SMP集群系统的可扩放性分析

Analysis on Scalability of SMP Cluster System

随着并行计算技术的快速发展和SMP集群的普及,可扩放性已经成为并行应用程序设计和实现方面最重要的性能之一。但传统的可扩放性评价准则不能对SMP集群的可扩放性进行较精准的评价。为此,在分析SMP集群中处理器集合的特性和传统等效率模型并掌握其优缺点以及分析并行计算速度的基础上,给出了一种适合SMP集群系统效率的定义,并基于该定义提出了一种新的可扩放性评价准则(改进的等效率可扩放性评价准则)。该新准则可用来评价并行算法和SMP集群相结合的可扩放性。为验证所提出评价准则的有效性,在集群平台上运行矩阵乘法程序进行了相关的扩放性实验。实验结果表明,所提出的新评价准则对算法和并行机的最优匹配有指导作用,同时有助于对并行算法的设计和改进。

With the rapid development of parallel computing technology and the popularity of SMP clusters, scalability has become one of the most important performance for parallel application program design and implementation. However, the traditional scalability metric cannot evaluate the scalability of SMP cluster accurately. Therefore, the definition has been proposed to describe the efficiency of SMP cluster system based on the analysis of the property of the set of processors and the ISO-efficiency function in SMP duster, understanding their merits and deficiencies, as well as the analysis of the parallel execution speed. Then a new metric,improved ISO-efficiency metric, has been proposed, which can be used to measure and evaluate the scalability of parallel algorithms and SMP cluster. The extension exper- iments have been carded out on the cluster platforms by running the program for the matrix multiplication algorithm in order to validate the effectiveness of the new evaluation criteria. Experimental results show that the new evaluation criterion has guided the optimal matc- hing of the parallel application and the parallel machines. Thus it is helpful to the design and the improvement of parallel algorithms.