Workshop: 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS
Authors: Hadi Sahin and Jeronimo Bezerra (Florida International University); Renata frez (Rede Nacional de Ensino e Pesquisa, Rio de Janeiro); and Italo Brito, Vasilka Chergarova, Luis Marin Vera, Gretel De la Peña, and Julio Ibarra (Florida International University)
Abstract: Bidirectional Forwarding Detection (BFD) and Border Gateway Protocol (BGP) flaps are key indicators of network instability. Frequent up/down transitions can increase latency, cause packet loss, and lead to Service Level Agreement (SLA) violations by repeatedly disrupting traffic forwarding and routing convergence. Root causes of the flaps can range from physical link failures to protocol misconfigurations and policy violations, making diagnosis costly and time-consuming, as operators must correlate evidence across devices, interfaces, protocols, logs, and performance metrics. Large Language Models (LLMs) offer a promising alternative, as they can integrate and reason across heterogeneous operational evidence to accelerate root cause identification and reduce mean time to resolution (MTTR). However, their effectiveness depends on explicit diagnostic reasoning, structured explanations, and validation of inferred causes against operational evidence. API-based LLMs also raise privacy concerns, while locally deployed open source alternatives require careful evaluation of their diagnostic accuracy and operational performance. In this paper, we present an LLM-based prototype for automated root cause analysis of BGP and BFD flap events. The prototype uses locally deployed open-source LLMs to correlate protocol logs with topology, service state, and forwarding information, identifying likely root causes and affected network entities. We evaluate the approach through controlled failure experiments in a Mininet-based testbed and through operational analysis of events from a production BGP/BFD testbed. The results show that prompt design substantially affects root-cause accuracy, with structured, domain-specific prompts enabling several medium-sized models to achieve perfect accuracy. Entity-extraction performance varies more widely across prompts and models, while model selection also affects inference latency. These findings suggest that locally deployed LLMs can support accurate, cost-effective, and privacy-preserving analysis of BGP and BFD flap events.
Back to 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS Archive Listing
Back to Full Workshop Archive Listing