Dear all,
having done some runs on ARCHER now, hypre’s BoomerAMG is doing very well at lowest order. On a full node (=24 cores) I get for a problem with 81920 x 64 = 5E6 cells (all times including setup times, which are ~2/3 or the total runtime):
matrixfree, geometric multigrid: 11.15
PETSc, GAMG (old results): 13.80
PETSc, BoomerAMG: 8.41
This is for the BoomerAMG standard parameters, I haven’t attempted any tuning, since looking back through my notes from two years ago this will only give maybe 10% more performance (it could, however, make a difference for the scalbility). Note also that I use the default smoother (which is symmetric-SOR/Jacobi), as I don’t think I can use block ILU with BoomerAMG. In the tests we did in the past with hypre we did the same, i.e. use a pointwise smoother in BoomerAMG. Since the predominant coupling is in the vertical direction, AMG will first coarsen in this direction until the problem is isotropic, and then a pointwise smoother works fine. This is very different to what’s done in the tensor-product geometric multigrid, which coarses in the horizontal only and so needs the columnwise smoother. This might also explain why the GAMG (which did use ILU) gives worse performance than BoomerAMG.
Unfortunately it doesn’t look good at higher order order. While the matrixfree solver converges as before, once I use BoomerAMG the convergence rate is close to 1.0
I will try tweaking the parameters (there are lots!) to see if that helps.
Eike