CARBON-Lyapunov: Carbon-Aware Edge-Cloud Collaboration With Lyapunov Optimization for Agricultural Consumer Electronics

초록

Consumer agricultural cameras (smartphones and UAV payloads) increasingly execute vision models under strict bandwidth, energy, and carbon constraints. We propose Carbon-Lyapunov (CLy), a carbon-aware, resource-efficient controller for edge-cloud collaboration that jointly optimizes placement (device/edge/cloud), model-level actuation (early-exit depth, quantization, DVFS), and timing under dynamic carbon signals. The problem is formulated as drift-plus-penalty optimization and solved via distributed ADMM with closed-form local policies, ensuring lightweight scalability. CLy is supported by theoretical guarantees: convex relaxations yield strong duality, Lyapunov analysis proves queue stability with an O(1/V) optimality gap, and regret bounds confirm robustness under non-stationary environments. Empirical evaluation using accuracy-cost profiles from an early-exit Convolutional Neural Network (CNN) trained on PlantVillage and real carbon-intensity traces shows that CLy achieves 100% Service Level Objective (SLO) hit rate, median latency near zero, and p90 latency below 15 ms, while consuming approximate to 494 J and emitting approximate to 384 g CO2 across 4800 tasks. Comparative results confirm that CLy tracks the Pareto frontier between latency, energy, and emissions. Overall, this work demonstrates that integrating carbon-awareness with model-level levers and formal optimization yields practical, provable, and reproducible benefits for agricultural vision systems without requiring new field data.

키워드