<div dir="ltr">Rolf,<div><br></div><div>Here is an attempt to simplify your example for discussion. Given a shared memory window, shr_mem_win, with buffer, shr_mem_buf:</div><div><br></div><font face="courier new, monospace">MPI_Win_fence(MPI_MODE_NOSTORE | MPI_MODE_NOPUT | MPI_MODE_NOPRECEDE | MPI_MODE_NOSUCCEED, shr_mem_win );<br>
</font><div><span style="font-family:'courier new',monospace">shr_mem_buf[...] = ...;</span><br></div><div><font face="courier new, monospace">MPI_Win_fence(MPI_MODE_NOPUT | MPI_MODE_NOPRECEDE | MPI_MODE_NOSUCCEED, shr_mem_win);</font></div>
<div><br></div><div>Right now, Fence assertions don't say anything special about shared memory windows:</div><div><br></div><div><img src="cid:ii_1460024758965ab0" alt="Inline image 1" width="445" height="170"><br></div>
<div><br></div>NOPRECEDE/SUCCEED are defined in terms of MPI RMA function calls, and do not cover load/store. Thus, Rolf's usage appears to be correct per the current text. In the MPICH fence implementation, src/mpid/ch3/src/ch3u_rma_sync.c:935 we have:<br>
<br><font face="courier new, monospace">if (!(assert & MPI_MODE_NOSUCCEED)) win_ptr->fence_issued = 1;</font>
<div><font face="courier new, monospace"><br></font></div>Because of this check, we don't actually start an active target epoch on the first fence in the example above. On the second fence, we therefore don't perform the necessary synchronization, leading to incorrect output in Rolf's example.<div>
<br></div><div>Question to WG: Do we need to update the fence assertions to better define interaction with local load/store accesses and remote stores?</div><div><br></div><div>If not, then Rolf's code is correct and we need to modify the check above in MPICH to something like:</div>
<div><font face="courier new, monospace"><br></font></div><div><font face="courier new, monospace">if (!(assert & MPI_MODE_NOSUCCEED) || win_ptr->create_<span class="">flavor</span> == MPI_WIN_FLAVOR_SHARED</font><span style="font-family:'courier new',monospace">)</span></div>
<div><span style="font-family:'courier new',monospace"> win_ptr->fence_issued = 1;</span></div><div><font face="courier new, monospace"><br></font></div><div> ~Jim.<br></div></div><div class="gmail_extra"><br>
<br><div class="gmail_quote">On Tue, Apr 8, 2014 at 12:02 PM, Rolf Rabenseifner <span dir="ltr"><<a href="mailto:rabenseifner@hlrs.de" target="_blank">rabenseifner@hlrs.de</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Jim,<br>
<br>
I'm now sure, that mpich has a bug with assertions on shared memory windows.<br>
<br>
in the example, rcv_buf_left and rcv_buf_right are the windows.<br>
the only accesses to these rcv_buf_... are stores from remote<br>
and fully local loads.<br>
Both accesses are done in different epochs surrounded by MPI_Win_fence.<br>
<br>
According to your interpretation (which is really okay)<br>
all fences can use all possible assertions (!!!),<br>
except after the remote stores, MPI_MODE_NOSTORE cannot be used.<br>
<br>
I updated the example and mpich is executing it wrong.<br>
<br>
Please check it yourself on your installation:<br>
halo_1sided_store_win_alloc_shared_w-a-2-cray.c<br>
<br>
Without the assertions, all works:<br>
halo_1sided_store_win_alloc_shared_w-a-2NO-cray.c<br>
<br>
Could you verify that mpich has a bug?<br>
<br>
Additionally, I would recommend that we add after MPI-3.0 p451:33<br>
<br>
Note that in shared memory windows (allocated with<br>
MPI_WIN_ALLOCATE_SHARED), there is no difference<br>
between remote store accesses and local store accesses<br>
to the window.<br>
<br>
This would help to understand that "the local window<br>
was not updated by stores" does not mean "by local stores",<br>
see p452:1 and p452:9.<br>
<br>
Is it a good idea?<br>
<br>
Best regards<br>
Rolf<br>
<div class=""><br>
<br>
----- Original Message -----<br>
> From: "Jim Dinan" <<a href="mailto:james.dinan@gmail.com">james.dinan@gmail.com</a>><br>
> To: "MPI WG Remote Memory Access working group" <<a href="mailto:mpiwg-rma@lists.mpi-forum.org">mpiwg-rma@lists.mpi-forum.org</a>><br>
> Sent: Friday, March 21, 2014 8:14:22 PM<br>
> Subject: Re: [mpiwg-rma] Problems with RMA synchronization in combination with load/store shared memory accesses<br>
><br>
><br>
><br>
</div><div class="">> Rolf,<br>
><br>
><br>
> This line is incorrect: MPI_Win_fence(MPI_MODE_NOSTORE +<br>
</div>> MPI_MODE_NOPRECEDE, win_ rcv_buf _left );<br>
<div class="HOEnZb"><div class="h5">><br>
><br>
> You need to do a bitwise OR of the assertions (MPI_MODE_NOSTORE |<br>
> MPI_MODE_NOPRECEDE).<br>
><br>
> In halo_1sided_store_win_alloc_shared.c, you are doing stores within<br>
> the epoch, so MPI_MODE_NOSTORE looks like an incorrect assertion on<br>
> the closing fence.<br>
><br>
> Following the Fence epoch, you are reading from the left/right recv<br>
> buffers. That also needs to be done within an RMA epoch, if you are<br>
> reading non-local data.<br>
><br>
> ~Jim.<br>
><br>
><br>
><br>
> On Fri, Feb 21, 2014 at 6:07 AM, Rolf Rabenseifner <<br>
> <a href="mailto:rabenseifner@hlrs.de">rabenseifner@hlrs.de</a> > wrote:<br>
><br>
><br>
> Dear member of the RMA group and especially the mpich developers,<br>
><br>
> I have real problems with the new shared memory in MPI-3.0,<br>
> i.e., the load/stores together with the RMA synchronization<br>
> causes wrong execution results.<br>
><br>
> The attached<br>
> 1sided_halo_C_mpich_problems_rabenseifner.tar.gz or .zip<br>
> contains<br>
><br>
> - 1sided/halo_1sided_put_win_alloc.c<br>
><br>
> The basis that works. It uses MPI_Put and MPI_Win_fence for<br>
> duplex left/right halo communication.<br>
><br>
> - 1sided/halo_1sided_store_win_alloc_shared.c<br>
><br>
> This is the same, but a shared memory window is used and<br>
> the MPU_Put is substituted by storing the data in the<br>
> neighbors window. Same MPI_Win_fence with same assertions.<br>
><br>
> This does not work, although I'm sure that my assertions are<br>
> correct.<br>
><br>
> Known possibilities:<br>
> - I'm wrong and was not able to understand the assertions<br>
> on MPI-3.0 p452:8-19.<br>
> - I'm wrong because it is invalid to use the MPI_Win_fence<br>
> together with the shared memory windows.<br>
> - mpich has a bug.<br>
> (The first two possibilities are the reason, why I use this<br>
> Forum email list)<br>
><br>
> - 1sided/halo_1sided_store_win_alloc_shared_w-a-cray.c<br>
><br>
> This is a work-around-for Cray that works on our Cray<br>
> and does not use MPI_MODE_NOPRECEDE and MPI_MODE_NOSUCCEED.<br>
> It also runs on another mpich installation.<br>
><br>
> - 1sided/halo_1sided_store_win_alloc_shared_pscw.c<br>
><br>
> Here, MPI_Win_fence is substituted by Post-Start-Complete-Wait<br>
> and it does not work for any assertions.<br>
><br>
> Same possibilities as above.<br>
><br>
> - 1sided/halo_1sided_store_win_alloc_shared_query.c<br>
> - 1sided/halo_1sided_store_win_alloc_shared_query_w-a-cray.c<br>
><br>
> Same as halo_1sided_store_win_alloc_shared.c<br>
> but non-contigues windows are used.<br>
> Same problems as above.<br>
><br>
> - 1sided/halo_1sided_store_win_alloc_shared_othersync.c<br>
><br>
> This version uses the synchronization according to<br>
> #413 and it is tested and works on two platforms.<br>
><br>
> Best regards<br>
> Rolf<br>
><br>
> --<br>
> Dr. Rolf Rabenseifner . . . . . . . . . .. email <a href="mailto:rabenseifner@hlrs.de">rabenseifner@hlrs.de</a><br>
> High Performance Computing Center (HLRS) . phone <a href="tel:%2B%2B49%280%29711%2F685-65530" value="+4971168565530">++49(0)711/685-65530</a><br>
> University of Stuttgart . . . . . . . . .. fax <a href="tel:%2B%2B49%280%29711%20%2F%20685-65832" value="+4971168565832">++49(0)711 / 685-65832</a><br>
> Head of Dpmt Parallel Computing . . . <a href="http://www.hlrs.de/people/rabenseifner" target="_blank">www.hlrs.de/people/rabenseifner</a><br>
> Nobelstr. 19, D-70550 Stuttgart, Germany . . . . (Office: Room 1.307)<br>
><br>
> _______________________________________________<br>
> mpiwg-rma mailing list<br>
> <a href="mailto:mpiwg-rma@lists.mpi-forum.org">mpiwg-rma@lists.mpi-forum.org</a><br>
> <a href="http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma" target="_blank">http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma</a><br>
><br>
><br>
> _______________________________________________<br>
> mpiwg-rma mailing list<br>
> <a href="mailto:mpiwg-rma@lists.mpi-forum.org">mpiwg-rma@lists.mpi-forum.org</a><br>
> <a href="http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma" target="_blank">http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma</a><br>
<br>
--<br>
Dr. Rolf Rabenseifner . . . . . . . . . .. email <a href="mailto:rabenseifner@hlrs.de">rabenseifner@hlrs.de</a><br>
High Performance Computing Center (HLRS) . phone <a href="tel:%2B%2B49%280%29711%2F685-65530" value="+4971168565530">++49(0)711/685-65530</a><br>
University of Stuttgart . . . . . . . . .. fax <a href="tel:%2B%2B49%280%29711%20%2F%20685-65832" value="+4971168565832">++49(0)711 / 685-65832</a><br>
Head of Dpmt Parallel Computing . . . <a href="http://www.hlrs.de/people/rabenseifner" target="_blank">www.hlrs.de/people/rabenseifner</a><br>
Nobelstr. 19, D-70550 Stuttgart, Germany . . . . (Office: Room 1.307)<br>
</div></div><br>_______________________________________________<br>
mpiwg-rma mailing list<br>
<a href="mailto:mpiwg-rma@lists.mpi-forum.org">mpiwg-rma@lists.mpi-forum.org</a><br>
<a href="http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma" target="_blank">http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma</a><br></blockquote></div><br></div>