Keep your Communities Clean: Exploring the Routing Message Impact of BGP Communities

TL;DR

This paper investigates how BGP communities, used for policy and traffic management, unintentionally cause increased routing update messages due to their transitive nature, impacting network efficiency and stability.

Contribution

It characterizes the handling of community-related updates across real-world BGP implementations and quantifies their impact on message proliferation over a decade.

Findings

25% of announcements modify communities without changing AS path in 2020

Communities cause increased update messages at tagging and neighboring ASes

63% of community attributes are revealed only through global withdrawals on a single day

Abstract

BGP communities are widely used to tag prefix aggregates for policy, traffic engineering, and inter-AS signaling. Because individual ASes define their own community semantics, many ASes blindly propagate communities they do not recognize. Prior research has shown the potential security vulnerabilities when communities are not filtered. This work sheds light on a second unintended side-effect of communities and permissive propagation: an increase in unnecessary BGP routing messages. Due to its transitive property, a change in the community attribute induces update messages throughout established routes, just updating communities. We ground our work by characterizing the handling of updates with communities, including when filtered, on multiple real-world BGP implementations in controlled laboratory experiments. We then examine 10 years of BGP messages observed in the wild at two route collector systems. In 2020, approximately 25% of all announcements modify the community attribute, but retain the AS path of the most recent announcement; an additional 25% update neither community nor AS path. Using predictable beacon prefixes, we demonstrate that communities lead to an increase in update messages both at the tagging AS and at neighboring ASes that neither add nor filter communities. This effect is prominent for geolocation communities during path exploration: on a single day, 63% of all unique community attributes are revealed exclusively due to global withdrawals.