Telektronikk 2/3.2001
1 Background
Two major trends have dominated the telecom-
munications industry over the past decade, viz.
wireless cellular systems such as GSM and
packet switched services over IP. The two tech-
nologies have largely evolved independently of
each other and the services are basically sepa-
rated in the networks.
This split approach is however set to change for
technological and commercial reasons. On the
technology side, high speed packet switching is
an integral part of third generation cellular sys-
tems, e.g. UMTS, and packet switching is be-
coming deployed in fixed access networks, e.g.
CATV. On the commercial side, competition
eats into the margins of traditional wireline ser-
vices while the growth of packet switched ser-
vices exposes the costs associated with separate
networks.
There is thus a growing need for an integrated
services network. This is by no means a new
idea, in fact it has been a vision at least since
the seventies, although the technologies have
differed: ISDN over STM in the seventies, B-
ISDN over ATM in the eighties and everything
over IP in the nineties. The visions have how-
ever remained visions, albeit for different rea-
sons. ISDN/STM was made obsolete by the
development of the computer industry, B-ISDN/
ATM was seen as too costly and too complicated
and all-IP was questioned in relation to quality
of service.
2 MPLS
The favourite candidate for service integration is
MPLS [1]. The fundamental idea behind MPLS
can be characterised as enhancing IP with some
quality of service concepts from ATM. From
this point of view, a primary feature of MPLS is
the ability to perform traffic engineeringand a
secondary feature is quality of service control[2].
IP networks typically use OSPF or a similar pro-
tocol to find shortest routes between points. The
fact that there is only one such route from any
node to any other node may lead to situations
where links on shortest routes are congested
while other links remain idle. Traffic engineer-
ing in MPLS essentially means that traffic flows
can be controlled in order to balance link loads.
Moreover, classical IP networks typically sup-
port only one service class, viz. best effort.
Although proposals such as IntServ and DiffServ
have been around for some time, they have not
yet gained widespread acceptance, let alone
deployment. IntServ suffers from scalability
problems which makes it unsuitable for back-
bone networks, while the ability of DiffServ to
provide quality of service is doubted. Quality of
The Design of Optimal Multi-Service
MPLS Networks
ÅKE ARVIDSSON AND ANTHONY KRZESINSKI
Anthony Krzesinski (55) obtain-
ed the MSc from the University
of Cape Town and the PhD from
Cambridge University, UK. In
1972 he joined the Shell Re-
search Laboratory in Amster-
dam where he worked on the
development of mathematical
models to predict the perfor-
mance of computer systems. In
1975 he joined the Department
of Computer Science at Univer-
sity of Stellenbosch. He is
presently a Professor of Com-
puter Science at the University
of Stellenbosch. His research
interests centre on the perfor-
mance evaluation of communi-
cation networks.
[email protected]
Åke Arvidsson (43) obtained the
MSc and PhD from the Lund
Institute of Technology, Sweden.
Between 1990 and 1992 he
worked at Bond University and
the University of Adelaide in
Australia. From 1993 to 1995
he was with the Lund Institute
of Technology and in 1995 he
joined the Blekinge Institute of
Technology, Sweden, as Profes-
sor of Teletraffic Theory. Since
1998 he is on leave and works
for Ericsson Core Network
Development as Technical
Expert on Data Traffic Theory.
His professional interests in-
clude traffic modelling, perfor-
mance evaluation, network
architecture and load control.
[email protected]
Multiprotocol label switching (MPLS) extends the IP destination-based routing protocols to provide new
and scalable routing capabilities in connectionless networks using relatively simple packet forwarding
mechanisms.
MPLS networks carry traffic aggregates on virtual connections called label switched paths (LSPs). The
first part of this paper examines under what circumstances it is advantageous to design dedicated LSPs
for individual origin-destination pairs and service classes. We show that separate LSPs in most realistic
cases are likely to be the preferred mode of operation.
We next consider path selection and bandwidth allocation in multi-service MPLS networks in order to
optimise the overall network quality of service. The optimisation is based upon the constrained optimisa-
tion of a non-linear objective function. We present a model of an MPLS network and a computationally
efficient algorithm called XFG to find and capacitate optimal LSPs. The algorithm is based on a band-
width market where bandwidth prices determine the allocation of bandwidth to LSPs. The XFG algo-
rithm is applied to compute optimal LSPs for a 55 node network model carrying 6 service classes.
The results above are limited to service classes typically supported by UDP, e.g. conversational voice
and streaming video, where the notation of equivalent bandwidth can be applied. This is, however, not
the case for service classes typically supported by TCP, e.g. interactive or background data, because of
the responsiveness of the protocol. We therefore extend our work to incorporate these types of traffic
and apply the XFG algorithm to compute optimal LSPs for a network of 8 nodes and 2 service classes.
Finally we use core networks of third generation cellular mobile systems as an example to show how
the method can be generalised to any multi-service network and we also discuss how to include virtual
private networks.
116
