Problem with MPI I/O with np>1 and large filetypes

[jsquyres]

Per https://www.mail-archive.com/users@lists.open-mpi.org/msg31021.html, Nils Moschuering (github ID unknown) is running into a problem with MPI I/O running on more than 1 rank using very large filetypes.

See the linked email thread for much more detail, including a reproducer program.

@edgargabriel is following up. I am simply filing this issue just for tracking purposes.

[ggouaillardet]

@edgargabriel so far, #3439 fixes the issue when ompio is used (i tested both local and NFS filesystem)
romio314 is fine on a local filesystem, but crashes on NFS (and this is likely a romio issue i will investigate)
could you please give it a try and let me know the outcome (if you can test other parallel filesystems, that would be great)

[ggouaillardet]

Refs pmodels/mpich#2617

[edgargabriel]

Thanks @ggouaillardet , I will try to check it out later. It matches my observation, that ompio crashed in a datatype_release, and if I just simply uncommented those lines, the test finished to the end. That being said, did you have a chance to check whether the output for the last test that was supposed to be erroneous, is truly wrong?

[ggouaillardet]

@edgargabriel the output is 100% consistent, and as far as i can tell, it looks correct.
(and once again, i only tested local filesystem and NFS)

[edgargabriel]

@ggouaillardet I finally found the time to test your patch. I can confirm that the test passes now, thank you! As far as I can tell, I can confirm that the output also looks correct.

[edgargabriel]

@ggouaillardet has this fix been committed to master and the branches?

[ggouaillardet]

@edgargabriel there were two issues

• one that is specific to ROMIO, and even if i guess this is not the one you are interested in, this was fixed in upstream MPICH, and then back ported to all branches from v2.0.x (v1.10 has an older romio and the backport is not really trivial)
• regarding the datatype issue, @bosilca revamped that part and committed this into master.
  there are discussions whether we should PR tov3.x (this breaks ABI compatibility, and the change is a bit large). @bwbarrett was a decision made about that ?
  this change cannnot backported into the v2 branches (since it breaks ABI) and nothing has been done so far. @bosilca shall i issue PR similar to opal/datatype: fix handling of unused opt_desc #3439 but as one-off commits for the v2 branches ?

[bwbarrett]

We're going to pull the ddt change for 3.0.