Developing a Benchmark for Evaluating the Performance of Parallel Computers

This paper discusses the development of a portable suite of benchmarking programs for parallel computers. Comparative measurement of the performance of parallel computing systems has been limited because of the great diversity of architectures and of processor interconnection schemes. One solution is to translate benchmark codes into a consistent and portable parallel language. This paper reports on progress in developing such a portable suite of benchmarks. An extensive introduction to parallel computing is included as an appendix, to provide a thorough understanding of the factors complicating development of the performance suite. Key to the development was the use of p4, a library of tools developed at Argonne National Laboratory. The benchmark codes were translated successfully using p4 and were run on a variety of parallel machines. Conclusions and suggestions for future work are given