Content Distribution Networks in the mobile age

Abstract

Mobile applications and websites rely significantly on third party service providers to, among others, accelerate performance and reduce costs. An interesting recent deployment strategy is for content and service providers to leverage the services of multiple CDN providers in order to ensure maximal coverage and uptime. They may further use a managed DNS service, to ensure optimal server selection and load balancing among several CDNs. On top of that, peering agreements may allow some CDNs to deploy their infrastructure inside an ISP, giving them an advantage when serving this ISP’s customers. All of the above create a complex ecosystem, which affects the increasingly popular segment of the mobile apps market, but which thus far has not been studied in-depth. Our main goal is to empirically evaluate, through an active measurement campaign, the robustness and deployment strategies of various CDN providers, as well as their interconnection with ISPs in a mobile scenario. Past measurements have shown that CDN performance is impacted by ISP configuration [1,2]. We currently attempt to dig deeper and investigate questions such as: 1) whether the replica selection strategies of the various players have any inefficiencies, 2) how the choice of mobile technology affects performance and 3) how user mobility affects performance.

In order to answer the above, we first have to identify CDN related addresses that are popular among mobile apps. We make use of the dataset created by Lumen, a crowdsourced mobile app that intercepts and analyzes mobile traffic in user space with real user- and network- stimuli, to identify popular URLs. Then, we use heuristics and third party databases to isolate the CDN related URLs. Each URL is then used as a target of the active measurement campaign which is performed over the MONROE testbed. MONROE is a distributed mobile testbed with hundred of fixed and mobile nodes present in four major European markets (Sweden, Norway, Spain, Italy) and multihomed to up to three mobile operators. It allows for large scale experiments, while being free from the technical and ethical limitations of the current most popular alternative - crowdsourced approaches. For each test URL we periodically perform a series of basic tests, which include traceroutes, pings and attempting to fetch small objects. These are enough to characterize most aspects of the performance of a CDN. Moreover, we can assess mobile specific use cases more closely. For example, mobility scenarios in the city (nodes located in local buses) and country level (nodes located in intercity trains), roaming scenarios (SIM cards of one operator located in the home network and abroad) and study the effect of different technologies (HSPA, LTE, etc.) on CDN performance. Another interesting aspect is how the selection of replicas varies over time in static nodes.

[1] Safari Khatouni, Ali, et al. "Speedtest-like Measurements in 3G/4G Networks: the MONROE Experience." (ITC 2017). [2] Alay, Özgü, et al. “An Open Platform for Experimentation with Commercial Mobile Broadband Networks” (MOBICOM 2017)