Source-linked AI summary
Knowledge-Defined Networking
Albert Mestres, Alberto Rodriguez-Natal, Josep Carner, Pere Barlet-Ros, Eduard Alarcón, Marc Solé, Victor Muntés, David Meyer, Sharon Barkai, Mike J Hibbett, Giovani Estrada, Khaldun Ma`ruf, Florin Coras, Vina Ermagan, Hugo Latapie, Chris Cassar, John Evans, Fabio Maino, Jean Walrand, Albert Cabellos
TL;DR
The paper addresses the limited prototyping and deployment of AI-based network control, especially the Knowledge Plane. It proposes Knowledge-Defined Networking, combining SDN, network analytics, and ML, and presents use cases with preliminary evidence of feasibility while identifying representative-training-set challenges.
Problem
AI-based network control, including the proposed Knowledge Plane, had not been extensively prototyped or deployed, while distributed networks complicate learning for network-wide control.
Method
The paper combines SDN’s centralized control and network analytics’ centralized network view with ML and DL in a Knowledge-Defined Networking paradigm.
Results
The paper presents networking use cases and preliminary experimental evidence illustrating KDN’s applicability, benefits, and technical feasibility.
Takeaways & Limitations
KDN provides a paradigm for automated network control by combining Software-Defined Networking, Network Analytics, and Machine Learning.
Takeaways & Limitations
Representative training sets remain an open challenge because model accuracy depends on network characteristics and training-set size, while ML models lack deterministic guarantees and can be difficult for humans to understand.
Abstract
from arXiv · showhide
The research community has considered in the past the application of Artificial Intelligence (AI) techniques to control and operate networks. A notable example is the Knowledge Plane proposed by D.Clark et al. However, such techniques have not been extensively prototyped or deployed in the field yet. In this paper, we explore the reasons for the lack of adoption and posit that the rise of two recent paradigms: Software-Defined Networking (SDN) and Network Analytics (NA), will facilitate the adoption of AI techniques in the context of network operation and control. We describe a new paradigm that accommodates and exploits SDN, NA and AI, and provide use cases that illustrate its applicability and benefits. We also present simple experimental results that support its feasibility. We refer to this new paradigm as Knowledge-Defined Networking (KDN).
I. INTRODUCTION
The paper revisits AI-based network control, explaining that distributed network complexity limited earlier Knowledge Plane adoption. It argues that SDN centralization and network analytics provide the control and network-wide visibility needed for KDN.
- I. INTRODUCTION: Earlier Knowledge Plane proposals promised automated operation and recommendations but had not yet been extensively prototyped or deployed.The Knowledge Plane was intended to recognize network conditions, act automatically, and explain or suggest actions.
- I. INTRODUCTION: Distributed network nodes have only partial system visibility and control, making learning for network-wide actions complex.SDN’s logically centralized control plane can ease this complexity by providing a logical single point with knowledge of the whole network.
- I. INTRODUCTION: Network telemetry supplies real-time packet- and flow-level information plus configuration and network-state data to centralized analytics platforms.These technologies provide a richer network view than conventional network management approaches.
- I. INTRODUCTION: KDN combines SDN’s centralized control with network analytics’ centralized view so ML and DL can learn network knowledge and control the network.The paradigm combines SDN, telemetry, Network Analytics, and the Knowledge Plane.
- I. INTRODUCTION: The paper describes KDN, presents networking use cases and early feasibility experiments, and discusses open research challenges.The experimental results are provided for some use cases rather than across the entire paradigm.
II. A KNOWLEDGE PLANE FOR SDN ARCHITECTURES
KDN adds a Knowledge Plane to SDN’s traditional planes, using learned behavioral models and reasoning to support network decisions. The existing planes provide the data, control, management, and telemetry context that the Knowledge Plane exploits.
- II. A KNOWLEDGE PLANE FOR SDN ARCHITECTURES: Adding a Knowledge Plane to SDN’s traditional three planes produces the Knowledge-Defined Networking paradigm.The paper presents Fig. 1 as an overview of KDN’s functional planes.
- II. A KNOWLEDGE PLANE FOR SDN ARCHITECTURES: The Data Plane stores, forwards, and processes packets through programmable forwarding hardware that is unaware of the rest of the network.Other planes populate forwarding tables and update configuration.
- II. A KNOWLEDGE PLANE FOR SDN ARCHITECTURES: The logically centralized SDN Control Plane exchanges operational state and programs forwarding elements through a southbound interface.The data plane operates at packet time scales, whereas the control plane typically operates at flow time scales.
- II. A KNOWLEDGE PLANE FOR SDN ARCHITECTURES: The Management Plane provisions devices, defines topology, monitors network operation, and maintains historical telemetry and network-state records.In SDN, these management functions are usually handled by the controller as well.
- II. A KNOWLEDGE PLANE FOR SDN ARCHITECTURES: The Knowledge Plane integrates behavioral models and reasoning for decision making, learning network behavior and sometimes operating the network automatically.It uses the control and management planes to obtain a rich view of and control over the network.
III. KNOWLEDGE-DEFINED NETWORKING
KDN uses an operational control loop in which analytics feed ML models whose knowledge informs network decisions and control actions. The paradigm supports automation, optimization, recommendation, validation, and estimation through SDN-based actuation.
- III. KNOWLEDGE-DEFINED NETWORKING: KDN’s control loop supports automation, recommendation, optimization, validation, and estimation.The paradigm draws on black-box optimization, neural networks in feedback control, and autonomic self-* architectures.
- III. KNOWLEDGE-DEFINED NETWORKING: The KDN operational loop gathers forwarding and controller information into an analytics platform to build a complete network view.It monitors data-plane elements in real time and queries the SDN controller for control and management state.
- III. KNOWLEDGE-DEFINED NETWORKING: Analytics data feed ML and DL algorithms that learn network behavior and generate knowledge such as a network model.The platform uses both current and historical data.
- III. KNOWLEDGE-DEFINED NETWORKING: Supervised learning relates network variables to operation using labeled data, unsupervised learning discovers data structure without labels, and reinforcement learning seeks actions leading to optimal configurations.These approaches support different forms of model building, discovery, and decision learning.
- III. KNOWLEDGE-DEFINED NETWORKING: KDN applications differ according to whether the network operator participates in the decision-making process.The paper summarizes these application types in Table I.
- III. KNOWLEDGE-DEFINED NETWORKING: Intent-driven languages translate high-level network directives into specific imperative control actions through the centralized SDN controller.The controller can render these directives because it has global network visibility and can actuate on network devices.
- III. KNOWLEDGE-DEFINED NETWORKING: The controller pushes parsed control actions to forwarding devices through southbound protocols or management protocols such as NETCONF.Forwarding elements then operate using the updated state and configuration.
IV. USE-CASES
The paper uses networking scenarios to illustrate KDN’s applications and ML benefits, with early experiments for two representative use cases. The datasets are publicly available through the cited repository.
- IV. USE-CASES: The use cases illustrate KDN applications and ML benefits for common networking problems.Two representative use cases also receive early experimental results supporting technical feasibility.
A. Routing in an Overlay Network
The paper applies ML to hidden-underlay overlay networks, where operators must optimize routing using limited control over edge links. It models underlay behavior from observed traffic and routing inputs to improve overlay routing.
- A. Routing in an Overlay Network: An ANN proof of concept estimates hidden-underlay path delay so the resulting model can later support improved overlay routing.The use case targets a network whose underlay is hidden from the overlay.
- A. Routing in an Overlay Network: Overlay operators control traffic paths only indirectly by distributing traffic across edge-node links.The underlay’s topology and configuration are hidden, limiting direct routing control.
- A. Routing in an Overlay Network: Finding optimal per-link policies is difficult because edge nodes lack underlay routing control while global performance depends on their coordinated choices.Efficient link use is the overlay operator’s main means of influencing underlay paths.
- A. Routing in an Overlay Network: Building an explicit underlay model is challenging because its topology and routing policy are unknown, and theoretical models may not capture the scenario’s complexity.The hidden administrative domain prevents the overlay administrator from directly observing key underlay details.
- A. Routing in an Overlay Network: ML can model the hidden underlay by learning correlations between overlay inputs—routing policy and traffic—and resulting performance.The paper expresses this relationship as f(routing policy, traffic) = performance.
1) Experimental Results:
The experiment tests whether an ANN can learn hidden-underlay behavior from overlay traffic and routing information. Its accuracy improves with training data, reaching roughly 1% relative error with 3,000 training samples.
- 1) Experimental Results:: The simulated topology contains 12 overlay nodes, 19 underlay elements, and 72 links, while the KP observes only traffic-sending and traffic-receiving overlay nodes.The underlay remains hidden from the KP perspective.
- 1) Experimental Results:: The ANN uses overlay traffic between node pairs and per-link traffic ratios as inputs, predicting average path delays without underlay information.Training uses 9,600 samples and validation uses 300 separate samples.
- 1) Experimental Results:: Roughly 1% relative error is achieved with 3,000 training samples when predicting average delay across paths.The error is averaged over all paths in each of 300 test samples.
- 1) Experimental Results:: Mean Squared Error decreases as the training-set size increases, following the typical exponential-delay pattern described for machine learning.Figure 4 plots MSE against training-data-set size.
B. Resource Management in an NFV scenario
The NFV use case models how VNF behavior and resource consumption depend on complex traffic and configuration factors. ANN-based models are evaluated for CPU consumption and support VNF placement optimization without modifying the network.
- B. Resource Management in an NFV scenario: VNF placement can substantially affect overall virtualized-network performance, making NFV resource management more complex than placement in mostly static data centers.The challenge arises because placement changes the performance of the virtualized network.
- B. Resource Management in an NFV scenario: NFV placement models face complexity rather than missing information because VNF behavior depends on many available factors, including topology, resource use, configuration, and traffic.This makes accurate modeling challenging even when the relevant information is collected.
- B. Resource Management in an NFV scenario: KDN uses ML to characterize VNF behavior from collected analytics and model resource requirements without modifying the network.The resulting model can support VNF placement and overall network-performance optimization.
- B. Resource Management in an NFV scenario: The experiment models CPU consumption for OVS firewall, OVS switch, and Snort configurations using traffic features extracted in 20-second batches.The input contains 86 traffic features, and CPU consumption is the output.
- B. Resource Management in an NFV scenario: The models use 600 training samples for the OVS firewall and 900 for Snort and the OVS switch, with separate test samples.The study first examines single-feature predictions before evaluating the full feature set.
- B. Resource Management in an NFV scenario: Single-feature models are inaccurate and nonlinear, while the full-feature model achieves reasonable relative-error accuracy across the evaluated VNFs.The firewall uses number of flows as the single predictor, whereas Snort uses number of packets.
C. Knowledge extraction from network logs
Network logs support KDN monitoring and troubleshooting, while unsupervised learning can correlate log events to discover knowledge for open-loop administration or closed-loop automation.
- Network devices report events such as link failures and packet losses, which operators use to monitor network health and troubleshoot issues.
- Within KDN, unsupervised learning can correlate log events and discover new knowledge for network operation.
- Discovered knowledge can support administrators through an open-loop approach or be handled automatically through a closed-loop solution using the Intent interface.
- Table II presents examples of knowledge discovery using network logging and unsupervised learning.
D. Short and long-term network planning
KDN uses historical analytics data to improve network planning by learning demand–load relationships and estimating when upgrades are required. This addresses planning processes that can otherwise produce over-provisioning.
- Network planning responds to increasing traffic loads and service requirements by designing topology, selecting hardware, and setting traffic policies.
- Expert-managed capacity models and forecasts are prone to errors, so network planning typically results in over-provisioning.
- KDN can develop a network model from historical analytics data to support network planning.
- By learning the relation between client or service counts and network load, KDN can estimate when a network upgrade is required.
V. CHALLENGES AND DISCUSSION
The paper identifies challenges that KDN introduces, including new machine-learning mechanisms, representative training data, new expertise, and standardized datasets. It also advocates public training sets for computer-network AI research.
- KDN requires adapting existing machine-learning mechanisms or developing new ones because networking introduces a new ML application domain.
- Graph-based network topologies remain difficult to model because only preliminary ML approaches have addressed systems represented through graphs.
- KDN models may lack guarantees and human interpretability, making representative training sets and their relationship to accuracy an important research question.
- KDN expands networking expertise requirements to include machine-learning and broader artificial-intelligence skills.
- Progress in ML depends heavily on standardized datasets for developing and benchmarking AI algorithms.
- The paper advocates public training sets from experimental networks for computer-network AI research and development.
VI. CONCLUSIONS
KDN combines SDN, Network Analytics, and Machine Learning to provide automated network control, while its full vision still requires addressing important challenges through interdisciplinary research.
- KDN combines Software-Defined Networking, Network Analytics, and Machine Learning to provide automated network control.
- The paper presents use-cases and preliminary experimental evidence illustrating KDN’s feasibility and advantages.
- Achieving KDN’s complete vision requires addressing important challenges through collaboration across Artificial Intelligence, Network Science, and Computer Networks.