[mpich-discuss] IBM BG/Q collectives

Rob Latham robl at mcs.anl.gov
Thu Nov 19 14:29:37 CST 2015


On 11/19/2015 10:38 AM, Jen B wrote:
> Thank you for your quick response, Jeff!
>
> I was wondering the values for "short" and "large" message sizes,
> they've been used throughout the paper but not mentioned explicitly. Do
> you happen to know if these values are explicitly mentioned somewhere?
>
> Also, if I get the last algorithm name used for the MPI collective
> routine, is there a way I can for sure tell which algorithm is being
> used? For example, is there any name formatting convention that is used.
>
> Sorry for so many questions. Any help with any of these questions will
> be greatly appreciated!

'short' and 'large' refers to those messages smaller than /greater than 
the EAGER_THRESHOLD.

https://www.alcf.anl.gov/user-guides/tuning-mpi-bgq has some resources 
and documentation. The  "application red book" is probably what you 
want.  Here's a relevant quote:

By default, the MPI send operations use the rendezvous protocol, instead 
of the eager protocol, for messages larger than 2048 bytes.

(there are also immediate and short messages)

==rob

>
> Thank you!
> Jen
>
>
> On Wed, Nov 18, 2015 at 6:06 PM, Jeff Hammond <jeff.science at gmail.com
> <mailto:jeff.science at gmail.com>> wrote:
>
>     A binary tree is the implementation of last resort.
>
>     See these papers for details of the MPI collective implementations
>     on Blue Gene:
>     - http://dx.doi.org/10.1177/1094342014552086
>     - http://dx.doi.org/10.1109/HOTI.2009.12
>     (this is not an inclusive list)
>
>     http://lists.mpich.org/pipermail/discuss/2015-March/003754.html has
>     some related info.
>
>     Jeff
>
>     On Wed, Nov 18, 2015 at 12:53 PM, Jen B <jenbna at gmail.com
>     <mailto:jenbna at gmail.com>> wrote:
>
>         Hello,
>
>         I have a couple of questions about BG/Q collectives and would
>         appreciate any help.
>         I was wondering if Do IBM BG/Q MPI collectives use binomial tree?
>         Once I find the name of the algorithm used by the MPI collective
>         routine using MPIX_Get_last_algorithm_name, is there a quick
>         reference or guide to check how it is implemented?
>
>         Thanks!
>         Jen
>
>
>         _______________________________________________
>         discuss mailing list discuss at mpich.org <mailto:discuss at mpich.org>
>         To manage subscription options or unsubscribe:
>         https://lists.mpich.org/mailman/listinfo/discuss
>
>
>
>
>     --
>     Jeff Hammond
>     jeff.science at gmail.com <mailto:jeff.science at gmail.com>
>     http://jeffhammond.github.io/
>
>     _______________________________________________
>     discuss mailing list discuss at mpich.org <mailto:discuss at mpich.org>
>     To manage subscription options or unsubscribe:
>     https://lists.mpich.org/mailman/listinfo/discuss
>
>
>
>
> _______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss
>

-- 
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA
_______________________________________________
discuss mailing list     discuss at mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss


More information about the discuss mailing list