- fastvariations and service classes with similar
time scale characteristics, it has been sug-
gested that the gain obtained from multiplex-
ing tends to be outweighed by increasing
overhead costs [12]. - Moreover, as indicated above, a partitioned
scheme permits more aggressive multiplexing
than an integrated one since the former avoids
the compromises of a joint traffic model, but
permits the use of more accurate single class
traffic models.
The reason behind the first point is that the set of
LSPs may be re-designed as required. The pro-
posal in [7] requires that ELRs monitor offered
traffic during intervals of tMtime units, with
new intervals commencing every tUthtime unit.
Traffic estimates are forwarded to an NMC
which computes updated LSPs and analyses the
result. If implementing the new design appears
profitable, the necessary information is sent back
to the ELRs and the design is implemented by
an LDP. Transmitting traffic information to the
NMC, computing and analysing the design,
returning results to the nodes, and implementing
the design is assumed to take a total of tEtime
units. There is a trade-off between the resources
spent on management actions such as altering
LSPs and the associated increase in carried traf-
fic. To compare the two and optimise the strat-
egy, it is proposed in [7] to associate a profit for
carried traffic and a cost CTfor each updating
attempt (transmission of data to the NMC, com-
puting and analysing a design), and a cost CIfor
implementing a new design. Other cost models
and similar proposals which are independent of
the NMC are discussed in [8, 9] and in several
papers in [10].
The second point simply says that the statistics
on which multiplexing relies must apply to all
classes. This means, for example, that it must be
possible and meaningful to buffer one class dur-
ing the busy periods of another class. The early
work reported in [11] showed that the multiplex-
ing gains obtained from mixing voice and data
are limited to quality of service improvements
but do not impact decisions regarding engineer-
ing. The work considered the SENET concept
where bandwidth is divided in time between
voice and data. The boundary between voice
segments and data segments could either be
movable or fixed depending on whether band-
width reserved for, but not used by, voice was
made available to data or not. The work in [11]
showed that the gains obtained from the mov-
able boundary were limited by the fact that the
dynamics of voice (connection holding times)
are very slow compared to the dynamics of data
(packet inter-arrival times). The fact that the
bandwidth for data is controlled by voice means
that data will experience good service when few
connections are in progress but poor service
when many connections are in progress. In fact,
the work showed that during the latter intervals
data traffic will be congested to the extent that
the service appears useless. This means that,
because of the different time scales, there is no
statistical gain to be obtained from multiplexing
the two services.The third point refers to the fact that full multi-
plexing requires link-by-link processing whereas
only end node processing is required for LSPs.
Depending on the relative costs of transmission
and processing, more powerful links to support
LSPs may be cheaper than more powerful nodes
to support full multiplexing. As a simple exam-
ple [12], consider two flows f 1 and f 2 of the same
service type with traffic demands ρ 1 and ρ 2
respectively. Let f 1 traverse a route from o 1 to
d 1 and f 2 from o 2 to d 2 which both span hhops.
Assume that the two routes have one physical
link lin common and that it is the bottleneck link
in the sense that its bandwidth Cldetermines the
grade of service offered to the two flows. With-
out LSPs, all of the bandwidth Blof lis available
to f 1 and f 2 whereas, with LSPs, f 1 has access to
a capacity C 1 and f 2 to a capacity C 2 , with C 1 +
C 2 = Bl. Moreover, without LSPs, it takes h
rounds of processing and control signalling to
perform CAC (one per physical link) and each
packet must be analysed htimes whereas, with
LSPs, it takes one hop of processing and control
signalling to perform CAC (on the logical link)
and each packet must be analysed once. In addi-
tion, a “once-and-for-all” cost of h= 5 hops of
processing and management signalling is
required to establish an LSP the cost of which
is depreciated over the expected life time of the
LSP T= 10 connection holding times. The
advantage of not having LSPs is a higher degree
of statistical multiplexing whereas the advantage
of LSPs is less processing and control signalling.
Figure 1 shows the overhead cost relative to traf-
fic gain at which the two advantages even out
for the case where the loss without LSPs is fixed
to 1 %. It is seen that, e.g. when f 1 and f 2 amount
to 100 erlangs LSPs will be preferred if the cost
of signalling and processing per connection ex-
ceeds 0.1 % of the revenue per connection. More
elaborate examples in [12], which consider a set
of realistic networks with a multitude of flows of
different magnitudes, show similar results.The fact that the arguments are in favour of sep-
aration does not mean that all service classes
should necessarily have resources of their own.
On the contrary, service classes with similar
statistical characteristics and similar quality of