Game-based Congestion-aware Adaptive Routing (GCAR) for Proactive Thermal-aware 3D Network-on-Chip Systems

Abstract

Because of the stacking dies and heterogeneous thermal conduction, the three dimensional Network-on-Chip (3D NoC) suffers from more serious thermal problem. The thermal issue limits the performance gain of 3D integration and results in lower reliability of the system. To ensure the thermal safety, the Proactive Dynamic Thermal Management (PDTM) is proven as an efficient way to mitigate the performance impact during the temperature control period. The PDTM involves the different clock frequency assignment to each NoC node based on the information of temperature prediction results. Therefore, the PDTM can early control the system temperature before the system temperature achieves the thermal emergency, which helps to mitigate the performance impact. However, due to the different clock frequency assignment, the heterogeneous packet processing speed will lead to unbalanced traffic distribution and significant performance degradation. Although many congestion-aware adaptive routing algorithms can detour the packets away from the traffic congested regions, these approaches only adopt the information of buffer congestion status to deliver the packets, which makes the packet transmission become static and results in large thermal stress in the traffic non-congested routing regions. To increase the routing path diversity during the temperature control period, we apply the Game Theory to propose a Game-based Congestion-aware Adaptive Routing (GCAR) in this paper. The GCAR will distribute the packet delivery based on the Nash Equilibrium property in Game Theory. The experimental results show that the proposed GCAR will improve 35% -- 66% system performance compared with the previous related works.