A COC Oriented Parallel Computing Strategy for Irregular Applications

Sparse and unstructured computations are widely involved in scientific and engineering applications. It means that data arrays could be indexed indirectly through the values of other arrays or non-affine subscripts. Data access pattern would not be known until runtime. So far all the parallel computing strategies for this kind of irregular problem are single network topology oriented, which cannot fully exploit the advantages of modern hierarchical computing architecture, like grid. We proposed a hybrid parallel computing strategy RP, shorted for "Replicated and Partially-shared", to improve the performance of irregular applications in the COC (Cluster of Clusters) environment. A detailed comparison is made between our strategy and other traditional models, along with experimental results demonstrating its effectiveness. A class of practical irregular applications employed RP strategy could obtain much shorter execution time and better scalability in heterogeneous network based computation environment.