"This lecture covers an interesting technique for synchronizing parallel processors. In all but one architectural structure, synchronization forces serial execution. The only exception proposed to date is a structure called "a combining network," that permits a functional serialization to occur without actually forcing processors to serialize their actual execution. The lecture examines the problem, illustrates the combining network, and then demonstrates how this network in a parallel computer can support an extremely efficient parallel imple-mentation of Quicksort."
University Video Communications: Distinguished Lectures