Automatic load balanced paritioning strategies for PDE computations

Abstract

In this paper we study the partitioning and allocation of computations associated with the numerical solution of partial differential equations (PDEs). Strategies for the mapping of such computations to parallel MIMD architectures can be applied to different levels of the solution process. We introduce and study heuristic approaches defined on the associated geometric data structures (meshes). Specifically, we study methods for decomposing finite element and finite difference meshes into balanced, nonoverlapping subdomains which guarantee minimum communication and synchronization among the underlying associated subcomputations. Two types of algorithms are considered: clustering techniques based on sequential orderings of the discrete geometric data and optimization based techniques involving geometric or graphical metric criteria. These algorithms support the automatic mode of a geometry decomposition tool developed in the parallel ELLPACK environment which is implemented under X11-window systems. A brief description of this tool is presented.