End-To-End IP Service Quality and Mobility: - Lecture #3
End-To-End IP Service Quality and Mobility: - Lecture #3
Vilho Risnen
Agenda
Goal of this lecture. End-to-end service quality requirements. Link to service event characteristics. Analysis of mobile services.
Vilho Risnen
Jan 13th Jan 20th Jan 27th Feb 3rd (Feb 10th) (Feb 17th) (Feb 24th) (Mar 3rd) (Mar 10th) (Mar 17th) (Mar 31st) (Apr 7th)
Vilho Risnen
Vilho Risnen
Medium
AMR 12.2
AMR 7.95
Medium
Tight
Loose
Vilho Risnen
HTTP run on top of TCP TCP throughput considerations need to be taken into account.
Vilho Risnen
Loose
Tight
quality evaluation, Proc. ICC2001] [Ramjee et al., Adaptive playout mechanisms, Proc. Infocom94.]
MOS
Operation of basic RTP media 6,0% 5,0% client: 4,0% Receive a sample (with 3,0% timestamp TS). 2,0% 1,0% Play out at time TS + . 0,0% 0 10 20 30 40 Dejittering buffer can use the Jitter Buffer Size (ms) following operations: Shift buffer in time (affects e2e 5.0 No errors 4.5 delay). 4.0 [Lakaniemi et al., Resize dejittering buffer 3.5 Subjective VoIP speech 3.0 (affects e2e delay).
Maximum Loss (%)
2.5 2.0 1.5 1.0
[101329-6]
Dynamic JB Delay (ms) Static JB Delay (ms)
50
CIU
Total
CIL
Dom estic
International
TCP throughput
Basic TCP: flow control based on advertised window size in ACKs. TCP throughput model of Padhye et al: Based on Reno variant. Model with loss rate and RTT as parameters. RTT and throughput consistency enhance overall throughput.
W 1 B( p) min max , RTT 3bp 2bp p 1 + 32 p 2 RTT + T0 min1,3 8 3
B=throughput p=probability that packet is lost W=window size RTT=round-trip delay b=number packets acknowledged by ACK T0=length of the first timeout period
[J. Padhye, V. Firoiu, D. Towsley and J. Kurose, Modeling TCP Reno Performance, IEEE/ACM Transactions on Networking, April 2000]
Vilho Risnen
Vilho Risnen
On temporal correlations
Latency: E2e latency can be measured for a service event or part thereof -> di. Measurement for a set of service events -> time series {di}. Delay jitter = temporal correlation of {di}. Packet loss: Service event is lost / not lost -> {li}. Measurement for a set of service events -> {li}. Loss correlation = temporal correlation of {li}. Throughput consistency: Temporal correlation of {bi}.
[K.Lai, M. Baker, Measuring bandwidth, in Proc. Infocom99, p. 235 ff.]
Vilho Risnen
Availability
Availability is measured on aggregate service level as per cents of total time. Depending on service, availability may be defined in different ways: All service events belonging to the service instance must be present. Some service events may be optional. Example: ACME kryptonite detectors Seeing marketing video considered essential for service event, ditto for other service events. Availability = 100% - (V||P||C)
V = video unavailability P = payment unavailability C= confirmation unavailability
Vilho Risnen
For other kinds of services, interactivity is important for service instantiation. Browsing. Streaming. In all cases, service instantiation time should be engineered for carrier-grade services. Consistency important: e.g., 95% percentile for service instantiation time distribution = X seconds.
Vilho Risnen
Latency
Inherent latency requirements: PDUs of IP media streams of conferencing applications. Designed latency: Interactive IP control traffic (RTSP, SIP, HTTP GET, WAP). Designed throughput: Data traffic (FTP, HTML).
3 (WIDE BAND) 2 (NARROWBAND) 2H (HIGH) < 100 ms 2M (MEDI UM) < 150 ms 2A
(ACCEP TABLE)
< 100 ms
< 400 ms
[101329-1]
Vilho Risnen
[101329-7]
Event loss
Overall effect of event loss depends on L4 and L2 protocol. TCP: retransmissions hidden from application up to maximum retry limit. UDP: IP-level packet losses visible to end application. L2 reliability. Loss can also take place because of bit errors (cf. next slide). PDU loss usually degrades service quality of all services. For media streams, correlated packet loss degrades end user experienced quality even with loss concealment schemes. VoIP: lower audio quality. Packet video: skipped / distorted frames.
[V. Risnen, On end-to-end analysis of packet loss, Computer Communications, in press.]
Vilho Risnen
BER
Bit Error Rate (BER) most important for wireless links. UDP: checksum computed over header & data. If checksum does not match one computed from L4 SDU, datagram rejected. RT data on wireless links: checking of headers needed, retransmissions not good for end-to-end latency. Higher bits of audio samples are more important than lower bits. UDPLite: checksum covers header + predefined number of the highest bits of payload.
Throughput consistency
Minimum token rate required by media streams of real-time IP applications: IP telephony. Streaming. Throughput consistency also yields higher overall throughput for TCP-based data transfer. For long data transfers, overall throughput most important => effect of individual bursts of lower throughput.
Vilho Risnen
[G.109]
User's satisfaction
Vilho Risnen
Data transfer
Desirable end user experience: Availability designed. Data transfer commences quickly. Overall duration predictable. Conclusions: Designed delay for small messages in UL direction. Downlink:
Designed delay, relatively consistent. Throughput relatively consistent. Packet loss designed, relatively consistent.
Vilho Risnen
10
Interactive applications
Desirable end user experience: Availability relatively high (designed). Replies to requests take place interactively. Duration for downloading predictable. Conclusions: Availability needs special attention. Designed delay for small messages in UL direction. Downlink:
Designed delay, relatively consistent. Throughput relatively consistent. Packet loss designed, relatively consistent.
Vilho Risnen
Vilho Risnen
11
Streaming
Desirable end user experience: Availability high. Service instantiation interactive. Media quality high (more important for audio than for video). Conclusions: Availability needs special attention. Designed delay and loss performance for RTSP. Latency for media streams relatively small. Constant token rate desirable
Packet loss allowed if retransmissions possible.
Multimedia conferencing
Desirable end user experience: Availability high. Service instantiation quick. Media quality high (more important for audio than for video). Conclusions: Availability needs special attention. High delay and loss performance for SIP signalling. Latency for media streams small. Minimum token rate required for audio component, desirable for video component. BER allowable for media streams.
Vilho Risnen
12
Summary
Service quality requirements: service type inherent vs. end user experience related. Most important service quality characteristics: Availability. Service instantiation time. Throughput consistency. E2e latency. E2e event loss. (BER). Temporal correlations. Modelling of end user experience.
Vilho Risnen
13