Limited energy availability is among the most challenging considerations developers face for heterogeneous systems and is critical for battery-powered devices. For complex systems composed of mechanical and computational units, such as drones and mobile robots, more than half of the power consumption can be due to the computational operations. Critically, these systems are often composed of many components, interacting concurrently to achieve specific functionality. As a result, power prediction and estimation can be a challenging task, especially if different computational units, such as CPU and GPU, should be modeled. In this paper, we focus on limited energy availability for mobile heterogeneous devices powered by a battery and present a coarse-grained computation-oriented energy modeling approach. Our approach predicts the energy consumption of a set of software components, in a specific configuration, executed according to a given scheduling policy. The model, determined numerically from several empirical power samples, describes the energy consumed by a software configuration and can be used for energy-aware planning and optimization from a computational point of view. It can potentially support a complex embedded system in maximizing the level of autonomy while minimizing power consumption and preserving the most appropriate amount of battery charge by finding the right rate of quality of service. Our approach is supported and validated by the design and implementation of a profiling tool. The tool abstracts computational energy behavior and describes the current battery drain as a function of all the admissible configurations.

中文翻译：

---

异构并行嵌入式系统的面向粗粒度计算的能量建模

有限的能源可用性是开发人员在异构系统中面临的最具挑战性的考虑之一，对于电池供电的设备至关重要。对于由机械和计算单元组成的复杂系统，例如无人机和移动机器人，一半以上的功耗可能来自计算操作。至关重要的是，这些系统通常由许多组件组成，同时交互以实现特定功能。因此，功率预测和估计可能是一项具有挑战性的任务，尤其是在应该对不同的计算单元（例如 CPU 和 GPU）进行建模的情况下。在本文中，我们关注由电池供电的移动异构设备的有限能源可用性，并提出了一种面向粗粒度计算的能源建模方法。我们的方法预测一组软件组件的能耗，在特定配置中，根据给定的调度策略执行。该模型由多个经验功率样本以数值方式确定，描述了软件配置消耗的能量，可用于从计算角度进行能量感知规划和优化。它可以潜在地支持复杂的嵌入式系统，以最大限度地提高自治水平，同时最大限度地降低功耗，并通过找到合适的服务质量率来保持最合适的电池电量。我们的方法得到了分析工具的设计和实施的支持和验证。该工具抽象了计算能量行为，并将当前电池消耗描述为所有允许配置的函数。