Techniques for overheating detection and sensor allocation in a real dual-core processor

Current processor families widely deploy on-chip thermal sensors to continuously monitor the real-time thermal behavior. However, on-chip thermal sensors are inevitably accompanied by a variety of noise sources such as fabrication randomness and environmental uncertainty, which directly impact the reliability of dynamic thermal management (DTM). In this paper, the problems of sensor allocation for overheating detection are formulated as constrained optimization problems, when the sensor observations have been corrupted by noise. Moreover, a lightweight sensor allocation scheme (called LSAS) based on the custom-built genetic algorithm is proposed to significantly improve the overheating detection performance with an approximate linear execution time. Based on the LSAS and greedy optimization techniques, a hybrid algorithm for local overheating detection is also proposed to identify the optimal sensor allocation for each individual processor block. Meanwhile, an infrared temperature measurement setup is developed to capture the thermal traces of a 45 nm dual-core AMD Athlon X2 5000 processor. The obtained realistic temperature data are used to verify the performance. Experimental results show that the LSAS can achieve the overheating detection probability by up to 0.93 with an overhead of ten sensors.