一句话：提出 LucRETius，以运行时事件为公共特征做迁移学习，把一台机器上的功耗模型快速移植到另一台机器。
开发能耗感知的应用是推动可持续计算的重要手段，但开发者与部署者面临一个根本性难题：如何把为一台机器构建的功耗模型移植到另一台机器上？本文提出 LucRETius，一种基于迁移学习的可移植功耗建模方法：在源机器上预训练的功耗模型充当 teacher，帮助在目标机器上快速构建 student 模型。其核心洞察在于，应用运行时这一层能够抽象掉机器特有的细节，因此即便两台机器的硬件与软件系统配置不同，仍可由一组统一的运行时事件作为迁移学习的公共特征。作者在 4 台机器上进行了双向迁移评估，对 36 个真实 JVM 应用的能耗预测中位百分比误差仅为 1.01%–1.75%；与从零开始训练相比，LucRETius 可带来 8.07 倍的加速。
Developing energy-aware applications is an important approach to promoting sustainable computing. This paper addresses a fundamental and challenging problem faced by developers and deployers: how to port a power model built for one machine to another? We present LucRETius, a novel approach toward portable power modeling through transfer learning, where a pre-trained power model on one machine can serve as a teacher to help construct a student model on another machine rapidly. The key insight that enables transfer learning is that the layer of application runtimes can abstract away the machine-specific details, so that two machines—despite different hardware and software system configurations—are unified with a common set of runtime events that can serve as features for transfer learning. We evaluate LucRETius through bi-directional transfers across 4 machines, and we show the power models built by LucRETius have a median percentage error of 1.01%-1.75% when used for predicting the energy consumption of 36 real-world JVM-based applications. Compared with training from scratch, LucRETius can lead to a speed up of 8.07 ×.