摘要
通过对基于MPI编程模型实现的开源有限元计算分析软件在多核集群计算平台中的程序性能的分析,找出程序瓶颈及其原因,实现了基于MPI编程模型的并行程序在多核计算环境中的性能优化。根据程序性能瓶颈的分析,提出了基于MPI/OpenMP混合并行编程模型的大规模线性/非线性方程组求解和多线程多进程同时进行消息通信的两种程序性能优化方案。不同计算规模的实验结果表明,在多核集群计算平台中,MPI/OpenMP混合编程模型实现的大规模非线性方程组求解器相对于单纯基于MPI编程模型实现的并行程序,其性能有2倍到3倍的提升;多线程多进程同时消息传递的优化方案虽然对程序能够起到性能优化作用,但是对解决程序消息通信瓶颈的问题不是最好的方法。两个方案总体性能分析结果表明,基于MPI/OpenMP混合编程模型实现的并行程序,在多核集群计算平台中能够更好地发挥硬件系统的计算能力。
Through the performance analysis of an open source finite element software based on MPI program model on multi-core cluster platform,some performance bottlenecks were founded.Based on the performance bottleneck analysis,two optimization plans based on MPI/OpenMP hybrid parallel program model were proposed,one of them resolves the inefficiency in solving linear or nonlinear system equations,and the other one elevates processes communication performance.Experiment results show that hybrid parallel solver can efficiently promote the pure MPI based parallel program performance,as up to 3 times.The multi-thread multi-process communication plan can do some optimization,but is not the best solution in this case.The overall optimized performance analysis indicates that on multi-core cluster computing platform,MPI/OpenMP parallel program model can more efficiently utilize hardware system computation resource.
出处
《计算机科学》
CSCD
北大核心
2012年第1期305-310,共6页
Computer Science
