Characterisation of radio access network slicing scenarios with 5G QoS provisioning

Received January 14, 2020, accepted March 6, 2020, date of publication March 13, 2020, date of current version March 24, 2020. Digital Object Identifier 10.1109/ACCESS.2020.2980685 Characterization of Radio Access Network Slicing Scenarios With 5G QoS Provisioning IRENE VILÀ , JORDI PÉREZ-ROMERO , (Member, IEEE), ORIOL SALLENT , AND ANNA UMBERT Department of Signal Theory and Communications, Universitat Politècnica de Catalunya (UPC), 08034 Barcelona, Spain Corresponding author: Irene Vilà ([email protected]) This work was supported in part by the Spanish Research Council and FEDER funds through SONAR 5G Grant under Grant TEC2017-82651-R, and in part by the Secretariat for Universities and Research of the Ministry of Business and Knowledge of the Government of Catalonia under Grant 2019FI_B1 00102. ABSTRACT 5G systems are envisaged to support a wide range of application scenarios with variate requirements. To handle this heterogeneity, 5G architecture includes network slicing capabilities that facilitate the partitioning of a single network infrastructure into multiple logical networks on top of it, each tailored to a given use case and provided with appropriate isolation and Quality of Service (QoS) characteristics. Network slicing also enables the use of multi-tenancy networks, in which the same infrastructure can be shared by multiple tenants by associating one slice to each tenant, easing the cost-effective deployment and operation of future 5G networks. Concerning the Radio Access Network (RAN), slicing is particularly challenging as it implies the configuration of multiple RAN behaviors over a common pool of radio resources. In this context, this work presents a Markov model for RAN slicing capable of characterizing diverse Radio Resource Management (RRM) strategies in multi-tenant and multi-service 5G scenarios including both guaranteed and non-guaranteed bit rate services. The proposed model captures the fact that different radio links from diverse users can experience distinct spectral efficiencies, which enables an accurate modeling of the randomness associated with the actual resource requirements. The model is evaluated in a multi-tenant scenario in urban micro cell and rural macro cell environments to illustrate the impact of the considered RRM polices in the QoS provisioning. INDEX TERMS Markov processes, radio access networks, RAN slicing, radio resource management, quality of service. I. INTRODUCTION The forthcoming Fifth Generation (5G) systems target the simultaneous support of a wide variety of application scenarios and vertical industries (e.g. automotive, utilities, smart cities, high-tech manufacturing) with distinct and variate requirements (i.e. high data rates, low latency, high mobility) [1]. 5G will enable both the evolution of the current business models and the emergence of new ones. Partnerships at multiple-layers will be established, ranging from sharing the infrastructure to exposing specific network capabilities as an end to end service and integrating partners’ services into the 5G system through a rich and software oriented capability set [2]. In this context, 5G systems need to be provided with the flexibility and configurability required to satisfy the The associate editor coordinating the review of this manuscript and approving it for publication was Usama Mir . foreseen diversity. With this purpose, one key feature of the 5G system architecture is network slicing, which is based on Software-Defined Networking (SDN) and Network Function Virtualization (NFV) technologies [3] and allows the sharing of a common infrastructure among diverse end-to-end (self-contained) logical networks (i.e. network slices), each tailored for a given use case [4]. Each network slice can be provided with the appropriated isolation and optimized characteristics for a particular application. Therefore, network slicing enables the use of multitenancy [5], in which multiple tenants, e.g. communication providers or mobile virtual network operators (MVNO), can share the common infrastructure to provide services to their own users, resulting in reductions of capital and operational costs. Network slicing support is especially challenging for the Radio Access Network (RAN) [6], [7], which is the most resource-demanding (and costliest) part of the mobile network. The reason is that the radio spectrum is a limited 51414 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ VOLUME 8, 2020 I. Vilà et al.: Characterization of RAN Slicing Scenarios With 5G QoS Provisioning resource and the level of isolation required by the different slices can compromise the efficiency of the radio resources usage. Consequently, one of the major research problems in the field is the definition of novel Radio Resource Management (RRM) strategies, or the adaptation of existing ones, that allow both the implementation of slicing at the RAN and the fulfilment of the users’ Quality of Service (QoS) requirements [8]. The first steps in the standardization process of the system architecture and functional aspects to support network slicing in both the 5G Core Network (5GC) and in the Next-Generation RAN (NG-RAN) have been conducted by 3GPP [9], [10]. This includes the definition of the 5G New Radio (NR) interface and the 5G QoS model. Moreover, implementation aspects of network slicing in the NG-RAN have been studied from multiple angles, ranging from virtualization techniques and programmable platforms with slice-aware traffic differentiation and protection mechanisms [11]–[13] to algorithms for dynamic resource sharing across slices [14]. Similarly, [15] analyses the RAN slicing problem in a multi-cell network in relation to RRM functionalities and [16] proposes an adaptation algorithm for resource allocation, which is based on the deviations from requirements. In turn, [17] proposes a set of vendor-agnostic configuration descriptors intended to characterize the features, policies and resources to be put in place across the radio protocol layers of a NG-RAN node for the realization of concurrent RAN slices. Also, [18] proposes a procedure to establish the level of centralization of different RRM functions while [19] presents an adaptive inter-slice TDD allocation algorithm, where resources are assigned by minimizing costs and interferences. Some other works focus on the network slice admission control for slices requests that need to support a given number of users for a certain time, such as [20], [21], which target to optimize the infrastructure providers’ revenue, or [22], which optimizes the network utilization by incorporating traffic forecasting capabilities. Moreover, frameworks for the realization and study of network slicing have been implemented and tested such as in [23], which designs a system for the management of RAN slices and the provisioning of radio resources to slices based on their requirements in Long Term Evolution (LTE) technology, considering its extension to 5G NR. In the above context, this paper tackles the RAN slicing problem from a modeling perspective by proposing and developing a Markov model characterization of RAN slicing in multi-tenant and multi-service scenarios. Markovian approaches have been widely used to characterize the utilization of resources in many fields, such as in mobility [24], cloud computing [25], Call Admission Control (CAC) scheme for 3G [26] or for heterogeneous networks Radio Access Technologies (RAT) policies [27]. More recently, works in the field of 5G exploit Markov modeling to approach a proactive resource allocation scheme in highly mobile networks [28], the management of Admission Control (AC) for handoff requests between small cell and macro cell domains [29], the computation of the estimated spectrum requirement [30] and the management of slices’ creation [31]. Markov chain models have also been considered in [32] for spectrum sharing schemes and primary/secondary scenarios [33]–[35]. Our recent works [36], [37] introduced a first approach to the use of Markov chains models to characterize different RRM policies for RAN slicing at different layers of the protocol stack. Similarly, [36], [37] considered the 5G QoS model, which embraces prioritization among traffic flows. This paper constitutes a step forward in the establishment of a wider range of relationships among the different dimensions of the RAN slicing problem, including aspects of the radio environment, services types and configurations, traffic scenarios, etc. Specifically, this paper includes several novelties and advances: (i) The model allows the definition of both Guaranteed Bit Rate (GBR) and non-Guaranteed Bit Rate (non-GBR) services in terms of its corresponding 5G QoS parameters such as the Allocation and Retention Priority (ARP) and the 5G QoS Identifier (5QI) parameters. (ii) RRM policies at the different radio protocol layers have been properly characterized to support both types of services (i.e. GBR and non-GBR) and to provide slicing capabilities, so that isolation between slices is achieved both in the admission of users and the allocation of resources. (iii) In terms of the resource allocation, a new statistical model has been developed that allows capturing 5G scenarios with variate radio propagation conditions (i.e. diverse spectral efficiencies can be perceived by the different users in the system). In particular, the presented statistical model for layer 2 allows deriving the probability density function (pdf) of both the required resources and the assigned resources in multi-tenant and multi-service scenarios based on the QoS requirements. From these pdfs, different performance indicators are extracted. (iv) The proposed model is evaluated in a scenario comprised of two tenants providing both GBR and non-GBR services over urban micro cell and rural macro cell environments, where different performance metrics of interest (e.g. blocking probability, occupancy, throughput) are assessed and the relationships between the different parameters are analyzed. The rest of the paper is organized as follows. Section II presents the system model, describing the analytical Markov chain approach considered. In Sections III and IV the slicingaware AC policy performed at layer 3 and the radio resource allocation procedure at layer 2 are characterized, respectively. Section V presents different performance metrics that can be extracted from the presented model. Section VI describes two example scenarios considered for 5G RAN slicing for the subsequent validation of the model and analysis of the performance results at state and system levels, as well as a discussion of the impact of introducing a new tenant. Finally, Section VII summarizes the conclusions. II. SYSTEM MODEL A scenario comprised of a common radio infrastructure shared by Ntenants is assumed. Each tenant operates in a VOLUME 8, 2020 51415 I. Vilà et al.: Characterization of RAN Slicing Scenarios With 5G QoS Provisioning FIGURE 1. System model conceptual scheme. RAN slice, which has already been created and deployed by the Operations, Administration and Management (OAM) system among the RAN infrastructure by means of the Network Slice Subnet Management Function (NSSMF) [38]. The n-th tenant provides Mnservice types, which can be either GBR or non-GBR services. The service type indicator Ts,ntakes value 0 if the s-th service of the n-th tenant is GBR and 1 if it is non-GBR. The QoS profile is given by the GBR value (i.e., the bit rate to be provided to the user of a GBR service, also referred to as Guaranteed Flow Bit Rate (GFBR) in 5G 3GPP’s terminology), the ARP indicator [9], which defines the relative importance of the service requesting for resources and starts from 1 (highest priority) onwards (for successive lower priority services), and the priority level associated with the 5QI. Therefore, for GBR services, the QoS profile is characterized by the guaranteed GBRs,n, the ARPs,nand the 5QI priority level PLs,nfor s=1,...,Mnand n=1,...,N. In the case of non-GBR services, the GBRs,nis set to 0, as no data rate is guaranteed. The considered scenario is depicted in Fig. 1. It is comprised of a gNB, which is the NG-RAN node operating the 5G NR interface, composed of a cell with a certain bandwidth subdivided in Physical Resource Blocks (PRB) of bandwidth B. Hence, the cell has a number of available PRBs Nava at layer 1 to serve the User Equipment (UE) traffic demands. Fig. 1 also illustrates the different layers of the radio interface protocol stack at the gNB that determine how the user plane information and the control plane signaling is transferred between the UE and the gNB. Specifically, the transfer of the user plane information (e.g. IP packets associated to the services of the UE) is carried out through the Service Data Adaptation Protocol (SDAP), Packet Data Convergence Protocol (PDCP), Radio Link Control (RLC), Medium Access Control (MAC) and physical (PHY) layers. In turn, the control plane signaling can be either generated at the Radio Resource Control (RRC) layer (e.g. for measurement reporting) or at upper layer Non Access Stratum (NAS) protocols for signaling between the UE and the 5GC network (e.g. for session establishment). In both cases, signaling is transferred between UE and gNB using the PDCP, RLC, MAC and PHY layers. Further details on the functionalities of each layer can be found in [39]. In order to efficiently use the radio resources and ensure the QoS requirements of the users in the system when transferring the information through the different layers of the protocol stack, the gNB includes a set of RRM functionalities, namely the AC function at Layer 3 (L3) and the resource allocation at Layer 2 (L2), which are the focus of this paper and are briefly described in the following. Whenever a new session of a service (i.e. a new QoS flow in 3GPP terminology) is established for transferring user data through the 5G system for a given UE, the 5GC network requests the gNB to set up resources to support this QoS flow at the radio interface. The 5GC network provides the gNB with the slice identifier of the tenant, i.e. the Single Network Slice Selection Assistance Information (S-NSSAI) [9], and the QoS parameters of the service. At the gNB, the SDAP layer maps the QoS flow to a Data Radio Bearer (DRB) that enables data transfer through the different layers of the radio 51416 VOLUME 8, 2020 I. Vilà et al.: Characterization of RAN Slicing Scenarios With 5G QoS Provisioning interface protocol stack according to the expected QoS [10]. For this reason, and to make sure that the cell will have sufficient resources to support the requested QoS, the AC function at L3 is required to determine the acceptance or rejection of the new DRB in accordance with its QoS parameters and the available capacity. Let us consider that each UE is provided with a single DRB and that, as a result of the AC function, the number of admitted DRB, and thus users, of the s-th service of the n-th tenant in the cell is us,n. Then, the resource allocation function is associated to the MAC layer at L2 and is in charge of dynamically assigning the Nava available PRBs in the cell among the admitted DRBs, thus determining how the data of these DRBs travels through the physical layer. In order to configure the multi-tenant behaviour of the AC and resource allocation functions, the OAM provides the RRM policy information to the gNB [38], including guidance for the split of radio resources among the different RAN slices. Specifically, this paper assumes that the OAM provides each gNB with the per-tenant parameters required to configure the RRM functionalities, as described in the detailed models for the L3 admission control and L2 resource allocation functions that are given in Sections III and IV, respectively. Assuming that users generate sessions of exponential duration according to a Poisson arrival process, the dynamic evolution of the number of admitted users of each service type and tenant can be characterized in general by a Continuous Time Markov Chain (CTMC) with (M1+M2+...+MN)-dimensional states. Let us define S(u1,1,...,uM1,1,u1,2,...,uM2,2,...,u1,N,...,uMN,N)as the state in which u1,1,...,uM1,1,u1,2,...,uM2,2,...,u1,N,...,uMN,Nusers are admitted in the system. The state space is defined as: S= {S(u1,1,...,uMN,N)|us,n≤Umax,s,n}(1) where Umax,s,nis the maximum allowed number of users of the s-th service of the n-th tenant, which is established for hardware limitation purposes (processor, memory, power). It is worth mentioning that the AC can further restrict the number of users per service to a value lower than Umax,s,n depending on how the AC policy is specified. Transitions between states occur due to session arrivals or session departures. In this respect, it is considered that session arrivals are generated with rate λs,nfor the s-th service of the n-th tenant, while the average session duration of this service is 1/µs,n. Moreover, since AC in L3 is in charge of admitting or rejecting users’ requests depending on the system occupation, it also affects the transitions between states. In this respect, let us define ACs,n (u1,1,...,uMN,N)as the binary AC indicator for the arrivals of the s-th service and n-th tenant, taking the value 1 if the new service request is accepted and 0 otherwise. It is worth mentioning that, since transitions can only occur due to the admission of a new session or the finalization of an existing session, they can only increase or decrease the number of admitted users in one unit, meaning that transitions are only possible between neighboring states. Therefore, the transition rate qx,yfrom a state xto another state y6= xis given by: qx,y=                λs,nACs,n (u1,1,...,uMN,N)if x=S(u1,1,...,us,n,...,uMN,N)and y=S(u1,1,...,us,n+1,...,uMN,N) us,nµs,nif x=S(u1,1,...,us,n.,..,uMN,N)and y=Su1,1,...,us,n−1,...,uMN,N) 0 otherwise (2) where transitions to states with one more user (i.e. first condition in (2)) depend on the ACs,n (u1,1,...,uMN,N)indicator and λs,n while the transitions to a state with one less user (i.e. second condition in (2)) depend on the number of users of the service that decreases and µs,n. The rest of transitions from/to states with more or less than one user are not allowed (i.e. third condition in (2)). Fig. 2 illustrates a state transition diagram with all the possible transitions to/from a given state S(u1,1,...,us,n,...,uMN,N) when increasing or decreasing one user of each of the services. From the state transition diagram, the general Steady-State Balance Equation (SSBE) is given by: P(u1,1,...,us,n,...,uMN,N)[X s,n us,nµs,n +X s,n S(u1,1,...,us,n+1,...,uMN,N)∈S λs,nACs,n (u1,1,...,us,n,...,uMN,N)] =X s,n P(u1,1,...,us,n−1,...,uMN,N)λs,nACs,n (u1,1,...,us,n−1,...,uMN,N) +X s,n S(u1,1,...,us,n+1,...,uMN,N)∈S P(u1,1,...,us,n+1,...,uMN,N)(us,n+1)µs,n (3) where P(u1,1,...,us,n,...,uMN,N)corresponds to the steady-state probability of being in state S(u1,1,...,us,n,...,uMN,N). When the SSBEs are obtained for all the states, the steady-state probabilities can be computed by using numerical methods capable of solving the system of equations composed by the different SSBEs and the following normalization constraint: X S(u1,1,...,uMN,N)∈S P(u1,1,...,uMN,N)=1 (4) III. ADMISSION CONTROL AT LAYER 3 The AC function at L3 decides on the acceptance or rejection of users initiating new sessions depending on their requested QoS and the already admitted users in the cell. Since the QoS requirements have different nature for GBR and non-GBR services, the considered AC policy behaves differently in each case. For GBR services (Ts,n=0) the admission or rejection decision of a new user from the s-th service of the n-th tenant considers its requested GBRs,nand ARPs,nparameters and VOLUME 8, 2020 51417 I. Vilà et al.: Characterization of RAN Slicing Scenarios With 5G QoS Provisioning FIGURE 2. State transition diagram with all the transitions from/to a given state S(u1,1,...,us,n,...,uMN,N)with u1,1,...,us,n,...,uMN,Nusers. those of the already admitted GBR users of the same tenant, together with a per-tenant capacity threshold Cmax,ndefined as the maximum aggregate GBR that can be admitted for tenant n. This threshold Cmax,nis provided by the OAM as the RRM policy information to be enforced by the AC. Correspondingly, the new GBR user can be admitted if the aggregate GBR considering both the new user and the already accepted users of tenant nwith higher or equal priority than the new user (i.e. with ARP lower or equal than ARPs,n) does not exceed the threshold Cmax,n. It is worth noting that, in practice, Cmax,ncould be dynamically adjusted, e.g. through a Self-Organizing Network (SON) function [40], to account for the spectral efficiency conditions experienced by the users in the cell. In contrast, for non-GBR services (Ts,n=1) the system is not committed to guarantee any GBR value. Therefore, the AC in this case only checks that the number of admitted users does not exceed the maximum threshold Umax,s,n. Based on these considerations, the AC decision for a new user of the s-th service of the n-th tenant is given by: ACs,n (u1,1,...,uMN,N) =                          1 if (Ts,n=0 and Mn P s0=1 ARPs0,n≤ARPs,n Ts0,n=0 us0,n·GBRs0,n+GBRs,n≤Cmax,n) or (Ts,n=1 and (us,n+1) ≤Umax,s,n) 0 otherwise (5) Note that the considered AC function is performed independently for each tenant, in the sense that the admission of a new user belonging to a certain tenant only depends on this tenant’s parameters and occupation. Also, notice that the time scale at which AC is triggered depends on the session generation rate of the users λs,n, which will typically be in the order of seconds. IV. MODEL OF RADIO RESOURCE ALLOCATION AT LAYER 2 From the perspective of the Markov model, admissions at L3 and session finalizations generate state transitions. In turn, within the sojourn time of a given state, the resource allocation at L2 defines how the Nava available PRBs in the cell are assigned to the admitted users in a given state. In practice, this is performed by the Packet Scheduler (PS), which dictates in a short time basis (millisecond time scale) the user allocation of PRBs to transmit data packets by considering the QoS constraints and the actual radio conditions associated with each user. In this paper, the PS behavior is characterized in accordance with the sojourn time of the Markov model states (i.e., typically few seconds) and, consequently, all the short time scale components of the PS (e.g. fast fading) will be averaged out in the resource allocation model. Fig. 3 shows the considered resource allocation process. It is performed independently for each tenant nand assuming a maximum number of PRBs, Nth,n, to be allocated to this tenant, which is also determined by the OAM and provided to the gNB as RRM policy information. As depicted in Fig. 3, the process firstly allocates PRBs among the admitted users of GBR services. For this purpose, let us denote as arp(1, n), arp(2, n),..., arp(Mn,n) the list of ARP 51418 VOLUME 8, 2020 I. Vilà et al.: Characterization of RAN Slicing Scenarios With 5G QoS Provisioning FIGURE 3. Resource allocation procedure. values in increasing order for the GBR services of tenant n, i.e. starting by arp (1,n)=mins|Ts,n=0ARPs,nand ending with arp (Mn,n)=maxs|Ts,n=0ARPs,n. Then, the process firstly allocates the PRBs required by all the GBR services with arp(1, n), then with arp(2 ,n), and so on. As long as there are available PRBs to serve the users of an ARP value, resources are assigned. Instead, when not enough PRBs are available (i.e. there is congestion), the procedure proportionally distributes the available PRBs among the users of that ARP value. Other criteria might be adopted for handling persistent congestion situations such as the use of congestion control functions. However, they are out of the scope of this work. After this process, the remaining PRBs are allocated among the admitted users of non-GBR services. Both phases are further developed in the following. A. RESOURCE ALLOCATION TO GBR SERVICES For a given state x=S(u1,1,...,uMN,N), the target here is to model the aggregate number of assigned PRBs to the users of a given GBR service sof tenant n, denoted as ax,s,n. Provided that the us,nusers of this service experience variable propagation conditions, this is a random variable that should be characterized statistically through its pdf, denoted as fax,s,n(k). In order to obtain fax,s,n(k), the first stage is to formulate the statistical distribution of the number of required PRBs by the us,nusers. Then, the procedure of Fig. 3 is needed to allocate the Nth,navailable PRBs among the Mnservice types, considering that the required PRBs of each service can be different and that the ARPs establish a prioritization among services. The number of required PRBs by a user of the GBR service sof tenant nis given by: Nreq,s,n=GBRs,n Seff ·B(6) where Seff is the spectral efficiency, measured in b/s/Hz, and Bis the PRB bandwidth. Since Seff fluctuates depending on the propagation conditions that users experience when moving around the cell, it is treated as a random variable. Therefore, based on measurements collected from the different users, it is possible to derive the pdf of the random variable Y=1/(Seff ·B), denoted as fY(y). This pdf is obtained by gathering samples of the wideband Channel Quality Indicator (CQI) distribution [41]. The CQI is an integer index that indicates the modulation and coding scheme available to user in accordance to its experienced propagation conditions. The wideband CQI distribution is computed by the gNB from the CQI reports provided by the different users, and it is reported to a Management and Data Analytics Function (MDAF) [42] with a given periodicity (e.g. 15 minutes). The MDAF gathers samples of this distribution and averages them for a longer time period in order to get the adequate statistical validity to be representative of the cell conditions. Then, the averaged distribution of the CQI indices can be directly mapped to the distribution of the Seff according to the tables in section 5.2.2 of [43]. From the distribution of the spectral efficiency, fSeff (y), the pdf of Y,fY(y) can be extracted based on fSeff (y) as: fY(y)=fSeff 1 y·B1 y2B(7) VOLUME 8, 2020 51419 I. Vilà et al.: Characterization of RAN Slicing Scenarios With 5G QoS Provisioning Correspondingly, the number of required resources Nreq,s,n is another random variable, whose pdf is: fNreq,,s,n(k)=fYk GBRs,n1 GBRs,n (8) Assuming that each user experiences independent propagation conditions, the pdf of the aggregate number of required PRBs rx,s,nof the s-th service of the n-th tenant in state xcan be computed as: frx,s,n(r)=1 GBRs,nus,n ·fYr GBRs,n∗. . . ∗fYr GBRs,n | {z } us,n (9) where ∗represents the convolution operator. As seen in Fig. 3, the allocation of resources to GBR users is performed in accordance with the priorities established by the ARP values of each service in the tenant, treating all the services with the same arp(m,n) together. Correspondingly, let us denote as Rarp(m,n)=[rx,s0,n| ARPs0,n=arp(m,n),Ts0,n=0] the vector of required resources for the GBR services of tenant nwith the same arp(m,n).Considering independence between requirements of different services, the joint pdf of the rx,s,nvariables in Rarp(m,n)is: fRarp(m,n)Rarp(m,n)= Mn Y s=1 ARPs,n=arp(m,n) Ts,n=0 frx,s,n(rx,s,n) (10) In turn, let us denote as Zarp(m,n)the random variable representing the aggregated number of PRBs already assigned to the GBR services in tenant nof ARP lower than arp(m,n) and fZarp(m,n)(z) its pdf. For arp(1 ,n), since no resources have been yet assigned, fZarp(1,n)=δ(z), where δ(·) is the Dirac delta function. The pdf fax,s,n(k) of the s-th service of the n-th tenant with ARPs,n=arp(m,n) can be computed by conditioning the value of assigned resources ax,s,nto the requirements Rarp(m,n)of the services with the same arp(m,n) and to the already assigned resources by GBR services Zarp(m,n). Then, by applying the law of the total probability, this yields to: fax,s,n(k)= ∞ Z0 ··· ∞ Z0 fax,s,n|Rarp(m,,n),Zarp(m,n)(k|Rarp(m,n),z) ·fZarp(m,n)(z)·fRarp(m,n)(Rarp(m,n))· ·dz · Mn Y s0=1 ARPs0,n=arp(m,n) Ts0,n=0 drx,s0,n(11) where fax,s,n|Rarp(m,n),Zarp(m,n)(k|Rarp(m,n),z) is the pdf of ax,s,nconditioned to Rarp(m,n)and Zarp(m,n), which is given by: fax,s,n|Rarp(m,n),Zarp(m,n)(k|Rarp(m,n),z) =                                δ(k−rx,s,n) if Mn P s0=1 ARPs0,n=arp(m,n) Ts0,n=0 rx,s0,n≤Nth,n−z δ(k−αarp(m,n)rx,s,n) if Mn P s0=1 ARPs0,n=arp(m,n) Ts0,n=0 rx,s0,n>Nth,n−z (12) The first condition in (12) considers the case that the number of available resources for the tenant is higher than the required resources given by Rarp(m,n)vector. Then, each user gets the required resources, i.e. the aggregate resources are ax,s,n=rx,s,n. In turn, the second condition in (12) considers the case that there are not sufficient resources for the tenant to fulfil the requirements of all the users of arp(m,n) (i.e. there is congestion). In this case, the assigned resources are proportionally distributed among these users as ax,s,n=αarp(m,n)·rx,s,nwith: αarp(m,n)=max            Nth,n−z Mn P s=1 ARPs,n=arp(m,n) Ts,n=0 rx,s,n ,0            (13) Once the computation of fax,s,n(k) for each of the services with arp(m,n) is completed, the pdfs for the subsequent ARP value arp(m +1,n) need to be derived. Then, the pdf fZarp(m+1,n)(z) is computed based on fZarp(m,n)(z) and the aggregate resources that have been assigned to the services with arp(m,n), denoted by the random variable Aarp(m,n). This yields to: fZarp(m+1,n)(z)= +∞ Z0 fZarp(m,n)(z0)·fAarp(m,n)|Zarp(m,n)(z−z0|z0)·dz0 (14) where fAarp(m,n)|Zarp(m,n)(t|z) is the pdf of Aarp(m,n)conditioned to Zarp(m,n)given by: fAarp(m,n)|Zarp(m,n)(t|z)=fKarp(m,n)(t)·H(t,Nth,n−z) +δ(t−(Nth,n−z)) · ∞ Z Nth,n−z fKarp(m,n)(k)dk (15) H(x,y)=(1x<y 0x≥y(16) 51420 VOLUME 8, 2020 I. Vilà et al.: Characterization of RAN Slicing Scenarios With 5G QoS Provisioning where fKarp(m,n)(z) is the pdf of the aggregate required resources for all the GBR services of arp(m,n), given by: fKarp(m,n)(t)=frx,1,1(t)∗. . . ∗frx,s,n(t) | {z } s=1,...,Mn,ARPs,n=arp(m,n),Ts,n=0 (17) Note that, for the case m=Mn, which corresponds to the maximum ARP value among the GBR services in tenant n, expression (14) gives the pdf of the random variable ZGBR,n that represents the aggregate PRBs assigned to all the GBR services in that tenant, i.e. fZGBR,n(z)= +∞ Z0 fZarp(Mn,n)(z0)·fAarp(Mn,n)|Zarp(Mn,n)(z−z0|z0)·dz0 (18) B. RESOURCE ALLOCATION TO NON-GBR SERVICES As seen in Fig. 3, the remaining PRBs after the allocation to GBR users of tenant nare distributed among the users of non-GBR services of the tenant. This distribution is carried out proportionally based on the 5QI priority level and the number of admitted users us,nof each service, with the proportional constant σs,n, defined as: σs,n=us,n·(1/PLs,n) Mn P s0=1 Ts0,n=1 us0,n·(1/PLs0n) (19) Consequently, the assigned PRBs to the non-GBR service sof tenant nare given by ax,s,n=(Nth,n−ZGBR,n)·σs,n. The pdf of ax,s,ncan be obtained by using the pdf of the assigned resources after GBR allocation fZGBR (z) as: fax,s,n(k)=1 σs,n fZGBR,n(Nth,n−k σs,n ) (20) V. PERFORMANCE METRICS Based on the steady-state probabilities, this section develops the different performance metrics of interest for the evaluation of the considered RRM strategies. A. BLOCKING PROBABILITY Blocking states are those in which the acceptance of a new user of a given service is not possible. Specifically, the set of blocking states for users of the s-th service of the n-th tenant is denoted as Sb s,n, defined as: Sb s,n= {S(u1,1,..,uMN,N)∈S|ACs,n (u1,1,...,uMN,N)=0}(21) Similarly, the set of blocking states for the n-th tenant, Sb n, are those states in which the acceptance of one user from any of the services of this tenant is not possible. Therefore, it is defined as the intersection of the sets of blocking states for the services of this tenant, i.e. Sb n=Sb 1,n∩Sb 2,n∩. . . ∩ Sb Mn,n. Similarly, the set of all blocking states in the system Sbis expressed as the intersection of the set of blocking states of each tenant/service. Based on the blocking states, the blocking probability computed per service and per tenant is given by: Pb s,n=X S(u1,1,...,uMN,N)∈Sb s,n P(u1,1,...,uMN,N)(22) This can be easily extended to compute the blocking probability per tenant or the global blocking probability by considering Sb nor Sb, respectively, in the summation of (20). B. OCCUPANCY METRICS Given the steady-state probabilities P(u1,1,...,uMN,N), it is also possible to compute different metrics that provide information about the occupancy of the system. The average number of admitted users Us,nof the s-th service of the n-th tenant is given by: Us,n=X S(u1,1,...,uMN,N)∈S us,n·P(u1,1,...,uMN,N)(23) The average number of admitted users per tenant can be computed by adding the average number of users per service, i.e. Un=U1,n+U2,n+. . .+UMn,n. Similarly, the global system average number of admitted users Uwould be computed as the sum of the average number of users for all services and tenants. Another system occupancy metric that can be obtained from the model is the average PRB utilization as,naggregated per each service. The average aggregated PRB utilization in a given state x=S(u1,1,...,uMN,N)for the s-th service and the n-th tenant ax,s,nis: ax,s,n=Z∞ 0 k·fax,s,n(k)·dk (24) Then, the system average aggregated PRB utilization per service as,ncan be computed by considering all the states ax,s,nand the steady-state probabilities as: as,n=X S(u1,1,...,uMN,N)∈S aS(u1,1,...,uMN,N),s,n·P(u1,1,...,uMN,N)(25) Accordingly, the average PRB utilization per tenant would result from an=a1,n+a2,n+. . . +aMn,nwhile the global system PRB utilization acan be computed by adding the average PRB utilization of each tenant. Then, the average normalized PRB utilization of the s-th service of the n-th tenant ωs,nis expressed as: ωs,n=as,n Nava (26) Similarly to the other metrics, the average normalised PRB utilization per tenant ωnand the global average normalized PRB utilization ωresult from adding the average normalized PRB utilization of the tenant’s services or all the services, respectively. VOLUME 8, 2020 51421 I. Vilà et al.: Characterization of RAN Slicing Scenarios With 5G QoS Provisioning C. AVERAGE AGGREGATED THROUGHPUT For a state x=S(u1,1,...,uMN,N), the average aggregated throughput for the s-th service of the n-th tenant Thx,s,nis computed differently for GBR and non-GBR services. In the case of GBR services, Thx,s,nis computed as: Thx,s,n= ∞ Z0 ··· ∞ Z0 k·fThrx,s,n|Rarp(m,,n),Zarp(m,n)(k|Rarp(m,n),z) ·fZarp(m,n)(z)·fRarp(m,n)(Rarp(m,n)) ·dz · Mn Y s0=1 ARPs0,n=arp(m,n) Ts0,n=0 drx,s0,n·dk (27) where fThx,s,n|Rarp(m,n),Zarp(m,n)(k|Rarp(m,n),z) is the pdf of the aggregated throughput Thx,s,nfor the GBR service sfrom tenant nconditioned to Rarp(m,n)and Zarp(m,n), given by: fThx,s,n|Rarp(m,n),Zarp(m,n)(k|Rarp(m,n),z) =                                δ(k−us,n·GBRs,n) if Mn P s0=1 ARPs0,n=arp(m,n) Ts0,n=0 rx,s0,n≤Nth,n−z δ(k−αarp(m,n)·us,n·GBRs,n) if Mn P s0=1 ARPs0,n=arp(m,n) Ts0,n=0 rx,s0,n>Nth,n−z (28) Expression (28) follows the same principle as (12), considering that, when the number of available PRBs is higher than the number of required PRBs (first condition in (28)), all users get the required PRBs and thus the throughput of each user is GBRs,n. Instead, when there are not sufficient resources (second condition in (28)), the assigned PRBs are just a fraction αarp(m,n)of the required ones and thus the throughput of each user is αarp(m,n)·GBRs,n· Regarding non-GBR services and considering that the number of assigned PRBs to the non-GBR users depends on the spare PRBs after the allocation to GBR users, the average aggregated throughput Thx,s,nfor non-GBR users is independent of its actual spectral efficiency and is given by: Thx,s,n=ax,s,n·Seff ·B(29) where Seff is the average of the spectral efficiency Seff , derived from measurements in a similar way as variable Y. The system average aggregated throughput Ths,nfor the sth service of the n-th tenant can be computed by considering Thx,s,nfor all services and the steady-state probability as: Ths,n=X S(u1,1,...,uMN,N)∈S ThS(u1,1,...,uMN,N),s,n·P(u1,1,...,uMN,N) (30) The average aggregated throughput for the n-th tenant Thnand the average global system aggregated throughput Th result from the summation of the average aggregated throughputs of the tenant’s services or all the services in the system, respectively. D. DEGRADATION PROBABILITY Degradation occurs when congestion is reached and some GBR admitted users cannot be assigned with their required resources to provide GBRs,n. Instead, they are assigned with a lower number of resources according to the considered resource allocation criteria. For a given state x, the degradation probability of the GBR service s-th of the n-th tenant with ARPs,n=arp(m,n) can be obtained by the following expression: pdeg,x,s,n= ∞ Z0 ··· ∞ Z0 δ(k−αarp(m,n)rx,s,n) ·H(Nth,n−z, Mn X s0=1 ARPs0,n=arp(m,n) Ts0,n=0 rx,s0,n)·fZarp(m,n)(z) ·fRarp(m,n)(Rarp(m,n))·dz· Mn Y s0=1 ARPs0,n=arp(m,n) Ts0,n=0 drx,s0,ndk (31) where H(·) is defined in (16) and allows considering in the computation only those cases in which there are not enough resources to satisfy the service requirements (i.e. the second condition in (12)). Based on the degradation at state level and the steady-state probabilities, the degradation probability of the s-th of the n-th tenant at system level is given by: Pdeg s,n=X x∈S pdeg,x,s,n·P(u1,1,...,uMN,N)(32) Note that the degradation probability is not computed for non-GBR services, as GBRs,nis not established for them. VI. PERFORMANCE EVALUATION This section presents a performance analysis of the proposed analytical model. Firstly, the detailed characteristics of the considered scenarios are presented. Then, the model validation, the analysis of the scenarios at both state and global system levels and a discussion of the impact of introducing a new tenant follow. A. CONSIDERED SCENARIO The scenario under test considers N=2 tenants, referred to Tenant 1 and Tenant 2. Tenant 1 provides M1=3 different services while Tenant 2 provides M2=2 services. For each of the services, the QoS parameters summarised 51422 VOLUME 8, 2020 I. Vilà et al.: Characterization of RAN Slicing Scenarios With 5G QoS Provisioning Solutions A and B, the obtained Us,nand Ths,nof each service is the same, so Table 7 does not present separate results for each solution. Besides, it is also observed that very small differences are obtained in the results of the UMi and RMa environments. VII. CONCLUSION This paper has presented a Markov model that characterizes the resource sharing in RAN slicing scenarios, where multiple tenants provide GBR and non-GBR services. The model can include diverse RRM functions in terms of admission control at layer 3 and radio resource allocation at layer 2. In terms of admission control, which determines the transition probabilities between the different states in the model, a slice-aware admission control policy has been selected. The model has also been provided with a slicing-aware resource allocation procedure that considers variable propagation conditions by deriving the pdfs of both the required and assigned resources according to the service’s QoS parameters. The effect of the considered RRM functions has been studied by evaluating different performance metrics (blocking probability, degradation probability, throughput, occupation, etc.) in a scenario considering two tenants providing GBR and non-GBR services in both urban micro cell and rural macrocell environments. Based on the considered scenario, the analytical model’s suitability has been validated given the low percentage errors obtained (i.e. maximum relative errors of 3%) when comparing the model’s results with the ones obtained with a system level simulator. The performance analysis conducted at state and system levels, as well as the analysis of the introduction of a new tenant, has revealed that (i) the considered admission control policy and resource allocation procedure are able to achieve isolation between the different slices, so that overload situations in one slice do not affect the performance of GBR users of the other slice while preserving the maximum capacity allowed to each of the slices; (ii) ARP priorities are respected by providing better performance to those GBR services with lower ARP (higher priority); (iii) GBR services are provided with negligible degradation rates, which implies that the requested GBR values are provided to the admitted users in the system; (iv) non-GBR services are provided with the remaining resources after the allocation to GBR services according to its priority level, so that better performance is given to those non-GBR services with higher priority (lower priority level); (v) The model is able to capture the radio propagation effects, enabling the analysis of different performance metrics in 5G environments of interest; (vi) The introduction of new tenants into the system can be performed by re-configuring the maximum capacity and maximum PRBs to be provided to each of the tenants, although the re-configuration of these values may impact on the performance of already operating tenants if the total amount of PRBs is not modified. Given the potential of the proposed model, a possible future extension of the resource allocation procedure at layer 2 can be the joint optimization of the performance of different tenants by distributing the unused PRBs of one tenant among the existing users of other tenants. REFERENCES [1] R. Hattachi, Ed., ‘‘5G White paper, version 1,’’ NGMN Alliance, Boston, MA, USA, White Paper, Feb. 2015. [Online]. Available: https://www.ngmn.org/fileadmin/ngmn/content/images/news/ngmn_news/ NGMN_5G_White_Paper_V1_0.pdf [2] X. Zhou, R. Li, T. Chen, and H. Zhang, ‘‘Network slicing as a service: Enabling enterprises’ own software-defined cellular networks,’’ IEEE Commun. Mag., vol. 54, no. 7, pp. 146–153, Jul. 2016. [3] J. Ordonez-Lucena, P. Ameigeiras, D. Lopez, J. J. Ramos-Munoz, J. Lorca, and J. Folgueira, ‘‘Network slicing for 5G with SDN/NFV: Concepts, architectures, and challenges,’’ IEEE Commun. Mag., vol. 55, no. 5, pp. 80–87, May 2017. [4] P. Rost, A. Banchs, I. Berberana, M. Breitbach, M. Doll, H. Droste, C. Mannweiler, M. A. Puente, K. Samdanis, and B. Sayadi, ‘‘Mobile network architecture evolution toward 5G,’’ IEEE Commun. Mag., vol. 54, no. 5, pp. 84–91, May 2016. [5] K. Samdanis, X. Costa-Perez, and V. Sciancalepore, ‘‘From network sharing to multi-tenancy: The 5G network slice broker,’’ IEEE Commun. Mag., vol. 54, no. 7, pp. 32–39, Jul. 2016. [6] P. Rost, C. Mannweiler, D. S. Michalopoulos, C. Sartori, V. Sciancalepore, N. Sastry, O. Holland, S. Tayade, B. Han, D. Bega, D. Aziz, and H. Bakker, ‘‘Network slicing to enable scalability and flexibility in 5G mobile networks,’’ IEEE Commun. Mag., vol. 55, no. 5, pp. 72–79, May 2017. [7] S. E. Elayoubi, S. B. Jemaa, Z. Altman, and A. Galindo-Serrano, ‘‘5G RAN slicing for verticals: Enablers and challenges,’’ IEEE Commun. Mag., vol. 57, no. 1, pp. 28–34, Jan. 2019. [8] D. M. Gutierrez-Estevez, O. Bulakci, M. Ericson, A. Prasad, E. Pateromichelakis, J. Belschner, P. Arnold, and G. Calochira, ‘‘RAN enablers for 5G radio resource management,’’ in Proc. IEEE Conf. Standards Commun. Netw. (CSCN), Helsinki, Finland, Sep. 2017, pp. 1–6. [9] System Architecture for the 5G System; Stage 2 (Release 15), document 3GPP TS 23.501 V16.0.2, Apr. 2019. [10] NR; NR and NG-RAN Overall Description; Stage 2 (Release 15), document 3GPP TS 38.300 V15.5.0, Apr. 2019. [11] R. Kokku, R. Mahindra, H. Zhang, and S. Rangarajan, ‘‘NVS: A substrate for virtualizing wireless resources in cellular networks,’’ IEEE/ACM Trans. Netw., vol. 20, no. 5, pp. 1333–1346, Oct. 2012. [12] X. Costa-Perez, J. Swetina, T. Guo, R. Mahindra, and S. Rangarajan, ‘‘Radio access network virtualization for future mobile carrier networks,’’ IEEE Commun. Mag., vol. 51, no. 7, pp. 27–35, Jul. 2013. [13] A. Ksentini and N. Nikaein, ‘‘Toward enforcing network slicing on RAN: Flexibility and resources abstraction,’’ IEEE Commun. Mag., vol. 55, no. 6, pp. 102–108, Jun. 2017. [14] P. Caballero, A. Banchs, G. de Veciana, and X. Costa-Pérez, ‘‘Multi-tenant radio access network slicing: Statistical multiplexing of spatial loads,’’ IEEE/ACM Trans. Netw., vol. 25, no. 5, pp. 3044–3058, Oct. 2017. [15] O. Sallent, J. Pérez-Romero, R. Ferrús, and R. Agustí, ‘‘On radio access network slicing from a radio resource management perspective,’’ IEEE Wireless Commun., vol. 24, no. 5, pp. 166–174, Oct. 2017. [16] B. Khodapanah, A. Awada, I. Viering, D. Oehmann, M. Simsek, and G. P. Fettweis, ‘‘Fulfillment of service level agreements via slice-aware radio resource management in 5G networks,’’ in Proc. IEEE 87th Veh. Technol. Conf. (VTC Spring), Porto, Portugal, Jun. 2018, pp. 1–6. [17] R. Ferrús, O. Sallent, J. Pérez-Romero, and R. Agustí, ‘‘On 5G radio access network slicing: Radio interface protocol features and configuration,’’ IEEE Commun. Mag., vol. 56, no. 5, pp. 184–192, May 2018. [18] E. Pateromichelakis and C. Peng, ‘‘Selection and dimensioning of slicebased RAN controller for adaptive radio resource management,’’ in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), San Francisco, CA, USA, Mar. 2017, pp. 1–6. [19] E. Pateromichelakis and K. Samdanis, ‘‘A graph coloring based inter-slice resource management for 5G dynamic TDD RANs,’’ in Proc. IEEE Int. Conf. Commun. (ICC), Kansas City, MO, USA, May 2018, pp. 1–6. [20] D. Bega, M. Gramaglia, A. Banchs, V. Sciancalepore, K. Samdanis, and X. Costa-Perez, ‘‘Optimising 5G infrastructure markets: The business of network slicing,’’ in Proc. IEEE INFOCOM Conf. Comput. Commun., Atlanta, GA, USA, May 2017, pp. 1–9. [21] D. Bega, M. Gramaglia, A. Banchs, V. Sciancalepore, and X. Costa-Perez, ‘‘A machine learning approach to 5G infrastructure market optimization,’’ IEEE Trans. Mobile Comput., vol. 19, no. 3, pp. 498–512, Mar. 2019. VOLUME 8, 2020 51429 I. Vilà et al.: Characterization of RAN Slicing Scenarios With 5G QoS Provisioning [22] V. Sciancalepore, K. Samdanis, X. Costa-Perez, D. Bega, M. Gramaglia, and A. Banchs, ‘‘Mobile traffic forecasting for maximizing 5G network slicing resource utilization,’’ in Proc. IEEE INFOCOM Conf. Comput. Commun., Atlanta, GA, USA, May 2017, pp. 1–9. [23] X. Foukas, M. K. Marina, and K. Kontovasilis, ‘‘Orion: RAN slicing for a flexible and cost-effective multi-service mobile network architecture,’’ in Proc. 23rd Annu. Int. Conf. Mobile Comput. Netw. (MobiCom), 2017, pp. 127–140. [24] J. Epperlein and J. Marecek, ‘‘Resource allocation with population dynamics,’’ in Proc. 55th Annu. Allerton Conf. Commun., Control, Comput. (Allerton), Monticello, IL, USA, Oct. 2017, pp. 1293–1300. [25] S. Jagannatha, N. S. Shravan, and S. Kavya, ‘‘Cost performance analysis: Usage of resources in cloud using Markov-chain model,’’ in Proc. 4th Int. Conf. Adv. Comput. Commun. Syst. (ICACCS), Coimbatore, Indian, Jan. 2017, pp. 1–8. [26] H. Y. Ng, K. T. Ko, and K. F. Tsang, ‘‘3G mobile network call admission control scheme using Markov chain,’’ in Proc. 9th Int. Symp. Consum. Electron. (ISCE), Jun. 2005, pp. 276–280. [27] X. Gelabert, J. Pérez-Romero, O. Sallent, and R. Agustí, ‘‘A Markovian approach to radio access technology selection in heterogeneous multiaccess/multiservice wireless networks,’’ IEEE Trans. Mobile Comput., vol. 7, no. 10, pp. 1257–1270, Oct. 2008. [28] V. V. Paranthaman, Y. Kirsal, G. Mapp, P. Shah, and H. X. Nguyen, ‘‘Exploring a new proactive algorithm for resource management and its application to wireless mobile environments,’’ in Proc. IEEE 42nd Conf. Local Comput. Netw. (LCN), Singapore, Oct. 2017, pp. 539–542. [29] S. Al-Rubaye, A. Al-Dulaimi, J. Cosmas, and A. Anpalagan, ‘‘Call admission control for non-standalone 5G ultra-dense networks,’’ IEEE Commun. Lett., vol. 22, no. 5, pp. 1058–1061, May 2018. [30] Y. Kim and S. Park, ‘‘Analytical calculation of spectrum requirements for LTE-A using the probability distribution on the scheduled resource blocks,’’ IEEE Commun. Lett., vol. 22, no. 3, pp. 602–605, Mar. 2018. [31] B. Han, D. Feng, and H. D. Schotten, ‘‘A Markov model of slice admission control,’’ IEEE Netw. Lett., vol. 1, no. 1, pp. 2–5, Mar. 2019. [32] M. N. Patwary, R. Abozariba, and M. Asaduzzaman, ‘‘Multi-operator spectrum sharing models under different cooperation schemes for next generation cellular networks,’’ in Proc. IEEE 86th Veh. Technol. Conf. (VTC-Fall), Toronto, ON, Canada, Sep. 2017, pp. 1–7. [33] S. Lin, L. Kong, Q. Gao, M. K. Khan, Z. Zhong, X. Jin, and P. Zeng, ‘‘Advanced dynamic channel access strategy in spectrum sharing 5G systems,’’ IEEE Wireless Commun., vol. 24, no. 5, pp. 74–80, Oct. 2017. [34] K. B. Ali, M. S. Obaidat, F. Zarai, and L. Kamoun, ‘‘Markov model-based adaptive CAC scheme for 3GPP LTE femtocell networks,’’ in Proc. IEEE Int. Conf. Commun. (ICC), London, U.K., Jun. 2015, pp. 6924–6928. [35] A. Jee, S. Hoque, B. Talukdar, and W. Arif, ‘‘Analysis of link maintenance probability for cognitive radio ad hoc networks,’’ in Proc. 5th Int. Conf. Signal Process. Integr. Netw. (SPIN), Noida, India, Feb. 2018, pp. 385–389. [36] I. Vila, O. Sallent, A. Umbert, and J. Pérez-Romero, ‘‘Guaranteed bit rate traffic prioritisation and isolation in multi-tenant radio access networks,’’ in Proc. IEEE 23rd Int. Workshop Comput. Aided Modeling Design Commun. Links Netw. (CAMAD), Barcelona, Spain, Sep. 2018. [37] I. Vila, O. Sallent, A. Umbert, and J. Pérez-Romero, ‘‘An analytical model for multi-tenant radio access networks supporting guaranteed bit rate services,’’ IEEE Access, vol. 7, pp. 57651–57662, Apr. 2019. [38] Management and orchestration; Provisioning; (Release 16), document 3GPP TS 28.531 v16.2.0 Jun. 2019. [39] E. Dahlman, S. Parkvall, and J. Sköld, 5G NR: The Next Generation Wireless Access Technology. New York, NY, USA: Academic, 2018. [40] J. Pérez-Romero, O. Sallent, R. Ferrús, and R. Agustí, ‘‘Self-optimised admission control for multitenant radio access networks,’’ in Proc. IEEE 28th Annu. Int. Symp. Pers., Indoor, Mobile Radio Commun. (PIMRC), Montreal, QC, Canada, Oct. 2017, pp. 1–5. [41] Management and Orchestration; 5G Performance Measurements (Release 16), document 3GPP TS 28.552 v16.2.0 Jun. 2019. [42] Management and Orchestration; Architecture Framework (Release 16), document 3GPP TS 28.533 v15.1.0, Jun. 2019. [43] NR; Physical Layer Procedures for Data (Release 15), document 3GPP TS 38.214 v15.6.0, Jun. 2019. [44] Study on Channel Model for Frequencies From 0.5 to 100 GHz (Release 15), document 3GPP TS 38.214 v15.6.0, Mar. 2019. [45] NR; Base Station (BS) Radio Transmission and Reception (Release 15), document 3GPP TS 38.104 v15.5.0, Apr. 2019. [46] NR; Physical Channels and Modulation (Release 15), document 3GPP TS 38.211 v15.3.0, Sep. 2018. [47] Study on New Radio Access Technology: Radio Frequency (RF) and coExistence Aspects (Release 14), document 3GPP TR 38.803 v14.2.0, Sep. 2017. [48] W. J. Stewart, Introduction to the Numerical Solution of Markov Chains. Princeton, NJ, USA: Princeton Univ. Press, 1994. IRENE VILÀ received the B.E. degree in telecommunication systems engineering and the M.E. degree in telecommunication engineering from the Universitat Politècnica de Catalunya (UPC), Barcelona, in 2015 and 2017, respectively. She is currently pursuing the Ph.D. degree with the Mobile Communication Research Group (GRCM), Department of Signal Theory and Communications (TSC), UPC, supported with an FI AGAUR Grant by the Government of Catalunya. In 2018, she joined the Mobile Communication Research Group (GRCM), Department of Signal Theory and Communications (TSC), UPC. Her current research interests include RAN Slicing, radio resource management, software defined networking (SDN), and network function virtualization (NFV), concepts to be included in new 5G technologies. JORDI PÉREZ-ROMERO (Member, IEEE) is currently a Professor with the Department of Signal Theory and Communications, Universitat Politècnica de Catalunya (UPC), Barcelona, Spain. He has been working in the field of wireless communication systems, with particular focus on radio resource management, cognitive radio networks, and network optimization. He has been involved in different European Projects and in projects for private companies. He has published more than 200 articles in international journals and conferences. ORIOL SALLENT is currently a Professor with the Universitat Politècnica de Catalunya (UPC). He has participated in a wide range of European projects with diverse responsibilities as the Workpackage Leader and a Coordinator partner and contributed to standardization bodies, such as 3GPP, IEEE, and ETSI. He has published more than 200 articles mostly in IEEE journals and conferences. His research interests include cognitive management in cognitive radio networks, self-organizing networks, radio network optimization, and QoS provisioning in heterogeneous wireless networks. ANNA UMBERT received the Engineering and Ph.D. degrees in telecommunications from the Universitat Politècnica de Catalunya (UPC), in 1998 and 2004, respectively. In 2001, she joined UPC as an Assistant Professor, and became an Associate Professor, in 2017, which is her current status. Since 1997, she has been participating in several projects founded by both public and private organizations. She has published more than 50 articles in international journals and conferences. Her research interests are focused in radio resource and QoS management in the context of heterogeneous wireless networks, cognitive management in cognitive radio networks, dynamic spectrum access and management, self-organized networks, and network optimization. 51430 VOLUME 8, 2020