[petsc-users] Optimized mode
Emmanuel Ayala
juaneah at gmail.com
Tue Sep 17 15:54:25 CDT 2019
Thanks for the tip!
Regards.
El mar., 17 de sep. de 2019 a la(s) 15:40, Mark Adams (mfadams at lbl.gov)
escribió:
> I am suspicious that he gets the exact same answer with a debug build.
>
> You might try -O2 (and -O1, and -O0, which should be the same as your
> debug build).
>
> On Tue, Sep 17, 2019 at 2:21 PM Emmanuel Ayala via petsc-users <
> petsc-users at mcs.anl.gov> wrote:
>
>> OK, thanks for the clarification! :D
>>
>> Regards.
>>
>> El mar., 17 de sep. de 2019 a la(s) 13:16, Smith, Barry F. (
>> bsmith at mcs.anl.gov) escribió:
>>
>>>
>>> In parallel the order of operations will always change with different
>>> number of processes; even with the same number the orders will be changed
>>> based on order of arrival of data from other processes; so identical runs
>>> can produce different results.
>>>
>>> Barry
>>>
>>>
>>> > On Sep 17, 2019, at 12:17 PM, Emmanuel Ayala <juaneah at gmail.com>
>>> wrote:
>>> >
>>> > Hi, thanks for the quick reply.
>>> >
>>> >
>>> > El lun., 16 de sep. de 2019 a la(s) 23:40, Smith, Barry F. (
>>> bsmith at mcs.anl.gov) escribió:
>>> >
>>> > What do you mean by 2.0e-4 units ? If you mean the last 4 digits
>>> may differ in the two solutions,
>>> >
>>> > yes, that is the meaning
>>> >
>>> > yes that is completely normal. How many digits you lose depends on the
>>> order of the operations and the condition number of the matrix and and for
>>> elasticity that will very easily be greater than 10^4
>>> >
>>> > OK. I understand that two solution can produce slightly different
>>> results. So, I guess, the optimized mode produce slightly different
>>> results because internally something change (order of some operations,
>>> maybe) when the number of process change?.
>>> >
>>> > Barry
>>> >
>>> > From:
>>> https://pdfs.semanticscholar.org/dccf/d6daa35fc9d585de1f927c58cc29c4cd0bab.pdf
>>> >
>>> > We conclude this section by noting the need for care in interpreting
>>> the forward error. Experiments in [24] show that simply changing the order
>>> of evaluation of an inner product in the substitution algorithm for
>>> solution of a triangular system can change the forward error in the
>>> computed solution by orders of magnitude. This means, for example, that it
>>> is dangerous to compare different codes or algorithms solely in terms of
>>> observed forward errors.
>>> >
>>> > Thank you very much!
>>> > Best regards.
>>> >
>>> > > On Sep 16, 2019, at 11:13 PM, Emmanuel Ayala via petsc-users <
>>> petsc-users at mcs.anl.gov> wrote:
>>> > >
>>> > > Hi everyone,
>>> > >
>>> > > I have a code for elastic linear analysis which involves DMDA,
>>> SuperLu and matrix operations. I have a MPI vector which contains the nodal
>>> displacements, when I run the code in debug mode i get the same value for
>>> the norm (2 or Inf), for different number of process. But when I use the
>>> optimized mode, I have small variations for the same norm depending on the
>>> number of process.
>>> > >
>>> > > -Debug mode: the same norm value for any number of processes.
>>> > > -Optimized mode: the norm value changes with the number of
>>> processes. The variations are around 2.0e-4 units.
>>> > >
>>> > > This is normal?
>>> > >
>>> > > For my optimized mode I used the next configuration
>>> > >
>>> > > ./configure --with-debugging=0 COPTFLAGS='-O3 -march=native
>>> -mtune=native' CXXOPTFLAGS='-O3 -march=native -mtune=native' FOPTFLAGS='-O3
>>> -march=native -mtune=native' --download-mpich --download-superlu_dist
>>> --download-metis --download-parmetis --download-cmake
>>> --download-fblaslapack=1 --with-cxx-dialect=C++11
>>> > >
>>> > > Best regards.
>>> >
>>>
>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20190917/e7fc61f6/attachment.html>
More information about the petsc-users
mailing list