Correlation-graph-based temperature sensor allocation for thermal-aware network-on-chip systems

Abstract

The increasing power density and time-varying workload distribution lead to runtime thermal hotspot in the modern multicore systems. To solve the thermal problems of the multicore systems, many dynamic thermal managements (DTMs) were proposed in recent years. These DTMs rely on the temperature sensing accuracy of the on-chip thermal sensors. However, due to the physical or economical constrains, the number of on-chip sensors is often limited, and the allocation of these sensors significantly affects the accuracy of full-chip temperature estimation. Hence, the main challenge is to determine the appropriate locations for thermal sensor placement. However, searching the proper locations for thermal sensor placement is an NP-hard problem. Besides, the time-varying temperature distribution of the multicore system further worsens the hardness to find the optimal sensor allocations. In this paper, we analyze the temperature correlation graph to propose a thermal sensor allocation algorithm. Compared with the conventional thermal sensor allocation methods, the proposed method can reduce 94% average error and 84% maximum error of full-chip temperature estimation.