This is why I'm writing a non-one-sided implementation of PetscSF (and that will become the default).
It will likely be faster even on MPICH due to the way one-sided is actually implemented. The only work-around for MatTranspose is to fall back to the old code that didn't preallocate correctly and was much slower. On Wed, Oct 17, 2012 at 4:57 PM, Barry Smith <bsmith at mcs.anl.gov> wrote: > I agree but if you start using SF in "normal" parts of PETSc OpenMPI is > constantly going to cause these emails and that annoys both users and us; > better to just turn off one-sided for OpenMPI and you can have configure > report that in a nice box if you like. -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20121017/8f77d070/attachment-0001.html>
