Re: Major regression on hackbench with SLUB

["Björn Steinbrink" <B.Steinbrink@gmx.de>]

On 2007.12.08 17:16:24 -0500, Steven Rostedt wrote:
> 
> Hi Linus,
> 
> > On Fri, 7 Dec 2007, Linus Torvalds wrote:
> > >
> > > Can you do one run with oprofile, and see exactly where the cost is? It
> > > should hopefully be pretty darn obvious, considering your timing.
> 
> The results are here:
> 
> http://people.redhat.com/srostedt/slub/results/slab.op
> http://people.redhat.com/srostedt/slub/results/slub.op

Hm, you seem to be hitting the "another_slab" stuff in __slab_alloc
alot. I wonder if !node_match triggers too often. We always start with
the per cpu slab, if that one is on the wrong node, you'll always hit
that "another_slab" path.

After searching for way too long (given that I have no clue about that
stuff anyway and just read the code out of curiousness), I noticed that
the the cpu_to_node stuff on x86_64 seems to be initialized to 0xff
(arch/x86/mm/numa_64.c), and Google brought me this dmesg output [1],
which, AFAICT, shows that the per cpu slab setup is done _before_
cpu_to_node is correctly setup. That would lead to the per cpu slabs all
having node == 0xff, which looks pretty bad.

Disclaimer: I read the slub/numa/$WHATEVER_I_SAW_THERE for the first
time, so this might be total bull ;-)

Björn

[1] http://linux.derkeiler.com/Mailing-Lists/Kernel/2007-10/msg04648.html