Search

Mobile systems will have to support multiple AI-based applications, each leveraging heterogeneous data sources through DNN architectures collaboratively executed within the network. To minimize the cost of the AI inference task subject to requirements on latency, quality, and - crucially - reliability of the inference process, it is vital to optimize (i) the set of sensors/data sources and (ii) the DNN architecture, (iii) the network nodes executing sections of the DNN, and (iv) the resources to use. To this end, we leverage dynamic gated neural networks with branches, and propose a novel algorithmic strategy called Quantile-constrained Inference (QIC), based upon quantile-Constrained policy optimization. QIC makes joint, high-quality, swift decisions on all the above aspects of the system, with the aim to minimize inference energy cost. We remark that this is the first contribution connecting gated dynamic DNNs with infrastructure-level decision making. We evaluate QIC using a dynamic gated DNN with stems and branches for optimal sensor fusion and inference, trained on the RADIATE dataset offering Radar, LiDAR, and Camera data, and real-world wireless measurements. Our results confirm that QIC matches the optimum and outperforms its alternatives by over 80%.

Pruning neural networks, i.e., removing some of their parameters whilst retaining their accuracy, is one of the main ways to reduce the latency of a machine learning pipeline, especially in resource- and/or bandwidth-constrained scenarios. In this context, the pruning technique, i.e., how to choose the parameters to remove, is critical to the system performance. In this paper, we propose a novel pruning approach, called FlexRel and predicated upon combining training-time and inference-time information, namely, parameter magnitude and relevance, in order to improve the resulting accuracy whilst saving both computational resources and bandwidth. Our performance evaluation shows that FlexRel is able to achieve higher pruning factors, saving over 35% bandwidth for typical accuracy targets.

The increasing pervasiveness of intelligent mobile applications requires to exploit the full range of resources offered by the mobile-edge-cloud network for the execution of inference tasks. However, due to the heterogeneity of such multi-tiered networks, it is essential to make the applications' demand amenable to the available resources while minimizing energy consumption. Modern dynamic deep neural networks (DNN) achieve this goal by designing multi-branched architectures where early exits enable sample-based adaptation of the model depth. In this paper, we tackle the problem of allocating sections of DNNs with early exits to the nodes of the mobile-edge-cloud system. By envisioning a 3-stage graph-modeling approach, we represent the possible options for splitting the DNN and deploying the DNN blocks on the multi-tiered network, embedding both the system constraints and the application requirements in a convenient and efficient way. Our framework -- named Feasible Inference Graph (FIN) -- can identify the solution that minimizes the overall inference energy consumption while enabling distributed inference over the multi-tiered network with the target quality and latency. Our results, obtained for DNNs with different levels of complexity, show that FIN matches the optimum and yields over 65% energy savings relative to a state-of-the-art technique for cost minimization.

Intelligent reflecting surfaces (IRSs) have several prominent advantages, including improving the level of wireless communication security and privacy. In this work, we focus on the latter aspect and introduce a strategy to counteract the presence of passive eavesdroppers overhearing transmissions from a base station towards legitimate users that are facilitated by the presence of IRSs. Specifically, we envision a transmission scheme that cycles across a number of IRS-to-user assignments, and we select them in a near-optimal fashion, thus guaranteeing both a high data rate and a good secrecy rate. Unlike most of the existing works addressing passive eavesdropping, the strategy we envision has low complexity and is suitable for scenarios where nodes are equipped with a limited number of antennas. Through our performance evaluation, we highlight the trade-off between the legitimate users' data rate and secrecy rate, and how the system parameters affect such a trade-off., Comment: arXiv admin note: text overlap with arXiv:2108.00149

The existing work on the distributed training of machine learning (ML) models has consistently overlooked the distribution of the achieved learning quality, focusing instead on its average value. This leads to a poor dependability}of the resulting ML models, whose performance may be much worse than expected. We fill this gap by proposing DepL, a framework for dependable learning orchestration, able to make high-quality, efficient decisions on (i) the data to leverage for learning, (ii) the models to use and when to switch among them, and (iii) the clusters of nodes, and the resources thereof, to exploit. For concreteness, we consider as possible available models a full DNN and its compressed versions. Unlike previous studies, DepL guarantees that a target learning quality is reached with a target probability, while keeping the training cost at a minimum. We prove that DepL has constant competitive ratio and polynomial complexity, and show that it outperforms the state-of-the-art by over 27% and closely matches the optimum.

In next-generation networks, cells will be replaced by a collection of points-of-access (PoAs), with overlapping coverage areas and/or different technologies. Along with a promise for greater performance and flexibility, this creates further pressure on network management algorithms, which must make joint decisions on (i) PoA-to-user association and (ii) PoA management. We solve this challenging problem through an efficient and effective solution concept called Cluster-then-Match (CtM). Importantly, CtM makes human-centric decisions, where pure network performance is balanced against metrics like energy consumption and electromagnetic field exposure, which concern all humans in the network area -- including those who are not network users. Through our performance evaluation, which leverages detailed models for EMF exposure estimation and standard-specified signal propagation models, we show that CtM outperforms state-of-the-art network management schemes, including those utilizing machine learning, reducing energy consumption by over 80%.

The virtual reality (VR) and human-computer interaction (HCI) combination has radically changed the way users approach a virtual environment, increasing the feeling of VR immersion, and improving the user experience and usability. The evolution of these two technologies led to the focus on VR locomotion and interaction. Locomotion is generally controller-based, but today hand gesture recognition methods were also used for this purpose. However, hand gestures can be stressful for the user who has to keep the gesture activation for a long time to ensure locomotion, especially continuously. Likewise, in Head Mounted Display (HMD)-based virtual environment or Spherical-based system, the use of classic controllers for the 3D scene interaction could be unnatural for the user compared to using hand gestures such \eg pinching to grab 3D objects. To address these issues, we propose a user study comparing the use of the classic controllers (six-degree-of-freedom (6-DOF) or trackballs) in HMD and spherical-based systems, and the hand tracking and gestures in both VR immersive modes. In particular, we focused on the possible differences between spherical-based systems and HMD in terms of the level of immersion perceived by the user, the mode of user interaction (controller and hands), on the reaction of users concerning usefulness, easiness, and behavioral intention to use.

Convergence bounds are one of the main tools to obtain information on the performance of a distributed machine learning task, before running the task itself. In this work, we perform a set of experiments to assess to which extent, and in which way, such bounds can predict and improve the performance of real-world distributed (namely, federated) learning tasks. We find that, as can be expected given the way they are obtained, bounds are quite loose and their relative magnitude reflects the training rather than the testing loss. More unexpectedly, we find that some of the quantities appearing in the bounds turn out to be very useful to identify the clients that are most likely to contribute to the learning process, without requiring the disclosure of any information about the quality or size of their datasets. This suggests that further research is warranted on the ways -- often counter-intuitive -- in which convergence bounds can be exploited to improve the performance of real-world distributed learning tasks.

To make machine learning (ML) sustainable and apt to run on the diverse devices where relevant data is, it is essential to compress ML models as needed, while still meeting the required learning quality and time performance. However, how much and when an ML model should be compressed, and {\em where} its training should be executed, are hard decisions to make, as they depend on the model itself, the resources of the available nodes, and the data such nodes own. Existing studies focus on each of those aspects individually, however, they do not account for how such decisions can be made jointly and adapted to one another. In this work, we model the network system focusing on the training of DNNs, formalize the above multi-dimensional problem, and, given its NP-hardness, formulate an approximate dynamic programming problem that we solve through the PACT algorithmic framework. Importantly, PACT leverages a time-expanded graph representing the learning process, and a data-driven and theoretical approach for the prediction of the loss evolution to be expected as a consequence of training decisions. We prove that PACT's solutions can get as close to the optimum as desired, at the cost of an increased time complexity, and that, in any case, such complexity is polynomial. Numerical results also show that, even under the most disadvantageous settings, PACT outperforms state-of-the-art alternatives and closely matches the optimal energy cost.

Deploying services efficiently while satisfying their quality requirements is a major challenge in network slicing. Effective solutions place instances of the services' virtual network functions (VNFs) at different locations of the cellular infrastructure and manage such instances by scaling them as needed. In this work, we address the above problem and the very relevant aspect of sub-slice reuse among different services. Further, unlike prior art, we account for the services' finite lifetime and time-varying traffic load. We identify two major sources of inefficiency in service management: (i) the overspending of computing resources due to traffic of multiple services with different latency requirements being processed by the same virtual machine (VM), and (ii) the poor packing of traffic processing requests in the same VM, leading to opening more VMs than necessary. To cope with the above issues, we devise an algorithm, called REShare, that can dynamically adapt to the system's operational conditions and find an optimal trade-off between the aforementioned opposite requirements. We prove that REShare has low algorithmic complexity and is asymptotic 2-competitive under a non-decreasing load. Numerical results, leveraging real-world scenarios, show that our solution outperforms alternatives, swiftly adapting to time-varying conditions and reducing service cost by over 25%.

Open Radio Access Network (O-RAN) is an emerging paradigm, whereby virtualized network infrastructure elements from different vendors communicate via open, standardized interfaces. A key element therein is the RAN Intelligent Controller (RIC), an Artificial Intelligence (AI)-based controller. Traditionally, all data available in the network has been used to train a single AI model to be used at the RIC. This paper introduces, discusses, and evaluates the creation of multiple AI model instances at different RICs, leveraging information from some (or all) locations for their training. This brings about a flexible relationship between gNBs, the AI models used to control them, and the data such models are trained with. Experiments with real-world traces show how using multiple AI model instances that choose training data from specific locations improve the performance of traditional approaches following the hoarding strategy.

In cell-free wireless networks, multiple connectivity options and technologies are available to serve each user. Traditionally, such options are ranked and selected solely based on the network performance they yield; however, additional information such as electromagnetic field (EMF) exposure could be considered. In this work, we explore the trade-offs between network performance and EMF exposure in a typical indoor scenario, finding that it is possible to significantly reduce the latter with a minor impact on the former. We further find that surrogate models represent an efficient and effective tool to model the network behavior.

In the mobile-edge-cloud continuum, a plethora of heterogeneous data sources and computation-capable nodes are available. Such nodes can cooperate to perform a distributed learning task, aided by a learning controller (often located at the network edge). The controller is required to make decisions concerning (i) data selection, i.e., which data sources to use; (ii) model selection, i.e., which machine learning model to adopt, and (iii) matching between the layers of the model and the available physical nodes. All these decisions influence each other, to a significant extent and often in counter-intuitive ways. In this paper, we formulate a problem addressing all of the above aspects and present a solution concept called RightTrain, aiming at making the aforementioned decisions in a joint manner, minimizing energy consumption subject to learning quality and latency constraints. RightTrain leverages an expanded-graph representation of the system and a delay-aware Steiner tree to obtain a provably near-optimal solution while keeping the time complexity low. Specifically, it runs in polynomial time and its decisions exhibit a competitive ratio of $2(1+\epsilon)$, outperforming state-of-the-art solutions by over 50%. Our approach is also validated through a real-world implementation.

We experimentally investigate the sensing properties of an open cavity operating in the THz regime and realized by employing as mirrors two thin vanadium dioxide (VO2) films grown on silicon parallel plates and separated by a variable length. The phase transition of VO2 is used to control the behavior of the system between two different responses: a high transmission mode to the incident radiation (VO2 in the insulating state) and a high sensitivity to tiny changes in the cavity refractive index (VO2 in the conducting state). In the first state, the low loss regime enables to adjust the cavity length and easily optimize the resonances due to the Fabry-Perot (FP) effect in the Si plates and in the cavity volume. The activation of the metallic-like state instead, by damping the FP oscillations in the plates, promotes the onset of a comb-like spectrum that can be exploited as a versatile tool for accurate sensing applications. Using both an analytical model and full-wave simulations, we estimate the device response to variation in the refractive index of the cavity volume, showing that the proposed structure can achieve sensitivity values among the highest reported for THz sensors.

Traditionally, distributed machine learning takes the guise of (i) different nodes training the same model (as in federated learning), or (ii) one model being split among multiple nodes (as in distributed stochastic gradient descent). In this work, we highlight how fog- and IoT-based scenarios often require combining both approaches, and we present a framework for flexible parallel learning (FPL), achieving both data and model parallelism. Further, we investigate how different ways of distributing and parallelizing learning tasks across the participating nodes result in different computation, communication, and energy costs. Our experiments, carried out using state-of-the-art deep-network architectures and large-scale datasets, confirm that FPL allows for an excellent trade-off among computational (hence energy) cost, communication overhead, and learning performance.

Millimeter-wave communication is widely seen as a promising option to increase the capacity of vehicular networks, where it is expected that connected cars will soon need to transmit and receive large amounts of data. Due to harsh propagation conditions, mmWave systems resort to narrow beams to serve their users, and such beams need to be configured according to traffic demand and its spatial distribution, as well as interference. In this work, we address the beam management problem, considering an urban vehicular network composed of gNBs. We first build an accurate, yet tractable, system model and formulate an optimization problem aiming at maximizing the total network data rate while accounting for the stochastic nature of the network scenario. Then we develop a graph-based model capturing the main system characteristics and use it to develop a belief propagation algorithmic framework, called CRAB, that has low complexity and, hence, can effectively cope with large-scale scenarios. We assess the performance of our approach under real-world settings and show that, in comparison to state-of-the-art alternatives, CRAB provides on average a 50% improvement in the amount of data transferred by the single gNBs and up to 30% better user coverage.

Intelligent reflecting surfaces (IRSs) have several prominent advantages, including improving the level of wireless communications security and privacy. In this work, we focus on this aspect and envision a strategy to counteract the presence of passive eavesdroppers overhearing transmissions from a base station towards legitimate users. Unlike most of the existing works addressing passive eavesdropping, the strategy we consider has low complexity and is suitable for scenarios where nodes are equipped with a limited number of antennas. Through our performance evaluation, we highlight the trade-off between the legitimate users' data rate and secrecy rate, and how the system parameters affect such a trade-off.

Assisted driving for connected cars is one of the main applications that 5G-and-beyond networks shall support. In this work, we propose an assisted driving system leveraging the synergy between connected vehicles and the edge of the network infrastructure, in order to envision global traffic policies that can effectively drive local decisions. Local decisions concern individual vehicles, e.g., which vehicle should perform a lane-change manoeuvre and when; global decisions, instead, involve whole traffic flows. Such decisions are made at different time scales by different entities, which are integrated within an edge-based architecture and can share information. In particular, we leverage a queuing-based model and formulate an optimization problem to make global decisions on traffic flows. To cope with the problem complexity, we then develop an iterative, linear-time complexity algorithm called Bottleneck Hunting (BH). We show the performance of our solution using a realistic simulation framework, integrating a Python engine with ns-3 and SUMO, and considering two relevant services, namely, lane change assistance and navigation, in a real-world scenario. Results demonstrate that our solution leads to a reduction of the vehicles' travel times by 66 in the case of lane change assistance and by 20 for navigation, compared to traditional, local-coordination approaches., Comment: arXiv admin note: text overlap with arXiv:2008.09336

Many machine learning (ML) techniques suffer from the drawback that their output (e.g., a classification decision) is not clearly and intuitively connected to their input (e.g., an image). To cope with this issue, several explainable ML techniques have been proposed to, e.g., identify which pixels of an input image had the strongest influence on its classification. However, in distributed scenarios, it is often more important to connect decisions with the information used for the model training and the nodes supplying such information. To this end, in this paper we focus on federated learning and present a new methodology, named node liability in federated learning (NL-FL), which permits to identify the source of the training information that most contributed to a given decision. After discussing NL-FL's cost in terms of extra computation, storage, and network latency, we demonstrate its usefulness in an edge-based scenario. We find that NL-FL is able to swiftly identify misbehaving nodes and to exclude them from the training process, thereby improving learning accuracy.

We consider the optimization of a smart radio environment where meta-surfaces are employed to improve the performance of multiuser wireless networks working at sub-THz frequencies. Motivated by the extreme sparsity of the THz channel we propose to model each meta-surface as an electronically steerable reflector, by using only two parameters, regardless of its size. This assumption, although suboptimal in a general multiuser setup, allows for a significant complexity reduction when optimizing the environment and, despite its simplicity, is able to provide high communication rates. We derive a set of asymptotic results providing insight on the system behavior when both the number of antennas at the transmitter and the meta-surfaces area grow large. For the optimization we propose an algorithm based on the Newton-Raphson method and a simpler, yet effective, heuristic approach based on a map associating meta-surfaces and users. Through numerical results we provide insights on the system behavior and we assess the performance limits of the network in terms of supported users and spatial density of the meta-surfaces.

As we move from 5G to 6G, edge computing is one of the concepts that needs revisiting. Its core idea is still intriguing: instead of sending all data and tasks from an end user's device to the cloud, possibly covering thousands of kilometers and introducing delays that are just owed to limited propagation speed, edge servers deployed in close proximity to the user, e.g., at some 5G gNB, serve as proxy for the cloud. Yet this promising idea is hampered by the limited availability of such edge servers. In this paper, we discuss a way forward, namely the virtual edge computing (V-Edge) concept. V-Edge bridges the gap between cloud, edge, and fog by virtualizing all available resources including the end users' devices and making these resources widely available using well-defined interfaces. V-Edge also acts as an enabler for novel microservices as well as cooperative computing solutions. We introduce the general V-Edge architecture and we characterize some of the key research challenges to overcome, in order to enable wide-spread and even more powerful edge services.

5G networks are going to support a variety of vertical services, with a diverse set of key performance indicators (KPIs), by using enabling technologies such as software-defined networking and network function virtualization. It is the responsibility of the network operator to efficiently allocate the available resources to the service requests in such a way to honor KPI requirements, while accounting for the limited quantity of available resources and their cost. A critical challenge is that requests may be highly varying over time, requiring a solution that accounts for their dynamic generation and termination. With this motivation, we seek to make joint decisions for request admission, resource activation, VNF placement, resource allocation, and traffic routing. We do so by considering real-world aspects such as the setup times of virtual machines, with the goal of maximizing the mobile network operator profit. To this end, first, we formulate a one-shot optimization problem which can attain the optimum solution for small size problems given the complete knowledge of arrival and departure times of requests over the entire system lifespan. We then propose an efficient and practical heuristic solution that only requires this knowledge for the next time period and works for realistically-sized scenarios. Finally, we evaluate the performance of these solutions using real-world services and large-scale network topologies. {Results demonstrate that our heuristic solution performs better than a state-of-the-art online approach and close to the optimum.

The traditional approach to distributed machine learning is to adapt learning algorithms to the network, e.g., reducing updates to curb overhead. Networks based on intelligent edge, instead, make it possible to follow the opposite approach, i.e., to define the logical network topology em around the learning task to perform, so as to meet the desired learning performance. In this paper, we propose a system model that captures such aspects in the context of supervised machine learning, accounting for both learning nodes (that perform computations) and information nodes (that provide data). We then formulate the problem of selecting (i) which learning and information nodes should cooperate to complete the learning task, and (ii) the number of iterations to perform, in order to minimize the learning cost while meeting the target prediction error and execution time. After proving important properties of the above problem, we devise an algorithm, named DoubleClimb, that can find a 1+1/|I|-competitive solution (with I being the set of information nodes), with cubic worst-case complexity. Our performance evaluation, leveraging a real-world network topology and considering both classification and regression tasks, also shows that DoubleClimb closely matches the optimum, outperforming state-of-the-art alternatives.

Under the federated learning paradigm, a set of nodes can cooperatively train a machine learning model with the help of a centralized server. Such a server is also tasked with assigning a weight to the information received from each node, and often also to drop too-slow nodes from the learning process. Both decisions have major impact on the resulting learning performance, and can interfere with each other in counterintuitive ways. In this paper, we focus on edge networking scenarios and investigate existing and novel approaches to such model-weighting and node-dropping decisions. Leveraging a set of real-world experiments, we find that popular, straightforward decision-making approaches may yield poor performance, and that considering the quality of data in addition to its quantity can substantially improve learning.

Single-task UAVs are increasingly being employed to carry out surveillance, parcel delivery, communication support, and other specific tasks. When the geographical area of operation of single-task missions is common, e.g., in post-disaster recovery scenarios, it is more efficient to have multiple tasks carried out as part of a single UAV mission. In these scenarios, the UAVs' equipment and mission plan must be carefully selected to minimize the carried load and overall resource consumption. In this paper, we investigate the joint planning of multitask missions leveraging a fleet of UAVs equipped with a standard set of accessories enabling heterogeneous tasks. To this end, an optimization problem is formulated yielding the optimal joint planning and deriving the resulting quality of the delivered tasks. In addition, two heuristic solutions are developed for large-scale environments to cope with the increased complexity of the optimization framework. The joint planning is applied to a specific scenario of a flood in the San Francisco area. Results show the effectiveness of the proposed heuristic solutions, which provide good performance and allow for drastic savings in the computational time required to plan the UAVs' trajectories with respect to the optimal approach, thus enabling prompt reaction to the emergency events., Comment: arXiv admin note: text overlap with arXiv:1905.05453

Automotive services for connected vehicles are one of the main fields of application for new-generation mobile networks as well as for the edge computing paradigm. In this paper, we investigate a system architecture that integrates the distributed vehicular network with the network edge, with the aim to optimize the vehicle travel times. We then present a queue-based system model that permits the optimization of the vehicle flows, and we show its applicability to two relevant services, namely, lane change/merge (representative of cooperative assisted driving) and navigation. Furthermore, we introduce an efficient algorithm called Bottleneck Hunting (BH), able to formulate high-quality flow policies in linear time. We assess the performance of the proposed system architecture and of BH through a comprehensive and realistic simulation framework, combining ns-3 and SUMO. The results, derived under real-world scenarios, show that our solution provides much shorter travel times than when decisions are made by individual vehicles., Comment: GLOBECOM 2020

Mmwave bands are being widely touted as a very promising option for future 5G networks, especially in enabling such networks to meet highly demanding rate requirements. Accordingly, the usage of these bands is also receiving an increasing interest in the context of 5G vehicular networks, where it is expected that connected cars will soon need to transmit and receive large amounts of data. Mmwave communications, however, require the link to be established using narrow directed beams, to overcome harsh propagation conditions. The advanced antenna systems enabling this also allow for a complex beam design at the base station, where multiple beams of different widths can be set up. In this work, we focus on beam management in an urban vehicular network, using a graph-based approach to model the system characteristics and the existing constraints. In particular, unlike previous work, we formulate the beam design problem as a maximum-weight matching problem on a bipartite graph with conflicts, and then we solve it using an efficient heuristic algorithm. Our results show that our approach easily outperforms advanced methods based on clustering algorithms.

Millimeter-wave (mmwave) communication represents a potential solution to capacity shortage in vehicular networks. However, effective beam alignment between senders and receivers requires accurate knowledge of the vehicles' position for fast beam steering, which is often impractical to obtain in real time. We address this problem by leveraging the traffic signals regulating vehicular mobility: as an example, we may coordinate beams with red traffic lights, as they correspond to higher vehicle densities and lower speeds. To evaluate our intuition, we propose a tractable, yet accurate, mmwave communication model accounting for both the distance and the heading of vehicles being served. Using such a model, we optimize the beam design and define a low-complexity, heuristic strategy. For increased realism, we consider as reference scenario a large-scale, real-world mobility trace of vehicles in Luxembourg. The results show that our approach closely matches the optimum and always outperforms static beam design based on road topology alone. Remarkably, it also yields better performance than solutions based on real-time mobility information.

Machine learning has emerged as a promising paradigm for enabling connected, automated vehicles to autonomously cruise the streets and react to unexpected situations. A key challenge, however, is to collect and select real-time and reliable information for the correct classification of unexpected, and often rare, situations that may happen on the road. Indeed, the data generated by vehicles, or received from neighboring vehicles, may be affected by errors or have different levels of resolution and freshness. To tackle this challenge, we propose an active learning framework that, leveraging the information collected through onboard sensors as well as received from other vehicles, effectively deals with scarce and noisy data. In particular, given the available information, our solution selects the data to add to the training set by trading off between two essential features, namely, quality and diversity. The results, obtained using real-world data sets, show that the proposed method significantly outperforms state-of-the-art solutions, providing high classification accuracy at the cost of a limited bandwidth requirement for the data exchange between vehicles.

One of the main use cases for advanced cellular networks is represented by massive Internet-of-things (MIoT), i.e., an enormous number of IoT devices that transmit data toward the cellular network infrastructure. To make cellular MIoT a reality, data transfer and control procedures specifically designed for the support of IoT are needed. For this reason, 3GPP has introduced the Control Plane Cellular IoT optimization, which foresees a simplified bearer instantiation, with the Mobility Management Entity (MME) handling both control and data traffic. The performance of the MME has therefore become critical, and properly scaling its computational capability can determine the ability of the whole network to tackle MIoT effectively. In particular, considering virtualized networks and the need for an efficient allocation of computing resources, it is paramount to characterize the MME performance as the MIoT traffic load changes. We address this need by presenting compact, closed-form expressions linking the number of IoT sources with the rate at which bearers are requested, and such a rate with the delay incurred by the IoT data. We show that our analysis, supported by testbed experiments and verified through large-scale simulations, represents a valuable tool to make effective scaling decisions in virtualized cellular core networks.