MPI issue?

[Chengcheng-Xiao]

Dear all,

I'd like to share my recent experience in case it may help others facing a similar issue in the future. I successfully resolved the problem by recompiling VASP with the patch available at https://github.com/Chengcheng-Xiao/VASP2WAN90_v2_fix. After recompiling, it's important to note that you should run VASP using only a SINGLE CPU core to generate mmn, amn, and spn. I encountered the same underestimation issue when attempting to generate mmn and amn using multiple CPUs. While I'm not an expert, I suspect this might be related to the use of openmpi or some other factor. Therefore, if you're struggling to obtain correct results, I recommend trying this patch and SINGLE CPU approach (note: reading converged WAVECAR and CHGCAR makes the SINGLE CPU approach affordable). Good luck!

Regards,
Hongyang

Originally posted by @ma455173220 in wannier-berri/wannier-berri#274 (comment)

[Chengcheng-Xiao]

@ma455173220 Thanks for reporting!

I encountered the same underestimation issue when attempting to generate mmn and amn using multiple CPUs.

Humm, this is weird, are you sure the .mmn and .amn files generated with multiple CPUs are different from the ones generated by a single CPU? It would be awesome if you could send me those files so I can take a look.

I suspect this might be related to the use of openmpi or some other factor.

I'm actually using intel MPI (which follows MPICH), not sure if that could be the cause... Will investigate.

[ma455173220]

Dear Dr. Xiao,

I can confirm that using .mmn and .amn files generated with multiple CPUs yields different results from those generated by a single CPU. I've shared the files on OneDrive:
https://unsw-my.sharepoint.com/:f:/g/personal/z5011330_ad_unsw_edu_au/EjtbuYhGYTNLht6--9OvzlwBFufNR6oIAls27S09uyCl8w?e=P8Dot1

I apologize for their large size due to the large KPOINTS, which made it impractical to upload here. It's worth noting that I cannot be certain if this issue affects all OpenMPI users. According to the IT staff at the supercomputer facility I used, the OpenMPI implementation on that particular supercomputer had been optimized and I found it also caused some issues with VASP and CP2K in the past. Therefore, it's possible that this issue is specific to the OpenMPI version on the supercomputer I utilized.

Regards,
Hongyang