problem with magnetic calculation using vasp with pgi compiler

[mariella_ippolito1]

Dear Vasp support,
I'm still experiencing problems with some magnetic calculations when running vasp/6.1.1 built with Pgi on both Power9 (with or without using GPUs) and x86-64 architectures. The energy doesn't converge properly. On Power9 I ran the calculations with both Pgi/19.1 and hpc-sdk/2010 and the result was the same.
The same calculations run fine on x86-64 architecture using the executable obtained with Intel compiler.
I attach the input files and the outputs obtained on both Power9 and x86-64 architecture, in this last case I attach the outputs obtained with both intel and pgi compiler.

I hope you can help me in solving this issue.

Best regards,
Mariella Ippolito


​​​​​​--
PhD Mariella Ippolito
CINECA SuperComputing Applications and Innovation Department - SCAI
via dei Tizii 6, 00185 Roma - ITALY
Phone: +39 0644486719
Skype: mariellaippolito
[ http://www.hpc.cineca.it/ | www.hpc.cineca.it ]

[mariella_ippolito1]

Sorry, there has been some confusion about the versions of vasp.
Running my magnetic calculations with vasp/6.1.1 (the last version available for download) I obtain divergent energies, on both Power9 and x86 architectures.
Using 6.1.2 the energies don't diverge anymore but the energy values are still different with respect vasp/5.4.4, and the number of iterations for reaching the convergence is double. This is true on both Power9 and x86, and in this last case the results are the same using intel or pgi compiler.
I attach the outputs obtained on x86 with both vasp/6.1.2 (with pgi and intel) and vasp/5.4.4

Thus the problem is not due to the pgi compiler but probably there are still some problems in vasp/6.1.2 related to the constrained magnetic moment.

[henrique_miranda]

A bug is present in the constrained magnetic moments calculation in vasp.6.1.0 and vasp.6.1.1 versions.

You can see the bug report here:
forum/viewtopic.php?f=3&t=17843&p=19248 ... ned#p19248

This information is also reported in the download page in the VASP portal.

Note that the OUTCAR files you sent me are for vasp.6.1.1 19Jun20.
Does the same issue occur with vasp.6.1.2?

[mariella_ippolito1]

vasp/6.1.2 is not available for download.
I used the source files that Martijn Marsman gave me on 2 July 2020 (I guess that it is 6.1.2 version) where the bug related to constramag.F had already been removed.
The output I attached related to the calculations on Power9 was obtained with that vasp version.
This vasp version give the same problems on x86 architecture using both intel or pgi compiler.
The calculation run fine using vasp/5.4.4.

[ferenc_karsai]

VASP 6.1.2 should be available, but only supervisors (licencse holders) can download.

[mariella_ippolito1]

I've installed vasp/6.1.2 and I run again this calculation with the same result obtained with the older vasp version.
I attach the output obtained with vasp/6.1.2 (the one obtained with 6.1.1 was the same).
I remind you that the job runs fine on x86 if I use vasp compiled with intel, on the contrary there are the same problems found on power9 also on x86 when using the pgi vasp version. The outputs obtained on x86 (with both intel and pgi) are also attached.