A More Effective Scalability Model for Parallel System
Yu Wang · Chinese Journal of Computers · 2001
The scalability is an important metric which reflects the performance changing property of the parallel system when the machine size and the problem size are scaled. In this paper, we first analyze the characteristics of three existing scalability models for parallel systems, i.e., the iso\|speed model, iso\|efficiency model and iso ratio of parallel overhead to computation model, prove the equivalence of the three conditions on which each model is defined, i.e., the isospeed condition, the isoefficiency condition and the iso ratio of parallel overhead to computation condition, and point out the limitations of these models when used to study the scalability of parallel systems. We then propose a new scalability model which reflects the scalability of parallel systems more directly and precisely. We prove the relationship of our model with the three formerly mentioned models, and applied our models to the investigation of issuses on scalability with three types of parallel systems characterized respectively by the Amdahl speedup model, the Gustafson speedup model, and a more realistic speedup model given by us. Case studies indicate that using our model, it is more efficient to solve the following problems: (1)investigating the scalability quantitatively; (2)comprehensively reflecting the effects of factors from program, machine and runtime environment on scalability; (3)providing suggestions on how to keep and improve the scalability of parallel systems.