Analysis And Optimization Of The Communication Mechanism For MPICH On The Cluster Of Workstation

MPICH is the most widely used implementation of MPI standard. It is a parallel function library which bases on message-passing, and it can be ported on variable computer systems. MPICH is a kind of open source software, but there is little information which can introduce how they implemented. The MPICH features high computing capability and porting property, but as for the cluster, it performs high communication delay. So, the communication optimization of MPICH plays a very important role in the improvement of the efficiency that the parallel program. By analyzing the source code of MPICH from the perspective of communicatory implementation, the MPICH communication optimization is achieved and an improved parallel function library is designed.MPICH creates the TCP connections between processes rather than nodes when communicating. Experimental results show that the TCP connections between the processes of every node increase the communication expenses and do nothing good to the communication optimization.It can be reduced to one TCP connection obviously. So, we take the method of creating one communicating management process on every node to take charge of all the TCP connections.In this way, not only the structure of the connections between processes is optimized, but also the communication expense especially the communication expense between processes on different nodes is decreased.In cluster of workstation, the task initiation is serially, when the user starts the MPI processes on the master node, the master node start tasks on the remote nodes in turn. In this situation, the communication expenses increases quickly with the increasing of the number of nodes. A parallel task initiation based on Doubling Technique is proposed in the thesis, and an experimental verification is done. The parallel task initiation takes the diffused tree structure to start tasks on nodes. In this way, the communication expense is decreased and we get a higher communication performance.At last, this thesis realizes an archetype and actualizing details of the parallel function library according to the methods of communication optimization mentioned above. And the implemental principle is described through idiographic examples.