inconsistent analytical stress tensor in cp2k mpi runs

[Severin Sylla]

Dear cp2k community,

I am calculating the analytical stress tensor of a cubic diamond 3x3x3 supercell using the popt/psmp version of cp2k.
I observe that the diagonal elements of the analytical stress tensor depend on the number of mpi processes. Only when running cp2k with one single mpi task, the analytical stress tensor roughly agrees with the numerical one.
I encountered this issue with cp2k 8.1 + openmpi 4.0 (compiled with gcc 9.2), and also with intel mpi 2018.3 (intel compiler suite 18.0).

Any suggestions on how to fix/avoid this inconsistency would be appreciated. Thanks!

In the attachment, you can find the input file, an example of an output file, and the computed diagonal elements of the stress tensor for different number of mpi processes. The problem does not affect the off-diagonal entries.

Kind Regards,
Severin

[Lucas Lodeiro]

Just a thing... It is weird the values of numeric stress tensor. In a cubic cell as diamond, the stress tensor has to be isotropic, or at least, the difference on the diagonal terms must be very low and relative to noise and convergence error. In this case the difference are huge... there is some problem related with parameters or implementation (more difficult).

--
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/CAOkt55U2ey15yqevddgs2MmHntQC7pVKisce5zzxkMsVK3hErA%40mail.gmail.com.

[Krack Matthias (PSI)]

Hi Severin

 

Do you observe the same behavior with LONG_RANGE_CORRECTION disabled?

 

Best regards

 

Matthias

 

--

[Severin Sylla]

Hi Matthias,

Thanks for your reply. The inconsistency disappears when I disable LONG_RANGE_CORRECTION.

Kind regards,

Severin

[Krack Matthias (PSI)]

Hi Severin

 

Thanks for testing. So it looks like a bug which appears when the LONG_RANGE_CORRECTION is enabled for DFT-D3 which is fortunately not the default. I have filed that bug as issue #1462. Thanks for reporting.

 

Best regards

 

Matthias

 

To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/5af6f64b-d707-4d23-9c7f-c2954af2a094n%40googlegroups.com.