Research On Virtual Network Slicing Based On Deep Learning And Distributed Computing

Network virtualization enables traditional Internet service providers(ISPs)to evolve into two types of entities,i.e.,infrastructure providers(InPs)and service providers(SPs),which effectively alleviates the structural rigidity of datacenter interconnections(DCIs).With this technology,InP can dynamically and adaptively provide SPs(i.e.,tenants)with IT resources in the datacenters(DCs)and bandwidth resources in the inter-DC network,thus meeting time-varying requirements of heterogeneous network services economically and efficiently.However,as a key technology to realize network virtualization,virtual network embedding(VNE)has been proven NP-hard,so how to effectively serve VNT requests of tenants is still a great challenge.Currently,most of the investigations try to solve VNE in the centralized manner,which leads to long computation time,or even the optimal solution cannot be obtained in large-scale scenarios.In addition,the existing studies only build VNE schemes based on the current network state,which are not forward-looking.Therefore,this dissertation revisits the problem of VNE in DCIs and realizes the distributed VNT slicing service framework combined with deep learning.The main idea is to get tenants involved in VNE calculation.Specifically,InP leverages deep reinforcement learning(DRL)to price and advertise the substrate resources,motivating tenants to request resources in a load-balanced manner.Meanwhile,the tenant’s task is to calculate their VNE schemes independently and distributedly according to the information advertised by InP.The proposed distributed computing improves the scalability of the VNE algorithm,while DRL focuses on the impact of current decisions on the future state of the network,making the algorithm more forward-looking.So as to ensure that the service framework achieves higher efficiency and scalability on the promise of comparable performance as the centralized approach.The main contributions of this dissertation are summarized as follows.1.The scheduling process of resources between InP and tenants is modeled as a sequential decision problem,and a simple deterministic pricing function is designed to assist InP and determine the strategy.Then,this dissertation studies how to resolve resource conflicts among the distributedly-calculated VNE schemes,builds a conflict graph(CG)to transform the conflict-free VNE selection into finding the maximum weighted independent set(MWIS)in the CG,and designs a polynomial-time approximation algorithm to solve the problem under certain constraints.Approximation analysis and simulation results reveal that the proposed algorithm has significant advantage on time-efficiency.2.To address the resource limitations of deterministic pricing function,this dissertation designs the resource pricing and management modules based on deep deterministic policy gradient(DDPG),so that VNT requests are evenly distributed in the substrate network.Thus,the blocking probability is reduced and the network utility is improved.Meanwhile,an auto-encoder(AE)based compressor is developed to extract the feature of VNT requests,which generalizes the DRL’s state and enhances its adaptability.Extensive simulation results demonstrate that compared with the centralized service framework relying solely on InP for VNE calculation,the service framework based on DRL and distributed computing provides significantly shorter computation time and comparable blocking performance for VNT requests,as well as better scalability and stability.