Side_1_360

Besides the bandwidth cost, other contributions
come from basic router functions (backplane,
etc.) and processing in routers and call handlers,
see Figure 10. Thus, corresponding contributions
have to be incorporated in the cost model. The
traffic flows may contribute differently to the
various cost components. For instance, a flow
not bothering the call handler is likely to have
lower cost associated with control (e.g. accep-
tance control).

Referring to the illustration in Figure 11 three
basic contributions to the cost can be identified
as described in the following:

Transmission/link/bandwidth cost, Ztq;reflecting
the cost of links between routers as well as ter-
mination units within the routers. One unit of
transmission capacity between locations iand j
with capacity Bij, has a cost Cij. Each transmis-
sion link terminates with a module in location i
that has a cost Ci. The module may connect Niof
these transmission links. The same relations
apply for location j. Assuming a traffic aggre-
gate qhaving capacity demand Bq, the band-
width cost for that aggregate carried from ito j
is calculated as:

The relation between traffic load in a flow
aggregate and the required capacity may not be
simple. Again this depends on the service types.
For service involving call handler, an effective
bandwidth measure may be needed, for instance
when doing admission control. For some of
these services, a blocking probability measure
may well be introduced. Thus, multirate block-
ing formulae could be applied, abstractly giving
the vector of blocking probabilities, Pbby:

Pb= f(A,R,C)

where the load, A, and the characteristics, R, of
all flows are to be supported by the capacity C.
A challenge is being able to find the needed
capacity, C. This could be done iteratively or
assuming approximate relationships. In case
measurement-based algorithms are used, these
parameters may vary. To some extent this could
be captured by the effective bandwidth (consid-
ering Aand R), e.g. see [COST242].

For some “pipe” services a fixed capacity is
specified between ingress and egress points (vir-
tual leased capacity service) and this bandwidth
value can then be used. Similar arguments might
be relevant when a certain level of overbooking
is introduced.

The one-to-any service type may have more

elastic behaviour, implying that actual relations

between traffic load and capacity are not strict.

Then a minimum capacity could be obtained

assuming some minimum effective bandwidth

value (e.g. minimum acceptable throughput).

Switching cost, Zsq;reflecting the effort

requested for transferring packets from an input

to an output port. This cost component is

assumed to be proportional to the traffic load,

Ab, and the effective bandwidth, EBb, for flows

of type b. That is, for a traffic aggregate qcarry-

ing several types:

where αand βare cost factors. These are chosen

to reflect actual router implementations. A cen-

tral point is to express the effective bandwidth

for the traffic flows. For some flows the charac-

teristics may be less influenced by the network

load, e.g. for inelastic traffic. Then an effective

bandwidth measure could be estimated for the

original traffic source characterisation (mean

bitrate, peak bitrate, etc.). For other flow types,

in particular those using TCP flow control, the

resulting characteristics for an individual flow

will depend on the network state. Thus, a base

estimate could be assumed depending on termi-

nal capabilities, duration of the transfer (due to

slow-start behaviour) acceptable delay, etc. An

important observation is that some “statistical”

effects have to be considered during the network

design task. One argument is that several inde-

pendent sources will be behind the traffic facing

the network. Another factor is that several of the

values used would frequently be attached with

uncertainties, in particular when a future net-

work is to be designed. A more detailed exami-

nation of effective rates of such sources could

be done when the resulting network design has

been found. The switching cost is frequently

associated with operations taking place per

packet, like exercising traffic handling mecha-

nisms (policer, marker, etc.).

Control cost, Zcq;reflecting the processing

requested for establishing/releasing a connec-

tion. Control cost will commonly be related to

call handler functionality and other mechanisms

implying exchange of information and configu-

ration of relevant traffic handling functions,

like policer, marker and shaper. It is frequently

assumed that each hop contributes to this cost.

That is, if no individual packets but rather aggre-

gates are examined (e.g. due to the use of LSP

through-connection) the control cost is likely to

be lower for that hop. For a call handler three

types of processing steps; connection request,

establishment and rejection, are considered.

Ztq=

Bq

Bij

Cij+

Bq

NiBij

Ci+

Bq

NjBij

Cj

Zsq=

∑

b∈q

Zsb=

∑

b∈q

Ab(α·EBb+β)