Views of mixed-mode computing and network evaluation

Abstract

Trade-offs between the SIMD and MIMD models of architecture for parallelism are presented. Mixed-mode parallelism, where a machine can switch between the SIMD and MIMD modes of parallelism at instruction-level granularity with generally negligible overhead, is discussed. Advantages and disadvantages of mixed-mode parallelism and an example of a mixed-mode parallel algorithm are given. The relationship of mixed-mode processing to high-performance heterogeneous computing is overviewed. Difficulties involved with evaluating interconnection networks for parallel machines are then considered. There are a myriad of metrics that have been used in the literature. The problems involved with choosing the most appropriate metric or weighted set of metrics, and performing "fair" comparisons, are explored.<>