Automation in Network Function Virtualization (NFV) management has become increasingly vital as networks grow in scale and complexity. Traditional NFV management, which relies on predefined policies and manual interventions, struggles to keep up with ...
Automation in Network Function Virtualization (NFV) management has become increasingly vital as networks grow in scale and complexity. Traditional NFV management, which relies on predefined policies and manual interventions, struggles to keep up with dynamic network conditions, often resulting in slow and inefficient responses to failures. Moreover, the growing diversity of NFV deployments further limits the flexibility and scalability of existing solutions. Considering these limitations, we explore Large Language Model (LLM) agent approaches by harnessing the reasoning and decision-making capabilities of LLMs, and present NFVAgent, an LLM-driven NFV recovery framework.
In NFVAgent, the LLM directly analyzes operation issues and derives appropriate recovery actions, thereby enhancing resilience under dynamic network conditions. While this LLM agent appears structurally feasible, its practical effectiveness is limited by the fact that LLMs lack prior training on diverse NFV environments. To overcome this, NFVAgent is built on Retrieval-augmented Generation (RAG) and advances autonomous recovery by incorporating a continuously evolving NFV knowledge base, updated through interactions with a separate testbed environment. This structured knowledge integration of NFVAgent enables the LLM to dynamically adopt relevant domain-specific knowledge and refine its contextual understanding on NFV failures and potential recovery strategies.
We evaluated NFVAgent in 500 failure cases across four NFV environments, specifically configured with several industry standards such as Open Network Automation Platform (ONAP), European Telecommunications Standards Institute (ETSI), and Open Radio Access Network (O-RAN), demonstrating its effectiveness in enhancing NFV resilience across diverse network conditions and failure events. NFVAgent achieves up to 99.8% pass rate, outperforming policy-based methods with an average improvement of 37.7% in multi-tenant and 16% in single-tenant environments.