There’s an openblas psi4 conda package available through the following command, so you can gauge the size savings. Be aware, though, of this issue. 400+ test cases are fine, including interaction energies with a variety of functionals (dft-bench-interaction). But there’s a lot of functionals messing up ionization energies with openblas, so I wouldn’t start production calcs just now.
conda create -n p4nomkl psi4 nomkl -c psi4/label/nomkl -c psi4/label/dev
thanks @loriab, it works fine for my test case, the container size is 818MB compared to 1GB of the mkl container (after pruning).
I wonder if I should try to base the Docker image on Alpine instead of Debian.
I wanted to bring https://github.com/psi4/psi4/issues/1009#issuecomment-390869229 to your attention, if you weren’t already following. We’re going to recommend OpenBLAS be used at your own risk and after testing. Also the conda openblas (which was called in that package I built for you) is the worst of all possible threading configurations. I wonder if the 20% size reduction you get for MKL --> OpenBLAS is really worth it. Feel free to comment on that GH issue — it’s the center for OpenBLAS discussion at the moment.
yeah, I agree it is not worth it.