Re: [LKP] [mm] 3484b2de949: -46.2% aim7.jobs-per-min

[Mel Gorman <mgorman@suse.de>]

On Thu, Mar 05, 2015 at 01:34:59PM +0800, Huang Ying wrote:
> Hi, Mel,
> 
> On Sat, 2015-02-28 at 15:30 +0800, Huang Ying wrote:
> > On Sat, 2015-02-28 at 01:46 +0000, Mel Gorman wrote:
> > > On Fri, Feb 27, 2015 at 03:21:36PM +0800, Huang Ying wrote:
> > > > FYI, we noticed the below changes on
> > > > 
> > > > git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
> > > > commit 3484b2de9499df23c4604a513b36f96326ae81ad ("mm: rearrange zone fields into read-only, page alloc, statistics and page reclaim lines")
> > > > 
> > > > The perf cpu-cycles for spinlock (zone->lock) increased a lot.  I suspect there are some cache ping-pong or false sharing.
> > > > 
> > > 
> > > Are you sure about this result? I ran similar tests here and found that
> > > there was a major regression introduced near there but it was commit
> > > 05b843012335 ("mm: memcontrol: use root_mem_cgroup res_counter") that
> > > cause the problem and it was later reverted.  On local tests on a 4-node
> > > machine, commit 3484b2de9499df23c4604a513b36f96326ae81ad was within 1%
> > > of the previous commit and well within the noise.
> > 
> > After applying the below debug patch, the performance regression
> > restored.  So I think we can root cause this regression to be cache line
> > alignment related issue?
> > 
> > If my understanding were correct, after the 3484b2de94, lock and low
> > address area free_area are in the same cache line, so that the cache
> > line of the lock and the low address area of free_area will be switched
> > between MESI "E" and "S" state because it is written in one CPU (page
> > allocating with free_area) and frequently read (spinning on lock) in
> > another CPU.
> 
> What do you think about this?
> 

My attention is occupied by the automatic NUMA regression at the moment
but I haven't forgotten this. Even with the high client count, I was not
able to reproduce this so it appears to depend on the number of CPUs
available to stress the allocator enough to bypass the per-cpu allocator
enough to contend heavily on the zone lock. I'm hoping to think of a
better alternative than adding more padding and increasing the cache
footprint of the allocator but so far I haven't thought of a good
alternative. Moving the lock to the end of the freelists would probably
address the problem but still increases the footprint for order-0
allocations by a cache line.

-- 
Mel Gorman
SUSE Labs