Hi folks, I'm trying to solve a 3D problem in parallel using an AMG preconditioner, but the performance is bad. I'm wondering if I can get some advice from someone who has experience choosing and tuning multigrid preconditioners, particularly for problems in 3D.
For context: - The problem is vector-valued (5 components) and elliptic(ish) - I use GMRES w/BoomerAMG preconditioner for asymmetric matrices, everything is default - In 2D, scaling is good even down to 5,000 DoFs per processor, and number of iterations is independent of problem size - In 3D the scaling is bad. Adding more processors after about 50,000 DoFs/processor actually slows the program down and sometimes gives memory errors. - Taking away the preconditioner in 3D gives a ~20x speedup at 16 processors, and strong scaling is linear. However, the number of iterations increases with problem size. Given all that, I have some questions: 1. Why might it be that the memory and wall-time scaling is so bad in 3D? 2. Are there any examples lying around deal.II of folks using multigrid for 3D problems in parallel? All the tutorials that I looked at were in 2D, including the 2 examples used in the distributed computing paper. 3. Might there be an easy way to fix it while still using BoomerAMG? I know the Hypre documentation <https://hypre.readthedocs.io/en/latest/solvers-boomeramg.html> gives some parameter recommendations, but I'm not so sure (1) how to set those options via deal.II (I think they are not available via the AdditionalData interface), or (2) whether those will work. Does anyone have experience with this? 4. Might the Trilinos AMG preconditioner work any better for this problem by default? And if not, is there a systematic way to tune it (particularly using a deal.II interface) to work better for a 3D problem? 5. Might the in-house GMG methods work better? And if so, do the matrix-free methods stand a chance of performing better even if I have to use some complicated functions in the matrix assembly (for my weak form I have a 5-component nonlinear function of my solution vector which has to be inverted via Newton's method across my domain). If more context on the particular problem would be helpful I can certainly give details. Mostly I'm looking for intuition, general suggestions, or pointers to good references. Any help is much appreciated. Kind regards, Lucas -- 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 dealii+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/dealii/7bcfdab4-ea5e-47ea-82f5-c78bc9e3ffd8n%40googlegroups.com.
