Thermal status and workload prediction using support vector regression

Because knowing information about the currently running workload and the thermal status of the processor is of importance for more adequate planning and allocating resources in microprocessor environments, we propose in this paper using support vector regression (SVR) to predict future processor thermal status as well as the currently running workload. We build two generalized SVR models trained with data from monitoring hardware performance counters collected from running SPEC2006 benchmarks. The first model predicts the Central Processing Unit´s thermal status in Celsius with a percentage error of less than 10%. The second model predicts the current workload with a percentage error of 0.08% for a heterogeneous training set of 6 different integer and floating point benchmark workloads. Cross validation for the two models show the effectiveness of our approach and motivate follow on research.