基于HPL测试的集群系统性能分析与优化  

Performance Analysis and Optimization for Cluster System Based on HPL Benchmark

作　　者：王晓英[1] 都志辉[1] 

机构地区：[1]清华大学计算机与科学系高性能计算研究所,北京100084

出　　处：《计算机科学》2005年第11期231-234,共4页Computer Science

基　　金：清华大学985基金;华为基金;IBM基金(JC2001027)

摘　　要：集群系统以很高的性价比和良好的可扩展性而成为当今高性能研究的一大新热点,如何评价与优化集群系统的特性也成为一个很关键的问题。本文通过采用全世界TOP500计算机排名所用的 HPL基准测试,对集群的浮点运算性能进行评测和分析,根据在集群系统上进行的HPL基准测试的结果分析总结了影响其整体性能的重要因素,并对这些方面一一给出具体的分析,提出了针对集群系统具有实际意义和研究价值的优化方向和优化方法。Cluster system has become increasingly significant in High Performance Computing（ HPC） research field because of its high performance/cost ratio and good scalability. How to evaluate and optimize the performance of cluster system becomes a key issue in this field. The World TOP500 computer rank list uses HPL as a standard benchmark, from which the optimized performance of clusters on floating-point operations can be obtained. Specific tests are carried out for a typical cluster system, and then some most important factors which affect the performance of the whole sys- tem are studied. With the analysis of these factors in detail, this article proposed some methods of optimization for cluster systems, which have both practical significance and researching value.

关 键 词：HPL测试 集群系统 可扩展性 浮点运算 软件包 

分 类 号：TP31[自动化与计算机技术—计算机软件与理论]