VNF Deployment And Dynamic Scaling Based On Cost And Demand Awareness

Network Function Virtualization(NFV)is an important change in telecommunication service provisioning.It enables the decoupling of network element functions and dedicated hardware devices.How to economically place and chain Virtual Network Functions(VNFs)according to the requirements of Service Functions Chains(SFCs)is one of the challenges for NFV orchestration.In addition,when VNF is placed,how to dynamically adjust the VNF instances to adapt to the changing network traffic and deal with the burst network peak is another challenge of current NFV orchestration.At present,the researches of VNF deployment mainly focuses on the optimization of deployment cost.Most of the research focuses on optimizing the number of VNF instances to reduce the instance cost,but it will lead to the growth of forwarding path and increase the forwarding bandwidth cost.How to balance these two costs is a problem to be solved.In terms of virtual resources dynamic scaling,the current research is mainly based on the threshold of passive scaling.In this way,the response lag and oscillation of scaling are often caused.In view of the above background and problems,this thesis proposes VNF deployment algorithm based on cost-aware and instance sharing,and virtual resource dynamic scaling algorithm based on LSTM and Priority-Deep Q-learning.The main work is as follows:(1)Proposing a VNF deployment algorithm based on cost-aware and instance sharing.The algorithm considers the deployment of VNF from the perspective of sharing VNF instances,so as to improve resource utilization and reduce the cost of total deployment.Firstly,this thesis proposes a mixed integer linear programming(MILP)model for VNF deployment.And our model can be configured dynamically according to different deployment preferences.Then,a heuristic algorithm is proposed.The algorithm sorts the physical nodes in the network topology through the centrality of graph theory,and selects the candidate deployment nodes of these shared VNF instances."Then,the algorithm selects an optimal deployment path to place and link VNF instances through Markov decision process,so as to ensure the optimization of total deployment cost.The results show that the difference of the deployment costs between our algorithm and the optimal solution is less than 3%.However,the execution time can be significantly reduced by 57.4%.(2)Proposing a virtual resource dynamic scaling algorithm based on LSTM and Priority-Deep Q-learning.In this thesis,a dynamic adaptive method of virtual resources is proposed,which combines network traffic prediction with resource capacity expansion and reduction decision.The algorithm uses long short-term memory(LSTM)neural network to analyze the network traffic in the previous hour and predict the network traffic in the next 10 minutes,so that the scaling decision algorithm can respond in advance to deal with the arrival of traffic peak.Priority deep Q-learning(Priority-DQN)is used in the decision-making process.Through the continuous interaction and feedback with the environment,Priority-DQN algorithm can make the correct scaling decision with the changing trend of traffic,and optimize the resource allocation before the peak traffic,so as to ensure the quality of service.Through simulation evaluation,the decision-making accuracy of this scaling algorithm can be stabilized above 92%.To sum up,the cost and demand aware VNF deployment algorithm and the dynamic scaling method proposed in this thesis can effectively optimize the total deployment cost in the deployment process,and can continuously learn online in the dynamic scaling process to adapt to the changing network traffic mode.