Performance Test and Analysis of Alltoall Collective Communication on Domestic Hundred Trillion Times Cluster System
Rao Li · 2010
As rapid development of the high performance computers,more and more cores are used and thus lead to more and more communication which debases the perfor-mance of parallel applications greatly.In the test of the performance of Dawning 5000Aby a kind of Fast Fouler Transform(HFFT),we found out that the huge overhead time of MPI_Alltoall is the bottleneck of HFFT.Thus,this paper aimed to test and analyze the leading Alltoall algorithm on Dawning 5000and Deepcomp 7000hoping to do a favor to further collective communication optimization.In this paper,the leading Alltoall algorithms such as 2D_Mesh,3D_Mesh,Bruck,MPICH native,Pair,recursive doubling,Ring,LAM/MPI simple were recounted and tested with different message size and core numbers.The conclusion is that for short message MPICH native and Bruck performs well on Dawning 5000Awhile the lower time consuming algorithms on Deepcomp 7000are LAM/MPI simple and Bruck;when the message size is medium and large,the best choice for Dawning 5000Ais Ring while the optimal algorithm on Deepcomp 7000is Pair.