Research On Data Center Network Traffic Optimization Based On Deep Reinforcement Learning

With the rapid development of cloud computing technology,the service scale of data center is constantly expanding and the service types are increasingly diverse.In order to meet user needs and improve the Quality of Service(Qos),data center networks(DCN)based on new network architectures are widely deployed.However,there still exist many shortcomings in current center traffic optimization algorithms when dealing with massive traffic.This dissertation focuses on the following issues:(1)Parameters of DCN traffic optimization algorithm are bound to specific network scenarios,and performance declines after network characteristics change;(2)The DCN power-saving algorithms put network devices into sleep in a fixed period,resulting in network QoS declines with the neglect of network traffic changes;(3)The virtual machine placement algorithms proposed for the cloud computing scenario are not suitable for the live video streaming decoding process,which is easy to cause network congestion,resulting in the increase of video streaming delay.Utilizing artificial intelligence methods to optimize the DCN traffic has become an important research trend,which has important theoretical significance and practical application value for simplifying network management,reducing operation cost and improving network performance.In this dissertation,the deep reinforcement learning algorithm is introduced into the DCN traffic optimization process,and conducts researches around the intelligent control and scheduling strategies.The main innovations are listed as follows:The existing DCN flow scheduling algorithms are bound to specific network scenes and the application scenarios are limited.We propose a DCN QoS optimization algorithm based on deep reinforcement learning.Firstly,the network resources and state information are unified into the network model,then the fine-grained traffic changes are dynamically perceived through the short-term and long-term memory,and finally the dynamic traffic scheduling strategy satisfying QoS objectives is generated based on deep reinforcement learning.The experimental results show that,compared with the existing QoS optimization algorithms,the proposed algorithm can not only generate the traffic scheduling strategy in accordance with the network scenario,but also improve the load balancing degree by 22.7% and network throughput rate by 8.2% while ensuring the end-to-end transmission delay and packet loss rate.To improve the power efficiency,existing DCN energy-saving algorithms put too many network devices into sleep,resulting in flow missing the completion time deadlines which leads to data center profits reduction.We propose a DCN energy-saving algorithm based on deep reinforcement learning.The algorithm aims to improve the profit of the DCN.The deep neural network is introduced to perceive the network traffic changes.Under the constraint of ensuring the delay-sensitive flow completion time,flows are dynamically consolidated onto some network devices and unused network devices are put into sleep.Through training,it gradually reduces the energy consumption of DCN and improves the income of DCN.Through training,it gradually generates energy-saving strategies that conform to the network traffic characteristics.Experimental results show that compared with the existing power-saving algorithms,the proposed algorithm reduces the flow completion time of delay-sensitive flows by approximately 6.7% and increases the overall profit by approximately 17.8% at the cost of a relatively increased energy consumption of 4.3%.The existing live video streaming optimization algorithms ignores the internal transmission efficiency of DCN,resulting in increased live video streaming delay.We propose a live video steaming optimization algorithm based on deep reinforcement learning.The algorithm utilizes deep neural network to perceive the load state of the network and the physical server in real time.Through training,it gradually generates virtual machine placement strategies that conform to the live streaming scenarios,so as to avoid the hot links congestion caused by the virtual machines placed nearby.Experimental results show that compared with the existing virtual machine placement algorithm,the proposed algorithm can improve the load balance by about 12.8%,reduce the video transmission delay by about 13.2%.And by cooperating with the bit rate adaptive algorithm,it further improves the user experience quality by about 7.2%...