This is a forked version of Remy which adds support for reinforcement learning (using code from the repository async_deep_reinforce). The results were published in Bachl, Maximilian, Tanja Zseby, and Joachim Fabini. "Rax: Deep Reinforcement Learning for Congestion Control." ICC 2019-2019 IEEE International Conference on Communications (ICC). IEEE, 2019.

All files with unicorn in the file name were added for this purpose.

The license for Remy can be found at debian/copyright. All added unicorn files are licensed under the same license. The license for async_deep_reinforce can be found at async_deep_reinforce/LICENSE.txt.

The modified version remy-unicorn can be launched like remy as follows:

learning_rate=-4.5 reward_type=no_cutoff src/remy-unicorn cf=config/2_2_really_small_buffer_0.cfg

This launches a simulation with 2 senders, an RTT of 100 ms, a buffer of one tenth of the bandwidth delay product, a throughput/packet-loss-tradeoff of 2, a stochastic packet loss probability of 0 and a link speed of 20 Mbit/s.

-- Maximilian Bachl

Remy is an optimization tool to develop new TCP congestion-control schemes, given prior knowledge about the network it will encounter and an objective to optimize for.

It is described further at the Web site for TCP ex Machina. A research paper on Remy was published at the ACM SIGCOMM 2013 annual conference.

Basic usage:

• Remy requires a C++11 compiler to compile, e.g. gcc 4.6 or contemporary clang++. You will also need the Google protobuf-compiler and the Boost C++ libraries.

• From the version-control repository checkout, run ./autogen.sh, ./configure, and make to build Remy.

• Run ./remy to design a RemyCC (congestion-control algorithm) for the default scenario, with link speed drawn uniformly between 10 and 20 Mbps, minRTT drawn uniformly between 100 and 200 ms, the maximum degree of multiplexing drawn uniformly between 1 and 32, and each sender "on" for an exponentially-distributed amount of time (mean 5 s) and off for durations drawn from the same distribution.

• Use the of= argument to have Remy save its RemyCCs to disk. It will save every step of the iteration.

• Use the if= argument to get Remy to read previous RemyCCs as the starting point for optimization.

• The sender-runner tool will execute saved RemyCCs. The filename should be set with a if= argument. It also accepts link= to set the link speed (in packets per millisecond), rtt= to set the RTT, and nsrc= to set the maximum degree of multiplexing.

If you have any questions regarding Remy, please visit Remy's Web site or e-mail remy at mit dot edu.

-- Keith Winstein