Re: [Qemu-devel][PATCH,RFC] Zero cluster dedup

[Laurent Vivier]

Le mercredi 03 septembre 2008 à 14:47 +0200, Kevin Wolf a écrit :
[...]
> > For the
> > cluster granularity issue, it is meaning less for dedup (zero or other).
> > The deduped clusters are rarely de-duped to sequential clusters and
> > anyhow a dedup write operation is just changing the meta-data of the
> > block (logical offset l2 entry + reference count of the physical
> > cluster) so no large write optimization can be done.
> 
> This is still the L2 table which has to be written each time. Say we
> have a 64k write, then you could save 15 L2 table writes if qcow_dedup
> could handle zeroing multiple contiguous clusters at once. This could
> matter for performance when wiping lots of data.
> 
> Laurent, do you have an opinion on this? You should know this stuff
> better than me as you did the according patches.

I agree with you.

Shahar, I think you should make some performance benchmarks with
"cache=off" with and without your patch, just to see what happens.

Laurent
-- 
----------------- address@hidden  ------------------
  "La perfection est atteinte non quand il ne reste rien à
ajouter mais quand il ne reste rien à enlever." Saint Exupéry