基于系统抽样的并行程序性能特征分析方法及其实现  

Research and Implementation of Parallel Program Profiler Based on System-Sampling

作　　者：陈永然[1] 窦文华[1] 钱悦[1] 齐星云[1] 

机构地区：[1]国防科学技术大学计算机学院,长沙410073

出　　处：《计算机研究与发展》2007年第10期1694-1701,共8页Journal of Computer Research and Development

基　　金：国家自然科学基金项目(40245023);计算物理国家重点实验室基金项目(51479040103KG0201)

摘　　要：程序性能特征分析是理解程序行为的基础,对识别程序性能瓶颈、了解软硬件资源利用状况具有重要作用.特别在大规模并行系统的性能评价中,受时间和空间的约束无法分析完整应用性能特征.一个有效的方法是通过抽样的方法分析应用程序部分代码的性能特征,以此代表完整应用的性能特征.分析了Profiler程序负载来源,提出了基于抽样的程序性能特征分析方法,并基于该方法实现了性能特征分析器SamplePro.与其他方法比较,基于系统抽样的程序性能特征方法在最小样本容量下得到最优的分析结果,仅需抽样分析1%～3%的程序指令就能实现小于3%的分析误差.Analyzing the performance of program is the basis of understanding the behaviors of programs.It is important to find the performance bottlenecks and learn the utilization of hardware and software resource.In the performance evaluation of parallel computing system,the performance characters of the application cannot be analyzed completely because of the restriction of the time and the space.An effective method is analyzing the performance characters of part of the application codes,and regarding them as the performance characters of the whole applications.In this paper,a method of analyzing the performance of the programs based on sampling is put forward.Furthermore,a profiler named SamplePro based on sample theory is developed.Compared with other methods under the condition of the same sample error,this technique can effectively reduce the needed instruction numbers,which means only 1%～3% instructions are used to achieve less than 3% analyzing error.

关 键 词：并行程序 程序行为 代码监控 应用特征 系统抽样 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]