12:27 p.m.
FYI, we noticed a -6.3% regression of unixbench.score due to commit:
commit 5c0a85fad949212b3e059692deecdeed74ae7ec7 ("mm: make faultaround produce old
ptes")
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
in testcase: unixbench
on test machine: lituya: 16 threads Haswell High-end Desktop (i7-5960X 3.0G) with 16G
memory
with following parameters: cpufreq_governor=performance/nr_task=1/test=shell8
Details are as below:
-------------------------------------------------------------------------------------------------->
=========================================================================================
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
commit:
4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
5c0a85fad949212b3e059692deecdeed74ae7ec7
4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
---------------- --------------------------
fail:runs %reproduction fail:runs
| | |
3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
%stddev %change %stddev
\ | \
14321 ± 0% -6.3% 13425 ± 0% unixbench.score
1996897 ± 0% -6.1% 1874635 ± 0% unixbench.time.involuntary_context_switches
1.721e+08 ± 0% -6.2% 1.613e+08 ± 0% unixbench.time.minor_page_faults
758.65 ± 0% -3.0% 735.86 ± 0% unixbench.time.system_time
387.66 ± 0% +5.4% 408.49 ± 0% unixbench.time.user_time
5950278 ± 0% -6.2% 5583456 ± 0% unixbench.time.voluntary_context_switches
1960642 ± 0% -11.4% 1737753 ± 0% cpuidle.C1-HSW.usage
5851 ± 0% -43.8% 3286 ± 1% proc-vmstat.nr_active_file
46185 ± 0% -21.2% 36385 ± 2% meminfo.Active
23404 ± 0% -43.8% 13147 ± 1% meminfo.Active(file)
4109 ± 5% -19.6% 3302 ± 4% slabinfo.pid.active_objs
4109 ± 5% -19.6% 3302 ± 4% slabinfo.pid.num_objs
94603 ± 0% -5.7% 89247 ± 0% vmstat.system.cs
8976 ± 0% -2.5% 8754 ± 0% vmstat.system.in
3.38 ± 2% +11.8% 3.77 ± 0% turbostat.CPU%c3
0.24 ±101% -86.3% 0.03 ± 54% turbostat.Pkg%pc3
66.53 ± 0% -1.7% 65.41 ± 0% turbostat.PkgWatt
2061 ± 1% -8.5% 1886 ± 0% sched_debug.cfs_rq:/.exec_clock.stddev
737154 ± 5% +10.8% 817107 ± 3% sched_debug.cpu.avg_idle.max
133057 ± 5% -33.2% 88864 ± 11% sched_debug.cpu.avg_idle.min
181562 ± 8% +15.9% 210434 ± 3% sched_debug.cpu.avg_idle.stddev
0.97 ± 7% +19.0% 1.16 ± 8% sched_debug.cpu.clock.stddev
0.97 ± 7% +19.0% 1.16 ± 8% sched_debug.cpu.clock_task.stddev
248.06 ± 11% +31.0% 324.94 ± 8% sched_debug.cpu.cpu_load[1].max
55.65 ± 14% +28.1% 71.30 ± 8% sched_debug.cpu.cpu_load[1].stddev
233.38 ± 10% +34.4% 313.56 ± 8% sched_debug.cpu.cpu_load[2].max
49.79 ± 15% +35.6% 67.50 ± 9% sched_debug.cpu.cpu_load[2].stddev
233.25 ± 12% +29.9% 302.94 ± 6% sched_debug.cpu.cpu_load[3].max
46.56 ± 8% +12.2% 52.25 ± 6% sched_debug.cpu.cpu_load[3].min
48.51 ± 15% +31.4% 63.76 ± 7% sched_debug.cpu.cpu_load[3].stddev
238.44 ± 12% +19.0% 283.69 ± 3% sched_debug.cpu.cpu_load[4].max
49.56 ± 9% +13.4% 56.19 ± 4% sched_debug.cpu.cpu_load[4].min
48.22 ± 13% +20.1% 57.93 ± 5% sched_debug.cpu.cpu_load[4].stddev
14792 ± 30% +71.9% 25424 ± 17% sched_debug.cpu.curr->pid.avg
42862 ± 1% +42.6% 61121 ± 0% sched_debug.cpu.curr->pid.max
19466 ± 10% +35.4% 26351 ± 9% sched_debug.cpu.curr->pid.stddev
1067 ± 6% -14.9% 909.35 ± 4% sched_debug.cpu.ttwu_local.stddev
To reproduce:
git clone git://git.kernel.org/pub/scm/linux/kernel/git/wfg/lkp-tests.git
cd lkp-tests
bin/lkp install job.yaml # job file is attached in this email
bin/lkp run job.yaml
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.
Thanks,
Xiaolong
7:51 p.m.
On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel test robot wrote:
FYI, we noticed a -6.3% regression of unixbench.score due to commit:
commit 5c0a85fad949212b3e059692deecdeed74ae7ec7 ("mm: make faultaround produce old
ptes")
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
in testcase: unixbench
on test machine: lituya: 16 threads Haswell High-end Desktop (i7-5960X 3.0G) with 16G
memory
with following parameters: cpufreq_governor=performance/nr_task=1/test=shell8
Details are as below:
-------------------------------------------------------------------------------------------------->
=========================================================================================
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
commit:
4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
5c0a85fad949212b3e059692deecdeed74ae7ec7
4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
---------------- --------------------------
fail:runs %reproduction fail:runs
| | |
3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
%stddev %change %stddev
\ | \
14321 ± 0% -6.3% 13425 ± 0% unixbench.score
1996897 ± 0% -6.1% 1874635 ± 0%
unixbench.time.involuntary_context_switches
1.721e+08 ± 0% -6.2% 1.613e+08 ± 0% unixbench.time.minor_page_faults
758.65 ± 0% -3.0% 735.86 ± 0% unixbench.time.system_time
387.66 ± 0% +5.4% 408.49 ± 0% unixbench.time.user_time
5950278 ± 0% -6.2% 5583456 ± 0% unixbench.time.voluntary_context_switches
That's weird.
I don't understand why the change would reduce number or minor faults.
It should stay the same on x86-64. Rise of user_time is puzzling too.
Hm. Is reproducible? Across reboot?
--
Kirill A. Shutemov
5:21 p.m.
"Kirill A. Shutemov" <kirill.shutemov(a)linux.intel.com> writes:
On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel test robot wrote:
>
> FYI, we noticed a -6.3% regression of unixbench.score due to commit:
>
> commit 5c0a85fad949212b3e059692deecdeed74ae7ec7 ("mm: make faultaround produce
old ptes")
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
>
> in testcase: unixbench
> on test machine: lituya: 16 threads Haswell High-end Desktop (i7-5960X 3.0G) with 16G
memory
> with following parameters: cpufreq_governor=performance/nr_task=1/test=shell8
>
>
> Details are as below:
>
-------------------------------------------------------------------------------------------------->
>
>
>
=========================================================================================
> compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
>
> commit:
> 4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
> 5c0a85fad949212b3e059692deecdeed74ae7ec7
>
> 4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
> ---------------- --------------------------
> fail:runs %reproduction fail:runs
> | | |
> 3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
> %stddev %change %stddev
> \ | \
> 14321 . 0% -6.3% 13425 . 0% unixbench.score
> 1996897 . 0% -6.1% 1874635 . 0%
unixbench.time.involuntary_context_switches
> 1.721e+08 . 0% -6.2% 1.613e+08 . 0% unixbench.time.minor_page_faults
> 758.65 . 0% -3.0% 735.86 . 0% unixbench.time.system_time
> 387.66 . 0% +5.4% 408.49 . 0% unixbench.time.user_time
> 5950278 . 0% -6.2% 5583456 . 0%
unixbench.time.voluntary_context_switches
That's weird.
I don't understand why the change would reduce number or minor faults.
It should stay the same on x86-64. Rise of user_time is puzzling too.
unixbench runs in fixed time mode. That is, the total time to run
unixbench is fixed, but the work done varies. So the minor_page_faults
change may reflect only the work done.
Hm. Is reproducible? Across reboot?
Yes. LKP will run every benchmark after reboot via kexec. We run 3
times for both the commit and its parent. The result is quite stable.
You can find the standard deviation in percent is near 0 across
different runs. Here is another comparison with profile data.
=========================================================================================
compiler/cpufreq_governor/debug-setup/kconfig/nr_task/rootfs/tbox_group/test/testcase:
gcc-4.9/performance/profile/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
commit:
4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
5c0a85fad949212b3e059692deecdeed74ae7ec7
4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
---------------- --------------------------
%stddev %change %stddev
\ | \
14056 ± 0% -6.3% 13172 ± 0% unixbench.score
6464046 ± 0% -6.1% 6071922 ± 0% unixbench.time.involuntary_context_switches
5.555e+08 ± 0% -6.2% 5.211e+08 ± 0% unixbench.time.minor_page_faults
2537 ± 0% -3.2% 2455 ± 0% unixbench.time.system_time
1284 ± 0% +5.8% 1359 ± 0% unixbench.time.user_time
19192611 ± 0% -6.2% 18010830 ± 0% unixbench.time.voluntary_context_switches
7709931 ± 0% -11.0% 6860574 ± 0% cpuidle.C1-HSW.usage
6900 ± 1% -43.9% 3871 ± 0% proc-vmstat.nr_active_file
40813 ± 1% -77.9% 9015 ±114% softirqs.NET_RX
111331 ± 1% -13.3% 96503 ± 0% meminfo.Active
27603 ± 1% -43.9% 15486 ± 0% meminfo.Active(file)
93169 ± 0% -5.8% 87766 ± 0% vmstat.system.cs
19768 ± 0% -1.7% 19437 ± 0% vmstat.system.in
6.22 ± 0% +10.3% 6.86 ± 0% turbostat.CPU%c3
0.02 ± 20% -85.7% 0.00 ±141% turbostat.Pkg%pc3
68.99 ± 0% -1.7% 67.84 ± 0% turbostat.PkgWatt
1.38 ± 5% -42.0% 0.80 ± 5%
perf-profile.cycles-pp.page_remove_rmap.unmap_page_range.unmap_single_vma.unmap_vmas.exit_mmap
0.83 ± 4% +28.8% 1.07 ± 21%
perf-profile.cycles-pp.release_pages.free_pages_and_swap_cache.tlb_flush_mmu_free.tlb_finish_mmu.exit_mmap
1.55 ± 3% -10.6% 1.38 ± 2%
perf-profile.cycles-pp.unmap_single_vma.unmap_vmas.exit_mmap.mmput.flush_old_exec
1.59 ± 3% -9.8% 1.44 ± 3%
perf-profile.cycles-pp.unmap_vmas.exit_mmap.mmput.flush_old_exec.load_elf_binary
389.00 ± 0% +32.1% 514.00 ± 8% slabinfo.file_lock_cache.active_objs
389.00 ± 0% +32.1% 514.00 ± 8% slabinfo.file_lock_cache.num_objs
7075 ± 3% -17.7% 5823 ± 7% slabinfo.pid.active_objs
7075 ± 3% -17.7% 5823 ± 7% slabinfo.pid.num_objs
0.67 ± 34% +86.4% 1.24 ± 30% sched_debug.cfs_rq:/.runnable_load_avg.min
-9013 ± -1% +14.4% -10315 ± -9% sched_debug.cfs_rq:/.spread0.avg
83127 ± 5% +16.9% 97163 ± 8% sched_debug.cpu.avg_idle.min
17777 ± 16% +66.6% 29608 ± 22% sched_debug.cpu.curr->pid.avg
50223 ± 10% +49.3% 74974 ± 0% sched_debug.cpu.curr->pid.max
22281 ± 13% +51.8% 33816 ± 6% sched_debug.cpu.curr->pid.stddev
251.79 ± 5% -13.8% 217.15 ± 5% sched_debug.cpu.nr_uninterruptible.max
-261.12 ± -2% -13.4% -226.03 ± -1% sched_debug.cpu.nr_uninterruptible.min
221.14 ± 3% -14.7% 188.60 ± 1% sched_debug.cpu.nr_uninterruptible.stddev
1.94e+11 ± 0% -5.8% 1.827e+11 ± 0% perf-stat.L1-dcache-load-misses
3.496e+12 ± 0% -6.5% 3.268e+12 ± 0% perf-stat.L1-dcache-loads
2.262e+12 ± 1% -5.5% 2.137e+12 ± 0% perf-stat.L1-dcache-stores
9.711e+10 ± 0% -3.7% 9.353e+10 ± 0% perf-stat.L1-icache-load-misses
8.051e+08 ± 0% -8.8% 7.343e+08 ± 1% perf-stat.LLC-load-misses
7.184e+10 ± 1% -5.6% 6.78e+10 ± 0% perf-stat.LLC-loads
5.867e+08 ± 2% -7.0% 5.456e+08 ± 0% perf-stat.LLC-store-misses
1.524e+10 ± 1% -5.6% 1.438e+10 ± 0% perf-stat.LLC-stores
2.711e+12 ± 0% -6.3% 2.539e+12 ± 0% perf-stat.branch-instructions
5.948e+10 ± 0% -3.9% 5.715e+10 ± 0% perf-stat.branch-load-misses
2.715e+12 ± 0% -6.4% 2.542e+12 ± 0% perf-stat.branch-loads
5.947e+10 ± 0% -3.9% 5.713e+10 ± 0% perf-stat.branch-misses
1.448e+09 ± 0% -9.3% 1.313e+09 ± 1% perf-stat.cache-misses
1.931e+11 ± 0% -5.8% 1.818e+11 ± 0% perf-stat.cache-references
58882705 ± 0% -5.8% 55467522 ± 0% perf-stat.context-switches
17037466 ± 0% -6.1% 15999111 ± 0% perf-stat.cpu-migrations
6.732e+09 ± 1% +90.7% 1.284e+10 ± 0% perf-stat.dTLB-load-misses
3.474e+12 ± 0% -6.6% 3.245e+12 ± 0% perf-stat.dTLB-loads
1.215e+09 ± 0% -5.5% 1.149e+09 ± 0% perf-stat.dTLB-store-misses
2.286e+12 ± 0% -5.8% 2.153e+12 ± 0% perf-stat.dTLB-stores
3.511e+09 ± 0% +20.4% 4.226e+09 ± 0% perf-stat.iTLB-load-misses
2.317e+09 ± 0% -6.8% 2.16e+09 ± 0% perf-stat.iTLB-loads
1.343e+13 ± 0% -6.0% 1.263e+13 ± 0% perf-stat.instructions
5.504e+08 ± 0% -6.2% 5.163e+08 ± 0% perf-stat.minor-faults
8.09e+08 ± 1% -9.0% 7.36e+08 ± 1% perf-stat.node-loads
5.932e+08 ± 0% -8.7% 5.417e+08 ± 1% perf-stat.node-stores
5.504e+08 ± 0% -6.2% 5.163e+08 ± 0% perf-stat.page-faults
Best Regards,
Huang, Ying
6:41 p.m.
"Huang, Ying" <ying.huang(a)intel.com> writes:
"Kirill A. Shutemov"
<kirill.shutemov(a)linux.intel.com> writes:
> On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel test robot wrote:
>>
>> FYI, we noticed a -6.3% regression of unixbench.score due to commit:
>>
>> commit 5c0a85fad949212b3e059692deecdeed74ae7ec7 ("mm: make faultaround
produce old ptes")
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
>>
>> in testcase: unixbench
>> on test machine: lituya: 16 threads Haswell High-end Desktop (i7-5960X 3.0G) with
16G memory
>> with following parameters: cpufreq_governor=performance/nr_task=1/test=shell8
>>
>>
>> Details are as below:
>>
-------------------------------------------------------------------------------------------------->
>>
>>
>>
=========================================================================================
>> compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
>>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
>>
>> commit:
>> 4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
>> 5c0a85fad949212b3e059692deecdeed74ae7ec7
>>
>> 4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
>> ---------------- --------------------------
>> fail:runs %reproduction fail:runs
>> | | |
>> 3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
>> %stddev %change %stddev
>> \ | \
>> 14321 . 0% -6.3% 13425 . 0% unixbench.score
>> 1996897 . 0% -6.1% 1874635 . 0%
unixbench.time.involuntary_context_switches
>> 1.721e+08 . 0% -6.2% 1.613e+08 . 0% unixbench.time.minor_page_faults
>> 758.65 . 0% -3.0% 735.86 . 0% unixbench.time.system_time
>> 387.66 . 0% +5.4% 408.49 . 0% unixbench.time.user_time
>> 5950278 . 0% -6.2% 5583456 . 0%
unixbench.time.voluntary_context_switches
>
> That's weird.
>
> I don't understand why the change would reduce number or minor faults.
> It should stay the same on x86-64. Rise of user_time is puzzling too.
unixbench runs in fixed time mode. That is, the total time to run
unixbench is fixed, but the work done varies. So the minor_page_faults
change may reflect only the work done.
> Hm. Is reproducible? Across reboot?
And FYI, there is no swap setup for test, all root file system including
benchmark files are in tmpfs, so no real page reclaim will be
triggered. But it appears that active file cache reduced after the
commit.
111331 ± 1% -13.3% 96503 ± 0% meminfo.Active
27603 ± 1% -43.9% 15486 ± 0% meminfo.Active(file)
I think this is the expected behavior of the commit?
Best Regards,
Huang, Ying
6:58 p.m.
On Wed, Jun 08, 2016 at 04:41:37PM +0800, Huang, Ying wrote:
"Huang, Ying" <ying.huang(a)intel.com> writes:
> "Kirill A. Shutemov" <kirill.shutemov(a)linux.intel.com> writes:
>
>> On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel test robot wrote:
>>>
>>> FYI, we noticed a -6.3% regression of unixbench.score due to commit:
>>>
>>> commit 5c0a85fad949212b3e059692deecdeed74ae7ec7 ("mm: make faultaround
produce old ptes")
>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
>>>
>>> in testcase: unixbench
>>> on test machine: lituya: 16 threads Haswell High-end Desktop (i7-5960X 3.0G)
with 16G memory
>>> with following parameters:
cpufreq_governor=performance/nr_task=1/test=shell8
>>>
>>>
>>> Details are as below:
>>>
-------------------------------------------------------------------------------------------------->
>>>
>>>
>>>
=========================================================================================
>>> compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
>>>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
>>>
>>> commit:
>>> 4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
>>> 5c0a85fad949212b3e059692deecdeed74ae7ec7
>>>
>>> 4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
>>> ---------------- --------------------------
>>> fail:runs %reproduction fail:runs
>>> | | |
>>> 3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
>>> %stddev %change %stddev
>>> \ | \
>>> 14321 . 0% -6.3% 13425 . 0% unixbench.score
>>> 1996897 . 0% -6.1% 1874635 . 0%
unixbench.time.involuntary_context_switches
>>> 1.721e+08 . 0% -6.2% 1.613e+08 . 0%
unixbench.time.minor_page_faults
>>> 758.65 . 0% -3.0% 735.86 . 0% unixbench.time.system_time
>>> 387.66 . 0% +5.4% 408.49 . 0% unixbench.time.user_time
>>> 5950278 . 0% -6.2% 5583456 . 0%
unixbench.time.voluntary_context_switches
>>
>> That's weird.
>>
>> I don't understand why the change would reduce number or minor faults.
>> It should stay the same on x86-64. Rise of user_time is puzzling too.
>
> unixbench runs in fixed time mode. That is, the total time to run
> unixbench is fixed, but the work done varies. So the minor_page_faults
> change may reflect only the work done.
>
>> Hm. Is reproducible? Across reboot?
>
And FYI, there is no swap setup for test, all root file system including
benchmark files are in tmpfs, so no real page reclaim will be
triggered. But it appears that active file cache reduced after the
commit.
111331 ± 1% -13.3% 96503 ± 0% meminfo.Active
27603 ± 1% -43.9% 15486 ± 0% meminfo.Active(file)
I think this is the expected behavior of the commit?
Yes, it's expected.
After the change faularound would produce old pte. It means there's more
chance for these pages to be on inactive lru, unless somebody actually
touch them and flip accessed bit.
I wounder if this regression can attributed to cost of setting accessed
bit. It looks too high, but who knows.
I don't have time to do testing myself right now. I will put this on todo
list.
--
Kirill A. Shutemov
10:49 a.m.
"Kirill A. Shutemov" <kirill.shutemov(a)linux.intel.com> writes:
On Wed, Jun 08, 2016 at 04:41:37PM +0800, Huang, Ying wrote:
> "Huang, Ying" <ying.huang(a)intel.com> writes:
>
> > "Kirill A. Shutemov" <kirill.shutemov(a)linux.intel.com> writes:
> >
> >> On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel test robot wrote:
> >>>
> >>> FYI, we noticed a -6.3% regression of unixbench.score due to commit:
> >>>
> >>> commit 5c0a85fad949212b3e059692deecdeed74ae7ec7 ("mm: make
faultaround produce old ptes")
> >>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
master
> >>>
> >>> in testcase: unixbench
> >>> on test machine: lituya: 16 threads Haswell High-end Desktop (i7-5960X
3.0G) with 16G memory
> >>> with following parameters:
cpufreq_governor=performance/nr_task=1/test=shell8
> >>>
> >>>
> >>> Details are as below:
> >>>
-------------------------------------------------------------------------------------------------->
> >>>
> >>>
> >>>
=========================================================================================
> >>>
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
> >>>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
> >>>
> >>> commit:
> >>> 4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
> >>> 5c0a85fad949212b3e059692deecdeed74ae7ec7
> >>>
> >>> 4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
> >>> ---------------- --------------------------
> >>> fail:runs %reproduction fail:runs
> >>> | | |
> >>> 3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
> >>> %stddev %change %stddev
> >>> \ | \
> >>> 14321 . 0% -6.3% 13425 . 0% unixbench.score
> >>> 1996897 . 0% -6.1% 1874635 . 0%
unixbench.time.involuntary_context_switches
> >>> 1.721e+08 . 0% -6.2% 1.613e+08 . 0%
unixbench.time.minor_page_faults
> >>> 758.65 . 0% -3.0% 735.86 . 0%
unixbench.time.system_time
> >>> 387.66 . 0% +5.4% 408.49 . 0% unixbench.time.user_time
> >>> 5950278 . 0% -6.2% 5583456 . 0%
unixbench.time.voluntary_context_switches
> >>
> >> That's weird.
> >>
> >> I don't understand why the change would reduce number or minor faults.
> >> It should stay the same on x86-64. Rise of user_time is puzzling too.
> >
> > unixbench runs in fixed time mode. That is, the total time to run
> > unixbench is fixed, but the work done varies. So the minor_page_faults
> > change may reflect only the work done.
> >
> >> Hm. Is reproducible? Across reboot?
> >
>
> And FYI, there is no swap setup for test, all root file system including
> benchmark files are in tmpfs, so no real page reclaim will be
> triggered. But it appears that active file cache reduced after the
> commit.
>
> 111331 . 1% -13.3% 96503 . 0% meminfo.Active
> 27603 . 1% -43.9% 15486 . 0% meminfo.Active(file)
>
> I think this is the expected behavior of the commit?
Yes, it's expected.
After the change faularound would produce old pte. It means there's more
chance for these pages to be on inactive lru, unless somebody actually
touch them and flip accessed bit.
I wounder if this regression can attributed to cost of setting accessed
bit. It looks too high, but who knows.
From perf profile, the time spent in page_fault and its children
functions are almost same (7.85% vs 7.81%). So the time spent in page
fault and page table operation itself doesn't changed much. So, you
mean CPU may be slower to load the page table entry to TLB if accessed
bit is not set?
I don't have time to do testing myself right now. I will put this
on todo
list.
Which kind of test your want to do? I want to check whether I can help.
Best Regards,
Huang, Ying
11:02 a.m.
On Sat, Jun 11, 2016 at 5:49 PM, Huang, Ying <ying.huang(a)intel.com> wrote:
From perf profile, the time spent in page_fault and its children
functions are almost same (7.85% vs 7.81%). So the time spent in page
fault and page table operation itself doesn't changed much. So, you
mean CPU may be slower to load the page table entry to TLB if accessed
bit is not set?
So the CPU does take a microfault internally when it needs to set the
accessed/dirty bit. It's not architecturally visible, but you can see
it when you do timing loops.
I've timed it at over a thousand cycles on at least some CPU's, but
that's still peanuts compared to a real page fault. It shouldn't be
*that* noticeable, ie no way it's a 6% regression on its own.
Linus
7:02 p.m.
Linus Torvalds <torvalds(a)linux-foundation.org> writes:
On Sat, Jun 11, 2016 at 5:49 PM, Huang, Ying
<ying.huang(a)intel.com> wrote:
>
> From perf profile, the time spent in page_fault and its children
> functions are almost same (7.85% vs 7.81%). So the time spent in page
> fault and page table operation itself doesn't changed much. So, you
> mean CPU may be slower to load the page table entry to TLB if accessed
> bit is not set?
So the CPU does take a microfault internally when it needs to set the
accessed/dirty bit. It's not architecturally visible, but you can see
it when you do timing loops.
I've timed it at over a thousand cycles on at least some CPU's, but
that's still peanuts compared to a real page fault. It shouldn't be
*that* noticeable, ie no way it's a 6% regression on its own.
I done some simple counting, and found that about 3.15e9 PTE are set to
old during the test after the commit. This may interpret the user_time
increase as below, because these accessed bit microfault is accounted as
user time.
387.66 . 0% +5.4% 408.49 . 0% unixbench.time.user_time
I also make a one line debug patch as below on top of the commit to set
the PTE to young unconditionally, which recover the regression.
modified mm/filemap.c
@@ -2193,7 +2193,7 @@ repeat:
if (file->f_ra.mmap_miss > 0)
file->f_ra.mmap_miss--;
addr = address + (page->index - vmf->pgoff) * PAGE_SIZE;
- do_set_pte(vma, addr, page, pte, false, false, true);
+ do_set_pte(vma, addr, page, pte, false, false, false);
unlock_page(page);
atomic64_inc(&old_pte_count);
goto next;
Best Regards,
Huang, Ying
11:38 p.m.
On Mon, Jun 13, 2016 at 05:02:15PM +0800, Huang, Ying wrote:
Linus Torvalds <torvalds(a)linux-foundation.org> writes:
> On Sat, Jun 11, 2016 at 5:49 PM, Huang, Ying <ying.huang(a)intel.com> wrote:
>>
>> From perf profile, the time spent in page_fault and its children
>> functions are almost same (7.85% vs 7.81%). So the time spent in page
>> fault and page table operation itself doesn't changed much. So, you
>> mean CPU may be slower to load the page table entry to TLB if accessed
>> bit is not set?
>
> So the CPU does take a microfault internally when it needs to set the
> accessed/dirty bit. It's not architecturally visible, but you can see
> it when you do timing loops.
>
> I've timed it at over a thousand cycles on at least some CPU's, but
> that's still peanuts compared to a real page fault. It shouldn't be
> *that* noticeable, ie no way it's a 6% regression on its own.
I done some simple counting, and found that about 3.15e9 PTE are set to
old during the test after the commit. This may interpret the user_time
increase as below, because these accessed bit microfault is accounted as
user time.
387.66 . 0% +5.4% 408.49 . 0% unixbench.time.user_time
I also make a one line debug patch as below on top of the commit to set
the PTE to young unconditionally, which recover the regression.
With this patch, meminfo.Active(file) is almost same unlike previous
experiment?
modified mm/filemap.c
@@ -2193,7 +2193,7 @@ repeat:
if (file->f_ra.mmap_miss > 0)
file->f_ra.mmap_miss--;
addr = address + (page->index - vmf->pgoff) * PAGE_SIZE;
- do_set_pte(vma, addr, page, pte, false, false, true);
+ do_set_pte(vma, addr, page, pte, false, false, false);
unlock_page(page);
atomic64_inc(&old_pte_count);
goto next;
Best Regards,
Huang, Ying
9:42 a.m.
Minchan Kim <minchan(a)kernel.org> writes:
On Mon, Jun 13, 2016 at 05:02:15PM +0800, Huang, Ying wrote:
> Linus Torvalds <torvalds(a)linux-foundation.org> writes:
>
> > On Sat, Jun 11, 2016 at 5:49 PM, Huang, Ying <ying.huang(a)intel.com>
wrote:
> >>
> >> From perf profile, the time spent in page_fault and its children
> >> functions are almost same (7.85% vs 7.81%). So the time spent in page
> >> fault and page table operation itself doesn't changed much. So, you
> >> mean CPU may be slower to load the page table entry to TLB if accessed
> >> bit is not set?
> >
> > So the CPU does take a microfault internally when it needs to set the
> > accessed/dirty bit. It's not architecturally visible, but you can see
> > it when you do timing loops.
> >
> > I've timed it at over a thousand cycles on at least some CPU's, but
> > that's still peanuts compared to a real page fault. It shouldn't be
> > *that* noticeable, ie no way it's a 6% regression on its own.
>
> I done some simple counting, and found that about 3.15e9 PTE are set to
> old during the test after the commit. This may interpret the user_time
> increase as below, because these accessed bit microfault is accounted as
> user time.
>
> 387.66 . 0% +5.4% 408.49 . 0% unixbench.time.user_time
>
> I also make a one line debug patch as below on top of the commit to set
> the PTE to young unconditionally, which recover the regression.
With this patch, meminfo.Active(file) is almost same unlike previous
experiment?
Yes. meminfo.Active(file) is almost same of that of the parent commit of
the first bad commit.
Best Regards,
Huang, Ying
>
> modified mm/filemap.c
> @@ -2193,7 +2193,7 @@ repeat:
> if (file->f_ra.mmap_miss > 0)
> file->f_ra.mmap_miss--;
> addr = address + (page->index - vmf->pgoff) * PAGE_SIZE;
> - do_set_pte(vma, addr, page, pte, false, false, true);
> + do_set_pte(vma, addr, page, pte, false, false, false);
> unlock_page(page);
> atomic64_inc(&old_pte_count);
> goto next;
>
> Best Regards,
> Huang, Ying
10:52 p.m.
On Sat, Jun 11, 2016 at 06:02:57PM -0700, Linus Torvalds wrote:
On Sat, Jun 11, 2016 at 5:49 PM, Huang, Ying
<ying.huang(a)intel.com> wrote:
>
> From perf profile, the time spent in page_fault and its children
> functions are almost same (7.85% vs 7.81%). So the time spent in page
> fault and page table operation itself doesn't changed much. So, you
> mean CPU may be slower to load the page table entry to TLB if accessed
> bit is not set?
So the CPU does take a microfault internally when it needs to set the
accessed/dirty bit. It's not architecturally visible, but you can see
it when you do timing loops.
I've timed it at over a thousand cycles on at least some CPU's, but
that's still peanuts compared to a real page fault. It shouldn't be
*that* noticeable, ie no way it's a 6% regression on its own.
Looks like setting accessed bit is the problem.
Withouth mkold:
Score: 1952.9
Performance counter stats for './Run shell8 -c 1' (3 runs):
468,562,316,621 cycles:u (
+- 0.02% )
4,596,299,472 dtlb_load_misses_walk_duration:u
( +- 0.07% )
5,245,488,559 itlb_misses_walk_duration:u (
+- 0.10% )
189.336404566 seconds time elapsed ( +-
0.01% )
With mkold:
Score: 1885.5
Performance counter stats for './Run shell8 -c 1' (3 runs):
503,185,676,256 cycles:u (
+- 0.06% )
8,137,007,894 dtlb_load_misses_walk_duration:u
( +- 0.85% )
7,220,632,283 itlb_misses_walk_duration:u (
+- 1.40% )
189.363223499 seconds time elapsed ( +-
0.01% )
We spend 36% more time in page walk only, about 1% of total userspace time.
Combining this with page walk footprint on caches, I guess we can get to
this 3.5% score difference I see.
I'm not sure if there's anything we can do to solve the issue without
screwing relacim logic again. :(
--
Kirill A. Shutemov
4:11 p.m.
On Mon, Jun 13, 2016 at 5:52 AM, Kirill A. Shutemov
<kirill.shutemov(a)linux.intel.com> wrote:
On Sat, Jun 11, 2016 at 06:02:57PM -0700, Linus Torvalds wrote:
>
> I've timed it at over a thousand cycles on at least some CPU's, but
> that's still peanuts compared to a real page fault. It shouldn't be
> *that* noticeable, ie no way it's a 6% regression on its own.
Looks like setting accessed bit is the problem.
Ok. I've definitely seen it as an issue, but never to the point of
several percent on a real benchmark that wasn't explicitly testing
that cost.
I reported the excessive dirty/accessed bit cost to Intel back in the
P4 days, but it's apparently not been high enough for anybody to care.
We spend 36% more time in page walk only, about 1% of total userspace
time.
Combining this with page walk footprint on caches, I guess we can get to
this 3.5% score difference I see.
I'm not sure if there's anything we can do to solve the issue without
screwing relacim logic again. :(
I think we should say "screw the reclaim logic" for now, and revert
commit 5c0a85fad949 for now.
Considering how much trouble the accessed bit is on some other
architectures too, I wonder if we should strive to simply not care
about it, and always leaving it set. And then rely entirely on just
unmapping the pages and making the "we took a page fault after
unmapping" be the real activity tester.
So get rid of the "if the page is young, mark it old but leave it in
the page tables" logic entirely. When we unmap a page, it will always
either be in the swap cache or the page cache anyway, so faulting it
in again should be just a minor fault with no actual IO happening.
That might be less of an impact in the end - yes, the unmap and
re-fault is much more expensive, but it presumably happens to much
fewer pages.
What do you think?
Linus
6:26 p.m.
On Mon, Jun 13, 2016 at 11:11:05PM -0700, Linus Torvalds wrote:
On Mon, Jun 13, 2016 at 5:52 AM, Kirill A. Shutemov
<kirill.shutemov(a)linux.intel.com> wrote:
> On Sat, Jun 11, 2016 at 06:02:57PM -0700, Linus Torvalds wrote:
>>
>> I've timed it at over a thousand cycles on at least some CPU's, but
>> that's still peanuts compared to a real page fault. It shouldn't be
>> *that* noticeable, ie no way it's a 6% regression on its own.
>
> Looks like setting accessed bit is the problem.
Ok. I've definitely seen it as an issue, but never to the point of
several percent on a real benchmark that wasn't explicitly testing
that cost.
I reported the excessive dirty/accessed bit cost to Intel back in the
P4 days, but it's apparently not been high enough for anybody to care.
> We spend 36% more time in page walk only, about 1% of total userspace time.
> Combining this with page walk footprint on caches, I guess we can get to
> this 3.5% score difference I see.
>
> I'm not sure if there's anything we can do to solve the issue without
> screwing relacim logic again. :(
I think we should say "screw the reclaim logic" for now, and revert
commit 5c0a85fad949 for now.
Okay. I'll prepare the patch.
Considering how much trouble the accessed bit is on some other
architectures too, I wonder if we should strive to simply not care
about it, and always leaving it set. And then rely entirely on just
unmapping the pages and making the "we took a page fault after
unmapping" be the real activity tester.
So get rid of the "if the page is young, mark it old but leave it in
the page tables" logic entirely. When we unmap a page, it will always
either be in the swap cache or the page cache anyway, so faulting it
in again should be just a minor fault with no actual IO happening.
That might be less of an impact in the end - yes, the unmap and
re-fault is much more expensive, but it presumably happens to much
fewer pages.
What do you think?
Well, we cannot do this for anonymous memory. No swap -- no swap cache, if
I read code correctly.
I guess it's doable for file mappings. Although I would expect regressions
in other benchmarks. IIUC, it would require page unmapping to propogate
page to active list, which is suboptimal.
And implications for page_idle is not clear to me.
Rik, Mel, any comments?
--
Kirill A. Shutemov
2:07 a.m.
On Tue, 2016-06-14 at 11:26 +0300, Kirill A. Shutemov wrote:
On Mon, Jun 13, 2016 at 11:11:05PM -0700, Linus Torvalds wrote:
>
> On Mon, Jun 13, 2016 at 5:52 AM, Kirill A. Shutemov
> <kirill.shutemov(a)linux.intel.com> wrote:
> >
> > On Sat, Jun 11, 2016 at 06:02:57PM -0700, Linus Torvalds wrote:
> > >
> > >
> > > I've timed it at over a thousand cycles on at least some CPU's,
> > > but
> > > that's still peanuts compared to a real page fault. It
> > > shouldn't be
> > > *that* noticeable, ie no way it's a 6% regression on its own.
> > Looks like setting accessed bit is the problem.
> Ok. I've definitely seen it as an issue, but never to the point of
> several percent on a real benchmark that wasn't explicitly testing
> that cost.
>
> I reported the excessive dirty/accessed bit cost to Intel back in
> the
> P4 days, but it's apparently not been high enough for anybody to
> care.
>
> >
> > We spend 36% more time in page walk only, about 1% of total
> > userspace time.
> > Combining this with page walk footprint on caches, I guess we can
> > get to
> > this 3.5% score difference I see.
> >
> > I'm not sure if there's anything we can do to solve the issue
> > without
> > screwing relacim logic again. :(
> I think we should say "screw the reclaim logic" for now, and revert
> commit 5c0a85fad949 for now.
Okay. I'll prepare the patch.
>
> Considering how much trouble the accessed bit is on some other
> architectures too, I wonder if we should strive to simply not care
> about it, and always leaving it set. And then rely entirely on just
> unmapping the pages and making the "we took a page fault after
> unmapping" be the real activity tester.
>
> So get rid of the "if the page is young, mark it old but leave it
> in
> the page tables" logic entirely. When we unmap a page, it will
> always
> either be in the swap cache or the page cache anyway, so faulting
> it
> in again should be just a minor fault with no actual IO happening.
>
> That might be less of an impact in the end - yes, the unmap and
> re-fault is much more expensive, but it presumably happens to much
> fewer pages.
>
> What do you think?
Well, we cannot do this for anonymous memory. No swap -- no swap
cache, if
I read code correctly.
I guess it's doable for file mappings. Although I would expect
regressions
in other benchmarks. IIUC, it would require page unmapping to
propogate
page to active list, which is suboptimal.
And implications for page_idle is not clear to me.
Rik, Mel, any comments?
We can clear the accessed/young bit when anon pages are moved
from the active to the inactive list.
Reclaim does not care about the young bit on active anon pages
at all. For anon pages it uses a two hand clock algorithm, with
only pages on the inactive list being cared about.
For file pages, I believe we do look at the young bit on mapped
pages when they reach the end of the inactive list. Again, we
only care about the young bit on inactive pages.
One option may be to count on actively used file pages actually
being on the active list, and always set the young bit on ptes
when the page is already active.
Then we can let reclaim do its thing with the smaller number of
pages that are on the inactive list, while doing the faster thing
for pages that are on the active list.
Does that make sense?
--
All Rights Reversed.
12:03 a.m.
On 06/14/2016 08:11 AM, Linus Torvalds wrote:
On Mon, Jun 13, 2016 at 5:52 AM, Kirill A. Shutemov
<kirill.shutemov(a)linux.intel.com> wrote:
> On Sat, Jun 11, 2016 at 06:02:57PM -0700, Linus Torvalds wrote:
>>
>> I've timed it at over a thousand cycles on at least some CPU's, but
>> that's still peanuts compared to a real page fault. It shouldn't be
>> *that* noticeable, ie no way it's a 6% regression on its own.
>
> Looks like setting accessed bit is the problem.
Ok. I've definitely seen it as an issue, but never to the point of
several percent on a real benchmark that wasn't explicitly testing
that cost.
I reported the excessive dirty/accessed bit cost to Intel back in the
P4 days, but it's apparently not been high enough for anybody to care.
> We spend 36% more time in page walk only, about 1% of total userspace time.
> Combining this with page walk footprint on caches, I guess we can get to
> this 3.5% score difference I see.
>
> I'm not sure if there's anything we can do to solve the issue without
> screwing relacim logic again. :(
I think we should say "screw the reclaim logic" for now, and revert
commit 5c0a85fad949 for now.
Considering how much trouble the accessed bit is on some other
architectures too, I wonder if we should strive to simply not care
about it, and always leaving it set. And then rely entirely on just
unmapping the pages and making the "we took a page fault after
unmapping" be the real activity tester.
So get rid of the "if the page is young, mark it old but leave it in
the page tables" logic entirely. When we unmap a page, it will always
either be in the swap cache or the page cache anyway, so faulting it
in again should be just a minor fault with no actual IO happening.
That might be less of an impact in the end - yes, the unmap and
re-fault is much more expensive, but it presumably happens to much
fewer pages.
FWIW, something like that is what Martin did for s390 3 years ago.
We now use invalidation and page faults to implement the *young
functions in pgtable.h (basically using a SW young bit). This
helped us to get rid of the storage keys (which contain the HW
reference bit). The performance did not seem to suffer.
See commit 0944fe3f4a323f436180d39402cae7f9c46ead17
s390/mm: implement software referenced bits
What do you think?
Your proposal would be to do the software tracking via
invalidation/fault part of the generic mm code and not to hide it
in the architecture backend. Correct?
Linus
6:57 p.m.
On Wed, Jun 08, 2016 at 11:58:11AM +0300, Kirill A. Shutemov wrote:
On Wed, Jun 08, 2016 at 04:41:37PM +0800, Huang, Ying wrote:
> "Huang, Ying" <ying.huang(a)intel.com> writes:
>
> > "Kirill A. Shutemov" <kirill.shutemov(a)linux.intel.com> writes:
> >
> >> On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel test robot wrote:
> >>>
> >>> FYI, we noticed a -6.3% regression of unixbench.score due to commit:
> >>>
> >>> commit 5c0a85fad949212b3e059692deecdeed74ae7ec7 ("mm: make
faultaround produce old ptes")
> >>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
master
> >>>
> >>> in testcase: unixbench
> >>> on test machine: lituya: 16 threads Haswell High-end Desktop (i7-5960X
3.0G) with 16G memory
> >>> with following parameters:
cpufreq_governor=performance/nr_task=1/test=shell8
> >>>
> >>>
> >>> Details are as below:
> >>>
-------------------------------------------------------------------------------------------------->
> >>>
> >>>
> >>>
=========================================================================================
> >>>
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
> >>>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
> >>>
> >>> commit:
> >>> 4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
> >>> 5c0a85fad949212b3e059692deecdeed74ae7ec7
> >>>
> >>> 4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
> >>> ---------------- --------------------------
> >>> fail:runs %reproduction fail:runs
> >>> | | |
> >>> 3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
> >>> %stddev %change %stddev
> >>> \ | \
> >>> 14321 . 0% -6.3% 13425 . 0% unixbench.score
> >>> 1996897 . 0% -6.1% 1874635 . 0%
unixbench.time.involuntary_context_switches
> >>> 1.721e+08 . 0% -6.2% 1.613e+08 . 0%
unixbench.time.minor_page_faults
> >>> 758.65 . 0% -3.0% 735.86 . 0%
unixbench.time.system_time
> >>> 387.66 . 0% +5.4% 408.49 . 0% unixbench.time.user_time
> >>> 5950278 . 0% -6.2% 5583456 . 0%
unixbench.time.voluntary_context_switches
> >>
> >> That's weird.
> >>
> >> I don't understand why the change would reduce number or minor faults.
> >> It should stay the same on x86-64. Rise of user_time is puzzling too.
> >
> > unixbench runs in fixed time mode. That is, the total time to run
> > unixbench is fixed, but the work done varies. So the minor_page_faults
> > change may reflect only the work done.
> >
> >> Hm. Is reproducible? Across reboot?
> >
>
> And FYI, there is no swap setup for test, all root file system including
> benchmark files are in tmpfs, so no real page reclaim will be
> triggered. But it appears that active file cache reduced after the
> commit.
>
> 111331 ± 1% -13.3% 96503 ± 0% meminfo.Active
> 27603 ± 1% -43.9% 15486 ± 0% meminfo.Active(file)
>
> I think this is the expected behavior of the commit?
Yes, it's expected.
After the change faularound would produce old pte. It means there's more
chance for these pages to be on inactive lru, unless somebody actually
touch them and flip accessed bit.
Hmm, tmpfs pages should be in anonymous LRU list and VM shouldn't scan
anonymous LRU list on swapless system so I really wonder why active file
LRU is shrunk.
12:34 a.m.
On Tue, Jun 14, 2016 at 05:57:28PM +0900, Minchan Kim wrote:
On Wed, Jun 08, 2016 at 11:58:11AM +0300, Kirill A. Shutemov wrote:
> On Wed, Jun 08, 2016 at 04:41:37PM +0800, Huang, Ying wrote:
> > "Huang, Ying" <ying.huang(a)intel.com> writes:
> >
> > > "Kirill A. Shutemov" <kirill.shutemov(a)linux.intel.com>
writes:
> > >
> > >> On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel test robot wrote:
> > >>>
> > >>> FYI, we noticed a -6.3% regression of unixbench.score due to
commit:
> > >>>
> > >>> commit 5c0a85fad949212b3e059692deecdeed74ae7ec7 ("mm: make
faultaround produce old ptes")
> > >>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
master
> > >>>
> > >>> in testcase: unixbench
> > >>> on test machine: lituya: 16 threads Haswell High-end Desktop
(i7-5960X 3.0G) with 16G memory
> > >>> with following parameters:
cpufreq_governor=performance/nr_task=1/test=shell8
> > >>>
> > >>>
> > >>> Details are as below:
> > >>>
-------------------------------------------------------------------------------------------------->
> > >>>
> > >>>
> > >>>
=========================================================================================
> > >>>
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
> > >>>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
> > >>>
> > >>> commit:
> > >>> 4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
> > >>> 5c0a85fad949212b3e059692deecdeed74ae7ec7
> > >>>
> > >>> 4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
> > >>> ---------------- --------------------------
> > >>> fail:runs %reproduction fail:runs
> > >>> | | |
> > >>> 3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
> > >>> %stddev %change %stddev
> > >>> \ | \
> > >>> 14321 . 0% -6.3% 13425 . 0% unixbench.score
> > >>> 1996897 . 0% -6.1% 1874635 . 0%
unixbench.time.involuntary_context_switches
> > >>> 1.721e+08 . 0% -6.2% 1.613e+08 . 0%
unixbench.time.minor_page_faults
> > >>> 758.65 . 0% -3.0% 735.86 . 0%
unixbench.time.system_time
> > >>> 387.66 . 0% +5.4% 408.49 . 0%
unixbench.time.user_time
> > >>> 5950278 . 0% -6.2% 5583456 . 0%
unixbench.time.voluntary_context_switches
> > >>
> > >> That's weird.
> > >>
> > >> I don't understand why the change would reduce number or minor
faults.
> > >> It should stay the same on x86-64. Rise of user_time is puzzling too.
> > >
> > > unixbench runs in fixed time mode. That is, the total time to run
> > > unixbench is fixed, but the work done varies. So the minor_page_faults
> > > change may reflect only the work done.
> > >
> > >> Hm. Is reproducible? Across reboot?
> > >
> >
> > And FYI, there is no swap setup for test, all root file system including
> > benchmark files are in tmpfs, so no real page reclaim will be
> > triggered. But it appears that active file cache reduced after the
> > commit.
> >
> > 111331 ± 1% -13.3% 96503 ± 0% meminfo.Active
> > 27603 ± 1% -43.9% 15486 ± 0% meminfo.Active(file)
> >
> > I think this is the expected behavior of the commit?
>
> Yes, it's expected.
>
> After the change faularound would produce old pte. It means there's more
> chance for these pages to be on inactive lru, unless somebody actually
> touch them and flip accessed bit.
Hmm, tmpfs pages should be in anonymous LRU list and VM shouldn't scan
anonymous LRU list on swapless system so I really wonder why active file
LRU is shrunk.
Hm. Good point. I don't why we have anything on file lru if there's no
filesystems except tmpfs.
Ying, how do you get stuff to the tmpfs?
--
Kirill A. Shutemov
9:52 a.m.
"Kirill A. Shutemov" <kirill(a)shutemov.name> writes:
On Tue, Jun 14, 2016 at 05:57:28PM +0900, Minchan Kim wrote:
> On Wed, Jun 08, 2016 at 11:58:11AM +0300, Kirill A. Shutemov wrote:
> > On Wed, Jun 08, 2016 at 04:41:37PM +0800, Huang, Ying wrote:
> > > "Huang, Ying" <ying.huang(a)intel.com> writes:
> > >
> > > > "Kirill A. Shutemov" <kirill.shutemov(a)linux.intel.com>
writes:
> > > >
> > > >> On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel test robot
wrote:
> > > >>>
> > > >>> FYI, we noticed a -6.3% regression of unixbench.score due to
commit:
> > > >>>
> > > >>> commit 5c0a85fad949212b3e059692deecdeed74ae7ec7 ("mm:
make faultaround produce old ptes")
> > > >>>
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
> > > >>>
> > > >>> in testcase: unixbench
> > > >>> on test machine: lituya: 16 threads Haswell High-end Desktop
(i7-5960X 3.0G) with 16G memory
> > > >>> with following parameters:
cpufreq_governor=performance/nr_task=1/test=shell8
> > > >>>
> > > >>>
> > > >>> Details are as below:
> > > >>>
-------------------------------------------------------------------------------------------------->
> > > >>>
> > > >>>
> > > >>>
=========================================================================================
> > > >>>
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
> > > >>>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
> > > >>>
> > > >>> commit:
> > > >>> 4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
> > > >>> 5c0a85fad949212b3e059692deecdeed74ae7ec7
> > > >>>
> > > >>> 4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
> > > >>> ---------------- --------------------------
> > > >>> fail:runs %reproduction fail:runs
> > > >>> | | |
> > > >>> 3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
> > > >>> %stddev %change %stddev
> > > >>> \ | \
> > > >>> 14321 . 0% -6.3% 13425 . 0% unixbench.score
> > > >>> 1996897 . 0% -6.1% 1874635 . 0%
unixbench.time.involuntary_context_switches
> > > >>> 1.721e+08 . 0% -6.2% 1.613e+08 . 0%
unixbench.time.minor_page_faults
> > > >>> 758.65 . 0% -3.0% 735.86 . 0%
unixbench.time.system_time
> > > >>> 387.66 . 0% +5.4% 408.49 . 0%
unixbench.time.user_time
> > > >>> 5950278 . 0% -6.2% 5583456 . 0%
unixbench.time.voluntary_context_switches
> > > >>
> > > >> That's weird.
> > > >>
> > > >> I don't understand why the change would reduce number or minor
faults.
> > > >> It should stay the same on x86-64. Rise of user_time is puzzling
too.
> > > >
> > > > unixbench runs in fixed time mode. That is, the total time to run
> > > > unixbench is fixed, but the work done varies. So the
minor_page_faults
> > > > change may reflect only the work done.
> > > >
> > > >> Hm. Is reproducible? Across reboot?
> > > >
> > >
> > > And FYI, there is no swap setup for test, all root file system including
> > > benchmark files are in tmpfs, so no real page reclaim will be
> > > triggered. But it appears that active file cache reduced after the
> > > commit.
> > >
> > > 111331 . 1% -13.3% 96503 . 0% meminfo.Active
> > > 27603 . 1% -43.9% 15486 . 0% meminfo.Active(file)
> > >
> > > I think this is the expected behavior of the commit?
> >
> > Yes, it's expected.
> >
> > After the change faularound would produce old pte. It means there's more
> > chance for these pages to be on inactive lru, unless somebody actually
> > touch them and flip accessed bit.
>
> Hmm, tmpfs pages should be in anonymous LRU list and VM shouldn't scan
> anonymous LRU list on swapless system so I really wonder why active file
> LRU is shrunk.
Hm. Good point. I don't why we have anything on file lru if there's no
filesystems except tmpfs.
Ying, how do you get stuff to the tmpfs?
We put root file system and benchmark into a set of compressed cpio
archive, then concatenate them into one initrd, and finally kernel use
that initrd as initramfs.
Best Regards,
Huang, Ying
10:13 a.m.
On Thu, Jun 16, 2016 at 07:52:26AM +0800, Huang, Ying wrote:
"Kirill A. Shutemov" <kirill(a)shutemov.name> writes:
> On Tue, Jun 14, 2016 at 05:57:28PM +0900, Minchan Kim wrote:
>> On Wed, Jun 08, 2016 at 11:58:11AM +0300, Kirill A. Shutemov wrote:
>> > On Wed, Jun 08, 2016 at 04:41:37PM +0800, Huang, Ying wrote:
>> > > "Huang, Ying" <ying.huang(a)intel.com> writes:
>> > >
>> > > > "Kirill A. Shutemov"
<kirill.shutemov(a)linux.intel.com> writes:
>> > > >
>> > > >> On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel test robot
wrote:
>> > > >>>
>> > > >>> FYI, we noticed a -6.3% regression of unixbench.score due
to commit:
>> > > >>>
>> > > >>> commit 5c0a85fad949212b3e059692deecdeed74ae7ec7
("mm: make faultaround produce old ptes")
>> > > >>>
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
>> > > >>>
>> > > >>> in testcase: unixbench
>> > > >>> on test machine: lituya: 16 threads Haswell High-end
Desktop (i7-5960X 3.0G) with 16G memory
>> > > >>> with following parameters:
cpufreq_governor=performance/nr_task=1/test=shell8
>> > > >>>
>> > > >>>
>> > > >>> Details are as below:
>> > > >>>
-------------------------------------------------------------------------------------------------->
>> > > >>>
>> > > >>>
>> > > >>>
=========================================================================================
>> > > >>>
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
>> > > >>>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
>> > > >>>
>> > > >>> commit:
>> > > >>> 4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
>> > > >>> 5c0a85fad949212b3e059692deecdeed74ae7ec7
>> > > >>>
>> > > >>> 4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
>> > > >>> ---------------- --------------------------
>> > > >>> fail:runs %reproduction fail:runs
>> > > >>> | | |
>> > > >>> 3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
>> > > >>> %stddev %change %stddev
>> > > >>> \ | \
>> > > >>> 14321 . 0% -6.3% 13425 . 0%
unixbench.score
>> > > >>> 1996897 . 0% -6.1% 1874635 . 0%
unixbench.time.involuntary_context_switches
>> > > >>> 1.721e+08 . 0% -6.2% 1.613e+08 . 0%
unixbench.time.minor_page_faults
>> > > >>> 758.65 . 0% -3.0% 735.86 . 0%
unixbench.time.system_time
>> > > >>> 387.66 . 0% +5.4% 408.49 . 0%
unixbench.time.user_time
>> > > >>> 5950278 . 0% -6.2% 5583456 . 0%
unixbench.time.voluntary_context_switches
>> > > >>
>> > > >> That's weird.
>> > > >>
>> > > >> I don't understand why the change would reduce number or
minor faults.
>> > > >> It should stay the same on x86-64. Rise of user_time is
puzzling too.
>> > > >
>> > > > unixbench runs in fixed time mode. That is, the total time to
run
>> > > > unixbench is fixed, but the work done varies. So the
minor_page_faults
>> > > > change may reflect only the work done.
>> > > >
>> > > >> Hm. Is reproducible? Across reboot?
>> > > >
>> > >
>> > > And FYI, there is no swap setup for test, all root file system
including
>> > > benchmark files are in tmpfs, so no real page reclaim will be
>> > > triggered. But it appears that active file cache reduced after the
>> > > commit.
>> > >
>> > > 111331 . 1% -13.3% 96503 . 0% meminfo.Active
>> > > 27603 . 1% -43.9% 15486 . 0% meminfo.Active(file)
>> > >
>> > > I think this is the expected behavior of the commit?
>> >
>> > Yes, it's expected.
>> >
>> > After the change faularound would produce old pte. It means there's
more
>> > chance for these pages to be on inactive lru, unless somebody actually
>> > touch them and flip accessed bit.
>>
>> Hmm, tmpfs pages should be in anonymous LRU list and VM shouldn't scan
>> anonymous LRU list on swapless system so I really wonder why active file
>> LRU is shrunk.
>
> Hm. Good point. I don't why we have anything on file lru if there's no
> filesystems except tmpfs.
>
> Ying, how do you get stuff to the tmpfs?
We put root file system and benchmark into a set of compressed cpio
archive, then concatenate them into one initrd, and finally kernel use
that initrd as initramfs.
I see.
Could you share your 4 full vmstat(/proc/vmstat) files?
old:
cat /proc/vmstat > before.old.vmstat
do benchmark
cat /proc/vmstat > after.old.vmstat
new:
cat /proc/vmstat > before.new.vmstat
do benchmark
cat /proc/vmstat > after.new.vmstat
IOW, I want to see stats related to reclaim.
Thanks.
8:27 a.m.
Minchan Kim <minchan(a)kernel.org> writes:
On Thu, Jun 16, 2016 at 07:52:26AM +0800, Huang, Ying wrote:
> "Kirill A. Shutemov" <kirill(a)shutemov.name> writes:
>
> > On Tue, Jun 14, 2016 at 05:57:28PM +0900, Minchan Kim wrote:
> >> On Wed, Jun 08, 2016 at 11:58:11AM +0300, Kirill A. Shutemov wrote:
> >> > On Wed, Jun 08, 2016 at 04:41:37PM +0800, Huang, Ying wrote:
> >> > > "Huang, Ying" <ying.huang(a)intel.com> writes:
> >> > >
> >> > > > "Kirill A. Shutemov"
<kirill.shutemov(a)linux.intel.com> writes:
> >> > > >
> >> > > >> On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel test
robot wrote:
> >> > > >>>
> >> > > >>> FYI, we noticed a -6.3% regression of unixbench.score
due to commit:
> >> > > >>>
> >> > > >>> commit 5c0a85fad949212b3e059692deecdeed74ae7ec7
("mm: make faultaround produce old ptes")
> >> > > >>>
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
> >> > > >>>
> >> > > >>> in testcase: unixbench
> >> > > >>> on test machine: lituya: 16 threads Haswell High-end
Desktop (i7-5960X 3.0G) with 16G memory
> >> > > >>> with following parameters:
cpufreq_governor=performance/nr_task=1/test=shell8
> >> > > >>>
> >> > > >>>
> >> > > >>> Details are as below:
> >> > > >>>
-------------------------------------------------------------------------------------------------->
> >> > > >>>
> >> > > >>>
> >> > > >>>
=========================================================================================
> >> > > >>>
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
> >> > > >>>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
> >> > > >>>
> >> > > >>> commit:
> >> > > >>> 4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
> >> > > >>> 5c0a85fad949212b3e059692deecdeed74ae7ec7
> >> > > >>>
> >> > > >>> 4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
> >> > > >>> ---------------- --------------------------
> >> > > >>> fail:runs %reproduction fail:runs
> >> > > >>> | | |
> >> > > >>> 3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
> >> > > >>> %stddev %change %stddev
> >> > > >>> \ | \
> >> > > >>> 14321 . 0% -6.3% 13425 . 0%
unixbench.score
> >> > > >>> 1996897 . 0% -6.1% 1874635 . 0%
unixbench.time.involuntary_context_switches
> >> > > >>> 1.721e+08 . 0% -6.2% 1.613e+08 . 0%
unixbench.time.minor_page_faults
> >> > > >>> 758.65 . 0% -3.0% 735.86 . 0%
unixbench.time.system_time
> >> > > >>> 387.66 . 0% +5.4% 408.49 . 0%
unixbench.time.user_time
> >> > > >>> 5950278 . 0% -6.2% 5583456 . 0%
unixbench.time.voluntary_context_switches
> >> > > >>
> >> > > >> That's weird.
> >> > > >>
> >> > > >> I don't understand why the change would reduce number
or minor faults.
> >> > > >> It should stay the same on x86-64. Rise of user_time is
puzzling too.
> >> > > >
> >> > > > unixbench runs in fixed time mode. That is, the total time
to run
> >> > > > unixbench is fixed, but the work done varies. So the
minor_page_faults
> >> > > > change may reflect only the work done.
> >> > > >
> >> > > >> Hm. Is reproducible? Across reboot?
> >> > > >
> >> > >
> >> > > And FYI, there is no swap setup for test, all root file system
including
> >> > > benchmark files are in tmpfs, so no real page reclaim will be
> >> > > triggered. But it appears that active file cache reduced after
the
> >> > > commit.
> >> > >
> >> > > 111331 . 1% -13.3% 96503 . 0% meminfo.Active
> >> > > 27603 . 1% -43.9% 15486 . 0%
meminfo.Active(file)
> >> > >
> >> > > I think this is the expected behavior of the commit?
> >> >
> >> > Yes, it's expected.
> >> >
> >> > After the change faularound would produce old pte. It means there's
more
> >> > chance for these pages to be on inactive lru, unless somebody actually
> >> > touch them and flip accessed bit.
> >>
> >> Hmm, tmpfs pages should be in anonymous LRU list and VM shouldn't scan
> >> anonymous LRU list on swapless system so I really wonder why active file
> >> LRU is shrunk.
> >
> > Hm. Good point. I don't why we have anything on file lru if there's no
> > filesystems except tmpfs.
> >
> > Ying, how do you get stuff to the tmpfs?
>
> We put root file system and benchmark into a set of compressed cpio
> archive, then concatenate them into one initrd, and finally kernel use
> that initrd as initramfs.
I see.
Could you share your 4 full vmstat(/proc/vmstat) files?
old:
cat /proc/vmstat > before.old.vmstat
do benchmark
cat /proc/vmstat > after.old.vmstat
new:
cat /proc/vmstat > before.new.vmstat
do benchmark
cat /proc/vmstat > after.new.vmstat
IOW, I want to see stats related to reclaim.
Hi,
The /proc/vmstat for the parent commit (parent-proc-vmstat.gz) and first
bad commit (fbc-proc-vmstat.gz) are attached with the email.
The contents of the file is more than the vmstat before and after
benchmark running, but are sampled every 1 seconds. Every sample begin
with "time: <time>". You can check the first and last samples. The
first /proc/vmstat capturing is started at the same time of the
benchmark, so it is not exactly the vmstat before the benchmark running.
Best Regards,
Huang, Ying
3:41 p.m.
On Thu, Jun 16, 2016 at 03:27:44PM -0700, Huang, Ying wrote:
Minchan Kim <minchan(a)kernel.org> writes:
> On Thu, Jun 16, 2016 at 07:52:26AM +0800, Huang, Ying wrote:
>> "Kirill A. Shutemov" <kirill(a)shutemov.name> writes:
>>
>> > On Tue, Jun 14, 2016 at 05:57:28PM +0900, Minchan Kim wrote:
>> >> On Wed, Jun 08, 2016 at 11:58:11AM +0300, Kirill A. Shutemov wrote:
>> >> > On Wed, Jun 08, 2016 at 04:41:37PM +0800, Huang, Ying wrote:
>> >> > > "Huang, Ying" <ying.huang(a)intel.com> writes:
>> >> > >
>> >> > > > "Kirill A. Shutemov"
<kirill.shutemov(a)linux.intel.com> writes:
>> >> > > >
>> >> > > >> On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel
test robot wrote:
>> >> > > >>>
>> >> > > >>> FYI, we noticed a -6.3% regression of
unixbench.score due to commit:
>> >> > > >>>
>> >> > > >>> commit 5c0a85fad949212b3e059692deecdeed74ae7ec7
("mm: make faultaround produce old ptes")
>> >> > > >>>
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
>> >> > > >>>
>> >> > > >>> in testcase: unixbench
>> >> > > >>> on test machine: lituya: 16 threads Haswell
High-end Desktop (i7-5960X 3.0G) with 16G memory
>> >> > > >>> with following parameters:
cpufreq_governor=performance/nr_task=1/test=shell8
>> >> > > >>>
>> >> > > >>>
>> >> > > >>> Details are as below:
>> >> > > >>>
-------------------------------------------------------------------------------------------------->
>> >> > > >>>
>> >> > > >>>
>> >> > > >>>
=========================================================================================
>> >> > > >>>
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
>> >> > > >>>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
>> >> > > >>>
>> >> > > >>> commit:
>> >> > > >>> 4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
>> >> > > >>> 5c0a85fad949212b3e059692deecdeed74ae7ec7
>> >> > > >>>
>> >> > > >>> 4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
>> >> > > >>> ---------------- --------------------------
>> >> > > >>> fail:runs %reproduction fail:runs
>> >> > > >>> | | |
>> >> > > >>> 3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
>> >> > > >>> %stddev %change %stddev
>> >> > > >>> \ | \
>> >> > > >>> 14321 . 0% -6.3% 13425 . 0%
unixbench.score
>> >> > > >>> 1996897 . 0% -6.1% 1874635 . 0%
unixbench.time.involuntary_context_switches
>> >> > > >>> 1.721e+08 . 0% -6.2% 1.613e+08 . 0%
unixbench.time.minor_page_faults
>> >> > > >>> 758.65 . 0% -3.0% 735.86 . 0%
unixbench.time.system_time
>> >> > > >>> 387.66 . 0% +5.4% 408.49 . 0%
unixbench.time.user_time
>> >> > > >>> 5950278 . 0% -6.2% 5583456 . 0%
unixbench.time.voluntary_context_switches
>> >> > > >>
>> >> > > >> That's weird.
>> >> > > >>
>> >> > > >> I don't understand why the change would reduce
number or minor faults.
>> >> > > >> It should stay the same on x86-64. Rise of user_time
is puzzling too.
>> >> > > >
>> >> > > > unixbench runs in fixed time mode. That is, the total
time to run
>> >> > > > unixbench is fixed, but the work done varies. So the
minor_page_faults
>> >> > > > change may reflect only the work done.
>> >> > > >
>> >> > > >> Hm. Is reproducible? Across reboot?
>> >> > > >
>> >> > >
>> >> > > And FYI, there is no swap setup for test, all root file
system including
>> >> > > benchmark files are in tmpfs, so no real page reclaim will
be
>> >> > > triggered. But it appears that active file cache reduced
after the
>> >> > > commit.
>> >> > >
>> >> > > 111331 . 1% -13.3% 96503 . 0% meminfo.Active
>> >> > > 27603 . 1% -43.9% 15486 . 0%
meminfo.Active(file)
>> >> > >
>> >> > > I think this is the expected behavior of the commit?
>> >> >
>> >> > Yes, it's expected.
>> >> >
>> >> > After the change faularound would produce old pte. It means
there's more
>> >> > chance for these pages to be on inactive lru, unless somebody
actually
>> >> > touch them and flip accessed bit.
>> >>
>> >> Hmm, tmpfs pages should be in anonymous LRU list and VM shouldn't
scan
>> >> anonymous LRU list on swapless system so I really wonder why active
file
>> >> LRU is shrunk.
>> >
>> > Hm. Good point. I don't why we have anything on file lru if there's
no
>> > filesystems except tmpfs.
>> >
>> > Ying, how do you get stuff to the tmpfs?
>>
>> We put root file system and benchmark into a set of compressed cpio
>> archive, then concatenate them into one initrd, and finally kernel use
>> that initrd as initramfs.
>
> I see.
>
> Could you share your 4 full vmstat(/proc/vmstat) files?
>
> old:
>
> cat /proc/vmstat > before.old.vmstat
> do benchmark
> cat /proc/vmstat > after.old.vmstat
>
> new:
>
> cat /proc/vmstat > before.new.vmstat
> do benchmark
> cat /proc/vmstat > after.new.vmstat
>
> IOW, I want to see stats related to reclaim.
Hi,
The /proc/vmstat for the parent commit (parent-proc-vmstat.gz) and first
bad commit (fbc-proc-vmstat.gz) are attached with the email.
The contents of the file is more than the vmstat before and after
benchmark running, but are sampled every 1 seconds. Every sample begin
with "time: <time>". You can check the first and last samples. The
first /proc/vmstat capturing is started at the same time of the
benchmark, so it is not exactly the vmstat before the benchmark running.
Thanks for the testing!
nr_active_file was shrunk 48% but the vaule itself is not huge so
I don't think it affects performance a lot.
There was no reclaim activity for testing. :(
pgfault, 6% reduced. Given that, pgalloc/free reduced 6%, too
because unixbench was time fixed mode and 6% regressed so no
doubt.
No interesting data.
It seems you tested it with THP, maybe always mode?
I'm so sorry but could you test it with disabling CONFIG_TRANSPARENT_HUGEPAGE=n
again? it might you already did.
Is it still 6% regressed with disabling THP?
nr_free_pages -6663 -6461 96.97%
nr_alloc_batch 2594 4013 154.70%
nr_inactive_anon 112 112 100.00%
nr_active_anon 2536 2159 85.13%
nr_inactive_file -567 -227 40.04%
nr_active_file 648 315 48.61%
nr_unevictable 0 0 0.00%
nr_mlock 0 0 0.00%
nr_anon_pages 2634 2161 82.04%
nr_mapped 511 530 103.72%
nr_file_pages 207 215 103.86%
nr_dirty -7 -6 85.71%
nr_writeback 0 0 0.00%
nr_slab_reclaimable 158 328 207.59%
nr_slab_unreclaimable 2208 2115 95.79%
nr_page_table_pages 268 247 92.16%
nr_kernel_stack 143 80 55.94%
nr_unstable 1 1 100.00%
nr_bounce 0 0 0.00%
nr_vmscan_write 0 0 0.00%
nr_vmscan_immediate_reclaim 0 0 0.00%
nr_writeback_temp 0 0 0.00%
nr_isolated_anon 0 0 0.00%
nr_isolated_file 0 0 0.00%
nr_shmem 131 131 100.00%
nr_dirtied 67 78 116.42%
nr_written 74 84 113.51%
nr_pages_scanned 0 0 0.00%
numa_hit 483752446 453696304 93.79%
numa_miss 0 0 0.00%
numa_foreign 0 0 0.00%
numa_interleave 0 0 0.00%
numa_local 483752445 453696304 93.79%
numa_other 1 0 0.00%
workingset_refault 0 0 0.00%
workingset_activate 0 0 0.00%
workingset_nodereclaim 0 0 0.00%
nr_anon_transparent_hugepages 1 0 0.00%
nr_free_cma 0 0 0.00%
nr_dirty_threshold -1316 -1274 96.81%
nr_dirty_background_threshold -658 -637 96.81%
pgpgin 0 0 0.00%
pgpgout 0 0 0.00%
pswpin 0 0 0.00%
pswpout 0 0 0.00%
pgalloc_dma 0 0 0.00%
pgalloc_dma32 60130977 56323630 93.67%
pgalloc_normal 457203182 428863437 93.80%
pgalloc_movable 0 0 0.00%
pgfree 517327743 485181251 93.79%
pgactivate 2059556 1930950 93.76%
pgdeactivate 0 0 0.00%
pgfault 572723351 537107146 93.78%
pgmajfault 0 0 0.00%
pglazyfreed 0 0 0.00%
pgrefill_dma 0 0 0.00%
pgrefill_dma32 0 0 0.00%
pgrefill_normal 0 0 0.00%
pgrefill_movable 0 0 0.00%
pgsteal_kswapd_dma 0 0 0.00%
pgsteal_kswapd_dma32 0 0 0.00%
pgsteal_kswapd_normal 0 0 0.00%
pgsteal_kswapd_movable 0 0 0.00%
pgsteal_direct_dma 0 0 0.00%
pgsteal_direct_dma32 0 0 0.00%
pgsteal_direct_normal 0 0 0.00%
pgsteal_direct_movable 0 0 0.00%
pgscan_kswapd_dma 0 0 0.00%
pgscan_kswapd_dma32 0 0 0.00%
pgscan_kswapd_normal 0 0 0.00%
pgscan_kswapd_movable 0 0 0.00%
pgscan_direct_dma 0 0 0.00%
pgscan_direct_dma32 0 0 0.00%
pgscan_direct_normal 0 0 0.00%
pgscan_direct_movable 0 0 0.00%
pgscan_direct_throttle 0 0 0.00%
zone_reclaim_failed 0 0 0.00%
pginodesteal 0 0 0.00%
slabs_scanned 0 0 0.00%
kswapd_inodesteal 0 0 0.00%
kswapd_low_wmark_hit_quickly 0 0 0.00%
kswapd_high_wmark_hit_quickly 0 0 0.00%
pageoutrun 0 0 0.00%
allocstall 0 0 0.00%
pgrotated 0 0 0.00%
drop_pagecache 0 0 0.00%
drop_slab 0 0 0.00%
numa_pte_updates 0 0 0.00%
numa_huge_pte_updates 0 0 0.00%
numa_hint_faults 0 0 0.00%
numa_hint_faults_local 0 0 0.00%
numa_pages_migrated 0 0 0.00%
pgmigrate_success 0 0 0.00%
pgmigrate_fail 0 0 0.00%
compact_migrate_scanned 0 0 0.00%
compact_free_scanned 0 0 0.00%
compact_isolated 0 0 0.00%
compact_stall 0 0 0.00%
compact_fail 0 0 0.00%
compact_success 0 0 0.00%
compact_daemon_wake 0 0 0.00%
htlb_buddy_alloc_success 0 0 0.00%
htlb_buddy_alloc_fail 0 0 0.00%
unevictable_pgs_culled 0 0 0.00%
unevictable_pgs_scanned 0 0 0.00%
unevictable_pgs_rescued 0 0 0.00%
unevictable_pgs_mlocked 0 0 0.00%
unevictable_pgs_munlocked 0 0 0.00%
unevictable_pgs_cleared 0 0 0.00%
unevictable_pgs_stranded 0 0 0.00%
thp_fault_alloc 22731 21604 95.04%
thp_fault_fallback 0 0 0.00%
thp_collapse_alloc 1 0 0.00%
thp_collapse_alloc_failed 0 0 0.00%
thp_split_page 0 0 0.00%
thp_split_page_failed 0 0 0.00%
thp_deferred_split_page 22731 21604 95.04%
thp_split_pmd 0 0 0.00%
thp_zero_page_alloc 0 0 0.00%
thp_zero_page_alloc_failed 0 0 0.00%
balloon_inflate 0 0 0.00%
balloon_deflate 0 0 0.00%
balloon_migrate 0 0 0.00%
5:26 a.m.
Minchan Kim <minchan(a)kernel.org> writes:
On Thu, Jun 16, 2016 at 03:27:44PM -0700, Huang, Ying wrote:
> Minchan Kim <minchan(a)kernel.org> writes:
>
> > On Thu, Jun 16, 2016 at 07:52:26AM +0800, Huang, Ying wrote:
> >> "Kirill A. Shutemov" <kirill(a)shutemov.name> writes:
> >>
> >> > On Tue, Jun 14, 2016 at 05:57:28PM +0900, Minchan Kim wrote:
> >> >> On Wed, Jun 08, 2016 at 11:58:11AM +0300, Kirill A. Shutemov
wrote:
> >> >> > On Wed, Jun 08, 2016 at 04:41:37PM +0800, Huang, Ying wrote:
> >> >> > > "Huang, Ying" <ying.huang(a)intel.com>
writes:
> >> >> > >
> >> >> > > > "Kirill A. Shutemov"
<kirill.shutemov(a)linux.intel.com> writes:
> >> >> > > >
> >> >> > > >> On Mon, Jun 06, 2016 at 10:27:24AM +0800, kernel
test robot wrote:
> >> >> > > >>>
> >> >> > > >>> FYI, we noticed a -6.3% regression of
unixbench.score due to commit:
> >> >> > > >>>
> >> >> > > >>> commit
5c0a85fad949212b3e059692deecdeed74ae7ec7 ("mm: make faultaround produce old
ptes")
> >> >> > > >>>
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
> >> >> > > >>>
> >> >> > > >>> in testcase: unixbench
> >> >> > > >>> on test machine: lituya: 16 threads Haswell
High-end Desktop (i7-5960X 3.0G) with 16G memory
> >> >> > > >>> with following parameters:
cpufreq_governor=performance/nr_task=1/test=shell8
> >> >> > > >>>
> >> >> > > >>>
> >> >> > > >>> Details are as below:
> >> >> > > >>>
-------------------------------------------------------------------------------------------------->
> >> >> > > >>>
> >> >> > > >>>
> >> >> > > >>>
=========================================================================================
> >> >> > > >>>
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
> >> >> > > >>>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
> >> >> > > >>>
> >> >> > > >>> commit:
> >> >> > > >>> 4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
> >> >> > > >>> 5c0a85fad949212b3e059692deecdeed74ae7ec7
> >> >> > > >>>
> >> >> > > >>> 4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
> >> >> > > >>> ---------------- --------------------------
> >> >> > > >>> fail:runs %reproduction
fail:runs
> >> >> > > >>> | | |
> >> >> > > >>> 3:4 -75% :4
kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
> >> >> > > >>> %stddev %change
%stddev
> >> >> > > >>> \ | \
> >> >> > > >>> 14321 . 0% -6.3% 13425 . 0%
unixbench.score
> >> >> > > >>> 1996897 . 0% -6.1% 1874635 . 0%
unixbench.time.involuntary_context_switches
> >> >> > > >>> 1.721e+08 . 0% -6.2% 1.613e+08 . 0%
unixbench.time.minor_page_faults
> >> >> > > >>> 758.65 . 0% -3.0% 735.86 . 0%
unixbench.time.system_time
> >> >> > > >>> 387.66 . 0% +5.4% 408.49 . 0%
unixbench.time.user_time
> >> >> > > >>> 5950278 . 0% -6.2% 5583456 . 0%
unixbench.time.voluntary_context_switches
> >> >> > > >>
> >> >> > > >> That's weird.
> >> >> > > >>
> >> >> > > >> I don't understand why the change would
reduce number or minor faults.
> >> >> > > >> It should stay the same on x86-64. Rise of
user_time is puzzling too.
> >> >> > > >
> >> >> > > > unixbench runs in fixed time mode. That is, the
total time to run
> >> >> > > > unixbench is fixed, but the work done varies. So
the minor_page_faults
> >> >> > > > change may reflect only the work done.
> >> >> > > >
> >> >> > > >> Hm. Is reproducible? Across reboot?
> >> >> > > >
> >> >> > >
> >> >> > > And FYI, there is no swap setup for test, all root file
system including
> >> >> > > benchmark files are in tmpfs, so no real page reclaim
will be
> >> >> > > triggered. But it appears that active file cache reduced
after the
> >> >> > > commit.
> >> >> > >
> >> >> > > 111331 . 1% -13.3% 96503 . 0%
meminfo.Active
> >> >> > > 27603 . 1% -43.9% 15486 . 0%
meminfo.Active(file)
> >> >> > >
> >> >> > > I think this is the expected behavior of the commit?
> >> >> >
> >> >> > Yes, it's expected.
> >> >> >
> >> >> > After the change faularound would produce old pte. It means
there's more
> >> >> > chance for these pages to be on inactive lru, unless somebody
actually
> >> >> > touch them and flip accessed bit.
> >> >>
> >> >> Hmm, tmpfs pages should be in anonymous LRU list and VM
shouldn't scan
> >> >> anonymous LRU list on swapless system so I really wonder why active
file
> >> >> LRU is shrunk.
> >> >
> >> > Hm. Good point. I don't why we have anything on file lru if
there's no
> >> > filesystems except tmpfs.
> >> >
> >> > Ying, how do you get stuff to the tmpfs?
> >>
> >> We put root file system and benchmark into a set of compressed cpio
> >> archive, then concatenate them into one initrd, and finally kernel use
> >> that initrd as initramfs.
> >
> > I see.
> >
> > Could you share your 4 full vmstat(/proc/vmstat) files?
> >
> > old:
> >
> > cat /proc/vmstat > before.old.vmstat
> > do benchmark
> > cat /proc/vmstat > after.old.vmstat
> >
> > new:
> >
> > cat /proc/vmstat > before.new.vmstat
> > do benchmark
> > cat /proc/vmstat > after.new.vmstat
> >
> > IOW, I want to see stats related to reclaim.
>
> Hi,
>
> The /proc/vmstat for the parent commit (parent-proc-vmstat.gz) and first
> bad commit (fbc-proc-vmstat.gz) are attached with the email.
>
> The contents of the file is more than the vmstat before and after
> benchmark running, but are sampled every 1 seconds. Every sample begin
> with "time: <time>". You can check the first and last samples. The
> first /proc/vmstat capturing is started at the same time of the
> benchmark, so it is not exactly the vmstat before the benchmark running.
>
Thanks for the testing!
nr_active_file was shrunk 48% but the vaule itself is not huge so
I don't think it affects performance a lot.
There was no reclaim activity for testing. :(
pgfault, 6% reduced. Given that, pgalloc/free reduced 6%, too
because unixbench was time fixed mode and 6% regressed so no
doubt.
No interesting data.
It seems you tested it with THP, maybe always mode?
Yes. With following in kconfig.
CONFIG_TRANSPARENT_HUGEPAGE=y
CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS=y
I'm so sorry but could you test it with disabling
CONFIG_TRANSPARENT_HUGEPAGE=n
again? it might you already did.
Is it still 6% regressed with disabling THP?
Yes. I disabled THP via
echo never > /sys/kernel/mm/transparent_hugepage/enabled
echo never > /sys/kernel/mm/transparent_hugepage/defrag
The regression is the same as before.
=========================================================================================
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase/thp_defrag/thp_enabled:
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench/never/never
commit:
4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
5c0a85fad949212b3e059692deecdeed74ae7ec7
4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
---------------- --------------------------
%stddev %change %stddev
\ | \
14332 ± 0% -6.2% 13438 ± 0% unixbench.score
6662206 ± 0% -6.2% 6252260 ± 0% unixbench.time.involuntary_context_switches
5.734e+08 ± 0% -6.2% 5.376e+08 ± 0% unixbench.time.minor_page_faults
2527 ± 0% -3.2% 2446 ± 0% unixbench.time.system_time
1291 ± 0% +5.4% 1361 ± 0% unixbench.time.user_time
19875455 ± 0% -6.3% 18622488 ± 0% unixbench.time.voluntary_context_switches
6570355 ± 0% -11.9% 5787517 ± 0% cpuidle.C1-HSW.usage
17257 ± 34% -59.1% 7055 ± 7%
latency_stats.sum.ep_poll.SyS_epoll_wait.entry_SYSCALL_64_fastpath
5976 ± 0% -43.0% 3404 ± 0% proc-vmstat.nr_active_file
45729 ± 1% -22.5% 35439 ± 1% meminfo.Active
23905 ± 0% -43.0% 13619 ± 0% meminfo.Active(file)
8465 ± 3% -29.8% 5940 ± 3% slabinfo.pid.active_objs
8476 ± 3% -29.9% 5940 ± 3% slabinfo.pid.num_objs
3.46 ± 0% +12.5% 3.89 ± 0% turbostat.CPU%c3
67.09 ± 0% -2.1% 65.65 ± 0% turbostat.PkgWatt
96090 ± 0% -5.8% 90479 ± 0% vmstat.system.cs
9083 ± 0% -2.7% 8833 ± 0% vmstat.system.in
467.35 ± 78% +416.7% 2414 ± 45% sched_debug.cfs_rq:/.MIN_vruntime.avg
7477 ± 78% +327.7% 31981 ± 39% sched_debug.cfs_rq:/.MIN_vruntime.max
1810 ± 78% +360.1% 8327 ± 40% sched_debug.cfs_rq:/.MIN_vruntime.stddev
467.35 ± 78% +416.7% 2414 ± 45% sched_debug.cfs_rq:/.max_vruntime.avg
7477 ± 78% +327.7% 31981 ± 39% sched_debug.cfs_rq:/.max_vruntime.max
1810 ± 78% +360.1% 8327 ± 40% sched_debug.cfs_rq:/.max_vruntime.stddev
-10724 ± -7% -12.0% -9433 ± -3% sched_debug.cfs_rq:/.spread0.avg
-17721 ± -4% -9.8% -15978 ± -2% sched_debug.cfs_rq:/.spread0.min
90355 ± 9% +14.1% 103099 ± 5% sched_debug.cpu.avg_idle.min
0.12 ± 35% +325.0% 0.52 ± 46% sched_debug.cpu.cpu_load[0].min
21913 ± 2% +29.1% 28288 ± 14% sched_debug.cpu.curr->pid.avg
49953 ± 3% +30.2% 65038 ± 0% sched_debug.cpu.curr->pid.max
23062 ± 2% +30.1% 29996 ± 4% sched_debug.cpu.curr->pid.stddev
274.39 ± 5% -10.2% 246.27 ± 3% sched_debug.cpu.nr_uninterruptible.max
242.73 ± 4% -13.5% 209.90 ± 2% sched_debug.cpu.nr_uninterruptible.stddev
Best Regards,
Huang, Ying
nr_free_pages -6663 -6461 96.97%
nr_alloc_batch 2594 4013 154.70%
nr_inactive_anon 112 112 100.00%
nr_active_anon 2536 2159 85.13%
nr_inactive_file -567 -227 40.04%
nr_active_file 648 315 48.61%
nr_unevictable 0 0 0.00%
nr_mlock 0 0 0.00%
nr_anon_pages 2634 2161 82.04%
nr_mapped 511 530 103.72%
nr_file_pages 207 215 103.86%
nr_dirty -7 -6 85.71%
nr_writeback 0 0 0.00%
nr_slab_reclaimable 158 328 207.59%
nr_slab_unreclaimable 2208 2115 95.79%
nr_page_table_pages 268 247 92.16%
nr_kernel_stack 143 80 55.94%
nr_unstable 1 1 100.00%
nr_bounce 0 0 0.00%
nr_vmscan_write 0 0 0.00%
nr_vmscan_immediate_reclaim 0 0 0.00%
nr_writeback_temp 0 0 0.00%
nr_isolated_anon 0 0 0.00%
nr_isolated_file 0 0 0.00%
nr_shmem 131 131 100.00%
nr_dirtied 67 78 116.42%
nr_written 74 84 113.51%
nr_pages_scanned 0 0 0.00%
numa_hit 483752446 453696304 93.79%
numa_miss 0 0 0.00%
numa_foreign 0 0 0.00%
numa_interleave 0 0 0.00%
numa_local 483752445 453696304 93.79%
numa_other 1 0 0.00%
workingset_refault 0 0 0.00%
workingset_activate 0 0 0.00%
workingset_nodereclaim 0 0 0.00%
nr_anon_transparent_hugepages 1 0 0.00%
nr_free_cma 0 0 0.00%
nr_dirty_threshold -1316 -1274 96.81%
nr_dirty_background_threshold -658 -637 96.81%
pgpgin 0 0 0.00%
pgpgout 0 0 0.00%
pswpin 0 0 0.00%
pswpout 0 0 0.00%
pgalloc_dma 0 0 0.00%
pgalloc_dma32 60130977 56323630 93.67%
pgalloc_normal 457203182 428863437 93.80%
pgalloc_movable 0 0 0.00%
pgfree 517327743 485181251 93.79%
pgactivate 2059556 1930950 93.76%
pgdeactivate 0 0 0.00%
pgfault 572723351 537107146 93.78%
pgmajfault 0 0 0.00%
pglazyfreed 0 0 0.00%
pgrefill_dma 0 0 0.00%
pgrefill_dma32 0 0 0.00%
pgrefill_normal 0 0 0.00%
pgrefill_movable 0 0 0.00%
pgsteal_kswapd_dma 0 0 0.00%
pgsteal_kswapd_dma32 0 0 0.00%
pgsteal_kswapd_normal 0 0 0.00%
pgsteal_kswapd_movable 0 0 0.00%
pgsteal_direct_dma 0 0 0.00%
pgsteal_direct_dma32 0 0 0.00%
pgsteal_direct_normal 0 0 0.00%
pgsteal_direct_movable 0 0 0.00%
pgscan_kswapd_dma 0 0 0.00%
pgscan_kswapd_dma32 0 0 0.00%
pgscan_kswapd_normal 0 0 0.00%
pgscan_kswapd_movable 0 0 0.00%
pgscan_direct_dma 0 0 0.00%
pgscan_direct_dma32 0 0 0.00%
pgscan_direct_normal 0 0 0.00%
pgscan_direct_movable 0 0 0.00%
pgscan_direct_throttle 0 0 0.00%
zone_reclaim_failed 0 0 0.00%
pginodesteal 0 0 0.00%
slabs_scanned 0 0 0.00%
kswapd_inodesteal 0 0 0.00%
kswapd_low_wmark_hit_quickly 0 0 0.00%
kswapd_high_wmark_hit_quickly 0 0 0.00%
pageoutrun 0 0 0.00%
allocstall 0 0 0.00%
pgrotated 0 0 0.00%
drop_pagecache 0 0 0.00%
drop_slab 0 0 0.00%
numa_pte_updates 0 0 0.00%
numa_huge_pte_updates 0 0 0.00%
numa_hint_faults 0 0 0.00%
numa_hint_faults_local 0 0 0.00%
numa_pages_migrated 0 0 0.00%
pgmigrate_success 0 0 0.00%
pgmigrate_fail 0 0 0.00%
compact_migrate_scanned 0 0 0.00%
compact_free_scanned 0 0 0.00%
compact_isolated 0 0 0.00%
compact_stall 0 0 0.00%
compact_fail 0 0 0.00%
compact_success 0 0 0.00%
compact_daemon_wake 0 0 0.00%
htlb_buddy_alloc_success 0 0 0.00%
htlb_buddy_alloc_fail 0 0 0.00%
unevictable_pgs_culled 0 0 0.00%
unevictable_pgs_scanned 0 0 0.00%
unevictable_pgs_rescued 0 0 0.00%
unevictable_pgs_mlocked 0 0 0.00%
unevictable_pgs_munlocked 0 0 0.00%
unevictable_pgs_cleared 0 0 0.00%
unevictable_pgs_stranded 0 0 0.00%
thp_fault_alloc 22731 21604 95.04%
thp_fault_fallback 0 0 0.00%
thp_collapse_alloc 1 0 0.00%
thp_collapse_alloc_failed 0 0 0.00%
thp_split_page 0 0 0.00%
thp_split_page_failed 0 0 0.00%
thp_deferred_split_page 22731 21604 95.04%
thp_split_pmd 0 0 0.00%
thp_zero_page_alloc 0 0 0.00%
thp_zero_page_alloc_failed 0 0 0.00%
balloon_inflate 0 0 0.00%
balloon_deflate 0 0 0.00%
balloon_migrate 0 0 0.00%
10:06 a.m.
On Fri, Jun 17, 2016 at 12:26:51PM -0700, Huang, Ying wrote:
Minchan Kim <minchan(a)kernel.org> writes:
> On Thu, Jun 16, 2016 at 03:27:44PM -0700, Huang, Ying wrote:
>> Minchan Kim <minchan(a)kernel.org> writes:
>>
>> > On Thu, Jun 16, 2016 at 07:52:26AM +0800, Huang, Ying wrote:
>> >> "Kirill A. Shutemov" <kirill(a)shutemov.name> writes:
>> >>
>> >> > On Tue, Jun 14, 2016 at 05:57:28PM +0900, Minchan Kim wrote:
>> >> >> On Wed, Jun 08, 2016 at 11:58:11AM +0300, Kirill A. Shutemov
wrote:
>> >> >> > On Wed, Jun 08, 2016 at 04:41:37PM +0800, Huang, Ying
wrote:
>> >> >> > > "Huang, Ying" <ying.huang(a)intel.com>
writes:
>> >> >> > >
>> >> >> > > > "Kirill A. Shutemov"
<kirill.shutemov(a)linux.intel.com> writes:
>> >> >> > > >
>> >> >> > > >> On Mon, Jun 06, 2016 at 10:27:24AM +0800,
kernel test robot wrote:
>> >> >> > > >>>
>> >> >> > > >>> FYI, we noticed a -6.3% regression of
unixbench.score due to commit:
>> >> >> > > >>>
>> >> >> > > >>> commit
5c0a85fad949212b3e059692deecdeed74ae7ec7 ("mm: make faultaround produce old
ptes")
>> >> >> > > >>>
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
>> >> >> > > >>>
>> >> >> > > >>> in testcase: unixbench
>> >> >> > > >>> on test machine: lituya: 16 threads
Haswell High-end Desktop (i7-5960X 3.0G) with 16G memory
>> >> >> > > >>> with following parameters:
cpufreq_governor=performance/nr_task=1/test=shell8
>> >> >> > > >>>
>> >> >> > > >>>
>> >> >> > > >>> Details are as below:
>> >> >> > > >>>
-------------------------------------------------------------------------------------------------->
>> >> >> > > >>>
>> >> >> > > >>>
>> >> >> > > >>>
=========================================================================================
>> >> >> > > >>>
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase:
>> >> >> > > >>>
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench
>> >> >> > > >>>
>> >> >> > > >>> commit:
>> >> >> > > >>>
4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
>> >> >> > > >>>
5c0a85fad949212b3e059692deecdeed74ae7ec7
>> >> >> > > >>>
>> >> >> > > >>> 4b50bcc7eda4d3cc
5c0a85fad949212b3e059692de
>> >> >> > > >>> ----------------
--------------------------
>> >> >> > > >>> fail:runs %reproduction
fail:runs
>> >> >> > > >>> | |
|
>> >> >> > > >>> 3:4 -75%
:4 kmsg.DHCP/BOOTP:Reply_not_for_us,op[#]xid[#]
>> >> >> > > >>> %stddev %change
%stddev
>> >> >> > > >>> \ |
\
>> >> >> > > >>> 14321 . 0% -6.3% 13425
. 0% unixbench.score
>> >> >> > > >>> 1996897 . 0% -6.1% 1874635
. 0% unixbench.time.involuntary_context_switches
>> >> >> > > >>> 1.721e+08 . 0% -6.2% 1.613e+08
. 0% unixbench.time.minor_page_faults
>> >> >> > > >>> 758.65 . 0% -3.0% 735.86
. 0% unixbench.time.system_time
>> >> >> > > >>> 387.66 . 0% +5.4% 408.49
. 0% unixbench.time.user_time
>> >> >> > > >>> 5950278 . 0% -6.2% 5583456
. 0% unixbench.time.voluntary_context_switches
>> >> >> > > >>
>> >> >> > > >> That's weird.
>> >> >> > > >>
>> >> >> > > >> I don't understand why the change would
reduce number or minor faults.
>> >> >> > > >> It should stay the same on x86-64. Rise of
user_time is puzzling too.
>> >> >> > > >
>> >> >> > > > unixbench runs in fixed time mode. That is,
the total time to run
>> >> >> > > > unixbench is fixed, but the work done varies.
So the minor_page_faults
>> >> >> > > > change may reflect only the work done.
>> >> >> > > >
>> >> >> > > >> Hm. Is reproducible? Across reboot?
>> >> >> > > >
>> >> >> > >
>> >> >> > > And FYI, there is no swap setup for test, all root
file system including
>> >> >> > > benchmark files are in tmpfs, so no real page
reclaim will be
>> >> >> > > triggered. But it appears that active file cache
reduced after the
>> >> >> > > commit.
>> >> >> > >
>> >> >> > > 111331 . 1% -13.3% 96503 . 0%
meminfo.Active
>> >> >> > > 27603 . 1% -43.9% 15486 . 0%
meminfo.Active(file)
>> >> >> > >
>> >> >> > > I think this is the expected behavior of the
commit?
>> >> >> >
>> >> >> > Yes, it's expected.
>> >> >> >
>> >> >> > After the change faularound would produce old pte. It
means there's more
>> >> >> > chance for these pages to be on inactive lru, unless
somebody actually
>> >> >> > touch them and flip accessed bit.
>> >> >>
>> >> >> Hmm, tmpfs pages should be in anonymous LRU list and VM
shouldn't scan
>> >> >> anonymous LRU list on swapless system so I really wonder why
active file
>> >> >> LRU is shrunk.
>> >> >
>> >> > Hm. Good point. I don't why we have anything on file lru if
there's no
>> >> > filesystems except tmpfs.
>> >> >
>> >> > Ying, how do you get stuff to the tmpfs?
>> >>
>> >> We put root file system and benchmark into a set of compressed cpio
>> >> archive, then concatenate them into one initrd, and finally kernel use
>> >> that initrd as initramfs.
>> >
>> > I see.
>> >
>> > Could you share your 4 full vmstat(/proc/vmstat) files?
>> >
>> > old:
>> >
>> > cat /proc/vmstat > before.old.vmstat
>> > do benchmark
>> > cat /proc/vmstat > after.old.vmstat
>> >
>> > new:
>> >
>> > cat /proc/vmstat > before.new.vmstat
>> > do benchmark
>> > cat /proc/vmstat > after.new.vmstat
>> >
>> > IOW, I want to see stats related to reclaim.
>>
>> Hi,
>>
>> The /proc/vmstat for the parent commit (parent-proc-vmstat.gz) and first
>> bad commit (fbc-proc-vmstat.gz) are attached with the email.
>>
>> The contents of the file is more than the vmstat before and after
>> benchmark running, but are sampled every 1 seconds. Every sample begin
>> with "time: <time>". You can check the first and last samples.
The
>> first /proc/vmstat capturing is started at the same time of the
>> benchmark, so it is not exactly the vmstat before the benchmark running.
>>
>
> Thanks for the testing!
>
> nr_active_file was shrunk 48% but the vaule itself is not huge so
> I don't think it affects performance a lot.
>
> There was no reclaim activity for testing. :(
>
> pgfault, 6% reduced. Given that, pgalloc/free reduced 6%, too
> because unixbench was time fixed mode and 6% regressed so no
> doubt.
>
> No interesting data.
>
> It seems you tested it with THP, maybe always mode?
Yes. With following in kconfig.
CONFIG_TRANSPARENT_HUGEPAGE=y
CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS=y
> I'm so sorry but could you test it with disabling CONFIG_TRANSPARENT_HUGEPAGE=n
> again? it might you already did.
> Is it still 6% regressed with disabling THP?
Yes. I disabled THP via
echo never > /sys/kernel/mm/transparent_hugepage/enabled
echo never > /sys/kernel/mm/transparent_hugepage/defrag
The regression is the same as before.
Still, 6% user_time regression and there is no difference with previous
experiment which enabled THP so I agree the regression is caused by just
access bit setting from CPU side, which is rather surprising to me.
I don't know how unixbench shell scripts testing touch the memory
but it should be per-page overhead and 6% regression for that is too heavy.
Anyway, at least, it would be better to notice it to x86 maintainer.
Thanks for the test!
=========================================================================================
compiler/cpufreq_governor/kconfig/nr_task/rootfs/tbox_group/test/testcase/thp_defrag/thp_enabled:
gcc-4.9/performance/x86_64-rhel/1/debian-x86_64-2015-02-07.cgz/lituya/shell8/unixbench/never/never
commit:
4b50bcc7eda4d3cc9e3f2a0aa60e590fedf728c5
5c0a85fad949212b3e059692deecdeed74ae7ec7
4b50bcc7eda4d3cc 5c0a85fad949212b3e059692de
---------------- --------------------------
%stddev %change %stddev
\ | \
14332 ± 0% -6.2% 13438 ± 0% unixbench.score
6662206 ± 0% -6.2% 6252260 ± 0%
unixbench.time.involuntary_context_switches
5.734e+08 ± 0% -6.2% 5.376e+08 ± 0% unixbench.time.minor_page_faults
2527 ± 0% -3.2% 2446 ± 0% unixbench.time.system_time
1291 ± 0% +5.4% 1361 ± 0% unixbench.time.user_time
19875455 ± 0% -6.3% 18622488 ± 0% unixbench.time.voluntary_context_switches
6570355 ± 0% -11.9% 5787517 ± 0% cpuidle.C1-HSW.usage
17257 ± 34% -59.1% 7055 ± 7%
latency_stats.sum.ep_poll.SyS_epoll_wait.entry_SYSCALL_64_fastpath
5976 ± 0% -43.0% 3404 ± 0% proc-vmstat.nr_active_file
45729 ± 1% -22.5% 35439 ± 1% meminfo.Active
23905 ± 0% -43.0% 13619 ± 0% meminfo.Active(file)
8465 ± 3% -29.8% 5940 ± 3% slabinfo.pid.active_objs
8476 ± 3% -29.9% 5940 ± 3% slabinfo.pid.num_objs
3.46 ± 0% +12.5% 3.89 ± 0% turbostat.CPU%c3
67.09 ± 0% -2.1% 65.65 ± 0% turbostat.PkgWatt
96090 ± 0% -5.8% 90479 ± 0% vmstat.system.cs
9083 ± 0% -2.7% 8833 ± 0% vmstat.system.in
467.35 ± 78% +416.7% 2414 ± 45% sched_debug.cfs_rq:/.MIN_vruntime.avg
7477 ± 78% +327.7% 31981 ± 39% sched_debug.cfs_rq:/.MIN_vruntime.max
1810 ± 78% +360.1% 8327 ± 40% sched_debug.cfs_rq:/.MIN_vruntime.stddev
467.35 ± 78% +416.7% 2414 ± 45% sched_debug.cfs_rq:/.max_vruntime.avg
7477 ± 78% +327.7% 31981 ± 39% sched_debug.cfs_rq:/.max_vruntime.max
1810 ± 78% +360.1% 8327 ± 40% sched_debug.cfs_rq:/.max_vruntime.stddev
-10724 ± -7% -12.0% -9433 ± -3% sched_debug.cfs_rq:/.spread0.avg
-17721 ± -4% -9.8% -15978 ± -2% sched_debug.cfs_rq:/.spread0.min
90355 ± 9% +14.1% 103099 ± 5% sched_debug.cpu.avg_idle.min
0.12 ± 35% +325.0% 0.52 ± 46% sched_debug.cpu.cpu_load[0].min
21913 ± 2% +29.1% 28288 ± 14% sched_debug.cpu.curr->pid.avg
49953 ± 3% +30.2% 65038 ± 0% sched_debug.cpu.curr->pid.max
23062 ± 2% +30.1% 29996 ± 4% sched_debug.cpu.curr->pid.stddev
274.39 ± 5% -10.2% 246.27 ± 3% sched_debug.cpu.nr_uninterruptible.max
242.73 ± 4% -13.5% 209.90 ± 2% sched_debug.cpu.nr_uninterruptible.stddev
Best Regards,
Huang, Ying
2281
days inactive
2295
days old
22 comments
8 participants
participants (8)
-
Christian Borntraeger -
Huang, Ying -
kernel test robot -
Kirill A. Shutemov -
Kirill A. Shutemov -
Linus Torvalds -
Minchan Kim -
Rik van Riel