MPI_ALLTOALL实现算法的改进与性能分析

Improvement and performance analysis of MPI_ALLTOALL algorithm

通过分析全互换通信中4种算法的性能,提出了一种改进算法。该改进算法递归倍增的创建子进程,通过增加通信进程数目来减少通信次数。对比分析改进算法与成对互换算法的通信次数,改进算法的通信次数是成对互换算法的一半。实验机群是在版本为MPICH2-1.0.8并行环境下测试,结果表明,消息类型为长消息通信且通信进程数为2的幂次方时,该改进算法比成对互换算法的性能更优。

Through analyzing the performance of four algorithm about all-to-all,an improved algorithm is presented.The improved algorithm creates the subprogram in a recursive doubling manner and reduces the number of communications by increasing the number of processes.The number of communications between the improved algorithmand the pairwise exchanges are comparatively analyzed,and the improved algorithm is half as much as the latter.Experimental clusters are tested in MPICH2-1.0.8 parallel environment,results show that the improved algorithm is better than pairwise exchanges when the type of information is long and the communication process is the power of two.