Reinforcement learning approach to implementation of individual controllers in data centre control system

Abstract

Contemporary data centres consume electricity on an industrial scale and require control to improve energy efficiency and maintain high availability. The article proposes an idea and structure of the framework supporting development and validation of the multi-agent control for the energy-efficient data centre. The framework comprises two subsystems: the modelling toolbox and the controlling toolbox. This work focuses on such essential components of the controlling toolbox, as an individual controller. The reinforcement learning approach is applied to the controllers’ implementation. The server fan controller, named SF agent, is implemented based on the framework infrastructure and reinforcement learning approach. The agent’s capability of energy-saving is demonstrated.