HPL performance prevision to intending system improvement

Abstract

HPL is a parallel Linpack benchmark package widely adopted in massive cluster system performance test. On HPL data layout among processors, a law to determine block size NB theoretically, which breaks through dependence on trial-and-error experiments, is found based on in-depth analysis of blocked parallel solution algorithm of linear algebra equations and implementation mechanics in HPL. According to that law, an emulation model to toughly estimate HPL execution time is constructed. Verified by real system, the model is used to do some scientific prevision on the benefits to Linpack test brought by intending system improvement, such as respectively memory size increase, communication bandwidth increase and so on. It is expected to conduce to direct system improvement on optimizing HPL test in the future. © Springer-Verlag Berlin Heidelberg 2004.