Linux ext3 performance (Re: fsync() takes about 0.06 second ...)

[Mika Iisakkila]

Jeremy Rumpf wrote:
> I would be really curious for you to run your tests with ext3 running in 
> "writeback" journaling mode.

Well, I've got this laptop which I'd actually _want_ to reinstall
from scratch, so I made some unscientific, but brave experiments :-)

Test platform:

P3 / 700MHz / 128 MB, laptop IDE (rather slow)
Linux kernel 2.4.19
Berkeley DB 4.0.14
Cyrus IMAPD 2.1.9
	mboxlist.db = skiplist
	subs.db = flat
	seen.db = skiplist
	duplicate.db = db3-nosync
File system: ext3 + internal journal of default size

Test software: http://www.jplanglois.com/products/imapcp/
copying folders of various sizes from another IMAP account

With the default mode (data=ordered), performance was rather
weak, with maybe 2-3 messages copied per second. All messages
seemed to go "right through" to the disk. Kernel buffers
didn't grow much, and the hard disk rattled in unison with
the copy progress.

Next I tried data=writeback. Performance was abysmal, often
slowing down to maybe one message per two seconds. See above
for buffers and noise.

Data=journal was the clear winner; copy speed with average
size text messages (archive of this list, actually) stayed
mostly at a nice 20+ messages per second. Kernel buffering
actually seemed to do something useful this time, growing
to 20-30 megabytes and the hard disk light didn't stay on
continuously.

The copy source, which also ran the copy program, was no screamer,
it's UW-IMAPD running on a PPRo/200 on Linux. I don't know
which one actually was the bottleneck in the last case.

On the brighter side, the file system of the laptop didn't get
corrupted :-) I still don't know if data=journal is completely safe
yet on 2.4.19, though... Anyone know for _sure_?

--mika