Virne: A Comprehensive Benchmark for RL-based Network Resource Allocation in NFV

TL;DR

Virne tackles the lack of standardized evaluation for RL-based NFV-RA by delivering a comprehensive, gym-style benchmarking framework that supports cloud, edge, and 5G scenarios. It unifies problem definitions, provides a modular RL-based implementation pipeline, and offers extensive evaluation criteria including solvability, generalization, and scalability. Through extensive experiments, it demonstrates that dual-graph neural network policies (e.g., PPO-DualGAT, PPO-DualGCN) consistently achieve superior solution quality and robustness, while highlighting the practical trade-offs with computation time and topology density. The framework also explores emerging network requirements such as heterogeneous resources and latency constraints, delivering actionable guidance for future RL-based NFV-RA research and deployment. Overall, Virne enables reproducible, data-driven progress in ML-enabled network resource optimization across diverse network paradigms.

Abstract

Resource allocation (RA) is critical to efficient service deployment in Network Function Virtualization (NFV), a transformative networking paradigm. Recently, deep Reinforcement Learning (RL)-based methods have been showing promising potential to address this complexity. However, the lack of a systematic benchmarking framework and thorough analysis hinders the exploration of emerging networks and the development of more robust algorithms while causing inconsistent evaluation. In this paper, we introduce Virne, a comprehensive benchmarking framework for the NFV-RA problem, with a focus on supporting deep RL-based methods. Virne provides customizable simulations for diverse network scenarios, including cloud, edge, and 5G environments. It also features a modular and extensible implementation pipeline that supports over 30 methods of various types, and includes practical evaluation perspectives beyond effectiveness, such as scalability, generalization, and scalability. Furthermore, we conduct in-depth analysis through extensive experiments to provide valuable insights into performance trade-offs for efficient implementation and offer actionable guidance for future research directions. Overall, with its diverse simulations, rich implementations, and extensive evaluation capabilities, Virne could serve as a comprehensive benchmark for advancing NFV-RA methods and deep RL applications. The code is publicly available at https://github.com/GeminiLight/virne.

Figures (13)

• Figure 1: A brief illustration of the NFV-RA problem.
• Figure 2: The architecture of Virne benchmark. Virne offers a streamlined workflow for supporting comprehensive experimentation of NFV-RA algorithms. The process begins with customizing simulation configurations that define network scenarios and conditions. Then, the network system is instantiated, triggering a series of service request events to process. At each event, the selected NFV-RA algorithm interacts with the system to resolve the instance. For each simulation, both the processing and final results are automatically recorded for subsequent analysis.
• Figure 3: A unified pipeline of gym-style environment and RL-based NFV-RA methods in Virne.
• Figure 4: Training curves of PPO-DualGAT and its variations. Each point represents the R2C ratio over all VN Requests within a single training simulation round.
• Figure 5: Results on the solvability study. For each size of VN, we highlight the worst-performing solver in green and the best-performing solver in red.