OPTIMIZATION OF LARGE MULTICLASS (NON-PRODUCT-FORM) QUEUEING NETWORKS USING PERTURBATION ANALYSIS.

Abstract

The recent perturbation analysis approach to discrete event systems is applied to large multiclass queueing networks. We consider such networks, with limited buffer sizes and general service time distributions. We derive the sensitivity of network performance, with respect to various decision parameters, by observing a single sample path (Monte Carlo experiment) of the system. For large, complex networks, with many parameters, this provides an extremely efficient tool for optimizing the system performance, and a substantial savings in experiment cost (simulation time, or actual system time).