Title: RL-MEOPS: reinforcement learning-driven memory energy optimisation via hybrid slack-aware control

Authors: Sai-Feng Zeng

Addresses: Department of Computer and Communication, Hunan Institute of Engineering, Xiangtan, 411104, China

Abstract: Modern datacentres demand memory energy optimisation without compromising performance, yet existing methods struggle with dynamic workloads and delayed responsiveness. To address this, we propose RL-MEOPS, a reinforcement learning-enhanced framework that dynamically balances energy efficiency and performance guarantees via hybrid slack-driven control. Our approach integrates an RL-based predictive model with runtime performance slack constraints, enabling proactive memory frequency scaling and powerdown state transitions during idle intervals. A safety layer enforces strict slack adherence, while dual-mode operation maximises additive energy savings. Evaluations across compute-intensive, memory-intensive, and balanced workloads demonstrate RL-MEOPS's superiority over state-of-the-art baselines in terms of energy savings and other metrics. The framework's adaptability to workload phase shifts and minimal runtime overhead highlight its practicality for deployment in energy-constrained systems.

Keywords: cloud computing; resource virtualisation; virtual machine; scheduling algorithm.

DOI: 10.1504/IJCC.2026.152356

International Journal of Cloud Computing, 2026 Vol.15 No.1, pp.17 - 34

Received: 03 Mar 2025
Accepted: 18 Aug 2025
Published online: 16 Mar 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article