Install on cray with mpi

GROMACS version: 2022.5
GROMACS modification: No

Hi,

I would like to install gmx mpi for multidir simulations within a single node. Although it is not comparable, but it seems OK on my local box, but I fail to have a good build on our Cray. I use gnu prg env, gcc, cuda, the same modules as for building the threaded version. Plus openmpi.
With the system’s openmpi (5) the gmx_mpi crashes right after start.
With my compiled openmpi it runs, but 5-6 times slower than the threaded version (I tested different cmdline arguments to increase performance). This is not a big surprise for me, since I do not think that my openmpi is cuda-aware.

If I want to tune and force to have a better openmpi build, I always fail - it always fails: sometimes with missing libraries during compile time, sometimes with missing libraries at run time. I have tried mpich, too.

Do you have experiences on Cray? Which modules should I use to have a smooth build with high performance?

Sorry, I do not know what specific info to provide to help answering.
Thanks, Tamas