain name system, or DNS, comes
to the rescue by creating a way to convert hierarchical text names to IP addresses. Thus, for example, one can
type [Link] instead of [Link]. Virtually all Internet software uses the same basic library
calls to convert DNS names to actual addresses.
1.11 DNS

21

An Introduction to Computer Networks, Release 1.0

One thing DNS makes possible is changing a websites IP address while leaving the name alone. This
allows moving a site to a new provider, for example, without requiring users to learn anything new. It
is also possible to have several different DNS names resolve to the same IP address, and through some
modest trickery have the http (web) server at that IP address handle the different DNS names as completely
different websites.
DNS is hierarchical and distributed; indeed, it is the classic example of a widely distributed database.
In looking up [Link] three different DNS servers may be queried: for [Link], for
[Link], and for .edu. Searching a hierarchy can be cumbersome, so DNS search results are normally
cached locally. If a name is not found in the cache, the lookup may take a couple seconds. The DNS
hierarchy need have nothing to do with the IP-address hierarchy.
Besides address lookups, DNS also supports a few other kinds of searches. The best known is probably
reverse DNS, which takes an IP address and returns a name. This is slightly complicated by the fact that
one IP address may be associated with multiple DNS names, so DNS must either return a list, or return one
name that has been designated the canonical name.

1.12 Transport
Think about what types of communications one might want over the Internet:
Interactive communications such as via ssh or telnet, with long idle times between short bursts
Bulk file transfers
Request/reply operations, eg to query a database or to make DNS requests
Real-time voice traffic, at (without compression) 8KB/sec, with constraints on the variation in delivery time (known as jitter; see 19.11.3 RTP Control Protocol for a specific numeric interpretation)
Real-time video traffic. Even with substantial compression, video generally requires much more
bandwidth than voice
While separate protocols might be used for each of these, the Internet has standardized on the Transmission
Control Protocol, or TCP, for the first three (though there are periodic calls for a new protocol addressing
the third item above), and TCP is sometimes pressed into service for the last two. TCP is thus the most
common transport layer for application data.
The IP layer is not well-suited to transport. IP routing is a best-effort mechanism, which means packets
can and do get lost sometimes. Data that does arrive can arrive out of order. The sender has to manage
division into packets; that is, buffering. Finally, IP only supports sending to a specific host; normally, one
wants to send to a given application running on that host. Email and web traffic, or two different web
sessions, should not be commingled!
TCP extends IP with the following features:
reliability: TCP numbers each packet, and keeps track of which are lost and retransmits them after a
timeout, and holds early-arriving out-of-order packets for delivery at the correct time. Every arriving
data packet is acknowledged by the receiver; timeout and retransmission occurs when an acknowledgment isnt received by the sender within a given time.

22

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

connection-orientation: Once a TCP connection is made, an application sends data simply by writing
to that connection. No further application-level addressing is needed.
stream-orientation: The application can write 1 byte at a time, or 100KB at a time; TCP will buffer
and/or divide up the data into appropriate sized packets.
port numbers: these provide a way to specify the receiving application for the data, and also to
identify the sending application.
throughput management: TCP attempts to maximize throughput, while at the same time not contributing unnecessarily to network congestion.
TCP endpoints are of the form host,port; these pairs are known as socket addresses, or sometimes as just
sockets though the latter refers more properly to the operating-system objects that receive the data sent to
the socket addresses. Servers (or, more precisely, server applications) listen for connections to sockets they
have opened; the client is then any endpoint that initiates a connection to a server.
When you enter a host name in a web browser, it opens a TCP connection to the servers port 80 (the standard
web-traffic port), that is, to the server socket with socket-address server,80. If you have several browser
tabs open, each might connect to the same server socket, but the connections are distinguishable by virtue
of using separate ports (and thus having separate socket addresses) on the client end (that is, your end).
A busy server may have thousands of connections to its port 80 (the web port) and hundreds of connections
to port 25 (the email port). Web and email traffic are kept separate by virtue of the different ports used. All
those clients to the same port, though, are kept separate because each comes from a unique host,port pair. A
TCP connection is determined by the host,port socket address at each end; traffic on different connections
does not intermingle. That is, there may be multiple independent connections to [Link],80. This is
somewhat analogous to certain business telephone numbers of the operators are standing by type, which
support multiple callers at the same time to the same number. Each call is answered by a different operator
(corresponding to a different cpu process), and different calls do not overhear each other.
TCP uses the sliding-windows algorithm, 6 Abstract Sliding Windows, to keep multiple packets en route
at any one time. The window size represents the number of packets simultaneously en route; if the window
size is 10, for example, then at any one time 10 packets are out there (perhaps 5 data packets and 5 returning
acknowledgments). As each acknowledgment arrives, the window slides forward and the data packet 10
packets ahead is sent. For example, consider the moment when the ten packets 20-29 are in transit. When
ACK[20] is received, Data[30] is sent, and so now packets 21-30 are in transit. When ACK[21] is received,
Data[31] is sent, so packets 22-31 are in transit.
Sliding windows minimizes the effect of store-and-forward delays, and propagation delays, as these then
only count once for the entire windowful and not once per packet. Sliding windows also provides an automatic, if partial, brake on congestion: the queue at any switch or router along the way cannot exceed the
window size. In this it compares favorably with constant-rate transmission, which, if the available bandwidth falls below the transmission rate, always leads to a significant percentage of dropped packets. Of
course, if the window size is too large, a sliding-windows sender may also experience dropped packets.
The ideal window size, at least from a throughput perspective, is such that it takes one round-trip time to send
an entire window, so that the next ACK will always be arriving just as the sender has finished transmitting the
window. Determining this ideal size, however, is difficult; for one thing, the ideal size varies with network
load. As a result, TCP approximates the ideal size. The most common TCP strategy that of so-called TCP
Reno is that the window size is slowly raised until packet loss occurs, which TCP takes as a sign that it
has reached the limit of available network resources. At that point the window size is reduced to half its
1.12 Transport

23

An Introduction to Computer Networks, Release 1.0

previous value, and the slow climb resumes. The effect is a sawtooth graph of window size with time,
which oscillates (more or less) around the optimal window size. For an idealized sawtooth graph, see
13.1.1 The Steady State; for some real (simulation-created) sawtooth graphs see 16.4.1 Some TCP Reno
cwnd graphs.
While this window-size-optimization strategy has its roots in attempting to maximize the available bandwidth, it also has the effect of greatly limiting the number of packet-loss events. As a result, TCP has come
to be the Internet protocol charged with reducing (or at least managing) congestion on the Internet, and
relatedly with ensuring fairness of bandwidth allocations to competing connections. Core Internet routers
at least in the classical case essentially have no role in enforcing congestion or fairness restrictions at all.
The Internet, in other words, places responsibility for congestion avoidance cooperatively into the hands of
end users. While cheating is possible, this cooperative approach has worked remarkably well.
While TCP is ubiquitous, the real-time performance of TCP is not always consistent: if a packet is lost,
the receiving TCP host will not turn over anything further to the receiving application until the lost packet
has been retransmitted successfully; this is often called head-of-line blocking. This is a serious problem
for sound and video applications, which can discretely handle modest losses but which have much more
difficulty with sudden large delays. A few lost packets ideally should mean just a few brief voice dropouts
(pretty common on cell phones) or flicker/snow on the video screen (or just reuse of the previous frame);
both of these are better than pausing completely.
The basic alternative to TCP is known as UDP, for User Datagram Protocol. UDP, like TCP, provides port
numbers to support delivery to multiple endpoints within the receiving host, in effect to a specific process on
the host. As with TCP, a UDP socket consists of a host,port pair. UDP also includes, like TCP, a checksum
over the data. However, UDP omits the other TCP features: there is no connection setup, no lost-packet
detection, no automatic timeout/retransmission, and the application must manage its own packetization.
The Real-time Transport Protocol, or RTP, sits above UDP and adds some additional support for voice and
video applications.

1.13 Firewalls
One problem with having a program on your machine listening on an open TCP port is that someone may
connect and then, using some flaw in the software on your end, do something malicious to your machine.
Damage can range from the unintended downloading of personal data to compromise and takeover of your
entire machine, making it a distributor of viruses and worms or a steppingstone in later break-ins of other
machines.
A strategy known as buffer overflow has been the basis for a great many total-compromise attacks. The idea
is to identify a point in a server program where it fills a memory buffer with network-supplied data without
careful length checking; almost any call to the C library function gets(buf) will suffice. The attacker
then crafts an oversized input string which, when read by the server and stored in memory, overflows the
buffer and overwrites subsequent portions of memory, typically containing the stack-frame pointers. The
usual goal is to arrange things so that when the server reaches the end of the currently executing function,
control is returned not to the calling function but instead to the attackers own payload code located within
the string.
A firewall is a program to block connections deemed potentially risky, eg those originating from outside
the site. Generally ordinary workstations do not ever need to accept connections from the Internet; client

24

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

machines instead initiate connections to (better-protected) servers. So blocking incoming connections works
pretty well; when necessary (eg for games) certain ports can be selectively unblocked.
The original firewalls were routers. Incoming traffic to servers was often blocked unless it was sent to one
of a modest number of open ports; for non-servers, typically all inbound connections were blocked. This
allowed internal machines to operate reasonably safely, though being unable to accept incoming connections
is sometimes inconvenient. Nowadays per-machine firewalls in addition to router-based firewalls are
common: you can configure your machine not to accept inbound connections to most (or all) ports regardless
of whether software on your machine requests such a connection. Outbound connections can, in many cases,
also be prevented.

1.14 Network Address Translation

A related way of providing firewall protection for ordinary machines is Network Address Translation, or
NAT. This technique also conserves IP addresses. Instead of assigning each host at a site a real IP address,
just one address is assigned to a border router. Internal machines get internal IP addresses, typically of the
form or 192.168.x.y or 10.x.y.z (addresses beginning with [Link]/16 or with [Link]/8 cannot be used
on the Internet). Inbound connections to the internal machines are banned, as with some firewalls. When
an internal machine wants to connect to the outside, the NAT router intercepts the connection, allocates a
new port, and forwards the connection on as if it came from that new port and its own IP address. The
remote machine responds, and the NAT router remembers the connection and forwards data to the correct
internal host, rewriting the address and port fields of the incoming packets. Done properly, NAT improves
the security of a site. It also allows multiple machines to share a single IP address, and so is popular with
home users who have a single broadband connection but who wish to use multiple computers. A typical
NAT router sold for residential or small-office use is commonly simply called a router, or (somewhat
more precisely) a residential gateway.
Suppose the NAT router is NR, and internal hosts A and B each connect from port 3000 to port 80 on
external hosts C and D, respectively. Here is what NRs NAT table might look like. No column for NRs IP
address is given, as it is fixed.
remote host
C
D

remote port
80
80

outside source port

3000
3000

inside host
A
B

inside port
3000
3000

A packet to C from A,3000 would be rewritten by NR so that the source was NR,3000. A packet from
C,80 addressed to NR,3000 would be rewritten and forwarded to A,3000. Similarly, a packet from
D,80 addressed to NR,3000 would be rewritten and forwarded to B,3000; the NAT table takes into
account the sending socket address as well as the destination.
Now suppose B opens a connection to C,80, also from inside port 3000. This time NR must remap the
port number, because that is the only way to distinguish between packets from C,80 to A and to B. The
new table is
remote host
C
D
C

remote port
80
80
80

outside source port

3000
3000
3001

1.14 Network Address Translation

inside host
A
B
B

inside port
3000
3000
3000

25

An Introduction to Computer Networks, Release 1.0

Typically NR would not create TCP connections between itself and C,80 and D,80; the NAT table does
forwarding but the endpoints of the connection are still at the inside hosts. However, NR might very well
monitor the TCP connections to know when they have closed, and so can be removed from the table.
It is common for Voice-over-IP (VoIP) telephony using the SIP protocol (RFC 3261) to prefer to use UDP
port 5060 at both ends. If a VoIP server is outside the NAT router (which must be the case as the server
must generally be publicly visible) and a telephone is inside, likely port 5060 will pass through without
remapping, though the telephone will have to initiate the connection. But if there are two phones inside, one
of them will appear to be connecting to the server from an alternative port.
VoIP systems run into a much more serious problem with NAT, however. A call ultimately between two
phones is typically first negotiated between the phones respective VoIP servers. Once the call is set up, the
servers would prefer to step out of the loop, and have the phones exchange voice packets directly. The SIP
protocol was designed to handle this by having each phone report to its respective server the UDP socket
(IP address,port pair) it intends to use for the voice exchange; the servers then report these phone sockets
to each other, and from there to the opposite phones. This socket information is rendered incorrect by NAT,
however, certainly the IP address and quite likely the port as well. If only one of the phones is behind a
NAT firewall, it can initiate the voice connection to the other phone, but the other phone will see the voice
packets arriving from a different socket than promised and will likely not recognize them as part of the call.
If both phones are behind NAT firewalls, they will not be able to connect to one another at all. The common
solution is for the VoIP server of a phone behind a NAT firewall to remain in the communications path,
forwarding packets to its hidden partner.
If a site wants to make it possible to allow connections to hosts behind a NAT router or other firewall, one
option is tunneling. This is the creation of a virtual LAN link that runs on top of a TCP connection
between the end user and one of the sites servers; the end user can thus appear to be on one of the organizations internal LANs; see 3.1 Virtual Private Network. Another option is to open up a specific port: in
essence, a static NAT-table entry is made connecting a specific port on the NAT router to a specific internal
host and port (usually the same port). For example, all UDP packets to port 5060 on the NAT router might
be forwarded to port 5060 on internal host A, even in the absence of any prior packet exchange.
NAT routers work very well when the communications model is of client-side TCP connections, originating
from the inside and with public outside servers as destination. The NAT model works less well for peerto-peer networking, where your computer and a friends, each behind a different NAT router, wish to
establish a connection. NAT routers also often have trouble with UDP protocols, due to the tendency for
such protocols to have the public server reply from a different port than the one originally contacted. For
example, if host A behind a NAT router attempts to use TFTP (11.3 Trivial File Transport Protocol, TFTP),
and sends a packet to port 69 of public server C, then C is likely to reply from some new port, say 3000,
and this reply is likely to be dropped by the NAT router as there will be no entry there yet for traffic from
C,3000.

1.15 IETF and OSI

The Internet protocols discussed above are defined by the Internet Engineering Task Force, or IETF (under
the aegis of the Internet Architecture Board, or IAB, in turn under the aegis of the Internet Society,
ISOC). The IETF publishes Request For Comment or RFC documents that contain all the formal Internet
standards; these are available at [Link] (note that, by the time a document appears
here, the actual comment-requesting period is generally long since closed). The five-layer model is closely

26

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

associated with the IETF, though is not an official standard.

RFC standards sometimes allow modest flexibility. With this in mind, RFC 2119 declares official understandings for the words MUST and SHOULD. A feature labeled with MUST is an absolute requirement
for the specification, while the term SHOULD is used when
there may exist valid reasons in particular circumstances to ignore a particular item, but the
full implications must be understood and carefully weighed before choosing a different course.
The original ARPANET network was developed by the US governments Defense Advanced Research
Projects Agency, or DARPA; it went online in 1969. The National Science Foundation began NSFNet in
1986; this largely replaced ARPANET. In 1991, operation of the NSFNet backbone was turned over to
ANSNet, a private corporation. The ISOC was founded in 1992 as the NSF continued to retreat from the
networking business.
Hallmarks of the IETF design approach were David Clarks declaration
We reject: kings, presidents and voting.
We believe in: rough consensus and running code.
and RFC Editor Jon Postels aphorism
Be liberal in what you accept, and conservative in what you send.
There is a persistent though false notion that the distributed-routing architecture of IP was due to a US
Department of Defense mandate that the original ARPAnet be able to survive a nuclear attack. In fact, the
developers of IP seemed unconcerned with this. However, Paul Baran did write, in his 1962 paper outlining
the concept of packet switching, that
If [the number of stations] is made sufficiently large, it can be shown that highly survivable
system structures can be built even in the thermonuclear era.
In 1977 the International Organization for Standardization, or ISO, founded the Open Systems Interconnection project, or OSI, a process for creation of new network standards. OSI represented an attempt at the
creation of networking standards independent of any individual government.
The OSI project is today perhaps best known for its seven-layer networking model: between Transport and
Application were inserted the Session and Presentation layers. The Session layer was to handle sessions
between applications (including the graceful closing of Transport-layer connections, something included
in TCP, and the re-establishment of broken Transport-layer connections, which TCP could sorely use),
and the Presentation layer was to handle things like defining universal data formats (eg for binary numeric
data, or for non-ASCII character sets), and eventually came to include compression and encryption as well.
However, most of the features at this level have ended up generally being implemented in the Application
layer of the IETF model with no inconvenience. Indeed, there are few examples in the TCP/IP world of data
format, compression or encryption being handled as a sublayer of the application, in which the sublayer
is responsible for connections to the Transport layer (SSL/TLS might be an example for encryption; applications read and write data directly to the SSL endpoint). Instead, applications usually read and write data
directly from/to the TCP layer, and then invoke libraries to handle things like data conversion, compression
and encryption.
OSI has its own version of IP and TCP. The IP equivalent is CLNP, the ConnectionLess Network Protocol,
although OSI also defines a connection-oriented protocol CMNS. The TCP equivalent is TP4; OSI also
defines TP0 through TP3 but those are for connection-oriented networks.

1.15 IETF and OSI

27

An Introduction to Computer Networks, Release 1.0

It seems clear that the primary reasons the OSI protocols failed in the marketplace were their ponderous
bureaucracy for protocol management, their principle that protocols be completed before implementation
began, and their insistence on rigid adherence to the specifications to the point of non-interoperability. In
contrast, the IETF had (and still has) a two working implementations rule for a protocol to become a
Draft Standard. From RFC 2026:
A specification from which at least two independent and interoperable implementations from different
code bases have been developed, and for which sufficient successful operational experience has been
obtained, may be elevated to the Draft Standard level. [emphasis added]
This rule has often facilitated the discovery of protocol design weaknesses early enough that the problems
could be fixed. The OSI approach is a striking failure for the waterfall design model, when competing
with the IETFs cyclic prototyping model. However, it is worth noting that the IETF has similarly been
unable to keep up with rapid changes in html, particularly at the browser end; the OSI mistakes were mostly
evident only in retrospect.
Trying to fit protocols into specific layers is often both futile and irrelevant. By one perspective, the RealTime Protocol RTP lives at the Transport layer, but just above the UDP layer; others have put RTP into the
Application layer. Parts of the RTP protocol resemble the Session and Presentation layers. A key component
of the IP protocol is the set of various router-update protocols; some of these freely use higher-level layers.
Similarly, tunneling might be considered to be a Link-layer protocol, but tunnels are often created and
maintained at the Application layer.
A sometimes-more-successful approach to understanding layers is to view them instead as parts of a
protocol graph. Thus, in the following diagram we have two protocols at the transport layer (UDP and
RTP), and one protocol (ARP) not easily assigned to a layer.

28

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

1.16 Berkeley Unix

Though not officially tied to the IETF, the Berkeley Unix releases became de facto reference implementations for most of the TCP/IP protocols. 4.1BSD (BSD for Berkeley Software Distribution) was released in
1981, 4.2BSD in 1983, 4.3BSD in 1986, 4.3BSD-Tahoe in 1988, 4.3BSD-Reno in 1990, and 4.4BSD in
1994. Descendants today include FreeBSD, OpenBSD and NetBSD. The TCP implementations TCP Tahoe
and TCP Reno (13 TCP Reno and Congestion Management) took their names from the corresponding
4.3BSD releases.

1.17 Epilog
This completes our tour of the basics. In the remaining chapters we will expand on the material here.

1.18 Exercises
1. Give forwarding tables for each of the switches S1-S4 in the following network with destinations A, B,
C, D. For the next_hop column, give the neighbor on the appropriate link rather than the interface number.
A

S1

S2

S3

S4

2. Give forwarding tables for each of the switches S1-S4 in the following network with destinations A, B,
C, D. Again, use the neighbor form of next_hop rather than the interface form. Try to keep the route to
each destination as short as possible. What decision has to be made in this exercise that did not arise in the
preceding exercise?
A

S1

S2

S4

S3

3. Consider the following arrangement of switches and destinations. Give forwarding tables (in neighbor
form) for S1-S4 that include default forwarding entries; the default entries should point toward S5. Eliminate all table entries that are implied by the default entry (that is, if the default entry is to S3, eliminate all
other entries for which the next hop is S3).
A

S1
D

S3

S2

S4

S5

1.16 Berkeley Unix

29

An Introduction to Computer Networks, Release 1.0

4. Four switches are arranged as below. The destinations are S1 through S4 themselves.
S1

S2

S4

S3

(a). Give the forwarding tables for S1 through S4 assuming packets to adjacent nodes are sent along the
connecting link, and packets to diagonally opposite nodes are sent clockwise.
(b). Give the forwarding tables for S1 through S4 assuming the S1S4 link is not used at all, not even for
S1S4 traffic.

5. Suppose we have switches S1 through S4; the forwarding-table destinations are the switches themselves. The table
S2: S1,S1 S3,S3 S4,S3
S3: S1,S2 S2,S2 S4,S4
From the above we can conclude that S2 must be directly connected to both S1 and S3 as its table lists them
as next_hops; similarly, S3 must be directly connected to S2 and S4.

(a). Must S1 and S4 be directly connected? If so, explain; if not, give a network in which there is no direct
link between them, consistent with the tables above.
(b). Now suppose S3s table is changed to the following. In this case must S1 and S4 be directly
connected? Why or why not?
S3: S1,S4 S2,S2 S4,S4
While the table for S4 is not given, you may assume that forwarding does work correctly. However, you
should not assume that paths are the shortest possible; in particular, you should not assume that each switch
will always reach its directly connected neighbors by using the direct connection.
6. (a) Suppose a network is as follows, with the only path from A to C passing through B:
...

...

Explain why a single routing loop cannot include both A and C.

(b). Suppose a routing loop follows the path A S1 S2 ... Sn A, where none of the Si are equal to
A. Show that all the Si must be distinct. (A corollary of this is that any routing loop created by datagramforwarding either involves forwarding back and forth between a pair of adjacent switches, or else involves
an actual graph cycle in the network topology; linear loops of length greater than 1 are impossible.)
7. Consider the following arrangement of switches:
S1

30

S4

S10

1 An Overview of Networks

An Introduction to Computer Networks, Release 1.0

S2

S5

S11

S3

S6

S12

Suppose S1-S6 have the forwarding tables below. For each destination A,B,C,D,E,F, suppose a packet is
sent to the destination from S1. Give the switches it passes through, including the initial switch S1, up until
the final switch S10-S12.
S1: (A,S4), (B,S2), (C,S4), (D,S2), (E,S2), (F,S4)
S2: (A,S5), (B,S5), (D,S5), (E,S3), (F,S3)
S3: (B,S6), (C,S2), (E,S6), (F,S6)
S4: (A,S10), (C,S5), (E,S10), (F,S5)
S5: (A,S6), (B,S11), (C,S6), (D,S6), (E,S4), (F,S2)
S6: (A,S3), (B,S12), (C,S12), (D,S12), (E,S5), (F,S12)
8. In the previous exercise, the routes taken by packets A-D are reasonably direct, but the routes for E and F
are rather circuitous.
Some routing applications assign weights to different links, and attempt to choose a path with the lowest
total link weight.

(a). Assign weights to the seven links S1S2, S2S3, S1S4, S2S5, S3S6, S4S5 and S5S6 so that
destination Es route in the previous exercise becomes the optimum (lowest total link weight) path.
(b). Assign (different!) weights to the seven links that make destination Fs route in the previous exercise
optimal.

Hint: you can do this by assigning a weight of 1 to all links except to one or two bad links; the bad links
get a weight of 10. In each of (a) and (b) above, the route taken will be the route that avoids all the bad
links. You must treat (a) entirely differently from (b); there is no assignment of weights that can account for
both routes.
9. Suppose we have the following three Class C IP networks, joined by routers R1R4. Give the forwarding
table for each router. For networks directly connected to a router (eg 200.0.1/24 and R1), include the network
in the table but list the next hop as direct.
R1

R4

R3

200.0.1/24

R2

200.0.2/24

200.0.3/24

1.18 Exercises

31

An Introduction to Computer Networks, Release 1.0

32

1 An Overview of Networks

2 ETHERNET

We now turn to a deeper analysis of the ubiquitous Ethernet LAN protocol. Current user-level Ethernet today
(2013) is usually 100 Mbps, with Gigabit Ethernet standard in server rooms and backbones, but because
Ethernet speed scales in odd ways, we will start with the 10 Mbps formulation. While the 10 Mbps speed is
obsolete, and while even the Ethernet collision mechanism is largely obsolete, collision management itself
continues to play a significant role in wireless networks.

2.1 10-Mbps classic Ethernet

The original Ethernet specification was the 1976 paper of Metcalfe and Boggs, [MB76]. The data rate was
10 megabits per second, and all connections were made with coaxial cable instead of todays twisted pair. In
its original form, an Ethernet was a broadcast bus, which meant that all packets were, at least at the physical
layer, broadcast onto the shared medium and could be seen, theoretically, by all other nodes. If two nodes
transmitted at the same time, there was a collision; proper handling of collisions was an important part of
the access-mediation strategy for the shared medium. Data was transmitted using Manchester encoding; see
4.1.3 Manchester.
The linear bus structure could be modified with repeaters (below), into an arbitrary tree structure, though
loops remain something of a problem even with todays Ethernet.
Whenever two stations transmitted at the same time, the signals would collide, and interfere with one another; both transmissions would fail as a result. In order to minimize collision loss, each station implemented
the following:
1. Before transmission, wait for the line to become quiet
2. While transmitting, continually monitor the line for signs that a collision has occurred; if a collision
happens, then cease transmitting
3. If a collision occurs, use a backoff-and-retransmit strategy
These properties can be summarized with the CSMA/CD acronym: Carrier Sense, Multiple Access, Collision Detect. (The term carrier sense was used by Metcalfe and Boggs as a synonym for signal sense;
there is no literal carrier frequency to be sensed.) It should be emphasized that collisions are a normal event
in Ethernet, well-handled by the mechanisms above.
IEEE 802 Network Standards
The IEEE network standards all begin with 802: 802.3 is Ethernet, 802.11 is Wi-Fi, 802.16 is WiMAX,
and there are many others. One sometimes encounters the claim that 802 represents the date of an early
meeting: February 1980. However, the IEEE has a continuous stream of standards (with occasional
gaps): 799: Handling and Disposal of Transformer PCBs, 800: D-C Aircraft Rotating Machines, 803:
Recommended Practice for Unique Identification in Power Plants, etc.

33

An Introduction to Computer Networks, Release 1.0

Classic Ethernet came in version 1 [1980, DEC-Intel-Xerox], version 2 [1982, DIX], and IEEE 802.3. There
are some minor electrical differences between these, and one rather substantial packet-format difference. In
addition to these, the Berkeley Unix trailing-headers packet format was used for a while.
There were three physical formats for 10 Mbps Ethernet cable: thick coax (10BASE-5), thin coax (10BASE2), and, last to arrive, twisted pair (10BASE-T). Thick coax was the original; economics drove the successive
development of the later two. The cheaper twisted-pair cabling eventually almost entirely displaced coax, at
least for host connections.
The original specification included support for repeaters, which were in effect signal amplifiers although
they might attempt to clean up a noisy signal. Repeaters processed each bit individually and did no buffering.
In the telecom world, a repeater might be called a digital regenerator. A repeater with more than two ports
was commonly called a hub; hubs allowed branching and thus much more complex topologies.
Bridges later known as switches came along a short time later. While repeaters act at the bit layer,
a switch reads in and forwards an entire packet as a unit, and the destination address is likely consulted
to determine to where the packet is forwarded. Originally, switches were seen as providing interconnection (bridging) between separate Ethernets, but later a switched Ethernet was seen as one large virtual
Ethernet. We return to switching below in 2.4 Ethernet Switches.
Hubs propagate collisions; switches do not. If the signal representing a collision were to arrive at one port of
a hub, it would, like any other signal, be retransmitted out all other ports. If a switch were to detect a collision
one one port, no other ports would be involved; only packets received successfully are ever retransmitted
out other ports.
In coaxial-cable installations, one long run of coax snaked around the computer room or suite of offices;
each computer connected somewhere along the cable. Thin coax allowed the use of T-connectors to attach
hosts; connections were made to thick coax via taps, often literally drilled into the coax central conductor.
In a standalone installation one run of coax might be the entire Ethernet; otherwise, somewhere a repeater
would be attached to allow connection to somewhere else.
Twisted-pair does not allow mid-cable attachment; it is only used for point-to-point links between hosts,
switches and hubs. In a twisted-pair installation, each cable runs between the computer location and a
central wiring closest (generally much more convenient than trying to snake coax all around the building).
Originally each cable in the wiring closet plugged into a hub; nowadays the hub has likely been replaced by
a switch.
There is still a role for hubs today when one wants to monitor the Ethernet signal from A to B (eg for
intrusion detection analysis), although some switches now also support a form of monitoring.
All three cable formats could interconnect, although only through repeaters and hubs, and all used the same
10 Mbps transmission speed. While twisted-pair cable is still used by 100 Mbps Ethernet, it generally needs
to be a higher-performance version known as Category 5, versus the 10 Mbps Category 3.
Here is the format of a typical Ethernet packet (DIX specification):

The destination and source addresses are 48-bit quantities; the type is 16 bits, the data length is variable up
to a maximum of 1500 bytes, and the final CRC checksum is 32 bits. The checksum is added by the Ethernet
34

2 Ethernet

An Introduction to Computer Networks, Release 1.0

hardware, never by the host software. There is also a preamble, not shown: a block of 1 bits followed by a
0, in the front of the packet, for synchronization. The type field identifies the next higher protocol layer; a
few common type values are 0x0800 = IP, 0x8137 = IPX, 0x0806 = ARP.
The IEEE 802.3 specification replaced the type field by the length field, though this change never caught on.
The two formats can be distinguished as long as the type values used are larger than the maximum Ethernet
length of 1500 (or 0x05dc); the type values given in the previous paragraph all meet this condition.
Each Ethernet card has a (hopefully unique) physical address in ROM; by default any packet sent to this
address will be received by the board and passed up to the host system. Packets addressed to other physical
addresses will be seen by the card, but ignored (by default). All Ethernet devices also agree on a broadcast
address of all 1s: a packet sent to the broadcast address will be delivered to all attached hosts.
It is sometimes possible to change the physical address of a given card in software. It is almost universally
possible to put a given card into promiscuous mode, meaning that all packets on the network, no matter
what the destination address, are delivered to the attached host. This mode was originally intended for
diagnostic purposes but became best known for the security breach it opens: it was once not unusual to find
a host with network board in promiscuous mode and with a process collecting the first 100 bytes (presumably
including userid and password) of every telnet connection.

2.1.1 Ethernet Multicast

Another category of Ethernet addresses is multicast, used to transmit to a set of stations; streaming video
to multiple simultaneous viewers might use Ethernet multicast. The lowest-order bit in the first byte of
an address indicates whether the address is physical or multicast. To receive packets addressed to a given
multicast address, the host must inform its network interface that it wishes to do so; once this is done, any
arriving packets addressed to that multicast address are forwarded to the host. The set of subscribers to a
given multicast address may be called a multicast group. While higher-level protocols might prefer that the
subscribing host also notifies some other host, eg the sender, this is not required, although that might be the
easiest way to learn the multicast address involved. If several hosts subscribe to the same multicast address,
then each will receive a copy of each multicast packet transmitted.
If switches (below) are involved, they must normally forward multicast packets on all outbound links, exactly
as they do for broadcast packets; switches have no obvious way of telling where multicast subscribers might
be. To avoid this, some switches do try to engage in some form of multicast filtering, sometimes by snooping
on higher-layer multicast protocols. Multicast Ethernet is seldom used by IPv4, but plays a larger role in
IPv6 configuration.
The second-to-lowest-order bit of the Ethernet address indicates, in the case of physical addresses, whether
the address is believed to be globally unique or if it is only locally unique; this is known as the Universal/Local bit. For real Ethernet physical addresses, the multicast and universal/local bits of the first byte
should both be 0.
When (global) Ethernet IDs are assigned by the manufacturer, the first three bytes serve to indicate the
manufacturer. They are allocated by the IEEE, and are officially known as organizationally unique identifiers. These can be looked up at any of several sites on the Internet to identify the manufacturer associated
with any given Ethernet address; the official IEEE site is [Link]/develop/regauth/oui/[Link]
(OUIs must be entered here without colons).

2.1 10-Mbps classic Ethernet

35

An Introduction to Computer Networks, Release 1.0

As long as the manufacturer involved is diligent in assigning the second three bytes, every manufacturerprovided Ethernet address should be globally unique. Lapses, however, are not unheard of.

2.1.2 The LAN Layer

The LAN layer, at its upper end, supplies to the network layer a mechanism for addressing a packet and
sending it from one station to another. At its lower end, it handles interactions with the physical layer. The
LAN layer covers packet addressing, delivery and receipt, forwarding, error detection, collision detection
and collision-related retransmission attempts.
In IEEE protocols, the LAN layer is divided into the media access control, or MAC, sublayer and a
higher logical link control, or LLC, sublayer for higher-level flow-control functions that today have moved
largely to the transport layer. For example, the HDLC protocol ([Link] HDLC) supports sliding windows
(6.2 Sliding Windows) as an option, as did the early X.25 protocol. ATM, 3.8 Asynchronous Transfer
Mode: ATM, also supports some higher-level functions, though not sliding windows.
Because the LLC layer is so often insignificant, and because the most well-known LAN-layer functions are
in fact part of the MAC sublayer, it is common to identify the LAN layer with its MAC sublayer, especially
for IEEE protocols where the MAC layer has official standing. In particular, LAN-layer addresses are
perhaps most often called MAC addresses.
Generally speaking, much of the operation of the LAN/MAC layer takes place in the network card. Host
systems (including drivers) are, for example, generally oblivious to collisions (although they may query the
card for collision statistics). In some cases, eg with Wi-Fi rate scaling (3.3.3 Dynamic Rate Scaling), the
host-system driver may get involved.

2.1.3 The Slot Time and Collisions

The diameter of an Ethernet is the maximum distance between any pair of stations. The actual total length
of cable can be much greater than this, if, for example, the topology is a star configuration. The maximum
allowed diameter, measured in bits, is limited to 232 (a sample budget for this is below). This makes
the round-trip-time 464 bits. As each station involved in a collision discovers it, it transmits a special jam
signal of up to 48 bits. These 48 jam bits bring the total above to 512 bits, or 64 bytes. The time to send
these 512 bits is the slot time of an Ethernet; time intervals on Ethernet are often described in bit times but
in conventional time units the slot time is 51.2 sec.
The value of the slot time determines several subsequent aspects of Ethernet. If a station has transmitted
for one slot time, then no collision can occur (unless there is a hardware error) for the remainder of that
packet. This is because one slot time is enough time for any other station to have realized that the first
station has started transmitting, so after that time they will wait for the first station to finish. Thus, after one
slot time a station is said to have acquired the network. The slot time is also used as the basic interval for
retransmission scheduling, below.
Conversely, a collision can be received, in principle, at any point up until the end of the slot time. As a result,
Ethernet has a minimum packet size, equal to the slot time, ie 64 bytes (or 46 bytes in the data portion). A
station transmitting a packet this size is assured that if a collision were to occur, the sender would detect it
(and be able to apply the retransmission algorithm, below). Smaller packets might collide and yet the sender
not know it, ultimately leading to greatly reduced throughput.

36

2 Ethernet

An Introduction to Computer Networks, Release 1.0

If we need to send less than 46 bytes of data (for example, a 40-byte TCP ACK packet), the Ethernet packet
must be padded out to the minimum length. As a result, all protocols running on top of Ethernet need to
provide some way to specify the actual data length, as it cannot be inferred from the received packet size.
As a specific example of a collision occurring as late as possible, consider the diagram below. A and B are
5 units apart, and the bandwidth is 1 byte/unit. A begins sending helloworld at T=0; B starts sending just
as As message arrives, at T=5. B has listened before transmitting, but As signal was not yet evident. A
doesnt discover the collision until 10 units have elapsed, which is twice the distance.

Here are typical maximum values for the delay in 10 Mbps Ethernet due to various components. These
are taken from the Digital-Intel-Xerox (DIX) standard of 1982, except that point-to-point link cable is
replaced by standard cable. The DIX specification allows 1500m of coax with two repeaters and 1000m
of point-to-point cable; the table below shows 2500m of coax and four repeaters, following the later IEEE
802.3 Ethernet specification. Some of the more obscure delays have been eliminated. Entries are one-way
delay times, in bits. The maximum path may have four repeaters, and ten transceivers (simple electronic
devices between the coax cable and the NI cards), each with its drop cable (two transceivers per repeater,
plus one at each endpoint).
Ethernet delay budget
item
coax
transceiver cables
transceivers
repeaters
encoders

length
2500M
500M

delay, in bits
110 bits
25 bits
40 bits, max 10 units
25 bits, max 4 units
20 bits, max 10 units

explanation (c = speed of light)

23 meters/bit (.77c)
19.5 meters/bit (.65c)
4 bits each
6+ bits each (DIX [Link])
2 bits each (for signal generation)

The total here is 220 bits; in a full accounting it would be 232. Some of the numbers shown are a little high,
but there are also signal rise time delays, sense delays, and timer delays that have been omitted. It works out
fairly closely.
Implicit in the delay budget table above is the length of a bit. The speed of propagation in copper is about
0.77c, where c=3108 m/sec = 300 m/sec is the speed of light in vacuum. So, in 0.1 microseconds (the

2.1 10-Mbps classic Ethernet

37

An Introduction to Computer Networks, Release 1.0

time to send one bit at 10 Mbps), the signal propagates approximately 0.77c10-7 = 23 meters.
Ethernet packets also have a maximum packet size, of 1500 bytes. This limit is primarily for the sake
of fairness, so one station cannot unduly monopolize the cable (and also so stations can reserve buffers
guaranteed to hold an entire packet). At one time hardware vendors often marketed their own incompatible
extensions to Ethernet which enlarged the maximum packet size to as much as 4KB. There is no technical
reason, actually, not to do this, except compatibility.
The signal loss in any single segment of cable is limited to 8.5 db, or about 14% of original strength.
Repeaters will restore the signal to its original strength. The reason for the per-segment length restriction
is that Ethernet collision detection requires a strict limit on how much the remote signal can be allowed to
lose strength. It is possible for a station to detect and reliably read very weak remote signals, but not at the
same time that it is transmitting locally. This is exactly what must be done, though, for collision detection
to work: remote signals must arrive with sufficient strength to be heard even while the receiving station is
itself transmitting. The per-segment limit, then, has nothing to do with the overall length limit; the latter is
set only to ensure that a sender is guaranteed of detecting a collision, even if it sends the minimum-sized
packet.

2.1.4 Exponential Backoff Algorithm

Whenever there is a collision the exponential backoff algorithm operating at the MAC layer is used
to determine when each station will retry its transmission. Backoff here is called exponential because the
range from which the backoff value is chosen is doubled after every successive collision involving the same
packet. Here is the full Ethernet transmission algorithm, including backoff and retransmissions:
1. Listen before transmitting (carrier detect)
2. If line is busy, wait for sender to stop and then wait an additional 9.6 microseconds (96 bits). One
consequence of this is that there is always a 96-bit gap between packets, so packets do not run together.
3. Transmit while simultaneously monitoring for collisions
4. If a collision does occur, send the jam signal, and choose a backoff time as follows: For transmission
N, 1N10 (N=0 represents the original attempt), choose k randomly with 0 k < 2N . Wait k slot
times (k51.2 sec). Then check if the line is idle, waiting if necessary for someone else to finish,
and then retry step 3. For 11N15, choose k randomly with 0 k < 1024 (= 210 )
5. If we reach N=16 (16 transmission attempts), give up.
If an Ethernet sender does not reach step 5, there is a very high probability that the packet was delivered
successfully.
Exponential backoff means that if two hosts have waited for a third to finish and transmit simultaneously,
and collide, then when N=1 they have a 50% chance of recollision; when N=2 there is a 25% chance, etc.
When N10 the maximum wait is 52 milliseconds; without this cutoff the maximum wait at N=15 would
be 1.5 seconds. As indicated above in the minimum-packet-size discussion, this retransmission strategy
assumes that the sender is able to detect the collision while it is still sending, so it knows that the packet
must be resent.
In the following diagram is an example of several stations attempting to transmit all at once, and using the
above transmission/backoff algorithm to sort out who actually gets to acquire the channel. We assume we
have five prospective senders A1, A2, A3, A4 and A5, all waiting for a sixth station to finish. We will
38

2 Ethernet

An Introduction to Computer Networks, Release 1.0

assume that collision detection always takes one slot time (it will take much less for nodes closer together)
and that the slot start-times for each station are synchronized; this allows us to measure time in slots. A solid
arrow at the start of a slot means that sender began transmission in that slot; a red X signifies a collision. If
a collision occurs, the backoff value k is shown underneath. A dashed line shows the station waiting k slots
for its next attempt.

At T=0 we assume the transmitting station finishes, and all the Ai transmit and collide. At T=1, then, each
of the Ai has discovered the collision; each chooses a random k<2. Let us assume that A1 chooses k=1, A2
chooses k=1, A3 chooses k=0, A4 chooses k=0, and A5 chooses k=1.
Those stations choosing k=0 will retransmit immediately, at T=1. This means A3 and A4 collide again, and
at T=2 they now choose random k<4. We will Assume A3 chooses k=3 and A4 chooses k=0; A3 will try
again at T=2+3=5 while A4 will try again at T=2, that is, now.
At T=2, we now have the original A1, A2, and A5 transmitting for the second time, while A4 trying again
for the third time. They collide. Let us suppose A1 chooses k=2, A2 chooses k=1, A5 chooses k=3, and A4
chooses k=6 (A4 is choosing k<8 at random). Their scheduled transmission attempt times are now A1 at
T=3+2=5, A2 at T=4, A5 at T=6, and A4 at T=9.
At T=3, nobody attempts to transmit. But at T=4, A2 is the only station to transmit, and so successfully
seizes the channel. By the time T=5 rolls around, A1 and A3 will check the channel, that is, listen first, and
wait for A2 to finish. At T=9, A4 will check the channel again, and also begin waiting for A2 to finish.
A maximum of 1024 hosts is allowed on an Ethernet. This number apparently comes from the maximum
range for the backoff time as 0 k < 1024. If there are 1024 hosts simultaneously trying to send, then,
once the backoff range has reached k<1024 (N=10), we have a good chance that one station will succeed in
seizing the channel, that is; the minimum value of all the random ks chosen will be unique.
This backoff algorithm is not fair, in the sense that the longer a station has been waiting to send, the lower
its priority sinks. Newly transmitting stations with N=0 need not delay at all. The Ethernet capture effect,
below, illustrates this unfairness.

2.1 10-Mbps classic Ethernet

39

An Introduction to Computer Networks, Release 1.0

2.1.5 Capture effect

The capture effect is a scenario illustrating the potential lack of fairness in the exponential backoff algorithm.
The unswitched Ethernet must be fully busy, in that each of two senders always has a packet ready to
transmit.
Let A and B be two such busy nodes, simultaneously starting to transmit their first packets. They collide.
Suppose A wins, and sends. When A is finished, B tries to transmit again. But A has a second packet, and
so A tries too. A chooses a backoff k<2 (that is, between 0 and 1 inclusive), but since B is on its second
attempt it must choose k<4. This means A is favored to win. Suppose it does.
After that transmission is finished, A and B try yet again: A on its first attempt for its third packet, and B
on its third attempt for its first packet. Now A again chooses k<2 but B must choose k<8; this time A is
much more likely to win. Each time B fails to win a given backoff, its probability of winning the next one
is reduced by about 1/2. It is quite possible, and does occur in practice, for B to lose all the backoffs until
it reaches the maximum of N=16 attempts; once it has lost the first three or four this is in fact quite likely.
At this point B simply discards the packet and goes on to the next one with N reset to 1 and k chosen from
{0,1}.
The capture effect can be fixed with appropriate modification of the backoff algorithm; the Binary Logarithmic Arbitration Method (BLAM) was proposed in [MM94]. The BLAM algorithm was considered for the
then-nascent 100 Mbps Fast Ethernet standard. But in the end a hardware strategy won out: Fast Ethernet
supports full-duplex mode which is collision-free (see 2.2 100 Mbps (Fast) Ethernet, below). While
full-duplex mode is not required for using Fast Ethernet, it was assumed that any sites concerned enough
about performance to be worried about the capture effect would opt for full-duplex.

2.1.6 Hubs and topology

Ethernet hubs (multiport repeaters) change the topology, but not the fundamental constraints. Hubs allow
much more branching; typically, each station in the office now has its own link to the wiring closet. Loops
are still forbidden. Before inexpensive switches were widely available, 10BASE-T (twisted pair Ethernet)
used hubs heavily; with twisted pair, a device can only connect to the endpoint of the wire. Thus, typically,
each host is connected directly to a hub. The maximum diameter of an Ethernet consisting of multiple
segments, joined by hubs, is constrained by the round-trip-time, and the need to detect collisions before the
sender has completed sending, as before. However, twisted-pair links are required to be much shorter, about
100 meters.

2.1.7 Errors
Packets can have bits flipped or garbled by electrical noise on the cable; estimates of the frequency with
which this occurs range from 1 in 104 to 1 in 106 . Bit errors are not uniformly likely; when they occur,
they are likely to occur in bursts. Packets can also be lost in hubs, although this appears less likely. Packets
can be lost due to collisions only if the sending host makes 16 unsuccessful transmission attempts and gives
up. Ethernet packets contain a 32-bit CRC error-detecting code (see 5.4.1 Cyclical Redundancy Check:
CRC) to detect bit errors. Packets can also be misaddressed by the sending host, or, most likely of all, they
can arrive at the receiving host at a point when the receiver has no free buffers and thus be dropped by a
higher-layer protocol.

40

2 Ethernet

An Introduction to Computer Networks, Release 1.0

2.1.8 CSMA persistence

A carrier-sense/multiple-access transmission strategy is said to be nonpersistent if, when the line is busy,
the sender waits a randomly selected time. A strategy is p-persistent if, after waiting for the line to clear,
the sender sends with probability p1. Ethernet uses 1-persistence. A consequence of 1-persistence is that,
if more than one station is waiting for line to clear, then when the line does clear a collision is certain.
However, Ethernet then gracefully handles the resulting collision via the usual exponential backoff. If N
stations are waiting to transmit, the time required for one station to win the backoff is linear in N.
When we consider the Wi-Fi collision-handling mechanisms in 3.3 Wi-Fi, we will see that collisions cannot
be handled quite as cheaply: for one thing, there is no way to detect a collision in progress, so the entire
packet-transmission time is wasted. In the Wi-Fi case, p-persistence is used with p<1.
An Ethernet broadcast storm was said to occur when there were too many transmission attempts, and most
of the available bandwidth was tied up in collisions. A properly functioning classic Ethernet had an effective
bandwidth of as much as 50-80% of the nominal 10Mbps capacity, but attempts to transmit more than this
typically resulted in successfully transmitting a good deal less.

2.1.9 Analysis of Classic Ethernet

How much time does Ethernet waste on collisions? A paradoxical attribute of Ethernet is that raising the
transmission-attempt rate on a busy segment can reduce the actual throughput. More transmission attempts
can lead to longer contention intervals between packets, as senders use the transmission backoff algorithm
to attempt to acquire the channel. What effective throughput can be achieved?
It is convenient to refer to the time between packet transmissions as the contention interval even if there is
no actual contention, even if the network is idle. Thus, a timeline for Ethernet always consists of alternating
packet transmissions and contention intervals:

As a first look at contention intervals, assume that there are N stations waiting to transmit at the start of the
interval. It turns out that, if all follow the exponential backoff algorithm, we can expect O(N) slot times
before one station successfully acquires the channel; thus, Ethernets are happiest when N is small and there
are only a few stations simultaneously transmitting. However, multiple stations are not necessarily a severe
problem. Often the number of slot times needed turns out to be about N/2, and slot times are short. If N=20,
then N/2 is 10 slot times, or 640 bytes. However, one packet time might be 1500 bytes. If packet intervals
are 1500 bytes and contention intervals are 640 byes, this gives an overall throughput of 1500/(640+1500)
= 70% of capacity. In practice, this seems to be a reasonable upper limit for the throughput of classic
shared-media Ethernet.

2.1 10-Mbps classic Ethernet

41

An Introduction to Computer Networks, Release 1.0

[Link] The ALOHA models

Another approach to analyzing the Ethernet contention interval is by using the ALOHA model that was a
precursor to Ethernet. In the ALOHA model, stations transmit packets without listening first for a quiet line
or monitoring the transmission for collisions (this models the situation of several ground stations transmitting
to a satellite; the ground stations are presumed unable to see one another). Similarly, during the Ethernet
contention interval, stations transmit one-slot packets under what are effectively the same conditions (we
return to this below).
The ALOHA model yields very similar throughput values to the O(N) model of the previous section. We
make, however, a rather artificial assumption: that there are a very large number of active senders, each
transmitting at a very low rate.
To model the success rate of ALOHA, assume all the packets are the same size and let T be the time to send
one (fixed-size) packet; T represents the Aloha slot time. We will find the transmission rate that optimizes
throughput.
The core assumption of this model is that that a large number N of hosts are transmitting, each at a relatively
low rate of s packets/slot. Denote by G the average number of transmission attempts per slot; we then have
G = Ns. We will derive an expression for S, the average rate of successful transmissions per slot, in terms of
G.
If two packets overlap during transmissions, both are lost. Thus, a successful transmission requires everyone
else quiet for an interval of 2T: if a sender succeeds in the interval from t to t+T, then no other node can have
tried to begin transmission in the interval tT to t+T. The probability of one station transmitting during an
interval of time T is G = Ns; the probability of the remaining N1 stations all quiet for an interval of 2T is
(1s)2(N1) . The probability of a successful transmission is thus
S = Ns*(1s)2(N1)
= G(1G/N)2N
Math Warning
Finding the limit of G(1G/N)2N and finding the maximum of Ge-2G realistically requires a little background in calculus. However, these are not central to applying the model.
As N gets large, the second line approaches Ge-2G . The function S = G e-2G has a maximum at G=1/2,
S=1/2e. The rate G=1/2 means that, on average, a transmission is attempted every other slot; this yields the
maximum successful-transmission throughput of 1/2e. In other words, at this maximum attempt rate G=1/2,
we expect about 2e1 slot times worth of contention between successful transmissions. What happens to
the remaining GS unsuccessful attempts is not addressed by this model; presumably some higher-level
mechanism (eg backoff) leads to retransmissions.
A given throughput S<1/2e may be achieved at either of two values for G; that is, a given success rate may
be due to a comparable attempt rate or else due to a very high attempt rate with a similarly high failure rate.

42

2 Ethernet

An Introduction to Computer Networks, Release 1.0

[Link] ALOHA and Ethernet

The relevance of the Aloha model to Ethernet is that during one Ethernet slot time there is no way to detect
collisions (they havent reached the sender yet!) and so the Ethernet contention phase resembles ALOHA
with an Aloha slot time T of 51.2 microseconds. Once an Ethernet sender succeeds, however, it continues
with a full packet transmission, which is presumably many times longer than T.
The average length of the contention interval, at the maximum throughput calculated above, is 2e1 slot
times (from ALOHA); recall that our model here supposed many senders sending at very low individual
rates. This is the minimum contention interval; with lower loads the contention interval is longer due to
greater idle times and with higher loads the contention interval is longer due to more collisions.
Finally, let P be the time to send an entire packet in units of T; ie the average packet size in units of T. P is
thus the length of the packet phase in the diagram above. The contention phase has length 2e1, so the
total time to send one packet (contention+packet time) is 2e1+P. The useful fraction of this is, of course, P,
so the effective maximum throughput is P/(2e1+P).
At 10Mbps, T=51.2 microseconds is 512 bits, or 64 bytes. For P=128 bytes = 2*64, the effective bandwidth
becomes 2/(2e-1+2), or 31%. For P=512 bytes=8*64, the effective bandwidth is 8/(2e+7), or 64%. For
P=1500 bytes, the model here calculates an effective bandwidth of 80%.
These numbers are quite similar to our earlier values based on a small number of stations sending constantly.

2.2 100 Mbps (Fast) Ethernet

In all the analysis here of 10 Mbps Ethernet, what happens when the bandwidth is increased to 100 Mbps,
as is done in the so-called Fast Ethernet standard? If the network physical diameter remains the same, then
the round-trip time will be the same in microseconds but will be 10-fold larger measured in bits; this might
mean a minimum packet size of 640 bytes instead of 64 bytes. (Actually, the minimum packet size might
be somewhat smaller, partly because the jam signal doesnt have to speed up at all, and partly because
some of the numbers in the 10 Mbps delay budget above were larger than necessary, but it would still be
large enough that a substantial amount of bandwidth would be consumed by padding.) The designers of Fast
Ethernet felt this was impractical.
However, Fast Ethernet was developed at a time (~1995) when reliable switches (below) were widely available, and longer networks could be formed by chaining together shorter ones with switches. So instead
of increasing the minimum packet size, the decision was made to ensure collision detectability by reducing
the network diameter instead. The network diameter chosen was a little over 400 meters, with reductions to
account for the presence of hubs. At 2.3 meters/bit, 400 meters is 174 bits, for a round-trip of 350 bits.
This 400-meter number, however, may be misleading: by far the most popular Fast Ethernet standard is
100BASE-TX which uses twisted-pair copper wire (so-called Category 5, or better), and in which any individual cable segment is limited to 100 meters. The maximum 100BASE-TX network diameter allowing
for hubs is just over 200 meters. The 400-meter distance does apply to optical-fiber-based 100BASE-FX
in half-duplex mode, but this is not common.
The 100BASE-TX network-diameter limit of 200 meters might seem small; it amounts in many cases to
a single hub with multiple 100-meter cable segments radiating from it. In practice, however, such star
configurations can easily be joined with switches. As we will see below in 2.4 Ethernet Switches, switches
partition an Ethernet into separate collision domains; the network-diameter rules apply to each domain
2.2 100 Mbps (Fast) Ethernet

43

An Introduction to Computer Networks, Release 1.0

separately but not to the aggregated whole. In a fully switched (that is, no hubs) 100BASE-TX LAN, each
collision domain is simply a single twisted-pair link, subject to the 100-meter maximum length.
Fast Ethernet also introduced the concept of full-duplex Ethernet: two twisted pairs could be used, one
for each direction. Full-duplex Ethernet is limited to paths not involving hubs, that is, to single station-tostation links, where a station is either a host or a switch. Because such a link has only two potential senders,
and each sender has its own transmit line, full-duplex Ethernet is collision-free.
Fast Ethernet uses 4B/5B encoding, covered in 4.1.4 4B/5B.
Fast Ethernet 100BASE-TX does not particularly support links between buildings, due to the networkdiameter limitation. However, fiber-optic point-to-point links are quite effective here, provided full-duplex
is used to avoid collisions. We mentioned above that the coax-based 100BASE-FX standard allowed a
maximum half-duplex run of 400 meters, but 100BASE-FX is much more likely to use full duplex, where
the maximum cable length rises to 2,000 meters.

2.3 Gigabit Ethernet

If we continue to maintain the same slot time but raise the transmission rate to 1000 Mbps, the network
diameter would now be 20-40 meters. Instead of that, Gigabit Ethernet moved to a 4096-bit (512-byte) slot
time, at least for the twisted-pair versions. Short frames need to be padded, but this padding is done by the
hardware. Gigabit Ethernet 1000Base-T uses so-called PAM-5 encoding, below, which supports a special
pad pattern (or symbol) that cannot appear in the data. The hardware pads the frame with these special
patterns, and the receiver can thus infer the unpadded length as set by the host operating system.
However, the Gigabit Ethernet slot time is largely irrelevant, as full-duplex (bidirectional) operation is almost
always supported. Combined with the restriction that each length of cable is a station-to-station link (that is,
hubs are no longer allowed), this means that collisions simply do not occur and the network diameter is no
longer a concern.
There are actually multiple Gigabit Ethernet standards (as there are for Fast Ethernet). The different standards apply to different cabling situations. There are full-duplex optical-fiber formulations good for many
miles (eg 1000Base-LX10), and even a version with a 25-meter maximum cable length (1000Base-CX),
which would in theory make the original 512-bit slot practical.
The most common gigabit Ethernet over copper wire is 1000BASE-T (sometimes incorrectly referred to
as 100BASE-TX. While there exists a TX, it requires Category 6 cable and is thus seldom used; many
devices labeled TX are in fact 1000BASE-T). For 1000BASE-T, all four twisted pairs in the cable are used.
Each pair transmits at 250 Mbps, and each pair is bidirectional, thus supporting full-duplex communication.
Bidirectional communication on a single wire pair takes some careful echo cancellation at each end, using
a circuit known as a hybrid that in effect allows detection of the incoming signal by filtering out the
outbound signal.
On any one cable pair, there are five signaling levels. These are used to transmit two-bit symbols
(4.1.4 4B/5B) at a rate of 125 symbols/sec, for a data rate of 250 bits/sec. Two-bit symbols in theory only require four signaling levels; the fifth symbol allows for some redundancy which is used for error
detection and correction, for avoiding long runs of identical symbols, and for supporting a special pad symbol, as mentioned above. The encoding is known as 5-level pulse-amplitude modulation, or PAM-5. The
target bit error rate (BER) for 1000BASE-T is 10-10 , meaning that the packet error rate is less than 1 in 106 .

44

2 Ethernet

An Introduction to Computer Networks, Release 1.0

In developing faster Ethernet speeds, economics plays at least as important a role as technology. As new
speeds reach the market, the earliest adopters often must take pains to buy cards, switches and cable known
to work together; this in effect amounts to installing a proprietary LAN. The real benefit of Ethernet,
however, is arguably that it is standardized, at least eventually, and thus a site can mix and match its cards
and devices. Having a given Ethernet standard support existing cable is even more important economically;
the costs of replacing cable often dwarf the costs of the electronics.

2.4 Ethernet Switches

Switches join separate physical Ethernets (or Ethernets and token rings). A switch has two or more Ethernet
interfaces; when a packet is received on one interface it is retransmitted on one or more other interfaces.
Only valid packets are forwarded; collisions are not propagated. The term collision domain is sometimes
used to describe the region of an Ethernet in between switches; a given collision propagates only within its
collision domain. All the collision-detection rules, including the rules for maximum network diameter, apply
only to collision domains, and not to the larger virtual Ethernets created by stringing collision domains
together with switches.
As we shall see below, a switched Ethernet offers much more resistance to eavesdropping than a nonswitched (eg hub-based) Ethernet.
Like simpler unswitched Ethernets, the topology for a switched Ethernet is in principle required to be loopfree, although in practice, most switches support the spanning-tree loop-detection protocol and algorithm,
below, which automatically prunes the network topology to make it loop-free.
And while a switch does not propagate collisions, it must maintain a queue for each outbound interface in
case it needs to forward a packet at a moment when the interface is busy; on occasion packets are lost when
this queue overflows.
Ethernet switches use datagram forwarding as described in 1.4 Datagram Forwarding. They start out with
empty forwarding tables, and build them through a learning process. If a switch does not have an entry for
a particular destination, it will fall back to broadcast: it will forward the packet out every interface other
than the one on which the packet arrived. The availability of fallback-to-broadcast is what makes it possible
for Ethernet switches to learn their forwarding tables without any switch-to-switch communication.
A switch learns address locations as follows: for each interface, the switch maintains a table of physical
addresses that have appeared as source addresses in packets arriving via that interface. The switch thus
knows that to reach these addresses, if one of them later shows up as a destination address, the packet needs
to be sent only via that interface. Specifically, when a packet arrives on interface I with source address S
and destination unicast address D, the switch enters S,I into its forwarding table.
To actually deliver the packet, the switch also looks up D in the forwarding table. If there is an entry D,J
with J=I that is, D is known to be reached via interface J then the switch forwards the packet out interface
J. If J=I, that is, the packet has arrived on the same interfaces by which the destination is reached, then the
packet does not get forwarded at all; it presumably arrived at interface I only because that interface was
connected to a shared Ethernet segment that also either contained D or contained another switch that would
bring the packet closer to D. If there is no entry for D, the switch must forward the packet out all interfaces
J with J=I; this represents the fallback to broadcast. As time goes on, this fallback to broadcast is needed
less and less often.

2.4 Ethernet Switches

45

An Introduction to Computer Networks, Release 1.0

If the destination address D is the broadcast address, or, for many switches, a multicast address, broadcast
is required.

In the diagram above, each switchs tables are indicated by listing near each interface the destinations known
to be reachable by that interface. The entries shown are the result of the following packets:
A sends to B; all switches learn where A is
B sends to A; this packet goes directly to A; only S3, S2 and S1 learn where B is
C sends to B; S4 does not know where B is so this packet goes to S5; S2 does know where B is so the
packet does not go to S1.
Switches do not automatically discover directly connected neighbors; S1 does not learn about A until A
transmits a packet.
Once all the switches have learned where all (or most of) the hosts are, packet routing becomes optimal. At
this point packets are never sent on links unnecessarily; a packet from A to B only travels those links that
lie along the (unique) path from A to B. (Paths must be unique because switched Ethernet networks cannot
have loops, at least not active ones. If a loop existed, then a packet sent to an unknown destination would be
forwarded around the loop endlessly.)
Switches have an additional advantage in that traffic that does not flow where it does not need to flow is
much harder to eavesdrop on. On an unswitched Ethernet, one host configured to receive all packets can
eavesdrop on all traffic. Early Ethernets were notorious for allowing one unscrupulous station to capture,
for instance, all passwords in use on the network. On a fully switched Ethernet, a host physically only sees
the traffic actually addressed to it; other traffic remains inaccessible.
Typical switches have room for table with 104 - 106 entries, though maxing out at 105 entries may be more
common; this is usually enough to learn about all hosts in even a relatively large organization. A switched
Ethernet can fail when total traffic becomes excessive, but excessive total traffic would drown any network
(although other network mechanisms might support higher bandwidth). The main limitations specific to
switching are the requirement that the topology must be loop-free (thus disallowing duplicate paths which
might otherwise provide redundancy), and that all broadcast traffic must always be forwarded everywhere.
As a switched Ethernet grows, broadcast traffic comprises a larger and larger percentage of the total traffic,
and the organization must at some point move to a routing architecture (eg as in 7.6 IP Subnets).
One of the differences between an inexpensive Ethernet switch and a pricier one is the degree of internal
parallelism it can support. If three packets arrive simultaneously on ports 1, 2 and 3, and are destined for
respective ports 4, 5 and 6, can the switch actually transmit the packets simultaneously? A simple switch
likely has a single CPU and a single memory bus, both of which can introduce transmission bottlenecks.
46

2 Ethernet

An Introduction to Computer Networks, Release 1.0

For commodity five-port switches, at most two simultaneous transmissions can occur; such switches can
generally handle that degree of parallelism. It becomes harder as the number of ports increases, but at some
point the need to support full parallel operation can be questioned; in many settings the majority of traffic
involves one or two server or router ports. If a high degree of parallelism is in fact required, there are various
architectures known as switch fabrics that can be used; these typically involve multiple simple processor
elements.

2.5 Spanning Tree Algorithm

In theory, if you form a loop with Ethernet switches, any packet with destination not already present in the
forwarding tables will circulate endlessly; naive switches will actually do this.
In practice, however, loops allow a form of redundancy if one link breaks there is still 100% connectivity
and so are desirable. As a result, Ethernet switches have incorporated a switch-to-switch protocol to
construct a subset of the switch-connections graph that has no loops and yet allows reachability of every
host, known as a spanning tree. The switch-connections graph is the graph with nodes consisting of both
switches and of the unswitched Ethernet segments and isolated individual hosts connected to the switches.
Multi-host Ethernet segments are most often created via Ethernet hubs (repeaters). Edges in the graph
represent switch-segment and switch-switch connections; each edge attaches to its switch via a particular,
numbered interface. The goal is to disable redundant (cyclical) paths while remaining able to deliver to any
segment. The algorithm is due to Radia Perlman, [RP85].
Once the spanning tree is built, all packets are sent only via edges in the tree, which, as a tree, has no loops.
Switch ports (that is, edges) that are not part of the tree are not used at all, even if they would represent
the most efficient path for that particular destination. If a given segment connects to two switches that
both connect to the root node, the switch with the shorter path to the root is used, if possible; in the event
of ties, the switch with the smaller ID is used. The simplest measure of path cost is the number of hops,
though current implementations generally use a cost factor inversely proportional to the bandwidth (so larger
bandwidth has lower cost). Some switches permit other configuration here. The process is dynamic, so if
an outage occurs then the spanning tree is recomputed. If the outage should partition the network into two
pieces, both pieces will build spanning trees.
All switches send out regular messages on all interfaces called bridge protocol data units, or BPDUs (or
Hello messages). These are sent to the Ethernet multicast address [Link], from the Ethernet
physical address of the interface. (Note that Ethernet switches do not otherwise need a unique physical
address for each interface.) The BPDUs contain
The switch ID
the ID of the node the switch believes is the root
the path cost to that root
These messages are recognized by switches and are not forwarded naively. Bridges process each message,
looking for
a switch with a lower ID (thus becoming the new root)
a shorter path to the existing root
an equal-length path to the existing root, but via a switch or port with a lower ID (the tie-breaker rule)

2.5 Spanning Tree Algorithm

47

An Introduction to Computer Networks, Release 1.0

When a switch sees a new root candidate, it sends BPDUs on all interfaces, indicating the distance. The
switch includes the interface leading towards the root.
Once this process is complete, each switch knows
its own path to the root
which of its ports any further-out switches will be using to reach the root
for each port, its directly connected neighboring switches
Now the switch can prune some (or all!) of its interfaces. It disables all interfaces that are not enabled by
the following rules:
1. It enables the port via which it reaches the root
2. It enables any of its ports that further-out switches use to reach the root
3. If a remaining port connects to a segment to which other segment-neighbor switches connect as well,
the port is enabled if the switch has the minimum cost to the root among those segment-neighbors, or,
if a tie, the smallest ID among those neighbors, or, if two ports are tied, the port with the smaller ID.
4. If a port has no directly connected switch-neighbors, it presumably connects to a host or segment, and
the port is enabled.
Rules 1 and 2 construct the spanning tree; if S3 reaches the root via S2, then Rule 1 makes sure S3s port
towards S2 is open, and Rule 2 makes sure S2s corresponding port towards S3 is open. Rule 3 ensures that
each network segment that connects to multiple switches gets a unique path to the root: if S2 and S3 are
segment-neighbors each connected to segment N, then S2 enables its port to N and S3 does not (because
2<3). The primary concern here is to create a path for any host nodes on segment N; S2 and S3 will create
their own paths via Rules 1 and 2. Rule 4 ensures that any stub segments retain connectivity; these would
include all hosts directly connected to switch ports.

2.5.1 Example 1: Switches Only

We can simplify the situation somewhat if we assume that the network is fully switched: each switch port
connects to another switch or to a (single-interface) host; that is, no repeater hubs (or coax segments!) are
in use. In this case we can dispense with Rule 3 entirely.
Any switch ports directly connected to a host can be identified because they are silent; the switch never
receives any BPDU messages on these interfaces because hosts do not send these. All these host port ends
up enabled via Rule 4. Here is our sample network, where the switch numbers (eg 5 for S5) represent their
IDs; no hosts are shown and interface numbers are omitted.

S1 has the lowest ID, and so becomes the root. S2 and S4 are directly connected, so they will enable the
interfaces by which they reach S1 (Rule 1) while S1 will enable its interfaces by which S2 and S4 reach it
48

2 Ethernet

An Introduction to Computer Networks, Release 1.0

(Rule 2).
S3 has a unique lowest-cost route to S1, and so again by Rule 1 it will enable its interface to S2, while by
Rule 2 S2 will enable its interface to S3.
S5 has two choices; it hears of equal-cost paths to the root from both S2 and S4. It picks the lower-numbered
neighbor S2; the interface to S4 will never be enabled. Similarly, S4 will never enable its interface to S5.
Similarly, S6 has two choices; it selects S3.
After these links are enabled (strictly speaking it is interfaces that are enabled, not links, but in all cases here
either both interfaces of a link will be enabled or neither), the network in effect becomes:

2.5.2 Example 2: Switches and Segments

As an example involving switches that may join via unswitched Ethernet segments, consider the following
network; S1, S2 and S3, for example, are all segment-neighbors via their common segment B. As before, the
switch numbers represent their IDs. The letters in the clouds represent network segments; these clouds may
include multiple hosts. Note that switches have no way to detect these hosts; only (as above) other switches.

Eventually, all switches discover S1 is the root (because 1 is the smallest of {1,2,3,4,5,6}). S2, S3 and S4
are one (unique) hop away; S5, S6 and S7 are two hops away.

2.5 Spanning Tree Algorithm

49

An Introduction to Computer Networks, Release 1.0

Algorhyme
I think that I shall never see
a graph more lovely than a tree.
A tree whose crucial property
is loop-free connectivity.
A tree that must be sure to span
so packet can reach every LAN.
First, the root must be selected.
By ID, it is elected.
Least-cost paths from root are traced.
In the tree, these paths are placed.
A mesh is made by folks like me,
then bridges find a spanning tree.
Radia Perlman
For the switches one hop from the root, Rule 1 enables S2s port 1, S3s port 1, and S4s port 1. Rule 2
enables the corresponding ports on S1: ports 1, 5 and 4 respectively. Without the spanning-tree algorithm
S2 could reach S1 via port 2 as well as port 1, but port 1 has a smaller number.
S5 has two equal-cost paths to the root: S5S4S1 and S5S3S1. S3 is the switch with the
lower ID; its port 2 is enabled and S5 port 2 is enabled.
S6 and S7 reach the root through S2 and S3 respectively; we enable S6 port 1, S2 port 3, S7 port 2 and S3
port 3.
The ports still disabled at this point are S1 ports 2 and 3, S2 port 2, S4 ports 2 and 3, S5 port 1, S6 port 2
and S7 port 1.
Now we get to Rule 3, dealing with how segments (and thus their hosts) connect to the root. Applying Rule
3,
We do not enable S2 port 2, because the network (B) has a direct connection to the root, S1
We do enable S4 port 3, because S4 and S5 connect that way and S4 is closer to the root. This enables
connectivity of network D. We do not enable S5 port 1.
S6 and S7 are tied for the path-length to the root. But S6 has smaller ID, so it enables port 2. S7s
port 1 is not enabled.
Finally, Rule 4 enables S4 port 2, and thus connectivity for host J. It also enables S1 port 2; network F has
two connections to S1 and port 2 is the lower-numbered connection.
All this port-enabling is done using only the data collected during the root-discovery phase; there is no
additional negotiation. The BPDU exchanges continue, however, so as to detect any changes in the topology.
If a link is disabled, it is not used even in cases where it would be more efficient to use it. That is, traffic
from F to B is sent via B1, D, and B5; it never goes through B7. IP routing, on the other hand, uses the

50

2 Ethernet

An Introduction to Computer Networks, Release 1.0

shortest path. To put it another way, all spanning-tree Ethernet traffic goes through the root node, or along
a path to or from the root node.
The traditional (IEEE 802.1D) spanning-tree protocol is relatively slow; the need to go through the treebuilding phase means that after switches are first turned on no normal traffic can be forwarded for ~30
seconds. Faster, revised protocols have been proposed to reduce this problem.
Another issue with the spanning-tree algorithm is that a rogue switch can announce an ID of 0, thus likely
becoming the new root; this leaves that switch well-positioned to eavesdrop on a considerable fraction of
the traffic. One of the goals of the Cisco Root Guard feature is to prevent this; another goal of this and
related features is to put the spanning-tree topology under some degree of administrative control. One likely
wants the root switch, for example, to be geographically at least somewhat centered.

2.6 Virtual LAN (VLAN)

What do you do when you have different people in different places who are logically tied together? For
example, for a while the Loyola University CS department was split, due to construction, between two
buildings.
One approach is to continue to keep LANs local, and use IP routing between different subnets. However,
it is often convenient (printers are one reason) to configure workgroups onto a single virtual LAN, or
VLAN. A VLAN looks like a single LAN, usually a single Ethernet LAN, in that all VLAN members will
see broadcast packets sent by other members and the VLAN will ultimately be considered to be a single IP
subnet (7.6 IP Subnets). Different VLANs are ultimately connected together, but likely only by passing
through a single, central IP router.
VLANs can be visualized and designed by using the concept of coloring. We logically assign all nodes on
the same VLAN the same color, and switches forward packets accordingly. That is, if S1 connects to red
machines R1 and R2 and blue machines B1 and B2, and R1 sends a broadcast packet, then it goes to R2 but
not to B1 or B2. Switches must, of course, be told the color of each of their ports.

In the diagram above, S1 and S3 each have both red and blue ports. The switch network S1-S4 will deliver
traffic only when the source and destination ports are the same color. Red packets can be forwarded to the
blue VLAN only by passing through the router R, entering Rs red port and leaving its blue port. R may
apply firewall rules to restrict redblue traffic.

2.6 Virtual LAN (VLAN)

51

An Introduction to Computer Networks, Release 1.0

When the source and destination ports are on the same switch, nothing needs to be added to the packet; the
switch can keep track of the color of each of its ports. However, switch-to-switch traffic must be additionally
tagged to indicate the source. Consider, for example, switch S1 above sending packets to S3 which has nodes
R3 (red) and B3 (blue). Traffic between S1 and S3 must be tagged with the color, so that S3 will know to
what ports it may be delivered. The IEEE 802.1Q protocol is typically used for this packet-tagging; a 32-bit
color tag is inserted into the Ethernet header after the source address and before the type field. The first
16 bits of this field is 0x8100, which becomes the new Ethernet type field and which identifies the frame as
tagged.
Double-tagging is possible; this would allow an ISP to have one level of tagging and its customers to have
another level.

2.7 Epilog
Ethernet dominates the LAN layer, but is not one single LAN protocol: it comes in a variety of speeds
and flavors. Higher-speed Ethernet seems to be moving towards fragmenting into a range of physical-layer
options for different types of cable, but all based on switches and point-to-point linking; different Ethernet
types can be interconnected only with switches. Once Ethernet finally abandons physical links that are
bi-directional (half-duplex links), it will be collision-free and thus will no longer need a minimum packet
size.
Other wired networks have largely disappeared (or have been renamed Ethernet). Wireless networks,
however, are here to stay, and for the time being at least have inherited the original Ethernets collisionmanagement concerns.

2.8 Exercises
1. Simulate the contention period of five Ethernet stations that all attempt to transmit at T=0 (presumably
when some sixth station has finished transmitting), in the style of the diagram in 2.1.4 Exponential Backoff
Algorithm. Assume that time is measured in slot times, and that exactly one slot time is needed to detect a
collision (so that if two stations transmit at T=1 and collide, and one of them chooses a backoff time k=0,
then that station will transmit again at T=2). Use coin flips or some other source of randomness.
2. Suppose we have Ethernet switches S1 through S3 arranged as below. All forwarding tables are initially
empty.
S1

S2

S3

(a). If A sends to B, which switches see this packet?

(b). If B then replies to A, which switches see this packet?
(c). If C then sends to B, which switches see this packet?
(d). If C then sends to D, which switches see this packet?

52

2 Ethernet

An Introduction to Computer Networks, Release 1.0

3. Suppose we have the Ethernet switches S1 through S4 arranged as below. All forwarding tables are
empty; each switch uses the learning algorithm of 2.4 Ethernet Switches.
B
S4
A

S1

S2

S3

Now suppose the following packet transmissions take place:

A sends to B
B sends to A
C sends to B
D sends to A
For each switch, list what source addresses (eg A,B,C,D) it has seen (and thus what nodes it has learned the
location of).
4. In the switched-Ethernet network below, find two packet transmissions so that, when a third transmission
AD occurs, the packet is delivered to B (that is, it is forwarded out all ports of S2), but is not similarly
delivered to C. All forwarding tables are initially empty, and each switch uses the learning algorithm of
2.4 Ethernet Switches.

S1

S2

S3

Hint: Destination D must be in S3s forwarding table, but must not be in S2s.
5. Given the Ethernet network with learning switches below, with (disjoint) unspecified parts represented
by ?, explain why it is impossible for a packet sent from A to B to be forwarded by S1 only to S2, but to be
forwarded by S2 out all of S2s other ports.

?
|
S1

?
|
S2

6. In the diagram of 2.4 Ethernet Switches, suppose node D is connected to S5, and, with the tables as
shown below the diagram, D sends to B.

(a). Which switches will see this packet, and thus learn about D?
(b). Which of the switches in part (a) do not already know where B is and will use fallback-to-broadcase
(ie, will forward the packet out all non-arrival interfaces)?

7. Suppose two Ethernet switches are connected in a loop as follows; S1 and S2 have their interfaces 1 and
2 labeled. These switches do not use the spanning-tree algorithm.
2.8 Exercises

53

An Introduction to Computer Networks, Release 1.0

Suppose A attempts to send a packet to destination B, which is unknown. S1 will therefore forward the
packet out interfaces 1 and 2. What happens then? How long will As packet circulate?
8. The following network is like that of 2.5.1 Example 1: Switches Only, except that the switches are
numbered differently. Again, the ID of switch Sn is n, so S1 will be the root. Which links end up pruned
by the spanning-tree algorithm, and why?
S1

S4

S6

S3

S5

S2

9. Suppose you want to develop a new protocol so that Ethernet switches participating in a VLAN all keep
track of the VLAN color associated with every destination. Assume that each switch knows which of its
ports (interfaces) connect to other switches and which may connect to hosts, and in the latter case knows the
color assigned to that port.

(a). Suggest a way by which switches might propagate this destination-color information to other switches.
(b). What happens if a port formerly reserved for connection to another switch is now used for a host?

54

2 Ethernet

3 OTHER LANS

In the wired era, one could get along quite well with nothing but Ethernet and the occasional long-haul pointto-point link joining different sites. However, there are important alternatives out there. Some, like token
ring, are mostly of historical importance; others, like virtual circuits, are of great conceptual importance but
so far of only modest day-to-day significance.
And then there is wireless. It would be difficult to imagine contemporary laptop networking, let alone mobile
devices, without it. In both homes and offices, Wi-Fi connectivity is the norm. A return to being tethered by
wires is almost unthinkable.

3.1 Virtual Private Network

Suppose you want to connect to your workplace network from home. Your workplace, however, has a
security policy that does not allow outside IP addresses to access essential internal resources. How do you
proceed, without leasing a dedicated telecommunications line to your workplace?
A virtual private network, or VPN, provides a solution; it supports creation of virtual links that join farflung nodes via the Internet. Your home computer creates an ordinary Internet connection (TCP or UDP)
to a workplace VPN server (IP-layer packet encapsulation can also be used; see 7.11 Mobile IP). Each
end of the connection is associated with a software-created virtual network interface; each of the two
virtual interfaces is assigned an IP address. When a packet is to be sent along the virtual link, it is actually
encapsulated and sent along the original Internet connection to the VPN server, wending its way through the
commodity Internet; this process is called tunneling. To all intents and purposes, the virtual link behaves
like any other physical link.
Tunneled packets are usually encrypted as well as encapsulated, though that is a separate issue. One example
of a tunneling protocol is to treat a TCP home-workplace connection as a serial line and send packets over
it back-to-back, using PPP with HDLC; see [Link] HDLC and RFC 1661.
At the workplace side, the virtual network interface in the VPN server is attached to a router or switch; at
the home users end, the virtual network interface can now be assigned an internal workplace IP address.
The home computer is now, for all intents and purposes, part of the internal workplace network.
In the diagram below, the users regular Internet connection is via hardware interface eth0. A connection is
established to Site As VPN server; a virtual interface tun0 is created on the users machine which appears
to be a direct link to the VPN server. The tun0 interface is assigned a Site-A IP address. Packets sent via
the tun0 interface in fact travel over the original connection via eth0 and the Internet.

55

An Introduction to Computer Networks, Release 1.0

After the VPN is set up, the home hosts tun0 interface appears to be locally connected to Site A, and thus
the home host is allowed to connect to the private area within Site A. The home hosts forwarding table will
be configured so that traffic to Site As private addresses is routed via interface tun0.
VPNs are also commonly used to connect entire remote offices to headquarters. In this case the remote-office
end of the tunnel will be at that offices local router, and the tunnel will carry traffic for all the workstations
in the remote office.
To improve security, it is common for the residential (or remote-office) end of the VPN connection to use
the VPN connection as the default route for all traffic except that needed to maintain the VPN itself. This
may require a so-called host-specific forwarding-table entry at the residential end to allow the packets that
carry the VPN tunnel traffic to be routed correctly via eth0. This routing strategy means that potential
intruders cannot access the residential host and thus the workplace internal network through the original
residential Internet access. A consequence is that if the home worker downloads a large file from a nonworkplace site, it will travel first to the workplace, then back out to the Internet via the VPN connection, and
finally arrive at the home.

3.2 Carrier Ethernet

Carrier Ethernet is a leased-line point-to-point link between two sites, where the subscriber interface at each
end of the line looks like Ethernet (in some flavor). The physical path in between sites, however, need not
have anything to do with Ethernet; it may be implemented however the carrier wishes. In particular, it will
be (or at least appear to be) full-duplex, it will be collision-free, and its length may far exceed the maximum
permitted by any IEEE Ethernet standard.
Bandwidth can be purchased in whatever increments the carrier has implemented. The point of carrier
Ethernet is to provide a layer of abstraction between the customers, who need only install a commodity
Ethernet interface, and the provider, who can upgrade the link implementation at will without requiring
change at the customer end.
In a sense, carrier Ethernet is similar to the widespread practice of provisioning residential DSL and cable
routers with an Ethernet interface for customer interconnection; again, the actual link technologies may not
56

3 Other LANs

An Introduction to Computer Networks, Release 1.0

look anything like Ethernet, but the interface will.

A carrier Ethernet connection looks like a virtual VPN link, but runs on top of the providers internal network
rather than the Internet at large. Carrier Ethernet connections often provide the primary Internet connectivity
for one endpoint, unlike Internet VPNs which assume both endpoints already have full Internet connectivity.

3.3 Wi-Fi
Wi-Fi is a trademark denoting any of several IEEE wireless-networking protocols in the 802.11 family,
specifically 802.11a, 802.11b, 802.11g, 802.11n, and 802.11ac. Like classic Ethernet, Wi-Fi must deal
with collisions; unlike Ethernet, however, Wi-Fi is unable to detect collisions in progress, complicating the
backoff and retransmission algorithms. Wi-Fi is designed to interoperate freely with Ethernet at the logical
LAN layer; that is, Ethernet and Wi-Fi traffic can be freely switched from the wired side to the wireless side.
Band Width
To radio engineers, band width means the frequency range used by a signal, not data rate; in keeping
with this we will in this section and 3.4 WiMAX use the term data rate instead of bandwidth. We
will use the terms channel width or width of the frequency band for the frequency range. All else
being equal, the data rate achievable with a radio signal is proportional to the channel width.
Generally, Wi-Fi uses the 2.4 GHz ISM (Industrial, Scientific and Medical) band used also by microwave
ovens, though 802.11a uses a 5 GHz band, 802.11n supports that as an option and the new 802.11ac has
returned to using 5 GHz exclusively. The 5 GHz band has reduced ability to penetrate walls, often resulting
in a lower effective range. Wi-Fi radio spectrum is usually unlicensed, meaning that no special permission
is needed to transmit but also that others may be trying to use the same frequency band simultaneously; the
availability of unlicensed channels in the 5 GHz band continues to evolve.
The table below summarizes the different Wi-Fi versions. All bit rates assume a single spatial stream;
channel widths are nominal.
IEEE name
802.11a
802.11b
802.11g
802.11n
802.11ac

maximum bit rate

54 Mbps
11 Mbps
54 Mbps
65-150 Mbps
78-867 Mbps

frequency
5 GHz
2.4 GHz
2.4 GHz
2.4/5 GHz
5 GHz

channel width
20 MHz
20 MHz
20 MHz
20-40 MHz
20-160 MHz

The maximum bit rate is seldom achieved in practice. The effective bit rate must take into account, at a
minimum, the time spent in the collision-handling mechanism. More significantly, all the Wi-Fi variants
above use dynamic rate scaling, below; the bit rate is reduced up to tenfold (or more) in environments with
higher error rates, which can be due to distance, obstructions, competing transmissions or radio noise. All
this means that, as a practical matter, getting 150 Mbps out of 802.11n requires optimum circumstances; in
particular, no competing senders and unimpeded line-of-sight transmission. 802.11n lower-end performance
can be as little as 10 Mbps, though 40-100 Mbps (for a 40 MHz channel) may be more typical.
The 2.4 GHz ISM band is divided by international agreement into up to 14 officially designated channels,
each about 5 MHz wide, though in the United States use may be limited to the first 11 channels. The 5 GHz

3.3 Wi-Fi

57

An Introduction to Computer Networks, Release 1.0

band is similarly divided into 5 MHz channels. One Wi-Fi sender, however, needs several of these official
channels; the typical 2.4 GHz 802.11g transmitter uses an actual frequency range of up to 22 MHz, or up
to five channels. As a result, to avoid signal overlap Wi-Fi use in the 2.4 GHz band is often restricted to
official channels 1, 6 and 11. The end result is that unrelated Wi-Fi transmitters can and do interact with and
interfere with each other.
The United States requires users of the 5 GHz band to avoid interfering with weather and military applications in the same frequency range. Once that is implemented, however, there are more 5 MHz channels at
this frequency than in the 2.4 GHz ISM band, which is one of the reasons 802.11ac can run faster (below).
Wi-Fi designers can improve speed through a variety of techniques, including
improved radio modulation techniques
improved error-correcting codes
smaller guard intervals between symbols
increasing the channel width
allowing multiple spatial streams via multiple antennas
The first two in this list seem by now to be largely tapped out; the third reduces the range but may increase
the data rate by 11%.
The largest speed increases are obtained by increasing the number of 5 MHz channels used. For example,
the 65 Mbps bit rate above for 802.11n is for a nominal frequency range of 20 MHz, comparable to that
of 802.11g. However, in areas with minimal competition from other signals, 802.11n supports using a 40
MHz frequency band; the bit rate then goes up to 135 Mbps (150 Mbps with a smaller guard interval). This
amounts to using two of the three available 2.4 GHz Wi-Fi bands. Similarly, the wide range in 802.11ac
bit rates reflects support for using channel widths ranging from 20 MHz up to 160 MHz (32 5-MHz official
channels).
For all the categories in the table above, additional bits are used for error-correcting codes. For 802.11g
operating at 54 Mbps, for example, the actual raw bit rate is (4/3)54 = 72 Mbps, sent in symbols consisting
of six bits as a unit.

3.3.1 Multiple Spatial Streams

The latest innovation in improving Wi-Fi data rates is to support multiple spatial streams, through an antenna
technique known as multiple-input-multiple output, or MIMO. To use N streams, both sender and receiver
must have N antennas; all the antennas use the same frequency channels but each transmitter antenna sends
a different data stream. While the antennas are each more-or-less omnidirectional, careful signal analysis
at the receiver can in principle recover each data stream. In practice, overall data-rate improvement over a
single antenna can be considerably less than N.
The 802.11n standard allows for up to four spatial streams, for a theoretical maximum bit rate of 600 Mbps.
802.11ac allows for up to eight spatial streams, for an even-more-theoretical maximum of close to 7 Gbps.
MIMO support is sometimes described with an ABC notation, eg 332, where A and B are the number
of transmitting and receiving antennas and C min(A,B) is the number of spatial streams.

58

3 Other LANs

An Introduction to Computer Networks, Release 1.0

3.3.2 Wi-Fi and Collisions

We looked extensively at the 10 Mbps Ethernet collision-handling mechanisms in 2.1 10-Mbps classic
Ethernet, only to conclude that with switches and full-duplex links, Ethernet collisions are rapidly becoming
a thing of the past. Wi-Fi, however, has brought collisions back from obscurity. While there is a largelycollision-free mode for Wi-Fi operation (3.3.7 Wi-Fi Polling Mode), it is not commonly used, and collision
management has a significant impact on ordinary Wi-Fi performance.
Wi-Fi, like almost any other wireless protocol, simply cannot detect collisions in progress. This has to do
with the relative signal strength of the remote signal at the local transmitter. Along a wire-based Ethernet
the remote signal might be as weak as 1/100 of the transmitted signal but that 1% received signal is still
detectable during transmission. However, with radio the remote signal might easily be as little as 1/100,000
of the transmitted signal, and it is simply overwhelmed during transmission. Thus, while the Ethernet
algorithm was known as CSMA/CD, where CD stood for Collision Detection, Wi-Fi uses CSMA/CA, where
CA stands for Collision Avoidance.
The basic collision-avoidance algorithm is as follows. There are three parameters applicable here (values
are for 802.11b/g in the 2.4 GHz band); the value we call IFS is more formally known as DIFS (D for
distributed; see 3.3.7 Wi-Fi Polling Mode).
slot time: 20 sec
IFS, the normal InterFrame Spacing: 50 sec
SIFS, the short IFS: 10 sec
A sender wanting to send a new data packet waits the IFS time after first sensing the medium to see if it is
idle. If no other traffic is seen in this interval, the station may then transmit immediately. However, if other
traffic is sensed, the sender waits for the other traffic to finish and then for one IFS time after. If the station
had to wait for other traffic, even briefly, it must then do an exponential backoff even for its first transmission
attempt; other stations, after all, are likely also waiting, and avoiding an initial collision is strongly preferred.
The initial backoff is to choose a random k<25 = 32 (recall that classic Ethernet in effect chooses an initial
backoff of k<20 = 1; ie k=0). The prospective sender then waits k slot times. While waiting, the sender
continues to monitor for other traffic; if any other transmission is detected, then the sender suspends the
backoff-wait clock. The clock resumes when the other transmission has completed and one followup idle
interval of length IFS has elapsed. Under this rule, the backoff-wait clock can expire only when the clock is
running and therefore the medium is idle, at which point the transmission begins.
If there is a collision, the station retries, after doubling the backoff range to 64, then 128, 256, 512, 1024
and again 1024. If these seven attempts all fail, the packet is discarded and the sender starts over.
In one slot time, radio signals move 6,000 meters; the Wi-Fi slot time unlike that for Ethernet has nothing
to do with the physical diameter of the network. As with Ethernet, though, the Wi-Fi slot time represents
the fundamental unit for backoff intervals.
Finally, we note that, unlike Ethernet collisions, Wi-Fi collisions are a local phenomenon: if A and B
transmit simultaneously, a collision occurs at node C only if the signals of A and B are both strong enough
at C to interfere with one another. It is possible that a collision occurs at station C midway between A and
B, but not at station D that is close to A. We return to this in [Link] Hidden-Node Problem.
Because Wi-Fi cannot detect collisions directly, the protocol adds link-layer ACK packets (unrelated to
the later TCP ACK), at least for unicast transmission. Once a station has received a packet addressed to it,
3.3 Wi-Fi

59

An Introduction to Computer Networks, Release 1.0

it waits for time SIFS and sends the ACK; at the instant when the end of the SIFS interval is reached, the
receiver will be the only station authorized to send. Any other stations waiting the longer IFS period will see
the ACK before the IFS time has elapsed and will thus not interfere with the ACK; similarly, any stations
with a running backoff-wait clock will continue to have that clock suspended.
[Link] Wi-Fi RTS/CTS
Wi-Fi stations optionally also use a request-to-send/clear-to-send (RTS/CTS) protocol. Usually this is
used only for larger packets; often, the RTS/CTS threshold (the size of the largest packet not sent using RTS/CTS) is set (as part of the Access Point configuration) to be the maximum packet size, effectively
disabling this feature. The idea here is that a large packet that is involved in a collision represents a significant waste of potential throughput; for large packets, we should ask first.
The RTS packet which is small is sent through the normal procedure outlined above; this packet includes
the identity of the destination and the size of the data packet the station desires to transmit. The destination
station then replies with CTS after the SIFS wait period, effectively preventing any other transmission after
the RTS. The CTS packet also contains the data-packet size. The original sender then waits for SIFS after
receiving the CTS, and sends the packet. If all other stations can hear both the RTS and CTS messages, then
once the RTS and CTS are sent successfully no collisions should occur during packet transmission, again
because the only idle times are of length SIFS and other stations should be waiting for time IFS.
[Link] Hidden-Node Problem
Consider the diagram below. Each station has a 100-meter range. Stations A and B are 150 meters apart and
so cannot hear one another at all; each is 75 meters from C. If A is transmitting and B senses the medium in
preparation for its own transmission, as part of collision avoidance, then B will conclude that the medium is
idle and will go ahead and send.

However, C is within range of both A and B. If A and B transmit simultaneously, then from Cs perspective
a collision occurs. C receives nothing usable. We will call this a hidden-node collision as the senders A
and B are hidden from one another; the general scenario is known as the hidden-node problem.
Note that node D receives only As signal, and so no collision occurs at D.

60

3 Other LANs

An Introduction to Computer Networks, Release 1.0

The hidden-node problem can also occur if A and B cannot receive one anothers transmissions due to a
physical obstruction such as a radio-impermeable wall:

One of the rationales for the RTS/CTS protocol is the prevention of hidden-node collisions. Imagine that,
instead of transmitting its data packet, A sends an RTS packet, and C responds with CTS. B has not heard
the RTS packet from A, but does hear the CTS from C. A will begin transmitting after a SIFS interval, but B
will not hear As transmission. However, B will still wait, because the CTS packet contained the data-packet
size and thus, implicitly, the length of time all other stations should remain idle. Because RTS packets are
quite short, they are much less likely to be involved in collisions themselves than data packets.
[Link] Wi-Fi Fragmentation
Conceptually related to RTS/CTS is Wi-Fi fragmentation. If error rates or collision rates are high, a sender
can send a large packet as multiple fragments, each receiving its own link-layer ACK. As we shall see in
5.3.1 Error Rates and Packet Size, if bit-error rates are high then sending several smaller packets often
leads to fewer total transmitted bytes than sending the same data as one large packet.
Wi-Fi packet fragments are reassembled by the receiving node, which may or may not be the final destination.
As with the RTS/CTS threshold, the fragmentation threshold is often set to the size of the maximum packet.
Adjusting the values of these thresholds is seldom necessary, though might be appropriate if monitoring
revealed high collision or error rates. Unfortunately, it is essentially impossible for an individual station
to distinguish between reception errors caused by collisions and reception errors caused by other forms of
noise, and so it is hard to use reception statistics to distinguish between a need for RTS/CTS and a need for
fragmentation.

3.3.3 Dynamic Rate Scaling

Wi-Fi senders, if they detect transmission problems, are able to reduce their transmission bit rate in a process
known as rate scaling or rate control. The idea is that lower bit rates will have fewer noise-related errors,
and so as the error rate becomes unacceptably high perhaps due to increased distance the sender should
fall back to a lower bit rate. For 802.11g, the standard rates are 54, 48, 36, 24, 18, 12, 9 and 6 Mbps. Senders
attempt to find the transmission rate that maximizes throughput; for example, 36 Mbps with a packet loss
rate of 25% has an effective throughput of 36 75% = 27 Mbps, and so is better than 24 Mbps with no
losses.
Senders may update their bit rate on a per-packet basis; senders may also choose different bit rates for
different recipients. For example, if a sender sends a packet and receives no confirming link-layer ACK, the

3.3 Wi-Fi

61

An Introduction to Computer Networks, Release 1.0

sender may fall back to the next lower bit rate. The actual bit-rate-selection algorithm lives in the particular
Wi-Fi driver in use; different nodes in a network may use different algorithms.
The earliest rate-scaling algorithm was Automatic Rate Fallback, or ARF, [KM97]. The rate decreases after
two consecutive transmission failures (that is, the link-layer ACK is not received), and increases after ten
transmission successes.
A significant problem for rate scaling is that a packet loss may be due either to low-level random noise
(white noise, or thermal noise) or to a collision (which is also a form of noise, but less random); only in the
first case is a lower transmission rate likely to be helpful. If a larger number of collisions is experienced, the
longer packet-transmission times caused by the lower bit rate may increase the frequency of hidden-node
collisions. In fact, a higher transmission rate (leading to shorter transmission times) may help; enabling the
RTS/CTS protocol may also help.
Signal Strength
Most Wi-Fi drivers report the received signal strength. Newer drivers use the IEEE Received Channel
Power Indicator convention; the RCPI is an 8-bit integer proportional to the absolute power received
by the antenna as measured in decibel-milliwatts (dBm). Wi-Fi values range from -10 dBm to -90 dBm
and below. For comparison, the light from the star Polaris delivers about -97 dBm to one eye on a good
night; Venus typically delivers about -73 dBm. A GPS satellite might deliver -127 dBm to your phone.
(Inspired by Wikipedia on DBm.)
A variety of newer rate-scaling algorithms have been proposed; see [JB05] for a summary. One, ReceiverBased Auto Rate (RBAR, [HVB01]), attempts to incorporate the signal-to-noise ratio into the calculation of
the transmission rate. This avoids the confusion introduced by collisions. Unfortunately, while the signalto-noise ratio has a strong theoretical correlation with the transmission bit-error rate, most Wi-Fi radios
will report to the host system the received signal strength. This is not the same as the signal-to-noise ratio,
which is harder to measure. As a result, the RBAR approach has not been quite as effective in practice as
might be hoped.
The Collision-Aware Rate Adaptation algorithm (CARA, [KKCQ06]) attempts (among other things) to infer
that a packet was lost to a collision rather than noise if, after one SIFS interval following the end of the packet
transmission, no link-layer ACK has been received and the channel is still busy. This will detect collisions,
of course, only with longer packets.
Because the actual data in a Wi-Fi packet may be sent at a rate not every participant is close enough to
receive correctly, every Wi-Fi transmission begins with a brief preamble at the minimum bit rate. Link-layer
ACKs, too, are sent at the minimum bit rate.

3.3.4 Wi-Fi Configurations

There are two common Wi-Fi configurations: infrastructure and ad hoc. The latter includes individual
Wi-Fi-equipped nodes communicating informally; for example, two laptops can set up an ad hoc connection
to transfer data at a meeting. Ad hoc connections are often used for very simple networks not providing
Internet connectivity. Complex ad hoc networks are, however, certainly possible; see 3.3.8 MANETs.
The more-common infrastructure configuration involves designated access points to which individual WiFi stations must associate before general communication can begin. The association process carried
62

3 Other LANs

An Introduction to Computer Networks, Release 1.0

out by an exchange of special management packets may be restricted to stations with hardware (LAN)
addresses on a predetermined list, or to stations with valid cryptographic credentials. Stations may regularly
re-associate to their Access Point, especially if they wish to communicate some status update.
Access Points
Generally, a Wi-Fi access point has special features; Wi-Fi-enabled station devices like phones and
workstations do not act as access points. However, it may be possible to for a station device to become
an access point if the access-point mode is supported by the underlying radio hardware and if suitable
drivers can be found. Under linux, the hostapd package is one option.
Stations in an infrastructure network communicate directly only with their access point. If B and C share
access point A, and B wishes to send a packet to C, then B first forwards the packet to A and A then forwards
it to C. While this introduces a degree of inefficiency, it does mean that the access point and its associated
nodes automatically act as a true LAN: every node can reach every other node. In an ad hoc network, by
comparison, it is quite common for two nodes to be able to reach each other only by forwarding through an
intermediate third node; this is in fact exactly the hidden-node scenario.
Finally, Wi-Fi is by design completely interoperable with Ethernet; if station A is associated with access
point AP, and AP also connects via (cabled) Ethernet to station B, then if A wants to send a packet to B it
sends it using AP as the Wi-Fi destination but with B also included in the header as the actual destination.
Once it receives the packet by wireless, AP acts as an Ethernet switch and forwards the packet to B.
While this forwarding is transparent to senders, the Ethernet and Wi-Fi LAN header formats are entirely
different.

The above diagram illustrates an Ethernet header and the Wi-Fi header for a typical data packet not using
Wi-Fi quality-of-service features. The Ethernet type field usually moves to an IEEE Logical Link Control
header in the Wi-Fi region labeled data. The receiver and transmitter addresses are the MAC addresses
of the nodes receiving and transmitting the (unicast) packet; these may each be different from the ultimate
destination and source addresses. In infrastructure mode one of the receiver or transmitter addresses is the
access point; in typical situations either the receiver is the destination or the sender is the transmitter.

3.3.5 Wi-Fi Roaming

Wi-Fi access points are generally identified by their SSID, an administratively defined string such as
linksys or loyola. These are periodically broadcast by the access point in special beacon packets (though

3.3 Wi-Fi

63

An Introduction to Computer Networks, Release 1.0

for pseudo-security reasons beacon packets can be suppressed). Large installations can create roaming access among multiple access points by assigning all the access points the same SSID. An individual station
will stay with the access point with which it originally associated until the signal strength falls below a certain level, at which point it will seek out other access points with the same SSID and with a stronger signal.
In this way, a large area can be carpeted with multiple Wi-Fi access points, so as to look like one large Wi-Fi
domain.
In order for this to work, traffic to wireless node B must find Bs current access point AP. This is done in
much the same way as, in a wired Ethernet, traffic finds a laptop that has been unplugged, carried to a new
building, and plugged in again. The distribution network is the underlying wired network (eg Ethernet) to
which all the access points connect. If the distribution network is a switched Ethernet supporting the usual
learning mechanism (2.4 Ethernet Switches), then Wi-Fi location update is straightforward. Suppose B
is a wireless node that has been exchanging packets via the distribution network with C (perhaps a router
connecting B to the Internet). When B moves to a new access point, all it has to do is send any packet over
the LAN to C, and the Ethernet switches involved will then learn the route through the switched Ethernet
from C to Bs current AP, and thus to B.
This process may leave other switches not currently communicating with B still holding in their forwarding tables the old location for B. This is not terribly serious, but can be avoided entirely if, after moving, B
sends out an Ethernet broadcast packet.
Ad hoc networks also have SSIDs; these are generated pseudorandomly at startup. Ad hoc networks have
beacon packets as well; all nodes participate in the regular transmission of these via a distributed algorithm.

3.3.6 Wi-Fi Security

Wi-Fi traffic is visible to anyone nearby with an appropriate receiver; this eavesdropping zone can be expanded by use of a larger antenna. Because of this, Wi-Fi security is important, and Wi-Fi supports several
types of traffic encryption.
The original Wired-Equivalent Privacy, or WEP, encryption standard, contained a fatal (and now-classic)
flaw. Bytes of the key could be could be broken that is, guessed sequentially. Knowing bytes 0 through
i1 would allow an attacker to guess byte i with a relatively small amount of data, and so on through the
entire key. See 21.6.8 Wi-Fi WEP Encryption Failure for details.
WEP has been replaced with Wi-Fi Protected Access, or WPA; the current version is WPA2. WPA2 encryption is believed to be quite secure, although there was a vulnerability in the associated Wi-Fi Protected
Setup protocol.
Key management, however, can be a problem. Many smaller sites use pre-shared key mode, known as
WPA-Personal, in which a single key is entered into the Access Point (ideally not over the air) and into
each connecting station. The key is usually a hash of a somewhat longer passphrase. The use of a common
key for multiple stations makes changing the key, or revoking the key for a particular user, difficult.
A more secure approach is WPA-Enterprise; this allows each station to have a separate key, and also for
each station key to be recognized by all access points. It is part of the IEEE 802.1X security framework
(technically so is WPA-Personal, though a much smaller part). In this model, the client node is known as
the supplicant, the Access Point is known as the authenticator, and there is a third system involved known
as the authentication server. One authentication server can support multiple access points.

64

3 Other LANs

An Introduction to Computer Networks, Release 1.0

To begin the association process, the supplicant contacts the authenticator using the Extensible Authentication Protocol, or EAP, with what amounts to a request to associate to that access point. EAP is a generic
message framework meant to support multiple specific types of authentication; see RFC 3748 and RFC
5247. The EAP request is forwarded to an authentication server, which may exchange (via the authenticator) several challenge/response messages with the supplicant. EAP is usually used in conjunction with
the RADIUS (Remote Authentication Dial-In User Service) protocol (RFC 2865), which is a specific (but
flexible) authentication-server protocol. WPA-Enterprise is sometimes known as 802.1X mode, EAP mode
or RADIUS mode.
One peculiarity of EAP is that EAP communication takes place before the supplicant is given an IP address
(in fact before the supplicant has completed associating itself to the access point); thus, a mechanism must
be provided to support exchange of EAP packets between supplicant and authenticator. This mechanism
is known as EAPOL, for EAP Over LAN. EAP messages between the authenticator and the authentication
server, on the other hand, can travel via IP; in fact, sites may choose to have the authentication server hosted
remotely.
Once the authentication server (eg RADIUS server) is set up, specific per-user authentication methods can
be entered. This can amount to username,password pairs, or some form of security certificate, or often
both. The authentication server will generally allow different encryption protocols to be used for different
supplicants, thus allowing for the possibility that there is not a common protocol supported by all stations.
When this authentication strategy is used, the access point no longer needs to know anything about what
authentication protocol is actually used; it is simply the middleman forwarding EAP packets between the
supplicant and the authentication server. The access point allows the supplicant to associate into the network
once it receives permission to do so from the authentication server.

3.3.7 Wi-Fi Polling Mode

Wi-Fi also includes a polled mechanism, where one station (the Access Point) determines which stations
are allowed to send. While it is not often used, it has the potential to greatly reduce collisions, or even
eliminate them entirely. This mechanism is known as Point Coordination Function, or PCF, versus the
collision-oriented mechanism which is then known as Distributed Coordination Function. The PCF name
refers to the fact that in this mode it is the Access Point that is in charge of coordinating which stations get
to send when.
The PCF option offers the potential for regular traffic to receive improved throughput due to fewer collisions.
However, it is often seen as intended for real-time Wi-Fi traffic, such as voice calls over Wi-Fi.
The idea behind PCF is to schedule, at regular intervals, a contention-free period, or CFP. During this
period, the Access Point may
send Data packets to any receiver
send Poll packets to any receiver, allowing that receiver to reply with its own data packet
send a combination of the two above (not necessarily to the same receiver)
send management packets, including a special packet marking the end of the CFP
None of these operations can result in a collision (unless an unrelated but overlapping Wi-Fi domain is
involved).

3.3 Wi-Fi

65

An Introduction to Computer Networks, Release 1.0

Stations receiving data from the Access Point send the usual ACK after a SIFS interval. A data packet from
the Access Point addressed to station B may also carry, piggybacked in the Wi-Fi header, a Poll request to
another station C; this saves a transmission. Polled stations that send data will receive an ACK from the
Access Point; this ACK may be combined in the same packet with the Poll request to the next station.
At the end of the CFP, the regular contention period or CP resumes, with the usual CSMA/CA strategy.
The time interval between the start times of consecutive CFP periods is typically 100 ms, short enough to
allow some real-time traffic to be supported.
During the CFP, all stations normally wait only the Short IFS, SIFS, between transmissions. This works
because normally there is only one station designated to respond: the Access Point or the polled station.
However, if a station is polled and has nothing to send, the Access Point waits for time interval PIFS (PCF
Inter-Frame Spacing), of length midway between SIFS and IFS above (our previous IFS should now really
be known as DIFS, for DCF IFS). At the expiration of the PIFS, any non-Access-Point station that happens
to be unaware of the CFP will continue to wait the full DIFS, and thus will not transmit. An example of
such a CFP-unaware station might be one that is part of an entirely different but overlapping Wi-Fi network.
The Access Point generally maintains a polling list of stations that wish to be polled during the CFP. Stations
request inclusion on this list by an indication when they associate or (more likely) reassociate to the Access
Point. A polled station with nothing to send simply remains quiet.
PCF mode is not supported by many lower-end Wi-Fi routers, and often goes unused even when it is available. Note that PCF mode is collision-free, so long as no other Wi-Fi access points are active and within
range. While the standard has some provisions for attempting to deal with the presence of other Wi-Fi
networks, these provisions are somewhat imperfect; at a minimum, they are not always supported by other
access points. The end result is that polling is not quite as useful as it might be.

3.3.8 MANETs
The MANET acronym stands for mobile ad hoc network; in practice, the term generally applies to ad hoc
wireless networks of sufficient complexity that some internal routing mechanism is needed to enable full
connectivity. The term mesh network is also used. While MANETs can use any wireless mechanism, we
will assume here that Wi-Fi is used.
MANET nodes communicate by radio signals with a finite range, as in the diagram below.

66

3 Other LANs

An Introduction to Computer Networks, Release 1.0

Each nodes radio range is represented by a circle centered about that node. In general, two MANET nodes
may be able to communicate only by relaying packets through intermediate nodes, as is the case for nodes
A and G in the diagram above.
In the field, the radio range of each node may not be very circular, due to among other things signal reflection and blocking from obstructions. An additional complication arises when the nodes (or even just
obstructions) are moving in real time (hence the mobile of MANET); this means that a working route may
stop working a short time later. For this reason, and others, routing within MANETs is a good deal more
complex than routing in an Ethernet. A switched Ethernet, for example, is required to be loop-free, so there
is never a choice among multiple alternative routes.
Note that, without successful LAN-layer routing, a MANET does not have full node-to-node connectivity
and thus does not meet the definition of a LAN given in 1.9 LANs and Ethernet. With either LAN-layer or
IP-layer routing, one or more MANET nodes may serve as gateways to the Internet.
Note also that MANETs in general do not support broadcast, unless the forwarding of broadcast messages
throughout the MANET is built in to the routing mechanism. This can complicate the assignment of IP
addresses; the common IPv4 mechanism we will describe in 7.8 Dynamic Host Configuration Protocol
(DHCP) relies on broadcast and so usually needs some adaptation.
Finally, we observe that while MANETs are of great theoretical interest, their practical impact has been
modest; they are almost unknown, for example, in corporate environments. They appear most useful in
emergency situations, rural settings, and settings where the conventional infrastructure network has failed
or been disabled.
[Link] Routing in MANETs
Routing in MANETs can be done either at the LAN layer, using physical addresses, or at the IP layer with
some minor bending (below) of the rules.

3.3 Wi-Fi

67

An Introduction to Computer Networks, Release 1.0

Either way, nodes must find out about the existence of other nodes, and appropriate routes must then be
selected. Route selection can use any of the mechanisms we describe later in 9 Routing-Update Algorithms.
Routing at the LAN layer is much like routing by Ethernet switches; each node will construct an appropriate
forwarding table. Unlike Ethernet, however, there may be multiple paths to a destination, direct connectivity
between any particular pair of nodes may come and go, and negotiation may be required even to determine
which MANET nodes will serve as forwarders.
Routing at the IP layer involves the same issues, but at least IP-layer routing-update algorithms have always
been able to handle multiple paths. There are some minor issues, however. When we initially presented
IP forwarding in 1.10 IP - Internet Protocol, we assumed that routers made their decisions by looking
only at the network prefix of the address; if another node had the same network prefix it was assumed to be
reachable directly via the LAN. This model usually fails badly in MANETs, where direct reachability has
nothing to do with addresses. At least within the MANET, then, a modified forwarding algorithm must be
used where every address is looked up in the forwarding table. One simple way to implement this is to have
the forwarding tables contain only host-specific entries as were discussed in 3.1 Virtual Private Network.
Multiple routing algorithms have been proposed for MANETs. Performance of a given algorithm may
depend on the following factors:
The size of the network
Whether some nodes have agreed to serve as routers
The degree of node mobility, especially of routing-node mobility if applicable
Whether the nodes are under common administration, and thus may agree to defer their own transmission interests to the common good
per-node storage and power availability

3.4 WiMAX
WiMAX is a wireless network technology standardized by IEEE 802.16. It supports both stationary subscribers (802.16d) and mobile subscribers (802.16e). The stationary-subscriber version is often used to
provide residential Internet connectivity, in both urban and rural areas. The mobile version is sometimes
referred to as a fourth generation or 4G networking technology; its similar primary competitor is known
as LTE. WiMAX is used in many mobile devices, from smartphones to traditional laptops with wireless
cards installed.
As in the sidebar at the start of 3.3 Wi-Fi we will use the term data rate for what is commonly called
bandwidth to avoid confusion with the radio-specific meaning of the latter term.
WiMAX can use unlicensed frequencies, like Wi-Fi, but its primary use is over licensed radio spectrum.
WiMAX also supports a number of options for the width of its frequency band; the wider the band, the
higher the data rate. Wider bands also allow the opportunity for multiple independent frequency channels.
Downlink (base station to subscriber) data rates can be well over 100 Mbps (uplink rates are usually smaller).
Like Wi-Fi, WiMAX subscriber stations connect to a central access point, though the WiMAX standard
prefers the term base station which we will use henceforth. Stationary-subscriber WiMAX, however, operates on a much larger scale. The coverage radius of a WiMAX base station can be tens of kilometers

68

3 Other LANs

An Introduction to Computer Networks, Release 1.0

if larger antennas are provided, versus less (sometimes much less) than 100 meters for Wi-Fi; mobilesubscriber WiMAX might have a radius of one or two kilometers. Large-radius base stations are typically
mounted in towers. Subscriber stations are not generally expected to be able to hear other stations; they
interact only with the base station. As WiMAX distances increase, the data rate is reduced.
As with Wi-Fi, the central contention problem is how to schedule transmissions of subscriber stations so
they do not overlap; that is, collide. The base station has no difficulty broadcasting transmissions to multiple
different stations sequentially; it is the transmissions of those stations that must be coordinated. Once a
station completes the network entry process to connect to a base station (below), it is assigned regular
(though not necessarily periodic) transmission slots. These transmission slots may vary in size over time;
the base station may regularly issue new transmission schedules.
The centralized assignment of transmission intervals superficially resembles Wi-Fi PCF mode (3.3.7 WiFi Polling Mode); however, assignment is not done through polling, as propagation delays are too large
(below). Instead, each WiMAX subscriber station is told in effect that it may transmit starting at an assigned
time T and for an assigned length L. The station synchronizes its clock with that of the base station as part
of the network entry process.
Because of the long distances involved, synchronization and transmission protocols must take account of
speed-of-light delays. The round-trip delay across 30 km is 200 sec which is ten times larger than the basic
Wi-Fi SIFS interval; at 160 Mbps, this is the time needed to send 4 KB. If a station is to transmit so that its
message arrives at the base station at a certain time, it must actually begin transmission early by an amount
equal to the one-way station-to-base propagation delay; a special ranging mechanism allows stations to
figure out this delay.
A subscriber station begins the network-entry connection process to a base station by listening for the base
stations transmissions (which may be organized into multiple channels); these message streams contain
regular management messages containing, among other things, information about available data rates in
each direction.
Also included in the base stations message stream is information about start times for ranging intervals.
The station waits for one of these intervals and sends a range-request message to the base station. These
ranging intervals are open to all stations attempting network entry, and if another station transmits at the
same time there will be a collision. However, network entry is only done once (for a given base station) and
so the likelihood of a collision in any one ranging interval is small. An Ethernet/Wi-Fi-like exponentialbackoff process is used if a collision does occur. Ranging intervals are the only times when collisions can
occur; afterwards, all station transmissions are scheduled by the base station.
If there is no collision, the base station responds, and the station now knows the propagation delay and thus
can determine when to transmit so that its data arrives at the base station exactly at a specified time. The
station also determines its transmission signal strength from this ranging process.
Finally, and perhaps most importantly, the station receives from the base station its first timeslot for a
scheduled transmission. These timeslot assignments are included in regular uplink-map packets broadcast
by the base station. Each stations timeslot includes both a start time and a total length; lengths are in the
range of 2 to 20 ms. Future timeslots will be allocated as necessary by the base station, in future uplink-map
packets. Scheduled timeslots may be periodic (as is would be appropriate for voice) or may occur at varying
intervals. WiMAX stations may request any of several quality-of-service levels and the base station may
take these requests into account when determining the schedule. The base station also creates a downlink
schedule, but this does not need to be communicated to the subscriber stations; the base station simply uses
it to decide what to broadcast when to the stations. When scheduling the timeslots, the base station may also
3.4 WiMAX

69

An Introduction to Computer Networks, Release 1.0

take into account availability of multiple transmission channels and of directional antennas.
Through the uplink-map schedules and individual ranging, each station transmits so that one transmission
finishes arriving just before the next transmission begins arriving, as seen from the perspective of the base
station. Only minimal guard intervals need be included between consecutive transmissions. Two (or
more) consecutive transmissions may in fact be in the air simultaneously, as far-away stations need to
begin transmitting early so their signals will arrive at the base station at the expected time. The following
diagram illustrates this for stations separated by relatively large physical distances.

Mobile stations will need to update their ranging information regularly, but this can be done through future
scheduled transmissions. The distance to the base station is used not only for the mobile stations transmission timing, but also to determine its power level; signals from each mobile station, no matter where located,
should arrive at the base station with about the same power.
When a station has data to send, it includes in its next scheduled transmission a request for a longer transmission interval; if the request is granted, the station may send the data (or at least some of the data) in its
next scheduled transmission slot. When a station is done transmitting, its timeslot shrinks to the minimum,
and may be scheduled less frequently as well, but it does not disappear. Stations without data to send remain
connected to the base station by sending empty messages during these slots.

3.5 Fixed Wireless

This category includes all wireless-service-provider systems where the subscribers location does not
change. Often, but not always, the subscriber will have an outdoor antenna for improved reception and
range. Fixed-wireless systems can involve relay through satellites, or can be terrestrial.

70

3 Other LANs

An Introduction to Computer Networks, Release 1.0

3.5.1 Terrestrial Wireless

Terrestrial wireless also called terrestrial broadband or fixed-wireless broadband involves direct (nonsatellite) radio communication between subscribers and a central access point. Access points are usually
tower-mounted and serve multiple subscribers, though single-subscriber point-to-point microwave links
also exist. A multi-subscriber access point may serve an area with radius up to several tens of miles, depending on the technology, though more common ranges are under ten miles. WiMAX 802.16d is one form
of terrestrial wireless, but there are several others. Frequencies may be either licensed or unlicensed. Unlicensed frequency bands are available at around 900 MHz, 2.4 GHz, and 5 GHz. Nominally all three bands
require that line-of-sight transmission be used, though that requirement becomes stricter as the frequency
increases. Lower frequencies tend to be better at seeing through trees and other obstructions.

3.5 Fixed Wireless

71

An Introduction to Computer Networks, Release 1.0

Trees vs Signal

Photo of the author attempting to improve his 2.4 GHz terrestrial-wireless signal via tree trimming.
Terrestrial fixed wireless was originally popularized for rural areas, where residential density is too low for
economical cable connections. However, some fixed-wireless ISPs now operate in urban areas, often using
WiMAX. One advantage of terrestrial fixed-wireless in remote areas is that the antennas covers a much
smaller geographical area than a satellite, generally meaning that there is more data bandwidth available per
user and the cost per megabyte is much lower.
Outdoor subscriber antennas often use a parabolic dish to improve reception; sizes range from 10 to 50 cm
in diameter. The size of the dish may depend on the distance to the central tower.
While there are standardized fixed-wireless systems, such as WiMAX, there are also a number of proprietary alternatives, including systems from Trango and Canopy. Fixed-wireless systems might, in fact, be
considered one of the last bastions of proprietary LAN protocols. This lack of standardization is due to a
72

3 Other LANs

An Introduction to Computer Networks, Release 1.0

variety of factors; two primary ones are the relatively modest overall demand for this service and the the fact
that most antennas need to be professionally installed by the ISP to ensure that they are properly mounted,
aligned, grounded and protected from lightning.

3.5.2 Satellite Internet

An extreme case of fixed wireless is satellite Internet, in which signals pass through a satellite in geosynchronous orbit (35,786 km above the earths surface). Residential customers have parabolic antennas typically from 70 to 100 cm in diameter, larger than those used for terrestrial wireless but smaller than the dish
antennas used at access points. Transmitter power is typically 1-2 watts, remarkably low for a signal that
travels 35,786 km.
The primary problem associated with satellite Internet is very long RTTs. The the speed-of-light round-trip
propagation delay is about 500 ms to which must be added queuing delays for the often-backlogged access
point (my own personal experience suggested that RTTs of close to 1,000 ms were the norm). These long
delays affect real-time traffic such as VoIP and gaming, but as we shall see in 14.11 The Satellite-Link TCP
Problem bulk TCP transfers also perform poorly with very long RTTs. To provide partial compensation for
the TCP issue, many satellite ISPs provide some sort of acceleration for bulk downloads: a web page, for
example, would be downloaded rapidly by the access point and streamed to the satellite and back down to
the user via a proprietary mechanism. Acceleration, however, cannot help interactive connections such as
VPNs.
Another common feature of satellite Internet is a low daily utilization cap, typically in the hundreds of
megabytes. Utilization caps are directly tied to the cost of maintaining satellites, but also to the fact that one
satellite covers a great deal of ground, and so its available capacity is shared by a large number of users.

3.6 Token Ring

A significant part of the previous chapter was devoted to classic Ethernets collision mechanism for supporting shared media access. After that, it may come as a surprise that there is a simple multiple-access
mechanism that is not only collision-free, but which supports fairness in the sense that if N stations wish to
send then each will receive 1/N of the opportunities.
That method is Token Ring. Actual implementations come in several forms, from Fiber-Distributed Data
Interface (FDDI) to so-called IBM Token Ring. The central idea is that stations are connected in a ring:

Packets will be transmitted in one direction (clockwise in the ring above). Stations in effect forward most
packets around the ring, although they can also remove a packet. (It is perhaps more accurate to think

3.6 Token Ring

73

An Introduction to Computer Networks, Release 1.0

of the forwarding as representing the default cable connectivity; non-forwarding represents the stations
momentarily breaking that connectivity.)
When the network is idle, all stations agree to forward a special, small packet known as a token. When a
station, say A, wishes to transmit, it must first wait for the token to arrive at A. Instead of forwarding the
token, A then transmits its own packet; this travels around the network and is then removed by A. At that
point (or in some cases at the point when A finishes transmitting its data packet) A then forwards the token.
In a small ring network, the ring circumference may be a small fraction of one packet. Ring networks
become large at the point when some packets may be entirely in transit on the ring. Slightly different
solutions apply in each case. (It is also possible that the physical ring exists only within the token-ring
switch, and that stations are connected to that switch using the usual point-to-point wiring.)
If all stations have packets to send, then we will have something like the following:
A waits for the token
A sends a packet
A sends the token to B
B sends a packet
B sends the token to C
C sends a packet
C sends the token to D
...
All stations get an equal number of chances to transmit, and no bandwidth is wasted on collisions.
One problem with token ring is that when stations are powered off it is essential that the packets continue
forwarding; this is usually addressed by having the default circuit configuration be to keep the loop closed.
Another issue is that some station has to watch out in case the token disappears, or in case a duplicate token
appears.
Because of fairness and the lack of collisions, IBM Token Ring was once considered to be the premium LAN
mechanism. As such, a premium price was charged (there was also the matter of licensing fees). But due
to a combination of lower hardware costs and higher bitrates (even taking collisions into account), Ethernet
eventually won out.
There was also a much earlier collision-free hybrid of 10 Mbps Ethernet and Token Ring known as Token
Bus: an Ethernet physical network (often linear) was used with a token-ring-like protocol layer above
that. Stations were physically connected to the (linear) Ethernet but were assigned identifiers that logically
arranged them in a (virtual) ring. Each station had to wait for the token and only then could transmit a
packet; after that it would send the token on to the next station in the virtual ring. As with real Token
Ring, some mechanisms need to be in place to monitor for token loss.
Token Bus Ethernet never caught on. The additional software complexity was no doubt part of the problem,
but perhaps the real issue was that it was not necessary.

74

3 Other LANs

An Introduction to Computer Networks, Release 1.0

3.7 Virtual Circuits

Before we can get to our final LAN example, ATM, we need to detour briefly through virtual circuits.
Virtual circuits are The Road Not Taken by IP.
Virtual-circuit switching is an alternative to datagram switching, which was introduced in Chapter 1. In
datagram switching, routers know the next_hop to each destination, and packets are addressed by destination. In virtual-circuit switching, routers know about end-to-end connections, and packets are addressed
by a connection ID.
Before any packets can be sent, a connection needs to be established first. For that connection, the route is
computed and then each link along the path is assigned a connection ID, traditionally called the VCI, for
Virtual Circuit Identifier. In most cases, VCIs are only locally unique; that is, the same connection may use
a different VCI on each link. The lack of global uniqueness makes VCI allocation much simpler. Although
the VCI keeps changing along a path, the VCI can still be thought of as identifying the connection. To send
a packet, the host marks the packet with the VCI assigned to the hostrouter1 link.
Packets arrive at (and depart from) switches via one of several ports, which we will assume are numbered
beginning at 0. Switches maintain a connection table indexed by VCI,port pairs; unlike a forwarding
table, the connection table has a record of every connection through that switch at that particular moment.
As a packet arrives, its inbound VCIin and inbound portin are looked up in this table; this yields an outbound
VCIout ,portout pair. The VCI field of the packet is then rewritten to VCIout , and the packet is sent via
portout .
Note that typically there is no source address information included in the packet (although the sender can
be identified from the connection, which can be identified from the VCI at any point along the connection).
Packets are identified by connection, not destination. Any node along the path (including the endpoints) can
in principle look up the connection and figure out the endpoints.
Note also that each switch must rewrite the VCI. Datagram switches never rewrite addresses (though they
do update hopcount/TTL fields). The advantage to this rewriting is that VCIs need be unique only for a
given link, greatly simplifying the naming. Datagram switches also do not make use of a packets arrival
interface.
As an example, consider the network below. Switch ports are numbered 0,1,2,3. Two paths are drawn in,
one from A to F in red and one from B to D in green; each link is labeled with its VCI number in the same
color.
We will construct virtual-circuit connections between
A and F (shown above in red)
A and E
A and C
B and D (shown above in green)
A and F again (a separate connection)
The following VCIs have been chosen for these connections. The choices are made more or less randomly
here, but in accordance with the requirement that they be unique to each link. Because links are generally

3.7 Virtual Circuits

75

An Introduction to Computer Networks, Release 1.0

taken to be bidirectional, a VCI used from S1 to S3 cannot be reused from S3 to S1 until the first connection
closes.
A to F: A 4

S1 6

S2 4 S4 8

S5 5 F; this path goes from S1 to S4 via S2

A to E: A 5 S1 6 S3 3 S4 8 E; this path goes, for no particular reason, from S1 to S4

via S3, the opposite corner of the square
A to C: A 6

S1 7

S3 3

B to D: B 4

S3 8 S1 7

A to F: A 7

S1 8

C
S2 8

S2 5 S4 9

D
S5 2 F

One may verify that on any one link no two different paths use the same VCI.
We now construct the actual VCI,port tables for the switches S1-S4, from the above; the table for S5 is left
as an exercise. Note that either the VCIin ,portin or the VCIout ,portout can be used as the key; we cannot
have the same pair in both the in columns and the out columns. It may help to display the port numbers for
each switch, as in the upper numbers in following diagram of the above red connection from A to F (lower
numbers are the VCIs):

Switch S1:
VCIin
4
5
6
8
7

portin
0
0
0
1
0

VCIout
6
6
7
7
8

portout
2
1
1
2
2

connection
AF #1
AE
AC
BD
AF #2

Switch S2:

76

3 Other LANs

An Introduction to Computer Networks, Release 1.0

VCIin
6
7
8

portin
0
0
0

VCIout
4
8
5

portout
1
2
1

connection
AF #1
BD
AF #2

VCIout
3
3
8

portout
2
1
3

connection
AE
AC
BD

VCIout
8
8
9

portout
2
1
2

connection
AF #1
AE
AF #2

Switch S3:
VCIin
6
7
4

portin
3
3
0

Switch S4:
VCIin
4
3
5

portin
3
0
3

The namespace for VCIs is small, and compact (eg contiguous). Typically the VCI and port bitfields can be
concatenated to produce a VCI,Port composite value small enough that it is suitable for use as an array
index. VCIs work best as local identifiers. IP addresses, on the other hand, need to be globally unique, and
thus are often rather sparsely distributed.
Virtual-circuit switching offers the following advantages:
connections can get quality-of-service guarantees, because the switches are aware of connections and
can reserve capacity at the time the connection is made
headers are smaller, allowing faster throughput
headers are small enough to allow efficient support for the very small packet sizes that are optimal for
voice connections. ATM packets, for instance, have 48 bytes of data; see below.
Datagram forwarding, on the other hand, offers these advantages:
Routers have less state information to manage.
Router crashes and partial connection state loss are not a problem.
If a router or link is disabled, rerouting is easy and does not affect any connection state. (As mentioned
in Chapter 1, this was Paul Barans primary concern in his 1962 paper introducing packet switching.)
Per-connection billing is very difficult.
The last point above may once have been quite important; in the era when the ARPANET was being developed, typical daytime long-distance rates were on the order of $1/minute. It is unlikely that early TCP/IP
protocol development would have been as fertile as it was had participants needed to justify per-minute
billing costs for every project.
It is certainly possible to do virtual-circuit switching with globally unique VCIs say the concatenation of
source and destination IP addresses and port numbers. The IP-based RSVP protocol (19.6 RSVP) does
exactly this. However, the fast-lookup and small-header advantages of a compact namespace are then lost.
Note that virtual-circuit switching does not suffer from the problem of idle channels still consuming resources, which is an issue with circuits using time-division multiplexing (eg shared T1 lines)

3.7 Virtual Circuits

77

An Introduction to Computer Networks, Release 1.0

3.8 Asynchronous Transfer Mode: ATM

ATM is a LAN mechanism intended to accommodate real-time traffic as well as bulk data transfer. It was
particularly intended to support voice. A significant source of delay in voice traffic is the packet fill time: at
DS0 speeds, a voice packet fills at 8 bytes/ms. If we are sending 1KB packets, this means voice is delayed
by about 1/8 second, meaning in turn that when one person stops speaking, the earliest they can hear the
others response is 1/4 second later. Voice delay also can introduce an annoying echo. When voice is sent
over IP (VoIP), one common method is to send 160 bytes every 20 ms.
ATM took this small-packet strategy even further: packets have 48 bytes of data, plus 5 bytes of header. Such
small packets are often called cells. To manage such a small header, virtual-circuit routing is a necessity. IP
packets of such small size would likely consume more than 50% of the bandwidth on headers, if the LAN
header were included.
Aside from reduced voice fill-time, other benefits to small cells are reduced store-and-forward delay and
minimal queuing delay, at least for high-priority traffic. Prioritizing traffic and giving precedence to highpriority traffic is standard, but high-priority traffic is never allowed to interrupt transmission already begun
of a low-priority packet. If you have a high-priority voice cell, and someone else has a 1500-byte packet
just started, your cell has to wait about 30 cell times, because 1500 bytes is about 30 cells. However, if their
low-priority traffic is instead made up of 30 cells, you have only to wait for their first cell to finish; the delay
is 1/30 as much.
ATM also made the decision to require fixed-size cells. The penalty for one partially used cell among many
is small. Having a fixed cell size simplifies hardware design, and, in theory, allows it easier to design for
parallelism.
Unfortunately, ATM also chose to mandate no cell reordering. This means cells can use a smaller sequencenumber field, but also makes parallel switches much harder to build. A typical parallel switch design might
involve forwarding incoming cells to any of several input queues; the queues would then handle the VCI
lookups in parallel and forward the cells to the appropriate output queues. With such an architecture, avoiding reordering is difficult. It is not clear to what extent the no-reordering decision was related to the later
decline of ATM in the marketplace.
ATM cells have 48 bytes of data and a 5-byte header. The header contains up to 28 bits of VCI information,
three type bits, one cell-loss priority, or CLP, bit, and an 8-bit checksum over the header only. The
VCI is divided into 8-12 bits of Virtual Path Identifier and 16 bits of Virtual Channel Identifier, the latter
supposedly for customer use to separate out multiple connections between two endpoints. Forwarding is by
full switching only, and there is no mechanism for physical (LAN) broadcast.

3.8.1 ATM Segmentation and Reassembly

Due to the small packet size, ATM defines its own mechanisms for segmentation and reassembly of larger
packets. Thus, individual ATM links in an IP network are quite practical. These mechanisms are called
ATM Adaptation Layers, and there are four of them: AALs 1, 2, 3/4 and 5 (AAL 3 and AAL 4 were once
separate layers, which merged). AALs 1 and 2 are used only for voice-type traffic; we will not consider
them further.
The ATM segmentation-and-reassembly mechanism defined here is intended to apply only to large data; no
cells are ever further subdivided. Furthermore, segmentation is always applied at the point where the data

78

3 Other LANs

An Introduction to Computer Networks, Release 1.0

enters the network; reassembly is done at exit from the ATM path. IPv4 fragmentation, on the other hand,
applies conceptually to IP packets, and may be performed by routers within the network.
For AAL 3/4, we first define a high-level wrapper for an IP packet, called the CS-PDU (Convergence
Sublayer - Protocol Data Unit). This prefixes 32 bits on the front and another 32 bits (plus padding) on the
rear. We then chop this into as many 44-byte chunks as are needed; each chunk goes into a 48-byte ATM
payload, along with the following 32 bits worth of additional header/trailer:
2-bit type field:
10: begin new CS-PDU
00: continue CS-PDU
01: end of CS-PDU
11: single-segment CS-PDU
4-bit sequence number, 0-15, good for catching up to 15 dropped cells
10-bit MessageID field
CRC-10 checksum.
We now have a total of 9 bytes of header for 44 bytes of data; this is more than 20% overhead. This did not
sit well with the IP-over-ATM community (such as it was), and so AAL 5 was developed.
AAL 5 moved the checksum to the CS-PDU and increased it to 32 bits from 10 bits. The MID field was
discarded, as no one used it, anyway (if you wanted to send several different types of messages, you simply
created several virtual circuits). A bit from the ATM header was taken over and used to indicate:
1: start of new CS-PDU
0: continuation of an existing CS-PDU
The CS-PDU is now chopped into 48-byte chunks, which are then used as the entire body of each ATM
cell. With 5 bytes of header for 48 bytes of data, overhead is down to 10%. Errors are detected by the
CS-PDU CRC-32. This also detects lost cells (impossible with a per-cell CRC!), as we no longer have any
cell sequence number.
For both AAL3/4 and AAL5, reassembly is simply a matter of stringing together consecutive cells in order
of arrival, starting a new CS-PDU whenever the appropriate bits indicate this. For AAL3/4 the receiver
has to strip off the 4-byte AAL3/4 headers; for AAL5 the receiver has to verify the CRC-32 checksum once
all cells are received. Different cells from different virtual circuits can be jumbled together on the ATM
backbone, but on any one virtual circuit the cells from one higher-level packet must be sent one right after
the other.
A typical IP packet divides into about 20 cells. For AAL 3/4, this means a total of 200 bits devoted to CRC
codes, versus only 32 bits for AAL 5. It might seem that AAL 3/4 would be more reliable because of this,
but, paradoxically, it was not! The reason for this is that errors are rare, and so we typically have one or at
most two per CS-PDU. Suppose we have only a single error, ie a single cluster of corrupted bits small enough
that it is likely confined to a single cell. In AAL 3/4 the CRC-10 checksum will fail to detect that error (that
is, the checksum of the corrupted packet will by chance happen to equal the checksum of the original packet)
with probability 1/210 . The AAL 5 CRC-32 checksum, however, will fail to detect the error with probability
1/232 . Even if there are enough errors that two cells are corrupted, the two CRC-10s together will fail to

3.8 Asynchronous Transfer Mode: ATM

79

An Introduction to Computer Networks, Release 1.0

detect the error with probability 1/220 ; the CRC-32 is better. AAL 3/4 is more reliable only when we have
errors in at least four cells, at which point we might do better to switch to an error-correcting code.
Moral: one checksum over the entire message is often better than multiple shorter checksums over parts of
the message.

3.9 Epilog
Along with a few niche protocols, we have focused primarily here on wireless and on virtual circuits. Wireless, of course, is enormously important: it is the enabler for mobile devices, and has largely replaced
traditional Ethernet for home and office workstations.
While it is sometimes tempting (in the IP world at least) to write off ATM as a niche technology, virtual
circuits are a serious conceptual alternative to datagram forwarding. As we shall see in 19 Quality of
Service, IP has problems handling real-time traffic, and virtual circuits offer a solution. The Internet has
so far embraced only small steps towards virtual circuits (such as MPLS, 19.12 Multi-Protocol Label
Switching (MPLS)), but they remain a tantalizing strategy.

3.10 Exercises
1. Suppose remote host A uses a VPN connection to connect to host B, with IP address [Link]. As
normal Internet connection is via device eth0 with IP address [Link]; As VPN connection is via device
ppp0 with IP address [Link]. Whenever A wants to send a packet via ppp0, it is encapsulated and
forwarded over the connection to B at [Link].

(a). Suppose As IP forwarding table is set up so that all traffic to [Link] uses eth0 and all traffic to
anywhere else uses ppp0. What happens if an intruder M attempts to open a connection to A at [Link]?
What route will packets from A to M take?
(b). Suppose As IP forwarding table is (mis)configured so that all outbound traffic uses ppp0. Describe
what will happen when A tries to send a packet.

2. Suppose remote host A wishes to use a TCP-based VPN connection to connect to host B, with IP address
[Link]. However, the VPN software is not available for host A. Host A is, however, able to run that
software on a virtual machine V hosted by A; A and V have respective IP addresses [Link] and [Link]
on the virtual network connecting them. V reaches the outside world through network address translation
(1.14 Network Address Translation), with A acting as Vs NAT router. When V runs the VPN software,
it forwards packets addressed to B the usual way, through A using NAT. Traffic to any other destination it
encapsulates over the VPN.
Can A configure its IP forwarding table so that it can make use of the VPN? If not, why not? If so, how? (If
you prefer, you may assume V is a physical host connecting to a second interface on A; A still acts as Vs
NAT router.)

80

3 Other LANs

An Introduction to Computer Networks, Release 1.0

3. Token Bus was a proprietary Ethernet-based network. It worked like Token Ring in that a small token
packet was sent from one station to the next in agreed-upon order, and a station could transmit only when it
had just received the token.

(a). If the data rate is 10 Mbps and the token is 64 bytes long (the 10-Mbps Ethernet minimum packet size),
how long does it take on average to send a packet on an idle network with 40 stations? Ignore the
propagation delay and the gap Ethernet requires between packets.
(b). Repeat part (a) assuming the tokens are only 16 bytes long.
(c). Sketch a protocol by which stations can sort themselves out to decide the order of token transmission;
that is, an order of the stations S0 ... Sn-1 where station Si sends the token to station S(i+1) mod n .

4. The IEEE 802.11 standard states transmission of the ACK frame shall commence after a SIFS period,
without regard to the busy/idle state of the medium; that is, the ACK sender does not listen first for an idle
network. Give a scenario in which the Wi-Fi ACK frame would fail to be delivered in the absence of this
rule, but succeed with it. Hint: this is another example of the hidden-node problem, [Link] Hidden-Node
Problem.
5. Suppose the average contention interval in a Wi-Fi network (802.11g) is 64 SlotTimes. The average
packet size is 1 KB, and the data rate is 54 Mbps. At that data rate, it takes about (81000)/54 = 148 sec
to transmit a packet.

(a). How long is the average contention interval, in sec?

(b). What fraction of the total potential bandwidth is lost to contention? (See 2.1.9 Analysis of Classic
Ethernet for a similar example).

6. WiMAX subscriber stations are not expected to hear one another at all. For Wi-Fi non-access-point
stations in an infrastructure (access-point) setting, on the other hand, listening to other non-access-point
transmissions is encouraged.

(a). List some ways in which Wi-Fi non-access-point stations in an infrastructure (access-point) network do
sometimes respond to packets sent by other non-access-point stations. The responses need not be in the
form of transmissions.
(b). Explain why Wi-Fi stations cannot be required to respond as in part (a).

7. Suppose WiMAX subscriber stations can be moving, at speeds of up to 33 meters/sec (the maximum
allowed under 802.16e).

(a). How much earlier (or later) can one subscriber packet arrive? Assume that the ranging process updates
the stations propagation delay once a minute. The speed of light is about 300 meters/sec.

3.10 Exercises

81

An Introduction to Computer Networks, Release 1.0

(b). With 5000 senders per second, how much time out of each second must be spent on guard intervals
accommodating the early/late arrivals above? You will need to double the time from part (a), as the base
station cannot tell whether the signal from a moving subscriber will arrive earlier or later.

8. [SM90] contained a proposal for sending IP packets over ATM as N cells as in AAL-5, followed by one
cell containing the XOR of all the previous cells. This way, the receiver can recover from the loss of any
one cell. Suppose N=20 here; with the SM90 mechanism, each packet would require 21 cells to transmit;
that is, we always send 5% more. Suppose the cell loss-rate is p (presumably very small). If we send 20
cells without the SM90 mechanism, we have a probability of about 20p that any one cell will be lost, and we
will have to retransmit the entire 20 again. This gives an average retransmission amount of about 20p extra
packets. For what value of p do the with-SM90 and the without-SM90 approaches involve about the same
total number of cell transmissions?
9. In the example in 3.7 Virtual Circuits, give the VCI table for switch S5.
10. Suppose we have the following network:
A

S1

S2

S3

S4

The virtual-circuit switching tables are below. Ports are identified by the node at the other end. Identify all
the connections. Give the path for each connection and the VCI on each link of the path.
Switch S1:
VCIin
1
2
3

portin
A
A
A

VCIout
2
2
3

portout
S3
S2
S2

VCIout
1
3
4

portout
B
S4
S4

VCIout
2
2

portout
S4
C

VCIout
2
3
1

portout
S2
S3
D

Switch S2:
VCIin
2
2
3

portin
S4
S1
S1

Switch S3:
VCIin
2
3

portin
S1
S4

Switch S4:
VCIin
2
3
4

82

portin
S3
S2
S2

3 Other LANs

An Introduction to Computer Networks, Release 1.0

11. Suppose we have the following network:

A

S1

S2

S3

S4

Give virtual-circuit switching tables for the following connections. Route via a shortest path.

(a). AD
(b). CB, via S4
(c). BD
(d). AD, via whichever of S2 or S3 was not used in part (a)

12. Below is a set of switches S1 through S4. Define VCI-table entries so the virtual circuit from A to B
follows the path
A S1 S2 S4 S3 S1 S2 S4 S3 B
That is, each switch is visited twice.
A

S1

S2

S3

S4

3.10 Exercises

83

An Introduction to Computer Networks, Release 1.0

84

3 Other LANs

4 LINKS

At the lowest (logical) level, network links look like serial lines. In this chapter we address how packet
structures are built on top of serial lines, via encoding and framing. Encoding determines how bits and
bytes are represented on a serial line; framing allows the receiver to identify the beginnings and endings of
packets.
We then conclude with the high-speed serial lines offered by the telecommunications industry, T-carrier and
SONET, upon which almost all long-haul point-to-point links that tie the Internet together are based.

4.1 Encoding and Framing

A typical serial line is ultimately a stream of bits, not bytes. How do we identify byte boundaries? This is
made slightly more complicated by the fact that, beneath the logical level of the serial line, we generally
have to avoid transmitting long runs of identical bits, because the receiver may simply lose count; this is
the clock synchronization problem (sometimes called the clock recovery problem). This means that, one
way or another, we cannot always just send the desired bits sequentially; for example, extra bits are often
inserted to break up long runs. Exactly how we do this is the encoding mechanism.
Once we have settled the transmission of bits, the next step is to determine how the receiver identifies the start
of each new packet. Ethernet packets are separated by physical gaps, but for most other link mechanisms
packets are sent end-to-end, with no breaks. How we tell when one packet stops and the next begins is the
framing problem.
To summarize:
encoding: correctly recognizing all the bits in a stream
framing: recognizing packet boundaries
These are related, though not the same.
For long (multi-kilometer) electrical serial lines, among other things we in addition want the average voltage
to be zero; that is, we want no DC component. We will mostly concern ourselves here, however, only with
lines short enough for this not to be a major concern.

4.1.1 NRZ
NRZ (Non-Return to Zero) is perhaps the simplest encoding; it corresponds to direct bit-by-bit transmission
of the 0s and 1s in the data. We have two signal levels, lo and hi, we set the signal to one or the other
of these depending on whether the data bit is 0 or 1, as in the diagram below. Note that in the diagram the
signal bits have been aligned with the start of the pulse representing that signal value.

85

An Introduction to Computer Networks, Release 1.0

NRZ replaces an earlier RZ (Return to Zero) encoding, in which hi and lo corresponded to +1 and -1, and
between each pair of pulses corresponding to consecutive bits there was a brief return to the 0 level.
One drawback to NRZ is that we cannot distinguish between 0-bits and a signal that is simply idle. However,
the more serious problem is the lack of synchronization: during long runs of 0s or long runs of 1s, the
receiver can lose count, eg if the receivers clock is running a little fast or slow. The receivers clock can
and does resynchronize whenever there is a transition from one level to the other. However, suppose bits
are sent at one per s, the sender sends 5 1-bits in a row, and the receivers clock is running 10% fast. The
signal sent is a 5-s hi pulse, but when the pulse ends the receivers clock reads 5.5 s due to the clock
speedup. Should this represent 5 1-bits or 6 1-bits?

4.1.2 NRZI
An alternative that helps here (though not obviously at first) is NRZI, or NRZ Inverted. In this encoding,
we represent a 0-bit as no change, and a 1-bit as a transition from lo to hi or hi to lo:

Now there is a signal transition aligned above every 1-bit; a 0-bit is represented by the lack of a transition.
This solves the synchronization problem for runs of 1-bits, but does nothing to address runs of 0-bits.
However, NRZI can be combined with techniques to minimize runs of 0-bits, such as 4B/5B (below).

4.1.3 Manchester
Manchester encoding sends the data stream using NRZI, with the addition of a clock transition between
each pair of consecutive data bits. This means that the signaling rate is now double the data rate, eg 20
MHz for 10Mbps Ethernet (which does use Manchester encoding). The signaling is as if we doubled the
bandwidth and inserted a 1-bit between each pair of consecutive data bits, removing this extra bit at the
receiver:

86

4 Links

An Introduction to Computer Networks, Release 1.0

All these transitions mean that the longest the clock has to count is 1 bit-time; clock synchronization is
essentially solved, at the expense of the doubled signaling rate.

4.1.4 4B/5B
In 4B/5B encoding, for each 4-bit nybble of data we actually transmit a designated 5-bit symbol, or code,
selected to have enough 1-bits. A symbol in this sense is a digital or analog transmission unit that decodes
to a set of data bits; the data bits are not transmitted individually.
Specifically, every 5-bit symbol used by 4B/5B has at most one leading 0-bit and at most two trailing 0-bits.
The 5-bit symbols corresponding to the data are then sent with NRZI, where runs of 1s are safe. Note that
the worst-case run of 0-bits has length three. Note also that the signaling rate here is 1.25 times the data
rate. 4B/5B is used in 100-Mbps Ethernet, 2.2 100 Mbps (Fast) Ethernet. The mapping between 4-bit data
values and 5-bit symbols is fixed by the 4B/5B standard:
data
0000
0001
0010
0011
0100
0101
0110
0111
1000
1001
1010

symbol
11110
01001
10100
10101
01010
01011
01110
01111
10010
10011
10110

data
1011
1100
1101
1110
1111
IDLE
HALT
START
END
RESET
DEAD

symbol
10111
11010
11011
11100
11101
11111
00100
10001
01101
00111
00000

There are more than sixteen possible symbols; this allows for some symbols to be used for signaling rather
than data. IDLE, HALT, START, END and RESET are shown above, though there are others. These can be
used to include control and status information without fear of confusion with the data. Some combinations
of control symbols do lead to up to four 0-bits in sequence; HALT and RESET have two leading 0-bits.
10-Mbps and 100-Mbps Ethernet pads short packets up to the minimum packet size with 0-bytes, meaning
that the next protocol layer has to be able to distinguish between padding and actual 0-byte data. Although
100-Mbps Ethernet uses 4B/5B encoding, it does not make use of special non-data symbols for packet
padding. Gigabit Ethernet uses PAM-5 encoding (2.3 Gigabit Ethernet), and does use special non-data
symbols (inserted by the hardware) to pad packets; there is thus no ambiguity at the receiving end as to
where the data bytes ended.
The choice of 5-bit symbols for 4B/5B is in principle arbitrary; note however that for data from 0100 to
1101 we simply insert a 1 in the fourth position, and in the last two we insert a 0 in the fourth position. The
4.1 Encoding and Framing

87

An Introduction to Computer Networks, Release 1.0

first four symbols (those with the most zeroes) follow no obvious pattern, though.

4.1.5 Framing
How does a receiver tell when one packet stops and the next one begins, to keep them from running together?
We have already seen the following techniques for addressing this framing problem: determining where
packets end:
Interpacket gaps (as in Ethernet)
4B/5B and special bit patterns
Putting a length field in the header would also work, in principle, but seems not to be widely used. One
problem with this technique is that restoring order after desynchronization can be difficult.
There is considerable overlap of framing with encoding; for example, the existence of non-data bit patterns
in 4B/5B is due to an attempt to solve the encoding problem; these special patterns can also be used as
unambiguous frame delimiters.
[Link] HDLC
HDLC (High-level Data Link Control) is a general link-level packet format used for a number of applications, including Point-to-Point Protocol (PPP) (which in turn is used for PPPoE PPP over Ethernet which
is how a great many Internet subscribers connect to their ISP), and Frame Relay, still used as the low-level
protocol for delivering IP packets to many sites via telecommunications lines. HDLC supports the following
two methods for frame separation:
HDLC over asynchronous links: byte stuffing
HDLC over synchronous links: bit stuffing
The basic encapsulation format for HDLC packets is to begin and end each frame with the byte 0x7E, or, in
binary, 0111 1110. The problem is that this byte may occur in the data as well; we must make sure we dont
misinterpret such a data byte as the end of the frame.
Asynchronous serial lines are those with some sort of start/stop indication, typically between bytes; such
lines tend to be slower. Over this kind of line, HDLC uses the byte 0x7D as an escape character. Any
data bytes of 0x7D and 0x7E are escaped by preceding them with an additional 0x7D. (Actually, they are
transmitted as 0x7D followed by (original_byte xor 0x20).) This strategy is fundamentally the same as that
used by C-programming-language character strings: the string delimiter is and the escape character is \.
Any occurrences of or \ within the string are escaped by preceding them with \.
Over synchronous serial lines (typically faster than asynchronous), HDLC generally uses bit stuffing. The
underlying bit encoding involves, say, the reverse of NRZI, in which transitions denote 0-bits and lack of
transitions denote 1-bits. This means that long runs of 1s are now the problem and runs of 0s are safe.
Whenever five consecutive 1-bits appear in the data, eg 011111, a 0-bit is then inserted, or stuffed, by the
transmitting hardware (regardless of whether or not the next data bit is also a 1). The HDLC frame byte of
0x7E = 0111 1110 thus can never appear as encoded data, because it contains six 1-bits in a row. If we had
0x7E in the data, it would be transmitted as 0111 11010.
The HDLC receiver knows that
88

4 Links

An Introduction to Computer Networks, Release 1.0

six 1-bits in a row marks the end of the packet

when five 1-bits in a row are seen, followed by a 0-bit, the 0-bit is removed
Example:
Data:
011110 0111110 01111110
Sent as: 011110 01111100 011111010 (stuffed bits in bold)
Note that bit stuffing is used by HDLC to solve two unrelated problems: the synchronization problem where
long runs of the same bit cause the receiver to lose count, and the framing problem, where the transmitted
bit pattern 0111 1110 now represents a flag that can never be mistaken for a data byte.
[Link] B8ZS
While insertion of an occasional extra bit or byte is no problem for data delivery, it is anathema to voice
engineers; extra bits upset the precise 64 Kbps DS-0 rate. As a result, long telecom lines prefer encodings
that, like 4B/5B, do not introduce timing fluctuations. Very long (electrical) lines also tend to require
encodings that guarantee a long-term average voltage level of 0 (versus 0.5 if half the bits are 1 v and half
are 0 v in NRZ); that is, the signal must have no DC component.
The AMI (Alternate Mark Inversion) technique eliminates the DC component by using three voltage levels,
nominally +1, 0 and -1; this ternary encoding is also known as bipolar. Zero bits are encoded by the 0
voltage, while 1-bits take on alternating values of +1 and -1 volts. Thus, the bits 011101 might be encoded
as 0,+1,-1,+1,0,-1, or, more compactly, 0++0. Over a long run, the +1s and the 1s cancel out.
Plain AMI still has synchronization problems with long runs of 0-bits. The solution used on North American
T1 lines (1.544 Mbps) is known as B8ZS, for bipolar with 8-zero substitution. The sender replaces any run
of 8 zero bits with a special bit-pattern, either 000+0+ or 000+0+. To decide which, the sender checks
to see if the previous 1-bit sent was +1 or 1; if the former, the first pattern is substituted, if the latter then the
second pattern is substituted. Either way, this leads to two instances of violation of the rule that consecutive
1-bits have opposite sign. For example, if the previous bit were +, the receiver sees

This double-violation is the clue to the receiver that the special pattern is to be removed and replaced with
the original eight 0-bits.

4.2 Time-Division Multiplexing

Classical circuit switching means a separate wire for each connection. This is still in common use for
residential telephone connections: each subscriber has a dedicated wire to the Central Office. But a separate
physical line for each connection is not a solution that scales well.

4.2 Time-Division Multiplexing

89

An Introduction to Computer Networks, Release 1.0

Once upon a time it was not uncommon to link computers with serial lines, rather than packet networks.
This was most often done for file transfers, but telnet logins were also done this way. The problem with this
approach is that the line had to be dedicated to one application (or one user) at a time.
Packet switching naturally implements multiplexing (sharing) on links; the demultiplexer is the destination
address. Port numbers allow demultiplexing of multiple streams to same destination host.
There are other ways for multiple channels to share a single wire. One approach is frequency-division
multiplexing, or putting each channel on a different carrier frequency. Analog cable television did this.
Some fiber-optic protocols also do this, calling it wavelength-division multiplexing.
But perhaps the most pervasive alternative to packets is the voice telephone systems time division multiplexing, or TDM, sometimes prefixed with the adjective synchronous. The idea is that we decide on a
number of channels, N, and the length of a timeslice, T, and allow each sender to send over the channel for
time T, with the senders taking turns in round-robin style. Each sender gets to send for time T at regular
intervals of NT, thus receiving 1/N of the total bandwidth. The timeslices consume no bandwidth on headers
or addresses, although sometimes there is a small amount of space dedicated to maintaining synchronization
between the two endpoints. Here is a diagram of sending with N=8:

Note, however, that if a sender has nothing to send, its timeslice cannot be used by another sender. Because
so much data traffic is bursty, involving considerable idle periods, TDM has traditionally been rejected for
data networks.

4.2.1 T-Carrier Lines

TDM, however, works extremely well for voice networks. It continues to work when the timeslice T is
small, when packet-based approaches fail because the header overhead becomes unacceptable. Consider for
a moment the telecom Digital Signal hierarchy. A single digitized voice line in North America is one 8-bit
sample every 1/8,000 second, or 64 Kbps; this is known as a DS0 channel. A T1 line the lowest level of
the T-carrier hierarchy and known at the logical level as a DS1 line represents 24 DS0 lines multiplexed
via TDM, where each channel sends a single byte at a time. Thus, every 1/8,000 of a second a T1 line carries
24 bytes of user data, one byte per channel (plus one bit for framing), for a total of 193 bits. This gives a
raw line speed of 1.544 Mbps.
Note that the per-channel frame size here is a single byte. There is no efficient way to send single-byte
packets. The advantage to the single-byte approach is that it greatly reduces the latency across the line. The
biggest source of delay in packet-based digital voice lines is the packet fill time at the senders end: the
sender generates voice data at a rate of 8 bytes/ms, and a packet cannot be sent until it is full. For a 1KB
packet, thats about a quarter second. For standard Voice-over-IP or VoIP channels, RTP is used with 160
bytes of data sent every 20 ms; for ATM, a 48-byte packet is sent every 6 ms. But the fill-time delay for a
call sent over a T1 line is 0.125 ms, which is negligible (to be fair, 6 ms and even 20 ms turn out to be pretty
negligible in terms of call quality). The T1 one-byte-at-a-time strategy also means that T1 multiplexers
need to do essentially no buffering, which might have been important back in 1962 when T-carrier was
introduced.
90

4 Links

An Introduction to Computer Networks, Release 1.0

The next most common T-carrier / Digital Signal line is perhaps T3/DS3; this represents the TDM multiplexing of 28 DS1 signals. The problem is that some individual DS1s may run a little slow, so an elaborate pulse
stuffing protocol has been developed. This allows extra bits to be inserted at specific points, if necessary, in
such a way that the original component T1s can be exactly recovered even if there are clock irregularities.
The pulse-stuffing solution did not scale well, and so T-carrier levels past T3 were very rarely used.
While T-carrier was originally intended as a way of bundling together multiple DS0 channels on a single
high-speed line, it also allows providers to offer leased digital point-to-point links with data rates in almost
any multiple of the DS0 rate.

4.2.2 SONET
SONET stands for Synchronous Optical NETwork; it is the telecommunications industrys standard mechanism for very-high-speed TDM over optical fiber. While there is now flexibility regarding the the optical
part, the synchronous part is taken quite seriously indeed, and SONET senders and receivers all use very
precisely synchronized clocks (often atomic). The actual bit encoding is NRZI.
Due to the frame structure, below, the longest possible run of 0-bits is ~250 bits (~30 bytes), but is usually
much less. Accurate reception of 250 0-bits requires a clock accurate to within (at a minimum) one part in
500, which is generally within reach. This mechanism solves most of the clock-synchronization problem,
though SONET also has a resynchronization protocol in case the receiver gets lost.
The primary reason for SONETs accurate clocking, however, is not the clock-synchronization problem as
we have been using the term, but rather the problem of demultiplexing and remultiplexing multiple component bitstreams in a setting in which some of the streams may run slow. One of the primary design goals for
SONET was to allow such multiplexing without the need for pulse stuffing, as is used in the Digital Signal
hierarchy. SONET tributary streams are in effect not allowed to run slow (although SONET does provide
for occasional very small byte slips, below). Furthermore, as multiple SONET streams are demultiplexed
at a switching center and then remultiplexed into new SONET streams, synchronization means that none of
the streams falls behind or gets ahead.
The basic SONET format is known as STS-1. Data is organized as a 9x90 byte grid. The first 3 bytes of
each row (that is, the first three columns) form the frame header. Frames are not addressed; SONET is a
point-to-point protocol and a node sends a continuous sequence of frames to each of its neighbors. When the
frames reach their destination, in principle they need to be fully demultiplexed for the data to be forwarded
on. In practice, there are some shortcuts to full demultiplexing.

4.2 Time-Division Multiplexing

91

An Introduction to Computer Networks, Release 1.0

The actual bytes sent are scrambled: the data is XORed with a standard, fixed pseudorandom pattern before
transmission. This introduces many 1-bits, on which clock resynchronization can occur, with a high degree
of probability.
There are two other special columns in a frame, each guaranteed to contain at least one 1-bit, so the maximum run of data bytes is limited to ~30; this is thus the longest run of possible 0s.
The first two bytes of each frame are 0xF628. SONETs frame-synchronization check is based on verifying
these byte values at the start of each frame. If the receiver is ever desynchronized, it begins a frame resynchronization procedure: the receiver searches for those 0xF628 bytes at regular 810-byte (6480-bit)
spacing. After a few frames with 0xF628 in the right place, the receiver is very sure it is looking at the
synchronization bytes and not at a data-byte position. Note that there is no evident byte boundary to a
SONET frame, so the receiver must check for 0xF628 beginning at every bit position.
SONET frames are transmitted at a rate of 8,000 frames/second. This is the canonical byte sampling rate
for standard voice-grade (DS0, or 64 Kbps) lines. Indeed, the classic application of SONET is to transmit
multiple DS0 voice calls using TDM: within a frame, each data byte position is given over to one voice
channel. The same byte position in consecutive frames constitutes one byte every 1/8000 seconds. The
basic STS-1 data rate of 51.84 Mbps is exactly 810 bytes/frame 8 bits/byte 8000 frames/sec.
To a customer who has leased a SONET-based channel to transmit data, a SONET link looks like a very fast
bitstream. There are several standard ways of encoding data packets over SONET. One is to encapsulate the
data as ATM cells, and then embed the cells contiguously in the bitstream. Another is to send IP packets
encoded in the bitstream using HDLC-like bit stuffing, which means that the SONET bytes and the IP bytes
may no longer correspond. The advantage of HDLC encoding is that it makes SONET re-synchronization
vanishingly infrequent. Most IP backbone traffic today travels over SONET links.
Within the 990-byte STS-1 frame, the payload envelope is the 987 region nominally following the three
header columns; this payload region has its own three reserved columns meaning that there are 84 columns
(984 bytes) available for data. This 987-byte payload envelope can float within the physical 990byte frame; that is, if the input frames are running slow then the output physical frames can be transmitted
at the correct rate by letting the payload frames slip backwards, one byte at a time. Similarly, if the input
frames are arriving slightly too fast, they can slip forwards by up to one byte at a time; the extra byte is
stored in a reserved location in the three header columns of the 990 physical frame.
Faster SONET streams are made by multiplexing slower ones. The next step up is STS-3, an STS-3 frame
is three STS-1 frames, for 9270 bytes. STS-3 (or, more properly, the physical layer for STS-3) is also
called OC-3, for Optical Carrier. Beyond STS-3, faster lines are multiplexed combinations of four of the
next-slowest lines. Here are some of the higher levels:
STS
STS-1
STS-3
STS-12
STS-48
STS-192
STS-768

STM
STM-0
STM-1
STM-4
STM-16
STM-64
STM-256

bandwidth
51.84 Mbps
155.52 Mbps
622.08 Mbps (=12*51.84, exactly)
2.488 Gbps
9.953 Gbps
39.8 Gbps

Faster SONET lines have been defined, but a simpler way to achieve very high data rates over optical fiber is
to use wavelength-division multiplexing (that is, frequency-division multiplexing at optical frequencies);
this means we have separate SONET channels at different wavelengths of light.

92

4 Links

An Introduction to Computer Networks, Release 1.0

SONET provides a wide variety of leasing options at various bandwidths. High-volume customers can lease
an entire STS-1 or larger unit. Alternatively, the 84 columns of an STS-1 frame can be divided into seven
virtual tributary groups, each of twelve columns; these groups can be leased individually or in multiples,
or be further divided into as few as three columns (which works out to be just over the T1 data rate).
[Link] Other Optical Fiber

Optical Fiber and Lightning

One of the advantages of of optical fiber (particularly at mountaintop observatories) is its resistance to
damage and interference from lightning. Some fiber-optic cables do, however, have metal jackets to
add strength or to resist animals; lightning resistance must be researched carefully.
SONET is primarily, though not exclusively, used by telecommunications carriers. There are also multiple
fiber-optic alternatives for smaller-scale operations that generally do not involve SONETs rigid synchronization mechanisms and voice-oriented framing. These are often part of the Ethernet family although they
may have little except their bitrate in common with one another or with Ethernet over copper wire. Another
standard that supports optical fiber links is so-called Fibre Channel, though that too also now supports
copper.
Distances supported by fiber-optic Ethernet range from hundreds of meters to tens of kilometers. Generally,
the long-haul links require the use of full duplex to avoid the collision-detection (slot time) requirement that
a sender continue to transmit for one full Ethernet-segment RTT, but full duplex is common at high Ethernet
speeds even for short links. For 100 Mbps Ethernet, fiber-optic standards include 100BASE-FX, 100BASESX, 100BASE-LX and 100BASE-BX. The latter supports distances up to 40 km; the limiting factor is the
need for signal regeneration. 1000-Gbit Ethernet optical standards include 1000BASE-SX, 1000BASE-LX,
1000BASE-BX and 1000BASE-EX; the latter again supports up to 40 km. These forms of Ethernet are
often deployed in residential fiber Internet connections, although in some cases the last hundred meters or
so may still involve copper.
At speeds of 10 Gbps, long-range optical fiber alternatives include 10GBASE-ER and 10GBASE-ZR, supportking 40 and 80 km respectively; 10GBASE-ZR is based on SONET STM-64 standards. It is worth
noting that fiber is often used for short links as well at 1 Gbps and 10 Gbps speeds; at 10 Gbps some copperlink standards support distances of only a few meters. As of 2014, several 10-Gbps Ethernet physical-layer
standards come from industry consortiums rather than the IEEE.

4.3 Epilog
This completes our discussion of common physical links. Perhaps the main takeaway point is that transmitting bits over any distance is not quite as simple as it may appear; simple NRZ transmission is not effective.

4.4 Exercises
1. What is encoded by the following NRZI signal? The first two bits are shown.

4.3 Epilog

93

An Introduction to Computer Networks, Release 1.0

2. Argue that sending 4 0-bits via NRZI requires a clock accurate to within 1 part in 8. Assume that the
receiver resynchronizes its clock whenever a 1-bit transition is received, but that otherwise it attempts to
sample a bit in the middle of the bits timeslot.
3.(a) What bits are encoded by the following Manchester-encoded sequence?

(b). Why is there no ambiguity as to whether the first transition is a clock transition or a data (1-bit)
transition?
(c). Give an example of a signal pattern consisting of an NRZI encoding of 0-bits and 1-bits that does not
contain two consecutive 0-bits and which is not a valid Manchester encoding of data. Such a pattern could
thus could be used as a special non-data marker.

4. What three ASCII letters (bytes) are encoded by the following 4B/5B pattern? (Be careful about uppercase
vs lowercase.)
010110101001110101010111111110
5.(a) Suppose a device is forwarding SONET STS-1 frames. How much clock drift, as a percentage, on
the incoming line would mean that the output payload envelopes must slip backwards by one byte per three
physical frames? | (b). In 4.2.2 SONET it was claimed that sending 250 0-bits required a clock accurate to
within 1 part in 500. Describe how a SONET clock might meet the requirement of part (a) above, and yet
fail at this second requirement. (Hint: in part (a) the requirement is a long-term average).

94

4 Links

5 PACKETS

In this chapter we address a few abstract questions about packets, and take a close look at transmission
times. We also consider how big packets should be, and how to detect transmission errors. These issues are
independent of any particular set of protocols.

5.1 Packet Delay

There are several contributing sources to the delay encountered in transmitting a packet. On a LAN, the most
significant is usually what we will call bandwidth delay: the time needed for a sender to get the packet onto
the wire. This is simply the packet size divided by the bandwidth, after everything has been converted to
common units (either all bits or all bytes). For a 1500-byte packet on 100 Mbps Ethernet, the bandwidth
delay is 12,000 bits / (100 bits/sec) = 120 sec.
There is also propagation delay, relating to the propagation of the bits at the speed of light (for the transmission medium in question). This delay is the distance divided by the speed of light; for 1,000 m of Ethernet
cable, with a signal propagation speed of about 230 m/sec, the propagation delay is about 4.3 sec. That is,
if we start transmitting the 1500 byte packet of the previous paragraph at time T=0, then the first bit arrives
at a destination 1,000 m away at T = 4.3 sec, and the last bit is transmitted at 120 sec, and the last bit
arrives at T = 124.3 sec.
Minimizing Delay
Back in the last century, gamers were sometimes known to take advantage of players with slow (as in
dialup) links; an opponent could be eliminated literally before he or she could respond. As an updated
take on this, some financial-trading firms have set up microwave-relay links between trading centers,
say New York and Chicago, in order to reduce delay. In computerized trading, milliseconds count. A
direct line of sight from New York to Chicago which we round off to 1200 km takes about 4 ms
in air, where signals propagate at essentially the speed of light c = 300 km/ms. But fiber is slower;
even an absolutely straight run would take 6 ms at glass fibers propagation speed of 200 km/ms. In
the presence of high-speed trading, this 2 ms savings is of considerable financial significance.
Bandwidth delay, in other words, tends to dominate within a LAN.
But as networks get larger, propagation delay begins to dominate. This also happens as networks get faster:
bandwidth delay goes down, but propagation delay remains unchanged.
An important difference between bandwidth delay and propagation delay is that bandwidth delay is proportional to the amount of data sent while propagation delay is not. If we send two packets back-to-back, then
the bandwidth delay is doubled but the propagation delay counts only once.
The introduction of switches leads to store-and-forward delay, that is, the time spent reading in the entire
packet before any of it can be retransmitted. Store-and-forward delay can also be viewed as an additional
bandwidth delay for the second link.

95

An Introduction to Computer Networks, Release 1.0

Finally, a switch may or may not also introduce queuing delay; this will often depend on competing traffic.
We will look at this in more detail in 14 Dynamics of TCP Reno, but for now note that a steady queuing delay
(eg due to a more-or-less constant average queue utilization) looks to each sender more like propagation
delay than bandwidth delay, in that if two packets are sent back-to-back and arrive that way at the queue,
then the pair will experience only a single queuing delay.

5.1.1 Delay examples

Case 1: A

Propagation delay is 40 sec

Bandwidth is 1 byte/sec (1 mB/sec, 8 Mbit/sec)
Packet size is 200 bytes (200 sec bandwidth delay)
Then the total one-way transmit time is 240 sec = 200 sec + 40 sec
Case 2: A

Like the previous example except that the propagation delay is increased to 4 ms
The total transmit time is now 4200 sec = 200 sec + 4000 sec
Case 3: A

We now have two links, each with propagation delay 40 sec; bandwidth and packet size as in
Case 1
The total transmit time for one 200-byte packet is now 480 sec = 240 + 240. There are two propagation
delays of 40 sec each; A introduces a bandwidth delay of 200 sec and R introduces a store-and-forward
delay (or second bandwidth delay) of 200 sec.
Case 4: A

The same as 3, but with data sent as two 100-byte packets

The total transmit time is now 380 sec = 3x100 + 2x40. There are still two propagation delays, but there is
only 3/4 as much bandwidth delay because the transmission of the first 100 bytes on the second link overlaps
with the transmission of the second 100 bytes on the first link.

96

5 Packets

An Introduction to Computer Networks, Release 1.0

These ladder diagrams represent the full transmission; a snapshot state of the transmission at any one
instant can be obtained by drawing a horizontal line. In the middle, case 3, diagram, for example, at no
instant are both links active. Note that sending two smaller packets is faster than one large packet. We
expand on this important point below.
Now let us consider the situation when the propagation delay is the most significant component. The crosscontinental US roundtrip delay is typically around 50-100 ms (propagation speed 200 km/ms in cable, 5,00010,000 km cable route, or about 3-6000 miles); we will use 100 ms in the examples here. At 1.0 Mbit, 100ms
is about 12KB, or eight full-sized Ethernet packets. At this bandwidth, we would have four packets and four
returning ACKs strung out along the path. At 1.0 Gbit, in 100ms we can send 12,000 KB, or 800 Ethernet
packets, before the first ACK returns.
At most non-LAN scales, the delay is typically simplified to the round-trip time, or RTT: the time between
sending a packet and receiving a response.
Different delay scenarios have implications for protocols: if a network is bandwidth-limited then protocols
are easier to design. Extra RTTs do not cost much, so we can build in a considerable amount of back-andforth exchange. However, if a network is delay-limited, the protocol designer must focus on minimizing
extra RTTs. As an extreme case, consider wireless transmission to the moon (0.3 sec RTT), or to Jupiter (1
hour RTT).
At my home I formerly had satellite Internet service, which had a roundtrip propagation delay of ~600 ms.
This is remarkably high when compared to purely terrestrial links.
When dealing with reasonably high-bandwidth large-scale networks (eg the Internet), to good approximation most of the non-queuing delay is propagation, and so bandwidth and total delay are effectively
independent. Only when propagation delay is small are the two interrelated. Because propagation delay
dominates at this scale, we can often make simplifications when diagramming. In the illustration below, A
sends a data packet to B and receives a small ACK in return. In (a), we show the data packet traversing
several switches; in (b) we show the data packet as if it were sent along one long unswitched link, and in (c)
we introduce the idealization that bandwidth delay (and thus the width of the packet line) no longer matters.
(Most later ladder diagrams in this book are of this type.)

5.1 Packet Delay

97

An Introduction to Computer Networks, Release 1.0

The bandwidth delay product (usually involving round-trip delay, or RTT), represents how much we can
send before we hear anything back, or how much is pending in the network at any one time if we send
continuously. Note that, if we use RTT instead of one-way time, then half the pending packets will be
returning ACKs. Here are a few values
RTT
1 ms
100 ms
100 ms

bandwidth
10 Mbps
1.5 Mbps
600 Mbps

bandwidth delay
1.2 KB
20 KB
8,000 KB

5.2 Packet Delay Variability

For many links, the bandwidth delay and the propagation delay are rigidly fixed quantities, the former by the
bandwidth and the latter by the speed of light. This leaves queuing delay as the major source of variability.
This state of affairs lets us define RTTnoLoad to be the time it takes to transmit a packet from A to B, and
receive an acknowledgment back, with no queuing delay.
While this is often a reasonable approximation, it is not necessarily true that RTTnoLoad is always a fixed
quantity. There are several possible causes for RTT variability. On Ethernet and Wi-Fi networks there is an
initial contention period before transmission actually begins. Although this delay is related to waiting for
other senders, it is not exactly queuing delay, and a packet may encounter considerable delay here even if it
ends up being the first to be sent. For Wi-Fi in particular, the uncertainty introduced by collisions into packet
delivery times even with no other senders competing can complicate higher-level delay measurements.
It is also possible that different packets are routed via slightly different paths, leading to (hopefully) minor
variations in travel time, or are handled differently by different queues of a parallel-processing switch.
A links bandwidth, too, can vary dynamically. Imagine, for example, a T1 link comprised of the usual
24 DS0 channels, in which all channels not currently in use by voice calls are consolidated into a single
data channel. With eight callers, the data bandwidth would be cut by a third from 24 DS0 to 16 DS0.
98

5 Packets

An Introduction to Computer Networks, Release 1.0

Alternatively, perhaps routers are allowed to reserve a varying amount of bandwidth for high-priority traffic,
depending on demand, and so the bandwidth allocated to the best-effort traffic can vary. Perceived link
bandwidth can also vary over time if packets are compressed at the link layer, and some packets are able to
be compressed more than others.
Finally, if mobile nodes are involved, then the distance and thus the propagation delay can change. This can
be quite significant if one is communicating with a wireless device that is being taken on a cross-continental
road trip.
Despite these sources of fluctuation, we will usually assume that RTTnoLoad is fixed and well-defined, especially when we wish to focus on the queuing component of delay.

5.3 Packet Size

How big should packets be? Should they be large (eg 64 KB) or small (eg 48 bytes)?
The Ethernet answer to this question had to do with equitable sharing of the line: large packets would not
allow other senders timely access to transmit. In any network, this issue remains a concern.
On the other hand, large packets waste a smaller percentage of bandwidth on headers. However, in most of
the cases we will consider, this percentage does not exceed 10% (the VoIP/RTP example in 1.3 Packets is
an exception).
It turns out that if store-and-forward switches are involved, smaller packets have much better throughput.
The links on either side of the switch can be in use simultaneously, as in Case 4 of 5.1.1 Delay examples.
This is a very real effect, and has put a damper on interest in support for IP jumbograms. The ATM
protocol (intended for both voice and data) pushes this to an extreme, with packets with only 48 bytes of
data and 5 bytes of header.
As an example of this, consider a path from A to B with four switches and five links:
A

R1

R2

R3

R4

Suppose we send either one big packet or five smaller packets. The relative times from A to B are illustrated
in the following figure:

5.3 Packet Size

99

An Introduction to Computer Networks, Release 1.0

The point is that we can take advantage of parallelism: while the R4B link above is handling packet 1,
the R3R4 link is handling packet 2 and the R2R3 link is handling packet 3 and so on. The five smaller
packets would have five times the header capacity, but as long as headers are small relative to the data, this
is not a significant issue.
The sliding-windows algorithm, used by TCP, uses this idea as a continuous process: the sender sends a
continual stream of packets which travel link-by-link so that, in the full-capacity case, all links may be in
use at all times.

5.3.1 Error Rates and Packet Size

Packet size is also influenced, to a modest degree, by the transmission error rate. For relatively high error
rates, it turns out to be better to send smaller packets, because when an error does occur then the entire
packet containing it is lost.
For example, suppose that 1 bit in 10,000 is corrupted, at random, so the probability that a single bit is
transmitted correctly is 0.9999 (this is much higher than the error rates encountered on real networks). For a
1000-bit packet, the probability that every bit in the packet is transmitted correctly is (0.9999)1000 , or about
90%. For a 10,000-bit packet the probability is (0.9999)10,000 37%. For 20,000-bit packets, the success
rate is below 14%.
Now suppose we have 1,000,000 bits to send, either as 1000-bit packets or as 20,000-bit packets. Nominally
this would require 1,000 of the smaller packets, but because of the 90% packet-success rate we will need to
retransmit 10% of these, or 100 packets. Some of the retransmissions may also be lost; the total number of
packets we expect to need to send is about 1,000/90% 1,111, for a total of 1,111,000 bits sent. Next, let
us try this with the 20,000-bit packets. Here the success rate is so poor that each packet needs to be sent on

100

5 Packets

An Introduction to Computer Networks, Release 1.0

average seven times; lossless transmission would require 50 packets but we in fact need 750 = 350 packets,
or 7,000,000 bits.
Moral: choose the packet size small enough that most packets do not encounter errors.
To be fair, very large packets can be sent reliably on most cable links (eg TDM and SONET). Wireless,
however, is more of a problem.

5.3.2 Packet Size and Real-Time Traffic

There is one other concern regarding excessive packet size. As we shall see in 19 Quality of Service, it
is common to commingle bulk traffic on the same links with real-time traffic. It is straightforward to give
priority to the real-time traffic in such a mix, meaning that a router does not begin forwarding a bulk-traffic
packet if there are any real-time packets waiting (we do need to be sure in this case that real-time traffic
will not amount to so much as to starve the bulk traffic). However, once a bulk-traffic packet has begun
transmission, it is impractical to interrupt it.
Therefore, one component of any maximum-delay bound for real-time traffic is the transmission time for
the largest bulk-traffic packet; we will call this the largest-packet delay. As a practical matter, most IPv4
packets are limited to the maximum Ethernet packet size of 1500 bytes, but IPv6 has an option for so-called
jumbograms up to 2 MB in size. Transmitting one such packet on a 100 Mbps link takes about 1/6 of a
second, which is likely too large for happy coexistence with real-time traffic.

5.4 Error Detection

The basic strategy for packet error detection is to add some extra bits formally known as an error-detection
code that will allow the receiver to determine if the packet has been corrupted in transit. A corrupted packet
will then be discarded by the receiver; higher layers do not distinguish between lost packets and those never
received. While packets lost due to bit errors occur much less frequently than packets lost due to queue
overflows, it is essential that data be received accurately.
Intermittent packet errors generally fall into two categories: low-frequency bit errors due to things like
cosmic rays, and interference errors, typically generated by nearby electrical equipment. Errors of the latter
type generally occur in bursts, with multiple bad bits per packet. Occasionally, a malfunctioning network
device will introduce bursty errors as well.
Networks v Refrigerators
At Loyola we once had a workstation used as a mainframe terminal that kept losing its connection.
We eventually noticed that the connection dropped every time the office refrigerator kicked on. Sure
enough, the cable ran directly behind the fridge; rerouting it solved the problem.
The simplest error-detection mechanism is a single parity bit; this will catch all one-bit errors. There is,
however, no straightforward generalization to N bits! That is, there is no N-bit error code that catches all
N-bit errors; see exercise 9.
The so-called Internet checksum, used by IP, TCP and UDP, is formed by taking the ones-complement sum
of the 16-bit words of the message. Ones-complement is an alternative way of representing signed integers
5.4 Error Detection

101

An Introduction to Computer Networks, Release 1.0

in binary; if one adds two positive integers and the sum does not overflow the hardware word size, then
ones-complement and the now-universal twos-complement are identical. To form the ones-complement sum
of 16-bit words A and B, first take the ordinary twos-complement sum A+B. Then, if there is an overflow
bit, add it back in as low-order bit. Thus, if the word size is 4 bits, the ones-complement sum of 0101 and
0011 is 1000 (no overflow). Now suppose we want the ones-complement sum of 0101 and 1100. First we
take the exact sum and get 1|0001, where the leftmost 1 is an overflow bit past the 4-bit wordsize. Because
of this overflow, we add this bit back in, and get 0010.
The ones-complement numeric representation has two forms for zero: 0000 and 1111 (it is straightforward
to verify that any 4-bit quantity plus 1111 yields the original quantity; in twos-complement notation 1111
represents -1, and an overflow is guaranteed, so adding back the overflow bit cancels the -1 and leaves us
with the original number). It is a fact that the ones-complement sum is never 0000 unless all bits of all the
summands are 0; if the summands add up to zero by coincidence, then the actual binary representation will
be 1111. This means that we can use 0000 in the checksum to represent checksum not calculated, which
the UDP protocol used to permit.
Ones-complement
Long ago, before Loyola had any Internet connectivity, I wrote a primitive UDP/IP stack to allow me
to use the Ethernet to back up one machine that did not have TCP/IP to another machine that did. We
used private IP addresses of the form 10.0.0.x. I set as many header fields to zero as I could. I paid
no attention to how to implement ones-complement addition; I simply used twos-complement, for the
IP header only, and did not use a UDP checksum at all. Hey, it worked.
Then we got a real Class B address block [Link]/16, and changed IP addresses. My software no
longer worked. It turned out that, in the original version, the IP header bytes were all small enough
that when I added up the 16-bit words there were no carries, and so ones-complement was the same as
twos-complement. With the new addresses, this was no longer true. As soon as I figured out how to
implement ones-complement addition properly, my backups worked again.
There is another way to look at the (16-bit) ones-complement sum: it is in fact the remainder upon dividing
the message (seen as a very long binary number) by 216 - 1. This is similar to the decimal casting out nines
rule: if we add up the digits of a base-10 number, and repeat the process until we get a single digit, then that
digit is the remainder upon dividing the original number by 10-1 = 9. The analogy here is that the message
is looked at as a very large number written in base-216 , where the digits are the 16-bit words. The process
of repeatedly adding up the digits until we get a single digit amounts to taking the ones-complement
sum of the words.
A weakness of any error-detecting code based on sums is that transposing words leads to the same sum, and
the error is not detected. In particular, if a message is fragmented and the fragments are reassembled in the
wrong order, the ones-complement sum will likely not detect it.
While some error-detecting codes are better than others at detecting certain kinds of systematic errors (for
example, CRC, below, is usually better than the Internet checksum at detecting transposition errors), ultimately the effectiveness of an error-detecting code depends on its length. Suppose a packet P1 is corrupted
randomly into P2, but still has its original N-bit error code EC(P1). This N-bit code will fail to detect the
error that has occurred if EC(P2) is, by chance, equal to EC(P1). The probability that two random N-bit
codes will match is 1/2N (though a small random change in P1 might not lead to a uniformly distributed
random change in EC(P1); see the tail end of the CRC section below).

102

5 Packets

An Introduction to Computer Networks, Release 1.0

This does not mean, however, that one packet in 2N will be received incorrectly, as most packets are errorfree. If we use a 16-bit error code, and only 1 packet in 100,000 is actually corrupted, then the rate at which
corrupted packets will sneak by is only 1 in 100,000 65536, or about one in 6 109 . If packets are 1500
bytes, you have a good chance (90+%) of accurately transferring a terabyte, and a 37% chance (1/e) at ten
terabytes.

5.4.1 Cyclical Redundancy Check: CRC

The CRC error code is based on long division of polynomials, which tends to sound complicated but which
in fact reduces all the long-division addition/subtraction operations to the much-simpler XOR operation.
We treat the message, in binary, as a giant polynomial m(X), using the bits of the message as successive
coefficients (eg 10011011 = X7 + X4 + X3 + X + 1). We standardize a divisor polynomial p(X) of degree 32
(at least for CRC-32 codes); the full specification of a given CRC code requires giving this polynomial. We
divide m(X) by p(X); our checksum is the remainder r(X), of maximum degree 31 (that is, 32 bits).
This is a reasonably secure hash against real-world network corruption, in that it is very hard for systematic errors to result in the same hash code. However, CRC is not secure against intentional corruption;
given msg1, there are straightforward mathematical means for tweaking the last bytes of msg2 so that so
CRC(msg1) = CRC(msg2)
As an example of CRC, suppose that the CRC divisor is 1011 (making this a CRC-3 code) and the message
is 0110 0001 1110. Here is the division:
1110 0110
1011

0110 0001 1100

101 1
011 10
10 11
01 010
1 011
0 0011 11
10 11
01 000
1 011
0 0110

The remainder, at the bottom, is 110; this is the CRC code.

CRC is easily implemented in hardware, using bit-shifting registers. Fast software implementations are also
possible, usually involving handling the bits one byte at a time, with a precomputed lookup table with 256
entries.
If we randomly change enough bits in packet P1 to create P2, then CRC(P1) and CRC(P2) are effectively
independent random variables, with probability of a match 1 in 2N where N is the CRC length. However, if
we change just a few bits then the change is not so random. In particular, for many CRC codes (that is, for
many choices of the underlying polynomial p(X)), changing up to three bits in P1 to create a new message
5.4 Error Detection

103

An Introduction to Computer Networks, Release 1.0

P2 guarantees that CRC(P1) = CRC(P2). For the Internet checksum, this is not guaranteed even if we know
only two bits were changed.
Finally, there are also secure hashes, such as MD-5 and SHA-1 and their successors. Nobody knows
(or admits to knowing) how to produce two messages with same hash here. However, these secure-hash
codes are generally not used in network error-correction as they take considerable time to compute; they are
generally used only for secure authentication and other higher-level functions.

5.4.2 Error-Correcting Codes

If a link is noisy, we can add an error-correction code (also called forward error correction) that allows the
receiver in many cases to figure out which bits are corrupted, and fix them. This has the effect of improving
the bit error rate at a cost of reducing throughput. Error-correcting codes tend to involve many more bits than
are needed for error detection. Typically, if a communications technology proves to have an unacceptably
high bit-error rate (such as wireless), the next step is to introduce an error-correcting code to the protocol.
This generally reduces the virtual bit-error rate (that is, the error rate as corrected) to acceptable levels.
Perhaps the easiest error-correcting code to visualize is 2-D parity, for which we need O(N1/2 ) additional
bits. We take NN data bits and arrange them into a square; we then compute the parity for every column,
for every row, and for the entire square; this is 2N+1 extra bits. Here is a diagram with N=4, and with even
parity; the column-parity bits (in blue) are in the bottom (fifth) row and the row-parity bits (also in blue)
are in the rightmost (fifth) column. The parity bit for the entire 44 data square is the light-blue bit in the
bottom right corner.

Now suppose one bit is corrupted; for simplicity, assume it is one of the data bits. Then exactly one columnparity bit will be incorrect, and exactly one row-parity bit will be incorrect. These two incorrect bits mark
the column and row of the incorrect data bit, which we can then flip to the correct state.
We can make N large, but an essential requirement here is that there be only a single corrupted bit per square.
We are thus likely either to keep N small, or to choose a different code entirely that allows correction of
multiple bits. Either way, the addition of error-correcting codes can easily increase the size of a packet
significantly; some codes double or even triple the total number of bits sent.
The Hamming code is another popular error-correction code; it adds O(log N) additional bits, though if N is
large enough for this to be a material improvement over the O(N1/2 ) performance of 2-D parity then errors
must be very infrequent. If we have 8 data bits, let us number the bit positions 0 through 7. We then write
each bits position as a binary value between 000 and 111; we will call these the position bits of the given
data bit. We now add four code bits as follows:
1. a parity bit over all 8 data bits

104

5 Packets

An Introduction to Computer Networks, Release 1.0

2. a parity bit over those data bits for which the first digit of the position bits is 1 (these are positions 4,
5, 6 and 7)
3. a parity bit over those data bits for which the second digit of the position bits is 1 (these are positions
010, 011, 110 and 111, or 2, 3, 6 and 7)
4. a parity bit over those data bits for which the third digit of the position bits is 1 (these are positions
001, 011, 101, 111, or 1, 3, 5 and 7)
We can tell whether or not an error has occurred by the first code bit; the remaining three code bits then tell
us the respective three position bits of the incorrect bit. For example, if the #2 code bit above is correct, then
the first digit of the position bits is 0; otherwise it is one. With all three position bits, we have identified the
incorrect data bit.
As a concrete example, suppose the data word is 10110010. The four code bits are thus
1. 0, the (even) parity bit over all eight bits
2. 1, the parity bit over the second half, 10110010
3. 1, the parity bit over the bold bits: 10110010
4. 1, the parity bit over these bold bits: 10110010
If the received data+code is now 10111010 0111, with the bold bit flipped, then the fact that the first code
bit is wrong tells the receiver there was an error. The second code bit is also wrong, so the first bit of the
position bits must be 1. The third code bit is right, so the second bit of the position bits must be 0. The
fourth code bit is also right, so the third bit of the position bits is 0. The position bits are thus binary 100, or
4, and so the receiver knows that the incorrect bit is in position 4 (counting from 0) and can be flipped to the
correct state.

5.5 Epilog
The issues presented here are perhaps not very glamorous, and often play a supporting, behind-the-scenes
role in protocol design. Nonetheless, their influence is pervasive; we may even think of them as part of the
underlying physics of the Internet.
As the early Internet became faster, for example, and propagation delay became the dominant limiting
factor, protocols were often revised to limit the number of back-and-forth exchanges. A classic example is
the Simple Mail Transport Protocol (SMTP), amended by RFC 1854 to allow multiple commands to be sent
together called pipelining instead of individually.
While there have been periodic calls for large-packet support in IPv4, and IPv6 protocols exist for jumbograms in excess of a megabyte, these are very seldom used, due to the store-and-forward costs of large
packets as described in 5.3 Packet Size.
Almost every LAN-level protocol, from Ethernet to Wi-Fi to point-to-point links, incorporates an errordetecting code chosen to reflect the underlying transportation reliability. Ethernet includes a 32-bit CRC
code, for example, while Wi-Fi includes extensive error-correcting codes due to the noisier wireless environment. The Wi-Fi fragmentation option ([Link] Wi-Fi Fragmentation) is directly tied to 5.3.1 Error
Rates and Packet Size.

5.5 Epilog

105

An Introduction to Computer Networks, Release 1.0

5.6 Exercises
1. Suppose a link has a propagation delay of 20 sec and a bandwidth of 2 bytes/sec.

(a). How long would it take to transmit a 600-byte packet over such a link?
(b). How long would it take to transmit the 600-byte packet over two such links, with a store-and-forward
switch in between?

2. Suppose the path from A to B has a single switch S in between: A

delay of 60 sec and a bandwidth of 2 bytes/sec.

B. Each link has a propagation

(a). How long would it take to send a single 600-byte packet from A to B?
(b). How long would it take to send two back-to-back 300-byte packets from A to B?
(c). How long would it take to send three back-to-back 200-byte packets from A to B?

3. Repeat parts (a) and (b) of the previous exercise, except change the per-link propagation delay from 60
sec to 600 sec.
4. Again suppose the path from A to B has a single switch S in between: A
width and propagation delays are as follows:
link
A S
S B

bandwidth
5 bytes/sec
3 bytes/sec

B. The per-link band-

propagation delay
24 sec
13 sec

(a). How long would it take to send a single 600-byte packet from A to B?
(b). How long would it take to send two back-to-back 300-byte packets from A to B? Note that, because
the S B link is slower, packet 2 arrives at S from A well before S has finished transmitting packet 1 to B.

5. Suppose in the previous exercise, the AS link has the smaller bandwidth of 3 bytes/sec and the SB
link has the larger bandwidth of 5 bytes/sec. Now how long does it take to send two back-to-back 300-byte
packets from A to B?
6. Suppose we have five links, A R1
R2 R3
R4
bytes/ms. Assume we model the per-link propagation delay as 0.

B. Each link has a bandwidth of 100

(a). How long would it take a single 1500-byte packet to go from A to B?

(b). How long would it take five consecutive 300-byte packets to go from A to B?

The diagram in 5.3 Packet Size may help.

106

5 Packets

An Introduction to Computer Networks, Release 1.0

7. Suppose there are N equal-bandwidth links on the path between A and B, as in the diagram below, and
we wish to send M consecutive packets.
S1

...

SN-1

Let BD be the bandwidth delay of a single packet on a single link, and let PD be the propagation delay on a
single link. Show that the total bandwidth delay is (M+N-1)BD, and the total propagation delay is NPD.
Hint: When does the Mth packet leave A? What is its total transit time to B? Why do no packets have to wait
at any Si for the completion of the transmission of an earlier packet?
8. Repeat the analysis in 5.3.1 Error Rates and Packet Size to compare the probable total number of bytes
that need to be sent to transmit 107 bytes using

(a). 1,000-byte packets

(b). 10,000-byte packets

Assume the bit error rate is 1 in 16 105 , making the error rate per byte about 1 in 2 105 .
9. In the text it is claimed there is no N-bit error code that catches all N-bit errors for N2 (for N=1, a
parity bit works). Prove this claim for N=2. Hint: pick a length M, and consider all M-bit messages with a
single 1-bit. Any such message can be converted to any other with a 2-bit error. Show, using the Pigeonhole
Principle, that for large enough M two messages m1 and m2 must have the same error code, that is, e(m1 ) =
e(m2 ). If this occurs, then the error code fails to detect the error that converted m1 into m2 .
10. In the description in the text of the Internet checksum, the overflow bit was added back in after each
ones-complement addition. Show that the same final result will be obtained if we add up the 16-bit words
using 32-bit twos-complement arithmetic (the normal arithmetic on all modern hardware), and then add the
upper 16 bits of the sum to the lower 16 bits. (If there is an overflow at this last step, we have to add that
back in as well.)
11. Suppose a message is 110010101. Calculate the CRC-3 checksum using the polynomial X3 + 1, that is,
find the 3-bit remainder using divisor 1001.
12. The CRC algorithm presented above requires that we process one bit at a time. It is possible to do the
algorithm N bits at a time (eg N=8), with a precomputed lookup table of size 2N . Complete the steps in the
following description of this strategy for N=3 and polynomial X3 + X + 1, or 1011.
13. Consider the following set of bits sent with 2-D even parity; the data bits are in the 44 upper-left block
and the parity bits are in the rightmost column and bottom row. Which bit is corrupted?
1

5.6 Exercises

107

An Introduction to Computer Networks, Release 1.0

14. (a) Show that 2-D parity can detect any three errors.
(b). Find four errors that cannot be detected by 2-D parity.
(c). Show that that 2-D parity cannot correct all two-bit errors. Hint: put both bits in the same row or
column.

15. Each of the following 8-bit messages with 4-bit Hamming code contains a single error. Correct the
message.

(a). 10100010 0111

(b). 10111110 1011

16. (a) What happens in 2-D parity if the corrupted bit is in the parity column or parity row?
(b). In the following 8-bit message with 4-bit Hamming code, there is an error in the code portion. How
can this be determined?
1001 1110 0100

108

5 Packets

6 ABSTRACT SLIDING WINDOWS

In this chapter we take a general look at how to build reliable data-transport layers on top of unreliable
lower layers. This is achieved through a retransmit-on-timeout policy; that is, if a packet is transmitted
and there is no acknowledgment received during the timeout interval then the packet is resent. As a class,
protocols where one side implements retransmit-on-timeout are known as ARQ protocols, for Automatic
Repeat reQuest.
In addition to reliability, we also want to keep as many packets in transit as the network can support. The
strategy used to achieve this is known as sliding windows. It turns out that the sliding-windows algorithm
is also the key to managing congestion; we return to this in 13 TCP Reno and Congestion Management.
The End-to-End principle, 12.1 The End-to-End Principle, suggests that trying to achieve a reliable transport layer by building reliability into a lower layer is a misguided approach; that is, implementing reliability
at the endpoints of a connection as is described here is in fact the correct mechanism.

6.1 Building Reliable Transport: Stop-and-Wait

Retransmit-on-timeout generally requires sequence numbering for the packets, though if a network path is
guaranteed not to reorder packets then it is safe to allow the sequence numbers to wrap around surprisingly
quickly (for stop-and-wait, a single-bit sequence number will work). However, as the no-reordering hypothesis does not apply to the Internet at large, we will assume conventional numbering. Data[N] will be the Nth
data packet, acknowledged by ACK[N].
In the stop-and-wait version of retransmit-on-timeout, the sender sends only one outstanding packet at a
time. If there is no response, the packet may be retransmitted, but the sender does not send Data[N+1] until
it has received ACK[N]. Of course, the receiving side will not send ACK[N] until it has received Data[N];
each side has only one packet in play at a time. In the absence of packet loss, this leads to the following:

109

An Introduction to Computer Networks, Release 1.0

6.1.1 Packet Loss

Lost packets, however, are a reality. The left half of the diagram below illustrates a lost Data packet, where
the sender is the host sending Data and the Receiver is the host sending ACKs. The receiver is not aware of
the loss; it sees Data[N] as simply slow to arrive.

110

6 Abstract Sliding Windows

An Introduction to Computer Networks, Release 1.0

The right half of the diagram, by comparison, illustrates the case of a lost ACK. The receiver has received
a duplicate Data[N]. We have assumed here that the receiver has implemented a retransmit-on-duplicate
strategy, and so its response upon receipt of the duplicate Data[N] is to retransmit ACK[N].
As a final example, note that it is possible for ACK[N] to have been delayed (or, similarly, for the first
Data[N] to have been delayed) longer than the timeout interval. Not every packet that times out is actually
lost!

In this case we see that, after sending Data[N], receiving a delayed ACK[N] (rather than the expected
ACK[N+1]) must be considered a normal event.
In principle, either side can implement retransmit-on-timeout if nothing is received. Either side can also
implement retransmit-on-duplicate; this was done by the receiver in the second example above but not by
the sender in the third example (the sender received a second ACK[N] but did not retransmit Data[N+1]).

6.1 Building Reliable Transport: Stop-and-Wait

111

An Introduction to Computer Networks, Release 1.0

At least one side must implement retransmit-on-timeout; otherwise a lost packet leads to deadlock as the
sender and the receiver both wait forever. The other side must implement at least one of retransmit-onduplicate or retransmit-on-timeout; usually the former alone. If both sides implement retransmit-on-timeout
with different timeout values, generally the protocol will still work.

6.1.2 Sorcerers Apprentice Bug

Sorcerers Apprentice
The Sorcerers Apprentice bug is named for the legend in which the apprentice casts a spell
on a broom to carry water, one bucket at a time. When the basin is full, the apprentice
chops the broom in half, only to find both halves carrying water. See Disneys Fantasia,
[Link] at around T = 5:35.
A strange thing happens if one side implements retransmit-on-timeout but both sides implement retransmiton-duplicate, as can happen if the implementer takes the naive view that retransmitting on duplicates is
safer; the moral here is that too much redundancy can be the Wrong Thing. Let us imagine that an
implementation uses this strategy, and that the initial ACK[3] is delayed until after Data[3] is retransmitted
on timeout. In the following diagram, the only packet retransmitted due to timeout is the second Data[3]; all
the other duplications are due to the retransmit-on-duplicate strategy.

112

6 Abstract Sliding Windows

An Introduction to Computer Networks, Release 1.0

All packets are sent twice from Data[3] on. The transfer completes normally, but takes double the normal
bandwidth.

6.1.3 Flow Control

Stop-and-wait also provides a simple form of flow control to prevent data from arriving at the receiver faster
than it can be handled. Assuming the time needed to process a received packet is less than one RTT, the
stop-and-wait mechanism will prevent data from arriving too fast. If the processing time is slightly larger
than RTT, all the receiver has to do is to wait to send ACK[N] until Data[N] has not only arrived but also
been processed, and the receiver is ready for Data[N+1].
For modest per-packet processing delays this works quite well, but if the processing delays are long it
introduces a new problem: Data[N] may time out and be retransmitted even though it has successfully been
received; the receiver cannot send an ACK until it has finished processing. One approach is to have two kinds
of ACKs: ACKWAIT [N] meaning that Data[N] has arrived but the receiver is not yet ready for Data[N+1],
and ACKGO [N] meaning that the sender may now send Data[N+1]. The receiver will send ACKWAIT [N]
when Data[N] arrives, and ACKGO [N] when it is done processing it.
Presumably we want the sender not to time out and retransmit Data[N] after ACKWAIT [N] is received, as
a retransmission would be unnecessary. This introduces a new problem: if the subsequent ACKGO [N] is
lost and neither side times out, the connection is deadlocked. The sender is waiting for ACKGO [N], which
is lost, and the receiver is waiting for Data[N+1], which the sender will not send until the lost ACKGO [N]
arrives. One solution is for the receiver to switch to a timeout model, perhaps until Data[N+1] is received.
TCP has a fix to the flow-control problem involving sender-side polling; see 12.16 TCP Flow Control.

6.2 Sliding Windows

Stop-and-wait is reliable but it is not very efficient (unless the path involves neither intermediate switches
nor significant propagation delay; that is, the path involves a single LAN link). Most links along a multi-hop
stop-and-wait path will be idle most of the time. During a file transfer, ideally we would like zero idleness
(at least along the slowest link; see 6.3 Linear Bottlenecks).
We can improve overall throughput by allowing the sender to continue to transmit, sending Data[N+1]
(and beyond) without waiting for ACK[N]. We cannot, however, allow the sender get too far ahead of the
returning ACKs. Packets sent too fast, as we shall see, simply end up waiting in queues, or, worse, dropped
from queues. If the links of the network have sufficient bandwidth, packets may also be dropped at the
receiving end.
Now that, say, Data[3] and Data[4] may be simultaneously in transit, we have to revisit what ACK[4] means:
does it mean that the receiver has received only Data[4], or does it mean both Data[3] and Data[4] have
arrived? We will assume the latter, that is, ACKs are cumulative: ACK[N] cannot be sent until Data[K] has
arrived for all KN.
The sender picks a window size, winsize. The basic idea of sliding windows is that the sender is allowed
to send this many packets before waiting for an ACK. More specifically, the sender keeps a state variable
last_ACKed, representing the last packet for which it has received an ACK from the other end; if data
packets are numbered starting from 1 then initially last_ACKed = 0.

6.2 Sliding Windows

113

An Introduction to Computer Networks, Release 1.0

Window Size
In this chapter we will assume winsize does not change. TCP, however, varies winsize up and down
with the goal of making it as large as possible without introducing congestion; we will return to this in
13 TCP Reno and Congestion Management.
At any instant, the sender may send packets numbered last_ACKed + 1 through last_ACKed + winsize; this
packet range is known as the window. Generally, if the first link in the path is not the slowest one, the sender
will most of the time have sent all these.
If ACK[N] arrives with N > last_ACKed (typically N = last_ACKed+1), then the window slides forward; we
set last_ACKed = N. This also increments the upper edge of the window, and frees the sender to send more
packets. For example, with winsize = 4 and last_ACKed = 10, the window is [11,12,13,14]. If ACK[11]
arrives, the window slides forward to [12,13,14,15], freeing the sender to send Data[15]. If instead ACK[13]
arrives, then the window slides forward to [14,15,16,17] (recall that ACKs are cumulative), and three more
packets become eligible to be sent. If there is no packet reordering and no packet losses (and every packet
is ACKed individually) then the window will slide forward in units of one packet at a time; the next arriving
ACK will always be ACK[last_ACKed+1].
Note that the rate at which ACKs are returned will always be exactly equal to the rate at which the slowest
link is delivering packets. That is, if the slowest link (the bottleneck link) is delivering a packet every
50 ms, then the receiver will receive those packets every 50 ms and the ACKs will return at a rate of
one every 50 ms. Thus, new packets will be sent at an average rate exactly matching the delivery rate;
this is the sliding-windows self-clocking property. Self-clocking has the effect of reducing congestion by
automatically reducing the senders rate whenever the available fraction of the bottleneck bandwidth is
reduced.
Here is a video of sliding windows in action, with winsize = 5. The second link, RB, has a capacity of five
packets in transit either way; the AR link has a capacity of one packet in transit either way.
The packets in the very first RTT represent connection setup. This particular video also demonstrates TCP
Slow Start: in the first data-packet RTT, two packets are sent, and in the second data RTT four packets
are sent. The full window size of five is reached in the third data RTT. For the rest of the connection, at
any moment (except those instants where packets have just been received) there are five packets in flight,
either being transmitted on a link as data, or being transmitted as an ACK, or sitting in a queue (this last
does not happen in the video).

6.2.1 Bandwidth Delay

As indicated previously (5.1 Packet Delay), the bandwidth RTT product represents the amount of data
that can be sent before the first response is received. This is generally the optimum size for the window size.
There is, however, one catch: if a sender chooses winsize larger than the bandwidth RTT product, then the
RTT simply grows due to queuing delays to the point that bandwidth RTT matches the chosen winsize.
That is, a connections own traffic can inflate RTTactual to well above RTTnoLoad ; see [Link] Case 3: winsize
= 6 below for an example. For this reason, a sender is often more interested in bandwidth RTTnoLoad , or,
at the very least, the RTT before the senders own packets had begun contributing to congestion.
We will sometimes refer to the bandwidth RTTnoLoad product as the transit capacity of the route. As
114

6 Abstract Sliding Windows

An Introduction to Computer Networks, Release 1.0

will become clearer below, a winsize smaller than this means underutilization of the network, while a larger
winsize means each packet spends time waiting in a queue somewhere.
Below are simplified diagrams for sliding windows with window sizes of 1, 4 and 6, each with a path
bandwidth of 6 packets/RTT (so bandwidth RTT = 6 packets). The diagram shows the initial packets sent
as a burst; these then would be spread out as they pass through the bottleneck link so that, after the first
burst, packet spacing is uniform. (Real sliding-windows protocols such as TCP generally attempt to avoid
such initial bursts.)

With winsize=1 we send 1 packet per RTT; with winsize=4 we always average 4 packets per RTT. To put
this another way, the three window sizes lead to bottle-neck link utilizations of 1/6, 4/6 and 6/6 = 100%,
respectively.
While it is tempting to envision setting winsize to bandwidth RTT, in practice this can be complicated;
neither bandwidth nor RTT is constant. Available bandwidth can fluctuate in the presence of competing
traffic. As for RTT, if a sender sets winsize too large then the RTT is simply inflated to the point that
bandwidth RTT matches winsize; that is, a connections own traffic can inflate RTTactual to well above
RTTnoLoad . This happens even in the absence of competing traffic.
6.2 Sliding Windows

115

An Introduction to Computer Networks, Release 1.0

6.2.2 The Receiver Side

Perhaps surprisingly, sliding windows can work pretty well with the receiver assuming that winsize=1, even
if the sender is in fact using a much larger value. Each of the receivers in the diagrams above receives
Data[N] and responds with ACK[N]; the only difference with the larger sender winsize is that the Data[N]
arrive faster.
If we are using the sliding-windows algorithm over single links, we may assume packets are never reordered,
and a receiver winsize of 1 works quite well. Once switches are introduced, however, life becomes more
complicated (though some links may do link-level sliding-windows for per-link throughput optimization).
If packet reordering occurs, it is common for the receiver to use the same winsize as the sender. This
means that the receiver must be prepared to buffer a full window full of packets. If the window is
[11,12,13,14,15,16], for example, and Data[11] is delayed, then the receiver may have to buffer Data[12]
through Data[16].
Like the sender, the receiver will also maintain the state variable last_ACKed, though it will not be
completely synchronized with the senders version. At any instant, the receiver is willing to accept
Data[lastACKed+1] through Data[lastACKed+winsize]. For any but the first of these, the receiver must
buffer the arriving packet. If Data[lastACKed+1] arrives, then the receiver should consult its buffers and
send back the largest cumulative ACK it can for the data received; for example, if the window is [11-16]
and Data[12], Data[13] and Data[15] are in the buffers, then on arrival of Data[11] the correct response is
ACK[13]. Data[11] fills the gap, and the receiver has now received everything up through Data[13]. The
new receive window is [14-19], and as soon as the ACK[13] reaches the sender that will be the new send
window as well.

6.2.3 Loss Recovery Under Sliding Windows

Suppose winsize = 4 and packet 5 is lost. It is quite possible that packets 6, 7, and 8 may have been received.
However, the only (cumulative) acknowledgment that can be sent back is ACK[4]; the sender does not
know how much of the windowful made it through. Because of the possibility that only Data[5] (or more
generally Data[lastACKed+1]) is lost, and because losses are usually associated with congestion, when we
most especially do not wish to overburden the network, the sender will usually retransmit only the first lost
packet, eg packet 5. If packets 6, 7, and 8 were also lost, then after retransmission of Data[5] the sender
will receive ACK[5], and can assume that Data[6] now needs to be sent. However, if packets 6-8 did make
it through, then after retransmission the sender will receive back ACK[8], and so will know 6-8 do not need
retransmission and that the next packet to send is Data[9].
Normally Data[6] through Data[8] would time out shortly after Data[5] times out. After the first timeout, however, sliding windows protocols generally suppress further timeout/retransmission responses until
recovery is more-or-less complete.
Once a full timeout has occurred, usually the sliding-windows process itself has ground to a halt, in that there
are usually no packets remaining in flight. This is sometimes described as pipeline drain. After recovery,
the sliding-windows process will have to start up again. Most implementations of TCP, as we shall see later,
implement a mechanism (fast recovery) for early detection of packet loss, before the pipeline has fully
drained.

116

6 Abstract Sliding Windows

An Introduction to Computer Networks, Release 1.0

6.3 Linear Bottlenecks

Consider the simple network path shown below, with bandwidths shown in packets/ms. The minimum
bandwidth, or path bandwidth, is 3 packets/ms.

The slow links are R2R3 and R3R4. We will refer to the slowest link as the bottleneck link; if there are
(as here) ties for the slowest link, then the first such link is the bottleneck. The bottleneck link is where the
queue will form. If traffic is sent at a rate of 4 packets/ms from A to B, it will pile up in an ever-increasing
queue at R2. Traffic will not pile up at R3; it arrives at R3 at the same rate by which it departs.
Furthermore, if sliding windows is used (rather than a fixed-rate sender), traffic will eventually not queue up
at any router other than R2: data cannot reach B faster than the 3 packets/ms rate, and so B will not return
ACKs faster than this rate, and so A will eventually not send data faster than this rate. At this 3 packets/ms
rate, traffic will not pile up at R1 (or R3 or R4).
There is a significant advantage in speaking in terms of winsize rather than transmission rate. If A sends to
B at any rate greater than 3 packets/ms, then the situation is unstable as the bottleneck queue grows without
bound and there is no convergence to a steady state. There is no analogous instability, however, if A uses
sliding windows, even if the winsize chosen is quite large (although a large-enough winsize will overflow
the bottleneck queue). If a sender specifies a sending window size rather than a rate, then the network will
converge to a steady state in relatively short order; if a queue develops it will be steadily replenished at the
same rate that packets depart, and so will be of fixed size.

6.3.1 Simple fixed-window-size analysis

We will analyze the effect of window size on overall throughput and on RTT. Consider the following network
path, with bandwidths now labeled in packets/second.

We will assume that in the backward BA direction, all connections are infinitely fast, meaning zero
delay; this is often a good approximation because ACK packets are what travel in that direction and they
are negligibly small. In the AB direction, we will assume that the AR1 link is infinitely fast, but
the other four each have a bandwidth of 1 packet/second (and no propagation-delay component). This
makes the R1R2 link the bottleneck link; any queue will now form at R1. The path bandwidth is 1
packet/second, and the RTT is 4 seconds.
As a roughly equivalent alternative example, we might use the following:

6.3 Linear Bottlenecks

117

An Introduction to Computer Networks, Release 1.0

with the following assumptions: the CS1 link is infinitely fast (zero delay), S1S2 and S2D each
take 1.0 sec bandwidth delay (so two packets take 2.0 sec, per link, etc), and ACKs also have a 1.0 sec
bandwidth delay in the reverse direction.
In both scenarios, if we send one packet, it takes 4.0 seconds for the ACK to return, on an idle network. This
means that the no-load delay, RTTnoLoad , is 4.0 seconds.
(These models will change significantly if we replace the 1 packet/sec bandwidth delay with a 1-second
propagation delay; in the former case, 2 packets take 2 seconds, while in the latter, 2 packets take 1 second.
See exercise 4.)
We assume a single connection is made; ie there is no competition. Bandwidth delay here is 4 packets (1
packet/sec 4 sec RTT)
[Link] Case 1: winsize = 2
In this case winsize < bandwidthdelay (where delay = RTT). The table below shows what is sent by A and
each of R1-R4 for each second. Every packet is acknowledged 4 seconds after it is sent; that is, RTTactual
= 4 sec, equal to RTTnoLoad ; this will remain true as the winsize changes by small amounts (eg to 1 or 3).
Throughput is proportional to winsize: when winsize = 2, throughput is 2 packets in 4 seconds, or 2/4 =
1/2 packet/sec. During each second, two of the routers R1-R4 are idle. The overall path will have less than
100% utilization.
Time
T
0
1
2
3
4
5
6
7
8
9

A
sends
1,2

3
4

5
6

R1
queues
2

R1
sends
1
2

3
4

5
6

R2
sends
1
2

3
4

R3
sends

1
2

3
4

R4
sends

1
2

3
4

B
ACKs

1
2

3
4

Note the brief pile-up at R1 (the bottleneck link!) on startup. However, in the steady state, there is no
queuing. Real sliding-windows protocols generally have some way of minimizing this initial pileup.
[Link] Case 2: winsize = 4
When winsize=4, at each second all four slow links are busy. There is again an initial burst leading to a brief
surge in the queue; RTTactual for Data[4] is 7 seconds. However, RTTactual for every subsequent packet is 4
seconds, and there are no queuing delays (and nothing in the queue) after T=2. The steady-state connection
118

6 Abstract Sliding Windows

An Introduction to Computer Networks, Release 1.0

throughput is 4 packets in 4 seconds, ie 1 packet/second. Note that overall path throughput now equals the
bottleneck-link bandwidth, so this is the best possible throughput.
T
0
1
2
3
4
5
6
7
8

A sends
1,2,3,4

R1 queues
2,3,4
3,4
4

5
6
7
8
9

R1 sends
1
2
3
4
5
6
7
8
9

R2 sends

R3 sends

R4 sends

B ACKs

1
2
3
4
5
6
7
8

1
2
3
4
5
6
7

1
2
3
4
5
6

1
2
3
4
5

At T=4, R1 has just finished sending Data[4] as Data[5] arrives from A; R1 can begin sending packet 5
immediately. No queue will develop.
Case 2 is the congestion knee of Chiu and Jain [CJ89], defined here in 1.7 Congestion.
[Link] Case 3: winsize = 6
T
0
1
2
3
4
5
6
7
8
9
10

A sends
1,2,3,4,5,6

7
8
9
10
11
12
13

R1 queues
2,3,4,5,6
3,4,5,6
4,5,6
5,6
6,7
7,8
8,9
9,10
10,11
11,12
12,13

R1 sends
1
2
3
4
5
6
7
8
9
10
11

R2 sends

R3 sends

R4 sends

B ACKs

1
2
3
4
5
6
7
8
9
10

1
2
3
4
5
6
7
8
9

1
2
3
4
5
6
7
8

1
2
3
4
5
6
7

Note that packet 7 is sent at T=4 and the acknowledgment is received at T=10, for an RTT of 6.0 seconds. All
later packets have the same RTTactual . That is, the RTT has risen from RTTnoLoad = 4 seconds to 6 seconds.
Note that we continue to send one windowful each RTT; that is, the throughput is still winsize/RTT, but RTT
is now 6 seconds.
One might initially conjecture that if winsize is greater than the bandwidthRTTnoLoad product, then the
entire window cannot be in transit at one time. In fact this is not the case; the sender does usually have the
entire window sent and in transit, but RTT has been inflated so it appears to the sender that winsize equals
the bandwidthRTT product.
In general, whenever winsize > bandwidthRTTnoLoad , what happens is that the extra packets pile up at
a router somewhere along the path (specifically, at the router in front of the bottleneck link). RTTactual is
inflated by queuing delay to winsize/bandwidth, where bandwidth is that of the bottleneck link; this means
winsize = bandwidthRTTactual . Total throughput is equal to that bandwidth. Of the 6 seconds of RTTactual
in the example here, a packet spends 4 of those seconds being transmitted on one link or another because

6.3 Linear Bottlenecks

119

An Introduction to Computer Networks, Release 1.0

RTTnoLoad =4. The other two seconds, therefore, must be spent in a queue; there is no other place for packets
wait. Looking at the table, we see that each second there are indeed two packets in the queue at R1.
If the bottleneck link is the very first link, packets may begin returning before the sender has sent the entire
windowful. In this case we may argue that the full windowful has at least been queued by the sender, and
thus has in this sense been sent. Suppose the network, for example, is

where, as before, each link transports 1 packet/sec from A to B and is infinitely fast in the reverse direction.
Then, if A sets winsize = 6, a queue of 2 packets will form at A.

6.3.2 RTT Calculations

We can make some quantitative observations of sliding windows behavior, and about queue utilization. First,
we note that RTTnoLoad is the physical travel time (subject to the limitations addressed in 5.2 Packet
Delay Variability); any time in excess of RTTnoLoad is spent waiting in a queue somewhere. Therefore, the
following holds regardless of competing traffic, and even for isolated packets:
1. queue_time = RTTactual RTTnoLoad
When the bottleneck link is saturated, that is, is always busy, the number of packets actually in transit (not
queued) somewhere along the path will always be bandwidth RTTnoLoad .
Second, we always send one windowful per actual RTT, assuming no losses and each packet is individually
acknowledged. This is perhaps best understood by consulting the diagrams above, but here is a simple nonvisual argument: if we send Data[N] at time TD , and ACK[N] arrives at time TA , then RTT = TA TD , by
definition. At time TA the sender is allowed to send Data[N+winsize], so during the RTT interval TD T
< TA the sender must have sent Data[N] through Data[N+winsize-1]; that is, winsize many packets in time
RTT. Therefore (whether or not there is competing traffic) we always have
2. throughput = winsize/RTTactual
where throughput is the rate at which the connection is sending packets.
In the sliding windows steady state, where throughput and RTTactual are reasonably constant, the average
number of packets in the queue is just throughputqueue_time (where throughput is measured in packets/sec):
3. queue_usage = throughput (RTTactual RTTnoLoad )
= winsize (1 RTTnoLoad /RTTactual )
To give a little more detail making the averaging perhaps clearer, each packet spends time (RTTactual
RTTnoLoad ) in the queue, from equation 1 above. The total time spent by a windowful of packets is winsize
(RTTactual RTTnoLoad ), and dividing this by RTTactual thus gives the average number of packets in the
queue over the RTT interval in question.
In the presence of competing traffic, the throughput referred to above is simply the connections current
share of the total bandwidth. It is the value we get if we measure the rate of returning ACKs. If there is

120

6 Abstract Sliding Windows

An Introduction to Computer Networks, Release 1.0

no competing traffic and winsize is below the congestion knee winsize < bandwidth RTTnoLoad then
winsize is the limiting factor in throughput. Finally, if there is no competition and winsize bandwidth
RTTnoLoad then the connection is using 100% of the capacity of the bottleneck link and throughput is equal
to the bottleneck-link physical bandwidth. To put this another way,
4. RTTactual = winsize/bottleneck_bandwidth
queue_usage = winsize bandwidth RTTnoLoad
Dividing the first equation by RTTnoLoad , and noting that bandwidth RTTnoLoad = winsize - queue_usage
= transit_capacity, we get
5. RTTactual /RTTnoLoad = winsize/transit_capacity = (transit_capacity + queue_usage) / transit_capacity
Regardless of the value of winsize, in the steady state the sender never sends faster than the bottleneck
bandwidth. This is because the bottleneck bandwidth determines the rate of packets arriving at the far
end, which in turn determines the rate of ACKs arriving back at the sender, which in turn determines the
continued sending rate. This illustrates the self-clocking nature of sliding windows.
We will return in 14 Dynamics of TCP Reno to the issue of bandwidth in the presence of competing traffic.
For now, suppose a sliding-windows sender has winsize > bandwidth RTTnoLoad , leading as above to a
fixed amount of queue usage, and no competition. Then another connection starts up and competes for the
bottleneck link. The first connections effective bandwidth will thus decrease. This means that bandwidth
RTTnoLoad will decrease, and hence the connections queue usage will increase.

6.3.3 Graphs at the Congestion Knee

Consider the following graphs of winsize versus
1. throughput
2. delay
3. queue utilization

The critical winsize value is equal to bandwidth RTTnoLoad ; this is known as the congestion knee. For
winsize below this, we have:
throughput is proportional to winsize

6.3 Linear Bottlenecks

121

An Introduction to Computer Networks, Release 1.0

delay is constant
queue utilization in the steady state is zero
For winsize larger than the knee, we have
throughput is constant (equal to the bottleneck bandwidth)
delay increases linearly with winsize
queue utilization increases linearly with winsize
Ideally, winsize will be at the critical knee. However, the exact value varies with time: available bandwidth
changes due to the starting and stopping of competing traffic, and RTT changes due to queuing. Standard
TCP makes an effort to stay well above the knee much of the time, presumably on the theory that maximizing
throughput is more important than minimizing queue use.
The power of a connection is defined to be throughput/RTT. For sliding windows below the knee, RTT is
constant and power is proportional to the window size. For sliding windows above the knee, throughput is
constant and delay is proportional to winsize; power is thus proportional to 1/winsize. Here is a graph, akin
to those above, of winsize versus power:

6.3.4 Simple Packet-Based Sliding-Windows Implementation

Here is a pseudocode outline of the receiver side of a sliding-windows implementation. Winsize is W; we
abbreviate lastACKed by LA. Thus, the next packet expected is LA+1 and the window is LA+1 ... LA+W.
We have a pool EA of early arrivals: packets buffered for future use. When packet Data[M] arrives:
if MLA or M>LA+W, ignore the packet
if M>LA+1, put the packet into EA.
if M==LA+1,
output the packet (that is, Data[LA+1])
LA = LA+1 (slide window forward by 1)
while (Data[LA+1] is in EA) {
output Data[LA+1]
LA = LA+1

122

6 Abstract Sliding Windows

An Introduction to Computer Networks, Release 1.0

}
send ACK[LA]
There are a couple details omitted.
A possible implementation of EA is as an array of packet objects, of size W. We always put packet Data[K]
into position K % W.
At any point between packet arrivals, Data[LA+1] is not in EA, but some later packets may be present.
For the sender side, we begin by sending a full windowful of packets Data[1] through Data[W], and setting
LA=0. When ACK[M] arrives:
if MLA or M>LA+W, ignore the packet
otherwise:
set K = LA+1
set LA = M, the new bottom edge of the window
for (i=K; iLA; i++) send Data[i]
Note that new ACKs may arrive while we are in the loop at that last line. We assume here that the sender
stolidly sends what it may send and only after that does it start to process arriving ACKs. Some implementations may take a more asynchronous approach, perhaps with one thread processing arriving ACKs and
incrementing LA and another thread sending everything it is allowed to send.

6.4 Epilog
This completes our discussion of the sliding-windows algorithm in the abstract setting. We will return to
concrete implementations of this in 11.3.2 TFTP Stop-and-Wait (stop-and-wait) and in 12.13 TCP Sliding
Windows; the latter is one of the most important mechanisms on the Internet.

6.5 Exercises
1. Sketch a ladder diagram for stop-and-wait if Data[3] is lost the first time it is sent. Continue the diagram
to the point where Data[4] is successfully transmitted. Assume an RTT of 1 second, no sender timeout (but
the sender retransmits on duplicate), and a receiver timeout of 2 seconds.
2. Suppose a stop-and-wait receiver has an implementation flaw. When Data[1] arrives, ACK[1] and ACK[2]
are sent, separated by a brief interval; after that, the receiver transmits ACK[N+1] when Data[N] arrives,
rather than the correct ACK[N].

(a). Assume the sender responds to each ACK as it arrives. What is the first ACK[N] that it will be able to
determine is incorrect? Assume no packets are lost.
(b). Is there anything the transmitter can do to detect this receiver problem earlier?

3.
A

Create a table as in 6.3.1

Simple fixed-window-size analysis for the original
R1 R2
R3 R4
B network with winsize = 8. As in the text examples, assume 1

6.4 Epilog

123

An Introduction to Computer Networks, Release 1.0

packet/sec bandwidth delay for the R1R2, R2R3, R3R4 and R4B links. The AR link and
all reverse links (from B to A) are infinitely fast. Carry out the table for 10 seconds.
B. The A
4. Create a table as in 6.3.1 Simple fixed-window-size analysis for a network A R1 R2
R1 ink is infinitely fast; the R1R2 and R2B each have a 1-second propagation delay, in each direction,
and zero bandwidth delay (that is, one packet takes 1.0 sec to travel from R1 to R2; two packets also take
1.0 sec to travel from R1 to R2). Assume winsize=6. Carry out the table for 8 seconds. Note that with zero
bandwidth delay, multiple packets sent together will remain together until the destination.
5. Suppose RTTnoLoad = 4 seconds and the bottleneck bandwidth is 1 packet every 2 seconds.

(a). What window size is needed to remain just at the knee of congestion?
(b). Suppose winsize=6. How many packets are in the queue, at the steady state, and what is RTTactual ?

6. Create a table as in 6.3.1 Simple fixed-window-size analysis for a network A R1

R2 R3 B.
The AR1 link is infinitely fast. The R1R2 and R3B links have a bandwidth delay of 1 packet/second
with no additional propagation delay. The R2R3 link has a bandwidth delay of 1 packet / 2 seconds, and no
propagation delay. The reverse BA direction (for ACKs) is infinitely fast. Assume winsize = 6. Carry
out the table for 10 seconds. Note that in this exercise you will need to show the queue for both R1 and
R2. (If you carry out the pattern at least partially until T=18, you can verify that RTTactual for packet 8 is as
calculated in the previous exercise. You will need more than 10 packets, but fewer than 16; the use of hex
labels A, B, C for packets 10, 11, 12 is a convenient notation.)
Hint: The column for R2 sends (or, more literally, R2 is in the process of sending) should look like this:
T
0
1
2
3
4
5
6
...

R2 sends
1
1
2
2
3
3
...

7. Argue that, if A sends to B using sliding windows, and in the path from A to B the slowest link is not the
first link out of A, then eventually A will have the entire window outstanding (except at the instant just after
each new ACK comes in).
8. Suppose RTTnoLoad is 50 ms and the available bandwidth is 2,000 packets/sec. Sliding windows is used
for transmission.

(a). What window size is needed to remain just at the knee of congestion?
(b). If RTTactual rises to 60 ms (due to use of a larger winsize), how many packets are in a queue at any one
time?
(c). What value of winsize would lead to RTTactual = 60 ms?
(d). What value of winsize would make RTTactual rise to 100 ms?
124

6 Abstract Sliding Windows

An Introduction to Computer Networks, Release 1.0

9. Suppose winsize=4 in a sliding-windows connection, and assume that while packets may be lost, they
are never reordered (that is, if two packets P1 and P2 are sent in that order, and both arrive, then they arrive
in that order). Show that if Data[8] is in the receivers window (meaning that everything up through Data[4]
has been received and acknowledged), then it is no longer possible for even a late Data[0] to arrive at the
receiver. (A consequence of the general principle here is that in the absence of reordering we can replace
the packet sequence number with (sequence_number) mod (2winsize+1) without ambiguity.)
10. Suppose winsize=4 in a sliding-windows connection, and assume as in the previous exercise that while
packets may be lost, they are never reordered. Give an example in which Data[8] is in the receivers window
(so the receiver has presumably sent ACK[4]), and yet Data[1] legitimately arrives. (Thus, the late-packet
bound in the previous exercise is the best possible.)
11. Suppose the network is A R1
R2 B, where the AR1 ink is infinitely fast and the R1R2
link has a bandwidth of 1 packet/second each way, for an RTTnoLoad of 2 seconds. Suppose also that A
begins sending with winsize = 6. By the analysis in [Link] Case 3: winsize = 6, RTT should rise to
winsize/bandwidth = 6 seconds. Give the RTTs of the first eight packets. How long does it take for RTT to
rise to 6 seconds?

6.5 Exercises

125

An Introduction to Computer Networks, Release 1.0

126

6 Abstract Sliding Windows

7 IP VERSION 4

There are multiple LAN protocols below the IP layer and multiple transport protocols above, but IP itself
stands alone. The Internet is essentially the IP Internet. If you want to run your own LAN somewhere or if
you want to run your own transport protocol, the Internet backbone will still work for you. But if you want
to change the IP layer, you may encounter difficulty. (Just talk to the IPv6 people, or the IP-multicasting or
IP-reservations groups.)
IP is, in effect, a universal routing and addressing protocol. The two are developed together; every node
has an IP address and every router knows how to handle IP addresses. IP was originally seen as a way to
interconnect multiple LANs, but it may make more sense now to view IP as a virtual LAN overlaying all
the physical LANs.
A crucial aspect of IP is its scalability. As the Internet has grown to ~109 hosts, the forwarding tables are
not much larger than 105 (perhaps now 105.5 ). Ethernet, in comparison, scales poorly.
Furthermore, IP, unlike Ethernet, offers excellent support for multiple redundant links. If the network
below were an IP network, each node would communicate with each immediate neighbor via their shared
direct link. If, on the other hand, this were an Ethernet network with the spanning-tree algorithm, then one
of the four links would simply be disabled completely.

The IP network service model is to act like a LAN. That is, there are no acknowledgments; delivery is
generally described as best-effort. This design choice is perhaps surprising, but it has also been quite fruitful.
Currently the Internet uses (almost exclusively) IP version 4, with its 32-bit address size. As the Internet has
run out of new large blocks of IPv4 addresses (1.10 IP - Internet Protocol), there is increasing pressure to
convert to IPv6, with its 128-bit address size. Progress has been slow, however, and delaying tactics such
as IPv4-address markets and NAT have allowed IPv4 to continue. Aside from the major change in address
structure, there are relatively few differences in the routing models of IPv4 and IPv6. We will study IPv4 in
this chapter and IPv6 in the following.
If you want to provide a universal service for delivering any packet anywhere, what else do you need besides
routing and addressing? Every network (LAN) needs to be able to carry any packet. The protocols spell
out the use of octets (bytes), so the only possible compatibility issue is that a packet is too large for a given
network. IP handles this by supporting fragmentation: a network may break a too-large packet up into
units it can transport successfully. While IP fragmentation is inefficient and clumsy, it does guarantee that
any packet can potentially be delivered to any node.

7.1 The IPv4 Header

The IPv4 Header needs to contain the following information:

127

An Introduction to Computer Networks, Release 1.0

destination and source addresses

indication of ipv4 versus ipv6
a Time To Live (TTL) value, to prevent infinite routing loops
a field indicating what comes next in the packet (eg TCP v UDP)
fields supporting fragmentation and reassembly.
Here is how it is all laid out in the header. Each row is 32 bits wide.

The IP header, and basics of IP protocol operation, were defined in RFC 791; some minor changes have
since occurred. Most of these changes were documented in RFC 1122, though the DS field was defined in
RFC 2474 and the ECN bits were first proposed in RFC 2481.
The Version field is, for IPv4, the number 4: 0100. The IHL field represents the total IP Header Length, in
32-bit words; an IP header can thus be at most 15 words long. The base header takes up five words, so the
IP Options can consist of at most ten words. If one looks at IPv4 packets using a packet-capture tool that
displays the packets in hex, the first byte will most often be 0x45.
The Differentiated Services (DS) field is used by the Differentiated Services suite to specify preferential handling for designated packets, eg those involved in VoIP or other real-time protocols. The Explicit
Congestion Notification bits are there to allow routers experiencing congestion to mark packets, thus indicating to the sender that the transmission rate should be reduced. We will address these in 14.8.2 Explicit
Congestion Notification (ECN). These two fields together replace the old 8-bit Type of Service field.
The Total Length field is present because an IP packet may be smaller than the minimum LAN packet
size (see Exercise 1) or larger than the maximum (if the IP packet has been fragmented over several LAN
packets. The IP packet length, in other words, cannot be inferred from the LAN-level packet size. Because
the Total Length field is 16 bits, the maximum IP packet size is 216 bytes. This is probably much too large,
even if fragmentation were not something to be avoided.
The second word of the header is devoted to fragmentation, discussed below.
The Time-to-Live (TTL) field is decremented by 1 at each router; if it reaches 0, the packet is discarded. A
typical initial value is 64; it must be larger than the total number of hops in the path. In most cases, a value of
32 would work. The TTL field is there to prevent routing loops always a serious problem should they occur
from consuming resources indefinitely. Later we will look at various IP routing-table update protocols and
how they minimize the risk of routing loops; they do not, however, eliminate it. By comparison, Ethernet
headers have no TTL field, but Ethernet also disallows cycles in the underlying topology.
128

7 IP version 4

An Introduction to Computer Networks, Release 1.0

The Protocol field contains a value to indicate if the body of the IP packet represents a TCP packet or a
UDP packet, or, in unusual cases, something else altogether.
The Header Checksum field is the Internet checksum applied to the header only, not the body. Its only
purpose is to allow the discarding of packets with corrupted headers. When the TTL value is decremented
the router must update the header checksum. This can be done algebraically by adding a 1 in the correct
place to compensate, but it is not hard simply to re-sum the 8 halfwords of the average header.
The Source and Destination Address fields contain, of course, the IP addresses. These would be updated
only by NAT firewalls.
One option is the Record Route option, in which routers are to insert their own IP address into the IP
header option area. Unfortunately, with only ten words available, there is not enough space to record most
longer routes (but see 7.9.1 Traceroute and Time Exceeded, below). Another option, now deprecated as
security risk, is to support source routing. The sender would insert into the IP header option area a list of IP
addresses; the packet would be routed to pass through each of those IP addresses in turn. With strict source
routing, the IP addresses had to represent adjacent neighbors; no router could be used if its IP address were
not on the list. With loose source routing, the listed addresses did not have to represent adjacent neighbors
and ordinary IP routing was used to get from one listed IP address to the next. Both forms are essentially
never used, again for security reasons: if a packet has been source-routed, it may have been routed outside
of the at-least-somewhat trusted zone of the Internet backbone.

7.2 Interfaces
IP addresses (IPv4 and IPv6) are, strictly speaking, assigned not to hosts or nodes, but to interfaces. In the
most common case, where each node has a single LAN interface, this is a distinction without a difference. In
a room full of workstations each with a single Ethernet interface eth0 (or perhaps Ethernet adapter
Local Area Connection), we might as well view the IP address assigned to the interface as assigned
to the workstation itself.
Each of those workstations, however, likely also has a loopback interface (at least conceptually), providing
a way to deliver IP packets to other processes on the same machine. On many systems, the name localhost
resolves to the IPv4 address [Link] (although the IPv6 address ::1 is also used). Delivering packets to
the localhost address is simply a form of interprocess communication; a functionally similar alternative is
named pipes. Loopback delivery avoids the need to use the LAN at all, or even the need to have a LAN.
For simple client/server testing, it is often convenient to have both client and server on the same machine,
in which case the loopback interface is convenient and fast. On unix-based machines the loopback interface
represents a genuine logical interface, commonly named lo. On Windows systems the interface may not
represent an actual entity, but this is of practical concern only to those interested in sniffing all loopback
traffic; packets sent to the loopback address are still delivered as expected.
Workstations often have special other interfaces as well. Most recent versions of Microsoft Windows have
a Teredo Tunneling pseudo-interface and an Automatic Tunneling pseudo-interface; these are both intended
(when activated) to support IPv6 connectivity when the local ISP supports only IPv4. The Teredo protocol
is documented in RFC 4380.
When VPN connections are created, as in 3.1 Virtual Private Network, each end of the logical connection
typically terminates at a virtual interface (one of these is labeled tun0 in the diagram of 3.1 Virtual Private

7.2 Interfaces

129

An Introduction to Computer Networks, Release 1.0

Network). These virtual interfaces appear, to the systems involved, to be attached to a point-to-point link
that leads to the other end.
When a computer hosts a virtual machine, there is almost always a virtual network to connect the host and
virtual systems. The host will have a virtual interface to connect to the virtual network. The host may act as
a NAT router for the virtual machine, hiding that virtual machine behind its own IP address, or it may act
as an Ethernet switch, in which case the virtual machine will need an additional public IP address.
Whats My IP Address?
This simple-seeming question is in fact not very easy to answer, if by my IP address one means the IP
address assigned to the interface that connects directly to the Internet. One strategy is to find the address
of the default router, and then iterate through all interfaces (eg with the Java NetworkInterface
class) to find an IP address with a matching network prefix. Unfortunately, finding the default router
is hard to do in an OS-independent way, and even then this approach can fail if the Wi-Fi and Ethernet
interfaces both are assigned IP addresses on the same network, but only one is actually connected.
Many workstations have both an Ethernet interface and a Wi-Fi interface. Both of these can be used simultaneously (with different IP addresses assigned to each), either on the same IP network or on different IP
networks.
Routers always have at least two interfaces on two separate LANs. Generally this means a separate IP
address for each interface, though some point-to-point interfaces can be used without being assigned an IP
address (7.10 Unnumbered Interfaces).
Finally, it is usually possible to assign multiple IP addresses to a single interface. Sometimes this is done
to allow two IP networks to share a single LAN; the interface might be assigned one IP address for each
IP network. Other times a single interface is assigned multiple IP addresses that are on the same LAN; this
is often done so that one physical machine can act as a server (eg a web server) for multiple distinct IP
addresses corresponding to multiple distinct domain names.
While it is important to be at least vaguely aware of all these special cases, we emphasize again that in most
ordinary contexts each end-user workstation has one IP address that corresponds to a LAN connection.

7.3 Special Addresses

A few IP addresses represent special cases.
While the standard IPv4 loopback address is [Link], any IP address beginning with 127 can serve as a
loopback address. Logically they all represent the current host. Most hosts are configured to resolve the
name localhost to [Link]. However, any loopback address eg. [Link] should work, eg with
ping.
Private addresses are IP addresses intended only for site internal use, eg either behind a NAT firewall or
intended to have no Internet connectivity at all. If a packet shows up at any non-private router (eg at an
ISP router), with a private IP address as either source or destination address, the packet should be dropped.
Three standard private-address blocks have been defined:
[Link]/8

130

7 IP version 4

An Introduction to Computer Networks, Release 1.0

[Link]/12
[Link]/16
The last block is the one from which addresses are most commonly allocated by DHCP servers
(7.8.1 DHCP and the Small Office) built into NAT routers.
Broadcast addresses are a special form of IP address intended to be used in conjunction with LAN-layer
broadcast. The most common forms are broadcast to this network, consisting of all 1-bits, and broadcast
to network D, consisting of Ds network-address bits followed by all 1-bits for the host bits. If you try to
send a packet to the broadcast address of a remote network D, the odds are that some router involved will
refuse to forward it, and the odds are even higher that, once the packet arrives at a router actually on network
D, that router will refuse to broadcast it. Even addressing a broadcast to ones own network will fail if the
underlying LAN does not support LAN-level broadcast (eg ATM).
The highly influential early Unix implementation Berkeley 4.2 BSD used 0-bits for the broadcast bits, instead of 1s. As a result, to this day host bits cannot be all 1-bits or all 0-bits in order to avoid confusion with
the IP broadcast address. One consequence of this is that a Class C network has 254 usable host addresses,
not 256.

7.3.1 Multicast addresses

Finally, IP multicast addresses remain as the last remnant of the Class A/B/C strategy: multicast addresses
are Class D, with first byte beginning 1110 (meaning that the first byte is, in decimal, 224-239). Multicasting
means delivering to a specified set of addresses, preferably by some mechanism more efficient than sending
to each address individually. A reasonable goal of multicast would be that no more than one copy of the
multicast packet traverses any given link.
Support for IP multicast requires considerable participation by the backbone routers involved. For example,
if hosts A, B and C each connect to different interfaces of router R1, and A wishes to send a multicast packet
to B and C, then it is up to R1 to receive the packet, figure out that B and C are the intended recipients, and
forward the packet twice, once for Bs interface and once for Cs. R1 must also keep track of what hosts have
joined the multicast group and what hosts have left. Due to this degree of router participation, backbone
router support for multicasting has not been entirely forthcoming. A discussion of IP multicasting appears
in 19 Quality of Service.

7.4 Fragmentation
If you are trying to interconnect two LANs (as IP does), what else might be needed besides Routing and
Addressing? IP explicitly assumes all packets are composed on 8-bit bytes (something not universally true
in the early days of IP; to this day the RFCs refer to octets to emphasize this requirement). IP also defines
bit-order within a byte, and it is left to the networking hardware to translate properly. Neither byte size nor
bit order, therefore, can interfere with packet forwarding.
There is one more feature IP must provide, however, if the goal is universal connectivity: it must accommodate networks for which the maximum packet size, or Maximum Transfer Unit, MTU, is smaller than
the packet that needs forwarding. Otherwise, if we were using IP to join Token Ring (MTU = 4KB, at least

7.4 Fragmentation

131

An Introduction to Computer Networks, Release 1.0

originally) to Ethernet (MTU = 1500B), the token-ring packets might be too large to deliver to the Ethernet side, or to traverse an Ethernet backbone en route to another Token Ring. (Token Ring, in its day, did
commonly offer a configuration option to allow Ethernet interoperability.)
So, IP must support fragmentation, and thus also reassembly. There are two major strategies here: per-link
fragmentation and reassembly, where the reassembly is done at the opposite end of the link (as in ATM), and
path fragmentation and reassembly, where reassembly is done at the far end of the path. The latter approach
is what is taken by IP, partly because intermediate routers are too busy to do reassembly (this is as true today
as it was thirty years ago), and partly because IP fragmentation is seen as the strategy of last resort.
An IP sender is supposed to use a different value for the IDENT field for different packets, at least up until
the field wraps around. When an IP datagram is fragmented, the fragments keep the same IDENT field, so
this field in effect indicates which fragments belong to the same packet.
After fragmentation, the Fragment Offset field marks the start position of the data portion of this fragment
within the data portion of the original IP packet. Note that the start position can be a number up to 216 , the
maximum IP packet length, but the FragOffset field has only 13 bits. This is handled by requiring the data
portions of fragments to have sizes a multiple of 8 (three bits), and left-shifting the FragOffset value by 3
bits before using it.
As an example, consider the following network, where MTUs are excluding the LAN header:

Suppose A addresses a packet of 1500 bytes to B, and sends it via the LAN to the first router R1. The packet
contains 20 bytes of IP header and 1480 of data.
R1 fragments the original packet into two packets of sizes 20+976 = 996 and 20+504=544. Having 980
bytes of payload in the first fragment would fit, but violates the rule that the sizes of the data portions be
divisible by 8. The first fragment packet has FragOffset = 0; the second has FragOffset = 976.
R2 refragments the first fragment into three packets as follows:
first: size = 20+376=396, FragOffset = 0
second: size = 20+376=396, FragOffset = 376
third: size = 20+224 = 244 (note 376+376+224=976), FragOffset = 752.
R2 refragments the second fragment into two:
first: size = 20+376 = 396, FragOffset = 976+0 = 976
second: size = 20+128 = 148, FragOffset = 976+376=1352
R3 then sends the fragments on to B, without reassembly.
Note that it would have been slightly more efficient to have fragmented into four fragments of sizes 376,
376, 376, and 352 in the beginning. Note also that the packet format is designed to handle fragments of
different sizes easily. The algorithm is based on multiple fragmentation with reassembly only at the final
destination.
Each fragment has its IP-header Total Length field set to the length of that fragment.

132

7 IP version 4

An Introduction to Computer Networks, Release 1.0

We have not yet discussed the three flag bits. The first bit is reserved, and must be 0. The second bit is the
Dont Fragment bit. If it is set to 1 by the sender then a router must not fragment the packet and must
drop it instead; see 12.12 Path MTU Discovery for an application of this. The third bit is set to 1 for all
fragments except the final one (this bit is thus set to 0 if no fragmentation has occurred). The third bit tells
the receiver where the fragments stop.
The receiver must take the arriving fragments and reassemble them into a whole packet. The fragments
may not arrive in order unlike in ATM networks and may have unrelated packets interspersed. The
reassembler must identify when different arriving packets are fragments of the same original, and must
figure out how to reassemble the fragments in the correct order; both these problems were essentially trivial
for ATM.
Fragments are considered to belong to the same packet if they have the same IDENT field and also the same
source and destination addresses and same protocol.
As all fragment sizes are a multiple of 8 bytes, the receiver can keep track of whether all fragments have
been received with a bitmap in which each bit represents one 8-byte fragment chunk. A 1 KB packet could
have up to 128 such chunks; the bitmap would thus be 16 bytes.
If a fragment arrives that is part of a new (and fragmented) packet, a buffer is allocated. While the receiver
cannot know the final size of the buffer, it can usually make a reasonable guess. Because of the FragOffset
field, the fragment can then be stored in the buffer in the appropriate position. A new bitmap is also allocated,
and a reassembly timer is started.
As subsequent fragments arrive, not necessarily in order, they too can be placed in the proper buffer in the
proper position, and the appropriate bits in the bitmap are set to 1.
If the bitmap shows that all fragments have arrived, the packet is sent on up as a completed IP packet. If, on
the other hand, the reassembly timer expires, then all the pieces received so far are discarded.
TCP connections usually engage in Path MTU Discovery, and figure out the largest packet size they can
send that will not entail fragmentation (12.12 Path MTU Discovery). But it is not unusual, for example,
for UDP protocols to use fragmentation, especially over the short haul. In the Network File Sharing (NFS)
protocol, for example, UDP is used to carry 8KB disk blocks. These are often sent as a single 8+ KB IP
packet, fragmented over Ethernet to five full packets and a fraction. Fragmentation works reasonably well
here because most of the time the packets do not leave the Ethernet they started on. Note that this is an
example of fragmentation done by the sender, not by an intermediate router.
Finally, any given IP link may provide its own link-layer fragmentation and reassembly; we saw in
3.8.1 ATM Segmentation and Reassembly that ATM does just this. Such link-layer mechanisms are, however, generally invisible to the IP layer.

7.5 The Classless IP Delivery Algorithm

Recall from Chapter 1 that any IP address can be divided into a net portion IPnet and a host portion IPhost ;
the division point was determined by whether the IP address was a Class A, a Class B, or a Class C. We
also hinted in Chapter 1 that the division point was not always so clear-cut; we now present the delivery
algorithm, for both hosts and routers, that does not assume a globally predeclared division point of the input
IP address into net and host portions. We will, for the time being, punt on the question of forwarding-table

7.5 The Classless IP Delivery Algorithm

133

An Introduction to Computer Networks, Release 1.0

lookup and assume there is a lookup() method available that, when given a destination address, returns
the next_hop neighbor.
Instead of class-based divisions, we will assume that each of the IP addresses assigned to a nodes interfaces
is configured with an associated length of the network prefix; following the slash notation of 1.10 IP Internet Protocol, if B is an address and the prefix length is k = kB then the prefix itself is B/k. As usual, an
ordinary host may have only one IP interface, while a router will always have multiple interfaces.
Let D be the given IP destination address; we want to decide if D is local or nonlocal. The host or router
involved may have multiple IP interfaces, but for each interface the length of the network portion of the
address will be known. For each network address B/k assigned to one of the hosts interfaces, we compare
the first k bits of B and D; that is, we ask if D matches B/k.
If one of these comparisons yields a match, delivery is local; the host delivers the packet to its final
destination via the LAN connected to the corresponding interface. This means looking up the LAN
address of the destination, if applicable, and sending the packet to that destination via the interface.
If there is no match, delivery is nonlocal, and the host passes D to the lookup() routine of the
forwarding table and sends to the associated next_hop (which must represent a physically connected
neighbor). It is now up to lookup() routine to make any necessary determinations as to how D
might be split into Dnet and Dhost .
The forwarding table is, abstractly, a set of network addresses now also with lengths each of the form
B/k, with an associated next_hop destination for each. The lookup() routine will, in principle, compare
D with each table entry B/k, looking for a match (that is, equality of the first k bits). As with the interfaces
check above, the net/host division point (that is, k) will come from the table entry; it will not be inferred
from D or from any other information borne by the packet. There is, in fact, no place in the IP header to
store a net/host division point, and furthermore different routers along the path may use different values of k
with the same destination address D. In 10 Large-Scale IP Routing we will see that in some cases multiple
matches in the forwarding table may exist; the longest-match rule will be introduced to pick the best match.
Here is a simple example for a router with immediate neighbors A-E:
destination
[Link]/16
[Link]/24
[Link]/24
[Link]/24
[Link]/24

next_hop
A
B
C
D
E

The IP addresses [Link] and [Link] both route to A. The addresses [Link], [Link] and
[Link] route to B, C and D respectively. Finally, [Link] matches both A and E, but the E match is
longer so the packet is routed that way.
The forwarding table may also contain a default entry for the next_hop, which it may return in cases when
the destination D does not match any known network. We take the view here that returning such a default
entry is a valid result of the routing-table lookup() operation, rather than a third option to the algorithm
above; one approach is for the default entry to be the next_hop corresponding to the destination [Link]/0,
which does indeed match everything (use of this would definitely require the above longest-match rule,
though).
Default routes are hugely important in keeping leaf forwarding tables small. Even backbone routers some-

134

7 IP version 4

An Introduction to Computer Networks, Release 1.0

times expend considerable effort to keep the network address prefixes in their forwarding tables as short as
possible, through consolidation.
Routers may also be configured to allow passing quality-of-service information to the lookup() method,
as mentioned in Chapter 1, to support different routing paths for different kinds of traffic (eg bulk file-transfer
versus real-time).
For a modest exception to the local-delivery rule described here, see below in 7.10 Unnumbered Interfaces.

7.6 IP Subnets
Subnets were the first step away from Class A/B/C routing: a large network (eg a class A or B) could be
divided into smaller IP networks called subnets. Consider, for example, a typical Class B network such
as Loyola Universitys (originally [Link]/16); the underlying assumption is that any packet can be
delivered via the underlying LAN to any internal host. This would require a rather large LAN, and would
require that a single physical LAN be used throughout the site. What if our site has more than one physical
LAN? Or is really too big for one physical LAN? It did not take long for the IP world to run into this
problem.
Subnets were first proposed in RFC 917, and became official with RFC 950.
Getting a separate IP network prefix for each subnet is bad for routers: the backbone forwarding tables now
must have an entry for every subnet instead of just for every site. What is needed is a way for a site to appear
to the outside world as a single IP network, but for further IP-layer routing to be supported inside the site.
This is what subnets accomplish.
Subnets introduce hierarchical routing: first we route to the primary network, then inside that site we route
to the subnet, and finally the last hop delivers to the host.
Routing with subnets involves in effect moving the IPnet division line rightward. (Later, when we consider
CIDR, we will see the complementary case of moving the division line to the left.) For now, observe that
moving the line rightward within a site does not affect the outside world at all; outside routers are not even
aware of site-internal subnetting.
In the following diagram, the outside world directs traffic addressed to [Link]/16 to the router R. Internally, however, the site is divided into subnets. The idea is that traffic from [Link]/24 to [Link]/24
is routed, not switched; the two LANs involved may not even be compatible. Most of the subnets shown
are of size /24, meaning that the third byte of the IP address has become part of the network portion of the
subnets address; one /20 subnet is also shown. RFC 950 would have disallowed the subnet with third byte
0, but having 0 for the subnet bits generally does work.

7.6 IP Subnets

135

An Introduction to Computer Networks, Release 1.0

What we want is for the internal routing to be based on the extended network prefixes shown, while externally continuing to use only the single routing entry for [Link]/16.
To implement subnets, we divide the sites IP network into some combination of physical LANs the subnets
, and assign each a subnet address: an IP network address which has the sites IP network address as prefix.
To put this more concretely, suppose the sites IP network address is A, and consists of n network bits (so
the site address may be written with the slash notation as A/n); in the diagram above, A/n = [Link]/16.
A subnet address is an IP network address B/k such that:
The address B/k is within the site: the first n bits of B are the same as A/ns
B/k extends A/n: kn
An example B/k in the diagram above is [Link]/24. (There is a slight simplification here in that subnet
addresses do not absolutely have to be prefixes; see below.)
We now have to figure out how packets will be routed to the correct subnet. For incoming packets we could
set up some proprietary protocol at the entry router to handle this. However, the more complicated situation
is all those existing internal hosts that, under the class A/B/C strategy, would still believe they can deliver
via the LAN to any site host, when in fact they can now only do that for hosts on their own subnet. We need
a more general solution.
We proceed as follows. For each subnet address B/k, we create a subnet mask for B consisting of k 1-bits
followed by enough 0-bits to make a total of 32. We then make sure that every host and router in the site
knows the subnet mask for every one of its interfaces. Hosts usually find their subnet mask the same way
they find their IP address (by static configuration if necessary, but more likely via DHCP, below).
Hosts and routers now apply the IP delivery algorithm of the previous section, with the proviso that, if a
subnet mask for an interface is present, then the subnet mask is used to determine the number of address bits
rather than the Class A/B/C mechanism. That is, we determine whether a packet addressed to destination
D is deliverable locally via an interface with subnet address B/k and corresponding mask M by comparing
D&M with B&M, where & represents bitwise AND; if the two match, the packet is local. This will generally

136

7 IP version 4

An Introduction to Computer Networks, Release 1.0

involve a match of more bits than if we used the Class A/B/C strategy to determine the network portion of
addresses D and B.
As stated previously, given an address D with no other context, we will not be able to determine the network/host division point in general (eg for outbound packets). However, that division point is not in fact
what we need. All that is needed is a way to tell if a given destination host address D belongs to the current
subnet, say B; that is, we need to compare the first k bits of D and B where k is the (known) length of B.
In the diagram above, the subnet mask for the /24 subnets would be [Link]; bitwise ANDing any IP
address with the mask is the same as extracting the first 24 bits of the IP address, that is, the subnet portion.
The mask for the /20 subnet would be [Link] (240 in binary is 1111 0000).
In the diagram above none of the subnets overlaps or conflicts: the subnets [Link]/24 and
[Link]/24 are disjoint. It takes a little more effort to realize that [Link]/20 does not overlap
with the others, but note that an IP address matches this network prefix only if the first four bits of the third
byte are 0001, so the third byte itself ranges from decimal 32 to decimal 63 = binary 0001 1111.
Note also that if host A = [Link] wishes to send to destination D = [Link], and A is not subnetaware, then delivery will fail: A will infer that the interface is a Class B, and therefore compare the first
two bytes of A and D, and, finding a match, will attempt direct LAN delivery. But direct delivery is now
likely impossible, as the subnets are not joined by a switch. Only with the subnet mask will A realize that
its network is [Link]/24 while Ds is [Link]/24 and that these are not the same. A would still be
able to send packets to its own subnet. In fact A would still be able to send packets to the outside world:
it would realize that the destination in that case does not match [Link]/16 and will thus forward to its
router. Hosts on other subnets would be the only unreachable ones.
Properly, the subnet address is the entire prefix, eg [Link]/24. However, it is often convenient to
identify the subnet address with just those bits that represent the extension of the site IP-network address;
we might thus say casually that the subnet address here is 65.
The class-based IP-address strategy allowed any host anywhere on the Internet to properly separate any
address into its net and host portions. With subnets, this division point is now allowed to vary; for example,
the address [Link] divides into 147.126 | 65.48 outside of Loyola, but into 147.126.65 | 48 inside.
This means that the net-host division is no longer an absolute property of addresses, but rather something
that depends on where the packet is on its journey.
Technically, we also need the requirement that given any two subnet addresses of different, disjoint subnets,
neither is a proper prefix of the other. This guarantees that if A is an IP address and B is a subnet address
with mask M (so B = B&M), then A&M = B implies A does not match any other subnet. Regardless of
the net/host division rules, we cannot possibly allow subnet [Link]/20 to represent one LAN while
[Link]/24 represents another; the second subnet address block is a subset of the first. (We can,
and sometimes do, allow the first LAN to correspond to everything in [Link]/20 that is not also in
[Link]/24; this is the longest-match rule.)
The strategy above is actually a slight simplification of what the subnet mechanism actually allows: subnet
address bits do not in fact have to be contiguous, and masks do not have to be a series of 1-bits followed by
0-bits. The mask can be any bit-mask; the subnet address bits are by definition those where there is a 1 in
the mask bits. For example, we could at a Class-B site use the fourth byte as the subnet address, and the
third byte as the host address. The subnet mask would then be [Link]. While this generality was
once sometimes useful in dealing with legacy IP addresses that could not easily be changed, life is simpler
when the subnet bits precede the host bits.

7.6 IP Subnets

137

An Introduction to Computer Networks, Release 1.0

7.6.1 Subnet Example

As an example of having different subnet masks on different interfaces, let us consider the division of a
class-C network into subnets of size 70, 40, 25, and 20. The subnet addresses will of necessity have different
lengths, as there is not room for four subnets each able to hold 70 hosts.
A: size 70
B: size 40
C: size 25
D: size 20
Because of the different subnet-address lengths, division of a local IP address LA into net versus host on
subnets cannot be done in isolation, without looking at the host bits. However, that division is not in fact
what we need. All that is needed is a way to tell if the local address LA belongs to a given subnet, say B;
that is, we need to compare the first n bits of LA and B, where n is the length of Bs subnet mask. We do
this by comparing LA&M to B&M, where M is the mask corresponding to n. LA&M is not necessarily the
same as LAnet , if LA actually belongs to one of the other subnets. However, if LA&M = B&M, then LA
must belong subnet B, in which case LA&M is in fact LAnet .
We will assume that the sites IP network address is [Link]/24. The first three bytes of each subnet address
must match 200.0.0. Only some of the bits of the fourth byte will be part of the subnet address, so we will
switch to binary for the last byte, and use both the /n notation (for total number of subnet bits) and also add
a vertical bar | to mark the separation between subnet and host.
Example: [Link] | 00 0000 / 26
Note that this means that the 0-bit following the 1-bit in the fourth byte is significant in that for a subnet
to match, it must match this 0-bit exactly. The remaining six 0-bits are part of the host portion.
To allocate our four subnet addresses above, we start by figuring out just how many host bits we need in
each subnet. Subnet sizes are always powers of 2, so we round up the subnets to the appropriate size. For
subnet A, this means we need 7 host bits to accommodate 27 = 128 hosts, and so we have a single bit in the
fourth byte to devote to the subnet address. Similarly, for B we will need 6 host bits and will have 2 subnet
bits, and for C and D we will need 5 host bits each and will have 8-5=3 subnet bits.
We now start choosing non-overlapping subnet addresses. We have one bit in the fourth byte to choose for
As subnet; rather arbitrarily, let us choose this bit to be 1. This means that every other subnet address must
have a 0 in the first bit position of the fourth byte, or we would have ambiguity.
Now for Bs subnet address. We have two bits to work with, and the first bit must be 0. Let us choose
the second bit to be 0 as well. If the fourth byte begins 00, the packet is part of subnet B, and the subnet
addresses for C and D must therefore not begin 00.
Finally, we choose subnet addresses for C and D to be 010 and 011, respectively. We thus have
subnet
A
B
C
D

138

size
128
64
32
32

address bits in fourth byte

1
00
010
011

host bits in 4th byte

7
6
5
5

decimal range
128-255
0-63
64-95
96-127

7 IP version 4

An Introduction to Computer Networks, Release 1.0

As desired, none of the subnet addresses in the third column is a prefix of any other subnet address.
The end result of all of this is that routing is now hierarchical: we route on the site IP address to get to a
site, and then route on the subnet address within the site.

7.6.2 Links between subnets

Suppose the Loyola CS department subnet ([Link]/24) and a department at some other site, we will
say [Link]/24, install a private link. How does this affect routing?
Each department router would add an entry for the other subnet, routing along the private link. Traffic
addressed to the other subnet would take the private link. All other traffic would go to the default router.
Traffic from the remote department to [Link]/24 would take the long route, and Loyola traffic to
[Link]/24 would take the long route.
Subnet anecdote
A long time ago I was responsible for two hosts, abel and borel. One day I was informed that machines
in computer lab 1 at the other end of campus could not reach borel, though they could reach abel.
Machines in lab 2, adjacent to lab 1, however, could reach both borel and abel just fine. What was the
difference?
It turned out that borel had a bad (/16 instead of /24) subnet mask, and so it was attempting local
delivery to the labs. This should have meant it could reach neither of the labs, as both labs were on
a different subnet from my machines; I was still perplexed. After considerably more investigation,
it turned out that between abel/borel and the lab building was a bridge-router: a hybrid device that
properly routed subnet traffic, but which also forwarded Ethernet packets directly, the latter feature
apparently for the purpose of backwards compatibility. Lab 2 was connected directly to the bridgerouter and thus appeared to be on the same LAN as borel, despite the apparently different subnet; lab
1 was connected to its own router R1 which in turn connected to the bridge-router. Lab 1 was thus, at
the LAN level, isolated from abel and borel.
Moral 1: Switching and routing are both great ideas, alone. But switching mixed with routing is not.
Moral 2: Test thoroughly! The reason the problem wasnt noticed earlier was that previously borel
communicated only with other hosts on the same subnet and with hosts outside the university entirely.
Both of these worked with the bad subnet mask; it is different-subnet local hosts that are the problem.
How would nearby subnets at either endpoint decide whether to use the private link? Classical link-state or
distance-vector theory (9 Routing-Update Algorithms) requires that they be able to compare the privatelink route with the going-around-the-long-way route. But this requires a global picture of relative routing
costs, which, as we shall see, almost certainly does not exist. The two departments are in different routing
domains; if neighboring subnets at either end want to use the private link, then manual configuration is likely
the only option.
..index:: single: switching, vs subnets single: subnets, vs switching

7.6 IP Subnets

139

An Introduction to Computer Networks, Release 1.0

7.6.3 Subnets versus Switching

A frequent network design question is whether to have many small subnets or to instead have just a few (or
even only one) larger subnet. With multiple small subnets, IP routing would be used to interconnect them;
the use of larger subnets would replace much of that routing with LAN-layer communication, likely Ethernet switching. Debates on this route-versus-switch question have gone back and forth in the networking
community, with one aphorism summarizing a common view:
Switch when you can, route when you must
This aphorism reflects the idea that switching is faster, cheaper and easier to configure, and that subnet
boundaries should be drawn only where necessary.
Ethernet switching equipment is indeed generally cheaper than routing equipment, for the same overall
level of features and reliability. And switching requires no configuration at all, while to implement subnets
not only must the subnets be created by hand but one must also set up and configure the routing-update
protocols. However, the price difference between switching and routing is not always significant in the big
picture, and the configuration involved is often straightforward.
Somewhere along the way, however, switching has acquired a reputation often deserved for being faster
than routing. It is true that routers have more to do than switches: they must decrement TTL, update the
header checksum, and attach a new LAN header. But these things are relatively minor: a larger reason many
routers are slower than switches may simply be that they are inevitably asked to serve as firewalls. This
means deep inspection of every packet, eg comparing every packet to each of a large number of firewall
rules. The firewall may also be asked to keep track of connection state. All this drives down the forwarding
rate, as measured in packets-per-second. The industry has come up with the term Layer 3 Switch to
describe routers that in effect do not support all the usual firewall bells and whistles; such streamlined
routers may also be able to do most of the hard work in specialized hardware, another source of speedup. (It
should be no surprise that Layer 3 switching is usually marketed in terms of the hardware speedup, not in
terms of the reduced flexibility.)
Switching scales remarkably well, but it does have limitations. First, broadcast packets must be forwarded
throughout a switched network; they do not, however, pass to different subnets. Second, LAN networks do
not like redundant links (that is, loops); while one can rely on the spanning-tree algorithm to eliminate these,
that algorithm too becomes less efficient at larger scales.

7.7 Address Resolution Protocol: ARP

If a host or router A finds that the destination IP address D = DIP matches the network address of one
of its interfaces, it is to deliver the packet via the LAN. This means looking up the LAN address DLAN
corresponding to DIP .
One approach would be via a special server, but the spirit of early IP development was to avoid such servers,
for both cost and reliability issues. Instead, the Address Resolution Protocol (ARP) is used. This is our first
protocol that takes advantage of the existence of LAN-level broadcast; on LANs without physical broadcast
(such as ATM), some other mechanism (usually involving a server) must be used.
The basic idea of ARP is that the host A sends out a broadcast ARP query or who-has DIP ? request, which
includes As own IP and LAN addresses. All hosts on the LAN receive this message. The host for whom the
message is intended, D, will recognize that it should reply, and will return an ARP reply or is-at message
140

7 IP version 4

An Introduction to Computer Networks, Release 1.0

containing DLAN . Because the original request contained ALAN , Ds response can be sent directly to A, that
is, unicast.

Additionally, all hosts maintain an ARP cache, consisting of IP,LAN address pairs for other hosts on the
network. After the exchange above, A has DIP ,DLAN in its table; anticipating that A will soon send it a
packet to which it needs to respond, D also puts AIP ,ALAN into its cache.
ARP-cache entries eventually expire. The timeout interval used to be on the order of 10 minutes, but linux
systems now use a much smaller timeout (~30 seconds observed in 2012). Somewhere along the line, and
probably related to this shortened timeout interval, repeat ARP queries about a timed-out entry are first sent
unicast, not broadcast, to the previous Ethernet address on record. This cuts down on the total amount of
broadcast traffic; LAN broadcasts are, of course, still needed for new hosts. The ARP cache on a linux
system can be examined with the command ip -s neigh; the corresponding windows command is arp
-a.
The above protocol is sufficient, but there is one further point. When A sends its broadcast who-has D?
ARP query, all other hosts C check their own cache for an entry for A. If there is such an entry (that is, if
AIP is found there), then the value for ALAN is updated with the value taken from the ARP message; if there
is no pre-existing entry then no action is taken. This update process serves to avoid stale ARP-cache entries,
which can arise is if a host has had its Ethernet card replaced.

7.7.1 ARP Finer Points

Most hosts today implement self-ARP, or gratuitous ARP, on startup (or wakeup): when station A starts up
it sends out an ARP query for itself : who-has A?. Two things are gained from this: first, all stations that
had A in their cache are now updated with As most current ALAN address, in case there was a change, and
second, if an answer is received, then presumably some other host on the network has the same IP address
as A.
Self-ARP is thus the traditional IPv4 mechanism for duplicate address detection. Unfortunately, it does
not always work as well as might be hoped; often only a single self-ARP query is sent, and if a reply is
received then frequently the only response is to log an error message; the host may even continue using
the duplicate address! If the duplicate address was received via DHCP, below, then the host is supposed to
notify its DHCP server of the error and request a different IPv4 address.
RFC 5227 has defined an improved mechanism known as Address Conflict Detection, or ACD. A host
using ACD sends out three ARP queries for its new IP address, spaced over a few seconds and leaving the
ARP field for the senders IP address filled with zeroes. This last step means that any other host with that
IP address in its cache will ignore the packet, rather than update its cache. If the original host receives no
replies, it then sends out two more ARP queries for its new address, this time with the ARP field for the
senders IP address filled in with the new address; this is the stage at which other hosts on the network will

7.7 Address Resolution Protocol: ARP

141

An Introduction to Computer Networks, Release 1.0

make any necessary cache updates. Finally, ACD requires that hosts that do detect a duplicate address must
discontinue using it.
It is also possible for other stations to answer an ARP query on behalf of the actual destination D; this is
called proxy ARP. An early common scenario for this was when host C on a LAN had a modem connected
to a serial port. In theory a host D dialing in to this modem should be on a different subnet, but that requires
allocation of a new subnet. Instead, many sites chose a simpler arrangement. A host that dialed in to Cs
serial port might be assigned IP address DIP , from the same subnet as C. C would be configured to route
packets to D; that is, packets arriving from the serial line would be forwarded to the LAN interface, and
packets sent to CLAN addressed to DIP would be forwarded to D. But we also have to handle ARP, and as
D is not actually on the LAN it will not receive broadcast ARP queries. Instead, C would be configured to
answer on behalf of D, replying with DIP ,CLAN . This generally worked quite well.
Proxy ARP is also used in Mobile IP, for the so-called home agent to intercept traffic addressed to the
home address of a mobile device and then forward it (eg via tunneling) to that device. See 7.11 Mobile
IP.
One delicate aspect of the ARP protocol is that stations are required to respond to a broadcast query. In the
absence of proxies this theoretically should work quite well: there should be only one respondent. However,
there were anecdotes from the Elder Days of networking when a broadcast ARP query would trigger an
avalanche of responses. The protocol-design moral here is that determining who is to respond to a broadcast
message should be done with great care. (RFC 1122 section 3.2.2 addresses this same point in the context
of responding to broadcast ICMP messages.)
ARP-query implementations also need to include a timeout and some queues, so that queries can be resent
if lost and so that a burst of packets does not lead to a burst of queries. A naive ARP algorithm without these
might be:
To send a packet to destination DIP , see if DIP is in the ARP cache. If it is, address the packet
to DLAN ; if not, send an ARP query for D
To see the problem with this approach, imagine that a 32KB packet arrives at the IP layer, to be sent over
Ethernet. It will be fragmented into 22 fragments (assuming an Ethernet MTU of 1500 bytes), all sent at
once. The naive algorithm above will likely send an ARP query for each of these. What we need instead is
something like the following:
To send a packet to destination DIP :
If DIP is in the ARP cache, send to DLAN and return
If not, see if an ARP query for DIP is pending.
If it is, put the current packet in a queue for D.
If there is no pending ARP query for DIP , start one,
again putting the current packet in the (new) queue for D
We also need:
If an ARP query for some CIP times out, resend it (up to a point)
If an ARP query for CIP is answered, send off any packets in Cs queue

7.7.2 ARP security

Suppose A wants to log in to secure server S, using a password. How can B (for Bad) impersonate S?
142

7 IP version 4

An Introduction to Computer Networks, Release 1.0

Here is an ARP-based strategy, sometimes known as ARP Spoofing. First, B makes sure the real S is down,
either by waiting until scheduled downtime or by launching a denial-of-service attack against S.
When A tries to connect, it will begin with an ARP who-has S?. All B has to do is answer, S is-at B.
There is a trivial way to do this: B simply needs to set its own IP address to that of S.
A will connect, and may be convinced to give its password to B. B now simply responds with something
plausible like backup in progress; try later, and meanwhile use As credentials against the real S.
This works even if the communications channel A uses is encrypted! If A is using the SSH protocol
(21.9.1 SSH), then A will get a message that the other sides key has changed (B will present its own
SSH key, not Ss). Unfortunately, many users (and even some IT departments) do not recognize this as a
serious problem. Some organizations especially schools and universities use personal workstations with
frozen configuration, so that the filesystem is reset to its original state on every reboot. Such systems may
be resistant to viruses, but in these environments the user at A will always get a message to the effect that
Ss credentials are not known.

7.7.3 ARP Failover

Suppose you have two front-line servers, A and B (B for Backup), and you want B to be able to step in if A
freezes. There are a number of ways of achieving this, but one of the simplest is known as ARP Failover.
First, we set AIP = BIP , but for the time being B does not use the network so this duplication is not a problem.
Then, once B gets the message that A is down, it sends out an ARP query for AIP , including BLAN as the
source LAN address. The gateway router, which previously would have had AIP ,ALAN in its ARP cache,
updates this to AIP ,BLAN , and packets that had formerly been sent to A will now go to B. As long as B is
trafficking in stateless operations (eg html), B can pick up right where A left off.

7.7.4 Detecting Sniffers

Finally, here is an interesting use of ARP to detect Ethernet password sniffers (generally not quite the issue
it once was, due to encryption and switching). To find out if a particular host A is in promiscuous mode,
send an ARP who-has A? query. Address it not to the broadcast Ethernet address, though, but to some
nonexistent Ethernet address.
If promiscuous mode is off, As network interface will ignore the packet. If promiscuous mode is on, As
network interface will pass the ARP request to A itself, which is likely then to answer it.
Alas, linux kernels reject at the ARP-software level ARP queries to physical Ethernet addresses other than
their own. However, they do respond to faked Ethernet multicast addresses, such as [Link] or
[Link].

7.8 Dynamic Host Configuration Protocol (DHCP)

DHCP is the most common mechanism by which hosts are assigned their IP addresses. DHCP started as a
protocol known as Reverse ARP (RARP), which evolved into BOOTP and then into its present form. It is
documented in RFC 2131.

7.8 Dynamic Host Configuration Protocol (DHCP)

143

An Introduction to Computer Networks, Release 1.0

Recall that ARP is based on the idea of someone broadcasting an ARP query for a host, containing the hosts
IP address, and the host answering it with its LAN address. DHCP involves a host, at startup, broadcasting a
query containing its own LAN address, and having a server reply telling the host what IP address is assigned
to it. The DHCP response is likely to contain several other essential startup options as well, including:
IP address
subnet mask
default router
DNS Server
These four items are a pretty standard minimal network configuration.
Default Routers and DHCP
If you lose your default router, you cannot communicate. Here is something that used to happen to me,
courtesy of DHCP:
1. I am connected to the Internet via Ethernet, and my default router is via my Ethernet interface
2. I connect to my institutions wireless network.
3. Their DHCP server sends me a new default router on the wireless network. However, this default
router will only allow access to a tiny private network, because I have neglected to complete the
Wi-Fi network registration process.
4. I therefore disconnect from the wireless network, and my wireless-interface default router goes
away. However, my system does not automatically revert to my Ethernet default-router entry;
DHCP does not work that way. As a result, I will have no router at all until the next scheduled
DHCP lease renegotiation, and must fix things manually.
The DHCP server has a range of IP addresses to hand out, and maintains a database of which IP address has
been assigned to which LAN address. Reservations can either be permanent or dynamic; if the latter, hosts
typically renew their DHCP reservation periodically (typically one to several times a day).

7.8.1 DHCP and the Small Office

If you have a large network, with multiple subnets, a certain amount of manual configuration is inevitable.
What about, however, a home or small office, with a single line from an ISP? A combination of NAT
(1.14 Network Address Translation) and DHCP has made autoconfiguration close to a reality.
The typical home/small-office router is in fact a NAT firewall coupled with a switch (usually also coupled
with a Wi-Fi access point); we will refer to this as a NAT router. One specially designated port, the external
port, connects to the ISPs line, and uses DHCP to obtain an IP address for that port. The other, internal,
ports are connected together by an Ethernet switch; these ports as a group are connected to the external port
using NAT translation. If wireless is supported, the wireless side is connected directly to the internal ports.
Isolated from the Internet, the internal ports can thus be assigned an arbitrary non-public IP address block,
eg [Link]/24. The NAT router contains a DCHP server, usually enabled by default, that will hand out
IP addresses to everything connecting from the internal side.
Generally this works seamlessly. However, if a second NAT router is also connected to the network (sometimes attempted to extend Wi-Fi range, in lieu of a commercial Wi-Fi repeater), one then has two operating
144

7 IP version 4

An Introduction to Computer Networks, Release 1.0

DHCP servers on the same subnet. This often results in chaos, as two different hosts may be assigned the
same IP address, or a hosts IP address may suddenly change if it gets a new IP address from the other server.
Disabling one of the DHCP servers fixes this.
While omnipresent DHCP servers have made IP autoconfiguration work out of the box in many cases, in
the era in which IP was designed the need for such servers would have been seen as a significant drawback in
terms of expense and reliability. IPv6 has an autoconfiguration strategy (8.7.2 Stateless Autoconfiguration
(SLAAC)) that does not require DHCP, though DHCPv6 may well end up displacing it.

7.8.2 DHCP and Routers

It is often desired, for larger sites, to have only one or two DHCP servers, but to have them support multiple
subnets. Classical DHCP relies on broadcast, which isnt forwarded by routers, and even if it were, the
DHCP server would have no way of knowing on what subnet the host in question was actually located.
This is generally addressed by DHCP Relay (sometimes still known by the older name BOOTP Relay).
The router (or, sometimes, some other node on the subnet) receives the DHCP broadcast message from a
host, and notes the subnet address of the arrival interface. The router then relays the DHCP request, together
with this subnet address, to the designated DHCP Server; this relayed message is sent directly (unicast), not
broadcast. Because the subnet address is included, the DHCP server can figure out the correct IP address to
assign.
This feature has to be specially enabled on the router.

7.9 Internet Control Message Protocol

The Internet Control Message Protocol, or ICMP, is a protocol for sending IP-layer error and status messages. It is, like IP, host-to-host, and so they are never delivered to a specific port, even if they are sent
in response to an error related to something sent from that port. In other words, individual UDP and TCP
connections do not receive ICMP messages, even when it would be helpful to get them.
Here are the ICMP types, with subtypes listed in the description. The list has been pruned to include only
the most common values.

7.9 Internet Control Message Protocol

145

An Introduction to Computer Networks, Release 1.0

Type
Echo Request
Echo Reply

Destination Unreachable

Source Quench
Redirect Message
Router Solicitation
Time Exceeded
Bad IP Header or Parameter

Description
ping queries
ping responses
Destination network unreachable
Destination host unreachable
Destination port unreachable
Fragmentation required but DF flag set
Network administratively prohibited
Congestion control
Redirect datagram for the network
Redirect datagram for the host
Redirect for TOS and network
Redirect for TOS and host
Router discovery/selection/solicitation
TTL expired in transit
Fragment reassembly time exceeded
Pointer indicates the error
Missing a required option
Bad length

ICMP is perhaps best known for Echo Request/Reply, on which the ping tool is based. Ping remains very
useful for network troubleshooting: if you can ping a host, then the network is reachable, and any problems
are higher up the protocol chain. Unfortunately, ping replies are blocked by default by many firewalls, on
the theory that revealing even the existence of computers is a security risk. While this may be an appropriate
decision, it does significantly impair the utility of ping. Most routers do still pass ping requests, but some
site routers block them.
Source Quench was used to signal that congestion has been encountered. A router that drops a packet due to
congestion experience was encouraged to send ICMP Source Quench to the originating host. Generally the
TCP layer would handle these appropriately (by reducing the overall sending rate), but UDP applications
never receive them. ICMP Source Quench did not quite work out as intended, and was formally deprecated
by RFC 6633. (Routers can inform TCP connections of impending congestion by using the ECN bits.)
The Destination Unreachable type has a large number of subtypes:
Network unreachable: some router had no entry for forwarding the packet, and no default route
Host unreachable: the packet reached a router that was on the same LAN as the host, but the host
failed to respond to ARP queries
Port unreachable: the packet was sent to a UDP port on a given host, but that port was not open.
TCP, on the other hand, deals with this situation by replying to the connecting endpoint with a reset
packet. Unfortunately, the UDP Port Unreachable message is sent to the host, not to the application on
that host that sent the undeliverable packet, and so is close to useless as a practical way for applications
to be informed when packets cannot be delivered.
Fragmentation required but DF flag set: a packet arrived at a router and was too big to be forwarded
without fragmentation. However, the Dont Fragment bit in the IP header was set, forbidding fragmentation. Later we will see how TCP uses this option as part of Path MTU Discovery, the process
of finding the largest packet we can send to a specific destination without fragmentation. The basic

146

7 IP version 4

An Introduction to Computer Networks, Release 1.0

idea is that we set the DF bit on some of the packets we send; if we get back this message, that packet
was too big.
Administratively Prohibited: this is sent by a router that knows it can reach the network in question,
but has configured to drop the packet and send back Administratively Prohibited messages. A router
can also be configured to blackhole messages: to drop the packet and send back nothing.

7.9.1 Traceroute and Time Exceeded

The traceroute program uses ICMP Time Exceeded messages. A packet is sent to the destination (often
UDP to an unused port), with the TTL set to 1. The first router the packet reaches decrements the TTL to
0, drops it, and returns an ICMP Time Exceeded message. The sender now knows the first router on the
chain. The second packet is sent with TTL set to 2, and the second router on the path will be the one to
return ICMP Time Exceeded. This continues until finally the remote host returns something, likely ICMP
Port Unreachable.
Many routers no longer respond with ICMP Time Exceeded messages when they drop packets. For the
distance value corresponding to such a router, traceroute reports ***.
Traceroute assumes the path does not change. This is not always the case, although in practice it is seldom
an issue.
Route Efficiency
Once upon a time (~2001), traceroute showed that traffic from my home to the office, both in the
Chicago area, went through the MAE-EAST Internet exchange point, outside of Washington DC. That
inefficient route was later fixed. A situation like this is typically caused by two higher-level providers
who did not negotiate sufficient Internet exchange points.
Traceroute to a nonexistent site works up to the point when the packet reaches the Internet backbone: the
first router which does not have a default route. At that point the packet is not routed further (and an ICMP
Destination Network Unreachable should be returned).

7.9.2 Redirects
Most non-router hosts start up with an IP forwarding table consisting of a single (default) router, discovered
along with their IP address through DHCP. ICMP Redirect messages help hosts learn of other useful routers.
Here is a classic example:

A is configured so that its default router is R1. It addresses a packet to B, and sends it to R1. R1 receives
the packet, and forwards it to R2. However, R1 also notices that R2 and A are on the same network, and
so A could have sent the packet to R2 directly. So R1 sends an appropriate ICMP redirect message to A
(Redirect Datagram for the Network), and A adds a route to B via R2 to its own forwarding table.
7.9 Internet Control Message Protocol

147

An Introduction to Computer Networks, Release 1.0

7.9.3 Router Solicitation

These ICMP messages are used by some router protocols to identify immediate neighbors. When we look
at routing-update algorithms, below, these are where the process starts.

7.10 Unnumbered Interfaces

We mentioned in 1.10 IP - Internet Protocol and 7.2 Interfaces that some devices allow the use of pointto-point IP links without assigning IP addresses to the interfaces at the ends of the link. Such IP interfaces
are referred to as unnumbered; they generally make sense only on routers. It is a firm requirement that the
node (ie router) at each endpoint of such a link has at least one other interface that does have an IP address;
otherwise, the node in question would be anonymous.
The diagram below shows a link L joining routers R1 and R2, which are connected to subnets [Link]/24
and [Link]/24 respectively. The endpoint interfaces of L, both labeled link0, are unnumbered.

The endpoints of L could always be assigned private IP addresses (7.3 Special Addresses), such as [Link]
and [Link]. To do this we would need to create a subnet; because the host bits cannot be all 0s or all 1s,
the minimum subnet size is four (eg [Link]/30). Furthermore, the routing protocols to be introduced in
9 Routing-Update Algorithms will distribute information about the subnet throughout the organization or
routing domain, meaning care must be taken to ensure that each links subnet is unique. Use of unnumbered links avoids this.
If R1 were to originate a packet to be sent to (or forwarded via) R2, the standard strategy is for it to treat
its link0 interface as if it shared the IP address of its Ethernet interface eth0, that is, [Link]; R2
would do likewise. This still leaves R1 and R2 violating the IP local-delivery rule of 7.5 The Classless
IP Delivery Algorithm; R1 is expected to deliver packets via local delivery to [Link] but has no interface
that is assigned an IP address on the destination subnet [Link]/24. The necessary dispensation, however,
is granted by RFC 1812. All that is necessary by way of configuration is that R1 be told R2 is a directly
connected neighbor reachable via its link0 interface. On linux systems this might be done with the ip
route command on R1 as follows:
ip route
The linux ip route command illustrated here was tested on a virtual point-to-point link created with
ssh and pppd; the link interface name was in fact ppp0. While the command appeared to work as
advertised, it was only possible to create the link if endpoint IP addresses were assigned at the time
of creation; these were then removed with ip route del and then re-assigned with the command
shown here.

148

7 IP version 4

An Introduction to Computer Networks, Release 1.0

ip route add [Link] dev link0

Because L is a point-to-point link, there is no destination LAN address and thus no ARP query.

7.11 Mobile IP
In the original IPv4 model, there was a strong if implicit assumption that each IP host would stay put. One
role of an IP address is simply as a unique endpoint identifier, but another role is as a locator: some prefix
of the address (eg the network part, in the class-A/B/C strategy, or the provider prefix) represents something
about where the host is physically located. Thus, if a host moves far enough, it may need a new address.
When laptops are moved from site to site, it is common for them to receive a new IP address at each
location, eg via DHCP as the laptop connects to the local Wi-Fi. But what if we wish to support devices like
smartphones that may remain active and communicating while moving for thousands of miles? Changing
IP addresses requires changing TCP connections; life (and application development) might be simpler if a
device had a single, unchanging IP address.
One option, commonly used with smartphones connected to some so-called 3G networks, is to treat the
phones data network as a giant wireless LAN. The phones IP address need not change as it moves within
this LAN, and it is up to the phone provider to figure out how to manage LAN-level routing, much as is
done in 3.3.5 Wi-Fi Roaming.
But Mobile IP is another option, documented in RFC 5944. In this scheme, a mobile host has a permanent
home address and, while roaming about, will also have a temporary care-of address, which changes from
place to place. The care-of address might be, for example, an IP address assigned by a local Wi-Fi network,
and which in the absence of Mobile IP would be the IP address for the mobile host. (This kind of care-of
address is known as co-located; the care-of address can also be associated with some other device known
as a foreign agent in the vicinity of the mobile host.) The goal of Mobile IP is to make sure that the mobile
host is always reachable via its home address.
To maintain connectivity to the home address, a Mobile IP host needs to have a home agent back on the
home network; the job of the home agent is to maintain an IP tunnel that always connects to the devices
current care-of address. Packets arriving at the home network addressed to the home address will be forwarded to the mobile device over this tunnel by the home agent. Similarly, if the mobile device wishes to
send packets from its home address that is, with the home address as IP source address it can use the
tunnel to forward the packet to the home agent.
The home agent may use proxy ARP (7.7.1 ARP Finer Points) to declare itself to be the appropriate
destination on the home LAN for packets addressed to the home (IP) address; it is then straightforward for
the home agent to forward the packets.
An agent discovery process is used for the mobile host to decide whether it is mobile or not; if it is, it then
needs to notify its home agent of its current care-of address.
There are several forms of packet encapsulation that can be used for Mobile IP tunneling, but the default one
is IP-in-IP encapsulation, defined in RFC 2003. In this process, the entire original IP packet (with header
addressed to the home address) is used as data for a new IP packet, with a new IP header (the outer header)
addressed to the care-of address.

7.11 Mobile IP

149

An Introduction to Computer Networks, Release 1.0

A special value in the IP-header Protocol field indicates that IP-in-IP tunneling was used, so the receiver
knows to forward the packet on using the information in the inner header. The MTU of the tunnel will be
the original MTU of the path to the care-of address, minus the size of the outer header.

7.12 Epilog
At this point we have concluded the basic mechanics of IPv4. Still to come is a discussion of how IP
routers build their forwarding tables. This turns out to be a complex topic, divided into routing within single
organizations and ISPs 9 Routing-Update Algorithms and routing between organizations 10 LargeScale IP Routing.
But before that, in the next chapter, we compare IPv4 with IPv6, now twenty years old but still seeing very
limited adoption. The biggest issue fixed by IPv6 is IPv4s lack of address space, but there are also several
other less dramatic improvements.

7.13 Exercises
1. Suppose an Ethernet packet represents a TCP acknowledgment; that is, the packet contains an IP header
and a 20-byte TCP header but nothing else. Is the IP packet here smaller than the Ethernet minimum-packet
size, and, if so, by how much?
2. How can a receiving host tell if an arriving IP packet is unfragmented?
3. How long will it take the IDENT field of the IP header to wrap around, if the sender host A sends a stream
of packets to host B as fast as possible? Assume the packet size is 1500 bytes and the bandwidth is 600
Mbps.
4. The following diagram has routers A, B, C, D and E; E is the border router connecting the site to
the Internet. All router-to-router connections are via Ethernet-LAN /24 subnets with addresses of the form
200.0.x. Give forwarding tables for each of A, B, C and D. Each table should include each of the listed
subnets and also a default entry that routes traffic toward router E.
200.0.5

200.0.6

200.0.7

200.0.8

Internet

200.0.9
C

150

7 IP version 4

An Introduction to Computer Networks, Release 1.0

200.0.10

5. (This exercise is an attempt at modeling Internet-2 routing.) Suppose sites S1 ... Sn each have a single
connection to the standard Internet, and each site Si has a single IP address block Ai . Each sites connection
to the Internet is through a single router Ri ; each Ri s default route points towards the standard Internet. The
sites also maintain a separate, higher-speed network among themselves; each site has a single link to this
separate network, also through Ri . Describe what the forwarding tables on each Ri will have to look like so
that traffic from one Si to another will always use the separate higher-speed network.
6. For each IP network prefix given (with length), identify which of the subsequent IP addresses are part of
the same subnet.

(a). [Link]/23: [Link], [Link], [Link], [Link]

(b). [Link]/22: [Link], [Link], [Link], [Link]
(c). [Link]/18: [Link], [Link], [Link], [Link]

7. Suppose that the subnet bits below for the following five subnets A-E all come from the beginning of the
fourth byte of the IP address; that is, these are subnets of a /24 block.
A: 00
B: 01
C: 110
D: 111
E: 1010

(a). What are the sizes of each subnet, and the corresponding decimal ranges? Count the addresses with
host bits all 0s or with host bits all 1s as part of the subnet.
(b). How many IP addresses in the class-C block do not belong to any of the subnets A, B, C, D and E?

8. In 7.7 Address Resolution Protocol: ARP it was stated that, in newer implementations, repeat ARP
queries about a timed out entry are first sent unicast, in order to reduce broadcast traffic. Why is this unicast
approach likely to succeed most of the time? Can you give an example of a situation in which the unicast
query would fail, but a followup broadcast query would succeed?
9. Suppose A broadcasts an ARP query who-has B?, receives Bs response, and proceeds to send B a
regular IP packet. If B now wishes to reply, why is it likely that A will already be present in Bs ARP cache?
Identify a circumstance under which this can fail.
10. Suppose A broadcasts an ARP request who-has B, but inadvertently lists the physical address of
another machine C instead of its own (that is, As ARP query has IPsrc = A, but LANsrc = C). What will
happen? Will A receive a reply? Will any other hosts on the LAN be able to send to A? What entries will
be made in the ARP caches on A, B and C?
7.13 Exercises

151

An Introduction to Computer Networks, Release 1.0

152

7 IP version 4

8 IP VERSION 6

What has been learned from experience with IPv4? First and foremost, more than 32 bits are needed for
addresses; the primary motive in developing IPv6 was the specter of running out of IPv4 addresses (something which, at the highest level, has already happened; see the discussion at the end of 1.10 IP - Internet
Protocol). Another important issue is that IPv4 requires a modest amount of effort at configuration; IPv6
was supposed to improve this.
By 1990 the IETF was actively interested in proposals to replace IPv4. A working group for the so-called
IP next generation, or IPng, was created in 1993 to select the new version; RFC 1550 was this groups
formal solicitation of proposals. In July 1994 the IPng directors voted to accept a modified version of the
Simple Internet Protocol, or SIP (unrelated to the Session Initiation Protocol), as the basis for IPv6.
SIP addresses were originally to be 64 bits in length, but in the month leading up to adoption this was
increased to 128. 64 bits would probably have been enough, but the problem is less the actual number than
the simplicity with which addresses can be allocated; the more bits, the easier this becomes, as sites can be
given relatively large address blocks without fear of waste. A secondary consideration in the 64-to-128 leap
was the potential to accommodate now-obsolete CLNP addresses, which were up to 160 bits in length (but
compressible).
IPv6 has to some extent returned to the idea of a fixed division between network and host portions: in most
ordinary-host cases, the first 64 bits of the address is the network portion (including any subnet portion)
and the remaining 64 bits represent the host portion. While there are some configuration alternatives here,
and while the IETF occasionally revisits the issue, at the present time the 64/64 split seems here to stay.
Routing, however, can, as in IPv4, be done on different prefixes of the address at different points of the
network. Thus, it is misleading to think of IPv6 as a return to Class A/B/C address allocation.
IPv6 is now twenty years old, and yet usage remains quite modest. However, the shortage in IPv4 addresses
has begun to loom ominously; IPv6 adoption rates may rise quickly if IPv4 addresses begin to climb in
price.

8.1 The IPv6 Header

The IPv6 fixed header looks like the following; at 40 bytes, it is twice the size of the IPv4 header. The
fixed header is intended to support only what every packet needs: there is no support for fragmentation, no
header checksum, and no option fields. However, the concept of extension headers has been introduced
to support some of these as options; some IPv6 extension headers are described below. Some fixed-header
header fields have been renamed from their IPv4 analogues: the IPv4 TTL is now the IPv6 Hop_Limit (still
decremented by each router with the packet discarded when it reaches 0), and the IPv4 DS field has become
the IPv6 Traffic Class.

153

An Introduction to Computer Networks, Release 1.0

The Flow Label is new. RFC 2460 states that it

may be used by a source to label sequences of packets for which it requests special handling by
the IPv6 routers, such as non-default quality of service or real-time service.
Senders not actually taking advantage of any quality-of-service options are supposed to set the Flow Label
to zero.
When used, the Flow Label represents a sender-computed hash of the source and destination addresses, and
perhaps the traffic class. Routers can use this field as a way to look up quickly any priority or reservation
state for the packet. All packets belonging to the same flow should have the same Routing Extension header,
below. Note that the Flow Label will in general not include any information about the source and destination
port numbers, except that only some of the connections between a pair of hosts may make use of this field.
A flow, as the term is used here, is one-way; the return traffic belongs to a different flow. Historically,
the term flow has been used at various other scales: a single TCP connection, multiple related TCP
connections, or even all traffic from a particular subnet (eg the computer-lab flow).

8.2 IPv6 addresses

IPv6 addresses are written in eight groups of four hex digits, separated by colons, and with leading 0s
optionally removed, eg
[Link]
If an address contains a long run of 0s, as would be the case if the IPv6 address had an embedded IPv4
address, then when writing the address the string :: may be used to represent however many blocks of
0000 as are needed to create an address of the correct length; to avoid ambiguity this can be used only once.
Also, embedded IPv4 addresses may continue to use the . separator:
::FFFF:[Link]

154

8 IP version 6

An Introduction to Computer Networks, Release 1.0

The above is an example of the standard IPv6 format for representing IPv4 addresses. A separate representation of IPv4 addresses, with the FFFF block replaced by 0-bits, is used for tunneling IPv6 traffic over
IPv4. The IPv6 loopback address is ::1 (that is, 127 0-bits followed by a 1-bit).
Network address prefixes may be written with the / notation, as in IPv4:
[Link]/60
RFC 3513 suggested that initial IPv6 unicast-address allocation be initially limited to addresses beginning
with the bits 001, that is, the 2000::/3 block (20 in binary is 0010 0000).
Generally speaking, IPv6 addresses consist of a 64-bit network prefix (perhaps including subnet bits) and a
64-bit host identifier. See 8.5 Network Prefixes.

8.2.1 Host identifier

If desired, the second 64 bits of an IPv6 address can be a host identifier derived from the LAN address.
To create the host identifier from a 48-bit Ethernet address, for example, 0xFFFE is inserted between the
first three bytes and the last three bytes, to get 64 bits in all. The seventh bit of the first byte (the Ethernet
universal/local flag) is then set to 1. The result is officially known as the Modified EUI-64 Identifier,
where EUI stands for Extended Unique Identifier; details can be found in RFC 4291. For example, for a
host with Ethernet address [Link], the EUI-64 address would be [Link] (note the
leading 00 becomes 02 when the seventh bit is turned on).

8.2.2 Link-local addresses

IPv6 defines link-local addresses, intended to be used only on a single LAN (and never routed). These begin
with the link-local prefix of the ten bits 1111 1110 10 followed by 54 more zero bits; that is, FE80::/64. The
final 64 bits are the host identifier for the link interface in question, above. The Ethernet link-local address of
the machine in the previous paragraph with Ethernet address [Link] is fe80::2a0:ccff:fe24:b0e4.
When sending to a link-local address, one must include somewhere an identification of which link to use.
IPv4 addresses, on the other hand, almost always implicitly identify the link by virtue of the network prefix.
See 8.10.1 ping6 and 8.10.2 TCP connections with link-local addresses for examples of sending to linklocal addresses.
Once the link-local address is created, it must pass the duplicate-address detection test before being used;
see 8.7.1 Duplicate Address Detection.

8.2.3 Anycast addresses

IPv6 also introduces anycast addresses. An anycast address might be assigned to each of a set of routers (in
addition to each routers own unicast address); a packet addressed to this anycast address would be delivered
to only one member of this set. Note that this is quite different from multicast addresses; a packet addressed
to the latter is delivered to every member of the set.
It is up to the local routing infrastructure to decide which member of the anycast group would receive the
packet; normally it would be sent to the closest member. This allows hosts to send to any of a set of
routers, rather than to their designated individual default router.
8.2 IPv6 addresses

155

An Introduction to Computer Networks, Release 1.0

8.3 IPv6 Multicast

IPv6 has moved away from LAN-layer broadcast, instead providing a wide range of LAN-layer multicast
groups. (Note that LAN-layer multicast is straightforward; it is general IP-layer multicast (19.5 Global
IP Multicast) that is problematic. See 2.1.1 Ethernet Multicast for the Ethernet implementation.) This
switch to multicast is intended to limit broadcast traffic in general, though many switches still propagate
LAN multicast traffic everywhere, like broadcast.
An IPv6 multicast address is one beginning with the eight bits 1111 1111; numerous specific such addresses,
and even classes of addresses, have been defined. For actual delivery, IPv6 multicast addresses correspond
to LAN-layer (eg Ethernet) multicast addresses through a well-defined static correspondence; specifically,
if x, y, z and w are the last four bytes of the IPv6 multicast address, in hex, then the corresponding Ethernet
multicast address is [Link]x:y:z:w (RFC 2464). A typical IPv6 host will need to join (that is, subscribe to)
several Ethernet multicast groups.
The IPv6 multicast address with the broadest scope is all-nodes, with address FF02::1; the corresponding
Ethernet multicast address is [Link]. This essentially corresponds to IPv4s LAN broadcast,
though the use of LAN multicast here means that non-IPv6 hosts should not see packets sent to this address.
Another important IPv6 multicast address is FF02::2, the all-routers address. This is meant to be used to
reach all routers, and routers only; ordinary hosts do not subscribe.
Generally speaking, IPv6 nodes on Ethernets send LAN-layer Multicast Listener Discovery (MLD) messages to multicast groups they wish to start using; these messages allow multicast-aware Ethernet switches
to optimize forwarding so that only those hosts that have subscribed to the multicast group in question
will receive the messages. Otherwise switches are supposed to treat multicast like broadcast; worse, some
switches may simply fail to forward multicast packets to destinations that have not explicitly opted to join
the group.

8.4 IPv6 Extension Headers

In IPv4, the IP header contained a Protocol field to identify the next header; usually UDP or TCP. All IPv4
options were contained in the IP header itself. IPv6 has replaced this with a scheme for allowing an arbitrary
chain of IPv6 headers. The IPv6 Next Header field can indicate that the following header is UDP or TCP,
but can also indicate one of several IPv6 options. These optional, or extension, headers include:
Hop-by-Hop options header
Destination options header
Routing header
Fragment header
Authentication header
Mobility header
Encapsulated Security Payload header
These extension headers must be processed in order; the recommended order for inclusion is as above. Most
of them are intended for processing only at the destination host; the routing header is an exception.

156

8 IP version 6

An Introduction to Computer Networks, Release 1.0

8.4.1 Hop-by-Hop Options Header

This consists of a set of type,value pairs which are intended to be processed by each router on the path. A
tag in the type field indicates what a router should do if it does not understand the option: drop the packet,
or continue processing the rest of the options. The only Hop-by-Hop options provided by RFC 2460 were
for padding, so as to set the alignment of later headers.
RFC 2675 later defined a Hop-by-Hop option to support IPv6 jumbograms: datagrams larger than 65,535
bytes. The need for such large packets remains unclear, in light of 5.3 Packet Size. IPv6 jumbograms are
not meant to be used if the underlying LAN does not have an MTU larger than 65,535 bytes; the LAN world
is not currently moving in this direction.

8.4.2 Destination Options Header

This is very similar to the Hop-by-Hop Options header. It again consists of a set of type,value pairs,
and the original specification only defined options for padding. The Destination header is intended to be
processed at the destination, before turning over the packet to the transport layer.

8.4.3 Routing Header

The routing header contains a list of IPv6 addresses through which the packet should be routed. These
do not have to be contiguous: if the list to be visited en route to destination D is R1,R2,...,Rn, then the
IPv6 option header contains R2,R3,...,Rn,D and the initial destination address is R1; R1 then updates this
header to R1,R3,...,Rn,D (that is, the old destination R1 and the current next-router R2 are swapped), and
sends the packet on to R2. This continues on until Rn addresses the packet to the final destination D. The
header contains a Segments Left pointer indicating the next address to be processed, so that when the packet
arrives at D the Routing Header contains the routing list R1,R3,...,Rn This is, in general principle, very
much like IPv4 Loose Source routing. Note, however, that routers between the listed routers R1...Rn do not
need to examine this header; they process the packet based only on its current destination address.

8.4.4 Fragment Header

IPv6 supports limited IPv4-style fragmentation via the Fragment Header. This header contains a 13-bit
Fragment Offset field, which contains as in IPv4 the 13 high-order bits of the actual 16-bit offset of the
fragment. This header also contains a 32-bit Identification field; all fragments of the same packet must carry
the same value in this field.
IPv6 fragmentation is done only by the original sender; routers along the way are not allowed to fragment
or re-fragment a packet. Sender fragmentation would occur if, for example, the sender had an 8KB IPv6
packet to send via UDP, and needed to fragment it to accommodate the 1500-byte Ethernet MTU.
If a packet needs to be fragmented, the sender first identifies the unfragmentable part, consisting of the
IPv6 fixed header and any extension headers that must accompany each fragment (these would include
Hop-by-Hop and Routing headers). These unfragmentable headers are then attached to each fragment.
IPv6 also requires that every link on the Internet have an MTU of at least 1280 bytes beyond the LAN
header; link-layer fragmentation and reassembly can be used to meet this MTU requirement (which is what
ATM links do).
8.4 IPv6 Extension Headers

157

An Introduction to Computer Networks, Release 1.0

Generally speaking, fragmentation should be avoided at the application layer when possible. UDP-based
applications that attempt to transmit filesystem-sized (usually 8 KB) blocks of data remain persistent users
of fragmentation.

8.5 Network Prefixes

We have been assuming that an IPv6 address, at least as seen by a host, is composed of a 64-bit network
prefix and a 64-bit host portion. As of this writing this is a requirement; RFC 4291 (IPv6 Addressing
Architecture) states:
For all unicast addresses, except those that start with the binary value 000, Interface IDs are
required to be 64 bits long....
This /64 requirement is occasionally revisited by the IETF, but we will assume here that it remains in force.
This is a departure from IPv4, where the host/subnet division point has depended, since the development of
subnets, on local configuration.
Note that while the net/host division point is fixed, routers may still use CIDR and may still route on any
shorter prefix than /64.
A typical IPv6 site may involve a variety of specialized network prefixes, including the link-local prefix
and prefixes for anycast and multicast addresses. Private IPv6 address prefixes, corresponding to IPv4s
[Link]/8, may also be in use.
The most important class of 64-bit network prefixes, however, are those supplied by a provider or other
address-numbering entity, and which represent the first half of globally routable IPv6 addresses. These are
the prefixes that will be visible to the outside world.
IPv6 customers will typically be assigned a relatively large block of addresses, eg /48 or /56. The former
allows 6448 = 16 bits for local subnet specification within a 64-bit network prefix; the latter allows 8
subnet bits. These subnet bits are as in IPv4 supplied through router configuration. The closest IPv6
analogue to the IPv4 subnet mask is that all network prefixes are supplied to hosts with an associated length,
although that length will often always be 64 bits.
Many sites will have only a single externally visible address block. However, some sites may be multihomed
and thus have multiple independent address blocks.
Sites may also have private unique local address prefixes, corresponding to IPv4 private address blocks like
[Link]/16 and [Link]/8. They are officially called Unique Local Unicast Addresses and are defined
in RFC 4193; these replace an earlier site-local address plan formally deprecated in RFC 3879. The first 8
bits of such a prefix are 1111 1101 (FD00::/8); the related prefix 1111 1100 (FC00::/8) is reserved for future
use. The last 16 bits of the 64-bit prefix represent the subnet ID, and are assigned either administratively or
via autoconfiguration. The 40 bits in between, from bit 8 up to bit 48, represent the Global ID. A site is to
set the Global ID to a pseudorandom value.
The resultant prefix is almost certainly globally unique, although it is not supposed to be routed and a site
would generally not admit any packets from the outside world addressed to a destination with the Global
ID as prefix. One rationale for choosing unique Global IDs for each site is to accommodate potential later
mergers of organizations without the need for renumbering; this has been a chronic problem for sites using
private IPv4 address blocks. Another justification is to accommodate VPN connections from other sites. For

158

8 IP version 6

An Introduction to Computer Networks, Release 1.0

example, if I use IPv4 block [Link]/8 at home, and connect using VPN to a site also using [Link]/8, it is
possible that my printer will have the same IPv4 address as their application server.

8.6 Neighbor Discovery

IPv6 Neighbor Discovery, or ND, is a protocol that replaces (and combines) several IPv4 tools, most notably
ARP, ICMP redirects and most non-address-assignment parts of DHCP. The original specification for ND is
in RFC 2461, later updated by RFC 4861. ND provides the following services:
Finding a local hosts LAN address, given its IPv6 address
Finding the local router(s)
Finding the set of network address prefixes that can be reached via local delivery (IPv6 allows there
to be more than one)
Determining that some neighbors are now unreachable
Detecting duplicate IPv6 addresses

8.6.1 Router Advertisement

IPv6 routers periodically send Router Advertisement packets; these are also sent in response to Router
Solicitation requests sent by hosts to the all-routers multicast group. Router Advertisement packets serve to
identify the routers; this process is sometimes called Router Discovery. These RA packets also contain a list
of all network address prefixes in use on the LAN. These packets may contain other important information
about the LAN as well, such as an agreed-on MTU; much of the information included in RA packets is, in
IPv4, sent via DHCP.
These IPv6 router messages represent a marked change from IPv4, in which routers need not send anything
besides forwarded packets, and in which hosts typically discover their default routers via DHCP. IPv6 routers
may even interact directly with ordinary hosts, if the SLAAC configuration mechanism (below) is used. To
become an IPv4 router, a node need only enable IPv4 forwarding in the kernel. To become an IPv6 router,
by comparison, in addition to forwarding a node MUST (RFC 4294) also run software to support Router
Advertisement; the linux radvd package is one option here.

8.6.2 Prefix Discovery

Closely related to Router Discovery is the Prefix Discovery process by which hosts learn what IPv6
network-address prefixes, above, are valid on the network. It is also where hosts learn which prefixes are
considered to be local to the hosts LAN, and thus reachable at the LAN layer instead of requiring router
assistance for delivery. IPv6, in other words, does not limit determination of whether delivery is local to the
IPv4 mechanism of having a node check a destination address against each of the network-address prefixes
assigned to the nodes interfaces.
Even IPv4 allows two IPv4 network prefixes to share the same LAN (eg a private one [Link]/24 and a
public one [Link]/24), but a consequence of IPv4 routing is that two such subnets can only reach one

8.6 Neighbor Discovery

159

An Introduction to Computer Networks, Release 1.0

another via a router on the LAN, even though they should in principle be able to communicate directly. IPv6
drops this restriction.
The Router Advertisement packets sent by the router should contain a complete list of valid network-address
prefixes, as the Prefix Information option. In simple cases this list may contain a single externally routable
64-bit prefix. If a particular LAN is part of multiple (overlapping) physical subnets, the prefix list will
contain an entry for each subnet; these 64-bit prefixes will themselves likely have a common prefix of length
N<64. For multihomed sites the prefix list may contain multiple unrelated prefixes corresponding to the
different address blocks. Finally, private and local IPv6 address prefixes may also be included.
Each prefix will have an associated lifetime; nodes receiving a prefix from an RA packet are to use it only
for the duration of this lifetime. On expiration (and likely much sooner) a node must obtain a newer RA
packet with a newer prefix list. The rationale for inclusion of the prefix lifetime is ultimately to allow sites
to easily renumber; that is, to change providers and switch to a new network-address prefix provided by a
new router. Each prefix is also tagged with a bit indicating whether it can be used for autoconfiguration, as
in 8.7.2 Stateless Autoconfiguration (SLAAC) below.

8.6.3 Neighbor Solicitation

Neighbor Solicitation messages are the IPv6 analogues of IPv4 ARP requests. These are essentially queries
of the form who has IPv6 address X? While ARP requests were broadcast, IPv6 Neighbor Solicitation
messages are sent to the solicited-node multicast address, which at the LAN layer usually represents a
rather small multicast group. This address is FF02::0001:x.y.z.w, where x, y, z and w are the low-order 32
bits of the IPv6 address the sender is trying to look up. Each IPv6 host on the LAN will need to subscribe to
all the solicited-node multicast addresses corresponding to its own IPv6 addresses (normally this is not too
many).
Neighbor Solicitation messages are repeated regularly, but followup verifications are initially sent to the
unicast LAN address on file (this is common practice with ARP implementations, but is optional). Unlike
with ARP, other hosts on the LAN are not expected to eavesdrop on the initial Neighbor Solicitation message.
The target hosts response to a Neighbor Solicitation message is called Neighbor Advertisement; a host
may also send these unsolicited if it believes its LAN address may have changed.
The analogue of Proxy ARP is still permitted, in that a node may send Neighbor Advertisements on behalf
of another. The most likely reason for this is that the node receiving proxy services is a mobile host
temporarily remote from the home LAN. Neighbor Advertisements sent as proxies have a flag to indicate
that, if the real target does speak up, the proxy advertisement should be ignored.
Hosts may also use the Authentication Header or the Encapsulated Security Payload Header to supply digital
signatures for ND packets. If a node is statically configured to require such checks, then the IPv6 analogue
of ARP spoofing (7.7.2 ARP security) can be prevented.
Finally, IPv4 ICMP Redirect messages have also been moved in IPv6 to the Neighbor Discovery protocol.

8.7 IPv6 Host Address Assignment

IPv6 provides two competing ways for hosts to obtain their full IP addresses. One is DHCPv6, based
on IPv4s DHCP (7.8 Dynamic Host Configuration Protocol (DHCP)). In addition to DHCPv6, IPv6
also supports StateLess Address AutoConfiguration, or SLAAC. The original idea behind SLAAC was
160

8 IP version 6

An Introduction to Computer Networks, Release 1.0

to support complete plug-and-play network setup: hosts on a completely isolated LAN could talk to one
another out of the box, and if a router was introduced connecting the LAN to the Internet, then hosts would
be able to determine unique, routable addresses from information available from the router.
In the early days of IPv6 development, in fact, DHCPv6 may have been intended only for address assignments to routers and servers, with SLAAC meant for ordinary hosts. In that era, it was still common for
IPv4 addresses to be assigned statically, via per-host configuration files. RFC 4862 states that SLAAC is
to be used when a site is not particularly concerned with the exact addresses hosts use, so long as they are
unique and properly routable.
SLAAC and DHCPv6 evolved to some degree in parallel. While SLAAC solves the autoconfiguration problem quite neatly, at this point DHCPv6 solves it just as effectively, and provides for greater administrative
control. For this reason, SLAAC may end up less widely deployed. On the other hand, SLAAC gives hosts
greater control over their IPv6 addresses, and so may end up offering hosts a greater degree of privacy by
allowing endpoint management of the use of private and temporary addresses (below).
When a host first begins the Neighbor Discovery process, it receives a Router Advertisement packet. In this
packet are two special bits: the M (managed) bit and the O (other configuration) bit. The M bit is set to
indicate that DHCPv6 is available on the network for address assignment. The O bit is set to indicate that
DHCPv6 is able to provide additional configuration information (eg the name of the DNS server) to hosts
that are using SLAAC to obtain their addresses.

8.7.1 Duplicate Address Detection

Whenever an IPv6 host obtains a unicast address a link-local address, an address created via SLAAC,
an address received via DHCPv6 or a manually configured address it goes through a duplicate-address
detection (DAD) process. The host sends one or more Neighbor Solicitation messages (that is, like an ARP
query), as in 8.6 Neighbor Discovery, asking if any other host has this address; if anyone answers, then
the address is a duplicate. As with IPv4 ACD (7.7.1 ARP Finer Points), but not as with the original IPv4
self-ARP, the source-IP-address field of this NS message is set to a special unspecified value; this allows
other hosts to recognize it as a DAD query.
Because this NS process may take some time, and because addresses are in fact almost always unique, RFC
4429 defines an optimistic DAD mechanism. This allows limited use of an address before the DAD process
completes; in the meantime, the address is marked as optimistic.
Outside the optimistic-DAD interval, a host is not allowed to use an IPv6 address if the DAD process has
failed. RFC 4862 in fact goes further: if a host with an established address receives a DAD query for that
address, indicating that some other host wants to use that address, then the original host should discontinue
use of the address.

8.7.2 Stateless Autoconfiguration (SLAAC)

To obtain an address via SLAAC, defined in RFC 4862, a host first generates its link-local address, as above
in 8.2.2 Link-local addresses, appending the standard 64-bit link-local prefix to its 64-bit host identifier
(8.2.1 Host identifier) derived from the LAN address. The host must then ensure that its newly configured
link-local address is in fact unique; it uses DAD (above) to verify this. Assuming no duplicate is found, then
at this point the host can talk to any other hosts on the same LAN, eg to figure out where the printers are.

8.7 IPv6 Host Address Assignment

161

An Introduction to Computer Networks, Release 1.0

The next step is to see if there is a router available. The host sends a Router Solicitation (RS) message
to the all-routers multicast address. A router if present should answer with a Router Advertisement
(RA) message that also contains a Prefix Information option; that is, a list of IPv6 network-address prefixes
(8.6.2 Prefix Discovery). The RA message will mark with a flag those prefixes eligible for use with
SLAAC; if no prefixes are so marked, then SLAAC should not be used. All prefixes will also be marked
with a lifetime, indicating how long the host may continue to use the prefix; once the prefix expires, the host
must obtain a new one via a new RA message.
The host chooses an appropriate prefix, stores the prefix-lifetime information, and, in the original version
of SLAAC, appends the prefix to the front of its host identifier to create what should now be a routable
address. The prefix length plus the host-identifier length must equal 128 bits; in the most common case each
is 64 bits. The address so formed must now be verified through the duplicate-address-detection mechanism
above.
An address generated in this way will, because of the embedded host identifier, uniquely identify the host
for all time. This includes identifying the host even when it is connected to a new network and is given
a different network prefix. Therefore, RFC 4941 defines a set of privacy extensions to SLAAC: optional
mechanisms for the generation of alternative host identifiers, based on pseudorandom generation using the
original LAN-address-based host identifier as a seed value. The probability of two hosts accidentally
choosing the same host identifier in this manner is very small; the Neighbor Solicitation mechanism with
DAD must, however, still be used to verify that the address is in fact unique. DHCPv6 also provides an
option for temporary address assignments, also to improve privacy, but one of the potential advantages of
SLAAC is that this process is entirely under the control of the end system.
Regularly (eg every few hours, or less) changing the host portion of an IPv6 address will make external
tracking of a host slightly more difficult. However, for a residential site with only a handful of hosts, a
considerable degree of tracking may be obtained simply by using the common 64-bit prefix.
In theory, if another host B on the LAN wishes to contact host A with a SLAAC-configured address containing the original host identifier, and B knows As IPv6 address AIPv6 , then B might extract As LAN address
from the low-order bits of AIPv6 . This was never actually allowed, however, even before the RFC 4941
privacy options, as there is no way for B to know that As address was generated via SLAAC at all. B would
always find As LAN address through the usual process of IPv6 Neighbor Solicitation.
A host using SLAAC may receive multiple network prefixes, and thus generate for itself multiple addresses.
RFC 6724 defines a process for a host to determine, when it wishes to connect to destination address D,
which of its own multiple addresses to use. For example, if D is a site-local address, not globally visible,
then the host will likely want to use an address that is also site-local. RFC 6724 also includes mechanisms
to allow a host with a permanent public address (eg corresponding to a DNS entry) to prefer alternative
temporary or privacy addresses for outbound connections.
At the end of the SLAAC process, the host knows its IPv6 address (or set of addresses) and its default router.
In IPv4, these would have been learned through DHCP along with the identity of the hosts DNS server; one
concern with SLAAC is that there is no obvious way for a host to find its DNS server. One strategy is to fall
back on DHCPv6 for this. However, RFC 6106 now defines a process by which IPv6 routers can include
DNS-server information in the RA packets they send to hosts as part of the SLAAC process; this completes
the final step of the autoconfiguration process.
How to get DNS names for SLAAC-configured IPv6 hosts into the DNS servers is an entirely separate
issue. One approach is simply not to give DNS names to such hosts. In the NAT-router model for IPv4
autoconfiguration, hosts on the inward side of the NAT router similarly do not have DNS names (although
162

8 IP version 6

An Introduction to Computer Networks, Release 1.0

they are also not reachable directly, while SLAAC IPv6 hosts would be reachable). If DNS names are needed
for hosts, then a site might choose DHCPv6 for address assignment instead of SLAAC. It is also possible to
figure out the addresses SLAAC would use (by identifying the host-identifier bits) and then creating DNS
entries for these hosts. Hosts can also use Dynamic DNS (RFC 2136) to update their own DNS records.

8.7.3 DHCPv6
The job of the DHCPv6 server is to tell an inquiring host its network prefix(es) and also supply a 64-bit host-identifier.
Hosts begin the process by sending a DHCPv6 request to the
All_DHCP_Relay_Agents_and_Servers multicast IPv6 address FF02::1:2 (versus the broadcast address for
IPv4). As with DHCPv4, the job of a relay agent is to tag a DHCP request with the correct current subnet, and then to forward it to the actual DCHPv6 server. This allows the DHCP server to be on a different subnet from the requester. Note that the use of multicast does nothing to diminish the need for
relay agents; use of the multicast group does not necessarily identify a requesters subnet. In fact, the
All_DHCP_Relay_Agents_and_Servers multicast address scope is limited to the current link; relay agents
then forward to the actual DHCP server using the site-scoped address All_DHCP_Servers.
Hosts using SLAAC to obtain their address can still use a special Information-Request form of DHCPv6 to
obtain their DNS server and any other static DHCPv6 information.
Clients may ask for temporary addresses. These are identified as such in the DHCPv6 request, and are
handled much like permanent address requests, except that the client may ask for a new temporary address
only a short time later. When the client does so, a different temporary address will be returned; a repeated
request for a permanent address, on the other hand, would usually return the same address as before.
When the DHCPv6 server returns a temporary address, it may of course keep a log of this address. The
absence of such a log is one reason SLAAC may provide a greater degree of privacy. Another concern
is that the DHCPv6 temporary-address sequence might have a flaw that would allow a remote observer to
infer a relationship between different temporary addresses; with SLAAC, a host is responsible itself for the
security of its temporary-address sequence and is not called upon to trust an external entity.
A DHCPv6 response contains a list (perhaps of length 1) of IPv6 addresses. Each separate address has an
expiration date. The client must send a new request before the expiration of any address it is actually using;
unlike for DHCPv4, there is no separate address lease lifetime.
In DHCPv4, the host portion of addresses typically comes from address pools representing small ranges
of integers such as 64-254; these values are generally allocated consecutively. A DHCPv6 server, on the
other hand, should take advantage of the enormous range (264 ) of possible host portions by allocating values
more sparsely, through the use of pseudorandomness. This makes it very difficult for an outsider who knows
one of a sites host addresses to guess the addresses of other hosts. Some DHCPv6 servers, however, do not
yet support this; such servers make the SLAAC approach more attractive.

8.7.4 Manual Configuration

Like IPv4 addresses, IPv6 addresses can also be set up purely by manual configuration. In theory, this would
be done only in the absence of an IPv6 router, in which case a unique-local prefix (8.5 Network Prefixes)
would likely be appropriate. See 8.10.3 Manual address configuration for one example. While it might

8.7 IPv6 Host Address Assignment

163

An Introduction to Computer Networks, Release 1.0

be convenient to distribute only the /64 prefix via manual configuration, and have SLAAC supply the loworder 64 bits, this option is not described in the SLAAC RFCs and seems not to be available in common
implementations.

8.8 Globally Exposed Addresses

Perhaps the most striking difference between a contemporary IPv4 network and an IPv6 network is that on
the former, many hosts are likely to be hidden behind a NAT router (1.14 Network Address Translation).
On an IPv6 network, on the other hand, every host is likely to be globally visible to the IPv6 world (though
NAT may still be used to allow connectivity to legacy IPv4 servers).
Legacy IPv4 NAT routers provide a measure of each of privacy, security and nuisance. Privacy in IPv6 can
be handled, as above, through private or temporary addresses.
The degree of security provided via NAT is largely if not entirely due to the fact that all connections must
be initiated from the inside; no packet from the outside is allowed through the NAT firewall unless it is a
response to a packet sent from the inside. This feature, however, can also be implemented via a conventional firewall (IPv4 or IPv6); one might need to configure the firewall with a list of inside addresses (or
subnets) meant to be globally visible. Furthermore, given such a conventional firewall, it would then be
straightforward to modify it so as to support limited and regulated connections from the outside world as
desired; an analogous modification of a NAT router is quite difficult. (It remains true, however, that as of
this writing consumer-grade IPv6 firewalls of the type described here do not really exist.) Finally, one of
the major reasons for hiding IPv4 addresses is that with IPv4 it is easy to map a /24 subnet by pinging or
otherwise probing each of the 254 possible hosts; such mapping may reveal internal structure. In IPv6 such
mapping is impractical as a /64 subnet has 264 18 quintillion hosts.
As for nuisance, NAT has always broken protocols that involve negotiation of new connections (eg TFTP, or
SIP, used by VoIP); IPv6 should make these much easier to manage.

8.9 ICMPv6
RFC 4443 defines an updated version of the ICMP protocol. It includes an IPv6 version of ICMP Echo
Request / Echo Reply, upon which the ping command is based. It also handles the error conditions below;
this list is somewhat cleaner than the corresponding ICMPv4 list:
Destination Unreachable
In this case, one of the following numeric codes is returned:
0. No route to destination, returned when a router has no next_hop entry.
1. Communication with destination administratively prohibited, returned when a router has a
next_hop entry, but declines to use it for policy reasons. Codes 5 and 6 are special cases; these
more-specific codes are returned when appropriate.
2. Beyond scope of source address, returned when a router is, for example, asked to route a packet
to a global address, but the return address is site-local. In IPv4, when a host with a private address
attempts to connect to a global address, NAT is almost always involved.

164

8 IP version 6

An Introduction to Computer Networks, Release 1.0

3. Address unreachable, a catchall category for routing failure not covered by any other message. An
example is if the packet was successfully routed to the last_hop router, but Neighbor Discovery failed
to find a LAN address corresponding to the IPv6 address.
4. Port unreachable, returned when, as in ICMPv4, the destination host does not have the requested
UDP port open.
5. Source address failed ingress/egress policy, see code 1.
6. Reject route to destination, see code 1.
Packet Too Big
This is like ICMPv4s Fragmentation Required but DontFragment flag sent; IPv6 however has no routerbased fragmentation.
Time Exceeded
This is used for cases where the Hop Limit was exceeded, and also where source-based fragmentation was
used and the fragment-reassembly timer expired.
Parameter Problem
This is used when there is a malformed entry in the IPv6 header, eg an unrecognized Next Header value.

8.10 Routerless Connection Examples

Most IPv6 networks require at least one IPv6 router; both SLAAC and DHCPv6 configuration requires
this. However, link-local addresses are quite serviceable, as long as one remembers that the interface must
be specified; manually configured addresses are another option. Here are some examples. One practical
problem with link-local addresses is that application documentation describing how to include a specification
of the interface is sometimes sparse.

8.10.1 ping6
We will start with the linux version of ping6, the IPv6 analogue of the familiar ping command. It is
used to send ICMPv6 Echo Requests. The ping6 command supports an option to specify the interface (-I
eth0); as noted above, this is mandatory when sending to link-local addresses.
ping6 ::1: This allows me to ping my own loopback address.
ping6 -I eth0 ff02::1: This pings the all-nodes multicast group on interface eth0. I get these answers:
64 bytes from fe80::3e97:eff:fe2c:2beb (this is the host I am pinging from)
64 bytes from fe80::2a0:ccff:fe24:b0e4 (another linux host)
My VoIP phone on the same subnet but apparently supporting IPv4 only remains mute.
ping6 -I eth0 fe80::6267:20ff:fe72:8960: This pings the link-local address of the other linux host answering
the previous query. Note the ff:fe in the host identifier. Also note the flipped seventh bit of the two bytes
02a0; the other linux host has Ethernet address [Link].

8.10 Routerless Connection Examples

165

An Introduction to Computer Networks, Release 1.0

8.10.2 TCP connections with link-local addresses

The next step is to create a TCP connection. Some commands, like ping6 above, may provide for a way of
specifying the interface as an option. Failing that, many linux systems allow appending %interface to the
address (which unfortunately usually is required to be in numeric form). The following, for example, allows
me to connect using ssh to the other linux host above:
ssh fe80::2a0:ccff:fe24:b0e4%eth0
That the ssh service was listening for IPv6 connections can be verified on that host by
netstat -a | grep -i tcp6
That the ssh connection actually used IPv6 can be verified by, say, use of a network sniffer like WireShark
(for which the filter expression [Link] == 6 may be useful).
If the connection fails, but ssh works for IPv4 connections and shows as listening in the tcp6 list from the
netstat command, a blocked port is a likely suspect. In this case the commands ufw and ip6tables
--list may be useful.

8.10.3 Manual address configuration

The use of manually configured addresses, 8.7.4 Manual Configuration, is also possible; this avoids the
need to specify an interface. The first step is to pick a unique-local prefix, eg [Link]/64 (note
that this particular prefix does not meet the randomness rules for unique-local address prefixes). Low-order
bits can then be added and addresses assigned manually; on linux this is done with:
host1: ip -6 address add [Link] dev eth0
host2: ip -6 address add [Link] dev eth0
Now on host1 the command
ssh [Link]
should work, again assuming ssh is listening for IPv6 connections. Because the address here is not linklocal, /etc/host entries may also be created, eg the following on host1:
[Link] host2-6
Now to connect all that should be needed is
ssh host2-6

8.10.4 Node Information Messages

In addition to ICMPv6, IPv6 also has Node Information (NI) Messages, defined in RFC 4620. One form of
NI query allows a host to be asked directly for its name; this is accomplished in IPv4 via DNS reverse-name
lookups. Other NI queries allow a host to be asked for its other IPv6 addresses, or for its IPv4 addresses.

166

8 IP version 6

An Introduction to Computer Networks, Release 1.0

8.11 IPv6-to-IPv4 connectivity

What happens if you switch to IPv6 completely, perhaps because your ISP has run out of IPv4 addresses?
Ideally you will only need to talk to IPv6 servers. For example, the DNS name [Link] corresponds to an IPv4 address, but also to an IPv6 address. If there is not IPv6 connectivity between you and
the IPv6 server you are trying to reach, tunneling of IPv6 traffic over IPv4 can be used.
But what do you do if you have only an IPv6 address and want to reach an IPv4-only server? IPv4 and IPv6
cannot directly interoperate; while IPv6 could have been modified to support an interoperability feature, a
change of such magnitude in IPv4 is politically and economically impossible. Thus, you will have to find
some sort of translator, such as an IPv6-to-IPv4 NAT router. RFC 2766 defines an IPv6-to-IPv4 flavor of
NAT that also includes any necessary translation of the higher transport layer as well. The NAT translator
will have an IPv4 address (otherwise it cannot talk to other IPv4 nodes), but that address can be shared
among multiple IPv6 clients.

8.12 Epilog
IPv4 has run out of large address blocks, as of 2011. IPv6 has reached a mature level of development. Most
common operating systems provide excellent IPv6 support.
Yet conversion has been slow. Many ISPs still provide limited (to nonexistent) support, and inexpensive
IPv6 firewalls to replace the ubiquitous consumer-grade NAT routers do not really exist. Time will tell how
all this evolves. However, while IPv6 has now been around for twenty years, top-level IPv4 address blocks
disappeared just three years ago. It is quite possible that this will prove to be just the catalyst IPv6 needs.

8.13 Exercises
1. Each IPv6 address is associated with a specific solicited-node multicast address. Explain why, on a
typical Ethernet, if the original IPv6 host address was obtained via SLAAC then the LAN multicast group
corresponding to the hosts solicited-node multicast addresses is likely to be small, in many cases consisting
of one host only. (Packet delivery to small LAN multicast groups can be much more efficient than delivery
to large multicast groups.)
(b). What steps might a DHCPv6 server take to ensure that, for the IPv6 addresses it hands out, the LAN
multicast groups corresponding to the host addresses solicited-node multicast addresses will be small?
2. If an attacker sends a large number of probe packets via IPv4, you can block them by blocking the
attackers IP address. Now suppose the attacker uses IPv6 to launch the probes; for each probe, the attacker
changes the low-order 64 bits of the address. Can these probes be blocked efficiently? If so, what do you
have to block? Might you also be blocking other users?
3. Suppose someone tried to implement ping6 so that, if the address was a link-local address and no interface
was specified, the ICMPv6 Echo Request was sent out all non-loopback interfaces. Could the end result be
different than conventional ping6 with the correct interface supplied? If so, how likely is this?
4. Create an IPv6 ssh connection as in 8.10 Routerless Connection Examples. Examine the connections
packets using WireShark or the equivalent. Does the TCP handshake (12.3 TCP Connection Establishment)
look any different over IPv6?
8.11 IPv6-to-IPv4 connectivity

167

An Introduction to Computer Networks, Release 1.0

5. Create an IPv6 ssh connection using manually configured addresses as in 8.10.3 Manual address
configuration. Again use WireShark or the equivalent to monitor the connection. Is DAD (8.7.1 Duplicate
Address Detection) used?

168

8 IP version 6

9 ROUTING-UPDATE ALGORITHMS

How do IP routers build and maintain their forwarding tables?

Ethernet bridges can always fall back on broadcast, so they can afford to build their forwarding tables
incrementally, putting a host into the forwarding table only when that host is first seen as a sender. For
IP, there is no fallback delivery mechanism: forwarding tables must be built before delivery can succeed.
While manual table construction is possible, it is not practical.
In the literature it is common to refer to router-table construction as routing algorithms. We will avoid
that term, however, to avoid confusion with the fundamental datagram-forwarding algorithm; instead, we
will call these routing-update algorithms.
The two classes of algorithms we will consider here are distance-vector and link-state. Both assume that
consistent information is available as to the cost of each link (eg that the two routers at opposite ends of each
link know this cost). This requirement classifies these algorithms as interior routing-update algorithms: the
routers involved are internal to a larger organization or other common administrative regime that has an
established policy on how to assign link weights. The set of routers following a common policy is known as
a routing domain or (from the BGP protocol) an autonomous system. The simplest link-weight strategy
is to give each link a cost of 1; link costs can also be based on bandwidth, propagation delay, financial cost,
or administrative preference value.
In the following chapter we will look at the Border Gateway Protocol, or BGP, in which no link-cost calculations are made, because links may be between unrelated organizations which have no agreement in place
on determining link cost.
Generally, all these algorithms apply to IPv6 as well as IPv4, though specific protocols of course may need
modification.
Finally, we should point out that from the early days of the Internet, routing was allowed to depend not just
on the destination, but also on the quality of service (QoS) requested; thus, forwarding table entries are
strictly speaking not destination, next_hop but rather destination, QoS, next_hop. Originally, the Type
of Service field in the IPv4 header (7.1 The IPv4 Header) could be used to specify QoS (often then called
ToS). Packets could request low delay, high throughput or high reliability, and could be routed accordingly.
In practice, the Type of Service field was rarely used, and was eventually taken over by the DS field and
ECN bits. The first three bits of the Type of Service field, known as the precedence bits, remain available,
however, and can still be used for QoS routing purposes (see the Class Selector PHB of 19.7 Differentiated
Services for examples of these bits). See also RFC 2386.
In much of the following, we are going to ignore QoS information, and assume that routing decisions
are based only on the destination. See, however, the first paragraph of 9.5 Link-State Routing-Update
Algorithm, and also 9.6 Routing on Other Attributes.

9.1 Distance-Vector Routing-Update Algorithm

Distance-Vector is the simplest routing-update algorithm, used by the Routing Information Protocol, or RIP.
On unix-based systems the process in charge of this is often called routed (pronounced route-d).

169

An Introduction to Computer Networks, Release 1.0

Routers identify their router neighbors (through some sort of neighbor-discovery mechanism), and
add a third column to their forwarding tables for cost; table entries are thus of the form
destination,next_hop,cost. The simplest case is to assign a cost of 1 to each link (the hopcount metric); it is also possible to assign more complex numbers.
Each router then reports the destination,cost portion of its table to its neighboring routers at regular
intervals (these table portions are the vectors of the algorithm name). It does not matter if neighbors
exchange reports at the same time, or even at the same rate.
Each router also monitors its continued connectivity to each neighbor; if neighbor N becomes unreachable
then its reachability cost is set to infinity.
Actual destinations in IP would be networks attached to routers; one router might be directly connected
to several such destinations. In the following, however, we will identify all a routers directly connected
networks with the router itself. That is, we will build forwarding tables to reach every router. While
it is possible that one destination network might be reachable by two or more routers, thus breaking our
identification of a router with its set of attached networks, in practice this is of little concern. See exercise 4
for an example in which networks are not identified with adjacent routers.

9.1.1 Distance-Vector Update Rules

Let A be a router receiving a report D,cD from neighbor N at distance cN . Note that this means A can
reach D via N with cost c = cD + cN . A updates its own table according to the following three rules:
1. New destination: D is a previously unknown destination. A adds D,N,cD +cN
2. Lower cost: D is a known destination with entry D,M,c, but c > cD +cN . A updates the entry for D
to D,N,cD +cN . It is possible that M=N, meaning that N is now reporting a distance decrease to D.
3. Next_hop increase: A has a previous entry D,N,c, but c < cD +cN . A updates the entry for D to
D,N,cD +cN . N is As next_hop to D, and N is now reporting a distance increase.
The first two rules are for new destinations and a shorter path to existing destinations. In these cases, the cost
to each destination monotonically decreases (at least if we consider all unreachable destinations as being at
distance ). Convergence is automatic, as the costs cannot decrease forever.
The third rule, however, introduces the possibility of instability, as a cost may also go up. It represents the
bad-news case, in that neighbor N has learned that some link failure has driven up its own cost to reach D,
and is now passing that bad news on to A, which routes to D via N.
The next_hop-increase case only passes bad news along; the very first cost increase must always come from
a router discovering that a neighbor N is unreachable, and thus updating its cost to N to . Similarly, if
router A learns of a next_hop increase to destination D from neighbor B, then we can follow the next_hops
back until we reach a router C which is either the originator of the cost= report, or which has learned of
an alternative route through one of the first two rules.

9.1.2 Example 1
For our first example, no links will break and thus only the first two rules above will be used. We will start
out with the network below with empty forwarding tables; all link costs are 1.

170

9 Routing-Update Algorithms

An Introduction to Computer Networks, Release 1.0

After initial neighbor discovery, here are the forwarding tables. Each node has entries only for its directly
connected neighbors:
A: B,B,1 C,C,1 D,D,1
B: A,A,1 C,C,1
C: A,A,1 B,B,1 E,E,1
D: A,A,1 E,E,1
E: C,C,1 D,D,1
Now let D report to A; it sends records A,1 and E,1. A ignores Ds A,1 record, but E,1 represents a
new destination; A therefore adds E,D,2 to its table. Similarly, let A now report to D, sending B,1 C,1
D,1 E,2 (the last is the record we just added). D ignores As records D,1 and E,2 but As records
B,1 and C,1 cause D to create entries B,A,2 and C,A,2. A and Ds tables are now, in fact, complete.
Now suppose C reports to B; this gives B an entry E,C,2. If C also reports to E, then Es table will have
A,C,2 and B,C,2. The tables are now:
A: B,B,1 C,C,1 D,D,1 E,D,2
B: A,A,1 C,C,1 E,C,2
C: A,A,1 B,B,1 E,E,1
D: A,A,1 E,E,1 B,A,2 C,A,2
E: C,C,1 D,D,1 A,C,2 B,C,2
We have two missing entries: B and C do not know how to reach D. If A reports to B and C, the tables will
be complete; B and C will each reach D via A at cost 2. However, the following sequence of reports might
also have occurred:
E reports to C, causing C to add D,E,2
C reports to B, causing B to add D,C,3
In this case we have 100% reachability but B routes to D via the longer-than-necessary path BCED.
However, one more report will fix this: suppose A reports to B. B will received D,1 from A, and will
update its entry D,C,3 to D,A,2.
Note that A routes to E via D while E routes to A via C; this asymmetry was due to indeterminateness in the
order of initial table exchanges.
If all link weights are 1, and if each pair of neighbors exchange tables once before any pair starts a second
exchange, then the above process will discover the routes in order of length, ie the shortest paths will be the
first to be discovered. This is not, however, a particularly important consideration.

9.1 Distance-Vector Routing-Update Algorithm

171

An Introduction to Computer Networks, Release 1.0

9.1.3 Example 2
The next example illustrates link weights other than 1. The first route discovered between A and B is the
direct route with cost 8; eventually we discover the longer ACDB route with cost 2+1+3=6.

The initial tables are these:

A: C,C,2 B,B,8
B: A,A,8 D,D,3
C: A,A,2 D,D,1
D: B,B,3 C,C,1
After A and C exchange, A has D,C,3 and C has B,A,10. After C and D exchange, C updates its B,A,10
entry to B,D,4 and D adds A,C,3; D receives Cs report of B,10 but ignores it. Now finally suppose
B and D exchange. D ignores Bs route to A, as it has a better one. B, however, gets Ds report A,3 and
updates its entry for A to A,D,6. At this point the tables are as follows:
A: C,C,2 B,B,8 D,C,3
B: A,D,6 D,D,3
C: A,A,2 D,D,1 B,D,4
D: B,B,3 C,C,1 A,C,3
We have two more things to fix before we are done: A has an inefficient route to B, and B has no route to
C. The first will be fixed when C reports B,4 to A; A will replace its route to B with B,C,6. The second
will be fixed when D reports to B; if A reports to B first then B will temporarily add the inefficient route
C,A,10; this will change to C,D,4 when Ds report to B arrives. If we look only at the AB route, B
discovers the lower-cost route to A once, first, C reports to D and, second, after that, D reports to B; a similar
sequence leads to As discovering the lower-cost route.

9.1.4 Example 3
Our third example will illustrate how the algorithm proceeds when a link breaks. We return to the first
diagram, with all tables completed, and then suppose the DE link breaks. This is the bad-news case: a
link has broken, and is no longer available; this will bring the third rule into play.

172

9 Routing-Update Algorithms

An Introduction to Computer Networks, Release 1.0

We shall assume, as above, that A reaches E via D, but we will here assume contrary to Example 1 that
C reaches D via A (see exercise 3.5 for the original case).
Initially, upon discovering the break, D and E update their tables to E,-, and D,-, respectively
(whether or not they actually enter into their tables is implementation-dependent; we may consider this
as equivalent to removing their entries for one another; the - as next_hop indicates there is no next_hop).
Eventually D and E will report the break to their respective neighbors A and C. A will apply the bad-news
rule above and update its entry for E to E,-,. We have assumed that C, however, routes to D via A, and
so it will ignore Es report.
We will suppose that the next steps are for C to report to E and to A. When C reports its route D,2 to E,
E will add the entry D,C,3, and will again be able to reach D. When C reports to A, A will add the route
E,C,2. The final step will be when A next reports to D, and D will have E,A,3. Connectivity is restored.

9.1.5 Example 4
The previous examples have had a global perspective in that we looked at the entire network. In the next
example, we look at how one specific router, R, responds when it receives a distance-vector report from its
neighbor S. Neither R nor S nor we have any idea of what the entire network looks like. Suppose Rs table
is initially as follows, and the SR link has cost 1:
destination
A
B
C
D

next_hop
S
T
S
U

cost
3
4
5
6

S now sends R the following report, containing only destinations and its costs:
destination
A
B
C
D
E

cost
2
3
5
4
2

R then updates its table as follows:

9.1 Distance-Vector Routing-Update Algorithm

173

An Introduction to Computer Networks, Release 1.0

destination
A
B
C
D
E

next_hop
S
T
S
S
S

cost
3
4
6
5
3

reason
No change; S probably sent this report before
No change; Rs cost via S is tied with Rs cost via T
Next_hop increase
Lower-cost route via S
New destination

Whatever Ss cost to a destination, Rs cost to that destination via S is one greater.

9.2 Distance-Vector Slow-Convergence Problem

There is a significant problem with Distance-Vector table updates in the presence of broken links. Not only
can routing loops form, but the loops can persist indefinitely! As an example, suppose we have the following
arrangement, with all links having cost 1:

Now suppose the DA link breaks:

If A immediately reports to B that D is no longer reachable (distance = ), then all is well. However, it is
possible that B reports to A first, telling A that it has a route to D, with cost 2, which B still believes it has.
This means A now installs the entry D,B,3. At this point we have what we called in 1.6 Routing Loops a
linear routing loop: if a packet is addressed to D, A will forward it to B and B will forward it back to A.
Worse, this loop will be with us a while. At some point A will report D,3 to B, at which point B will
update its entry to D,A,4. Then B will report D,4 to A, and As entry will be D,B,5, etc. This process
is known as slow convergence to infinity. If A and B each report to the other once a minute, it will take
2,000 years for the costs to overflow an ordinary 32-bit integer.

9.2.1 Slow-Convergence Fixes

The simplest fix to this problem is to use a small value for infinity. Most flavors of the RIP protocol use
infinity=16, with updates every 30 seconds. The drawback to so small an infinity is that no path through
the network can be longer than this; this makes paths with weighted link costs difficult. Cisco IGRP uses a
variable value for infinity up to a maximum of 256; the default infinity is 100.
There are several well-known other fixes:

174

9 Routing-Update Algorithms

An Introduction to Computer Networks, Release 1.0

[Link] Split Horizon

Under split horizon, if A uses N as its next_hop for destination D, then A simply does not report to N that
it can reach D; that is, in preparing its report to N it first deletes all entries that have N as next_hop. In the
example above, split horizon would mean B would never report to A about the reachability of D because A
is Bs next_hop to D.
Split horizon prevents all linear routing loops. However, there are other topologies where it cannot prevent
loops. One is the following:

Suppose the A-D link breaks, and A updates to D,-,. A then reports D, to B, which updates its
table to D,-,. But then, before A can also report D, to C, C reports D,2 to B. B then updates to
D,C,3, and reports D,3 back to A; neither this nor the previous report violates split-horizon. Now As
entry is D,B,4. Eventually A will report to C, at which point Cs entry becomes D,A,5, and the numbers
keep increasing as the reports circulate counterclockwise. The actual routing proceeds in the other direction,
clockwise.
Split horizon often also includes poison reverse: if A uses N as its next_hop to D, then A in fact reports
D, to N, which is a more definitive statement that A cannot reach D by itself. However, coming up with
a scenario where poison reverse actually affects the outcome is not trivial.
[Link] Triggered Updates
In the original example, if A was first to report to B then the loop resolved immediately; the loop occurred
if B was first to report to A. Nominally each outcome has probability 50%. Triggered updates means that
any router should report immediately to its neighbors whenever it detects any change for the worse. If A
reports first to B in the first example, the problem goes away. Similarly, in the second example, if A reports
to both B and C before B or C report to one another, the problem goes away. There remains, however, a
small window where B could send its report to A just as A has discovered the problem, before A can report
to B.
[Link] Hold Down
Hold down is sort of a receiver-side version of triggered updates: the receiver does not use new alternative
routes for a period of time (perhaps two router-update cycles) following discovery of unreachability. This
gives time for bad news to arrive. In the first example, it would mean that when A received Bs report D,2,
it would set this aside. It would then report D, to B as usual, at which point B would now report D,
back to A, at which point Bs earlier report D,2 would be discarded. A significant drawback of hold down
is that legitimate new routes are also delayed by the hold-down period.

9.2 Distance-Vector Slow-Convergence Problem

175

An Introduction to Computer Networks, Release 1.0

These mechanisms for preventing slow convergence are, in the real world, quite effective. The Routing
Information Protocol (RIP, RFC 2453) implements all but hold-down, and has been widely adopted at
smaller installations.
However, the potential for routing loops and the limited value for infinity led to the development of alternatives. One alternative is the link-state strategy, 9.5 Link-State Routing-Update Algorithm. Another
alternative is Ciscos Enhanced Interior Gateway Routing Protocol, or EIGRP, 9.4.2 EIGRP. While part of
the distance-vector family, EIGRP is provably loop-free, though to achieve this it must sometimes suspend
forwarding to some destinations while tables are in flux.

9.3 Observations on Minimizing Route Cost

Does distance-vector routing actually achieve minimum costs? For that matter, does each packet incur the
cost its sender expects? Suppose node A has a forwarding entry D,B,c, meaning that A forwards packets
to destination D via next_hop B, and expects the total cost to be c. If A sends a packet to D, and we follow
it on the actual path it takes, must the total link cost be c? If so, we will say that the network has accurate
costs.
The answer to the accurate-costs question, as it turns out, is yes for the distance-vector algorithm, if we
follow the rules carefully, and the network is stable (meaning that no routing reports are changing, or, more
concretely, that every update report now circulating is based on the current network state); a proof is below.
However, if there is a routing loop, the answer is of course no: the actual cost is now infinite. The answer
would also be no if As neighbor B has just switched to using a longer route to D than it last reported to A.
It turns out, however, that we seek the shortest route not because we are particularly trying to save money
on transit costs; a route 50% longer would generally work just fine. (AT&T, back when they were the
Phone Company, once ran a series of print advertisements claiming longer routes as a feature: if the direct
path was congested, they could still complete your call by routing you the long way round.) However, we
are guaranteed that if all routers seek the shortest route and if the network is stable then all paths are
loop-free, because in this case the network will have accurate costs.
Here is a simple example illustrating the importance of global cost-minimization in preventing loops. Suppose we have a network like this one:

Now suppose that A and B use distance-vector but are allowed to choose the shortest route to within 10%. A
would get a report from C that D could be reached with cost 1, for a total cost of 21. The forwarding entry
via C would be D,C,21. Similarly, A would get a report from B that D could be reached with cost 21, for
a total cost of 22: D,B,22. Similarly, B has choices D,C,21 and D,A,22.
If A and B both choose the minimal route, no loop forms. But if A and B both use the 10%-overage rule,
they would be allowed to choose the other route: A could choose D,B,22 and B could choose D,A,22.

176

9 Routing-Update Algorithms

An Introduction to Computer Networks, Release 1.0

If this happened, we would have a routing loop: A would forward packets for D to B, and B would forward
them right back to A.
As we apply distance-vector routing, each router independently builds its tables. A router might have some
notion of the path its packets would take to their destination; for example, in the case above A might believe
that with forwarding entry D,B,22 its packets would take the path ABCD (though in distance-vector
routing, routers do not particularly worry about the big picture). Consider again the accurate-cost question
above. This fails in the 10%-overage example, because the actual path is now infinite.
We now prove that, in distance-vector routing, the network will have accurate costs, provided
each router selects what it believes to be the shortest path to the final destination, and
the network is stable, meaning that further dissemination of any reports would not result in changes
To see this, suppose the actual route taken by some packet from source to destination, as determined by
application of the distributed algorithm, is longer than the cost calculated by the source. Choose an example
of such a path with the fewest number of links, among all such paths in the network. Let S be the source,
D the destination, and k the number of links in the actual path P. Let Ss forwarding entry for D be D,N,c,
where N is Ss next_hop neighbor. To have obtained this route through the distance-vector algorithm, S must
have received report D,cD from N, where we also have the cost of the SN link as cN and c = cD + cN . If
we follow a packet from N to D, it must take the same path P with the first link deleted; this sub-path has
length k-1 and so, by our hypothesis that k was the length of the shortest path with non-accurate costs, the
cost from N to D is cD . But this means that the cost along path P, from S to D via N, must be cD + cN = c,
contradicting our selection of P as a path longer than its advertised cost.
There is one final observation to make about route costs: any cost-minimization can occur only within
a single routing domain, where full information about all links is available. If a path traverses multiple
routing domains, each separate routing domain may calculate the optimum path traversing that domain. But
these local minimums do not necessarily add up to a globally minimal path length, particularly when
one domain calculates the minimum cost from one of its routers only to the other domain rather than to a
router within that other domain. Here is a simple example. Routers BR1 and BR2 are the border routers
connecting the domain LD to the left of the vertical dotted line with domain RD to the right. From A to B,
LD will choose the shortest path to RD (not to B, because LD is not likely to have information about links
within RD). This is the path of length 3 through BR2. But this leads to a total path length of 3+8=11 from
A to B; the global minimum path length, however, is 4+1=5, through BR1.

In this example, domains LD and RD join at two points. For a route across two domains joined at only a
single point, the domain-local shortest paths do add up to the globally shortest path.

9.3 Observations on Minimizing Route Cost

177

An Introduction to Computer Networks, Release 1.0

9.4 Loop-Free Distance Vector Algorithms

It is possible for routing-update algorithms based on the distance-vector idea to eliminate routing loops
and thus the slow-convergence problem entirely. We present brief descriptions of two such algorithms.

9.4.1 DSDV
DSDV, or Destination-Sequenced Distance Vector, was proposed in [PB94]. It avoids routing loops by
the introduction of sequence numbers: each router will always prefer routes with the most recent sequence
number, and bad-news information will always have a lower sequence number then the next cycle of corrected information.
DSDV was originally proposed for MANETs (3.3.8 MANETs) and has some additional features for traffic
minimization that, for simplicity, we ignore here. It is perhaps best suited for wired networks and for small,
relatively stable MANETs.
DSDV forwarding tables contain entries for every other reachable node in the system. One successor of
DSDV, Ad Hoc On-Demand Distance Vector routing or AODV, allows forwarding tables to contain only
those destinations in active use; a mechanism is provided for discovery of routes to newly active destinations.
See [PR99] and RFC 3561.
Under DSDV, each forwarding table entry contains, in addition to the destination, cost and next_hop, the current sequence number for that destination. When neighboring nodes exchange their distance-vector reachability reports, the reports include these per-destination sequence numbers.
When a router R receives a report from neighbor N for destination D, and the report contains a sequence
number larger than the sequence number for D currently in Rs forwarding table, then R always updates to
use the new information. The three cost-minimization rules of 9.1.1 Distance-Vector Update Rules above
are used only when the incoming and existing sequence numbers are equal.
Each time a router R sends a report to its neighbors, it includes a new value for its own sequence number,
which it always increments by 2. This number is then entered into each neighbors forwarding-table entry for
R, and is then propagated throughout the network via continuing report exchanges. Any sequence number
originating this way will be even, and whenever another nodes forwarding-table sequence number for R is
even, then its cost for R will be finite.
Infinite-cost reports are generated in the usual way when former neighbors discover they can no longer reach
one another; however, in this case each node increments the sequence number for its former neighbor by 1,
thus generating an odd value. Any forwarding-table entry with infinite cost will thus always have an odd
sequence number. If A and B are neighbors, and As current sequence number is s, and the AB link breaks,
then B will start reporting A at distance with sequence number s+1 while A will start reporting its own
new sequence number s+2. Any other node now receiving a report originating with B (with sequence number
s+1) will mark A as having cost , but will obtain a valid route to A upon receiving a report originating
from A with new (and larger) sequence number s+2.
The triggered-update mechanism is used: if a node receives a report with some destinations newly marked
with infinite cost, it will in turn forward this information immediately to its other neighbors, and so on. This
is, however, not essential; bad and good reports are distinguished by sequence number, not by relative
arrival time.

178

9 Routing-Update Algorithms

An Introduction to Computer Networks, Release 1.0

It is now straightforward to verify that the slow-convergence problem is solved. After a link break, if there is
some alternative path from router R to destination D, then R will eventually receive Ds latest even sequence
number, which will be greater than any sequence number associated with any report listing D as unreachable.
If, on the other hand, the break partitioned the network and there is no longer any path to D from R, then the
highest sequence number circulating in Rs half of the original network will be odd and the associated table
entries will all list D at cost . One way or another, the network will quickly settle down to a state where
every destinations reachability is accurately described.
In fact, a stronger statement is true: not even transient routing loops are created. We outline a proof. First,
whenever router R has next_hop N for a destination D, then Ns sequence number for D must be greater
than or equal to Rs, as R must have obtained its current route to D from one of Ns reports. A consequence
is that all routers participating in a loop for destination D must have the same (even) sequence number s for
D throughout. This means that the loop would have been created if only the reports with sequence number
s were circulating. As we noted in 9.1.1 Distance-Vector Update Rules, any application of the next_hopincrease rule must trace back to a broken link, and thus must involve an odd sequence number. Thus, the
loop must have formed from the sequence-number-s reports by the application of the first two rules only.
But this violates the claim in Exercise 10.
There is one drawback to DSDV: nodes may sometimes briefly switch to routes that are longer than optimum
(though still correct). This is because a router is required to use the route with the newest sequence number,
even if that route is longer than the existing route. If A and B are two neighbors of router R, and B is closer
to destination D but slower to report, then every time Ds sequence number is incremented R will receive
As longer route first, and switch to using it, and Bs shorter route shortly thereafter.
DSDV implementations usually address this by having each router R keep track of the time interval between
the first arrival at R of a new route to a destination D with a given sequence number, and the arrival of the
best route with that sequence number. During this interval following the arrival of the first report with a new
sequence number, R will use the new route, but will refrain from including the route in the reports it sends
to its neighbors, anticipating that a better route will soon arrive.
This works best when the hopcount cost metric is being used, because in this case the best route is likely
to arrive first (as the news had to travel the fewest hops), and at the very least will arrive soon after the first
route. However, if the networks cost metric is unrelated to the hop count, then the time interval between
first-route and best-route arrivals can involve multiple update cycles, and can be substantial.

9.4.2 EIGRP
EIGRP, or the Enhanced Interior Gateway Routing Protocol, is a once-proprietary Cisco distance-vector
protocol that was released as an Internet Draft in February 2013. As with DSDV, it eliminates the risk of
routing loops, even ephemeral ones. It is based on the distributed update algorithm (DUAL) of [JG93].
EIGRP is an actual protocol; we present here only the general algorithm. Our discussion follows [CH99].
Each router R keeps a list of neighbor routers NR , as with any distance-vector algorithm. Each R also
maintains a data structure known (somewhat misleadingly) as its topology table. It contains, for each
destination D and each N in NR , an indication of whether N has reported the ability to reach D and, if so, the
reported cost c(D,N). The router also keeps, for each N in NR , the cost cN of the link from R to N. Finally,
the forwarding-table entry for any destination can be marked passive, meaning safe to use, or active,
meaning updates are in process and the route is temporarily unavailable.

9.4 Loop-Free Distance Vector Algorithms

179

An Introduction to Computer Networks, Release 1.0

Initially, we expect that for each router R and each destination D, Rs next_hop to D in its forwarding table
is the neighbor N for which the following total cost is a minimum:
c(D,N) + cN
Now suppose R receives a distance-vector report from neighbor N1 that it can reach D with cost c(D,N1 ).
This is processed in the usual distance-vector way, unless it represents an increased cost and N1 is Rs
next_hop to D; this is the third case in 9.1.1 Distance-Vector Update Rules. In this case, let C be Rs
current cost to D, and let us say that neighbor N of R is a feasible next_hop (feasible successor in Ciscos
terminology) if Ns cost to D (that is, c(D,N)) is strictly less than C. R then updates its route to D to use the
feasible neighbor N for which c(D,N) + cN is a minimum. Note that this may not in fact be the shortest path;
it is possible that there is another neighbor M for which c(D,M)+cM is smaller, but c(D,M)C. However,
because Ns path to D is loop-free, and because c(D,N) < C, this new path through N must also be loop-free;
this is sometimes summarized by the statement one cannot create a loop by adopting a shorter route.
If no neighbor N of R is feasible which would be the case in the DAB example of 9.2 Distance-Vector
Slow-Convergence Problem, then R invokes the DUAL algorithm. This is sometimes called a diffusion
algorithm as it invokes a diffusion-like spread of table changes proceeding away from R.
Let C in this case denote the new cost from R to D as based on N1 s report. R marks destination D as
active (which suppresses forwarding to D) and sends a special query to each of its neighbors, in the form
of a distance-vector report indicating that its cost to D has now increased to C. The algorithm terminates
when all Rs neighbors reply back with their own distance-vector reports; at that point R marks its entry for
D as passive again.
Some neighbors may be able to process Rs report without further diffusion to other nodes, remain passive,
and reply back to R immediately. However, other neighbors may, like R, now become active and continue
the DUAL algorithm. In the process, R may receive other queries that elicit its distance-vector report; as
long as R is active it will report its cost to D as C. We omit the argument that this process and thus the
network must eventually converge.

9.5 Link-State Routing-Update Algorithm

Link-state routing is an alternative to distance-vector. It is often though certainly not always considered
to be the routing-update algorithm class of choice for networks that are sufficiently large, such as those of
ISPs.
In distance-vector routing, each node knows a bare minimum of network topology: it knows nothing about
links beyond those to its immediate neighbors. In the link-state approach, each node keeps a maximum
amount of network information: a full map of all nodes and all links. Routes are then computed locally
from this map, using the shortest-path-first algorithm. The existence of this map allows, in theory, the
calculation of different routes for different quality-of-service requirements. The map also allows calculation
of a new route as soon as news of the failure of the existing route arrives; distance-vector protocols on the
other hand must wait for news of a new route after an existing route fails.
Link-state protocols distribute network map information through a modified form of broadcast of the status
of each individual link. Whenever either side of a link notices the link has died (or if a node notices that
a new link has become available), it sends out link-state packets (LSPs) that flood the network. This
broadcast process is called reliable flooding; in general broadcast protocols work poorly with networks that
have even small amounts of topological looping (that is, redundant paths). Link-state protocols must ensure
180

9 Routing-Update Algorithms

An Introduction to Computer Networks, Release 1.0

that every router sees every LSP, and also that no LSPs circulate repeatedly. (The acronym LSP is used by
a link-state implementation known as IS-IS; the preferred acronym used by the Open Shortest Path First
(OSPF) implementation is LSA, where A is for advertisement.) LSPs are sent immediately upon link-state
changes, like triggered updates in distance-vector protocols except there is no race between bad news
and good news.
It is possible for ephemeral routing loops to exist; for example, if one router has received a LSP but another
has not, they may have an inconsistent view of the network and thus route to one another. However, as soon
as the LSP has reached all routers involved, the loop should vanish. There are no race conditions, as with
distance-vector routing, that can lead to persistent routing loops.
The link-state flooding algorithm avoids the usual problems of broadcast in the presence of loops by having
each node keep a database of all LSP messages. The originator of each LSP includes its identity, information
about the link that has changed status, and also a sequence number. Other routers need only keep in their
databases the LSP packet with the largest sequence number; older LSPs can be discarded. When a router
receives a LSP, it first checks its database to see if that LSP is old, or is current but has been received before;
in these cases, no further action is taken. If, however, an LSP arrives with a sequence number not seen
before, then in typical broadcast fashion the LSP is retransmitted over all links except the arrival interface.
As an example, consider the following arrangement of routers:

Suppose the AE link status changes. A sends LSPs to C and B. Both these will forward the LSPs to D;
suppose Bs arrives first. Then D will forward the LSP to C; the LSP traveling CD and the LSP traveling
DC might even cross on the wire. D will ignore the second LSP copy that it receives from C and C will
ignore the second copy it receives from D.
It is important that LSP sequence numbers not wrap around. (Protocols that do allow a numeric field to wrap
around usually have a clear-cut idea of the active range that can be used to conclude that the numbering has
wrapped rather than restarted; this is harder to do in the link-state context.) OSPF uses lollipop sequencenumbering here: sequence numbers begin at -231 and increment to 231 -1. At this point they wrap around
back to 0. Thus, as long as a sequence number is less than zero, it is guaranteed unique; at the same time,
routing will not cease if more than 231 updates are needed. Other link-state implementations use 64-bit
sequence numbers.
Actual link-state implementations often give link-state records a maximum lifetime; entries must be periodically renewed.

9.5.1 Shortest-Path-First Algorithm

The next step is to compute routes from the network map, using the shortest-path-first algorithm. Here is
our example network:

9.5 Link-State Routing-Update Algorithm

181

An Introduction to Computer Networks, Release 1.0

The shortest path from A to D is A-B-C-D, which has cost 3+4+2=9.

We build the set R of all shortest-path routes iteratively. Initially, R contains only the 0-length route to the
start node; one new destination and route is added to R at each stage of the iteration. At each stage we have
a current node, representing the node most recently added to R. The initial current node is our starting
node, in this case, A.
We will also maintain a set T, for tentative, of routes to other destinations. This is also initialized to empty.
At each stage, we find all nodes which are immediate neighbors of the current node and which do not
already have routes in the set R. For each such node N, we calculate the length of the route from the start
node to N that goes through the current node. We see if this is our first route to N, or if the route improves
on any route to N already in T; if so, we add or update the route in T accordingly.
At the end of this process, we choose the shortest path in T, and move the route and destination node to R.
The destination node of this shortest path becomes the next current node. Ties can be resolved arbitrarily,
but note that, as with distance-vector routing, we must choose the minimum or else the accurate-costs
property will fail.
We repeat this process until all nodes have routes in the set R.
For the example above, we start with current = A and R = {A,A,0}. The set T will be {B,B,3, C,C,10,
D,D,11}. The shortest entry is B,B,3, so we move that to R and continue with current = B.
For the next stage, the neighbors of B without routes in R are C and D; the routes from A to these through B
are C,B,7 and D,B,12. The former is an improvement on the existing T entry C,C,10 and so replaces
it; the latter is not an improvement over D,D,11. T is now {C,B,7, D,D,11}. The shortest route in T is
that to C, so we move this node and route to R and set C to be current.
For the next stage, D is the only non-R neighbor; the path from A to D via C has entry D,B,9, an improvement over the existing D,D,11 in T. The only entry in T is now D,B,9; this is the shortest and thus we
move it to R.
We now have routes in R to all nodes, and are done.
Here is another example, again with links labeled with costs:

182

9 Routing-Update Algorithms

An Introduction to Computer Networks, Release 1.0

We start with current = A. At the end of the first stage, B,B,3 is moved into R, T is {D,D,12}, and
current is B. The second stage adds C,B,5 to T, and then moves this to R; current then becomes C. The
third stage introduces the route (from A) D,B,10; this is an improvement over D,D,12 and so replaces it
in T; at the end of the stage this route to D is moved to R.
A link-state source node S computes the entire path to a destination D. But as far as the actual path that a
packet sent by S will take to D, S has direct control only as far as the first hop N. While the accurate-cost
rule we considered in distance-vector routing will still hold, the actual path taken by the packet may differ
from the path computed at the source, in the presence of alternative paths of the same length. For example,
S may calculate a path SNAD, and yet a packet may take path SNBD, so long as the NAD and
NBD paths have the same length.
Link-state routing allows calculation of routes on demand (results are then cached), or larger-scale calculation. Link-state also allows routes calculated with quality-of-service taken into account, via straightforward
extension of the algorithm above.

9.6 Routing on Other Attributes

There is sometimes a desire to route on attributes other than the destination, or the destination and QoS
bits. This kind of routing is decidedly nonstandard, and is generally limited to the imposition of very local
routing policies. It is, however, often available.
On linux systems this is part of the Linux Advanced Routing facility, often grouped with some advanced
queuing features known as Traffic Control; the combination is referred to as LARTC.
As a simple example of what can be done, suppose a site has two links L1 and L2 to the Internet, where L1
is faster and L2 serves more as a backup; in such a configuration the default route to the Internet will often
be via L1. A site may wish to route some outbound traffic via L2, however, for any of the following reasons:
the traffic may involve protocols deemed lower in priority (eg email)
the traffic may be real-time traffic that can benefit from reduced competition on L2
the traffic may come from lower-priority senders; eg some customers within the site may be relegated
to using L2 because they are paying less
In the first two cases, routing might be based on the destination port numbers; in the third, it might be based
on the source IP address.
Note that nothing can be done in the inbound direction unless L1 and L2 lead to the same ISP, and even
there any special routing would be at the discretion of that ISP.
The trick with LARTC is to be compatible with existing routing-update protocols; this would be a problem if
the kernel forwarding table simply added columns for other packet attributes that neighboring non-LARTC
9.6 Routing on Other Attributes

183

An Introduction to Computer Networks, Release 1.0

routers knew nothing about. Instead, the forwarding table is split up into multiple dest, next_hop (or dest,
QoS, next_hop) tables. One of these tables is the main table, and is the table that is updated by routingupdate protocols interacting with neighbors. Before a packet is forwarded, administratively supplied rules
are consulted to determine which table to apply; these rules are allowed to consult other packet attributes.
The collection of tables and rules is known as the routing policy database.
As a simple example, in the situation above the main table would have an entry default, L1 (more precisely,
it would have the IP address of the far end of the L1 link instead of L1 itself). There would also be another
table, perhaps named slow, with a single entry default, L2. If a rule is created to have a packet routed
using the slow table, then that packet will be forwarded via L2. Here is one such linux rule, applying to
traffic from host [Link]:
ip rule add from [Link] table slow

Now suppose we want to route traffic to port 25 (the SMTP port) via L2. This is harder; linux provides no
support here for routing based on port numbers. However, we can instead use the iptables mechanism to
mark all packets destined for port 25, and then create a routing-policy rule to have such marked traffic
use the slow table. The mark is known as the forwarding mark, or fwmark; its value is 0 by default. The
fwmark is not actually part of the packet; it is associated with the packet only while the latter remains
within the kernel.
iptables --table mangle --append PREROUTING \\
--protocol tcp --source-port 25 --jump MARK --set-mark 1
ip rule add fwmark 1 table slow

Consult the applicable man pages for further details.

The iptables mechanism can also be used to set the IP precedence bits, so that a single standard IP forwarding
table can be used, though support for the IP precedence bits is limited.

9.7 Epilog
At this point we have concluded the basics of IP routing, involving routing within large (relatively) homogeneous organizations such as multi-site corporations or Internet Service Providers. Every router involved
must agree to run the same protocol, and must agree to a uniform assignment of link costs.
At the very largest scales, these requirements are impractical. The next chapter is devoted to this issue of
very-large-scale IP routing, eg on the global Internet.

9.8 Exercises
1. Suppose the network is as follows, where distance-vector routing update is used. Each link has cost 1,
and each router has entries in its forwarding table only for its immediate neighbors (so As table contains
B,B,1, D,D,1 and Bs table contains A,A,1, C,C,1).
A

184

9 Routing-Update Algorithms

An Introduction to Computer Networks, Release 1.0

(a). Suppose each node creates a report from its initial configuration and sends that to each of its neighbors.
What will each nodes forwarding table be after this set of exchanges? The exchanges, in other words, are
all conducted simultaneously; each node first sends out its own report and then processes the reports
arriving from its two neighbors.
(b). What will each nodes table be after the simultaneous-and-parallel exchange process of part (a) is
repeated a second time?
Hint: you do not have to go through each exchange in detail; the only information added by an exchange is
additional reachability information.

2. Now suppose the configuration of routers has the link weights shown below.
A

2
D

C
12

(a). As in the previous exercise, give each nodes forwarding table after each node exchanges with its
immediate neighbors simultaneously and in parallel.
(b). How many iterations of such parallel exchanges will it take before C learns to reach F via B; that is,
before it creates the entry F,B,11? Count the answer to part (a) as the first iteration.

3. Suppose a router R has the following distance-vector table:

destination
A
B
C
D
E

cost
5
6
7
8
9

next hop
R1
R1
R2
R2
R3

R now receives the following report from R1; the cost of the RR1 link is 1.
destination
A
B
C
D
E
F

9.8 Exercises

cost
4
7
7
6
8
8

185

An Introduction to Computer Networks, Release 1.0

Give Rs updated table after it processes R1s report. For each entry that changes, give a brief explanation,
in the style of 9.1.5 Example 4.
3.5. At the start of Example 3 (9.1.4 Example 3), we changed Cs routing table so that it reached D via A
instead of via E: Cs entry D,E,2 was changed to D,A,2. This meant that C had a valid route to D at the
start.
How might the scenario of Example 3 play out if Cs table had not been altered? Give a sequence of reports
that leads to correct routing between D and E.
4. In the following exercise, A-D are routers and the attached networks N1-N6, which are the ultimate
destinations, are shown explicitly. Routers still exchange distance-vector reports with neighboring routers,
as usual. If a router has a direct connection to a network, you may report the next_hop as direct, eg, from
As table, N1,direct,0
A

N1

N3
N6

N5

N4
N2

(a). Give the initial tables for A through D, before any distance-vector exchanges.
(b). Give the tables after each router A-D exchanges with its immediate neighbors simultaneously and in
parallel.
(c). At the end of (b), what networks are not known by what routers?

5. Suppose A, B, C, D and E are connected as follows. Each link has cost 1, and so each forwarding table
is uniquely determined; Bs table is A,A,1, C,C,1, D,A,2, E,C,2. Distance-vector routing update is
used.

Now suppose the DE link fails, and so D updates its entry for E to E,-,.

(a). Give As table after D reports E, to A

(b). Give Bs table after A reports to B
(c). Give As table after B reports to A; note that B has an entry E,C,2
(d). Give Ds table after A reports to D.

186

9 Routing-Update Algorithms

An Introduction to Computer Networks, Release 1.0

6. Consider the network in [Link] Split Horizon:, using distance-vector routing updates.

D,A,2

D,A,2

Suppose the DA link breaks and then

A reports D, to B (as before)
C reports D,2 to B (as before)
A now reports D, to C (instead of B reporting D,3 to A)

(a). What reports (a pair should suffice) will lead to the formation of the routing loop?
(b). What (single) report will eliminate the possibility of the routing loop?

7. Suppose the network of 9.2 Distance-Vector Slow-Convergence Problem is changed to the following.
Distance-vector update is used; again, the AD link breaks.
D,D,1
A

D,E,2
B

(a). Explain why Bs report back to A, after A reports D,-,, is now valid.
(b). Explain why hold down ([Link] Hold Down) will delay the use of the new route ABED.

8. Suppose the routers are A, B, C, D, E and F, and all link costs are 1. The distance-vector forwarding
tables for A and F are below. Give the network with the fewest links that is consistent with these tables.
Hint: any destination reached at cost 1 is directly connected; if X reaches Y via Z at cost 2, then Z and Y
must be directly connected.
As table
dest
B
C
D
E
F

cost
1
1
2
2
3

next_hop
B
C
C
C
B

Fs table

9.8 Exercises

187

An Introduction to Computer Networks, Release 1.0

dest
A
B
C
D
E

cost
3
2
2
1
1

next_hop
E
D
D
D
E

9. (a) Suppose routers A and B somehow end up with respective forwarding-table entries D,B,n and
D,A,m, thus creating a routing loop. Explain why the loop may be removed more quickly if A and B both
use poison reverse with split horizon, versus if A and B use split horizon only.
(b). Suppose the network looks like the following. The AB link is extremely slow.
A
C

Suppose A and B send reports to each other advertising their routes to D, and immediately afterwards the
CD link breaks and C reports to A and B that D is unreachable. After those unreachability reports are
processed, A and Bs reports sent to each other before the break finally arrive. Explain why the network is
now in the state described in part (a).
10. Suppose the distance-vector algorithm is run on a network and no links break (so by the last paragraph
of 9.1.1 Distance-Vector Update Rules the next_hop-increase rule is never applied).

(a). Prove that whenever A is Bs next_hop to destination D, then As cost to D is strictly less than Bs.
Hint: assume that if this claim is true, then it remains true after any application of the rules in
9.1.1 Distance-Vector Update Rules. If the lower-cost rule is applied to B after receiving a report from A,
resulting in a change to Bs cost to D, then one needs to show As cost is less than Bs, and also Bs new
cost is less than that of any neighbor C that uses B as its next_hop to D.
(b). Use (a) to prove that no routing loops ever form.

11. Give a scenario illustrating how a (very temporary!) routing loop might form in link-state routing.
12. Use the Shortest Path First algorithm to find the shortest path from A to E in the network below. Show
the sets R and T, and the node current, after each step.

13. Suppose you take a laptop, plug it into an Ethernet LAN, and connect to the same LAN via Wi-Fi. From
laptop to LAN there are now two routes. Which route will be preferred? How can you tell which way traffic
is flowing? How can you configure your OS to prefer one path or another?

188

9 Routing-Update Algorithms

10 LARGE-SCALE IP ROUTING

In the previous chapter we considered two classes of routing-update algorithms: distance-vector and linkstate. Each of these approaches requires that participating routers have agreed not just to a common protocol,
but also to a common understanding of how link costs are to be assigned. We will address this further below
in 10.6 Border Gateway Protocol, BGP, but the basic problem is that if one site prefers the hop-count
approach, assigning every link a cost of 1, while another site prefers to assign link costs in proportion to
their bandwidth, then path cost comparisons between the two sites simply cannot be done. In general, we
cannot even translate costs from one site to another, because the paths themselves depend on the cost
assignment strategy.
The term routing domain is used to refer to a set of routers under common administration, using a common
link-cost assignment. Another term for this is autonomous system. While use of a common routingupdate protocol within the routing domain is not an absolute requirement for example, some subnets may
internally use distance-vector while the sites backbone routers use link-state we can assume that all
routers have a uniform view of the sites topology and cost metrics.
One of the things included in the term large-scale IP routing is the coordination of routing between multiple routing domains. Even in the earliest Internet there were multiple routing domains, if for no other
reason than that how to measure link costs was (and still is) too unsettled to set in stone. However, another
component of large-scale routing is support for hierarchical routing, above the level of subnets; we turn to
this next.

10.1 Classless Internet Domain Routing: CIDR

CIDR is the mechanism for supporting hierarchical routing in the Internet backbone. Subnetting moves the
network/host division line further rightwards; CIDR allows moving it to the left as well. With subnetting,
the revised division line is visible only within the organization that owns the IP network address; subnetting
is not visible outside. CIDR allows aggregation of IP address blocks in a way that is visible to the Internet
backbone.
When CIDR was introduced in 1993, the following were some of the justifications for it, all relating to the
increasing size of the backbone IP forwarding tables, and expressed in terms of the then-current Class A/B/C
mechanism:
The Internet is running out of Class B addresses (this happened in the mid-1990s)
There are too many Class Cs (the most numerous) for backbone forwarding tables to be efficient
Eventually IANA (the Internet Assigned Numbers Authority) will run out of IP addresses (this happened in 2011)
Assigning non-CIDRed multiple Class Cs in lieu of a single Class B would have helped with the first point
in the list above, but made the second point worse.
Ironically, the current (2013) very tight market for IP address blocks is likely to lead to larger and larger
backbone IP forwarding tables, as sites are forced to use multiple small address blocks instead of one large
block.
189

An Introduction to Computer Networks, Release 1.0

By the year 2000, CIDR had essentially eliminated the Class A/B/C mechanism from the backbone Internet,
and had more-or-less completely changed how backbone routing worked. You purchased an address block
from a provider or some other IP address allocator, and it could be whatever size you needed, from /27 to
/15.
What CIDR enabled is IP routing based on an address prefix of any length; the Class A/B/C mechanism of
course used fixed prefix lengths of 8, 16 and 24 bits. Furthermore, CIDR allows different routers, at different
levels of the backbone, to route on prefixes of different lengths.
CIDR was formally introduced by RFC 1518 and RFC 1519. For a while there were strategies in place to
support compatibility with non-CIDR-aware routers; these are now obsolete. In particular, it is no longer
appropriate for large-scale routers to fall back on the Class A/B/C mechanism in the absence of CIDR
information; if the latter is missing, the routing should fail.
The basic strategy of CIDR is to consolidate multiple networks going to the same destination into a single
entry. Suppose a router has four class Cs all to the same destination:
[Link]/24 foo
[Link]/24 foo
[Link]/24 foo
[Link]/24 foo
The router can replace all these with the single entry
[Link]/22 foo
It does not matter here if foo represents a single ultimate destination or if it represents four sites that just
happen to be routed to the same next_hop.
It is worth looking closely at the arithmetic to see why the single entry uses /22. This means that the first 22
bits must match [Link]; this is all of the first and second bytes and the first six bits of the third byte. Let
us look at the third byte of the network addresses above in binary:
200.7.000000 00.0/24 foo
200.7.000000 01.0/24 foo
200.7.000000 10.0/24 foo
200.7.000000 11.0/24 foo
The /24 means that the network addresses stop at the end of the third byte. The four entries above cover
every possible combination of the last two bits of the third byte; for an address to match one of the entries
above it suffices to begin 200.7 and then to have 0-bits as the first six bits of the third byte. This is another
way of saying the address must match [Link]/22.
Most implementations actually use a bitmask, eg [Link].00 (in hex) rather than the number 22; note 0xFC
= 1111 1100 with 6 leading 1-bits, so [Link].00 has 8+8+6=22 1-bits followed by 10 0-bits.
The IP delivery algorithm of 7.5 The Classless IP Delivery Algorithm still works with CIDR, with the
understanding that the routers forwarding table can now have a network-prefix length associated with any
entry. Given a destination D, we search the forwarding table for network-prefix destinations B/k until we
find a match; that is, equality of the first k bits. In terms of masks, given a destination D and a list of table
entries prefix,mask = B[i],M[i], we search for i such that (D & M[i]) = B[i].
It is possible to have multiple matches, and responsibility for avoiding this is much too distributed to be

190

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

declared illegal by IETF mandate. Instead, CIDR introduced the longest-match rule: if destination D
matches both B1 /k1 and B2 /k2 , with k1 < k2 , then the longer match B2 /k2 match is to be used. (Note that if
D matches two distinct entries B1 /k1 and B2 /k2 then either k1 < k2 or k2 < k1 ).

10.2 Hierarchical Routing

Strictly speaking, CIDR is simply a mechanism for routing to IP address blocks of any prefix length; that is,
for setting the network/host division point to an arbitrary place within the 32-bit IP address.
However, by making this network/host division point variable, CIDR introduced support for routing on
different prefix lengths at different places in the backbone routing infrastructure. For example, top-level
routers might route on /8 or /9 prefixes, while intermediate routers might route based on prefixes of length
14. This feature of routing on fewer bits at one point in the Internet and more bits at another point is exactly
what is meant by hierarchical routing.
We earlier saw hierarchical routing in the context of subnets: traffic might first be routed to a class-B site
[Link]/16, and then, within that site, to subnets such as [Link]/24, [Link]/24, etc. But with
CIDR the hierarchy can be much more flexible: the top level of the hierarchy can be much larger than the
customer level, lower levels need not be administratively controlled by the higher levels (as is the case
with subnets), and more than two levels can be used.
CIDR is an address-block-allocation mechanism; it does not directly speak to the kinds of policy we might
wish to implement with it. Here are three possible applications; the latter two involve hierarchical routing:
Application 1 (legacy): CIDR allows IANA to allocate multiple blocks of Class C, or fragments of a
Class A, to a single customer, so as to require only a single forwarding-table entry
Application 2 (legacy): CIDR allows opportunistic aggregation of routes: a router that sees the four
200.7.x.0/24 routes above in its table may consolidate them into a single entry.
Application 3 (current): CIDR allows huge provider blocks, with suballocation by the provider. This
is known as provider-based routing.
Application 4 (hypothetical): CIDR allows huge regional blocks, with suballocation within the region, somewhat like the original scheme for US phone numbers with area codes. This is known as
geographical routing.
Hierarchical routing does introduce one new wrinkle: the routes chosen may no longer be globally optimal.
Suppose, for example, we are trying to route to prefix [Link]/16. Suppose at the top level that traffic
to [Link]/8 is routed to router R1, and R1 then routes traffic to [Link]/16 to R2. A packet sent
to [Link] by an independent router R3 would always pass through R1, even if there were a shorter
path R3R4R2 that bypassed R1. This is addressed in further detail below in 10.4.3 Provider-Based
Hierarchical Routing.
We can avoid this by declaring that CIDR will only be used when it turns out that, in the language of the
example of the preceding paragraph, the best path to reach [Link]/16 is always through R1. But this is
seldom the case, and is even less likely to remain the case as new links are added. Such a policy also defeats
much of the potential benefit of CIDR at reducing router forwarding-table size by supporting the creation
of arbitrary-sized address blocks and then routing to them as a single unit. The Internet backbone might
be much happier if all routers simply had to maintain a single entry [Link]/8, R1, versus 256 entries
200.x.0.0/16, R1 for all but one value of x.
10.2 Hierarchical Routing

191

An Introduction to Computer Networks, Release 1.0

10.3 Legacy Routing

Back in the days of NSFNet, the Internet backbone was a single routing domain. While most customers did
not connect directly to the backbone, the intervening providers tended to be relatively compact, geographically that is, regional and often had a single primary routing-exchange point with the backbone. IP
addresses were allocated to subscribers directly by the IANA, and the backbone forwarding tables contained
entries for every site, even the Class Cs.
Because the NSFNet backbone and the regional providers did not necessarily share link-cost information,
routes were even at this early point not necessarily globally optimal; compromises and approximations were
made. However, in the NSFNet model routers generally did find a reasonable approximation to the shortest
path to each site referenced by the backbone tables. While the legacy backbone routing domain was not
all-encompassing, if there were differences between two routes, at least the backbone portions the longest
components would be identical.

10.4 Provider-Based Routing

In provider-based routing, large CIDR blocks are allocated to large-scale providers. The different providers
each know how to route to one another. Subscribers (usually) obtain their IP addresses from within their
providers blocks; thus, traffic from the outside is routed first to the provider, and then, within the providers
routing domain, to the subscriber. We may even have a hierarchy of providers, so packets would be routed
first to the large-scale provider, and eventually to the local provider. There may no longer be a central
backbone; instead, multiple providers may each build parallel transcontinental networks.
Here is a simpler example, in which providers have unique paths to one another. Suppose we have providers
P0, P1 and P2, with customers as follows:
P0: customers A,B,C
P1: customers D,E
P2: customers F,G
We will also assume that each provider has an IP address block as follows:
P0: [Link]/8
P1: [Link]/8
P2: [Link]/8
Let us now allocate addresses to the customers:
A: [Link]/16
B: [Link]/16
C: [Link]/20 (16 = 0001 0000)

D: [Link]/16
E: [Link]/16

192

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

F: [Link]/16
G: [Link]/16

The routing model is that packets are first routed to the appropriate provider, and then to the customer.
While this model may not in general guarantee the shortest end-to-end path, it does in this case because
each provider has a single point of interconnection to the others. Here is the network diagram:

With this diagram, P0s forwarding table looks something like this:
destination
[Link]/16
[Link]/16
[Link]/20
[Link]/8
[Link]/8

next_hop
A
B
C
P1
P2

That is, P0s table consists of

one entry for each of P0s own customers
one entry for each other provider
If we had 1,000,000 customers divided equally among 100 providers, then each providers table would have
only 10,099 entries: 10,000 for its own customers and 99 for the other providers. Without CIDR, each
providers forwarding table would have 1,000,000 entries.
Even if we have some additional secondary links, that is, additional links that do not create alternative paths between providers, the routing remains relatively straightforward. Shown here are the private
customer-to-customer links CD and EF; these are likely used only by the customers they connect. Two
customers are multi-homed; that is, they have connections to alternative providers: AP1 and EP2. Typically, though, while A and E may use these secondary links all they want for outbound traffic, their respective
inbound traffic would still go through their primary providers P0 and P1 respectively.

10.4 Provider-Based Routing

193

An Introduction to Computer Networks, Release 1.0

10.4.1 Internet Exchange Points

The long links joining providers in these diagrams are somewhat misleading; providers do not always like
sharing long links and the attendant problems of sharing responsibility for failures. Instead, providers often connect to one another at Internet eXchange Points or IXPs; the link P0
P1 might actually be
P0
IXP P1, where P0 owns the left-hand link and P1 the right-hand. IXPs can either be third-party
sites open to all providers, or private exchange points. The term Metropolitan Area Exchange, or MAE,
appears in the names of the IXPs MAE-East, originally near Washington DC, and MAE-West, originally in
San Jose, California; each of these is now actually a set of IXPs. MAE in this context is now a trademark.

10.4.2 CIDR and Staying Out of Jail

Suppose we want to change providers. One way we can do this is to accept a new IP-address block from the
new provider, and change all our IP addresses. The paper Renumbering: Threat or Menace [LKCT96] was
frequently cited at least in the early days of CIDR as an intimation that such renumbering was inevitably
a Bad Thing. In principle, therefore, we would like to allow at least the option of keeping our IP address
allocation while changing providers.
An address-allocation standard that did not allow changing of providers might even be a violation of the
US Sherman Antitrust Act; see American Society of Mechanical Engineers v Hydrolevel Corporation, 456
US 556 (1982). The IETF thus had the added incentive of wanting to stay out of jail, when writing the
CIDR standard so as to allow portability between providers (actually, antitrust violations usually involve
civil penalties).
The CIDR longest-match rule turns out to be exactly what we (and the IETF) need. Suppose, in the diagrams
above, that customer C wants to move from P0 to P1, and does not want to renumber. What routing changes
need to be made? One solution is for P0 to add a route [Link]/20, P1 that routes all of Cs traffic to
P1; P1 will then forward that traffic on to C. P1s table will be as follows, and P1 will use the longest-match
rule to distinguish traffic for its new customer C from traffic bound for P0.
destination
[Link]/8
[Link]/8
[Link]/16
[Link]/16
[Link]/20

194

next_hop
P0
P2
D
E
C

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

This does work, but all Cs inbound traffic except for that originating in P1 will now be routed through
Cs ex-provider P0, which as an ex-provider may not be on the best of terms with C. Also, the routing is
inefficient: Cs traffic from P2 is routed P2P0P1 instead of the more direct P2P1.
A better solution is for all providers other than P1 to add the route [Link]/20, P1. While traffic to
[Link]/8 otherwise goes to P0, this particular sub-block is instead routed by each provider to P1. The
important case here is P2, as a stand-in for all other providers and their routers: P2 routes [Link]/8 traffic
to P0 except for the block [Link]/20, which goes to P1.
Having every other provider in the world need to add an entry for C is going to cost some money, and, one
way or another, C will be the one to pay. But at least there is a choice: C can consent to renumbering (which
is not difficult if they have been diligent in using DHCP and perhaps NAT too), or they can pay to keep their
old address block.
As for the second diagram above, with the various private links (shown as dashed lines), it is likely that the
longest-match rule is not needed for these links to work. As private link to P1 might only mean that
A can send outbound traffic via P1
P1 forwards As traffic to A via the private link
P2, in other words, is still free to route to A via P0. P1 may not advertise its route to A to anyone else.

10.4.3 Provider-Based Hierarchical Routing

With provider-based routing, the route taken may no longer be end-to-end optimal; we have replaced the
problem of finding an optimal route from A to B with the two problems of finding an optimal route from
A to Bs provider, and then from that provider entry point to B; the second strategy may not yield the same
result. This second strategy mirrors the two-step routing process of first routing on the address bits that
identify the provider, and then routing on the address bits including the subscriber portion.
Consider the following example, in which providers P1 and P2 have three interconnection links (r1s1,
r2s2 and r3s3), each with cost 1. We assume that P1s costs happen to be comparable with P2s costs.

The globally shortest path between A and B is via the r2s2 crossover, with total length 6+1+5=12. However,
traffic from A to B will be routed by P1 to its closest crossover to P2, namely the r3s3 link. The total path is
2+1+8+5=16. Traffic from B to A will be routed by P2 via the r1s1 crossover, for a length of 2+1+7+6=16.

10.4 Provider-Based Routing

195

An Introduction to Computer Networks, Release 1.0

This routing strategy is sometimes called hot-potato routing; each provider tries to get rid of any traffic (the
potatoes) as quickly as possible, by routing to the closest exit point.
Not only are the paths taken inefficient, but the AB and BA paths are now asymmetric. This can be
a problem if forward and reverse timings are critical, or if one of P1 or P2 has significantly more bandwidth
or less congestion than the other. In practice, however, route asymmetry is of little consequence.
As for the route inefficiency itself, this also is not necessarily a significant problem; the primary reason
routing-update algorithms focus on the shortest path is to guarantee that all computed paths are loop-free.
As long as each half of a path is loop-free, and the halves do not intersect except at their common midpoint,
these paths too will be loop-free.
The BGP MED value ([Link] MULTI_EXIT_DISC) offers an optional mechanism for P1 to agree that
AB traffic should take the r1s1 crossover. This might be desired if P1s network were better and
customer A was willing to pay extra to keep its traffic within P1s network as long as possible.

10.5 Geographical Routing

The classical alternative to provider-based routing is geographical routing; the archetypal model for this is
the telephone area code system. A call from anywhere in the US to Loyola Universitys main switchboard,
773-274-3000, would traditionally be routed first to the 773 area code in Chicago. From there the call would
be routed to the north-side 274 exchange, and from there to subscriber 3000. A similar strategy can be used
for IP routing.
Geographical addressing has some advantages. Figuring out a good route to a destination is usually straightforward, and close to optimal in terms of the path physical distance. Changing providers never involves
renumbering (though moving may). And approximate IP address geolocation (determining a hosts location
from its IP address) is automatic.
Geographical routing has some minor technical problems. First, routing may be inefficient between immediate neighbors A and B that happen to be split by a boundary for larger geographical areas; the path might
go from A to the center of As region to the center of Bs region and then to B. Another problem is that some
larger sites (eg large corporations) are themselves geographically distributed; if efficiency is the goal, each
office of such a site would need a separate IP address block appropriate for its physical location.
But the real issue with geographical routing is apparently the business question of who carries the traffic.
The provider-based model has a very natural answer to this: every link is owned by a specific provider. For
geographical IP routing, my local provider might know at once from the prefix that a packet of mine is to be
delivered from Chicago to San Francisco, but who will carry it there? My provider might have to enter into
different traffic contracts for multiple different regions. If different local providers make different arrangements for long-haul packet delivery, the routing efficiency (at least in terms of table size) of geographical
routing is likely lost. Finally, there is no natural answer for who should own those long inter-region links. It
may be useful to recall that the present area-code system was created when the US telephone system was an
AT&T monopoly, and the question of who carried traffic did not exist.
That said, the top five Regional Internet Registries represent geographical regions (usually continents), and
provider-based addressing is below that level. That is, the IANA handed out address blocks to the geographical RIRs, and the RIRs then allocated address blocks to providers.
At the intercontinental level, geography does matter: some physical link paths are genuinely more expensive

196

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

than other (shorter) paths. It is much easier to string terrestrial cable than undersea cable. However, within
a continent physical distance does not always matter as much as might be supposed. Furthermore, a large
geographically spread-out provider can always divide up its address blocks by region, allowing internal
geographical routing to the correct region.
Here is a diagram of IP address allocation as of 2006: [Link]

10.6 Border Gateway Protocol, BGP

In 9 Routing-Update Algorithms, we considered interior routing-update protocols. For both DistanceVector and Link State methods, the per-link cost played an essential role: by trying to minimize the cost, we
were assured that no routing loops would be present in a stable network.
But when systems under different administration (eg two large ISPs) talk to each other, comparing metrics
simply does not work. Each sides metrics may be based on any of the following:
hopcount
bandwidth
cost
congestion
One providers metric may even use larger numbers for better routes (though if this is done then total path
costs, or preference, cannot be obtained simply by adding the per-link values). Any attempt at comparison
of values from the different sides is a comparison of apples and oranges.
The Border Gateway Protocol, or BGP, is assigned the job of handling exchange of routing information
between neighboring independent organizations; this is sometimes called exterior routing. The current
version is BGP-4, documented in RFC 4271. The BGP term for a routing domain under coordinated administration, and using one consistent link-cost metric throughout, is Autonomous System, or AS. That said,
all that is strictly required is that all BGP routers within an AS have the same consistent view of routing, and
in fact some Autonomous Systems do run multiple routing protocols and may even use different metrics at
different points. As indicated above, BGP does not support the exchange of link-cost information between
Autonomous Systems.
BGP also has a second goal, in addition to the purely technical problem of finding routes in the absence of
cost information: BGP also provides support for policy-based routing; that is, for making routing decisions
based on managerial or administrative input (perhaps regarding who is paying what for the traffic carried).
Every non-leaf site (and some large leaf sites) has one or more BGP speakers: the routers that run BGP.
If there is more than one, they must remain coordinated with one another so as to present a consistent view
of the sites connections and advertisements; this coordination process is sometimes called internal BGP
to distinguish it from the communication with neighboring Autonomous Systems. The latter process is then
known as external BGP.
The BGP speakers of a site are often not the busy border routers that connect directly to the neighboring AS,
though they are usually located near them and are often on the same subnet. Each interconnection point with
a neighboring AS generally needs its own BGP speaker. Connections between BGP speakers of neighboring
Autonomous Systems sometimes called BGP peers are generally configured administratively; they are
not subject to a neighbor discovery process like that used by most interior routers.
10.6 Border Gateway Protocol, BGP

197

An Introduction to Computer Networks, Release 1.0

The BGP speakers must maintain a database of all routes received, not just of the routes actually used.
However, the speakers exchange with neighbors only the routes they (and thus their AS) use themselves;
this is a firm BGP rule.
The current BGP standard is RFC 4271.

10.6.1 AS-paths
At its most basic level, BGP involves the exchange of lists of reachable destinations, like distance-vector
routing without the distance information. But that strategy, alone, cannot avoid routing loops. BGP solves
the loop problem by having routers exchange, not just destination information, but also the entire path used
to reach each destination. Paths including each router would be too cumbersome; instead, BGP abbreviates
the path to the list of ASs traversed; this is called the AS-path. This allows routers to make sure their routes
do not traverse any AS more than once, and thus do not have loops.
As an example of this, consider the network below, in which we consider Autonomous Systems also to be
destinations. Initially, we will assume that each AS discovers its immediate neighbors. AS3 and AS5 will
then each advertise to AS4 their routes to AS2, but AS4 will have no reason at this level to prefer one route
to the other (BGP does use the shortest AS-path as part of its tie-breaking rule, but, before falling back on
that rule, AS4 is likely to have a commercial preference for which of AS3 and AS5 it uses to reach AS2).

Also, AS2 will advertise to AS3 its route to reach AS1; that advertisement will contain the AS-path
AS2,AS1. Similarly, AS3 will advertise this route to AS4 and then AS4 will advertise it to AS5. When
AS5 in turn advertises this AS1-route to AS2, it will include the entire AS-path AS5,AS4,AS3,AS2,AS1,
and AS2 would know not to use this route because it would see that it is a member of the AS-path. Thus,
BGP is spared the kind of slow-convergence problem that traditional distance-vector approaches were subject to.
It is theoretically possible that the shortest path (in the sense, say, of the hopcount metric) from one host to
another traverses some AS twice. If so, BGP will not allow this route.
AS-paths potentially add considerably to the size of the AS database. The number of paths a site must keep
track of is proportional to the number of ASs, because there will be one AS-path to each destination AS.
(Actually, an AS may have to record many times that many AS-paths, as an AS may hear of AS-paths that
it elects not to use.) Typically there are several thousand ASs in the world. Let A be the number of ASs.
Typically the average length of an AS-path is about log(A), although this depends on connectivity. The
amount of memory required by BGP is
CAlog(A) + KN,
where C and K are constants.
The other major goal of BGP is to allow some degree of administrative input to what, for interior routing,

198

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

is largely a technical calculation (though an interior-routing administrator can set link costs). BGP is the
interface between large ISPs, and can be used to implement contractual agreements made regarding which
ISPs will carry other ISPs traffic. If ISP2 tells ISP1 it has a good route to destination D, but ISP1 chooses
not to send traffic to ISP2, BGP can be used to implement this.
Despite the exchange of AS-path information, temporary routing loops may still exist. This is because BGP
may first decide to use a route and only then export the new AS-path; the AS on the other side may realize
there is a problem as soon as the AS-path is received but by then the loop will have at least briefly been in
existence. See the first example below in 10.6.8 Examples of BGP Instability.
BGPs predecessor was EGP, which guaranteed loop-free routes by allowing only a single route to any AS,
thus forcing the Internet into a tree topology, at least at the level of Autonomous Systems. The AS graph
could contain no cycles or alternative routes, and hence there could be no redundancy provided by alternative
paths. EGP also thus avoided having to make decisions as to the preferred path; there was never more than
one choice. EGP was sometimes described as a reachability protocol; its only concern was whether a given
network was reachable.

10.6.2 AS-Paths and Route Aggregation

There is some conflict between the goal of reporting precise AS-paths to each destination, and of consolidating as many address prefixes as possible into a single prefix (single CIDR block). Consider the following
network:

Suppose AS2 has paths

path=AS2, destination 200.0.0/23
path=AS2,AS3, destination 200.0.2/24
path=AS2,AS4, destination 200.0.3/24
If AS2 wants to optimize address-block aggregation using CIDR, it may prefer to aggregate the three destinations into the single block 200.0.0/22. In this case there would be two options for how AS2 reports its
routes to AS1:
Option 1: report 200.0.0/22 with path AS2. But this ignores the ASs AS3 and AS4! These
are legitimately part of the AS-paths to some of the destinations within the block 200.0.0/22; loop
detection could conceivably now fail.
Option 2: report 200.0.0/22 with path AS2,AS3,AS4, which is not a real path but which does
include all the ASs involved. This ensures that the loop-detection algorithm works, but artificially
inflates the length of the AS-path, which is used for certain tie-breaking decisions.
As neither of these options is desirable, the concept of the AS-set was introduced. A list of Autonomous
Systems traversed in order now becomes an AS-sequence. In the example above, AS2 can thus report net
200.0.0/22 with
10.6 Border Gateway Protocol, BGP

199

An Introduction to Computer Networks, Release 1.0

AS-sequence=AS2
AS-set={AS3,AS4}
AS2 thus both achieves the desired aggregation and also accurately reports the AS-path length.
The AS-path can in general be an arbitrary list of AS-sequence and AS-set parts, but in cases of simple
aggregation such as the example here, there will be one AS-sequence followed by one AS-set.
RFC 6472 now recommends against using AS-sets entirely, and recommends that aggregation as above be
avoided.

10.6.3 Transit Traffic

It is helpful to distinguish between two kinds of traffic, as seen from a given AS. Local traffic is traffic that
either originates or terminates at that AS; this is traffic that belongs to that AS. At leaf sites (that is, sites
that connect only to their ISP and not to other sites), all traffic is local.
The other kind of traffic is transit traffic; the AS is forwarding it along on behalf of some nonlocal party.
For ISPs, most traffic is transit traffic. A large almost-leaf site might also carry a small amount of transit
traffic for one particular related (but autonomous!) organization.
The decision as to whether to carry transit traffic is a classic example of an administrative choice, implemented by BGPs support for policy-based routing.

10.6.4 BGP Filtering and Policy Routing

As stated above, one of the goals of BGP is to support policy routing; that is, routing based on managerial
or administrative concerns in addition to technical ones. A BGP speaker may be aware of multiple routes to
a destination. To choose the one route that we will use, it may combine a mixture of optimization rules and
policy rules. Some examples of policy rules might be:
do not use AS13 as we have an adversarial relationship with them
do not allow transit traffic
BGP implements policy through filtering rules that is, rules that allow rejection of certain routes at three
different stages:
1. Import filtering is applied to the lists of routes a BGP speaker receives from its neighbors.
2. Best-path selection is then applied as that BGP speaker chooses which of the routes accepted by the
first step it will actually use.
3. Export filtering is done to decide what routes from the previous step a BGP speaker will actually
advertise. A BGP speaker can only advertise paths it uses, but does not have to advertise every such
path.
While there are standard default rules for all these (accept everything imported, use simple tie-breakers,
export everything), a site will usually implement at least some policy rules through this filtering process (eg
prefer routes through the ISP we have a contract with).

200

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

As an example of import filtering, a site might elect to ignore all routes from a particular neighbor, or to
ignore all routes whose AS-path contains a particular AS, or to ignore temporarily all routes from a neighbor
that has demonstrated too much recent route instability (that is, rapidly changing routes). Import filtering
can also be done in the best-path-selection stage. Finally, while it is not commonly useful, import filtering
can involve rather strange criteria; for example, in 10.6.8 Examples of BGP Instability we will consider
examples where AS1 prefers routes with AS-path AS3,AS2 to the strictly shorter path AS2.
The next stage is best-path selection, for which the first step is to eliminate AS-paths with loops. Even if the
neighbors have been diligent in not advertising paths with loops, an AS will still need to reject routes that
contain itself in the associated AS-path.
The next step in the best-path-selection stage, generally the most important in BGP configuration, is to assign
a local_preference, or weight, to each route received. An AS may have policies that add a certain amount to
the local_preference for routes that use a certain AS, etc. Very commonly, larger sites will have preferences
based on contractual arrangements with particular neighbors. Provider ASs, for example, will in general
prefer routes that go through their customers, as these are cheaper. A smaller ISP that connects to two
or more larger ones might be paying to route almost all its outbound traffic through a particular one of the
two; its local_preference values will then implement this choice. After BGP calculates the local_preference
value for every route, the routes with the best local_preference are then selected.
Domains are free to choose their local_preference rules however they wish. Some choices may lead to
instability, below, so domains are encouraged to set their rules in accordance with some standard principles,
also below.
In the event of ties two routes to the same destination with the same local_preference a first tie-breaker
rule is to prefer the route with the shorter AS-path. While this superficially resembles a shortest-path algorithm, the real work should have been done in administratively assigning local_preference values.
Local_preference values are communicated internally via the LOCAL_PREF path attribute, below. They
are not shared with other Autonomous Systems.
The final significant step of the route-selection phase is to apply the Multi_Exit_Discriminator value; we
postpone this until below. A site may very well choose to ignore this value entirely. There may then be
additional trivial tie-breaker rules; note that if a tie-breaker rule assigns significant traffic to one AS over
another, then it may have significant economic consequences and shouldnt be considered trivial. If this
situation is detected, it would probably be addressed in the local-preferences phase.
After the best-path-selection stage is complete, the BGP speaker has now selected the routes it will use. The
final stage is to decide what rules will be exported to which neighbors. Only routes the BGP speaker will
use that is, routes that have made it to this point can be exported; a site cannot route to destination D
through AS1 but export a route claiming D can be reached through AS2.
It is at the export-filtering stage that an AS can enforce no-transit rules. If it does not wish to carry transit
traffic to destination D, it will not advertise D to any of its AS-neighbors.
The export stage can lead to anomalies. Suppose, for example, that AS1 reaches D and AS5 via AS2, and
announces this to AS4.

10.6 Border Gateway Protocol, BGP

201

An Introduction to Computer Networks, Release 1.0

Later AS1 switches to reaching D via AS3, but A is forbidden by policy to announce AS3-paths to AS4.
Then A must simply withdraw the announcement to AS4 that it could reach D at all, even though the route
to D via AS2 is still there.

10.6.5 BGP Path attributes

BGP supports the inclusion of various path attributes when exchanging routing information. Attributes
exchanged with neighbors can be transitive or non-transitive; the difference is that if a neighbor AS does
not recognize a received path attribute then it should pass it along anyway if it is marked transitive, but not
otherwise. Some path attributes are entirely local, that is, internal to the AS of origin. Other flags are used
to indicate whether recognition of a path attribute is required or optional, and whether recognition can be
partial or must be complete.
The AS-path itself is perhaps the most fundamental path attribute. Here are a few other common attributes:
[Link] NEXT_HOP
This mandatory external attribute allows BGP speaker B1 of AS1 to inform its BGP peer B2 of AS2 what
actual router to use to reach a given destination. If B1, B2 and AS1s actual border router R1 are all on the
same subnet, B1 will include R1s IP address as its NEXT_HOP attribute. If B1 is not on the same subnet
as B2, it may not know R1s IP address; in this case it may include its own IP address as the NEXT_HOP
attribute. Routers on AS2s side will then look up the immediate next hop they would use as the first step
to reach B1, and forward traffic there. This should either be R1 or should lead to R1, which will then route
the traffic properly (not necessarily on to B1).

202

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

[Link] LOCAL_PREF
If one BGP speaker in an AS has been configured with local_preference values, used in the best-pathselection phase above, it uses the LOCAL_PREF path attribute to share those preferences with all other
BGP speakers at a site.
[Link] MULTI_EXIT_DISC
The Multi-Exit Discriminator, or MED, attribute allows one AS to learn something of the internal structure
of another AS, should it elect to do so. Using the MED information provided by a neighbor has the potential
to cause an AS to incur higher costs, as it may end up carrying traffic for longer distances internally; MED
values received from a neighboring AS are therefore only recognized when there is an explicit administrative
decision to do so.
Specifically, if an autonomous system AS1 has multiple links to neighbor AS2, then AS1 can, when advertising an internal destination D to AS2, have each of its BGP speakers provide associated MED values so
that AS2 can know which link AS1 would prefer that AS2 use to reach D. This allows AS2 to route traffic
to D so that it is carried primarily by AS2 rather than by AS1. The alternative is for AS2 to use only the
closest gateway to AS1, which means traffic is likely carried primarily by AS1.
MED values are considered late in the best-path-selection process; in this sense the use of MED values is a
tie-breaker when two routes have the same local_preference.
As an example, consider the following network (from 10.4.3 Provider-Based Hierarchical Routing, with
providers now replaced by Autonomous Systems); the numeric values on links are their relative costs. We
will assume that border routers R1, R2 and R3 are also AS1s BGP speakers.

In the absence of the MED, AS1 will send traffic from A to B via the R3S3 link, and AS2 will return the
traffic via S1R1. These are the links that are closest to R and S, respectively, representing AS1 and AS2s
desire to hand off the outbound traffic as quickly as possible.
However, AS1s R1, R2 and R3 can provide MED values to AS2 when advertising destination A, indicating
a preference for AS2AS1 traffic to use the rightmost link:
R1: destination A has MED 200
R2: destination A has MED 150

10.6 Border Gateway Protocol, BGP

203

An Introduction to Computer Networks, Release 1.0

R3: destination A has MED 100

If this is done, and AS2 abides by this information, then AS2 will route traffic from B to A via the S3R3
link; that is, via the link with the lowest MED value. Note the importance of fact that AS2 is allowed to
ignore the MED; use of it may shift costs from AS1 to AS2!
The relative order of the MED values for R1 and R2 is irrelevant, unless the R3S3 link becomes disabled,
in which case the numeric MED values above would mean that AS2 should then prefer the R2S2 link for
reaching A.
There is no way to use MED values to cause AB traffic to take the R2S2 link; that link is the minimal-cost
link only in the global sense, and the only way to achieve global cost minimization is for the two ASs to
agree to use a common distance metric and a common metric-based routing algorithm, in effect becoming
one AS. While AS1 does provide different numeric MED values for the three cross-links, they are used only
in ranking precedence, not as numeric measures of cost (though they are sometimes derived from that).
As a hypothetical example of why a site might use the MED, suppose site B in the diagram above wants
its customers to experience high-performance downloads. It contracts with AS2, which advertises superior
quality in the downloads experienced by users connecting to servers in its domain. In order to achieve
this superior quality, it builds a particularly robust network S1SS2S3. It then agrees to accept MED
information from other providers so that it can keep outbound traffic in its own network as long as possible,
instead of handing it off to other networks of lower quality. AS2 would thus want BA traffic to travel via
S3 to maximize the portion of the path under its own control.
In the example above, the MED values are used to decide between multiple routes to the same destination
that all pass through the same AS, namely AS1. Some BGP implementations allow the use of MED values to
decide between different routes through different neighbor ASs. The different neighbors must all have the
same local_preference values. For example, AS2 might connect to AS3 and AS4 and receive the following
BGP information:
AS3: destination A has MED 200
AS4: destination A has MED 100
Assuming AS2 assigns the same local_preference to AS3 and AS4, it might be configured to use these MED values as
On Cisco routers, the always-compare-med command is used to create this behavior.
MED values are not intended to be used to communicate routing preferences to non-neighboring ASs.
Additional information on the use of MED values can be found in RFC 4451.
[Link] COMMUNITY
This is simply a tag to attach to routes. Routes can have multiple tags corresponding to membership in multiple communities. Some communities are defined globally; for example, NO_EXPORT and
NO_ADVERTISE. A route marked with one of these two communities will not be shared further. Other
communities may be relevant only to a particular AS.
The importance of communities is that they allow one AS to place some of its routes into specific categories
when advertising them to another AS; the categories must have been created and recognized by the receiving
AS. The receiving AS is not obligated to honor the community memberships, of course, but doing so has

204

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

the effect of allowing the original AS to configure itself without involving the receiving AS in the process. Communities are often used, for example, by (large) customers of an ISP to request specific routing
treatment.
A customer would have to find out from the provider what communities the provider defines, and what their
numeric codes are. At that point the customer can place itself into the providers community at will.
Here are some of the community values once supported by a no-longer-extant ISP that we shall call AS1.
The full community value would have included AS1s AS-number.
value
90
100
105
110
990
991

action
set local_preference used by AS1 to 90
set local_preference used by AS1 to 100, the default
set local_preference used by AS1 to 105
set local_preference used by AS1 to 110
the route will not leave AS1s domain; equivalent to NO_EXPORT
route will only be exported to AS1s other customers

10.6.6 BGP Policy and Transit Traffic

Perhaps the most common source of simple policy decisions is whether a site wants to accept transit traffic.
As a first example, let us consider the case of configuring a private link, such as the dashed link1 below
between friendly but unaffiliated sites A and B:

If A and B fully advertise link1, by exporting to their respective ISPs routes to each other, then ISP1 (paid
by A) may end up carrying much of Bs traffic or ISP2 (paid by B) may end up carrying much of As traffic.
Economically, these options are not desirable unless fully agreed to by both parties. The primary issue here
is the use of the ISP1A link by B, and the ISP2B link by A; use of the shared link1 might be a secondary
issue depending on the relative bandwidths and A and Bs understandings of appropriate uses for link1.
Three common options A and B might agree to regarding link1 are no-transit, backup, and load-balancing.
For the no-transit option, A and B simply do not export the route to their respective ISPs at all. This is done
via export filtering. If ISP1 does not know A can reach B, it will not send any of Bs traffic to A.
For the backup option, the intent is that traffic to A will normally arrive via ISP1, but if the ISP1 link is down
then As traffic will be allowed to travel through ISP2 and B. To achieve this, A and B can export their link1route to each other, but arrange for ISP1 and ISP2 respectively to assign this route a low local_preference
value. As long as ISP1 hears of a route to B from its upstream provider, it will reach B that way, and will
not advertise the existence of the link1 route to B; ditto ISP2. However, if the ISP2 route to B fails, then As
upstream provider will stop advertising any route to B, and so ISP1 will begin to use the link1 route to B

10.6 Border Gateway Protocol, BGP

205

An Introduction to Computer Networks, Release 1.0

and begin advertising it to the Internet. The link1 route will be the primary route to B until ISP2s service is
restored.
A and B must convince their respective ISPs to assign the link1 route a low local_preference; they cannot
mandate this directly. However, if their ISPs recognize community attributes that, as above, allow customers
to influence their local_preference value, then A and B can use this to create the desired local_preference.
For outbound traffic, A and B will need a way to send through one another if their own ISP link is down.
One approach is to consider their default-route path (eg to [Link]/0) to be a concrete destination within BGP.
ISP1 advertises this to A, using As interior routing protocol, but so does B, and A has configured things so
Bs route has a higher cost. Then A will route to [Link]/0 through ISP1 that is, will use ISP1 as its default
route as long as it is available, and will switch to B when it is not.
For inbound load balancing, there is no easy fix, in that if ISP1 and ISP2 both export routes to A, then A has
lost all control over how other sites will prefer one to the other. A may be able to make one path artificially
appear more expensive, and keep tweaking this cost until the inbound loads are comparable. Outbound
load-balancing is up to A and B.
Another basic policy question is which of the two available paths site (or regional AS) A uses to reach site
D, in the following diagram. B and C are Autonomous Systems.

How can A express preference for B over C, assuming B and C both advertise to A their routes to D?
Generally A will use a local_preference setting to make the carrier decision for AD traffic, though it is
D that makes the decision for the DA traffic. It is possible (though not customary) for one of the transit
providers to advertise to A that it can reach D, but not advertise to D that it can reach A.
Here is a similar diagram, showing two transit-providing Autonomous Systems B and C connecting at
Internet exchange points IXP1 and IXP2.

B and C each have routers within each IXP. B would probably like to make sure C does not attempt to save
on its long-haul transit costs by forwarding AD traffic over to B at IXP1, and DA traffic over to
B at IXP2. B can avoid this problem by not advertising to C that it can reach A and D. In general, transit
providers are often quite careful about advertising reachability to any other AS for whom they do not intend
to provide transit service, because to do so may implicitly mean getting stuck with that traffic.
If B and C were both to try try to get away with this, a routing loop would be created within IXP1! But in
that case in Bs next advertisement to C at IXP1, B would state that it reaches D via AS-path C (or C,D

206

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

if D were a full-fledged AS), and C would do similarly; the loop would not continue for long.

10.6.7 BGP Relationships

Arbitrarily complex policies may be created through BGP, and, as we shall see in the following section,
convergence to a stable set of routes is not guaranteed. However, most AS policies are determined by a few
simple AS business relationships, as identified in [LG01].
Customer-Provider: A provider contracts to transit traffic from the customer to the rest of the
Internet. A customer may have multiple providers, for backup or for more efficient routing, but
a customer does not accept transit traffic between the providers.
Siblings: Siblings are ISPs that have a connection between them that they intend as a backup
route or for internal traffic. Two siblings may or may not have the same upstream ISP provider
as parent. The siblings do not use their connection for transit traffic except when one of their
primary links is down.
Peers: Peers are two providers that agree to exchange all their customer traffic with each other;
thus, peers do use their connection for transit traffic. Generally the idea is for the interconnection to be seen as equally valuable by both parties (eg because the parties exchange comparable
volumes of traffic); in such a case there would likely be no exchange of cash, but if the volume
flow is significantly asymmetric then compensation can certainly be negotiated.
These business relationships can be described in terms of or even inferred from what routes are accepted
[LG01]. Every AS has, first of all, its customer routes, that is, the routes of its direct customers, and its
customers customers, etc. These might be referred to as the ISPs own routes. Other routes are learned from
peers or, sometimes, providers. Essentially every AS exports its customer routes to all its AS neighbors; the
issue is with provider routes and peer routes. Here are the rules of [LG01]:
Customers do not export to their providers routes they have learned from peers or other providers.
This would make them into transit providers, which is not their job.
Providers do export provider/peer routes to their customers (though a provider may just provide a
single consolidated default route to a single-homed customer).
Peers do not export peer or provider routes to other peers; that is, if ISP P1 has peers P2 and P3, and
P1 learns of a route to destination D through P2, then P1 does not export that route to P3. Instead,
P3 would be expected to have its own relationship with P2, or be a customer of a provider that had a
relationship with P2. Peers do not (usually) provide transit services to third parties.
Siblings do export provider/peer routes to each other, though likely with an understanding about
preference values. If S1 and S2 are two siblings with a mutual-backup relationship, and S1 loses its
normal connectivity and must rely on the S1S2 link, then S2 must know all about S1s routes.
All ASs do export their own routes, and their customers routes, to customers, siblings, peers and
providers.
There may be more complicated variations: if a regional provider is a customer of a large transit backbone,
then the backbone might only announce routes listed in transit agreement (rather than all routes, as above).
There is a supposition here that the regional provider has multiple connections, and has contracted with that
particular transit backbone only for certain routes.

10.6 Border Gateway Protocol, BGP

207

An Introduction to Computer Networks, Release 1.0

Following these rules creates a simplified BGP world. Special cases for special situations have the potential
to introduce non-convergence or instability.
The so-called tier-1 providers are those that are not customers of anyone; these represent the top-level
backbone providers. Each tier-1 AS must, as a rule, peer with every other tier-1 AS.
A consequence of the use of the above classification and attendant export rules is the no-valley theorem
[LG01]: if every AS has BGP policies consistent with the scheme above, then when we consider the full ASpath from A to B, there is at most one peer-peer link. Those to the left of the peer-peer link are (moving from
left to right) either customerprovider links or siblingsibling links; that is, they are non-downwards
(ie upwards or level). To the right of the peer-peer link, we see providercustomer or siblingsibling
links; that is, these are non-upwards. If there is no peer-peer link, then we can still divide the AS-path into
a non-downwards first part and a non-upwards second part.
The above constraints are not quite sufficient to guarantee convergence of the BGP system to a stable set
of routes. To ensure convergence in the case without sibling relationships, it is shown in [GR01] that the
following simple local_preference rule suffices:
If AS1 gets two routes r1 and r2 to a destination D, and the first AS of the r1 route is a customer
of AS1, and the first AS of r2 is not, then r1 will be assigned a higher local_preference value
than r2.
More complex rules exist that allow for cases when the local_preference values can be equal; one such rule
states that strict inequality is only required when r2 is a provider route. Other straightforward rules handle
the case of sibling relationships, eg by requiring that siblings have local_preference rules consistent with the
use of their shared connection only for backup.
As a practical matter, unstable BGP arrangements appear rare on the Internet; most actual relationships and
configurations are consistent with the rules above.

10.6.8 Examples of BGP Instability

What if the normal rules regarding BGP preferences are not followed? It turns out that BGP allows
genuinely unstable situations to occur; this is a consequence of allowing each AS a completely independent
hand in selecting preference functions. Here are two simple examples, from [GR01].
Example 1: A stable state exists, but convergence to it is not guaranteed. Consider the following network
arrangement:

We assume AS1 prefers AS-paths to destination D in the following order:

AS2,AS0, AS0

208

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

That is, AS2,AS0 is preferred to the direct path AS0 (one way to express this preference might be
prefer routes for which the AS-PATH begins with AS2; perhaps the AS1AS0 link is more expensive).
Similarly, we assume AS2 prefers paths to D in the order AS1,AS0, AS0. Both AS1 and AS2 start out
using path AS0; they advertise this to each other. As each receives the others advertisement, they apply
their preference order and therefore each switches to routing Ds traffic to the other; that is, AS1 switches
to the route with AS-path AS2,AS0 and AS2 switches to AS1,AS0. This, of course, causes a routing
loop! However, as soon as they export these paths to one another, they will detect the loop in the AS-path
and reject the new route, and so both will switch back to AS0 as soon as they announce to each other the
change in what they use.
This oscillation may continue indefinitely, as long as both AS1 and AS2 switch away from AS0 at the
same moment. If, however, AS1 switches to AS2,AS0 while AS2 continues to use AS0, then AS2 is
stuck and the situation is stable. In practice, therefore, eventual convergence to a stable state is likely.
AS1 and AS2 might choose not to export their D-route to each other to avoid this instability.
Example 2: No stable state exists. This example is from [VGE00]. Assume that the destination D is attached
to AS0, and that AS0 in turn connects to AS1, AS2 and AS3 as in the following diagram:

AS1-AS3 each have a direct route to AS0, but we assume each prefers the AS-path that takes their clockwise
neighbor; that is, AS1 prefers AS3,AS0 to AS0; AS3 prefers AS2,AS0 to AS0, and AS2 prefers
AS1,AS0 to AS0. This is a peculiar, but legal, example of input filtering.
Suppose all adopt AS0, and advertise this, and AS1 is the first to look at the incoming advertisements.
AS1 switches to the route AS3,AS0, and announces this.
At this point, AS2 sees that AS1 uses AS3,AS0; if AS2 switches to AS1 then its path would be
AS1,AS3,AS0 rather than AS1,AS0 and so it does not make the switch.
But AS3 does switch: it prefers AS2,AS0 and this is still available. Once it makes this switch, and
advertises it, AS1 sees that the route it had been using, AS3,AS0, has become AS3,AS1,AS0. At this
point AS1 switches back to AS0.
Now AS2 can switch to using AS1,AS0, and does so. After that, AS3 finds it is now using AS2,AS1,AS0
and it switches back to AS0. This allows AS1 to switch to the longer route, and then AS2 switches back
to the direct route, and then AS3 gets the longer route, then AS2 again, etc, forever rotating clockwise.

10.7 Epilog
CIDR was a deceptively simple idea. At first glance it is a straightforward extension of the subnet concept,
moving the net/host division point to the left as well as to the right. But it has ushered in true hierarchical
10.7 Epilog

209

An Introduction to Computer Networks, Release 1.0

routing, most often provider-based. While CIDR was originally offered as a solution to some early crises in
IPv4 address-space allocation, it has been adopted into the core of IPv6 routing as well.
Interior routing using either distance-vector or link-state protocols is neat and mathematical. Exterior
routing with BGP is messy and arbitrary. Perhaps the most surprising thing about BGP is that the Internet
works as well as it does, given the complexity of provider interconnections. The business side of routing
almost never has an impact on ordinary users. To an extent, BGP works well because providers voluntarily limit the complexity of their filtering preferences, but that seems to be largely because the business
relationships of real-world ISPs do not seem to require complex filtering.

10.8 Exercises
1. Consider the following IP forwarding table that uses CIDR. IP address bytes are in hexadecimal, so each
hex digit corresponds to four address bits.
destination
[Link]/12
81.3c.0.0/16
81.3c.50.0/20
[Link]/12
[Link]/14

next_hop
A
B
C
D
E

For each of the following IP addresses, indicate to what destination it is forwarded.

(i) 81.3b.15.49
(ii) 81.3c.56.14
(iii) 81.3c.85.2e
(iv) 81.4a.35.29
(v) [Link]
(vi) 81.43.01.c0
2. Consider the following IP forwarding table, using CIDR. As in exercise 1, IP address bytes are in hexadecimal.
destination
[Link]/2
[Link]/2
[Link]/2
C0.0.0.0/2

next_hop
A
B
C
D

(a). To what next_hop would each of the following be routed? 63.b1.82.15, 9e.00.15.01, [Link]
(b). Explain why every IP address is routed somewhere, even though there is no default entry.

3. Give an IPv4 forwarding table using CIDR that will route all Class A addresses to next_hop A, all
Class B addresses to next_hop B, and all Class C addresses to next_hop C.

210

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

4. Suppose a router using CIDR has the following entries. Address bytes are in decimal except for the third
byte, which is in binary.
destination
37.119.0000 0000.0/18
37.119.0100 0000.0/18
37.119.1000 0000.0/18
37.119.1100 0000.0/18

next_hop
A
A
A
B

These four entries cannot be consolidated into a single /16 entry, because they dont all go to the same
next_hop. How could they be consolidated into two entries?

5. Suppose P, Q and R are ISPs with respective CIDR address blocks (with bytes in decimal) [Link]/8, [Link]/8 and
A: [Link]/16
B: [Link]/16
Q has customers C and D and assigns them address blocks as follows:
C: [Link]/16
D: [Link]/16

(a). Give forwarding tables for P, Q and R assuming they connect to each other and to each of their own
customers.
(b). Now suppose A switches from provider P to provider Q, and takes its address block with it. Give the
forwarding tables for P, Q and R; the longest-match rule will be needed to resolve conflicts.
(c). Now suppose in addition to A switching from P to Q, C switches from provider Q to provider R. Give
the forwarding tables.

6. Suppose P, Q and R are ISPs as in the previous problem. P and R do not connect directly; they route traffic
to one another via Q. In addition, customer B is multi-homed and has a secondary connection to provider R;
customer D is also multi-homed and has a secondary connection to provider P. R and P use these secondary
connections to send to B and D respectively; however, these secondary connections are not advertised to
other providers. Give forwarding tables for P, Q and R.
7. Consider the following network of providers P-S, all using BGP. The providers are the horizontal lines;
each provider is its own AS.

10.8 Exercises

211

An Introduction to Computer Networks, Release 1.0

(a). What routes to network NS will P receive, assuming there is no export filtering? For each route, list the
AS-path.
(b). What routes to network NQ will P receive? For each route, list the AS-path.
(c). Suppose R uses export filtering so as not to advertise to P any of its routes except those that involve S
in their AS-path. What routes to network NR will P receive, with AS-paths?

8. Consider the following network of Autonomous Systems AS1 through AS6, which double as destinations.
When AS1 advertises itself to AS2, for example, the AS-path it provides is AS1.
AS1

AS2

AS4

AS5

AS3
:
:
:
AS6

(a). If neither AS3 nor AS6 exports their AS3AS6 link to their neighbors AS2 and AS5 to the left, what
routes will AS2 receive to reach AS5? Specify routes by AS-path.
(b). What routes will AS2 receive to reach AS6?
(c). Suppose AS3 exports to AS2 its link to AS6, but AS6 continues not to export the AS3AS6 link to
AS5. How will AS5 now reach AS2? How will AS2 now reach AS6? Assume that there are no local
preferences in use in BGP best-path selection, and that the shortest AS-path wins.

9. Suppose that Internet routing in the US used geographical routing, and the first 12 bits of every IP
address represent a geographical area similar in size to a telephone area code. Megacorp gets the prefix
[Link]/16, based geographically in Chicago, and allocates subnets from this prefix to its offices in all 50
states. Megacorp routes all its internal traffic over its own network.

(a). Assuming all Megacorp traffic must enter and exit in Chicago, what is the route of traffic to and from
the San Diego office to a client also in San Diego?

212

10 Large-Scale IP Routing

An Introduction to Computer Networks, Release 1.0

(b). Now suppose each office has its own link to a local ISP, but still uses its [Link]/16 IP addresses.
Now what is the route of traffic between the San Diego office and its neighbor?
(c). Suppose Megacorp gives up and gets a separate geographical prefix for each office. What must it do to
ensure that its internal traffic is still routed over its own network?

10. Suppose we try to use BGPs strategy of exchanging destinations plus paths as an interior routing-update
strategy, perhaps replacing distance-vector routing. No costs or hop-counts are used, but routers attach to
each destination a list of the routers used to reach that destination. Routers can also have route preferences,
such as prefer my link to B whenever possible.
(a). Consider the network of 9.2 Distance-Vector Slow-Convergence Problem:
D

The DA link breaks, and B offers A what it thinks is its own route to D. Explain how exchanging path
information prevents a routing loop here.
(b). Suppose the network is as below, and initially each router knows about itself and its immediately
adjacent neighbors. What sequence of router announcements can lead to A reaching F via
ADEBCF, and what individual router preferences would be necessary? (Initially, for example,
A would reach B directly; what preference might make it prefer ADEB?)

(c). Explain why this method is equivalent to using the hopcount metric with either distance-vector or
link-state routing, if routers are not allowed to have preferences and if the router-path length is used as a
tie-breaker.

10.8 Exercises

213

An Introduction to Computer Networks, Release 1.0

214

10 Large-Scale IP Routing

11 UDP TRANSPORT

The standard transport protocols riding above the IP layer are TCP and UDP. As we saw in Chapter 1, UDP
provides simple datagram delivery to remote sockets, that is, to host,port pairs. TCP provides a much
richer functionality for sending data, but requires that the remote socket first be connected. In this chapter,
we start with the much-simpler UDP, including the UDP-based Trivial File Transfer Protocol.
We also review some fundamental issues any transport protocol must address, such as lost final packets and
packets arriving late enough to be subject to misinterpretation upon arrival. These fundamental issues will
be equally applicable to TCP connections.

11.1 User Datagram Protocol UDP

RFC 1122 refers to UDP as almost a null protocol; while that is something of a harsh assessment, UDP
is indeed fairly basic. The two features it adds beyond the IP layer are port numbers and a checksum on
the data. The UDP header consists of the following:

The port numbers are what makes UDP into a real transport protocol: with them, an application can now
connect to an individual server process (that is, the process owning the port number in question), rather
than simply to a host.
UDP is unreliable, in that there is no UDP-layer attempt at timeouts, acknowledgment and retransmission;
applications written for UDP must implement these. As with TCP, a UDP host,port pair is known as a
socket (though UDP ports are considered a separate namespace from TCP ports). UDP is also unconnected,
or stateless; if an application has opened a port on a host, any other host on the Internet may deliver packets
to that host,port socket without preliminary negotiation.
UDP packets use the 16-bit Internet checksum (5.4 Error Detection) on the data. While it is seldom done
now, the checksum can be disabled and the field set to the all-0-bits value, which never occurs as an actual
ones-complement sum.
UDP packets can be dropped due to queue overflows either at an intervening router or at the receiving host.
When the latter happens, it means that packets are arriving faster than the receiver can process them. Higherlevel protocols that define ACK packets (eg UDP-based RPC, below) typically include some form of flow
control to prevent this.
UDP is popular for local transport, confined to one LAN. In this setting it is common to use UDP as the
transport basis for a Remote Procedure Call, or RPC, protocol. The conceptual idea behind RPC is that
one host invokes a procedure on another host; the parameters and the return value are transported back and
forth by UDP. We will consider RPC in greater detail below, in 11.4 Remote Procedure Call (RPC); for

215

An Introduction to Computer Networks, Release 1.0

now, the point of UDP is that on a local LAN we can fall back on rather simple mechanisms for timeout and
retransmission.
UDP is well-suited for request-reply semantics beyond RPC; one can use TCP to send a message and get
a reply, but there is the additional overhead of setting up and tearing down a connection. DNS uses UDP,
largely for this reason. However, if there is any chance that a sequence of request-reply operations will be
performed in short order then TCP may be worth the overhead.
UDP is also popular for real-time transport; the issue here is head-of-line blocking. If a TCP packet is lost,
then the receiving host queues any later data until the lost data is retransmitted successfully, which can take
several RTTs; there is no option for the receiving application to request different behavior. UDP, on the
other hand, gives the receiving application the freedom simply to ignore lost packets. This approach is very
successful for voice and video, where small losses simply degrade the received signal slightly, but where
larger delays are intolerable. This is the reason the Real-time Transport Protocol, or RTP, is built on top
of UDP rather than TCP. It is common for VoIP telephone calls to use RTP and UDP.

11.1.1 QUIC
Sometimes UDP is used simply because it allows new or experimental protocols to run entirely as user-space
applications; no kernel updates are required, as would be the case with TCP changes. Google has created
a protocol named QUIC (Quick UDP Internet Connections, [Link]/quic) in this category, though
QUIC also takes advantage of UDPs freedom from head-of-line blocking. For example, one of QUICs
goals includes supporting multiplexed streams in a single connection (eg for the multiple components of a
web page). A lost packet blocks its own stream until it is retransmitted, but the other streams can continue
without waiting. Because QUIC supports error-correcting codes (5.4.2 Error-Correcting Codes), a lost
packet might not require any waiting at all; this is another feature that would be difficult to add to TCP.
QUIC also eliminates the extra RTT needed for setting up a TCP connection.
QUIC provides support for advanced congestion control, currently (2014) including a UDP analog of TCP
CUBIC (15.11 TCP CUBIC). QUIC does this at the application layer but new congestion-control mechanisms within TCP often require client operating-system changes even when the mechanism lives primarily
at the server end. QUIC represents a promising approach to using UDPs flexibility to support innovative
or experimental transport-layer features. The downside of QUIC is its nonstandard programming interface,
but note that Google can achieve widespread web utilization of QUIC simply by distributing the client side
in its Chrome browser.

11.1.2 UDP Simplex-Talk

One of the early standard examples for socket programming is simplex-talk. The client side reads lines of
text from the users terminal and sends them over the network to the server; the server then displays them on
its terminal. Simplex here refers to the one-way nature of the flow; duplex talk is the basis for Instant
Messaging, or IM. Even at this simple level we have some details to attend to regarding the data protocol:
we assume here that the lines are sent with a trailing end-of-line marker. In a world where different OSs
use different end-of-line marks, including them in the transmitted data can be problematic. However, when
we get to the TCP version, if arriving packets are queued for any reason then the embedded end-of-line
character will be the only thing to separate the arriving data into lines.
As with almost every Internet protocol, the server side must select a port number, which with the servers IP

216

11 UDP Transport

An Introduction to Computer Networks, Release 1.0

address will form the socket address to which clients connect. Clients must discover that port number or
have it written into their application code. Clients too will have a port number, but it is largely invisible.
On the server side, simplex-talk must do the following:
ask for a designated port number
create a socket, the sending/receiving endpoint
bind the socket to the socket address, if this is not done at the point of socket creation
receive packets sent to the socket
for each packet received, print its sender and its content
The client side has a similar list:
look up the servers IP address, using DNS
create an anonymous socket; we dont care what the clients port number is
read a line from the terminal, and send it to the socket address server_IP,port
[Link] The Server
We will start with the server side, presented here in Java. We will use port 5432; the socket-creation and
port-binding operations are combined into the single operation new DatagramSocket(destport).
Once created, this socket will receive packets from any host that addresses a packet to it; there is no need
for preliminary connection. We also need a DatagramPacket object that contains the packet data and
source IP_address,port for arriving packets. The server application does not acknowledge anything sent to
it, or in fact send any response at all.
The server application needs no parameters; it just starts. (That said, we could make the port number a
parameter, to allow easy change. The port we use here, 5432, has also been adopted by PostgreSQL for TCP
connections.) The server accepts both IPv4 and IPv6 connections; we return to this below.
Though it plays no role in the protocol, we will also have the server time out every 15 seconds and display
a message, just to show how this is done; implementations of real protocols essentially always must arrange
when attempting to receive a packet to time out after a certain interval with no response. The file below is
at udp_stalks.java.
/* simplex-talk server, UDP version */
import [Link].*;
import [Link].*;
public class stalks {
static public int destport = 5432;
static public int bufsize = 512;
static public final int timeout = 15000; // time in milliseconds
static public void main(String args[]) {
DatagramSocket s;
// UDP uses DatagramSockets

11.1 User Datagram Protocol UDP

217

An Introduction to Computer Networks, Release 1.0

try {
s = new DatagramSocket(destport);
}
catch (SocketException se) {
[Link]("cannot create socket with port " + destport);
return;
}
try {
[Link](timeout);
// set timeout in milliseconds
} catch (SocketException se) {
[Link]("socket exception: timeout not set!");
}
// create DatagramPacket object for receiving data:
DatagramPacket msg = new DatagramPacket(new byte[bufsize], bufsize);
while(true) { // read loop
try {
[Link](bufsize); // max received packet size
[Link](msg);
// the actual receive operation
[Link]("message from <" +
[Link]().getHostAddress() + "," + [Link]() + ">");
} catch (SocketTimeoutException ste) {
// receive() timed out
[Link]("Response timed out!");
continue;
} catch (IOException ioe) {
// should never happen!
[Link]("Bad receive");
break;
}
String str = new String([Link](), 0, [Link]());
[Link](str);
// newline must be part of str
}
[Link]();
} // end of main
}

[Link] UDP and IP addresses

The line s = new DatagramSocket(destport) creates a DatagramSocket object bound to the
given port. If a host has multiple IP addresses, packets sent to that port to any of those IP addresses will
be delivered to the socket, including localhost (and in fact all IPv4 addresses between [Link] and
[Link]) and the subnet broadcast address (eg [Link]). If a client attempts to connect to the
subnet broadcast address, multiple servers may receive the packet (in this we are perhaps fortunate that the
stalk server does not reply).
Alternatively, we could have used
s = new DatagramSocket(int port, InetAddress local_addr)

in which case only packets sent to the host and port through the hosts specific IP address local_addr
would be delivered. It does not matter here whether IP forwarding on the host has been enabled. In the
218

11 UDP Transport

An Introduction to Computer Networks, Release 1.0

original C socket library, this binding of a port to (usually) a server socket was done with the bind() call.
To allow connections via any of the hosts IP addresses, the special IP address INADDR_ANY is passed to
bind().
When a host has multiple IP addresses, the standard socket library does not provide a way to find out to
which these an arriving UDP packet was actually sent. Normally, however, this is not a major difficulty. If
a host has only one interface on an actual network (ie not counting loopback), and only one IP address for
that interface, then any remote clients must send to that interface and address. Replies (if any, which there
are not with stalk) will also come from that address.
Multiple interfaces do not necessarily create an ambiguity either; the easiest such case to experiment with
involves use of the loopback and Ethernet interfaces (though one would need to use an application that,
unlike stalk, sends replies). If these interfaces have respective IPv4 addresses [Link] and [Link],
and the client is run on the same machine, then connections to the server application sent to [Link] will
be answered from [Link], and connections sent to [Link] will be answered from [Link]. The
IP layer sees these as different subnets, and fills in the IP source-address field according to the appropriate
subnet. The same applies if multiple Ethernet interfaces are involved, or if a single Ethernet interface is
assigned IP addresses for two different subnets, eg [Link] and [Link].
Life is slightly more complicated if a single interface is assigned multiple IP addresses on the same subnet,
eg [Link] and [Link]. Regardless of which address a client sends its request to, the servers reply
will generally always come from one designated address for that subnet, eg [Link]. Thus, it is possible
that a legitimate UDP reply will come from a different IP address than that to which the initial request was
sent.
If this behavior is not desired, one approach is to create multiple server sockets, and to bind each of the
hosts network IP addresses to a different server socket.
[Link] The Client
Next is the Java client version udp_stalkc.java. The client any client must provide the name of the
host to which it wishes to send; as with the port number this can be hard-coded into the application but is
more commonly specified by the user. The version here uses host localhost as a default but accepts any
other hostname as a command-line argument. The call to [Link](desthost)
invokes the DNS system, which looks up name desthost and, if successful, returns an IP address.
([Link]() also accepts addresses in numeric form, eg [Link], in which case
DNS is not necessary.) When we create the socket we do not designate a port in the call to new
DatagramSocket(); this means any port will do for the client. When we create the DatagramPacket
object, the first parameter is a zero-length array as the actual data array will be provided within the loop.
A certain degree of messiness is introduced by the need to create a BufferedReader object to handle
terminal input.
// simplex-talk CLIENT in java, UDP version
import [Link].*;
import [Link].*;
public class stalkc {
static public BufferedReader bin;

11.1 User Datagram Pr