I think the default is mainly due to the fact that only minimal
testing has been done of the more optimized path. -- Rob
On Jan 5, 2011, at 5:57 AM, Julian Kunkel wrote:
Hi,
thanks for all your responses.
Now I understand that there is no additional buffering.
Each op had 512-bytes each (likely to encapsulate 64 in memory
noncontig double values) and there were 125 operations. Sorry if I did
not explain that properly.
I rechecked and I used MPICH 1.3.1 and not 1.3.2, sorry for that.
Now I used gdb to trace the behavior, it calls:
156 ADIO_WriteStrided(fh, buf, count, datatype,
file_ptr_type,
which ends in ADIOI_PVFS2_WriteStrided:
/* Use classic list I/O if no hints given base case */
ADIOI_PVFS2_OldWriteStrided(...)
The 64 matches the define MAX_OL_COUNT 64 in ad_pvfs2_io_list.c.
I have seen that the code before checks:
/* Use list I/O in the base case */
if (fd->hints->fs_hints.pvfs2.listio_write == ADIOI_HINT_ENABLE) {
ret = ADIOI_PVFS2_WriteStridedListIO(fd, buf, count, datatype,
file_ptr_type, offset, status, error_code);
return;
}
I checked where the HINT_ENABLE is set, it is set in:
ADIOI_PVFS2_SetInfo
However there is the code & comment:
/* disable the aggressive strided optimizations by default */
...
fd->hints->fs_hints.pvfs2.listio_write = ADIOI_HINT_DISABLE;
It seems to be enabled only if the hint "romio_pvfs2_listio_write" is
set to "enable",
so no wonder that the list-I/O is used.
The same for the read path.
Is there a particular reason to disable this useful feature by
default, or is it a experimental feature?
Is there some kind of (end-user) documentation about this behavior?
Unfortunately, Google does not find useful links if I google for
romio_pvfs2_listio_write.
I wonder because I guess normal users might not know about these
optimizations, they might expect that these features are enabled by
default (because they show some strength of PVFS/Orangefs).
Regards,
Julian
_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers