Research On SDN Migration Strategy From A Dynamic And Global Perspective

Software Defined Network(SDN)is the key technology of 5G network and one of the most potential next-generation network solutions due to its separation of control and forwarding and network programmability.The research popularity is growing,and more and more companies choose to research and implement SDN in backbone networks,bearer networks,and access networks.This paper focuses on SDN solutions for backbone networks.However,the construction of SDN is not a revolutionary process,but an evolutionary process from the traditional IP network to an SDN network,which requires multiple steps.The entire dynamic upgrade process needs to solve three proble ms(3W):which switches to upgrade(Which),the switch upgrade sequence(When)and how to deploy the controller(ho W),to complete the upgrade process from the traditional IP network to the hybrid SDN,and then to the pure SDN network.The optimal dynamic upgrade strategy not only needs to ensure the network programmability benefits brought by the upgraded switches and the network performance brought by the deployed controllers in each step,called the local dual-objective,but also needs to ensure the performance of the pure SDN which is continuous operated after the upgrade,called the global single objective.However,the local goal of each step and the final global goal have complex effects in both the positive and negative directions of the time sequence,which makes it difficult to obtain the optimal dynamic upgrade strategy by considering the above three aspects(3W)at the same time.This paper discovers and studies the above-mentioned optimal dynamic upgrade problem.The first innovative work of the paper is to consider the 3W problem at the same time,and model the optimal dynamic upgrade problem as a dual-objective dynamic optimization model.During the entire upgrade sequence(when),simultaneously optimize the benefits of switches upgrade switch(which)and controller deployment(how).By introducing a degradation factor that changes with time,the transformation process from a local dual-objective to a global single objective is described.Different from the existing dual-objective optimization model for joint deployment of switches and controllers,the dual-objective proposed in this paper is dynamic.Different from the classic dynamic optimization model,this optimization model is dynamic in both directions of the timing sequence.The second innovative work of the paper is to transform the dual-objective dynamic optimization model into a series of single-objective integer linear programming proble ms which can be solved directly by introducing a reverse penalty term to consider both local and global benefits during single-step upgrades.Then a heuristic algorithm is proposed.By improving the genetic algorithm,a genetic algorithm based on SDN matching relationship(SM-GA)and a genetic algorithm based on complementary crossover mutation(CCM-GA)are obtained to solve the penalty term and the optimal upgrade strategy respectively.The third innovative work is to regard the upgrade process as a Markov decision process and propose a solution algorithm based on reinforcement learning.The algorithm automatically learns the impact of global returns on single-step local returns through learning,instead of defining through experience like heuristic algorith ms,and empirical definitions may not be able to obtain the optimal value.The third innovative work is to regard the upgrade process as a Markov decision process and propose a solution algorithm based on reinforcement learning.Compared with the heuristic algorithm,the solution algorithm based on reinforcement learning does not need to define the impact of the global benefits on the local decision,but learns the optimal value automatically through learning,while the empirical definitions may not be able to obtain the optimal value.The paper uses real network topology data to conduct simulation experiments on the above proble ms and algorith ms,and compares them with the baselines.The results show that the solution algorith ms proposed in this paper better achieve the balance between the local and the global goal.Besides,the pure SDN obtained by the upgrade has better performance.In the large-scale real network topology,the solution algorithm based on deep reinforcement learning can obtain better global benefits than the heuristic algorithm based on the improved genetic algorithm.