A decoupled multicomputer architecture with optical full interconnection

Abstract

Decoupled architectures have been proposed in sequential processing to increase performance when memory access delays are a bottle-neck. Communication delays are expected to be the major bottle-neck in future large-scale parallel computers. Although optical technology is often used to improve the throughput of each connection, such communication systems still have significant data transfer delays. We believe that providing more parallel connections, by increasing the density of topologies, is a more effective approach to reduce the communication bottle-neck. We are currently evaluating the first prototype of a new decoupled multicomputer architecture supporting static allocation and static scheduling of data processing tasks. The decoupling of communication and computation in this prototype not only supports efficient allocation of data processing resources, but at the same time also allows efficient scheduling of data transfers on the interconnection system to minimize the effect of data transfer delays.