Reinforcement Learning for Dynamic Resource Allocation in Cloud Data Centers
Reinforcement Learning for Dynamic Resource Allocation in Cloud Data Centers
Yashar Abbasloo Soumeeh1
1) University Lecturer, Senior Data Center Engineer, Karaj Fire Department and Safety Services Organization
IOI Code
:
ACCD-ACAZFE
Publication :
10th International Conference on Applied Researches in Science & Engineering (carseconf.com)
Abstract :
Abstract
Energy is a major operating cost of cloud data centers, yet many servers run far below capacity for much of the day. Deciding how many servers to keep active at each moment is therefore a central control problem: too few servers violate service-level agreements (SLAs), whereas too many waste energy. Static provisioning and hand-tuned threshold rules are simple but do not adapt to changing workloads. This paper studies reinforcement learning (RL) as an alternative for dynamic server-pool sizing. We formulate the problem as a Markov decision process whose cost combines energy, unmet demand and server state changes, and in which a one-slot boot delay forces the controller to act in anticipation of demand. A tabular Q-learning agent is trained in a simulated cluster of 12 servers under a diurnal workload with random bursts and is compared with four baselines: two static provisioning levels and two reactive utilization-threshold policies. Averaged over three independent training runs and 100 unseen evaluation days, the RL policy lowers the total daily cost by about 37% relative to peak provisioning and by 5.5% relative to a 70% threshold policy. It comes within 0.5% of the best threshold policy, whose threshold had to be tuned manually. Preliminary experiments also suggest that the results are sensitive to the state design and the learning-rate schedule. Finally, we discuss the conditions that determine whether such agents can be deployed in production, including public-sector IT environments: safe exploration, simulator fidelity, sample efficiency and explainability.
Keywords : reinforcement learning resource allocation cloud data centers q learning