[Gnu-arch-users] Re: stat performance on AFS [was Re: [PATCH] arch speedups on big trees]

[Miles Bader]

On Wed, Jan 07, 2004 at 02:02:08AM -0600, Charles Duffy wrote:
> > Does AFS not special-case stats of files that are in the cache?!?
> 
> Put it this way: Doing an inode-signature-based operation twice in a row
> on our home network (in many respects a scaled-down model of the network
> at work) results in quite a bit of a clatter from the AFS server both
> times, whereas in the case of a read()-based operation done twice, the
> second run has no server involvement at all.

It may be be that AFS _does_ special case this operation, but that doing a
read-less scan of a tree's inode doesn't actually put anything in the cache
(since no data is actually used), so doing it a second time is exactly as
slow as the first.  But since a read-based `stupid diff' fills the cache the
first time, the second go is much faster.  That would mean that it's 
2 * MEDIUM-SLOW operations versus REALLY-SLOW + FAST...

My vague recollection is that the a bulk-stat protocol was at least in
consideration for AFS (this was in the 80's); does it have that yet?

> "Files that are in the cache" is something of a misnomer -- the cache
> contains chunks rather than complete files; hence, one can't simply stat
> the cache's copy of the file and expect it to represent the correct
> values.

The AFS filesystem code should certainly know enough to keep track of that
info, though!

-miles
-- 
In New York, most people don't have cars, so if you want to kill a person, you
have to take the subway to their house.  And sometimes on the way, the train
is delayed and you get impatient, so you have to kill someone on the subway.
  [George Carlin]