Research And Implementation Of OpenMPI Communication Mechanism For Multicore Processor System

With the rapid development of science and technology, more and more disciplines began to use scientific computing and numerical simulation tools to solve major problems. High performance computing affects the national defense and national security directly and has become an important factor to measure a nation's comprehensive strength. The technology of multi-core processor is widely used in high performance computing system in order to overcome the bottleneck of the processor performance. MPI, as the main parellel programming mode and enviroment for parallel and distributed systems, the performance of its implementation directly affects the efficiency of high performance computing. Open MPI is an open source, high performance MPI implementation, which uses component-based architecture, not only provideing a stable platform for third-party researchers, but also increasing the flexibility of the platform. Open MPI has a good performance in inter-node massage passing and intra-node short message passing, but poor in intra-node large message passing.This thesis proposes two optimizations in order to improve the performance of intra-node large message passing of Open MPI in multi-core architecture. We design and implement a message passing component KBDT based on a kernel module memory mapping. It shows that, KBDT component can effectively improve the intra-node large message passing performance of Open MPI, and has a better scalability than the SM component. The main works are as follows:First, we study the intra-node communication protocols and analyze their strengths and weaknesses. Analyze the structure of the current widely used MPI implementation MPICH2 and the implementation of its nemesis channel which is a high performance communication subsystem of MPICH2.Second, we analyze the communication structure of Open MPI, and study the shared memory based message passing protocol in SM; evaluate and analyze the performance of SM components comparing with nemesis of MPICH2 and propose two optimization solutions.Third, we design a massage passing component KBDT for the intra-node communication of Open MPI. Implement a prototype of KBDT using the kernel-based memory mapping mechanism and message passing protocols.Fourth, we evaluate the performance of KBDT component comparing with SM. The experiments contain point-to-point communication tests, collective communication test and NPB benchmark evaluation. The results illustrate the high performance of KBDT component, especially for the large message passing. It effectively improves the intra-node communication performance in Open MPI and provides an efficient MPI implementation platform for message passing in multi-core systems.