Analyze And Optimize The Protocol Of Cluster Based On MPI

The Cluster technology, which is the newly appear, through the high performance network those computers connect as a single aggregate, every node has its own memory, and I/O device and operation system, managed with single system mode. It is main purpose is to provide highly available, highly extended and highly robust. It provide high performance and quickly and reliable sevice.The protocol used in cluster is based on TCP/IP protocol, so it maybe has the condition that the performance of running a parallel program is less than on single computer, for the communication time is larger than computer time. So the user is more care the performance of communication when the balance between the lower rate of error in the net of cluster and the performance was changed. How to get the lower latency of communication machine to get rid of the bottle-neck domino effect in cluster. It is the key of this paper, the factors which caused the efficiency to be lower have been analyzed, and measure the time spend on the every layer in protocol. for consider of the protocol of TCP,It has many overhead,mainly as those three parts, the overhead of protocol, the overhead of data copy, including copy data from user mode to system mode, the other one is the operation system overhead, including the allocation of memory, switch of progress, deal with Interrupt. In this paper, we focus on the overhead of protocol and do partly modification, change the buffer size and send window size to decrease the time waiting for in send buffer and wait for allocating memory. And disable the checksum in software, let it to be down by hardware.The System based on Ethernet, through analyses of kernel in linux and modify. Last rebuild it. Do the experience with the tradition protocol of TCP/IP, it has validated the validity of protocol. Be careful of the code of network part, don't change the interface of code, be sure of the usability of system.