Delay performance of some scheduling strategies in an input queuing ATM switch with multiclass bursty traffic

We consider an N x N nonblocking, space division, input queuing asynchronous transfer mode (ATM) cell switch, and a class of Markovian models for cell arrivals on each of its inputs, The traffic at each input comprises geometrically distributed bursts of cells, each burst destined for a particular output, The inputs differ in the burstiness of the offered traffic, with burstiness being characterized in terms of the average burst length, We analyze burst delays in the situation in which some inputs receive traffic with low burstiness and others receive traffic with higher burstiness. Three policies for head-of-the-line contention resolution are studied: two static priority policies [viz,, shorter-expected-burst-length-first (SEBF), longer-expected-burst-length-first (LEBF)] and random selection (RS), Direct queuing analysis is used to obtain approximations for asymptotic (as N --> infinity) high and low priority mean burst delays with the priority policies, Simulation is used for obtaining mean burst delays for finite N and for the random selection policy, Numerical results show that, as the traffic burstiness increases, the asymptotic analysis can serve as a good approximation only for large switch sizes, Qualitative performance comparisons based on the asymptotic analysis are, however, found to continue to hold for finite switch sizes. It is found that the SEBF policy yields the best delay performance over a wide range of loads, while RS lies in between, SEBF drastically reduces the delay of the less bursty traffic (e,g,, distributed computing traffic) while only slightly increasing the delay of the more bursty traffic, e,g,, variable bit rate (VER) video, LEBF causes severe degradation in the delay of less bursty traffic, while only marginally improving the delays of the more bursty traffic, RS can be an adequate compromise if there is no prior knowledge of input traffic burstiness.