A Language-Independent Set of Benchmarks for Parallel Procesors

Benchmarks are essential for objective comparison of computer performance. Established scientific computing benchmarks (LINPACK, Whetstones, Livermore Kernels, etc.) are most appropriate for sequential machines running Fortran. As architectures become increasingly parallel, these benchmarks are decreasingly representa-tive of the actual performance achievable on large-scale scientific applications. We present a set of six benchmarks which are language-independent, representative of mainstream scientific computing, and have parallel content approaching 100 % as parameters grow large. They are designed to fit a wide variety of multiprocessing approaches. The memory, operation, and time complexity of each benchmark problem are analyzed. We intend to promote collection and publication of performance figures for a large set of both experimental and commercial computers, with strict controls to insure integrity and verifiability of results