New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
wrong attributes on MPI_win_create() on big-endian systems? #1446
Comments
can you please confirm the failure occurs in the mpi4py test suite ? |
Correct. This is on the mpi4py test suite |
I will double check that tomorrow |
@amckinstry i was able to reproduce the issue with a C program, but i could not (yet) build mpi4py on a big endian box. |
…and MPI_WIN_MODEL Thanks Alastair McKinstry for the report.
Thanks. Submitting an experimental version through our build infra now. Should have confirmation by Monday |
@amckinstry Sweet. We're actually debating this fix over on open-mpi/ompi-release#1018 -- I see what @ggouaillardet did as a good code cleanup, but I confess to not understanding why it would have fixed your bug. If you can confirm / deny that it actually fixed the bug, that would be great. Thank you! |
@ggouaillardet Any chance you could send me the full list of failing tests? |
@ggouaillardet @jsquyres The patch is OK as long as A better implementation of the attribute engine should also handle the case of predefined attributes whose value is a C A quick workaround that would not require changing too much the codebase would be to asume that either @ggouaillardet I'm a little busy as to write a full patch myself, but assuming I was clear enough above, you could write it in 5 minutes and I'll happily review it. |
@ggouaillardet Apply the following patch to mpi4py sources from release tarball, I expect it to workaround the issue. PS: I wrote that workaround in mpi4py assuming that if Open MPI 1.10.3 is ever released, it will include a proper fix for this issue. Otherwise, the issue will pop-up again and we will need to update the preprocessor version check guard. |
I've tested the patch on s390x, and unfortunately it segfaults: testIsThreadMain (test_threads.TestMPIThreads) ... ok I'm still awaiting an update of openmpi (experimental) on the test regards On 13/03/2016 12:48, Lisandro Dalcin wrote:
Alastair McKinstry, alastair@sceal.ie, mckinstry@debian.org, https://diaspora.sceal.ie/u/amckinstry |
@amckinstry Do you have shell access to the machine? Any chance you can run it under valgrind? Otherwise, could you add a quick printf() for the value of the |
@dalcinl I was unable to test that today. anyway, what about
that being said, I have no idea what will happen if/when attrval is free'd |
Just to confirm: I got access to an s390x with the patched openmpi regards On 11/03/2016 14:07, Jeff Squyres wrote:
Alastair McKinstry, alastair@sceal.ie, mckinstry@debian.org, https://diaspora.sceal.ie/u/amckinstry |
@amckinstry @dalcinl the fix has landed into the v1.10 and v2.x branches, which it means it will be available when Open MPI 1.10.3 and 2.0.0 are released i will then close this issue from now |
I'm debugging a test issue on mpi4py (2.0.0) linked against libopenmpi 1.10.2 in Debian.
This fails on a units on mpi_win_* on some archs.
In particular s390x, sparc64, ppc64, but works on ppc64el.
This implies a big-endian problem. looking at mpi4py, it appears that certain attrs are being set to 0;
unit (disp_unit) is 0, not 1 in the testAttributes() test (after MPI_Win_Create, with unit=1 AFAIK)
testCreateFlavor() fails as MPI.WIN_CREATE_FLAVOR is 0 not one of MPI.WIN_FLAVOR_CREATE, MPI.WIN_FLAVOR_ALLOCATE, MPI.WIN_FLAVOR_DYNAMIC, MPI.WIN_FLAVOR_SHARED,)
( or MPI.KEYVAL_INVALID).
I'm not sure at this stage whether the bug is in openmpi or mpi4py, as I don't see any mpi_win_* tests being run on build (the debian build autodetects and runs make check/test). What tests do you suggest to check out the openmpi 1.10.2 builds on these archs?
The text was updated successfully, but these errors were encountered: