Energy Consumption Management System Of Data Center Network Based On Deep Reinforcement Learning

Data center networks usually use rich link topology and multi-path routing to ensure the business stability in peak traffic.Research shows that the load of data center network usually accounts for 5%-25%of the actual network energy consumption.This design makes the data center waste too much energy in the case of less traffic.Deep reinforcement learning is an optimal control method to find strategy model.It optimizes the agent strategy by continuously interacting with the environment and obtaining action feedback to achieve autonomous learning of the algorithm and ultimately maximize the benefits.In recent years,as a powerful development direction,deep reinforcement learning has made great achievements in robotics,games and other fields.At the same time,with the development of next generation network and the maturity of data plane and control plane,control plane can control data plane and collect data plane information in real time,which meets the closed loop that can obtain timely feedback required by deep reinforcement learning.This inspired us whether we can combine deep reinforcement learning with data center network energy saving.Therefore,we propose to apply deep reinforcement learning to network closed-loop system to realize network energy saving.The main idea is that through training,the deep reinforcement learning algorithm can sleep some links and switches without causing network congestion in the data center,so as to achieve energy saving.The thesis builds a closed-loop system of deep reinforcement learning and data center network,and achieves training convergence under the two architectures of Spine-Leaf and Fat-Tree and different load conditions,which confirms the feasibility of deep reinforcement learning for network energy saving.According to the energy-saving calculation formula given by us,In the case of no burst traffic,it can save energy by 34.4%under the condition of low link load;under high load conditions,it can save energy by 17.2%;it can save energy of 17.2%?34.4%under the traffic model of day and night mode(that is,traffic peak during the day and traffic decline at night).Fat-Tree architecture can save energy by 36%under low load.