Iain Duff
Rutherford Appleton Laboratory & CERFACS
UK & France
email: I.S.Duff@rl.ac.uk
Abstract:
We first review the various levels of parallelism that are available in the direct solution of large sparse linear systems. These can be summarized as being at a system, matrix, and loop level. We then examine how partitioning can be used to exploit corse-grained parallelism and show that good performance can be obtained on machines with up to 16 processors, that are very commonly used in engineering calculations. We will also briefly consider iterative as well as direct methods in this latter study.