Parallel Sorting in a Ring Network of Processors

A parallel implementation of selection sorting algorithms is presented that uses a ring-connected array of processors, in which each processor has the same amount of memory. The scheme allows all processors to have the same amount of memory with a small fragmentation loss. Uniformity in the size of the memories provides advantages from the viewpoint of manufacture, maintenance, inventory, item placement, and many other aspects of system design. These memories can, together, function as an interleaved memory for a general-purpose computer. Hence, it is possible to build a medium-scale system in which parallel-sort is implemented as a primitive, like block-move, without a large increment in hardware complexity and cost. An analysis is presented that shows how to improve performance by overlapping comparisons with memory accesses and how to achieve performance improvement by applying carry-look-ahead techniques to comparison. >