| Network virtualization can abstract and encapsulate network resources to realize the construction of network services with specific functions and the on-demand deployment and flexible deployment of resources among the cloud,network,and edge.Virtual network embedding technology effectively deploys several virtual network services with different structures to the underlying network.The emergence of new applications and new requirements has brought new scenarios for virtual network embedding.If the virtual network embedding scheme is formulated and deployed indiscriminately,it will reduce the quality of service.For example,the network scenario in the peak period caused by the large-scale increase of user demand,compared with the low-peak situation,it pays more attention to the stability of communication and ensures the reception of requests.The emerging application scenarios with less end system resources,such as Internet of vehicles and visual detection,have high requirements for rapid embedding and deployment of resources.Based on this,this thesis studies the virtual network embedding algorithm for these two representative network scenarios.The main work is as follows:1)In the peak network scenario,aiming at the unreasonable resource allocation caused by the insufficient consideration of node location in the existing virtual network mapping algorithms,a two-stage group teaching virtual network mapping algorithm based on weighted K-shell decomposition is proposed.According to the weighted kshell decomposition method,the underlying network is preprocessed,and then the links are embedded along the shortest path between nodes.In combination with the intelligent optimization strategies of grouping,teaching,self-learning and mutual learning of the grouping teaching model,the coordinated mapping of nodes and links is realized,so as to further improve the quality of the solution.The simulation results show that the algorithm can comprehensively consider the location level of nodes in the network,and reduce the unnecessary overhead of link resources,so as to ensure the rational use of resources in peak network scenarios,and improve the request acceptance rate and revenue to cost ratio of virtual network.2)In the network scenario with insufficient end resources and high requirements for fast embedding,aiming at the problem that the existing virtual network mapping algorithm takes too long in the process of learning and exploration,a NoisyNet-DQN virtual network mapping algorithm based on end-cloud collaboration is proposed.Using the idea of "end-cloud collaboration",the model training for virtual network embedding is carried out in the cloud,and then the trained model is delivered to the end-side.When the online virtual network request from the end-side arrives,it can be inferred and applied directly based on the trained model.The NoisyNet-DQN technology is selected as the deep reinforcement learning training model,and five network attributes containing the global topology information of the network are extracted as the input of the model to improve the learning and exploration efficiency of neural network.The simulation results show that the end-cloud collaboration idea of the algorithm can effectively shorten the running time of processing virtual network requests,and ensure the quality of the embedding strategy to a certain extent. |
