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 roughly 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.
声明:本文内容及配图由入驻作者撰写或者入驻合作网站授权转载。文章观点仅代表作者本人,不代表电子发烧友网立场。文章及其配图仅供工程师学习之用,如有内容侵权或者其他违规问题,请联系本站处理。 举报投诉