Understanding the three-way statistical relationship between traffic,
capacity and realized performance is vital not only for adequately
provisioning IP network resources to satisfy anticipated demand but also for
defining the service model and congestion control mechanisms that determine
how those resources are shared. In expressing demand it is particularly
important to know what set of traffic characteristics has a significant
impact on performance and needs therefore to be monitored and controlled.
This set clearly depends on the adopted network architecture: defined
service classes, scheduling mechanisms, admission controls, traffic routing
algorithms,...
We argue that an essential criterion in designing the network architecture
is that the resulting traffic-capacity-performance relationship be as simple
as possible. In other words, resource sharing schemes should be such that
performance is largely insensitive to detailed traffic characteristics. This
in not true, for example, of schemes relying on significant buffering in the
network to smooth out momentary rate overloads since delays can vary widely
depending on the correlation structure of the packet arrival process
(particularly when the latter is long range dependent).
Two resource sharing schemes which do meet the insensitivity criterion are
so-called bufferless multiplexing for open-loop controlled streaming traffic
and fair bandwidth sharing for closed-loop controlled elastic traffic. We
review the powerful insensitivity results available for these multiplexing
schemes and suggest that they indeed constitute the preferred basis for a
network architecture offering guaranteed quality of service. Necessary
congestion control mechanisms are described and their implementation is
discussed.
Copyright. All Rights Reserved.