Re: [PATCH v3] sched/cache: Remove the old cache group footprint on exec

[Chen Yu <yu.c.chen@intel.com>]

On Thu, Oct 15, 2026 at 01:36:04PM +0800, Jemmy Wong wrote:

[ ... ]

> @@ -3965,36 +3974,6 @@ static void task_numa_placement(struct task_struct *p)
>  				ng->total_faults += diff;
>  				group_faults += ng->faults[mem_idx];
>  			}
> -#ifdef CONFIG_SCHED_CACHE
> -			/*
> -			 * Per task p->numa_faults[mem_idx] converges,
> -			 * so the accumulation of each task's faults
> -			 * converges too - Given the number of threads,
> -			 * it cannot overflow an unsigned long.

Above comments can be kept for future reference, with minor adjustment:

 s/it cannot overflow an unsigned long/it cannot overflow an atomic_long_t

I launched a full hackbench test on a Xeon with multiple LLCs per node. There
is no obvious difference except for some false-positive regress with large
run-to-run variance. This patch looks good to me in terms of turning the
footprint into an accurate counter. Feel free to add:

Reviewed-by: Chen Yu <yu.c.chen@intel.com>
Tested-by: Chen Yu <yu.c.chen@intel.com>

thanks,
Chenyu



First run:
=========================================
Hackbench comparison
BASE: cas_on_baseline
TEST: cas_on_patched
=========================================
MODE     G  FD |         BASE(s) |         TEST(s) |   DIFF% | RESULT
------- -- ----+-----------------+-----------------+---------+----------
process 16   1 |      0.178/1.7% |      0.166/0.6% |   6.74% | IMPROVED
process 16  16 |      4.871/9.1% |      5.230/3.8% |  -7.37% | REGRESSED
process 16   2 |      1.574/3.0% |     1.296/27.5% |  17.66% | IMPROVED
process 16  20 |      6.502/7.8% |      6.262/1.5% |   3.69% | IMPROVED
process 16   4 |      2.296/5.0% |     2.138/13.0% |   6.88% | IMPROVED
process 16   8 |      2.056/1.1% |     1.982/12.3% |   3.60% | IMPROVED
process  1   1 |      2.040/0.8% |      2.010/4.5% |   1.47% | IMPROVED
process  1  16 |      4.837/3.5% |     4.189/12.7% |  13.40% | IMPROVED
process  1   2 |      3.708/5.0% |     2.768/14.3% |  25.35% | IMPROVED
process  1  20 |      4.552/3.5% |      4.505/2.2% |   1.03% | IMPROVED
process  1   4 |     2.696/17.0% |     3.092/21.8% | -14.69% | REGRESSED
process  1   8 |      5.474/3.4% |      5.372/9.3% |   1.86% | IMPROVED
process  2   1 |      1.089/3.4% |      1.115/2.3% |  -2.39% | REGRESSED
process  2  16 |     4.985/13.9% |      5.246/5.1% |  -5.24% | REGRESSED
process  2   2 |      1.063/7.2% |      0.998/3.2% |   6.11% | IMPROVED
process  2  20 |     5.091/13.6% |      5.081/3.1% |   0.20% | IMPROVED
process  2   4 |     3.935/18.2% |     3.438/23.2% |  12.63% | IMPROVED
process  2   8 |      4.791/7.0% |      4.925/3.8% |  -2.80% | REGRESSED
process  4   1 |     0.690/18.3% |      0.586/1.9% |  15.07% | IMPROVED
process  4  16 |      5.878/3.5% |      6.119/1.9% |  -4.10% | REGRESSED
process  4   2 |      0.534/6.4% |      0.512/2.7% |   4.12% | IMPROVED
process  4  20 |      5.839/3.1% |      5.644/5.0% |   3.34% | IMPROVED
process  4   4 |      4.012/1.5% |     3.386/12.0% |  15.60% | IMPROVED
process  4   8 |     4.737/11.0% |      4.703/3.9% |   0.72% | IMPROVED
process  8   1 |      0.317/0.9% |      0.322/0.3% |  -1.58% | REGRESSED
process  8  16 |      3.671/5.4% |      3.922/4.8% |  -6.84% | REGRESSED
process  8   2 |     0.921/35.8% |     0.812/36.0% |  11.83% | IMPROVED
process  8  20 |      4.878/6.9% |      4.769/1.8% |   2.23% | IMPROVED
process  8   4 |     2.969/10.2% |     3.500/16.5% | -17.88% | REGRESSED
process  8   8 |      4.698/2.6% |      4.263/4.7% |   9.26% | IMPROVED
threads 16   1 |      0.192/0.5% |      0.189/0.0% |   1.56% | IMPROVED
threads 16  16 |      5.430/4.7% |      5.685/1.2% |  -4.70% | REGRESSED
threads 16   2 |     0.362/17.4% |     0.577/29.6% | -59.39% | REGRESSED
threads 16  20 |      6.210/3.6% |      5.523/6.2% |  11.06% | IMPROVED
threads 16   4 |     2.678/11.0% |      3.173/5.6% | -18.48% | REGRESSED
threads 16   8 |      2.347/3.3% |      2.394/8.7% |  -2.00% | REGRESSED
threads  1   1 |      2.131/1.5% |      2.260/2.5% |  -6.05% | REGRESSED
threads  1  16 |      0.379/2.4% |      0.353/2.5% |   6.86% | IMPROVED
threads  1   2 |      2.050/0.9% |      2.003/0.6% |   2.29% | IMPROVED
threads  1  20 |      0.347/2.6% |      0.324/0.9% |   6.63% | IMPROVED
threads  1   4 |      1.342/1.0% |      1.333/1.1% |   0.67% | IMPROVED
threads  1   8 |      0.730/2.5% |      0.732/1.4% |  -0.27% | REGRESSED
threads  2   1 |      1.144/1.2% |      1.167/1.2% |  -2.01% | REGRESSED
threads  2  16 |      0.286/1.4% |     0.987/71.3% | -245.10% | REGRESSED
threads  2   2 |      1.035/2.4% |      1.040/0.5% |  -0.48% | REGRESSED
threads  2  20 |     1.016/73.2% |     2.038/43.5% | -100.59% | REGRESSED
threads  2   4 |      0.647/0.9% |      0.647/1.2% |   0.00% | SAME
threads  2   8 |      0.338/1.8% |      0.335/1.5% |   0.89% | IMPROVED
threads  4   1 |      0.658/1.4% |      0.614/4.4% |   6.69% | IMPROVED
threads  4  16 |      6.434/1.4% |      6.194/1.5% |   3.73% | IMPROVED
threads  4   2 |      0.522/2.1% |      0.521/1.7% |   0.19% | IMPROVED
threads  4  20 |      5.361/1.4% |      6.060/1.6% | -13.04% | REGRESSED
threads  4   4 |      0.324/1.5% |      0.335/1.8% |  -3.40% | REGRESSED
threads  4   8 |     0.946/35.4% |     0.854/61.1% |   9.73% | IMPROVED
threads  8   1 |      0.351/5.7% |      0.364/1.9% |  -3.70% | REGRESSED
threads  8  16 |      3.818/4.8% |      4.109/4.9% |  -7.62% | REGRESSED
threads  8   2 |      0.294/6.5% |      0.276/1.4% |   6.12% | IMPROVED
threads  8  20 |      4.640/5.2% |      4.788/9.1% |  -3.19% | REGRESSED
threads  8   4 |     1.548/57.0% |     0.880/11.0% |  43.15% | IMPROVED
threads  8   8 |      5.028/2.2% |      5.161/2.9% |  -2.65% | REGRESSED

Then I pick the following test, which seems to have a big regression with a
large stderr%:
threads 2 16 | 0.286/1.4% | 0.987/71.3% | -245.10% | REGRESSED
threads 2 20 | 1.016/73.2% | 2.038/43.5% | -100.59% | REGRESSED
threads 16 2 | 0.362/17.4% | 0.577/29.6% | -59.39% | REGRESSED
threads 16 4 | 2.678/11.0% | 3.173/5.6% | -18.48% | REGRESSED


The following is the second run:
=========================================
MODE     G  FD |         BASE(s) |         TEST(s) |   DIFF% | RESULT
------- -- ----+-----------------+-----------------+---------+----------
threads  2  16 |     0.915/29.6% |     1.376/66.1% | -50.38% | REGRESSED
threads  2  20 |     2.154/23.0% |     1.299/50.7% |  39.69% | IMPROVED
threads 16   2 |     0.320/18.8% |     0.756/26.3% | -136.25% | REGRESSED
threads 16   4 |      4.139/1.9% |     3.174/18.2% |  23.31% | IMPROVED

So the data seems to be quite unstable when the system is saturated, which
might be another issue that I'll investigate soon.