On Thu, Oct 25, 2012 at 11:40 AM, Roy Stogner <royst...@ices.utexas.edu>wrote:
> Here I'm not sure. We don't have any explicit transient problems in
> the libMesh examples, do we? That's really not the kind of problem
> everyone had in mind when libMesh was designed, but it's probably what
> we want to look at if we want to benchmark the parts of the code that
> sometimes get lost the noise when a matrix assembly and linear solver
> get involved.
JFNK is very close to an Explicit Transient solve... we have to compute a
residual vector for every residual evaluation. Because of this, I've been
following this thread... we DO see the time for computing values, gradients
and (especially) second derivatives of our variables showing up in
profiling. When we get up over 2,000 or 3,000 variables (yes, we do run
simulations with this many!) this can actually grow to about 40% of the
runtime! (depending on what else we're doing of course). Even for
"simple" 3D thermo-mechanics (4 variables) the time to compute quadrature
point quantities can consume a decent chunk of the runtime and always show
up in profiling. I've optimized the HELL out of it (including using
performance analysis tools from Intel to get the loops just right so they
vectorize as well as possible) but there is no where to go without changing
the representation on the libMesh end...
In summary: we would definitely be interested in helping on this front...
Derek
------------------------------------------------------------------------------
Everyone hates slow websites. So do we.
Make your web apps faster with AppDynamics
Download AppDynamics Lite for free today:
http://p.sf.net/sfu/appdyn_sfd2d_oct
_______________________________________________
Libmesh-devel mailing list
Libmesh-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/libmesh-devel
