Improving Datacenter Performance using MPTCP

CS244: Win 2013

The goal of this project is to use Mininet to reproduce the results of: Improving Datacenter Performance and Robustness with Multipath TCP by Raiciu et al. which appeared in SIGCOMM'11.

Authors: Brian O'Connor & Ryan Wilson

Preliminaries

This project has the following dependencies:

• Mininet 2.0 http://mininet.github.com/
• An MPTCP-enabled Linux kernel http://mptcp.info.ucl.ac.be/
• RiplPox (and its dependencies) https://github.com/brandonheller/riplpox
• Matplotlib http://matplotlib.org/index.html
• TermColor https://pypi.python.org/pypi/termcolor
• Sysstat sudo apt-get install sysstat

Starting with our image on EC2 (Recommended)

This is the easiest way to get started.

We have made the following image available on EC2 [US West (Oregon)]: CS244-Mininet-MPTCP

You should use a c1.xlarge instance for the experiments.

If you use a c1.medium instance, make sure to turn off queue monitoring and don't use a Fat Tree topology with k > 6, unless your goal is to max out the CPU.

For more detailed instructions, follow the guide here: http://www.stanford.edu/class/cs244/2013/ec2.html

Starting with the CS244 image on EC2

Start will the following EC2 image: CS244-Win13-Mininet

Then, follow the instructions here: https://github.com/bocon13/mptcp_setup

Starting with a fresh Ubuntu 12.10 image on EC2

You will need to install Mininet, Matplotlib, TermColor, and sysstat on your own.

We have made the following scripts to install MPTCP and RiplPox (+ dependencies): https://github.com/bocon13/mptcp_setup

Setting up the Test

Clone the code repository

git clone https://github.com/bocon13/datacenter_mptcp.git

Initialize the util submodule and clone it

cd datacenter_mptcp

git submodule init

git submodule update

If not using "CS244-Mininet-MPTCP" image, patch and install iperf:

./iperf_patch/build-patched-iperf.sh

Running the test

Use the following command to run the experiments:

sudo ./run_mptcp.sh

This will run the one-to-one workload (used in the paper) on a Fat Tree topology (k=4) with queue monitoring disabled.

Test options

Our test is capable of running various sized FatTree topologies and workloads.

You can specify them in the run command: sudo ./run_mptcp.sh <size of fat tree> <workload> [qmon]

For <size of fat tree>, valid values are even integers >= 4. (default = 4)

For <workload>, valid values are:

• one_to_one (default)
• one_to_several
• all_to_all

To enable queue monitoring, add qmon as the third argument. It is disabled by default.

Examples:

• Run a Fat Tree topology (k=6) with one_to_one workload: sudo ./run_mptcp.sh 6
• Run a Fat Tree topology (k=8) with all_to_all workload with queue monitoring disabled: sudo ./run_mptcp.sh 8 all_to_all
• Run all workloads on various Fat Tree topologies: k = {4, 6, 8, 10, 12} with queue monitoring disabled: sudo ./run_mptcp.sh all all
• Run a Fat Tree topology (k=4) with one_to_one workload with queue monitoring enabled: sudo ./run_mptcp.sh 4 one_to_one qmon

IMPORTANT: If you plan to run the same test twice (the same workload on the same topology), make sure you run the command between tests: sudo rm -rf results/ft<size of fat tree>/<workload>

Observing your results

The results are graphs that are saved to the plots sub-directory.

Easiest way to view results

Start a webserver on your EC2 instance (ensure that port 8000 is open):
Check out the Modify Security Group Settings in http://www.stanford.edu/class/cs244/2013/ec2.html

python -m SimpleHTTPServer &> /dev/null &

Navigate to your server in your favorite web browser:

http://<your ec2 instance name>:8000/