The Good, the Bad, and the Ugly of BGP Routing Modeling: Confounding Factors, Selective Announcements and Location-aware Simulations

Kastanakis, Savvas and Giotsas, Vasileios and Suri, Neeraj (2025) The Good, the Bad, and the Ugly of BGP Routing Modeling: Confounding Factors, Selective Announcements and Location-aware Simulations. PhD thesis, Lancaster University.

Text (2025kastanakisphd)
2025kastanakisphd.pdf - Published Version
Available under License Creative Commons Attribution.
Download (1MB)

Abstract

Over the past quarter-century, substantial research has been devoted to the accurate inference/prediction of AS-level paths on the Internet. Literature, though, has shown that BGP simulations can be highly inaccurate making it hard to take ”in-vitro” evaluations at face value. This PhD thesis investigates the complexities of this long standing inference problem by exploring various confounding factors, analyzing the evolution of interdomain routing policies, and examining the potential of geolocation-aware inference models.  As a first step, we identify and quantify the confounding factors to accurate AS-path inference through passive measurements and controlled experiments. From our analysis, two factors are highlighted: (a) the complex routing policies implemented by ASes, and (b) the geolocation agnostic BGP best path selection process.  To delve deeper into the first confounding factor, we explore the evolution of the interdomain routing policies over the last 20 years. We show that while the phenomenon of selective announcements is persistent, not only across time, but also across networks, the influence of AS relationships on path selection has diminished. This underscores the importance of frequent BGP policy inference to keep pace with the evolving landscape of AS connectivity.  As we investigate the second confounding factor, we explore the role of geospatial attributes in geolocation-based routing paradigms such as IP anycast routing. Our analysis shows that 84.06% of anycast ASes rely on selective announcements, a phenomenon largely influenced by the geolocation-agnostic nature of the BGP best path selection process. Additionally, we demonstrate that anycast ASes follow different routing patterns per geographical regions.  Our work supports the need for more flexible routing models and can aid in the understanding of a variety of interdomain routing applications, such as the measurement of the RPKI adoption, fine-grained interdomain policy learning, interdomain routing verification, privacy-preserving routing and studying routing attacks.