Benchmarking Parallel Processing SystemsA

| Increasing demands for performance (in the sense of processing power) in scientiic applications lead to a seemingly never ending quest for \peak performance". A variety of architectural concepts supporting parallel processing have been proposed and several commercial machines are (more or less successful) on the market. While the theoretical peak performance reported by hardware vendors has surpassed the Teraaop/s mark, it has been observed and documented in many case studies that the actual performance achieved with parallel computers when running real applications is far oo from these optimum values. This gap between peak performance and actual attainable performance raises the question of new techniques for evaluating and comparing the performance of parallel architectures. In this paper, we give a survey of benchmarks which have been proposed for the evaluation of parallel architectures. We discuss their advantages and disadvantages.