Temperature Aware Scheduling in Multicore Systems

Abstract

Power density in microprocessors is increasing rapidly, which results in higher operating temperature, and systems are prone to overheating. Dynamic thermal management has been used to dissipate heat, reduce the operating temperature to avoid thermal emergencies but is not aware of the application behavior. Several techniques like thread migration, Dynamic Voltage Scaling (DVS), Dynamic Voltage Frequency Scaling (DVFS), clock gating etc. were used in the past to resolve the problem of thermal emergencies but are reactive to the increased chip temperature. In this thesis, we propose a proactive dynamic thermal management method that is based on grouping of the applications by thermal behavior. In the experiments, offline data is used as the primary resource for the categorization of running application. The temperature of a core is predicted by the rate of temperature change at the core proportional to the difference in current and steady state temperature. We evaluate this method in a multicore system running several benchmarks. The experimental results show a decrease of 1.2 -1.6 C in the average peak temperature of the CPU with a little performance overhead as compared to the standard Linux scheduler.