Novel Feature Selection Algorithm for Thermal Prediction Model

Demand for more computing power grows steadily, which leads to increase the integration density of modern processors that rise thermal hotspots. A proactive thermal management algorithm tries to avoid exceeding a thermal threshold by exploiting the control decisions and using the thermal prediction model. In this paper, we proposed a new approach to build a thermal prediction model for a multicore processor using a multilayer perceptron (MLP). We generate a data set composed of system state samples gathered during Standard Performance Evaluation Corporation benchmark execution. Thereafter, a compilation method is applied to extract the new features from the data set. These features are categorized into two behavioral and reflective groups. The first group allows the model to track the current thermal behavior of the processor, whereas the second one helps the model to predict the control response temperature. Finally, a smaller set of input features is selected by a new proposed feature selection algorithm. The evaluation shows that the mean prediction error of the proposed model is about 0.5 degrees C-0.6 degrees C with 0.6 degrees C-0.7 degrees C standard deviations from 2- to 5-s prediction distances. The results show the superiority of our model and feature selection algorithm in comparison with the counterparts.