Replicate Figure 7 of the NSDI paper How Hard Can It Be? Designing and Implementing a Deployable Multipath TCP.
This work was done by CJ Cullen and Stephen Barber for CS 244.
The associated blog post is located on the Reproducing Network Research blog.
This experiment is published on Amazon EC2 (Oregon) under CS244-Spr14-MPTCP-Trusty
. We used
a c3.large instance - results may vary depending on the instance.
Running the experiment is simple - in the pa3 directory, simply do
sudo ./run_all.sh
python -m SimpleHTTPServer
This runs the base experiment, as well as the sensitivity analysis for jitter over 3G.
After the script finishes running (~25 minutes), you'll find several png files
fig7.png
A faithful reproduction of Figure 7 from the paper.fig7_j150.png
Figure 7 with 3G jitter raised to 150msfig7_j200.png
Figure 7 with 3G jitter raised to 200msfig7_j250.png
Figure 7 with 3G jitter raised to 250ms
A single experiment can be run with
sudo python run.py [optional args, see -h]
python plot.py [optional filename suffix]
Our base setup was Ubuntu 14.04 LTS. On top of this, a custom kernel and mininet need to be installed, along with the code for the experiment.
We used a snapshot of the MPTCP kernel dated May 19, 2014, and applied a patch
(add_sysctls.patch
) to add support for sysctls that disable mechanisms 1 and 2
from the NSDI paper. The patch applies cleanly against commit
2c005d975695b925f21248a11a5d1afed392ee2e
in the mptcp_trunk
branch of the
MPTCP kernel git repository.
The following kernel config settings must be set in order to use MPTCP:
CONFIG_MPTCP=y
CONFIG_MPTCP_PM_ADVANCED=y
CONFIG_MPTCP_FULLMESH=y
CONFIG_MPTCP_NDIFFPORTS=y
CONFIG_DEFAULT_MPTCP_PM="fullmesh"
A sample kernel config (config-3.14.0-mptcp
) is included for convenience.
A patch (mininet-loss-rate.patch
) is included for mininet to add support for
non-integer link loss rates.
This should apply cleanly against commit
5797f5852ec2edf31ce3e0986b6ba15e95ecfd37
in the master
branch of
mininet. After applying the patch, mininet
should be installed with the util/install.sh script.
run.py
accepts arguments for bandwidth, latency, jitter, and loss rates of
both WiFi and 3G links. Run python run.py -h
to see available arguments.
client.c
has defined constants for TCP send buffer size (default 200 KB) and
the time to run the experiment (default 100 seconds).
server.c
has a defined constant for the TCP receive buffer size (default 30
KB).