FYI to Christoph and Julien, I think that alltoallv/w operations with zero counts are deserving special consideration too … why would we require a barrier ? ; allowing a barrier is acceptable … with neighbors, there is little conceptual difference
ZjQcmQRYFpfptBannerStart
This Message Is From an External Sender
This message came from outside your organization.
 
ZjQcmQRYFpfptBannerEnd
FYI to Christoph and Julien, I think that alltoallv/w operations with zero counts are deserving special consideration too … why would we require a barrier ? ; allowing a barrier is acceptable …  with neighbors, there is little conceptual difference from zero counts for some sources and sinks, so we must be consistent.

The term synchronizing collective means
— must 
Or
— may 
?? That is confusing. 

Clearly barrier must… what about persistent allreduce implemented via 1-sided inside … is there for sure a barrier effect ?

I am worried about this detail to define partitioned collectives too.    

Tony 

Anthony Skjellum, PhD
skjellum@gmail.com
Cell: +1-205-807-4968




On Tue, Mar 18, 2025 at 9:15 AM Christoph Niethammer via mpiwg-hw-topology <mpiwg-hw-topology@lists.mpi-forum.org> wrote:
Hello everyone, During the preparation for the FRM, a question was raised regarding whether and how the semantics of collectives apply to neighborhood collectives. [1] This question stemmed from remark 18) in the summary table in Appendix A. 2,
ZjQcmQRYFpfptBannerStart
This Message Is From an External Sender
This message came from outside your organization.
 
ZjQcmQRYFpfptBannerEnd
Hello everyone,

During the preparation for the FRM, a question was raised regarding
whether and how the semantics of collectives apply to neighborhood collectives. [1]
This question stemmed from remark 18) in the summary table in Appendix A.2,
which was used to apply the semantics of collectives to neighbourhood
collectives at this place.

The discussions expanded during the March meeting to the question of whether remark 18,
stating that "all-to-all collectives *must* synchronize (with some restrictions)",
is correct. [2]

I would like to invite you all to discuss these two issues in more detail during a 
Topologies WG meeting on March 27th at 15:00, as part of our regular WG slot.

Best regards
Christoph Niethammer

[1] https://urldefense.us/v3/__https://github.com/mpi-forum/mpi-issues/issues/863__;!!G_uCfscf7eWS!edo6G3QE7JbjtZTTusbpg5yIgnpGmlkZf6PJXEhK0BIFo_p25xzgJ-iVjVZhEuL7z3exLA62R0MjXhpEVDjC5QAigTtN4tiyDa1MnE8$
[2] https://urldefense.us/v3/__https://github.com/mpi-forum/mpi-issues/issues/971__;!!G_uCfscf7eWS!edo6G3QE7JbjtZTTusbpg5yIgnpGmlkZf6PJXEhK0BIFo_p25xzgJ-iVjVZhEuL7z3exLA62R0MjXhpEVDjC5QAigTtN4tiyWCbeaNg$

--

Dr.-Ing. Christoph Niethammer
High Performance Computing Center Stuttgart (HLRS)
Nobelstrasse 19
70569 Stuttgart

Tel: ++49(0)711-685-87203
email: christoph.niethammer@hlrs.de
https://urldefense.us/v3/__https://www.hlrs.de/people/christoph-niethammer__;!!G_uCfscf7eWS!edo6G3QE7JbjtZTTusbpg5yIgnpGmlkZf6PJXEhK0BIFo_p25xzgJ-iVjVZhEuL7z3exLA62R0MjXhpEVDjC5QAigTtN4tiyVz0SIxo$
_______________________________________________
mpiwg-hw-topology mailing list
mpiwg-hw-topology@lists.mpi-forum.org
https://lists.mpi-forum.org/mailman/listinfo/mpiwg-hw-topology