Research On Parallel Programming And Optimization Techniques For General-purpose Multi-core Clusters

Along with the emergence of multi-core processor architecture and the gradual popularization of symmetric multiprocessor and cluster architecture,new multi-core clusterarchitecture arises,and for its higher performance-price-ratio and better scalability,it becomes a research hot focus in the field of high-performance computing.Multi-core processor architecture and high-speed Internet sustain rapid development,multi-core cluster structure has also been improving,and developers will face a more complex programming environment.In this paper,the author takes the multi-core cluster as a platform,and aiming at its unique architecture,explores its programming models,mapping mechanism,implementation mechanism,programming methods and optimization strategiesto which it applies.In this paper,the author first introduces the research background of the subject and the development story of parallel computing;second,introduces the main architecture of the parallel computer,wherein their main characteristics are analyzed and the abstract diagram of the system structure is given;third,introduces the three programming models of parallel programming and analyzing their characteristics respectively;at last,introduces two main parallel programming models in a multi-core parallel programming environment: a shared memoryprogramming modelOpenMP and a message delivery programming model MPI,wherein the characteristics of these two programming models and their execution model are analyzed in depth,including their execution mode andthe advantages and disadvantages of programming,and puts forward corresponding optimization strategies.In this paper,the author focuses on the study of a hybridprogramming;introducesan OpenMP/MPI programming model and analyzes the advantage of the hybrid programming;aiming at the architecture of multi-core cluster,proposes an impact on the performance of the cluster from a mapping mechanism of ahybrid programming model and a multi-core cluster architecture;andprovides an implementation mechanism and an optimization strategy of the mixed model programming design;and proposes the problems that needed attention in the mixed programming process.In this paper,the hardware and software environment of the dawn TC5000 cluster is introduced during the experimentation process.The author focuses on the study that conjugate gradient algorithm is used to solve the linear equations which is implemented in parallel in the multi-core cluster and carries out relevant experiments;through the comparisonand analysis of the experimental data,obtains a fact that compared to the single parallel programming model,the mixed programming model can matches its unique architecture more intuitively and has better performance,and the number of processes in the hybrid model is significantly reduced and is equivalent to the number of nodes rather than the total number of processors in the MPI model,and the most efficient number of accelerator threads is equal to the number of processors within the node.In addition,by researching and analyzing the factors that affect the performance of MPI/OpenMP programming model on multi-core clusters,the optimization strategies and methodsfor improving performance are proposed.At last,the above-mentioned researches are summarized,and future work is prospected.