| In recent years,due to the rapid development of digital information technology,the rising data services such as HD video and VR have seen explosive growth and also put forward the higher performance demands on mobile communication infrastructure network.Against this backdrop,the fifth generation mobile communication(5G)system came into being to solve the future massive terminal access and complex requirements of end-to-end service quality.To achieve these goals,5G makes many changes to the carrier network to support the broader application scenarios and higher communication standards.The data center interconnection(DCI)as the core in the carrier network becomes the key to provide cloud computing and edge computing services to support the new 5G data services.However,the traditional way to construct a DCI network for 5G applications is facing severe challenges.To meet the end-to-end performance requirements of future 5G application,the DCI network needs new architecture and technologies,including super-large transmission capacity,centralized control plane and intelligent control algorithm.In response to the above challenges,this paper explores the architecture and key technologies of the new DCI network for 5G applications.Based on the newly-developed optical network technologies,this paper proposes the architecture of cognitive data center interconnection network(CDCIN),and adopts the following technologies: 1)construct the underlying physical network by using space division multiplexing(SDM)transmission to improve the network capacity;2)use the software-defined optical network(SDON)architecture with the centralized control plane to support high quality and efficient network slicing;3)introduce cognitive plane and intelligent algorithm to improve the ability of adaptive processing;4)in order to verify the technicals proposed in this paper,we designed the physical topology structure of CDCIN based on SDM optical transmission,and conducted a lot of simulations.Because of the different intrinsic mechanism of SDM fiber,there are different SDM transmission paradigms.According to the coupling effect in SDM fiber,the SDM transmission can be divided into three different paradigms: independent transmission(InT),joint transmission(JoT)and fractional joint transmission(FJoT).In the process of data transmission,each transmission paradigm has different resource allocation.In InT,each transmission element can be treated as an individual channel for spatial resource allocation,and packets can be allocated to different channels for transmission.In JoT,al the spatial elements are coupled through the single fiber link so that all the spatial elements need to be treated as a single entity at the reception.In FJoT,The spatial elements are divided into several non-overlapping,independently-transmitted subgroups.Each subgroup is transmitted jointly and considered as the JoT.These SDM transmission paradigms bring different network performance,including throughput,packet drop ratio,and average transmission delay.The DCIN based SDM enabaled networking should consider the performance of each paradigm.In this paper,we build the mathematical model of different SDM paradigms and use queuing theory and markov chain to deduce the network performance under stable state.Simulation results show that 1)InT has the highest transmission efficiency with the same spatial resource;2)JoT can provide the highest transmission rate but can not guarantee the lowest average transmission delay;3)In the case of different spatial resource,FJoT can provide the best transmission performance with finer transmission granularity;4)the size of the average optical packet has an important impact on the network performance,and SDM transmission is more conducive to the construction of an all-optical network with flexible transmission granularity.The above results reveal the capability and efficiency of different transmission paradigms in SDM enabled optical networks,providing an insightful guideline for design of such networks.The traditional control layer adopts the link-state routing protocols which allocate the same kind of business to the fixed transmission path.This routing algorithm does not take into account the change of load in the transmission path.When the load on this shortest path is heavy,a large number of packets are discarded which results in the high packet drop ratio,so the tranditonal way can not guarantee the end-to-end service quality of 5G business.Based on this,an adaptive routing algorithm based on deep learning is proposed.The cognitive plane can obtain the load data of network nodes.We use deep neural network(DNN)to learn the data samples of packet drop ratio in different cases of load distribution.After enough training,the DNN can output the optimal transmission path corresponding to the current network traffic distribution,so as to realize the dynamic adjustment of the transmission path and thus avoid network congestion.Simulations verify the effectiveness of the deep learning based adaptive routing algorithm.Compared with traditional routing protocols,the deep learning based adaptive routing algorithm shows better packet drop ratio performance and congestion avoidance ability.Network slicing is an important guarantee for end-to-end transmission delay of different 5G business.The current 5G network slicing adopts the scheme of dividing the channel bandwidth in the physical layer,which can quickly realize the carrier isolation of the 5G business,but does not necessarily guarantee the end-to-end transmission delay.This is because the slicing scheme in the physical layer does not take into account the change of traffic load in the network layer.When the load in the slicing network is serious,it is difficult for the network to guarantee the transmission delay of delay-sensitive services.Therefore,a dynamic slicing scheme in network layer is proposed.The cognitive plane is used to obtain the load data of network nodes,and a slicing scheme suitable for the current network state is obtained through the combination training of multiple DNNs.Simulation results show that the dynamic slicing algorithm in the network layer has better transmission delay performance than the traditional slicing scheme in physical layer. |
PDF Full Text Request