一种等性能面积的并行计算可扩展性度量方法

A Novel Scalability Metric for Parallel Computing

可扩展性是并行计算的一个重要性能指标,但传统可扩展性度量机制只是试图从某一个侧面反映并行计算可扩展情况,难以全面度量并行计算系统综合性能.深入和全面地研究可扩展性度量机制,从众多性能指标中精选一组关键参数,对其进行归一化处理,然后用Kiviat图的面积来刻画并行计算的综合性能,由此给出一种新的等性能面积的并行计算扩展性度量机制,并进一步分析新度量机制和传统机制的关系.最后,应用新度量机制分析矩阵相乘算法在LogP计算机模型上的扩展性,并在集群平台上实际运行程序进行扩展性实验,进一步验证新机制的有效性.新度量机制对于指导并行计算体系结构完善,以及并行算法设计和调优有重要意义.

Scalability is a very important performance metric of parallel computing, but the traditional scalability metrics only try to reflect the parallel computing scalability from one side, which is difficult to fully measure its comprehensive performance. This paper studies scalability metrics deeply and fully. We choose a group of key metrics from lots of performance parameters and normalize them, then characterize the overall performance of parallel computing by the area of Kiviat graph which is posed by the group of key parameters. Thereby, we propose a novel scalability metric about iso-area of performance for parallel computing which measures the scalability of parallel computing by comparing the initial area of the Kiviat graph with the extended one. And the relationship between the new metric and the traditional ones is further analyzed. Finally, the novel metric is applied to address the scalability of the matrix multiplication algorithm under LogP model, and the experiments about extension are carried out on cluster platforms by running the program for the algorithm in order to further validate the effectiveness of the new metric. It is significant to improve parallel computing architecture and to tune parallel algorithm design.