[petsc-dev] why does this inanity exist?

[Lisandro Dalcin]

Initial implementation, still WIP, I'll fire the PR once I get
dalcinl/refactor-ts-adapt merged into master

https://bitbucket.org/petsc/petsc/commits/9d2aaf5d9cfa4509075a23b366eed8b8cf407a27?at=dalcinl/ts-adapt-default-type

On 30 April 2017 at 23:27, Barry Smith <bsmith at mcs.anl.gov> wrote:
>
>> On Apr 30, 2017, at 3:19 PM, Jed Brown <jed at jedbrown.org> wrote:
>>
>> Proposal: add ts->default_adapt_type.  TSCreate_* sets this.  Remove
>> TSGetAdapt from TSCreate and move it to TSSetUp and TSSetFromOptions.
>> Make TSGetAdapt call TSAdaptSetType(adapt,ts->default_adapt_type).  Now
>> there is no "setbyuser", but TS implementations can select their
>> preferences and those will be ignored if the user calls TSGetAdapt and
>> changes it themselves.  And -ts_adapt_type none always works.
>
>   I don't see any immediate flaws in this approach.
>
>>
>> Barry Smith <bsmith at mcs.anl.gov> writes:
>>
>>>> On Apr 30, 2017, at 12:53 PM, Lisandro Dalcin <dalcinl at gmail.com> wrote:
>>>>
>>>> On 30 April 2017 at 19:09, Barry Smith <bsmith at mcs.anl.gov> wrote:
>>>>>
>>>>>> On Apr 30, 2017, at 8:39 AM, Lisandro Dalcin <dalcinl at gmail.com> wrote:
>>>>>>
>>>>>> On 30 April 2017 at 05:53, Barry Smith <bsmith at mcs.anl.gov> wrote:
>>>>>>> TS already had mechanisms to turn on or off adaptivity (presumably) so why would an individual type need to reproduce this functionality except to drive users insane?
>>>>>>
>>>>>> We do not have a clear mechanism for a TS subtype to decide on whether
>>>>>> adaptivity should be turn on or off by default. The rationale to turn
>>>>>> it off by default are more or less described in other related email:
>>>>>> the error estimator is bad quality (or lower order than the one of the
>>>>>> LTE of the method).
>>>>>
>>>>>  Why isn't the default just switched when the subtype (with a bad adapter) is selected by the user?
>>>>>
>>>>>  The reason is that the user may have already manually turned on the adapter and you don't want PETSc switching it secretly later?
>>>>>
>>>>
>>>> Yes, indeed. We have the usual fight between explicit API calls in
>>>> code interacting with SetType() and SetFromOptions().
>>>>
>>>> However, after my refactoring of TSAdapt, maybe we are closer to get
>>>> something that works better.
>>>>
>>>>
>>>>>  Do we need then to have a flag in TS - adaptsetbyuser? and only change the default if that flag is not on?
>>>>>
>>>>>  I prefer not to have setbyuser flags in PETSc but sometimes it seems we cannot live without them.
>>>>>
>>>>
>>>> I don't like any setbyuser flags either.
>>>>
>>>>>>
>>>>>> Please note that THETA and ALPHA do use the standard TSAdapt
>>>>>> infrastructure, the only thing this "use adapt" flag does is to turn
>>>>>> it off by default. If you consider it appropriate, I can remove all
>>>>>> that code for the "use adapt" flag in THETA and ALPHA, users can then
>>>>>> turn off adaptivity with `-ts_adapt_type none`.
>>>>>
>>>>>  Yes, I don't understand why users should not just use -ts_adapt_type none.
>>>>>
>>>>
>>>> These "use adapt" flags were just an attempt to not surprise users
>>>> with TSTHETA changing previous behavior. However, as the error
>>>> estimator is very conservative, I don't really see a problem with
>>>> having adaptivity on by default in THETA and ALPHA.
>>>>
>>>> So Barry, if you are for it, and Jed has no objections, I can fire a
>>>> PR tomorrow removing these crappy "use adapt flags" in favor of  the
>>>> -ts_adapt_type approach.
>>>
>>>   Yes.
>>>
>>>
>>>
>>>>
>>>> @Stefano, this should fix for good your previous complaints/questions
>>>> about "How do I turn on adaptivity on these methods?".
>>>>
>>>> --
>>>> Lisandro Dalcin
>>>> ============
>>>> Research Scientist
>>>> Computer, Electrical and Mathematical Sciences & Engineering (CEMSE)
>>>> Extreme Computing Research Center (ECRC)
>>>> King Abdullah University of Science and Technology (KAUST)
>>>> http://ecrc.kaust.edu.sa/
>>>>
>>>> 4700 King Abdullah University of Science and Technology
>>>> al-Khawarizmi Bldg (Bldg 1), Office # 0109
>>>> Thuwal 23955-6900, Kingdom of Saudi Arabia
>>>> http://www.kaust.edu.sa
>>>>
>>>> Office Phone: +966 12 808-0459
>



-- 
Lisandro Dalcin
============
Research Scientist
Computer, Electrical and Mathematical Sciences & Engineering (CEMSE)
Extreme Computing Research Center (ECRC)
King Abdullah University of Science and Technology (KAUST)
http://ecrc.kaust.edu.sa/

4700 King Abdullah University of Science and Technology
al-Khawarizmi Bldg (Bldg 1), Office # 0109
Thuwal 23955-6900, Kingdom of Saudi Arabia
http://www.kaust.edu.sa

Office Phone: +966 12 808-0459