Parallel partitioning with Zoltan: Is hypergraph partitioning worth it?

Graph partitioning is an important and well studied problem in combinatorial scientific computing, and is commonly used to reduce communication in parallel computing. Different models (graph, hypergraph) and objectives (edge cut, boundary vertices) have been proposed. Hypergraph partitioning has become increasingly popular over the last decade. Its main strength is that it accurately captures communication volume, but it is slower to compute than graph partitioning. We present an empirical study of the Zoltan parallel hypergraph and graph (PHG) partitioner on graphs from the 10th DIMACS implementation challenge and some directed (nonsymmetric) graphs. We show that hypergraph partitioning is superior to graph partitioning on directed graphs (nonsymmetric matrices), where the communication volume is reduced in several cases by over an order of magnitude, but has no significant benefit on undirected graphs (symmetric matrices) using current parallel software tools.