On using crowd-sourced network measurements for performance prediction

Geo-location-based bandwidth prediction together with careful download scheduling for mobile clients can be used to minimize download times, reduce energy usage, and improve streaming performance. Although crowd-sourced measurements provide an important prediction tool, little is known about the prediction accuracy and improvements such datasets can provide. In this paper we use a large-scale crowd-sourced dataset from Bredbandskollen, Sweden's primary speedtest service, to evaluate the prediction accuracy and achievable performance improvements with such data. We first present a scalable performance map methodology that allows fast insertion/retrieval of geo-sparse measurements, and use this methodology to characterize the Bredbandskollen usage. Second, we analyze the bandwidth variations and predictability of the download speeds observed within and across different locations, when accounting for various factors. Third, we evaluate the relative performance improvements achievable by users leveraging different subsets of measurements (capturing effects of limited sharing or filtering based on operator, network technology, or both) when predicting opportune locations to perform downloads. Our results are encouraging for both centralized and peer-to-peer performance map solutions. For example, most measurements are done in locations with many measurements and good prediction accuracy, and further improvements are possible through filtering (e.g., based on operator and technology) or limited information sharing.