Energy-efficient cloud task scheduling using deep learning

Abstract

This thesis presents a deep learning-based approach for energy-efficient cloud workflow scheduling by modeling scheduling as a multi-objective reinforcement learning problem that simultaneously optimizes makespan, energy consumption, and Quality of Service (QoS) . The work first introduces GNN-Flat, a baseline Graph Neural Network (GNN) scheduler that represents workflow dependencies using Directed Acyclic Graph (DAG) structures and demonstrates the feasibility of applying graphbased learning to cloud task scheduling. Building on insights from this implementation, the research proposes 2SD-GAT, a two-stage deep reinforcement learning scheduler based on Graph Attention Networks (GAT) , which forms the core contribution of the thesis. The 2SD-GAT architecture separates task selection and resource allocation decisions while leveraging attention mechanisms to capture inter-task dependencies and preference-based reward optimization, enabling improved Pareto trade-offs across competing objectives. Extensive experiments using synthetic and real-world datasets show superior performance compared with heuristic and learning-based baselines. The results show that the proposed approach achieves a 26.8% hypervolume gain with a 4.5x lower Inverted Generational Distance (IGD), along with a makespan improvement of 14.08%. Finally, the research bridges theory and practice by developing a pluggable RL-based scheduling framework, integrations with Apache Airflow and Kubernetes, and supporting tools including workflow engines, execution components, and a web-based GUI, demonstrating how the proposed scheduler can be deployed in realistic cloud orchestration environments.