Internet Research Task Force M. Li Internet-Draft C. Zhou Intended status: Informational D. Chen Expires: 8 January 2025 China Mobile 7 July 2024 Data Generation and Optimization for Digital Twin Network Performance Modeling draft-li-nmrg-dtn-data-generation-optimization-02 Abstract Digital Twin Network (DTN) can be used as a secure and cost-effective environment for network operators to evaluate network performance in various what-if scenarios. Recently, AI models, especially neural networks, have been applied for DTN performance modeling. The quality of deep learning models mainly depends on two aspects: model architecture and data. This memo focuses on how to improve the model from the data perspective. Requirements Language The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 [RFC2119]. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at https://datatracker.ietf.org/drafts/current/. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on 8 January 2025. Copyright Notice Copyright (c) 2024 IETF Trust and the persons identified as the document authors. All rights reserved. Li, et al. Expires 8 January 2025 [Page 1] Internet-Draft Data Generation and Optimization for DTN July 2024 This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/ license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License. Table of Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2 2. Acronyms & Abbreviations . . . . . . . . . . . . . . . . . . 3 3. Requirements . . . . . . . . . . . . . . . . . . . . . . . . 3 4. Framework of Data Generation and Optimization . . . . . . . . 4 4.1. Data Generation Stage . . . . . . . . . . . . . . . . . . 5 4.2. Data Optimization Stage . . . . . . . . . . . . . . . . . 6 5. Data Generation . . . . . . . . . . . . . . . . . . . . . . . 6 5.1. Network Topology . . . . . . . . . . . . . . . . . . . . 6 5.2. Routing Policy . . . . . . . . . . . . . . . . . . . . . 7 5.3. Traffic Matrix . . . . . . . . . . . . . . . . . . . . . 7 6. Data Optimization . . . . . . . . . . . . . . . . . . . . . . 8 6.1. Seed Sample Selection Phase . . . . . . . . . . . . . . . 8 6.2. Incremental Optimization Phase . . . . . . . . . . . . . 9 7. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 9 8. Security Considerations . . . . . . . . . . . . . . . . . . . 10 9. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 10 10. References . . . . . . . . . . . . . . . . . . . . . . . . . 10 10.1. Informative References . . . . . . . . . . . . . . . . . 10 10.2. Normative References . . . . . . . . . . . . . . . . . . 10 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 11 1. Introduction Digital twin is a virtual instance of a physical system (twin) that is continually updated with the latter's performance, maintenance, and health status data throughout the physical system's life cycle. Digital Twin Network (DTN) is a digital twin that is used in the context of networking [I-D.irtf-nmrg-network-digital-twin-arch]. DTN can be used as a secure and cost-effective environment for network operators to evaluate network performance in various what-if scenarios. Recently, AI models, especially neural networks, have been applied for DTN performance modeling. The quality of AI models mainly depends on two aspects: model architecture and data. This memo focuses on the impact of training data on the model. The quality of training data will directly affect the accuracy and generalization ability of the model. This memo Li, et al. Expires 8 January 2025 [Page 2] Internet-Draft Data Generation and Optimization for DTN July 2024 focuses on how to design data generation and optimization methods for DTN performance modeling, which can generate simulated network data to solve the problem of practical data shortage and select high- quality data from various data sources. Using high-quality data for training can improve the accuracy and generalization ability of the model. 2. Acronyms & Abbreviations DTN: Digital Twin Network AI: Artificial Intelligence AIGC: AI-Generated Content ToS: Type of Service OOD: Out-of-Distribution FIFO: First In First Out SP: Strict Priority WFQ: Weighted Fair Queuing DRR: Deficit Round Robin BFS: Breadth-First Search CBR: Constant Bit Rate 3. Requirements Performance modeling is vital in DTN, which is involved in typical network management scenarios such as planning, operation, optimization, and upgrade. Recently, some studies have applied AI models to DTN performance modeling, such as RouteNet [RouteNet] and MimicNet [MimicNet]. AI is a data-driven technology whose performance heavily depends on data quality. Network data sources are diverse and of varying quality, making it difficult to directly serve as training data for DTN performance models: * Practical data from production networks: Data from production networks usually have high value, but the quantity, type, and accuracy are limited. Moreover, it is not practical in production networks to collect data under various configurations; Li, et al. Expires 8 January 2025 [Page 3] Internet-Draft Data Generation and Optimization for DTN July 2024 * Network simulators: Network simulators (e.g., NS-3 and OMNeT++) can be used to generate simulated network data, which can solve the problems of quantity, diversity, and accuracy to a certain extent. However, simulation is usually time-consuming. In addition, there are usually differences between simulated data and practical data from production networks, which hinders the application of trained models to production networks; * Generative AI models: With the development of AI-Generated Content (AIGC) technology, generative AI models (e.g., GPT and LLaMA) can be used to generate simulated network data, which can solve the problems of quantity and diversity to a certain extent. However, the accuracy of the data generated by generative AI models is limited and often has gaps with practical data from production networks. Therefore, data generation and optimization methods for DTN performance modeling are needed, which can generate simulated network data to solve the problem of practical data shortage and select high- quality data from multi-source data. High-quality data meets the requirements of high accuracy, diversity, and fitting the actual situation of practical data. Training with high-quality data can improve the accuracy and generalization of DTN performance models. 4. Framework of Data Generation and Optimization The framework of data generation and optimization for DTN performance modeling is shown in Figure 1, which includes two stages: the data generation stage and the data optimization stage. Li, et al. Expires 8 January 2025 [Page 4] Internet-Draft Data Generation and Optimization for DTN July 2024 Data generation Data optimization +---------------------------+ +-------------------------------------+ | | | | | +---------+ | | +---------+ | | | | | | +----------+ | | | | | Network | | | | Practical| | Easy | | | | topology| +-----------+ | | | data | | samples | | | | | | | | | +-----+----+ | | | | | | | Network | | | | | | +--------+ | | | | | simulator | | | +-----v----+ | | | | | | | Routing | | | | | | | | Hard | | High | | | | policy +-> +-+-+-> Candidate+-> samples +-> quality| | | | | | | | | | data | | | | data | | | | | | Generative| | | | | | | | | | | | | | AI model | | | +----------+ | | +--------+ | | | Traffic | | | | | | OOD | | | | matrix | +-----------+ | | | samples | | | | | Data generator| | | (remove)| | | +---------+ | | | | | | Network | | +---------+ | | configuration | | Data selection | | | | | +---------------------------+ +-------------------------------------+ Figure 1: Framework of Data Generation and Optimization for DTN Performance Modeling 4.1. Data Generation Stage The data generation stage aims to generate candidate data (simulated network data) to solve the problem of the shortage of practical data from production networks. This stage first generates network configurations and then imports them into data generators to generate the candidate data. * Network configurations: Network configurations typically include network topology, routing policy, and traffic matrix. These configurations need to be diverse to cover as many scenarios as possible. Topology configurations include the number and structure of nodes and edges, node buffers' size and scheduling strategy, link capacity, etc. Routing policy determines the path of a packet from the source to the destination. The traffic matrix describes the traffic entering/leaving the network, which includes the traffic's source, destination, time and packet size distribution, Type of Service (ToS), etc. Li, et al. Expires 8 January 2025 [Page 5] Internet-Draft Data Generation and Optimization for DTN July 2024 * Data generators: Data generators can be network simulators (e.g., NS-3 and OMNeT++) and/or the generative AI models (e.g., GPT and LLaMA). Network configurations are imported into data generators to generate candidate data. 4.2. Data Optimization Stage The data optimization stage aims to optimize the candidate data from various sources to select high-quality data. * Candidate data: Candidate data includes simulated network data generated in the data generation stage and the practical data from production networks. * Data selection: The data selection module investigates the candidate data to filter out the easy, hard, and Out-of- Distribution (OOD) samples. Hard examples refer to samples that are difficult for the model to accurately predict. During the training process, exposing the model to more hard examples will enable it to perform better on such samples later on. Then the easy samples and hard samples are considered valid samples and added to the training data. OOD samples are considered invalid and removed. * High-quality data: High-quality data needs to meet the requirements of high accuracy, diversity, and fitting the actual situation of practical data, which can be verified by expert knowledge (such as the ranges of delay, queue utilization, link utilization, and average port occupancy). 5. Data Generation This section will describe how to generate network configurations, including network topology, routing policy, and traffic matrix. Then these configurations will be imported into data generators to generate the candidate data. 5.1. Network Topology Network topologies are generated using the Power-Law Out-Degree algorithm, where parameters are set according to real-world topologies in the Internet Topology Zoo. Li, et al. Expires 8 January 2025 [Page 6] Internet-Draft Data Generation and Optimization for DTN July 2024 When the flow rate exceeds the link bandwidth or the bandwidth set for the flow, the packet is temporarily stored in the node buffer. A larger node buffer size means a larger delay and possibly a lower packet loss rate. The node scheduling policy determines the time and order of packet transmission, which is randomly selected from the policies such as First In First Out (FIFO), Strict Priority (SP), Weighted Fair Queuing (WFQ), and Deficit Round Robin (DRR). A larger link capacity means a smaller delay and less congestion. To cover diverse link loads to get good coverage of possible scenarios, we set the link capacity to be proportional to the total average bandwidth of the flows passing through the link. 5.2. Routing Policy Routing policy plays a crucial role in routing protocols, which determines the path of a packet from the source to the destination. * Default: We set the weight of all links in the topology to be the same, that is, equal to 1. Then we use the Dijkstra algorithm to generate the shortest path configuration. Dijkstra algorithm uses Breadth-First Search (BFS) to find the single source shortest path in a weighted digraph. * Variants: We randomly select some links (the same link can be chosen more than once) and add a small weight to them. Then we use the Dijkstra algorithm to generate a series of variants of the default shortest path configuration based on the weighted graph. These variants can add some randomness to the routing configuration to cover longer paths and larger delays. 5.3. Traffic Matrix The traffic matrix is very important for network performance modeling. The traffic matrix can be regarded as a network map, which describes the traffic entering/leaving the network, including the source, destination, distribution of the traffic, etc. We generate traffic matrix configurations with variable traffic intensity to cover low to high loads. The parameters packet sizes, packet size probabilities, and ToS are generated according to the validation dataset analysis to have similar distributions. The arrival of packets for each source-destination pair is modeled using one of the time distributions such as Poisson, Constant Bit Rate (CBR), and ON-OFF. Li, et al. Expires 8 January 2025 [Page 7] Internet-Draft Data Generation and Optimization for DTN July 2024 6. Data Optimization This section will describe how to optimize the data from various sources to filter out high-quality data, which includes the seed sample selection phase and incremental optimization phase. Candidate data includes simulated network data generated in the data generation stage and real data from production networks. Data optimization supports a variety of selection strategies, including high fidelity, high coverage, etc. High fidelity means that the selected data can fit the real data (e.g., having similar topologies, routing policies, traffic models, etc.), and high coverage means that the selected data can cover as many scenarios as possible. 6.1. Seed Sample Selection Phase In the seed sample selection phase, high-quality seed samples are selected through the following steps to provide high-quality initial samples for the incremental optimization phase. STEP 1: Training feature extraction model and feature extraction. (1.1) The training data D' is selected from the candidate data D according to the selection strategy. For the high fidelity strategy, the real data is used as the training data D'; for the high coverage strategy, the real data and simulated data are used together as the training data D'. (1.2) Feature extraction model E is trained using the training data D'. Feature extraction model E is a network performance evaluation model that can be used to evaluate performance indicators such as delay, jitter and packet loss (such as RouteNet). (1.3) Use the feature extraction model E obtained in STEP (1.2) to extract the feature of the training data D' obtained in STEP (1.1). A network can be defined as a set of flow F, queue Q, and link L. The link state SF (such as link utilization), queue state SQ (such as port occupation), and flow state SL (such as delay, throughput, packet loss, etc.) are taken as features. Each sample in the training data D' is converted to a feature vector [SF,SQ,SL]. STEP 2: Clustering. Cluster the training data D' after feature extraction. Clustering (such as K-means and DBSCAN) is an unsupervised machine learning technique that can automatically discover the natural groups in the data, divide the data into multiple clusters, and the samples in the same cluster have similarities. Li, et al. Expires 8 January 2025 [Page 8] Internet-Draft Data Generation and Optimization for DTN July 2024 Repeat STEP 3 and STEP 4 until all clusters have been traversed. STEP 3: Calculating cluster centers and nearest neighbors. (3.1) Calculate cluster centers. The method of calculating cluster centers is determined according to the clustering algorithm used in STEP 2. For example, using K-means clustering algorithm, the cluster center is calculated by finding the average of all data points in the cluster. These cluster centers are added to the seed dataset DS. (3.2) Calculate k nearest neighbors of each cluster center and add them to the seed dataset DS. Suitable nearest neighbor calculation methods can be used, such as Euclidean distance, cosine distance, etc. STEP 4: Expert knowledge verification. (4.1) Expert knowledge can be used to verify the validity of samples through the range of indicators such as delay, queue occupation, and link utilization. If the verification passed, go to STEP 3. Otherwise, go to STEP (4.2). (4.2) Randomly select m samples from the seed dataset DS and remove them. Calculate the nearest neighbors of the removed m samples, add them to the seed data set DS, and go to STEP (4.1). 6.2. Incremental Optimization Phase The seed samples are taken as the initial training dataset. The filter model investigates the remaining candidate samples to filter out the easy, hard and OOD samples. Then the easy samples and hard samples are added to the training dataset. These processes are repeated to iteratively optimize the filter model and the training data until the high-quality data meets the constraints. 7. Discussion Several topics related to data generation and optimization for DTN performance modeling require further discussion. * Data generation methods: 1) Generate configurations that cover enough scenarios and scale from small to large networks. 2) Choose data generators that consider accuracy, speed, fidelity, etc. 3) Use data augmentation technology to expand the training data by using a small amount of practical data to generate similar data through prior knowledge. Li, et al. Expires 8 January 2025 [Page 9] Internet-Draft Data Generation and Optimization for DTN July 2024 * Data optimization methods: 1) Select data from multi-source candidate data, including hard sample mining, OOD detection, etc. 2) Verify whether the data quality meets the requirements. * Deployment: 1) Time/space complexity and explainability of the data generation and optimization methods. 2) Provide feedback for data collection to form a closed loop. 8. Security Considerations TBD 9. IANA Considerations This document has no requests to IANA. 10. References 10.1. Informative References [I-D.irtf-nmrg-network-digital-twin-arch] Zhou, C., Yang, H., Duan, X., Lopez, D., Pastor, A., Wu, Q., Boucadair, M., and C. Jacquenet, "Network Digital Twin: Concepts and Reference Architecture", Work in Progress, Internet-Draft, draft-irtf-nmrg-network-digital- twin-arch-05, 4 March 2024, . [MimicNet] Zhang, Q. Zhang., NG, K. K.W. NG., Kazer, C. W. Kazer., Yan, S. Yan., Sedoc, J. Sedoc., and V. Liu. Liu, "MimicNet: Fast Performance Estimates for Data Center Networks with Machine Learning. In ACM SIGCOMM 2021 Conference (SIGCOMM ’21).", August 2021. [RouteNet] Rusek, K. Rusek., Suárez-Varela, J. Suárez-Varela., Almasan, P. Almasan., Barlet-Ros, P. Barlet-Ros., and A. Cabellos-Aparicio. Cabellos-Aparicio, "RouteNet: Leveraging Graph Neural Networks for network modeling and optimization in SDN. IEEE Journal on Selected Areas in Communication (JSAC), vol. 38, no. 10", October 2020. 10.2. Normative References [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, . Li, et al. Expires 8 January 2025 [Page 10] Internet-Draft Data Generation and Optimization for DTN July 2024 Authors' Addresses Mei Li China Mobile Beijing 100053 China Email: limeiyjy@chinamobile.com Cheng Zhou China Mobile Beijing 100053 China Email: zhouchengyjy@chinamobile.com Danyang Chen China Mobile Beijing 100053 China Email: chendanyang@chinamobile.com Li, et al. Expires 8 January 2025 [Page 11]