revamping, reintroducing OpenMPAssembler
6 months ago by
dolfin used to have a multi-threaded assembler that was removed I think because it did not play well with a lot of examples or PETSc. Does anyone know or is working on a strategy to implement a new OpenMP assembler for fenics? I would find it useful for solving large nonlinear problems (1e6 nodes) without having to rewrite my code to use MPI. I'd be happy to try implementing one if the fenics experts think it's not a lost cause, although I do not quite know where to start. I am also interested in any workaround in PETSc or Eigen that might speed up the assembly in large systems without using MPI. I tried building Eigen and openblas with openmp and linking dolfin to the multi-threaded libopenblas but that does not seem to speed things up.
Thanks for any suggestion,
Please login to add an answer/comment or follow this question.