will-it-scale.per_process_ops -15% regression on v6.16-rc1

will-it-scale.per_process_ops -15% regression on v6.16-rc1

[D, Suneeth]

Greetings,

We have observed ~(15-16)% regression on the will-it-scale-page_fault3 
variant when the testcase was run in process mode on the kernel 
v6.16-rc5. Further investigation showed that the regression was sparked 
in v6.16-rc1 and has propagated all its way through v6.16-rc5.

Below are the test parameters and Machine configuration that have been 
tested with:-

testcase: will-it-scale
compiler: gcc-13
test machine: 256 threads 1 sockets AMD EPYC 9754 128-Core Processor @ 
2.2GHz [Bergamo] with 258G memory
Test params:

         nr_task: [1 8 128 192 256]
	mode: process
	test: page_fault3
	cpufreq_governor: performance

stable version (v6.15)	%diff 	per_process_ops	kernel_rc_ver
----------------------  -----   --------------- -------------
492057			-15%	416927		v6.16-rc1
492057			-16%	414140		v6.16-rc2
492057			-15%	419158		v6.16-rc3
492057			-15%	420476		v6.16-rc4
492057			-15%	416334		v6.16-rc5

Also had a run with latest stable

v6.15	%diff 	v6.15.5
-----   -----   -------
492057	 1%	494990


Recreation steps:

1) git clone https://github.com/antonblanchard/will-it-scale.git
2) git clone https://github.com/intel/lkp-tests.git
3) cd will-it-scale && git apply 
lkp-tests/programs/will-it-scale/pkg/will-it-scale.patch
4) make
5) python3 ./runtest.py page_fault3 32 process 0 0 1 8 128 192 256

NOTE: [5] is specific to machine's architecture. starting from 1 is the 
array of no.of tasks that you'd wish to run the testcase which here is 
no.cores per CCX, per NUMA node/ per Socket, nr_threads.

Currently bisection is under progress b/w v6.15 and v6.16-rc1 to spot 
the culprit commit.

Thanks & Regards,
Suneeth D

Re: will-it-scale.per_process_ops -15% regression on v6.16-rc1

[D, Suneeth]

Hello,

On 7/10/2025 5:21 PM, D, Suneeth wrote:
> Greetings,
> 
> We have observed ~(15-16)% regression on the will-it-scale-page_fault3 
> variant when the testcase was run in process mode on the kernel 
> v6.16-rc5. Further investigation showed that the regression was sparked 
> in v6.16-rc1 and has propagated all its way through v6.16-rc5.
> 
> Below are the test parameters and Machine configuration that have been 
> tested with:-
> 
> testcase: will-it-scale
> compiler: gcc-13
> test machine: 256 threads 1 sockets AMD EPYC 9754 128-Core Processor @ 
> 2.2GHz [Bergamo] with 258G memory
> Test params:
> 
>          nr_task: [1 8 128 192 256]
>      mode: process
>      test: page_fault3
>      cpufreq_governor: performance
> 
> stable version (v6.15)    %diff     per_process_ops    kernel_rc_ver
> ----------------------  -----   --------------- -------------
> 492057            -15%    416927        v6.16-rc1
> 492057            -16%    414140        v6.16-rc2
> 492057            -15%    419158        v6.16-rc3
> 492057            -15%    420476        v6.16-rc4
> 492057            -15%    416334        v6.16-rc5
> 
> Also had a run with latest stable
> 
> v6.15    %diff     v6.15.5
> -----   -----   -------
> 492057     1%    494990
> 
> 
> Recreation steps:
> 
> 1) git clone https://github.com/antonblanchard/will-it-scale.git
> 2) git clone https://github.com/intel/lkp-tests.git
> 3) cd will-it-scale && git apply 
> lkp-tests/programs/will-it-scale/pkg/will-it-scale.patch
> 4) make
> 5) python3 ./runtest.py page_fault3 32 process 0 0 1 8 128 192 256
> 
> NOTE: [5] is specific to machine's architecture. starting from 1 is the 
> array of no.of tasks that you'd wish to run the testcase which here is 
> no.cores per CCX, per NUMA node/ per Socket, nr_threads.
> 
> Currently bisection is under progress b/w v6.15 and v6.16-rc1 to spot 
> the culprit commit.
> 

The bisection got me landed onto 
7ac67301e82f02b77a5c8e7377a1f414ef108b84 as the culprit commit. This was 
the same commit that was causing a regression in lmbench3 
micro-benchmark as well earlier for which the detailed discussion can be 
found in [1].

[1] 
https://lore.kernel.org/lkml/f59ef632-0d11-4ae7-bdad-d552fe1f1d78@amd.com/

> Thanks & Regards,
> Suneeth D
>