we have μacc= 2,048 kbps and φ= 60 kbytes.
The data and acknowledgement packets are of
size 1,500 and 40 bytes respectively. The link
propagation delay τ, which includes nodal pro-
cessing delays, is for reasons of simplicity fixed
to 100 ms and the maximum window size wmax
= 64. The edge routers have (logically) separate
buffers for each aggregate. All buffers are of size
5 packets and the revenue θearned per bit car-
ried is 1 and 3 for domestic and business users
respectively. The request arrival rates differ
between O-D pairs as shown in Table 2, but
not between service classes.The XFG algorithm requires 3 minutes of execu-
tion on a Pentium III 1.5 GHz processor to com-
pute the solution. The algorithm constructs 128
transit routes: 19,462 units of bandwidth are
assigned to the 40 (2 ×20) direct routes and
14,955 units of capacity to the transit routes.
The routes contain 34,417 units of capacity.The lengths of the LSPs and their capacity
assignments are shown in Table 3. Using the
same notation as in Table 1, Table 3 shows that
28,678 units of capacity are assigned to the 120shortest routes (the shortest routes have nor-
malised length 0) and that 19,462 units of capac-
ity are assigned to the 40 direct routes (the direct
routes have un-normalised length 1). It is also
seen that 85 % of the routes are constructed on
the shortest and the shortest-but-one paths con-
necting the O-D pairs: 90 % of the capacity is
configured on these shortest routes.Some more results are shown in Table 4. The
simple TCP model predicts that the average
packet loss probability eover all aggregates is
1.9 %. In more detail, the packet loss probability
for aggregates of domestic users is 5.4 %, where-
as aggregates of business users, with their faster
access links and larger data requests, experience
a 0.1 % packet loss probability.It is also seen that the soft blocking factor αis
3.6 % for domestic users and 1.1 % for business
users. The effective rates at which users get new
data delivered is 46 kbps for domestic users and
2,026 kbps for business users which result in
average download times of 5.2 s and 0.2 s
respectively.The algorithm found 168 LSPs, of which 88 are
used by domestic users and 80 by business users.
An aggregate is said to have an LSP multiplicity
of Kif it is supported by KLSPs. The table
shows that the LSP multiplicity is higher for
domestic users than for business users. As
shown in (18), a higher multiplicity will give a
higher delay. Domestic users are less sensitive
than business users to such a delay because of
the lower speeds of their access links. The algo-
rithm exploits this relative insensitivity, which
results in different multiplicities for the two ser-
vice classes. LSP bandwidth is the average band-
width per LSP and as expected, it is seen that
business class users consume more bandwidth,
partly because of the longer files and partly
because of their faster access links. The last
observation is confirmed by the LSP loads,
which are seen to be 37 % and 24 % for domes-
tic and business users respectively.The buffers at the ingress nodes of the MPLS
network can be used to reduce the packet loss
probabilities at the expense of increasing the
buffer delay. Such a change will also affect the
optimal choices of routes and bandwidths for the
LSPs. Table 4 shows the effect of increasing the
buffer space on the LSPs for domestic users,
which suffer from a high packet loss, from 5
to 10.The result is not only a lower packet loss proba-
bility for domestic users, but the soft blocking
factors decrease and the effective rates increase
for both types of users. The latter is a conse-
quence of the fact that a complete redesign isFigure 5 The core NSF
networkNode 123456781 –6719523
2 7–243311569
3 825– 43718711
4 133–4711
5 11 33 39 5 – 24 9 15
6 51416221– 4 6
7 256184–2
8 381011262–Table 2 Request rate matrix
for both domestic users and
business users21876345Palo Alto
San DiegoHoustonArgonneCollege
ParkAtlantaPrincetonCambridge