Building a Conda environment for Horovod

How to get started with distributed training of DNNs using Horovod.

What is Horovod?

Horovod is an open-source distributed training framework for TensorFlow , Keras , PyTorch , and Apache MXNet . Originally developed by Uber for in house use, Horovod was open sourced a couple of years ago and is now an official Linux Foundation AI (LFAI) project.

In this post I describe how I build Conda environments for my deep learning projects when I am using Horovod to enable distributed training across multiple GPUs (either on the same node or spread across multuple nodes). If you like my approach then you can make use of the template repository on GitHub to get started with your next Horovod data science project!

Installing the NVIDIA CUDA Toolkit

First thing you need to do is to install the appropriate version of the NVIDIA CUDA Toolkit on your workstation. I am using NVIDIA CUDA Toolkit 10.1 (documentation) which works with all three deep learning frameworks that are currently supported by Horovod.

Why not just use the cudatoolkit package?

Typically when installing PyTorch, TensorFlow, or Apache MXNet with GPU support using Conda you simply add the appropriate version of the cudatoolkit package to your environment.yml file.

Unfortunately, for the moment at least, the cudatoolkit package available from conda-forge does not include NVCC which is required in order to use Horovod with either PyTorch, TensorFlow, or MXNet as you need to compile extensions.

What about the cudatoolkit-dev package?

While there are cudatoolkit-dev packages available from conda-forge that do include NVCC, I have had difficult getting these packages to consistently install properly. Some of the available builds require manual intervention to accept license agreements making these builds unsuitable for installing on remote systems (which is critical functionality). Other builds seems to work on Ubuntu but not on other flavors of Linux.

I would encourage you to try adding cudatoolkit-dev to your environment.yml file and see what happens! The package is well maintained so perhaps it will become more stable in the future.

Use the nvcc_linux-64 meta-pacakge!

The most robust approach to obtain NVCC and still use Conda to manage all the other dependencies is to install the NVIDIA CUDA Toolkit on your system and then install a meta-package nvcc_linux-64 from conda-forge which configures your Conda environment to use the NVCC installed on the system together with the other CUDA Toolkit components installed inside the Conda environment. For more details on this package I recommend reading through the issue threads on GitHub .

The environment.yml file

I prefer to specify as many dependencies as possible in the Conda environment.yml file and only specify dependencies in requirements.txt that are not available via Conda channels. Check the official Horovod installation guide for details of required dependencies.

Channel Priority

I use the recommended channel priorities. Note that conda-forge has priority over defaults .

name: nullchannels:
  - pytorch
  - conda-forge
  - defaults

Dependencies

There are a few things worth noting about the dependencies.

1. Even though I have installed the NVIDIA CUDA Toolkit manually I still use Conda to manage the other required CUDA components such as cudnn and nccl (and the optional cupti ).
2. I use two meta-pacakges, cxx-compiler and nvcc_linux-64 , to make sure that suitable C, and C++ compilers are installed and that the resulting Conda environment is aware of the manually installed CUDA Toolkit.
3. Horovod requires some controller library to coordinate work between the various Horovod processes. Typically this will be some MPI implementation such as OpenMPI . However, rather than specifying the openmpi package directly I instead opt for mpi4py Conda package which provides a cuda-aware build of OpenMPI (assuming it is supported by your hardware).
4. Horovod also support that Gloo collective communications library that can be used in place of MPI. I include cmake in order to insure that the Horovod extensions for Gloo are built.

Below are the core required dependencies. The complete environment.yml file is available on GitHub .

dependencies:
  - bokeh=1.4
  - cmake=3.16 # insures that Gloo library extensions will be built
  - cudnn=7.6
  - cupti=10.1
  - cxx-compiler=1.0 # insures C and C++ compilers are available
  - jupyterlab=1.2
  - mpi4py=3.0 # installs cuda-aware openmpi
  - nccl=2.5
  - nodejs=13
  - nvcc_linux-64=10.1 # configures environment to be "cuda-aware"
  - pip=20.0
  - pip:
    - mxnet-cu101mkl==1.6.* # MXNET is installed prior to horovod
    - -r file:requirements.txt
  - python=3.7
  - pytorch=1.4
  - tensorboard=2.1
  - tensorflow-gpu=2.1
  - torchvision=0.5

The requirements.txt File

The requirements.txt file is where all of the pip dependencies, including Horovod itself, are listed for installation. In addition to Horovod I typically will also use pip to install JupyterLab extensions to enable GPU and CPU resource monitoring via jupyterlab-nvdashboard and Tensorboard support via jupyter-tensorboard .

horovod==0.19.*
jupyterlab-nvdashboard==0.2.*
jupyter-tensorboard==0.2.*# make sure horovod is re-compiled if environment is re-built
--no-binary=horovod

Note the use of the --no-binary option at the end of the file. Including this option insures that Horovod will be re-built whenever the Conda environment is re-built.

The complete requirements.txt file is available on GitHub .

Building 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 ./env of 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_ALLREDUCE=NCCL
export HOROVOD_GPU_BROADCAST=NCCL
conda env create --prefix $ENV_PREFIX --file environment.yml --force

By default Horovod will try and build extensions for all detected frameworks. See the Horovod documentation on environment variables for the details on additional environment variables that can be set prior to building Horovod.

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

conda activate $ENV_PREFIX

The postBuild File

If you wish to use any JupyterLab extensions included in the environment.yml and requirements.txt files, then you may need to rebuild the JupyterLab application.

For simplicity, I typically include the instructions for re-building JupyterLab in a postBuild script. Here is what this script looks like for my Horovod environments.

jupyter labextension install --no-build @pyviz/jupyterlab_pyviz
jupyter labextension install --no-build jupyterlab-nvdashboard 
jupyter labextension install --no-build jupyterlab_tensorboard
jupyter serverextension enable jupyterlab_sql --py --sys-prefix
jupyter lab build

Use the following commands to source the postBuild script.

conda activate $ENV_PREFIX # optional if environment already active
. postBuild

Wrapping it all up in a Bash script

I typically wrap these commands into a shell script create-conda-env.sh . Running the shell script will set the Horovod build variables, create the Conda environment, activate the Conda environment, and built JupyterLab with any additional extensions.

#!/bin/bash --loginset -eexport ENV_PREFIX=$PWD/env
export HOROVOD_CUDA_HOME=$CUDA_HOME
export HOROVOD_NCCL_HOME=$ENV_PREFIX
export HOROVOD_GPU_ALLREDUCE=NCCL
export HOROVOD_GPU_BROADCAST=NCCLconda env create --prefix $ENV_PREFIX --file environment.yml --force
conda activate $ENV_PREFIX
. postBuild

I typically put scripts inside a ./bin directory in my project root directory. 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 the deep learning frameworks TensorFlow, PyTorch, Apache MXNet, and the contollers MPI and Gloo 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
    [X] MXNetAvailable Controllers:
    [X] MPI
    [X] GlooAvailable Tensor Operations:
    [X] NCCL
    [ ] DDL
    [ ] CCL
    [X] MPI
    [X] Gloo

Listing the contents of the Conda environment

To see the full list of packages installed into the environment run the following command.

conda activate $ENV_PREFIX # optional if environment already active
conda list

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

However, whenever I add new dependencies I prefer to re-run the Bash script which will re-build both the Conda environment and JupyterLab.

./bin/create-conda-env.sh

Summary

Finding a reproducible process for building Horovod extensions for my deep learning projects was tricky. Key to my solution is the use of meta-packages from conda-forge to insure that the appropriate compilers are installed and that the resulting Conda environment is aware of the system installed NVIDIA CUDA Toolkit. The second key is to use the --no-binary flag in the requirements.txt file to insure that Horovod is re-built whenever the Conda environment is re-built.

If you like my approach then you can make use of the template repository on GitHub to get started with your next Horovod data science project!