Ctrl+K
Checkout, Frequently Asked Questions!

Using conda on Shaheen III#

Users can opt to maintain their own software using conda pack manager on Shaheen III.

Installing package manager#

The following steps demonstrate the installation of Conda on Shaheen III Lustre. It is recommended that users install their Miniconda in high IOPS tier of Shaheen III filesystem. It is exempted from periodic purging and is designed to keep up with large number of I/O operations d. The following is a one-off step

mkdir -p $MY_SW && cd $MY_SW
bash /sw/sources/miniconda/conda24.1.2-python3.12.1/Miniconda3-latest-Linux-x86_64.sh -b -s -p $MY_SW/miniconda3-amd64 -u
source $MY_SW/miniconda3-amd64/bin/activate
conda install -y -c conda-forge mamba
conda deactivate

The commands above will install conda package manager’s base environment in $MY_SW/miniconda3-amd64.

Important

Any use of conda package from here on will require user to activate the base or target environment. The following command achieves this:

source $MY_SW/miniconda3-amd64/bin/activate

Optional step#

Installing elaborate conda environments can be time consuming. To accelerate the installation process, mamba package manager can be used. It is a reimplementation of the conda package manager in C++. Simply install mamba in miniconda3’s base environment.

conda install -y -c conda-forge mamba

Creating new environments#

Conda can be activated using the following command:

source $MY_SW/miniconda3-amd64/bin/activate

The following example demonstrates the installation of a Conda envrionment named myenv on Shaheen III using user’s own miniconda installation. The intent is to install Python 3.9 in the environment myenv.

mamba  create --prefix /scratch/$USER/iops/sw/envs/myenv -c conda-forge python=3.9

The Conda environment myenv can be activated as follows:

conda activate $MY_SW/envs/myenv

Important

conda package cache on Shaheen III is redirected to $MY_SW/cache. Clearing this cache directory is the responsibility of user to avoid failures when creating new environments due to lack of space in $SCRATCH_IOPS directory.

conda clean --all

Installing complex environments#

conda environments can sometimes come with a lot of dependencies. This causes them to take long time to install. It is recommended to create such environments as a SLURM batch job instead of running interactively.

One prerequisite for this is to have an environment.yaml file listing all the required software and preferred channels to search these packages. Below is an example environment file:

name: pytorch
channels:
- pytorch
- conda-forge
dependencies:
- python=3.9
- pytorch=2.2.0
- torchvision=0.17.0
- torchaudio=2.2.0
- cpuonly
- jupyterlab
- notebook
- ipykernel
- nb_conda_kernels
- nodejs
- tensorboard
- tensorboardx
- pip
- pip:
    - ipython
    - ipywidgets
    - ipyparallel
    - matplotlib
    - bokeh==2.4.3
    - jupyterlab_nvdashboard
    - pytorch-lightning

The example jobscript will look something as below:

#!/bin/bash
#SBATCH --time 01:0:0
#SBATCH --partition=ppn

source $MY_SW/miniconda3-amd64/bin/activate

mamba env create -f environment.yaml -p $MY_SW/envs/pytorch

And can be submitted to SLURM with the following command.

sbatch job.slurm

Running jobs with conda environments#

Applications installed as conda packages can be run as batch workloads on SLURM. To do this, the jobscripts needs to activate the right environment and then launch the application. Here is an examples jobscript which queries the version and installation location of Pytorch installed the in :ref: previous section <conda_package_install_batch>

#!/bin/bash
#SBATCH -t 00:10:00
#SBATCH -p workq

source $MY_SW/miniconda3-amd64/bin/activate $MY_SW/envs/pytorch

python -c 'import torch; print("Pytorch Version:",torch.__version__)'
python -c 'import torch; print("Pytorch location:",torch.__file__)'

Expected output is:

Pytorch Version: 2.2.0
Pytorch location: </path/to/env>/lib/python3.9/site-packages/torch/__init__.py
On this page