Variance reduction applied to product form multiclass queuing networks

Abstract

Performance of product-form multiclass queuing networks can be determined from normalization constants. For large models, the evaluation of these performance metrics is not possible because of the required amount of computer resources (either by using normalization constants or by using MVA approaches). Such large models can be evaluated with Monte Carlo summation and integration methods. This article proposes two cluster sampling Monte Carlo techniques to deal with such models. First, for a particular type of network, we propose a variance reduction technique based on antithetic variates. It leads to an improvement of Ross, Tsang and Wang's algorithm which is designed to analyze the same family of models. Second, for a more general class of models, we use a mixture of Monte Carlo and quasi-Monte Carlo methods to improve the estimate with respect to Monte Carlo alone.