Simulative performance analysis of ATM switching fabrics

Abstract

Asynchronous Transfer Mode (ATM) has emerged as the major technology paradigm for the implementation of the Broadband Integrated Digital Services Network (B?ISDN). In ATM, all forms of information such as voice, video and data are transmitted in small fixed?sized packets called cells. The successful deployment of ATM?based broadband networks presupposes the availability of appropriate fast?packet switching methods and technology. These switches should not only operate at high speeds but also satisfy stringent performance criteria. ATM switches for such networks are within the reach of current technology. The performance evaluation of these switches is necessary in order to economically dimension the network and meet various Quality of Service constraints. Simulation is a popular and flexible performance?analysis tool which can be employed for a careful and accurate modelling of broadband networks. In this thesis we address two issues which arise during the performance evaluation of ATM switching fabrics. We first consider the problem of characterising the traffic which appears at the input ports of a switch. It is well known that the assumed traffic model has a profound effect on the behaviour of ATM switch buffer queues; e.g. for the same average load, different traffic types can result in entirely different cell?loss probabilities and mean waiting times. Simple renewal traffic models such as Bernoulli, which do not explicitly take into account temporal dependence, yield overly optimistic performance estimates. We have generated compact and realistic models for voice, video, data and multimedia traffic. A single voice source has been shown to be well?represented by an ON?OFF Markov model. We have used Transform?Expand?Sample (TES) models for modelling coded video such as MPEG and JPEG streams. TES models have been found to capture the marginal distribution as well as the autocorrelation structure in video streams to a high degree of accuracy. Autoregressive models have been applied to video source modelling, especially for video teleconferencing. However, these have been found to be unsuitable for detailed traffic studies. A scheme which combines autoregressive models with a Markov?chain model has been proposed and implemented. We have considered self?similar models for Ethernet data. We have created both source models, using heavy?tailed ON?OFF sources, as well as aggregate models, using Fractional Gaussian Noise. We estimate cell?loss probability, mean waiting time and delay quantiles. We next consider the problem of efficiently estimating very low cell?loss probabilities. Conventional Monte Carlo simulation, in such cases, is prohibitively expensive in terms of computer time and even infeasible. We have explored the use of a variance?reduction technique called Importance Sampling to increase the efficiency of simulation. In literature, simple and practical Importance Sampling techniques can be found for output?queuing systems. Regenerative simulation has been combined with Importance Sampling to estimate steady?state metrics. However, these techniques cannot be directly applied to shared?memory?based systems. We have successfully employed a simple biasing scheme to speedily determine the performance of shared?memory systems in terms of cell?loss probability.