159

1 Introduction

Introducing high capacity links in IP-based net-

works, with differentiated QoS, allows delivery

of services with highly variable characteristics.

As we know the IP protocol provides statistical

multiplexing between user applications that may

generate packets of highly variable lengths. For

typical real time services the main QoS parame-

ters are end-to-end delay and jitter, and we must

be aware that the main contributions to these

parameters will come from low capacity links

(in the access network). This justifies the need to

put special emphasis on the link layer protocol

and the multiplexing structure in the access net-

work.

The access network will encompass a variety of

different access technologies that are currently

available. These can be divided according to

• Fixed access, or


• Mobile access.

With the recent advances in access technology

the fixed access may be a mixture of one or

more different types such as Asymmetric Digital

Subscriber Line (ADSL), Very high speed Digi-

tal Subscriber Line (VDSL), Coax and optical

fibre, all having very different physical charac-

teristics. The logical structure of the access net-

work may therefore be very different.

For mobile access the radio medium has limited

bandwidth implying that the available bitrate for

each user will be limited.

The link layer protocol structure in the access

network will therefore be very different depend-

ing on the actual physical technologies applied.

In Universal Mobile Telecommunication System

Terrestrial Radio Access Network (UTRAN) the

current link protocols are based on ATM (AAL2

or AAL5), however, there is a common trend to

try to minimise the use of circuit-like protocols

and instead deploy IP also in the radio network.

The multiplexing of IP packets over ATM has

some desired features due to the fact that the

ATM cells are rather short and have fixed

lengths, thereby avoiding large delay variation

due to long packets.

Traditionally there has been quite a strict distinc-

tion between the access network and the core

(transit) network, where the access is defined as

the part of the network from the subscriber to the

local exchange. By increasing the line speed by

introducing different active components this def-

inition of where the access network ends and

where the core (transit) network starts are not

directly valuable any more. In IP networks the

definition seems to be more flexible on the basis

of more functional distinctions. Usually one will

define the core network as the part of the net-

work where DiffServ and/or MPLS is deployed.

By the increased line speeds it is however an

interesting question to find out how ‘far’ out in

the ‘old access network’ the DiffServ model

(and possible MPLS) is effective.

2 A Model Evaluating the IP

Multiplexing Problem

for Low Capacity Links

Multiplexing traffic of different types on the IP

level may cause delay and jitter problems if

these traffic types share a link with rather low

capacity. The main cause for this delay and jitter

is the variation in the packet lengths for the dif-

ferent traffic types. While typical real time traf-

fic like voice will emit packets of a small fixed

size, the typical data application may generate

packets that are quite long. Due to this mismatch

in packet size between different applications the

queuing delay for typical real time traffic may

IP Multiplexing for Low Capacity Links?

OLAV ØSTERBØ

In this paper we address the well known multiplexing problem in IP-networks containing links with low

capacity. Due to the highly variable packet lengths real time traffic may experience transmission with

unacceptable delays and jitter that may cause degraded quality.

Even if priority mechanisms are implemented in the routers (e.g. DiffServ is implemented), this will not

completely solve the problem unless some kind of fragmentation of long IP packets is performed.

To study this negative multiplexing effect we have taken a non-preemptive priority queuing model, which

will give the best performance for the high priority traffic classes if no fragmentation is performed. As a

second model describing the effect of fragmentation we have taken a non-preemptive priority queuing

model with batch arrivals where the size of a batch corresponds to the number of fragments an IP

packet will consist of. The numerical examples presented show that the critical link capacity lies around

2 Mbit/s if the maximum packet length is limited to 1500 bytes.

Olav Østerbø (48) received his
MSc in Applied Mathematics
from the University of Bergen
1980 and joined Telenor R&D
the same year. His main inter-
ests include teletraffic modelling
and performance analysis of
various aspects of telecom net-
works. Activities in recent years
have been related to dimension-
ing and performance analysis of
ATM networks and now moving
towards IP networks, where the
main focus is on modelling and
control of different parts of next
generation IP-based networks.

olav-norvald.osterbo
@telenor.com

Telektronikk 2/3.2001