| With the gradual expansion of the network scale,the proportion of Software-Defined Networking(SDN)in the network is also increasing,and failure recovery,as an important guarantee of high reliability of SDN,has been a popular area of research.At this stage,SDN failure recovery methods can be divided into two main categories,one is the proactive method,which directly completes failure recovery in the data plane by configuring alternative paths in advance,as the recovery process does not require the participation of the control plane,the recovery speed is fast.The other is a reactive approach,which uses the control plane to calculate new available paths in conjunction with the network state at the time of failure,then guiding the traffic to complete the recovery,which is slower but has a better network state after recovery.The main research and contributions of this paper are as follows:、1.Existing research on proactive failure recovery methods focuses on how alternative paths are generated,while there is no effective research on alternative path selection methods at the time of failure.To address this problem,this paper uses reinforcement learning combined with the traffic engineering characteristics of the network to decide the alternative path priority for each flow,and switch the available alternative paths according to the priority when the flow encounters a failure,so the proactive approach is network-state“aware” even if it is done entirely in the data plane.This results in a greater recovery success rate and improved recovery quality compared to existing proactive failure recovery methods,while this paper also designs a holistic combined proactive and reactive SDN failure recovery method with the help of reinforcement learning to determine the intervention timing of reactive methods.2.In a large network,the decision of alternative path priorities for each traffic flow to improve failure recovery can suffer from scalability problems due to the large decision space and over-occupation of forwarding table entries.To address these two problems,this paper uses the idea of partitioning to divide a large network into several parts by clustering,which is carefully designed so that each cluster only needs to decide on alternative path priorities within itself,thus reducing the decision space and allowing the processing of different clusters to be executed in parallel,solving the problem of excessive complexity of the algorithm.The clustering approach also reduces the size of the forwarding table entries by merging inter-cluster paths.3.For scenarios where too many nodes and traffic flows can only be recovered by reactive failure recovery methods,the problem is that the recovery time is too long,resulting in the degradation of network service quality.This paper addresses this problem by introducing the concept of “horizontal slicing” where resources are isolated from each other,enabling the controller to calculate paths in parallel,thus solving the problem of slow recovery by reactive methods.The introduction of “horizontal slice” requires the selection of appropriate slices for traffic flows and suffers from link capacity “fragmentation”,both of which are addressed by combining reinforcement learning and dynamic slice resource adaptation.4.Finally,detailed experiments are conducted on real network topologies and different traffic characteristics for all our proposed failure recovery methods,and current mainstream methods are compared.The experimental results show that methods proposed in this paper have excellent results. |
