QueuePilot

A reinforcement learning-based active queue management (AQM) that enables small buffers in backbone routers.

Setup

QueuePilot was tested on Ubuntu 20.04 and python 3.9. A basic configuration includes a source host and a destination host, connected through two routing PCs.

The file environment.yml contains the python dependencies.

sudo apt install -y ethtool openssh-server openssh-client
conda env create -f environment.yml

QueuePilot uses BCC to create and load eBPF code and get access to kernel space. Use BCC's installation instructions to build BCC.

Source and destination hosts

1. Install dependencies: sudo apt install -y ethtool openssh-server nuttcp
2. Build iperf3 from iperf3 repository
3. Copy the scripts from scripts/src and scripts/dst respectively.
4. Create ssh key-file to connect from the router to the source and destination hosts. Changing ECN properties and creating qdisc requires elevated permissions.

Update all topology parameters in gym_lr/rl_src/system_configuration.py

Usage

eBPF requires elevated permissions, so the following should run with sudo or as root. This is in addition to the source/destination hosts commands, so consider your system's security.

conda activate queue_pilot

Train: python train.py [train label]

Test: python test_all_envs.py <test label> <checkpoint_path> Addtional tests can be found in test.py

Output files are available in QueuePilot/logs/. TensorBoard log of the training run is in ~/ray_results/

Contact

Please use GitHub's issues.