On Tue, Dec 01, 2009 at 09:33:37AM -0800, Jeremy Allison wrote:
> On Tue, Dec 01, 2009 at 05:56:59PM +0100, Björn JACKE wrote:
> > Hi,
> > bug 6942 beamed by attention to this article:
> > http://software.intel.com/en-us/articles/windows-client-cifs-behavior-can-slow-linux-nas-performance/
> > It might be a good idea to change the default of strict allocate to "yes".
> > This
> > makes out-of-quota situation less error prone and scenarios like in that
> > article
> > will improve, too.
> > Has anyone strong feelings against that change of defaults?
> Yes, don't do this. It's improves performance for some
> cases, but kills performance for others. It's an ext3-only
> fix as well, which doesn't work for ext4. Admittedly ext3
> is our most common filesystem for now, but that won't always
> be the case.
> > In addition to that: When available, we could do allocation with
> > posix_fallocate(), which has the additional benefit that with recent
> > Linux
> > kernels and advanced filesystems, allocation is being done at
> > kernel/filesystem
> > layer via fallocate with quasi zero overhead.
> Look at https://bugzilla.samba.org/show_bug.cgi?id=6942
> where there is a patch I did a while ago (modified since
> then I think) that changes strict allocate to a tri-state
> where the setting "partial" does most of what you want.
FYI. I think the "strict allocate" patch might be worthwhile
for 3.5.0. Volker, what do you think ? I know some of our
OEM's are already using it.