[petsc-dev] PetscIntialize docs

[Jed Brown]

"Mark F. Adams" <mfadams at lbl.gov> writes:
> Ah yes, I was being simplistic and trying to avoid all rc files.  So
> all we need to do is add -skip_petsrc to the the user specified file,
> or the command line, and this will avoid trying to get the home
> directory.

Or use 'next' or 'jed/avoid-getpwuid', which never calls that crazy function.

Vitali said there was another independent problem that sequences of
MPI_Bcast could deliver packets in the wrong order and hang or crash.
You can work around that by exporting PAMID_COLLECTIVES=0 to fall back
on the MPICH reference implementations (some of which reportedly perform
better than the "optimized" implementations).  He suggested that
communicating the rc file in one round instead of two might help avoid
the MPI_Bcast bug.  I made branch 'jed/options-insert-bcast-bgq', which
you can use, e.g.,

  $ git checkout --detach next
  $ git merge origin/jed/options-insert-bcast-bgq
  $ make

I don't know how to reproduce the problem or identify how frequently it
occurs, so I only tested that this branch functions correctly, but not
whether it reduces the occurrence of triggering the BG/Q bug.  I'm
skeptical, but I think the branch probably makes sense to merge because
the code is no more complicated and it does less rounds of
communication, so it's a bit easier to step through in a debugger.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 835 bytes
Desc: not available
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20130905/b553d398/attachment.sig>