Abstract: Dynamic Voltage and Frequency Scaling is essential for enhancing energy efficiency in mobile platforms. However, traditional heuristic-based governors are increasingly inadequate for managing the complexity of heterogeneous System-on-Chip designs and diverse application workloads. Although reinforcement learning approaches offer improved performance, their poor generalization capability and reliance on extensive retraining for each hardware and application combination leads to significant deployment costs. In this work, we observe that device and application metadata inherently encapsulate valuable knowledge for DVFS, presenting an opportunity to overcome these limitations. We formulate DVFS for heterogeneous devices and applications as a multi-task reinforcement learning problem. We introduce MetaDVFS, which is a metadata-guided framework that systematically leverages metadata to discover and transfer shared knowledge across DVFS tasks. MetaDVFS can output a set of DVFS models with significant generalization capability for various applications of heterogeneous devices. Evaluations on five Google Pixel devices running six applications show that MetaDVFS achieves up to 17% improvement in Performance-Power Ratio and up to 26% improvement in Quality of Experience. Compared to state-of-the-art methods, MetaDVFS delivers 70.8% faster adaptation and 5.8-27.6% higher performance over standalone device-application specific training, while avoiding negative transfer effects. These results establish MetaDVFS as an effective and scalable solution for DVFS deployment in heterogeneous mobile environments.
Introduction: The Generalization Crisis in Mobile Power Management
Dynamic Voltage and Frequency Scaling (DVFS) is a critical component of mobile power management, responsible for balancing processor performance with energy consumption. However, modern mobile systems are facing a "generalization crisis." Traditional heuristic-based governors, such as Android's schedutil, are increasingly inadequate for managing the complexity of contemporary heterogeneous System-on-Chip (SoC) designs and the highly dynamic nature of modern application workloads.
While Reinforcement Learning (RL) offers a more sophisticated alternative, standard RL models are notoriously brittle. A model optimized for a specific application on one device often fails when deployed on a different hardware-software combination. For manufacturers, the cost of retraining custom models for every possible device-app pair is prohibitive. The core challenge lies in building a DVFS system that can learn from one scenario and rapidly adapt to another.
MetaDVFS: A Metadata-Guided Framework
To address these challenges, we introduce MetaDVFS, a framework that treats frequency scaling as a multi-task reinforcement learning problem. MetaDVFS systematically leverages device and application metadata—such as chip process nodes, core counts, and target frame rates—to discover and transfer shared knowledge across tasks.
The framework is built on three key pillars:
- Liquid Neural Network (LNN) Backbone: MetaDVFS utilizes LNNs, which are inspired by biological neurons. LNNs excel at handling the continuous-time dynamics and noisy, irregularly sampled data characteristic of mobile workloads, making them more robust than traditional Recurrent Neural Networks (RNNs).
- Metadata-driven Task Definition: The system employs a tree-based clustering algorithm to group device-application pairs based on shared metadata. This strategic clustering ensures that knowledge is only transferred between relevant tasks, preventing "negative transfer" where learning from one context might degrade performance in another.
- MAML-based Meta-Training: Using Model-Agnostic Meta-Learning (MAML), the framework learns a "meta-model" that can be rapidly adapted to new, unseen devices or applications with minimal local data.
Experimental Evaluation and Results
We evaluated MetaDVFS on a real-world testbed comprising five generations of Google Pixel devices (Pixel 3, 4, 6, 8, and 9) and six popular applications representing diverse workload profiles, including video apps (TikTok, Kwai), interactive apps (Taobao), and graphics-intensive benchmarks (3DMark).
Key Performance Metrics
- Efficiency Gains: MetaDVFS achieved up to a 17% improvement in the Performance-Power Ratio (PPW) and a 26% boost in Quality of Experience (QoE) compared to default Android governors.
- Rapid Adaptation: The system reached optimal performance in just 3.5 ± 1.1 minutes, representing a 70.8% faster adaptation rate than standard incremental training methods, which typically require nearly 12 minutes.
- Cross-Architecture Robustness: The metadata-guided approach maintained stable performance even when transitioning between vastly different chip architectures, such as Qualcomm Snapdragon and Google Tensor.
Conclusion
MetaDVFS demonstrates that metadata is a powerful proxy for runtime behavior, enabling the creation of scalable and adaptable DVFS solutions. By combining LNNs with intelligent metadata-driven clustering and meta-learning, the framework overcomes the generalization limitations of traditional RL. These results establish MetaDVFS as an effective solution for deploying high-performance, energy-efficient power management across the diverse landscape of modern mobile devices. Future work will explore extending this metadata-guided approach to other system-level optimizations, such as thermal management and network scheduling.
Summary updated