On 11/5/25 11:37, Nihar Bhardwaj Darbhamulla wrote:
In the parallel setup, I am testing with 214788 dofs, with 1, 2,4,8 and 16
processes, and the timings are attached below. As can be seen, the assembly
scales perfectly well with doubling the processes, but it is the solve which
ends up taking quite long. The direct solve is based on the following sub-
sizes of dofs:
M1 - 66049
M2 - 132098
M3 - 16641
None of these sizes are prohibitively expensive for a Direct Solver in
parallel. Given this premise, which directions should I potentially probe into.
Nihar:
without having looked at your code in detail:
* These are very small problems. I would not expect that a parallel solver can
provide any kind of speedup on a problem this small -- you will be spending
all of your time in communication and none in actual computations. You'll want
to test with much larger problems -- a good rule of thumb is that parallel
solvers only scale if you have perhaps 50k or 100k unknowns per MPI process;
you should apply that to the smallest of the blocks you have, M3. So you'll
have to make it at least ten times larger to even be relevant to 2 processes.
* I believe that in your code, you compute the sparse direct LU decomposition
every time the vmult() function is called for the preconditioner. That's
inefficient. You will want to compute the decomposition only once, in the
constructor of the DirectInverseMatrix, and then only *apply* the
decomposition in each call to vmult.
* The results you show indicate that your method leads to a roughly constant
number of GMRES iterations. That's a good start. In order to find out which
part of your implementation is leading to the growth in run time, put timer
sections into the various parts of your preconditioner so that you get
finer-grained information about your algorithm's individual components.
Best
W.
--
------------------------------------------------------------------------
Wolfgang Bangerth email: [email protected]
www: http://www.math.colostate.edu/~bangerth/
--
The deal.II project is located at http://www.dealii.org/
For mailing list/forum options, see
https://groups.google.com/d/forum/dealii?hl=en
---
You received this message because you are subscribed to the Google Groups "deal.II User Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To view this discussion visit
https://groups.google.com/d/msgid/dealii/8ce37dcb-e4e0-4090-bf05-aefa87d9dc62%40colostate.edu.
