Software support for distributed and parallel computing

Abstract

This dissertation addresses creating portable and efficient parallel programs for scientific computing. Both of these aspects are important. Portability means the program can execute on any parallel machine. Efficiency means there is little or no penalty for using our solution instead of hand-coded, architecture-specific programs. Although parallel programming is necessarily more difficult than sequential programming, it is currently more complicated than it has to be. The Filaments package provides fine-grain parallelism and a shared memory programming model. It can be viewed as a "least common denominator" for parallel scientific computing. Fine-grain parallelism supports any number (even thousands) of threads, and shared memory provides a natural programming model. Consequently, the combination allows the programmer to concentrate on the application and not the architecture of the target machine. The Filaments package makes extensive use of run-time decision making. Run-time decision making has several advantages. First, it is often possible to make a better decision because more information is available at run time. Second, run-time decision making can obviate the need for complex, often intractable, static analysis. Moreover, run-time decision making leads to much of the package's efficiency.