Parallel processing with a sorting network
Josef G. Krammer · 2002
Consideration is given to the parallel execution of algorithms with global and irregular data dependencies on a regular and locally connected processor array. The associated communication problems are solved by the use of a two-dimensional sorting algorithm, and a multiprocessor based on a two-dimensional sorting network is proposed. In this architecture a one-dimensional arrangement of processors performs all required control and arithmetic operations, whereas the sorter solves complex data transfer problems and utilizes its storage capability as a memory for data elements. The algorithms for sparse matrix computations, mapped onto this architecture, show that the utilization of the processors is of O(1) or only slightly less, depending on the sorting algorithm used.>