강화학습 기반 모바일 기기의 DVFS 정책

MasterThis thesis presents a power management policy that exploits reinforcement learning to increase the power efficiency of mobile device systems. Proposed policy based on reinforcement learning predicts a system's characteristics and learns power management controls to adapt to the system's variations. The policy determines the state of the system by considering the thread-level parallelism and the number of clock cycles per CPU cluster. Therefore, the policy can flexibly manage the system power regardless of the application scenarios and can achieve lower energy per QoS (EPQ) compared to previous dynamic voltage/frequency scaling governors; it was 45% lower than that of interactive, 40% lower than that of ondemand, 72% lower than that of performance, 53% lower than that of mm-centric, and 10% lower than that of pw-min. Although the EPQ of powersave governor is almost the same as the proposed one, it causes a lot of QoS violations so that it is difficult to use on real mobile devices. Furthermore, I reduced the run-time overhead by implementing the proposed policy as hardware. The hardware-implemented policy can make a decision 3.92 times faster than the software-implemented policy