Skip to content
/ iroko Public

A platform to test reinforcement learning policies in the datacenter setting.

License

Notifications You must be signed in to change notification settings

dcgym/iroko

Repository files navigation

Iroko: The Data Center RL Gym

DISCLAIMER: This project is still very early stage research. It is not stable, well tested, and changes quickly. If you want to use this project, be warned.

Iroko is an open source project that is focused on providing openAI compliant gyms. The aim is to develop machine learning algorithms that address data center problems and to fairly evaluate solutions again traditional techniques.

A concrete description is available in our Arxiv paper. A more elaborate version is presented in this master's thesis. There is also a published workshop paper on the topic:

Iroko: A Framework to Prototype Reinforcement Learning for Data Center Traffic Control. Fabian Ruffy, Michael Przystupa, Ivan Beschastnikh. Workshop on ML for Systems at NIPS 2018.

Requirements

The data center emulator makes heavy uses of Linux tooling and its networking features. It operates most reliably on a recent Linux kernel (4.15+) and is written in Python 3.6+. The supported platform is Ubuntu (at least 16.04 is required). Using the emulator requires full sudo access.

Package Dependencies

  • GCC or Clang and the build-essentials are required.
  • git for version control
  • libnl-route-3-dev to compile the traffic managers
  • ifstat and tcpdump to monitor traffic
  • python3 and python3-setuptools to build Python packages and run the emulator

Python Dependencies

The generator supports only Python3. pip3 can be used to install the packages.

  • numpy for matrix operations
  • gym to install openAI gym
  • seaborn, pandas and matplotlib to generate plots
  • gevent for lightweight threading

Mininet Dependencies

The datacenter networks are emulated using Mininet. At minimum Mininet requires the installation of

  • openvswitch-switch, cgroup-bin, help2man

Ray Dependencies

The emulator uses Ray to implement and evaluate reinforcement learning algorithms. Ray's dependencies include:

  • Pip: tensorflow, setproctitle, psutil, opencv-python, lz4
  • Apt: libsm6, libxext6, libxrender-dev

Goben Dependencies

The emulator generates and measures traffic using Goben. While an amd64 binary is already provided in the repository, the generator submodule can also be compiled using Go 1.11. The contrib/ folder contains a script to install Goben locally.

Installation

A convenient, self-contained way to install the emulator is to run the ./install.sh. It will install most dependencies locally via Poetry.

Run

To test the emulator you can run sudo -E python3 run_basic.py. This is the most basic usage example of the Iroko environment.

run_ray.py contains examples on how to use Ray with this project.

benchmark.py is a test suite, which runs multiple tests in sequence and produces a comparison plot at the end of all runs.