Date:	Thu, 14 Sep 2000 17:03:18 +0100
From:	"Stephen C. Tweedie" <sct@redhat.com>
To:	linux-fsdevel@vger.kernel.org
Subject: ext3 metadata-only journaling, snapshot#1

Hi,

Quick update --- ext3 metadata-only journaling is working, after a
fashion.  It is not recovery-proof just yet, so it isn't releasable,
but the current code is complete except for the need for certain log
records to be made when metadata gets deleted (to prevent old log
entries from overwriting data on a crash).

Current straight dd performance for the first cut, totally untuned
code: 

	169.8365 MB in 21.4378 secs, 7.9223 MB/sec

versus

	412.2214 MB in 43.8004 secs, 9.4114 MB/sec

for ext2.  There's still a way to go, but this is without any
performance tuning at all so far.

dbench shows poorer performance: only about 2/3rds of the performance
of ext2.  This appears to be due to the large amount of caching and
deleting going on in the dbench test --- the stricter ordering
requirements of ext3 make it a lot harder to do all of the truncate
optimisations that ext2 can make.

Right now, we have bdflush() trying to sync data buffers as well as
the journal thread, and that is probably the cause of a certain amount
of the performance degradation.  I'll have to look closely now at the
possibility of delegating more of the writeback functionality to the
filesystem to avoid overlapping syncs happening --- this will
definitely require extra VM-level support.

If anyone really wants to see the code, have a look at the 0.0.3b
tarball and 0.0.4a patch in

	ftp.uk.linux.org:/pub/linux/sct/fs/jfs/test/

The amount of changed code is relatively short --- the jfs layer
already has most of the support, and I was surprised at how much of
that support actually worked once I started hammering on it. :-) 

Warning: it's called DONTUSE for a reason!  The fact that it completed
several straight dbench runs without a problem does not mean that I
trust this code yet...

Cheers,
 Stephen
-
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@vger.kernel.org