Authors : Maria A. Kartawidjaja
Abstract: The demand for solving large and complex problems in a relative short time has motivated scientists to improve computational performance. The performance of a parallel system can be evaluated through the scalability analysis under different constraints, such as speedup and efficiency. Although, there is no common definition for scalability in general scalability describes the ability of a parallel system to utilize an increasing number of processors. In this study, we analyze the scalability of a parallel matrix multiplication on a cluster based environment using Dimensionless Universal Scaling Diagram (DUSD) proposed by Hockney in terms of speedup and compare it with Gustafsons scaled speedup model. We conclude that Hockneys concept provides us with a sophisticated way in analyzing the scalability of parallel system that extends Gustafsons speedup concept, nevertheless Hockneys model does not consider the overlapping between computation and communication and thus does not provide the actual speedup value. Furthermore, Hockneys concept concerns only with speedup on optimum number of processors, while in reality we utilize whatever number of processors available. Gustafsons speedup model, which is merely a function of program parameters is easy to implement and provide the actual value of speedup but it does not take into account hardware parameters that actually affect the parallel performance.
Maria A. Kartawidjaja , 2010. Analyzing Scalability of Parallel Matrix Multiplication Using Dusd. Asian Journal of Information Technology, 9: 78-84.