Hi Katharina, while I agree that it is difficult to integrate with the meshworker due to point 1, I think that implementing point 2 (and maybe 3) into the meshworker would have been a nice addition for everyone. Assembling vectors without locks might give a nice performance boost on multi-core machines.
I wonder how hard it would be to replace the dependency on FEValues inside the meshworker to an arbitrary template class. How hard would it be to fit your work into the meshworker then? On Sun, Jun 24, 2012 at 10:54 AM, Katharina Kormann <[email protected]> wrote: > It is certainly true that MeshWorker has a very flexible interface for faces > and multigrid that should be reused whenever possible. In fact, Martin has > looked at this at some point but we decided that the focus was quite > different and that it would take a bigger effort to integrate our work to the > MeshWorker. Essentially, there are three choices in the MeshWorker framework > that do not fit with our strategy: > > 1. The MeshWorker is centered around the usage of FEValuesBase objects for > all local evaluations. While this is very general (all of the deal.II > assembly code gets channeled through it), it turned out to be too slow for > our purposes: When you only assemble vectors like in a matrix-free approach, > FEValues eats just too much time. > > 2. The parallel loop over cells for MeshWorker is based WorkStream which does > local computations in parallel but serializes writing into vectors or > matrices. This, again, was severely impacting performance, in particular on > multicore machines. > > 3. The looping in deal.II is based on the topology of the mesh. Again, this > does not really fit well with parallel algorithms where one wants to use > vectorization which processes several cells with a single CPU instruction. > > In order to be able to support DG, it might be advantageous to integrate the > MeshWorker. However, one would at least need to find a way to implement the > caching of the face data efficiently. -- Timo Heister http://www.math.tamu.edu/~heister/ _______________________________________________ dealii mailing list http://poisson.dealii.org/mailman/listinfo/dealii
