Re: [x86, sched] 2a0abc5969: phoronix-test-suite.stress-ng.SystemVMessagePassing.bogo_ops_s -14.1% regression

[Giovanni Gherdovich <ggherdovich@suse.cz>]

On Mon, 2020-11-16 at 15:07 +0800, kernel test robot wrote:
> Greeting,
> 
> FYI, we noticed a -14.1% regression of phoronix-test-suite.stress-ng.SystemVMessagePassing.bogo_ops_s due to commit:
> 
> 
> commit: 2a0abc59699896f03bf6f16efb8a3a490511216f ("x86, sched: Add support for frequency invariance on SKYLAKE_X")
> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git master
> 
> 
> in testcase: phoronix-test-suite
> on test machine: 96 threads Intel(R) Xeon(R) Gold 6252 CPU @ 2.10GHz with 192G memory
> with following parameters:
> 
> 	test: stress-ng-1.2.2
> 	option_a: System V Message Passing
> 	cpufreq_governor: performance
> 	ucode: 0x5002f01

I haven't checked this regression yet, but when a "schedutil" change like
this patch triggers a regression when the "performance" governor is used,
the trouble might come in some roundabout way, eg. from cache alignment issues.

The patch defines a number of (not necessarily aligned) per-cpu variables, and
there was a recent thread where Boris Petkov, Tony Luck and Mel Gorman debugged
something that had similar characteristics:

"RE: [LKP] Re: [x86/mce] 1de08dccd3: will-it-scale.per_process_ops -14.1% regression"
https://lore.kernel.org/lkml/6e996ad05e434a6fb13f069ee72b876b@intel.com/

I'll study this problem with that in mind.


Giovanni