Rework MPI Info FAPL preserve PR to use VFD 'ctl' operations #3782
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
jhendersonHDF
added
Merge - To 1.14
Priority - 3. Low 🔽
jhendersonHDF
requested review from
lrknox,
derobins,
byrnHDF,
fortnern,
qkoziol,
vchoi-hdfgroup,
bmribler,
glennsong09,
mattjala and
brtnfld
as code owners
| @@ -143,6 +142,45 @@ H5FD_mpi_get_comm(H5FD_t *file) | |||
| FUNC_LEAVE_NOAPI(ret_value) | |||
| } /* end H5FD_mpi_get_comm() */ | |||
|
|
|||
| /*------------------------------------------------------------------------- | |||
| * Function: H5FD_mpi_get_comm | |||
| * | |||
byrnHDF
approved these changes
Nov 1, 2023
fortnern
approved these changes
Nov 2, 2023
jhendersonHDF
added a commit
to jhendersonHDF/hdf5
that referenced
this pull request
Dec 7, 2023
lrknox
pushed a commit
that referenced
this pull request
Dec 7, 2023
* Preserve MPI-I/O file hints when fapl is closed (#3755) * Fix for issue #3025: Save the MPI info in the file struct so H5Fget_access_plist() can retrieve it from there. * Add compression tests for subfiling (#3769) * Fix typo in comment (#3775) * Fixed a file handle leak in the core VFD (#3779) When opening a file with the core VFD and a file image, if the file already exists, the file check would leak the POSIX file handle. Fixes GitHub issue #635 * Fix a format string warning in the C++ examples (#3776) * Cancel running GitHub workflows on push to same PR (#3772) * Cancel running GitHub workflows on push to same PR * Remove github.sha from workflow concurrency groups * Print some messages in parallel tests on MPI rank 0 only (#3785) Avoids overly verbose output from all processes emitting progress, etc. info. * Avoid attempted use of NULL pointer in parallel compression code (#3786) The parallel compression test code tests for the case where all MPI ranks have no selection in a dataset when writing to it. Add an early exit to the code to avoid attempting to use a NULL pointer due to there being no work to do. * Don't install h5tools_test_utils test program on system (#3793) * Add Doxygen to H5FDsplitter.h (#3794) * H5FD_CURR_SPLITTER_VFD_CONFIG_VERSION * H5FD_SPLITTER_PATH_MAX * H5FD_SPLITTER_MAGIC * H5FD_splitter_vfd_config_t * H5Pset_fapl_splitter() * H5Pget_fapl_splitter() * Update Doxygen initializers & identifiers in VFDs (#3795) * Add Doxygen for all H5FD_<VFD> initializers * Add Doxygen for all H5FD_<VFD>_VALUE values * Mark H5FD_<vfd>_init() calls private in Doxygen * Fix memory corruption in 'MPI I/O FAPL preserve' test (#3806) * Fix usage of h5_clean_files in t_pflush2.c (#3807) * Fix parallel driver check in h5_fixname_real (#3808) * Fix a couple usages of MPI_Info_get (#3809) * Remove H5system.c warning on Windows oneAPI. (#3812) * Add processing of NVHPC flags in linux-gnulibc1 file (#3804) * Disable testing as tests are failing the same as in CMake * Use the current toolchain for examples as default (#3810) * Fix misc. warnings from GCC when compiling with -fsanitize=undefined (#3787) * Set NVHPC maximum optimization level to -O1 for now (#3800) * Set NVHPC maximum optimization level to -O1 for now Compiling HDF5 with NVHPC 23.5 - 23.9 results in test failures in 4 different test files that need to be resolved. Since those tests pass with an optimization level of -O1 (and -O0) and it is currently unclear whether the test failures are due to issues in HDF5 or issues in the 'nvc' compiler, set the maximum optimization level for NVHPC to -O1 until the test failures are resolved. * Disable nvhpc Java testing in CMake and amend known issues * Re-enable testing of Autotools nvhpc * Update some doxygen links to local refs (#3814) * Rework MPI Info FAPL preserve PR to use VFD 'ctl' operations (#3782) * Removed the use of C wrappers from H5P APIs. (#3824) * fix seg fault on frontier/cray * fix seg fault on frontier/cray * fix seg fault on frontier/cray * removed the use of h5pclose_c * removed the use of h5pclose_c * Fortran Wrappers H5VLnative_addr_to_token_f and H5VLnative_token_to_address_f (#3801) * Added H5VLnative_addr_to_token_f and H5VLnative_token_to_address_f * Added H5VLnative_addr_to_token_f and H5VLnative_token_to_address_f tests * Create test for H5Pget_dxpl_mpio (#3825) * Create test and add to testphdf5 * Renamed h5fuse.sh to h5fuse (#3834) * provide an alternative to mapfile for older bash * Disable FP exceptions in H5T init code (#3837) The H5T floating-point datatype initialization code can raise exceptions when handling signaling NaNs. This change disables FE_INVALID exceptions during initialization. Also removes the -ieee=full change for NAG Fortran as that shouldn't be necessary anymore. Fixes #3831 * Add intel oneapi windows build to CI CMake (#3836) * Remove printf format warning on Windows oneAPI. (#3838) * Correct ENV variables (#3841) * Remove Autotools sed hack (#3848) configure.ac contains a sed line that cleans up incorrect library flags which was added to paper over some bugs in earlier versions of the Autotools. These issues are not a problem with the current versions of the Autootols. The sed line causes problems on MacOS, so it has been removed. Fixes #3843 * Make filter unregister callbacks safe for VOL connectors (#3629) * Make filter callbacks use top-level API functions When using VOL connectors, H5I_iterate may not provide valid object pointers to its callback. This change keeps existing functionality in H5Zunregister() without using potentially unsafe pointers. * Filter callbacks use internal API * Skip MPI work on non-native VOL * Add extra space in comments for consistency (#3852) * Add extra space in comments for consistency * uncomment tfloatsattrs test * Update Actions badges to link to relevant workflow (#3850) * Add CMake long double cross-compile defaults (#3683) HDF5 performs a couple of checks at build time to see if long double values can be converted correctly (IBM's Power architecture uses a special format for long doubles). These checks were performed using TRY_RUN, which is a problem when cross-compiling. These checks now use default values appropriate for most non-Power systems when cross-compiling. The cache values can be pre-set if necessary, which will preempt both the TRY_RUN and the default. Affected values: H5_LDOUBLE_TO_LONG_SPECIAL (default no) H5_LONG_TO_LDOUBLE_SPECIAL (default no) H5_LDOUBLE_TO_LLONG_ACCURATE (default yes) H5_LLONG_TO_LDOUBLE_CORRECT (default yes) H5_DISABLE_SOME_LDOUBLE_CONV (default no) Fixes GitHub #3585 * Updates for building and testing VOL connectors * Fix issue with HDF5_VOL_ALLOW_EXTERNAL CMake variable * Initialize parallel testing with MPI_THREAD_MULTIPLE when testing API * Add CMake variable to allow specifying a VOL connector's package name * Remove call to MPI_Init in serial API tests While previously necessary, it now interferes with VOL connectors that may need to be initialized with MPI_THREAD_MULTIPLE * Fixes for CI and presets (#3853) * Change dest for doxygen (#3856) * Implement selection vector I/O with collective chunk filling (#3826) * Changes for ECP-344: Implement selection vector I/O with collective chunk filling. Also fix a bug in H5FD__mpio_write_vector() to account for fixed size optimization when computing max address. * Fixes based on PR review comments: For H5Dchunk.c: fix H5MM_xfree() For H5FDmpio.c: 1) Revert the fix to H5FD__mpio_write_vector() 2) Apply the patch from Neil on the proper length of s_sizes reported by H5FD__mpio_vector_build_types() * Put back the logic of dividing up the work among all the mpi ranks similar to the original H5D__chunk_collective_fill() routine. * Add a test to verify the fix for the illegal reference problem in H5FD__mpio_write_vector(). * Do not publish compression headers or docs (#3865) * Fix typo: look -> loop (#3866) * Moved the README to markdown and expanded its overview of the files, file generation, and other Fortran wrapper development practices as mentioned in the HDF5 architectural document. I added a new figure and included the SVG file and the original xfig file it was generated from. (#3862) * Add HDF5_DISABLE_TESTS_REGEX option to skip tests (#3859) * Fix typo in error message for `MPI_Type_dup`. (#3867) * Complete the `if command line option` sentence. (#3868) * Fix h5dump segmentation fault when --vfd-value option is used (#3873) * Updated URL in funding.yml (#3882) Using new shortened URL, might look better. * Remove unused variable from unmerged changes * Update src/H5Tinit_float.c
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This doesn't fix the test failure with nvhpc and intel classic, but brings the changes more in line with how we retrieve the MPI_Comm object currently by using the VFD 'ctl' callback. The test failure with nvhpc goes away with an optimization level of -O1 or lower, so I still suspect that's due to an nvhpc problem and not something in the library. It appears that Nvidia also does or was building HDF5 with -O1 from https://forums.developer.nvidia.com/t/help-configuring-hdf5-with-nvhpc-no-version-information-available/183413.