Evolutionary design of group communication schedules for interconnection networks

Abstract

As chip multiprocessors are quickly penetrating new application areas in network and media processing, their interconnection architectures become a subject of optimization. Group communications are frequently used in many parallel algorithms and if their overhead is excessive, performance degrades rapidly with a processor count. This paper deals with the design of a new application-specific Bayesian Optimization Algorithm (BOA) and the use of BOA and HSGA (Hybrid parallel Genetic Simulated Annealing) to design optimal communication algorithms for an arbitrary topology of the interconnection network. Each of these algorithms is targeted for a different switching technique. The group communication schedules were designed for an asymmetrical AMP network on one hand and for the benchmark hypercube network on the other, using SF (Store-and-Forward) and WH (Wormhole) switching. © Springer-Verlag Berlin Heidelberg 2005.