Performance Analysis of Distributed Iterative Linear Solvers

Abstract

The solution of large, sparse systems of linear equations is an inherent part of many computational methods in science and engineering. For such systems, iterative methods are often more attractive than direct methods because of their small (and constant) memory requirements. Also, the performance of iterative solvers can easily be improved by using distributed systems. A common performance characteristic of distributed applications is their speedup which is usually defined as the ratio of the execution time of an application on a single processor to the execution time of the same workload on a P–processor system. The paper estimates the speedup of distributed linear iterative solvers, analyzes the influence of different communication schemes on the speedup, and compares the estimates with the measurements of real distributed programs.