Machine learning-based dynamic reconfiguration algorithm for reconfigurable NoCs

Abstract

Hybrid optical-electro Network-on-Chip (HOE_NoC) is a disruptive technology that can provide high bandwidth and low latency for global communication. However, optical links suffers with a problem of large static power consumption in network. For different applications, traffic distribution in space and time may differ largely. Therefore, it is necessary to dynamically provide optical link bandwidth to network for higher power efficiency under all traffic distribution. In this paper, we propose a machine learning-based dynamic reconfiguration algorithm for reconfigurable NoCs (RHOE_NoC) to reduce the static power. With machine learning prediction technique, we reconfigure the optical nodes dynamically to adapt different traffic demands while maintaining higher performance. Experimental results shown that as compared to electronic network latency has been reduced by 51%, while throughput has been improved by 14% for 64 node network architecture and energy consumption has been reduced by 26%. We have also compared RHOE_NoC with HOE_NoC without reconfiguration, results show that static energy consumption has been reduced by about 28%.