Dynamic spectral partitioning

Abstract

Recursive Spectral Bisection (RSB) has been proposed as an efficient heuristic algorithm for partitioning problem for unstructured problems about six years ago, and has since become a widely used tool when implementing finite element applications on parallel processors. However, many applications have a dynamically changing communication pattern, e.g. unstructured grid problems with adaptive refinement. In this talk we will outline a method which combines the advantages of RSB with the capability of quickly updating a partitioning, if the underlyng grid is dynamically changing. We call this method dynamic spectral partitioning (DSP). We will discuss the derivation of DSP, and motivate it with applications from structural analysis and computational chemistry. Then, using an adaptive grid refinement 3d Navier-Stokes code on the IBM SP-2, we will demonstrate how DSP can be applied to the solution of practical problems.