Abstract :
[en] Service providers rely on replication to improve service performance and reliability, placing instances in multiple locations and redirecting clients to nearby copies. Anycast is a common mechanism used for redirecting clients in a variety of domains from naming to CDNs and video streaming. IP anycast offers a method for making a service IP address available to a routing system from several locations at once, and clients' requests are directed based on BGP routing policies. For operators, IP anycast offers an economic, scalable, and simple approach to replicated services; BGP provides considerable robustness, adapting to changes in service and network availability. For clients, however, the mapping can be suboptimal, unstable, and seemingly chaotic, as routing policies have not only technical motivations, and routing changes can silently shift traffic from one site to another with a consequent loss of shared state and potential performance impact. Given its wide deployment and interesting tradeoffs, IP anycast has been the focus of much recent measurement work. All prior studies have, nevertheless, focused on wired networks despite the growing dominance of mobile Internet. Today, the number of mobile subscriptions is over 7.4 billion, and users spend over 2x times more hours browsing on their smartphones than on any other device, with the corresponding increase on cellular traffic. We present early results on the first study of anycast performance for mobile users. Our evaluation focuses on two distinct anycast services, K-and F-Root, each providing part the DNS Root zone. Both services are widely replicated with publicly available site locations and unicast IP addresses that allow us to evaluate the relative performance of anycast routing to its "optimal" (in terms of unicast) site location. We collected active measurements from geographically distributed clients on both cellular and WiFi networks from September 2016 until April 2017, using the Aqualab’s ALICE engine [1]. In each experiment clients launched ping and traceroute measurements towards the root servers’ anycast addresses, as well as to five chosen unicast addresses determined to be the closest to the client in terms of geographic distance, at an hourly rate. Clients also recorded their geographic location, anonymized to a 10 km² area. Our findings show that mobile clients are routed to suboptimal replicas in terms of geographical distance, more frequently while on a cellular connection than on WiFi, with a significant impact on perceived service performance. The phenomenon seems to be more pronounced for K-Root than for F-Root. A possible explanation for the long distances would be that our cellular clients are simply far away from all the available replicas. However, our investigations demonstrate that this is not necessarily the case. Finally, we start to explore the root causes for anycast anomalies in cellular networks. We reveal three classes of anomalies: distant client packet gateways, poor anycast routing within Tier-1 networks, and improper routing out of cellular networks. [1] http://aqualab.cs.northwestern.edu/projects/261-alice
