An Enhanced Dynamic Weighted Incremental Technique for QoS Support in NoC

Abstract

Providing Quality-of-Service (QoS) in many-core network-on-chip (NoC) platforms is critical due to the high level of resource sharing in such systems. This article presents a hard-built Equality-of-Service (EoS) and Differential-Service (DS) as subsets of QoS in NoC using weighted round-robin arbitration policy. In the proposed technique, packets can be injected with variable initial weights. An enhanced dynamic weight incremental technique is proposed that automatically increases the weights according to the contention degree that packets face along their paths. The proposed technique provides EoS for all packets that are injected with equal initial weights. Furthermore, the router’s input port bandwidth is shared among the passing flows according to the portion of packets’ initial weights in the presence of contention (saturation). This provides DS in the network. The area and timing overhead of the proposed technique is insignificant (<%3). However, it achieves a remarkable improvement in the performance against baseline routers. The simulation using synthetic traffic patterns shows up to 67% average EoS improvement in both network throughput and worst-case delay (WCD). Two case studies are conducted for observing the DS. The first case study shows how a hotspot traffic bandwidth can be shared differently among other nodes according to the flows’ initial weights. The second case study explores the DS in an experiment running traffic patterns generated from several real applications task graphs. This case study shows when all packets of a specific application are injected with larger initial weights than the others, its WCD can be reduced up to 50% compared to the time when all applications’ packets are injected with equal initial weights.