A Cluster Testbed for SCI-based Parallel Processing

* Dr. Rosen is with the Naval Air Warfare Center, Aircraft Division, Warminster, PA Abstract A critical factor in the design of any parallel processing computer system is the interconnection network between processors. Latency, throughput and other network factors weigh heavily on the performance of any parallel program. In this paper, the average latency and sustained throughput for a ten-node, SPARC-based SCI cluster is tested and measured based on variations in message size, cluster or ring size, and transfer method. These results are compared with corresponding TCP/IP-based Ethernet values in order to help gauge the potential of COTS SCI clusters for supporting parallel processing environments and applications.