Estimating the effective performance of program parallelization on shared memory MIMD multiprocessors

Abstract

One of the problems the parallelization of programs poses is the evaluation of the quality of the parallelization performed. We show that the disruption generated by the target machine, which we call noise, has a non-negligible influence on the sequential as well as parallel execution times. The effects on the efficiency, which depends on these execution times, are investigated. We show that this performance criterion gives an altered picture of the quality of the parallelization. We present a set of criteria for estimating the effective performance of the parallelization of programs on one hand and the machine generated drop in efficiency on the other hand. The theoretical results exposed are illustrated by experimental measures conducted on the Encore Multimax.