Re: [linus:master] [mm] 9890ecab6a: vm-scalability.throughput 3.8% regression

[Ankur Arora]

kernel test robot <oliver.sang@xxxxxxxxx> writes:

> Hello,
>
> kernel test robot noticed a 3.8% regression of vm-scalability.throughput on:
>

[ ... ]

> testcase: vm-scalability
> config: x86_64-rhel-9.4
> compiler: gcc-14
> test machine: 224 threads 4 sockets Intel(R) Xeon(R) Platinum 8380H CPU @ 2.90GHz (Cooper Lake) with 192G memory
> parameters:
>
> 	runtime: 300s
> 	size: 8T
> 	test: anon-w-seq-mt
> 	cpufreq_governor: performance

This test exercises the THP sequential zeroing path.

>      15142            -4.2%      14512        vm-scalability.time.system_time
>      16939            +6.1%      17975        vm-scalability.time.user_time

stime drops because folio_zero_user() is more efficient. But utime goes
up because of higher user miss rate since folio_zero_user() is now clearing
sequentially instead of the earlier left right fashion:

>      61.69            +9.8       71.51        perf-stat.i.cache-miss-rate%
>  6.147e+08           +10.7%  6.805e+08        perf-stat.i.cache-misses
>  9.904e+08            -4.7%  9.436e+08        perf-stat.i.cache-references
>       2.17            +8.2%       2.35        perf-stat.i.cpi


I had noted similar behaviour with anon-w-seq-hugetlb in 93552c9a3350:

   vm-scalability/anon-w-seq-hugetlb: this workload runs with 384 processes
   (one for each CPU) each zeroing anonymously mapped hugetlb memory which
   is then accessed sequentially.

                            stime                utime

   discontiguous-page      1739.93 ( +- 6.15% )  1016.61 ( +- 4.75% )
   contiguous-page         1853.70 ( +- 2.51% )  1187.13 ( +- 3.50% )
   batched-pages           1756.75 ( +- 2.98% )  1133.32 ( +- 4.89% )
   neighbourhood-last      1725.18 ( +- 4.59% )  1123.78 ( +- 7.38% )

  Both stime and utime largely respond somewhat expectedly. There is a
  fair amount of run to run variation but the general trend is that the
  stime drops and utime increases. There are a few oddities, like
  contiguous-page performing very differently from batched-pages.

  As such this is likely an uncommon pattern where we saturate the memory
  bandwidth (since all CPUs are running the test) and at the same time
  are cache constrained because we access the entire region.


Ankur

> cb431accb36e51b6 9890ecab6ad9c0d3d342469f3b6
> ---------------- ---------------------------
>          %stddev     %change         %stddev
>              \          |                \
>       0.08 ±  3%      +8.7%       0.09 ±  3%  vm-scalability.free_time
>     357969            -6.6%     334511        vm-scalability.median
>  1.034e+08            -3.8%   99382138        vm-scalability.throughput
>     634243           -13.6%     548120 ±  6%  vm-scalability.time.involuntary_context_switches
>   12706518            -6.6%   11872543        vm-scalability.time.minor_page_faults
>      15142            -4.2%      14512        vm-scalability.time.system_time
>      16939            +6.1%      17975        vm-scalability.time.user_time
>     251227            -6.8%     234071        vm-scalability.time.voluntary_context_switches
>  1.791e+10            -6.6%  1.674e+10        vm-scalability.workload
>       0.30            -7.5%       0.28        turbostat.IPC
>       9203            -5.5%       8693        vmstat.system.cs
>       0.08            +0.0        0.08        mpstat.cpu.all.soft%
>      25.14            +1.5       26.62        mpstat.cpu.all.usr%
>       3.13           +18.3%       3.71        perf-stat.i.MPKI
>   6.22e+10            -6.6%   5.81e+10        perf-stat.i.branch-instructions
>      61.69            +9.8       71.51        perf-stat.i.cache-miss-rate%
>  6.147e+08           +10.7%  6.805e+08        perf-stat.i.cache-misses
>  9.904e+08            -4.7%  9.436e+08        perf-stat.i.cache-references
>       9303            -5.2%       8823        perf-stat.i.context-switches
>       2.17            +8.2%       2.35        perf-stat.i.cpi
>     598.97            -4.6%     571.28        perf-stat.i.cpu-migrations
>   1.95e+11            -6.6%  1.822e+11        perf-stat.i.instructions
>       0.47            -7.2%       0.43        perf-stat.i.ipc
>      43153            -6.5%      40334        perf-stat.i.minor-faults
>      43153            -6.5%      40335        perf-stat.i.page-faults
>       3.16           +18.5%       3.74        perf-stat.overall.MPKI
>       0.02            +0.0        0.03        perf-stat.overall.branch-miss-rate%
>      62.11           +10.1       72.19        perf-stat.overall.cache-miss-rate%
>       2.19            +8.3%       2.37        perf-stat.overall.cpi
>     692.89            -8.6%     633.07        perf-stat.overall.cycles-between-cache-misses
>       0.46            -7.6%       0.42        perf-stat.overall.ipc
>  6.121e+10            -6.8%  5.705e+10        perf-stat.ps.branch-instructions
>  6.054e+08           +10.5%  6.689e+08        perf-stat.ps.cache-misses
>  9.747e+08            -4.9%  9.266e+08        perf-stat.ps.cache-references
>       9124            -5.6%       8613        perf-stat.ps.context-switches
>     583.66            -4.9%     555.21        perf-stat.ps.cpu-migrations
>  1.919e+11            -6.8%  1.789e+11        perf-stat.ps.instructions
>      42389            -6.7%      39549        perf-stat.ps.minor-faults
>      42389            -6.7%      39549        perf-stat.ps.page-faults
>  5.812e+13            -6.5%  5.434e+13        perf-stat.total.instructions
>      40.26           -40.3        0.00        perf-profile.calltrace.cycles-pp.clear_subpage.folio_zero_user.vma_alloc_anon_folio_pmd.__do_huge_pmd_anonymous_page.__handle_mm_fault
>      40.76            -2.1       38.66        perf-profile.calltrace.cycles-pp.folio_zero_user.vma_alloc_anon_folio_pmd.__do_huge_pmd_anonymous_page.__handle_mm_fault.handle_mm_fault
>      42.59            -2.0       40.61        perf-profile.calltrace.cycles-pp.asm_exc_page_fault.do_access
>      42.54            -2.0       40.57        perf-profile.calltrace.cycles-pp.do_user_addr_fault.exc_page_fault.asm_exc_page_fault.do_access
>      42.54            -2.0       40.57        perf-profile.calltrace.cycles-pp.exc_page_fault.asm_exc_page_fault.do_access
>      42.40            -2.0       40.43        perf-profile.calltrace.cycles-pp.handle_mm_fault.do_user_addr_fault.exc_page_fault.asm_exc_page_fault.do_access
>      42.32            -2.0       40.36        perf-profile.calltrace.cycles-pp.__handle_mm_fault.handle_mm_fault.do_user_addr_fault.exc_page_fault.asm_exc_page_fault
>      42.23            -2.0       40.27        perf-profile.calltrace.cycles-pp.__do_huge_pmd_anonymous_page.__handle_mm_fault.handle_mm_fault.do_user_addr_fault.exc_page_fault
>      41.70            -2.0       39.74        perf-profile.calltrace.cycles-pp.vma_alloc_anon_folio_pmd.__do_huge_pmd_anonymous_page.__handle_mm_fault.handle_mm_fault.do_user_addr_fault
>       0.76            -0.0        0.72        perf-profile.calltrace.cycles-pp.vma_alloc_folio_noprof.vma_alloc_anon_folio_pmd.__do_huge_pmd_anonymous_page.__handle_mm_fault.handle_mm_fault
>       0.72            -0.0        0.68        perf-profile.calltrace.cycles-pp.__alloc_frozen_pages_noprof.alloc_pages_mpol.vma_alloc_folio_noprof.vma_alloc_anon_folio_pmd.__do_huge_pmd_anonymous_page
>       0.67            -0.0        0.64        perf-profile.calltrace.cycles-pp.get_page_from_freelist.__alloc_frozen_pages_noprof.alloc_pages_mpol.vma_alloc_folio_noprof.vma_alloc_anon_folio_pmd
>       0.72            -0.0        0.69        perf-profile.calltrace.cycles-pp.alloc_pages_mpol.vma_alloc_folio_noprof.vma_alloc_anon_folio_pmd.__do_huge_pmd_anonymous_page.__handle_mm_fault
>       0.56            -0.0        0.54        perf-profile.calltrace.cycles-pp.prep_new_page.get_page_from_freelist.__alloc_frozen_pages_noprof.alloc_pages_mpol.vma_alloc_folio_noprof
>       0.00            +0.8        0.76 ±  2%  perf-profile.calltrace.cycles-pp.asm_sysvec_apic_timer_interrupt.folio_zero_user.vma_alloc_anon_folio_pmd.__do_huge_pmd_anonymous_page.__handle_mm_fault
>      30.25            +1.2       31.46        perf-profile.calltrace.cycles-pp.do_rw_once
>      40.49           -40.5        0.00        perf-profile.children.cycles-pp.clear_subpage
>      42.61            -2.0       40.63        perf-profile.children.cycles-pp.asm_exc_page_fault
>      42.55            -2.0       40.58        perf-profile.children.cycles-pp.exc_page_fault
>      42.54            -2.0       40.57        perf-profile.children.cycles-pp.do_user_addr_fault
>      42.40            -2.0       40.43        perf-profile.children.cycles-pp.handle_mm_fault
>      42.33            -2.0       40.36        perf-profile.children.cycles-pp.__handle_mm_fault
>      42.23            -2.0       40.27        perf-profile.children.cycles-pp.__do_huge_pmd_anonymous_page
>      41.70            -2.0       39.74        perf-profile.children.cycles-pp.vma_alloc_anon_folio_pmd
>      40.83            -1.9       38.92        perf-profile.children.cycles-pp.folio_zero_user
>      63.93            -1.2       62.77        perf-profile.children.cycles-pp.do_access
>       0.95            -0.0        0.91        perf-profile.children.cycles-pp.__alloc_frozen_pages_noprof
>       0.78            -0.0        0.74        perf-profile.children.cycles-pp.vma_alloc_folio_noprof
>       0.95            -0.0        0.92        perf-profile.children.cycles-pp.alloc_pages_mpol
>       0.79            -0.0        0.76        perf-profile.children.cycles-pp.get_page_from_freelist
>       0.63            -0.0        0.60        perf-profile.children.cycles-pp.prep_new_page
>      40.31            +2.5       42.80        perf-profile.children.cycles-pp.do_rw_once
>      39.77           -39.8        0.00        perf-profile.self.cycles-pp.clear_subpage
>       9.54            -0.3        9.23        perf-profile.self.cycles-pp.do_access
>       0.55            -0.0        0.53        perf-profile.self.cycles-pp.prep_new_page
>      38.35            +2.6       40.96        perf-profile.self.cycles-pp.do_rw_once
>       0.36 ±  2%     +38.0       38.32        perf-profile.self.cycles-pp.folio_zero_user
>
>
>
>
> Disclaimer:
> Results have been estimated based on internal Intel analysis and are provided
> for informational purposes only. Any difference in system hardware or software
> design or configuration may affect actual performance.