Reducing Consumed Data Volume in Bandwidth Measurements via a Machine Learning Approach

Abstract

Measurements that determine the available download and upload bandwidth of an end-user Internet connection (so-called speed tests) are typically performed by maximizing the utilization of the connection for a fixed time interval. Especially in broadband connections, such tests consume a huge amount of data volume during their execution. As a result, only a few tests can be performed per month on mobile connections with limited data volumes, since otherwise a significant portion of the volume is used for tests or additional costs are incurred. To reduce the required average data volume of these tests, we present a novel approach with a dynamic test duration based on a machine learning model. We train this model via a supervised learning process, using the recorded data of real speed tests executed by end-users in cellular 4G networks. The evaluation of the resulting method suggests that the amount of saved data volume is significant, while the deviation of the determined bandwidth (compared to a usual test with fixed duration) is negligible.