A Hybrid SDN/NFV Architecture for Future LTE Networks

Transcription

1 A Hybrid SDN/NFV Architecture for Future LTE Networks Ali Tawbeh, Haidar Safa, and Ahmad R. Dhaini Department of Computer Science American University of Beirut Beirut, Lebanon {at58, hs33, Abstract In the LTE EPC, many network entities and interfaces have to be maintained and updated regularly. Moreover, to accommodate more users, new hardware must be integrated, although rarely used. To address these challenges, the EPC can be moved to the cloud using two modern technologies: SDN and NFV. In this paper, we study the impact of integrating these novel technologies on LTE networks. We propose a hybrid approach for selecting whether to apply NFV or SDN on each gateway at a given time while minimizing the network load taking into consideration key network parameters such as the number of active datacenters, the deployment city population, the intensity at a given time, the QoS class identifier (QCI), and the delay budget. We formulate the SDN decomposition/nfv virtualization selection as an optimization problem where the objective is to minimize the network load subject to a set of constraints. Our results show that our proposed solution is more responsive to the dynamic state of the network such that for a given gateway, at a certain time slot, an SDN decomposition might be the optimal choice; while at another time slot with a different network state, the NFV architecture might be more suitable. Index Terms LTE, Evolved Packet Core, SDN Decomposition, NFV Virtualization. I. INTRODUCTION Long Term Evolution (LTE) comprises two main entities: the evolved UMTS terrestrial radio access network (E- UTRAN) and the Evolved Packet Core (EPC) [12]. The EPC is mainly composed of the home subscriber server (HSS), the packet data network gateway (PGW), the service gateway (SGW), and the mobility management entity (MME). The PGW handles the communication between the EPC and packet data networks. The SGW plays the role of a router of data from the enb to the correspondent PGW. The LTE EPC has a static fit-to-purpose architecture. Each hardware component in the EPC is employed to perform one specific task, which means adding more functionalities will require integrating more hardware entities in the core network. Also, accommodating a large number of users during peak times is done via duplicating entities that may not be necessarily used during idle times. In this context, upgrading the EPC will result in more complex architecture and protocols leading to an increase in the capital and operational expenditures; meanwhile the average revenue per subscriber is expected to keep decreasing [15]. Motivated by the core principles of two promising technologies, namely Network functions Virtualization (NFV) and Software Defined Networking (SDN), moving the EPC to the cloud helps achieve cost reduction for the benefit of enlarging revenues margin [9], [13]. However, despite their potential in building robust, reliable and high performance service delivery networks, little attention has been given to the impact of virtualizing LTE EPS using NFV and SDN on the network load and data plane. In addition, even though integrating NFV in mobile core networks may be advantageous, it requires steering all data traffic to a datacenter where the functions are virtualized; this adds extra load on the transport network and results in longer delay on the data plane depending on the location of data centers. The load overhead in the network increases proportionally with the amount of control that SDN is granted due to the signaling between the controller and the data forwarding elements. In this paper, we study the impact of integrating SDN and NFV in LTE networks and propose a hybrid architecture for SGW and PGW gateways, which applies both the SDN decomposition and NFV concept on each gateway while minimizing the network load. The remainder of the paper is organized as follows. In Section II, we briefly introduce SDN and NFV and discuss some related work. In Section III, we present the proposed approach and formulate the selection problem between SDN decomposition and NFV virtualization as an optimization problem where the objective is to minimize the network load subject to a set of constraints that include the number of active datacenters, the population of the city of deployment, the intensity at the given time, the QoS class identifier (QCI) and the volume of generated traffic in addition to the packet delay budget. In Section IV, we evaluate the performance of the proposed approach. Finally, we present our conclusions and future work in Section V. II. BACKGROUND AND RELATED WORK In SDN, the infrastructure layer consists of programmable switches and routers that perform data forwarding. On top of the infrastructure layer and below the application layer, there is a control layer that administers the infrastructure layer via an open protocol interface such as OpenFlow [11]. The application layer implements customized business applications and network services. This separation abstracts the infrastructure for network services and applications so as the network appears as a single logical switch maintained by centralized /17/$ IEEE

2 software-based SDN controllers [10]. NFV changes the current practice of buying and installing new hardware to integrate new functions or services, by implementing network functions as software that can be run on general purpose servers with the ability of instantiating and moving instances of these functions between different datacenters at different locations as required, without the need of installing new hardware [13]. NFV and SDN can be implemented separately or be combined together to achieve greater value. By separating control and data planes, SDN can enhance performance, and introduce flexibility and simplicity in resolving compatibility issues and maintenance through programmability and centralized control. In turn, NFV can provide the infrastructure as virtualized functions implemented as software instances that can be connected and controlled using SDN [13]. Most related work [1] [8] lack analyzing the impact of virtualization on EPC s performance. A lightweight mobile cloud Offloading Architecture is presented in [8], which utilizes the virtualized resources hosted in datacenters only to offload the network traffic within the mobile core. A qualitative analysis of the benefits of SDN and NFV to the mobile core are discussed in [1]. Additionally, [6] studied the influence of SDN and NFV on the gateways and datacenter placement within the mobile core, however considering only uniform traffic demands that should not vary with respect to time and intensity of the area of deployment. This is a major limitation in the approach. In fact, demands are strongly correlated with time variations and intensity of the area being served. A mobile network architecture with virtual components and SDN control, which offers a fine-granular control of the available resources was proposed in [7]. NFV and SDN concepts were applied on the high volume data-plane within the mobile core network. Even though both [6] and [7] proposed applying NFV on SGW and PGW, however only [6] suggested decomposing the gateways between control plane and data plane using SDN. Both works quantify and minimize the network load, but unlike [6] which assumes uniform demands, [7] considers time-dependent and population dependent demands. Also, [7] proposed models that take into consideration datacenter available resources and save power, but only considering full virtualization; i.e., all the gateway s functions are implemented in a datacenter and the gateway is replaced by a basic SDN networking element that steers traffic to different datacenters. The drawback of such an architecture is the impact on delay-critical functions. Moreover, both [6] and [7] did not consider the characteristics of the bearers being established despite the fact that the bearer packet delay budget and resource requirements may differ depending on the QCI of the bearer and its other QoS parameters. Finally, both approaches do not take into account datacenter available resources and locations which will certainly affect the total load in the network, and increase data plane delay. III. PROPOSED SOLUTION In this section, we propose a hybrid architecture where SDN decomposition and NFV virtualization are applied on every gateway. We also use the LTE QCI to determine the delay budget for each set of bearers. We then formulate the data center placement problem as an optimization problem and describe the computation of the parameters. A. Hybrid Network Architecture Figure 1 illustrates the proposed hybrid architecture. Here, interconnected gateways are replaced by networking elements (NE), each connected to a datacenter. The left NE replaces an SGW, while the right NE replaces a PGW. L1 denotes the length of the path between the NE replacing the SGW and the datacenter, L2 is the length of the path between the two NEs, and L3 is the length of the path between the NE replacing the PGW and the datacenter. Fig. 1: Hybrid architecture The application of SDN decomposition on a gateway necessitates enhancing the NE to support the gateway data plane functions, such as GTP. The control plane functions of each gateway are implemented as software instances in the datacenter, depicted as SGW-C and PGW-C. The NEs are connected to the control plane instances via the SDN controller, namely CTR (which also resides in the datacenter), via the dashed links which are dedicated to transport control messages and exchange flow tables rules (no user data on this link). In the NFV scenario, the data (SGW-U and PGW-U) and control planes (SGW-C and PGW-C) functions of each gateway are implemented as software instances in the datacenter, so as the NEs are only used to forward packets from/to the datacenter and between them. The NEs are connected to the user plane instances in the datacenter using solid lines links to transport user data plane packets. The hybrid architecture can be achieved by implementing a single control plane instance for each gateway since both the SDN and NFV require running the control part as software application in the datacenter. While there are two implementations for the data plane functions, the one in the datacenter gets activated in case of NFV deployment, while the other one is implemented in the NEs and gets activated in case of SDN deployment. For example, in metropolitan (crowded) areas, if an SGW is deployed, in peak times (e.g., Monday afternoon), the NFV deployment can be used to minimize the network load; however, this does not imply that the virtualized deployment will be used at all times. The hybrid architecture allows for the activation of the SDN decomposed deployment during in light-loaded periods (e.g., late at night) to decrease the average packet delay.

3 B. QoS Consideration In LTE systems, data is transported using bearers according to certain quality of service (QoS) requirements. The QoS requirements are enforced via a QoS Class Identifier (QCI), which maps the bearer to four metrics: the resource type (i.e., guaranteed-bit-rate, GBR or non-gbr), packet error/loss rate, packet delay budget, and QCI priority. The number of QCI values is nine [14]. Our proposed hybrid architecture employs SDN decomposition and NFV on each gateway, which gives a more granular control over the network and makes it responsive to the dynamic state of the network. Namely, for a given gateway, at a certain time slot, an SDN decomposition may be the optimal choice for the current network state (to decrease delay); however, in another time slot, the state of the network may change and thus the NFV architecture may be more suitable (to decrease delay). The inclusion of QCI-enabled demands may have one deployment suitable for a set of demands, but it might not be the case for others. This asserts the need of a hybrid architecture, which, for a given gateway, at a given time, a set of QCIs may be operating on one deployment, while the other set will be operating on the other deployment. C. Problem Formulation Our objective is to minimize the total network load by finding the optimal data centers placement in each time slot for every demand given certain packet delay budget constraints. Optimal datacenter placement means that in every time slot, for each set of bearers having the same QCI of the same demand, we must find the optimal location of the datacenter and choose which deployment to activate, given a certain delay budget. The possible datacenter locations are where the operator already has a deployment. We adopt the QCI standard classes in order to classify the data flows of each demand since it helps to identify the threshold of packet delay and to estimate the traffic volume when combining it with other QoS parameters, namely the MBR, GBR, UE-AMBR and APN-AMBR. We define the following notations: Q: the set of standardized QCI values (from 1 to 9). C: the set of datacenter locations. D: the set of demands. T : the set of time slots. P : the set of paths. K: the number of data centers. In our model, Q is the set of QoS standardized classes that can be assigned to bearers in the LTE s core network. P is the set of all possible paths that a set of bearers might take between the SGW and the corresponding PGW. In total, there are four possible paths: 1) Between a virtualized SGW and a virtualized PGW. 2) Between a virtualized SGW and a decomposed PGW. 3) Between a decomposed SGW and a virtualized PGW. 4) Between a decomposed SGW and a decomposed PGW. Our goal is to minimize the total network load by choosing for each QCI q of bearers established on demand d, a datacenter c with a path p at each time slot t. The problem can be formulated as follows: minimize δ q,c,d,t,p N q,c,d,t,p q Q c C d D t T p P where δ q,c,d,t,p is a binary variable that is set to one if at time t, the bearers of QCI q of the demand d are assigned for the datacenter in location c on the path p. N q,c,d,t,p is a pre-calculated load for the combination q, c, d, t and p. The constraints of the minimization problem are: δ c = K (1) c C δ q,c,d,t,p δ c q Q, d D, c C, t T (2) p P c C p P δ q,c,d,t,p = 1 q Q, d D, t T (3) c C p P δ q,c,d,t,p L q,c,d,t,p L budget q Q, d D, t T Constraint (1) ensures that K datacenters are under operation such that δ c is a binary variable that determines whether a datacenter c is selected to be under operation. Constraint (2) ensures that in case a datacenter c is chosen, a path p P can be selected for bearers of QCI q Q, which are established for demand d in time slot t. Constraint (3) forces the selection of a single path p and a single datacenter c for each QCI q for demand d in time slot t. The traffic delay budget is met by constraint (4); for QCI q s bearers of demand d in time slot t, the delay produced by choosing a datacenter c and path p, namely L q,c,d,t,p, must remain under q s delay budget. D. Calculating the Problem Parameters The next step is to quantify the pre-calculated problem parameters, namely the network load N and the latency, for each combination of QCI q, datacenter location c, demand d, time slot t, and path p. 1) Calculating network load: Similar to [7], the traffic of a city ct CT, where CT is the set of considered cities in time slot, t, can be represented as the product of the intensity in time slot t, denoted i(t), and the population of the city, p(ct): f(ct, t) = i(t) p(ct). The traffic, T R, at an SGW, which is equivalent to the traffic caused by a demand since a demand is defined between each SGW and its PGW, is expressed as: T R d,t = T R SGW (t) = f(ct, time ct,sgw (t)) b ct,sgw, ct CT (5) where f(ct, time ct,sgw (t)) is the traffic of city ct at time t, and b ct,sgw is a boolean value that is set to 1 if and only if city ct is covered by the considered SGW. However, this formula ignores that the demand is composed of bearers belonging to different QCIs; thus, not all of them have the same impact on the total network load. Also the formula does not take (4)

4 into account the extra load added by the control plane when adopting a path p where one or both gateways are decomposed. Therefore, we reformulate Eq. (5) to reflect the impact of each QCI q on the network load by integrating the average bit rate of bearers belonging to q, denoted as BR avgq. To account for the load added by the control plane of the SDN decomposition, we define a coefficient α denoting the SDN control volume percentage of the traffic generated by an SGW. This percentage depends on the protocol adopted by the operator. The load added by SDN control plane also depends on the chosen path due to the fact that when choosing a path where both gateways (the SGW and the PGW) are decomposed, the amount of control messages is roughly double the amount of when a single gateway is decomposed. When none of the gateways is decomposed, the SDN control messages are absent. Therefore depending on the chosen path p and the SDN control volume α, the traffic must be multiplied by a coefficient β p,α, where β can be expressed as follows: β p,α = 1 + γ α (6) where γ is the number of gateways decomposed in path p. Consequently, the traffic volume generated by the bearers of QCI q, taking the path p, on the demand d, at a time t, is expressed as follows: T R q,d,t,p = ( ct CT f(ct, time ct,sgw(t)) b ct,sgw +BR avgq ) β p,α (7) BR avgq is the average bit rate of all bearers constituting demand d and having a QCI q; this value can be computed based on MBR and GBR value. We add the bit rate to the equation in order to account for the considered QCI. Therefore the traffic of two different QCIs will have two different values. The QCI with greater bit rate will have a greater traffic value. Consequently, the network load N q,c,d,t,p can be expressed as follows: N q,c,d,t,p = T R q,d,t,p length c,d,p (8) where length c,d,p is the length of the path p for demand d passing through the datacenter at location c (i.e., the distance of the path packets take). Depending on the path p, length c,d,p can be calculated as follows: If both SGW and PGW are virtualized, then the communication between the NEs will happen through the datacenter. Thus, length c,d,p = L1 + L3. If the SGW is virtualized and the PGW is decomposed then the NE replacing the SGW needs to communicate with the data center when receiving packets while the communication between the NEs happens through the direct link since PGW is decomposed and therefore the data plane de[ployed in its NE is activated. Thus, length c,d,p = 2 L1 + L2. If the SGW is decomposed and the PGW is virtualized then the length of the path is the same as in the previous case, however L3 will be multiplied by two instead of L1. Thus, length c,d,p = 2 L3 + L2. If both SGW and PGW are decomposed, then the communication between the NEs will happen directly between them. Thus, length c,d,p = L2. Since real deployment depicting the routing process between gateways has not been yet reported in the literature, we abstract the path lengths between the gateways and use euclidean distances; this does not affect the selection of a path versus another. 2) Calculating network delay: T proc is the time needed to process the packet at each network node, i.e., at the SGW and the PGW, which is subject to: 1) the demand since it specifies the involved gateways and the cities connected to them which affects the number of established bearers, 2) the chosen path since it determines whether each gateway is virtualized or decomposed, and 3) the time slot since the number of active bearers varies according to time. In our performance evaluation, we use T proc values that were estimated in [6] for virtualized gateways and decomposed gateways. T prop is the propagation delay time on each link between the SGW and the PGW, which is subject to: 1) the demand since it determines the involved gateways and hence their location, 2) the datacenter location, and 3) the chosen path because it determines how to calculate the length. The total network delay between the access edge, i.e. SGW, and the IP domains gateway, i.e. PGW, can be expressed as L c,d,t,p = T procd,t,p + T propc,d,p. IV. PERFORMANCE EVALUATION To evaluate the performance of our solution under real network scenarios, we developed a simulation model that depicts the US mobile core gateways based on the US LTE coverage map [16] and the US population map [17]. We also use traffic delay and network load as performance metrics. The clustered topology used for performance evaluation is based on the one presented in [6], [7]. It is composed of 4 PGWs represented as red rectangles, and 18 SGWs represented as green rectangles. We assume that there exists a demand between each SGW and PGW pair; thus, we have D = 18 demands in total. Each gateway is identified by a unique ID number. The distances between the gateways are measured using the measure distance map on FreeMapTools [18]. Finally, we assume that any two gateways can be interconnected (i.e., the network is meshed). This assumption is important as it enables the gateways to connect to the PGW when it is selected as a datacenter. To decrease the optimization problem solver s running time so as to retrieve results instantly, we obtain the values of the parameters N q,c,d,t,p and L q,c,d,t,p offline; this is done via calculating them for all the combinations of QCIs, datacenters locations, demands, time slots and paths. In LTE, there are 9 QCIs, thus Q = 9; there are 22 possible datacenter locations (18 SGWs and 4 PGWs), thus C = 22. We assume T s unit

5 is in hours; thus to cover a full day, T = 24. Finally, there are four possible paths for each demand as described in the previous section; thus, P = 4. A. Traffic and Network Load Measurement To calculate the total traffic on each SGW, we used the population sizes presented in Table I; these are extracted from the US coverage map via summing the population sizes of all cities situated in the vicinity of each SGW. To obtain the network usage for every time slot, we followed the daily traffic intensity presented in [19]. SGWs sgw1 sgw2 sgw3 sgw4 sgw5 sgw6 Population (in Millions) SGWs sgw7 sgw8 sgw9 sgw10 sgw11 sgw12 Population (in Millions) SGWs sgw13 sgw14 sgw15 sgw16 sgw17 sgw18 Population (in Millions) TABLE I: Population of each SGW For evaluating Eq. (7), we estimate the average bit rate of each QCI q per the example services for each QCI value reported in [14]. Based on these values, we estimate the bit rate of each QCI value as shown in Table II. QCI BR avg 64kbps 384kbps 16kbps QCI BR avg 20Mbps 1kbps 19Mbps QCI BR avg 384kbps 20Mbps 20Mbps TABLE II: Estimated bit rate of each QCI value To calculate β p,α, we set the value of α to 10% [6]; γ is computed as detailed in Eq. 6. Finally, to compute the network load N q,c,d,t,p for each parameters combination, we compute the path length length c,d,p for each combination based on the assumed US core network. B. Delay Measurement As aforementioned, the traffic delay between the mobile core gateways L c,d,t,p is computed as the sum of T prop (which is distance over speed) and T proc, which is computed based on the work reported in [7], summarized in Table III. The table shows a correlation between the number of established bearers and the processing delay for a virtualized gateway. Indeed, a higher number of bearers results in longer processing delay. However, the processing delay for decomposed gateways remains constant. No. of bearers k 10k bits/sec 1 M 10 M 100 M 1 G packets/sec K 83K Virtualized GW T proc 62 µs 83 µs 109 µs 132 µs Decomposed GW T proc 15 µs 15 µs 15 µs 15 µs TABLE III: Average processing delay The number of bearers depends on the population of the cities within the demand s gateway coverage range and the intensity of the considered time slot. For example, in crowded cities, at peak times, the number of established bearers is higher than at off peak times. Also the number of bearers established in cities, is higher than its value in suburbs. Therefore, we compute the total number of bearers at time t for a population pop as follows: B t,pop = intensity(t) pop σ where σ is a parameter used to normalize the output of the equation; it can be determined empirically. In our simulations, we set σ = 500 because this value gives a number of bearers proportional to the population used. To generate a number of bearers for each QCI of each demand we divide the total number over the number of QCIs. To do so, we split B t,pop into nine random numbers, and then assign one random number to each QCI value. LTE specifies delay budgets for each QCI based on the latency between user s UE and the server running the service; thus the specified values are relatively high. In our simulations, we used a fixed value of 4.95ms determined empirically on the presumed core network topology, in a way that any smaller value will cause constraint (4) not to be met, resulting in infeasible model. Relaxation on that value was then applied such that the delay budget values for each QCI are proportional to the ones in [14] but normalized to the order of the fixed value. For example, the values 50ms, 100ms, 150ms and 300ms, were mapped to the values 4.95ms, 5.0ms, 5.1ms and 5.3ms, respectively. C. Results We first consider a topology with a single data center. Fig. 2 shows the topology of the US core network after running the optimization problem at time slot 15 (between 3:00 pm and 4:00 pm) where the intensity is at its highest value (0.86). The chosen datacenter is PGW with ID 2 circled in orange. The figure also shows the path taken by each demand for QCI 3. We observe that all the QCIs of each demand took the same path, so for simplicity we only depict the paths of QCI 3. The path of each demand is represented with a different color. For the SGWs that were not originally connected to the PGW selected as datacenter, the path goes from the SGW to the datacenter then back to the PGW that it is originally connected to it. For example, in Fig. 2, the path of the demand generated by SGW 18 goes to PGW 2, which is selected as a datacenter, then goes back to PGW 4, which SGW 18 is originally connected to. For the SGWs that were originally connected to the PGW that is chosen as the datacenter, there is a single line going from the SGW to the PGW. This is, for example, the case of SGW 11 in Fig. 2, where the traffic goes directly to the PGW that is connected to (i.e., PGW 2), which is selected as a datacenter. Regarding the types of paths, each demand was either virtualized (both gateways are virtualized), represented by solid line, or decomposed (both gateways are decomposed), represented by dashed lines. We conclude that the farthest SGWs from the datacenter (i.e., the SGWs that belong to other PGWs than the one chosen as the datacenter) have taken the full decomposed path, while the nearest SGWs (9)

6 IEEE ICC 2017 Communications Software, Services, and Multimedia Applications Symposium have taken the virtualized path. This result is expected and can be justified as follows: the demands from the farthest SGWs will face higher propagation delay, therefore to remain within the delay budget, the compensation happens by choosing the decomposed path since it requires lower processing delay than the processing delay required by a virtualized path. additional control layer, at the expense of increasing the traffic delay. In our future work, we aim to study how the objective function might vary with respect to time for different number of datacenters. Also, we want to analyze the variation of runtime with respect to constraints, time slots, number of demands, and number of active datacenters. Moreover, we plan to add further mobile core network components such as the MME, with control plane delay budgets. R EFERENCES Fig. 2: The new US core network topology at time slot 3:00-4:00 pm, for QCI 3 and with one datacenter Fig. 3 illustrates the effect of increasing the number of datacenters to two. It shows that the new datacenter is PGW 1, it is chosen instead of PGW 3 or 4 since PGW 1 s SGWs are much more populated than the others because they serve big cities like San Francisco or New York, allowing its SGWs to connect to it via taking virtualized paths to minimize the total network load as much as possible. Fig. 3: The new US core network topology at time slot 3:00-4:00 pm, for QCI 3 and with two datacenter V. C ONCLUSIONS AND F UTURE W ORK In this paper, we studied the problem of virtualizing the LTE EPC using SDN and NFV. We proposed a hybrid architecture that applies both technologies on each gateway, and finds the optimal path for each set of bearers with the same QCI between each connected SGW and PGW without impairing the QoS requirements. Our simulation results showed that the closest SGWs to the datacenter took virtualized paths while the farthest took SDN decomposed paths. This asserts the fact that SDN decomposition decreases the network delay while it increases the total network load; in contrast, an NFV gateway does not increase the network load due to the absence of an [1] A. Basta, W. Kellerer, M.Hoffmann, K. Hoffmann, and E-D. Schmidt. A Virtual SDN-Enabled LTE EPC Architecture: A Case Study for S/P-Gateways Functions. In the proc. of 2013 IEEE SDN for Future Networks and Services (SDN4FNS 2013), November [2] J. Kempf, B. Johansson, S. Pettersson, H. Lning, and T. Nilsson. Moving the Mobile Evolved Packet Core to the Cloud. In the proc. of IEEE 8th International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob 2012), October [3] K. Pentikousis, Y. Wang, W. Hu. MobileFlow: Toward Software-Defined Mobile Networks. IEEE Communications Magazine, 51(7), pp , July [4] X. Jiny, L. Erran Li, L. Vanbevery, and J. Rexford. SoftCell: Scalable and Flexible Cellular Core Network Architecture. In the proc. of 9th ACM Conference on Emerging networking experiments and technologies, December [5] S. Ben Hadj Said et al., New Control Plane in 3GPP LTE/EPC Architecture for On-Demand Connectivity Service. In the proc. of 2nd IEEE International Conference on Cloud Networking (CloudNet 2013), November [6] A. Basta et al., Applying NFV and SDN to LTE mobile core gateways, the functions placement problem. In the 4th workshop on All things cellular: operations, applications, and challenges, August [7] A. Basta et al., SDN and NFV Dynamic Operation of LTE EPC Gateways for Time-varying Traffic Patterns. In the proc. of 6th International Conference on Mobile Networks and Management, September [8] A. Banerjee el al., MOCA: A Lightweight Mobile Cloud Offloading Architecture. In Proceedings of the 8th ACM international workshop on Mobility in the evolving internet architecture - MobiArch 13, [9] A. Hakiri, A. Gokhale, P. Berthou, D.C. Schmidt, and T. Gayraud. Software-Defined Networking: Challenges and research opportunities for Future Internet. Computer Networks, 75, pp , December [10] Software-Defined Networking: The New Norm for Networks, white paper. ONF, April [11] OpenFlow Switch Specification: Version 1.2 (Wire Protocol 0x03). ONF, December [12] Christopher Cox. An Introduction to LTE: LTE, LTE-advanced, SAE, and 4G Mobile Communications. John Wiley & Sons, [13] H. Hawilo, A. Shami, M. Mirahmadi, and R. Asal. NFV: State of the Art, Challenges, and Implementation in Next Generation Mobile Networks. IEEE Network, 28(6), pp , November [14] M. Alasti, B. Neekzad, J. Hui, and R. Vannithamby. Quality of service in WiMAX and LTE networks. IEEE Communication Magazine, 48(5), pp , May [15] Tellabs end of profit study executive summary. Tellabs Tech. Rep., January [16] LTE Coverage Map [17] LTE Community Facts. facts.xhtml [18] Free Map Tools Measure distance map [19] C. Rossi et al., In Proc. of the 9th ACM conference on Emerging networking experiments and technologies - CoNEXT 13, December 2013.