Software-Defined Access Network (SDAN)

Ken Kerpez and George Ginis

ASSIA, Inc.
Redwood City, California, USA
[email protected]

Abstract—Control-plane functions are migrating from infrastructure, and retail service providers responsible for
dedicated network equipment into software running on interfacing to customers and providing services. While
commodity hardware. The Software-Defined Access Network SDAN is highly useful in multi-operator environments; it
(SDAN) concept is introduced here that extends the benefits of also benefits a single operator with multiple networks and
Software-Defined Networking (SDN) into broadband access.
vendors. SDAN can encompass any type of broadband and
The SDAN virtualizes access-network control and management
functions for broadband access, to enable network extend into home network control and diagnostics, and even
optimizations, streamline operations, and encourage innovative into heterogeneous networks.
services creation, particularly in multi-operator environments. Elements of the SDAN concept are already advancing.
This paper identifies software-definable control and ETSI NFV is examining access network virtualization,
management functions for broadband access, and presents including moving complex processing from DSL Access
some specific network optimizations using the SDAN. Multiplexers (DSLAMs) into the network, and multi-
tenancy [2]. The ITU-T standardizes management primitives
Keywords-Broadband, SDN, NFV, Vectoring, DSL for optical and copper access, and the ITU-T SG11/Q4
I. INTRODUCTION Q.SBAN project is working on scenarios and signaling
requirements for software-defined Broadband Access
Software Defined Networking (SDN) virtualizes Networks (SBAN). The Broadband Forum (BBF) is active
configuration and control of network elements (NE), in SDN and NFV for access networks. In the UK, the
generally with a controller [1] located in the cloud or data Network Interoperability Consultative Committee (NICC)
center communicating with network elements via a protocol has initiated a study on sharing data for the purpose of DSL
such as OpenFlow. SDN focuses on control of network Dynamic Spectrum Management (DSM) [5][6]; potentially
routing and switching, and touches on network an SDAN component. The recent ATIS report on SDN and
management; but the SDN concept is expanding. Network NFV provides an overview of other related standards [7].
Functions Virtualization (NFV) will migrate many network This paper summarizes and extends these concepts.
functions into servers and the cloud, including access (see
use case 9 in [2]). SDN and NFV are useful for multi- A. SDAN Architecture
provider networks [3][4].
There are many control and management functions Management
required for broadband access, which, similar to SDN, can Controller A System Controller B
migrate from embedded firmware in dedicated network
equipment into software controllers running on commodity
hardware in a private or public cloud, saving costs while
Common Interface
increasing capabilities. This paper extends SDN concepts
into the realm of broadband access by presenting the
concept of the Software-Defined Access Network (SDAN). SDAN Abstraction Layer
The SDAN is built on a common control plane that Proprietary
virtualizes the infrastructure, separating the control plane Interfaces
from the data plane. The SDAN provides a platform to
implement many types of network optimizations. Network
control and management is programmable, which allows Network Elements
open innovation and introduction of new and agile services. Figure 1. SDAN implementation via an abstraction layer.
The SDAN provides a common interface and a unified
touch point for policy, control, and management. SDAN uses a logical “controller” located in the cloud,
Competitive multi-operator broadband is common in and moves control and management functions from NEs
many territories, particularly for DSL services. into the controller. SDAN also provides a common interface
Virtualization of broadband access control and management to the controller functions that can be accessed by multiple
in multi-operator environments is a theme of the SDAN. operators in competitive environments. This interface can be
Competitive environments generally have a wholesale constructed via an abstraction layer consisting of adapters
network provider responsible for the underlying between existing management and control interfaces such as

978-1-4799-3001-2/14/$31.00 ©2014 IEEE

SNMP and a standardized common interface, MIB, or data virtualization are extended into control, diagnostics, and
model as shown in Figure 1. optimization of access networks and premises networks.
A common interface may also be explicitly Local-loop unbundling (LLU) using DSL has pushed
standardized as a set of messages, schemas, or APIs. Data down prices and enabled broadband proliferation in many
and commands can cross the common interface in real-time. countries. With new super-fast broadband rollouts, fiber
A retail service provider can request data or control actions; penetrates deeper in the access network, with vectored
these are then interpreted, translated and executed through VDSL deployed from cabinets for only the last few hundred
the SDAN controller and abstraction layer. The abstraction meters. It is economically challenging to deploy multiple
layer may also limit the frequency or numbers of messages small DSLAMs at such cabinets, making LLU difficult. The
and the admissible ranges of values. situation is similar with fiber-to-the-premises (FTTP) or
Implementation may be centralized or distributed. In cable architectures with small nodes. The cost and
scenarios with multiple retail service providers, a innovation benefits of competition could soon disappear.
centralized SDAN architecture divides the control and However, the SDAN can virtualize the physical access
management systems with multi-tenancy, abstracting the network infrastructure, allowing retailers to use virtual
access network into multiple logical access networks. A unbundling with nearly the same levels of diagnostics,
distributed SDAN architecture shares data and control control, and optimization as physical unbundling.
functions through a standardized common interface.
Separation of Competition with
Retail provider B
Control and Data Infrastructure
Management Systems
Retail provider A
Management Planes Monopoly
Services
Systems
Controller
Cloud Virtual
Common Interface Resources Unbundling
SDAN
Arbitration

Data on
Abstraction Layer Abstraction Layer Network Services
Controller Authorization
performance,
status, test, Services Request management
Control
and Optimizations Innovation
and diagnostics, Central Configuration

Wholesale
Database
Real-time
provider
Equipment Equipment
Consumer
or third party
Interface 1 Interface 2 Service Creation
Involvement

Data Plane
Figure 4. Drivers behind the SDAN concept.
Access Nodes (ANs)
Consumers can be empowered by opening SDAN
functions through a consumer-device interface such as an
CPE
intuitive app, which releases consumer choice of services
Figure 2. SDAN for multi-operator environments. and service qualities. Consumers can further be informed
about their service quality, and can then rebalance their
Figure 2 expands Figure 3 to depict how the SDAN can service choices in an informed positive feedback loop. As
enable disparate retail providers on a single wholesale consumers become more absorbed by Internet services, they
architecture. A logically centralized system performs want high-level diagnostics on their connections, and
orchestration, authorizes and arbitrates requests for data and especially want to rapidly resolve service affecting troubles.
control, and implements an abstraction layer that interfaces With a simple interface, consumers can be enticed into
between equipment interfaces, the centralized functions, and value-added and real-time services, such as requesting a
the common interface used by the retail providers. In Figure temporary speed boost.
2, retail provider A allows its service to be defined and
controlled by the wholesale provider, while retail provider B III. SDAN BENEFITS
defines services and performs control itself. Network optimizations and operational improvements
offered by the SDAN benefit wholesale and retail operators
II. CONFLUENCE
by increasing the overall broadband footprint. SDAN-
A number of trends are converging to gestate and empowered operators can increase their total number of
sustain the SDAN, as depicted in Figure 4 and described access lines (including unbundled) to boost revenue, and
further in this section. Cloud infrastructure can be leveraged also lower their costs.
to support many network functions with low-cost cloud Streamlined and automated operations provide benefits
computing in architectures which also support more to wholesale infrastructure providers, retail service
advanced network configuration and management. providers, vendors, and consumers. OpEx is lowered for the
Economies of scale offered by cloud resources and server providers, a standardized common interface simplifies
vendors’ requirements, and most importantly consumers can

2
get better service and benefit from unlimited possibilities in optimization to approximately double the speeds of the non-
applications innovation. The SDAN enables new service vectored lines.
concepts to be easily trialed and implemented, allowing
innovation and creativity to flourish. IV. SDAN NETWORK ELEMENT CONTROL
Multi-line optimizations using shared broadband data A number of control functions that are currently
enhances the performance of both incumbent and performed by broadband NEs can be migrated into the
competitor’s lines, and enhanced maintenance improves SDAN controller. This includes relatively slowly-varying
customer satisfaction. Multi-line optimizations include scheduling, administrative, and policy functions as well as
bandwidth and class-of-service assignments on PON or real-time control and optimization.
cable modem networks. A common interface can allow real- Figure 6 gives a simplified view of broadband access
time access to performance monitoring and fault data. Faults network elements. The aggregation network is the part of
that occur simultaneously on multiple lines owned by the network connecting the Access Nodes to the BNG. The
different providers can be correlated and fixed with a single Access Node (AN), such as a DSLAM, Optical Line
dispatch, instead of fixing each line separately. Terminal (OLT), or Cable Modem Termination System
(CMTS); is the network termination for the last mile
A. DSM Example
connection to the customer.
With DSL or other systems using copper transmission,
sharing data among multiple operators can enable network
optimizations via Dynamic Spectrum Management (DSM) Private Internet
[5][6] that often double or triple speeds. Real-time Backbone
variations can be correlated on multiple lines, for example a Controller,
DSL sending higher power may correlate to a neighboring Management
DSL receiving crosstalk that causes errors. An example of Systems
DSM optimization with the SDAN is presented in Figure 5. Broadband Network
Gateway (BNG)

Average bit rate, Non-vectored lines

50
Ethernet
45
Aggregation
Downstream Bit Rate (Mbps)

40
Switch
35
30
25 Access
20 Node (AN)
15 With SDAN
10 No SDAN Residential
5 Gateway (RG)
0
300 325 350 375 400 425 450 475 500 525 550 575 Customer Premises
Loop Length (m) Equipment (CPE)
Figure 5. Example of DSM optimization with the SDAN.
Figure 6. Broadband access Network Elements (NE).

In this DSM example, downstream bit rates of VDSL2 A. Virtualized access network functions
were calculated via simulation with and without an SDAN
architecture implementing DSM on the basis of data shared Network Functions Virtualization (NFV) uses
between the vectored and non-vectored lines. This assumed commodity hardware to implement Virtualized Network
VDSL2 Profile 17a, 0.5 mm cable, and a single network Functions (VNFs) through software virtualization
endpoint. The line lengths are uniformly spread from 300m techniques. Features can be software-defined and re-use of
to 575m, with up to 25 lines in the cable binder, and an service components with “service chaining” allows rapid
average of 15% of lines in the cable are active and are changes in service definitions. Upgrading network elements
equally likely to be vectored or non-vectored. With SDAN, via software decreases hardware obsolescence and is
the lines share data and participate in joint optimization of operationally easier to implement. Many functions of
transmit spectra using the Iterative Waterfilling (IWF) broadband networks such as authorization, advanced
optimization technique [5]. Without SDAN the non- diagnostics, setting forwarding rules, etc. can be virtualized.
vectored lines are limited to transmit only below 2 MHz, the B. Aggregation Network
maximum static spectrum allocation that ensures the
Broadband access aggregation networks are currently
vectored lines achieve 100 Mbps at 575 m. Figure 5 shows
closed systems managed by a single provider that are glued
that using the SDAN for DSM data sharing allows network
together with static layer 2 configurations. The SDAN can
dynamically manage these settings and allow multiple

3
providers to create new networking methods. Broadband performance of the premises network and devices. The
aggregation networks generally employ some type of layer- network quality can be managed from the customer’s end of
2 logical network separation such as stacked virtual local the line, which can ease broadband self-install. CPE can be
area networks (VLANs) or tunneling, which can be directly specifically enhanced to extract diagnostics data from the
subsumed by SDN configurations. SDN can enable retailers customer premises, interact with the SDAN, and automate
to lease various quantities of backhaul bandwidth at various re-configuration.
SLA levels. Bandwidth per subscriber and bandwidth per Wireless premises networks make good examples. A
DSLAM are selectable with defined QoS levels. cloud-based controller can assign resources such as
frequency bands and time slots to femtocells, small cells,
C. Broadband Network Gateway (BNG)
and base stations; coordinating resource assignments across
The BNG mediates between the aggregation network such heterogeneous networks. Resources can be controlled
and other networks such as the Internet, or service-specific in near real-time, with tradeoffs between users managed to
networks such as voice or video The BNG is an IP edge ensure fairness. WiFi can similarly be controlled, with
router where bandwidth and QoS policies may be applied. channel assignments and even station associations optimally
BNG virtualization, using virtualization to offer multiple allocated across multiple WiFi access points, similar to the
logically separate BNG instances to multiple retail way closed enterprise networks can be controlled now.
providers, has already been trialed. CPE can also apply cross-layer optimizations and cross
D. Access Node domain (access and premises networks) optimizations for
further improvements.
The Access Node - the DSLAM, OLT, CMTS, etc. – is
at the heart of broadband access. With SDAN, each retail
V. SDAN AND NETWORK OPTIMIZATION
service provider can use their own virtual access node.
Virtualizing the access node would be similar to server Network optimizations can be supported by the SDAN, with
virtualization, allocating virtual machines under the some examples encompassing:
supervision of a hypervisor. Each virtual access node could • Automated fault and performance diagnostics and
only consume a limited amount of resources (physical ports, reconfigurations to lower wholesalers’ operational
processing, network bandwidth, etc.) which is allocated so expenses.
that they don’t conflict with each other. Virtualized access • Diagnostics by the consumer.
nodes, however, may only appear in the long term. An • Creation of real-time services such as bandwidth on
abstraction layer between the access nodes and management demand.
systems could separate network resources between retailers • Services differentiation, services innovation, and
similar to a virtualized access node. service chaining.
E. Premises networks, CPE, and Consumers • DSL Dynamic Spectrum Management (DSM)
With “wires-only” a retail provider need lease only a • Fiber to the distribution Point (FTTdp) control
DSLAM port and the copper loop from a wholesaler; the functions
CPE and other broadband elements, functions, and This paper applies an example of network optimization
management can all be provided by the retailer. Using wires to the last element of this list. FTTdp architectures are
only in conjunction with the SDAN migrating management emerging that use metallic transmission only over the last
and control into a commonly accessible data center, can few hundred meters from a Distribution Point Unit (DPU),
allow virtual unbundling to operate in a way that is nearly extending fiber nearly to the customer while avoiding the
indistinguishable from physical unbundling. considerable cost of installing fiber into a customer
The residential gateway performs a wide range of premises. The DPU is a very small, low-power device which
functions, and some of these can migrate back into a “virtual needs to be energy efficient. Computationally complex
CPE” located in the network. Security functionality, control and management functions should be performed
firewall, enhanced diagnostics, and some traffic remotely instead of in the DPU, which can be supported by
conditioning can be performed in the virtual CPE; as well as the SDAN through an NFV abstraction layer [2]. Examples
enhanced services, such as parental control or virtual PBX. of such functions include service assignments, QoS policies,
Premises networks and devices are increasingly filtering, multicast group control, dynamic address
complex and difficult for the consumer to manage. provisioning, authentication, authorization, and accounting.
Automated diagnostics are needed, especially for self- Virtualization can also support multi-tenancy and enable
installation processes. One method of providing them is via virtual unbundling of some functions.
a smartphone app that communicates with the CPE to
extract diagnostics data from the CPE. The app then also A. Example: Vector precoding for FTTdp
communicates to a cloud-based SDAN system that analyzes The emerging ITU-T G.fast standard [8] is assumed for
the data and provides guidance, or automated re- FTTdp here; this uses multicarrier modulation but here each
configuration, to assist in repairing or enhancing the subcarrier is treated separately. Transmission across the

4
MIMO crosstalk channel is optimized by using a non-linear The above computations may lead to significant energy
precoder for crosstalk minimization (based on the structure increases at the precoder output, which are limited by
of the generalized decision feedback equalizer, GDFE) applying a modulo operation to each dimension of the
[9][10]. The input x is an L×1 vector representing the constellation, similar to Tomlinson-Harashima precoding.
transmit data sample at a single subcarrier. Each element xj is The modulo operation, as discussed in [10], may be
a complex value that represents one of the constellation expressed as follows:
points for the specific sub-carrier with index j. ⎢ x + d ⋅ M 1d / 2 ⎥
The channel matrix is modeled by H = H d ⋅ (I + C) , ΓM1d ( x) = x − d ⋅ M 1d ⋅ ⎢ ⎥
where H d is a L×L diagonal matrix whose elements define
⎣ d ⋅ M 1d ⎦
where M1d is the number of points per dimension of the
the direct channel frequency response for each of the L wire constellation, and d is the distance between constellation
pairs in the channel, I is an L×L identity matrix, and C a points in one dimension. When applying modulo arithmetic,
matrix with zero-value diagonal elements, and off-diagonal the precoder output samples may be expressed as
elements representing the normalized crosstalk channel x'1 = u1
coupling coefficients, and L is the number of lines in the
vector group (MIMO channel). The component (I + C) is ⎡ r ⎤
x'2 = ΓM 2 ⎢u 2 − 21 x'1 ⎥
the normalized channel matrix. r22 ⎦
Figure 7 shows the channel model for vectoring ⎣
operation with a non-linear GDFE. The GDFE is constructed #
using QR factorization of the complex conjugate of the
⎡ r r ⎤
channel matrix as described in [10]. The conjugate transpose x' L = ΓM L ⎢u L − L , L −1 x' L −1 − " − L,1 x'1 ⎥
of the normalized channel matrix is factored into two ⎣ rLL rLL ⎦
*
matrices, i.e. ( I + C) = QR , where R is an L×L upper where Mi, i = 1,…,L, identifies the per dimension
triangular matrix and Q is a unitary matrix, where constellation size for which modulo operation is to be
applied, ui are the end-user data samples, and rij are the
Q*Q = I . elements of the lower triangular matrix R*.
The GDFE precoder is formulated as follows. The
channel matrix output is n
u x' x y
y = Hx + n . y' ŷ û

y ' = H d−1y
y ' = ( I + C) x + n ' Figure 7. MIMO Channel Reference Model with GDFE.

ni
where ni′ = . By defining x = Qx ' , where x' is an L×1
hii The receiver input samples yi (channel output samples)
vector of the precoder output, and by observing that for line i are
(I + C) = R *Q* , it is seen that ⎡ i −1 rij ⎤ i −1

y ' = R * x ' + n' .

y 'i = rii ΓM i ⎢ui −
⎢⎣
∑
j =1 rii
x' j ⎥ +
⎥⎦
∑ r x'
j =1
ij j + n 'i

where R * is lower triangular. With respect to the end-user In each of the receivers, modulo operation is applied to
data vector u , the channel crosstalk is cancelled when y'i
*
the equalized sample as follows:
Su = R x' rii
where S is an L×L diagonal matrix whose elements are the ⎡ y' ⎤
diagonal elements of R * . The sample vector x' is obtained yˆ i = ΓM i ⎢ i ⎥
⎣ rii ⎦
through the inversion of the lower triangular matrix R * to
formulate the following precoder operation on the user data ⎡ ⎡ i −1 rij ⎤ i −1 rij n 'i ⎤
vector u , namely = ΓM i ⎢ΓM i ⎢ui −
⎢⎣
∑j =1 rii
x' j ⎥ + ∑r x' j +
rii ⎥
⎥
x'1 = u1 ⎣⎢ ⎦⎥ j =1 ii ⎦
r21
x '2 = u 2 − x'1
r22 As discussed in [10], observing
# that ΓM [ x ± y ] = ΓM {ΓM [ x] ± ΓM [ y ]} , the output of the
modulo operation is
rL, L −1 rL,1
x' L = u L − x' L −1 −" − x'1
rLL rLL

5
 ⎡ VI. SUMMARY AND THE ROAD AHEAD
n' ⎤
yˆ i = ΓM i ⎢ui + i ⎥ Network capacity and complexity are rapidly
⎣ rii ⎦ increasing, and the cost of cloud-based computing and
storage in large-scale data centers is decreasing with similar
The decision is made on the above sample. rapidity. Broadband service offerings and customer
The performance of each line varies with the line broadband behavior are becoming increasingly
ordering, since the diagonal values of the R* matrix do not sophisticated. Computation and storage on network
have the same value; and so the received SNR is not the elements is becoming prohibitively expensive by
same on each line. This results in unequal performance, so comparison, and network control functions will migrate into
that the ordering of the lines in the vector affects the the cloud wherever feasible. The SDAN leverages these
performance of the lines. The lines can be re-ordered simply trends to lower the costs and increase the capabilities of
by multiplying the input vector by a permutation matrix. In broadband network control. SDAN also lowers OpEx by
this way, the bit rates assigned on the different lines can be minimizing the number of “touch points” needed to manage
varied. the network.
An example was evaluated, for G.fast running on 10 This paper discussed the SDAN in general and
lines, transmitting at -76 dBm/Hz from 23 MHz to 212 MHz presented one network optimization scenario using the
on 51.75 kHz subcarriers, and with -140 dBm/Hz SDAN in detail. Many more types of network optimization
background noise. Measurements of the channel matrix of a could be performed using the SDAN; these should be
100 m cable taken by BT [11] were used here. Full identified and explored by creative researchers.
knowledge of the channel matrix was assumed with perfect
cancellation. Each simulation run performed 10,000
different orderings, and the maximum, average, and REFERENCES
minimum bit rate achieved by each line with all these re- [1] Open Networking Foundation,
orderings were found to bracket the range of assignable https://www.opennetworking.org/index.php.
speeds. [2] ETSI GS NFV, “Network Functions Virtualisation (NFV); Use
Cases,” V1.1.1, 2013-10.
Figure 8 plots a dozen cases of bit rates with a dozen http://www.etsi.org/deliver/etsi_gs/NFV/001_099/001/01.01.01_60/g
particular orderings, and also shows the minimum (696 s_NFV001v010101p.pdf
Mbps) and maximum (1215 Mbps) bit rates found from all [3] Sezer, S.; Scott-Hayward, S.; Chouhan, and P.K.; Fraser, B. “Are we
10,000 different orderings. ready for SDN? Implementation challenges for software-defined
networks,” IEEE Communications Magazine, Volume:51, Issue: 7,
July 2013
[4] S. Davy, et. al., “Challenges to Support Edge-as-a-Service,” IEEE
Communications Magazine, pp 132-138, January 2014.
[5] ATIS Std. 0900007 (2012), Dynamic Spectrum Management
Technical Report, Issue 2.
[6] NICC ND-1513 (2010), Report on Dynamic Spectrum Management
(DSM) Methods in the UK Access Network.
[7] Operational Opportunities and Challenges of SDN/NFV
Programmable Infrastructure, ATIS-I-0000044, 2013
[8] ITU-T draft G.9701, “Fast Access to Subscriber Terminals (FAST) –
Physical layer specification.”
[9] Shailendra Singh and M. Sorbara, “G.fast: Parameters Specific to
Figure 8. Precoder performance with different line orderings for G.fast.
Linear and Non-Linear Pre-coding for G.fast,” ITU-T contribution
2013-03-Q4-074, March 2013.
There is 75% speed increase from the minimum to the
maximum bit rate. However, it is not known a priori how the [10] G. Ginis and J. M. Cioffi, “A multi-user precoding scheme achieving
crosstalk cancellation with application to DSL Systems,” the Thirty-
ordering changes the bit rates; this can only be found by
Fourth Asilomar Conference on Signals, Systems, and Computers,
varying ordering and performing the computations as shown October 29 – November 1, 2000.
here. With SDAN, precoder coefficients can be computed
offline, and a search can then match user requirements to [11] BT, “G.fast: Release of BT cable measurements for use in
simulations,” ITU-T Q4a/15 Contribution 2012-11-4A-034, Chengdu,
assigned line speeds. Such an assignment can only be China, November 5-9, 2012.
computed and managed with the type of compute resources
available in the cloud via the SDAN; it cannot be performed
by the low-power FTTdp equipment itself. So this example
shows how the SDAN can improve user bit rate assignments
by 75%.