[gerbervdgraaf]

I have read that the MUltifrontal Massively Parallel sparse direct Solver (MUMPS) from http://mumps.enseeiht.fr/ might increase drasticly the calculation speed. Searching on this and related forums I did not find much about MUMPS to be used by OpenFOAM. Did anybody have a look at this? Is it worthwhile to consider MUMPS in the OpenFOAM environment and is it possible ot implement using the API for the C programming language?

Thanks,
Gerber

[gerbervdgraaf]

So far I have not received any comments, yet. I'd like to extend my question to other, similar, solvers like SuperLU. Is there any experience with these techniques in OpenFOAM? So far I have read that Mumps is faster for less than 50 - 60 CPU nodes.

Gerber van der Graaf

[Bocliff]

Gerber,

Multifrontal solvers are common in FEM and they are quite efficient version of Gauss elimination algorithm. These solvers are not iterative and not very suitable for non-linear systems of equations that are pervasive in CFD. The same goes for any other LU factorization algorithms. They are not used is CFD commonly and OF has no implementation of them that I know of.

Bocliff

[gerbervdgraaf]

Thanks for your reply. We are working here at UPC (Univ of Catalunya, Spain) with another Finite Volume CFD code (home brewed at our university) that uses MUMPS. It is much faster. Whether the results are as reliable / accurate as OpenFOAM remains to be seen.

[mirko]

Originally Posted by Bocliff

Gerber,

Multifrontal solvers are common in FEM and they are quite efficient version of Gauss elimination algorithm. These solvers are not iterative and not very suitable for non-linear systems of equations that are pervasive in CFD. The same goes for any other LU factorization algorithms. They are not used is CFD commonly and OF has no implementation of them that I know of.

Bocliff

I am new to this parallelisation business. Can you recommend some references where I can learn more about parallelism in CFD?

Thanks

[gerbervdgraaf]

Concerning Mumps the main web site is:
http://mumps.enseeiht.fr
The UserGuide can be obtained from the Documentation link

Concerning Scotch:
http://www.labri.fr/perso/pelegrin/scotch/

[arjun]

Originally Posted by gerbervdgraaf

Thanks for your reply. We are working here at UPC (Univ of Catalunya, Spain) with another Finite Volume CFD code (home brewed at our university) that uses MUMPS. It is much faster. Whether the results are as reliable / accurate as OpenFOAM remains to be seen.

with multifrontal solvers the main issue is memory as far i understand. Polyflow of ansys uses it and we have given up polyflow because we could not run anything seriously large with it.. Plus for large cases they were slow too.

They are very robust though.

As far as your question about OF goes, this olny OF developers can answer.

I personally definitely would like to learn ,understand and implement them they will be good for visco-elastic applications.

[mirko]

Originally Posted by gerbervdgraaf

Concerning Mumps the main web site is:
http://mumps.enseeiht.fr
The UserGuide can be obtained from the Documentation link

Concerning Scotch:
http://www.labri.fr/perso/pelegrin/scotch/

Thank you for the links

[amuzeshi]

Originally Posted by gerbervdgraaf

I have read that the MUltifrontal Massively Parallel sparse direct Solver (MUMPS) from http://mumps.enseeiht.fr/ might increase drasticly the calculation speed. Searching on this and related forums I did not find much about MUMPS to be used by OpenFOAM. Did anybody have a look at this? Is it worthwhile to consider MUMPS in the OpenFOAM environment and is it possible ot implement using the API for the C programming language?

Thanks,
Gerber

OpenFOAM library contains only iterative linear solvers implementations; however, it is possible to use direct solvers by linking with external libraries, such as Eigen , MUMPS , PETSc or Trilinos .

Ref: P. Cardiff et al., “An open-source finite volume toolbox for solid mechanics and fluid-solid interaction simulations,” arXiv:1808.10736 [physics], Aug. 2018.