tuning XFS for tiny files

tuning XFS for tiny files

[timotheus]

[-- Attachment #1: Type: text/plain, Size: 428 bytes --]


Hi. Is there a way to tune XFS filesystem parameters to better address
the usage pattern of 10000s of tiny files in directories such as:
    maildir directory
    mh mail directory
    ccache directory

My understanding is that XFS will always be much slower than reiserfs
with respect to deleting 10000s files; but that XFS might be possible to
tune toward more rapid read access of 10000s of tiny files.

Regards,
-timotheus

[-- Attachment #2: Type: application/pgp-signature, Size: 188 bytes --]

Re: tuning XFS for tiny files

[Eric Sandeen]

timotheus wrote:
> Hi. Is there a way to tune XFS filesystem parameters to better address
> the usage pattern of 10000s of tiny files in directories such as:
>     maildir directory
>     mh mail directory
>     ccache directory
> 
> My understanding is that XFS will always be much slower than reiserfs
> with respect to deleting 10000s files; but that XFS might be possible to
> tune toward more rapid read access of 10000s of tiny files.
> 
> Regards,
> -timotheus

Do you have a way to benchmark your load?

logv2 at mkfs time, and throwing in larger logbufs & logbsize as mount
options might help.

-Eric

Re: tuning XFS for tiny files

[Nathan Scott]

On Wed, 2007-07-18 at 15:52 -0500, Eric Sandeen wrote:
> 
> > Hi. Is there a way to tune XFS filesystem parameters to better
> address
> > the usage pattern of 10000s of tiny files in directories such as:
> >     maildir directory
> >     mh mail directory
> >     ccache directory
> > 
> > My understanding is that XFS will always be much slower than
> reiserfs
> > with respect to deleting 10000s files; but that XFS might be
> possible to
> > tune toward more rapid read access of 10000s of tiny files.
> > 
> > Regards,
> > -timotheus
> 
> Do you have a way to benchmark your load?
> 
> logv2 at mkfs time, and throwing in larger logbufs & logbsize as mount
> options might help.

For these kinds of workloads, you may have some joy using a mkfs run
with smaller blocksize (down to 512 bytes) and also larger directory
blocksize (up to 64k).  (e.g. mkfs.xfs -bsize=512 -nsize=16k ...)

cheers.

--
Nathan

Re: tuning XFS for tiny files

[Andi Kleen]

timotheus <timotheus@tstotts.net> writes:

> Hi. Is there a way to tune XFS filesystem parameters to better address
> the usage pattern of 10000s of tiny files in directories such as:
>     maildir directory
>     mh mail directory
>     ccache directory
> 
> My understanding is that XFS will always be much slower than reiserfs
> with respect to deleting 10000s files; but that XFS might be possible to
> tune toward more rapid read access of 10000s of tiny files.

-d agcount=1 at mkfs time might help (unless you have a lot of CPUs) 

-Andi

Re: tuning XFS for tiny files

[David Chinner]

On Thu, Jul 19, 2007 at 03:38:43PM +0200, Andi Kleen wrote:
> timotheus <timotheus@tstotts.net> writes:
> 
> > Hi. Is there a way to tune XFS filesystem parameters to better address
> > the usage pattern of 10000s of tiny files in directories such as:
> >     maildir directory
> >     mh mail directory
> >     ccache directory
> > 
> > My understanding is that XFS will always be much slower than reiserfs
> > with respect to deleting 10000s files; but that XFS might be possible to
> > tune toward more rapid read access of 10000s of tiny files.
> 
> -d agcount=1 at mkfs time might help (unless you have a lot of CPUs) 

Yeah, might help, but it's not good for being able to repair the
filesystem - repair will be unable to find a secondary superblock
to compare the primary against and abort.....

-d agcount is only good for science experiments, not production
systems ;)

Cheers,

Dave.
-- 
Dave Chinner
Principal Engineer
SGI Australian Software Group

Re: tuning XFS for tiny files

[Andi Kleen]

On Thu, Jul 19, 2007 at 11:14:01PM +1000, David Chinner wrote:
> On Thu, Jul 19, 2007 at 03:38:43PM +0200, Andi Kleen wrote:
> > timotheus <timotheus@tstotts.net> writes:
> > 
> > > Hi. Is there a way to tune XFS filesystem parameters to better address
> > > the usage pattern of 10000s of tiny files in directories such as:
> > >     maildir directory
> > >     mh mail directory
> > >     ccache directory
> > > 
> > > My understanding is that XFS will always be much slower than reiserfs
> > > with respect to deleting 10000s files; but that XFS might be possible to
> > > tune toward more rapid read access of 10000s of tiny files.
> > 
> > -d agcount=1 at mkfs time might help (unless you have a lot of CPUs) 
> 
> Yeah, might help, but it's not good for being able to repair the
> filesystem - repair will be unable to find a secondary superblock
> to compare the primary against and abort.....

Any reason why it aborts? It could just continue with a warning, couldn't it?

> -d agcount is only good for science experiments, not production
> systems ;)

XFS small file performance needs a lot of science.

-Andi