基于SMP机群的并行编程模型的研究应用与优化

发布时间：2018-05-31 03:57

本文选题：SMP机群 + MPI+OpenMP　；参考：《曲阜师范大学》2014年硕士论文

【摘要】：随着科学技术的不断发展，各个领域对于大规模数据处理的需求日益增长，这极大的促进了并行计算的发展，同时高性能计算机也因此越来越受到广泛的应用，其中SMP机群因为具有较高的性价比和良好的可扩展性以及可用性等特点备受欢迎，成为当前并行计算机领域中的主流结构，也成为了高性能计算领域的研究重点。本文的主要研究内容就是SMP机群以及适用于该体系结构的编程模型，并提出了优化措施。本文首先介绍了课题的研究背景以及并行计算技术和并行计算机体系结构的发展。随后介绍了几种当代流行的并行计算机系统，，对其结构和特点进行了详细分析，同时对当前主流的并行编程模型进行了介绍。接着对本文的重点进行了详细研究和分析：SMP机群及适用于SMP机群的并行编程模型。论文中对SMP机群的体系结构做了详细分析，SMP机群结合了SMP系统和机群系统的优点，它以SMP作为节点，内部采用共享存储的结构，而节点之间则采用分布存储的结构，在SMP机群上可以充分利用共享存储编程模型和分布存储编程模型进行编程。其中OpenMP和MPI作为两种编程模型的代表同样是本文的研究重点，针对SMP机群特殊的体系结构，介绍了一种MPI+OpenMP的混合编程模型，并给出了混合编程模型的结构、实现机制以及其优势，分析了MPI+粗粒度OpenMP和MPI+细粒度OpenMP的两种并行化方式的优缺点，同时提出了一些优化措施，进一步提高混合编程模型的性能。本文的实验部分首先介绍了实验的软硬件环境，然后介绍了实验部分进行性能评测的方法，实验重点是实现了N体问题的PP算法的并行程序设计。针对这个问题设计了相应的MPI程序和MPI+OpenMP混合程序，并对混合编程进行了优化，最后将程序在曙光TC5000机群上对程序进行了测试。通过测试获得程序的执行时间进而求得加速比，证明了在一定情况下混合编程模型是可以通过优化进一步提高性能的，对比分析得出在SMP机群上MPI+OpenMP混合编程模型要好于纯MPI编程模型。论文的最后对前面的主要研究内容进行了概括，并对下一步可进行的工作进行了展望。
[Abstract]:With the continuous development of science and technology, the demand for large-scale data processing in various fields is increasing, which greatly promotes the development of parallel computing, at the same time, high-performance computers are more and more widely used. Among them, SMP cluster is popular because of its high performance-price ratio, good scalability and usability. It has become the mainstream structure in the field of parallel computing, and has also become the focus of research in the field of high performance computing. The main research content of this paper is SMP cluster and programming model suitable for this architecture, and the optimization measures are proposed. This paper first introduces the research background and the development of parallel computing technology and parallel computer architecture. Then several popular parallel computer systems are introduced, their structure and characteristics are analyzed in detail, and the current mainstream parallel programming models are introduced. Then, the emphasis of this paper is studied and analyzed in detail, and the parallel programming model for SMP cluster is analyzed. In this paper, the architecture of SMP cluster is analyzed in detail, which combines the advantages of SMP system and cluster system. It takes SMP as the node, adopts the structure of shared storage, and adopts the structure of distributed storage among nodes. The shared storage programming model and the distributed storage programming model can be fully utilized in the SMP cluster. OpenMP and MPI as representatives of two programming models are also the focus of this paper. According to the special architecture of SMP cluster, a hybrid programming model of MPI OpenMP is introduced, and the structure of hybrid programming model is given. The implementation mechanism and its advantages are analyzed. The advantages and disadvantages of the two parallelization methods of MPI coarse-grained OpenMP and MPI fine-grained OpenMP are analyzed. At the same time, some optimization measures are proposed to further improve the performance of the hybrid programming model. In the experiment part, we first introduce the hardware and software environment of the experiment, and then introduce the method of performance evaluation in the experiment part. The key point of the experiment is to realize the parallel programming of PP algorithm for N-body problem. In order to solve this problem, the corresponding MPI program and MPI OpenMP hybrid program are designed and optimized. Finally, the program is tested on the dawning TC5000 cluster. The speedup ratio is obtained by testing the execution time of the program. It is proved that the hybrid programming model can further improve the performance through optimization under certain circumstances. Comparison and analysis show that MPI OpenMP hybrid programming model is better than pure MPI programming model on SMP cluster. At the end of the paper, the main research contents are summarized, and the future work is prospected.
【学位授予单位】：曲阜师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP311.11;TP338.6

【参考文献】