多核SMP集群并行计算的可扩放性研究

发布时间：2019-05-25 00:30

【摘要】：随着高性能计算技术的飞速发展,并行计算已经成为提高计算机系统性能的主要方法,并行计算的可扩放性也成为评价并行计算性能的主要指标之一。但随着高性能计算机系统规模的不断增长,并行计算系统的结构日益复杂,编程困难、存储量大、效率下降等一系列的问题大量出现,以上问题都从不同方面降低了并行计算的性能,从而对并行系统的正常运行产生了一定的影响。由此可见并行计算的性能不能随着机器规模的增加而提升。现如今可扩放性已成为并行处理中一个重要的研究问题,被越来越广泛地用来描述并行算法(并行程序)能否有效利用可扩充的处理器数的能力。因此,深入和全面地研究并行算法和并行机相结合的可扩放性是对并行计算性能进行合理、精准评价的关键。针对以上情况,本文主要研究多核SMP集群系统并行计算的可扩放性,可扩放性评价准则是研究并行计算可扩放性的基础,建立合适的可扩放性评价方法,以实现并行计算性能提高的目标。论文主要研究内容如下:首先,论文介绍了并行计算机体系结构和模型以及多核SMP集群系统的体系架构特点,然后详细介绍了可扩放性的概念以及相关技术,论文在深入分析典型的可扩放性评价准则的基础之上,指出了它们的适用范围和优缺点。其次,根据多核SMP集群系统体系结构特点,改进了传统等效率可扩放性评价准则。分析了多核SMP集群中的处理器集合存在不完全等价的问题,为了解决不完全等价的问题,本文以最优处理器集合的概念为基础,并对处理器集合的计算能力进行定义,改进了传统等效率可扩放性评价准则以适用于多核SMP集群系统,并分析了改进等效率可扩放性评价准则,得出传统评价准则是改进评价准则在一定条件下的一种形式,以及改进评价准则与并行执行时间之间的关系等重要结论。此外,在改进等效率可扩放性评价准则的基础上进一步研究了如何调节集群系统的节点间可扩放性,为了并行算法尽可能地充分利用增加的处理器,选择合适的网络互联结构,增加网络通信带宽,可以有效地减少通信延迟和计算延迟,进而提高系统的性能。因此提出集群系统中基于带宽的节点间可扩放调节系数概念,进一步优化评价准则,从理论和实验方面分析了节点间可扩放调节系数对系统性能的提升起到重要作用。最后,论文将曙光TC5000作为并行计算的实验环境,并在集群平台上实际运行矩阵相乘算法进行扩放性实验,进一步验证改进等效率可扩放性评价准则的有效性,对得到实验数据进行详细的分析。实验结果表明,此可扩放性评价准则较好的分析了多核SMP集群系统的可扩放性,对于指导并行计算体系结构完善,以及并行算法设计和改进具有重要的意义。
[Abstract]:With the rapid development of high performance computing technology, parallel computing has become the main method to improve the performance of computer systems, and the scalability of parallel computing has become one of the main indicators to evaluate the performance of parallel computing. However, with the continuous growth of the scale of high performance computer system, the structure of parallel computing system is becoming more and more complex, the programming is difficult, the storage capacity is large, the efficiency is declining and so on. The above problems reduce the performance of parallel computing from different aspects, which has a certain impact on the normal operation of parallel systems. It can be seen that the performance of parallel computing can not be improved with the increase of machine size. Nowadays, extensibility has become an important research problem in parallel processing, which is more and more widely used to describe whether parallel algorithms (parallel programs) can effectively make use of the number of expandable processors. Therefore, it is the key to evaluate the performance of parallel computing reasonably and accurately to study the extensibility of parallel algorithm and parallel computer deeply and comprehensively. In view of the above situation, this paper mainly studies the extensibility of parallel computing in multi-core SMP cluster system. The evaluation criterion of extensibility is the basis of studying the extensibility of parallel computing, and a suitable evaluation method of extensibility is established. In order to achieve the goal of improving the performance of parallel computing. The main research contents of this paper are as follows: firstly, the parallel computer architecture and model and the architecture characteristics of multi-core SMP cluster system are introduced, and then the concept of extensibility and related technologies are introduced in detail. On the basis of in-depth analysis of typical extensibility evaluation criteria, this paper points out their applicable scope, advantages and disadvantages. Secondly, according to the architecture characteristics of multi-core SMP cluster system, the traditional evaluation criterion of equal efficiency extensibility is improved. This paper analyzes the problem of incomplete equivalence of processor sets in multi-core SMP cluster. In order to solve the problem of incomplete equivalence, this paper is based on the concept of optimal processor set, and defines the computing power of processor set. The traditional evaluation criterion of equal efficiency extensibility is improved to apply to multi-core SMP cluster system, and the improved equal efficiency extensibility evaluation criterion is analyzed. It is concluded that the traditional evaluation criterion is a form of improved evaluation criterion under certain conditions. And some important conclusions, such as the relationship between the improved evaluation criteria and the parallel execution time, are also discussed. In addition, on the basis of improving the evaluation criterion of equal efficiency extensibility, this paper further studies how to adjust the inter-node extensibility of cluster system, and selects the appropriate network interconnection structure in order to make full use of the increased processor as much as possible in the parallel algorithm. By increasing the network communication bandwidth, the communication delay and computing delay can be effectively reduced, and the performance of the system can be improved. Therefore, the concept of inter-node expandable adjustment coefficient based on bandwidth in cluster system is put forward, and the evaluation criterion is further optimized. from the theoretical and experimental aspects, it is analyzed that the inter-node expandable adjustment coefficient plays an important role in improving the performance of the system. Finally, the dawning TC5000 is used as the experimental environment of parallel computing, and the expansibility experiment is carried out by running the matrix multiplication algorithm on the cluster platform to further verify the effectiveness of the improved equal efficiency extensibility evaluation criterion. The experimental data are analyzed in detail. The experimental results show that the extensibility evaluation criterion is a good analysis of the scalability of multi-core SMP cluster system, which is of great significance to guide the improvement of parallel computing architecture, as well as the design and improvement of parallel algorithms.
【学位授予单位】：曲阜师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP338.6

【参考文献】