International Journal of Science, Engineering and TechnologyAuthors: Dr.Girija D.K, Dr.Rashmi M, Dr.Divyashree J
Abstract: This study presents a carbon-aware scheduling framework for edge-centric AI workloads in a smart-campus setting. We formalize a multi-objective problem that minimizes energy use and carbon-weighted energy while satisfying latency service-level agreements through joint control of dynamic voltage/frequency scaling (DVFS), task placement across heterogeneous edge clusters and cloud, temporal deferral of non-urgent jobs, and adaptive model selection. A tractable decomposition combines convex deferral (water-filling over time-varying grid-intensity), min-cost flow for placement on a time-expanded graph with carbonized edge weights, DVFS tuning under queue-stability constraints, and a contextual bandit for model choice. Queueing performance is modeled with M/M/1 response times to enforce utilization caps and SLA feasibility. In a day-long synthetic campus case study with diurnal arrivals and carbon cycles, the proposed controller achieves ≈31% average energy reduction relative to a greedy low-latency baseline, with single-digit to tens-of-milliseconds latency penalties for interactive tasks; larger savings accrue for deferrable analytics via demand shaping. Ablations indicate DVFS and deferral deliver the largest gains, while model selection contributes incremental savings without harming accuracy targets. We discuss deployment guidance, limitations (forecast error, burstiness, hardware heterogeneity), and future directions including robust/MPC extensions and integration with campus microgrids. The results demonstrate that principled, carbon-aware control can materially improve sustainability without sacrificing user experience.
