*** WARNING in fm/cp_fm_elpa.F:522 :: Setting real_kernel for ELPA failed ***

912 views
Skip to first unread message

Max Nishibuchi

unread,
Feb 29, 2024, 11:18:01 AM2/29/24
to cp2k
Hello all, 

I have been trying to perform a geometric optimization on a polycarbonate structure, and I repeatedly see the following warning appear in the output file despite the relaxation continuing. 

 *** WARNING in fm/cp_fm_elpa.F:522 :: Setting real_kernel for ELPA failed ***

I am running the simulation using the most recent generic openmpi Docker container (available here, 2024.1_openmpi_generic_psmp). I have been able to converge geometric optimizations in the past while this warning was still present. 

Do I need to worry about this warning? It's not clear to me what role real_kernel plays in the simulation. I've attached my input file for the sake of replication. 
input.inp

Guilherme da Silva

unread,
Mar 2, 2024, 8:17:22 PM3/2/24
to cp2k
Hi,

I will share with you my little experience with this warning.

I believe it is just a matter of how elpa was compiled not being tuned to the machine you are running the simulation.The ELPA_KERNEL default is AUTO and somehow it didnt manage to choose one of the options to the target machine.

I tested with elpa or scalapack some single points and geometry optimizations and, even with that warning, results were the same.

If you are still not confident about that you can try two changes under the GLOBAL section:

1. Choose PREFERRED_DIAG_LIBRARY SCALAPACK -

2.
Change ELPA_KERNEL to GENERIC or GENERIC_SIMPLE.

In some tests I did, SCALAPACK or GENERIC were (very roughly) 0.8% slower than GENERIC_SIMPLE.


Kind regards

Max Nishibuchi

unread,
Mar 5, 2024, 11:39:47 AM3/5/24
to cp2k
Guilherme,

Thank you very much for sharing your experience with this warning. 

I'll do a few tests with the GLOBAL parameters you mentioned, but it's a relief to know that it shouldn't make a significant difference in the accuracy of the simulation results.

Have a great day!

Kind regards,

Max

Reply all
Reply to author
Forward
0 new messages