Parallel processing

Abstract

Parallel algorithms, based on simulated annealing, neural networks and genetic algorithms, for mapping irregular data to multicomputers are presented and compared. The three algorithms deviate from the sequential versions in order to achieve acceptable speed-ups. The parallel annealing and neural algorithms include communication schemes adapted to the properties of the mapping problem and of the algorithms themselves. These schemes arc found useful for providing both good solutions and reasonable execution times. The parallel genetic algorithm is based on a model of natural evolution. The three algorithms preserve the high quality solutions and the non-bias properties of their sequential counterparts. Further, the comparison results show their suitability for different requirements of mapping time and quality.