Adaptive Carbon-Aware Scheduling for Federated Cloud Systems with Time-Critical Workloads

Data centers have become significant contributors to global carbon emissions, and in federated cloud environments, their impact is amplified by rapid, region-specific fluctuations in grid carbon intensity. These dynamics make sustainable workload placement a complex multi-objective challenge. This study examines scheduling strategies that consider carbon exposure, deadline compliance, and load balancing using only current and historical telemetry data. Two research questions are addressed: (i) how effective are traditional real-time schedulers, Earliest-Deadline-First (EDF) and Least-Slack-Time (LST), when combined with greedy low-carbon cloud selection; and (ii) whether a contextual bandit can dynamically improve these trade-offs. We simulate ten European clouds with realistic carbon traces and time-varying stochastic workloads. Across randomized trials, greedy EDF demonstrates the greatest carbon efficiency. Meanwhile, the contextual bandit learns to balance sustainability and timeliness better, reducing deadline misses by nearly 50% and improving load distribution while keeping carbon emissions within a few percent of the optimal baseline.