ufs+softupdates / consistency

[David Schultz]

On Wed, Jan 26, 2005, Arne WXrner wrote:
> --- David Schultz <das at FreeBSD.ORG> wrote:
> > ext2fs mounted async does not provide consistency; in fact the
> > state of the disk can be almost arbitrarily inconsistent at any
> > given time.  Soft updates is supposed to provide performance
> > comparable to async writes without the inconsistency problem. 
> > I'm not sure what it is about your setup that causes such a
> > disparity. (Many factors such as the FS block size and ATA write
> > caching can make a big difference.)
> >
> Somebody in list freebsd-performance@ opened a thread "FreeBSD 5.3
> I/O Performance / Linux 2.6.10 | Continued Discussion", where he
> and one other state, that FreeBSD R4.11 beforms much better (twice
> faster) than FreeBSD R5.3.
> 
> Could the disparity I saw be caused by the SMPng project in R5.3?

It's certainly possible.

> > By the way, ext3fs uses journalling, which provides metadata
> > consistency in a very different way from soft updates.  You
> > might also want to experiment with that to see if it works
> better
> > for your workload.
> >
> Hmm... I do not understand this hint.
> 
> Does FreeBSD offer a journaling file system?

No, although an interest in UFS+journalling has been expressed by
someone who could make it happen.  Also, Jean-Sebastien Pedron has
been working on a port of ReiserFS to FreeBSD.

> > > Are we already trying to issue write order requests for the
> > > disc blocks (whose write order is arbitrary) sorted by sector
> > > number (in order to move the disc heads as less as possible)?
> > > The disc write cache could do that, but I disabled it in order
> > > to decrease the probability of inconsistency.
> > 
> > Hopefully you disabled it on both FreeBSD and Linux, so you're
> > comparing apples to apples...
> > 
> During the tests I enabled write cache in both settings, because I
> did not know how to turn it off in KNOPPIX...
> 
> Isn't it possible to simulate the hard disc write cache in kernel?

The write caching will make a big difference.  Sorry, but I don't
know how to turn it off under Linux.  Perhaps you could turn it on
under FreeBSD for the purpose of your performance analysis.