Dear Dan and Tony and all, based on your both answers, I would recommend to add an errata to 4.0, telling that MPI_Bsend should be used only with communicators in the world model. Then, in the future, there is any freedom to extend MPI_Bsend somehow the session model without the need to define any restriction or non-backward compatible changes. This would be a very small errata and MPI-4.1 is designed to contain such erratas about 4.0. Kind regards Rolf ----- Original Message -----

From: "Dan Holmes, MPI" <[email protected]> To: "MPI Sessions working group" <[email protected]> Cc: "Rolf Rabenseifner" <[email protected]>, "Anthony Skjellum" <[email protected]> Sent: Saturday, September 17, 2022 4:23:34 PM Subject: Re: [mpiwg-sessions] MPI_Buffer_attach in the sessions model

Hi Rolf & Tony,

With current text, MPI-4.0, all sessions and world model buffered send operations would use the same attached buffer. This is likely to be implemented correctly but it breaches the intent of separation of resources used by different sessions, so is not ideal.

A new buffer attach API could be designed to be more in keeping with the intent of sessions. A per-session buffer and a per-communicator buffer are both options to consider. There’s no GitHub issue or PR for this new API yet. Addressing this should fall under the remit of the sessions WG and will probably target MPI-5.0 at the earliest.

Thanks for raising this point.

Cheers, Dan.

Sent from my iPhone

...

On 17 Sep 2022, at 14:18, Tony Skjellum via mpiwg-sessions <[email protected]> wrote:

Rolf, you’ve discovered a problem with the spec… it’s global to attach a buffer, so implicitly associated with ‘world model’ and the communicators derived from MPI_Comm_world …

Each session is to be isolated resource wise … so, we have to make a new API when there is no world model .

The API would take a session key to align buffer attachment to a session in my opinion.

Use of the current api with a session appears erroneous. :-)

I am sure it would work in real implementations nonetheless.

Tony

Anthony Skjellum, PhD 205-807-4968

On Sep 17, 2022, at 8:42 AM, Rolf Rabenseifner via mpiwg-sessions <[email protected]> wrote:

Dear sessions-group,

is there any information about the usage of MPI_Buffer_attach, MPI_Buffer_detach and MPI_Bsend for applications that use the world model together with (several) sessions.

MPI-4.0, the function index entries of MPI_Buffer_attach and the corresponding texts do not give an answer, or do they?

Is it allowed or intended that two MPI_Bsend calls with two different communicators derived not from the same session or world may use portions of the same buffer that was attached with MPI_Buffer_attach?

Is there any issue or pull request already for this problem prepared? Or slides in previous meetings?

Best regards Rolf

-- Dr. Rolf Rabenseifner . . . . . . . . . .. . . [email protected] . High Performance Computing Center (HLRS) . . . ++49(0)711/685-65530 . University of Stuttgart . . . . . . www.hlrs.de/people/rabenseifner . Nobelstr. 19, 70569 Stuttgart, Germany _______________________________________________ mpiwg-sessions mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpiwg-sessions _______________________________________________ mpiwg-sessions mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpiwg-sessions

-- Dr. Rolf Rabenseifner . . . . . . . . . .. . . [email protected] . High Performance Computing Center (HLRS) . . . ++49(0)711/685-65530 . University of Stuttgart . . . . . . www.hlrs.de/people/rabenseifner . Nobelstr. 19, 70569 Stuttgart, Germany