Sparse Iterative Solvers for Shared Memory Parallelism

S.A.Salvini

The Numerical Algorithms Group (NAG) Ltd, Wilkinson House, Jordan Hill Road, Oxford OX2 8DR, United Kingdom


Abstract

SMP or multithreaded Parallelism is becoming increasingly important particularly with the emergence of OpenMP as a portable API across a variety of platforms. SMP parallelism allows the use of serial interfaces to parallel modules, thus hiding parallelism details from users and allowing the immediate and effective porting of codes from serial to parallel platforms.

We have followed this philosophy in developing algorithms for the iterative solution of sparse systems of equations on shared memory platforms. The aim of this paper is to introduce our work, to report on the performance and scalability that we have obtained on a range of platforms, to discuss related problems, such as the development and use of preconditioners, and to describe our current and future lines of enquiry.