Performance characterization and comparison of SIMD architectures

Abstract

A methodology is presented for evaluating and comparing the performance of Single Instruction Multiple Data (SIMD) architectures using the applications given to the machines by their users. Machine characterization, code characterization, execution time estimation and performance evaluation subsystems of the methodology are explained. Machine parameters for SIMD architectures are defined. Adaptation of the methodology for SIMD systems is explained. Measurement results obtained on MasPar are presented and its performance is compared to four MIMD systems.