A Deep Reinforcement Learning Approach to Optimizing Cloud Workload Migration

Abstract

Cloud data centers consume a significant amount of energy worldwide, prompting the need for intelligent resource management. Dynamic workload migration (moving virtual machine workloads between servers or to the cloud) can improve resource utilization and reduce energy consumption by consolidating loads onto fewer machines. However, live migration incurs performance overhead; migrating too frequently or at suboptimal times can degrade application performance. This paper proposes a novel AI-driven approach to optimize cloud workload migration decisions. We leverage deep reinforcement learning (RL) to autonomously learn when and where to live-migrate workloads in order to minimize energy use and operational costs while respecting performance constraints. The proposed method uses publicly available cloud workload traces to train and evaluate the RL agent’s decision-making. We design and implement the solution within a simulation environment, and extensive experiments show that our method significantly outperforms baseline heuristics in reducing energy consumption (by over 20%) and lowering service-level agreement (SLA) violations.