Guide to Partitioning Unstructured Meshes for Parallel Computing

Unstrucutured grids are used frequently in finite element or finite volume analysis. Unlike structured grids which are mostly applicable to finite difference schemes, unstructured grids require a list of the connectivity which specifies the way that a given set of vertices form the individual elements. To implement models that use an unstructured numerical decomposition on a distributed memory computer system, careful consideration is required when partitioning the initial grid. This Computational Science and Engineering (CSE) report discusses the process of partitioning an unstructured grid. First, we outline the algorithms behind graph partitioning and introduce the most commonly used applications for performing this work. Two widely used packages are introduced namely METIS and Scotch, together with how they may be implemented for use on HECToR and this section of the report is intended to be used as a quick start guide. Finally, we compare the efficiency of partitions produced from using these two packages by demonstrating their use in partitioning an unstructured grid for use with the CABARET finite volume code.