Understanding the performance of broadband networks through the statistical analysis of speed tests

Abstract

In this thesis we performed the statistical analysis of a dataset containing raw data captured through user-initiated Internet speed tests. The NDT dataset represents the largest publicly available source of raw measurements ever collected on the Internet. Using the statistical package R, we analyzed the dataset and quantified the impact that several factors have on speed measurements. Those factors include the distance between client and server (roundtrip time), the receive window, buffer sizes along the path, the effect that users who run the test multiple times have on aggregate variables, and the effect of test durations in the outcome. Further, we used the information to obtain a clearer picture on congestion, an important phenomenon which is not widely understood, answering questions such as the proportion of tests that encounter congestion, how many times they encountered congestion and what sort of congestion it was. Thanks to the analysis of the data, we have been able to reflect on the current landscape of broadband Internet in the United States, characterizing the service offered by the top six Internet Service Providers from different angles. The current thesis describes a procedure that allows us to analyze the network capacity, by excluding from the analysis tests that are limited by causes alien to the network itself, such as limitations on the receive window imposed by the operating system. Finally, some recommendations are offered that we believe would make the NDT test a more useful tool to gather information about the performance of broadband networks, and we come up with policy recommendations for the different stakeholders in the broadband arena.