Skip to content

KAUST-Academy/horovod-gpu-data-science-project

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

27 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

horovod-gpu-data-science-project

Repository containing scaffolding for a Python 3-based data science project that uses distributed, multi-gpu training with Horovod together with one of TensorFlow, PyTorch, or MXNET.

Creating a new project from this template

Simply follow the instructions to create a new project repository from this template.

Project organization

Project organization is based on ideas from Good Enough Practices for Scientific Computing.

  1. Put each project in its own directory, which is named after the project.
  2. Put external scripts or compiled programs in the bin directory.
  3. Put raw data and metadata in a data directory.
  4. Put text documents associated with the project in the doc directory.
  5. Put all Docker related files in the docker directory.
  6. Install the Conda environment into an env directory.
  7. Put all notebooks in the notebooks directory.
  8. Put files generated during cleanup and analysis in a results directory.
  9. Put project source code in the src directory.
  10. Name all files to reflect their content or function.

Installing NVIDIA CUDA Toolkit

Workstation

You will need to have the appropriate version of the NVIDIA CUDA Toolkit installed on your workstation. For this repo we are using NVIDIA CUDA Toolkit 10.1 (documentation).

After installing the appropriate version of the NVIDIA CUDA Toolkit you will need to set the following environment variables.

$ export CUDA_HOME=/usr/local/cuda-10.1
$ export PATH=$CUDA_HOME/bin:$PATH
$ export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH

Ibex

Ibex users do not neet to install NVIDIA CUDA Toolkit as the relevant versions have already been made available on Ibex by the Ibex Systems team. Users simply need to load the appropriate version using the module tool.

$ module load cuda/11.0.1

Building the Conda environment

After adding any necessary dependencies that should be downloaded via conda to the environment.yml file and any dependencies that should be downloaded via pip to the requirements.txt file you create the Conda environment in a sub-directory ./envof your project directory by running the following commands.

export ENV_PREFIX=$PWD/env
export HOROVOD_CUDA_HOME=$CUDA_HOME
export HOROVOD_NCCL_HOME=$ENV_PREFIX
export HOROVOD_GPU_OPERATIONS=NCCL
conda env create --prefix $ENV_PREFIX --file environment.yml --force

Once the new environment has been created you can activate the environment with the following command.

conda activate $ENV_PREFIX

Note that the ENV_PREFIX directory is not under version control as it can always be re-created as necessary.

For your convenience these commands have been combined in a shell script ./bin/create-conda-env.sh. The script should be run from the project root directory as follows.

./bin/create-conda-env.sh # assumes that $CUDA_HOME is set properly

Verifying the Conda environment

After building the Conda environment you can check that Horovod has been built with support for TensorFlow and MPI with the following command.

conda activate $ENV_PREFIX # optional if environment already active
horovodrun --check-build

You should see output similar to the following.

Horovod v0.19.1:

Available Frameworks:
    [X] TensorFlow
    [X] PyTorch
    [ ] MXNet

Available Controllers:
    [X] MPI
    [X] Gloo

Available Tensor Operations:
    [X] NCCL
    [ ] DDL
    [ ] CCL
    [X] MPI
    [X] Gloo  

Listing the full contents of the Conda environment

The list of explicit dependencies for the project are listed in the environment.yml file. To see the full lost of packages installed into the environment run the following command.

conda list --prefix $ENV_PREFIX

Updating the Conda environment

If you add (remove) dependencies to (from) the environment.yml file or the requirements.txt file after the environment has already been created, then you can re-create the environment with the following command.

$ conda env create --prefix $ENV_PREFIX --file environment.yml --force

Using Docker

In order to build Docker images for your project and run containers with GPU acceleration you will need to install Docker, Docker Compose and the NVIDIA Docker runtime.

Detailed instructions for using Docker to build and image and launch containers can be found in the docker/README.md.

About

Template repository for a Python 3-based data science project that uses Horovod.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published