Virne: A Comprehensive Benchmark for RL-based Network Resource Allocation in NFV
Tianfu Wang, Liwei Deng, Xi Chen, Junyang Wang, Huiguo He, Zhengyu Hu, Wei Wu, Leilei Ding, Qilin Fan, Hui Xiong

TL;DR
Virne is a comprehensive benchmarking framework designed for evaluating deep reinforcement learning methods in network resource allocation within NFV, supporting diverse scenarios and providing extensive analysis tools.
Contribution
Introduces Virne, a modular, extensible benchmarking framework for NFV-RA that supports over 30 methods and offers diverse evaluation perspectives.
Findings
Insights into performance trade-offs for different algorithms
Evaluation of scalability and generalization of RL-based methods
Guidance for future research directions in NFV-RA
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…
Peer Reviews
Decision·ICLR 2026 Poster
a) The Virne framework supports a wide variety of network scenarios and resource types, including heterogeneous resources and latency-sensitive environments. b) The authors conduct experiments across diverse topologies (e.g., Waxman, GEANT, BRAIN) and network scales, providing valuable insights into the strengths and weaknesses of different algorithms.
a) The authors use a fixed 50% random probability for interconnections between virtual nodes. A fixed interconnection probability may not accurately reflect real-world VN topologies, which often exhibit more structured (sequential or hierarchical) connectivity patterns. To better align with real-world NFV scenarios, the authors should consider adopting more dynamic and realistic VN topology generation methods (e.g., using real-world datasets). b) The computing and bandwidth resource configurati
The paper decomposes the implementation of RL into three parts, the RL algorithm, the neural network architecture, and the implementation techniques, and provides multiple design options for each component to investigate their underlying design principles, offering valuable insights into the application of RL in NFV-RA. The benchmarking framework proposed in the paper provides a comprehensive platform that includes diverse NFV application scenarios, implementations of over 30 RA algorithms, and
Considering that this is a unified benchmarking framework, assuming a one-to-one mapping from virtual networks to physical networks in the system model seems unreasonable. In general, multiple virtual nodes should be allowed to map to the same physical node to enable flexible deployment. There is a lack of research on emerging network architectures, such as Transformers and diffusion models. The paper repeatedly selects virtual nodes from the virtual network and maps them onto the physical net
+ VIRNE represents the most systematic and extensive benchmark for NFV-RA to date. It consolidates numerous algorithms, diverse simulation environments, and multiple evaluation perspectives into a unified, open, and reproducible framework. This kind of infrastructure contribution fills a long-standing need in the NFV and RL communities. + The modular design, well-documented implementation, and thorough experimental setup reflect a high level of technical maturity. The open-source release with de
- The paper’s contribution is primarily infrastructural rather than algorithmic. The RL formulations (e.g., MDP setup, PPO training, GNN encoders) follow standard designs without introducing new theoretical insights or model innovations. - While experiments are comprehensive, most results are reported as single averages without statistical significance tests or error margins. This weakens the strength of empirical claims. - Several sections (notably Sections 3–4) emphasize implementation details
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsSoftware-Defined Networks and 5G · IoT and Edge/Fog Computing · Cloud Computing and Resource Management
