Online Thermal Prediction in Different Thermal Phases for Multicore Systems

Increasing the number of cores to enhance computing power of processors leads to an increase in temperature for multicore systems. Thermal management is significant challenge in these processors. A proactive dynamic thermal management uses a thermal model to predict the temperature before processor temperature reaches the threshold. In this paper, some appropriate features for thermal model are read by using system measurement tools. Other features as historical and control features are created using the proposed processes. An online thermal model based on several thermal phase is proposed. For each phase, a neural network is used to forecast temperature. Different thermal phases are identified according to the parameters affecting the processor temperature using the adaptive resonance theory network. For each of the neural networks, the minimum number of proper features is selected based on the mutual information between the features. The proposed thermal model is able to detect new thermal phase at run time. Then, appropriate neural network is created for new phase. The proposed model has been evaluated to predict temperature for different time distances. The results shows the mean absolute error is less than 1 °C.