to client – server interactions. Hence, load bal-
ancing directing traffic on the different servers
may improve the overall performance. This is
sometimes called traffic directing, operating
on the application layer, ref. Box D.
- DiffServ issues. As DiffServ is more widely
deployed, adequate TE systems become more
critical to ensure that conditions in Service
Level Agreements (SLAs) are met. Service
classes (Class of Service) can be offered by
defining Per-Hop Behaviours (PHBs) along
the path, exercising DiffServ in the nodes, in
particular by configuring mechanisms like
traffic classification, marking, policing and
shaping (mainly in edge routers). A PHB is a
forwarding treatment including buffer man-
agement and scheduling. In addition the
amount of service capacity, e.g. bandwidth,
allocated to the different service classes has to
be decided upon. The following issues, from
[ID_tepri], give some requirements on TE in a
DiffServ/MPLS environment:
- An LSP should provide configurable maxi-
mum reservable bandwidth and/or buffer
for each supported service class.
- An LSR may provide configurable mini-
mum available bandwidth and/or buffer for
each class on each of its links.
- In order to perform constraint-based routing
on a per-class basis for LSPs, the routing
protocols should support extensions to
propagate per-class resource information.
When delay bounds is an issue, path selec-
tion algorithms for traffic trunks with
bounded delay requirement should take
delay constraints into account.
- When an LSR dynamically adjusts resource
allocation based on per-class LSP resource
requests, adjusting weights for the schedul-
ing algorithms should not adversely impact
delay and jitter characteristics.
- An LSR should provide configurable maxi-
mum allocation multiplier on a per-class
basis.
- Measurement-based admission control may
be used to improve resource usage, espe-
cially for classes not having strict loss or
delay/jitter requirements.
- Controlling the network. In order to see the
effect of having a TE system, the relevant
decisions must be introduced into the network.
Control mechanisms may be manual or auto-
matic, the latter being a goal for most. Net-
work control functions must be secure, reli-
able and stable, in particular during failure
situations.
5 TE Taxonomy
A taxonomy of TE systems is given in [ID_tepri]
in accordance with the following criteria:
- Time-dependentvs. state-dependentvs. event-
dependent: A static TE system implies that no
TE methods are applied on the time scale con-
sidered. Therefore, it is commonly assumed
that all TE schemes are dynamic (on the time
scale looked at). A time-dependent scheme is
based on timely variations in traffic patterns
and used to pre-program changes in the traffic
handling. A state-dependent scheme adapts
the traffic handling based on state of the net-
work, allowing for taking actual variations in
the traffic patterns into account. The state of a
network may be based on resource utilisation,
delay measures, etc. Accurate information
available is crucial for adaptive TE schemes.
This information has to be gathered and dis-
tributed to the relevant routers. A challenge is
to limit the amount of information that must
be exchanged between routers, still allowing
for sufficiently updated data in each of the
routers to make the traffic handling decisions.
Event-dependent schemes may lead to fewer
information exchanges compared to state-
dependent schemes. Then, certain events are
used as input when updating the traffic han-
dling, like traffic load crossing a threshold,
unsuccessful establishment of an LSP, etc.
- Offlinevs. online. In case changes in traffic
handling do not need to be done in real time,
the computations can be done offline, e.g.
allowing for more thorough searches over the
feasible solutions finding the better one to
apply. On the other hand, when traffic han-
dling is to adapt to changing traffic patterns, it
is to be done online. For online calculations,
relatively simple algorithms are applied lead-
ing to short response times until the updated
traffic handling can be activated. Still the
algorithm should present a solution that is
close to the optimal one.
- Centralisedvs. distributed. In a centralised
scheme a central function decides upon the
traffic handling in each of the routers. Then,
the central function has to collect and return
the information. In order to limit the overhead,
infrequent information exchanges are sought;
however, more frequent exchanges are asked
for to keep an accurate picture of the network
state in the central function. This results in a
classic trade-off problem, finding the time
