An unstructured grid partitioning method based on genetic algorithms

Abstract

A cost-effective method for the recursive bisection of two-dimensional unstructured grids into 2 n subdomains is introduced. The method is based on Genetic Algorithms (GAs) and is capable of generating evenly loaded disjoint mesh subsets with small interface length, that can be efficiently processed in parallel, on distributed memory platforms. Non-dimensional quotients, that express the load-balance and minimum communication requirements, are combined in a simple cost function controlled by the GA. The genetic operators are coupled with a single-pass multilevel scheme that allows for reduced size chromosomes to be processed. Upon convergence at any level, the GA continues to operate at the next finer grid using the same population of the uncoarsened current chromosomes. At the lowest level, a special refinement task, that permits local modifications only to the available partitions, is carried out. Crossover operators are standard, while a customised mutation operator, that takes into account the graph connectivity, is proposed. After completion of the mutation, a correction task enhances the convergence properties of the proposed method.