Submiited for Publication to Ieee Transactions on Parallel and Distributed Systems. Ultrafast Deeection Networks for Multiprocessors

Interconnection networks for shared memory multiprocessors are generally designed to have a network link bandwidth that is comparable to their insertion bandwidth requirement. In contrast, ultrafast networks possess a very high link bandwidth between nodes compared to the insertion bandwidth sought by the hosts. They are usually direct networks so a link to a host carries both its traac and through traac. In this paper we show that such ultrafast networks with a deeection routing protocol can easily match the processor-memory communication requirements of a multiprocessor. Deeection routing simpliies processing at routing nodes, and allows the network to adapt itself eeciently to non-uniformities in traac. We show that the system adjusts well to temporal variations in traac provided the average processor insertion bandwidth is less than what the network can provide. Our study of hot-spots shows that memory bandwidth constitutes the limiting bottleneck because deeections route non-hot spot traac away from congested nodes. Our model system is fully pipelined. It utilizes modiied deeection routing nodes to reduce deeections by temporal reordering of packets in ight, and to thereby allow a given link bandwidth to support a higher insertion rate.