对MPICH作业提交方法的改进和验证

IMPROVEMENT AND VERIFICATION OF JOB SUBMISSION IN MPICH

MPI(Message Passing Interface)是大规模集群和网格平台中最通用的编程环境,而MPICH是其中应用得最广的一种可移植的实现。在集群式系统中,通信时间取决于许多因素,如节点数、网络带宽、拓扑结构还有软件算法等。到目前为止关于程序层面上的通信模式被研究得很多,以期达到提高通信效率的目的,但是MPICH系统内部所需要的通信时间特别是作业提交过程所花费的时间往往为人们所忽略。分析了当前MPICH的作业提交方法,并提出了同步二叉树法、异步二叉树法和二倍扩散法等一系列改进算法,达到了减少通信时间,优化通信性能的目的。

MPI （ Message Passing Interface） is the most universal programming mode for large scale clusters and gird computing platforms, and MPICH is the most widely used portable realization of MPI. In collective systems, communication time depends on many factors such as the number of nodes, the bandwidth, the topology, and also the software algorithm. So far, communication modes upon the programming level are well researched in order to enhance communication efficiency. However, the communication time in MPICH system, especially the job submission time, is often neglected. In this paper it analyzes the current MPI job submission method, and presents a series improved methods including Synchronous Binary Tree method, Asynchronous Binary Tree method, and Double Diffusion method to achieve the targets of reducing communication time and optimizing communication performance.