Input-output-buffered broad-band packet-switch architectures with correlated input traffic

This paper investigates the robustness of the nonblocking asynchronous transfer mode (ATM) switch architecture with input-output buffering and back-pressure control under correlated input traffic conditions. Previous studies have concentrated on the performance of these architectures under random (uncorrelated) traffic conditions. These studies have shown that the input-output-buffered switch architecture with back-pressure control provides considerable performance improvement, in terms of maximum switch throughput, over pure input-buffered architectures, while requiring only a fairly small amount of output buffering. In this paper, traffic correlation is modelled by two parameters; namely, the mean active duration of the ON-OFF source feeding each input port, and the spacing between the cells generated within an active period. This spacing can be fixed or random. The effect of these parameters on the switch performance in terms of the cell-loss behaviour and the mean switching delay is investigated. The performance of different selection mechanisms needed to resolve output-port contention is evaluated, taking into consideration the performance enhancements obtained by incorporating different arbitration criteria within the selection mechanisms. Our study indicates that the burst-length factor significantly affects the buffering requirements of the switch to achieve a certain cell-loss probability, even if the switch is operated at fairly low utilization. The spacing between cells within a burst has a considerable effect on the cell-loss performance of the switch. Using the longer-input-queue criterion as the arbitration criterion between cells contending for the same output port results in smaller buffer sizes for given cell-loss probability and value of applied load, than is the case when the arbitration criterion is implemented according to the earlier arrival time at the head-of-line (HOL) position of the input queues. Since the longer-queue criterion does not involve time-stamping, as is the case with the HOL-time criterion, it is also the more attractive solution from an implementation point of view. This result confirms previously reported conclusions under the uniform random traffic case.