Addressing the Scalability of Network Digital Twins: A Network Sampling Approach

Preprint version. The final version of the paper is available at: https://ieeexplore.ieee.org/document/10814373 (M. Kellil, S. Ben Hadj Said, M. -T. Thi, C. Janneteau and A. Olivereau, "Addressing the Scalability of Network Digital Twins: A Network Sampling Approach," IEEE CNSM, Prague, Czech Republic, 2024, pp. 1-7, doi: 10.23919/CNSM62983.2024.10814373). Licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). This license applies only to the text and figures of this publication and does not grant any rights under any patent or patent application held by the authors. Addressing the Scalability of Network Digital Twins: A Network Sampling Approach Mounir Kellil Institute LIST, CEA Paris-Saclay University Palaiseau, France [email protected] Christophe Janneteau Institute LIST, CEA Paris-Saclay University Palaiseau, France [email protected] Siwar Ben Hadj Said Institute LIST, CEA Paris-Saclay University Palaiseau, France [email protected] Alexis Olivereau Institute LIST, CEA Paris-Saclay University Palaiseau, France [email protected] Minh-Thuyen Thi Institute LIST, CEA Paris-Saclay University Palaiseau, France [email protected] Abstract— With the increasing complexity of mobile communication networks like 5G/6G networks, characterized by the diversity of network devices, technologies, and applications, advanced network management solutions are needed to ensure real-time network optimization with risk-free decision-making operations (e.g., safe network reconfiguration). To achieve this objective, a Network Digital Twin (NDT) paradigm represents an attractive perspective, enabling the manipulation of the virtual counter part of a real communication network. Nonetheless, generating a Digital Twin of a complex network, comprising thousands of heterogeneous devices and highly dynamic network characteristics (resource usage, network topology, link quality, etc.) poses a serious scalability problem. This paper aims at addressing the scalability problem for the generation of a Digital Twin of a complex network like a 5G/6G network. In particular, the paper proposes a sampling approach in conjunction with a structured network information representation, as well as zoom-in/out operations to enable a modular generation of the NDT. Keywords—Network sampling, network management, digital twin I. INTRODUCTION Mobile communication networks are experiencing a paradigm shift thanks to the recent advances in 5G and future 6G networks, characterized by an exponential growth of communication network capacities and features, including network heterogeneity, large bandwidth, ultra-low latency, and decentralized decision-making operations. In the same vein, new generations of network applications have emerged, including , immersive applications, V2X, smart factory, etc.. The deployment of such applications is accelerated by the increasing integration and adoption of IoT devices and AI services in today’s mobile communication networks. Nonetheless, this mobile communication network ecosystem typically operates in a highly dynamic environment characterized by frequent network topology changes as well as to the unpredictable network resource consumption patterns. To ensure supervision, or more generally, management, whether fine-grained or synthetic, of such communication networks, various network management services like SNMP, CMIP, LwM2M, CoMI, and SDN can be used, alone or combined [1]. Furthermore, such services can adopt recent advances in automated network management, including the usage of machine learning techniques in order to accommodate the ever-increasing complexly of mobile communication networks [1][2]. Nonetheless, network management services do not intrinsically guarantee risk-free (or error-free) (re)configuration of complex mobile communication networks like 5G/6G networks. This is particularly true when it comes to performing real-time optimization of the communication network, network tests in operational mode (what-if analysis) [3][4][5], or proceeding with real-time network upgrade or extension. In such scenarios, any improper network (re)configuration operation (ex., misconfiguration of routes, underestimation of allocated resources, etc.) would result in network service interruption and require extra cost to set up failure recovery strategies and operations. To prevent this problem, a number of recent research initiatives explored the adoption of the Digital Twin concept to the communication networks. This concept is commonly referred to as Network Digital Twin (NDT) [4]-[7] . The NDT is a digital twin of the real (or physical) network, where a digital copy of the real network can be manipulated without risk, which makes it possible in particular to visualize or predict the state of the real network if this or that network configuration is to be applied. Despite the attractive perspective of deploying an NDT solution to facilitate the supervision of complex communication networks and streamline their effective management, ensuring this objective in a scalable fashion is not straightforward. Indeed, frequent NDT synchronizations with the real communication network poses a scalability problem when dealing with complex networks (e.g. too large number of network entities, highly dynamic topologies, large volume of information per node or per network link), especially when each network information is to be reported to at the NDT’s side. This paper aims at addressing the scalability problem when designing a digital twin for complex communication networks like 5G and future 6G networks. In particular, the paper proposes a network sampling concept in conjunction with a structured network information representation, as well as zoom-in/out operations, in order to ensure efficient adoption of NDTs in emerging mobile communication networks. The rest of the paper is organized as follows. Section II discusses various solutions for supporting NDT in emerging mobile communication networks. Section III describes both the network information structure, the network sampling concept, and the zoom-in/out modes. Section IV exploits the concepts described in section III to present an incremental sampling scheme, used to generate the full NDT of the real communication network. Section V presents preliminary evaluation results of the information overhead associated to the proposed sampling solution. Finally, section VI concludes the paper and discusses future research directions. II. STATE-OF-THE ART Mobile communication networks typically experience frequent topology changes, caused by various factors, including the presence of error-prone wireless channels, user behavior (mobility, network connection habits), routing protocols, resource allocation services, etc. Furthermore, such communication networks may be subject to various coexisting traffic patterns, ranging from (predictable) periodic reporting and control events (e.g. protocol-specific, telemetry, command/control applications) to streaming applications (e.g., audio/video, immersive environments, etc.), and encompassing event-driven data traffic (e.g., ambient applications, alert applications, etc.). Virtualizing such a complex communication network ecosystem to expose its real-time digital twin counterpart is inherently challenging, especially that the network states (e.g., topology, resource usage, etc.) are perpetually changing, hence the need for a scalable approach to solve the network digital twin problem. Various research initiatives have explored the adoption of digital twins for mobile communication networks like 5G and 6G (e.g., [5][6]-[8]). However, to our knowledge, few papers have considered the scalability problem, even though the question of modular implementation has been pointed out by a number of authors (e.g., [8][12]). Notably, the authors in [10], were the first to consider the scalability problem for implementing digital twins. Their work, though, focused on the smart manufacturing domain (digital twins of individual machines, production cells, and entire manufacturing facilities), where the communication component is seen as a means for generating a digital twin, and not as a target component for digital twin generation. The authors addressed the scalability problem for data acquisition and proposed a stepwise method that breaks down the implementation into distinct, manageable steps, allowing teams to follow a defined pathway from recognizing devices to establishing communication and integrating data. In [11], the authors present a digital twin system for mobile networks. The proposed system considers three key elements from the mobile network: mobile users, base stations, and wireless environments. For each element, a virtual version is generated and configured with real-world data by modeling its principles and parameters. These virtual elements are assembled to form the whole digital twin system. The paper addresses the scalability problem using four concepts: 1) generative models like GANs (Generative Adversarial Networks) and VAEs (Variational Autoencoders) to simulate the behavior of millions of mobile users and network components, 2) Parallel Processing techniques to handle simulations of a large number of mobile users simultaneously, 3) Modular digital twin creation for the virtualized elements, and 4) real-time adjustment of the digital twin based on simulation. A methodology for creating Mobile Network Digital Twins (MNDT) for 5G networks was proposed in [12] by modeling physical network elements and their interactions. Network elements include physical devices, communication links, operating environment (e.g., traffic patterns and user behavior), and various 5G-specific network functions. To address the scalability problem, the methodology emphasizes automated data acquisition and modeling, using agents that collect network information (topology and network parameters) and operational data (current status and performance metrics of devices) from the physical network. In sum, the review solutions addressed the NDT scalability problem using network modeling along with network agents. These agents are used for automated data acquisition and selfoptimization operations. Building on such concepts, our solution takes a step further by focusing on the information overhead problem associated to maintaining up-to-date information at the NDT’ side. This problem has not been sufficiently addressed by current approaches. In this paper, we propose a novel scheme for improving the data acquisition process and, incidentally, streamlining network modeling operations. To this aim, we propose a network sampling scheme. This scheme enables both selective and incremental data acquisition from the real communication network. Upon each network sampling phase, a zoom-in/out operation over the real communication network is performed, depending on whether more (zoom-in) or less (zoom-out) details need to be exposed for a specific segment (sample) of the real communication network. This zoom-in/out scheme is facilitated thanks to network information categorization (device and link) and structuring (object→resource(s)). Each object being a device or a link. The main advantage of the proposed solution is that it proceeds with a modular NDT generation, where each sampling phase targets a part of the communication network. In this paper, modularly is expressed both horizontally (size of the network sample, i.e., the number of objects composing the sample (i.e., devices and links)) and vertically (number of captured resources for each object of the sample). This solution is particularly useful when there is a need for saving network resources (bandwidth, memory, CPU, energy), while managing effectively the communication network (i.e. collect only the relevant network parameters). Additionally, this modular NDT generation also enables an incremental generation of the complete NDT, if necessary. III. SOLUTION DESCRIPTION For the sake of simplicity and clarity, figure 1 presents a simplified view of the NDT system in the context of a mobile communication network. This figure is consistent with the NDT reference architecture specified by the IETF and ITU-T, and which we discussed in a previous paper [17]. Within the NDT system, there is a key component called NDT manager, which is responsible for generating and updating the NDT associated to the real communication network or a part of it, called a sample. As explained in the next section, the NDT manager views each network sample as a set of objects. Figure 1 also shows that the NDT manger interacts with three functional components: the NDT prediction module (e.g., a machine learning agent [4]-[9]), the NDT GUI (to display the NDT [9]), and the network management module (a conventional network management system like SNMP, CMIP, LwM2M, SDN, etc.) [1][2]. Fig. 1. High-level view of an NDT system for a mobile communication network A. Network information representation To facilitate the identification of the network entity at the NDT level, the network information identification will follow a hierarchical structure: Sample→Object→Resource, adhering to an Object-Resource model similar to the one used in existing network data models, such as the LwM2M data model [13]. This structured representation of network information also aims to facilitate fine-grained supervision and configuration of the communication network at its NDT counterpart. In addition, a network object can be of two possible types: network device or network link. As shown in Figure 2, a network sample can be composed of one or more network objects (devices and/or network links), each with zero, one or more associated resources. A network resource represents any type of information that can characterize a network object. For a network device like a router, the associated resource could be a network interface’s IP address, CPU usage ratio, etc. For a network link, the resource could be the PHY communication technology, SNR, ratio of available bandwidth, types of transport protocols, etc. Fig. 2. Hierachical repsentation of network objects and associated resources Furthermore, we consider that, from the standpoint of a communication network, any external physical object (i.e., aside from a network object) that is effectively influencing, in one way or another, the behavior of the communication network (e.g., network user, building, vehicle, etc.) must have a signature on one or more network objects (device(s) and/or link(s)), where the signature could be expressed in the form of a resource value of the concerne network object(s). In light of this, any external object that does not impact the communication network will not have a signature in the set of objects captured on the NDT side. As a result, this external object will simply be ignored. B. Network sampling The choice of the network sample can be driven by different selection strategies. For instance, the selection may be random, or it may also be based on recent events that occurred in the network (e.g., hardware/software failures, abnormally low/high values of network parameters, etc.). At each sampling phase, a zoom-in or zoom-out operation is performed in parallel. The objective behind a zoom-in mode is to focus on a finegrained representation of the sample that is being collected by the NDT manager (e.g., collect all the available resources of each of the objects of the sample). This way, a set of nodes and links can be dynamically selected based on certain criteria (e.g., high CPU usage, critical links, anomaly analysis, etc.) to monitor their performance metrics more closely. On the other hand, a zoom-out mode is useful when a high-level view of a specific sample or the entire network is needed for several reasons. For instance, it helps to free up resources at the NDT manager level, capture the list of current network nodes and links, and obtain a snapshot of high-level trends or patterns at the sample or network scale (e.g., total CPU usage, average traffic throughput, etc.). In the network sampling phase, the NDT manager generates a request (request_desc) to fill in a local structure called sample descriptor (desc_S). An example of a sample descriptor is provided in figure 3. Fig. 3. Example of a sample descritor On the other hand, the request request_desc has a structured expression, enabling conditional collection of objects and associated resources for a given network sample. This can be achieved using a number of possible network management protocols like NETCONF [14], RESTCONF [15] with their YANG-based data model [16]. In addition, the request includes a flag, which indicates whether the network sample is to be generated following a zoom-in or a zoom-out mode. A simplified description of the request is provided hereafter, for illustrative purposes. The exact form of such a request is outside the scope of this paper. Sample i Network Object 1.1 Sample 1 ... Sample N ... ... Object 1.n ... Object N.1 Object N.m ... ... ... ... ... Resources of object 1.1 Resources of object 1.n Resources of object N.1 Resources of object N.m Object Attribut Description + Obj_ID Object identifier (e.g., IP address of the network device ) + Obj_Type Object type (device or link) + Res_nb Total number of resources associated to Obj_!D + Rel_obj Associated object (ID of the object (device or link) associated to Obj_ID). *Resource Attribut Description + Res_ID Identifier of resource + Res_type Type of resource (ex. CPU) + Res_value Value of resource (ex. CPU usage ratio) 1 Fig. 4. Example of Expression field (zoom-in (left) and zoom-out (right)) IV. APPLICATION OF NETWORK SAMPLING FOR THE GENERATION OF THE NDT OF A FULL NETWORK This section explains the incremental network sampling procedure that enables the NDT manager to incrementally generate the NDT (i.e., the real-time picture of a full communication network) using successive selective sampling operations in a zoom-in mode (request_desc with zoom_flag= zoom-in). A. Initial step – NDT initialization To perform the incremental sampling, the NDT manager is initialized with the list of all the objects of the full network, i.e., all the network devices and all the associated network links. Although not critical, this initialization phase enable the NDT manager to get a high-level view of the actual status (a first glance) of the real communication network. The initial step consists in transmitting a request_desc query from the NDT manager to the network manager. This query includes an Expression field asking for all the objects of the communication network (active devices and active links) and having a zoom_flag value set to zoom-out. An example of the Expression filed is provided hereafter. Fig. 5. Initial phase of Incremental Sampling - Example of Expression field The incremental NDT generation will take place through a set of successive network sampling operations. Each newly generated sample Si is chosen so that its sample descriptor desc_Si is different from the descriptor of any previous sample of the current incremental NDT generation procedure. This enables to ensure the convergence of the generation procedure towards the NDT of the complete communication network (cf. figure 6). Fig. 6. Incremental NDT generation – Illustrative figure In order to keep up to date all the already virtualized samples on the NDT’s side, the virtualized counterpart of each sample is regularly updated (e.g., on demand, upon the virtualization of a new network sample, or periodically) using predictive models, based, for instance, on machine learning algorithms. The detailed description of this prediction model will be addressed in a future work.The following figure shows a 6-step procedure for NDT generation of a full phyisical netwotk comprising 4 samples. Fig. 7. Abstract view of incremental NDT generation using a combined network sampling and sample prediction scheme. V. PERFORMANCE EVALUATION Recall that the primary objective of our solution is to help mitigate the overhead associated with keeping an up-to-date NDT representation despite the dynamism of the physical communication network. In this section, we present preliminary evaluation results of the information overhead when suing network sampling operations to generate the sample’s NDT. A. Trafic Model The simulation phase considers a traffic pattern following a Poisson distribution. For each time step, the number of traffic events, both on the node and on the link, is generated using the Poisson distribution. This way, each traffic event contributes to the CPU and memory usage of the nodes and the bandwidth usage of the links. The objective is to express a traffic model that determines the amount of traffic data processed by each node and the amount of data present in the communication link (network bandwidth usage). Higher traffic amount (higher λ values in the Poisson traffic model) can lead to more frequent updates in the network state (e.g., more frequent data processing and transmissions). A given value of λ means that, on average, there is λ traffic events per second, per node or link. Our simulation also considers three values of λ are considered in the simulation phase 0.1, 1 and 10. B. Network Model The performance evaluation considers a wireless communication network, comprising 100 nodes and a probabilistic creation of the link between nodes using a normal distribution with a link type-dependent mean and a standard deviation s of 0.05. This allows the link existence probability to vary probabilistically based on a normal distribution. The choice of parameters was made to ensure a reasonable and realistic representation of the network topology. The mean probability of creating a link between any two core nodes is 0.1. This indicates that, on average, each possible link between nodes has a 10% chance of existing. This choice balances the network between being overly sparse and overly dense. A standard deviation of 0.05 introduces variability in the link existence probability while keeping the values within a reasonable range. This variation helps in simulating real-world scenarios where link availability can fluctuate due to various factors such as network congestion, physical distance, or random failures. C. Information overhead in random network sampling In this section, the evaluation of the information overhead for network sampling will consider a sampling function that randomly selects 20 nodes from the network. In addition, let 𝑆𝑁 and 𝑆𝐿 be the node-specific information size and link-specific information size, respectively. For a 16-Byte device ID size, a 32-Byte link ID size, and a 10-byte resource field size, we get: 𝑆𝑁 = 46 bytes and 𝑆𝐿 = 62 bytes. Also, on a zoom-in operation, it is assumed that one object ID and three associated resources are captured per object of interest (device or link). Also, let: 𝑁𝑆 be the number of sampled nodes and 𝐿𝑆 be the number of links between the sampled nodes. As a result, the information overhead 𝐼𝑡 for the selective sampling is provided hereafter: 𝐼𝑡=(𝑁𝑆∙ 𝑆𝑁+ 𝐿𝑆∙ 𝑆𝐿) D. Information overhead in full network represetation Let N be the number of nodes in the network and L be the number of links in the network Then, the information overhead 𝐼𝑡 for the full network representation can be expressed as: 𝐼𝑡=(𝑁 ∙ 𝑆𝑁+ 𝐿 ∙ 𝑆𝐿) E. Simulation results First, it is worth noting that in our simulation we consider traffic event-driven data acquisition to evaluate the “worst case” for network information overhead evaluation (i.e., aggressive mode). Figures 8 and 9 show the network information overhead related to the NDT’s information acquisition, respectively in case of a naïve (full) information collection approach and a selective sampling approach. The results show that more frequent the traffic events, the higher the network information overhead. This can be justified by the fact that in our simulation, we correlate data acquisition to traffic events (λ) both in random sampling and full network representation cases. In addition, these preliminary results show the information overhead for data acquisition in case of full network representation is more than 16 times higher than that of selective sampling. Fig. 8. Information overhead – case of full NDT generation Fig. 9. Information overhead – case of random sampling (20 nodes per sample) VI. CONCLUSION AND FUTURE DIRECTIONS This paper addresses the scalability problem in implementing a network digital twin (NDT) solution for complex communication networks, which are characterized by highly dynamic network information, including various frequently changing parameters from both network devices and links. Because this dynamic network information should be continuously captured at the NDT counterpart, in order to maintain an up-to-date picture of the real communication network, a novel paradigm for network information representation and acquisition should be designed. To this aim, we proposed a network sampling scheme in conjunction with a hierarchical representation of all network information to enable a flexible, adaptive, and modular generation of the NDT. Our solution allows for a sliced representation of the communication network (horizontal slice and vertical slice). This can help audit the network by focusing the analysis on large-scale but high-level view communication network segment (horizontal slice) or a small scale but a detailed view of the different communication layers of the network segment. In the next step, we will evaluate the performance of our solution, considering more realistic scenarios, including user mobility modeling along with the integration of various traffic offloading patterns. In addition, the following issues will be addressed. Adaptive sampling: the proposed sampling paradigm can be extended to enable adaptive sampling. For instance, the highly dynamic network information will be more frequently reported that the relatively stable network information. ML-based sample selection and prediction: the proposed sampling paradigm can be expanded to enable both sample selection based on relevant network events like node or link failures, traffic congestion, etc. To this aim, ML techniques for event prediction and sampling selection can be explored, with a particular focus on those that can handle temporal dependencies, adapt to real-time data, and learn from evolving patterns in the network (e.g., RNN, GNN, and Autoencoders). Inter-sample correlation: considering inter-sample correlation (e.g., the correlation between two adjacent samples) is particularly important for predicting network samples on the NDT’s side. Federated Learning (FL) techniques can help aggregate learning from different samples of the real network, leading to a more accurate representation of each sample's NDT and, by extension, a more precise overall NDT representation. However, when dealing with inter-sample correlation in a distributed setting like FL, ensuring data synchronization across samples is critical to maintaining the effectiveness of the learned models. Accuracy-overhead tradeoff: one main challenge in implementing a scalable approach for NDT support in complex communication networks is to find the best tradeoff between the accuracy of the NDT representation and the intrinsic overhead of the NDT solution (i.e., bandwidth, memory, and CPU usage). One possible approach to address this problem is to split it into two sub-problems: instantaneous accuracy and average accuracy. Instantaneous accuracy could be more relevant in situations where immediate decisions are made based on an NDT snapshot (e.g., evidence of an imminent event in the real network), while average accuracy is key for long-term analysis or longterm forecast of the real communication network. REFERENCES [1] M. Aboubakar, M. Kellil, P. Roux, “A review of IoT network management: Current status and perspectives”, Elsevier JKSU Journal, Volume 34, Issue 7, July 2022, Pages 4163-4176. [2] E. Coronado et al., "Zero Touch Management: A Survey of Network Automation Solutions for 5G and 6G Networks," in IEEE Communications Surveys & Tutorials, vol. 24, no. 4, pp. 2535-2578. [3] E. Ak, B. Canberk, V. Sharma, O. A. Dobre and T. Q. Duong, "Whatif Analysis Framework for Digital Twins in 6G Wireless Network Management," 2024 International Wireless Communications and Mobile Computing (IWCMC), Ayia Napa, Cyprus, 2024, pp. 232-237. [4] C. Zhou et al., “Digital Twin Network: Concepts and Reference Architecture”, Internet Draft, July 2024, Work In Progress. [5] P. Almasan et al., “Network Digital Twin: Context, Enabling Technologies, and Opportunities,” IEEE Commun. Mag. 60(11): 22-27 (2022). [6] L. U. Khan et al., “Digital-Twin-Enabled 6G: Vision, Architectural Trends, and Future Directions,” IEEE Communications Magazine, pp. 74-80, 2022. [7] X. Lin et al., “6G Digital Twin Networks: From Theory to Practice”, CoRR abs/2212.02032 (2022). [8] M. Yaqoob et al., “Digital-Twin-Driven End-to-End Network Slicing Toward 6G," in IEEE Internet Computing, vol. 28, no. 2, pp. 47-55, March-April 2024, doi: 10.1109/MIC.2023.3332252. [9] H. Ahmadi et al., "Networked Twins and Twins of Networks: An Overview on the Relationship Between Digital Twins and 6G," in IEEE Communications Standards Magazine, vol. 5, no. 4, pp. 154-160, December 2021, doi: 10.1109/MCOMSTD.0001.2000041. [10] S. Dittmann et al., "Towards a scalable implementation of digital twins - A generic method to acquire shopfloor data”, Procedia CIRP, vol. 96, 2021, pp. 157-162. [11] J. Gong et al., “Demo: Scalable Digital Twin System for Mobile Networks with Generative AI”, Proceedings of the 21st Annual International Conference on Mobile Systems, Applications and Services, 2023, pp. 610–611. [12] M. Sanz Rodrigo et al., “Digital Twins for 5G Networks: A Modeling and Deployment Methodology" in IEEE Access, 2023, vol. 11, pp. 38112-38126, 2023. [13] “Lightweight Machine to Machine Technical Specification: Core”, Version: 1.1.1 - 2019 06 17, Open Mobile Alliance, OMA-TSLightweightM2M_Core-V1_1_1-20190617-A. [14] R. Enns et al., “Network Configuration Protocol (NETCONF)”, RFC 6242, June 2021. [15] A. Bierman et al., “ RESTCONF Protocol”, RFC 8040, January 2017. [16] M. Bjorklund, Ed., “The YANG 1.1 Data Modeling Language”, RFC 7950, August 2016. [17] S. B. Hadj Said, M. -T. Thi, M. Kellil and A. Olivereau, "On the Management of TSN Networks in 6G: A Network Digital Twin Approach," 2023 IEEE 28th International Conference on Emerging Technologies and Factory Automation (ETFA), Sinaia, Romania, 2023, pp. 1-7. [18] K. McCloghrie et al., “Structure of Management Information Version 2 (SMIv2)”, RFC 2578, April 1999. [19] M. Bjorklund, “The YANG 1.1 Data Modeling Language”, RFC 7950, August 2016.