Analyze an academic article.
nasstuxa
65035253.pdf
◆ Quantifying QoS Benefits in the Aggregation Network for Internet Video Services Bilgehan Erman and Z. John Zhao
As consumers enjoy new streaming media services from the Internet, network providers are being left with the bill for managing the traffic load that is growing at an explosive rate. For network providers, increasing the network capacity will be necessary, but it is often misleading to think that capacity increase alone can help deliver these new services cost effectively. Internet video and online gaming services can create revenue opportunities, but quality of service (QoS) assurance becomes an important requirement for offering these services with a service level agreement (SLA) commitment to both business partners and residential subscribers. In this paper, we first explain the network optimization challenges faced in the aggregation network related to the Internet-based service delivery scenarios, and then we describe our traffic model that quantitatively demonstrates the value of implementing dynamic QoS and deep packet inspection to enable new service offers and reduce capital expenditure in network capacity expansion. © 2011 Alcatel-Lucent.
residential broadband traffic load. Delivery of high
definition (HD) IP video at 3 Mb/s to 8 Mb/s is likely to
further increase the average residential broadband
traffic, which is now on the order of several hundred
kilobits per second. Unlike the IPTV that uses multi-
cast, Internet TV is often delivered as individual uni-
cast flows that linearly increase the total traffic
volume in the aggregation and the core networks. On
the other hand, it is quite uncertain what the typical
bandwidth usage per household will be even in the
near future. Infonetics Research points out that video
traffic is increasing anywhere from 50 to 100 percent
annually, according to network providers [9]. Many
interdependent variables will affect subscribers’ take
rate for these new video services such as how soon
Introduction Up until recently, watching television (TV) pro-
gramming relied mostly on terrestrial broadcasts, cable
networks, and satellite links. Advances in technology
now allow delivery of TV programming via the
Internet as Internet Protocol (IP) packet streams.
Although network provider–owned IP television
(IPTV) platforms have been in operation for some
time, watching TV from the Internet using ordinary
broadband data services has become a new trend only
recently.
As further discussed in the next section, there are
many types of video services delivered through the
Internet. In the context of this study, our focus has
been on the Internet TV service. Increasing use of
Internet TV is expected to have substantial impact on
Bell Labs Technical Journal 16(2), 63–78 (2011) © 2011 Alcatel-Lucent. Published by Wiley Periodicals, Inc. Published online in Wiley Online Library (wileyonlinelibrary.com) • DOI: 10.1002/bltj.20503
64 Bell Labs Technical Journal DOI: 10.1002/bltj
content producers’ and distributers’ business models
will shift towards online offers, how soon consumer
devices will widely support Internet TV, and how
cable and telecom companies will maintain TV ser-
vice fees. Since we are in the early phase of this new
trend, it is challenging for network providers to antici-
pate what the growth rate of broadband traffic passing
through their network will be in the coming years.
When the growth rate is unpredictable, then plan-
ning and sizing the network capacity become fairly
difficult.
Although the data traffic carried over the network
is increasing because of Internet video services, this
does not translate into a revenue increase for network
providers [13], which are also trying to secure their
place in the new ecosystem. There are indeed oppor-
tunities for network providers to increase quality of
experience (QoE) for consumers and to partner with
video content providers (VCPs) for services like con-
tent caching. If network providers get involved in the
delivery of Internet TV as a service offer, then control-
ling quality of service (QoS) for these media flows
becomes a necessity for managing high QoE. Controlling
QoS also becomes an important tool in battling the
uncertain rate of traffic growth. However, another chal-
lenge that network operators face is the publicly
debated perception that applying QoS treatment to
Internet services will infringe upon free and fair
Internet access for all and thus impede “net-neutrality.”
New regulations are expected to clarify the boundaries
and eliminate the industry-wide confusion.
Network providers will eventually have to
increase network capacity in response to anticipated
traffic growth. However, when the rate of traffic
growth is uncertain, determining adequate capacity
increase is challenging. Minimizing required network
capacity while delivering high service quality is not
an easy task for network providers. There are many
published works that study broadband traffic charac-
teristics [6, 15], QoS management [12], and also net-
work optimization for capital expenditure (CAPEX)
savings [3]. But it is hard to find published studies
that look at the investment impact of QoS control of
IP broadband traffic in the aggregation network. In
the book The Competitive Internet Service Provider [7],
author Oliver M. Heckmann provides a detailed
capacity expansion model for optimizing network effi-
ciency against a target QoS level. The model is based
on a periodic network capacity expansion plan where
congestion targets are set for each link and link capac-
ity is doubled whenever a pre-set threshold is crossed.
The model incorporates the investment capital cost
for link upgrades and a parameter representing
cost of congestion among other link capacity related
parameters. That cost of congestion parameter is likely
to pose the biggest challenge for use of this type of
Panel 1. Abbreviations, Acronyms, and Terms
3GPP—3rd Generation Partnership Project ADSL—Asynchronous digital subscriber line BNG—Broadband network gateway CAPEX—Capital expenditure DES—Discrete event simulation DPI—Deep packet inspection DSL—Digital subscriber line DSLAM—Digital subscriber line access
multiplexer FTTH—Fiber to the home HD—High definition HSI—High speed Internet IP—Internet Protocol IPTV—Internet Protocol television
ITU—International Telecommunication Union NPV—Net present value QCI—QoS class identifier QoE—Quality of experience QoS—Quality of service SLA—Service level agreement TCP—Transmission Control Protocol TS—Technical Specification TV—Television UDP—User Datagram Protocol VCP—Video content provider VDSL—Very high speed digital subscriber line VoD—Video-on-demand VoIP—Voice over IP
DOI: 10.1002/bltj Bell Labs Technical Journal 65
optimization model. For example, the following ques-
tion is not trivial to answer for network providers:
“What is the cost differential of subscribers experi-
encing 0.01 percent versus 0.001 percent packet loss
in an hour for HD video service?”
The value proposition presented in this paper
explains
• Why use of dynamic QoS control can help mini-
mize required network capacity,
• Why dynamic QoS control is essential to provide
service level agreement commitments, and
• How CAPEX savings from the use of dynamic
QoS control can be quantified.
The value proposition we present contrasts five
year capital expenditures for network capacity expan-
sion versus network cost savings due to dynamic QoS
control. The value proposition is derived in three
steps: 1) quantify traffic congestion in terms of QoS
metrics, 2) associate QoS metrics with required level
of link utilization, and 3) estimate required capital
investment for a given level of link utilization. The
quantification of traffic congestion is presented in
the “Aggregation Network Traffic Analysis” section,
and the other two steps are explained in the “Value
Proposition” section. To better explain our approach,
in the next two sections, we discuss the characteristics
of the Internet video services and describe the refer-
ence network architecture for our case study.
Internet-Based Services The realization of new Internet applications and
services can be attributed to an increase in broadband
access line speeds, advances in computer hardware,
new Internet-programming techniques, and improve-
ments in media codec processing. The applications
that rank high with respect to bandwidth usage
include IP video, peer-to-peer, and online storage and
backup. In terms of the impact on network efficiency,
planning, and operations, the most significant appli-
cations are those related to IP video because of their
high bandwidth consumption and tight performance
constraints.
From the end user’s perspective, common sources
of IP video include video clips, online movies, live TV
programs, Webcams, and video conferencing. Popular
examples of these sources include the YouTube*
Internet site, where users can watch short video clips;
Netflix*, a popular source of long-form online movies;
and BBC iPlayer*, which provides a popular catchup
TV service in the United Kingdom (U.K.). Other types
of IP video sources are based on the IPTV platform for
broadcast TV and video-on-demand (VoD) applications.
Table I provides detailed characteristics of IP video applications including the service domain that supplies
the application, its delivered content type, bandwidth
usage, rate of usage, the type of network transport
used, and examples of providers of the service.
The IPTV applications are often tightly integrated
within the network provider’s service domain and
thus carefully planned and engineered by the
provider. IPTV flows of the live TV and the VoD ser-
vices are managed with strict end-to-end QoS con-
trol. These flows are carried with higher priority
compared to the best-effort traffic where the Internet
services described above are transported [1]. As
explained later, our focus in this study is the differ-
entiation of select Internet service flows within the
best-effort traffic class. Therefore the IPTV flows are
excluded in the traffic analysis. For example, network
operators may provision separate link capacity for
IPTV service and the rest of the traffic on the same
physical link.
IP video applications can be classified in two broad
categories with different network performance
requirements: progressive video download and live
video streaming. Progressive video download is a
technique used with bounded, fixed size video con-
tent, such as YouTube clips. When the user plays the
video, the client application starts to download and
buffer the video content. The video does not start
to play until enough content has been downloaded to
avoid further pause during playout. When the video
starts to play, the download continues until the client
receives the entire content file. In this case, the play-
back rate and the video download speed are not
related. Since there is often several minutes worth of
buffering at the client side, it is acceptable for the con-
tent transmission to be bursty, i.e., for a large number
of video packets to be delivered at irregular intervals.
Thus there are no stringent playout timing concerns
for downloaded video in relation to the arrival timing
of transport packets. Transmission Control Protocol
66 Bell Labs Technical Journal DOI: 10.1002/bltj
(TCP) can be used for video transfer in this case with-
out worrying about the impact of data retransmis-
sions on QoE. The typical impairment experienced by
progressive video download is video freezing when
the playout buffer is exhausted. If adaptive steaming
techniques are used, video resolution may also
degrade as the congestion on the line increases.
In contrast to progressive download, live video
streaming intends to deliver a steady stream of
video packets at a fairly constant rate. The video
buffering occurs on the order of seconds in this case.
Therefore, the playback speed is closely tied to the
content transmission rate. With a steady flow of video
packets and small buffer sizes, the variance in band-
width use is much smaller compared to progressive
download. Under these circumstances, video packet
delivery using User Datagram Protocol (UDP), without
data retransmit at the transport layer, improves per-
formance [8], but packet loss then becomes a more
important performance measure, and it translates into
Table I. IP video applications and services.
Application Service Bandwidth Rate of or service domain Content usage usage Transport Example
Video clips Internet Bounded, Medium; bursty; High Progressive download, YouTube � 15 minutes server congestion HTTP, TCP, downlink
is a factor in average bandwidth
Online Internet Bounded, High; bursty; server Rapidly Progressive download, Hulu movies 30 min to congestion and increasing HTTP, TCP, downlink and shows 2 hours transport capacity
are factors on average bandwidth
Live TV- Internet Unbounded High, 500 kb/s to Increasing Unicast, UDP, BBC iPlayer sports, stream 8 Mb/s sustained downlink broadcast
Web-cams Internet Unbounded Low, � 300 kb/s Low Unicast, HTTP, TCP, EarthCam, stream UDP, uplink, and security cameras
downlink
Video Internet Unbounded 40 kb/s to 450 kb/s Increasing Unicast, HTTP, TCP, Skype conferencing stream (higher for high UDP, uplink, and
end) downlink
IPTV broad- Network Unbounded 2 Mb/s to 15 Mb/s Low Multicast, AT&T, U-verse cast TV provider stream UDP, downlink
IPTV VoD Network Bounded, High; bursty; � High Unicast, progressive AT&T, U-verse provider 30 min to 2 Mb/s; server download, TCP, UDP,
2 hours and network downlink capacity are planned and engineered
BBC iPlayer is a trademark of British Broadcasting Corporation. EarthCam is a registered trademark of EarthCam Inc. Hulu is a trademark of Hulu LLC. Skype is a registered trademark of Skype Limited. U-verse is a trademark of AT&T Intellectual Property II, L.P. YouTube is a registered trademark of Google, Inc.
HTTP—Hypertext Transfer Protocol IP—Internet Protocol IPTV—Internet Protocol television TCP—Transmission Control Protocol TV—Television VoD—Video on demand
DOI: 10.1002/bltj Bell Labs Technical Journal 67
various types of video quality impairments [4]. In the
rest of the document we refer to live streaming IP
video as “Internet TV,” which predominantly uses
UDP for transport. Since Internet TV streams are
delivered as best-effort traffic by default, the main
requirement in offering it as a service is the ability to
control its transmission quality.
Today only a small number of consumers have
Internet TV services, but their take rate and usage are
increasing rapidly. Since these services are delivered
directly from the Internet to the end user, network
providers do not have responsibility for the audio and
video quality that users experience. However, net-
work providers have a strong desire to partner with
content providers to position themselves in this new
media ecosystem [13]. When network providers part-
ner with VCPs in offering Internet TV and participate
in delivering the services to the user, the service
model will resemble the IPTV service that they offer
today: subscribers will have a certain level of expec-
tation for quality, availability, and reliability. When
network providers partner with VCPs to offer sub-
scription services to users, their contracts will be
subject to a service level agreement (SLA). Network
providers will have to deliver the service with mea-
surable QoS so that they can commit to SLAs. If the
network provider relies on spare link capacity rather
than strict QoS control for delivering Internet video
services, it can no longer guarantee QoS performance
goals for the service.
Services that are associated with SLAs, like Voice
over IP (VoIP) as an example, are delivered in a high
priority traffic class. When service flows are protected
by placing them in the high priority traffic class, the
characteristics of the cumulative traffic, such as bursti-
ness, are not a concern in terms of meeting the SLA
commitments. However, as in the case of Internet
video services, if the service flows are delivered in the
lowest priority traffic class, then such traffic charac-
teristics become important. For any premium service
delivered via the lowest priority traffic class, the
burstiness of ordinary Internet traffic under high uti-
lization will cause packet loss and excessive packet
transmission delays. This will be discussed in detail in
the “Aggregation Network Traffic Analysis” section.
Network Architecture Internet video flows today are delivered like any
other best-effort Internet traffic to subscribers with
broadband Internet access service. Figure 1 depicts the high level network architecture necessary to sup-
port such an Internet access service, also known as
high speed Internet (HSI) service. Here, the first mile
is the transport leg that directly connects a customer
premises device to the service provider equipment,
e.g., the access loop that connects a residential gate-
way to its access node (e.g., a digital subscriber line
access multiplexer (DSLAM)). The first mile and the
access node are commonly referred to as the access net-
work. The “second mile” is part of the aggregation
network that connects access nodes to the aggrega-
tion switches. The second mile and the aggregation
segment of the network are referred to as the aggre-
gation network. The broadband network gateway
(BNG) generally serves as an IP edge router that
aggregates all Internet traffic including the IP video
flows between the Internet video VCP and the users.
For high quality Internet video delivery, BNG also
plays a critical role in providing IP QoS and traffic
management at the subscriber level.
Basic HSI services are typically offered with lim-
ited SLA guarantees—peak downstream and upstream
speed being the primary ones. Premium HSI services
(e.g., for business customers) may have additional
quality measures, such as average latency, and packet
loss rate. Applications such as Internet video, which
are delivered via the HSI service, are not provided
with additional traffic quality prioritization within the
HSI’s best-effort traffic class. However, to deliver
Internet video flows with higher priority end-to-end
QoS, it is necessary to separate those flows from the
rest of the HSI traffic. Policy rules for doing this traf-
fic prioritization have to be applied at the user and
application levels. Traffic prioritization has two essen-
tial steps: the first is to identify the individual traffic
flows, and the second is to mark the traffic flows for
high QoS treatment. Here identifying IP traffic flows
with specific application types provides critical infor-
mation for dynamic traffic management. For traffic
that is coming from sources outside the network
provider’s domain, deep packet inspection (DPI) can
68 Bell Labs Technical Journal DOI: 10.1002/bltj
be used for maximum flexibility in identifying
Internet video flows. DPI equipment can be used for
many different purposes at different functional levels
ranging from simple stateful inspection of the trans-
port protocol headers to inspecting deeper packet con-
tent looking for application patterns. However, DPI
deployment is not always required if the IP video
flows can be identified and marked for traffic man-
agement by other methods.
Different segments of the network have different
traffic characteristics. The problem of predicting the
required network link capacity becomes more diffi-
cult in the parts of the network that are closer to the
subscriber. In the access network, the peak transmis-
sion rate—rather than average link utilization—is the
subject of greater focus because there are only a few
users on each line and traffic is extremely bursty [8].
On the aggregation network links, traffic from hun-
dreds of access lines smoothes out significantly but
not enough that a statistical model can be applied to
optimize utilization [6]. The aggregation network is
where the difficulty arises about how to treat Internet
TV service flows with high QoS while maintaining
high link utilization to reduce CAPEX.
Aggregation Network Traffic Analysis As part of network operations, traffic measure-
ments are typically collected in five to fifteen minute
intervals. Traffic data at that time scale will give an
indication of the mean traffic volume but it will not
provide information on how bursty the traffic is at
shorter time scales. As described in [2], traffic at dif-
ferent time scales has different statistical properties
and is used for different engineering purposes such
as capacity planning, traffic engineering, and perfor-
mance analysis. For example, monthly, weekly, and
daily traffic patterns show a “signal � noise” character-
istic that is represented by non-stationary models.
On the other hand, statistical characteristics of the
traffic in hours and smaller time scales can have sta-
tionary statistical characteristics. Traffic patterns at
these smaller time scales are important for perfor-
mance engineering. For the QoS-related performance
analysis, the time scale of interest for this study is in
the milliseconds range.
Figure 2 shows our discrete event simulation (DES) data that demonstrates how burstiness of traf-
fic smooths out from a 10 millisecond time scale to a
1 second time scale. This data is the aggregate traffic
High level architecture
Access loop
Access node
Ethernet switch
BNG
RG
VoD/BTV
VCP server
Internet
Access network Aggregation network
Second mileFirst mile IP core
BNG—Broadband network gateway BTV—Broadband television IP—Internet Protocol
RG—Residential gateway VCP—Video content provider VoD—Video on demand
Figure 1. Reference architecture.
DOI: 10.1002/bltj Bell Labs Technical Journal 69
generated from 300 nodes. The chart in Figure 3 shows the coefficient-of-variance of data, decreasing
from 0.154 to 0.094, as the time scale in traffic mea-
surement increases.
The level of aggregation of individual flows also
impacts the burstiness of the traffic. As flows aggre-
gate in the network, the burstiness of the traffic
decreases. The variance of broadband traffic at indi-
vidual user end points is naturally very high, but as
the traffic flows aggregate, burstiness smooths out sig-
nificantly. Figure 4 provides a measure of this
smoothing factor driven from our DES results. The
measurement results for 10, 100, 300, and 600 nodes
show how the peak-to-mean ratio rapidly converges
to 1.76 beyond 300 users. Other details of the simu-
lation are described in the “Discrete Event Simulation”
section.
There has been an ongoing debate for more than
a decade whether the Internet traffic is self-similar or
0
50
100
150
200
250
300
350
400
450
500
150 170 190 210 230 250 270 290 Seconds
M b
/s
10 ms 100 ms 1 sec Mean
Sec—Seconds
Figure 2. Same traffic at 10 ms, 100 ms, 1 second time scales.
0.154
0.000
0.020
0.040
0.060
0.080
0.100
0.120
0.140
0.160
0.180
10 100 1000
Time scale (milliseconds)
C o
e ff
ic ie
n t
o f
va ri
a n
ce
0.109 0.094
Figure 3. Variance of traffic data at increasing time scales.
1.76 1.54 1.33
1.00
10.00
100.00
0 200 400 600 800
Number of users
O ve
rp ro
vi si
o n
in g
f a ct
o r
Peak
0.01% loss
0.1% loss
Figure 4. Peak-to-mean ratio of traffic at different aggregation levels.
70 Bell Labs Technical Journal DOI: 10.1002/bltj
not—that is, in simple terms, if the burstiness of the
traffic transcends time scales. It has not been possible
to determine a unifying model to accurately repre-
sent the burstiness of the aggregate broadband traffic
because there are many factors that complicate its
analysis. Some of these factors include changes in the
application profiles, the access link speeds, and
the characteristics of session flows; time scale intervals
that are used as reference; and network properties
such as the aggregation speed-up ratio [14]. Based on
recent understandings, the widely accepted view is
that in the millisecond time scale range, broadband
data traffic does not exhibit self-similarity. Terdik and
Gyires present the most recent discussion on this
topic, summarizing decade-long prior work. They also
mention the widely referenced observation that traf-
fic above the seconds time scale shows long-range
dependence, which is described as “spikes riding on
ripples riding on swells” [17]. In terms of network
capacity planning, it is good if the traffic in the mil-
lisecond time scale does not exhibit self-similarity.
Otherwise it would mean that the aggregate traffic
will be burstier and QoS-related performance will be
more unpredictable.
As described in earlier sections, a broadband
access line can carry IP traffic flows for various appli-
cations with different characteristics. These different
traffic flows require different QoS treatment; for
example, VoIP flows require low latency and low
packet loss. Access network equipment applies differ-
ent priority treatment to individual flows to support
different service requirements. Evans and Filsfils pro-
vide a good summary of the QoS requirements of dif-
ferent service types in [5]. Our analysis is concerned
only with the best-effort traffic class, so the point in
question is whether dynamic QoS control is neces-
sary to provide premium Internet services or whether
SLA can be satisfied cost effectively with an increase
in capacity alone.
As shown in Figure 5, carrying premium Internet TV flows in the same traffic class as ordinary Inter-
net traffic may impair the media quality. Traffic bursts
on highly utilized links can result in excessive packet
delay jitter for all IP flows and cause packet drop in IP
video flows due to the real time constraints. If the
Internet TV traffic is given a higher priority than
the best-effort class, it will make traffic engineering
for the whole pipe easier since a stringent upper
bound for packet loss will no longer be required for
the best-effort class, and any bursts caused by regular
HSI traffic will no long impact the quality of experi-
ence. Otherwise, the challenge remains to anticipate
the traffic characteristics at the aggregation network.
It will also be a challenge to accurately estimate the
(b) Premium service differentiation with QoS(a) Premium service differentiation without dynamic QoS
Packet loss of other best-effort flows
Packet loss and excessive delay of premium media flows
QoS—Quality of service
Packets of premium Internet media flows
All other Internet traffic
Figure 5. Impact of congestion on premium service flows.
DOI: 10.1002/bltj Bell Labs Technical Journal 71
packet loss rate for that traffic pattern for a given link
capacity. Determining the expected packet loss rate
will allow network planners to estimate the target
link utilization level, leaving just enough excess
capacity in the link.
Discrete Event Simulation In the aggregation network, since the traffic charac-
teristics do not fit into a uniform pattern (and are con-
tinuously changing with use of new Internet
applications) it is not easy to represent this traffic with
stationary statistical models. Therefore DES becomes
a useful method to simulate traffic characteristics and
to quantify its properties. In our study we used the
OPNET SP Guru* Network Planner tool for the simu-
lation.
Two types of traffic were used in the model: 1)
average broadband traffic and 2) premium service
traffic. The average broadband traffic is the typical
Internet traffic generated by residential end users. The
premium service traffic represents HD Internet TV
traffic delivered over the Internet. The model simu-
lates multi-year traffic growth and includes an
increase in average traffic rates going forward as well
as an increase in the take rate for premium services.
The average broadband traffic per user at busy hour
started at 100 kb/s in Year 1 and grew to 384 kb/s in
Year 5. Premium service traffic started at 1.5 Mb/s
in Year 1 and grew to 7.6 Mb/s in Year 5. If we con-
sider the Internet video services that are currently
available, the assumed bandwidth for premium
Internet video in the model is low compared to typi-
cal IPTV transmission rates. However, over time, the
take rate for Internet video is expected to increase to
IPTV levels, while better encoder efficiencies are likely
to bring the transmission rates for both lower.
Nevertheless, these bandwidth assumptions did not
have relevance for absolute traffic volumes, but they
were rather used in relative terms for capital cost com-
parison under different traffic management schemes.
The model had 300 nodes representing individual
subscribers. Average broadband traffic was applied to
each of the 300 nodes. The premium traffic was applied
to a subset of nodes with increasing percentage over
five years. This increase represented the subscriber
take rate for premium services; values were 6 per-
cent, 15 percent, 24 percent, 31 percent, and 32 per-
cent from Year 1 to Year 5, respectively.
Custom OPNET tasks were defined to represent
each traffic type (average broadband and premium
services). A task is composed of a request and a
response, and it allows detailed definition of the sta-
tistical distribution of the traffic pattern. Both the
request and the response components have several
parameters such as inter-request time, packets per
request, response packet size, packets for response,
and inter-packet time. A mix of Pareto and exponential
statistical distributions were used to represent average
broadband traffic and premium video service traffic.
Twenty parameters had to be defined to represent
the two types of traffic for each year and few of the
parameters were changed to create the multi-year
simulation. Then for each year and traffic type, appli-
cation definitions were created, which refer to and
use the custom tasks. Finally, profiles were defined
that associate application definitions with individual
nodes.
Simulation Results The key result presented by the simulation was
the representation of packet-level traffic at the 10 mil-
lisecond time scale. From this detailed traffic data, it
was possible to quantify the association between traf-
fic peaks and the expected packet loss rate when links
are utilized at 100 percent capacity. Figure 6 shows a snapshot of the traffic trace with lines showing peak
traffic cutoff lines for 0.01 percent and 5 percent
packet loss rates. The results show that for traffic with
an average rate of 441 Mb/s, the link has to have more
than 655 Mb/s capacity for 0 percent packet loss,
575 Mb/s capacity for 0.01 percent packet loss, and
500 Mb/s capacity for 5 percent packet loss. The plot
in Figure 7 shows how the traffic peak is cut off for a range of packet loss rates.
From the simulation output, we calculate peak
traffic for 0.01 percent and 5 percent packet loss rates
for the five year interval. The expected IPTV packet
loss rate can be on the order of 0.0001 percent [16].
The International Telecommunication Union (ITU)
standards document Y.1541 suggests 0.01 percent
72 Bell Labs Technical Journal DOI: 10.1002/bltj
packet loss rate as the upper bound for IP based ser-
vices. The 3rd Generation Partnership Project (3GPP)
standards document TS 23.203 Release 9 identifies
various “packet error loss rates” for different QoS class
identifier (QCI) values; the upper bound for the loss
rate is given as 0.1 percent for conversational video
and 0.0001 percent for non-conversational video. We
picked 0.01 percent target as a conservative assump-
tion for non-conversational video in our study (more
stringent loss rates are likely to amplify results). A 5
percent packet loss was picked as the target for best-
effort traffic performance in relation to the measured
Internet packet loss rate in Europe [11].
Figure 8 shows the five year total traffic esti- mates. Three separate curves indicate peak, 0.01 per-
cent, and 5 percent packet loss rates. The differential
between the 0.01 percent loss curve and the 5 percent
loss curve shows the difference in operational target
level that dynamic QoS control would bring. If pre-
mium media flows are protected by placing them to a
separate, high priority traffic class, then the link can be
operated at a higher packet loss rate, which will only
impact the remaining best-effort Internet traffic—as it
is normally operated today. If the premium media flows
are not placed in a different traffic class, then the uti-
lization of the link has to be lowered so the whole best-
effort traffic stream (with the premium flows in it) has
to be operated at a 0.01 percent loss rate. As we will
see in the sections following, the differential between
200 400 600 800 1000 12000 4.5
4.8
5
5.2
5.4
5.6
5.8
6
6.2
6.4
6.6 � 108 2013 655 Mb peak, 441 Mb mean, 10 sec @10ms
5% pkt loss
pkt—Packet sec—Seconds
0.01% pkt loss
Figure 6. Packet loss levels marked on a sample traffic trace.
0%
5%
10% 15%
20%
25%
30%
35% 40%
45% 50%
0.00% 1.00% 2.00% 3.00% 4.00% 5.00%
Packet-loss
Figure 7. Traffic peak cutoff for a range of packet loss rates.
DOI: 10.1002/bltj Bell Labs Technical Journal 73
the 0.01 percent and 5 percent utilization levels is an
important instrument to defer capital investment.
We must emphasize that these results are depen-
dent on the traffic profiles that are generated for indi-
vidual end points in our simulation. Actual end point
traffic profiles will be different in the field, and they
will change as new Internet application patterns
emerge. These results are not used as absolute mea-
sures but for relative comparison of two different
engineering approaches, that is, dynamic QoS control
versus putting in excess link capacity early on.
Cost Models The value proposition of using dynamic QoS con-
trol relies on the findings of the traffic model as
described in the previous section and the three dif-
ferent cost models: the aggregation network capacity
cost model, the access network capacity cost model, and
the DPI network deployment cost model. Since the
scope of this paper is the aggregation network,
the aggregation network capacity cost model is
explained in detail.
Aggregation Network Capacity Cost Model In the “Aggregation Network Traffic Analysis” sec-
tion, we saw how traffic congestion is quantified in
terms of QoS metrics. In this section, we are going to
look at how link utilization is associated with conges-
tion, and how the required capital investment for link
capacity is estimated for a given level of utilization.
The capacity cost model estimates the required five
year capital investment for upgrading aggregation net-
work links—to provide enough capacity to meet a
given QoS metric. The model first calculates the link
capacity differential from the packet loss derived in the
previous section. Then it estimates the amount of new
equipment needed using the capacity differential. The
model has the following set of key parametric assump-
tions: a five year plan; a broadband subscriber growth
rate; average broadband traffic growth rate; premium
media service take rates, increasing values for each
year; 300 nodes per access node; and a 1 Gb/s link pro-
visioned between the access node and the aggregation
switch for best-effort Internet traffic.
As we have seen in the simulation results in
Figure 8, not using dynamic QoS control requires
decreasing the traffic load on links to satisfy a given
packet loss upper bound. Otherwise, a network
provider cannot give strict SLA guaranties for pre-
mium services just by increasing link capacity (if the
statistical model of the traffic is not known).
Packet loss impact on capacity
-
200
400
600
800
1,000
1,200
2010 2011 2012 2013 2014
M b
/s
Peak
0.01%
5.0%
Peak value differential for 0.01% and 5% packet loss
Figure 8. Total traffic on 1 Gb/s link, peak at different loss rates.
74 Bell Labs Technical Journal DOI: 10.1002/bltj
The target utilization differential between the two
cases at a given time (that is, using dynamic QoS
control with a high packet loss rate versus
maintaining an excess capacity) can also be viewed
as increasing the capacity to the same target level at
some time in the future, or delaying an upgrade, as
shown in Figure 9. This time interval levels-out as an average of 8.8 months in the model, or roughly
about six months in the last three years of the
model.
When 1 Gb/s aggregation network links in the
second mile are considered, the traffic estimates indi-
cate that the bulk of the link upgrades must occur
after Year 4. To represent the statistical uncertainty of
the increasing likelihood of an upgrade, we can arbi-
trarily assign a link upgrade probability curve: a
quadratic function. The same probability curve is then
shifted on the x-scale six months with the same
assigned probabilities for each year. Therefore, given
the number of total subscribers, subscribers per access
node, traffic per access node, and link upgrade proba-
bility, it is possible to estimate the total number of
links that must be upgraded each year. Finally, know-
ing the equipment cost allows us to calculate the
required five year capital investment. The model takes
price erosion and cost of capital into consideration.
These results are shown in the “Value Proposition”
section.
Access Network Capacity Cost Model For the access network, network providers are
under competitive pressure to upgrade their DSL lines
to increase bandwidth capacity. New Internet services
are also putting pressure on the access lines in terms
of capacity. Besides the ongoing transformation effort
to move legacy ADSL lines to VDSL and to fiber, using
techniques like channel-bonding (or pair-bonding)
also allows network providers to double the band-
width of existing lines [10]. However, unlike the
capacity expansion in the aggregation network,
deploying channel-bonding to upgrade millions of
access lines requires a large investment. To reduce
these network upgrade costs, dynamic QoS control
again becomes an effective technique, as it is in the
aggregation network. Deploying dynamic QoS con-
trol enables delivery of premium media flows by mak-
ing maximum use of the available bandwidth. Details
of the access network cost model are not covered here,
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0 12 24 36 48 Months
U ti
li za
ti o
n
DPI—Deep packet inspection
w/o DPI w DPI
Figure 9. Time interval for reaching the same link utilization level on a 1 Gb/s link.
DOI: 10.1002/bltj Bell Labs Technical Journal 75
but the results are summarized as part of the value
proposal.
The access network cost model has the following
input components: Five year subscriber projections
for ADSL to VDSL and fiber upgrades, access line dis-
tance to central office distribution statistics, and ADSL
and VDSL distance to bandwidth characteristics. Using
the input information, the model estimates the num-
ber of subscribers in the different line capacity range
for each access technology type for the next five years.
Then the model calculates the total cost of link
upgrades while taking into consideration the price
erosion, discount rate, customer sale discounts, labor
cost, labor salary increases, and equipment cost. The
results are shown in the “Value Proposition” section.
DPI Cost Model As discussed above, putting premium media
traffic flows on a separate traffic class first requires
identifying the individual traffic flows. DPI is often
used to mark premium media flows as the traffic com-
ing into the network from external sources. Various
network properties have to be considered in deter-
mining a right DPI solution for a provider’s network.
The cost of DPI deployment is impacted by many fac-
tors, like the embedded equipment in the network,
equipment models and vendors, and the communi-
cations technologies in use.
In our study, a five year DPI network deployment
cost model is used based on the Alcatel-Lucent
Multiservice-Integrated Service Adapter module (MS-
ISA) deployed on Alcatel-Lucent’s 7750 Service Router.
Inputs to this model include five year HSI subscriber
projections, premium service take rates, average high-
speed-Internet bandwidth, average premium services
bandwidth, price erosion, and engineering and instal-
lation costs. The model incorporates details on the
configuration and the cost of the router system that
hosts the MS-ISA modules. Besides network equip-
ment parameters, the model also includes mainte-
nance, floor space, and power cost parameters. The
model assumes that the Internet traffic for the con-
gested links is diverted for DPI, and such diversion
remains in place even after the link capacities are
upgraded. A further optimal approach here would
remove the traffic from DPI when link capacities are
upgraded. The results are shown in the “Value
Proposition” section.
Value Proposition The value proposition is built on the results of the
three cost models described so far. The two network
cost models quantify the savings from postponing net-
work equipment upgrades, and the DPI cost model
provides the required network investment for the
dynamic QoS control capability, which makes the cost
savings possible in the aggregation and the access net-
works. Figure 10 shows the net present value (NPV) of the overall savings in a planning horizon of five
years (with the cost of capital at 10 percent).
As shown in the figure, the strategy yields positive
NPV throughout the study period with 12 million
euros in the fifth year. Dynamic QoS delivers the
highest savings in the access network. Savings in
the aggregation network are also significant though
those alone do not cover the cost of DPI deployment.
As intuitively expected, the DPI cost is most sensitive
to the total volume estimated for premium Internet
service delivery, and thus we see the cost increasing
over the years as the service take rate increases. The
cumulative chart shows that cost savings accrue early
on and taper off in the later years. This is because the
capacity upgrade costs are postponed to later years.
The DPI cost, on the other hand, continues to increase
as the traffic estimates increase over the years—both
for ordinary Internet traffic and for the premium ser-
vices traffic. The decreasing NPV curve in the later
years is also indicative of this trend.
Since there are many variables in this type of
value analysis, savings will depend on the particular
network, current constraints, and future plans tied to
business goals. For example, in the access network
scenario that was analyzed, factors such as more rapid
migration from ADSL to fiber to the home (FTTH)
tend to lower the savings from use of DPI—since
excess bandwidth capacity provides enough room for
Internet video traffic. Therefore, it is hard to general-
ize rules from one such complex study and apply it to
other network scenarios.
76 Bell Labs Technical Journal DOI: 10.1002/bltj
Conclusion As the communications and the information
industries are going through significant transforma-
tion, network providers are also trying to adjust their
business interests. Until recently, the nature of wire-
line network connectivity allowed network providers
to maintain exclusive service contracts with their sub-
scribers, but the delivery of new Internet TV services
directly to the user is challenging traditional business
models. In this changing environment, one of the
many alternatives for network providers is to partner
with third party content providers. However, partner-
ships and new services also bring new requirements.
Two of these requirements at the top of the list are
the capability to offer SLA commitments and the
capability to deliver high QoE. It is a challenge to
meet these requirements while managing a network
traffic load that is increasing exponentially because of
new Internet services. The cost of the network capac-
ity upgrade is the first challenge that network opera-
tors face. In this study, we showed why dynamic QoS
control is important to offer Internet TV both for pro-
viding SLA commitments and for service assurance.
We were able to demonstrate that using dynamic QoS
techniques can lower the capital cost for network
capacity upgrades. We reached this conclusion by
building a correlation model of QoS targets, utiliza-
tion thresholds, and required network upgrades.
Through our network simulation model, we demon-
strated how traffic characteristics in the aggregation
network can impact service quality unless QoS man-
agement techniques are adopted by network
providers.
Acknowledgements We want to thank and acknowledge Paul Gagen,
Mohamed El-Sayed, and Vanita Katkar for their guid-
ance and contributions throughout this study.
*Trademarks BBC iPlayer is a trademark of British Broadcasting
Corporation. Netflix is a registered trademark of Netflix, Inc. SPGuru is a registered trademark of OPNET
Technologies. YouTube is a registered trademark of Google, Inc.
References [1] J. Babiarz, K. Chan, and F. Baker,
“Configuration Guidelines for DiffServ Service Classes,” IETF RFC 4594, Aug. 2006, �http://www.ietf.org/rfc/rfc4594/�.
[2] M. Crovella and B. Krishnamurthy, Internet Measurement: Infrastructure, Traffic &
€ 18,968
€ 8,784
€ (15,413)
€ 12,339
€ (20,000)
€ (15,000)
€ (10,000)
€ (5,000)
€ -
€ 5,000
€ 10,000
€ 15,000
€ 20,000
€ 25,000
2010 2011 2012 2013 2014
N P V
T h
o u
sa n
d s
CAPEX—Capital expenditures DPI—Deep packet inspection NPV—Net present value
1st Mile
2nd Mile
DPI cost
Net Present Value
Figure 10. NPV of CAPEX savings.
DOI: 10.1002/bltj Bell Labs Technical Journal 77
Applications, John Wiley & Sons, Chichester, Eng., Hoboken, NJ, 2006, pp. 216–233.
[3] M. L. El-Sayed, Y. Hu, S. Kulkarni, and N. Wilson, “Comparison of Transport Network Technologies for IPTV Distribution,” Bell Labs Tech. J., 11:2 (2006), 215–240.
[4] B. Erman and E. P. Matthews, “Analysis and Realization of IPTV Service Quality,” Bell Labs Tech. J., 12:4 (2008), 195–212.
[5] J. Evans and C. Filsfils, Deploying IP and MPLS QoS for Multiservice Networks: Theory and Practice, Morgan Kaufmann, San Francisco, CA, 2007, pp. 1–86.
[6] G. Hasslinger, J. Mende, R. Geib, T. Beckhaus, and F. Hartleb, “Measurement and Characteristics of Aggregated Traffic in Broadband Access Networks,” Proc. 20th Internat. Teletraffic Conf. (ITC-20) (Ottawa, Ont., Can., 2007), pp. 998–1010.
[7] O. Heckmann, The Competitive Internet Service Provider: Network Architecture, Interconnection, Traffic Engineering and Network Design, John Wiley & Sons, Chichester, Eng., Hoboken, NJ, 2006, pp. 273–294.
[8] T. Hossfeld and K. Leibnitz, “A Qualitative Measurement Survey of Popular Internet- Based IPTV Systems,” Proc. 2nd Internat. Conf. on Commun. and Electron. (ICCE ‘08) (Hoi An, Vnm., 2008), pp. 156–161.
[9] M. Howard, Video Traffic Behaves Erratically: CDNs to the Rescue?, Infonetics Research, Nov. 16, 2010.
[10] International Telecommunication Union, Telecommunication Standardization Sector, “ATM-Based Multi-Pair Bonding,” ITU-T G.998.1, Jan. 2005, �http://www.itu.int�.
[11] Internet Traffic Report, �http://www.internet trafficreport.com/�.
[12] J. Joung, J. Song, and S. S. Lee, “Flow-Based QoS Management Architectures for the Next Generation Network,” ETRI J., 30:2 (2008), 238–248.
[13] A. Kishore, OTT Video: Service Providers Face a Gathering Storm, Heavy Reading, vol. 8, no. 3, Mar. 2010.
[14] W. Lautenschlaeger and G. J. Eilenberger, “Dimensioning of Carrier Grade Packet Transport for Simple Network Operation,” Bell Labs Tech. J., 14:1 (2009), 147–160.
[15] G. Maier, A. Feldmann, V. Paxson, and M. Allman, “On Dominant Characteristics of Residential Broadband Internet Traffic,” Proc.
9th ACM SIGCOMM Internet Measurement Conf. (IMC ‘09) (Chicago, IL, 2009), pp. 90–102.
[16] Multimedia Research Group, IPTV Video Quality: QoS & QoE, Feb. 2007, p. 85.
[17] G. Terdik and T. Gyires, “Does the Internet Still Demonstrate Fractal Nature?,” Proc. 8th Internat. Conf. on Networks (ICN ‘09) (Gosier, Glp., 2009), pp. 30–34.
(Manuscript approved February 2011)
BILGEHAN ERMAN is a research engineer in the Alcatel- Lucent Bell Labs’ Network Planning, Performance, and Economic Analysis organization in Murray Hill, New Jersey. He holds a B.S. degree in electrical engineering and an M.S. degree in computer engineering
both from Middle East Technical University (METU) in Ankara, Turkey. His current interests and recent activities include business development, traffic, and policy management solutions in all-IP networks, and application enablement solutions for large network providers. His past activities and expertise include IP video quality modeling, distributed applications, and real time switching software architectures. Mr. Erman holds one patent and has six patents pending.
Z. JOHN ZHAO is a member of technical staff with the Network Modeling and Optimization Department at Alcatel-Lucent Bell Labs in Murray Hill, New Jersey, and a member of the Alcatel-Lucent Technical Academy. He holds a B.S. degree in automatic control
engineering from Nanjing Institute of Technology in China and M.S. and Ph.D. degrees in electrical engineering from Polytechnic University, Brooklyn, New York. His current work focuses on topology design, cost analysis, and scalability assessment of various data networks, primarily MPLS/IP and metro Ethernet networks for broadband access and mobile data services. Dr. Zhao is also involved in the development of innovative design and optimization methods and tools for next-generation network (NGN) deployment. His recent research interests include dynamic QoS policy and application modeling for emerging service realization in both wireline and wireless networks. ◆
Copyright of Bell Labs Technical Journal is the property of John Wiley & Sons, Inc. and its content may not be
copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written
permission. However, users may print, download, or email articles for individual use.
