On Thu, Mar 01, 2018 at 04:01:05PM -0800, Andrew Morton wrote:
> On Thu, 1 Mar 2018 14:28:44 +0800 Aaron Lu <aaron...@intel.com> wrote:
> > When freeing a batch of pages from Per-CPU-Pages(PCP) back to buddy,
> > the zone->lock is held and then pages are chosen from PCP's migratetype
> > list. While there is actually no need to do this 'choose part' under
> > lock since it's PCP pages, the only CPU that can touch them is us and
> > irq is also disabled.
> > Moving this part outside could reduce lock held time and improve
> > performance. Test with will-it-scale/page_fault1 full load:
> > kernel Broadwell(2S) Skylake(2S) Broadwell(4S) Skylake(4S)
> > v4.16-rc2+ 9034215 7971818 13667135 15677465
> > this patch 9536374 +5.6% 8314710 +4.3% 14070408 +3.0% 16675866 +6.4%
> > What the test does is: starts $nr_cpu processes and each will repeatedly
> > do the following for 5 minutes:
> > 1 mmap 128M anonymouse space;
> > 2 write access to that space;
> > 3 munmap.
> > The score is the aggregated iteration.
> But it's a loss for uniprocessor systems: it adds more code and adds an
> additional pass across a list.
Performance wise, I assume the loss is pretty small and can not
On my Sandybridge desktop, with will-it-scale/page_fault1/single process
run to emulate uniprocessor system, the score is(average of 3 runs):
base(patch 1/3): 649710
this patch: 653554 +0.6%
prefetch(patch 3/3): 650336 (in noise range compared to base)
On 4 sockets Intel Broadwell with will-it-scale/page_fault1/single
base(patch 1/3): 498649
this patch: 504171 +1.1%
prefetch(patch 3/3): 506334 +1.5% (compared to base)
It looks like we don't need to worry too much about performance for