Migration to the conda-forge or bioconda channel

[lohedges]

I've made a few minor updates to BioSimSpace such that all dependencies for Sire and BioSimSpace are now available through the conda-forge channel (other than the fkcombu tool, which we currently bundle with BioSimSpace, but could easily omit). This should hopefully make the eventual migration to conda-forge or bioconda fairly painless. At present the only wrinkle is OpenMM, which is still in testing at conda-forge. Once this is fully integrated into conda-forge we should then be able to build Sire on their platform (assuming no timeout issues).

Depending on conda-forge and bioconda policy we may still wish to host development builds on our own channel. (The information was somewhat ambiguous and I received conflicting advice on this when previously considering conda-forge.)

I'll update this thread as the migration process progresses.

[chryswoods]

Excellent - hopefully this will reduce the frequency of weird conda issues... ;-)

[lohedges]

Well, it took all of a day for the next set of issues:

• In the change from version 4.0.1 to 4.0.2 of openmpi-mpicc (which is used to compile GROMACS and PLUMED for the notebook server) no longer installs a shared library called libmpi_cxx.so.40, so GROMACS and PLUMED fail to find it. This means that we will need to rebuild the packages against the updated library. (I tried symlinking using each of the installed libmpi_* files, but none of these worked.)

• Conda's scoring metrics mean that you end up with version 2019.1.0 of Sire and BioSimSpace if you try to install on Linux. You have to specifically request the latest version of BioSimSpace to get it to work. This is because it tries to minimise, amongst other things, the number of packages installed which can increase if your dependencies (or their dependencies) change over time. If anyone knows of an easy way to force it to always install the most recent version/build of the requested package, then let me know. At present you need to manually specify the package name, which requires the user to go to our Anaconda Cloud page and look it up.

• I'm getting increasingly concerned by the terrible performance when installing and upgrading packages, particularly on macOS, where it can take upwards of 20 minutes on the "Solving environment" stage. (Which is far slower than downloading and unpacking our binary.) There are some discussions regarding the performance issues here.