This is a Pytorch implementation of the paper: A Reinforcement Learning Based Strategy for Dynamic Scheduling on Heterogeneous Platforms.

Requirements

• Python 3.7
• For the other packages, please refer to the requirements.txt.

Usage

The main script is train_new.py . You can reproduce the results by:

pip install -r requirements.txt
python train_new.py

Some packages (PyTorch, torch-geometric) may have to be installed separately.

Options

DAG types

It is possible to train on three types of DAG: Cholesky, QR, LU, like in the following examples:

python train_new.py -env_type chol
python train_new.py -env_type QR
python train_new.py -env_type LU

Number of tiles T

The parameter n controls the number of tiles T. As the graphs grow rapidly with the number of tiles, the training can become slow from 10 onwards.

python train_new.py -env_type QR -n 4

Number of CPUs and GPUs

To change the number of processors on which the scheduling is made:

python train_new.py -nCPU 2 -nGPU 2

Number of CPUs and GPUs

For other hyper-parameters concerning the RL algorithm or training, please look at the parser arguments at the top of the train_new file.