Settle problem only on MPI

GROMACS version: 2018, 2020, 2021
GROMACS modification: No

We are running into a problem in a relatively large simulation large box (6x6x12 nm^3), that we have thoroughly minimized (decreasing emtol from default value by an order of magnitude) and equilibrated in NVT ensemble with position restraints and then in the NPT ensemble until density was well converged. When we start our production problem, the simulation runs smoothly when using OpenMP parallelization (specifically, using up to 14 threads). However, as soon as we want to exploit MPI parallelization we run into a settle problem

step 1: One or more water molecules can not be settled.

and the simulation ultimately breaks after we get a problem related with the box

Warning: Triclinic box is too skewed.

We have reproduced the problem with different hardware and multiple Gromacs versions just in case.
Can you suggest any fixes? Is there any reason why this problem may show up depending on which type of parallelization we want to exploit?