dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Tue, Apr 07, 2009 at 01:51:54PM +0200, Jed Brown wrote:
> On Mon 2009-04-06 07:42, Matthew Knepley wrote:
> > On Mon, Apr 6, 2009 at 1:09 AM, Anders Logg <logg@xxxxxxxxx> wrote:
> > 
> >     On Sun, Apr 05, 2009 at 05:11:51PM -0500, Matthew Knepley wrote:
> >     > On Sun, Apr 5, 2009 at 2:54 PM, Anders Logg <logg@xxxxxxxxx> wrote:
> >     >
> >     >     Dirichlet BC would need to be added in/after each multiplication and
> >     >     it should be possible to build it into the mult() operator and make
> >     it
> >     >     efficient.
> >     >
> >     >
> >     > I still think the best way to handle this is to eliminate them, as I
> >     talked
> >     > about
> >     > last time at Simula.
> >     >
> >     >   Matt
> > 
> >     I think we're in position to do that now with the new FunctionSpace
> >     class. One could allow restrictions to be imposed, for example
> > 
> >      V.restrict(bc);
> > 
> >     But I haven't seen any good numbers to indicate it's worth the effort.
> >     How much better is it to eliminate, condition numbers etc for letting
> >     the constrained dofs just sit along?
> > 
> > 
> > Its not about performance, so much as
> > 
> >   1) Programming ease, ESPECIALLY with nonlinear problems
> > 
> >   2) Separation of storage organization from traversal organization
> > 
> >   3) Cache performance
> > 
> >   4) Interaction with external packages
> > 
> > I think I must have done a terrible job explaining this. I have always coded
> > both ways and always found that elimination is better.
> 
> I agree with all your points, but with caveats on 3.
> 
> There are much bigger gains in cache performance by interlacing the dofs
> so as to enable the use of inodes and blocked matrix formats (BAIJ).
> 
> If we want to use blocked formats with slip boundary conditions, we need
> to be able to leave the condition in.  A slip boundary condition imposes
> a Dirichlet condition on the normal component and stress conditions on
> the tangent components.  The global degrees of freedom can be written in
> a rotated coordinate frame so that the Dirichlet condition can be
> completely removed, but when using BAIJ we have to put something there.
> This can be done, complete with "zeroing the column", but it requires
> manipulations in function evaluation and when setting values in the
> matrix.  In some of my experiments, the performance benefits of BAIJ
> over AIJ+inodes justify this strategy.
> 
> If you have Dirichlet conditions on all components, then removing these
> dofs from the system really is better.

How much faster?

>  Also, if you are not using blocked matrix formats, you might as
> well remove all Dirichlet dofs for all the reasons Matt states.

It still doesn't make sense to me.

> >   1) Programming ease, ESPECIALLY with nonlinear problems

I think keeping the dofs is easier.

> >   2) Separation of storage organization from traversal organization

I don't really understand this point.

> >   3) Cache performance

I don't know anything about this.

> >   4) Interaction with external packages

Which packages? What makes keeping the dofs simpler is that you only
have one vector of dofs, no dofs to put back for example when
plotting (which is interaction with an external package).

-- 
Anders

Attachment: signature.asc
Description: Digital signature