[mpich-discuss] nonblocking collectives, progress thread

Rajeev Thakur thakur at mcs.anl.gov
Sat Apr 7 22:09:40 CDT 2012


I think the MPIX calls are not being profiled by Jumpshot and hence you are not seeing the MPIX_Iallreduce.

Rajeev

On Apr 7, 2012, at 2:42 AM, Pieter Ghysels wrote:

> Ah, yes, sorry for the confusion. I am also overlapping the
> MPIX_Iallreduce with other (one-sided) MPI calls. But these calls are
> buried in the library I use, and I'm not so familiar with one sided
> communication. But I was told that there are no calls to MPI_Win_create
> or MPI_Win_free, except at the very beginning and end of the program.
> 
> Is it possible to overlap different MPI calls like that? 
> 
> I configured MPICH2 with 
> --enable-threads=runtime
> and initialized with 
> MPI_Init_thread(argc, argv, MPI_THREAD_MULTIPLE, &thread_support);
> 
> 
> Pieter
> 
> On Fri, 2012-04-06 at 19:27 -0500, Jeff Hammond wrote:
>> One way to implement nonblocking allreduce would be to use
>> MPI_Accumulate, so it doesn't seem entirely nuts.  I presume you know
>> the MPICH2 implementation of MPIX_Iallreduce does not use RMA...
>> 
>> Jeff
>> 
>> On Fri, Apr 6, 2012 at 6:24 PM, Rajeev Thakur <thakur at mcs.anl.gov> wrote:
>>> You mean your program doesn't call MPI_Win_create and MPI_win_free but Jumpshot shows them? No, that's not normal.
>>> 
>>> Rajeev
>>> 
>>> On Apr 6, 2012, at 3:14 PM, Pieter Ghysels wrote:
>>> 
>>>> Dear,
>>>> 
>>>> I'm replacing calls to MPI_Allreduce by calls to MPIX_Iallreduce.
>>>> I'm compiling with -mpe=mpilog and look at the communication pattern
>>>> with jumpshot. The MPI_Allreduce's are gone now of course, but
>>>> mpe/jumpshot does not show MPIX_Iallreduce. Instead, where the
>>>> MPI_Allreduce used to be, now a lot of time is being spent in
>>>> MPI_Win_create and MPI_Win_free, where before the MPI_Allreduce was
>>>> taking place. Is this normal?
>>>> 
>>>> Can I configure MPICH2 to use a progress thread for the nonblocking
>>>> collectives, as in libNBC?
>>>> Do I have to use MPI_Init_thread()?
>>>> 
>>>> Pieter
>>>> 
>>>> _______________________________________________
>>>> 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
>>> 
>>> _______________________________________________
>>> 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
>> 
>> 
>> 
> 
> 
> _______________________________________________
> 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



More information about the mpich-discuss mailing list