Multi-Objective Cloud Workflow Scheduling With Two-Stage Deep Reinforcement Learning and Graph Neural Networks

Cloud workflow scheduling is a critical challenge in modern cloud computing, requiring intelligent task-to-resource assignment under dynamic conditions and conflicting objectives such as makespan, energy consumption, and QoS latency. Traditional methods often fall short in addressing complex task dependencies and multi-objective trade-offs. This paper proposes 2SD-GAT, a novel dynamic scheduling approach that combines Two-Stage Deep Reinforcement Learning (DRL) with Graph Neural Networks (GNNs) to optimize cloud workflow execution. The method models workflows as Directed Acyclic Graphs (DAGs) and uses a Proximal Policy Optimization (PPO)-based two-stage agent to sequentially select tasks and allocate virtual machines (VMs). A Graph Attention Network (GAT) encoder captures workflow structure, while a preference-based reward function enables trade-off exploration between makespan, energy consumption, and QoS latency. The agent is trained on synthetic and real-world datasets, Comparative evaluation against random, heuristic, and evolutionary baselines shows that the proposed approach consistently outperforms in hypervolume and IGD metrics while maintaining low decision latency. Notably, it achieves a 26.8% hypervolume gain with a 4.5x lower IGD on real-world cloud traces under dynamic conditions, demonstrating both effectiveness and scalability in real-time multi-objective cloud scheduling.