Self-similarity in a multi-stage queueing ATM switch fabric

Recent studies of digital network traffic have shown that arrival processes can be more accurately modeled as a statistically self-similar process than as a Poisson-based process. We present a simulation of a combination shared-output queueing ATM switch fabric, sourced by two models of self-similar input, namely, Pareto-distributed interarrival times and a Poisson-Zeta ON-OFF process. The effect of self-similarity on the average queue length and cell loss probability for this multi-stage queue is examined for varying load, buffer size, and internal speedup. The results using two self-similar input models are compared with each other and with Poisson interarrival times and an ON-OFF bursty traffic source with geometrically distributed burst lengths. The results show that at a high utilization and at a high degree of self-similarity, cell loss probability declines slowly with increasing buffer size and speedup, as compared to the decline using Poisson-based traffic