Graph Neural Network-Guided Reinforcement Learning with Hybrid Optimization for SLA-Sensitive Multi-Cloud Scheduling

Abstract

The scheduling of tasks in multi-cloud systems is complex due to workload heterogeneity, resource variability, and the need to meet Service Level Agreement (SLA) requirements. Researchers have found that traditional heuristic and metaheuristic methods fail to balance efficiency, cost, energy, and scalability in dynamic environments. To address these issues, this paper presents a Graph Neural Network (GNN)-assisted Deep Reinforcement Learning (DRL) model enhanced with hybrid optimization for SLA activity scheduling in multi-cloud systems. The structure utilizes workload-aware profiling based on the Workload Intensity Score (WIS), refinement of the workload heat score (WHS), and suitable task-VM mappings using GNN, along with adaptive scheduling via an actor-critic DRL agent. A Differentiable Surrogate Optimizer (DSO) ensures constraints are feasible, while a hybrid fallback method combining quantum firefly optimization (QFO) and an improved genetic algorithm (IGA) optimizes workload allocations during peak stress. Experimental results show that this approach reduces execution time by 30-40%, cuts SLA violations by 60%, increases throughput by 25-30%, and decreases energy consumption by 20-25% compared to baseline algorithms, demonstrating its effectiveness and scalability in multi-cloud environments.