如何测量SMP机群可扩放性(英文)  

How to Measure Scalability of SMP Cluster

作　　者：何家华[1] 陈国良[1] 单久龙[1] 

机构地区：[1]中国科学技术大学计算机科学技术系国家高性能计算中心(合肥),安徽合肥230027

出　　处：《软件学报》2004年第7期977-986,共10页Journal of Software

基　　金：国家高技术研究发展计划(863)~~

摘　　要：可扩放性是并行计算的一个重要性能标准,但是传统的可扩放性准则并不适用于SMP机群.如何测量SMP机群的可扩放性?试图提出该问题的一个解决方案.首先找出并验证问题的根源处理器集合不等价性.然后,采用处理器集合的观点来全面、正确地观察系统的行为,而并非像传统的做法那样仅仅使用处理器数来描述并行系统.通过引入性能参考因子的概念,扩展了传统的准则以适应SMP机群体系结构.实验结果显示,扩展后的度量准则适用于SMP机群,且具有较高的准确性.Scalability is an important performance criterion of parallel computing. However the conventional scalability metrics are not suitable for SMP(symmetric multiprocessor) cluster. How to measure scalability of SMP cluster? This paper proposes a solution to the problem. It first finds out and verifies the reason, nonequivalence of the processor sets and then it adopts the viewpoint of processor set to observe the behaviors of the system correctly and comprehensively instead of using only the number of processors to describe the parallel system. By introducing the concept of performance reference factor, it extends the conventional metrics to fit the SMP cluster architecture. As the experiments indicate, the extended metrics are applicable to the SMP cluster and have high precision.

关 键 词：并行计算 性能评测 SMP机群 可扩放性 

分 类 号：TP338[自动化与计算机技术—计算机系统结构]