diff options
author | John Mainzer <mainzer@hdfgroup.org> | 2004-08-31 22:12:07 (GMT) |
---|---|---|
committer | John Mainzer <mainzer@hdfgroup.org> | 2004-08-31 22:12:07 (GMT) |
commit | da9fd4bb1800ee29266a9b1b2201fcab4b7d94f4 (patch) | |
tree | 2e3f812b4e8f8c17ca8e36e72298cfc21fe77ef4 /README.txt | |
parent | 2202352d2840a4abbbd85ab654124194cade2b07 (diff) | |
download | hdf5-da9fd4bb1800ee29266a9b1b2201fcab4b7d94f4.zip hdf5-da9fd4bb1800ee29266a9b1b2201fcab4b7d94f4.tar.gz hdf5-da9fd4bb1800ee29266a9b1b2201fcab4b7d94f4.tar.bz2 |
[svn-r9178] Purpose:
Fix parallel bug reported by Thomas Guignon, in which different
processes became confused as to whether they were doing collective
or individual I/O.
Description:
When one process had a point selection, and another didn't, the
first concluded that the I/O was independant, while the second
presumed that it was collective. A hang resulted.
Solution:
Get all processes involved in an I/O to compare notes. If all
agree that the I/O is collective, they proceed with same. If
any think the I/O should be independant, all use independant
I/O.
Note that this is an interrim fix -- the correct solution is to
support collective I/O on point selections. This will take a
while.
Platforms tested:
copper
h5committested
Eirene (serial and parallel)
In the parallel test on Eirene, I encountered a bug in h5repacktst.
However the problem vanished on recompile. Since I couldn't reproduce
it elsewhere, I went ahead with the checkin.
Given my druthers, I would have liked to study the code more
carefully before this check-in. However, there is some time pressure.
The new code implementing the consensus check must not be executed
unless MPI is initialized, and there is a communicator associated
with the file. I think my guards against this case are adequate,
but if we run into a hang or an illegal instruction error, this
change should be suspect.
Misc. update:
Diffstat (limited to 'README.txt')
0 files changed, 0 insertions, 0 deletions