Fault-Tolerant Arbitration in Multichip Crossbar Switches

Abstract

A major problem in the design of large multichip crossbar networks is the latency associated with the centralized controller used to arbitrate requests. We present a solution to reduce this latency and also provide fault-tolerance by the use of multiple controllers that can operate concurrently. This paper concentrates on crossbars that are constructed using onesided crosspoint switches, as such designs can alleviate simultaneous switching noise. Two arbitration mechanisms are proposed that provide a range of tradeogs in controller complexity and hardware overheads. We derive a lower bound on the number of busses required for any distributed control scheme and compare it with simulation results for a 64 port network. The results show that besides reducing arbitration latencies, distributed control can provide almost non-blocking operation with little extra hardware, and with improved properties of fault-tolerance and graceful degradation.