Comparing methods to find and map anycast server locations
Locating and Enumerating Anycast: a Comparison of Two Approaches
Networking and Internet Architecture
Summary
Anycast lets internet services run from many places using one address, but it's hard to tell where these places are. The authors compare two ways to find and map anycast sites: one uses delays in network response times, and the other traces the route packets take. They find that tracing routes is a bit more precise for counting and locating sites but costs more in terms of network checks and sometimes guesses too many sites. This work helps in understanding the strengths and weaknesses of methods used to map important internet infrastructure.
What this means in practice
- •For network operators: Improve strategies for identifying and locating anycast servers to enhance service resilience and maintenance.
- •For internet infrastructure analysts: Use traceroute-based mapping to gain slightly more accurate geolocation data about anycast PoPs despite higher probing costs.
Authors
Remi Hendriks, Tim Betzer, Ben Du, Raffaele Sommese, Mattijs Jonker, Roland van Rijswijk-Deij
Abstract
Anycast allows for providing services from multiple, geographically distant Points of Presence (PoPs), using a single IP address, to, e.g., improve resilience. Due to its opaqueness, it is often unknown which addresses are provisioned using anycast and, if so, where the PoPs are located. As anycast is widely used for critical Internet infrastructures (e.g., the DNS) efforts have been made to map anycast deployments. The current state-of-the-art mapping technique, iGreedy, relies on latency-based measurements, and is adversely affected by noise caused by, e.g., network processing delays. Previous work has shown that traceroute can alternatively be used to detect anycast. As traceroute reveals the hops a packet traverses, it may also be used to locate sites using geolocation data for hops near the anycast PoPs. This paper is the first to assess the performance of the traceroute-based approach at scale, by targeting 14k prefixes from an anycast census. Using ground truth we show traceroute achieves a slight increase in enumeration and geolocation precision over iGreedy. However, it suffers from overestimating the number of PoPs and incurs a 4 times increase in probing cost, making it unattractive for anycast censuses.