Computational costs of fast stochastic simulation techniques for Markovian fluid models in multiservice networks

Abstract

Two accelerated simulation techniques, Importance Sampling and RESTART, are analyzed in the context of the class of Markovian fluid models, widely used in broadband communication networks. The different terms corresponding to their computational cost are exposed and evaluated for a sample case. Importance Sampling is found to be preferable (i.e. to present the lower computational cost) for the simulation of very rare events and large systems. A tentative criterion is proposed for the a priori selection of the computationally lighter simulation technique, based on the system size and on the probability of the event being estimated.