[Samba] performance regression between 30.14a and 3.0.20

[Jeremy Allison]

On Fri, Apr 28, 2006 at 09:40:39AM +1000, James Peach wrote:
> On 4/27/06, Jeremy Allison <jra at samba.org> wrote:
> > On Wed, Apr 26, 2006 at 01:20:23PM -0400, Greg Dickie wrote:
> > > OK all my gobbledy-gook stats aside..... sorry for being incoherent.
> > >
> > > Stracing reveals that files after 3.0.20b are opened with O_SYNC. Could
> > > that be the source of the problems? I'll try and find that and test it
> > > when I get home but the question is is it necessary?
> >
> > Greg - I LOVE YOU :-). That's almost certainly the problem.
> > Back in the Win9x protocols days (the openX style of calls)
> > we had the problem that Windows clients were setting the
> > "sync this data" flag for almost every open - completely
> > unneccessary. For 3.0.20x we changed to the NTcreateX
> > style of open, and mapped the create option FILE_WRITE_THROUGH
> > to the O_SYNC. I bet the Windows redirector in W2K, XP and W2K3
> > is just as dim, and is adding that on almost every open (I'm
> > on a plane right now so can't get to my vmware sessions
> > easily to check).
> 
> xdd explicitly sets this flag when it calls CreateFile. Testing with a
> different windows dd program gives more expected results.

The question is - what does netbench do ? And what does iometer do ?
I'm back home now (but *very* tired) so I'll get some answers soon.
It's definately something we needed to fix though.

Jeremy.