Hi,
I have uploaded ext3-0.0.5b.tar.gz to
ftp.*.kernel.org:/pub/linux/kernel/people/sct/fs/ext3/test/
and ftp.uk.linux.org:/pub/linux/sct/fs/jfs/test/
New in this release is a functionally complete implementation of
metadata-only journaling: data writes will no longer have to be
written to the journal.
Journaling still imposes some extra serialisation points between
interacting processes, so for some very heavy multi-process workloads,
performance will still be less than ext2's. No performance tuning has
been done on this release. However, I'm getting about 75% of ext2's
sustained performance for 50 client dbench stress tests now, and 100%
of ext2 performance for simple linear writes.
This release includes a new journal superblock format. You can
upgrade old filesystems to the new format by mounting with the option
"-o journal=update".
You have the option of three journaling modes:
-o data=journal
The old-style journaling: all data and metadata is
journaled without exception.
-o data=ordered
Data is not journaled, but IS forced to disk before
a transaction can commit, so after a crash, you will
never see old, stale disk blocks allocated to a file.
-o data=writeback
Data writeback is done completely asynchronously
with respect to transaction commits.
The default mode is "data=journal" for V1 superblocks and
"data=ordered" for V2. Counter-intuitively, ordered journaling is
probably faster than writeback for most workloads for now, since
writeback journaling allows the journal commit and the VM's data
writeback thread to run at the same time, causing extra disk
thrashing.
This is still a testing release --- it's unwise to commit production
data to it, but it is now in a state where testers are welcome to
start attacking it. Quota journaling is not implemented for the
ordered data mode, but other than that all functionality needed for
metadata-only journaling should be in place in this release.
Updated e2fsprogs for the new journal format will follow.
Cheers,
Stephen
