Thermal sensor allocation and full-system temperature characterization for thermal-aware mesh-based NoC system by using compressive sensing technique

Abstract

As the complexity of multicore system grows with respect to the technology development, the large power density of multicore systems makes the thermal problems become severer. To prevent the multicore systems from overheating, in the practical way, some thermal sensors, sensing temperature at few selected locations, are embedded in the system. Then, the sensing temperatures are used to perform dynamic thermal managements (DTMs). However, searching the appropriate locations for those number limited thermal sensors is an NP-hard problem, which is time-consuming to find the optimal sensor allocations. In this paper, we demonstrate that the on-chip thermal gradients lead to sparse signals in the frequency domain. Based on this sparse-signal characteristic, we first apply the compressive sensing theory to find the thermal sensor allocations. Compressive sensing technique aims to recover the original signal from the fewer samples than the required by the Nyquist theorem. By using low-complexity random thermal sensor allocation method and Stagewise Orthogonal Matching Pursuit (StOMP)-based full-system temperature characterization, we can reduce 49% average error of full-system temperature characterization. Besides, under a constraint of full-system temperature characterization error, the proposed method can reduce 53% average number of adopted thermal sensors.