Performance Analysis of Traffic Bottleneck Induced by Concentration at K-Ary N-Cube NoC Routers

Abstract

Mesh and torus are the most popular topologies suggested to implement Network-on-Chip (NoC). Concentration has been suggested to reduce networks complexity in terms of routers and wiring requirements. Several studies have examined the performance of concentrated mesh. This paper expands on these studies to re-evaluate the benefits of concentration in mesh and torus. Three alternative router configurations are proposed to investigate bottleneck performance due to concentration. Simulation results indicate that, under uniform traffic distribution, only large ratios of packet length-to-average hop count is in favor of concentrated mesh and torus. The Cmesh profits from its high channel bandwidth to offer better performance than Ctorus. The latter is unable to overcome the high packet serialization overhead even when using separate input ports at routers, one for each IP core, to reduce packet contention. Furthermore, non-local traffic at routers suffers more from contention than local traffic. Providing dedicated input ports, one for each IP core, at routers, reduces the average packet latency by 80% compared to a configuration with a single input port shared by all IP cores of the cluster, while only 20% improvement is achieved by adding a separate input port to service local traffic. However, when high traffic load is sent locally to the IP cores belonging to the same cluster as the source, a router configuration with multiple channels servicing local traffic greatly reduces the overall packet latency.