[aspect-devel] Using MUMPS through Trillions instead of PETSc

John Naliboff jbnaliboff at ucdavis.edu
Tue Feb 21 12:34:11 PST 2017

Hi all,

I recently tried a few models (similar to continental extension cookbook) with PETSc instead of Trillions.  

In the first test case I used a direct solver (MUMPS) for the stokes system and this resulted in quite a few less non-linear iterations (8 instead of ~ 50 on first time step) than with the iterative solver via Trillions.  Unfortunately, the model crashed on the second time step during the compositional field advection.

In the second test case with PETSc I used an iterative solve for the stokes system and the model crashed during the first stokes solve.  

I can provide more detail on the exact errors if this is of interest, but my interest in using PETSc was in fact to try out MUMPS.  Based on the first test case, it seems like it might be worth testing MUMPS further but not through PETSc.  

I saw deal.II has the capability to use MUMPS through Trillions through the SolverDirect class:
https://www.dealii.org/8.4.1/doxygen/deal.II/classTrilinosWrappers_1_1SolverDirect.html <https://www.dealii.org/8.4.1/doxygen/deal.II/classTrilinosWrappers_1_1SolverDirect.html>

However, the few discussions I’ve seen regarding using MUMPS through Trillions are not entirely encouraging :)
https://github.com/dealii/dealii/issues/409 <https://github.com/dealii/dealii/issues/409>
https://groups.google.com/forum/#!topic/dealii/inrgGe7tSFw <https://groups.google.com/forum/#!topic/dealii/inrgGe7tSFw>

So, my question is whether looking further into using MUMPS through Trillions is a rabbit hole I should be heading down.

In reality, my interest is not explicitly with MUMPS but rather a parallel direct solver option (MUMPS, PARDISO, etc).


John Naliboff
Assistant Project Scientist, CIG
Earth & Planetary Sciences Dept., UC Davis

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.geodynamics.org/pipermail/aspect-devel/attachments/20170221/eaedee0f/attachment.html>

More information about the Aspect-devel mailing list