[mpiwg-rma] Fence + threads
rabenseifner at hlrs.de
Sun Jun 22 03:06:37 CDT 2014
Are you sure that you want to use MPI_Fence?
MPI_Fence is collective because a Group of processes
(here threads in different proceeses) have done some
RMA and want to tell all others in the group that they did.
If they do not know to whom they want to tell some information,
and only these receivers want to be blocked,
then non-collective methods based on MPI_Lock/Unlock may be
If I understand right, then you have the following issues:
- the MPI_PUT is locally finished
- the data stored with MPI_PUT is remotely accessible
- some other thread are informed about that this
data is now available.
- some restrictions on who knows about
who needs the information, i.e., a fence over all
threads of all processes is the wrong way.
----- Original Message -----
> From: "Jed Brown" <jedbrown at mcs.anl.gov>
> To: "William Gropp" <wgropp at illinois.edu>, "MPI WG Remote Memory Access working group"
> <mpiwg-rma at lists.mpi-forum.org>
> Sent: Saturday, June 21, 2014 11:12:49 PM
> Subject: Re: [mpiwg-rma] Fence + threads
> William Gropp <wgropp at illinois.edu> writes:
> > There are lots of choices. Passive target can be used. Or a
> > single
> > thread can call MPI_Win_fence - all threads can call the RMA
> > communication routines.
> The point of my example was that the threads are doing divergent
> so wrapping a user mutex around MPI_Win_fence makes the algorithm
> synchronous and may result in idle/spinning threads.
> The argument that passive target can be used isn't compelling to me
> since it also applies in other cases that we use to justify active
> mpiwg-rma mailing list
> mpiwg-rma at lists.mpi-forum.org
Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner at hlrs.de
High Performance Computing Center (HLRS) . phone ++49(0)711/685-65530
University of Stuttgart . . . . . . . . .. fax ++49(0)711 / 685-65832
Head of Dpmt Parallel Computing . . . www.hlrs.de/people/rabenseifner
Nobelstr. 19, D-70550 Stuttgart, Germany . . . . (Office: Room 1.307)
More information about the mpiwg-rma