III: Small: Reducing The Graphical User Interface's Consumption of Power on Mobile Devices

随着移动设备获得新功能，它们的能源需求也随之增加。十年前，用户可能需要每周给移动设备充电一次；现在，每天给它们充电是很常见的。这种增加的能源需求部分是由于移动设备不断轮询，以确定是否已接收到输入，如手势、电子邮件消息或蜂窝电话。轮询，即使设备处于空闲状态和用户口袋中，也会不断唤醒CPU，防止其进入低功耗、节能状态。由于输入设备传统上在接收输入时是“静音”的，因此这种能源效率低下的“拉动”模型是必要的。因此，为了确保及时处理输入，CPU被迫每秒多次查询每个输入设备，以确定该设备是否正在保存输入。最近的技术进步使输入设备能够在输入到达时立即通知CPU。这一进步允许我们研究另一种“推送”模型，该模型在没有输入时将CPU置于睡眠状态，只有在实际接收到输入时才唤醒CPU。因此，当移动设备闲置在用户的口袋里时，CPU不再需要消耗电力。此外，较新的移动设备有多个CPU核心，其中一些较慢，功耗较低的核心，当CPU处于休眠状态时，它们会更快地唤醒。我们可以将引导输入到应用程序的软件放在其中一个低功耗的内核上，而应用程序本身则放在更快、更高功耗的内核上。这种布局有两个好处。首先，一些输入可以被忽略，如果一个输入可以在低功率核心上处理和丢弃，那么高功耗核心就不需要被唤醒。其次，如果输入必须由应用程序处理，则可以预热高功率核，同时收集有关输入的信息并为低功率核上的应用程序准备。为了实现这个推送模型，事件处理的控制点将从应用程序转移到内核中。这种新的布局将需要1)开发用于事件调度的内核例程，2)开发允许进程调度器知道哪些进程是输入驱动的内核级显示服务器，以及3)修改进程调度器，以便它可以根据需要取消输入驱动进程的调度。我们将首先为图形用户界面接收的输入开发一个推送模型，然后将其扩展到通用系统I/O，如传感器和GPS信号。我们计划修改Android运行时（ART），以便与推送模型一起工作，以便分析实际应用，以确定使用我们提议的模型可以节省多少电力，并可能减少延迟。我们的研究应该产生更广泛的影响，既减少移动设备需要充电的频率，又通过减少移动设备消耗的能量来减少对环境的影响。另一个更广泛的影响是，拟议的技术应该适用于蜂窝和Wi-Fi处理应用程序，也适用于台式电脑。

Mobile devices have increased their energy demand as they have acquired new capabilities. A decade ago, users might have had to re-charge their mobile devices once a week; now it is common to have to re-charge them daily. This increased energy demand is in part due to the constant polling that a mobile device performs to determine whether an input, such as a gesture, an email message, or a cellular call, has been received. The polling, even when the device is idle and in the user's pocket, constantly rouses the CPU and prevents it from being put into low power, energy-conserving states. This energy inefficient 'pull' model has been necessitated by the fact that input devices traditionally have been 'mute' when they receive an input. Thus in order to ensure that an input is promptly processed, the CPU is forced to query each input device multiple times per second to determine if the device is holding an input. Recent technological advances now enable input devices to immediately notify the CPU when an input has arrived. This advance allows us to investigate an alternative 'push' model that puts the CPU to sleep while no inputs are present and awakens the CPU only when an input is actually received. Hence the CPU no longer needs to consume power while the mobile device sits idle in the user's pocket. Additionally, newer mobile devices have multiple CPU cores, some of which are slower, lower power consuming cores that awaken faster when the CPU has been asleep. We can place the software that directs inputs to apps on one of these lower power consuming cores and the apps themselves on the faster, higher power consuming cores. This placement has two benefits. First, some inputs can be ignored, and if an input can be handled and discarded on the lower power core, then the higher power consuming cores do not need to be awakened. Second, if the input must be handled by an app, the higher power core can be warmed up while information about the input is gathered and staged for the app on the lower power core.In order to realize this push model, the locus of control for event handling will move from the application into the kernel. This new placement will necessitate 1) the development of kernel routines for event dispatching, 2) the development of a kernel-level display server that allows the process scheduler to know which processes are input driven, and 3) the modification of the process scheduler so that it can de-schedule input-driven processes as needed. We will first develop a push model for inputs received by graphical user interfaces and then extend it to generic system I/O such as sensors and GPS signals. We plan to modify the Android runtime (ART) to work with the push model in order to profile actual apps to determine how much power savings, and potentially reduced latency, can be achieved with our proposed model. Our research should have the desirable broader impact of both reducing the frequency with which mobile devices need to be recharged and reducing the environmental impact of mobile devices by reducing the amount of energy they consume. Another broader impact is that the proposed techniques should work for cellular and Wi-Fi handling apps and for desktop computers as well.