NGUYEN Laurent wrote:
Hello,
I'm trying to understand the difference between theses two options:
" --mca mpi_paffinity_alone 1 "
and
" --bind-to-core "
To me, it's the same thing (may be paffinity have maffinity in
addition) but the purpose af theses options is to bind MPI process to
processors. Which is the best to obtain performances?
They are essentially the same thing. The history is that
mpi_paffinity_alone came first. Then a group of us decided that we
wanted to be able to do more than just core binding (like bind-to-socket
and eventually others) so we decided a more descriptive name made
sense. I would suggest sticking with --bind-to-core just so thinks look
clear in your mpirun command. That and the possibility that one day
mpi_paffinity_alone is deprecated.
I have another question about processes bindings: I use rankfiles to
have a precise binding of my process on processor or socket (it useful
for hybrid computing like pthread or OpenMP). I read you don't want to
maintain this feature: how can I substitute it?
IMO, I say we do support it and will until we find a suitable
replacement (there is a proposal that Jeff and I have bounced around to
each other intermittently but have not brought it to the community).
There have been discussion about lack of testing of rankfiles and other
binding related items making it hard for some to feel comfortable about
putting code back that may have side negative side affects. But I don't
equate that to mean we will not support or take bug reports on the
rankfile feature.
--
Oracle
Terry D. Dontje | Principal Software Engineer
Developer Tools Engineering | +1.650.633.7054
Oracle * - Performance Technologies*
95 Network Drive, Burlington, MA 01803
Email terry.don...@oracle.com <mailto:terry.don...@oracle.com>
