HPL中条状矩阵分解的算法分析与优化  

Analysis and Optimization on Panel Matrix Factorization of HPL Benchmark

作　　者：林蓉芬[1] 王丹云[1] 

机构地区：[1]江南计算技术研究所,无锡214000

出　　处：《高性能计算技术》2013年第3期24-27,共4页

摘　　要：HPL是Linpack Benchmark的并行版本，是高性能并行计算机系统性能评测的标准测试程序。本文将HPL原有算法进行改进，使HPL中条状矩阵分解的离散访存量部分转化为连续访存量。在超级计算机领域越来越高的计算访存比趋势下，使HPL能更好地反映机器的浮点性能，具有重要的现实意义。HPL is the parallel version of Linpack Benchmark and the standard testing program for high performance parallel computer systems. This paper improves the panel matrix factorization of HPL, and changes the disperse memory access to continuum memory access. In Supercomputer field, the ratio of computation to memory access is higher and higher. It plays an important role in evaluating a machine＇s floating-point performance better.

关 键 词：HPL 条状矩阵 优化 性能评测 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]