Research On The Optimitation Method Of Group Communication On Cluster System

Cluster system simply is composed of various common commercial component by the way of network. Cluster system with its high scalability, high availability and cost-effective advantages gradually becomes the trend of high-performance computer architecture development. With the increasing number of information between cluster, the efficiency of cluster communication system, which is used of contact each nodes and transmit informations, is directly effect the whole cluster system performance, so optimization the communication of cluster became the focus of attention. Group communication is the most popular method of optizing the communication of cluster nodes, MPICH provide the special communication interface and appear all kinds of representative algorithms, likely Bruck algorithm,ring algorithm,recursive algorithm and pairwise exchange algorithm etc.This paper based on the method how to optimize group communication to study, mainly include:1.It analyzes the current of cluster communication system, introduces the development situation of group communication, and detailed exposition cluster system to use parallel computing.2.It specific analyzes four typical algorithms of group communication, ring algorithm,recursive algorithm,neighbor exchange algorithm and Bruck algorithm's thought and the communication performance, deeply analyze the group communication's mechanism based on MPICH2-1.0.8.Introduction the influence of improving group communication on the performance of cluster system, and summarizing optimization method.3.Through the analysis of four kinds of performance of the algorithm, respectively proposed optimization algorithm to all-to-all communication and allgather communication. The improved algorithm in all-to-all creates the subprogram in a recursive doubling manner, reduces the number of communications by increasing the number of processes. It analyzes comparatively the number of communications between the improved algorithm and the pairwise exchanges,and the improved algorithm is half as much as the latter. Allgather algorithm's optimization algorithm combine the ring algorithm and neighbor exchange algorithm is. The optimization algorithm requires the operation of the one-step ring algorithm, which may guarantee two data for each node machine to enhance the communication partial performance. Then, the new neighbor exchange algorithm is used to increase the number of the exchange data in a way of power of 2 to reduce the frequency of the communication. The testing results indicate that the optimization algorithm performs better than either neighbor exchange algorithm or ring algorithm with regard to the long information communication, especially when the number of communications nodes machine is power of 2, the optimization algorithm's performance can reach the best.