Research On Data Center Optical Interconnection Topology Based On Deep Reinforcement Learning

With the continuous increase in data traffic demand and the continuous in-crease in processing speed and link bandwidth requirements of next-generation high-performance data centers,traditional electrical switching networks can no longer efficiently carry the demand for high-speed data interaction.At the same time,with the rapid development of cloud computing,Internet of Things,and streaming media industries,the types of applications deployed in data centers are becoming increasingly diversified,and their traffic distribution characteris-tics vary greatly.Traditional data centers are based on fixed network connection structures and The routing strategy has poor adaptability and flexibility and cannot guarantee network performance.In recent years,in terms of switching technology,thanks to the transparen-cy of optical switching to link rate and data,the interconnection architecture based on optical-electric hybrid switching technology has broken the limitations of traditional electrical switching networks in terms of bandwidth and power consumption.Avoid the fast control and conflict resolution problems of all-optical switching.In terms of network control,technologies such as soft-ware-defined networking,deep reinforcement learning,and intention-driven networking have made considerable progress,enabling a unified intelligent de-cision-making controller to realize the underlying electrical and optical switch-ing equipment in the optical and electrical hybrid interconnection data center.The integrated management and control of the network becomes possible,and then the accurate scheduling of traffic is realized.Based on the above analysis,the photoelectric hybrid networking technol-ogy has flexible link connection characteristics,which provides the possibility for network topology reconstruction to adapt to various traffic distributions gen-erated by different applications and services.Therefore,this article aims at the poor adaptability of traditional data center electrical interconnection networks in dealing with business dynamic traffic,and optimizes the reconstruction of data center services from two aspects:routing optimization and topology optimiza-tion.In terms of optimization strategy,this paper uses the DDPG deep rein-forcement learning algorithm,through the iterative interaction with the OM-NeT++-based network simulation system and the deep reinforcement learning model,to let the model learn complex task control strategies,and realize the re-lationship between the topology structure and the business traffic distribution.Continuous training.At the same time,in order to realize the automatic control of network reconfiguration,this paper proposes a new network architecture,which uses the SDN controller to plan the network based on the optoelectronic hybrid network.The deep reinforcement learning model is based on the services collected by the SDN controller in real time.Traffic distribution information re-alizes automatic optimization and reconstruction of network topology,thereby improving network performance.The experimental results show that,for a given traffic intensity,the model trained by deep reinforcement learning can signifi-cantly reduce the average network delay and packet loss rate compared with the untrained model.