一种面向数据密集型应用的并行程序执行模型

Parallel Program Execution Model for Data-intensive Applications

随着各领域需要处理的数据量越来越大,数据密集型应用也变得越来越被重视.该文提出一种包含数据访存层次和访存冲突等信息的新并行程序执行模型PSRAM(h).针对数据密集型应用以访存为主的特点,PSRAM(h)模型将程序执行时间简化为访存时间,通过分析各程序子段的访存层次和数量来预测串行程序的执行时间,进而通过使用各线程执行时间的最大值来预测并行程序的执行时间.使用PSRAM(h)模型下对最典型的数据密集型应用矩阵向量乘进行分析,在龙芯3A处理器和Intel Xeon E5520处理器两个平台上的测试结果表明,PSRAM(h)模型分析结果与实测结果大部分情况下误差小于20%.由此可见,针对数据密集型应用,PSRAM(h)不但可以给出程序执行时间的下限,还可以有效的预测程序的执行时间.

With the amount of data in all areas need to be processed,data-intensive applications become more and more important.This paper proposes a new parallel program execution model PSRAM（h） considering the data memory hierarchy and memory access conflicts.Memory access is a major component of the date-intensive applications.The program execution time is simplified into the memory access time in model PSRAM（h）.The execution time of the serial program is predicted by analyzing the level and the number of the memory accesses in the program segments.And the execution time of the parallel program is predicted by using the maximum value of the thread execution time.Matrix-vector multiplication,w hich is the most typical data-intensive applications,is analyzed by the model.Test results on loongson-3A processor and Intel Xeon E5520 processor show that,in most case,the deviation betw een the PSRAM（h） model＇s results and w ith experimental results is less than 20%.Thus,PSRAM（h） model not only show s the low er limit of the program execution time,also can be used to predict program execution time effectively.