Research On Virtual Network Function Deployment Based On Machine Learning For Network Slicing

In order to cope with the explosive growth of data traffic and diverse service requirements in the future,5G network slicing technology emerged as the times require.It uses software defined network and network function virtualization technology to divide and reorganize limited physical resources into logically independent virtual network resources for use by various slices,thereby achieving centralized management of resources and providing better quality of service for tenants.However,due to the limited underlying resources,how to effectively and reasonably deploy the virtual network functions in the slice into the physical network,migrate and optimize the virtual network functions,and meet the corresponding quality of service(QoS)requirements,has become an urgent issue in virtualization research.This article focuses on the deployment of virtual network functions under 5G network slicing.The main contents and innovations are summarized as follows:(1)Aiming at the VNF deployment optimization problem caused by the dynamic arrival of network service requests,a virtual network function deployment algorithm based on reinforcement learning was proposed.The main work done is to establish a Markov decision model for multiobjective optimization,which jointly optimizes the total network cost and SFC end-to-end delay.The total network cost includes node operation cost,VNF deployment cost,and communication cost;Simulation results show that this algorithm can accelerate the training of neural networks,effectively meet the end-to-end delay of SFC,while reducing the deployment cost of SFC.(2)Due to the lack of effective prediction of resource requirements in existing research,resulting in poor service quality and increased migration costs,proactive prediction mechanisms are needed.Therefore,a dynamic migration algorithm for virtual network functions based on improved LSTM resource demand prediction is proposed.The algorithm first establishes a total cost model,which includes comprehensive bandwidth overhead,migration overhead,and energy consumption of the underlying physical server.LSTM is used to predict the resource requirements of SFC,backpropagation algorithm is used for network learning and training,and particle swarm optimization algorithm is applied to solve the VNF migration strategy.The simulation results show that the algorithm has higher prediction accuracy,less total system overhead,and fewer migration times in the prediction of CPU resource requirements for all VNFs in SFC.