Re: Nonlinear term MPI running does not end

[Konstantinos Poulios]

Dear Tetsuo,

Have you compiled GetFEM with the parallel version of mumps? In Ubuntu/Debian you must link to dmumps instead of dmumps_seq for example.

BR

Kostast

On Tue, May 18, 2021 at 2:09 PM Tetsuo Koyama <tkoyama010@gmail.com> wrote:

Dear Kostas

Thank you for your report.

I am happy that it runs well in your system.

I will organize the procedure that can reproduce this error. Please wait.

Best Regards Tetsuo

2021年5月18日(火) 18:10 Konstantinos Poulios <logari81@googlemail.com>:

Dear Tetsuo,

I could not confirm this issue. On my system the example runs well both on 1 and 2 processes (it doesn't scale well though)

BR

Kostas

On Sun, May 16, 2021 at 10:07 AM Tetsuo Koyama <tkoyama010@gmail.com> wrote:

Dear Kostas

I am looking inside the source code.
> if (generic_expressions.size()) {...}
Sorry it looks complex for me.

FYI. I found that MPI process 1 and 2 is different in the following line.
>    if (iter.finished(crit)) {
This is in the "Newton_with_step_control" function in getfem_model_solvers.h.

"crit" is calculated by rit = res / approx_eln and res and approx_eln is ...

$ mpirun -n 1 python demo_parallel_laplacian.py
res=1.31449e-11
approx_eln=6.10757
crit=2.15222e-12

$ mpirun -n 2 python demo_parallel_laplacian.py
res=6.02926
approx_eln=12.2151
crit=0.493588

res=0.135744
approx_eln=12.2151
crit=0.0111128

I am now trying to understand what is the correct residual value of  Newton(-Raphson) algorithm.
I will be glad if you have an opinion.

Best Regards Tetsuo

2021年5月11日(火) 19:28 Tetsuo Koyama <tkoyama010@gmail.com>:

Dear Kostas

> The relevant code is in the void model::assembly function in getfem_models.cc. The relevant code assembling the term you add with md.add_nonlinear_term(..) must be executed inside the if condition

>
> if (generic_expressions.size()) {...}

> You can have a look there and ask for further help if it looks too complex. You should also check if the test works when you run it with md.add_nonlinear_term but setting the number of MPI processes to one.

Thanks. I will check it. And the following command completed successfully..

$ mpirun -n 1 python demo_parallel_laplacian.py

So all we have to check is compare -n 1 with -n2 .

Best regards Tetsuo

2021年5月11日(火) 18:44 Konstantinos Poulios <logari81@googlemail.com>:

Dear Tetsuo,

The relevant code is in the void model::assembly function in getfem_models.cc. The relevant code assembling the term you add with md.add_nonlinear_term(..) must be executed inside the if condition

if (generic_expressions.size()) {...}

You can have a look there and ask for further help if it looks too complex. You should also check if the test works when you run it with md.add_nonlinear_term but setting the number of MPI processes to one.

BR

Kostas

On Tue, May 11, 2021 at 10:44 AM Tetsuo Koyama <tkoyama010@gmail.com> wrote:

Dear Kostas

Thank you for your reply.

> Interesting. In order to isolate the issue, can you also check with

> md.add_linear_term(..)

> ?
It ends when using md.add_linear_term(..).

It seems that it is a problem of md.add_nonlinear_term(..).

Is there a point which I can check?

Best regards Tetsuo.

2021年5月11日(火) 17:19 Konstantinos Poulios <logari81@googlemail.com>:

Dear Tetsuo,

Interesting. In order to isolate the issue, can you also check with
md.add_linear_term(..)

?

Best regards

Kostas

On Tue, May 11, 2021 at 12:22 AM Tetsuo Koyama <tkoyama010@gmail.com> wrote:

Dear GetFEM community

I am running MPI Parallelization of GetFEM.The running command is

$ git clone https://git.savannah.nongnu.org/git/getfem.git
$ cd getfem
$ bash autogen.sh
$ ./configure --with-pic --enable-paralevel=2
$ make
$ make install
$ mpirun -n 2 python demo_parallel_laplacian.py

The python script ends correctly. But when I changed the following linear term to nonlinear term the script did not end.

-md.add_Laplacian_brick(mim, 'u')
+md.add_nonlinear_term(mim, "Grad_u.Grad_Test_u")

Do you know the reason?
Best regards Tetsuo