High Throughput Networks for Petaflops Computing

Abstract

The smallest networks that can connect eight thousand processing elements and memory interfaces in a petaflops cryocomputer contain hundreds of thousands of 2/spl times/2 switching nodes. We have determined circuit costs, maximal throughput and average latency for feasible multistage banyan and multidimensional pruned ring mesh networks. Each can deliver 20000 single-word packets every 30 picoseconds, more than eight million gigabytes per second. Switching delays one-way through each network total 1 to 2 nanoseconds. Banyans have 2/3 the switching delays of the smallest meshes. However, banyan signal propagation delays are larger. The only candidate network needing less than 100 square meters in four connection layers is a pruned mesh of shape 18/spl times/18/spl times/55/spl times/55 with nearly one million nodes. The smallest banyan has one quarter as many nodes, but needs nearly twice the wiring area.