Re: [linux-next:master] [mm] f1a7941243: unixbench.score -5.1% regression

[Oliver Sang]

Hi, Shakeel Butt,

please be noted we noticed the regression is still existing after this commit
merged to mainline, we reported as
https://lore.kernel.org/all/202301301057.e55dad5b-oliver.sang@xxxxxxxxx/

On Tue, Dec 06, 2022 at 10:41:00AM +0800, Huang, Ying wrote:
> Shakeel Butt <shakeelb@xxxxxxxxxx> writes:
> 
> > On Sun, Dec 4, 2022 at 9:56 PM kernel test robot <yujie.liu@xxxxxxxxx> wrote:
> >>
> >> Greeting,
> >>
> >> FYI, we noticed a -5.1% regression of unixbench.score due to commit:
> >>
> > [...]
> >> 9cd6ffa60256e931 f1a7941243c102a44e8847e3b94
> >> ---------------- ---------------------------
> >>          %stddev     %change         %stddev
> >>              \          |                \
> >>       7917            -5.1%       7509        unixbench.score
> >
> > What is unixbench.score?
> 
> Should be benchmark throughput.
> 
> >>      10485           -12.1%       9216        unixbench.time.maximum_resident_set_size
> 
> This should reflect accuracy change of per_cpu_counter.
> 
> >>   37236706            -5.1%   35324104        unixbench.time.minor_page_faults
> 
> The reduction is same as benchmark score.  So I think this reflect the
> nature of time-bound testing (instead of workload-bound).
> 
> > For above two, is negative change good or bad?
> >
> >>       0.98 ą 20%      +0.7        1.64 ą 38%  perf-profile.calltrace.cycles-pp.link_path_walk.path_openat.do_filp_open.do_sys_openat2.__x64_sys_openat
> >>       2.12 ą 19%      +0.8        2.96 ą 13%  perf-profile.calltrace.cycles-pp.handle_mm_fault.do_user_addr_fault.exc_page_fault.asm_exc_page_fault
> >>       2.35 ą 13%      +0.9        3.28 ą 13%  perf-profile.calltrace.cycles-pp.__handle_mm_fault.handle_mm_fault.do_user_addr_fault.exc_page_fault.asm_exc_page_fault
> >>       0.14 ą 74%      +0.4        0.55 ą 32%  perf-profile.children.cycles-pp.do_task_dead
> >>       0.04 ą223%      +0.4        0.47 ą 49%  perf-profile.children.cycles-pp.__mmdrop
> >
> > Also how should I interpret the above perf-profiles?
> 
> It appears that the changes of handle_mm_fault() and __mmdrop() are
> related to the code of the commit?  That is, for this specific workloads
> (not so unpractical), the operations become slower?
> 
> Best Regards,
> Huang, Ying
>