Improving Temperature Prediction Accuracy Using Kalman and Particle Filtering Methods

Predicting the device temperature is crucial for high performance mobile devices since a high temperature reduces the device reliability and lifetime, and increases the power dissipation per processing activity. For these reasons, thermal models are used to predict the temperature and schedule the workloads according to these predictions. This means that more accurate predictions can improve the reliability, lifetime and energy-efficiency of devices. We introduce two different generic methods to extend a thermal model to improve the prediction accuracy. The first method is to extend a thermal model with a Kalman filter. This approach enables a device to adapt to environmental changes more easily and to reduce the effect of noise by combining sensor data and dynamic behavior of the system. However, it assumes every random variable to be normally distributed. The second method is to extend a thermal model with a particle filter. In addition to the ability of adapting better to environmental changes, this approach enables a device to approximate any arbitrary distribution to reduce the effect of noise. Both methods are applicable to any dynamic thermal model to improve its prediction accuracy. Our experimental results show that the new methods indeed improve the prediction accuracy.