I should also check how to specify to GROMACS which compiler (and MPI lib) to use, to avoid that it picks up icc/ifort that may be available on the system, and prefer it over GCC (when building with goolf)...
Behavior like this was reported by @ocaisa

you are likely looking for help from this:
http://www.hpcadvisorycouncil.com/pdf/GROMACS_Best_Practices.pdf
It seems you need something of the style
./configure --enable-mpi --with-fft=mkl CC=icc F77=ifort CXX=icpc,
yet watch out because around/before 4.6 gromacs transitioned to cmake build, so you'll need to adjust it.

It turns out that simple enabling MPI support in GROMACS goolf easyconfigs is not the way to go: the regression test hangs with goolf or goolfc when MPI is enabled, in the first test run, i.e. regressiontests-4.6.1/simple/angles1. The cause is unclear.

I cleaned up the GROMACS easyconfigs by introducing a custom easyblock for GROMACS, see easybuilders/easybuild-easyblocks#335. The easyconfigs are now explicit one of the following options:

• -seq: no OpenMP, no MPI
• -mt: OpenMP, no MPI
• -mpi: no OpenMPI, MPI
• -hybrid: both OpenMP and MPI

For goolf and goolfc, the -mpi and -hybrid easyconfigs are left out, since the regression test seems to be hanging there (and eventually leads to a broken build when the test times out.

well-tested on top of easybuilders/easybuild-easyblocks#335, going in...