Pascal,
You may try an easy way first, i.e., using runtime option. For example,
petsc/src/ksp/ksp/tutorials/ex2.c:
1) ./ex2 -pc_type lu -pc_factor_mat_solver_type mumps -help |grep mumps
-pc_factor_mat_solver_type <now mumps : formerly (null)>: Specific direct
solver to use (MatGetFactor)
-mat_mumps_icntl_1 <now 6 : formerly 6>: ICNTL(1): output stream for error
messages (None)
-mat_mumps_icntl_2 <now 0 : formerly 0>: ICNTL(2): output stream for
diagnostic printing, statistics, and warning (None)
-mat_mumps_icntl_3 <now 0 : formerly 0>: ICNTL(3): output stream for global
information, collected on the host (None)
-mat_mumps_icntl_4 <now 0 : formerly 0>: ICNTL(4): level of printing (0 to 4)
(None)
...
it tells you that '-mat_mumps_icntl_4 <#>' prints mumps internal info
2) ./ex2 -pc_type lu -pc_factor_mat_solver_type mumps -mat_mumps_icntl_4 2
MEMORY ESTIMATIONS ...
Estimations with standard Full-Rank (FR) factorization:
Total space in MBytes, IC factorization (INFOG(17)): 0
Total space in MBytes, OOC factorization (INFOG(27)): 0
Elapsed time in analysis driver= 0.0006
Entering DMUMPS 5.2.1 from C interface with JOB, N, NNZ = 2 56
250
executing #MPI = 1, without OMP
****** FACTORIZATION STEP ********
GLOBAL STATISTICS PRIOR NUMERICAL FACTORIZATION ...
Number of working processes = 1
ICNTL(22) Out-of-core option = 0
ICNTL(35) BLR activation (eff. choice) = 0
ICNTL(14) Memory relaxation = 20
INFOG(3) Real space for factors (estimated)= 556
INFOG(4) Integer space for factors (estim.)= 1135
Maximum frontal size (estimated) = 11
Number of nodes in the tree = 39
Memory allowed (MB -- 0: N/A ) = 0
Memory provided by user, sum of LWK_USER = 0
...
Hong
________________________________
From: petsc-users <[email protected]> on behalf of Barry Smith
<[email protected]>
Sent: Wednesday, August 12, 2020 11:17 PM
To: Kraft, Pascal (IANM) <[email protected]>
Cc: [email protected] <[email protected]>
Subject: Re: [petsc-users] Measuring memory consumption
Pascal,
Do
KSPGetPC(ksp,&pc);
PCFactorGetMatrix(pc,fact)
PetscViewerPushFormat(PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)fact)),PETSC_VIEWER_ASCII_INFO);
MatView(fact,PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)fact);
PetscViewerPopFormat(PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)fact));
after
SNES/KSPSolve
this should print the various MUMPS variables related to memory usage.
This is processed in /src/mat/impls/aij/mpi/mumps/mumps.c
If you want to process the values directly in the application code you can
look at this file and see how it is extracting the memory values from the mumps
variables and use the variables directly in your code.
Barry
On Aug 12, 2020, at 6:52 PM, Kraft, Pascal (IANM)
<[email protected]<mailto:[email protected]>> wrote:
Dear PETSc Users,
I use the MUMPS wrapper in PETSc (loaded from dealii). I know that MUMPS
computes a factorization based on the mulrifrontal method and since I have a
very memory-strapped problem, I would like to run some test on the memory
requirements for the factorizations across some sets of parameters in my
underlying FEM-problem.
My question is if there is a native way to check memory (as a measure of
fill-in) for the MUMPS factorization in PETSc. I would prefer not to go through
the operating system since that feels somewhat inconclusive to me. Does PETSc
provide such functionality in any way?
I would prefer to know the memory consumption for storing the factorization if
that is possible, I would also take the total memory consumption of the direct
solver if there is no other way, and, if nothing else is possible, I guess I
would have to go with the memory requirement of the whole application if all
else fails.
Is anyone aware of something I can do?
Kind regards,
Pascal
--------------------------------------------------------------------------
Karlsruhe Institute of Technology (KIT)
Institute for Applied and Numerical Mathematics
Kraft, Pascal
Research scientist
Englerstraße 2
Gebäude 20.30
76130, Germany
Phone: +49 721 608-42801
Mobile: +49 163 6927612
E-mail: pascal.kraft∂kit.edu<http://kit.edu/>
Web: www.math.kit.edu/ianm2/~kraft/de<http://www.math.kit.edu/ianm2/~kraft/de>
KIT – The Research University in the Helmholtz Association
Since 2010, the KIT has been certified as a family-friendly university.
