PaA4 Abstract

Title: Domain Decomposition Techniques for Solving Linear Systems (PaA4)

Authors: F.-X. Roux and J.P. Geschiere

Email: roux@onera.fr

Date: July 1994

Abstract:

This paper reports some studies made at ONERA for the implementation of domain decomposition methods for solving numerical engineering problems on a massively parallel MIMD system, the iPSC-860 with 128 processors, with message-passing programming environment. MCSPARSE is a parallel solver based on large grain parallelism, combined with medium and fine grain parallelism. Initially, MCSPARSE was specifically targeted for a multilevel memory platform, the Cedar architecture. Later, this version is used to develop a shared memory version, targeting the Cray Y-MP. The design of a distributed memory version of MCSPARSE is described together with its implementation on an iPSC/860. A comparative analysis is made of how the parallelism inherent in MCSPARSE is mapped onto the parallelism offered by each of the three target architectures, and how this relates to the performance of MCSPARSE on each of these platforms.

This report is available through