dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Thu, May 22, 2008 at 11:50:36AM +0200, Dag Lindbo wrote:
> Anders Logg wrote:
> > On Thu, May 22, 2008 at 11:26:41AM +0200, Dag Lindbo wrote:
> >> DOLFIN wrote:
> >>> One or more new changesets pushed to the primary dolfin repository.
> >>> A short summary of the last three changesets is included below.
> >>>
> >>> changeset:   4224:c5a7da38ff2c966e5f2483b5232993fb75ac6aa1
> >>> tag:         tip
> >>> user:        Anders Logg <logg@xxxxxxxxx>
> >>> date:        Thu May 22 10:19:48 2008 +0200
> >>> files:       dolfin/fem/DirichletBC.cpp dolfin/fem/DirichletBC.h
> >>> description:
> >>> Major cleanup of DirichletBC:
> >>>  - Store boundary facets in array, common for all different representations
> >>>  - BCs may now be specified in 3 differenet ways: SubDomain, MeshFunction, or as Mesh data
> >>>  - Should speed up repeated application of BCs
> >>>  - Tested on a few different applications, but please check that nothing broke.
> >>>    In particular, geometric and pointwise application has not been tested.
> >> Noting broke here!
> > 
> > ok, nice. Thanks for checking (Kristian also).
> > 
> >> However, I'm not seeing the speedup... (and I'm reapplying the dirichlet
> >> bcs every timestep). My code spends a good 30% of runtime applying
> >> dirichlet BCs (linear systems are pre-factorized so the solve step takes
> >> no time). Any hints?
> > 
> > The speedup may not be noticeable if a lot of time is spent on
> > evaluating the function. Would be interesting to profile.
> > 
> > The difference from before is that each time apply() is called, we
> > only iterate over the actual facets included in your subdomain of the
> > boundary, not the entire mesh and then skipping those facets which
> > don't belong to your boundary. Also, depending on which type of search
> > ("topological", "geometric", "pointwise"), the new implementation
> > should avoid initializing mesh facets.
> > 
> 
> Does this mean that there now is a distinct difference in performance
> between giving the subdomain as a class that inherits from SubDomain
> (overloading bool inside) and pre-computing a mesh function with
> subdomain markers? /Dag

There is some overhead but I don't know how big it is. The performance
for the three different ways to specify boundary conditions should be

  MeshData > MeshFunction > SubDomain

For the three methods for finding dofs on facets, the performance
should be

  topological > geometric > pointwise

-- 
Anders