HDF5 collective I/O on BlueGene

Hi BlueGene users in this group: I know BlueGene is not "officially
supported by the HDF group" but I also know there are more than a few
folks using HDF5 on that platform.

If you are using collective mode to write out an HDF5 dataset -- and
on that platform I cannot over-emphasize how important it is that you
do so -- you will likely encounter an unfortunate data-corrupting bug.
I've fixed this bug, but it will take some time to get the fix into a
real IBM update.

If your BlueGene/P is running "v1r3", you might have to disable
collective I/O until the fix for this bug gets pushed out. The
pending "v1r4" update (currently scheduled for september 2009) already
contains a fix.



Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA