I answered to the PR but I'll bring my comment here as well. In addition to
the performance implication, there might be a correctness implication here.
del_procs does not have to be called globally by all participating
processes in same time, and can be called with a subset of processes. As an
example, a correct implementation of MPI_Comm_disconnect could call
del_procs with all the processes that are becoming disconnected. Thus, a
global pmix_fence, by it's conceptual implication of a global
synchronization (due to the NULL parameter) might be the wrong thing to do.

  George.


On Mon, Mar 2, 2015 at 10:05 AM, Ralph Castain <[email protected]> wrote:

> It’s your code, so you are welcome to do this if you want. I’ll just point
> out that his is a really big hit in total execution time at scale as it
> will be done in addition to the barrier already performed in MPI_Finalize
>
> So you are going to do _two_ barriers during shutdown.
>
>
> > On Mar 2, 2015, at 5:43 AM, [email protected] wrote:
> >
> > This is an automated email from the git hooks/post-receive script. It was
> > generated because a ref change was pushed to the repository containing
> > the project "open-mpi/ompi".
> >
> > The branch, master has been updated
> >       via  fbb7c80312cbcd823346e89a56f5d83e8620c57c (commit)
> >       via  168c83ed9592120fd2199e8280b517ab0060e136 (commit)
> >      from  42f5a36ee3f1e400aa251804725b86192c9df9fa (commit)
> >
> > Those revisions listed above that are new to this repository have
> > not appeared on any other notification email; so we list those
> > revisions in full, below.
> >
> > - Log -----------------------------------------------------------------
> >
> https://github.com/open-mpi/ompi/commit/fbb7c80312cbcd823346e89a56f5d83e8620c57c
> >
> > commit fbb7c80312cbcd823346e89a56f5d83e8620c57c
> > Merge: 42f5a36 168c83e
> > Author: Mike Dubman <[email protected]>
> > Date:   Mon Mar 2 15:43:32 2015 +0200
> >
> >    Merge pull request #439 from alex-mikheev/topic/mxm_finalize_fix
> >
> >    OMPI/MXM: add out of band barrier at the end of del_procs
> >
> >
> >
> >
> https://github.com/open-mpi/ompi/commit/168c83ed9592120fd2199e8280b517ab0060e136
> >
> > commit 168c83ed9592120fd2199e8280b517ab0060e136
> > Author: Alex Mikheev <[email protected]>
> > Date:   Mon Mar 2 12:56:02 2015 +0200
> >
> >    OMPI/MXM: add out of band barrier at the end of del_procs
> >
> >    mxm shutdown requires out of band barrier
> >
> > diff --git a/ompi/mca/mtl/mxm/mtl_mxm.c b/ompi/mca/mtl/mxm/mtl_mxm.c
> > index 1a4e21a..ed4089a 100644
> > --- a/ompi/mca/mtl/mxm/mtl_mxm.c
> > +++ b/ompi/mca/mtl/mxm/mtl_mxm.c
> > @@ -617,6 +617,7 @@ int ompi_mtl_mxm_del_procs(struct
> mca_mtl_base_module_t *mtl, size_t nprocs,
> >             OBJ_RELEASE(endpoint);
> >         }
> >     }
> > +    opal_pmix.fence(NULL, 0);
> >     return OMPI_SUCCESS;
> > }
> >
> > diff --git a/ompi/mca/pml/yalla/pml_yalla.c
> b/ompi/mca/pml/yalla/pml_yalla.c
> > index 2cfa6ca..d53cb7c 100644
> > --- a/ompi/mca/pml/yalla/pml_yalla.c
> > +++ b/ompi/mca/pml/yalla/pml_yalla.c
> > @@ -240,6 +240,7 @@ int mca_pml_yalla_del_procs(struct ompi_proc_t
> **procs, size_t nprocs)
> >         PML_YALLA_VERBOSE(2, "disconnected from rank %ld",
> procs[i]->super.proc_name);
> >         procs[i]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_PML] = NULL;
> >     }
> > +    opal_pmix.fence(NULL, 0);
> >     return OMPI_SUCCESS;
> > }
> >
> >
> >
> > -----------------------------------------------------------------------
> >
> > Summary of changes:
> > ompi/mca/mtl/mxm/mtl_mxm.c     | 1 +
> > ompi/mca/pml/yalla/pml_yalla.c | 1 +
> > 2 files changed, 2 insertions(+)
> >
> >
> > hooks/post-receive
> > --
> > open-mpi/ompi
> > _______________________________________________
> > ompi-commits mailing list
> > [email protected]
> > http://www.open-mpi.org/mailman/listinfo.cgi/ompi-commits
>
> _______________________________________________
> devel mailing list
> [email protected]
> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel
> Searchable archives:
> http://www.open-mpi.org/community/lists/devel/2015/03/index.php

Reply via email to