[OpenAFS] vice partition sizes, and large disks

Jeffrey Hutzelman jhutz@cmu.edu
Wed, 20 Jun 2001 15:09:09 -0400 (EDT)


On Sat, 16 Jun 2001, Bill Zumach wrote:

> > In message <200106161507.PAA32863269@smtp6ve.mailsrvcs.net>,Bill Zumach write
> > s:
> > >That's a good start. Linux 2.2.18 is about 7,000 files, so if 3 users
> > >are un-taring at the same time we're up to about 21,000 files. Where I'm
> > >going with this is that you also want to check what happens when the log fil
> > ls
> > 
> > just some more silly tests (i did a umount/mount between these tests so
> > that should flush the log):
> > 
> >   1000 real        0.2 user        0.0 sys         0.1
> >   5000 real        1.8 user        0.0 sys         0.8
> >  10000 real        3.8 user        0.0 sys         1.5
> >  20000 real        8.3 user        0.1 sys         3.1
> >  50000 real       25.8 user        0.2 sys         8.0
> > 100000 real       56.5 user        0.5 sys        17.5
> > 
> > if you graph that, it seems pretty linear so i dont think that i am
> > filling up the log.  the test partition is 10G, which implies (according
> > to the man page) a 10M logfile.  cool eh?
> > _______________________________________________
> That looks good. I'd guess they're writing the transactions back frequently
> enough. Once the OS has a chance to properly order the blocks it writes to
> disk one can get this sort of performance for creates.
> 
> BTW, are you creating all these files in the same directory? I would not have
> thought the create times would start to be non-linear just becase of the
> time taken to look up file names in the directory.

Of course they should.  The lookup may involve a small time constant, but
it's still nonzero, and it's linear in the size of the directory.  What's
interesting is that the above numbers seem to should the reverse -- as the
number of files increases, the time per file goes _down_.  I can only
assume this is because either (1) optimization is being done when all
those updates are committed at once, or (2) the fixed time involved dwarfs
the per-file time, even at 100000 files.

-- Jeff