lkml.org 
[lkml]   [2009]   [Oct]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Latest vfs scalability patch
On Tue, Oct 06 2009, Nick Piggin wrote:
> On Tue, Oct 06, 2009 at 12:14:14PM +0200, Jens Axboe wrote:
> > On Tue, Oct 06 2009, Nick Piggin wrote:
> > > Hi,
> > >
> > > Several people have been interested to test my vfs patches, so rather
> > > than resend patches I have uploaded a rollup against Linus's current
> > > head.
> > >
> > > ftp://ftp.kernel.org/pub/linux/kernel/people/npiggin/patches/fs-scale/
> > >
> > > I have used ext2,ext3,autofs4,nfs as well as in-memory filesystems
> > > OK (although this doesn't mean there are no bugs!). Otherwise, if your
> > > filesystem compiles, then there is a reasonable chance of it working,
> > > or ask me and I can try updating it for the new locking.
> > >
> > > I would be interested in seeing any numbers people might come up with,
> > > including single-threaded performance.
> >
> > I gave this a quick spin on the 64-thread nehalem. Just a simple dbench
> > with 64 clients on tmpfs. The results are below. While running perf top
> > -a in mainline, the top 5 entries are:
> >
> > 2086691.00 - 96.6% : _spin_lock
> > 14866.00 - 0.7% : copy_user_generic_string
> > 5710.00 - 0.3% : mutex_spin_on_owner
> > 2837.00 - 0.1% : _atomic_dec_and_lock
> > 2274.00 - 0.1% : __d_lookup
> >
> > Uhm auch... It doesn't look much prettier for the patch kernel, though:
> >
> > 9396422.00 - 95.7% : _spin_lock
> > 66978.00 - 0.7% : copy_user_generic_string
> > 43775.00 - 0.4% : dput
> > 23946.00 - 0.2% : __link_path_walk
> > 17699.00 - 0.2% : path_init
> > 15046.00 - 0.2% : do_lookup
>
> Yep, this is the problem of the common-path lookup. Every dentry
> element in the path has its d_lock taken for every path lookup,
> so cwd dentry lock bounces a lot for dbench.
>
> I'm working on doing path traversal without any locks or stores
> to the dentries in the common cases, so that should basically
> be the last bit of the puzzle for vfs locking (although it can be
> considered a different type of problem than the global lock
> removal, but RCU-freed struct inode is important for the approach
> I'm taking, so I'm basing it on top of these patches).
>
> It's a copout, but you could try running multiple dbenches under
> different working directories (or actually, IIRC dbench does root
> based path lookups so maybe that won't help you much).

Yeah, it's hitting dentry->d_lock pretty hard so basically
spin-serialized at that point.

> > Anyway, below are the results. Seem very stable.
> >
> > throughput
> > ------------------------------------------------
> > 2.6.32-rc3-git | 561.218 MB/sec
> > 2.6.32-rc3-git+patch | 627.022 MB/sec
>
> Well it's good to see you got some improvement.

Yes, it's an improvement though the results are still pretty abysmal :-)

--
Jens Axboe



\
 
 \ /
  Last update: 2009-10-06 14:53    [W:0.418 / U:0.172 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site