Hypergraph partitioning satisfying dual constraints on vertex and edge weight

Abstract

In this paper we present a multilevel hypergraph partitioning algorithm which satisfies not only vertex weight constraint but also edge weight constraint. In our multilevel paradigm, two FM [1] variants are performed alternately in the refinement stage. The first FM variant aims to symmetrize edge weight of the two partitioning blocks and the other aims to minimize cut size under dual constraints on both edge and vertex weight. Inheritance of edge weights by vertices is considered in the coarsening stage. Experimental results on ISPD-98 [2] show that our algorithm satisfies both 10% and 2% dual constraints, producing solutions, under 10% dual constraints, with much more balanced edge weight than hMetis [3], MLPart [4], and FM in terms of imbalance proportion by 837.2%, 1039.1% and 762.9%, and reasonable cut size increment by 29.0%, 28.4% and −42.3%, run time increment by 109.2%, 61.8% and 43.7%, respectively.