On-line learning based dynamic thermal management for multicore systems

Power consumption of a high-end microprocessor increases very rapidly. High power consumption will lead to rapid increase in chip temperature as well. If temperature reaches beyond a certain level, chip operation becomes either slow or unreliable. Therefore various approaches for dynamic thermal management (DTM) have been proposed. In this paper, we propose a new application-oriented learning-based dynamic thermal management (LDTM) technique for a multi-core system. From repetitive executions of an application, we learn the thermal patterns of the chip, and we control the future temperature through DTM. When the predicted temperature may rise above a threshold value, we reduce the temperature by decreasing the operation frequency of the corresponding core. We implement our learning-based thermal management on an Intel's dual core system which is equipped with digital thermal sensors (DTS). The dynamic frequency scaling (DFS) is implemented to have three frequency steps on a Linux kernel. We carried out experiments using Phoronix Test Suite benchmarks for Linux. The peak temperature has been reduced by on average 7°C using our LDTM, and the overall average temperature reduced from 72°C to 65°C.