Re: [linus:master] [mm] d2136d749d: vm-scalability.throughput -7.1% regression

[Baolin Wang <baolin.wang@linux.alibaba.com>]

On 2024/6/20 15:38, Huang, Ying wrote:
> Baolin Wang <baolin.wang@linux.alibaba.com> writes:
> 
>> On 2024/6/20 10:39, kernel test robot wrote:
>>> Hello,
>>> kernel test robot noticed a -7.1% regression of
>>> vm-scalability.throughput on:
>>> commit: d2136d749d76af980b3accd72704eea4eab625bd ("mm: support
>>> multi-size THP numa balancing")
>>> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git master
>>> [still regression on linus/master
>>> 92e5605a199efbaee59fb19e15d6cc2103a04ec2]
>>> testcase: vm-scalability
>>> test machine: 128 threads 2 sockets Intel(R) Xeon(R) Gold 6338 CPU @ 2.00GHz (Ice Lake) with 256G memory
>>> parameters:
>>> 	runtime: 300s
>>> 	size: 512G
>>> 	test: anon-cow-rand-hugetlb
>>> 	cpufreq_governor: performance
>>
>> Thanks for reporting. IIUC numa balancing will not scan hugetlb VMA,
>> I'm not sure how this patch affects the performance of hugetlb cow,
>> but let me try to reproduce it.
>>
>>
>>> If you fix the issue in a separate patch/commit (i.e. not just a new version of
>>> the same patch/commit), kindly add following tags
>>> | Reported-by: kernel test robot <oliver.sang@intel.com>
>>> | Closes: https://lore.kernel.org/oe-lkp/202406201010.a1344783-oliver.sang@intel.com
>>> Details are as below:
>>> -------------------------------------------------------------------------------------------------->
>>> The kernel config and materials to reproduce are available at:
>>> https://download.01.org/0day-ci/archive/20240620/202406201010.a1344783-oliver.sang@intel.com
>>> =========================================================================================
>>> compiler/cpufreq_governor/kconfig/rootfs/runtime/size/tbox_group/test/testcase:
>>>     gcc-13/performance/x86_64-rhel-8.3/debian-12-x86_64-20240206.cgz/300s/512G/lkp-icl-2sp2/anon-cow-rand-hugetlb/vm-scalability
>>> commit:
>>>     6b0ed7b3c7 ("mm: factor out the numa mapping rebuilding into a new helper")
>>>     d2136d749d ("mm: support multi-size THP numa balancing")
>>> 6b0ed7b3c77547d2 d2136d749d76af980b3accd7270
>>> ---------------- ---------------------------
>>>            %stddev     %change         %stddev
>>>                \          |                \
>>>        12.02            -1.3       10.72 ±  4%  mpstat.cpu.all.sys%
>>>      1228757            +3.0%    1265679        proc-vmstat.pgfault
> 
> Also from other proc-vmstat stats,
> 
>       21770  36%      +6.1%      23098  28%  proc-vmstat.numa_hint_faults
>        6168 107%     +48.8%       9180  18%  proc-vmstat.numa_hint_faults_local
>      154537  15%     +23.5%     190883  17%  proc-vmstat.numa_pte_updates
> 
> After your patch, more hint page faults occurs, I think this is expected.
> 
> Then, tasks may be moved between sockets because of that, so that some
> hugetlb page access becomes remote?

After trying to reproduce this case, I also find that more hint page 
faults occur. And I think that is casued by changing 
"folio_ref_count(folio) != 1" to "folio_likely_mapped_shared(folio)", 
which results in scanning more exclusive pages, so I think this is 
expected from the previous discussion.

Yes, I think your analysis is correct, some hugetlb page accesses become 
remote due to task migration.