A network based model for heterogeneous parallel computation

Abstract

The computational requirements of science and engineering demand computational resources orders of magnitude of the current day sequential machines. Most of the research effort has been concentrated upon the creation of parallel algorithms and parallel axchitectures, but not enough research has been done on the software environment and parallel program portability. Unfortunately, no architecture can map all programs optimally and most of the programs perform poorly. Heterogeneous network computing offers a viable and cost-efficient alternative to actually constructing a heterogeneous element machine. We attempt to extract machine level parallelism in a given computational problem by selecting parts of a program that could be executed in parallel, and then selecting a machine that will be the most appropriate to execute that part of the program. In effect we could have n machines executing the same program but simultaneously working on a particular part that maps the best on that machine. Alternatively, different programs could execute on the three machines assuming there are minimal conflicts, that is, the jobs are well balanced. In this project we have developed a network based heterogeneous program execution environment. We adopted a master-slave approach to free the user from the rigors of network programming. At the heart of the system is a the master, a workstation, which controls the scheduling and execution of jobs on the different slaves. We selected three different machines each of them a supercomputer in its own different machines each of them a supercomputer in its own right, each of them belonging to a different classification of parallel machines. The machines we selected were, a hypercube(MIMD), an array processor(SIMD), and a vector processor(MISD). An existing physical problem was implemented to benchmark this approach.