CFDA parallel Method for Adaptive Unstructured Grids with Optimum Static Grid Repartitioning

Abstract

This chapter presents a technique for redressing load-imbalances, which arise in parallel computational fluid dynamics (CFD) methods using adapted unstructured grids. This technique is applicable to any CFD code parallelized through the multidomain concept and message-exchange protocols, like the parallel virtual machine (PVM). Data subsets—unstructured grid partitions—are defined using an efficient, genetic algorithm (GA) based tool and the adapted grid is repartitioned after each adaptation cycle. In order to combine the advantages of both grid adaptation and parallelization, without their side-effects, several load-balancing algorithms have been proposed. These can be classified in static grid repartitioning (SGR) and dynamic load-balancing (DLB) algorithms. A flow solver for adaptive unstructured grids is ported to a low-cost, distributed memory parallel platform using the multidomain technique and the PVM communication protocol. In order to redress load imbalances during the parallel execution, the so-called SGR algorithm, where one processor undertakes the repartitioning of the flow domain after each adaptation cycle, is proved to perform better than DLB, which is based on the migration of grid cells between processors. Despite its distinct sequential parts, the so-called SGR algorithm, yields high parallel efficiencies and presents certain advantages over DLB algorithms, which rely on the exchange of excess grid entities between processors.