Skip to content

Measuring application level latency in MPTCP

Notifications You must be signed in to change notification settings

evelinad/cs244-pa3

 
 

Repository files navigation

MPTCP application latency experiment over WiFi/3G

Purpose

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.

Running the experiment on Amazon EC2

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 150ms
  • fig7_j200.png Figure 7 with 3G jitter raised to 200ms
  • fig7_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]

Running the experiment from scratch

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.

Tweakable values

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).

About

Measuring application level latency in MPTCP

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published