VASP 6.5.0 segfaulting with OpenMP parallelization

[asrosen]

I am trying to use VASP 6.5.0 with Intel OneAPI 2024.2 and with OpenMP parallelization. While everything runs correctly with MPI (i.e. changing the number of nodes works as expected), the OpenMP parallelization (i.e. changing cpus-per-task in Slurm) results in a segfault with no clear explanation as to why. I remain unable to resolve the problem.

The following Slurm directives are fine:

Code: Select all

#SBATCH --nodes=1
#SBATCH --ntasks-per-node=112
#SBATCH --cpus-per-task=1   

The following Slurm directives are not fine:

Code: Select all

#SBATCH --nodes=1
#SBATCH --ntasks-per-node=56
#SBATCH --cpus-per-task=2   

All of the relevant files are attached to this post, except for the POTCAR. To summarize things, here is the seg fault info in slurm-493108.out:

Code: Select all

srun: error: tiger-g04c4n9: tasks 14,27,47: Segmentation fault (core dumped)
srun: Terminating StepId=493108.0
slurmstepd: error: *** STEP 493108.0 ON tiger-g04c4n9 CANCELLED AT 2025-01-28T14:28:42 ***
forrtl: error (78): process killed (SIGTERM)

The makefile.include is based off of makefile.include.oneapi_omp. All I did was change MKLROOT and enable the HDF5 flags, setting HDF5_ROOT as well.

Any suggestions would be greatly appreciated.

[alexey.tal]

Dear asrosen,

Thank you for your question.

This issue could be related to the OMP stack size. Could you try to set the following variable and run your test again?

Code: Select all

export OMP_STACKSIZE=512m

Best wishes,
Alexey