5G RAN Slicing: Dynamic Single Tenant Radio Resource Orchestration for eMBB Traffic within a Multi-Slice Scenario

Emerging 5G systems will need to seamlessly guarantee novel types of services in a multi-do-main ecosystem. New methodologies of network and infrastructure sharing facilitate the cooperation among the operators, exploiting the core and access sections of the system architecture. Network slicing (NS) is the operators' best technique for building and managing a network. Without NS, the 5G requirements in terms of flexibility, optimal resource utilization, and investment returns cannot materialize. Before slicing is commercially available, different sections of the 5G architecture should be modified to include NS. In this work, we present a novel dynamic radio access network slicing resource sharing method aimed to guarantee optimal service level agreements through the monitoring of each slice tenant's key performance indicators. The experiments are conducted following the 3GPP specifications, and the solution is validated using a testbed based on the main 5G functionalities.


IntroductIon
The evolution of radio access networks (RANs) in recent years is now at 5G.From the previous 4G technology, the Third Generation Partnership Program (3GPP) has added multiple functionalities to the RAN and core network (CN) to support new network elements and applications.Even though the Long-Term Evolution (LTE) technology introduced significant RAN improvements, the exponential growth of connected devices and innovative services during the last decade pushes mobile network operators (MNOs) to investigate new methods for improving the coverage, increase their systems' capacity, and reduce the services latency in mobile and fixed networks.As a consequence, in the next years, the main mobile infrastructure challenges for the MNO will be [1]: • Enabling 80-100 MHz of spectrum for the operator at 3.5 GHz, and around 1 GHz per operator at 26-28 GHz [2] • Definition of new market and business models through the deployment of new infrastructures and software solutions • Massive network expansion and upgrade of private solutions • Intensive infrastructure deployments, advanced indoor and outdoor distributed antenna systems • Advanced hardware and software (HW/SW) solutions for new services, with smart energy consumption and accurate system management The aforementioned novel capabilities are an opportunity for operators to further move the simple connectivity concept and explore new architectural solutions such as the coexistence with 4G networks and other access/core technologies to provide an unlimited, fast, reliable, and enhanced broadband experience to society, reaching data rates up to 1 Gb/s and latency less than 4 ms.Such novel system performance permits the operators to increase their portfolio with cloud and artificial intelligence (AI) solutions, massive Internet of Things (IoT) deployments, and advanced industrial applications for new vertical markets.
Earmarked as a prominent feature of 5G for enabling the aforementioned technological capabilities, the concept of network slicing (NS) has been introduced.This key technology enabler permits multiple vertical industries to execute their solutions on top of a shared infrastructure, where the service provider (SP) customizes the network capabilities (security, connection, processing power, storage, etc.) [3].From an SP point of view, this solution represents a new type of business model, known as NS as a service (NSaaS), which will be the future network trend of the coming years.MNOs, using an unique network slice type, are able to arrange the needs of multiple customers; also, different services belonging to multiple network slices may be gathered together and supplied as a single slice to a customer with diverse requirements.A typical example of a vertical market using this solution is autonomous driving, where telemetry and infotainment services are simultaneously provided in a single package.
Due to the stochastic behavior of networks, static allocation of network resources does not represent an efficient approach.In particular, when applied to wireless resources, due to bursts of traffic, user mobility, and the time-varying channel, only through resource overprovisioning techniques are we able to deal with multiple services, without exploiting the flexibility of short-medium-term fluctuations in terms of resource requirements [4].For this reason, dynamic NS represents the leading approach for future networks.With this technique, network resources are dynamically assigned to meet tailored performance requirements (e.g.capacity, latency, priority, security) through a seamless and virtually continuous network propagated across the 5G network architecture [5].This methodology will help operators to optimize resource usage, reduce cost, and accelerate time to market for innovative new service offerings.
To this end, this work illustrates a novel real-time end-to-end NS framework where multiple network slice instances are dynamically modeled according to the MNO's service level agreement (SLA) and users' quality of service (QoS) to meet the optimal radio resource sharing and service performance.The concept of NS was initially proposed for the 5G CN, while only in the last few years have 3GPP study items started to investigate the impact of slicing in the RAN part of the network.
To the best of our knowledge, our solution represents the first application of a dynamic NS solution through a joint evaluation of the slice tenant owner SLA and the real-time service performance from the user perspective.The main contributions of our solutions include: • Creation of a network traffic generator tool for testing different services in line with the testbed capabilities.The traffic requirements are proportionally scaled according to the performance of our real testbed.• Highly customizable configuration of the SP traffic requirements, slices isolation policies, processing resources for new slice instances, and logging.• Real-time analysis of the slice performance and resources configuration through joint evaluation of the user traffic trend and slice SLA.• Implementation of the dynamic NS solution using platform-independent interfaces for advanced scalability with third-party SW/ HW.The remainder of this article is organized as follows.We give an overview of the concept of NS and its properties.We illustrate the architecture and the optimization algorithm of our solution.We illustrate the experimental features and results, and concluding remarks are given.

ns FeAtures, technIques, And development
This section illustrates some basic considerations necessary for the SPs when a new slice-based system is instantiated, and a description of two NS implementation methodologies will provide a better understanding of our approach.To motivate the proposed system architecture of this work, a brief discussion behind the RAN advancements is performed.

requIrements
This subsection illustrates the principal requirements and considerations adopted by an SP when a new slice instance should be instantiated.In our solution, similar analysis is conducted for defining the initial configuration of each deployed slice.
For each tenant, the SP configures a network slice instance following some baseline principles [6]: • The network section: RAN, transport network (TN), and CN.Each section presents different requirements and performance, which should be carefully analyzed before instantiating a service.• The SLAs: Latency, guaranteed bit rate (GBR), non-guaranteed bit rate (N-GBR), availability, and packet loss are some parameters to evaluate when the slice is defined.• The type of vertical market application: Industry 4.0, vehicle-to-everything (V2X), smart cities, the Internet of Things (IoT), and so on.The customization of multiple slice granularities introduce many challenges, especially in terms of network management and orchestration.For this reason, the SP must have a complete vision of the network capabilities necessary to perform optimal network management and orchestration of the resources.

multI-provIder And multI-domAIn nsAAs
Since the main principles of 5G are scalability and flexibility, multiple SPs may share the same physical infrastructure and activate different services on top of a multi-domain ecosystem.Moreover, with the growth of software defined networking (SDN) and network function virtualization (NFV) technologies, the sharing of network resources has become a common practice among SPs [6].
While in a single-domain scenario, an SP is aware of the topology and available network resources, in a multi-domain scenario there is no management tool for sharing the topologies and resources information across SPs.Therefore, 5G introduces the exchange of information across these providers through a series of specific interfaces, as standardized in [6].Using IF1 interface, the tenant sends a request for the deployment of a service or a slice.To manage operations on top of a multi SP system, IF2 interface is specific for the communication among the orchestrators.Finally, IF3 interface facilitates the management of multi-domain networks through the separation of the technological solution from the vendor-specific infrastructure.
This solution facilitates the creation of intelligent 5G network management systems across different domains and connected optimization functions.This principle improves network management operations for multiple markets as well as the vision of a unique transparent infrastructure to the final service tenant.
The customization of multiple slice granularities introduce many challenges, especially in terms of network management and orchestration.For this reason, the SP must have a complete vision of the network capabilities necessary to perform optimal network management and orchestration of the resources.

stAtIc vs dynAmIc slIcIng
The major element underlying NS is the mechanism for resource allocation among slices.During the first trials on NS, 3GPP suggested that the base station resources are accurately divided based on predefined network policies [7].Multiple providers share the same infrastructure, and the resources are allocated according to the QoS requirements.
Resource overprovisioning is utilized against SLA violation, introducing as a side effect a reduction of the system performance due to the possible allocation of unusable resources.Moreover, the stochastic behavior of the network medium introduces complexity when it comes to allocating the resources of a new slice instance.
With radio dynamic NS, multiple tenants adjust their network capacities during different time periods, according to service and system variations.Machine learning and/or traffic forecasting techniques can be deployed to assist the slice provider in handling unexpected network situations and traffic pattern fluctuation, which would involve SLA violation.Concurrently, a reinforcement learning (RL) solution such as Q-Learning can efficiently determine the optimal slice admission policy that maximizes the MNO's revenue [8].Even though the Q-Learning method is capable of being executed in an online learning fashion with a much more reasonable computation cost, the training and decision phases may not be processed in time considering the complexity of 5G frame structure patterns (European Telecommunications Standards Institute, ETSI TS 138-211 V15.3.0), implying congestion and SLA violation in the RAN part and modest service quality.To overcome this issue, our solution introduces an extra resource tolerance pool to each slice able to amortize unexpected traffic peaks, as explained below.

rAn slIcIng evolutIon
The RAN has seen an exponential evolution over the past few years, and considerable effort will be needed as we enter a new phase of the mobile industry.
The future RAN network will be more intricate and diffused, reinforcing well-known solutions such as network hypervisors, virtual machines (VMs), and containers, while simultaneously exploiting novel technologies such as SDN, NFV, virtualization (vRAN), and cloud (C-RAN) [9].Network hypervisors are the network elements that abstract the physical infrastructure into logically isolated virtual network slices.A virtual machine (VM) enables the virtualization of a physical resource where each client can execute its own operating system (OS), and resources such as computing, storage, memory, and network are shared among VMs.From the combination of the aforementioned tools, containers are lightweight alternatives to hypervisor-based VMs.A physical server in containers is virtualized such that standalone applications and services can be instantiated on isolated servers.vRAN applies the features of NFV by virtualizing network functions (NFs), providing a higher degree of flexibility for the RAN section.A virtual RAN consists of a centralized pool of baseband units (BBUs), virtualized RAN control functions, and service delivery optimization platforms [10].Furthermore, vRAN permits a shared CN interface to assist multiple 5G New Radios (5G-NRs), facilitating the deployment of 5G street macro and small cells in areas characterized by different densities.C-RAN represents the network architecture that can be used to activate virtualized networks.It requires a high-capacity and low-latency access network to manage fronthaul traffic.
The combination of vRAN and C-RAN enables the division of the control plane (CP) and data plane (DP), reducing the decoupling complexity of the NFs from private hardware, and increasing the level of versatility of MNOs' networks needed for the commercialization of 5G.These elastic and scalable access and connectivity related functionalities are provided as a service to customers in a given geography area using 3GPP standard-based technologies.

system desIgn And mAnAgement system ArchItecture
The proposed dynamic RAN slicing solution architecture is illustrated in Fig. 1.Our solution consists of three macro layers (fetch, management, execution), nestled together following a bottom-up workflow.It is important to remark that our proposed solution is backward compatible with the existing 3GPP 5G stack, since it utilizes standardized protocols and functionalities to communicate with the main network architecture elements.Moreover, even though this work is focused on RAN slicing, using a suitable set of input data and output configuration file settings, it is possible to extend our approach to other network sectors (midhaul, backhaul) without downgrading the optimization capabilities of our model.
Fetch Phase: In this subsection, all the information from the data acquisition blocks are collected, sorted, and skimmed to define a network data model utilized as input for the upper layers.The NIC traffic block filters the incoming traffic according to the MNO's slice requirements, using variable window granularities.Different types of filters can be customized based on the operator's policies: packet size, colored traffic, IP subnet, and so on.
The CP acquisition block monitors the behavior of the served users to evaluate if the SLAs and custom user traffic requirements are guaranteed.As the current served users have priority, this block can dynamically modify the incoming services' acquisition rules.Through this methodology, new incoming users are rejected when the slice resources are saturated, or no resource sharing policy is available.Moreover, if the slice tenant decides to modify the SLA requirements after the service is instantiated, this block is responsible for reconfi guring the acquisition criteria for the new and served users, while maintaining seamless service.
To conclude this subsection, the channel acquisition block monitors the access and fronthaul parts to identify physical variations of the medium, which would affect not only the service acquisition rate, but also the confi guration of the slices in the upper layers.For example, if a channel degradation is identifi ed, the amount of resources for users must be redefined, together with the slice acceptance rates.
Management Phase: This part illustrates the leading block of the presented approach.As the central block, the Manager synchronizes and coordinates the tasks of all other blocks.Its implementation can be centralized or distributed, as most of the system components are virtualized.The correct placement of the Manager improves the system tolerance against failures and minimizes traffi c overloading in sensitive network nodes.For this reason, in this work we assume that the proposed framework is installed in the edge part of the network, which represents a strategic point for the management of multiple RAN aspects.As a minor task of the Manager, it is responsible for instantiating or deleting the slices, according to the slice owner decision policies and/or the system performance.
As first step, the Manager receives as input the real-time data scenario model from the fetch layer, the SP-specifi c settings, and per slice operator SLA.This information is encapsulated following a specific pattern, and forwarded to the Optimization Algorithm block, which returns the optimal slices parameterization for the next system processing window.For each slice of the MNO a container is defi ned with the amount of available resources and predefined scheduling policies.In Fig. 1, only two slice containers are represented, one eMBB and one mMTC, in order to keep our explanation aligned with the subsequent testing part.
When the identification of RAN resources changes, the Manager estimates a slice parameterization every time new data are received from the fetch phase.If the novel slice configuration diff ers from the current one in terms of required resources per slice, the Manager reconfigures the amount of resources assigned to each slice according to the service's need.From a practical perspective, this procedure corresponds to shifting the portion of resource blocks (RBs) among the slices, while preserving the SLA and QoS of each user.Following a tunable granularity, this operation can be executed dynamically, in accordance with the transmission time interval (TTI) of our system.The new slices parameterization is structured following the JavaScript Object Notation (JSON) syntax, and forwarded to the top tier of the architecture.
Execution Phase: The top system layer implements a set of RESTful-based application programming interfaces (APIs) for the exchange of control information between the system and third-party radio software.Once a new slice configuration is posted to the RAN, the POST block sends an acknowledgment (ACK) back to the Manager with the outcome of the operation.
As the final operation, the Manager calls the Update System Statistics block, which is responsible for updating the system variables required for the optimal processing of the forthcoming slices' confi gurations.Optionally, a log fi le can be provided to trace potential issues during the entire workfl ow.

optImIZAtIon AlgorIthm
Inside the Optimization Algorithm block, the system partitions the physical resources of each slice into three modalities, each one with a specific role, as illustrated in Fig. 2: • Support: The slice accepts incoming service requests, and part of its resources can be shared with other slices.• Conservative: The slice prioritizes its own traffi c by disabling sharing of resources with others.• Critical: The amount of allocated resources does not guarantee a complete SLA.The system evaluates if one exists or more slices in support mode are able to share part of their resources with the critical slice.The set of thresholds for each slice is defi ned considering the type of traffi c, isolation policies, custom confi gurations, slice SLA, and so on.Given a specifi c key performance indicator (KPI) for each slice, the goal of the optimization block is to guarantee the optimal performance and SLA through real-time sharing and balancing of the physical resources in the RAN part.For each slice, the tenant defi nes the SLA in terms of maximum guaranteed bit error rate (BER), minimum and average guaranteed data rate, maximum tolerable latency, and maximum percentage of rejected requests.For the estimation of radio resources per user, a joint reverse-engineering approach with the slice SLA is performed by the Manager to estimate the amount of physical resources needed for a given user data rate.Once the user is accepted, its assigned resources are further refi ned according to the system evolution.When the analysis of the real-time user traffic and slice SLA requires an amount of resources that exceed the second threshold, the slice shifts to critical mode, and the algorithm activates the resource sharing procedure.
The definition of a critical mode represents the innovation behind our dynamic management of the radio resources.Since the wireless channel is characterized by a stochastic behavior, the definition of a pool of extra resources handles unexpected traffic peaks, and the incoming users can be accepted and served while the algorithm performs resource balancing among the competing slices.This principle would be highly complex and inaccurate using a machine learning solution, since the training sets are difficult to model for real-time unpredictable RAN traffic, and the elaboration of an optimal slice resource configuration may not be computed in time according to the frame and subframe structure.
Through the illustrated slice mode division, the Manager is always aware of the traffic served by each slice and the amount of available resources, since through the comparison of the occupied radio resources from the served users with the slice maximum capacity and threshold values, it is possible to always estimate the current mode for each slice.This principle facilitates the identification of a slice that will most probably require additional resources without impacting the SLA.

sIngle tenAnt, two slIces: A cAse study system scenArIo
In order to assess the performance of our solution, an HW/SW experimental platform is deployed, as illustrated in Fig. 3.We focus our experiments in downlink traffic, and for the proof of concept we use two types of slices, eMBB and mMTC, as part of a single-tenant scenario.Since we want to evaluate our solution under heavy traffic condition, significant eMBB traffic is injected [11], while the primary role of the mMTC slice is to assist the supply of eMBB radio resources along the simulation.It is important to highlight that our system applies inter-slice resource prioritization only if negotiated during the SLA definition process.Otherwise, the first come first served (FCFS) policy is applied.Due to the current limitation in open source 5G standalone (SA) platforms, in this work, the performance of the dynamic NS solution is evaluated on top of an LTE-based testbed, equipped with 5G virtualization functionalities.Following, the main HW/ SW tools of our testbed are described, while an exhaustive illustration of all the system elements was presented in our previous work [12].Our platform is based on OpenAirInterface (OAI) [13], a flexible solution for 5G research implementing the 3GPP cellular stack on general-purpose processor architectures.The OAI radio section presents a series of interfaces for the interconnection of different third-party RF modules.Our scenario utilizes a USRP B210 software defined radio (SDR) provided by Ettus Research.This SDR guarantees a bandwidth of 10 MHz at 2.5 GHz in downlink, which corresponds to 50 RBs in the RAN.From the client side, a Raspberry Pi 4 Model B is equipped with a 4G/LTE HAT board, provided by Sixfab GmBH.
As shown in Fig. 3, we separated the RAN and CORE parts into two different machines.This implementation choice balances the processing workload of the system, and allows flexibility in terms of centralized or distributed management, and the deployment of multiple split options as standardized for 5G [14].
Table 1 summarizes the main parameters of the tested scenario.The initial slice division represents the percentage of RBs assigned to each slice given the total system capacity, while the thresholds refer to the maximum capacity of each slice mode.

perFormAnce AnAlysIs
Using the aforementioned scenario as the baseline of our experiments, we evaluate the efficiency of the proposed dynamic NS solution when a high load of eMBB traffic is injected, and as a   consequence the SM selects radio resources from the mMTC slice to balance the system until the SLAs are not fulfilled.This condition is confirmed through the evaluation of the final slice eMBB mode occupancy percentage (8 percent support, 24 percent conservative, and 68 percent critical), where the high presence of critical calls force the SM to perform resource migration among the slices during the entire simulation time.
Figure 4 illustrates the behavior of the eMBB slice during the entire simulation time.To correctly interpret the achieved results, the reader should take into account that the maximum downlink capacity using the OAI-based testbed is around 30 Mb/s.
The blue line indicates the injected eMBB high load traffic (average 20.62 Mb/s, standard deviation 3.83, variance 14.74), while the red line represents the slice eMBB capacity growth trend until the SLA is not reached.For every variation of the traffic flow, the SM evaluates if a new slice configuration must be applied.This procedure is displayed with the bar plot, where each bar represents the amount of RBs required by the incoming service to reach the optimal slice eMBB SLA.When the bar sequence has a growing trend, the SM increases the amount of RBs for the eMBB slice in the next slice configuration.Conversely, a decreasing bar sequence trend indicates that the current slice configuration correctly matches the served users traffic, and the SM can also decide on a partial release of RBs.
As stress downlink test for the slice eMBB, a series of flows are generated from the cloud network toward the users, with an average traffic load equal to 80 percent of the total system capacity.As expected, the consecutive resource optimization calls of the SM brings a continuous increment of the eMBB slice capacity (from 25 RBs of the initial setup through a final slice capacity of 44 RBs), with an occupation of 88 percent of the total system capacity before the end of the simulation.
With our solution, using an SM decision granularity of 3 s, after approximately 60 s the system reaches steady resource balancing, optimally configured for the type of injected traffic.Moreover, as the sharing policies take into account the performance of the complete set of slices within the system, the SLAs are guaranteed for both slices during the entire process until a final stable slice configuration.
With an average amount of received packets of 20.50 Mb/s, this experiment presents a packet error ratio (PER) equal to 0.005.The dynamic approach of our solution, as expected, slightly affects the PER, which is nevertheless acceptable if compared to the standardized eMBB requirements [15].
For every iteration, the impact on the slice eMBB capacity proportionally affects all the modes, as illustrated in Fig. 5.This proportional scaling of the RBs of each mode reduces the ping-pong effect among the slices, which appears when highly variable traffic is injected, and continued scaling up and down of the resources impacts the PER and management complexity.Without the division of each slice in different modes, the system would continue moving radio resources among the slices with the objective to balance the maximum capacity of each slice, defining a new slice configuration even for irrelevant traffic variations.
The results of Fig. 5 bring to light an intrinsic principle regarding the initial parameterization of the Support and Conservative eMBB slice modes.Allocating a tiny amount of RBs in Support (20 percent of the initial capacity) pushes the system into Conservative mode even when a small amount of traffic is injected in the system.This condition protects the eMBB RBs, since the RBs sharing functionality is disabled, as explained earlier.Moreover, even a reduced Conservative threshold benefits the eMBB slice, since the slice is more inclines to enter the Critical mode, requesting other slices to share part of their RBs.This observation highlights how the initial slice modes setup should be carefully investigated, taking into account the type of slice traffic, the slice isolation policies, and the specific tenant's requirements.
To conclude this section, an average jitter of 0.206 ms confirms the optimal configuration of the tested scenario, ensuring a high order modulation scheme during the entire simulation.The high customization degree of the SM parameters permits tackling multiple scenarios and simultaneously optimizing different types of slices.
A trade-off between slice resource assignation accuracy and a new slice configuration processing time should be carefully inspected.Depending on the type of traffic, this feature may have an impact on the slice throughput, latency, and responsiveness to real-time traffic pattern changes.

conclusIons
In this article, we present a complete solution for dynamic RAN slicing resource allocation where the optimal slice configuration is computed through a joint evaluation of the slice SLAs and the real-time evolution of the served users' traffic.Even though the experimental section is conducted using a single-tenant scenario with two slices (eMBB and mMTC), the proposed method can be extended to multi-tenant multi-slice environments.
The obtained results show how our solution is able to autonomously remap the radio resources in a few seconds while keeping a PER of 0.005 under a heavy traffic scenario.We have proved that a suitable configuration of the slice policies and system parameters guarantees optimal performance for different types of traffic, matching the scalability and flexibility properties of the 5G networks.
As future work, we will define an exhaustive formulation of the optimization problem where the probability of accepting/rejecting an incoming user connection request is done combining a state-independent model of the multi-slice scenario with the analysis of the system capabilities in terms of real-time resources availability and SLAs.The stochastic model outcomes will be used by the Manager to further improve the resource management and accelerate the decision process, while simultaneously reducing the risk of resource saturation of the system.Moreover, the testing of our solutions on top of an SA end-to-end open source 5G testbed is also one of our core priorities.
As future work, we will define an exhaustive formulation of the optimization problem where the probability of accepting/rejecting an incoming user connection request is done combining a state-independent model of the multi-slice scenario with the analysis of the system capabilities in terms of realtime resources availability and SLAs.

FIGURE 1 .
FIGURE 1. Framework structure of our solution for dynamic RAN slicing.

TABLE 1 .
System and simulation parameters.