I do like #397. The info key is something that could go in 3.1, and helps an existing usage pattern (e.g. in OSHMPI or ARMCI-MPI). Ticket #397 looks like something that we will have to target at MPI 4. IIUC, MPICH has started detecting this automatically without the info key, so this ticket may be moot. Hopefully an MPICH developer familiar with this optimization can comment. ~Jim. On Tue, Oct 14, 2014 at 10:58 AM, Jeff Hammond <[email protected]> wrote:

I contend that https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/397 is the right way to solve this problem. Using overlapping windows here is unscalable because if you have one win_create for every process in the node on top of one win_allocate_shared per node, you end up with ppn+1 windows to hack the effect of one win_allocate if my ticket is passed. This is pretty significant overhead if one wants to run OpenSHMEM on an Intel Xeon Phi...

And note that the time to setup all these windows properly is nontrivial. It has been measured in the context of another project. I can't share the details since it's not published and I'm not in a position to decide to share it at this point.

Jeff

On Tue, Oct 14, 2014 at 7:42 AM, Jim Dinan <[email protected]> wrote:

...

This is an info key that e.g. OSHMPI could use to tell MPI that the buffer being passed to MPI_Win_create was allocated via MPI_Win_allocate_shared. If this is already covered by another ticket, let me know the existing trac number and I will close this one. If we want to pursue this as a 3.0 erratum, we could utilize this new ticket.

~Jim.

On Tue, Oct 14, 2014 at 10:38 AM, Jeff Hammond <[email protected]> wrote:

...

I don't fully understand. Is this designed to make it easier to use the overlapping windows trick because we can get at shm inside of other

window

...

types? I have a ticket that solves that issue holistically already...

Sent from my iPhone

On Oct 14, 2014, at 7:17 AM, Jim Dinan <[email protected]> wrote:

Hi All,

Please see the ticket: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/460

For a new info key that was proposed by Mikhail Brinsky of Intel.

Thanks, ~Jim.

_______________________________________________ mpiwg-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma

_______________________________________________ mpiwg-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma

_______________________________________________ mpiwg-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma

-- Jeff Hammond [email protected] http://jeffhammond.github.io/ _______________________________________________ mpiwg-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma