summaryrefslogtreecommitdiffstats
path: root/src/H5Fmpio.c
diff options
context:
space:
mode:
authorRobert Kim Yates <rkyates@llnl.gov>1998-08-29 00:17:20 (GMT)
committerRobert Kim Yates <rkyates@llnl.gov>1998-08-29 00:17:20 (GMT)
commite4053f38c2b2732ffa48b7401581db2516b47951 (patch)
tree1dbfdeaa9d97bdc7ed7729ae890a7be83401a2db /src/H5Fmpio.c
parentc435f3387080614c68392293de0700191f5483d3 (diff)
downloadhdf5-e4053f38c2b2732ffa48b7401581db2516b47951.zip
hdf5-e4053f38c2b2732ffa48b7401581db2516b47951.tar.gz
hdf5-e4053f38c2b2732ffa48b7401581db2516b47951.tar.bz2
[svn-r632] Extended comment about elimination of redundant metadata writes.
Diffstat (limited to 'src/H5Fmpio.c')
-rw-r--r--src/H5Fmpio.c7
1 files changed, 7 insertions, 0 deletions
diff --git a/src/H5Fmpio.c b/src/H5Fmpio.c
index 7214d6b..4362b62 100644
--- a/src/H5Fmpio.c
+++ b/src/H5Fmpio.c
@@ -694,6 +694,13 @@ H5F_mpio_tas_allsame(H5F_low_t *lf, hbool_t newval )
* which is rather klugey.
* Would it be better to pass a parameter to low-level writes
* like H5F_block_write and H5F_low_write, instead? Or...???
+ * Also, when I created this mechanism I wanted to minimize
+ * the difference in behavior between the old way of doing things
+ * (i.e., all procs write) and the new way, so the writes are
+ * eliminated at the very lowest level, here in H5F_mpio_write.
+ * It may be better to rethink that, and short-circuit the writes
+ * at a higher level (e.g., at the points in the code where
+ * H5F_mpio_tas_allsame is called).
*
* Errors:
* IO WRITEERROR MPI_File_write_at failed.