dedalusproject / dedalus_conda

Conda recipes for building Dedalus

Dedalus Conda builds

Conda-forge feedstock

There is currently a Dedalus feedstock on conda-forge which we have recently updated to support the latest Dedalus v2 release. This feedstock builds Dedalus for Linux and macOS (x86 only) with mpich and openmpi support. This build is recommended for laptops, workstations, and clusters that do not require linking to custom C libraries.

Custom build scripts

The scripts in this repository will build a new Conda environment for Dedalus and will allow you to link to system-specific MPI/FFTW/HDF5 libraries, if desired. These scripts are tested nightly on Ubuntu and macOS (x86 only) via GitHub Actions.

dedalus_conda's Issues

Apple M1 runners

Since we've seen some different compatibility problems with conda-forge builds on mac between architectures, it would be great to add M1 runners when possible.

This is currently being tracked here: actions/runner-images#2187 (comment)

Matplotlib throws NSApplication error on osx

Reported here. This is coming from an error with matplotlib backends. It can be fixed by explicitly setting the backend to TkAgg with matplotlib.use('tkagg’).

The issue is being tracked here and a fix is slated for the matplotlib 3.1.0 release in early March.

Full conda build

It would obviously be great to have a full conda build instead of having to use the current conda script, but we currently do so to have the ability to link to local/cluster FFTW, MPI and HDF5 libraries.

However I've just seen there's a new section in the conda-forge docs on "Using External MPI Libraries". It looks like they've created dummy conda builds of mpich that pass through local mpich libraries. If in the future this is extended to include OpenMPI, IMPI, FFTW, and HDF5, it could enable us to build a full conda recipe without losing this functionality.

Creating this issue here just to keep this in mind.

MPI chosen non-deterministically when resolving environment packages

Building from the environment file does not seem to consistenly produce the same build. In particular, openmpi sometimes comes from conda-forge (v3) and sometimes from cryoem (v2), in which case conda-forge installs mpi and mpi4py with mpich, and things are all messed up. A work-around when things go wrong is to remove fftw-mpi from the environment file and add it by hand later.

install_conda.sh does not install mkl

If mkl is selected, the script builds against netlib blas/lapack.

Need compilers with openmpi

New builds of openmpi are built using the anaconda compilers. It looks like a separate package is needed now to make mpicc available to run. More info here. For now, I've pinned openmpi to an older version to avoid.