Network Slicing on 5G Vehicular Cloud Computing Systems

: Fifth generation Vehicular Cloud Computing (5G-VCC) systems support various services with strict Quality of Service (QoS) constraints. Network access technologies such as Long-Term Evolution Advanced Pro with Full Dimensional Multiple-Input Multiple-Output (LTE-A Pro FD-MIMO) and LTE Vehicle to Everything (LTE-V2X) undertake the service of an increasing number of vehicular users, since each vehicle could serve multiple passenger with multiple services. Therefore, the design of efﬁcient resource allocation schemes for 5G-VCC infrastructures is needed. This paper describes a network slicing scheme for 5G-VCC systems that aims to improve the performance of modern vehicular services. The QoS that each user perceives for his services as well as the energy consumption that each access network causes to user equipment are considered. Subsequently, the satisfactory grade of the user services is estimated by taking into consideration both the perceived QoS and the energy consumption. If the estimated satisfactory grade is above a predeﬁned service threshold, then the necessary Resource Blocks (RBs) from the current Point of Access (PoA) are allocated to support the user’s services. On the contrary, if the estimated satisfactory grade is lower than the aforementioned threshold, additional RBs from a Virtual Resource Pool (VRP) located at the Software Deﬁned Network (SDN) controller are committed by the PoA in order to satisfy the required services. The proposed scheme uses a Management and Orchestration (MANO) entity implemented by a SDN controller, orchestrating the entire procedure avoiding situations of interference from RBs of neighboring PoAs. Performance evaluation shows that the suggested method improves the resource allocation and enhances the performance of the offered services in terms of packet transfer delay, jitter, throughput and packet loss ratio.


Introduction
Nowadays, 5G Network Slicing is receiving increased attention from both the research and the industrial communities. Several research challenges of the 5G Network Slicing must be addressed, including the Radio Access Network (RAN) Virtualization [1] and the End-to-End Slice Orchestration and Management [2].
RAN Virtualization is one of the most important challenges that have arisen. Many solutions allow spectrum division while providing isolation of radio resources. However, the usage of the available resources is not always efficient and reliable. Another challenge • A three layer architecture is introduced allowing the optimal allocation of communication resources to both Guaranteed Bitrate (GBR) and non-Guaranteed Bitrate (non-GBR) services. • SDN controllers maintain Virtual Resource Pools allowing their underlying PoAs to commit the necessary of communication resources to support users' services. • The network slicing is performed considering the satisfaction grade of the user services.
• Fuzzy logic is applied for the estimation of the satisfaction grade of the user services by taking into consideration both the Quality of Service (QoS) of user services and the energy consumption of the user equipment.
The remainder of the paper is as follows: In Section 2 the related research literature is revised, while Section 3 presents the proposed scheme for performing network slicing on 5G-VCC architectures. Section 4 presents the simulation setup and Section 5 describes the evaluation results. Finally, Section 6 concludes the discussed work.

State of the Art
As the International Telecommunication Union (ITU) and the Fifth Generation Public Private Partnership (5G-PPP) have specified, 5G vehicular networks support three main categories of modern communications, namely the Enhanced Mobile Broadband, the Massive Machine-type Communications and the Critical Communications (or Low Latency Communications) [7]. In order to meet the new standards and requirements, many 5G architectures have been proposed, each serving a different use case. Furthermore, according to the Next Generation Mobile Network Alliance (NGMN), network architecture should be approached with more flexible and efficient techniques such as Network Slicing.
Initially, only Core Network (CN) slicing [8] was proposed in the academic and industrial domain. Thereafter, the NGMN also supported End-to-End (E2E) slicing [9] in order to enhance both the performance of the Radio Access Network (RAN) and the Core Network (CN). Specifically, the NGMN proposes a three-layer network management architecture, which consists of the Infrastructure Resource layer, the Business Enablement layer and the Business Application layer. According to the NGMN, a Network Slice represents an instance of a service consisting of a specific structure, configuration and workflow. Each Slice employs multiple instances of logical subnetworks. Each subnetwork fulfils the diverse functions and requirements requested by the particular service. The whole process is organized by an End-to-End (E2E) MANO entity [6]. Additionally, the architecture proposal of the 5G Infrastructure Public Private Partnership (5G-PPP) describes in detail the relations among the components of the 5G network. Specifically, the proposed architecture consists of five different layers namely the Infrastructure, the Network Function, the Orchestration, the Business Function and the Service Layer. As follows, in 5G-PPP the MANO entity consists of a separate layer. Furthermore, the Business Application layer of the NGNM architecture has been replaced in the 5G-PPP by two separate layers, namely the Business Function layer and the Service layer.
In Reference [10] the guidelines of the ITU, the NGMN and the 5G-PPP are considered and a three-layer framework for Network Slicing implementation is proposed. Specifically, the framework consists of the Service, the Network Function and the Infrastructure layers. The service layer can be approached in two ways: Either it defines the service level description by a set of SLA requirements and services, or it defines a composite structure identifying the functions and RATs that should be used by the slice. The difference between the two lies in the process of the slice generation: In the first scenario, the slice implementation necessitates the application of efficient operations to meet the requirements described. In the second scenario, the slice will be implemented in a simpler way, which, however, might be less effective.
Subsequently, the Network Function layer includes both the control and the user planes of the network architecture. Specifically, it includes every operation that has to do with the configuration of the network functions and data forwarding.
The Infrastructure layer refers to the physical network that includes the Radio Access Network (RAN) and the Core Network (CN). Also, it implements functionalities for the management and the allocation of the network resources.
Furthermore, in this framework a MANO entity coexists that converts use cases and service models into network slices using the network functions. The MANO entity links the use cases and service models to infrastructure resources, while at the same time it keeps configuring and monitoring each slice during its life cycle.
In addition to the aforementioned works, several network slicing schemes have been proposed from the research literature. Each scheme applies either non-QoS aware or QoS aware strategies to perform the resource allocation.
In Reference [15] three network slicing methods are studied. In the first one, which is called Static Allocation (SA) method, the required number of RBs is estimated for each slice considering its service constraints. Subsequently, each slice allocates its RBs to vehicular users, by applying the PF resource allocation algorithm. In the second method called Allocation of Ordered Slices (AOS) slices are prioritized. The requirements of slices with higher priority are processed first, while the PF algorithm is applied to allocate RBs to the users of each slice. Finally, the third method called Impartial Allocation (IA) is an extension of the AOS method. In this approach, each slice selects which RBs will be allocated to its users considering the channel quality of each RB. Thus, slices with higher priority allocate RBs with higher channel quality, increasing the quality of their services.
In Reference [16] the Reliable Software-Defined RAN (RSDR) scheme for network slicing which applies an improved version of the RR algorithm is proposed. A weigh for each slice is calculated in each TTI and the slices with higher weight gain priority against others to the allocation of RBs. For the allocation of RBs to the available slices, the improved version of the RR scheme is applied based on the priority weight of each slice.
Regarding the QoS aware strategies for resource allocation, several schemes have also been proposed optimizing the performance of modern network environments. Indicatively, the Modified Largest Weighted Delay First (MLWDF) [17] and the Exponential/PF (EXP/PF) [18] QoS aware algorithms extend the PF metric by taking into consideration network factors such as the head of line delay and the packet loss ratio.
In Reference [19] three types of slices are defined, namely the Fixed, the On-demand and the Dynamic slices. Firstly, dedicated communication resources are allocated to the Fixed slices considering the Service Layer Agreement (SLA) of each user. Subsequently, available communication resources are allocated to the On-demand slices based on their QoS constraints by applying an alternative version of the MLWDF algorithm where Equation (1) is used for the estimation of the scheduling metric. It has to be noted that the α i value is determined by Equation (2) where δ i is the target packet loss ratio and τ i is the delay constraint. Additionally, the D HOL,i parameter is the Head of Line (HOL) delay observed for the ith service slice, while the R i req k and theR i (t − 1) parameters represent the required throughput and the past average throughput of the slice. Finally, the remaining resources are allocated to the Dynamic slices using Equation (3).
Another QoS aware algorithm for resource allocation is called Frame Level Scheduler (FLS) [20,21]. It implements a two level QoS aware streategy. The upper level estimates the u i (z) quota of data that the ith real time flow must transmit at the zth frame to satisfy its QoS constraints using Equation (4). In this equation, q i (z) represents the queue length in the zth frame, C i the number of coefficients used and c i (y) the value of the yth coefficient. It should be noted that the coefficients are used to guarantee that the delay constraint for each real time flow is satisfied. The number C i of coefficients is estimated using Equation (5), where τ i represents the target delay and T f the frame length. Additionally, the coefficient value c i (y) is determined by Equation (6). Subsequently, the lower level applies the PF algorithm to allocate network resources to real time flows for transmitting their quota of data. The remaining resources are allocated to best effort flows. Furthermore, some modifications of the FLS algorithm are proposed to further improve the performance of real-time services [22] and to support Cross Carrier (CC) scheduling [23,24].
Finally, in [25] the Ultra Reliable Low Latency Communication for Autonomous Vehicular Networks (URLLC-AVN) scheme is proposed for performing network slicing in vehicular networks. Specifically, each RSU contains a set of RBs, which is divided into two subsets, namely the local RBs and the shared RBs. The local RBs are allocated only to vehicles that are connected to the specific RSU, while the shared RBs of each RSU are added to a virtual resource pool. The queue delay that each vehicle perceives for his services is monitored. If the estimated delay grade is lower than a predefined threshold, then only RBs from the current RSU are allocated to the vehicle. On the contrary, if the estimated delay is higher than the threshold, additional RBs from the virtual resource pool are committed in order for the available resources to be increased and thus decreasing the delay that the vehicle perceives.
In this work, design characteristics of both the FLSA-CC [23] and URLLC-AVN [25] algorithms are adopted to accomplish optimal allocation of the communication resources to user services. Furthermore, the proposed scheme implements a Fuzzy Inference System (FIS) [26] for the estimation of the satisfaction grade of each user service considering both QoS and energy aware criteria.

The Proposed Network Slicing Scheme
The proposed network slicing scheme aims to improve the allocation of the communication resources of the access network infrastructure to maximize the satisfaction of the QoS constraints of modern vehicular services. It consists of a three-layer stack performing the allocation of communication resources to user services. The proposed slicing scheme is deployed in a Software Defined 5G network architecture, with each network component implementing specific layers of the proposed stack. In the following subsections the proposed scheme and the network architecture are described.

The Layered Design of the Proposed Scheme
The proposed slicing scheme is based on the three-layer design of the FLSA-CC [23] scheduler presented in Figure 1. Two service groups are defined, namely the Guaranteed Bit Rate (GBR) containing services with strict QoS constraints and the non-Guaranteed Bit Rate (non-GBR) containing best effort services. The upper layer of the scheme evaluates the amount of RBs that should be committed for each GBR service to succeed its QoS constraints. Subsequently, the second layer allocates RBs obtained from the upper layer to the GBR services. Finally, the third layer has been added to allocate the remaining RBs to the non-GBR services. The following subsections describe the functionalities of each layer.

The Upper Layer of the Network Slicing Scheme
A set M u (t n ) of GBR services per user u require resources to satisfy their constraints during the t n TTI. For each m u (t n ) ∈ M u (t n ) service, the S estimated m,u (t n ) indicator is defined, determining the estimated satisfaction grade of user u during the upcoming t n TTI that is the next TTI for which RBs are going to be allocated.
The S estimated m,u (t n ) value are estimated using the Mamdani Fuzzy Inference System (MFIS) system described in [27], where the E estimated m,u and the Q estimated m,u parameters are considered as inputs expressed as Interval Valued Octagonal Fuzzy Numbers (IVOFNs) [28].  Table 1. Additionally, Table 2 presents the knowledge base of the MFIS which consists of the fuzzy rules considered for the estimation of the S estimated m,u parameter. In particular, the E estimated m,u parameter represents the communication energy consumption. It is calculated using Equation (7) [29], where λ b is the cell density in the area of the user and, thus, λ b · π · R 2 represents the number of PoAs connected to a user moving in an area with radius R. It has to be noted that in our case λ b · π · R 2 is assumed to be equal to 1. Additionally, ζ(t n ) ∈ (0, 1] is a power amplifier parameter, P tr (t n ) is the transmission power of the PoA in Watts, P bs (t n ) is the fixed power at PoA in Watts, P osc (t n ) is the local oscillator power in Watts, M(t n ) is the number of used antennas, P b (t n ) is the Radio Frequency (RF) chain power [30] at the base station in Watts, C 0 (t n ) is the energy per complex operation in Joules, B(t n ) is the available bandwidth in MHz, r d (t n ) is the spectral efficiency of the downlink channel measured in bps/Hz, P cod (t n ) is the channel coding power measured in Watt Gbit/sec , P dec (t n ) is the channel decoding power measured in Watt Gbit/sec , and P d (t n ) is the RF chain power at user equipment measured in Watts.  In particular, the th estimated m,u (t n ) is estimated using Equation (9), where µ(t) represents the estimated throughput per RB and r available is the number of the RBs that are available for allocation. is the required throughput of the m th service of user u. Additionally, the w th , w d ,w j and w pl in (8) represent the weights of the aforementioned parameters. In this work, the considered weights are calculated using the Fuzzy Analytic Network Process (FANP) method described in [27] which in this work is implemented using IVOFNs. factor, respectively. Thus, for each crisp value, two membership degrees are determined in the corresponding MF, one for the upper octagon and one for the lower octagon. Table 1 represents the linguistic terms and the corresponding IVOFNs of MF E , MF Q and MF S membership functions, which are equally distributed inside the domain [U min , U max ] = [0, 1], as described in [27]. Furthermore, Table 2 presents the considered fuzzy rule base which is used from the MFIS for producing the satisfaction chart.   values is close to 0, the user satisfaction is in low levels.  (13) and (14) for the E estimated m,u and Q estimated m,u parameters, respectively. In these equations, the E max and the Q max parameters represent the maximum values that have been observed for the E and the Q parameters from the instantiation of the system and up to the current time point, while at the same time E min = 0 and Q min = 0. 3.

The Middle Layer of the Network Slicing Scheme
In each TTI, the middle layer uses an improved version of the MLWDF-CC scheduler [23], called MLWDF with Energy Awareness (MLWDF-EA), to allocate to each GBR service the number of RBs estimated from the upper level. This scheduler extends the MLWDF-CC metric to take into consideration the estimated energy consumption of each RB. Specifically, for each ith GBR flow requiring RBs, the MLWDF-EA metric of the kth RB is estimated by Equation (15). The RB that maximizes the estimated value of the MLWDF-EA metric is allocated to the ith flow. It has to be noted that the α i value is determined by Equation (2). Additionally, the E estimated i,k factor is normalized in order to obtain values inside the range (0,1]. As follows, the use of the cross carrier QoS aware MLWDF-EA scheduler realizes improved resource distribution among the GBR services. Cross carrier scheduling is deemed necessary since both the LRBs available in each PoA and the SRBs available in the Virtual Resource Pool can belong to different carriers.

The Lower Layer of the Network Slicing Scheme
The third layer allocates available RBs that exist to the Virtual Resource Pool to best effort flows using Equation (16) of the PF-CC [33,34] algorithm, where d i k (t) represents the available throughput for the ith flow in the kth RB of the t TTI andR i,j (t − 1) denotes the past average throughput.

The Proposed Network Architecture
The network architecture ( Figure 3) includes K Points of Access (PoAs) that provide network access to the vehicular users. Each k ∈ K PoA (PoA k ) has a set of Resource Blocks (RBs), a Local Monitoring Module (LMM) and a Local Allocation Module (LAM). A variable RB rem k (t n ) indicates the remaining RBs of the kth PoA during the t n Transmission Time Interval (TTI). Additionally, the RBs of the PoA are organized into two subsets, which are called Local RBs (LRBs) and Shared RBs (SRBs).  A set of SDN controllers manipulate the PoAs. In particular, each SDN controller maintains a Virtual Resource Pool where the SRBs of its underlying PoAs are stored. Additionally, each SDN controller includes its Allocation Module (AM). Furthermore, a centralized SDN controller with a MANO entity orchestrates the entire network slicing procedure maintaining the necessary frequency reuse factor. Specifically, during the instantiation of the system, each PoA communicates with the MANO entity to identify which subset of its RBs can be considered as SRB in order to be stored to the corresponding Virtual Resource Pool. Thus, the channel interference that can occur from the assignment of the SRBs with similar frequencies to neighbouring PoAs is minimized.

SDN
Based on the layered design of the proposed scheme the allocation of RBs to user services is as follows. Initially the LMM module of each PoA implements the Upper layer of the scheme to monitor the satisfaction grade S estimated m,u (t n ) of each user service. If the estimated S estimated m,u (t n ) is higher than the predefined S threshold m,u threshold value, the service requirements can be satisfied from the remaining RBs that exist in the current PoA k . Thus the LAM module of the PoA implements the Middle and the Lower layers of the slicing scheme to allocate LRBs of the PoA k to both GBR and non-GBR services for the next TTI. The remaining RBs RB rem k (t n ) are sent back to the Virtual Resource Pool which is maintained from the corresponding SDN controller. However, if the estimated satisfaction grade S estimated m,u (t n ) is lower than the predefined S threshold m,u threshold value, the service requirements cannot be satisfied from the RB rem k (t n ). Thus, the AM of the SDN controller implements the Middle and the Lower layers of the scheme to allocate extra RBs from the Virtual Resource Pool to the user services. Regarding the Virtual Resource Pool the algorithm prefers to allocate RBs with similar sub frequencies for each user to minimize the number of antennas required for each vehicle, which is a factor that affects the energy consumption of the system.

Simulation Setup
In our experiments, the 5G Vehicular Cloud Computing topology presented in Figure 4 was considered. A mobility trace indicating the map of the city of Kastoria along with road traffic data was created using the Open Street Map (OSM) software [35]. Then, the mobility trace was used as input in the Simulator of Urban Mobility (SUMO) simulator [36] allowing the production of a realistic mobility pattern for the simulated vehicles including 52,843 vehicles in total, moving inside Kastoria city in a 24 h period. The average arrival rate of vehicles was equal to 0.611608796 vehicles per second, while their average departure rate was equal to 0.61025463 vehicles per second. The network topology was built upon the map, using the Network Simulator 3 (NS3) [37]. It included a Fog infrastructure, as well as a Cloud infrastructure.    Table 3 presents the 5QI value assigned to each service, along with the corresponding constrains of each 5QI as they are defined in the 5G-PPP specifications for 5G communications [32]. Each vehicle received one flow for each service and requires RBs in each TTI to satisfy its constraints. Furthermore, an OpenFlow SDN controller provided centralized control of the entire system, while its functionality was extended with a MANO entity that orchestrated the entire functionality of the system. Table 4 presents the simulation parameters.  During the network slicing process, the user satisfaction S estimated m,u (t n ) is obtained by a lookup to the MFIS Satisfaction Chart, using the estimated E estimated m,u (t n ) and Q estimated m,u (t n ) values. It should be noted that the services' weights used for the Q estimated m,u (t n ) estimation are calculated using the FANP method [27]. The criteria used include throughput, delay, jitter and packet loss. Indicativelly, Table 5 presents the FANP pairwise comparison matrix for the ANav service, while the estimated weights for each service are depicted in Figure 5. The weights for the estimation of the Q factor for each service Figure 5. The FANP weights used for the estimation of Q estimated m,u (t).

Experimental Results and Discussion
The proposed scheme is compared with the URLLC-AVN scheme described in [25], considering parameters such as the packet transfer delay, the jitter, the throughput and the packet loss ratio. In particular, Table 6 presents the average evaluation results achieved from each scheme during the 24-h simulation, indicating that the proposed scheme achieves improved system performance in all cases. Furthermore, Figure 6 presents the average packet transfer delay achieved from each scheme for every moment during the 24-h simulation. As can be observed both schemes satisfied the delay constraint defined from the 5G-PPP specifications for each service. However, the proposed scheme achieved lower delay values in all cases. Indicatively, for the CVo service slice, the packet transfer delays observed in the case of the proposed scheme were approximately 20 ms lower than the ones observed in the case of the URLLC-AVN scheme. Furthermore, in Figures 7 and 8 the two schemes are compared considering the jitter and the throughput factors, respectively. Similar to the packet transfer delay results, the proposed scheme achieved better results ensuring higher service quality in all cases. In particular, the proposed scheme attained up to 10 ms lower jitter and up to 50 kbps higher throughput in some cases. Similarly, in Figure 9 the two schemes are compared considering the Packet Loss Ratio (PLR) factor, which is a very critical parameter for the considered services. As it is observed, the proposed scheme satisfied the PLR constraint for the entire services, for every moment of the simulation, while at the same time the URLLC-AVN scheme did not satisfy this constraint in any case. Thus, in cases where the URLLC-AVN scheme was used, critical information may have been lost leading to undesirable situations in some cases (e.g., road accident in case of ANav services).     Both schemes were capable of using shared communication resources to satisfy the requirements of user services. As a result, in most cases plenty of SRBs were available for allocation from each corresponding Virtual Resource Pool. Figure 10 presents the average saturation level observed in each Virtual Resource Pool during the simulation. In particular, the saturation levels in the Virtual Resource Pools maintained in SDN Controller 1 and SDN Controller 2 were low, since both controllers manipulated LTE-A Pro FD-MIMO Macrocells with plenty of resources, along with the other cells. Additionally, the rest Virtual Resource Pools obtained approximately from 45% and up to 65% average saturation. It has to be noted that in all cases the proposed scheme resulted in slightly higher saturation of communication resources in each Virtual Resource Pool, since it committed additional resources considering more parameters in comparison with the ones considered from the URLLC-AVN, including the throughput, the delay, the jitter and the packet loss ratio. On the contrary, the URLLC-AVN scheme committed additional resources considering only the delay factor. Thus, the URLLC-AVN scheme stopped the commitment of additional resources when the delay factor was satisfied. However, the satisfaction of the packet transfer delay did not impose the satisfaction of the additional parameters considered in the proposed scheme. As a result, the proposed scheme enhanced the resource allocation in all cases. Finally, both schemes were compared considering the total energy consumption observed for each scheme during the 24 h simulation. Specifically, Figure 11 shows that the proposed scheme achieved up to 100 Joules lower energy consumption during the simulation. This energy reduction can be explained since the proposed scheme took into consideration the energy consumption factor to perform the resource allocation.

Total Energy Consumption
Proposed Scheme URLLC-AVN Scheme Figure 11. The total energy consumption observed for each scheme.

Conclusions
In this paper a network slicing scheme for 5G-VCC systems is described. The proposed scheme improves the performance of modern vehicular services. Specifically, the QoS that each user perceives for his services as well as the energy consumption that each access network causes to user equipment are considered. Subsequently, the satisfaction grade of the user services is estimated by taking into consideration both the QoS and the energy consumption factors. If the estimated satisfaction grade is lower than a predefined threshold, additional RBs from the neighbouring PoAs are committed in order for the available resources to be increased. The orchestration of the entire procedure is performed by a MANO entity which is implemented by a centralized SDN controller. Performance evaluation showed that the proposed scheme outperforms existing solutions in terms of packet transfer delay, jitter, throughput and packet loss ratio. Future work includes the deployment of Unmanned Aerial Vehicles (UAVs) to operate as aerial relays to offload part of the RSUs tasks. Additionally, in order to further improve the proposed network slicing architecture, the design recommendations and specifications of the upcoming Sixth Generation (6G) wireless networks as well as their novel services will be studied.