A Performance Study of Di usive vs . Remapped Load - BalancingSchemes

For a large class of irregular grid applications, the computational structure of the problem changes in an incremental fashion from one phase of the computation to another. Eventually, as the graph evolves, it becomes necessary to correct the partition in accordance with the structural changes in the computation. Two diierent types of schemes to accomplish this task have been developed recently. In one scheme, the graph is partitioned from scratch and then the resulting partition is remapped intelligently to the original partition. The second type of scheme uses a multi-level diiusion repartitioner. In this paper, we conduct a comparison study on repartitioning via intelligent remapping versus repartitioning by diiusion. We show that multilevel diiusion algorithms generally produce signiicantly lower data migration overhead for adap-tive graphs in which low magnitude localized or non-localized imbalances occur and for graphs in which high magnitude imbalances occurs globally throughout the domains than partitioning from scratch and remapping the resulting partition. We show that for the class of problems in which high magnitude imbalances occur in localized areas of the graph, partitioning from scratch and remapping the resulting partition will result in higher quality partitions than diffusion schemes, while obtaining data migration overheads which are comparable to those obtained by dif-fusive schemes. Finally, we show that the run times of the various schemes are all similar.