On Sun, Apr 27, 2025 at 7:04 PM Skjellum, Anthony <[email protected]> wrote:

Jeff, thank you, I will read more about UCC.

In the meantime, since I like MPI a lot, and intend to keep using it, and trying to make it better, I mentioned this because I view MPI_Info arguments is a substitute for polymorphism in the interface, as I know you know, and with AI/machine-learning, we might be able to generate wisdom/profile-guided optimization to fulfill these arguments. Is this not a good direction?

If somebody wants to do this, they can implement it on a per call site inside of the MPI library by running a short backtrace to identify the call site and selecting an algorithm for that call site based on a profile database. On the other hand, nobody cares if it's MPI or not when the standard API of interest to AI is pytorch.distributed.

And, since you probably know that I've brought up orthogonality of the APIs before many times in the meetings, I am worried that we have not made the more general operations like non-blocking have the ability to give assertions or info.

I would pay money for you and Jasper to have a public debate about adding versus subtracting features from MPI ;-) https://urldefense.us/v3/__https://github.com/mpi-forum/mpi-issues/issues/97...

I am mostly interested in more performance in MPI, that's why I added my comment to yours. Do you think it is not a good idea to orthogonalize to support optimization, or is there a better way to achieve higher performance in your opinion.

No, I don't care about orthogonality and I don't prioritize peak performance. If we want orthogonality, we should have put, put-with-notify, wait and test, and implement everything else in terms of those. We are so far from orthogonality and it's fine. We have 5+ kinds of send. We can implement alltoall with scatter or with send-recv or with win_create+fence+put, etc. etc. The MPI standard has so much linear dependency and that's actually fine, because it's solving the problems that our users have. See the following comment about peak performance...

Right now, MPI is slower than vendor-specific primitives, limiting its use on key applications. That also drives my thinking.

And this has always been true. It was true on Blue Gene, yet only the QCD folks wrote in anything but MPI. I regularly gave workshop talks about the merits of DCMF and PAMI, but nobody used them for anything except implementing GASNet and Charm++. I don't think peak performance is feasible for MPI and don't think it's a useful goal for the Forum. What we want is time- and system-averaged performance over all the HPC architectures across decades.

The use case I am thinking of is improved choice of algorithm, for collectives. Or knowing that buffers have specific semantics without the cost of testing.

Yes, allowing expert users to select the algorithm for each call site will show upside in some cases. On the other hand, it will ruin the performance portability of MPI apps, because the best collective algorithm is machine-dependent. At that point, you should just write non-portable code, e.g. UCC. Heck, selecting the best algorithm is not even performance portable across different versions of an MPI library on the same system. jeff

Tony

Anthony Skjellum, PhD Professor of Computer Science Director, Advanced Scalable Computing, Extreme Networks & Data (ASCEND) Center Tennessee Technological University email: [email protected] cell: +1-205-807-4968

------------------------------ *From:* Jeff Hammond <[email protected]> *Sent:* Sunday, April 27, 2025 8:33 AM *To:* Skjellum, Anthony <[email protected]> *Cc:* [email protected] < [email protected]> *Subject:* Re: [mpiwg-languages] static datatypes lifetimes

*External Email Warning*

*This email originated from outside the university. Please use caution when opening attachments, clicking links, or responding to requests.* ------------------------------ I think the API you want then is basically UCC. Prescribe everything for every operation.

Jeff

On Sun, Apr 27, 2025 at 6:10 PM Skjellum, Anthony <[email protected]> wrote:

Another issue: Info arguments are missing on non-blocking and blocking collectives.

Anthony Skjellum, PhD Professor of Computer Science Director, Advanced Scalable Computing, Extreme Networks & Data (ASCEND) Center Tennessee Technological University email: [email protected] cell: +1-205-807-4968

------------------------------ *From:* mpiwg-languages <[email protected]> on behalf of Jeff Hammond via mpiwg-languages < [email protected]> *Sent:* Saturday, April 26, 2025 11:53 PM *To:* [email protected] < [email protected]> *Cc:* Jeff Hammond <[email protected]>; [email protected] <[email protected]> *Subject:* Re: [mpiwg-languages] static datatypes lifetimes

*External Email Warning*

*This email originated from outside the university. Please use caution when opening attachments, clicking links, or responding to requests.* ------------------------------ Attributes on ops, datatypes and files is critical for multiple use cases. It’s ridiculous we don’t have them. The standard is inconsistent without. Jeff Sent from my iPhone On 27. Apr 2025, at 5. 51, Joseph Schuchart via mpiwg-languages <mpiwg-languages@ lists. mpi-forum. org> ZjQcmQRYFpfptBannerStart This Message Is From an External Sender This message came from outside your organization.

ZjQcmQRYFpfptBannerEnd Attributes on ops, datatypes and files is critical for multiple use cases. It’s ridiculous we don’t have them. The standard is inconsistent without.

Jeff

Sent from my iPhone

On 27. Apr 2025, at 5.51, Joseph Schuchart via mpiwg-languages < [email protected]> wrote:

 Unfortunately, there is a catch: MPI_COMM_SELF is only relevant/available/valid in the World Process Model (WPM), i. e. , if using `MPI_Init`/`MPI_Finalize`. In the Sessions process model, predefined communicators are not available. The life-time ZjQcmQRYFpfptBannerStart This Message Is From an External Sender This message came from outside your organization.

ZjQcmQRYFpfptBannerEnd

Unfortunately, there is a catch: MPI_COMM_SELF is only relevant/available/valid in the World Process Model (WPM), i.e., if using `MPI_Init`/`MPI_Finalize`. In the Sessions process model, predefined communicators are not available.

The life-time of datatypes is a known quirk and I think it was discovered after Sessions became part of the standard. They are not bound to any other MPI object and can survive complete shutdown of all sessions / the WPM. IIRC in Open MPI (but have to check), datatypes retain a reference on the internal MPI instance and it is the application's responsibility to free all MPI objects before shutdown. Once the last datatype/session/wpm is gone we release the instance.

I don't like the state of things there and it is problematic. For starters, it prevents complete session isolation (and the benefits that come with it, such as different threading levels). It's not clear to me how that can be rectified and I think the Forum is not clear on that either, which is why we ended up with this weird zombie state. If someone wants to open a ticket to start a discussion on this I'm happy to participate.

For the problem at hand though (as I understand it), maybe it's sufficient to add attributes to datatypes? I don't see why that would be a problem and if it helps with language adoption we have a good argument for it.

Cheers Joseph

On 4/25/25 17:24, Alfredo Correa via mpiwg-languages wrote:

...

Hi Sayan, On Fri, Apr 25, 2025 at 2: 03 PM Ghosh, Sayan <sayan. ghosh@ pnnl. gov> wrote: Consider finalize-delete-callback (this is what Alfredo is alluding to perhaps w. r. t datatype-attached-to-environment) – that seems to rely on MPI_COMM_SELF ZjQcmQRYFpfptBannerStart This Message Is From an External Sender This message came from outside your organization. ZjQcmQRYFpfptBannerEnd Hi Sayan,

On Fri, Apr 25, 2025 at 2:03 PM Ghosh, Sayan <[email protected]> wrote:

* Consider finalize-delete-callback (this is what Alfredo is alluding to perhaps w.r.t datatype-attached-to-environment) – that seems to rely on MPI_COMM_SELF callback (freeing comm-self triggers callback)

That is a good point. At first glance, attaching things to MPI_COMM_WORLD or MPI_COMM_SELF would have a similar effect to attaching stuff to the environment. I didn't think about this because I was reluctant to modify (in any way) either of these special communicators, in particular MPI_COMM_WORLD. But MPI_COMM_SELF might still be a good candidate; others can point out if there is a catch.

Thanks, Alfredo _ _

-- mpiwg-languages mailing list [email protected] https://urldefense.us/v3/__https://lists.mpi-forum.org/mailman/listinfo/mpiw... <https://urldefense.us/v3/__https://lists.mpi-forum.org/mailman/listinfo/mpiwg-languages__;!!G_uCfscf7eWS!ZuNmK6mbS8qplL9AjJRWQ3PTAbAGMusqRU86UgvAro-oRcRR3IP5L-WaNZYOuqTljFttBrcvO2SBSSC6N1BMLIOoHQbouh28UsDH$>

-- Jeff Hammond [email protected] https://urldefense.us/v3/__http://jeffhammond.github.io/__;!!G_uCfscf7eWS!aX...

-- Jeff Hammond [email protected] https://urldefense.us/v3/__http://jeffhammond.github.io/__;!!G_uCfscf7eWS!aX...