Conda / Mamba / Pixi

This exercise focuses on Conda environments (and their Pixi equivalents).

You can find the relevant documentation in the Conda / Mamba / Pixi section of the GSSI training material and in the “Conda-based environment managers” page on VSCDocs.

First steps

You can start with configuring Miniforge, creating a Conda environment and installing software packages in it. You can go about this by either applying the commands from the training material or by using the scripts included here (best done in an interactive job):

cd ${VSC_DATA}/gssi-training/exercises/conda_mamba_pixi
export GSSI_BASEDIR=/tmp
bash wrapper.sh miniforge

Note that the wrapper script:

• removes the ~/.conda directory (contains some cached settings)
• temporarily overwrites your ~/.condarc file

Other packages

The scripts and slides use the tblite package and its Python bindings as an example. You can substitute tblite(-python) with your software package of interest. Use a conda search command or browse the conda-forge package index to see whether it is indeed available as a Conda package.

Other Conda implementations

If interested, you can repeat this exercise with Miniconda, Micrombamba and/or Pixi (see the respective subdirectories).

Other features to try

• Take a look at the output of the conda info command.

• If your Conda environment includes BLAS (as will be the case if you installed tblite), find out which BLAS implementation got installed by looking at the output of conda list. It is likely to be OpenBLAS - you can try changing it to the Intel MKL implementation using conda install "libblas=*=*mkl".

• Create a separate Conda environment with the conda-tree package in it. Activate it and use it to analyze the dependency tree of your first Conda environment.

• Clean your package and other caches with conda clean --all.

Homework: microarchitectural optimizations

• Using one of the Conda implementations, create a new environment and install the latest version of the eigensolver package ELPA with Open MPI support (with all packages taken from the conda-forge channel).

• Execute elpa2_print_kernels and compare the output with what you get if you use the same command from a centrally installed module (e.g. ELPA/2023.05.001-foss-2023a).

• Explain why the Conda installation offers ELPA2 kernels with AVX2 but not with AVX512 instructions, even if conda info indicates your CPU would support it (tip: look at its conda-forge feedstock).