Network partitioning of data parallel computations

Partitioning data parallel computations across a network of heterogeneous workstations is a difficult problem for the user. We have developed a runtime partitioning method for choosing the number and type of processors to apply to a data parallel computation, and a decomposition of the data domain in order to achieve reduced completion time. The partitioning method utilizes information about the problem in the form of callback functions and uses a set of topology-specific communication functions to estimate communication costs. We show that the method makes effective partitioning decisions and has runtime overhead that is easily tolerated. In particular we show that for two implementations of a canonical stencil application, minimum elapsed times are obtained for a range of problem sizes on a network of heterogeneous workstations