An Automated Network Slicing at Edge with Software Defined Networking and Network Function Virtualization: A Federated Learning Approach

Network slicing allows heterogeneous applications can be launched across different domain using virtualized resources. The virtualized resources are created on the physical infrastructure. The orchestrator is essential for coordination of network slice management. The overhead in the slice orchestrator is reduced by distributed approach. The slice template act as a design specification template for the creation of network slices. This template can be predicted using federated learning, in which local models are trained with the generated data and global model is trained with local model parameters. Then the global parameters are updated in the local model for further learning. The federated model uses the Software Defined Networking (SDN) capability to learn the local model data distribution and hence enhance global SDN federated controller prediction accuracy for network slices. This process can be automated with the help of slice template with the predicted pattern. The parameters of the slice template are directly proportional to the performance of the slice orchestrator and prediction of future slice demands. The edge devices with its local model communicate with the global SDN federated model to satisfy the requirement of dynamic network slicing. The request on-demand services can be provided as virtual network function using network function virtualization.The optimal resource allocation for the requested slice can be done with statistical modeling of observed traffic and autoscaling can be carried out. Experimental studies reveal that the proposed network slicing with federated approach minimal response time with maximal orchestrator scalability.

This rest of the article is organized as follows: Sect. 2 explores the related literatures. Federated orchestrator and its related aspects of network slicing are covered in Sect. 3. Section 4 includes optimal resource allocation algorithm and its importance. Section 5  includes experimental analysis to evaluate the performance of the model. Finally, Sect. 6 draws the conclusion.

Related Works
Mobile edge computing significantly reduces the computational complexity on the centralized cloud servers and it is distributed over edge nodes. The intelligence in the edges of the cloud environment enhance the performance improvement in terms of throughput and delay [5]. Deep Reinforcement learning with FL provides optimized edge caching and communication. For training the local models in FL, federated averaging (FedAvg) and FedProx [6] can be considered.FedAvg utilizes stochastic gradient descent method for optimization. To achieve fairness among the devices of the cloud network, all the devices are weighed equally by the global model. FL model provides low latency, communication and preservation of privacy [7]. The unmanned aerial vehicles networks can use collaborative federated deep learning model for exchange of user sensitive information. Distributed ensemble learning [3] allows multiple learners to coordinate in the training of the model. This approach minimizes the likelihood of choosing the insufficient model parameters during the aggregation in the global model. To reduce the communication overhead further, the gradients can be compressed before transferring to the aggregator to build the global model and parameter updation [8].
Based on the output of the learning model, the network slice can be created on demand for the requested services. To maximize the utilization of resources of the wireless network, many optimization theories has been proposed. Prospect Theory [9] is applied to incorporate the user-centric prospects. The decision-making process is made with the help of Expected Utility Theory based on utility function and rational behavior on subjective perceptions. Network slice has to be mapped to the resource based on virtual network embedding. Multi-domain slice can be established based on recursive virtualization and hierarchical network abstraction [10]. In data center network, differentiated quality of service can be provisioned by automated network slicing [11]. For the deployment of multi-tenant applications, network slice must be automatically provisioned and reconfigured based on dynamics of the network as well as the user request. The centralized SDN controller slots the bandwidth of the network dynamically for the differentiated quality of services with reconfiguration of automated network slice. The orchestrator is the core part for network slice management [12]. The network slice descriptor contains the specifications of the network slice and contains descriptions on lifecycle management of network slice. Network requirement description and network resource description accomplish in the management of the network slice.
In network resource orchestration, service differentiation and massive connectivity can achieved by exploiting spatio-temporal traffic patterns [13]. Network Functions Virtualization (NFV) and SDN complements each other in implementing network slice. SDN enables network slicing using control plane functions while NFV provides and manages network slice lifecycle and orchestrate slice using virtual network functions. The heterogeneity and dynamic characteristics of the wireless network demands powerful tool to automate the network slice which is provided by deep learning models. URLLC applications like smart grid uses SDN for network slicing which reduces network downtime due to raise in network resilience [14]. Cross domain orchestration is expected in industrial applications [15]. For example, for wind power plant application, geographical locations are distributed over different administrative domain. Within the domain, SDN controller per network service provider is responsible for management and orchestration of network slice. On the next level higher level orchestrator is responsible for coordination of network service providers.

SDN in Network Slicing at the Edge
SDN and NFV enables faster creation of network slices with SDN controller as orchestrator. Edge computing allows low latency data collection, processing and analysis of mobile edge devices in the fog and cloud computing networks. The network slice lifecycle consists of following stages: Preparation, Commissioning, Instantiation, configuration and activation and Decommissioning. Network environment preparation and on boarding is enabled by SDN and commissioning and decommission is by NFV. In 5G era, critical network slicing is essential for many of its services and based on the application environment. It requires differentiated service requirements with varying Key Performance Indicators (KPI). For improved efficiency of the dynamic network, slice provision, placement, migration with dynamic resource allocation are key indicators. Multi-access Edge Computing (MEC) incorporates the traffic offloading between data plane and control plane that provokes traffic monitoring and analysis with the SDN controller. MEC also allows network programmability and intelligence in the edge devices (Table 1).
5G network needs-controlled slicing of virtual or physical resources into logical entities to support cross domain slices and slice isolation. Slice customization is done by SDN controller that abstract the network topology. The NFV capability in the data plane provides on demand tailored network functions with data forwarding. Further data intelligence can be embedded like federated learning for fine-grained slice creation and resource allocation. The Generic Slice Template (GST) is provided by the slice controller consist of latency, bandwidth, reliability and other functionalities. The GST is provided by the core cloud and the network specific slice template is provided by edge cloud as virtualized network function as shown in Fig. 3.

The Federated Orchestrator
Consider a set of services S = (1, 2, … N) with diverse set of resources. Network slicing is enabled with network virtualization by virtualizing physical infrastructure resources into Virtual Network Function (VNF). The VNF is composed of smaller sub-units called network function which realize the network functionality. The network function can be used by different service instances quickly on-demand. The network functions are mutually exclusive and independent of each other. Each network slice is composed of many network functions based on the need and network characteristics. The network slice can be launched independently without affecting the ongoing services. The network slicing architecture is based on SDN with OpenFlow architecture. The centralized SDN controller in the control plane can control the availability of resources and handles the request of resources and requirements. The orchestrator which controls the multiple slices can be designed as federated model as shown in Fig. 4.

3
Each SDN controller is responsible for particular domain like IoT, Transport, Health care. Education, etc. The local learning model in each SDN controller will be trained using its local generated data and the global model resides in the federated orchestrator. The  global model is updated and the parameters of the federated orchestrator are updated back to the local model for further performance improvement. This mechanism of distributed network slicing with federated orchestrator allows optimal resource allocation globally, privacy preserving, less communication overhead and support for heterogeneous resources. The SDN controller queries the federated orchestrator on receiving the request for creation of network slice. The federated orchestrator will also supervise the resource reservation and data path for the routing of resources. Figure 4 depicts the federated slicing framework with SDN controller as the slice orchestrator.The objective of the proposed work is to minimize the service response time which is dependent on communication delay and queuing delay. The smallest unit of work in the slice is described as task and size of it is denoted as t n for the service type n . Assume the bandwidth of the allocated channel is and arrival rate of services follow Poisson distribution, P( m ) with m is the expected arrival rate of the services. Since the radio access network is open, the noise level and channel gain have the influence on the delay which are denoted by m and m respectively. The communication delay can be expressed as The queuing delay is dependent on the processing power of the node and service arrival rate. The processing power is denoted by m and the queuing delay is denoted as The Eqs. (1) and (2) can be combined which results in overall response time for the n th type of service offered by the network.
The computational resources are associated with certain constraints. The constraint can be modeled as sum of allocated resources to the requested services, s ∈ S cannot exceed the threshold, .

Optimal Resource Allocation
Given the set of network slices, the services are provided by instances of virtual network function. Consider the service process rate S i and the total end-to-end delay is given by D t . The traffic pattern and service request follow Gaussian distribution G( , 2 ) . For the delay to escape from infinity, service rate should exceed mean arrival rate λ. With the expected accuracy, optimal resource allocation for the tenants can be exploited. The overall process is depicted in the Algorithm 1.

Automated Network Slice Resource Provision
Autoscaling is done on demand prediction based on the history of the request made for each resource type in particular domain. The slice load predictor uses statistical learning model [25] for the mean arrival rate of traffic pattern. Based on the statistics of the peak traffic rate and mean arrival rate, the future demand of the slice resources can be estimated and optimal resource provisioning can be done using the Algorithm 1. The process flow in the automated network slice is listed as below. Configuration of dynamic virtual link-The virtual link between the NFV and the subnetwork is setup through configuration file 8. Slice Notification and Usage-Once the slice is configured as virtual network function, it is notified and is used by the application 9. Collection of Slice Statistics-The performance statistics of the network slice is reported to the slice manager entity 10. Statistics to the Slice user-The slice usage and its performance attributes are given back to the tenant 11. Terminate and Notify-Once the process is completed, the slice is terminated and notified to the service layer.

Experimental Analysis
The automated network slicing is done with OpenDayLight SDN controller with Open-Stack [26] as the orchestrator. The OpenStack module is integrated with federated learning module for the distributed deep learning model. The performance of the model is evaluated using the response time for the slice request and the QoS guarantees. The various QoS parameters like throughput, delay and deep learning parameters. The traffic traces are collected using the SDN controlled and federator orchestrator is built using the statistical data of the SDN traces. The multi domain architecture is built by the segregation of the nodal parameters of the given topology of the SDN architecture and the type of application. The VNF can be crafted on demand based the slice request and optimal resource mapping is done with slice manager. The simulation parameters are shown in Table 2.
The SDN federated controller provides the global model for the slice demand prediction. The slice template parameters act as the features (Bandwidth, Latency, reliability, flow count). Based on the history of features, the prediction model for time series data, long short-term memory is used to forecast the future traffic. The communication overhead between the local model and the global model is reduced with the usage of OpenFlow PACKET_OUT message during the routing of flows. Figure 5 provides the training accuracy of the federated approach and centralized approach. It can be observed that the accuracy of the federated model is consistent with the training epochs of 97% because the local model parameters provide the capture of generic data patterns and its dependency that are incorporated through updation of model parameters. Further the test accuracy of the global model outperforms the local models as shown in Fig. 6. With optimal resource allocation for slice, the response time of slice is reduced gradually as shown in Fig. 7. Since the slice configurations are accurately predicted, only minimal and optimal resources are allocated, forthcoming demands can be served quickly and hence the reduced response time. However, if the length of the slice (time required to complete the given network task) is higher, the response time shoots up due to less availability of resources to meet the demands. With increase in time, the prediction accuracy improves with the current data being learnt by the federated model, the response time decreases to the extent to match the needs of the application with 5G communication constraints. The communication overhead also reduces with federated learning based on epochs. The updation of the model parameters is inversely proportional to the number of samples being updated as shown in Fig. 8.

Comparison of Different Prediction Models
In SDN network, the centralized controller is responsible for all the orchestration activities may results in single point of failure that degrades network performance. Hence distributed learning models like federated learning approach is proposed for efficient performance of uncertain wireless network. And, bandwidth utilization is more due to communication overhead in the centralized model. In Federated learning approach of network slicing, the edge nodes participate in the resource prediction and allocation both globally and locally. Only the learning parameters are shared between the local and global model with minimal bandwidth utilization. Global model resides in the controller whereas the local model in edge device learns from edge data and convergence of the model is based on the tuning of learning parameters.
The resource prediction for the network slice is achieved through VNF. The VNF used the resource template for optimal resource prediction and allocation. Various time series prediction models like regressive models, sequence models and sampling of prediction window can be applied for resource prediction and optimal allocation of resources. Table 3 provides comparative analysis of various prediction models with  It can be observed that the MSE is lower compared to other time series prediction models and with higher performance. The computational complexity measured in terms of memory and training depicts that the proposed model requires memory lesser than that of other sequence models. Since federated learning is a distributed learning approach, the learning is distributed among the local and global models. The distribution of learning models make the convergence of the time series prediction model to have a bit high training time compared to other models. There is always tradeoff between the memory and the computational time and hence the federated approach best suits the light weight application in mobile communication.
The performance of the slice orchestrator is given by key performance indicators [27] as follow.

Slice Deployment Time (SDT)
It is the time interval between the slice deployment request and the time at which the slice starts its operation. It is dependent on parameters of the slice template, orchestrator features and virtualized infrastructure.

Slice Deployment Time Scalability (SDTS)
The measure of the slice scalability with respect to deployment operation. When N slice requests are sent for deployment for the same slice template, SDTS is measures as where GSDT means Generic Slice Deployment Time.

Slice Termination Time (STT)
The time interval between the slice termination request and the time at which all the allocated resources are released for further usage. Larger value of STT indicates the reduced efficiency. In Eq. (5), the overall time taken for deployment of N slice is given by GSDT and the value of N is a greedy value. For large value of N , there will be less availability of resources and for small value of N , the scalability and orchestration usage are under determined.
For the experimental purpose, each client is allowed to create an instance of 20 different slices as the threshold. Based on these slice creation, deployment and termination time, the various performance indicators like SDT, SDTS and STT are derived. The comparative analysis of these values reveals the performance of the slice orchestrator, i.e. SDN controller with federated model for network slicing. The deployment time decreases gradually with increase in slice instance because the slice template prediction is more accurate towards growing of instance. This scenario is exhibited in Fig. 9 and it can be observed that the federated orchestration decreases the deployment time with rise in the scalability because of its distributed client architecture. Figure 10 provides the comparison of GSDT and STT. The termination time of the slice is lesser compared to the deployment time and thus reveals the good scalability of the slice orchestrator. Further the linear growth of slice termination time with increase in the number of instances provides the reliability of the slice allocation scheme with the optimal resources.

Conclusion
The automated network slicing with federated orchestrator provides efficient performance in multi-domain environment. The virtualized resources and virtualized network functions enabled by NFV and SDN provides the dynamic predicted slicing of network resources. The federated architecture with its distributed clients able to provide the accurate prediction of the slice requirements dynamically with minimal communication overhead. The optimal resource allocation and provision schemes triggers the scalability of the slice orchestrator. The OpenFlow protocol of SDN environment enables the fair distribution of the model parameters with federated SDN controller. The prediction model in the global controller provides the accurate forecasting of the network slice parameters and thus dynamic allocation of resource on demand is provisioned. The performance of the slice orchestrator is dependent on the slice template parameters and however been modelled suitability by incorporation of the learning model.The proposed framework can be further implemented with the proactive content caching mechanism in the mobile edge devices. Further hierarchical controller can be placed for slice management in the cross-domain environment for efficient slicing of resources.
Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. He has 3 years of industrial experience and 17 years of academic and research experiences. He has authored 6 books in his areas of interest. He has published 110 technical papers in International and National Journals and presented 107 papers in National and International Conferences. He has completed 12 Government of India funded projects and currently 5 projects are under progress. His PhD work on Wearable Electronics bagged National Award from ISTE and he has received 12 Awards in the National level. Ashok Kumar has 3 patents to his credit. He is a Member and in prestigious positions in various National Forums. He has visited many countries for institute industry collaboration and as a Keynote speaker. He has been an Invited speaker in 110 programmes. Also he has organized 56 programmes like conferences and seminars. He completed his graduate programme in Electrical and Electronics Engineering from University of Madras and his post-graduation from PSG College of Technology, India and Masters in Business Administration from IGNOU, New Delhi. After completion of his Graduate Degree, he joined as Project Engineer in Serval