[petsc-dev] GAMG and custom MatMults in smoothers

[Matthew Knepley]

On Mon, Jul 2, 2018 at 8:28 AM Vaclav Hapla <vaclav.hapla at erdw.ethz.ch>
wrote:

> 2. 7. 2018 v 15:05, Matthew Knepley <knepley at gmail.com>:
>
> On Mon, Jul 2, 2018 at 7:54 AM Vaclav Hapla <vaclav.hapla at erdw.ethz.ch>
> wrote:
>
>>
>>
>> 2. 7. 2018 v 14:48, Matthew Knepley <knepley at gmail.com>:
>>
>> On Mon, Jul 2, 2018 at 3:48 AM Vaclav Hapla <vaclav.hapla at erdw.ethz.ch>
>> wrote:
>>
>>> Barry wrote:
>>>
>>>   This could get ugly real fast, for example, for vector operations,
>>> there may be dozens of named vectors and each one gets its own logging?
>>> You'd have to make sure that only the objects you care about get named, is
>>> that possible?
>>>
>>>    I don't know if there is a good solution within the PETSc logging
>>> infrastructure to get what you want but maybe what you propose is the best
>>> possible.
>>>
>>>
>>> As I suggest, this behavior would be only triggered by a specific option.
>>>
>>> I think there are actually 4 strings which could be used as an event
>>> name suffix in log view:
>>> 1) name
>>> 2) prefix
>>> 3) type
>>> 4) custom string (set by something like PetscObjectSetLogViewSuffix)
>>> I think the best would be to let user choose by offering
>>> -log_view_by_{name,prefix,type,suffix}.
>>>
>>> For example, with -log_view_by_prefix, you could readily distinguish
>>> PCTelescope outer and inner apply, because you would see a separate
>>> "PCApply (telescope_)" event.
>>> With -log_view_by_type, you would see PCApply (telescope).
>>>
>>> I think this would be useful because the current class-wide events like
>>> MatMult or PCApply aggregate very different operations from which some are
>>> for free and some form hotspots.
>>>
>>>
>>> Stefano wrote:
>>>
>>> The issue with this sort of “dynamic” logging is that now PETSc requires
>>> PetscLogEvent created during the registration of the class, so that all the
>>> ranks in PETSC_COMM_WORLD have the same events registered.
>>> What you propose is not generally supported for this specific reason.
>>>
>>> Your “log_name” may work if users register their own classes (with their
>>> own LogEvents created properly), and currently we don’t have support (maybe
>>> I’m wrong) to add an “InitializePackage” method for the users’ registered
>>> classes.
>>>
>>>
>>> I don't agree. What I suggest is basically an ability to allow
>>> automatically created object-wise events, so it _can't_ be managed during
>>> the class registration. In presence of respective option, the event would
>>> be created during PetscLogEventBegin by taking the class-wide event's name,
>>> concatenating the suffix and registering a new event. The event id would be
>>> stored in the PetscObject structure.
>>>
>>>
>>> Matt wrote:
>>>
>>> As people have pointed out, this would not work well for Events.
>>> However, this is exactly what stages are for.
>>> Use separate stages for the different types of MatMult. I did this, for
>>> example, when looking at performance
>>> on different MG levels.
>>>
>>>
>>> Yes, performance on different MG levels is a nice use case. I don't
>>> understand how you inject stages into MatMults. To me it's exactly the same
>>> problem as with events - you have to define MatMult_custom where you take
>>> the original mult and wrap into PetscStageLogPush/Pop and then use
>>> MatSetOperation to redefine MatMult. Or do you mean something more elegant?
>>>
>>
>> You could do that, but usually I think of stages as being structural. I
>> think for your example I would push/pop the stage
>> inside your Mat operation wrapper (I don't see why you need another one),
>> and this behavior could be controlled with
>> another option so you could turn it off.
>>
>>
>> I meant hierarchies of typically Mats or PCs, where you don't define any
>> custom operations but compose together existing types (which should be
>> promoted I believe). So no "my" wrapper. As I wrote below:
>>
>>  Think e.g. of having additive MATCOMPOSITE wrapping multiplicative
>>> MATCOMPOSITE wrapping MATTRANSPOSE wrapping MATAIJ. You want to measure
>>> this MATAIJ instance's MatMult separately but you surely don't want to
>>> rewrite implementation of MatMult_Transpose or force yourself to use
>>> MATSHELL just to hang the events on MatMult*.
>>>
>>>
> Its not enough to make separate stages for additive MC, multiplicative MC,
> and MT? If you want stages for every single
> combination created dynamically, you can push another stage when each of
> these combinations is created using GetTag()
> or something like that. You could switch between these behaviors with an
> option.
>
>
> I'm not sure I understand. Do you mean registering and pushing/popping
> these stages in the user's code? You can surely call PetscStageLogRegister
> somewhere after PetscInitialize, but where do you place your
> PetscStageLogPush/Pop calls?
>

No. You would create a stage when the MATCOMPOSITE is created (or once when
any MATCOMPOSITE is created), and push/pop
on application.

  Matt


> Thanks
>
> Vaclav
>
>
> The reason I think this is preferable is that we do not mess with any
> logging infrastructure, we just use stages inside of other objects.
>
>   Thanks,
>
>      Matt
>
>
>> Thanks
>>
>> Vaclav
>>
>>
>>
>>   Matt
>>
>>
>>> Thanks
>>>
>>> Vaclav
>>>
>>>
>>>
>>> 29. 6. 2018 v 22:42, Smith, Barry F. <bsmith at mcs.anl.gov>:
>>>
>>>
>>>
>>> On Jun 29, 2018, at 9:33 AM, Vaclav Hapla <vaclav.hapla at erdw.ethz.ch>
>>> wrote:
>>>
>>>
>>>
>>> 22. 6. 2018 v 17:47, Smith, Barry F. <bsmith at mcs.anl.gov>:
>>>
>>>
>>>
>>> On Jun 22, 2018, at 5:43 AM, Pierre Jolivet <pierre.jolivet at enseeiht.fr>
>>> wrote:
>>>
>>> Hello,
>>> I’m solving a system using a MATSHELL and PCGAMG.
>>> The MPIAIJ Mat I’m giving to GAMG has a specific structure (inherited
>>> from the MATSHELL) I’d like to exploit during the solution phase when the
>>> smoother on the finest level is doing MatMults.
>>>
>>> Is there some way to:
>>> 1) decouple in -log_view the time spent in the MATSHELL MatMult and in
>>> the smoothers MatMult
>>>
>>>
>>> You can register a new event and then inside your MATSHELL MatMult()
>>> call PetscLogEventBegin/End on your new event.
>>>
>>>  Note that the MatMult() like will still contain the time for your
>>> MatShell mult so you will need to subtract it off to get the time for your
>>> non-shell matmults.
>>>
>>>
>>> In PERMON, we sometimes have quite complicated hierarchy of wrapped
>>> matrices and want to measure MatMult{,Transpose,Add,TransposeAdd}
>>> separately for particular ones. Think e.g. of having additive MATCOMPOSITE
>>> wrapping multiplicative MATCOMPOSITE wrapping MATTRANSPOSE wrapping MATAIJ.
>>> You want to measure this MATAIJ instance's MatMult separately but you
>>> surely don't want to rewrite implementation of MatMult_Transpose or force
>>> yourself to use MATSHELL just to hang the events on MatMult*.
>>>
>>> We had a special wrapper type just adding some prefix to the events for
>>> the given object but this is not nice. What about adding a functionality to
>>> PetscLogEventBegin/End that would distinguish based on the first
>>> PetscObject's name or option prefix? Of course optionally not to break guys
>>> relying on current behavior - e.g. under something like -log_view_by_name.
>>> To me it's quite an elegant solution working for any PetscObject and any
>>> event.
>>>
>>>
>>>   This could get ugly real fast, for example, for vector operations,
>>> there may be dozens of named vectors and each one gets its own logging?
>>> You'd have to make sure that only the objects you care about get named, is
>>> that possible?
>>>
>>>    I don't know if there is a good solution within the PETSc logging
>>> infrastructure to get what you want but maybe what you propose is the best
>>> possible.
>>>
>>>   Barry
>>>
>>>
>>> I can do that if I get some upvotes.
>>>
>>> Vaclav
>>>
>>>
>>> 2) hardwire a specific MatMult implementation for the smoother on the
>>> finest level
>>>
>>>
>>> In the latest release you do MatSetOperation() to override the normal
>>> matrix vector product with anything else you want.
>>>
>>>
>>> Thanks in advance,
>>> Pierre
>>>
>>> PS : here is what I have right now,
>>> MatMult              118 1.0 1.0740e+02 1.6 1.04e+13 1.6 1.7e+06 6.1e+05
>>> 0.0e+00 47100 90 98  0 47100 90 98  0 81953703
>>> […]
>>> PCSetUp                2 1.0 8.6513e+00 1.0 1.01e+09 1.7 2.6e+05 4.0e+05
>>> 1.8e+02  5  0 14 10 66   5  0 14 10 68 94598
>>> PCApply               14 1.0 8.0373e+01 1.1 9.06e+12 1.6 1.3e+06 6.0e+05
>>> 2.1e+01 45 87 72 78  8 45 87 72 78  8 95365211 // I’m guessing a lot of
>>> time here is being wasted in doing inefficient MatMults on the finest level
>>> but this is only speculation
>>>
>>> Same code with -pc_type none -ksp_max_it 13,
>>> MatMult               14 1.0 1.2936e+01 1.7 1.35e+12 1.6 2.0e+05 6.1e+05
>>> 0.0e+00 15100 78 93  0 15100 78 93  0 88202079
>>>
>>> The grid itself is rather simple (two levels, extremely aggressive
>>> coarsening),
>>> type is MULTIPLICATIVE, levels=2 cycles=v
>>> KSP Object: (mg_coarse_) 1024 MPI processes
>>> linear system matrix = precond matrix:
>>>   Mat Object: 1024 MPI processes
>>>     type: mpiaij
>>>     rows=775, cols=775
>>>     total: nonzeros=1793, allocated nonzeros=1793
>>>
>>> linear system matrix followed by preconditioner matrix:
>>> Mat Object: 1024 MPI processes
>>> type: shell
>>> rows=1369307136, cols=1369307136
>>> Mat Object: 1024 MPI processes
>>> type: mpiaij
>>> rows=1369307136, cols=1369307136
>>> total: nonzeros=19896719360, allocated nonzeros=19896719360
>>>
>>>
>>>
>>
>> --
>> What most experimenters take for granted before they begin their
>> experiments is infinitely more interesting than any results to which their
>> experiments lead.
>> -- Norbert Wiener
>>
>> https://www.cse.buffalo.edu/~knepley/ <http://www.caam.rice.edu/~mk51/>
>>
>>
>>
>
> --
> What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> -- Norbert Wiener
>
> https://www.cse.buffalo.edu/~knepley/ <http://www.caam.rice.edu/~mk51/>
>
>
>

-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener

https://www.cse.buffalo.edu/~knepley/ <http://www.caam.rice.edu/~mk51/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20180702/e16b0c42/attachment-0001.html>