Clustering and message distance trade-offs in torus-based Networks-on-Chip

Low-dimensional k-ary n-cubes were the most popular topologies in parallel computers, and they are still gaining a great interest in the context of on-chip networks. Clustered networks have been suggested to further reduce the diameter of low-dimensional k-ary n-cubes such as mesh and torus. This paper studies the performance merits of clustering with torus topology in terms of average message latency, throughput, and power consumption. Results of this study reveal that clustering increases implementation costs of the torus, the power consumption of the clustered-torus router is higher than that of the torus router under all traffic conditions. In terms of message latency and throughput, clustering degrades the performance of torus under uniform random traffic pattern. The torus can benefit from clustering only under high degree of communication locality.