SYSTEM AND METHOD FOR PREDICTING COMPUTING RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS

Description

16 SYSTEM AND METHOD FOR PREDICTING COMPUTING RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS FIELD OF THE INVENTION

[001] The invention shown in this patent application relates to the field of telecommunications, specifically showing a system and method applicable to virtualized networks, especially in environments with high demand for computational resources, such as telecommunications networks and data systems, aiming at resource optimization based on performance indicators. This prediction is based on Key Performance Indicators (KPIs) of the RAN and on monitoring CPU and RAM memory. The method obtains both these KPIs of network usage and network server resource usage, then constructing regression models regarding resource usage and KPIs, in order to obtain a relationship between these KPIs and server resource usage, thus predicting CPU and / or RAM usage in virtualized server components. This present invention relates to the field of optimizing the allocation of computational resources for telecommunications networks and wireless networks. FUNDAMENTALS OF THE INVENTION AND STATE OF THE ART

[002] The rapid expansion of 5G networks and the growing demand for high-speed and reliable connectivity have driven the need for greater virtualization within the Radio Access Network (RAN) [1]. In this context, Open Radio Access Network (Open RAN) technology emerges as a significant enabler for enhanced RAN virtualization, offering several advantages, such as reduced Total Cost of Ownership (TCO) [2], greater flexibility, and the possibility of integrating components from multiple vendors. The disaggregated and interoperable components of Open RAN allow network operators to have greater customization and optimization capabilities, making it a robust solution for modern telecommunications challenges [3], [4].

[003] The virtualization of Open RAN components is enabled by key technologies such as Network Functions Virtualization (NFV) and Software Defined Networking (SDN) [5]. NFV allows the deployment of network functions on virtual machines or Petition 870240110808, dated 12 / 27 / 2024, page 6 / 56 / 16 containers, which can be dynamically managed and reconfigured. On the other hand, SDN provides centralized control over network resources, enabling efficient programmability and optimized infrastructure management. The combination of these technologies enables the flexible and efficient provision of network services for running Open RAN components, ensuring high data rates, low latency, and reliable connectivity in cloud-based environments.

[004] However, the transition to a virtualized RAN presents challenges related to resource allocation. Inadequate management of computing resources can result in significant problems, such as degradation of Quality of Service (QoS), increased latency, and inefficient energy consumption [6]. In a cloud-based environment, the dynamic and unpredictable nature of network demands requires precise and efficient resource allocation strategies. Failures in this management can lead to underutilized resources, increased operating costs, and compromised network performance.

[005] In order to address the efficient use of computational resources in NFVs of Open RAN 5G, recent studies highlight the importance of predicting resource consumption based on Key Performance Indicators (KPIs) of the RAN [7]. Accurate prediction of resource usage from real measurements allows for the implementation of proactive and more effective management strategies. This approach is essential to maintain the performance and reliability of 5G networks, simultaneously optimizing resource utilization and energy efficiency. Furthermore, KPI monitoring combined with the use of machine learning-based models has demonstrated greater efficiency in predicting the consumption of computational resources and network parameters in the 5G core [8]. However, the problem of applying such a prediction methodology to the RAN remains an open question.

[006] Thus, the method shown in this invention seeks to predict the use of computational resources of an Open RAN network server according to the KPIs obtained from network resource usage, then constructing regression models according to these obtained data, in order to predict the use of the computational resources themselves in relation to network usage. In this way, we can optimize resource allocation. Petition 870240110808, dated 12 / 27 / 2024, page 7 / 56 / 16 real-time computing, thus improving network performance, and also reducing energy consumption and operational costs of network use.

[007] Regarding state-of-the-art documents, some of these were found. One of these is US2024196252, entitled “MANAGING RESOURCES IN A RADIO ACCESS NETWORK” and published on 06 / 13 / 2024, with a translated and reproduced abstract below: “A method is disclosed for managing resources in a Radio Access Network (RAN) of a communication network. The method, executed by a first node in the RAN, comprises obtaining a record of resource status information describing the use, during a historical time period, of RAN resources controlled by a second node in the RAN and predicting, based on the record obtained, resource status information describing the use of RAN resources controlled by the second node during a future time period.The method further comprises performing at least one of the following: using the predicted resource status information describing the use of RAN resources controlled by the second node in a process related to the management of RAN resources controlled by the first node, or sending, to a third node in the RAN, a representation of the predicted resource status information describing the use of RAN resources controlled by the second node.

[008] Another document found is US2024378506, entitled “Zero-Touch Deployment and Orchestration of Network Intelligence in Open RAN Systems” and published on 11 / 14 / 2024, with a summary translated and reproduced below: “Methods and systems are provided here for deploying and orchestrating network intelligence in an Open RAN, including receiving requests in a request collector, selecting one or more applicable ML (Machine Learning) / AI (Artificial Intelligence) models by an orchestration mechanism to satisfy the plurality of collected requests, assigning at least one Open RAN resource to execute each of the ML / AI models, automatically generating executable software components incorporating at least one of the ML / AI models by an orchestration mechanism, dispatching each executable software component to the assigned Open RAN resource, and instantiating, in the Open RAN resource,at least one of the ML / AI models embedded in the executable software component to configure Open RAN to satisfy the requests”. Petition 870240110808, dated 12 / 27 / 2024, page 8 / 56 / 16

[009] Another document found is CN118511639, entitled “Network aware computing resource management cases for O-RAN non-RTRIC” and published on 08 / 16 / 2024, with the abstract translated and reproduced below: “A system and method is provided for optimizing open radio access network (O-RAN) cloud resources using an rApp from a non-RT RIC.The method comprises: acquiring, by an rApp hosted on a non-RT RIC, O1 data received through an O1 interface of an SMO framework, the SMO framework for managing and orchestrating an O-cloud platform, the O1 interface for communicating with virtualized network functions (VNFs) hosted on a plurality of physical nodes of the O-cloud platform; the rApp obtains O2 data received through an O2 interface of the SMO framework, and the O2 interface is used to communicate with an infrastructure management service (IMS) and a deployment management service (DMS) of an O-cloud platform; and generating, by the rApp, a policy to optimize an O-cloud platform or a VNF through an O1 interface of an SMO anchoring function or an SMO framework, based on at least one of the acquired O1 data and one of the acquired O2 data. OBJECTIVES AND ADVANTAGES OF THE INVENTION

[010] The present invention aims to efficiently predict the consumption of computational resources based on KPIs of virtualized systems, optimizing resource allocation in 5G Open RAN networks. To this end, it intends to develop a predictive method that uses real data and models based on data analysis and machine learning to provide accurate consumption forecasts, promoting greater energy efficiency, cost reduction, and improved overall network performance. The objectives and advantages achieved with the invention include the following points: • Prediction of computational resource consumption: Creation of predictive models that estimate the use of computational resources based on RAN KPIs, providing accurate resource usage forecasts and enabling dynamic and efficient resource allocation in real time; • Identify critical variables and KPIs: Analyze and select the most relevant KPIs for the consumption of computing resources, highlighting those with the highest... Petition 870240110808, dated 12 / 27 / 2024, page 9 / 56 / 16 impact on performance, which ensures that the method is guided by meaningful data, increasing its accuracy and reliability; • Optimize energy efficiency and operational costs: Implement strategies that minimize waste of computing and energy resources, aligning with ESG standards, which reduces costs and promotes sustainability in Open RAN networks; • Supporting proactive resource management: Providing insights based on real-world experimental data to anticipate needs and avoid overload or underutilization, which facilitates the adoption of predictive strategies for efficient and scalable management.

[011] Some possible uses of the invention are demonstrated in the following areas: • Private 5G and 6G networks: Implementation in corporate networks that use Open RAN to optimize computing resources in high-demand systems; • Data center management: Forecasting hardware energy consumption and optimizing resources in virtualized IT infrastructures; • IoT and Industry 4.0: Application in industrial networks to predict and manage the consumption of computing resources in highly connected scenarios; • Hybrid telecommunications networks: Used in networks that combine dedicated hardware functions and virtualized functions; • Cloud gaming and video streaming: Optimizing computational consumption on platforms that rely on high-performance networks to deliver real-time content.

[012] The main objective is achieved through the integration of specific objectives, which together overcome the limitations of the state of the art. The use of KPIs and real data, combined with advanced forecasting models, solves the lack of effective predictive strategies currently observed. This approach avoids problems such as resource overload, excessive energy consumption, and high costs. Furthermore, by identifying critical variables and providing tools for proactive management, the invention enables more granular and efficient management, promoting greater reliability, flexibility, and sustainability for 5G Open RAN networks. Petition 870240110808, dated 12 / 27 / 2024, page 10 / 56 / 16 GENERAL DESCRIPTION OF THE INVENTION

[013] The present invention relates to a KPI-based predictive method for forecasting and optimizing the consumption of computational resources in virtualized 5G Open RAN network systems. This method uses experimental data collected from a real-world environment and applies advanced supervised machine learning models to accurately estimate the use of computational resources. Innovative aspects include the collection and identification of critical KPIs, the analysis of the importance of variables, and the application of predictive strategies for the proactive management of computational resources.

[014] The method of the invention consists of forecasting the consumption of computational resources in 5G Open RAN networks using a KPI-based method. The approach follows the main steps below: • Data Collection: RAN performance metrics are collected periodically from an environment that includes centralized (CU) and distributed (DU) units, with measurements taken at regular intervals and stored in a database; • Selection of Critical KPIs: A statistical analysis identifies the most relevant KPIs for resource forecasting, based on their significance with computing utilization metrics; • Model Training: Supervised machine learning models guided by computational resource metrics and RAN metrics (KPIs), model performance is measured using feature evaluation and importance metrics; • Prediction and Optimization: The trained models are used to predict the consumption of computing resources in real time, allowing proactive adjustments in network management to optimize efficiency, reduce costs, and improve performance; • Validation and Application: The method is validated in an experimental testbed, where the performance of the models is compared with the actual resource consumption, ensuring the robustness and applicability of the solution in practical scenarios. Petition 870240110808, dated 12 / 27 / 2024, page 11 / 56 / 16

[015] The main application of the invention is in the efficient management of computing resources in 5G Open RAN networks, especially in virtualized environments that use Network Function Virtualization (NFV) and Software-Defined Networking (SDN).

[016] The invention is thus based on three main pillars to achieve the defined objectives: • Use of real data: Collecting metrics in an experimental environment allows models to reflect real-world scenarios, increasing the applicability and accuracy of predictions; • Advanced modeling: The application of robust regression algorithms allows the creation of models that capture the complexity of the relationships between KPIs and computational consumption; • Proactive and granular management: Analysis of critical variables and real-time forecasting enable interventions that optimize resource use, reducing costs and maximizing efficiency.

[017] With these elements, the invention resolves the limitations of the prior art, such as the absence of accurate predictions and the lack of mechanisms to manage computational consumption in highly dynamic systems. Furthermore, its modular and flexible design enables application in various scenarios beyond the context of Open RAN networks. DESCRIPTION OF THE FIGURES

[018] The invention will be described in detail below, and for better understanding, reference will be made to the attached drawings, in which are represented: Figure 1: Diagram showing the network elements of the system used to implement the method and the connections between these elements; Figure 2: Flowchart showing the method for predicting computational consumption in 5G Open RAN networks; Figure 3: Sequence diagram showing sub-steps of a possible implementation of the first stage of the method, which includes Steps 1, 2, and 3; Petition 870240110808, dated 12 / 27 / 2024, page 12 / 56 / 16 Figure 4: Sequence diagram showing sub-steps of a possible implementation of the second stage of the method, which includes Step 4; Figure 5: Sequence diagram showing sub-steps of a possible implementation of the third stage of the method, which includes Steps 5, 6, and 7. DETAILED DESCRIPTION OF THE INVENTION

[019] The SYSTEM AND METHOD FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS shows a method to be implemented in a system showing a Radio Access Network, or RAN meaning Radio Access Network, implemented using Open RAN technology, or Open Radio Access Network, in order to predict the use of the computational resources of this network.

[020] An illustrative, but not limiting, example of an aspect of the invention relates particularly to the system components used for the implementation of the proposed method, with Figure 1 demonstrating how the invention's system integrates RAN data collection, the data being specifically Key Performance Indicators, or KPIs, of network usage, and computational resource metrics in a fifth-generation mobile network using disaggregation and standardization from Open RAN. The elements shown in the system in Figure 1 and used for the implementation of the method are as follows: • 5G (5G): An environment characterized by a small, medium, or large area, ranging from one to n radiation cells, in which user equipment (UE) connects and can transmit data over the internet; • User Equipment (UE): This refers to equipment designed for consumer use. It is any device used by an end user, such as a smartphone or other mobile, portable, or tablet device with access to mobile networks; • Server (SV): a hardware and / or software system designed to provide services, resources, or data to other devices or users, known as clients, on a network. Even though the server (SV) architecture, such as x86_64 or ARM, determines aspects like power consumption, cost, and performance, most can meet computing requirements in a flexible and efficient manner; Petition 870240110808, dated 12 / 27 / 2024, page 13 / 56 / 16 • Open RAN Stack (POR): Utilizes the disaggregation of the RAN into two main components, known as Distributed Unit (DU) and Centralized Unit (CU), and adopts open and standardized interfaces for the control and management of these elements, as well as the Radio Unit (RU). This approach enables the creation of SDN controllers capable of optimizing and automating the RAN through these interfaces; • 5G Core (C5G): The network core in mobile technologies such as this invention, 5G, is the central component responsible for managing, processing, and routing data and control in a telecommunications network. It connects end users to voice, data, and other functionalities, ensuring that mobile devices can access internal and external networks, such as the internet; • Auxiliary Host (HA): A host is any computer or machine connected to a network via a defined IP address and domain that is responsible for providing resources, information, and services to users. It can be a virtual machine, a notebook, a server, a dedicated development environment, and / or any environment capable of performing software processing to collect and export metrics to a database (DB); • Database (DB): A database is a software or hardware component designed to store, organize, retrieve, and manage data in a structured or unstructured way. It can be implemented in various formats, technologies, and architectures, adapting to different application requirements. The definition of a database can be generalized to meet any type of data storage, regardless of the model or technology used; • Computational Resource Predictor Model (MP): This is a supervised machine learning model that makes predictions about the use of computational resources based on the state of the RAN, made possible through key performance indicators.

[021] The entire method of collecting, storing, processing, training, and validating the predictive model is described at a high level in Figure 2. These are grouped into Petition 870240110808, dated 12 / 27 / 2024, page 14 / 56 / 16 three major stages (E1), (E2), (E3), described below, and each stage including a certain number of steps. In general terms, the steps of the method are: a) Step 1 (1): Auxiliary Host (HA) starts data collection, where key RAN performance indicators and server computing resource metrics (SV) are collected periodically from an environment that includes Centralized Units (CU) and Distributed Units (DU); b) Step 2 (2): User equipment (UE) connects to the network; c) Step 3 (3): Data storage, these being the key performance indicators and computational resource metrics, are stored in a database (DB); d) Step 4 (4): Key performance indicator analysis, where a statistical analysis is performed to identify relevant key performance indicators based on their significance with computational utilization metrics, analyzing each indicator individually, and if an indicator is not relevant for resource prediction, this indicator is excluded from its use in the prediction model (PM), and otherwise, this indicator is maintained for use in the model (PM); e) Step 5 (5): Data preprocessing, where the collected data are normalized and prepared for training the predictive models (PM); f) Step 6 (6): Training and validation of the predictor model (PM), where supervised machine learning models are trained using the computational resource metrics and the key performance indicators of the RAN, and then the performance of the models (PM) is evaluated using metrics for evaluating and the importance of relevant key performance indicators, referred to as features; g) Step 7 (7): Prediction of computational resource consumption, where trained models (MP) are used to predict computational resource consumption in real time; h) Step 8 (8): Resource Allocation Adjustments, where proactive adjustments are made to network management based on consumption forecasts. Petition 870240110808, dated 12 / 27 / 2024, page 15 / 56 / 16 computational resources in order to optimize efficiency, reduce costs and improve network performance.

[022] We note that Step 5 (5) of Data Preprocessing is optional, and does not need to be performed in certain implementations of the method.

[023] The first stage (E1) of the method for predicting the consumption of computational resources from performance indicators in virtualized networks for the system described above in this invention consists of collecting and storing data relating to the measurements of computational resource consumption and network performance measurements, which includes Steps 1 (1), 2 (2) and 3 (3) shown in Figure 2. In an illustrative, but not limiting, example of this aspect of the invention related to this stage, illustrated in Figure 3, we have the following substeps of Steps 1 (1), 2 (2) and 3 (3) that constitute the first stage (E1): • Regarding Step 1 (1), we have the following substeps: Substep (1S1) in which the Auxiliary Host (HA) requests data collection from the server (SV), and substep (1S2) in which the server (SV) confirms the start of data collection by the Auxiliary Host (HA); Substep (1S3) of requesting a key performance indicator from the Auxiliary Host (HA) to the server and substep (1S4) of the server (SV) sending the requested indicator to the Auxiliary Host (HA), and substep (1S5) of requesting a computational usage metric from the Auxiliary Host (HA) to the server and substep (1S6) of the server (SV) sending the requested metric to the Auxiliary Host (HA), these substeps (1S3), (1S4), (1S5), (1S6) in a loop until all indicators and metrics are obtained. • Regarding Step 2 (2), we have the following substeps: In a substep (2S1) a user device (UD) requests a connection to a radio unit (RU), and in the following substep (2S2) the radio unit (RU) confirms the establishment of the connection with the user device (UD). • Referring to Step 3 (3), we have the following substep: Petition 870240110808, dated 12 / 27 / 2024, page 16 / 56 / 16 In a substep (3S1) the Auxiliary Host (HA) stores a key performance indicator or computational usage metric in the database (DB), this is repeated in a loop until all indicators and metrics are stored.

[024] In an illustrative, but not limiting, example of implementing an aspect of the invention related to the first step (E1), RAN metrics are captured on the Auxiliary Host (HA) represented by a virtual machine running a real-time monitoring system such as, for example, Prometheus Server, which receives various key performance indicators from a native script of the Open RAN stack vendor (POR). Then, computational resource metrics are captured on the Prometheus Server, receiving computational resource metrics from the native Prometheus package, i.e., Node Exporter. In another illustrative, but not limiting, example, another way to capture computational resource data could use a script developed in Python with tools from the PSUTIL library, and then export the data as text in CSV format, or Comma-Separated Value, displaying a spreadsheet.Finally, data storage in the database (DB) occurs using data maintained on the Prometheus Server, utilizing a Python library, Prometheus_Pandas, capable of converting Prometheus data to CSV format, one of the formats compatible with the Python Pandas data analysis library. As an illustrative, but not limiting, example, one way to perform key performance indicator selection, database (DB) studies, and data preprocessing is with Pandas.

[025] The second stage (E2) of the method involves analyzing the significance of virtualized network performance metrics to determine which combinations of virtualized network performance metrics are most correlated with measures of computational resource consumption. This process of analyzing key performance indicators is important to maintain the model's accuracy and soundness for good performance, a stage known in the literature as feature engineering, a method capable of selecting significant variables. In an illustrative, but not limiting, example of this aspect of the invention related to the second stage (E2), which refers to Step 4 (4), the Petition 870240110808, dated 12 / 27 / 2024, page 17 / 56 / 16 Figure 4 demonstrates how to select some of the various key performance indicators showing RAN metrics according to their significance with computational resource metrics, the prediction target of the invention, with this Figure 4 showing the following substeps of Step 4 (4), which constitutes the second stage (E2): • A substep (4S1) in which the Auxiliary Host (HA) analyzes key performance indicators and computing resource usage located in the database (DB); • In a substep (4S2), a key performance indicator is selected; in a substep (4S3), the significance level between this indicator and the use of computational resources is calculated; and then, if the indicator has low significance in relation to the use of resources, a substep (4S4) is performed to exclude the use of this indicator in the predictive model (PM), and if the indicator has high significance in relation to the use of resources, a substep (4S5) is performed to indicate the use of this indicator in the predictive model (PM).

[026] In an illustrative, but not limiting, example of the aspect of this invention related to the second stage (E2), the significance between computational resource metrics and key performance indicators can be measured using the Pearson and Spearman coefficients. If the correlation measured by these coefficients is strong, the key performance indicator is selected to be in the model feature set (MP). Examples of key performance indicators that have a high linear relationship with computational resources are: downlink and uplink transfer rate, number of downlink and uplink transport blocks, number of users connected to the cell and / or the network as a whole, total use of downlink and uplink physical resource blocks, etc. Computational resource metrics will be predicted using this subset of key performance indicators with a high linear relationship, i.e., the server hardware usage (SV) varies linearly as the RAN state changes.Illustrative, but not limiting, examples of predictable computing resource metrics include: RAM usage, CPU usage (per socket, core, or thread), Hugepage usage, average load, etc. Petition 870240110808, dated 12 / 27 / 2024, page 18 / 56 / 16

[027] The third stage (E3) of the method involves a training, validation, and real prediction pipeline (PL) of computational resource consumption from estimator models based on virtualized network performance metrics. An illustrative, but not limiting, example of this aspect of the invention related to the third stage (E3) is shown in Figure 5, where a supervised machine learning pipeline (PL) processes new data (DD) at each time interval T. This pipeline (PL) operates in a continuous cycle, adjusting the model (MP) to preserve performance and generalization, evaluated through performance indicators common to this type of model (MP), and then making predictions (PR). In Figure 5, we have the following substeps referring to Steps 5 (5), 6 (6), and 7 (7) that constitute the third stage (E3): • Steps 5 (5), 6 (6), and 7 (7) and their substeps are within a loop, repeating themselves at each time interval T defined by the method user, and regarding Step 5 (5), we have the following substeps: In a substep (5S1), a pipeline (PL) requests preprocessing of the data (DD), and in a substep (5S2) it receives this preprocessed data (DD). • Referring to Step 6 (6), we have the following substeps: In a substep (6S1) the predictor model (MP) is trained, relating the key performance indicators and the use of computational resources, then in a substep (6S2) the model (MP) is validated; if the model (MP) is validated, a substep (6S3) is performed in which the pipeline (PL) obtains the updated model (MP), and if the model (MP) is not validated, a substep (6S4) is performed in which the pipeline (PL) adjusts the data (DD) and then a substep (6S5) in which the adjusted data (DD) are used to train the model (MP) again. • Referring to Step 7 (7), we have the following substeps: In a substep (7S1) the pipeline (PL) uses the model (MP) to perform prediction of computational resource usage (PR) in relation to Petition 870240110808, dated 12 / 27 / 2024, page 19 / 56 / 16 key performance indicators, and in a substep (7S2) the pipeline (PL) obtains these prediction (PR) results.

[028] In an illustrative, but not limiting, example of the aspect of this invention related to the third stage (E3), for the construction of the predictive model (PM) we can use linear regression models, since the prediction target of the invention has a high linear relationship with the model features, or the key performance indicators selected as relevant to the analysis. In an illustrative, but not limiting, example of how this can be implemented, it is possible to use functions from the Scikit-Learn library, such as Linear Regression and Random Forest Regressor, also for non-linear models, and XGBoost. With these models, good evaluation metrics such as mean squared error and coefficient of determination are achieved.

[029] With the prediction (PR) of the use of computational resources from the RAN state, it is possible to reduce operational costs by dynamically allocating hardware resources for the processing required by the mobile network stages, such as modulation, cryptography, control and signaling, etc.

[030] BIBLIOGRAPHICAL REFERENCES • W. Attaoui, E. Sabir, H. Elbiaze, and M. Guizani, “Vnf and cnf placement in 5g: Recent advances and future trends,” IEEE Transactions on Network and Service Management, vol. 20, no. 4, pp. 4698-4733, 2023. • M. Polese, M. Dohler, F. Dressler, M. Erol-Kantarci, R. Jana, R. Knopp, and T Melodia, “Empowering the 6G Cellular Architecture With Open RAN,” IEEE Journal on Selected Areas in Communications, vol. 42, no. 2, pp. 245-262, 2024. • M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T Melodia, “Understanding ORAN: Architecture, Interfaces, Algorithms, Security, and Research Challenges,” IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 1376-1411,2023. • O-RAN ALLIANCE e.V., “O-ran architecture description 11.0,” O-RAN ALLIANCE, Workgroup Report O-RAN.WG1.OAD-R003-v11.00, February 2024, wG1: Use Cases and Overall Architecture Workgroup. • S. Ramanathan, K. Kondepu, T Zhang, B. Mirkhanzadeh, M. Razo, M. Tacca, L. Valcarenghi, and A. Fumagalli, “Orchestrating virtualized core network migration in Petição 870240110808, de 27 / 12 / 2024, pág. 20 / 56 / 16 open roadm sdn-enabled network,” in 2021 International Conference on Optical Network Design and Modeling (ONDM), 2021, pp. 1-6. • N. R. Mohan and E. B. Raj, “Resource allocation techniques in cloud computing research challenges for applications,” in 2012 Fourth International Conference on Computational Intelligence and Communication Networks, 2012, pp. 556-560. • S. Barrachina-Mu~noz, M. Payar'o, and J. Mangues-Bafalluy, “Cloud-native 5g experimental platform with over-the-air transmissions and end-to-end monitoring,” in 2022 13th International Symposium on Communication Systems, Networks and Digital Signal Processing (CSNDSP), 2022, pp. 692-697. • C.-N. Nhu and M. Park, “Dynamic network slice scaling assisted by attentionbased prediction in 5g core network,” IEEE Access, vol. 10, pp. 72 955-72 972, 2022. Petição 870240110808, de 27 / 12 / 2024, pág. 21 / 56

Claims

1 / 5 CLAIMS 1) SYSTEM FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS, a system for predicting computational resource consumption in virtualized networks, characterized by including: a) Server (SV), configured to host an Open RAN Stack (POR) composed of a Centralized Unit (CU) and a Distributed Unit (DU); b) Auxiliary Host (HA), configured to collect and export RAN performance metrics and computational resource metrics; c) Database (DB), configured to store the collected metrics; d) Computational Resource Predictor Model (MP), configured to receive data from the Database (DB) for training, testing, and validation of machine learning models intended to predict computational resource consumption; e) Radio Unit (RU), configured to manage data traffic between User Equipment (UE) and the Open RAN Stack (POR). 2) A SYSTEM FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS, according to claim 1, characterized in that the Computational Resource Prediction Model (PM) implements machine learning algorithms to perform resource consumption predictions. 3) METHOD FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS, using the system of claims 1 and 2, characterized by including the following main steps: a) Step 1 (1): Auxiliary Host (HA) starts data collection, where key RAN performance indicators and server (SV) computing resource metrics are periodically collected from an environment that includes Centralized Units (CU) and Distributed Units (DU); Petition 870240110808, dated 12 / 27 / 2024, page 22 / 56 2 / 5 b) Step 2 (2): User Equipment (UE) connects to the network; c) Step 3 (3): Storage of key performance indicators and computing resource metrics in a database (DB);d) Step 4 (4): Key performance indicator analysis, where a statistical analysis is performed to identify relevant key performance indicators based on their significance with computational utilization metrics, analyzing each indicator individually, and if an indicator is not relevant for resource prediction, this indicator is excluded from its use in the prediction model (PM), and otherwise, this indicator is maintained for use in the model (PM); e) Step 6 (6): Training and validation of the predictive model (PM), where supervised machine learning models (PM) are trained using computational resource metrics and the key performance indicators of the RAN, and then the performance of the models (PM) is evaluated using metrics for evaluating and assessing the importance of relevant key performance indicators;f) Step 7 (7): Prediction of computational resource consumption, where trained models (PM) are used to predict computational resource consumption in real time; g) Step 8 (8): Resource Allocation Adjustments, where proactive adjustments are made to network management based on computational resource consumption predictions made through the predictive model (PM) in order to optimize efficiency, reduce costs and improve network performance. 4) METHOD FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS, according to claim 3, characterized by additionally performing a Data Preprocessing Step 5 (5), where the collected data are normalized and prepared for training the predictor models (PM). 5) METHOD FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS, of Petition 870240110808, of 12 / 27 / 2024, page 23 / 56 3 / 5 according to claim 3, characterized in that Step 1 (1) is carried out through the following substeps: • Substep (1S1) in which the Auxiliary Host (HA) requests data collection from the server (SV), and substep (1S2) in which the server (SV) confirms the start of data collection by the Auxiliary Host (HA); • Substep (1S3) of requesting a key performance indicator from the Auxiliary Host (HA) to the server and substep (1S4) of the server (SV) sending the requested indicator to the Auxiliary Host (HA), and substep (1S5) of requesting a computational usage metric from the Auxiliary Host (HA) to the server and substep (1S6) of the server (SV) sending the requested metric to the Auxiliary Host (HA), these substeps (1S3), (1S4), (1S5), (1S6) in a loop until all indicators and metrics are obtained;Step 2 (2) is performed through the following substeps: • In a substep (2S1) a user device (UD) requests a connection to a radio unit (RU), and in the following substep (2S2) the radio unit (RU) confirms the establishment of the connection with the user device (UD); and Step 3 (3) is performed through the following substeps: • In a substep (3S1) the Auxiliary Host (AH) stores a key performance indicator or computational usage metric in the database (DB), this is repeated in a loop until all indicators and metrics are stored. 6) METHOD FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS, according to claim 3, characterized in that Step 4 (4) is performed through the following substeps: • A substep (4S1) in which the auxiliary Host (HA) analyzes the key performance and computational resource usage indicators located in the database (DB); • A substep (4S2) selects a key performance indicator, a substep (4S3) in which the significance level between this indicator and the use of computational resources is calculated; and then if the indicator has low significance Petition 870240110808, dated 12 / 27 / 2024, p. 24 / 56 4 / 5 in relation to resource use, a sub-step (4S4) is performed to exclude the use of this indicator in the predictor model (MP), and if the indicator has high significance in relation to resource use, a sub-step (4S5) is performed to indicate the use of this indicator in the predictor model (MP). 7) METHOD FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS, according to claims 3 and 4, characterized in that Steps 5 (5), 6 (6), and 7 (7) and their substeps are within a loop, repeating themselves at each time interval T defined by the method user, and Step 5 (5) is performed through the following substeps: • In a substep (5S1), a pipeline (PL) requests preprocessing of the data (DD), and in a substep (5S2) receives this preprocessed data (DD); and Step 6 (6) is carried out through the following substeps: • In a substep (6S1) the predictor model (PM) is trained, relating the key performance indicators and the use of computational resources, then in a substep (6S2) the model (PM) is validated;If the model (MP) is validated, a substep (6S3) is performed in which the pipeline (PL) obtains the updated model (MP), and if the model (MP) is not validated, a substep (6S4) is performed in which the pipeline (PL) adjusts the data (DD), and then a substep (6S5) is performed in which the adjusted data (DD) is used to train the model (MP) again; and Step 7 (7) is performed through the following substeps: • In a substep (7S1) the pipeline (PL) uses the model (MP) to perform prediction of computational resource usage (PR) in relation to key performance indicators, and in a substep (7S2) the pipeline (PL) obtains these prediction results (PR). 8) METHOD FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS, according to claims 3 and 6, characterized in that Step 4 (4) uses calculation of Petition 870240110808, dated 12 / 27 / 2024, page 25 / 56 5 / 5 correlation coefficients to determine the relevance of key performance indicators. 9) METHOD FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS, according to claims 3 and 7, characterized in that Step 6 (6) of Training and validation of the predictive model (PM) includes the use of metrics such as mean squared error and coefficient of determination to evaluate the performance of the predictive models (PM). 10) METHOD FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS, according to claim 3, characterized by using the Prometheus Server real-time monitoring system, which receives various key performance indicators from a native script of the Open RAN (POR) stack vendor, and the capture of computational resource metrics in the Prometheus Server is carried out through the native Node Exporter package of Prometheus; and the storage in the database (DB) occurs with the data maintained in the Prometheus Server, and using a Python library Prometheus_Pandas to convert the Prometheus data to CSV format. 11) METHOD FOR PREDICTING COMPUTATIONAL RESOURCE CONSUMPTION BASED ON PERFORMANCE INDICATORS IN VIRTUALIZED NETWORKS, according to claim 3, characterized in that the capture of computational resource data uses a script developed in the Python language with tools from the PSUTIL library, and is then exported as text in CSV format. Petition 870240110808, dated 12 / 27 / 2024, p. 26 / 56

Images