15.1 Introduction

Most modern computer systems exhibit features that violate the assumptions required by product-form networks. Some system features that violate separable assumptions are: 1) blocking, 2) high service time variability at FCFS centers, 3) simultaneous resource possession, 4) process forking and synchronization, and 5) priority scheduling. These features cannot be modeled directly by product-form models. Thus, many algorithms have been developed to approximate the solution of models that incorporate non product-form features. These algorithms share same basic ideas. They accept a queuing network model that has some nonseparable features and transform it into an approximate network that obeys the BCMP conditions [8]. Some algorithms rely on models that use heuristic extensions to the approximate MVA equations [18, 39]. The approximate QN has a product-form solution and gives approximate measures on the performance of the original model. The input parameters of the original model are either transformed by means of an iterative process or by approximations that resemble the effects of the feature being modeled. Once the new parameters are obtained, an MVA algorithm is used to calculate the performance measures of the approximate network. Some kind of transformation may be required to map the results of the approximate model back to the original one. The following sections present approximate techniques for modeling non product-form queuing models.