tianhaotao / rltaskoffloading

RLTaskOffloading

Prerequisites

The code requires python3 (>=3.5) with the development headers. You'll also need system packages CMake, OpenMPI, graphviz and zlib. Those can be installed as follows

Ubuntu

sudo apt-get update && sudo apt-get install cmake libopenmpi-dev python3-dev zlib1g-dev
sudo apt-get install graphviz

Configure the virtual environment:

It is better to use a virtual environment (e.g., Anaconda) to run the code. About how to install anaconda, please refer to the official website: https://www.anaconda.com

Once you have anaconda installed, run

conda env create -f environment.yaml

To create the virtual environment. The current version of the code only supports TensorFlow 1.x (>=1.5).

Run the code

We implemented two DRL-based algorithms for task offloading: DRLTO and DDQNTO.

To train and evaluate DRLTO under different scenarios, run

# train and evaluate DRLTO with different number of tasks and LO target.
python train.py --algo DRLTO --scenario Number --goal LO --dependency True 
# train and evaluate DRLTO with different number of tasks and EE target.
python train.py --algo DRLTO --scenario Number --goal EE --dependency True 
# train and evaluate DRLTO with different transmission rate and LO target.
python train.py --algo DRLTO --scenario Trans --goal LO --dependency True 
# train and evaluate DRLTO with different transmission rate and EE target.
python train.py --algo DRLTO --scenario Trans --goal EE --dependency True 

To train DRLTO without considering task dependency, run

# train and evaluate DRLTO with different number of tasks and LO target without considering dependency
python train.py --algo DRLTO --scenario Number --goal LO --dependency False 

To train and evaluate DDQNTO under different scenarios, run

# train and evaluate DDQNTO with different number of tasks and LO target. In DDQNTO we do not consider the dependency.
python train.py --algo DDQNTO --scenario Number --goal LO --dependency False 
# train and evaluate DDQNTO with different number of tasks and EE target.
python train.py --algo DDQNTO --scenario Number --goal EE --dependency False 
# train and evaluate DDQNTO with different transmission rate and LO target.
python train.py --algo DDQNTO --scenario Trans --goal LO --dependency False 
# train and evaluate DDQNTO with different transmission rate and EE target.
python train.py --algo DDQNTO --scenario Trans --goal EE --dependency False 

The running results can be found in the log folder (default path of log folder is './log/Result')

To evaluate the heuristic algorithms, run

python evaluate_heuristic_algo.py --scenario Number --goal LO
python evaluate_heuristic_algo.py --scenario Number --goal EE
python evaluate_heuristic_algo.py --scenario Trans --goal LO
python evaluate_heuristic_algo.py --scenario Trans --goal EE

Related publication

If you are interested in this work, please cite the paper

@article{Wang2021Depedent,
  author={Wang, Jin and Hu, Jia and Min, Geyong and Zhan, Wenhan and Zomaya, Albert and Georgalas, Nektarios},
  journal={IEEE Transactions on Computers}, 
  title={Dependent Task Offloading for Edge Computing based on Deep Reinforcement Learning}, 
  year={2021},
  doi={10.1109/TC.2021.3131040}}