Using Anaconda for python environments

Python is a commonly used programming language for scientific workloads. Among the ways to use Python libraries, the Anaconda binary distribution is a very popular means for managing python packages and environments. On Mines’ HPC system users, we recommend they use our centrally managed anaconda installation for setting up Python enivronments.

We will go through how to setup several common anaconda environments.

Setting up base environment

For all environments listed below, we will need to setup a base environment. There are a few different ways to do this, with several advanced options that can be viewed here.

Loading the Python module

Wendian and Mio

module load apps/python3      

Creating and activating a conda environment under your home directory

To create a conda environment using a environment name of your choice requires the following command:

conda create --name name_of_your_environment

You then will be asked to confirm by typing ‘y’ and press enter.

This will create your new environment and will be saved under under $HOME/.conda/envs/name_of_your_environment

To activate the environment, type the command

conda activate name_of_your_environment

In your terminal, you should now see the following to the left of your username

(name_of_your_environment) username@hpc$

Creating conda environment in a custom location

To create a conda environment in a custom location, use the —prefix flag when using conda create:

conda create --prefix=/your/custom/path/

You will be asked to confirm by typing ‘y’ like above. To activate the environment, type the command

conda activate /your/custom/path

Appliciation-specific conda environment

All instructions here we be assumed to be created under `$HOME/.conda

NumPy + SciPy + Matplotlib

For this environment, you can set it up in one command when initially creating the environment:

conda create --name scientific_python numpy scipy matplotlib

Tensorflow (GPU)

For Tensorflow with NVIDIA GPU support, GPU support is included with Tensorflow:

create create --name tf_env tensorflow 

Tensorflow (CPU-only)

If you want tensorflow with CPU support only, use the following command when creating your environment:

conda create --name tf_cpu_only_env tensorflow-cpu

Cleaning up conda packages

After installing packages into a conda environment, there will many packages left as cache under $HOME/.conda/pkgs by default. These cached packages can be reused when creating new environments that depend on packages that were previously installed. However, this cache can become quite large and often cause you to reach the 20 GB limit of $HOME. There are few different options for cleaning up the package cache.

Remove all cached packages

conda clean --all

This will remove all cached packages under $HOME/.conda/pkgs

Proper mpi4py setup to use system MPI on Wendian/Mio

Load the required modules

module load apps/python3
module load compilers/gcc
module load mpi/openmpi/gcc-cuda

Set the environment variable CC to mpicc:

export CC=$(which mpicc)

You can double check it was set correctly by typing:

echo $CC

And you should see something like:

/sw/mpi/openmpi/4.1.4/gnu-cuda/bin/mpicc

Next, create your conda environment:

conda create -n mpi4py_test python=3.11 -y

and activate it:

conda activate mpi4py_test

Now install mpi4py using pip not conda:

pip install --no-cache-dir mpi4py numpy

You can double check it works by launching python and importing the module:

$ python
>>> from mpi4py import MPI
>>> MPI.Comm
<class 'mpi4py.MPI.Comm'>