Re: [Cooker] Dramatic poor performance for XFS delete (worse since 2008.1)

[Herton Ronaldo Krzesinski]

Em Monday 28 April 2008 11:17:49 Olivier LAHAYE escreveu:
> Recetly, I've discovered that XFS delete performance have been worse than
> ever.
> Doing the test below on ext3 and on xfs (same partition, same computer
> (after a mkfs) shows that ext3 is using buffers (disk blinks from times to 
> times) while xfs seems to write the disk on the fly (disk always lit).
>
> Question: are you aware of a patch in kernel that (for any reason) disabled
> the use of buffers in XFS?
>
> time ( for i in $(seq 1 10000); do echo $i | md5sum> xxx_$i; done; sync );
> time ( /bin/rm -f xxx_*; sync )
>
> Note that we studied disk parameters and set the same in both tests
> (hdparm -c3 -d1 -u1)
> and I mountex xfs filesystem using -ologbufs=8
>
> With absolutely no differences.
>
> Following the tweaking guide below did not change any results (while some
> improvements were expected)
> http://www.generation-nt.com/licence-3g-rejet-candidature-free-mobile-arcep
>-actualite-46044.html
>
> This leads me to think that now, the xfs filesystem doesn't use any buffer
> I/O. Am I wrong?

May be there is some regression in the kernel, but first I think your 
benchmark is wrong: you're using sync, eliminating any advantage of any 
buffering xfs has.

--
[]'s
Herton