Research And Application Of Hierarchical Parallel Algorithm Based On The MPI Environment

In recent years, the rapid development of computer technology greatly promoted the progress of the calculation science, the scope of the computer application is more and more extensive and the scale of problems dealing is increasingly larger. In order to solve these large amounts of practical problems, only by enhancing the performance of the computer hardware is far from meeting, and the enhancement also has certain restriction. Therefore an effective and popular solution is to adopt parallel computing to improve the computer computations. Then, the multi-core processor emerges as the times require. The advent of the multi-core processor has brought strong power to the calculation science, but also brought greater challenges to the computer software. The traditional parallel programming model is divided into two kinds, one kind is Shared Storage Model, mainly is the threads, another kind is the Message Passing Model, generally the distributed storage model. However, these models all have some defects being not good at promoting process performance. Basing on the analysis and comparison of the two kinds of parallel programming model, this paper is to seek an algorithm that matches and maps with the parallel computing system very well, and the algorithm can organize parallel assignments reasonably, greatly improving parallel computing performance.This paper firstly introduces the parallel computing and parallel thought, and then introduces the architecture of the parallel computer, and on this basis I analyze and compare all sorts of parallel computing programming environment. Then introduces MPI that is a kind of popular parallel programming model at present based on message passing, and writes the detail process of setup and configuration of MPICH2, which is the typical realization of MPI. Through the analysis of the MPI object semantics and the MPICH2 internal background, I describe the optimization of one aspect of a multithreaded MPI implementation and introduce how to minimize MPI resource contention in multithreaded multicore environments; the new solution improved the multithreaded messaging rate by 31% in my experiment platform. However, just relying on the optimization to MPI can not improve the performance of parallel computing very well, so I produces an idea combining this Message Passing Model with the Shared Storage Model, then this paper in detail introduces Threading Building Blocks that is one of the Shared Storage Model, and puts forward feasibility and inferiority of the combination of TBB and MPI. This paper provides a method when facing an actual problem, first, you should take the problem into MPI decomposition, in which the problem is decomposed into several subproblems their communication are not crowded, each subproblem is assigned to each node, the nodes communicate through the message passing manner that are coarse granularity; then takes each subproblem decomposed into several subtasks, these subtasks has the fine granularity by TBB in each node. Finally I summarized all the work in this paper, the experience got in the work and problems found in the search, and then I put forward the further studv.