Fast arbitration in dilated routers

Abstract

We present a technique, dynamic dilation, which reduces latency in dilated routing components without greatly affecting flexibility. Initially in order to achieve lower latency we limit dilation flexibility, extract parallelism from the arbitration process and pipeline the process into two cycles [ALM90]. For implementation reasons the allocation cycle is required to have some unused latency. Dynamic dilation takes advantage of this latency to allow the dilation of a component to be reconfigured each cycle based on incoming messages. This effectively recovers the flexibility lost by limiting dilation, while retaining extremely low latency.