TLDR: This research explores how AI, specifically Machine Learning and Deep Neural Networks, can be used to predict network latency and ensure Service-Level Agreements (SLAs) for customized network services. Unlike previous studies, this paper validates its approach on real, large-scale national testbeds (FIBRE-NG and Fabric), using a distributed database application (Cassandra). The findings demonstrate that AI models can accurately forecast performance by analyzing generic network and computing metrics, proving their effectiveness for intelligent, self-managing network orchestration in production environments.
Modern digital services demand networks that are not only fast and reliable but also intelligent and adaptable. This is where ‘network slicing’ comes into play, allowing different parts of a network to be customized for specific user needs, much like creating dedicated lanes on a digital highway. However, ensuring these customized network slices consistently meet their performance promises, known as Service-Level Agreements (SLAs), is a significant challenge, especially when using Artificial Intelligence (AI) to manage them.
Traditionally, many AI-driven network management solutions are tested in simulated environments or small-scale labs. While useful, these setups don’t fully capture the complexities and unpredictable nature of real-world, large-scale production networks. This research paper, titled AI-driven Orchestration at Scale: Estimating Service Metrics on National-Wide Testbeds, tackles this very issue by proposing and validating a method for AI-driven orchestration directly on national-wide testbeds.
The core of this study revolves around predicting network latency, a critical performance metric, using advanced AI techniques like Deep Neural Networks (DNNs) and basic Machine Learning (ML) algorithms. The goal is to embed these predictive capabilities into network orchestration architectures, enabling them to proactively manage and ensure SLA conformance.
The SFI2 Architecture and Data Collection
The researchers utilized the Slicing Future Internet Infrastructures (SFI2) architecture, a framework designed to manage AI-native network slices across diverse network environments. Within this architecture, a ‘Predictor Module’ is key. This module collects a wide array of data from the network, including:
- Application metrics: Details about the performance of the services running on the network slice (e.g., latency, operation rates from a distributed database like Cassandra).
- Infrastructure metrics: Information about the computational resources consumed by the network slice (e.g., CPU, RAM, and I/O usage of the host machines).
- Network metrics: Statistics about data flow, errors, and loss across the network interfaces.
This rich dataset, collected from real-time operations, is then used to train the AI models to predict the mean latency of read and write operations within the Cassandra database application.
Experiments on Nationwide Testbeds
To ensure the findings were relevant to real-world scenarios, the experiments were conducted on two large-scale, geographically distributed testbeds: FIBRE-NG and Fabric. These testbeds include nodes in various locations, even intercontinental ones like CERN. The researchers simulated realistic network conditions, including introducing controlled amounts of latency (jitter) and packet loss, to see how these factors impacted application performance and how well the AI models could predict it.
Key Findings and Model Performance
The study revealed several important insights:
- Network conditions significantly impact performance: Both network delay (jitter) and packet loss were found to have a substantial effect on the latency of database operations, highlighting the need for intelligent management in dynamic network environments.
- Basic ML models are effective: Algorithms like Random Forest (RF) and Decision Trees (DT) showed strong performance in predicting latency, often outperforming other basic ML models.
- Deep Neural Networks excel: Various DNN architectures, including ResNet, InceptionTimePlus, FCN, and TCN, demonstrated remarkable accuracy in forecasting latency. These models were particularly adept at learning patterns from the complex, time-series data generated in a live production network.
The low error rates achieved by both basic ML algorithms and DNNs suggest that these AI techniques can be effectively integrated into network slicing orchestrators. This integration would allow for ‘zero-touch orchestration,’ where the network can intelligently manage itself, proactively ensuring that services meet their agreed-upon performance levels without constant human intervention.
Also Read:
- Managing Wireless Networks with AI: A New Approach to RAN Automation
- Smarter XR: Predicting Head Movements for Better Human-Machine Interaction
Future Implications
This research provides a strong foundation for the future of AI-driven network management. By demonstrating the feasibility of training and deploying AI models on real, large-scale networks, it paves the way for more intelligent, self-managing network infrastructures that can adapt to diverse application requirements and dynamic network conditions. The findings suggest that monitoring easily collected, generic network and computing metrics is sufficient for accurate SLA forecasting, making this approach highly practical for widespread adoption.
