Research On The Resource Scheduling Algorithms For Network Slicing In 5G

As a key technology of 5G,network slicing enables 5G networks to meet the heterogeneous needs of various services flexibly and efficiently.Under Slice-as-a-Service business model,service providers provide users with network slices meeting service level agreements for profit.Therefore,service providers enhance their profits by making full use of limited network resources.In fact,resource demand of a network slice is timevarying rather than fixed.Moreover,network slices in different scenarios are interested in distinct network performance index,asking for distinct resource configuration.Timevarying and heterogeneous network resource requirements bring great challenges to service providers for network slice resource scheduling.For the heterogeneous network slice requests,a network slice scheduling system,named SLA-NS,is proposed in this thesis.Exploiting the time-varying characteristic of resource demand,SLA-NS makes full use of network resources for improving the request access rate and maximizing service provider revenue.The main modules of SLA-NS include network slice pricing,network slice resource allocation,network slice admission control,and network slice instantiation and configuration.The corresponding algorithms are designed in this thesis for the first three modules.The network slice pricing module customizes the optimal price strategy for network slice services with different quality leaving room for dynamic resource allocation.In this module,the interaction process between service provider and users is modeled as a two-layer game model.Then,the Stackelberg strategy is calculated as the best price strategy,so as to optimize the benefits of service providers.A network slicing resource allocation algorithm based on demand prediction is designed for the network slice resource allocation module to improve the resource utilization.According to the resource requirements of network slicing services,a network slice resource demand predictor with prediction preference is designed in this thesis.The predictor exploits the Encoder-Decoder LSTM to predict the resource demand of a slice,and uses an asymmetric loss function during network model training to avoid the predicted value being lower than the demand value.The resource allocation algorithm proposed in this thesis uses the predictor to predict the resource demand of each network slice for each time slot,thereby achieving dynamic resource allocation.Two efficient network slice admission control algorithms across time windows are proposed in this thesis to improve request acceptance rate,i.e.,a network slice admission control algorithm based on reinforcement learning and a network slice admission control algorithm based on statistical multiplexing.The former uses deep reinforcement learning to realize autonomous decision-making of admission control,while the latter exploits the resource demand statistics of network slices to realize network slice overbooking by means of statistical multiplexing.Both of them can make full use of network resources with guaranteed SLA,thereby increasing the request acceptance rate and service provider revenue.The simulation experiments of the above algorithms show that each algorithm proposed in this thesis outperform the comparison algorithms in relevant indicators,so that SLA-NS improves resource utilization and service provider revenue with guaranteed SLA.