Machine Learning Based Traffic Prediction And Load-balancing For Software Defined Network

With the rapid development of communication networks,the number of users and users' demands for traffic has increased dramatically.However,many problems such as complicated network deployment,vertical and closed network architecture and mismatch of total telecommunication revenues have posed great challenges to today's networks.The appearance and development of SDN(Software Defined Network)architecture have greatly promoted the evolution of communication network architecture.In SDN-based networks,the control plane and data plane are separated and no longer tightly coupled models,making the network more flexible and scalable.Distributed SDN architecture is a physically distributed and logically centralized implementation of SDN architecture.As an important architecture model of SDN paradigm,distributed SDN,which has received extensive attention from the research and industrial communities,can be applied to large-scale networks and has more flexible and stable control functions.However,distributed SDN can easily lead to load balancing problem.Load sharing difference easily occurs among multiple controllers,causing the significant performance degradation of individual high-load controllers.Meanwhile,the problem of unbalanced load among links tends to occur in a single SDN control domain.Load balancing between the links will affect the performance of the entire network.In particular,the end-equipment requirements on the network today,real network traffic inherently may drastically fluctuate in both space and time dimensions.In order to avoid network concussion and enhance the predictability of network decision-making,this paper first uses SVR to forecast the network traffic,and then based on the prediction,this paper studies the load balancing of distributed SDN network in the control plane and the data plane.The focus of this article is traffic prediction and load balancing in distributed SDN architecture based on machine learning.In mobile communication networks,traffic is a time-varying variable.In order to accommodate traffic surges and avoid congestion,existing networks retain a large amount of redundant link capacity.With the increase of network video streaming services in the future,the range of traffic sequence function will gradually increase.This method of link capacity redundancy will result in low network resource utilization.In response to this problem,we propose a predictive model to analyze the network characteristics,pre-configure the network in advance,and reduce unnecessary network concussion and resource waste.We collected one month's real-time data on the core network switch,preprocessed the collected raw data by using entropy analysis method and found the valid data related to traffic prediction from the complicated and redundant data.In our work,our dataset is nonlinear and small sample data.Therefore,we choose SVR algorithm as the core of the algorithm and use the improved Online-multiKernel SVR(OKSVR)analysis method to predict the link traffic in the network.Next,we discuss the problem of data plane load balancing in a distributed SDN architecture.In the SDN core network,the flows with the same source border router are converged into the same aggregate flow.Therefore,the aggregate flow generally enters core network from one border router in the network and from another border router out of the core network.This network-mode easily leads to network link load imbalance,thus affecting the performance of network transmission.To solve the above problems,we aim to minimize the maximum link utilization in the network and build a mathematical model based on multipath routing technology under the premise of QoS.We propose an auction-oriented multi-path load balancing algorithm Path routing strategy named MPLP(Multi-Path Load balance Policy),and the model was verified by simulation.The results show that the load balancing strategy can minimize the maximum link utilization in the network and guarantee the intra-domain delay while fully utilizing the network link resources.Last,we study the problem of load balancing on distributed SDN control plane.In order to effectively use network control resources and achieve network flexibility,SDN switch migration research is necessary.In this paper,aiming at the problem of relocation and reconfiguration of SDN switches,we takes the network time-varying,relocation overhead and load balancing into account,establishes a mathematical model and proposes a Dynamic Switch Migration Policy(DSMP)for switch load balancing.Through simulation and verification,we verify that DSMP improves the balance of the networks and guarantees the stability of the network,compared with the greed-based migration scheme and the stochastic migration scheme.This thesis summarizes the distributed SDN technology,and designs the load balancing model of data plane and control plane in distributed SDN networks.This provides a new idea for the load balancing in distributed SDN networks.