diff options
Diffstat (limited to 'release_docs/RELEASE.txt')
| -rw-r--r-- | release_docs/RELEASE.txt | 81 |
1 files changed, 80 insertions, 1 deletions
diff --git a/release_docs/RELEASE.txt b/release_docs/RELEASE.txt index 1734c01..f228e39 100644 --- a/release_docs/RELEASE.txt +++ b/release_docs/RELEASE.txt @@ -47,6 +47,14 @@ New Features Configuration: ------------- + - Improved support for Intel oneAPI + + * Separates the old 'classic' Intel compiler settings and warnings + from the oneAPI settings + * Uses `-check nouninit` in debug builds to avoid false positives + when building H5_buildiface with `-check all` + * Both Autotools and CMake + - Added new options for CMake and Autotools to control the Doxygen warnings as errors setting. @@ -127,6 +135,16 @@ New Features Library: -------- + - Added a simple cache to the read-only S3 (ros3) VFD + + The read-only S3 VFD now caches the first N bytes of a file stored + in S3 to avoid a lot of small I/O operations when opening files. + This cache is per-file and created when the file is opened. + + N is currently 16 MiB or the size of the file, whichever is smaller. + + Addresses GitHub issue #3381 + - Added new API function H5Pget_actual_selection_io_mode() This function allows the user to determine if the library performed @@ -159,7 +177,8 @@ New Features - Fortran async APIs H5A, H5D, H5ES, H5G, H5F, H5L and H5O were added. - Added Fortran APIs: - h5pset_selection_io_f, h5pget_selection_io_f + h5pset_selection_io_f, h5pget_selection_io_f, + h5pget_actual_selection_io_mode_f, h5pset_modify_write_buf_f, h5pget_modify_write_buf_f - Added Fortran APIs: @@ -219,6 +238,66 @@ Bug Fixes since HDF5-1.14.2 release =================================== Library ------- + - Fixed some issues with chunk index metadata not getting read + collectively when collective metadata reads are enabled + + When looking up dataset chunks during I/O, the parallel library + temporarily disables collective metadata reads since it's generally + unlikely that the application will read the same chunks from all + MPI ranks. Leaving collective metadata reads enabled during + chunk lookups can lead to hangs or other bad behavior depending + on the chunk indexing structure used for the dataset in question. + However, due to the way that dataset chunk index metadata was + previously loaded in a deferred manner, this could mean that + the metadata for the main chunk index structure or its + accompanying pieces of metadata (e.g., fixed array data blocks) + could end up being read independently if these chunk lookup + operations are the first chunk index-related operation that + occurs on a dataset. This behavior is generally observed when + opening a dataset for which the metadata isn't in the metadata + cache yet and then immediately performing I/O on that dataset. + This behavior is not generally observed when creating a dataset + and then performing I/O on it, as the relevant metadata will + usually be in the metadata cache as a side effect of creating + the chunk index structures during dataset creation. + + This issue has been fixed by adding callbacks to the different + chunk indexing structure classes that allow more explicit control + over when chunk index metadata gets loaded. When collective + metadata reads are enabled, the necessary index metadata will now + get loaded collectively by all MPI ranks at the start of dataset + I/O to ensure that the ranks don't unintentionally read this + metadata independently further on. These changes fix collective + loading of the main chunk index structure, as well as v2 B-tree + root nodes, extensible array index blocks and fixed array data + blocks. There are still pieces of metadata that cannot currently + be loaded collectively, however, such as extensible array data + blocks, data block pages and super blocks, as well as fixed array + data block pages. These pieces of metadata are not necessarily + read in by all MPI ranks since this depends on which chunks the + ranks have selected in the dataset. Therefore, reading of these + pieces of metadata remains an independent operation. + + - Fixed potential hangs in parallel library during collective I/O with + independent metadata writes + + When performing collective parallel writes to a dataset where metadata + writes are requested as (or left as the default setting of) independent, + hangs could potentially occur during metadata cache sync points. This + was due to incorrect management of the internal state tracking whether + an I/O operation should be collective or not, causing the library to + attempt collective writes of metadata when they were meant to be + independent writes. During the metadata cache sync points, if the number + of cache entries being flushed was a multiple of the number of MPI ranks + in the MPI communicator used to access the HDF5 file, an equal amount of + collective MPI I/O calls were made and the dataset write call would be + successful. However, when the number of cache entries being flushed was + NOT a multiple of the number of MPI ranks, the ranks with more entries + than others would get stuck in an MPI_File_set_view call, while other + ranks would get stuck in a post-write MPI_Barrier call. This issue has + been fixed by correctly switching to independent I/O temporarily when + writing metadata independently during collective dataset I/O. + - Fixed a bug with the way the Subfiling VFD assigns I/O concentrators During a file open operation, the Subfiling VFD determines the topology |