基于通信优化的Infiniband集群MPI作业加载*

针对如何缓解Infiniband集群中因通信冲突引起的MPI程序性能下降问题进行了研究，从系统管理的角度出发，提出了通过改变进程映射来优化MPI作业加载方案从而优化应用程序通信性能的方法，设计了用于评价MPI作业加载方案的通信性能损失系数（CPLR）指标，基于模拟退火算法设计了优化加载方案的搜索算法，并对所提出的指标和算法进行了实现和测试。测试结果表明，经过优化加载后的MPI程序在通信性能上有一定程度的提高。

MPI job launching based on communication optimization on Infiniband clusters

This paper unexpected performance loss of MPI parallel programs caused by the communication congestions on Infiniband clusters. Standing on the point of supercomputer system management and administration,proposed a method to get better communication performance for MPI applications on Infiniband clusters by launching MPI jobs with optimized process mapping schemes.It introduced the CPLR (communication performance loss rating) indicator to evaluate the quality of job launching scheme.Then it adopted simulated annealing algorithm to search the optimal job launching scheme according to the indicator.It presented detail experiments on the test platform. The results demonstrate that optimized launching scheme will benefit the communication performance of MPI applications.