基于NPB的并行机群系统可扩展性研究  

Parallel cluster system scalability research based on NPB

作　　者：苑野[1] 于永澔[1] 

机构地区：[1]哈尔滨工业大学基础与交叉科学研究院,哈尔滨150080

出　　处：《哈尔滨商业大学学报（自然科学版）》2016年第1期60-63,共4页Journal of Harbin University of Commerce：Natural Sciences Edition

基　　金：国家自然科学基金(51201051)

摘　　要：使用NPB并行基准测试程序能有效的测试出大型计算流体力学(CFD)应用中的计算和数据移动特征,这对大量科学与工程应用具有重要意义.在并行机群计算环境下,使用NPB基准测试程序(IS、EP、CG、MG、FT、BT、SP和LU)对该系统的性能进行了评测,实验表明,在一定规模下,NPB的8个基准测试程序随着处理器数目的提高,其系统性能和可扩展性的表现各不一致.EP程序的可扩展性较好,性能较稳定;而IS、CG、MG、BT和SP的可扩展性较差,性能较差;LU和FT在不同规模下,其可扩展性表现为不规律性.机群系统的可扩展性与应用程序的匹配程度还需要进一步提高.NPB parallel benchmark can effectively test a large computational fluid dynamics（ CFD） application in mobile computing and data features,which is important for a large number of science and engineering application. This paper used NPB benchmark（ IS,EP,MG,CG,FT,LU,BT and SP） to test and evaluate the performance of the system in the parallel cluster computing environment. With the increasing number of processors under a certain size,experiments showed that NPB’s eight benchmarks had the various inconsistencies of the system performance and scalability. EP program has better scalability,more stable performance. But IS,CG,MG,BT and SP programs had poor scalability and performance. In different size,the scalability and performance of LU and FT were irregular. The scalability of cluster system and the matching degree between the application program also need to be further improved.

关 键 词：NPB 性能评测 并行基准测试程序 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]