service control in MPLS essentially means that
bandwidth can be reserved for traffic flows.
MPLS networks are accessed through ELRs
and contain LSRs internally. Packets arriving
at ELRs are inspected with respect to destination
and classified into flows. The classification may
be extended to include other attributes such as
source, application class, etc. Flows are associ-
ated with unique flow labels which are used by
LSRs instead of IP addresses to switch/route
packets through the network. All packets of a
flow may thus follow the same path (a so-called
label switched path), but packets can also be
classified so that packets relating to different
sources or applications may follow different
paths even if the destinations are identical.
As can be seen there are several similarities
between LSPs and other well-known concepts,
e.g. VPs or VCs in ATM. The flow labels corre-
spond to VPIs or VCIs and the LSRs correspond
to ATM cross connects or ATM backbone
switches. On the other hand, VPs and VCs in
ATM are established through the management
system or by signalling, while LSPs are estab-
lished by the IP-based LDP [3], e.g. CR-LDP
[4] or RSVP-TE [5]. See also [6].
3 Multi-Service Networks
Different types of traffic have different require-
ments in terms of bandwidth consumption and
sensitivity to delay or loss of information. To
provide sufficient quality of service either the
amount of transmission resources must be set to
ensure that the most stringent requirements are
satisfied for all flows, or a discrimination mech-
anism must be used that allows each flow to
obtain its required quality of service.
The former approach is simple but is usually
regarded as economically viable only if the
traffic type with the most stringent requirements
is the dominant one in terms of traffic volume.
The major traffic types today are real time voice
and best effort data. Voice has high requirements
on delay whereas data has high requirements on
loss. Although voice is still dominant in many
networks, data is growing fast and is gradually
becoming dominant in most networks. More-
over, it is expected that in the not so distant
future video retrievals, the requirements of
which are typically between voice and data,
will make up a large part of the traffic in a net-
work. The conclusion is that it does not seem
realistic to satisfy the most stringent require-
ments for all traffic types, but some kind of
differentiated quality is required.
Clearly, discrimination mechanisms only work
within the limits given by the total capacity of a
system, but they cannot resolve problems related
to overload. To prevent the latter, overload pro-
tection in the form of connection admission con-
trol and packet policing is usually applied at the
edge of a network. To determine the rules
according to which these mechanisms should
operate, the impact of various loads is studied
by mathematical traffic models. The results are
often presented as a safe region of operation, and
the mechanisms are tuned to ensure that the load
stays within this region.The classical view of multi-service networks is
the integratedone where all service classes, ori-
gins and destinations are multiplexed together
directly on the physical network. The difficulty
with this approach is that the traffic models are
quite complex, and that combined models of all
traffic types are even more complex because of
the additional difficulties in combining the dif-
ferent kinds of models used for different service
classes. Considerable simplifications must thus
be made to obtain a tractable result. The conse-
quences are that the accuracy of a safe region
is questionable and additional margins must be
added, and that there is no direct link between
modelling errors and performance problems
since problems for one service class may be
related to modelling errors for any class or for
a combination of service classes.A contrasting view is the separatedone where
all service classes, origins and destinations are
supported by dedicated logical end-to-end links,
such as LSPs, which identify routes and reserve
bandwidth. In this approach transmission re-
sources are partitioned between various service
classes and node pairs. This means that admis-
sion controls operate on dedicated resources
based on single class traffic models and there is
no need for joint models of all classes. Conse-
quently, modelling errors for one class will not
impact other classes and a failure in meeting a
performance target for a certain class is cor-
rected by modifying the particular model in
question. Moreover, new classes can be added
to the network without reworking the complete
model of all classes and without impairing the
performance of other service classes.It may be argued that this way of partitioning
will be less efficient in exploiting statistical mul-
tiplexing gain. However,- slowvariations may be handled just as effi-
ciently by redistributing the resources between
the logical links [7, 8, 9, 10]; for - fastvariations and service classes with differ-
enttime scale characteristics, it has long been
known that there is little or no gain to be had
from multiplexing [11]; and for