Optimization Of MPI Communication Library On KD60 Platform

Large-scale cluster has been developed to multi-core era with the improvements in hardware architecture, multi-core architecture proposed new requirements to parallel computing. Communication plays an important role in parallel computing, enhancing communication efficiency also means high performance of the parallel computing.KD60 is a Teraflops multi-core cluster using the first domestic multi-core chips---loogson III. Message passing model and shared memory model are two commonly used parallel computing models and MPI is the most commonly used message passing model, MPICH2 communication library is an important realization of the MPI-2 standard.The main work of this paper is transplanting MPICH2 into the domestic high performance cluster KD60 and optimizing the performance of MPICH2. Several optimization methods were proposed to optimize the performance of point to point communications and collective communications, on the basis of the characteristics of the system architecture and network topology in KD60. The main work can be summarized as follows:1. Analyzes the characteristics of multi-core architecture and its memory hierarchy model, and points out the communication features existed in multi-core architecture, and then briefly introduce a new communication analysis model.2. Analyzes the structure of MPICH2 and the implementation of its internal modules, and also points out the deficiencies of the MPICH2 implementation under the multi-core architecture.3. Optimizes the performance of network communication between nodes and the performance of shared memory communication in a node, for the point to point communications in MPICH2, respectively. Use technology such as adjusting network parameters and TCP checksum offloading for inter-node communication optimization, use methods such as data tiling, single data copy and using loogson III's 128 bits memory access instructions to optimize the efficiency of memory copy function to optimize intra-node communication performance.4. For the collective communications in MPICH2, optimizes the communication paths of the original collective communication functions, according to the communication features mentioned in Article 1. And takes broadcast and global exchange for example to introduce the details of optimization algorithm for one to multi mode and multi to multi mode functions of collective communications.Experimental results show that the optimization methods used in this article can enhance the performance of MPICH2 communication library in different extent.