• 5G
• Edge Environments
• Foundation Models
• Network Function Virtualization (NFV)
• Time-Series Forecasting
• Virtualized Network Functions (VNFs)

The advent of 5G has revolutionized network architectures, driving the adoption of software-driven solutions to enhance flexibility, scalability, and cost-effectiveness. A key aspect of this transformation is “network softwarization”, which replaces rigid, hardware-based infrastructures with agile, software-defined architectures. This shift enables the deployment of Network Function Virtualization (NFV), where dedicated hardware is substituted with Virtualized Network Functions (VNFs) running on shared cloud and edge resources. By leveraging virtualization and software-based control, NFV enhances network adaptability and optimizes resource allocation.
However, as the scale of infrastructure and deployment of VNFs continues to grow, the challenges related to the efficient placement and workload prediction of VNFs become increasingly critical. These challenges require advanced forecasting techniques that are capable of predicting future demands to optimize resource allocation.
This thesis explores the potential of utilizing innovative, large-scale network technologies to predict a diverse set of workload metrics beyond traditional CPU usage, extending to transmitted bytes, received bytes, and memory usage, with the aim of improving NFV efficiency in edge environments. Using SNDZoo, an open benchmarking framework for software-defined networking and NFV research, we generated and collected monitoring data from virtual machines hosting VNFs. Additionally, we leveraged a real-world dataset from Vodafone, which encompasses CPU demand data from NFV infrastructures across multiple European countries.
To predict workload fluctuations, we employed two foundation models for time-series forecasting, TimesFM and Lag-LLAMA, demonstrating strong zero-shot capabilities. Both models were fine-tuned to further enhance predictive accuracy. Performance was evaluated using Mean Absolute Percentage Error (MAPE) and Continuous Ranked Probability Score (CRPS). Our findings reveal that CPU usage predictions achieved strong performance in both synthetic and real-world settings (SNDZoo and Vodafone). Moreover, incorporating network traffic metrics provides a more holistic and complete workload forecasting approach. The excellent predictive performance of these models underscores the potential of foundation models to improve NFV workload prediction and support more adaptive network management strategies.