Serial and parallel multilevel graph partitioning using fixed centers

Abstract

We present new serial and parallel algorithms for multilevel graph partitioning. Our algorithm has coarsening, partitioning and uncoarsening phases like other multilevel partitioning methods. However, we choose fixed nodes which are at least a specified distance away from each other and coarsen them with their neighbor nodes in the coarsening phase using various heuristics. Using this algorithm, it is possible to obtain theoretically and experimentally much more balanced partitions with substantially decreased total edge costs between the partitions than other algorithms. We also developed a parallel method for the fixed centered partitioning algorithm. It is shown that parallel fixed centered partitioning obtains significant speedups compared to the serial case.