Experimenting with a trace-driven simulator of a batch computing system

Abstract

Measures of the performance of multiprogrammed systems frequently exhibit considerable instability with respect to workload characteristics. As a result, an evaluator of a change to a multiprogrammed system should investigate the effect of this change under different workloads and draw conclusions by means of statistical methods. This paper describes how such an evaluation is performed with a trace-driven simulator of the Watson Research Center's batch computing system. Two simulation experiments are described, one dealing with the number of printers and the other with the number of initiators.