> On Jul 5, 2018, at 8:28 AM, Mark Adams <[email protected]> wrote: > > > Please share the results of your experiments that prove OpenMP does not > improve performance for Mark’s users. > > This obviously does not "prove" anything but my users use OpenMP primarily > because they do not distribute their mesh metadata.
I.e. my users have decided not to write scalable code. > They can not replicated the mesh on every core, > on large scale problems and shared memory allows them to survive. They have > decided to use threads as opposed to MPI shared memory. (Not a big deal, once > you decide not to use distributed memory the damage is done and NERSC seems > to be OMP centric so they can probably get better support for OMP than MPI > shared memory.) > > BTW, PETSc does support OMP, that is what I have been working on testing for > the last few weeks. First with Hypre (numerics are screwed up from an > apparent compiler bug or a race condition of some sort; it fails at higher > levels of optimization), and second with MKL kernels. The numerics are > working with MKL and we are working on packaging this up to deliver to a user > (they will test performance). >
