Utilizing Predictors for Efcient Thermal Management in Multiprocessor SoCs

Conventional thermal management techniques are reactive in nature; that is, they take action after temperature reaches a predetermined threshold value. Such approaches do not always minimize and balance the temperature on the chip, and furthermore, control temperature at a noticeable performance cost. In this work, we investigate how to use predictors for forecasting future temperature and workload dynamics, and propose proactive thermal management techniques for multiprocessor system-on-chips (MPSoCs). The predictors we study include autoregressive moving average (ARMA) modeling and look-up table based predictors. We implement several reactive and predictive thermal management techniques, and provide extensive evaluations on an UltraSPARC T1 system as well as on an architecture-level simulator. By means of dynamic prediction, proactive methods can achieve signicantly better thermal proles in comparison to their reactive counterparts, and the performance cost associated with dynamic thermal management can be mostly avoided. 1Utilizing Predictors for Efcient Therma