The increasingly growth of autonomous end devices applied to Internet of Things (IoT) environments and has led to the development of new microcontrollers with enhanced power modes. An operating system (OS) is usually implemented on top of these resource-constrained devices to enhance applications development and also manage system resources and power modes. Dynamic frequency scaling (DFS) techniques have long been implemented in many OSes, as well as tick-less scheduling strategies which are gaining importance to improve energy efficiency in new embedded devices. In this work, we performed a novel analysis of the power consumption achieved when combining DFS and tick-less scheduling with enhanced power modes. This analysis pointed to the existence of different optimal clock frequencies in terms of power consumption depending on the application being executed at any given time in those scenarios. Based on these results, we propose an adaptive frequency scaling (AFS) OS module based on a reinforcement learning algorithm to dynamically adapt the clock frequency to a value as close as possible to the optimal one in terms of power consumption in each time. Several experimental tests were performed to measure the power consumption when executing a set of different applications on top of a real-time OS running on a STM32L476 microcontroller. The results confirmed the hypothesis raised by the theoretical analysis and demonstrated that the proposed AFS module may improve the energy efficiency compared to the use of an arbitrary non-optimal static clock frequency, especially in applications with a dynamic behaviour. The proposed AFS module also allows guaranteeing real-time constraints under certain conditions, and it operates automatically and transparently to the application level, so a developer does not require any additional effort. As a result, AFS strategies raise as a valuable approach that may boost energy efficiency in future resource-constrained devices.

应用于物联网 (IoT) 环境的自主终端设备的日益增长，并导致开发出具有增强功率模式的新型微控制器。操作系统 (OS) 通常在这些资源受限的设备之上实现，以增强应用程序开发并管理系统资源和电源模式。动态频率缩放 (DFS) 技术早已在许多操作系统中实施，并且无滴答调度策略对于提高新嵌入式设备的能效越来越重要。在这项工作中，我们对将 DFS 和无滴答调度与增强功率模式相结合时实现的功耗进行了新颖的分析。该分析指出，在这些场景中，根据在任何给定时间执行的应用程序，在功耗方面存在不同的最佳时钟频率。基于这些结果，我们提出了一种基于强化学习算法的自适应频率缩放 (AFS) OS 模块，以动态地将时钟频率调整为尽可能接近每次功耗最佳值的值。在 STM32L476 微控制器上运行的实时操作系统之上执行一组不同的应用程序时，进行了多项实验测试以测量功耗。结果证实了理论分析提出的假设，并证明与使用任意非最佳静态时钟频率相比，所提出的 AFS 模块可以提高能源效率，尤其是在具有动态行为的应用中。提议的 AFS 模块还允许在某些条件下保证实时约束，并且它对应用程序级别自动透明地运行，因此开发人员不需要任何额外的工作。因此，AFS 策略成为一种有价值的方法，可以提高未来资源受限设备的能源效率。所以开发人员不需要任何额外的努力。因此，AFS 策略成为一种有价值的方法，可以提高未来资源受限设备的能源效率。所以开发人员不需要任何额外的努力。因此，AFS 策略成为一种有价值的方法，可以提高未来资源受限设备的能源效率。