Bufferless and Minimally-Buffered Deflection Routing

A conventional Network-on-Chip (NoC) router uses input buffers to store in-flight packets. These buffers improve performance, but consume significant power. It is possible to bypass these buffers when they are empty, reducing dynamic power, but static buffer power remains, and when buffers are utilized, dynamic buffer power remains as well. To improve energy efficiency, bufferless deflection routing removes input buffers, and instead uses deflection (misrouting) to resolve contention. Bufferless deflection routing is able to provide similar network performance to conventional buffered routing when the network carries light to moderate traffic, because deflections are relatively rare. However, at high network load, deflections cause unnecessary network hops, wasting power and reducing performance. In order to avoid some deflections and recover some performance, recent work has proposed to add a small buffer which holds only flits that contend with others and would have been deflected. This minimally-buffered deflection (MinBD) router improves performance relative to bufferless deflection routing without incurring the cost of a large buffer, because it can make more efficient use of a small buffer. The result is a router design which is more energy-efficient than prior buffered, bufferless, and hybrid router designs. Chris Fallin, Greg Nazario, Kevin Chang, Rachata Ausavarungnirun, and Onur Mutlu Carnegie Mellon University {cfallin,gnazario,kevincha,rausavar,onur}@cmu.edu