Abstract

The performance of a large parallel processing system can become limited by the serial nature of synchronization.¹ For example, divide-and-conquer parallel search algorithms partition work into separate processes that must all halt when a solution is found. This process requires synchronization of all the participating processors. However, when each processor has only one input and output port, such operations can only proceed serially; this represents a bottleneck in the performance. Previously, we have designed an optoelectronic multistage interconnection network for communication and parallel computing.^2,3 In this paper, we extend this design to add distributed processing capability within the network.

© 1991 Optical Society of America