Concurrent solutions of large sparse linear systems

Abstract

Computer-aided circuit analysis, or circuit simulation, is widely used in the area of circuit design. Circuit simulation programs, e.g. SPICE, create and solve systems of differential equations which describe the analyzed electronic circuit. The systems of differential equations are converted into nonlinear algebraic equations and solved through a sequence of linear approximations to the nonlinear equations. -- Parallel processing is a promising way to improve the performance of circuit simulation programs. Several attempts to port sequential codes into equivalent ones for shared-memory architectures have been reported in the literature. However, with the increasing popularity of message-passing systems, our project aims at the parallelization of a circuit simulation program on a network of workstations. A domain decomposition approach was implemented through a master-slave model on a cluster of SUN stations. The experiments show speedups over up to 8 workstations. This presentation discusses the implemented algorithm and provides an overview of some performance results.