This repository contains the artifact for the USENIX ATC '25 paper "Internet Connection Splitting: What's Old is New Again" by Gina Yuan, Thea Rossman, and Keith Winstein from Stanford University. The artifact consists of all the raw data from the paper, Jupyter notebooks for analyzing and plotting this data using the split throughput heuristic, and scripts for automating data collection for replicating the emulation experiments.
To get started with this artifact, we recommend exploring the Jupyter notebooks using the raw data. For further exploration, we then recommend running emulations and replicating some of the results of the emulation throughput experiments, or even evaluating your own congestion control schemes.
All experiments were run on CloudLab x86_64 rs630 nodes in the Massachusetts
cluster, using Ubuntu 22.04. Any similar Linux x86_64 machine would suffice.
Clone this repository into your home directory. Install dependencies, accepting
any prompts:
./deps/node0.sh
The instructions assume your working directory is ~/connection-splitting.
The Jupyter notebooks read and write raw data to ~/connection-splitting/data.
Initialize this directory with raw data:
tar xvf 2025-01-15-data.tar.gz
Follow the instructions in notebook/
to setup your Python kernel and access the notebook from your local machine.
Confirm that you are able to execute all (non-hidden) cells in the following notebooks to regenerate the figures and analysis in the paper:
parameter_exploration.ipynb- Figure 4: Characterize TCP congestion control schemes at 10 Mbit/s.
- Figure 5: Characterize QUIC congestion control schemes at 10 Mbit/s.
- Figures 8-12 in the Appendix: Heatmaps of the full parameter space.
network_path_analysis.ipynb- Table 3: Apply the heuristic to analyze TCP in theoretical split settings - FINDING 2.
- Figure 1 (heuristic only): Identify marquee network settings for Figure 1 - FINDING 1.
- Figure 6: Apply the heuristic to reason about QUIC in split settings - FINDING 3.
network_path_real_data.ipynb(optional)- Figure 1: Marquee network settings where TCP BBRv3 benefits significantly from a real connection-splitting PEP but TCP BBRv1 does not.
accuracy_analysis.ipynb(optional)- Figure 7: Heatmaps comparing measured and predicted split throughputs.
The network_path_analysis.ipynb notebook relies purely on cached data to apply
the split throughput heuristic. Feel free to analyze this data more yourself!
For further exploration, follow the detailed instructions below for how to replicate the emulation experiments.
Figure 6: The predicted behavior of each CCA based on end-to-end
behavior varies significantly by implementation.
The measurement study consists of two parts:
- Characterize various congestion control schemes in a wide parameter space of end-to-end network settings.
- Apply the split throughput heuristic to analyze these congestion control schemes in various split network settings.
The "Getting Started" instructions describe Part 2, and in this section we describe how to run the emulations for Part 1. The experiments also run emulations with a real TCP PEP to evaluate the accuracy of the heuristic.
These instructions assume you have followed the setup instructions in "Getting
Started" above. No more dependencies are needed unless you wish to evaluate TCP
BBRv2/3 (which require separate kernels) or Google QUIC (which uses up a lot
of disk space). However, instructions for installing these are available in
deps/.
Follow the "Example" instructions in emulation/
to a run a single emulation. The default network configuration has loss on the
path segment near the client who makes the HTTPS GET request.
First, delete or move the ~/connection-splitting/data
directory so the notebooks don't rely on the cached data. If a notebook cell
states that data is missing without running the emulation, change the
execute=False argument to True. Note that TCP BBRv2 and TCP BBRv3 require
that you are on the correct Linux kernel!
Collect heatmap data for a few congestion control schemes in
parameter_exploration.ipynb such as TCP BBRv1, picoquic BBRv3, Cloudflare
CUBIC, etc. We recommend changing the configurations in the third cell to the
following to run fewer trials while still getting a basic idea of the heatmap:
NUM_TRIALS = 1
PDF = False
LOSSES = [0, 1, 2, 3, 4]
DELAYS = [1, 20, 40, 60, 80, 100]
BWS = [10]
Each heatmap takes up to a few hours to collect with one trial. You can
interrupt the process, change execute=True to False, and execute the cell
to visualize intermediate progress. As long as you started the Jupyter notebook
in a detached shell like tmux, you can leave the process running in the
background.
Figure 4: Link rate utilization heatmaps at 10 Mbit/s.
To replicate the remaining figures, similarly re-execute the cells in
network_path_real_data.ipynb and accuracy_analysis.ipynb after moving
the cached data and changing execute=False to True. For the paper, we
executed TCP BBRv2 and TCP BBRv3 benchmarks on different nodes and rsynced
the data directories.
This work was supported in part by NSF grants 2045714 and 2039070, DARPA contract HR001120C0107, a Stanford School of Engineering Fellowship, a Sloan Research Fellowship, affiliate members and other supporters of Stanford DAWN, including Meta, Google, and VMware, as well as Cisco and SAP, and Huawei, Dropbox, Amazon, and the Mozilla Foundation. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors.