Its not really a good idea to mix active and passive synchronization (we may actually explicitly forbid it in the future). You can remove the calls to MPI_Win_fence () and still have correct synchronization. That said, you did find a bug in my bad synchronization detection because this is legal:
MPI_Win_fence (...); MPI_Win_lock (...); but this is not MPI_Win_fence (...); MPI_Put (...); /* MPI_Get, MPI_Accumulate, etc */ MPI_Win_lock (...); I will fix the bad synchronization detection in osc/rdma. -Nathan On Mon, May 02, 2016 at 06:49:15PM +0000, Palmer, Bruce J wrote: > Gilles, > > I downloaded and built openmpi-2.0.0rc2 and used that for the test. I get a > crash on more than 1 processor for the lock/unlock protocol with the error > message > > [node005:29916] *** An error occurred in MPI_Win_lock > [node005:29916] *** reported by process [3736862721,6] > [node005:29916] *** on win rdma window 3 > [node005:29916] *** MPI_ERR_RMA_SYNC: error executing rma sync > [node005:29916] *** MPI_ERRORS_ARE_FATAL (processes in this win will now > abort, > [node005:29916] *** and potentially your MPI job) > > and the request-based protocol hangs on the MPI_Rget call. The flush_local > protocol seems to work though. Unlike 1.8.3, the problems seem to occur no > matter what the value of NSIZE is. Should I try actually building 1.10 after > applying the patch to it? > > Bruce > > Message: 1 > Date: Mon, 2 May 2016 13:42:21 +0900 > From: Gilles Gouaillardet <gil...@rist.or.jp> > To: Open MPI Users <us...@open-mpi.org> > Subject: Re: [OMPI users] MPI Datatypes and RMA > Message-ID: <01c20fdf-c41b-96a8-6732-661745ddf...@rist.or.jp> > Content-Type: text/plain; charset="windows-1252"; Format="flowed" > > Bruce, > > > this issue was previously fixed on master and v2.x, but for some reasons, the > fix was not backported to v1.10 > > i made a PR at https://github.com/open-mpi/ompi-release/pull/1120/files > > in the mean time, feel free to manually apply the patch at > https://patch-diff.githubusercontent.com/raw/open-mpi/ompi-release/pull/1120.patch > > > Cheers, > > > Gilles > _______________________________________________ > users mailing list > us...@open-mpi.org > Subscription: https://www.open-mpi.org/mailman/listinfo.cgi/users > Link to this post: > http://www.open-mpi.org/community/lists/users/2016/05/29066.php
pgpKVTulDYUBB.pgp
Description: PGP signature
