Periodic boundary conditions demo fails with Epetra backend

[Anders Logg]

Unit tests and demos fail to run if Trilinos is installed but not PETSc:

http://www.fenicsproject.org:8080/builders/dolfin-linux64-exp/builds/917/steps/dolfin%20check/logs/demo.log
http://www.fenicsproject.org:8080/builders/dolfin-linux64-exp/builds/917/steps/dolfin%20check/logs/stdio

[Anders Logg (logg)]

Looks like the only remaining thing is periodic BC not working with Trilinos. From running the periodic demo:

Error: Unable to add block of values to Epetra matrix.
*** Reason: Did not manage to perform Epetra_FECrsMatrix::SumIntoGlobalValues.
*** Where: This error was encountered inside EpetraMatrix.cpp.

[Andre Massing (massing)]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 11/17/2011 09:40 PM, Anders Logg wrote:
> Looks like the only remaining thing is periodic BC not working
> with Trilinos. From running the periodic demo:
>
> Error: Unable to add block of values to Epetra matrix. *** Reason:
> Did not manage to perform Epetra_FECrsMatrix::SumIntoGlobalValues.
> *** Where: This error was encountered inside EpetraMatrix.cpp.
>

I wrote a somewhat lengthy email about that, explaining what goes
wrong and asking for opinons... you have even replied to it :)

The bottom line is that there are several related issues:
1. The sparsity pattern does not consider PeriodicBC thus there will
be an error in Trilinos when copying a slave row to a master row since
the matrix is filled / finalized.

2. Dirichlet and PeriodicBC calling and matrix.apply() while applying
bc. Every such call triggers a finalization of the matrix atm.

3. https://bugs.launchpad.net/dolfin/+bug/891452
(which for some reason does not show up in the bug overview, neither
has been sent to the list, strange.)

See the original post
https://bugs.launchpad.net/dolfin/+bug/864741/comments/2

for a more detailed explanation.
The point is that it seems like a more fundamental problem where I
haven't dared to dirty-hack around without opinions from the devs
/maintainers :)

- --
Andre

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.17 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iQEcBAEBAgAGBQJOxXuoAAoJEA79ggnbq9dm0ToIAJYp09Rkq9midizQnt8zw4gr
197meLuTEuBn7fN4FR9EV3yBqhAwg7Ay/4UW96Da9mt09JoY+CRKEWyzBVhD0TC9
jcDw7aroXZVJMZ9KgHWKiwyBA/7l+4SRXen6DXAneEKIfBJGnNFiul3HTgjBxFdR
ZG7sMELsvynDM7ryt2FpZl89qI34YnOwCNj9vZAXNe8TWA4qsHmc7BTmBWojLYBS
eStuEd73652epXBOmmQHZ47fZg83SxMUJgPkPbDKxgHBaM0sp9Q2X+f1lu5uY/wF
4pIOPIfb8bSA/90Ud3JBIinin43gW4GfL661DzldNlki8kbU5Zzob97wOnCHOpI=
=l8m5
-----END PGP SIGNATURE-----

[Anders Logg (logg)]

On Thu, Nov 17, 2011 at 10:24:56PM +0100, Andre Massing wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
> On 11/17/2011 09:40 PM, Anders Logg wrote:
> > Looks like the only remaining thing is periodic BC not working
> > with Trilinos. From running the periodic demo:
> >
> > Error: Unable to add block of values to Epetra matrix. *** Reason:
> > Did not manage to perform Epetra_FECrsMatrix::SumIntoGlobalValues.
> > *** Where: This error was encountered inside EpetraMatrix.cpp.
> >
>
> I wrote a somewhat lengthy email about that, explaining what goes
> wrong and asking for opinons... you have even replied to it :)

I know and I've read it carefully!

I think I responded and said that this needs to be addressed by
incorporating boundary conditions into the FunctionSpace by
implementing constrained spaces. I even linked to a blueprint.

Bottom line: we won't fix this now so I'll add a check in that demo
that it doesn't work with Epetra.

--
Anders

> The bottom line is that there are several related issues:
> 1. The sparsity pattern does not consider PeriodicBC thus there will
> be an error in Trilinos when copying a slave row to a master row since
> the matrix is filled / finalized.
>
> 2. Dirichlet and PeriodicBC calling and matrix.apply() while applying
> bc. Every such call triggers a finalization of the matrix atm.
>
> 3. https://bugs.launchpad.net/dolfin/+bug/891452
> (which for some reason does not show up in the bug overview, neither
> has been sent to the list, strange.)
>
> See the original post
> https://bugs.launchpad.net/dolfin/+bug/864741/comments/2
>
> for a more detailed explanation.
> The point is that it seems like a more fundamental problem where I
> haven't dared to dirty-hack around without opinions from the devs
> /maintainers :)

> - --
> Andre
>
>
>
>
>
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v2.0.17 (GNU/Linux)
> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
>
> iQEcBAEBAgAGBQJOxXuoAAoJEA79ggnbq9dm0ToIAJYp09Rkq9midizQnt8zw4gr
> 197meLuTEuBn7fN4FR9EV3yBqhAwg7Ay/4UW96Da9mt09JoY+CRKEWyzBVhD0TC9
> jcDw7aroXZVJMZ9KgHWKiwyBA/7l+4SRXen6DXAneEKIfBJGnNFiul3HTgjBxFdR
> ZG7sMELsvynDM7ryt2FpZl89qI34YnOwCNj9vZAXNe8TWA4qsHmc7BTmBWojLYBS
> eStuEd73652epXBOmmQHZ47fZg83SxMUJgPkPbDKxgHBaM0sp9Q2X+f1lu5uY/wF
> 4pIOPIfb8bSA/90Ud3JBIinin43gW4GfL661DzldNlki8kbU5Zzob97wOnCHOpI=
> =l8m5
> -----END PGP SIGNATURE-----

[Garth Wells (garth-wells)]

On 17 November 2011 21:37, Anders Logg <email address hidden> wrote:
> On Thu, Nov 17, 2011 at 10:24:56PM +0100, Andre Massing wrote:
>> -----BEGIN PGP SIGNED MESSAGE-----
>> Hash: SHA1
>>
>> On 11/17/2011 09:40 PM, Anders Logg wrote:
>> > Looks like the only remaining thing is periodic BC not working
>> > with Trilinos. From running the periodic demo:
>> >
>> > Error:  Unable to add block of values to Epetra matrix. *** Reason:
>> > Did not manage to perform Epetra_FECrsMatrix::SumIntoGlobalValues.
>> > *** Where:  This error was encountered inside EpetraMatrix.cpp.
>> >
>>
>> I wrote a somewhat lengthy email about that, explaining what goes
>> wrong and asking for opinons... you have even replied to it :)
>
> I know and I've read it carefully!
>
> I think I responded and said that this needs to be addressed by
> incorporating boundary conditions into the FunctionSpace by
> implementing constrained spaces. I even linked to a blueprint.
>

A blueprint on this has been about for coming up two years:

   https://blueprints.launchpad.net/dolfin/+spec/bcs-in-sparsitypattern

Garth

> Bottom line: we won't fix this now so I'll add a check in that demo
> that it doesn't work with Epetra.
>
> --
> Anders
>
>
>> The bottom line is that there are several related issues:
>> 1. The sparsity pattern does not consider PeriodicBC thus there will
>> be an error in Trilinos when copying a slave row to a master row since
>> the matrix is filled / finalized.
>>
>> 2. Dirichlet and PeriodicBC calling and matrix.apply() while applying
>> bc. Every such call triggers a finalization of the matrix atm.
>>
>> 3. https://bugs.launchpad.net/dolfin/+bug/891452
>> (which for some reason does not show up in the  bug overview, neither
>> has been sent to the list, strange.)
>>
>> See the original post
>> https://bugs.launchpad.net/dolfin/+bug/864741/comments/2
>>
>> for a more detailed explanation.
>> The point is that it seems like a more fundamental problem where I
>> haven't dared to dirty-hack around without opinions from the devs
>> /maintainers :)
>
>
>> - --
>> Andre
>>
>>
>>
>>
>>
>> -----BEGIN PGP SIGNATURE-----
>> Version: GnuPG v2.0.17 (GNU/Linux)
>> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
>>
>> iQEcBAEBAgAGBQJOxXuoAAoJEA79ggnbq9dm0ToIAJYp09Rkq9midizQnt8zw4gr
>> 197meLuTEuBn7fN4FR9EV3yBqhAwg7Ay/4UW96Da9mt09JoY+CRKEWyzBVhD0TC9
>> jcDw7aroXZVJMZ9KgHWKiwyBA/7l+4SRXen6DXAneEKIfBJGnNFiul3HTgjBxFdR
>> ZG7sMELsvynDM7ryt2FpZl89qI34YnOwCNj9vZAXNe8TWA4qsHmc7BTmBWojLYBS
>> eStuEd73652epXBOmmQHZ47fZg83SxMUJgPkPbDKxgHBaM0sp9Q2X+f1lu5uY/wF
>> 4pIOPIfb8bSA/90Ud3JBIinin43gW4GfL661DzldNlki8kbU5Zzob97wOnCHOpI=
>> =l8m5
>> -----END PGP SIGNATURE-----
>
> --
> You received this bug notification because you are a member of DOLFIN
> Core Team, which is subscribed to DOLFIN.
> https://bugs.launchpad.net/bugs/891716
>
> Title:
>  Unit tests fail to run if Trilinos is installed but not PETSc
>
> To manage notifications about this bug go to:
> https://bugs.launchpad.net/dolfin/+bug/891716/+subscriptions
>

[Andre Massing (massing)]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 11/17/2011 10:37 PM, Anders Logg wrote:
> On Thu, Nov 17, 2011 at 10:24:56PM +0100, Andre Massing wrote: On
> 11/17/2011 09:40 PM, Anders Logg wrote:
>>>> Looks like the only remaining thing is periodic BC not
>>>> working with Trilinos. From running the periodic demo:
>>>>
>>>> Error: Unable to add block of values to Epetra matrix. ***
>>>> Reason: Did not manage to perform
>>>> Epetra_FECrsMatrix::SumIntoGlobalValues. *** Where: This
>>>> error was encountered inside EpetraMatrix.cpp.
>>>>
>
> I wrote a somewhat lengthy email about that, explaining what goes
> wrong and asking for opinons... you have even replied to it :)
>
>> I know and I've read it carefully!

I was just a bit confused about the implied desired action :)

>
>> I think I responded and said that this needs to be addressed by
>> incorporating boundary conditions into the FunctionSpace by
>> implementing constrained spaces. I even linked to a blueprint.

I have seen that, nice. Marie had already registered on at
https://blueprints.launchpad.net/dolfin/+spec/constrained-spaces
so these can be merged or the older one can be marked as superseded?

>
>> Bottom line: we won't fix this now so I'll add a check in that
>> demo that it doesn't work with Epetra.

Ok, I had the same feeling...But you could also use crisscross grids
to fix it (just kidding ... )

- --
Andre

>
>> -- Anders
>
>
> The bottom line is that there are several related issues: 1. The
> sparsity pattern does not consider PeriodicBC thus there will be an
> error in Trilinos when copying a slave row to a master row since
> the matrix is filled / finalized.
>
> 2. Dirichlet and PeriodicBC calling and matrix.apply() while
> applying bc. Every such call triggers a finalization of the matrix
> atm.
>
> 3. https://bugs.launchpad.net/dolfin/+bug/891452 (which for some
> reason does not show up in the bug overview, neither has been sent
> to the list, strange.)
>
> See the original post
> https://bugs.launchpad.net/dolfin/+bug/864741/comments/2
>
> for a more detailed explanation. The point is that it seems like a
> more fundamental problem where I haven't dared to dirty-hack around
> without opinions from the devs /maintainers :)
>
>
>
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.17 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iQEcBAEBAgAGBQJOxYJAAAoJEA79ggnbq9dmTJkH/RrNfQpSUT3XOd7jy+sqsQtW
AnXpoW9k8EfAk6UO9NpEAge7mt808EVZKR7PaTQw9G34twSunz63Fze08agIpLYY
EWJx7mrubKuvl6DRkmJcsCh1FAeHw6mLHugnVizLyls3u5BVkrKB4KX9hLjI2vHM
QaI4YKwt3VRUJb6uBujzUoDdR0MezhpHUyj2E4bsV//URuVmq8R2xBb+Hx2QALMx
1NtQ0yWUpPuh7lINlQBC+X7lwx8pd/LAje67ffpKR+g2ldskAzfoklF0cvOkQvsa
loGhyC05WEs+jzfn+PMj3sJNoV4BIXJ4+AHwLVdhg9vXPY6C+T8ycWqvAqLuC58=
=J8Fg
-----END PGP SIGNATURE-----

[Anders Logg (logg)]

On Thu, Nov 17, 2011 at 10:53:04PM +0100, Andre Massing wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
>
>
> On 11/17/2011 10:37 PM, Anders Logg wrote:
> > On Thu, Nov 17, 2011 at 10:24:56PM +0100, Andre Massing wrote: On
> > 11/17/2011 09:40 PM, Anders Logg wrote:
> >>>> Looks like the only remaining thing is periodic BC not
> >>>> working with Trilinos. From running the periodic demo:
> >>>>
> >>>> Error: Unable to add block of values to Epetra matrix. ***
> >>>> Reason: Did not manage to perform
> >>>> Epetra_FECrsMatrix::SumIntoGlobalValues. *** Where: This
> >>>> error was encountered inside EpetraMatrix.cpp.
> >>>>
> >
> > I wrote a somewhat lengthy email about that, explaining what goes
> > wrong and asking for opinons... you have even replied to it :)
> >
> >> I know and I've read it carefully!
>
> I was just a bit confused about the implied desired action :)

No action required at the moment. :-)

> >> I think I responded and said that this needs to be addressed by
> >> incorporating boundary conditions into the FunctionSpace by
> >> implementing constrained spaces. I even linked to a blueprint.
>
> I have seen that, nice. Marie had already registered on at
> https://blueprints.launchpad.net/dolfin/+spec/constrained-spaces
> so these can be merged or the older one can be marked as superseded?

I'll take a look. I actually looked for that blueprint as I had a
vague memory that we should have a blueprint on that from before but
couldn't find it. Maybe I searched for "restricted" instead of
"constrained".

> >> Bottom line: we won't fix this now so I'll add a check in that
> >> demo that it doesn't work with Epetra.
>
> Ok, I had the same feeling...But you could also use crisscross grids
> to fix it (just kidding ... )

Criss-cross grids seem to fix many problems. There was a long review
process for one of the FEniCS book chapters regarding an issue where
the authors had used criss-cross grids to get around some problem
(locking if I remember correctly).

--
Anders

> - --
> Andre
>
>
>
> >
> >> -- Anders
> >
> >
> > The bottom line is that there are several related issues: 1. The
> > sparsity pattern does not consider PeriodicBC thus there will be an
> > error in Trilinos when copying a slave row to a master row since
> > the matrix is filled / finalized.
> >
> > 2. Dirichlet and PeriodicBC calling and matrix.apply() while
> > applying bc. Every such call triggers a finalization of the matrix
> > atm.
> >
> > 3. https://bugs.launchpad.net/dolfin/+bug/891452 (which for some
> > reason does not show up in the bug overview, neither has been sent
> > to the list, strange.)
> >
> > See the original post
> > https://bugs.launchpad.net/dolfin/+bug/864741/comments/2
> >
> > for a more detailed explanation. The point is that it seems like a
> > more fundamental problem where I haven't dared to dirty-hack around
> > without opinions from the devs /maintainers :)
> >
> >
> >
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v2.0.17 (GNU/Linux)
> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
>
> iQEcBAEBAgAGBQJOxYJAAAoJEA79ggnbq9dmTJkH/RrNfQpSUT3XOd7jy+sqsQtW
> AnXpoW9k8EfAk6UO9NpEAge7mt808EVZKR7PaTQw9G34twSunz63Fz...

[Marie Rognes (meg-simula)]

On 11/17/2011 10:53 PM, Andre Massing wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
>
> On 11/17/2011 10:37 PM, Anders Logg wrote:
>> On Thu, Nov 17, 2011 at 10:24:56PM +0100, Andre Massing wrote: On
>> 11/17/2011 09:40 PM, Anders Logg wrote:
>>>>> Looks like the only remaining thing is periodic BC not
>>>>> working with Trilinos. From running the periodic demo:
>>>>>
>>>>> Error: Unable to add block of values to Epetra matrix. ***
>>>>> Reason: Did not manage to perform
>>>>> Epetra_FECrsMatrix::SumIntoGlobalValues. *** Where: This
>>>>> error was encountered inside EpetraMatrix.cpp.
>>>>>
>> I wrote a somewhat lengthy email about that, explaining what goes
>> wrong and asking for opinons... you have even replied to it :)
>>
>>> I know and I've read it carefully!
> I was just a bit confused about the implied desired action :)
>
>>> I think I responded and said that this needs to be addressed by
>>> incorporating boundary conditions into the FunctionSpace by
>>> implementing constrained spaces. I even linked to a blueprint.
> I have seen that, nice. Marie had already registered on at
> https://blueprints.launchpad.net/dolfin/+spec/constrained-spaces
> so these can be merged or the older one can be marked as superseded?

I think they both deserve to be kept. Although they are related, the one
does not necessarily imply the other.

I'll target the constrained-spaces one to 1.1 too.

--
Marie

>>> Bottom line: we won't fix this now so I'll add a check in that
>>> demo that it doesn't work with Epetra.
> Ok, I had the same feeling...But you could also use crisscross grids
> to fix it (just kidding ... )
>
> - --
> Andre
>
>
>>> -- Anders
>>
>> The bottom line is that there are several related issues: 1. The
>> sparsity pattern does not consider PeriodicBC thus there will be an
>> error in Trilinos when copying a slave row to a master row since
>> the matrix is filled / finalized.
>>
>> 2. Dirichlet and PeriodicBC calling and matrix.apply() while
>> applying bc. Every such call triggers a finalization of the matrix
>> atm.
>>
>> 3. https://bugs.launchpad.net/dolfin/+bug/891452 (which for some
>> reason does not show up in the bug overview, neither has been sent
>> to the list, strange.)
>>
>> See the original post
>> https://bugs.launchpad.net/dolfin/+bug/864741/comments/2
>>
>> for a more detailed explanation. The point is that it seems like a
>> more fundamental problem where I haven't dared to dirty-hack around
>> without opinions from the devs /maintainers :)
>>
>>
>>
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v2.0.17 (GNU/Linux)
> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
>
> iQEcBAEBAgAGBQJOxYJAAAoJEA79ggnbq9dmTJkH/RrNfQpSUT3XOd7jy+sqsQtW
> AnXpoW9k8EfAk6UO9NpEAge7mt808EVZKR7PaTQw9G34twSunz63Fze08agIpLYY
> EWJx7mrubKuvl6DRkmJcsCh1FAeHw6mLHugnVizLyls3u5BVkrKB4KX9hLjI2vHM
> QaI4YKwt3VRUJb6uBujzUoDdR0MezhpHUyj2E4bsV//URuVmq8R2xBb+Hx2QALMx
> 1NtQ0yWUpPuh7lINlQBC+X7lwx8pd/LAje67ffpKR+g2ldskAzfoklF0cvOkQvsa
> loGhyC05WEs+jzfn+PMj3sJNoV4BIXJ4+AHwLVdhg9vXPY6C+T8ycWqvAqLuC58=
> =J8Fg
> -----END PGP SIGNATURE-----
>

[Anders Logg (logg)]

On Thu, Nov 17, 2011 at 10:53:04PM +0100, Andre Massing wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
>
>
> On 11/17/2011 10:37 PM, Anders Logg wrote:
> > On Thu, Nov 17, 2011 at 10:24:56PM +0100, Andre Massing wrote: On
> > 11/17/2011 09:40 PM, Anders Logg wrote:
> >>>> Looks like the only remaining thing is periodic BC not
> >>>> working with Trilinos. From running the periodic demo:
> >>>>
> >>>> Error: Unable to add block of values to Epetra matrix. ***
> >>>> Reason: Did not manage to perform
> >>>> Epetra_FECrsMatrix::SumIntoGlobalValues. *** Where: This
> >>>> error was encountered inside EpetraMatrix.cpp.
> >>>>
> >
> > I wrote a somewhat lengthy email about that, explaining what goes
> > wrong and asking for opinons... you have even replied to it :)
> >
> >> I know and I've read it carefully!
>
> I was just a bit confused about the implied desired action :)
>
> >
> >> I think I responded and said that this needs to be addressed by
> >> incorporating boundary conditions into the FunctionSpace by
> >> implementing constrained spaces. I even linked to a blueprint.
>
> I have seen that, nice. Marie had already registered on at
> https://blueprints.launchpad.net/dolfin/+spec/constrained-spaces
> so these can be merged or the older one can be marked as superseded?

And here is another one from last year:

https://blueprints.launchpad.net/dolfin/+spec/bcs-in-sparsitypattern

Seems like this is an issue that keeps coming up.

--
Anders

> >
> >> Bottom line: we won't fix this now so I'll add a check in that
> >> demo that it doesn't work with Epetra.
>
> Ok, I had the same feeling...But you could also use crisscross grids
> to fix it (just kidding ... )
>
> - --
> Andre
>
>
>
> >
> >> -- Anders
> >
> >
> > The bottom line is that there are several related issues: 1. The
> > sparsity pattern does not consider PeriodicBC thus there will be an
> > error in Trilinos when copying a slave row to a master row since
> > the matrix is filled / finalized.
> >
> > 2. Dirichlet and PeriodicBC calling and matrix.apply() while
> > applying bc. Every such call triggers a finalization of the matrix
> > atm.
> >
> > 3. https://bugs.launchpad.net/dolfin/+bug/891452 (which for some
> > reason does not show up in the bug overview, neither has been sent
> > to the list, strange.)
> >
> > See the original post
> > https://bugs.launchpad.net/dolfin/+bug/864741/comments/2
> >
> > for a more detailed explanation. The point is that it seems like a
> > more fundamental problem where I haven't dared to dirty-hack around
> > without opinions from the devs /maintainers :)
> >
> >
> >
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v2.0.17 (GNU/Linux)
> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
>
> iQEcBAEBAgAGBQJOxYJAAAoJEA79ggnbq9dmTJkH/RrNfQpSUT3XOd7jy+sqsQtW
> AnXpoW9k8EfAk6UO9NpEAge7mt808EVZKR7PaTQw9G34twSunz63Fze08agIpLYY
> EWJx7mrubKuvl6DRkmJcsCh1FAeHw6mLHugnVizLyls3u5BVkrKB4KX9hLjI2vHM
> QaI4YKwt3VRUJb6uBujzUoDdR0MezhpHUyj2E4bsV//URuVmq8R2xBb+Hx2QALMx
> 1NtQ0yWUpPuh7lINlQBC+X7lwx8pd/LAje67ffpKR+g2ldskAzfoklF0cvOkQvsa
> loGhyC05WEs+jzfn+PMj3sJNoV4BIXJ4+AHwLVdhg9vXPY6C+T8ycWqvAqLuC58=
> =J8Fg
> -----END PGP SIGNATURE-----

[Anders Logg (logg)]

On Thu, Nov 17, 2011 at 09:52:12PM -0000, Garth Wells wrote:
> On 17 November 2011 21:37, Anders Logg <email address hidden> wrote:
> > On Thu, Nov 17, 2011 at 10:24:56PM +0100, Andre Massing wrote:
> >> -----BEGIN PGP SIGNED MESSAGE-----
> >> Hash: SHA1
> >>
> >> On 11/17/2011 09:40 PM, Anders Logg wrote:
> >> > Looks like the only remaining thing is periodic BC not working
> >> > with Trilinos. From running the periodic demo:
> >> >
> >> > Error:  Unable to add block of values to Epetra matrix. *** Reason:
> >> > Did not manage to perform Epetra_FECrsMatrix::SumIntoGlobalValues.
> >> > *** Where:  This error was encountered inside EpetraMatrix.cpp.
> >> >
> >>
> >> I wrote a somewhat lengthy email about that, explaining what goes
> >> wrong and asking for opinons... you have even replied to it :)
> >
> > I know and I've read it carefully!
> >
> > I think I responded and said that this needs to be addressed by
> > incorporating boundary conditions into the FunctionSpace by
> > implementing constrained spaces. I even linked to a blueprint.
> >
>
> A blueprint on this has been about for coming up two years:
>
> https://blueprints.launchpad.net/dolfin/+spec/bcs-in-sparsitypattern

I just noticed (see other post). I've marked it as superseded by the
new blueprint.

--
Anders

> Garth
>
> > Bottom line: we won't fix this now so I'll add a check in that demo
> > that it doesn't work with Epetra.
> >
> >
> >
> >> The bottom line is that there are several related issues:
> >> 1. The sparsity pattern does not consider PeriodicBC thus there will
> >> be an error in Trilinos when copying a slave row to a master row since
> >> the matrix is filled / finalized.
> >>
> >> 2. Dirichlet and PeriodicBC calling and matrix.apply() while applying
> >> bc. Every such call triggers a finalization of the matrix atm.
> >>
> >> 3. https://bugs.launchpad.net/dolfin/+bug/891452
> >> (which for some reason does not show up in the  bug overview, neither
> >> has been sent to the list, strange.)
> >>
> >> See the original post
> >> https://bugs.launchpad.net/dolfin/+bug/864741/comments/2
> >>
> >> for a more detailed explanation.
> >> The point is that it seems like a more fundamental problem where I
> >> haven't dared to dirty-hack around without opinions from the devs
> >> /maintainers :)
> >
> >
> >> - --
> >> Andre
> >>
> >>
> >>
> >>
> >>
> >> -----BEGIN PGP SIGNATURE-----
> >> Version: GnuPG v2.0.17 (GNU/Linux)
> >> Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/
> >>
> >> iQEcBAEBAgAGBQJOxXuoAAoJEA79ggnbq9dm0ToIAJYp09Rkq9midizQnt8zw4gr
> >> 197meLuTEuBn7fN4FR9EV3yBqhAwg7Ay/4UW96Da9mt09JoY+CRKEWyzBVhD0TC9
> >> jcDw7aroXZVJMZ9KgHWKiwyBA/7l+4SRXen6DXAneEKIfBJGnNFiul3HTgjBxFdR
> >> ZG7sMELsvynDM7ryt2FpZl89qI34YnOwCNj9vZAXNe8TWA4qsHmc7BTmBWojLYBS
> >> eStuEd73652epXBOmmQHZ47fZg83SxMUJgPkPbDKxgHBaM0sp9Q2X+f1lu5uY/wF
> >> 4pIOPIfb8bSA/90Ud3JBIinin43gW4GfL661DzldNlki8kbU5Zzob97wOnCHOpI=
> >> =l8m5
> >> -----END PGP SIGNATURE-----
> >
> >
> > Title:
> >  Unit tests fail to run if Trilinos is installed but not PETSc
> >
> > To manage notifications about this bug go to:
> > https://bugs.launchpad.net/dolfin/+bug/891716/+subscriptions
> >
>

[Anders Logg (logg)]

On Thu, Nov 17, 2011 at 10:04:53PM -0000, Marie Rognes wrote:
> On 11/17/2011 10:53 PM, Andre Massing wrote:
> > -----BEGIN PGP SIGNED MESSAGE-----
> > Hash: SHA1
> >
> >
> > On 11/17/2011 10:37 PM, Anders Logg wrote:
> >> On Thu, Nov 17, 2011 at 10:24:56PM +0100, Andre Massing wrote: On
> >> 11/17/2011 09:40 PM, Anders Logg wrote:
> >>>>> Looks like the only remaining thing is periodic BC not
> >>>>> working with Trilinos. From running the periodic demo:
> >>>>>
> >>>>> Error: Unable to add block of values to Epetra matrix. ***
> >>>>> Reason: Did not manage to perform
> >>>>> Epetra_FECrsMatrix::SumIntoGlobalValues. *** Where: This
> >>>>> error was encountered inside EpetraMatrix.cpp.
> >>>>>
> >> I wrote a somewhat lengthy email about that, explaining what goes
> >> wrong and asking for opinons... you have even replied to it :)
> >>
> >>> I know and I've read it carefully!
> > I was just a bit confused about the implied desired action :)
> >
> >>> I think I responded and said that this needs to be addressed by
> >>> incorporating boundary conditions into the FunctionSpace by
> >>> implementing constrained spaces. I even linked to a blueprint.
> > I have seen that, nice. Marie had already registered on at
> > https://blueprints.launchpad.net/dolfin/+spec/constrained-spaces
> > so these can be merged or the older one can be marked as superseded?
>
> I think they both deserve to be kept. Although they are related, the one
> does not necessarily imply the other.
>
> I'll target the constrained-spaces one to 1.1 too.

Feel free to un-supersede and break up in pieces as you like, but the
one I added regards both constrained and restricted spaces and my
feeling is we can create a solution that works for both.

--
Anders

[Garth Wells (garth-wells)]

Renaming and splitting this bug report. Re-targeting periodic bcs (this report), which relates to accounting for bcs in the sparsity pattern, to 1.1.