search

RealVNF / distributed-drl-coordination

Distributed Online Service Coordination Using Deep Reinforcement Learning
19 stars 6 forks source link
acktr actor-critic coordination distributed-systems networking online-algorithm orchestration placement python reinforcement-learning routing scaling scheduling service service-management stable-baselines tensorflow

readme

Python package

Distributed Online Service Coordination Using Deep Reinforcement Learning

Self-learning and self-adaptive service coordination using deep reinforcement learning (DRL). Service coordination includes scaling and placement of chained service components as well as scheduling and routing of flows/requests through the placed instances. We train our proposed DRL approach offline in a centralized fashion and then deploy a distributed DRL agent at each node in the network. This fully distributed DRL approach only requires local observation and control and significantly outperforms existing state-of-the-art solutions.

Citation

If you use this code, please cite our paper:

@inproceedings{schneider2021distributed,
    title={Distributed Online Service Coordination Using Deep Reinforcement Learning},
    author={Schneider, Stefan and Qarawlus, Haydar and Karl, Holger},
    booktitle={IEEE International Conference on Distributed Computing Systems (ICDCS)},
    year={2021},
    organization={IEEE},
    note={to appear}
}

Installation

This package requires stable_baselines to work. Before installing, make sure the following packages are installed on the system.

# On Ubuntu
sudo apt-get update && sudo apt-get install cmake libopenmpi-dev python3-dev zlib1g-dev libgl1-mesa-glx libsm6 libxext6
# check your python3 version
python3 --version

If your Python version is neither 3.6 nor 3.7 (3.8+ does not support TensorFlow1, which is currently required by stable_baselines), manually install the correct version as described here:

sudo apt update
sudo apt install software-properties-common
sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt install python3.7 python3.7-dev

The package can then be installed as follows, requiring Python 3.6 or 3.7 :

# Create a python 3 virtual environment
python3 -m venv venv
# if python3 != 3.6 or 3.7, use the manually installed python3.7 instead (see above)

# Activate the virtual environment
source venv/bin/activate

# Update pip
pip install -U pip

# Install the DRL package and its requirements (-e is for dev install)
pip install -e .

On Windows 10+, you might need to install MPI separately: https://stackoverflow.com/a/54907810/2745116

Usage

The inputs available for the DRL agent are placed in the inputs folder.

The inputs folder is structured as follows:

.
├── Config:
│   ├── drl: contains the configuration YAML files for the DRL agent
│   └── simulator: contains the simulator configuration files defining the traffic patterns
├── networks: contains the network GraphML files.
├── services: contains the configuration files for the defined services and their components
└── traces: contains a trace to dynamically change the traffic pattern during simulations

Based on these inputs and after installation, a correct installation of the DRL agent can be checked as follows (ignore tensorflow warnings):

$ spr-rl --help

Usage: spr-rl [OPTIONS] NETWORK AGENT_CONFIG SIMULATOR_CONFIG SERVICES TRAINING_DURATION

  SPR-RL DRL Scaling and Placement main executable

Options:
  -s, --seed INTEGER       Set the agent's seed
  -t, --test TEXT          Path to test timestamp and seed
  -a, --append_test        test after training
  -m, --model_path TEXT    path to a model zip file
  -ss, --sim-seed INTEGER  simulator seed
  -b, --best               Select the best agent
  --help                   Show this message and exit.

To train an ACKTR DRL agent for 100,000 steps then test it immediately on the Abilene network, run the following example command from within this directory as follows:

$ spr-rl inputs/networks/abilene_1-5in-1eg/abilene-in1-rand-cap0-2.graphml inputs/config/drl/acktr/acktr_default_4-env.yaml inputs/config/simulator/mean-10.yaml inputs/services/abc-start_delay0.yaml 100000 -a

The results and trained model are saved in the results directory, which is created automatically.

Tensorboard

To visualize training progress, start tensorboard:

tensorboard --logdir tb

Then go to http://localhost:6006

Parallelization

The training of the agent can be parallelized via the GNU Parallel tool. A helper scripts is already provided in the utils folder. The inputs of the agent must be defined in the corresponding *.txt files inside the utils folder To run the parallel scripts: From the current directory, run the following command:

$ ./utils/parallel.sh

Related Projects