On Fri, Jan 12, 2007 at 11:28:21AM -0500, Shaya Potter wrote:
> Here's my optimization patch (quick and dirty, looks bigger than it is,
> most of it is changing indenting of already existing code within if()
> statements I added) w/ postmark numbers. I'm a little confused as now
> unionfs on backing store is faster than plain backing store (albiet not
> many test iterations and my methodology might be suspect, any thoughts?)
>
> so what do I do.
>
> 1) anywhere we check for a whiteout (lookup.c or inode.c), we don't
> check for it if the bindex we are searching in is bend/bopaque as, a
> whiteout shouldn't exist there.
The way you worded that it almost sounds like we are altering the semantics
a bit. It is more of a reduction, much like:
if (1 || foo) {}
will always be true.
> for lookup.c ( unionfs_lookup_backend() ) that is the major loop and
> basically skip whiteout search if bindex = bend
>
> for inode.c, since we are only looking in bstart for the whiteout, its
> if bstart = bend (or opaque if < bend and != -1)
>
> there's also some whiteout code in rename.c that I didn't get to yet
> (though probably have to).
Ok.
> I also addressed whiteout code in unlink.c, unionfs_unlink_whiteout()
> always created a whiteout, even in a case where there was nothing to
> whiteout. In those cases (I believe, didn't change any branching just
> changed a create_whiteout() to "goto out").
Makes sense.
> similarly for rmdir, one shouldn't create a whiteout if one doesn't have to.
Yep.
> and this is what I mean by having to change rename as it also probably
> creates whiteouts when it doesn't have to.
Ok.
> as can be seen from the attached pdf, the speedup in postmark is
> remarkable, though makes perfect sense. Also attaching the raw 1 run
> numbers
Very odd that the optimized case is faster than the base fs.
Thanks for the patch!
Josef "Jeff" Sipek.
--
The box said "Windows XP or better required". So I installed Linux.
_______________________________________________
unionfs mailing list
[email protected]
http://www.fsl.cs.sunysb.edu/mailman/listinfo/unionfs
