CarbonMobile: Carbon-Efficient Multiedge Workload Orchestration for Mobile IoT

Edge computing’s annual carbon footprint is projected to exceed 481MtCO2eq by 2025, yet existing carbon-aware orchestrators are pinned to coarse three-tier action spaces and make myopic per-task decisions. This paper presents CARBON-MOBILE, a carbon-efficient orchestrator whose scope isedge-offloaded mobile IoT— it routes each task to one ofEindividual edges or a single cloud region but does not schedule on-device execution. Two design ideas drive the gain. First, the per-edge action space lets the orchestrator exploit intra-city differences in grid carbon intensity. Second, a Lyapunov virtual queue converts a long-term carbon budget into a per-task penalty; the oracle drift-plus-penalty minimiser admits the standardO(1/V)–O(V )trade-off, while the learned PPO policy we actually deploy inherits no formal guarantee but empirically maintains stable queue dynamics over a four-hour stress test. Mobility traces drive task generation, but a feature ablation shows that explicit mobility features contribute less than 2% to the gain in our evaluated configuration; we therefore frame the system asmobility-compatiblerather than mobility-aware. On calibrated synthetic CI traces matching the training distribution, CARBONMOBILE-DRL reduces mean per-task carbon footprint by 72.9%, 56.8%, and 75.0% versus re-implementations of GreenScale (IoT-J 2025), the energy-aware AutoScale, and an edge-greedy heuristic respectively, with no statistically significant latency degradationin our evaluated workload. An out-of-distribution spatial-spread sweep shows the DRL gain is distribution-specific: when the CI spread shifts the training-free CARBONMOBILE-Greedy variant — which still beats the GreenScale Q-learning baseline by 26–53% — is the recommended fallback.