Dynamic load balancing is an essential tool for parallel adaptive computation. It has been an important research topic for more than a decade. The Zoltan library provides applications with a reusable, object-oriented interface to several load balancing algorithms, including coordinate bisection, octree/space filling curve methods, and multilevel graph partitioners. It allows run-time selection among balancing techniques, facilitating comparisons. It also provides a framework in which to modify existing algorithms or to implement new algorithms. We compare the performance of several available methods as mesh partitioners and as dynamic load balancers for an adaptive computation. Partition quality metrics and running times of the balancers and the solution process are measured for several methods. We also examine the effect of octant granularity for the Zoltan octree partitioning algorithm, and describe a modification of the recursive coordinate bisection algorithm that bisects along only one coordinate axis, to achieve a "slice" partitioning, designed to minimize interprocessor adjacency.
|Citation (BIBTEX)||Paper (PS; 1.1MB)||Paper (PDF; 268KB)|