RAN Slicing in Multi-MVNO Environment under Dynamic Channel Conditions

With the advent of Internet of Things (IoT), number of devices accessing the internet has been increasing exponentially. Ericsson has estimated that about 5 billion IoT devices will be connected to the internet and about 2.6 billion 5G subscriptions by the end of 2025 [1]. Efficient utilization of available spectrum resources becomes vital to accommodate this growth. Adding to this requirement, is the complexity of users having varied QoS requirements.

To address this complexity, Radio Access Network (RAN) slicing technology has been widely adopted by several industrial communities [2, 3]. With the help of RAN slicing, operators can perform service customization, isolation and multi-tenancy support on a common physical network infrastructure by enabling logical as well as physical separation of network resources [4]. This multi-tenancy support enables network operators (NO) to support multiple mobile virtual network operators (MVNOs) in the form of a slice. The Third-Generation Partnership Project (3GPP) has identified network slicing as one of the key technologies to achieve varied performance requirements — such as high throughput, low high security goals in 5G networks [5].

One of the key features of RAN slicing is that - MVNOs are assigned slices that are independent from one another [6]. That is, the allocation of the radio resources is up to the NO who can allocate them at will, based on the QoS requirement while ensuring complete isolation between slices. The NO we consider is based on Software Defined - Radio Access Network (SD-RAN) controller architecture comprising of a Slice Manager and MVNO specific scheduler. The NO architecture is formally introduced in Section III.

Once the QoS requirements for each MVNO is collected by NO, the core problem lies in allocation of scarce spectrum resources such that each MVNO’s QoS requirement is met for all its users. We consider spectrum resources as resource blocks (RB). This is a difficult problem because over provisioning of RB for a user, will result in wastage and under provisioning might not meet the QoS requirements. Therefore the design of efficient slicing algorithm to meet each MVNO user’s requirement is key for optimal usage of RB. Also, from a business standpoint, optimal usage of RB which will result in increased number of users served in a time slot and thereby increased number of MVNOs supported by fixed number of RB is of great interest.

One of the factors which influence the slicing decision, is the channel condition experienced by the RB during its path towards the users. In order to convey the channel information, each user in the network sends a CQI report back to the NO. Often in real world scenarios, the channel conditions do not remain the same. They keep varying with respect to time and frequency. In order to take into account of this dynamic channel condition, the users send the CQI report in regular intervals with its periodicity determined by the NO and in between this interval, the channel conditions are assumed to remain same [7]. In order to remain close to reality, we consider dynamic channel conditions in our work.

We illustrate the problem of RAN Slicing under dynamic channel conditions by considering minimum data rate per time slot for each user as a specification by MVNOs. Calculation of data rate for a user at a given time depends on MCS level chosen for the user by the NO at that time. Choice of MCS level in turn depends on the CQI report sent from the users of MVNO. Now, the problem we are addressing in this paper is, how do we create a channel conditions aware slice for each MVNO such that, maximum number of MVNO user’s minimum data rate requirement is met.

Even though RAN resource allocation issue has been studied extensively in the recent past [8],[6],[9], the problem of resource allocation to MVNOs under dynamic channel conditions is relatively new. This is discussed more in Section II. Design of efficient resource allocation/slicing enforcement algorithm is not trivial and is met with unique challenges:

• •

  Users maximization: Meeting the minimum data rate requirement for maximum number of MVNO users in the slice time slot. This can be achieved by choosing the optimal number of RB and the MCS level for each user.

• •

  Orthogonality: Each RB should be allocated to only one user across all MVNOs at a given time slot to avoid interference  [10],[11],[12].

• •

  Support advanced 5G technologies: The RB allocation should also facilitate in implementation of advanced 5G technologies such as CoMP and MIMO [9].

The aim of this work is to design, analyze and validate MCS aware RAN Slicing(MaRS) algorithm that take into consideration the challenges mentioned above. To summarize, this work makes the following contributions:

• •

  We formulate the MCS aware RAN Slicing (MaRS) problem as a Non-Linear Programming Optimization problem in Section V using the model developed in Section III. We will also prove the NP-Hardness of the MaRS problem.

• •

  We propose a solution for this problem using the greedy algorithm paradigm in Section VI.

• •

  We develop an upper performance bound for the MaRS algorithm in Section VII.

• •

  We provide an implementation of the proposed solution and carry out an exhaustive evaluation in Section VIII.

There has been significant work to address the problem of RAN slicing, especially in the recent past. There have been many excellent surveys on this topic  [4, 10, 13, 14, 12]. The authors in these surveys provide a comprehensive information regarding the work being done on this topic. Additionally, a book has been published on the topic of RAN Slicing where many slicing algorithms have been proposed [15]. Specifically [13] covers the advancements in RAN slicing which is based on the SDN architecture. The architecture considered in our work loosely follows the work covered in [13].

In the recent works, the RAN slicing problem [16, 17, 18] has been dealt by designing solutions using various theoretical means optimization [19, 20], game theory [21]. There has also been many advancements where several machine learning approaches have been used to address the RAN Slicing problem – {Reinforcement Learning [22], Deep Learning [23, 24, 25]}. These machine learning approaches are not suitable for deployment due to their huge data requirements for training and the time it requires to do so. Moreover accurate predictions of the channel conditions are required to make the slicing algorithm effective.

One of the key limitation of these works is that it does not show the actual deployment of RAN slices on top of a physical network. Although the authors in the paper [9] discuss RAN slicing policies and enforcement problem by considering fine grained control of resources, it falls short when we bring in dynamic channel conditions. Moreover, the problem formulation considers slice as allocation of certain percentage of resource blocks from a given pool without considering the underlying requirement for these slices.

One of the work which closely focuses on addressing RB allocation problem is [11]. The authors propose a RB partitioning algorithm which focuses on allocating RB to every MVNO by simultaneously maximizing the percentage of satisfied MVNOs while allocating the minimum amount of RB. However, the problem in [11] does not take into account the dynamic channel conditions.

Our work can be closely compared to [8]. The authors of [8] address the problem of RAN slicing by considering dynamic channel conditions in a SD-RAN based architecture. One of the key architectural differences between our work and [8] is the flexibility offered to the MVNO in the SD-RAN architecture. In [8], the authors consider individual slice managers for each slice but a common scheduler for all the users. This provides very less flexibility for MVNOs. In our work, we consider independent scheduler for all MVNOs. This allows MVNO, the option of choosing its users for scheduling at each time interval. Section III discusses this in detail.

In [26], the authors address the RAN Slicing problem for multiplexing eMBB and URLLC slices. Although the paper [26], considers the MCS selection in the design of the slicing algorithm, it again falls short in providing MVNO the flexibility in scheduling as the architecture considered in completely different.

The papers [27, 28] present a framework for LTE virtualization. The authors propose an architecture for virtualizing the LTE base stations (called eNodeB in LTE architecture) with the objective of having different operators sharing the same physical resources. The solution is based on a hypervisor (as in CPU virtualization), which hosts different virtual nodes, allocates the resources and is responsible of the spectrum sharing and data multiplexing. In [29], the framework from  [27, 28] is used to present a algorithm for scheduling physical RB for the virtual nodes. The main idea of this algorithm is that if a eNodeB is overloaded and a neighbor eNodeB has available resources, a user is selected to be migrated to the unloaded eNodeB. Although the concept of centralized control is similar to our work, the problem statement is completely different. In our work, we are addressing the problem of RAN slicing in a multi-MVNO environment as opposed to resource sharing.

In summary, our work addresses the shortcomings of these papers by providing more flexibility to the MVNOs, developing efficient slicing algorithm with dynamic channel conditions and carrying out a thorough validation.

We consider a NO administering a single $5G$ RAN base station $B$ and set of $\mathcal{M}=\{1,2...,$M$\}$ MVNOs as depicted in Fig. 1.The NO serves the MVNOs by creating virtual RAN slices built on top of underlying physical network $B$. We split NO functionally into Slice Manager and MVNO scheduler. This architecture lies inline with 5G RAN concepts, where the management and orchestration is implemented as a Software Defined Network (SDN). We adopt the architecture principle similar to  [8], [30], and include additional features to aid the proposed slicing procedure.

Once the NO collects the minimum data rate slice request from all MVNOs, it creates an instance of MVNO scheduler for each MVNO in the network.We define $\Lambda_{m}^{i}$ as the minimum data rate requirement for each user $i$ of MVNO $\forall i\in m,m\in\mathcal{M},\forall m$. MVNO Scheduler for all $m\in\mathcal{M}$ provides a scheduling order of users belonging to $m$, $\mathcal{U}_{m}$, to the Slice Manager. The Slice Manager, which has the CQI information for each user in the network, dynamically assigns the resources on $B$ to each MVNO slice based on this scheduling order sent by the MVNO. Advantage of this architecture is that it leaves the choice of scheduler implementation, up to the MVNO. Each MVNO may employ unique scheduling algorithm.

Since the BS follows $5G$ cellular technology, spectrum resources are organized as grids of RB, that span across both time and frequency domains [31]. Each RB represents the minimum spatio-temporal scheduling unit. Considering $N_{RB}$ and $T$ as the number of available subcarriers and temporal slot respectively, the set of available RB is $|R_{b}|=N_{RB}\times T$ in the physical RAN network for a certain bandwidth.

Implication of time slot T. Theoretically, the time slot $T$ can range from 1 TTI($t$) to $1000$’s of TTIs, depending upon how dynamically the slice manager wishes to operate resource slicing policy. Under realistic consideration and in lieu with next-generation O-RAN architecture [32], it is expected that the slicing manager will either reside in non-real-time Ran Intelligent Controller(RIC) or near-real-time RIC, which are respectively, in order of $>1$s and ($10-1000$ ms) time scales [33]. Thus, in our work, we consider that $T$ will be a large value, in range of several milliseconds. Further, we consider the user’s minimum data rate requirement is defined per time slot $T$.

Dynamic Channel Conditions. We consider the channel conditions to be dynamic in nature and may vary in frequency and time, but remain consistent within the time slot $T$. This is similar to aperiodic csi reporting [7]. Depending upon the channel condition obtained from CQI reports for users of the MVNOs being served, the Slice Manager determines a suitable MCS for transmission depending on each MVNO user’s minimum data rate requirement, out of $29$ MCS levels as per 5G 3GPP specification [31]. Let $\mathcal{C}$ denote the set of available MCS, i.e., $\mathcal{C}=\{0,1,\dots,28\}$. The MCS determines how much information (in bits) is modulated and coded in each RB by the BS. The higher the MCS is, the higher the modulation and coding rate is. That means, the maximum amount of information that can be transmitted on one RB also depends on the channel conditions. If the channel condition is poor and the NO uses a high MCS, then the information carried in the RB will not be successfully received and decoded. Therefore, the achievable data rate by an RB depends on both MCS level chosen by the NO as well as the channel condition for this RB.

Let $q_{u^{i}_{m}}^{r,t}$ denotes the maximum MCS that can be used for a certain RB $r$ to serve a user $u^{i}_{m}\in\mathcal{U}_{m}$ such that the information carried in RB can be successfully received by the user at TTI $t\in T$.

Let $v^{c,t}$ denote the modulation and coding rate for a resource block under MCS $c\in\mathcal{C}$ and $d_{u^{i}_{m}}^{r,c,t}$ denote the maximum achievable data rate by RB $r$ for the user $u^{i}_{m}$ under MCS $c\in\mathcal{C}$ at time $t\in T$. If $c\leq q_{u^{i}_{m}}^{r,t}$, the transmission would be successful and the achievable data rate is $v^{c,t}$. Otherwise, the transmission would be unsuccessful and the data would be lost. That is,

Finally, the NO also imposes a restriction on maximum throughput allowed per MVNO slice, namely $\overline{\Lambda_{m}}$, depending on channel conditions or business requirements. This restriction prevents any individual MVNO slice from overloading the network. An intuitive way of selecting $\Lambda_{m}^{i}$ may be from pure business perspective, i.e, whichever MVNO pays the most will get higher throughput. However, in general the choice of maximum throughput in a multi-MVNO, limited resources environment introduces a new problem that is out of the scope for this work.

Before we proceed with designing of the Slicing Algorithm, it is important to understand how different components in the SD-RAN architecture interact with each other to serve users of MVNO. In this section, we present the workflow for our SD-RAN architecture in Fig 1.

The NO communicates with several components before it assigns resources to a specific user of a MVNO. After a MVNO requests services from a NO, the task of allocating resources can be broadly divided into 4 steps.

In this section, we formulate the MCS-aware RAN Slicing (MaRS) problem as a Non Linear Optimization problem. The problem aims at determining the optimal set of resource blocks to be allocated to each MVNO $m\in\mathcal{M}$ in time slot $T$, such that maximum number of users can be served across MVNOs in $T$, by considering – (1) MVNO’s minimum bit rate requirement is met for each of its user, (2) Each MVNO scheduler’s unique user scheduling order is ensured, and (3) The total throughput per slice does not exceed the maximum allowable throughput set by the NO for that MVNO slice.

Notation. Let set $\mathcal{U}_{m}=\{u^{1}_{m},\dots u^{j}_{m},\dots u^{|U_{m}|}_{k}\}$ denote the scheduling order of all users belonging to MVNO $m\in\mathcal{M}$.

Decision variables. Let $u^{i}_{m}$ denotes whether a user $i$ belonging to MVNO $m$ can be served by the Slice Manager. Let $x^{r,i,t}_{m}$ denote the whether a certain RB $r\in R$ is allocated to any user $u^{i}_{m}$ in MVNO $m$ at TTI $t\in T$. Let $y^{c,t}_{i,m}$ denote whether a MCS level $c$ is chosen by a user $u^{i}_{m}$ at TTI $t$.

The maximization problem given in (2) targets to accommodate maximum number of users to satisfy the constraints. Constraint (3) indicates that a resource block can be allocated to one UE at any given time. Equation (4) indicates that MCS chosen for a user cannot be greater than maximum MCS supported by any resource block r at that time. Moreover, equation (5), ensures that a single MCS level is chosen for a user at time $t$. Constraint (6) ensures that scheduling order determined by MVNO scheduler is maintained in allocating resources. Equation (7) meets the minimum data rate requirement for each user belonging to a MVNO. Equation (8) indicated the maximum data rate achieved by the resources allocated to MVNO is under the maximum allowable throughput for MVNO.

In this section, we develop the MCS aware RAN slicing algorithm based on a greedy paradigm.

As proven in Theorem 1, the MaRS problem is NP-hard and it is not feasible to find a polynomial-time optimal solution. Therefore, it is vital to develop an upper bound for the objective function defined in Exp. (2). This upper bound can be used as a benchmark to measure the performance of the scheduling algorithm that we presented in section V.

When $R$, the maximum number of resource blocks in time $T$ is given, our problem aims to find a subset of $R$ for each user. Choice of this subset depends on $c$, the MCS selected for them. Note that, if we want to maximize Exp. (2), we need to find a subset which contains least amount of RBs.

Since we want to find an upper bound for objective function in Exp. (2) , let us consider a fictitious scenario of excellent channel condition for every user in time $T$. Therefore, every resource block $r\in R$, can support the maximum MCS value that a user can support during its allocation. That is,

We further consider that the MVNO scheduler always uses the maximum MCS for each user. That is,

We then proceed with the allocation of the resource blocks for each user following the slicing list, $\mathcal{L}^{t}$. In this fictitious scenario, the data rate achievable for each user in time $T$, is directly proportional to the number of resource blocks allocated to it at time $T$. Therefore, we can re-write constraints (7), (8) as,

We can see that, the criterion to meet each user’s minimum data rate requirement completely depends on number of resource blocks allocated to it. Since we assume the maximum MCS level for each resource block, any allocation of the resource block with MCS $d_{u^{i}_{m}}^{r,c,t}$, for a user to meet $\Lambda_{m}^{i}$ would use least amount of resource blocks. Therefore, if we use the least amount of resource blocks for every user, we can find the maximum users that can be supported by the given set of resource blocks $R$ for time slot $T$.

In this section, we developed a very intuitive based upper performance bound for the MaRS algorithm. In the following section, we perform simulations of this upper bound and compare the performance of MaRS algorithm with it.

In this section, we assess the performance of the MaRS algorithm proposed in Section VI. We evaluate MaRS algorithm in terms of its ability to achieve our objective function of maximizing the users served by varying various 5G network parameters. We use the upper bound developed in section VII as the benchmark for this purpose.

In this paper, we investigated the problem of RAN slicing in multi-MVNO environment with varied users having minimum data rate requirement as a specification for the users. First, we discussed the SD-RAN architecture and discussed its operation flow. Then, we formulated the MCS-aware RAN Slicing (MaRSP) problem as optimization problem with an objective function to increase the number of supported users at each time slot $T$. We proved that MaRSP problem is NP-Hard. Next, we developed the novel MCS-aware RAN Slicing (MaRS) algorithm where we maximize the data rate for each user at each TTI and assign resources to it based on a greedy paradigm. We also showed that the MaRS algorithm has a polynomial time complexity. Following that, we developed a upper performance bound for the MaRS algorithm by considering no frequency and time correlation. Finally, we carry out thorough evaluation of the MaRS algorithm under various network and channel scenarios. Results conclude that the proposed slicing algorithm achieves near-optimal performance when compared with the upper bound. Through various simulation settings, we have also shown that MaRS algorithm is easily scalable. In compliance with the O-RAN architecture, we have seen through results that MaRS algorithm can be applied to non-real-time and near-real-time RIC deployments. Using RB utilization as a metric, we have compared the performance of MaRS algorithm with other static allocation algorithms. We see that MaRS algorithm outperforms many static allocation algorithms by using least amount of resource blocks to serve minimum data rate requirement for each user.