Structured parallel computation

Main repercussions of the intertwinement of architecture, programming and algorithmics on the design of parallel computers are discussed, where for detailed solutions to practical problems is referred to the Delft Parallel Processor. Measures for task interactivity and for processor interconnectivity are defined in dependence of parameters of the system architecture of a parallel computer. It is sketched that speedup enlargement with preservation of compute-bound processing is possible by applying multi-level parallel computing. It is shown that a parallel computer fitted out with a `full interconnectivity´ interconnect and a dynamic processor clustering facility allows the application of the multi-newspaper protocol as an efficient and easy-to-program way to avoid transfer-bound processing. It is explained how opto-electrical full interconnectivity and dynamic processor clustering is realized in the Delft Parallel Processor. The impact of parallel algorithms on the structured design of parallel computers is demonstrated through a review of the design of some parallel algorithms for iterative and direct solving of sets of equations