임베디드 멀티코어 프로세서를 위한 점유율 제한 기법을 사용한 시스템 전력관리 기법 연구

Abstract

Modern mobile devices have become to be equipped with high-performance application processors (APs) including embedded central processing units (CPUs) and embedded graphics processing units (GPUs). Generally, higher computing performance causes more power consumption; therefore, most manufacturers have provided hardware interfaces to adjust the performance level of computing units such as CPUs and GPUs. Dynamic voltage and frequency scaling (DVFS) is a well-known power management technique to reduce the dynamic power consumption of CPUs. Most mobile devices adopt the OS-driven power management, where the operating system (OS) directly manages power management governors defining how to use DVFS, for embedded CPUs. Therefore, it is important to study the governing algorithm for the CPU power reduction. Among various power management governors using DVFS, the Ondemand governor is widely used because of its simple algorithm against the achievable power reduction. The Ondemand governor uses CPU utilization as a key metric, which shows how long tasks occupy the CPU in a given time. The CPU utilization may be a decent criterion to manage the system; however, it does not represent the quality of subjective satisfaction of end users while a mobile device is in use. Actually, there exists some gap between the physical performance level of hardware and the subjective satisfaction level of end users. If it is able to cut down the gap between the CPU performance level and the users’ satisfaction level, it would be feasible to save more CPU power without affecting the user experience referring to the perceived quality of system services to end users when using a mobile device. Several research works have tried to reflect the user experience in the power management for CPUs or GPUs. Despite their efforts, their works have a structure limitation that DVFS is the only method to adjust the CPU performance level. In case of DVFS, changing the CPU clock speed will influence all tasks running on the CPU. Namely, when the CPU clock speed is lowered to reduce the power consumption in running a certain target task, severe UX degradations of other tasks that are executed on the same CPU may occur. Therefore, it is necessary to have a method to control the CPU performance level without affecting all the tasks that are concurrently running. This dissertation introduces CPU bandwidth control as a power management technique, which is a software approach to adjust the CPU performance level while DVFS is a hardware technique. In this dissertation, CPU bandwidth control has studied to show its effects on the CPU performance level and the CPU power consumption. Finally, the user-centric governors for the CPU power management are proposed. The proposed governors make the following contributions. First, it is the first investigation on CPU bandwidth control as a power management technique. In other words, this dissertation shows how the CPU bandwidth control (BWC) can affect the CPU performance level and how much power can be saved. Second, the correlation between the CPU performance level and the frame rate is investigated in order to use the frame rate as a metric to predict the satisfaction level of end users. Finally, the novel user-centric system-level power management is proposed, and it is implemented on a real device. Several experimental results show that the proposed governor can reduce the CPU power consumption without severe UX degradations when running graphical applications at an embedded device.