tMPI error on using mpd deform

GROMACS version: 2021
GROMACS modification: Yes/No

I have a simple POPC membrane which runs well prior to trying to use deform to stretch it.
Only change is to add a line to the pcouple section:
pcoupl = Parrinello-Rahman ;; Berendsen ;
pcoupltype = surface-tension
tau_p = 2.0
ref_p = 1.0 01.0
compressibility = 0 4.5e-5
refcoord_scaling = com
deform = 1 1 0 0 0 0


grompp executes normally with and without the deform statement
mdrun throws an error: tMPI error: Invalid tMPI_Comm (in valid comm)

the cmake I used to compile is straightforward ( cmake 3.20, CUDA 11, gcc 9.3 )
cmake … -DGMX_BUILD_OWN_FFTW=ON -DREGRESSIONTEST_DOWNLOAD=ON -DGMX_GPU=CUDA -DGMX_USE_OPENCL=off

Do I need to recompile with MPI to use deform ( assuming deform is the correct parameter to stretch a membrane ) ?

thanks for any hints.
Paul

Firstly, mdp… deform is the correct parameter to use for strain development.

I am able to get the system to run by not using -ntmpi and -npme when starting a simulation. However without thses specificatons now only one gpu is used . This is likley which i related to the tMPI error.

Is deform somehow restricted to one gpu ?

Paul

Did anyone else look into this problem and can give some insight? I have the same error using the deform option.

Hi,

Are you using the 2022 release? Does the problem only appear if using multiple thread-MPI ranks, e.g. -ntmpi N N>=2. Does it happen with or without direct communication? What happens if you use a regular MPI build?

Cheers,
Szilárd

The issue with the setup in the original post is that a deformation speed of 1 nm per ps is insanely fast. Unless the system is a gas, it will explode due to instabilities. A deformation speed on the order of 1 nm per ns is more reasonable.

We were able to circumvent the error by using less threads (-nt 32), or, by using the 2020.4 version of Gromacs. Change in deformation speed (in my case usually 0.1 nm/ns) had no influence on the issue.

This issue has by now been identified as a bug. The issue is Deform option does not work in parallel (#4604) · Issues · GROMACS / GROMACS · GitLab
The fix is planned to be present in the 2022.4 release.

Note that we have also identified issues when using deform to simulate shear flows. The issue for this is The deform option is not suited for shear flow (#4607) · Issues · GROMACS / GROMACS · GitLab
I am working on a proper fix for this. This fix might just make the 2023 release.