You are on page 1of 11

A Method and Simulation Study: Network Dimensioning of the

Broadband Residential Ethernet-based Access Network


Zhuang Zhao, Qiong Liu Hao Guan
Multimedia Communication
& Network Engineering Research Center Nokia Research Center
Institute of Software, the Chinese Academy of Sciences Nokia (China) Investment CO., LTD
Dept.3 P. O. Box 8718, Beijing, China 100080 He Ping Li Dong Jie, Beijing, 100013 P.R. China
zhuangzh@isdn.iscas.ac.cn, liuq@isdn.iscas.ac.cn hao.guan@nokia.com

Abstract
Network dimensioning is an indispensable component of QoS provisioning in the residential multi-service IP-
based access networks. This paper discusses network dimensioning of the residential multi-service IP-based
access networks over Ethernet thoroughly, i.e. a Broadband Residential Ethernet-Based Access Network
(BREAN), and tries to find a pragmatic engineering method of network dimensioning. We propose that the
layer 2 Ethernet domain of the BREAN implements EF (PHB and PDB) and AF (PHB and PDB) services of
DiffServ so that we can deploy a same method of network dimensioning in the whole BREAN. This paper
investigates the past study efforts of network dimensioning, and expends or amends some study results with
DiffServ, such as the universal link model, the dimensioning for the elastic traffic and the stream traffic, traffic
matrix and user model. By systematic method, combining these indispensable components gives a rather
pragmatic engineering method of network dimensioning in the BREAN, offering PTOS, VoIP, TV, VoD and
Internet access, etc. We validate our method of network dimensioning by the simulations, and the simulation
results illustrate that our method can appropriately estimate the bandwidth requirements of the different
classes based on DiffServ framework, reasonably plan bandwidth capacity of the network links, and promise
multi-service QoS provisioning without overburden always in the BREAN.
Keywords: Residential Access Network; Ethernet; DiffServ; Network Dimensioning; Universal Link Model;
Traffic Matrix; User Model

1 Introduction in case network is congested or in overburden. Net-


work engineering has to resolve the trade-off be-
Nowadays, the broadband residential access infra- tween capacity and QoS requirements [2].
structure remains one of the major research areas Traffic theory currently plays a very minor role in
for the Next Generation Network (NGN). A great the design of the Internet. Network dimensioning is
deal of interest surrounds the deployment of IP generally based on simple rules of thumb while
technique in metropolitan area networks and access considerable effort is spent on the design of variety
networks currently, and there are also many research QoS mechanisms. Jim W. Roberts argued that traffic
efforts about Ethernet in the subscriber access net- theory, an essential component in the design of tradi-
work. Some standard organizations and associations tional telecommunications networks, should be in-
such as the IEEE study group on "Ethernet in the creasingly applied in the development of the multi-
First Mile" (EFM), Ethernet in the First Mile Alli- service Internet [3].
ance (EFMA), the Metro Ethernet Forum (MEF), Indeed, a BREAN with DiffServ-enabling can hardly
etc., are studying the relating issues. We presented a offer QoS provisioning based on per-flow traffic. So
broadband residential Ethernet-based access infra- network dimensioning or capacity planning of per-
structure in the past paper, and noted that combing class combining with QoS mechanisms is indispen-
network dimensioning, QoS mechanisms and pricing sable for the applications.
can offer multi-service QoS provisioning [1]. The rest of this paper is organized as follows. Sec-
Network dimensioning and QoS mechanisms are two tion 2 introduces architecture of the BREAN. Sec-
key components to provide multi-service QoS guar- tion 3 discusses QoS provisioning in L2 domain and
antees without overprovision. Network dimensioning L3 domain. Section 4 expands the universal link
makes the networks not to work with overburden al- model with DiffServ-enabling. Section 5 discusses
ways. QoS mechanisms provide QoS guaranteeing the dimensioning for the elastic traffics and the
for important applications or real-time applications stream traffics in the BREAN with DiffServ-enabling.
Section 6 gives traffic matrices and user model in switches are in charge of aggregating traffics in the
the BREAN with DiffServ-enabling. In section 7, we large-scale BREAN. QoS provisioning of the L2
give a rather pragmatic engineering method of net- domain is discussed in the next section thoroughly.
work dimensioning and validate our method by simu- The DS domain includes access router, edge router
lation study. Finally, we conclude this paper. and core router. These kinds of routers must offer
EF PHB and AF PHB defined by IETF DiffServ
Workgroup.
2 Architecture of the BREAN Service domain includes local service domain and
MAN/WAN service domain. For example, local ser-
The residential access network is a star-structured vice domain may offer TV and VoD, and MAN/WAN
topology network typically. Architecture of the service domain may offer Internet access and VoIP.
BREAN is illustrated in Figure 1. The BREAN con- The BREAN uses the Ethernet link, i.e. gigabit
sists of four components, i.e. L2 domain, L3 Diff- Ethernet, 100 Mbit/s Fast Ethernet or 10 Mbit/s
Serv (DS) domain, service domain and Customer Ethernet connecting between the different equip-
Premises Network (CPN). ments. In addition, exploiting IEEE 802.3ad link ag-
The L2 domain consists of multi-level Ethernet gregation or MPLA techniques in the BREAN can
switches. The first level Ethernet switch is in charge improve network availability.
of user accessing and the other level Ethernet

Local Service Domain


MAN/WAN Service Domain
Satellite
IP Telephony
Gateway
CATV
Network
P STN/ISDN
Network
Video
Database Antenna

Internet
Network TV Server
Management&
Video Billing Station
Server

Switch GE

Edge

Layer 3 Domain
Router
POS/GE

POS/GE POS/GE

Access Access
Router Router

GE GE GE
Layer 2 Domain

GE

Switch FE FE Switch
FE FE
... ...

Switch Switch Switch Switch

Residential Residential
Gateway Gateway
CPN

1
42
753
86
9 TV
*8#
TV PC
IP Phone

PC
Phone

GE: Gagabit Ethernet


FE: Fast Ethernet
POS: Packet Over SONET/SDH
CPN: Customer Premises Network

Figure 1. Architecture of the Broadband Residential Ethernet-Based Access Network


tween different traffic types, and implement QoS
provisioning in the carrier-grade residential access
3 QoS Provisioning Based on infrastructure. We propose that Ethernet offers
DiffServ Framework DiffServ-like features in the layer 2, such as classi-
fying, conditioning, queue scheduling and buffer
Adequate QoS provisioning of multi-service is a management, etc. Indeed, there are some Application
challenge in a converge network. Nowadays, in a Specific Integrated Circuit (ASIC) chip-set solutions
popular view, a network consists of backbone net- only to offer the limited functions of DiffServ at
work and access network. There is an agreement that present. We can easily implement these DiffServ-
backbone network and access network can deploy the like features of layer 2 by network processor. Thus,
different design methods of QoS provisioning re- we can implement QoS provisioning based on Diff-
spectively. So, we can solely consider the QoS Serv framework in the whole BREAN, i.e. the
provisioning in the BREAN. In addition, there is still BREAN can offer EF and AF services in L2 domain
a need to cooperate in handing traffic from access and L3 domain.
network to backbone network to offer end-to-end QoS provisioning of the L2 domain is based on Diff-
QoS guarantee for the applications. Serv-like framework where traffic entering a
Most often, industry players talk in terms of IntServ network is classified and possibly conditioned at the
being deployed at the edge of a network, while Diff- boundaries of the network, such as Residential
Serv is used in the network core. There are many Gateway (RG) and switch at the boundaries of the L2
study efforts about IntServ framework at the edge of domain, and assigned to different behaviour aggre-
a network, i.e. access network. In the carrier-grade gates. Each behaviour aggregate is identified by a
residential access networks, however, there are thou- single 802.1q priority type. Within the core of the
sands of residential users, and these users produce L2 domain, packets are forwarded according to the
thousands of flows of the applications. Scalability is per-hop behaviour associated with the 802.1q prior-
still an important design target in the BREAN. We ity type.
argue that DiffServ framework is an appropriate Clause 2.3.4 gives locations of traffic conditioners
choice in carrier-grade residential access infrastruc- and MF classifiers in RFC 2475 [4]. L2 domain can
ture. be regard as the source domain, and traffic source
We employ DiffServ architecture in the whole and intermediate node within a source domain may
BREAN. Obviously, the generic L3 domain deploys perform traffic classification and conditioning func-
DiffServ architecture, and offers EF and AF services. tions. At the boundary between a L2 domain and a L3
The differentiated services architecture [4] is based domain, traffic streams may be classified, marked,
on a simple model where traffic entering a network and otherwise conditioned on either end of a bound-
is classified and possibly conditioned at the bounda- ary link (the egress node of the upstream L2 domain
ries of the network, and assigned to different behav- or the DiffServ ingress node of the downstream L3
iour aggregates. Each behaviour aggregate is identi- domain). The SLA between the domains should
fied by a single DS code-point. Within the core of specify which domain has responsibility for mapping
the network, packets are forwarded according to the traffic streams to layer 2 differentiating service be-
per-hop behaviour associated with the DS code-point. haviour or DiffServ behaviour aggregates and condi-
Traditional Ethernet technique didnt provide tioning those aggregates in conformance with the
bounded delay and distinguish between different traf- appropriate Traffic Conditioning Agreement (TCA).
fic types. There are some study efforts to provi de In addition, there is a mapping from IEEE 802.1q
QoS provisioning of real-time applications over priority type to DS code-point at the boundary be-
Ethernet in the past couple of years. Ethernet tech- tween a L2 domain and a L3 domain.
nique has evolved from shared bus to switched mode
and from half duplex operation to full duplex opera-
tion. A dedicated segment is no longer a shared me- 4 Universal Link Model with
dium, there are no collisions and the full bandwidth DiffServ-Enabling
is available to a single node. IEEE Std. 802.1P and
802.1Q are presented to provide service differentia- Anto Riedl, et al presented a Universal Link Model
tion of Layer 2, and IEEE 802.1Q adds a 16-bit tag (ULM) [5], which can be used for dimensioning pur-
to the Layer 2 header, including three bits that can be poses. However, their ULM isnt fully suiting to the
used to classify priority. The standard specifies eight DiffServ-enabling network. We amend this ULM
different priorities. Queuing mechanism directly considering aggregate scheduling and give the details
contributes for implementing service differentiation, of this model. The ULM amended is illustrated in
and a switch typically implements multiple queues in Figure 2.
their input port interfaces or output port interfaces. The new link model adequately considers the charac-
However, Ethernet is short of systematic studies to teristics of the IP access network over DiffServ
assure bounded delay, effectively distinguish be- framework. It considers the different traffic classes
Bandwidth Partitioning stant and equal to C i . This pessimistic point of view
or Traffic Priorization
will lead to conservative dimensioning rules by pro-
Traffic viding lower bounds on performance parameters [22].
BestEffort Traffic
A specific queue or a system is described either by
an M/G/1 PS or M/G/R PS model [6]. Anton Riedl,

LinkCapacity
AF1Traffic
et al. gave the dimensioning procedure based on
AF2Traffic
AC TC M/G/R PS model for the elastic traffic [7]. They
(optional)
AF3Traffic considered the average file transfer time (for files of
AF4Traffic a certain size) as the relevant QoS criterion. From
the other perspective, they also considered an ave r-
AC TC
EF Traffic age throughput requirement for each flow. The appli-
RejectedFlow cations of the elastic traffic can select the two dif-
Dropped Packets
ferent criterions respectively.
AC:AdmissionControl
TC: Traffic Conditioning
Let us represent elastic flow arrivals by a Poisson
process with mean arrival rate i , a realistic assump-
Figure 2. Universal Link Model with DiffServ-Enable tion if we restrict our study to stationary busy peri-
separately and determines their necessary capacity ods, and denote by Li the mean flow size (in bytes)
shares independently like Anto Riedl, et al. We re- without specifying the flow size distribution. The
state the two points: offered load of class i traffic is
1. Fixed bandwidth, i.e., predictable capacity assign- i = i Li / C = i / C with respect to the allotted
ment to the individual classes without detrimental class i link capacity. J.W. Roberts, et al. [21] gave
interference can be assumed. Indeed, the problem is the expression for average flow throughput with ac-
a difficult task, but the QoS mechanisms designed cess rate constraint:
elaborately can solve it. However, potential multi- N 1
( R)n ( R) N
plexing gains in case of work-conserving systems (1 ) +
n! N!
are neglected. d = Ci (1 ) n=0
(1)
N 1
n
N
R(1 )
( R ) ( R)
2. Within an allocated bandwidth fraction, certain 2

n!
+
N!
[1+ N (1 )]
traffic streams can be treated preferentially. n=0

C
where = i , R = i and N = [R ] . (integral
i rmax
5 Dimensioning for Elastic part of R )
Traffic and Stream Traffic J.W. Roberts, et al. noted that capacity allocation
could be performed according to a target perform-
The applications are classified as the essential cate-
ance specified in terms of average flow throughput
gories of elastic and streaming flows. Streaming
when no admission control is applied, and in terms
flows are produced by audio and video applications.
of flow blocking probability when admission control
On the other hand, elastic flows result from the
is used.
transfer of digital documents. (Web pages, files,
MP3 tracks...)
5.2 Dimensioning for Stream Traffic
5.1 Dimensioning for Elastic Traffic Streaming flows are produced by audio and video ap-
plications. Different video compression schemes,
The elastic traffic is always carried by the TCP pro-
such as H.261, H.263, MPEG-1, MPEG-2, and
tocol. It is thus important that we have a queuing dis-
MPEG-4, are designed to meet different objectives
cipline of a simple elastic type, which works essen-
and therefore have different bit rates and stream
tially as Processor Sharing (PS) (i.e. not FIFO)
characteristics. The video characteristics are also
queues, so that the great files will not delay the small
affected by the video encoding control scheme used.
ones too much and so that there will be some fair-
For a given content and a given compression scheme,
ness between the files on each link [8]. In a network
constant bit rate (CBR) video maintains a streaming
with DiffServ-enabling, the elastic traffic applica-
rate that varies little over time. By contrast, variable
tions always are accommodated by AF PHB services.
bit rate (VBR) video traffic has been shown to be
Bandwidth allocation among different AF classes can
self-similar and may have a peak rate which is many
be implemented by the Weighted Fair Queuing
times the average. CBR is easier buffer management
(WFQ) packet service discipline, or other variants of
and on the network, but not bandwidth efficient.
WFQ. A fraction C i = iC of the total capacity
Whereas, VBR is bandwidth efficient, but is bursty
C will be supposed to be allocated to each class, the
and buffer management required. In the BREAN,
packets of which are handled in separate queues. We
Ethernet can provide more cheap bandwidth capacity
assume that each queue processing capacity is con-
than the other techniques. We sacrifice bandwidth
efficiency in order to simplify network design, and together with the traffic characteristics of each ser-
use CBR encoding control scheme for the stream vice type and the network structure determine the
flows. M. Grossglauser et al. investigated the per- traffic which is offered to the network. H.E. Hanra-
formance of CBR traffic in the context of large- han defined an individual user or average user in
scale networks, and listed several reasons that CBR terms of atomic elements corresponding to the pure
is interesting [23]. A Poissan stream is a good ap- uses of the pure services by typical users [13]. Thus
proximation to a superposition of many CBR a single real user has a profile expressed as a linear
streams with differing phases and bandwidths. array of percentages, containing the percentage use
Voice connections, in general, generate a stream of made of each of the N services. That is,
small packets of similar size (a few tens of bytes) at P = [S1, S 2 , ... , S N ] . (3)
relatively low bit rates, and exhibit the low bursti-
Each element of P represents the user's probability
ness [9]. Usually, fixed-rate speech coders are em-
of making use of each of a set of N services in the
ployed. Such coders generate a constant output bit
set. In addition, for an access network with DiffServ-
rate, independently of network conditions [10].
enabling, each priority type may consist of one or
Bonald et al. noted that if all flows are initially CBR
more services. Similarly, the priority types mean EF
and in all queues:
type (or multiple instances of EF) and AF groups
flow rates < service rate . (2) here. A more complex model would also have col-
They never acquire sufficient jitter to become worse umns introduced to represent the probability of a
for performance than a Poisson stream with MTU service being used at different time intervals.
packets [11]. So, dimensioning of CBR stream traf- Some services, such as TV programs, use multicast-
fics may simply deploy this engineering rule. ing mechanism to save substantial downstream ca-
Dimensioning for VBR stream traffic is a more pacity. Thus, the capacity requirements of each user
complicated task. The problems of admission con- are up to how many other users want to receive the
trol and capacity planning in a packet network may be same multicast traffic as well as their location in the
addressed by a concept known as the effective network. To deal with probabilistic models of the
bandwidth or equivalent bandwidth of a connec- sharing of multicast traffic become too computa-
tion [12], but it is outside scope of this paper. tionally demanding in an optimisation. Peter Wang et
al. gave a method to estimate the traffic volume of
multicasting with assuming that a 90/10 rule applies,
6 Traffic Matrix and User whereby 10 percent of the programs are requested
Model by 90 percent of the users, and conversely, the re-
maining 90 percent of the programs are requested by
6.1 Traffic Matrix 10 percent of the users [14]. We give a more generic
model for the total number of streams required by
In classical teletraffic theory, an origin-destination the multicasting applications on a link x :
traffic matrix for a particular service, such as te- Fx = min {(1-q)N,pM} + min{qN,(1-p)M} (4)
lephony, is applied to the network dimensioning. We
where M is the number of TV programs available and
adopt the simple origin-destination matrix to multi-
N is the number of active home viewers downstream
service access networks with DiffServ-enabling by
from link x . p is the fraction of popular TV pro-
adding an extra dimension for the different priority
grams, and q is the fraction of users viewing special
of traffics. The origin-destination matrix based on
interest TV programs, p, q [0,1] .
DiffServ architecture is:
Obviously, p or q is determined by the demograph-
V (i,j,k), i=1,...,r ; j=1,...,r ; k=1,...s , ics and the size of the daily audience. We can get the
where s is the total number of priority types. Noted, two values based on historical statistics and some
the priority types mean EF type (or multiple in- surveys for TV viewers. There are some studies to
stances of EF) and AF groups here. For a given prior- investigate the estimation of the size of a broadcast
ity k, a measure of the volume of traffic sourced by audience. With the convergence of TV and the Inter-
th
users of the k priority in the origin i and destined net it will be possible to provide real-time and pre-
for the destination j is Nijk . cise audience estimation [15]. Thus, we can choose
appropriate p and q for the realistic environment.
However, the discussion about broadcast audience
estimation is outside scope of this paper.
6.2 User Model

The user model is an essential ingredient in deter-


mining the network capacity for the traffic of-
fered to the network. Individual users make use of a
particular set of services at particular times and this,
7 A Pragmatic Engineering S 11
Method and Simulation Study S 21

...
N1
S 1x
7.1 A Pragmatic Engineering Method R 11

...
S 11
We present a pragmatic method of network dimen- S 2x

...
sioning of the BREAN with DiffServ-enabling below.
S 1x
Our method mainly considers the generic applica-
R 21 N2
tions, i.e. TV, VoD, VoIP and WWW. The residential S 11
access network is an asymmetric network, and down-

...
S 2x+1
stream traffic is much more than up-stream traffic
now. Thus, we dont consider the network dimen- S 1x

sioning for the up-stream traffic. R 12

...
S 11 N3
We assign EF1 for VoIP, EF2 for TV and EF3 for VoD.

...
S 2m
WWW browsing with guaranteed minimum capacity
is implemented by AF 1. We deploy admission con- S 1x

trol scheme for VoIP, TV and VoD, and traffic condi- S: Switch
R: Router
tioning is deployed for VoIP, TV, VoD and WWW N 1: PSTN; N2 : IP Netw ork; N 3 : the Network s Offering TV and VoD Service
browsing services. Noted, in the L2 domain, these Figure 3. An Example of Access Network Architecture
services are assigned the same mechanisms men-
tioned in Section 3. Noted, we only deploy WRED In Table 1, we give the assumptions about the user
mechanism for WWW browsing AF 1 queue. model. The real user model can be derived from his-
We consider TV and VoD using MPEG-2 based on torical statistics and some surveys.
CBR with 6 Mbps [1]. We deploy G.723.1 [16] with For pure VoD, we can simply deploy Equation 2 for
5.33Kbit/s. For an Ethernet LAN with IEEE 802.1q, network dimensioning. However, for TV programs
a 12 byte RTP header, an 8 byte UDP header, a 20 with multicasting mechanism, we need to combine
byte IP header, and a 30 byte 802.3 MAC header are Equation 2 with Equation 4 for network dimension-
appended to voice packets, namely, a total header ing. Obviously, p and q will seriously affect the
size H = 70 bytes . Mansour J. Karam investigated capacity planning for the multicasting applications.
the effects of packet size on voice delay and band- We can determine the two values based on historical
width utilization, and noted that G.723.1 encoder, us- statistics and some surveys for TV viewers. In
ing 30 ms formation time leads to a negligible waste addition, the links in the BREAN may choose the
in resources, and constitutes a good compromise in different values of p or q based on the aggregating
terms of delay and bandwidth utilization (that is, the level of traffic. For example, we can assume that all
recommended packet size is 20 bytes) [17]. Thus, a end users wholly watch different TV programs for a
total size of each voice packet is 90 bytes. In other link between S1x and S2x, because this link level
words, a voice call with G.723.1 encoder produces doesnt embody the statistical effect of watching
24Kbps traffic. behaviour for the audiences. Thus, p =0 , q=1 , and
An example of access network architecture is illus- the total number of streams on a link between S1x
trated in Figure 3. We assume each user connects the and S2x: F = min {N , M} . (5)
x
access Switch, i.e. S1x with 10Mbps Ethernet link, For the high-speed links aggregating more traffic,
and the Ethernet link between S1x and S2x is such as the links between Node S2x and Node R1x or
100Mbps, the link between S2x and R1x is 1000Mbps. between Node R1x and Node R21, we assume that
Next, we combine the user model and the traffic ma- there are 100 TV programs, and 20 percent of the
trices to discuss the network dimensioning of the programs are requested by 80 percent of the users,
BREAN with DiffServ-enable. and conversely, the remaining 80 percent of the pro-
For WWW browsing or FTP applications, we can de- grams are requested by 20 percent of the users.
ploy the procedure of network dimensioning for Hence the total number of streams is given by:
Fx =min {0.8N,20} + min {0.2N,80}.
elastic traffics mentioned in Section 5.1. We assume
(6)
no admission control is applied for WWW browsing
and FTP applications, capacity allocation could be So, choosing the values of the p elaborately is a vi-
performed according to a target performance speci- tal step for network dimensioning. In addition, for
fied in terms of average flow throughput. We assume stream traffic, there is a type of per-flow admission
that the access rate constraint of each users WWW control schemes, and the blocking probability de-
browsing is 1 Mbps and WWW browsing of each termined is used to dimension the network.
user is given 1 Mbps throughput guarantee. Accord- Medina, et al. presented taxonomy of IP traffic ma-
ing to Equation 1, we can compute the average flow trices [18]. The residential access network is a star-
throughput with access rate constraint and the Table structured topology network typically. Thus, we dis-
3 give the result below. cuss the Point-to-Point and Point-to-Multipoint traf-
fic matrices with link level or Switch/Router We, firstly, need to determine the total numbers of
level .for network dimensioning. residential customers supported by 100Mbps
We, firstly, need to determine the total numbers of Ethernet between Node S1x and Node S2x. We can
residential customers supported by 100Mbps derive the maximum value of N based on the N-home
Ethernet between Node S1x and Node S2x. We can aggregating traffics that sum up the traffics of the
derive the maximum value of N based on the N-home applications, i.e. VoIP, TV, VoD and Internet access.
aggregating traffics that sum up the traffics of the So, the Point-to-Point traffic matrix with link level
applications, i.e. VoIP, TV, VoD and Internet access. between Node S1x and Node S2x, i.e. V (S1x ,S2x ,K) ,
So, the Point-to-Point traffic matrix with link level is determined, where K is the number of priority
between Node S1x and Node S2x, i.e. V (S1x ,S2x ,K) , types. In other words, network dimensioning assigns
is determined, where K is the number of priority the capacity for the different EFx or AFx services.
types. In other words, network dimensioning assigns
the capacity for the different EFx or AFx services.
Table 1. An example of the user model with DiffServ-enabling
Priori- Applications User Model Average Busy N-Home Aggregated Traffics
ties Hours per-
Home Traffics
AF2 WWW, FTP 50% of homes with Null [0.5N] with 1Mbps throughput guar-
1Mbps throughput antee
guarantee
AF3 Pure VoD 25% of homes, 1.5Mbps N*1.5Mbps
MPEG-2 6 Mbps
AF4 TV One program /homes Null Fx =min {(1-q)N,pM } + min {qN,(1-p)M }
MPEG-2 6 Mbps with *6Mbps
100 TV programs
EF VoIP 25% of homes, bi- 5.8Kbps N*5.8Kbps
directional, 224 Kbps

Next, we use the Point-to-Multipoint traffic matrix router packet forwarding capability and buffer capac-
{ }
with link level, i.e. V (S, D1, D2 ,...Dn ,K) , for ity on QoS provisioning. In addition, we deploy non-
preemptive priority scheduling with 4 queues for
capacity planning of up-link from a high-level node S VoIP, TV, VoD and WWW respectively in both L2
{ }
to a set of lower-level nodes D1, D2 ,... Dn . Thus, domain and L3 domain.
Noted, we generated the WWW traffic by explicitly
we can determine the capacity of up-link for Node modelling the interactions of HTTP 1.1 with the fol-
S2x, R1x and R2x with DiffServ-enabling, respectively. lowing entities: size of requested index objects,
The other important task is to determine the capacity number and size of corresponding embedded objects
between edge router i.e. R21 and the MAN or WAN and think time between two successive down loads.
network. Distributions for all parameters are depicted in Table
In addition, the method of these static traffic ma- 2. This model generates a mean load around 800
trixes above is deployed for network dimensioning Kbps at sufficient bandwidth.
of a new access network and expansion of network
capacity later. There are some studies about dynamic Table 2 Distribution of WWW Traffics Parameter
traffic matrixes [19-20], and these methods may Parameter Distribu- Average Shape
amend the static traffic matrixes for network dimen- tion
sioning later. Size of Index Pareto 8000Byte 1.2
Obj.
Size of Em- Pareto 4000Byte 1.1
7.2 Simulation Study bedded Obj.

Our simulation environments are same as those as- Number of Pareto 20 1.5
sumptions about network topology, source traffics, Embedded
QoS mechanisms, etc. in section 7.1. In order to ac- Obj.
curately validate the effect of our method of network Think Time Pareto 30 sec 2.5
dimensioning, we also assume that bottleneck of the
BREAN only lies on link bandwidth capacity. In We simulate the BREAN with different aggregating
other words, we dont consider the impact of termi- traffics including source traffic and cross traffic,
nal equipment performance, server performance,
such as TV, VoD, VoIP and WWW. In Table 3, we cates the BREAN can well offer bounded delay and
give each flow throughput guarantee, allotted band- delay variation for TV, VoD and VoIP without packet
width for WWW browsing and the different links loss. The delay and delay variation of TV, VoD and
bandwidth utilization. Noted, source traffic and cross VoIP in case 1 are illustrated in Figure 4. In case 2,
traffic load of different simulation cases based on case3, case4, the delay and the delay variation of the
the computation of our method of network dimen- applications increase with an increase of traffic
sioning. For example, in case 1, we model 3 residen- loads we generated. But the BREAN can still offer
tial customers with WWW browsing and the other 3 bounded delay and delay variation without packet
residential customers with TV, VoD, VoIP and loss with high bandwidth utilization. These simula-
WWW respectively. We dont restate the details of tion results validate our method of network dimen-
source traffic and cross traffic loads in the other sioning and are in accord with the computation of
cases here. our method in Section 7.1. The delay and delay varia-
We give of the six simulation cases in the figures tion of TV, VoD and VoIP in case 4 are illustrated in
below. In case 1, we investigate the QoS provision- Figure 5.
ing with light burden, and the simulation result indi-
Table 3 Each Flow Throughput Guarantee, Allotted Bandwidth and Bandwidth Utilization
Case Allotted Bandwidth Each Flow Bandwidth Utilization of Bandwidth Utiliza-
for WWW Brows- Throughput the 100Mbps Link Be- tion of the other
ing (Mbps) Guarantee tween S1x and S2x 1000Mbps Link
(Mbps)
1 77.5 1 28.1% 2.81%
2 10 0.993 97.8% 99.85%
3 8.5 0.966 98.73% 100%
4 7 0.957 99.1% 100%
5 5.5 0.392 99.1% 100%
6 4 Null 99.1% 100%

In case 5, delay and delay variation for TV and VoD cant offer QoS provisioning for those real-time ap-
can still meet the QoS requirement, but the delay plications, and the delay and the delay variation of
maximum of VoIP reach 192ms, which is beyond the these applications degrades drastically. The delay
normal delay coverage of VoIP in BREAN. The delay and delay variation of TV, VoD and VoIP in case 6
and delay variation of TV, VoD and VoIP in case 5 are illustrated in Figure 7.
are illustrated in Figure 6. In case 6, the BREAN

Figure 4 Delay and Delay Variation of the Applications with Light Burden. Bandwidth Utilization of the
100Mbps Link Between S1x and S2x is 28.1%; Bandwidth Utilization of the other 1000Mbps Link is 2.81%. The
BREAN can well offer bounded delay and delay variation for TV, VoD and VoIP without packet loss.
Figure 5 Bounded Delay and Delay Variation of the Applications with High Bandwidth Utilization. Bandwidth
Utilization of the 100Mbps Link Between S1x and S2x is 99.1%; Bandwidth Utilization of the other 1000Mbps
Link is 100%. In this case, the bandwidth capacity allotted for www browsing is can approximately meet the re-
quirements of our network dimensioning method, and well meet the requirements of our network dimensioning
method for the other applications. The BREAN can still offer bounded delay and delay variation with an
increase without packet loss.

Figure 6 Degraded Delay and Delay Variation of the Applications with High Bandwidth Utilization. In this case
we generate more traffic loads, but bandwidth utilization of the 100Mbps Link Between S1x and S2x is same as
Figure 5. More traffic loads dont result in higher bandwidth utilization. The delay and delay variation for TV
and VoD can still meet the QoS requirement, but the delay maximum of VoIP reach 192ms, which is beyond the
normal delay coverage of VoIP in BREAN. In this case, bandwidth capacity allotted for www browsing cant
meet the requirements of our network dimensioning method, and each flow throughput guarantee of computa-
tion result is only 0.392Mbps.
Figure 7 Unacceptable Delay and Delay Variation of the Applications with Overburden. In case 6, more traffic
loads result in failing to offer QoS provisioning for those real-time applications, and the delay and the delay
variation of these applications degrades drastically.

If we also investigate the average throughput per-


formance of the WWW elastic flows combining
TV, VoD and VoIP, we can find the real reason of
Conclusion
performance degradation. From case 1 to case 4,
Combing network dimensioning and QoS mecha-
the networks allot enough bandwidth resource for
nisms is an appropriate way to implement QoS pro-
www flows, and approximately promise the ave r-
visioning in the BREAN over DiffServ framework.
age throughput performance of the WWW elastic
We discuss the main components of network dimen-
flows. Whereas, in case 5 and in case 6, we dont
sioning framework and give a rather pragmatic
allot enough bandwidth guaranteeing each flow
method of network dimensioning in the BREAN over
throughput for WWW browsing, www flows have
DiffServ, which offers PTOS, VoIP, TV, VoD and
a negative impact on the other higher priority
Internet access. The simulation studies validate our
flows, such as VoIP, TV and VoD. Specially, in
method of network dimensioning. The implementa-
case 6, www flows results in the severe perform-
tion of DiffServ in the BREAN, which offers deter-
ance degradation of the real-time applications in
ministic guarantees or statistical guarantees, is still
the BREAN. Indeed, the BREAN hasnt promised
in progress.
QoS provisioning of the real-time applications,
i.e. VoIP, TV and VoD, in case 6. So, the BREAN
must deploy traffic control and allot enough
bandwidth for the elastic applications, such as
Acknowledgement
WWW browsing, FTP, etc.
The authors would like to thank to the colleges in
The simulation results illustrate that if the BREAN
Multimedia Communication& Network Engineering
allots bandwidth capacity for each service based on
research center for their collaboration. Noted, this
this method, it can promise multi-service QoS provi-
research was supported in part by the Projects of
sioning without overburden always. In other words,
Development Plan of the State Key Fundamental Re-
when the traffic loads of the different priority ser-
search in China under grant GZ 995337. The Con-
vices in the BREAN are below bandwidth capacity
ference Participation is supported by Nokia Bridging
planned by our method of network dimensioning, the
the World Program.
BREAN can promise QoS provisioning of the real-
time applications combining with some QoS mecha-
nisms. We argue that the network dimensioning is
indispensable for the residential multi-service IP-
based access networks.
References [11] T. Bonald, A. Proutire, J. Roberts, "Statistical
performance guarantees for streaming flows
[1] Zhuang Zhao, Qiong Liu, and Zhi-mei Wu, "The using expedited forwarding", in the Proceedings
Broadband Residential Ethernet Access of IEEE Infocom, Anchorage, AK, USA, April
Infrastructure: Architecture, Service & Traffic 2001.
Mode, QoS Provisioning and Research Issues," [12] A. W. Berger and W. Whitt, Extending the
in the Proceedings of the 9th International effective bandwidth concept to networks with
Conference on Telecommunications (ICT priority classes, IEEE Communications
2002), Volume3, pp.267-276, Beijing, China, Magazine, August 1998, pp. 78-83.
June 2002. [13] H.E. Hanrahan, et al., Framework for Access
[2] Zafer Sahinoglu, et al., On multimedia Networks: Network Traffic Modeling for Next Generation
Self-Similar Traffic and Network Networks ITC 14th Specialists Seminar on
Performance, IEEE Communications. Mag., Access Networks and Systems,Girona, Spain,
January 1999, pp. 48-52. April 25-27, 2001.
[3] Jim W. Roberts, Traffic Theory and the Internet, [14] Peter Wang, et al., Ethernet Growing UP: The
IEEE Communications. Mag., January.2001, pp. Extensible Broadband Access Infrastructure,
94-99. ISSLS 2000, Stockholm. Sweden, June 2000.
[4] D S. Blake, et al., An Architecture for [15] Chuanhai Liu, and Jorg Nonnenmacher,
Differentiated Services, RFC 2475, BBN, Broadcast Audience Estimation, In the
December 1998.Anton Riedl, Thomas Proceedings of IEEE INFOCOM 2000, pages
Bauschert, Jochen Frings, A Framework for 421430, Tel-Aviv, Israel, April 2000.
Multi-Service IP Network Planning, in the [16] Recommendation G.723.1, Speech Coders:
Proceedings of the 10th International Dual Rate Speech Coder for Multimedia
Telecommunication Network Strategy and Communications Transmitting at 5.3 and 6.3
Planning Symposium (Networks 2002), Munich, kbit/s,ITU, March 1996.
Germany, June 2002. [[17] Mansour Karam and Fouad Tobagi, Analysis of
[5] A. Charny, J.-Y. Le Boudec, Delay Bounds in a the Delay and Jitter of Voice Traffic over the
Network with Aggregate Scheduling," in the Internet,In Proceedings of the IEEE Infocom
Proceedings of QoFIS'2000, September 25-26, 2001 Conference, Anchorage, Alaska USA,
2000, Berlin, Germany. April 2001.
[6] Anton Riedl, Maren Perske, Thomas Bauschert, 18] A. Medina, C. Fraleigh, N. Taft, S. Bhattacharyya,
Andreas Probst, Investigation of the M/G/R C. Diot, "A Taxonomy of IP Traffic Matrices," in
Processor Sharing Model for Dimensioning of the Proceedings of SPIE/ITCOM., Boston, USA,
IP Access Networks with Elastic Traffic,in the June 2002.
Proceedings of First Polish-German Teletraffic [19] Alberto Medina, N. Taft, K.Salamatian, S.
Symposium PGTS 2000,Dresden, Germany, Bhattacharyya, C. Diot, Traffic Matrix
September 2000. Estimation Techniques: Existing Techniques
[7] Anton Riedl, Maren Perske, Thomas Bauschert, Compared and New Directions, in the
Andreas Probst, Dimensioning of IP Access proceedings of SIGCOMM 2002, Pittsburgh,
Networks with Elastic Traffic, in the PA. August 2002.
Proceedings of the 9th International [20] Van der Zijpp, N.J., A Comparison Of Methods
Telecommunication Network Planning For Dynamic Origin-destination Matrix
Symposium (Networks 2000), Toronto, Estimation, in the proceedings of the 8th
CanadaSeptember 2000. IFAC/IFIP/IFORS Symposium Vol. 3, Chania,
[8] Karl Lindberger, Balancing Quality of Service, Greece, 16-18 June 1997, Elsevier, pp 1375-
pricing and utilization in multiservice networks 80.
with stream and elastic traffic, in the [21] J.W. Roberts, P. Olivier, IP Link Dimensioning
proceeding of ITC16, Edinburgh, Scotland, June in the Context of DiffServ,COST 257TD (00)
1999. 01, 2000.
[9] Mansour Karam and Fouad Tobagi, "Analysis of [22] D S. Blake, et al., An Architecture for
the Delay and Jitter of Voice Traffic over the Differentiated Services, RFC 2475, BBN,
Internet", Proc. of Infocom 2000, Tel Aviv, December 1998.
Israel, March 2000. [24] M. Grossglauser, et al, On CBR Service,Proc.
[10] C. Casetti, et al, Adaptive Voice over IP, IEEE Infocom '96, San Francisco, California,
Source Modeling and Encoding Workshop, March 1996.
Viareggio, Italian, July 1999.

You might also like