国产百万亿次机群系统Alltoall性能测试与分析  

Performance Test and Analysis of Alltoall Collective Communication on Domestic Hundred Trillion Times Cluster System

作　　者：饶立[1] 张云泉[1,2] 李玉成[1] 

机构地区：[1]中国科学院软件研究所并行计算实验室,北京100190 [2]中国科学院软件研究所计算机科学国家重点实验室,北京100190

出　　处：《计算机科学》2010年第8期186-188,207,共4页Computer Science

基　　金：国家自然科学基金(No.60303020);国家自然基金重点项目(No.60533020);国家863(No.2006AA01A102;No.2006AA01A125)资助

摘　　要：随着高性能计算机的应用和发展,并行应用程序所使用的处理器数越来越多,进程间的通信量也不断增多,这对应用程序的性能有很大影响。在采用一种快速傅里叶变换HFFT对曙光5000A进行性能测试时发现,MPI集合通信函数MPI Alltoall的巨大通信开销是并行程序设计的瓶颈。为此,对现有主流Alltoall算法在曙光5000A和深腾7000上进行性能测试与分析,以期对未来的Alltoall算法的优化工作做出贡献。利用不同消息长度和不同进程数测试了Alltoall函数多种算法的性能,这些算法包括二维网格算法、三维网格算法、Bruck算法、原始算法、成对交换算法、递归倍增算法、环算法以及LAM/MPI中的简单算法等。实验结果表明:消息长度较小时,在曙光5000A上采用原始算法和Bruck算法的性能较好,而在深腾7000上用时较少的算法是简单算法和Bruck算法;对于长消息,曙光5000A上最优的算法是环算法,深腾7000上成对交换性能最优。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 5000A by 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 5000 and Deepcomp 7000 hoping 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 5000A while the lower time consuming algorithms on Deepcomp 7000 are LAM/MPI simple and Bruck; when the message size is medium and large, the best choice for Dawning 5000A is Ring while the optimal algorithm on Deepcomp 7000 is Pair.

关 键 词：集合通信 Alltoall 曙光5000A 性能测试与分析 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]