[CP2K-user] 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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20210412/d1992365/attachment.htm>
-------------- next part --------------
#no. of mpi processes,	sigma_xx[GPa],	sigma_yy[GPa],	sigma_zz[GPa]
1   7.53150793685E-02	7.53151564596E-02	7.53151910798E-02
2   1.63976436219E-01	1.63976514203E-01	1.63976549382E-01
4   3.41299316516E-01	3.41299395659E-01	3.41299431171E-01
10  8.73268114771E-01	8.73268194762E-01	8.73268230413E-01
20  1.75988285103E+00	1.75988293119E+00	1.75988296700E+00

numerical stress tensor:

    5.77309220724E-02	9.45089923945E-02	7.37528216119E-02



-------------- next part --------------
A non-text attachment was scrubbed...
Name: System.inp
Type: application/octet-stream
Size: 22318 bytes
Desc: not available
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20210412/d1992365/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: System.out
Type: application/octet-stream
Size: 102748 bytes
Desc: not available
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20210412/d1992365/attachment-0001.obj>