dolfin team mailing list archive

[kent-and@xxxxxxxxx]

> Kent-Andre Mardal wrote:
>> BTW: the Assembler code that computes A, b and enforces BC
>> simultaneously in a symmetric way,
>
> Very nice. We can now implement some symmetric linear solvers.
>
> Do you keep the 'Dirichlet' dofs and zero the row and column, or do you
> eliminate them from the global system?

The Dirichlet dofs are kept. I thought that was simplest to do..
>
> is now faster than the standard
>> non-symmetric way of doing it:
>>
>> timings:
>> new     : 1.55
>> standard: 1.94
>>
>
> Which backend is this using? It could be saying more about the
> efficiency of the function ident(..), which I suspect is slow for
> uBlasMatrix, than the assembly process.
>

Good point. There is speed up for all backends that I have tried.
Here are the numbers:

 ~/local/src/dolfin/demo/pde/sym-dirichlet-bc >python speed-up-test.py  |
grep using
time for standard assembly  1.89726901054  using  uBLAS
time for new assembly       1.49513792992  using  uBLAS
time for standard assembly  1.97571492195  using  PETSc
time for new assembly       1.55099701881  using  PETSc
time for standard assembly  2.26576018333  using  Epetra
time for new assembly       1.86703896523  using  Epetra


Kent