Hierarchical macro-dataflow computation scheme

This paper proposes a hierarchical macro-dataflow computation scheme in a Fortran program. In addition to the usually used parallelism among loop iterations, the hierarchical macro-dataflow computation scheme allows us to hierarchically exploit the near fine grain parallelism among statements inside basic block, and the coarse grain parallelism among basic blocks, loops, or subroutines inside a main routine, subroutines, and loops. This paper describes the hierarchical definition of macrotasks, the parallelism extraction scheme among macrotasks defined inside a higher level layer macrotask and a scheduling scheme that assigns hierarchical macrotasks on hierarchical processor clusters. A result is also described of the performance of hierarchical macro-dataflow computation evaluated on a distributed shared memory and centralized memory multiprocessor system OSCAR.