In order to assess the performance of domains hosted in Cloud Service Providers (CSP), we run a dedicated measurements campaign on the MONROE measurements platform.
For this study we use nodes in 4 countries and the Mobile Network Operators (MNO) over which we measure are:
Norway: Telenor, Telia, Telia (SE)
Italy: Vodafone, Wind, TIM
Spain: Yoigo, Orange, Vodafone (IT)
Sweden: Telia, Telenor, 3
* MNOs written in bold letters are SIM cards that are roaming internationally (home country code in brackets).
We benefit from MONROE’s openness and capabilities to run long-lived active measurements in realistic but controlled scenarios in commercial MNOs.
DNS test: We run DNS lookups for each target mobile domain towards: 1) the default MNO DNS resolver, 2) Google's public resolver and 3) OpenDNS's public resolver. This will allow us to compare the quality of the responses and identify possible DNS-level inefficiencies during the replica selection. All our DNS queries include the EDNS edns-client-subnet flag as it may be used by CSPs like EdgeCast and Amazon services to locate the end user and improve the quality of replica selection. If we do not receive a reply, as it occurs for Vodafone (IT) possibly due to the presence of an in-path DNS proxy filtering those requests, we repeat the request without the flag.
Traffic performance test: We measure the TCP connection and TLS session establishment time, if applicable, towards the resolved IP addresses. We repeat the same experiment over TCP ports 80 and 443, in case the MNO performs traffic shaping or deploys TCP-splitting proxies. The TCP connection, as well as the TLS session establishment are triggered by an HTTP GET request for the favicon.ico object. The presence of the object in the server is not relevant as the handshakes are triggered regardless of its existence.
Network topology test: For each resolved IP address, we run UDP traceroutes to study CSP-MNO peering and topological relationships. We also run pings towards the same IP address.
We also log all the related metadata that are provided by the platform.
The container code used on the MONROE nodes to generate this dataset can be found at:
We run the aforementioned experiments in isolation from other experiments continuously over 4 weeks, from April 5, 2017 until May 6, 2017.
The combined results of the three tests for a given FQDN produce a "sample". The measurements are continuously run and the periodicity of the same sample varies between 4 and 24 hours. Thus, each sample has several instances across time of day and various radio conditions.