Multi-criteria Graph Partitioning with Scotch
Résumé
Load balancing parallel multi-physics simulations is a hard task often performed by solving a multi-criteria partitioning problem. The aim of this paper is to describe how this problem is solved in Scotch, explaining the various algorithmic choices performed. We also present a method to generate multi-criteria weight distributions for meshes corresponding to those obtained by Monte-Carlo particle transport simulations. This method is used on 5 meshes that serve to compare multi-criteria partitioning tools. A mesh corresponding to an industrial test case is also considered. In order to compare multi-criteria partitioning tools, we analyze their performance profiles. Results show that Scotch returns solutions of smaller edgecut than other partitioning tools such as MeTiS and PaToH, especially for the industrial test case.