Optimizing Modern On-Chip Networks Through Circuit-Switching

Current on-chip network designs utilize virtual channel flow control to realize better utilization of the link bandwidth through buffering and packet switching.However,VC flow control also has its shortcomings,that is,it is not only the main source of energy footprint but also increases the per-hop latency of the network.For parallel workloads,bandwidth of the network that has virtual channels is relatively sufficient,so the seemingly-outdated but extremely simple flow control mechanism,circuit switching,can be utilized when the bandwidth demand is not high,in order to optimize the energy efficiency of the network.This thesis proposes a hybrid switching flow control mechanism,which mixes both virtual channels and circuit switching,to provide a latency-competitive and energy-efficient network design.The main purpose of this proposal is to speed up the traffic forwarding process by allowing traffic in the network to be forwarded through the circuit-switched subnet as much as possible.This flow control mechanism under hybrid switching comes with two circuit set-up schemes,which are based on either global or local traffic profile.Configurations of the network under hybrid switching is designed from two perspectives: one is from the methods of circuit construction,including end-to-end circuits and route-to-route circuits,while the other is from the perspective of the number of subnets,including two,four and eight subnets.In the evaluation,eight patterns of synthetic traffic are used to obtain network performance under high injection load,while the performance under different applications are evaluated using the PARSEC benchmark suite.When compared to the on-chip network based on virtual channels,the average latency of the network traffic is up to 10% shorter while the energy footprint can be cut for up to 24% with the global traffic-based circuit set-up scheme.On the other hand,when the local traffic-based circuit set-up scheme is adopted,the latency of the network traffic can be reduced by up to 5% and its energy consumption can be up to 23% smaller.