This talk presents elastic buffers (EBs), an efficient flow-control scheme that uses the storage already present in pipelined channels in place of explicit input virtual-channel buffers (VCBs). With this approach, the channels themselves act as distributed FIFO buffers under congestion. Without VCBs, and hence virtual channels (VCs), deadlock prevention is achieved by duplicating physical channels. We develop a channel occupancy detector to apply universal globally adaptive load-balancing (UGAL) routing to load balance traffic in networks using EBs. Using EBs results in up to 12% (17% for low-swing channels) improvement in peak throughput per unit power compared to a network using VC flow control. The power saved by eliminating VCBs is used to make channels wider, providing increased throughput that more than offsets the degradation caused by increased blocking without VCs. EB networks have identical zero-load latency to VC networks and have considerably simpler router logic as a VC allocator is not required.
George Michelogiannakis is a PhD student at the Electrical Engineering department of Stanford University. His current research is on NoC flow-control and NoCs for large-scale CMPs.