Effective data distribution and parallelization of computations involving irregular data structures is a challenging task. We tackle the twin-problems in the context of computations involving block-sparse matrices. The programming model provides a global view of a distributed blocksparse matrix. Mechanisms are provided for the user to express the parallel tasks in the computation. Locality-aware load-balancing of the tasks ensures good scalability. The primitives are based on the Aggregate Remote Memory Copy Interface, and are inter-operable with the Global Arrays programming suite and MPI. Results are presented that demonstrate the utility of this approach.
Revised: April 19, 2006 |
Published: December 31, 2005
Citation
Krishnamoorthy S., J. Nieplocha, and P. Sadayappan. 2005.Data and Computation Abstractions for Dynamic and Irregular Computations. In Proceedings of the High Performance Computing-HiPC 2005. 12th International Conference. Published in Lecture Notes in Computer Science, 3769, 258-269. Berlin:Springer Verlag.PNNL-SA-46884.