Guanzhou Hu @ UW-Madison, CS740, Spring 2022

This is a minimal implementation of the DCTCP congestion control protocol (SIGCOMM paper, RFC 8257) with the ns-3 simulator. Tested on Ubuntu 20.04.

It turns out that ns-3.35 already ships with a default implementation of DCTCP. This repo holds my own minimal implementation based on my understanding of the original SIGCOMM paper.

Due to the design of ns-3, all new components and simulation scripts must be put within the ns-3.35 folder, be built by the default ./waf toolchain, and be run with ./waf --run. Hence, ns-3.35 source code is included in-place. I added my new DCTCP component into the source folder and modified the build scripts to involve the new stuff.

Files added or modified:

• ns-3.35/src/internet/model/tcp-dctcp-my.cc: my implementation of TcpDctcpMy component
• ns-3.35/src/internet/model/tcp-dctcp-my.h: mostly a duplicate of tcp-dctcp.h
• ns-3.35/src/internet/test/tcp-dctcp-my-test.cc: a duplicate of tcp-dctcp.test.cc
• ns-3.35/src/internet/wscript: modified to involve TcpDctcpMy stuff when building
• ns-3.35/scratch/trace-tcp-reno.cc: simple simulation experiment on TcpLinuxReno
• ns-3.35/scratch/trace-dctcp-my.cc: simple simulation experiment on TcpDctcpMy
• ns-3.35/scratch/paper-dctcp-my.cc: a duplicate of dctcp-example.cc

Build ns-3, including the new components added:

make clean && make

Run unit tests to ensure all PASSED:

make test

Two simple simulation scripts, trace-tcp-reno.cc and trace-dctcp-my.cc, have been provided to test the congestion window adjustment behavior of my DCTCP algorithm and the default Reno algorithm. The simulation involves 20 senders, 1 switch, and 1 receiver. Senders each connect to the switch through 1Gbps link, while the switch is connected to the receiver through 10Gbps link. Each sender inejcts a steady stream of packets at rate 1Gbps to the receiver.

Run the simulation:

make simulation

You will find trace-tcp-reno.log and trace-dctcp-my.log at the root directory of this repo once this finished. In both logs, the time region between 2 secs - 3 secs is the meaningful measurement window after convergence.

ns-3 ships with a much more complicated example simulation that mimicks the exepriment originally presented in the DCTCP paper. I also duplicated that one to paper-dctcp-my.cc.

Run the paper experiment:

make paper-expt     # may take long

You will find a bunch of .dat data logs under ns-3.35/ once this finished.