Hello Lawrence, thanks for the fast reply. i am thinking about a way to parallelize parts of our micromagnetic code (Magnum.Fe) which is based on FEniCS. it basically solves that LLG equation dm = m x m x h_eff, where m, h_eff are FEM functions and h_eff again depends on m in some (complicated) fashion. so i have a sequence of FEniCS problems which finally gives dm. Finally we use sundials to do time-integration. I would now like to parallelize at least parts of the problem using MPI. therefor each part of the code needs to support MPI. If I understand correctly all FEniCS operators should directly work and the only missing link is the 'external' time-integrator. if petsc-matshell could be used, the petsc time-integrators should allow parallel execution of the whole code directly. Could this work, or do you see any basic problems? as a first step i thought packing a single dolfin matrix in a petsc-matshell should be a nice example, but at the moment it does not run in parallel. thanks Florian On 03/10/2017 12:02 PM, Lawrence Mitchell wrote:
Hi Florian,
On 10 Mar 2017, at 10:47, Florian Bruckner <e0425375@gmail.com> wrote:
hi,
i am wondering if the method to interace PETSc (http://www.firedrakeproject.org/petsc-interface.html) would also work in parallel?
i tried a slightly modified code (since i am using dolfin at the moment), which works on one processor but fails in parallel. Is there an easy way to pack a sequence of dolfin / firedrake operators into one parallel petsc opterator (matshell) You can do whatever you like inside a MatShell, but you will have to handle parallelism yourself. Can you describe what it is you're trying to do? It may be that there is a better way.
Lawrence _______________________________________________ firedrake mailing list firedrake@imperial.ac.uk https://mailman.ic.ac.uk/mailman/listinfo/firedrake