Minimizing development overhead with partial parallelization

To demonstrate the evolutionary path for porting serial applications to the T3D as advocated by Cray Research, Inc., the authors implemented a heterogeneous, distributed, adaptive finite-element code on Cray´s Y-MP/T3D system. They implemented a serial adaptive finite-element code on the Y-MP and, rather than fully distributing the program on the T3D, modified only the computationally intensive portion for a massively parallel processing environment. A set of two communicating programs resulted in a Y-MP program that addresses mesh maintenance and modification, and a T3D program that distributes the data and computations for a given mesh. The speedups achieved over a serial code demonstrate that significant performance gains can result from focusing the parallelization of serial code on smaller components, rather than attempting to derive a completely parallel implementation. Thus, a satisfactory trade-off between implementation effort and optimal performance resulted. Although implemented in a Y-MP/T3D environment, the approach is generic and would work in numerous heterogeneous computing environments