This is a read only copy of the old FEniCS QA forum. Please visit the new QA forum to ask questions

how to make MUMPS run in parallel

0 votes

Hello all,

I am trying to solve a linear problem with the following code

solve(a == L, s, solver_parameters={'linear_solver': 'mumps'}

Previously, when I run the code with "python main.py", this solver runs in parallel automatically, I don't need to do anything else. But after I reinstalled my ubuntu system, the solver is not running in parallel anymore. How should I set the system up to make it run in parallel? I installed fenics following the instruction showing on official website for ubuntu. All parallel libraries were all installed.

asked Jun 21, 2017 by sin2cos FEniCS Novice (130 points)

1 Answer

0 votes

Use mpirun:

mpirun -n 4 python main.py

In this case 4 is the number of processes to run.

answered Jun 21, 2017 by johannr FEniCS Expert (17,350 points)

I know we can run fenics code with mpirun, but my problem is actually a little bit different from this. Previously, even I run with "python main.py", the solver MUMPS is also run in parallel. But now it is not the case. I just reinstalled my system and fenics with the same steps as what I did in my previous old system. I am not so sure why it is not running in parallel now, and I am curious how can I configure the mumps solver in fenics to run in paralell even without mpirun. To write code which can run under mpirun will require a little bit more skills in some cases, I am not prepared to step into that world yet.

commented Jun 21, 2017 by sin2cos FEniCS Novice (130 points)

How are you determining if it is running in parallel? In my experience, without mpirun, I only see one core being used.

commented Jun 26, 2017 by mwelland FEniCS User (8,410 points)

Sorry for answering a bit late.

I have a program which solves several pdes one after another. There is one PDE which always leads to a large linear problem since it is a mixed problem with a lot of unknowns. So I used mumps solver for this part. If you check with htop, all cores were fully occupied when this part of the code was run and it takes like 1 minutes to finish the calculation. This works very well with Fenics 2016.2.0, but recently, I reinstalled my system and updated my fenics to 2017.1.0. When I run my old code again, I found only one core is occupied and it takes about 20 minutes to solve it. In fenics 2016.2.0 with ubuntu 16.04, I am sure the mumps solver was working in parallel even you don't run with mpirun -np, now I can not figure out what has happedn with fenics 2017.1.0 or petsc, why mumps is not running in parallel by default.

commented Jul 4, 2017 by sin2cos FEniCS Novice (130 points)

Hmmm... maybe there has been a change in threading?
Take a look at here
In that case, I believe the thought was that the parallelism really came during assembly not the solver. If your code only took a minute to run, it might be hard to discern which took how long.

Looking at Change log, I see that "OpenMpAsssmebler" was removed... maybe that is the source of the change?

(Note: speculation. If there is a knowledgeable source please chime in!)

You might be SOL in avoiding running mpirun / mpiexec. Is your code a bash script? Maybe make an alias to avoid changing the script?

commented Jul 5, 2017 by mwelland FEniCS User (8,410 points)
...