Parallel primitives as tools for implementing parallel algorithms: Definition, design and implementation

This dissertation researches the general area of parallel computers and, in particular, how algorithms are implemented on these parallel systems. The basic intent of the research is to suggest better methods and procedures for implementing large problems onto parallel computer hardware configurations. Specifically, the concept of a parallel primitive tool is proposed, designed and implemented. Three research steps are presented: (1) A study of how parallel algorithms are currently being implemented; (2) The definition of parallel primitives as new tools for implementation; and (3) The design and implementation of example parallel primitives, using content-addressable memory. A matrix inversion problem is used as an implementation model, using a parallel primitive for finding a maximum value. The matrix inversion is benchmarked as serial, parallel, and simulated parallel primitive operations in both high-level and assembler languages. The final part of research is the design and analysis of a more complex parallel primitive, which also uses content addressable memory, to add a series of floating point numbers as a single operation. Timings are developed for the summing operation and compared with other parallel configurations. Parallel primitives are shown to offer some improvement in implementing parallel algorithms by simplifying program coding and reducing program execution times. And content addressable memory is shown to be effective for the design of parallel primitives.