[aspect-devel] Direct Solver in Aspect

[Wolfgang Bangerth]

On 05/12/2014 04:56 PM, Timo Heister wrote:
>> Does it work in parallel?
>
> It uses Amesos::Klu which according to the documentation doesn't work
> with more than process, but I used it like that before (on one machine
> though). Oops. Not sure what it does then. Amesos also has a parallel
> direct solver (paraklete). Maybe it picks that one automatically? If
> not, we could change our interface of SolverDirect.

Good question. There seem to be a number of options for parallel direct solves:
   http://trilinos.org/docs/r11.8/packages/amesos/doc/html/index.html
Amesos2 also seems to support SuperLU-MT and SuperLU-Dist:
   http://trilinos.sandia.gov/packages/amesos2/
I have no idea how difficult it would be to support any of this (I don't even 
recall who wrote the current interfaces...)


> On a single workstation it might be worth running with one mpi task
> and multi-threading for assembly and a direct solver, though.

Yes, that occurred to me as well after I sent the last email. We currently 
drive the MPI_InitFinalize class with only a single thread, but maybe we 
should be using
   n_cores/n_mpi_jobs_on_this_host
instead? This would give us exactly the behavior you describe, while reverting 
back to the current state on machines that we fully load with MPI processes...

W.


-- 
------------------------------------------------------------------------
Wolfgang Bangerth               email:            bangerth at math.tamu.edu
                                 www: http://www.math.tamu.edu/~bangerth/