Thank you very much for the fast answers and for the insight.
Indeed, I tried changing Richardson with Chebyshev (while keeping fixed all the rest), and the method did converge in 39 iterations. Richards and GMRES instead has a very slow convergence as you were expecting (around one thousand iterations for a four digit relative residual, at that point I terminated it), even if it had come to convergence it would have been a waste for an SPD system. Best, Fabio Il 18/03/22 23:14, Mark Adams ha scritto:
MG is not stable with richardson/jacobi, or at least it is almost not stable. Yea I would guess gmres will go forever because it does not care. I think CG exists when it gets a negative number for beta or whatever and it was probably -eps. That is my guess. On Fri, Mar 18, 2022 at 3:17 PM Barry Smith <[email protected]> wrote: The GAMG produced an indefinite operator. I don't know if there is a way to detect why this happened or how to stop it. You can try -ksp_type gmres and see how that goes since it can handle an indefinite preconditioner. > On Mar 18, 2022, at 11:53 AM, Fabio Durastante <[email protected]> wrote: > > For the default case: > > Linear solve converged due to CONVERGED_ATOL iterations 14 > > for the other it tells > > Linear solve did not converge due to DIVERGED_INDEFINITE_PC iterations 2 > > that again seems a bit strange to me, since this should be a symmetric V-cycle built on smoothed aggregation that should be definite (and symmetric). > > Fabio > > Il 18/03/22 16:47, Barry Smith ha scritto: >> Run with -ksp_converged_reason to have it print why it has stopped the iteration. >> >> >>> On Mar 18, 2022, at 11:44 AM, Fabio Durastante <[email protected]> wrote: >>> >>> Hi everybody, >>> >>> I'm trying to run the rotated anisotropy example ex54f using CG and GAMG as preconditioner, I run it with the command: >>> >>> mpirun -np 2 ./ex54f -ne 1011 \ >>> -theta 18.0 \ >>> -epsilon 100.0 \ >>> -pc_type gamg \ >>> -pc_gamg_type agg \ >>> -log_view \ >>> -log_trace \ >>> -ksp_view \ >>> -ksp_monitor \ >>> -ksp_type cg \ >>> -mg_levels_pc_type jacobi \ >>> -mg_levels_ksp_type richardson \ >>> -mg_levels_ksp_max_it 4 \ >>> -ksp_atol 1e-9 \ >>> -ksp_rtol 1e-12 >>> >>> But the KSP CG seems to stop just after two iterations: >>> >>> 0 KSP Residual norm 6.666655711717e-02 >>> 1 KSP Residual norm 9.859661350927e-03 >>> >>> I'm attaching the full log, the problem seems to appear when I modify the value of epsilon, if I leave it to the default (1.0) it prints >>> >>> 0 KSP Residual norm 5.862074869050e+00 >>> 1 KSP Residual norm 5.132711016122e-01 >>> 2 KSP Residual norm 1.198566629717e-01 >>> 3 KSP Residual norm 1.992885901625e-02 >>> 4 KSP Residual norm 4.919780086064e-03 >>> 5 KSP Residual norm 1.417045143681e-03 >>> 6 KSP Residual norm 3.559622318760e-04 >>> 7 KSP Residual norm 9.270786187701e-05 >>> 8 KSP Residual norm 1.886403709163e-05 >>> 9 KSP Residual norm 2.940634415714e-06 >>> 10 KSP Residual norm 5.015043022637e-07 >>> 11 KSP Residual norm 9.760219712757e-08 >>> 12 KSP Residual norm 2.320857464659e-08 >>> 13 KSP Residual norm 4.563772507631e-09 >>> 14 KSP Residual norm 8.896675476997e-10 >>> >>> that is very strange because the case with epsilon 1 should be easier. >>> >>> Any help with this would be great. >>> >>> Thank you very much, >>> >>> Fabio Durastante >>> <log_ex54f.txt>
