The Research Of Matrix Multiplication Efficiency Based On MPI

Parallel computing is main theme in21th and is considered as a standard to solve problems in many applicative areas as engineering design,science computing. With the development of high-speed network technology, Parallel computing has been laboratory-based transformation to calculation-based, which is characterized as cluster computing based on LAN and Grid computing based on Internet.Cluster computing has become a hot research topic and the main bend of the high-performance computer system,because it provides computing power in a flexible and very low-cost way.PC cluster in LAN that based On parallel computing environment of message passing interface(MPI) have large potential abilities in network parallelcomputing.But now,there’re many questions of matrix multiplication which based on the environment of MPI need to besolved by further research.This paper focused on the following aspects:Firstly,general theory of parallel programming. Through analyzing the factors impacting on the performance of parallel, I put forward a feasible way to improve the parallel performance within the local area network (LAN).Secondly,because the current primitives is not specifically for matrix transmission,there is large cormnunication overhead of matrix transport in parallel computing. Based on the above problems, this paper proposes a new matrix transport protocol to reduce transmission overhead.Thirdly, in order to reduce the parallel computing time while the load was kept balanced, this paper puts forward a new task scheduling algorithm..Finally,the experiment platform was construct based on LAN, the experiment results were analyzed.