An Energy Saving Control Strategy Based on Multi-Agent Q-Learning Algorithm for Data Center

Abstract

In recent years, the application of green renewable energy to data centers has become an important trend. Traditional solutions lack the consideration of matching tasks to renewable energy supplies. Therefore, in the face of diverse real-time computing tasks, how to reduce the total energy cost while ensuring the quality of service is an important challenge for the data center in the future. In this paper, our focus is on using the information on renewable energy supply and task characteristics as input states to assign tasks that maximize user satisfaction while meeting the minimum total cost of energy consumption. We consider the diversity of real-time tasks and design three different task types: the most crucial task, the crucial task and the non-crucial task. According to the different characteristics of these tasks, we propose a scheduling algorithm based on multi-agent, which uses multiple sets of agents with different initial positions to parallel search in different dimensions of the parameter space to find the optimal solution. To further optimize the algorithm, we eliminate the centralized noise solution based on the Pareto sorting method and sort the multiple optimal solutions to highlight the most suitable solution. The experimental results show that the proposed algorithm compared with other algorithms can reduce the total energy consumption by 11% and increase the customer satisfaction by 13% on average, and has better performance and applicability.