#1 2022-02-25 13:37:15

jonas loenartz

Member

Registered: 2021-10-01

Posts: 51

Strange behaviour of residuals with MPI version SM2021 (MUMPS solver)

Hello everyone,

I encountered strange behaviour running an analysis today when using MUMPS solver with the mpi-version.

With the following mpi-settings, everything goes as planned, and the study runs towards completion without errors:

• ncpus = 1; mpi_nbcpus = 2

• ncpus = 2; mpi_nbcpus = 2

But when I change it to:
ncpus = 1; mpi_nbcpus = 4

I get the following error message:

 ╔════════════════════════════════════════════════════════════════════════════════════════════════╗
 ║ <EXCEPTION> <FACTOR_57>                                                                        ║
 ║                                                                                                ║
 ║ Solver MUMPS:                                                                                  ║
 ║  The solution of the linear system is too vague:                                               ║
 ║  Computed error:  120.491                                                                      ║
 ║  Acceptable error:  1e-06 (RESI_RELA)                                                          ║
 ║                                                                                                ║
 ║  Advices:                                                                                      ║
 ║  One can increase the value of the key word SOLVER/RESI_RELA.                                  ║
 ╚════════════════════════════════════════════════════════════════════════════════════════════════╝

The machine I use to run the study has 4 cores.
Is the residual supposed to change based on the mpi-settings (to this degree)?
Is there anything I have to specify in the command-file considering distribution of the model in order to not get this error?

Maybe someone else has encountered this before, but I couldn't find a thread in the forum, concerning this kind of issue.

Jonas

PS: I attached the .comm and the .mess file, just in case

---

Attachments:
files.zip, Size: 13.28 KiB, Downloads: 62

sb1966

Member

Registered: 2015-03-16

Posts: 243

Re: Strange behaviour of residuals with MPI version SM2021 (MUMPS solver)

Hi,

I remember to have read somewhere (sorry for my poor memory) that

1. one should not use more than half the total no of available cores for MPI calculations.
2.The other system processes also need to have cores for running them.

Hope this helps.

Regards
Sukumar

jonas loenartz

Member

Registered: 2021-10-01

Posts: 51

Re: Strange behaviour of residuals with MPI version SM2021 (MUMPS solver)

Thank you for the answer, yes, the settings were kept from another machine with more cores, when I transferred the case. I tested it on another machine, and the behavior is the same.

But I noticed I made a mistake in my model. Some nodes were simultaniously defined as slave-nodes in multiple LIAISON_ interfaces (even though the results were plausible, when the calculation ran to completion).

When I fixed that, the mpi-settings don't seem to have an effect anymore. But nonetheless it would be interesting to know, what caused this kind of behavior while solving.

I will keep the thread open for another week, before I mark this issue as solved. Maybe someone has an explanation.

Jonas