Cluster Based Adaptive Multi-Voltage Scaling Dynamic Task Mapping for WNoC and HWNoC

Abstract Wireless Network-on-Chip (WNoC) and Hybrid Wireless Network-on-Chip (HWNoC) architectures are promising solutions for future high-performance computing systems. However, WNoC consumes significant power, while HWNoC experiences congestion over the wireless link. Several state-of-the-art task mapping algorithms have been proposed to reduce power consumption and congestion over wireless links. However, these existing task mapping algorithms face challenges related to hotspots creation, sub-optimal utilization of wireless links, and also overlook idle core power reduction strategy. Additionally, each of the existing task mapping algorithms is designed for a specific architecture, either WNoC or HWNoC. To address these challenges we propose a novel task mapping algorithm called Cluster-Based Adaptive Multi-Voltage Scaling (CB-AMS). This algorithm dynamically maps tasks to clusters while performing multi-voltage scaling based on workload to significantly reduce power consumption and congestion over wireless links. A new cluster selection strategy is also proposed in CB-AMS to address the hotspot creation issue. CB-AMS is designed to be used in both WNoC and HWNoC architecture. Experimental results show that CB-AMS significantly reduces power consumption by 41% for WNoC and by 15-20% for HWNoC compared to state-of-the-art task mapping algorithms. Experimental results also validate that CB-AMS achieves better congestion control in HWNoC architecture by reducing latency by 3.6-5.5% compared to existing task mapping algorithms. Our experimental analysis has demonstrated that CB-AMS outperforms the current algorithms and delivers significant power reduction and improved congestion control for both WNoC and HWNoC architectures.