| The rapid development of Internet technology contributes to the diversification of network application scenarios and users’ service requests.The requirements for various types of customised network services are different,including the functional requirements for on-demand combinations of firewalls,load balancers and intrusion detection systems,as well as the quality of service requirements such as network node resources,link bandwidth and so on.Constructing network services into the customizable Service Function Chains(SFCs),and effectively orchestrating the SFC can reduce network operating costs and improve the flexibility of network deployment.However,the resources of computing,storage and network are complex and heterogeneous,which poses challenges to the on-demand resource supply and flexible networking of customized SFC.In terms of network architecture,considering the original “triple binding”design of traditional networks and the coupling of service functions with hardware devices,there are various drawbacks here such as static and inflexible design with high cost of deployment.In terms of resource adaptation mechanism,the non-intelligent SFC resource adaptation algorithms have problems such as high complexity,single optimization objective,poor scalability,and so on,which results in them showing varying different degrees of performance degradation.The "three layers and three domains" architecture of Smart Integration Identifier Networking(SINET-I)vertically associates service layer and network layer of the existing Internet and creatively introduces the knowledge space.It provides differentiated and fine-grained network services for different business scenarios through intelligent scheduling strategies,and realizes the collaborative mobilization and intelligent integration of the service space and the network space with their own resources,in order to facilitate dynamic on-demand adjustment of service and network.To this end,based on the SINET-I resource adaptation mechanism,this dissertation constructs the research on key technologies for SFC-oriented resource intelligent adaptation in SINET-I,and carries out systematic research along the three steps of "Orchestration-Optimization-Elastic control".By using a series of research methods such as technical research,principle mechanism design,problem modeling,mathematical derivation/AI learning,simulation and experimental evaluation,the goal is to build a more intelligent,efficient,flexible and accurate adaptation mechanism as well as to comprehensively schedule network resources based on the entire network capabilities.Thus,complex network resources can be effectively managed and utilized.The main works and contributions are as follows.(1)To address the Virtualized Network Function(VNF)instantiation dilemma in SINET-I,an SFC orchestration mechanism based on hierarchical clustering is proposed.Firstly,the SFC orchestration problem is systematically modeled to form a multiobjective joint optimization problem that maximizes resource utilization and SFC flow admitted rate,and minimizes the number of VNF instances and end-to-end delay of SFC flows.Secondly,SFC orchestration mechanism based on hierarchical clustering is designed.At the network level,VNFs are clustered according to the correlation between VNF pairs,and VNFs belonging to the same cluster are deployed on the same network node as much as possible to reduce link bandwidth consumption.At the SFC flow level,SFC flows are clustered according to their end-to-end delay performance,and SFC flows belonging to the same cluster share VNF instances,which aims to avoid path stretch and improve the resource utilization of VNF instances.The designed mechanism solves the problems of high algorithm complexity and single optimization objective existing in the non-intelligent resource adaptation mechanism.Finally,the performance evaluation results verify the feasibility and effectiveness of the designed orchestration mechanism.(2)To address the joint orchestration of SFC deployment and routing in SINET-I,an SFC orchestration mechanism based on Deep Reinforcement Learning(DRL)framework is proposed.Firstly,the SFC orchestration problem is modeled as a Markov Decision Process(MDP)to capture dynamic network state transitions.Secondly,user-level Quality of Experience and network-level Quality of Service(Qo S)are factored into the formulaic reward function,and then an SFC orchestration algorithm based on long-term system performance profits is designed,which combines the perception ability of deep learning and the decision-making ability of RL and jointly optimizes the virtual node mapping problem at the network level and the virtual edge mapping problem at the SFC flow level.The designed mechanism solves the problem that the non-intelligent resource adaptation mechanism relies heavily on prior knowledge of network configuration and has poor scalability.Finally,the performance evaluation results verify the designed orchestration mechanism can better balance the performance in terms of SFC flow admitted rate,resource utilization,and end-to-end delay,etc.(3)To address the optimization problem of SFC traffic scheduling in SINET-I,a multi-path routing-based SFC optimization orchestration mechanism is proposed.Firstly,SFC orchestration problem is modeled as a stochastic optimization problem,and Lyapunov optimization theory is used to decouple the time coupling of optimal decisionmaking,and then a distributed SFC orchestration algorithm is proposed to improve the scalability of the algorithm.Furthermore,in order to respond to the degradation of network performance caused by dynamic changes in requirements during the orchestration process,and optimize traffic scheduling based on the initial orchestration results,a multi-path load balancing optimization algorithm based on DRL framework is designed.The algorithm adaptively allocates network traffic to multiple paths for parallel transmission according to the current network state.The designed optimization mechanism solves the problem of lower network availability caused by the single-pathbased transmission mechanism.Finally,the new type of routing and switching prototype system based on the SINET-I architecture is used for experimental evaluation.The performance evaluation results verify that the proposed algorithm has good performance in terms of throughput,resource utilization,queue depth and round-trip time,etc.(4)To address the elastic control problem of SFC flows in SINET-I,a virtual network resource elastic control based on DRL framework is proposed.Firstly,the virtual resource elastic control problem is modeled as an MDP model.Secondly,the elastic control mechanism based on single-agent DRL framework is designed to achieve the elastic optimization under the premise of guaranteeing fault-tolerant.Further,to avoid the singlepoint of failure problem of the controller,an elastic control mechanism based on multiagent DRL framework is designed to deliver heavy training tasks from the control plane to the data plane.The designed elastic control mechanism can meet the Qo S requirements of time-varying traffic with low network cost in the long-term system run.Finally,the performance evaluation results verify the performance of the proposed algorithms in terms of energy consumption,migration cost,system revenue and load balancing,etc.Meanwhile,the feasibility and effectiveness of the proposed algorithm are verified in the new type of routing and switching prototype system based on the SINET-I architecture. |
PDF Full Text Request