Degree Name

Doctor of Philosophy


A Broadband Integrated Services Digital Network based on ATM promises communication at higher rates and with greater flexibility than current circuit or packet switched networks. However the objective of carrying bursty and variable rate traffic and maintaining both high utilisations and an acceptable quality of service presents difficult traffic control issues.

This thesis attempts to deal with these problems in a systematic and unified fashion. Understanding and predicting the behaviour of statistically multiplexed traffic is identified as a fundamental concern. A state dependent discrete process (SDDP) is introduced as a means for representing both single sources and aggregated traffic streams. Exact steady state solution techniques are described for queues and multiplexers fed by such processes. Further analysis yields information about the temporal behaviour of these queues and their loss characteristics, and leads to several queueing approximation techniques.

Using these solution methods the performance of an ATM multiplexer is investigated. In particular the effects of source traffic parameters on attainable utilisation for a given long term loss constraint are evaluated. Performance is also examined in terms of overload duration and intensity statistics, and in terms of cell delay variation. Peak-to-link ratio and burstiness are identified as key parameters affecting utilisation and resource allocation. Temporal connection characterisitics such as burst length are found to have most impact on overload durations.

Drawing on these results the efficacy of preventative control schemes based upon a traffic contract is examined. The design of peak rate enforcers and in particular the trade-off between tolerance of cell delay variation and enforcement accuracy is investigated. The difficulties associated with estimating and enforcing mean rate are examined, and identified as a key weakness in a purely preventative control system. To alleviate these difficulties a range of additional traffic controls are considered.

The implementation of real time connection admission based on the zero buffer and virtual bandwidth type methods and the extension of these methods to environments with cell loss priority are discussed. The use of virtual paths for resource management and QOS control are explored. A number of potential resource management strategies are identified and compared on the basis of ideal performance limits and implementation complexity. A priority based scheme is found to give best network utilisation but it would be complex to implement. A three layered scheme based on policed virtual paths is found to be very attractive from an implementation point of view. As a result of this comparison a new hybrid resource management scheme is suggested.

Overall the thesis concludes that a traffic control system based on an enforceable traffic contract for peak and average rates can offer guaranteed quality of service (long term cell loss probability) assuming burst scale independence between sources, while obtaining close to optimal statistical gains. However there are two major difficulties with such a scheme: • it is difficult to predict average rate, and • statistical gains are very limited for traffic with a peak-to-link ratio greater than 1/10.

The first difficulty may be addressed by introducing new traffic controls such as in-call parameter renegotiation, traffic shaping and callpadding, and the second perhaps by intelligent multiplexing which takes advantage of the fact the some traffic may be scheduled rather than relying on its statistical nature.