DVFO: Learning-Based DVFS for Energy-Efficient Edge-Cloud Collaborative Inference

• DOI: 10.48550/arxiv.2306.01811
• Publication Date: 2023-01-01
• Authors: Ziyang Zhang; Yang Zhao; Huan Li; Jie Liu
• Keywords: Edge device; Frequency scaling

Due to limited resources on edge and different characteristics of deep neural network (DNN) models, it is a big challenge optimize DNN inference performance in terms energy consumption end-to-end latency devices. In addition the dynamic voltage frequency scaling (DVFS) technique, edge-cloud architecture provides collaborative approach for efficient inference. However, current methods have not optimized various compute Thus, we propose DVFO, novel DVFS-enabled framework, which co-optimizes DVFS offloading parameters via reinforcement learning (DRL). Specifically, DVFO automatically 1) CPU, GPU memory frequencies devices, 2) feature maps be offloaded cloud servers. addition, leverages thinking-while-moving concurrent mechanism accelerate DRL process, spatial-channel attention extract secondary importance workload offloading. This improves models under conditions. Extensive evaluations using two datasets six widely-deployed three heterogeneous devices show that significantly reduces by 33% average, compared state-of-the-art schemes. Moreover, achieves up 28.6%-59.1% reduction, while maintaining accuracy within 1% loss average.