Enhanced adaptive thermal-aware routing algorithm for network-on-chip

Abstract

Routers in 3D NoC are responsible for generating more heat than other components putting an extra strain on the chip cooling cost. Various methods have been suggested to balance temperature distribution, including thermal-aware routing. Thermal-aware adaptive routing is a viable remedy to reduce thermal hotspots by migrating load to the cooler areas of the chip, hence adopting longer and extended paths and suffering from traffic congestion in the network. Furthermore, routing algorithms fetch parameters from neighbouring nodes each time, causing an extra strain on the network. The objective of this study was to have a balance between path length and temperature, reduce the workload on thermally unstable paths and reduce control traffic overhead. This work presented an efficient thermal-aware adaptive routing. The proposed work could adaptively choose the next neighbour at each intermediate node, leading the packet closer to the destination. An effective thermalaware dynamic weighted adaptive routing was also proposed in this study. The dynamic weighted model had considered parameters related to congestion and thermal issues and provided a suitable balanced approach according to the current situation at each node. Furthermore, an interval-based record-keeping mechanism was proposed to record parameters of previously accessed nodes. Therefore, better, efficient and rapid routing decisions could be made. Results obtained from the simulations showed that the proposed routing algorithms had performed a 12-49% improvement in terms of global average delay under various synthetic traffic conditions compared to the state-of-the-art ATAR. The proposed techniques had observed 24-30% lower hop counts and considerable reduction in thermal profiling, along with up to 35-60% more valid records found from the history table compared to existing techniques. Overall, the proposed routing techniques have contributed to finding progressive routing paths and reducing control traffic overhead within the network.