From 5550fb330a00ef75b6ab6daa6d8dae8bcf832185 Mon Sep 17 00:00:00 2001 From: Jordan Henderson Date: Wed, 25 Jul 2018 13:16:41 -0500 Subject: RELEASE.txt updates for HDFFV-10467 and HDFFV-10509 --- release_docs/RELEASE.txt | 38 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 38 insertions(+) diff --git a/release_docs/RELEASE.txt b/release_docs/RELEASE.txt index 80be0ab..5ace8a6 100644 --- a/release_docs/RELEASE.txt +++ b/release_docs/RELEASE.txt @@ -197,6 +197,44 @@ Bug Fixes since HDF5-1.10.2 release (DER - 2018/02/26, HDFFV-10356) + - A bug was discovered in the parallel library which caused partial + parallel reads of filtered datasets to return incorrect data. The + library used the incorrect dataspace for each chunk read, causing + the selection used in each chunk to be wrong. + + The bug was not caught during testing because all of the current + tests which do parallel reads of filtered data read all of the data + using an H5S_ALL selection. Several tests were added which exercise + partial parallel reads. + + (JTH - 2018/07/16, HDFFV-10467) + + - A bug was discovered in the parallel library which caused parallel + writes of filtered datasets to trigger an assertion failure in the + file free space manager. + + This occurred when the filter used caused chunks to repeatedly shrink + and grow over the course of several dataset writes. The previous chunk + information, such as the size of the chunk and the offset in the file, + was being cached and not updated after each write, causing the next write + to the chunk to retrieve the incorrect cached information and run into + issues when reallocating space in the file for the chunk. + + (JTH - 2018/07/16, HDFFV-10509) + + - A bug was discovered in the parallel library which caused the + H5D__mpio_array_gatherv() function to allocate too much memory. + + When the function is called with the 'allgather' parameter set + to a non-true value, the function will receive data from all MPI + ranks and gather it to the single rank specied by the 'root' + parameter. However, the bug in the function caused memory for + the received data to be allocated on all MPI ranks, not just the + singular rank specified as the receiver. In some circumstances, + this would cause an application to fail due to the large amounts + of memory being allocated. + + (JTH - 2018/07/16, HDFFV-10467) Configuration ------------- -- cgit v0.12