Using conda to Run Python on the OSPool¶
The Anaconda/Miniconda distribution of Python is a common tool for installing and managing Python-based software and other tools.
When should you use Miniconda as an installation method in OSG? * Your software has specific conda-centric installation instructions. * The above is true and the software has a lot of dependencies. * You mainly use Python to do your work.
Notes on terminology:
- conda is a Python package manager and package ecosystem that exists in parallel with
- Miniconda is a slim Python distribution, containing the minimum amount of packages necessary for a Python installation that can use conda.
- Anaconda is a pre-built scientific Python distribution based on Miniconda that has many useful scientific packages pre-installed.
To create the smallest, most portable Python installation possible, we recommend starting with Miniconda and installing only the packages you actually require.
To use a Miniconda installation on OSG, create your installation environment on the submit server and send a zipped version to your jobs.
Install Miniconda and Package for Jobs¶
In this approach, we will create an entire software installation inside Miniconda and then use a tool called
conda pack to package it up for running jobs.
1. Create a Miniconda Installation¶
On the submit server, download the latest Linux miniconda installer and run it.
[alice@login05]$ wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh [alice@login05]$ sh Miniconda3-latest-Linux-x86_64.sh
Accept the license agreement and default options. At the end, you can choose whether or not to “initialize Miniconda3 by running conda init?” The default is no; you would then run the eval command listed by the installer to “activate” Miniconda. If you choose “no” you’ll want to save this command so that you can reactivate the Miniconda installation when needed in the future.
2. Create a conda "Environment" With Your Packages¶
(If you are using an
environment.yml file as described later, you should instead create the environment from your
environment.yml file. If you don’t have an
environment.yml file to work with, follow the install instructions in this section. We recommend switching to the
environment.yml method of creating environments once you understand the “manual” method presented here.)
Make sure that you’ve activated the base Miniconda environment if you haven’t already. Your prompt should look like this:
To create an environment, use the
conda create command and then activate the environment:
(base)[alice@login05]$ conda create -n env-name (base)[alice@login05]$ conda activate env-name
Then, run the
conda install command to install the different packages and software you want to include in the installation. How this should look is often listed in the installation examples for software (e.g. Qiime2, Pytorch).
(env-name)[alice@login05]$ conda install pkg1 pkg2
Some Conda packages are only available via specific Conda channels which serve as repositories for hosting and managing packages. If Conda is unable to locate the requested packages using the example above, you may need to have Conda search other channels. More detail are available at https://docs.conda.io/projects/conda/en/latest/user-guide/concepts/channels.html.
Packages may also be installed via
pip, but you should only do this when there is no
conda package available.
Once everything is installed, deactivate the environment to go back to the Miniconda “base” environment.
(env-name)[alice@login05]$ conda deactivate
For example, if you wanted to create an installation with
matplotlib and call the environment
py-data-sci, you would use this sequence of commands:
(base)[alice@login05]$ conda create -n py-data-sci (base)[alice@login05]$ conda activate py-data-sci (py-data-sci)[alice@login05]$ conda install pandas matplotlib (py-data-sci)[alice@login05]$ conda deactivate (base)[alice@login05]$
More About Miniconda¶
See the official conda documentation for more information on creating and managing environments with conda.
3. Create Software Package¶
Make sure that your job’s Miniconda environment is created, but deactivated, so that you’re in the “base” Miniconda environment:
Then, run this command to install the
conda pack tool:
(base)[alice@login05]$ conda install -c conda-forge conda-pack
y when it asks you to install.
conda pack to create a zipped tar.gz file of your environment (substitute the name of your conda environment where you see
env-name), set the proper permissions for this file using
chmod, and check the size of the final tarball:
(base)[alice@login05]$ conda pack -n env-name (base)[alice@login05]$ chmod 644 env-name.tar.gz (base)[alice@login05]$ ls -sh env-name.tar.gz
When this step finishes, you should see a file in your current directory named
4. Check Size of Conda Environment Tar Archive¶
The tar archive,
env-name.tar.gz, created in the previous step will be used as input for subsequent job submission. As with all job input files, you should check the size of this Conda environment file. If >100MB in size, you should NOT transfer the tar ball using
transfer_input_files from your home directory. Instead, you should plan to use OSG Connect's
/public folder, and a
stash:/// link, as described in this guide. Please contact a research computing facilitator at firstname.lastname@example.org if you have questions about the best option for your jobs.
More information is available at File Availability with Squid Web Proxy and Managing Large Data in HTC Jobs.
5. Create a Job Executable¶
The job will need to go through a few steps to use this “packed” conda environment; first, setting the
PATH, then unzipping the environment, then activating it, and finally running whatever program you like. The script below is an example of what is needed (customize as indicated to match your choices above).
#!/bin/bash # have job exit if any command returns with non-zero exit status (aka failure) set -e # replace env-name on the right hand side of this line with the name of your conda environment ENVNAME=env-name # if you need the environment directory to be named something other than the environment name, change this line ENVDIR=$ENVNAME # these lines handle setting up the environment; you shouldn't have to modify them export PATH mkdir $ENVDIR tar -xzf $ENVNAME.tar.gz -C $ENVDIR . $ENVDIR/bin/activate # modify this line to run your desired Python script and any other work you need to do python3 hello.py
6. Submit Jobs¶
In your submit file, make sure to have the following:
- Your executable should be the the bash script you created in step 5.
- Remember to transfer your Python script and the environment
tar.gzfile to the job. If the
tar.gzfile is larger than 100MB, please use the
stash:///file delivery mechanism as described above.
Specifying Exact Dependency Versions¶
An important part of improving reproducibility and consistency between runs is to ensure that you use the correct/expected versions of your dependencies.
When you run a command like
conda install numpy
conda tries to install the most recent version of
numpy For example,
1.22.3 was released on Mar 7, 2022. To install exactly this version of numpy, you would run
conda install numpy=1.22.3 (the same works for
pip if you replace
==). We recommend installing with an explicit version to make sure you have exactly the version of a package that you want. This is often called “pinning” or “locking” the version of the package.
If you want a record of what is installed in your environment, or want to reproduce your environment on another computer, conda can create a file, usually called
environment.yml, that describes the exact versions of all of the packages you have installed in an environment. This file can be re-used by a different conda command to recreate that exact environment on another computer.
To create an
environment.yml file from your currently-activated environment, run
[alice@login05]$ conda env export > environment.yml
environment.yml will pin the exact version of every dependency in your environment. This can sometimes be problematic if you are moving between platforms because a package version may not be available on some other platform, causing an “unsatisfiable dependency” or “inconsistent environment” error. A much less strict pinning is
[alice@login05]$ conda env export --from-history > environment.yml
which only lists packages that you installed manually, and does not pin their versions unless you yourself pinned them during installation. If you need an intermediate solution, it is also possible to manually edit
environment.yml files; see the conda environment documentation for more details about the format and what is possible. In general, exact environment specifications are simply not guaranteed to be transferable between platforms (e.g., between Windows and Linux). We strongly recommend using the strictest possible pinning available to you.
To create an environment from an
environment.yml file, run
[alice@login05]$ conda env create -f environment.yml
By default, the name of the environment will be whatever the name of the source environment was; you can change the name by adding a
-n \<name> option to the
conda env create command.
If you use a source control system like
git, we recommend checking your
environment.yml file into source control and making sure to recreate it when you make changes to your environment. Putting your environment under source control gives you a way to track how it changes along with your own code.
If you are developing software on your local computer for eventual use on the Open Science pool, your workflow might look like this:
- Set up a conda environment for local development and install packages as desired (e.g.,
conda create -n science; conda activate science; conda install numpy).
- Once you are ready to run on the Open Science pool, create an
environment.ymlfile from your local environment (e.g.,
conda env export > environment.yml).
- Move your
environment.ymlfile from your local computer to the submit machine and create an environment from it (e.g.,
conda env create -f environment.yml), then pack it for use in your jobs, as per Create Software Package. More information on conda environments can be found in their documentation.