Towards understanding city-level routing using BGP location communities

To understand Internet routing, it is crucial to go beyond the Autonomous System (AS) level to city-level routing, as it reveals the underlying physical locations. Border Gateway Protocol (BGP) location communities are the only routing data that provide this level of detail, by signalling where ASes peer with each other.

However, most BGP communities are not standardized or documented, leaving researchers and operators no reliable way to interpret them. We propose a method to infer their meaning.

This post presents highlights of our inference method, which exploits the spatial correlation between a network prefix's origin and the location of the router that attaches a location community, along with selected inference results.

You can find a more comprehensive exploration of the topic in our paper.

Direct inference is complicated

BGP communities are opaque values, so the locations they signal cannot be inferred directly from the community values themselves. For example, the communities and are defined by two different ASes (AS35280 and AS1299), yet both indicate the same location: Singapore. In our work, we rely on the origin of prefixes that are tagged with these communities.

However, using ground-truth location communities, we found that only about 4% of routes are tagged close (≤50 km) to their origin, while most are tagged far from their origin (>50 km, up to 20,000 km), complicating direct inference.

To address this, we developed filters to isolate prefixes tagged close to their origin for location inference, and additional mechanisms to improve inference accuracy.

Prefixes may be tagged at handoff

We infer the locations signalled in BGP communities using prefixes, AS-PATHs, and community attributes from RouteViews and RIPE RIS, mapping prefixes to geographic coordinates with the MaxMind geolocation database.

Prefixes tagged with the same community are clustered in two-dimensional geographic space (latitude and longitude) to identify spatial concentration corresponding to a physical location. The method builds on the intuition that prefixes may be tagged at handoff locations close to their destination, for example, when destinations are reachable through a single location or under cold-potato routing when multiple exits are available.

This is the first method to infer the geographic semantics of BGP location communities at city-level granularity using only passive routing data.

A measurable spatial signal

We successfully infer locations for 1,482 out of 1,595 communities (recall 93%) in our validation dataset. For 80% of these, the inferred location is within 70km of the ground truth.

Figure 1 illustrates the operation of our inference method. We randomly selected ten city-level communities defined by AS2914. Circles denote prefix clusters, with sizes proportional to the number of prefixes, normalized per community (colour).

We make three observations:

1. For each community, the largest cluster correctly identifies the city signalled by that community, demonstrating the accuracy of our approach.
2. We observe per-economy routing catchments, meaning that clusters originating in a given economy are learned through the corresponding tagging router located in that economy, as seen for Madrid, Milan, Sofia, Vienna, and Warsaw.
3. Some tagging routers learn clusters from distant regions. For example, clusters from Russia are learned via Oslo, while clusters from Israel are learned via Marseille. AS2914 has no local presence in those regions, and connectivity is likely provided through undersea cables or remote peering.

The results confirm that there is a measurable spatial signal between the origin of prefixes and the router that tags them with location communities, making passive inference of undocumented communities feasible in practice.

The method can perform poorly for communities with clusters from regions where the tagging AS has no local point of presence, causing tagging locations to be decoupled from prefix origins. Nevertheless, identifying these outliers has merit in its own right and invites further analysis.

Validation and scale

We validated our results using a ground-truth dataset of documented location communities drawn from Tier-1, content, eyeball, and carrier networks. For undocumented communities, we contacted network operators directly and observed similar inference accuracy.

While not yet operationally applicable, this work suggests that city-level interconnection information can be inferred at scale, opening the door to broader applications.

A deeper understanding of interconnection behaviour

This work was conducted in collaboration with the Center for Applied Internet Data Analysis (CAIDA) at UC San Diego, and the University of Liège. The datasets and code used in this work are publicly available on GitHub, and our paper has been published in the Proceedings of the Association for Computing Machinery on Networking, Volume 3.

While we developed a method to infer the locations signalled by communities, reliably identifying which communities actually encode city-level information remains an open problem. As part of a broader effort to understand undocumented BGP communities, we are developing methods to automatically identify and classify them.

Our observation of a spatial correlation between tagging routers and prefix origins can be leveraged to study geographic routing patterns, detect anomalies, and deepen our understanding of interconnection behaviour beyond the AS level.

Thomas Krenc is a researcher at Internet Initiative Japan (IIJ) Research Laboratory.

The views expressed by the authors of this blog are their own and do not necessarily reflect the views of APNIC. Please note a Code of Conduct applies to this blog.