Re: [NUMA Balancing] e39bb6be9f: will-it-scale.per_thread_ops 64.4% improvement

[Feng Tang <feng.tang@intel.com>]

Hi Linus,

On Fri, Apr 01, 2022 at 09:35:24AM -0700, Linus Torvalds wrote:
> On Fri, Apr 1, 2022 at 2:42 AM kernel test robot <oliver.sang@intel.com> wrote:
> >
> > FYI, we noticed a 64.4% improvement of will-it-scale.per_thread_ops due to commit:
> > e39bb6be9f2b ("NUMA Balancing: add page promotion counter")
> 
> That looks odd and unlikely.
> 
> That commit only modifies some page counting statistics. Sure, it
> could be another cache layout thing, and maybe it's due to the subtle
> change in how NUMA_PAGE_MIGRATE gets counted, but it still looks a bit
> odd.

We did a quick check about cache stuff by disabling HW cache prefetch
completely (writing 0xf to MSR 0x1a4), and the performance change
is almost gone:

  ee97347fe058d020     e39bb6be9f2b39a6dbaeff48436 
  ----------------     --------------------------- 
    134793            -1.4%     132867        will-it-scale.per_thread_ops

The test box is a Cascadelake machine with 4 nodes, and the similar trend
is found on a 2 nodes machine, that the commit has 55% improvement with
HW cache prefetch enabled, and has less than 1% change when disabled.

Though we still cannot pin-point the exact place affected.

Also per our experience, the patch changing vm statistics can easily
trigger strange performance bumps for micro-benchmarks like will-it-scale,
stress-ng etc.

Thanks,
Feng


>                Linus