Joint work with T.Basaruddin, B.A.A.NaziefIn solving complicated problems, parallel computing has become a potential approach for achieving high performance. The performance of a parallel system can be evaluated through the scalability analysis under different constraints, such as fixed load, fixed execution time and fixed processor efficiency. Although there is no common definition for scalability, in general it expresses the ability of a parallel system to preserve its performance with the increasing number of processors, measured in some metric.
In this paper we present a new approach of scalability analysis that was proposed by Hockney in 1995, named as DUSD (Dimensionless Universal Scaling Diagram) which discussed the scalability of a parallel system based on 3-parameter model of parallel hardware and parallel program. Hockney also introduced a very powerful concept, Computational Similarity, which allows a scalability analysis of large classes of seemingly different parallel applications, executed on any parallel architecture that can be described by the 3-parameter model. We investigate the strengths and weakness of the scalability analysis based on this DUSD approach compared to those of Amdahl's and Gustafson's. As our case study we chose Level 1 BLAS (vector inner product and vector update) and Level 2 BLAS (matrix-vector product) due to many numerical algorithms used to model physical problems have those subroutines as their core solver. Analyzing the scalability of these routines may help us in predicting the performance of the overall system.
We chose a cluster-based system as our research platform because of its widespread use and ability to provide high performance. Our experiments were conducted by using PVM (parallel virtual machine), a software framework that emulates parallel computer architecture on general-purpose computers in a networked environment.
We conclude that the scalability analysis based on DUSD concept extends those of Amdahl and Gustafson. However in some cases such as BLAS1, DUSD concept is not applicable unless we adopt certain assumptions. These assumptions to validate the DUSD analysis will also be discussed in this paper.