Re: Hot rb_next, setup_cluster_no_bitmap

[Chris Mason <chris.mason@oracle.com>]

Excerpts from Simon Kirby's message of 2011-08-03 19:10:59 -0400:
> On Wed, Aug 03, 2011 at 03:39:49PM -0700, Simon Kirby wrote:
> 
> > On Wed, Aug 03, 2011 at 03:06:55PM -0700, Simon Kirby wrote:
> > 
> > > I see Josef's 86d4a77ba3dc4ace238a0556541a41df2bd71d49 introduced the
> > > bitmaps list. I could try temporarily reverting this (some fixups needed)
> > > if anybody thinks my cache bouncing idea might be slightly possible.
> > 
> > I'll try the attached and see how the profile changes.
> 
> Hmm, I bound the SMP affinity of all of the btrfs processes to one CPU,
> and the page dirtying rate got slower, so I suspect the writes aren't
> really a big deal, and the problem is just that there is way too much
> walking going on after rsync has ran for a while and loads everything
> into memory.
> 
> Any ideas?

The current for-linus branch gets rid of all the bottlenecks in the
metadata blocks.  So now we're stuck with the bottlenecks in the
allocator.

There are a few simple things we can do here but Josef has a patch that
fixes delalloc reservations for inline extents that might help as a
first step.

-chris