On thermal sensor calibration and software techniques for many-core thermal management

The high power density of a many-core processor results in increased temperature which negatively impacts system reliability and performance. Dynamic thermal management applies thermal-aware techniques at run time to avoid overheating using temperature information collected from on-chip thermal sensors. Temperature sensing and thermal control schemes are two critical technologies for successfully maintaining thermal safety. In this dissertation, on-line thermal sensor calibration schemes are developed to provide accurate temperature information. Software-based dynamic thermal management techniques are proposed using calibrated thermal sensors. Due to process variation and silicon aging, on-chip thermal sensors require periodic calibration before use in DTM. However, the calibration cost for thermal sensors can be prohibitively high as the number of on-chip sensors increases. Linear models which are suitable for on-line calculation are employed to estimate temperatures at multiple sensor locations using performance counters. The estimated temperature and the actual sensor thermal profile show a very high similarity with correlation coefficient ~0.9 for SPLASH2 and SPEC2000 benchmarks. A calibration approach is proposed to combine potentially inaccurate temperature values obtained from two sources: thermal sensor readings and temperature estimations. A data fusion strategy based on Bayesian inference, which combines information from these two sources, is demonstrated. The result shows the strategy can effectively recalibrate sensor readings in response to inaccuracies caused by process variation and environmental noise. The average absolute error of the corrected sensor temperature readings is A dynamic task allocation strategy is proposed to address localized overheating in many-core systems. Our approach employs reinforcement learning, a dynamic machine learning algorithm that performs task allocation based on current temperatures and a prediction regarding which assignment will minimize the peak temperature. Our results show that the proposed technique is fast (scheduling performed in \u3c1 \u3ems) and can efficiently reduce peak temperature by up to 8 degree C in a 49-core processor (6% on average) versus a leading competing task allocation approach for a series of SPLASH-2 benchmarks. Reinforcement learning has also been applied to 3D integrated circuits to allocate tasks with thermal awareness