It’s important that research be accessible and reproducible so that results and evaluations can be verified. As part of Stanford’s graduate course on computer networks, students have been reproducing parts of network systems research papers for several years on the public Reproducing Network Research blog.
In order to contribute to this effort, my partner Jervis and I have recently recreated Figure 8 of the original BBR paper, and exploring the effects of several experimental parameters. We’ve posted our comments on Stanford’s Blog, along with an accompanying GitHub repository that contains the code and instructions used to recreate our results.
We found that BBR does indeed achieve higher throughputs than CUBIC in lossy networks, and show that this behavior holds true across several orders of magnitude of bottleneck bandwidths, a variety of routh trip times, and also on an LTE cellular link trace. We also observe that BBR performs better than several other TCP congestion control variants like vegas, westwood, bic, and reno.
Follow the discussion on our BBR development thread!