[Trilinos-Users] Teko and Navier-Stokes preconditioning

Bart Janssens bart.janssens at lid.kviv.be
Thu Mar 15 10:59:52 MDT 2012

Hi guys,

I'd like to use Teko to experiment with block preconditioners for a
linear system resulting from a finite element discretization of the
Navier-Stokes equations. Some questions came to mind, after a first
look at the problem:
- We currently don't use a block structure, i.e. our variables are
grouped u1,v1,p1, u2, v2, p2, ... From what I understand, Teko can in
fact convert to a blocked view from there. Is there a performance
penalty involved, i.e. would it be recommended to switch to a block
structure from  the start?
- A literature review seems to indicate that augmented lagrangian type
preconditioners perform best. Does anyone have experience with this
type of preconditioner, and know if it works as well as described (in
terms of total solve time, not just iteration count)? For reference, I
intend to implement [1].
- Building with ML_ENABLE_TekoSmoothers fails with missing headers
Teko_SmootherPreconditionerFactory.hpp and
Teko_MLPreconditionerFactory.hpp. Not sure if I actually need
ML_ENABLE_TekoSmoothers or what it does.

[1] http://dx.doi.org/10.1002/fld.3654 "Efficient augmented
Lagrangian-type preconditioning for the Oseen problem using Grad-Div
(pdf at http://www.math.tamu.edu/~heister/preprints/grad_div_prec.pdf)

Kind regards,


More information about the Trilinos-Users mailing list