Data Partitioning for Scientific Applications and Emerging Architectures.

Mesh adjacencies represent application data more completely than standard graphpartitioning models. • All mesh entities can be considered, while graph-partitioning models use only a subset of mesh adjacency information. • Any adjacency can be obtained in O(1) time (assuming use of a complete mesh adjacency structure). • Partition model supports efficient part adjacency queries. (See FASTMath’s Unstructured Mesh poster for more details.) Partition Improvement with Multiple Criteria • Diffusive procedure driven by application-defined priority list of mesh entity types to balance. Iteration over three stages: • Schedule load transfer from heavy parts to light parts • Select mesh elements on the part boundaries that will smooth the part boundary • Migrate selected mesh elements according to schedule