Intelligent congestion control for NoC architecture in Gem5 simulator

Abstract

Congestion in a network significantly impacts the performance of an NoC as there is a substantial increase in latency and power consumption. Machine Learning techniques aid in designing routing methods to keep the network cognizant of the traffic status. This paper presents a congestion-aware Q-routing algorithm based on the Q-learning model of reinforcement learning. The proposed algorithm enhances the network's performance in an NoC under heavy traffic conditions by routing the packets along a less congested path. Thus, it reduces the congestion in the network. This is possible as Q-learning allows the network to keep track of the local and non-local congestion by estimating Q-values. The Q-values guide a node in sending a data packet along an optimal path, thereby evading busy routes. The simulation done on the gem5 simulator with uniform link latency in the network exhibits that Q-routing performs better in a high-load environment than traditional XY and Odd-Even Routing methods, with a performance gain of 5.73% and 12.73%, respectively. The results for varied link latencies that were randomly assigned to create a practical congestion-probable scenario showed that the proposed method outperformed both the XY and Odd-Even routing algorithm with a respective performance gain of 7.38% and 15.19%.