Run-time probabilistic detection of miscalibrated thermal sensors in many-core systems

Many-core architectures use large numbers of small temperature sensors to detect thermal gradients and guide thermal management schemes. In this paper a technique to identify thermal sensors which are operating outside a required accuracy is described. Unlike previous on-chip temperature estimation approaches, our algorithms are optimized to run on-line while thermal management decisions are being made. The accuracy of a sensor is determined by comparing its readings to expected values from a probability distribution function determined from surrounding sensors. Experiments show that a sensor operating outside a desired accuracy can be identified with a detection rate of over 90% and an average false alarm rate of < 6%, with a confidence level of 90%. The run time of our method is shown to be around 3x lower than a recently-published temperature estimation method, enhancing its suitability for run-time implementation.