[petsc-users] Convergence of AMG

[Mark Adams]

The last two cases were cut off but that is OK.

Lets just do -pc_gamg_square_graph 0 and please add/do -ksp_view -
mg_levels_ksp_type richardson -pc_mg_levels 5 -gamg_est_ksp_type cg

Send the grep GAMG again, and the ksp_view output.

And then do this again with: -pc_gamg_agg_nsmooths

I think the eigen estimators are messed up on the coarse grids.


On Tue, Oct 30, 2018 at 12:39 PM Karin&NiKo <niko.karin at gmail.com> wrote:

> Ok, sorry for my misunderstanding  and thank you for the clarification.
>
> Le mar. 30 oct. 2018 13:55, Mark Adams <mfadams at lbl.gov> a écrit :
>
>> Nicolas,
>>
>> Smoothed aggregation is fine with shells. see the original SA paper (
>> https://link.springer.com/article/10.1007/BF02238511).
>>
>> The rotational modes, which are the non-trivial modes that must be
>> supplied, are used in the interpolation.
>>
>> Mark
>>
>> On Tue, Oct 30, 2018 at 5:22 AM Karin&NiKo <niko.karin at gmail.com> wrote:
>>
>>> Manav,
>>>
>>> How are interpolated the rotational degrees of freedom? AFAIK, when
>>> using smoothed aggregation, the interpolation process tries to mimic linear
>>> interpolation, which can be OK for the displacement DOF but is not for the
>>> rotational DOF using some plate and shell formulations.
>>> This can explain poor convergence of a multilevel approach, which needs
>>> to restrict and extrapolate the unknowns. In order to check this
>>> hypothesis, you can try a test case with zero rotations.
>>>
>>> Nicolas
>>>
>>> Le lun. 29 oct. 2018 à 22:13, Mark Adams via petsc-users <
>>> petsc-users at mcs.anl.gov> a écrit :
>>>
>>>> * the two level results tell us that MG is not doing well on the coarse
>>>> grids. So the coarse grids are the problem.
>>>>
>>>> * Do not worry about timing now. Get the math correct. The two level
>>>> solve is not meant to be a solution just a diagnostic so don't try to
>>>> optimize it by squaring the graph. Use -pc_gamg_square_graph 0.
>>>>
>>>> * It looks like you don't need 4 smoothing steps but lets keep it and
>>>> we can dial it back later.
>>>>
>>>> * This table is interesting. First, you had about 12 iterations earlier
>>>> and I think your rtol was tighter than the default (so the iteration could
>>>> should go down not up). Do you know what change here?
>>>>
>>>> Note, even though -mg_levels_ksp_max_it is not in the ksp_view it does
>>>> work. It is syntactic sugar to just add it to all levels like you did
>>>> manually.
>>>>
>>>> Anyway, these number look reasonable. It is interesting that 3 levels
>>>> ran well but the 4th level ran poorly. This implies we want to slow down
>>>> coarsening on these levels, but ...
>>>>
>>>> First can you please rerun this experiment with -pc_gamg_square_graph 0.
>>>>
>>>> Also, please run with -info. This is very noisy but you can grep on
>>>> "GAMG" and send that output to us (about 15 lines).
>>>>
>>>> Thanks,
>>>> Mark
>>>>
>>>>
>>>>
>>>> On Mon, Oct 29, 2018 at 3:34 PM Manav Bhatia <bhatiamanav at gmail.com>
>>>> wrote:
>>>>
>>>>> Barry,
>>>>>
>>>>>    Here are some quick numbers with the following options on 4 CPUs
>>>>> and 543,606 dofs:
>>>>>
>>>>> -mg_levels_ksp_max_it 4 -pc_gamg_square_graph 1 -pc_gamg_threshold 0.
>>>>>
>>>>>  #levels   |    #KSP Iters
>>>>> ———————————
>>>>>      2        |       18
>>>>>      3        |       18
>>>>>      4        |       40
>>>>>      5        |       59
>>>>>
>>>>> -Manav
>>>>>
>>>>>
>>>>> On Oct 29, 2018, at 2:06 PM, Smith, Barry F. <bsmith at mcs.anl.gov>
>>>>> wrote:
>>>>>
>>>>>
>>>>>  Exactly how much does it increase with number of levels? Send a chart
>>>>> number of levels and number of iterations. With say -mg_levels_ksp_maxit 4
>>>>>
>>>>>   Thanks
>>>>>
>>>>>   Barry
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On Oct 29, 2018, at 12:59 PM, Manav Bhatia <bhatiamanav at gmail.com>
>>>>> wrote:
>>>>>
>>>>> Thanks for the clarification.
>>>>>
>>>>> I also observed that the number of KSP iterations increases with an
>>>>> increase in the levels of AMG. Is this true, in general, for all/most
>>>>> applications?
>>>>>
>>>>> -Manav
>>>>>
>>>>> On Oct 29, 2018, at 12:53 PM, Jed Brown <jed at jedbrown.org> wrote:
>>>>>
>>>>> Manav Bhatia <bhatiamanav at gmail.com> writes:
>>>>>
>>>>> Thanks, Jed.
>>>>>
>>>>> The description says: “ Square the graph, ie. compute A'*A before
>>>>> aggregating it"
>>>>>
>>>>> What is A here?
>>>>>
>>>>>
>>>>> The original matrix, or its "graph" (your 6x6 blocks condensed to
>>>>> scalars).
>>>>>
>>>>> What is the impact of setting this to 0, which led to a very
>>>>> significant increase in the CPU time in my case?
>>>>>
>>>>>
>>>>> The aggregates are formed on the connectivity of your original matrix,
>>>>> so root nodes are aggregated only with their first neighbors, resulting
>>>>> in slower coarsening.
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20181030/51a40669/attachment.html>