Hi Joachim, Personally, I think it is easier to make sure we understand and agree on the execution ordering requirements at each MPI process first, and only then tackle the possible matching orderings that are permitted to result from that execution order. IIRC, the MPI_THREAD_CONCURRENT proposal mixes the two concerns (execution ordering and matching ordering) together and thereby invites ire from several angles at once. The "mpi_assert_allow_overtaking" is a great way to circumvent the matching ordering aspects of this topic, but it isn't a complete solution because it is permitted for the user to omit it (and for MPI to ignore it) and the outcome must still be well-defined, understandable, and useful. Best wishes, Dan. -----Original Message----- From: Joachim Jenke <[email protected]> Sent: 10 May 2023 11:07 To: Hybrid working group mailing list <[email protected]>; Holmes, Daniel John <[email protected]> Cc: Jeff Hammond <[email protected]> Subject: Re: [mpiwg-hybridpm] Meeting Tomorrow What is the difference of MPI_THREAD_CONCURRENT and MPI_THREAD_MULTIPLE + "mpi_assert_allow_overtaking"? If the message ordering is not important for the application, it should just set the assertion for the respective communicator and will get the same possible performance advantage as MPI_THREAD_CONCURRENT might provide. - Joachim Am 10.05.23 um 11:57 schrieb Jeff Hammond via mpiwg-hybridpm:

What I gathered from the live debate a few months back is that some people want a semantic that is different from my (and others') interpretation of MPI_THREAD_MULTIPLE.  Rather than fight forever about what MPI_THREAD_MULTIPLE means, why don't the people who want the more relaxed semantic just propose that as MPI_THREAD_CONCURRENT, as was discussed back in 2016.

I will not quarrel one bit with a new thread level, MPI_THREAD_CONCURRENT, that does what Maria's team wants, but I intend to fight until my dying breath against any change to the MPI standard that violates the "as if in some order" text.

Jeff

...

On 10. May 2023, at 12.50, Holmes, Daniel John <[email protected]> wrote:

Hi Jeff,

1. Yes 2. If only it were that simple

Your first quote is compromised by the example 11.17 that follows it: which order is the interleaved execution mimicking? MPI_SEND;MPI_RECV and MPI_RECV;MPI_SEND both result in deadlock (as stated in that example). That sentence needs "interpretation" - your wording is slightly better, but it needs to be something like "as if the stages of the operations were executed atomically in some order". The initiation and starting stages of both the send and receive operations are local and can happen without dependence on anything else (in particular, without dependence on each other). Once that has happened, both operations are enabled and must complete in finite time. The execution outcome is "as if" each of the blocking procedures were replaced with a nonblocking initiation and completion procedure pair and both initiation procedures were executed (in some order) before the completion procedures (in some order). The observation and clarification above is necessary, but not sufficient, for resolving the logically concurrent issue. It speaks to execution ordering, but not to message matching ordering. However, rather than mash everything together (we've been down that road, we know where it leads), we could consider the merits of just this adjustment on its own. We could call it "The two MPI thread-safety rules conflict with each other." Best wishes, Dan. *From:*Jeff Hammond <[email protected] <mailto:[email protected]>> *Sent:*10 May 2023 07:28 *To:*MPI Forum <[email protected] <mailto:[email protected]>> *Cc:*Holmes, Daniel John <[email protected] <mailto:[email protected]>> *Subject:*Re: [mpiwg-hybridpm] Meeting Tomorrow "All MPI calls are thread-safe, i.e., two concurrently running threads may make MPI calls and the outcome will be as if the calls executed in some order, even if their execution is interleaved." I'm going to continue to die on the hill that "as if executed in some order" constrains the implementation behavior to something equivalent to "MPI operations are initiated atomically" because otherwise one cannot be guaranteed that some ordering exists.  The text about logically concurrent merely explains the obvious to users: it is impossible to know in what order unsynchronized threads execute operations.  The previous sentence makes it clear what is meant by logically concurrent, and it is consistent with Chapter 11, i.e. it logically unordered: "...if the process is multithreaded, then the_semantics of thread execution may not define a relative order between two send operations executed by two distinct threads_. The operations are logically concurrent..."

I can't provide the full history of the Intel instruction ENQCMD <https://community.intel.com/legacyfs/online/drupal_files/managed/c5/15/architecture-instruction-set-extensions-programming-reference.pdf> but it appears to address the problem of a large number of semi-independent HW units initiating MPI operations in a manner compliant with the text above. As I have stated previously, it is possible to relax the constraint "as if executed in some order" with the addition of a new threading level, which Pavan proposed years ago as MPI_THREAD_CONCURRENT <https://github.com/mpiwg-sessions/sessions-issues/wiki/2016-06-07-forum#notes-from-meeting-bold--specific-work-to-do> (although details are impossible to find at this point). Jeff

On 9. May 2023, at 17.41, Holmes, Daniel John via mpiwg-hybridpm <[email protected] <mailto:[email protected]>> wrote:

Hi all,  Unfortunately, I am double-booked for tomorrow's HACC WG time slot - so my answer to the implied question below is "not yet".  The "logically concurrent isn't" issue #117 is now accepted and merged into MPI-4.1 (take a moment to celebrate!) - but it just says "here be dragons".  Do we care enough to defeat those dragons?  Argument FOR: as systems become more heterogeneous, an MPI process is likely to abstractly "contain" more semi-independent HW units that will want to communicate with other MPI processes, which will result in lots of logically concurrent MPI communication operations - exactly the territory in which these dragons live.  Argument AGAINST: we've been throwing brave warriors into this particular dragon fire for about a decade and we've only now convinced ourselves that the dragons do, in fact, exist. How many more volunteers do we have and do they have sufficiently pointy sticks?  Best wishes, Dan.  From: mpiwg-hybridpm <[email protected] <mailto:[email protected]>> On Behalf Of Jim Dinan via mpiwg-hybridpm Sent: 09 May 2023 15:15 To: Hybrid working group mailing list <[email protected] <mailto:[email protected]>> Cc: Jim Dinan <[email protected] <mailto:[email protected]>> Subject: [mpiwg-hybridpm] Meeting Tomorrow  Hi All,  We had to reschedule the topic planned for tomorrow's meeting, so the agenda is now open. Please let me know if you have a topic you'd like to discuss. If we don't have a topic ahead of time, we will cancel.  Thanks,  ~Jim. _______________________________________________ mpiwg-hybridpm mailing list [email protected] <mailto:[email protected]> https://lists.mpi-forum.org/mailman/listinfo/mpiwg-hybridpm <https://lists.mpi-forum.org/mailman/listinfo/mpiwg-hybridpm>

_______________________________________________ mpiwg-hybridpm mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpiwg-hybridpm

-- Dr. rer. nat. Joachim Jenke IT Center Group: High Performance Computing Division: Computational Science and Engineering RWTH Aachen University Seffenter Weg 23 D 52074 Aachen (Germany) Tel: +49 241 80- 24765 Fax: +49 241 80-624765 [email protected] www.itc.rwth-aachen.de