DRL-TaskMapping

A deep reinforcement learning method for solving task mapping problems.

Requirements

Containerized Environment (Recommended)

Ensure you meet the following system requirements:

• CUDA >= 10.2
• Docker >= 19.03
• NVIDIA Docker >= 2.0 or nvidia-container-toolkit

Bare Metal

• CUDA >= 10.2

• GNU Make >= v4.1

• CMake >= v3.8

• Python >= v3.6.5

• PIP >= v19.0

  • PyPI packages
  • numpy
  • tensorflow-gpu == 1.14.0
  • baseline

• Essential libraries and utilities

  • git, aria2, libboost-dev, pajeng, simgrid, libxrender1, libsm6

Installation

Download the DRL-TaskMapping Source Code

$ git clone https://github.com/NTHU-LSALAB/DRL-TaskMapping.git
$ cd DRL-TaskMapping
$ git submodule update --init --recursive --progress

Setting up the demo environment

Build the docker image

$ bash scripts/build.sh

Extract demo train/test cases

$ tar -xf data/testcases/sample-test.tar.xz -C data/testcases
$ tar -xf data/testcases/sample-train.tar.xz -C data/testcases

Launch the container

$ bash scripts/launch.sh

In the demo, we use a MPI program to explore the communication pattern. Compile it.

$ make -C /data/src

Training

Enter the DRL-TaskMapping and run the training script. The demo trains the model with only 1024 steps. Modify the num_timesteps parameter to train longer.

$ cd workspace/DRL-TaskMapping
$ bash scripts/train.sh

Inference

Run the play.sh to do the inference, the output will be logged at logs/<num_env>/<num_eval>/<checkpoint>/runtime-*

bash scripts/play.sh

Code Structure

DRL-TaskMapping
├── data
│   ├── src                # MPI application 
│   ├── testcases          # Communication pattern
│   └── xmldescs           # Architecture description
├── baselines              # Modified baseline library with our env
│   ├── scripts            # Demo scripts
│   ├── baselines          # Baselines library
│   └── ...
├── docker
│   └── Dockerfile         # Dockerfile
└── scripts                # Build & launch the Docker image