MPI_INIT & MPI_FINALIZE

[mpiforumbot]

Originally by snir on 2012-01-11 22:11:18 -0600

---

clean text about who needs to invoke them

[mpiforumbot]

Originally by dougmill on 2012-02-08 08:33:32 -0600

---

How about describing this in terms of the program execution instead of MPI or OS processes:

"Entity" is used to describe an execution-instance of code, for example a thread (or in the case of "non-threaded apps", an OS process).

There is exactly one call to MPI_INIT or MPI_INIT_THREAD for each entity that executes main() (for C/C++, or the PROGRAM block for FORTRAN) in an MPI Program. There is exactly one call to MPI_FINALIZE for each call to MPI_INIT/MPI_INIT_THREAD.

We can then state that it is up to the implementations (of both MPI and other paradigms that are interoperable with MPI, such as UPC) to ensure that MPI_INIT (et al.) have reasonable/intended results.

This should cover the case where a UPC implementation/configuration creates multiple threads within an OS process and each one calls main() in the user code (regardless of whether that is actually "main()" in the object code).

[mpiforumbot]

Originally by dougmill on 2012-02-09 09:16:57 -0600

---

With regard to the issue/example of UPC in it's "ultra threaded" mode, I see two "official" references to this on UPC web sites, both essentially say that the "-pthreads" option to upcc is not supported with UPC+MPI programs. This is good reading going forward to make sure we don't get de-railed with things like this.

Berkeley UPC with MPI

UPC Interoperability with MPI

[mpiforumbot]

Originally by snir on 2012-03-04 23:12:25 -0600

---

Attachment added: ticket 313.pdf (215.2 KiB)

[mpiforumbot]

Originally by snir on 2012-03-06 23:39:22 -0600

---

Attachment added: ticket 313 v2.pdf (252.3 KiB)

[mpiforumbot]

Originally by snir on 2012-05-01 21:06:36 -0500

---

Attachment added: 313.pdf (273.7 KiB)
voted text

[mpiforumbot]

Originally by snir on 2012-05-01 21:07:52 -0500

---

Text ready for vote

[mpiforumbot]

Originally by RolfRabenseifner on 2012-06-28 11:13:23 -0500

---

In the pdf, three times, the old name MPI_INFO_KEY is used instead of MPI_INFO_GET_ENV.

This ticket-0-change and a new pdf should be made before we do the 2nd vote.

[mpiforumbot]

Originally by RolfRabenseifner on 2012-07-17 17:37:45 -0500

---

The ticket was missing the MPI_INFO_GET_ENV as constant in the chap-appLang.

I added it in a new table named 'Environmental inquiry info keys'
above the existing table 'Environmental inquiry keys'.

Committed to approved/chap-appLang in r1370.

[mpiforumbot]

Originally by RolfRabenseifner on 2013-01-08 08:27:20 -0600

---

The ticket is not fully completed, or the forum rejected this change and it was not reported on this ticket.

This was also reported in
https://svn.mpi-forum.org/svn/mpi-forum-docs/trunk/meetings/2012-07-jul/mpi3-tickets.xlsx

The attachment:313.pdf contains the following change

• MPI-3.0 p358:5-7 reads

  All MPI programs must contain exactly one call to an MPI initialization routine:
  MPI_INIT or MPI_INIT_THREAD.
  Subsequent calls to any initialization routines are erroneous.

  -and 313.pdf proposed that it should read*

  All MPI programs must contain exactly one call to an MPI
  initialization routine: MPI_INIT or MPI_INIT_THREAD.
  Each MPI process must call an MPI initialization routine,
  MPI_INIT or MPI_INIT_THREAD, exactly once. Subsequent calls by that process to
  any initialization routines are erroneous.

All other changes in 313.pdf were done correctly in MPI-3.0.

The ticket owner may decide how to proceed.

[mpiforumbot]

Originally by RolfRabenseifner on 2013-01-08 10:35:36 -0600

---

Marc Snir: pls close it. I shall not be able to continue pushing it, in any case.

It is already closed. Therefore nothing to do.