基于Trace的并行性能模拟任务映射方法  

Task Mapping Method for Parallel Performance Simulation Based on Trace

作　　者：方建滨[1] 徐传福[1] 车永刚[1] 翁玉芬[1] 王正华[1] 

机构地区：[1]国防科技大学计算机学院并行与分布处理国家重点实验室,长沙410073

出　　处：《计算机工程》2010年第12期269-271,共3页Computer Engineering

基　　金：国家"863"计划基金资助项目(2007AA01Z116);国家自然科学基金资助项目(60603055)

摘　　要：针对Trace驱动的并行性能模拟问题,提出基于Trace信息指导的映射方法CO-LP3M。CO-LP3M利用从Trace中提取的目标应用程序的通信特征,以宿主机物理进程间通信次数最小化为目标,兼顾计算负载均衡,生成并行模拟任务到宿主机的映射。对HPL程序进行实验,结果表明CO-LP3M可有效提高并行模拟性能,相对于常见的映射方式,模拟性能最多提高14.7%。在此基础上给出CO-LP3M的扩展技术SCO-LP3M。Aiming at minimizing communication between host physical processes and balancing computation loads, this paper proposes a mapping method named Communication-Optimal LP to PP Mapping（CO-LpaM） which can obtain the communication pattem of object application programs based on trace generated in parallel performance simulation. Experimental results on High Performance Linpack（HPL） show that the method can increase simulation performance by up to 14.7% compared with common mapping methods. Based on CO-Lp3M, scalability technology named SCO-LP3M is given.

关 键 词：并行模拟 任务映射 通信优化 负载平衡 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]