Research On Software-Defined Networking Routing Optimization For Data Center

A data center is a complex system interconnecting hardware resources(large servers,storage devices,and network facilities)through a dedicated topology network.As the basic information infrastructure of countries and enterprises,data centers play an indispensable role in the economic,scientific,and technological fields,as well as in daily life.Data Center Network(DCN)plays a crucial role in the design of data centers.Many applications hosted in data centers cause intensive data interaction between servers,resulting in frequent traffic bursts.Due to the tight coupling between the control plane and data plane of traditional network architectures,routing algorithms cannot achieve network state information collection and forwarding path planning through a global perspective,resulting in low network utilization and being prone to severe congestion and delay.Therefore,traditional network architectures and routing algorithms will be challenging to meet the growing high-speed transmission requirements of various applications in data centers and become a bottleneck in developing data center networks.Software Defined Networking(SDN)is a new network architecture that has emerged recently.The SDN decouples the data plane from the control plane and achieves the collection of switch state information and data flow forwarding planning through the controller,providing the primary conditions for the deployment of artificial intelligence algorithms in the data center network.This thesis researches the application of artificial intelligence algorithms in the DCN routing optimization from generalization and convergence performance optimization.This thesis aims to use improved artificial intelligence algorithms to optimize network routing,solve network link congestion problems,improve network link utilization,and improve the quality of service(QoS)of DCN networks.The research content of this thesis is as follows:(1)In the electric domain network,there is a time difference between the network link information obtained by the SDN controller and real-time information,which results in the flow table being unable to utilize network resources effectively and may exacerbate network congestion under high load conditions.A routing optimization algorithm based on link-state prediction is proposed to address the above issues.Firstly,a prediction model of network link state value based on the Long Short Term Memory(LSTM)algorithm is designed.This prediction model obtains information on network link status by inputting historical observations of network links,including transmission delays between controllers and switches.Subsequently,the link prediction value is used as the Dijkstra weight to calculate the optimal path between hosts.Finally,an evaluation model for network link load balancing was constructed,and the effectiveness of the proposed algorithm was verified by simulating a GEANT network to collect a dataset of network link states.The results show that the algorithm proposed in this thesis eliminates the impact of transmission delay between controllers and switches on link load balancing and improves network link utilization in high-load network environments.Compared with the Open Shortest Path First(OSPF)algorithm and Q-learning-based reinforcement learning method,the algorithm has improved link load balancing performance by 23.7% and 11.7%,respectively.(2)In the above work,the LSTM algorithm requires a large amount of data sets for model training.Due to its inability to interact with the environment,its generalization ability is limited,making it challenging to optimize the performance of dynamic networks.To address the above issues,a routing decision model based on the Deep Reinforcement Learning(DRL)is designed to optimize the performance of SDN network link load balancing and Transmission delay.Meanwhile,in response to the problems of slow convergence speed and instability after convergence in traditional deep reinforcement learning algorithms,an improved deep deterministic policy gradient(DDPG)algorithm based on multi-step and priority experiential replay is proposed.Firstly,we use a multi-step approach to address the overestimation and underestimation issues of the DDPG algorithm.The multi-step method uses the maximum value of the n-step effect currently estimated by the neural network,reducing the possibility of positive errors in the Q-value function and effectively improving the convergence stability of the reinforcement learning model.Secondly,priority empirical sampling based on the Sum Tree binary tree is adopted to improve the convergence speed of the DDPG algorithm.Finally,the effectiveness of the proposed algorithm is verified by simulating dynamic network environments with different traffic intensities in the NSFNET topology.The experimental results show that the algorithm proposed in this paper can effectively improve the throughput of dynamic networks and reduce link transmission delay.Compared to the baseline algorithm,the performance is improved by more than 5%.Meanwhile,our algorithm significantly improves convergence speed and stability after convergence compared with existing DRL algorithms.(3)In the above work,the neural network of the DDPG algorithm cannot capture the global spatio-temporal correlation information of graph structure data.It results in limited convergence performance in network routing optimization applications.In response to the above issues,this thesis proposes a proximal strategy optimization algorithm based on attention mechanism and spatio-temporal correlation to improve the convergence performance of reinforcement learning models.Firstly,based on the Proximal Policy Optimization(PPO)algorithm,Gate Recurrent Unit(GRU)and Graph Attention Network(GAT)are used to extract temporal and spatial correlation features from state information,providing implicit information containing more environments for reinforcement learning decisions.Secondly,using jump connection methods to connect implicit and directly related information to multi-layer perceptrons improves the model’s learning efficiency and perceptual ability.Finally,static and dynamic traffic experiments verify the effectiveness of the proposed algorithm in different network topologies.Benefiting spatio-temporal correlation learning with a global perspective,the proposed algorithm performs better load balancing and congestion control than other reinforcement learning baseline algorithms under different traffic intensities and network topologies.Experimental results show that in static and dynamic traffic scenarios,the convergence performance of the proposed algorithm is improved by 9.02% and 15.07% compared to the baseline algorithm,respectively.(4)The routing method provides end-to-end services in optical domain networks through wavelength-selective routing.Routing algorithms based on electric domain networks cannot offer differentiated services to optical network scenarios,nor can they effectively handle the rapid changes in topology and link characteristics in optical network environments.Aiming at the transmission characteristics of optical network traffic in data center networks,a Deep Q-Network(DQN)algorithm based on a multi view attention mechanism was proposed to optimize the routing decision in optical transmission networks.Firstly,aiming at the transmission characteristics of optical domain network traffic,a DRL-based routing decision model is designed to find the optimal routing strategy for each source-destination traffic demand.Secondly,considering the sparse connection characteristics of optical network nodes,a multi-view attention network is used as a neural network model for the DQN.By calculating the attention weights of adjacent edges,reinforcement learning agents can consciously aggregate important network information and improve the model’s generalization ability.At the same time,the multi-view method is combined to improve the convergence speed and stability of the graph attention network model.Finally,a simulation routing experiment is designed based on Gym,and the algorithm’s load balancing and generalization capabilities are evaluated in different network topologies.Experimental results show that the algorithm proposed in this paper has good convergence performance and load balancing ability in routing optimization of optical transmission networks and can generalize new network structures,maintaining good decision-making ability even when some network nodes fail.