Limits to parallelism in scientific computing

The goal of our research is to decrease the execution time of scientific computing applications. We exploit the application\u27s inherent parallelism to achieve this goal. This exploitation is expensive as we analyze sequential applications and port them to parallel computers. Many scientifically computational problems appear to have considerable exploitable parallelism; however, upon implementing a parallel solution on a parallel computer, limits to the parallelism are encountered. Unfortunately, many of these limits are characteristic of a specific parallel computer. This thesis explores these limits.;We study the feasibility of exploiting the inherent parallelism of four NASA scientific computing applications. We use simple models to predict each application\u27s degree of parallelism at several levels of granularity. From this analysis, we conclude that it is infeasible to exploit the inherent parallelism of two of the four applications. The interprocessor communication of one application is too expensive relative to its computation cost. The input and output costs of the other application are too expensive relative to its computation cost. We exploit the parallelism of the remaining two applications and measure their performance on an Intel iPSC/2 parallel computer. We parallelize an Optimal Control Boundary Value Problem. This guidance control problem determines an optimal trajectory of a boat in a river. We parallelize the Carbon Dioxide Slicing technique which is a macrophysical cloud property retrieval algorithm. This technique computes the height at the top of a cloud using cloud imager measurements. We consider the feasibility of exploiting its massive parallelism on a MasPar MP-2 parallel computer. We conclude that many limits to parallelism are surmountable while other limits are inescapable.;From these limits, we elucidate some fundamental issues that must be considered when porting similar problems to yet-to-be designed computers. We conclude that the technological improvements to reduce the isolation of computational units frees a programmer from many of the programmer\u27s current concerns about the granularity of the work. We also conclude that the technological improvements to relax the regimented guidance of the computational units allows a programmer to exploit the inherent heterogeneous parallelism of many applications