On the modeling of shared resources with various lock granularities using queueing networks

Abstract

This paper presents a practical and noniterative method for evaluating queueing networks in which serialization delays are represented using measurable parameters. Two types of models are introduced to describe synchronization of accesses to shared resources using simulation and numerical methods, respectively. The model parameters are estimated from the values of measurable quantities. For the analytic models, the probabilities that a shared resource is locked, thereby causing other processes to wait for the resource to be accessible, are calculated, under certain assumptions, as functions of the probabilities of access and of the number of active transactions. The service times of lock servers are calculated as functions of the loads on all service centers. Performance measures applicable to the evaluation of computer systems are introduced and evaluated. An impact of various lock granularities on system performance is presented. An example involving a distributed file system and measurement data collected in a small business installation is given to compare performance measures provided by the simulation and analytic models.