On Tue 2008-11-18 12:21, Roy Stogner wrote:
> > We might actually also be able to do away with the _p_level as well
> > if we are willing to impose some maximum p-level, say 16 or
> > whatever...
> 
> I have a fantasy where we some day support spectral elements with
> FFT-based techniques to handle p in the 100s... but for now, 15 (4
> bits) would be more than adequate.

Having used p>1000 in 2D and p>100 in 3D, I don't think it's actually
useful.  The condition number goes like p^4 so you need to solve the
linear systems to much higher tolerance than in a lower-order method.
Modest order will easily let you hit machine precision and the memory
access is nicer anyway.  I do think matrix-free tensor products methods
are very useful, I just don't think p>15 will ever be important.

Anisotropic elements can be useful, for instance in stabilizing the
inf-sup constant for high aspect ratio flows.  If you store each part of
the anisotropic order explicitly per-element, then they would each need
some space so you're looking at 12 bits per element.  My 3D
implementation is completely different from libmesh and uses
(1+4f)*sizeof(void*) per element regardless of order (where f is the
number of approximation orders used---2 for a mixed Stokes system) plus
a constraint matrix (DoF map).  The connectivity needs to be present
during setup, but then can be discarded.

Jed

Attachment: pgp32mKZt2MUZ.pgp
Description: PGP signature

-------------------------------------------------------------------------
This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
Build the coolest Linux based applications with Moblin SDK & win great prizes
Grand prize is a trip for two to an Open Source event anywhere in the world
http://moblin-contest.org/redirect.php?banner_id=100&url=/
_______________________________________________
Libmesh-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/libmesh-devel

Reply via email to