Performance prediction of parallel systems with scalable specifications—methodology and case study

Abstract

This paper describes the general methodology of specifying parallel systems within the PAPS (Performance Analysis of Parallel Systems) toolset and presents a case study that shows the applicability and accuracy of the Petri net based performance prediction tools contained in the toolset. Parallel systems are specified in the PAPS toolset by separately defining the program workload, the hardware resources, and the mapping of the program to the hardware. The resource parameterization is described in detail for a multiprocessor computer with a store & forward communication network. The Gaussian elimination algorithm is taken as a workload example to demonstrate how regularly structured parallel algorithms are modelled with acyclic task graphs. Three different program specifications with various levels of model accuracy are developed and their parameterization is described. The predicted execution time is compared with the measured execution times of the real program on the parallel hardware. It is shown that the Petri net based performance prediction tools provide accurate performance predicitons.