Re: 2.6.6-mm5 oops mounting ext3 or reiserfs with -o barrier

[Jens Axboe <axboe@suse.de>]

On Sun, May 23 2004, Lorenzo Allegrucci wrote:
> On Sunday 23 May 2004 18:56, Jens Axboe wrote:
> 
> > > Untar, read, copy and remove the OpenOffice tarball, each test
> > > run with cold cache (mount/umount cycle).
> >
> > I understand that, I just don't see how you can call it a regression.
> > It's a given that barrier will be slower.
> 
> I'm sorry, I didn't know :)
> 
> I read from www.kerneltrap.org:
> 
> Request barriers, also known as write barriers, provide a mechanism for 

[snip]

Ah ok, I see the confusion! ext3 does nothing clever with barriers, it
just uses it for data integrity. A journalled fs is normally not safe to
run on write back cached drives if they aren't battery backed. You could
try reiser instead, it's has a more intelligent use of barriers.

> > > > but yes of course -o barrier=1 is going to
> > > > be slower than default + write back caching. What you should compare is
> > > > without barrier support and hdparm -W0 /dev/hdX, if -o barrier=1 with
> > > > caching on is slower then that's a regression :-)
> > >
> > > hdparm -W0 /dev/hda
> > >
> > > ext3 (-o barrier=0)
> > > untar		read		copy		remove
> > > 1m55.190s	0m27.633s	2m19.072s	0m21.348s
> > > 0m7.081s	0m1.189s	0m0.724s	0m0.083s
> > > 0m6.502s	0m3.244s	0m9.715s	0m1.633s
> > >
> > > ext3 (-o barrier=1)
> > > untar		read		copy		remove
> > > 1m55.358s	0m23.831s	2m16.674s	0m21.508s
> > > 0m7.153s	0m1.200s	0m0.748s	0m0.087s
> > > 0m6.775s	0m3.358s	0m9.985s	0m1.781s
> > >
> > >
> > > haparm -W1 /dev/hda
> > >
> > > ext3 (-o barrier=0)
> > > untar		read		copy		remove
> > > 0m55.405s	0m26.230s	1m28.765s	0m20.766s
> > > 0m7.195s	0m1.199s	0m0.773s	0m0.081s
> > > 0m6.502s	0m3.359s	0m9.672s	0m1.868s
> > >
> > > ext3 (-o barrier=1)
> > > untar		read		copy		remove
> > > 0m52.117s	0m28.502s	1m51.153s	0m25.561s
> > > 0m7.231s	0m1.209s	0m0.738s	0m0.071s
> > > 0m6.117s	0m3.191s	0m9.347s	0m1.635s
> >
> > Your results look a bit over the map, how many runs are your averaging
> > for each one?
> 
> Just one run, no averaging.
> Yes, it's not a scientific approach, but I have not enough time
> and this is my production machine :)
> By experience I can say that numbers between each run are quite
> stable and reproducible.

It just looks odd that eg reads vary as much as they do, and that -o
barrier=1 makes -W0 reads faster (and faster then -W1 even). remove
looks reasonable for -W1, but -W0 is still faster. That is _really_ odd.

-- 
Jens Axboe