MPI Alltoall通信在多核机群中的优化.pdfVIP

lSSN 1000-1239/CN 11-1777fTP 计算机研究与发展 Journal of Computer Research and Development 50(8): 1744-1754 , 2013 MPI Alltoall 通信在多核机群中的优化 李强1，2.3 孙凝晖1，2 霍志刚1 马捷1 1 (中国科学院计算技术研究所高性能计算机研究中心 北京 100190) 气中国科学院计算机系统结构重点实验室 北京 100190) 3 (中国科学院大学北京 100039) Oiqiang@ncic. ac. cn) Optimizing MPI Alltoall Communications in Multicore Clusters 1 2 3 1 2 1 1 Li Qiang • • , Sun Ninghui • , Huo Zhigang , and Ma Jie 1 (High Performance Comρuter Research Center , Institute 0f Computing Technology , Chinese Academy of Sciences , Beijing 100190) 2 (Key Laboratory of Comρuter System and Architecture , Chinese Academy of Scie肝ence 3 (University of Chi口nese Academy of Scieηces ， Beijing 100039) Abstract MPI Alltoall is an important collective operation. In multicore clusters , many processes run in a node. On the one hand , shared memory can be adopted to optimize Alltoall communications of small messages by leader-based schemes. However , as these schemes adopt a fixed number of leader processes , the optimal performance cant be obtained for all small messages. On the other hand , processes within a node contend for the same network resource. In Alltoall communications of large messages , many synchronization messages are used. Nevertheless. the contention makes their latency increase many times and the synchronization overhead cant be ingored. To solve these problems , two optimizations are presented. For small messages , the PLP method adopts changeable numbers of leader processes. For large messages , the LSS method reduces the number of synchronization messages from 3N to 2 ß. The evaluations prove two methods.