Predictive thermal management for hard real-time tasks

Abstract

Dynamic thermal management (DTM) techniques help a system to operate within a safe range of temperature by reducing the performance of the CPU dynamically when the system is too hot. Dynamic voltage scaling (DVS) and localized toggling are both DTM techniques. DVS is easier to use for real-time systems since the performance degradation can be controlled accurately so that tasks are still meeting their deadlines. Localized toggling changes architectural configurations of a CPU to a less optimized setting. Its performance degradation is harder to measure and to control accurately. In this paper, we propose a method which applies various localized toggling techniques to real-time systems while still able to meet task deadlines. Our method activates DTM when the temperature over the execution of a job is predicted to be too high at the start time. When DTM is activated, our method measures the performance degradation of different toggling techniques during the slack time to select the most effective technique and still be able to meet deadline. We use instructions per cycle (IPC) as the performance measure. Our method is evaluated on the SimpleScalar CPU simulator with Wattch, the energy simulator, and HotSpot, the thermal model.