Network-Distributed Accelerators

A computer architecture comprised of a cluster of small, programmable, computing nodes connected by and to a high-speed network is attractive for many defense and commercial applications. Unlike accelerators that attach to a host computer's memory bus, network-based accelerators can be shared by all of the hosts on the network, and can provide continued system operation when nodes or network links have failed. The fine-grain clusters provide very high performance per unit cost, weight, and power; the modularity allows a short system-development cycle; and the programmability allows the system to be modified to meet changing requirements. Myricom has adapted the packet-communication and routing technologies developed originally for scalable, concurrent computers to allow arbitrary-topology networks, long-distance links, and fault-tolerant operation. We propose to examine whether this technology can be applied to distributed computing between conventional computers and clusters of small nodes specialized for tasks such as video storage and retrieval, and signal and image processing. The results of this study will be preliminary hardware designs, a prototype programming system, and an evaluation of the suitability of these systems for multimedia-server and raster-image-processing applications. Anticipated Benefits: Raster-image processing for publishing and printing-on-demand applications, medical-image processing, and multimedia servers are commercial applications for which Myricom seeks to develop products and markets.