Analysis and elimination of noise-induced temperature error in processor thermal control

Abstract

Processor thermal control in real-time systems is crucial because overheated processors may result in critical performance degradation or even system failure. The main challenges in processor thermal control in real-time systems are as follows: (i) the need to satisfy both real-time and thermal constraints; (ii) uncertain system dynamics; and (iii) thermal sensor noise. The first two issues have been resolved in substantial studies while the issue of sensor noise has not been thoroughly addressed. In this paper, we experimentally identify that even a small zero-mean sensor noise can induce a significant steady-state error in processor thermal control. This steady-state temperature error is contrary to the intuition that zero-mean sensor noise normally induces zero-mean fluctuations around the target temperature. We rigorously analyze the phenomenon based on the stochastic averaging theory and quantify the error in a closed form in terms of noise statistics and system parameters. We propose a thermal control architecture for effectively eliminating the thermal error and tightly controlling the processor temperature. Through extensive experiments, we validate the proposed architecture.