Towards Understanding City-Level Routing Using BGP Location Communities

Key Takeaways

• Inferred city-level locations for 1,482 of 1,595 BGP communities (93% recall)
• 80% of inferred locations lie within 70 km of ground truth
• Only ~4% of routes tag locations near their origin (≤50 km)
• Distant tagging observed: Russian routes tagged via Oslo, Israeli via Marseille
• Method uses passive RouteViews/RIPE data and MaxMind geolocation clustering

Pulse Analysis

Understanding where traffic actually traverses the Internet has long been a blind spot for network engineers. While the Border Gateway Protocol (BGP) exposes autonomous system (AS) paths, it rarely reveals the physical points of presence that shape latency and resilience. BGP location communities—numeric tags attached to routes—promise city‑level granularity, yet most remain undocumented, leaving operators without a reliable map of inter‑city peering. This knowledge gap hampers capacity planning, regulatory compliance, and the detection of routing anomalies that can affect end‑user experience.

The APNIC‑CAIDA team tackled this challenge by exploiting a spatial correlation between a prefix’s origin and the router that appends a location community. Using passive feeds from RouteViews and RIPE RIS, they mapped prefixes to coordinates via the MaxMind database, then clustered tags in latitude‑longitude space. The approach filtered out long‑distance tags and focused on clusters that tightly surround a city, enabling inference of 1,482 communities with 93% recall. Impressively, 80% of these inferred locations fell within a 70‑kilometer radius of verified ground truth, proving that passive data alone can surface hidden geographic signals.

The implications extend beyond academic curiosity. Operators can now approximate where peers interconnect without relying on proprietary documentation, aiding latency‑sensitive services such as video streaming and financial trading. Researchers gain a new lens for studying routing policies, undersea cable utilization, and regional resilience. Although the method still struggles with ASes lacking local presence—producing distant tagging anomalies—it provides a foundation for automated classification of undocumented BGP communities. Future work will refine outlier detection and integrate real‑time feeds, moving toward actionable, city‑level routing intelligence for the broader Internet ecosystem.