I have trouble with OpenMPI 3.0 and ghost. OpenMPI now includes hwloc and will by default pin your processes in a completely hopeless way. There is the nice option –bind-to none, which resolves the issue for our builtin kernels, but with ghost I’m still missing a factor 10 in performance, and I’m quite sure that it’s because of OpenMPI. On our normal machines I resolved this by compiling OpenMPI without hwloc, but on the cluster we use Spack to install everything, and there hwloc is a reequired dependency of openmpi. Could you try to build ghost and phist with OpenMPI 3 and see if you can find a true fix for this?
I think the single selling point of ghost is performance out-of-the-box, so this is quite critical.