Abstract
<title>Abstract</title> <p>Federated cloud environments are gaining momentum for hosting latency-sensitive, geo-distributed, containerised applications, yet efficient scheduling in such environments remains a significant challenge. Current approaches use heuristic, rule-based, or single-objective techniques that prioritise latency, cost, or energy performance in isolation, thereby complicating SLO adherence, inter-cluster traffic, fairness, and migration stability. These limitations make them less suitable for more dynamic multi-region cloud environments with high variability and heterogeneities in workloads and resources. Motivated by this gap, this paper introduces an AI-based framework that performs multi-objective scheduling and load balancing in federated cloud environments. It incorporates a graph neural network (GNN)-based tail-latency predictor; energy and cost estimators; an NSGA-II-based Pareto optimiser; a migration-aware feasibility guard; policy distillation; and a contextual bandit for adaptive online policy selection. Our algorithm minimises SLO violations, energy consumption, operational cost, and inter-cluster traffic while maximising fairness and jointly optimising stability. Specifically, experimental evaluation based on Google Cluster Workload Traces and DeathStarBench microservice benchmarks shows that the proposed scheduler achieves the lowest SLO violation rate of 3.28%, the lowest energy consumption of 132.94 kWh, the lowest operational cost of $389.27, and the lowest inter-cluster traffic of 248.36 GB. Also achieved lower p95 and p99 latencies of 148.2 ms and 196.4 ms, respectively, 6 hotspot incidents, and a fairness index of 0.93. The results validate that the presented framework is a pragmatic, adaptive, and scalable solution for intelligent container-based federated cloud scheduling, with high utility for Next Generation Edge–Cloud Orchestration and Distributed Service Management.</p>