一种Nehalem平台上的MPI多级分段归约算法

A Hierarchical-segment Reduction Algorithm for Nehalem Systems in Threaded MPI

基于线程MPI环境提出一种适用于Nehalem平台长消息归约的多级分段归约算法(HSRA).HSRA考虑了Nehalem系统的体系结构特点,分处理器内归约和处理器外归约两个步骤实施节点内归约通信,在均匀分布计算负载的前提下仅需要较少的远端内存访问.首先在MPIActor的归约算法框架中设计、实现了HSRA算法,从访存角度分析了HSRA算法的开销,然后与单级分段和已有的另外三种节点内基于共享内存的归约算法进行比较;最后在真实系统上通过IMB(Intel MPI Benchmark)验证算法,实验结果表明,该算法是一种适用于在Nehalem系统中处理长消息节点内归约的高效算法.

A new intra-node reduction algorithm called Hierarchical-Segment Reduction Algorithm（HSRA） is proposed, which is for Nehalem systems based on threaded MPI environment. By considering the character of Nehalem micro architecture, HSRA imple- ments a intra-node reduction communication in two steps which refers as inter-processor reduction and outer-processor reduction, the design can balance computing loads with less remote memory access. First in MPIActor reduction algorithm framework implementing HSRA based on access and storation, then comparing of reduction algorithm with single segment and other three nodes, finally testing the algorithm in Intel MPI Benchmark. The experiment shows that HSRA is an effective algorithm for long message reduction on Nehalem systems.