We examine the ways in which congestion control schemes vary across different real network scenarios, and quantify the complexity of variations. Using measurement data from the Pantheon of Congestion Control, we measure the complexity of variation shown in congestion control scheme performance between different network scenarios. We show that certain network scenario relationships (such as, the difference between performance on one network path versus another) can be easily accounted for using simple, 2D affine transformations in throughput/delay space. This modeling ability gives credence to machine-learning approaches to congestion control, which rely on their being a consistent model that dictates congestion control performance. On the other hand, relationships such as that between an emulator and the path it is emulating cannot be modeled linearly, suggesting they are either random or very nonlinear, and showing a non-natural aspect of emulators which is not captured by previous evaluation methods..
Reproducing Network Research 