[mpich-discuss] One-sided communication: MPI_Win_lock inside MPI_Win_fence
Ziaul Haque Olive
mzh.olive at gmail.com
Tue May 29 17:33:55 CDT 2012
Hello Jim,
I was coalescing data communication. data is collected in a buffer instead
of sending. let say, size of this temporary buffer size is N. whenever the
buffer is filled up, I need to send all the data from buffer, so that it
can be reused in later iterations. but, without any synchronization it is
not possible to be sure that the buffer is free to reuse. so I used
lock/unlock. moreover, it is required at the start of each iteration that
all the data transfer from remote processes to local window has been
finished. So the fence is used.
the MPI_MODE_NOSUCCEED in the first fence was a typo. it is not present
in the original code.
Thanks,
Ziaul
On Tue, May 29, 2012 at 5:11 PM, Jim Dinan <dinan at mcs.anl.gov> wrote:
> Hi Ziaul,
>
> MPI_MODE_NOSUCCEED is incorrect in the first call to fence since RMA
> operations to succeed this synchronization call. Hopefully this is just a
> typo and you meant MPI_MODE_NOPRECEDE.
>
> In terms of the datatypes, the layout at the origin and target need not be
> the same. However, the basic unit types and the total number of elements
> must match for accumulate.
>
> In the example given below, you shouldn't need lock/unlock in addition to
> the fences. Can you refine this a little more to capture why this is
> needed in your application?
>
> Best,
> ~Jim.
>
>
> On 5/29/12 4:43 PM, Ziaul Haque Olive wrote:
>
>> Hello Rajeev,
>>
>> yes, there is a reason for my program. the I sent you was a
>> simplified version. the original one is little bit different. but this
>> code sometimes work correctly, sometimes do not.
>>
>> I have another question, about indexed data type and MPI_Accumulate,
>>
>> if the indexed data-type for target process contains indexes like,
>>
>> 2, 5, 3, 1 -> out of order
>> or
>> 2, 4, 5, 4, 2, 2 - out of order and repetition
>> or
>> 2, 3, 3, 3, 5 -> in order and repetition
>>
>> would these be a problem?
>>
>> Thanks,
>> Ziaul
>>
>> On Tue, May 29, 2012 at 4:32 PM, Rajeev Thakur <thakur at mcs.anl.gov
>> <mailto:thakur at mcs.anl.gov>> wrote:
>>
>> Nesting of synchronization epochs is not allowed. Is there a reason
>> to do it this way?
>>
>> Rajeev
>>
>> On May 29, 2012, at 4:28 PM, Ziaul Haque Olive wrote:
>>
>> > Hello Rajeev,
>> >
>> > The whole code is bit large, and the code is from graph500
>> benchmark. the bfs_one_sided.c. I am trying to transform it a bit.
>> here is a portion of the code,
>> >
>> > MPI_Win_fence(MPI_MODE_**NOSUCCEED, queue2_win);
>> >
>> > int ii=0,jj,count=1;
>> > acc_queue2_win_MPI_BOR_data[**ii] =
>> masks[VERTEX_LOCAL(w)/elts_**per_queue_bit%ulong_bits];
>> > acc_queue2_win_MPI_BOR_disp[**ii] =
>> VERTEX_LOCAL(w)/elts_per_**queue_bit/ulong_bits;
>> > acc_queue2_win_MPI_BOR_target[**ii] = VERTEX_OWNER(w);
>> >
>> > MPI_Datatype target_type;
>> > MPI_Type_indexed( count, blength ,
>> &acc_queue2_win_MPI_BOR_disp[**ii], MPI_UNSIGNED_LONG, &target_type);
>> > MPI_Type_commit(&target_type);
>> > int dest = acc_queue2_win_MPI_BOR_target[**ii];
>> > MPI_Win_lock(MPI_LOCK_**EXCLUSIVE, dest, 0, queue2_win );
>> >
>> > MPI_Accumulate(&acc_queue2_**win_MPI_BOR_data[ii], count,
>> MPI_UNSIGNED_LONG, dest, 0, 1,target_type, MPI_BOR, queue2_win);
>> >
>> >
>> > MPI_Win_unlock(dest, queue2_win );
>> > MPI_Type_free(&target_type);
>> >
>> > MPI_Win_fence(MPI_MODE_**NOSUCCEED, queue2_win);
>> >
>> >
>> > Let me know if it works.
>> >
>> > Thanks
>> > Ziaul
>> >
>> > On Tue, May 29, 2012 at 4:13 PM, Rajeev Thakur
>> <thakur at mcs.anl.gov <mailto:thakur at mcs.anl.gov>> wrote:
>> > Can you send the complete program if it is small.
>> >
>> > Rajeev
>> >
>> > On May 29, 2012, at 2:54 PM, Ziaul Haque Olive wrote:
>> >
>> > > for smaller number of processes like 4, i was getting correct
>> result, but for 8, it was providing incorrect result.
>> > >
>> > > I tried with and without lock/unlock. without lock/unlock
>> provides correct result all the time.
>> > >
>> > > Hello,
>> > >
>> > > I am getting incorrect result while using lock/unlock
>> synchronization inside fence. the pattern is as follows,
>> > >
>> > > MPI_Win_fence(win1);
>> > > ..........
>> > > MPI_Win_lock(exclusive, win1);
>> > >
>> > > MPI_Accumulate(MPI_BOR, win1);
>> > >
>> > > MPI_Win_unlock(win1);
>> > >
>> > > MPI_Win_fence(win1);
>> > >
>> > > is it invalid to use lock in this way?
>> > >
>> > > Thanks,
>> > > Ziaul.
>> > >
>> > > ______________________________**_________________
>> > > mpich-discuss mailing list mpich-discuss at mcs.anl.gov
>> <mailto:mpich-discuss at mcs.anl.**gov <mpich-discuss at mcs.anl.gov>>
>>
>> > > To manage subscription options or unsubscribe:
>> > > https://lists.mcs.anl.gov/**mailman/listinfo/mpich-discuss<https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss>
>> >
>> > ______________________________**_________________
>> > mpich-discuss mailing list mpich-discuss at mcs.anl.gov
>> <mailto:mpich-discuss at mcs.anl.**gov <mpich-discuss at mcs.anl.gov>>
>>
>> > To manage subscription options or unsubscribe:
>> > https://lists.mcs.anl.gov/**mailman/listinfo/mpich-discuss<https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss>
>> >
>> > ______________________________**_________________
>> > mpich-discuss mailing list mpich-discuss at mcs.anl.gov
>> <mailto:mpich-discuss at mcs.anl.**gov <mpich-discuss at mcs.anl.gov>>
>>
>> > To manage subscription options or unsubscribe:
>> > https://lists.mcs.anl.gov/**mailman/listinfo/mpich-discuss<https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss>
>>
>> ______________________________**_________________
>> mpich-discuss mailing list mpich-discuss at mcs.anl.gov
>> <mailto:mpich-discuss at mcs.anl.**gov <mpich-discuss at mcs.anl.gov>>
>>
>> To manage subscription options or unsubscribe:
>> https://lists.mcs.anl.gov/**mailman/listinfo/mpich-discuss<https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss>
>>
>>
>>
>>
>> ______________________________**_________________
>> mpich-discuss mailing list mpich-discuss at mcs.anl.gov
>> To manage subscription options or unsubscribe:
>> https://lists.mcs.anl.gov/**mailman/listinfo/mpich-discuss<https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss>
>>
> ______________________________**_________________
> mpich-discuss mailing list mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/**mailman/listinfo/mpich-discuss<https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20120529/a9f069d1/attachment.html>
More information about the mpich-discuss
mailing list