Automatic mapping and load balancing of pointer-based dynamic data structures on distributed memory machines

Describes an algorithm for automatically mapping and load balancing unstructured, dynamic data structures on distributed memory machines. The algorithm is intended to be embedded in a compiler for a parallel language (DYNO) for programming unstructured numerical computations. The result is that the mapping and load balancing are transparent to the programmer. The algorithm iterates over two basic steps: (1) It identifies groups of nodes ('pieces') that disproportionately contribute to the number of off-processor edges of the data structure and moves them to processors to which they are better connected. (2) It balances the loads by identifying groups of nodes ('flows') that can moved to adjacent processors without creating new pieces. The initial results are promising, giving good load balancing and a reasonably low number of inter-processor edges.<>