François Pellegrini

Applying Graph Partitioning Methods in Measurement-based Dynamic Load Balancing

Load imbalance in an application can lead to degradation of performance and a significant drop in system utilization. Achieving the best parallel efficiency for a program requires optimal load balancing which is an NP-hard problem. This paper explores the use of graph partitioning algorithms, traditionally used for partitioning physical domains/meshes, for measurement-based dynamic load balancing of parallel applications. In particular, we present repartitioning methods that consider the previous mapping to minimize dynamic migration costs. We also discuss a new imbalance reduction algorithm for graphs with heavily skewed load distributions. These algorithms are implemented in a graph partitioning toolbox called SCOTCH and we use CHARM++, a migratable objects based programming model, to experiment with various load balancing scenarios. To compare with different load balancing strategies based on graph partitioners, we have implemented METIS and ZOLTAN-based load balancers in CHARM++. We demonstrate the effectiveness of the new algorithms developed in SCOTCH in the context of the NAS BT solver and two micro-benchmarks. We show that SCOTCH based strategies lead to better performance compared to other existing partitioners, both in terms of the application execution time and fewer number of objects migrated.