AMDAHLS LAW AND THE STATISTICAL CONTENT OF THE NAS PARALLEL BENCHMARKS

The NAS Parallel Benchmarks have been developed at the NASA Ames Resea rch Center. In the last three years extensive performance data have be en reported for parallel machines both based on the NAS Parallel Bench marks [1, 2] and on LINPACK [3]. In this study we have used the report ed benchmark results and performed a number of statistical experiments . These included cluster, factor and regression analyses. We did this to find out how many of the NAS Parallel Benchmarks are - in a statist ical sense - necessary, to represent all the reported results. We also fitted Amdahl's Law to the data, to see whether it is meaningful to a pply more sophisticated performance models to the reported results. Al l statistical experiments were done for absolute performances as well as for the corresponding efficiencies. The analysis of Amdahl's Law wa s performed for both classes (Class A and B) of the NAS Parallel Bench marks.