Re: [PATCH 1/3] slub: set a criteria for slub node partial adding

[Shaohua Li <shaohua.li@intel.com>]

On Mon, 2011-12-12 at 12:14 +0800, Shi, Alex wrote:
> On Mon, 2011-12-12 at 10:43 +0800, Li, Shaohua wrote:
> > On Wed, 2011-12-07 at 15:28 +0800, David Rientjes wrote:
> > > On Wed, 7 Dec 2011, Shaohua Li wrote:
> > > 
> > > > interesting. I did similar experiment before (try to sort the page
> > > > according to free number), but it appears quite hard. The free number of
> > > > a page is dynamic, eg more slabs can be freed when the page is in
> > > > partial list. And in netperf test, the partial list could be very very
> > > > long. Can you post your patch, I definitely what to look at it.
> > > 
> > > It was over a couple of years ago and the slub code has changed 
> > > significantly since then, but you can see the general concept of the "slab 
> > > thrashing" problem with netperf and my solution back then:
> > > 
> > > 	http://marc.info/?l=linux-kernel&m=123839191416478
> > > 	http://marc.info/?l=linux-kernel&m=123839203016592
> > > 	http://marc.info/?l=linux-kernel&m=123839202916583
> > > 
> > > I also had a separate patchset that, instead of this approach, would just 
> > > iterate through the partial list in get_partial_node() looking for 
> > > anything where the number of free objects met a certain threshold, which 
> > > still defaulted to 25% and instantly picked it.  The overhead was taking 
> > > slab_lock() for each page, but that was nullified by the performance 
> > > speedup of using the alloc fastpath a majority of the time for both 
> > > kmalloc-256 and kmalloc-2k when in the past it had only been able to serve 
> > > one or two allocs.  If no partial slab met the threshold, the slab_lock() 
> > > is held of the partial slab with the most free objects and returned 
> > > instead.
> > With the per-cpu partial list, I didn't see any workload which is still
> > suffering from the list lock, 
> 
> The merge error that you fixed in 3.2-rc1 for hackbench regression is
> due to add slub to node partial head. And data of hackbench show node
> partial is still heavy used in allocation. 
The patch is already in base kernel, did you mean even with it you still
saw the list locking issue with latest kernel?