You are on page 1of 5

Voice Delay Considerations for DSL

Multimedia Access Network Design and Operation


S. Wright A. J. Vernon A. Whited
Architecture & Emerging Technology Ridge Consulting Group Inc., Architecture & Emerging Technology
BellSouth, Atlanta, GA Atlanta, GA BellSouth, Atlanta, GA
Steven.Wright@BellSouth.com ajv@RidgeConsultingGroup.com Albert.Whited@BellSouth.com

Abstract—This paper discusses the carrier perspective on through 802.3ad, or at an IP service layer using ECMP—each
baseline capabilities and technology tradeoffs in the capacity with different tradeoffs in terms of efficiency and end-user
planning methods in the context of the evolution of the DSL impact.
Access network. Capacity improvements are being made possible
by new modem technologies and deeper deployments (shorter Given the current industry focus, we assume the service
loop lengths). DSL networks are evolving from a service-specific provider wishes to offer a “triple play” suite of voice, data, and
(Internet Access) infrastructure to one that must support entertainment video services over a DSL infrastructure. In this
multiple services. Emerging multimedia service requirements paper we consider one illustrative deployment scenario where
(e.g., for voice) are considered from the perspective of the broadcast entertainment video is channelized separately from
additional design constraints (e.g., delay) that impact DSL the DSL access network (e.g. via satellite, hybrid Fiber Coax,
network design and its operation. Fiber to the Curb, MMDS), and video on demand or PVR
video content is delivered via a “best-effort” data service (e.g.
Keywords-DSL, NGN,QoS, Delay, Capacity Planning tcp, udp). Two QoS classes are assumed, with voice receiving
either preemptive or non-preemptive priority over best-effort
I. INTRODUCTION data.
Service providers, equipment vendors and others are The paper is organized as follows. Section II reviews some
expending significant effort to develop both a vision statement of the critical traffic demand considerations, focusing on delay
and an architectural framework for Next Generation Networks issues. Section III discusses some of the implications for DSL
(NGN) and services, resulting in standardization efforts in network capacity design, primarily through the consideration of
bodies such as ITU-T, 3GPP, ETSI, and ATIS (see e.g. [3]). VoIP as a multimedia service component. Conclusions
The NGN envisages a multi-provider, multi-access, QoS- regarding DSL link configuration, CODEC selection options,
enabled service environment based on IP connectivity, and implications of alternative broadcast delivery strategies are
including existing and evolving DSL network infrastructure. discussed in Section IV.
The NGN standards work introduces the notion of an IP
Connectivity Access Network (IPCAN) to provide access to II. DSL TRAFFIC DEMAND CONSIDERATIONS
NGN services. These NGN notions introduce additional
functionality to DSL access networks (e.g. Quality of Service A. DSL Network Delay Performance Objectives
(QoS), mobility and a Session Initiation Protocol (SIP) control
plane based on the IP Multimedia System (IMS)). While DSL network performance, as expressed through Quality of
IPCAN conformance criteria are still under discussion, DSL Service (QOS) metrics must be translated into appropriate
networks are expected to evolve in compliance. Quality of Experience (QOE) metrics to determine whether the
end user requirements are met. Mean Opinion Survey (MOS)
Recent work at the DSL Forum [1] has created a focus on scores are the accepted QOE metrics for voice and the E-model
the development of architectures and the selection of provides an empirical mechanism to translate delay metrics into
appropriate mechanisms to support QoS-enabled IP services. MOS scores [9]. The primary focus of this QoS analysis is
While future improvements to the bandwidth available on the delay. ITU Y.1541 [10] identifies end-to-end network
DSL line (e.g., ADSL, ADSL2, ADSL2+, VDSL2) may be performance objectives for IP networks, including delay
deployed by operators and may appear to diminish the need for parameters, and Y.1541 Appendix X provides some example
QoS mechanisms, any deployed bandwidth tends to be speech quality calculations for packet voice calls, albeit with a
consumed [2]. While strict QoS guarantees are not envisioned different reference model than this document. ITU G.114
in this environment, the QoS capabilities required are expected provides planning values for the delays associated with various
to address the stricter performance requirements for latency- transmission elements in the PSTN [8]. G.114 Appendix II
and jitter-sensitive multimedia services (e.g., Voice over provides guidance on one-way delay for packet voice systems
Internet Protocol (VoIP) and gaming [13]). Deeper for the intra-and inter-regional cases; however, it does not
deployments (that is, shortening copper loop lengths) and loop make recommendations on apportioning this delay. In the
bonding provide additional mechanisms to improve DSL regional case (defined to have a diameter less than 5000 km,
capacity. Bonding can occur at multiple levels in the protocol such as North America) the “mouth-to-ear” one-way delay
stack, for example, at the modems with G.Bond [11], at the should be less than 150 ms. In this paper we assume the voice
BRAS through MLPPP, at an Ethernet switching device call traverses an inter-exchange carrier as well as a local

0-7803-9319-8/05/$20.00 (c)2005 IEEE. 111


exchange carrier at the far end, and allocate the delay objective would limit the peak traffic load that the infrastructure would
equally across the three service providers. This results in a need to support.
50ms delay objective for voice service across the DSL network
(from the voice client to the PSTN gateway, as illustrated in C. Delay Model for DSL Access Network
Figure 4)
A model for the transport of voice in the access network and
While bounds exist for worst-case delays in packets the relevant delay components is shown in Figure 4. The
networks [7], engineering a network based on those bounds model assumes that the voice streams are
would certainly lead to cost-prohibitive over-provisioning. terminated/originated at a (PSTN) gateway or border proxy.
Instead of only considering an upper bound on worst-case Link delays, for example Duplink, include processing overheads
delay, the approach adopted here is based on the stochastic at the transmitting node, if any. In the core IP transport
analysis of the delay bound, as presented in [5], which network, voice transport delay is obviously a function of how
specifically accounts for voice having preemptive or non- the IP traffic is routed; therefore, the index h is used to
preemptive priority over data traffic (for example, voice is represent a router hop (a hop includes the processing and
transported in the Expedited Forwarding (EF) class in the IETF scheduling delays at the transmitting router). From Figure 4 it
DIFFSERV model). Following [5] and their arguments on delays
is quite easy to see that the equation governing the access
and the intelligibility of received speech, we adopt a statistical
network delay (D) of voice packets is:
upper- bound on the maximum delay for the voice services:
D = D pack + Denc + Dhome + DDSL
Pr[Access Network Delay ≥ 50 ms] < 10 −5 (1) (2)
+ Duplink + ∑D
h∈Path
h + Degress + D play + Ddec
B. Traffic Demand Characteristics
The data traffic demand is assumed to be derived from the For transport links, the delay on a link is the sum of the delays
requirement to support Internet access. Although the DSL due to transmission, queuing, interleaving, bonding,
Forum architecture permits multiple data classes of propagation, and node processing. It will be assumed that the
differentiated service traffic, we assume all data traffic is interleaving delay (Ilink), link bonding delay (Blink), propagation
treated as “Best Effort”. delay (Plink) and the node processing delay (Slink) are all
Voice traffic characteristics depend upon the CODEC negligible and thus the delay on a link is assumed to
selection; we consider both G.711 and G.729A CODECs without be Dlink = Tlink + Qlink . Also, to avoid potential packet loss,
the use of voice activity detection. Both schemes use an 8 kHz
the play-out (Dplay) and decode (Ddec) delays at the PSTN
sampling rate and a 20 ms packet formation time, but G.729A
uses 16-bit samples and 16:1 compression ratio, resulting in a gateway are assumed to equal the maximum queuing delay
voice data rate of 8 kbps (compared with 64 kbps for G.711). (Q*) [5]. Obviously, transmission delay (Tlink), and queuing
While this implies significant efficiency, G.729A does incur a delay (Qlink) depend on the length of the packets and the arrival
5 ms look-ahead delay penalty. process of the packets to the link, both of which are functions
of the type of CODEC used. Per the objectives discussed in
Entertainment video is primarily a unidirectional service Section II.A, , the delay must be bounded to ensure that
with significant bandwidth demands depending upon codec application performance is sufficient:
selection. Entertainment services are typically considered as
broadcast or multicast services, but emerging trends towards Denc + D pack + Thome + TDSL + Tuplink
(3)
Video-on-Demand (VoD) and the user oriented content search
capabilities demonstrated by Personal Video Recorders (PVRs)
+ ∑T
h∈Path
h + Tegress + 2Q * ≤ 50ms
seem to be changing the entertainment video service concept
towards a unicast model. Modeling video traffic sources can be D. Analysis and Example Results
fairly complex, (see [12] for a tutorial) depending on CODEC Considering the upstream toward the PSTN gateway, the
selection (e.g. MPEG-4), operational parameters (e.g. video transmission delays are dominated by the transmission delay of
profile), and system level service design choices (e.g. DSL line. In a worst-case scenario (256 kbps upstream data
Multilevel encoding for different video resolutions). rate), the sum of the transmission delays is approximately 10
Entertainment video is also sensitive to errors, with stringent ms. Accounting for CODEC-specific delays, Equation 4 implies
demands for the users QoE resulting in requirements for error the following constraint on maximum queuing delay:
concealment techniques in the codecs and error control
Q * ≤ 
mechanisms (e.g. FEC /Interleave). The predictive nature of 10 ms G.711 (4)
video frame construction in MPEG magnifies the effect of a 7.5 ms G.729
single packet error, potentially damaging multiple video frames
constructed at later time, expanding the error over several Analysis of queuing delay is based on the methodology
frames. presented in [5] and [6]. For worst-case delay analysis, it is
assumed that a voice packet will see the residual transmission
Of course, the traffic demands for the different multimedia time of a 1500-byte TCP payload, in addition to the intra-class
service components supported by the DSL network may be queuing delays resulting from the phase differences between
constrained by admission control – either a generic rate control the voice streams. For G.711 CODECs, Figures 1 and 2 provide
or a session control based on the NGN SIP architecture. This example complementary cumulative distribution functions for

112
the maximum queuing delays on a DSL link and a DS3 to gracefully constrain excess demand, and (b) provisioning
DSLAM uplink, respectively. In both figures, the parameter K capacity based on (i) well-defined traffic characteristics and
denotes the number of active voice streams. Results for demand assumptions and (ii) well-known theory to translate
G.729A CODECs and the other links in the network (see these into capacity requirements (e.g., Erlang tables [16]). As
Figure 4) can be similarly obtained, but those results are an example, consider a DSLAM with 100 lines each supporting
omitted for the sake of brevity. In the next section, these and a VOIP service with demand estimated at 0.1 ERL. At a blocking
similar results are applied to the issue of DSL network design. probability of 0.5%, the uplink capacity required for the 10 ERL
aggregate demand at the DSLAM is 21 circuits. The bandwidth
required for the voice circuits can be calculated using the
applicable overheads for the protocol stacks on the link (see
Figure 4). The bandwidth required for the voice service must
be compared with that available on the physical facility and
that required for other services. In this case we assume priority
for the voice service and elasticity of the best-effort data
service, but some minimum throughput capacity is still
required for the best effort data class. For a typical DS3 uplink,
21 G.711 voice streams (each requiring 106 kbps, including
link overheads) is well within the capacity available. Note that
efforts to shorten DSL loop lengths result in additional stages
of aggregation: if the DSL links were aggregated by an FTTC
implementation with 10 voice lines per node, the voice capacity
required at the node uplinks would be 5 circuits. Aggregation
Figure 1 DSL Link Queuing Delay CCDF: Preemptive characteristics of the data class are beyond the scope of this
(dashed) and Non-Preemptive (solid) paper. Greater dispersal of the aggregation reduces the node
size and consequently impacts the trunking efficiency. The
admission control mechanism should recognize each potential
choke point in the network and admit accordingly for proper
service operation.

B. Delay-Constrained CapacityPlanning for DSL


Bandwidth dimensioning is only a single facet of the
capacity planning process for packet-switched systems: delay-
sensitive applications, such as voice, require that the delay also
be engineered, and it is quite possible that delay may be the
more limiting factor. As an example consider a DSLAM
uplink carrying G.711 voice streams that is allocated 1ms for
voice queuing delay and the probability this budget is exceeded
cannot be greater than 10-5. Using the model and analysis
approach described earlier, Figure 3 plots the voice-stream
Figure 2 DS3 Link Queuing delay CCDF: Preemptive capacity as a function of the link type and illustrates that for
(dashed) and Non-Preemptive (solid) both preemptive and non-preemptive voice priority, delay
constraints are more limiting than bandwidth constraints; in
III. DSL NETWORK DESIGN METHODOLOGY fact, the service may not be feasible on some lower-speed links.
Access networks are blocking aggregation networks: the For packet-voice services, the end-to-end delay
aggregate bandwidth of subscriber interfaces is greater than the requirements must be accounted for when determining the
aggregate bandwidth of core network interfaces. Capacity maximum voice-stream capacity for a link and thresholds for
planning for Internet access services in DSL networks is all the links in the path must be determined. This is
typically based on an assumption of heavily asymmetric traffic complicated by two issues: (i) all links (except the egress link)
dominated by downloads. The asymmetry assumption may be are expected to support traffic other than voice, and the delay
challenged in future by changing data usage patterns, for requirements for these other applications could be much less
example, a growing emphasis on peer-peer communication stringent: in the absence of strict voice priority mechanisms,
applications rather than client-server download applications. both the aggregate utilization of the link and the data traffic
On the other hand, entertainment video services would make characteristics (such as potential self-similarity) must be
the aggregate traffic more asymmetric. addressed; (ii) a capacity planning process based on the sum of
individual link delay budgets equaling the end-to-end delay
A. Bandwidth-Constrained Capacity Planning for Voice requirement can potentially result in over-engineering (since
services over DSL the upper bound on worst-case end-to-end delay is the sum of a
series of random variables).
For the PSTN access networks, congestion was managed by
a combination of factors including (a) using admission control

113
The first issue is readily mitigated by implementing the C. Operational Considerations for Streaming Media over
necessary QoS mechanisms to ensure that voice is sufficiently DSL
insulated from the data traffic; for example, using Expedited Packet networks, by their nature, introduce variations in
Forwarding (EF) class in the IETF DIFFSERV model. From an delay depending on traffic conditions. Variance in frequency
operational perspective, delay-constrained capacity planning or delay variation—jitter—is typically addressed through
must make a trade-off between the complexity and expense of buffering, adding to the end-to-end delay of a service,
the planning process and the resulting capital savings from counteracting the effect of QoS.
optimum asset utilization. The second issue can thus be
addressed by managing this trade-off. As intimated by the Introduction of QoS-based real-time streaming services
partial set of results provided in the previous section, it is clear may necessitate service level agreements (SLAs) or objectives
that (in the upstream direction) queuing delays on the DSL link (SLOs). In both cases, time of day stamping is required with
will dominate. If the delay allocation to the other links could precision in order to establish a suitable measurement
be limited to approximately 1 ms or less, then capacity infrastructure to support the relevant delay metrics. Similarly,
planning process could be reduced to a simple set of devices beyond the domain of the network synchronization
engineering guidelines; for example: distribution may require some network reference in order to
properly ascertain one-way delay, enhance QoS, or provide a
Pr[ q~home ≥ 1 ms] < 10 −5 , timing reference for circuit-packet boundary devices or the
 10 ms G.711 enterprise edge.
Pr  q~dsl ≥  −5
 < 10 , (5)
  7 . 5 ms G . 729  Stable frequency distribution has typically been provided
by physical layer mechanisms in the TDM network. Such
Pr[ q~uplink ≥ 1 ms] < 10 −5 , mechanisms are not always available in packet networks with
H non-TDM interfaces (e.g. Ethernet), and are potentially
Pr[∑ q~h ≥ 1 ms] < 10 −5 expensive to retrofit. The Network Time Protocol and IEEE
h =1
1588 provide mechanisms to distribute Time of Day
Results obtained but not shown indicate that when voice is information, if stable local frequency references are available.
given priority over data on links with speeds at or above a DS3, Tolerances for time and frequency references for the SLA
a 1ms delay budget can be easily achieved without significantly measurement infrastructure are not yet standardized.
limiting voice-stream capacity below the bandwidth-based
constraint; thus, there is no impact on the utilization of those IV. CONCLUSIONS AND FURTHER WORK
assets (see Figure 3 for example capacities). Those results also
indicated that when non-DSL links are constrained to 1ms and In this paper, the implications for the management of voice
network topology limits the number of IP network hops to less delay in a DSL-based multimedia service offering has been
than five, the error that would be introduced by not accounting presented and an approach to capacity planning that makes a
for the end-to-end delay being a sum of random variables is trade-off between the operational expense associated with
negligible (from a network engineering perspective). Delay- complexity and the capital savings associated with optimal
constrained capacity planning can thus be simplified to the asset utilization has been proposed.
form presented in Equation 5. Whereas some assumptions were made to simplify the
number of elements contributing to delay, the methodology
highlights how other sources of delay, such as interleaving and
link bonding, can have a significant impact of the ability to
meet a 50 ms end-to-end delay target. When interleaving is
used, typical DSL equipments offer settings in the range 0-16
Voice Stream Capacity

ms and thus the additional delay cannot be ignored. Link


bonding arrangements such as G.Bond introduce a fixed delay
of less than 1 ms, but higher layer link bonding approaches are
expected to introduce greater delays as well as potential
interactions with link scheduling mechanisms supporting QoS.
Delay-Limited Delay-Limited While G.Bond may have negligible impact on delay,
Bandwidth-Limited
(Preemptive Priority) (Non-preemptive Priority)
alternative approaches may require further analysis.
The impact of preemption on the best effort data traffic
class is also a subject for further study. Alternative approaches
including MTU size adjustments are also being considered.
This paper focused on the capacity planning methods for the
Link Type case of one delay sensitive service (voice) and one delay
tolerant service (best effort). Further analysis is ongoing to
Figure 3 DSLAM Uplink: Delay-Constrained Link
support the multi-service scenarios with multiple-delay
Capacity
sensitive services (voice, video), with different traffic
characteristics. Entertainment services, in particular, raise

114
questions regarding capacity planning methods for supporting [2] S. Wright, T. Anschutz, "QoS requirements in DSL networks",
multicast multimedia streams. GLOBECOM 2003 - IEEE Global Telecommunications Conference,
vol. 22, no. 1, Dec 2003 pp. 4049-4053.
Efforts to improve peak bandwidths by reducing loop [3] ATIS, “ATIS Next Generation Network (NGN) Framework, Part I:
lengths will reduce the node size at the first aggregation point. NGN Definitions, Requirements and Architecture”, Issue 1.0 , Nov.
While efficient network operation is important, traditional 2004, Alliance for Telecommunication Industry Standards.
trunking efficiency metrics may be inappropriate in a multi- [4] S. Wright, H. Fahmy, A.J. Vernon, “Deployment challenges for
access/metro optical networks and services”, IEEE/OSA J. Lightwave
service environment. Rather than simply considering the Technology, Nov. 2004, vol. 22, no. 11, pp. 2606-2616.
bandwidth required at each stage of aggregation, it may be [5] M. Karam, F.A. Tobagi, “Analysis of the Delay and Jitter of Voice
more appropriate to consider distance sensitive metrics for Traffic over the Internet, INFOCOM 2001.
network efficiency. [6] S. Sharafeddine, N. Kongtong, Z. Dawry, “Capacity Allocation for
Voice over IP Networks Using Maximum Waiting Time Models”, ICT
In terms of future work, we note that in certain 2004.
circumstances the delay characteristic of service controls may [7] A. Charny, J.-Y. Le Boudec, “Delay Bounds in a Network with
actually be as or more constraining than the delay characteristic Aggregate Scheduling”, First International Workshop on Quality of
of the service packets themselves. For example, calls for Future Internet Services, Berlin, Germany, September 2000.
changes in per-flow policy may result in abandoned sessions or [8] ITU-T Recommendation G.114 (2003) "One-Way Transmission Time".
web-sites if the set-up time of an enabling policy is too long. [9] ITU-T Recommendation G.107 (2003) "The E-Model, a computational
Similarly, too long a network response to requests for joining model for use in transmission planning".
or leaving multicast streams may generate a very negative [10] ITU-T Recommendation Y.1541 (2002) "Network Performance
impression among subscribers. We note that anecdotally Objectives for IP-based services".
multicast video services hold as an objective channel change [11] Draft ITU-T Recommendation G.Bond, work in progress.
response times of less than one second. Channel change delays [12] P. Seeling, M.Reisslein, B.Kulapala, “Network Performance Evaluation
in excess of 1.5 seconds often result in a lowered MOS score using Frame Size and Quality Traces of Single Layer and Two-layer
for stream QoE and, subsequently, elevated subscriber churn. video: A tutorial”, IEEE Communications Surveys, 3rd Q. 2004, Vol. 6,
No..3, pp 58-78.
Whether there is a capacity constraint implied by such a
[13] S. Wright, S. Tischer, "Architectural Considerations in Online Game
signaling delay constraint is a subject for further study. Services over DSL Networks", in Multimedia Technologies and Services
Symposium, ICC 2004, Paris, France, June 2004.
ACKNOWLEDGMENT [14] IEEE Std 1588-2002 “IEEE Standard for a Precision Clock
Synchronization Protocol for Networked Measurment and Control
The authors wish to acknowledge the many helpful Systems”
comments and insights provided by colleagues and reviewers. [15] D.Mills “Network Time Protocol (Version 3) Specification,
Implementation and Analysis”, March 1992, RFC 1305
REFERENCES [16] J.Bellamy, “Digital Telephony”, 2nd Ed., Appendix D Traffic Tables, p.
541 J.Wiley & Sons, Inc., New York NY.1990
[1] DSL Forum, “DSL Evolution – architecture requirements for the support
of QoS enabled IP services”, TR-59 Rev. 1, Sept. 2003

Ddec Dplay Degress


∑D
h∈Path
h
Duplink Ddsl Dhome Denc+ Dpack

Gateway Client
Play-out
De- CO LAN DSLAM DSL Home
Decode IP DSLAM Modem Network Packetize Encode
packetize (Egress) Router Router Uplink Link
Buffer Network

Central Office Central Office Customer Premise

G.711 G.711
RTP 8 bytes 8 bytes RTP
20 bytes 20 bytes
UDP 12 bytes 40 bytes 12 bytes UDP
IP 20 bytes IP 20 bytes IP 20 bytes IP 20 bytes IP
MPLS (4 bytes) MPLS
L2TP 8 bytes L2TP PPP 2 bytes PPP PPP 2 bytes PPP
PPPoE 6 bytes PPPoE
802.3 MAC 29 bytes 802.3 HDLC 12 bytes HDLC HDLC 12 bytes HDLC Ethernet 16 bytes Ethernet 802.3 29 bytes 802.3 MAC
RFC 2684 8 bytes RFC 2684

AAL5 8 bytes AAL5


ATM 65 bytes ATM
PHY PHY PHY PHY PHY PHY PHY PHY PHY PHY
69 bytes 54 bytes 54 bytes 105 bytes 69 bytes

Figure 4: Voice Transport Model and Transport Protocol Overhead

115

You might also like