Robust Determination of WiFi Throughput Tests Being Indicative of Broadband Bottlenecks

Measurement of network speed, specifically bandwidth, has long been used as a key performance indicator for home broadband. Not only has it become a tool for detecting and diagnosing poor performance, but also for making investment decisions and measuring the quality of experience. However, current tools employ traditional techniques that consider wired measurements as the most accurate. Unfortunately, home users rarely have the capability to conduct reliable wired tests, instead being only able to measure using Wi-Fi. In particular, home wireless is often viewed as an unreliable indicator of network speed, leaving home users with little recourse to challenge the quality of broadband speed that is actually delivered.

In this thesis, we investigate the extent to which Wi-Fi-based tests are actually unreliable, and more importantly, to understand if one can accurately determine if the result was indicative of broadband as a bottleneck or if the measurement was limited by Wi-Fi. We also examine whether the accuracy of the tool is determined by the congestion control algorithm (CCA) and robust against specific use cases.

The results demonstrate that such a determination is eminently possible regardless of the CCA, and that it can be done drawing only on the feature and groups of features already reported by iPerf. We show through extensive experiments that goodness (the test was indicative of broadband speeds) or badness (the test was not `indicative of broadband speeds) can be captured with a precision of 92.4%, drawing only the median throughput and interquartile range with second-by-second windowing reported by iPerf. Finally, we illustrate that the classifier is robust against cross-traffic.