Estimating Computer Performance for Parallel Sparse QR Factorisation

Abstract

Performance estimates of a parallel computer during sparse matrix factorisation aid in the identification of overheads and the tuning of software. This paper proposes a technique which allows the computer parameters of computation speed, communication speed, latency and parallel efficiency to be estimated. The technique is based upon the use of mathematical models derived from a model problem in conjunction with experimental results. By combining the mathematical models with the experimental results, sets of simultaneous equations can be derived which can be solved for the above computer parameters. The technique is explained in the context of sparse QR factorisation.