The impact of duration and settings of TCP measurements on available bandwidth estimation in mobile networks

Today, crowdsourcing applications distribute the task of performance measurements in mobile networks over various mobile users to achieve an accurate benchmark. A common measurement methodology is an application on the customers' phones which injects probe packets into the network under test and derives estimates for selected network properties such as available bandwidth. Bandwidth measurements in cellular wireless networks, which are highly dynamic and offer only shared resources, are in general a resource demanding task. Current methods deploy tests of constant duration running data transfers over TCP. Their resource consumption scales directly with the available capacity of the link. In this work we take a closer look at the different chosen test durations and analyze the trade-off between accuracy and resources spent in various different configurations, e.g., numbers of parallel TCP connections, based on extensive measurements in real world operational LTE and UMTS networks conducted in 2016, Austria. In our experiments a minimum measurement duration of six seconds allows for estimating available bandwidth with an average error of less than 1%.