一种新的基于HPM并行计算性能数据采集方法  

A new parallel program performance gathering method based on HPM

作　　者：程克非[1] 罗江华[1] 李红波[2] 

机构地区：[1]重庆邮电大学计算机科学与技术学院,重庆400065 [2]重庆邮电大学人工智能研究所,重庆400065

出　　处：《重庆邮电大学学报（自然科学版）》2011年第1期80-84,共5页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：重庆市教委科学基金(KJ20090511);重庆邮电大学博士基金(A2006-80)~~

摘　　要：针对传统数据采集方法不足和Vtune的局限性,介绍了一种在多处理器环境下基于硬件性能监视器(hardware performance monitor,HPM)的并行计算性能数据采集方法,该方法基于集群和消息传递接口(message passinginterface,MPI)环境,利用应用程序性能接口(performance application programming interface,PAPI)采集CPU消耗的时钟周期数据。实验表明,采用该方法进行数据采样比传统方法精度高6个数量级,且跨平台性较Vtune更好,兼容更多的硬件和操作系统,更适用于多处理器环境下的并行程序性能分析。To avoid traditional data gathering method and Vtune＇s limitations, a parallel program performance data sampling method based on hardware performance monitor（HPM） was introduced under the circumstances of multi-processor. By clustering and message passing interface（MPI） technology, CPU cycles consumed by software with performance application programming interface（PAPI） can be gathered. Experiments show the method can get six orders of magnitude higher precision performance data than traditional one, better cross platform performance than Vtune, compatible with more hardware and operating systems, and more suitable for parallel program performance analyze in multi-proeessor environment.

关 键 词：数据采集 并行 消息传递接口(MPI) 应用程序性能接口(PAPI) 硬件性能监视器(HPM) 

分 类 号：TP302.7[自动化与计算机技术—计算机系统结构]