On Fri, Aug 21, 2020 at 10:02:59AM +0800, Feng Tang wrote:
> 1de08dccd383 x86/mce: Add a struct mce.kflags field
> 9554bfe403bd x86/mce: Convert the CEC to use the MCE notifier
>
> And strange thing is after using gcc9 and debian10 rootfs, with same commits
> the regression turns to a improvement,
How so?
> though the trend keeps, that if we
> changes the kflags from __u64 to __u32, the performance will be no change.
>
> Following is the comparing of regression, I also attached the perf-profile
> for old and new commit (let me know if you need more data)
>
>
> 9554bfe403bdfc08 1de08dccd383482a3e88845d355
> ---------------- ---------------------------
> %stddev %change %stddev
> \ | \
> 192362 -15.1% 163343 will-it-scale.287.processes
> 0.91 +0.2% 0.92 will-it-scale.287.processes_idle
> 669.67 -15.1% 568.50 will-it-scale.per_process_ops
This is the data from your previous measurement:
9554bfe403bdfc08 1de08dccd383482a3e88845d355
---------------- ---------------------------
%stddev %change %stddev
\ | \
668.00 -14.1% 573.75 will-it-scale.per_process_ops
If I'm reading it correctly, commit
1de08dccd383 ("x86/mce: Add a struct mce.kflags field")
is still the slower one vs
9554bfe403bd ("x86/mce: Convert the CEC to use the MCE notifier")
Or am I misreading it?
In any case, this really looks like what Tony said: this enlargement of
struct mce pushes some variable into a cacheline-misaligned placement,
causing it to bounce.
The $ 10^6 question is, which variable is that...
--
Regards/Gruss,
Boris.
SUSE Software Solutions Germany GmbH, GF: Felix Imendörffer, HRB 36809, AG
Nürnberg
