Semi-automatic Parallelisation of Unstructured Mesh Codes

Abstract

This chapter discusses the general parallelization strategy based on a decomposition of the mesh for unstructured mesh-based codes. The strategies embedded within the parallelization tools, Computer Aided Parallelization Tools (CAPTools), have evolved through experience in manual parallelizations. The use of CAPTools instead of a manual parallelization significantly reduces the time and effort required to parallelize such codes, typically from months to hours/days. In addition, the Single Program Multi Data (SPMD) code generated by CAPTools can be optimized, since the code is very similar to the original serial code and therefore easy to understand. Although the code generated is portable and still retains a high degree of parallel efficiency, it can also be optimized further for a particular architecture, for example, better cache usage. For unstructured mesh codes, two alternative partitioning strategies are commonly applied. When a direct solver dominates execution time, a cyclic decomposition of the system matrix may be required. When no such obstacle exists, a geometric decomposition of the unstructured mesh usually enables an effective parallelization. Both partitioning approaches are available within CAPTools. For mesh decomposition, the partition details are stored as a list, where each entry indicates the owning processor of that partitioned array element.