diff options
author | Richard Warren <Richard.Warren@hdfgroup.org> | 2017-10-13 12:58:46 (GMT) |
---|---|---|
committer | Richard Warren <Richard.Warren@hdfgroup.org> | 2017-10-13 12:58:46 (GMT) |
commit | 0c207cd2f84021526690bcbc4e0a8a49ea4cb8a4 (patch) | |
tree | 8cb858355dceffefc201f5dd1d35c0f9b9b7801e /release_docs/RELEASE.txt | |
parent | 087e0d6efa0fdc9e9656d61f5dc2efd62b169e07 (diff) | |
parent | 3dde6d0e32461f46630f814a2fdfbd4c813703bf (diff) | |
download | hdf5-0c207cd2f84021526690bcbc4e0a8a49ea4cb8a4.zip hdf5-0c207cd2f84021526690bcbc4e0a8a49ea4cb8a4.tar.gz hdf5-0c207cd2f84021526690bcbc4e0a8a49ea4cb8a4.tar.bz2 |
Merge pull request #704 in HDFFV/hdf5 from optimize_superblock_read to develop
* commit '3dde6d0e32461f46630f814a2fdfbd4c813703bf':
Updated the code and RELEASE.txt note per comments from John Mainzer
Try to address most of the issues raised by Dana in the code review
Made edits suggested by John as part of the code review
Update the MANIFEST and release_docs/RELEASE files
Add a test for parallel reads of independent files using MPI subgroups
Edits to the file open optimization and associated test code to bring them closer to the HDF5 library's unwritten coding standards. Also bug fix to repair a hang in testphdf5.
The initial coding for the superblock read optization
Diffstat (limited to 'release_docs/RELEASE.txt')
-rw-r--r-- | release_docs/RELEASE.txt | 25 |
1 files changed, 25 insertions, 0 deletions
diff --git a/release_docs/RELEASE.txt b/release_docs/RELEASE.txt index db946a0..db3e4e0 100644 --- a/release_docs/RELEASE.txt +++ b/release_docs/RELEASE.txt @@ -62,6 +62,31 @@ New Features Parallel Library: ----------------- + - Optimize parallel open/location of the HDF5 super-block + + Previous releases of PHDF5 required all parallel ranks to + search for the HDF5 superblock signature when opening the + file. As this is accomplished more or less as a synchronous + operation, a large number of processes can experience a + slowdown in the file open due to filesystem contention. + + As a first step in improving the startup/file-open performance, + we allow MPI rank 0 of the associated MPI communicator to locate + the base offset of the super-block and then broadcast that result + to the remaining ranks in the parallel group. Note that this + approach is utilized ONLY during file opens which employ the MPIO + file driver in HDF5 by previously having called H5Pset_fapl_mpio(). + + HDF5 parallel file operations which do not employ multiple ranks + e.g. specifiying MPI_COMM_SELF (whose MPI_Comm_size == 1) + as opposed to MPI_COMM_WORLD, will not be affected by this + optimization. Conversely, parallel file operations on subgroups + of MPI_COMM_WORLD are allowed to be run in parallel with each + subgroup operating as an independant collection of processes. + + (RAW – 2017/10/10, HDFFV-10294) + + - Large MPI-IO transfers Previous releases of PHDF5 would fail when attempting to |