Re: [mm, thp] 85b9f46e8e: vm-scalability.throughput -8.7% regression

From: Huang, Ying
Date: Mon Oct 19 2020 - 23:23:55 EST


David Rientjes <rientjes@xxxxxxxxxx> writes:

> On Sun, 4 Oct 2020, kernel test robot wrote:
>
>> Greeting,
>>
>> FYI, we noticed a -8.7% regression of vm-scalability.throughput due to commit:
>>
>>
>> commit: 85b9f46e8ea451633ccd60a7d8cacbfff9f34047 ("mm, thp: track fallbacks due to failed memcg charges separately")
>> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git master
>>
>>
>> in testcase: vm-scalability
>> on test machine: 104 threads Skylake with 192G memory
>> with following parameters:
>>
>> runtime: 300s
>> size: 1T
>> test: lru-shm
>> cpufreq_governor: performance
>> ucode: 0x2006906
>>
>> test-description: The motivation behind this suite is to exercise functions and regions of the mm/ of the Linux kernel which are of interest to us.
>> test-url: https://git.kernel.org/cgit/linux/kernel/git/wfg/vm-scalability.git/
>>
>>
>>
>> If you fix the issue, kindly add following tag
>> Reported-by: kernel test robot <rong.a.chen@xxxxxxxxx>
>>
>>
>> Details are as below:
>> -------------------------------------------------------------------------------------------------->
>>
>>
>> To reproduce:
>>
>> git clone https://github.com/intel/lkp-tests.git
>> cd lkp-tests
>> bin/lkp install job.yaml # job file is attached in this email
>> bin/lkp run job.yaml
>>
>> =========================================================================================
>> compiler/cpufreq_governor/kconfig/rootfs/runtime/size/tbox_group/test/testcase/ucode:
>> gcc-9/performance/x86_64-rhel-8.3/debian-10.4-x86_64-20200603.cgz/300s/1T/lkp-skl-fpga01/lru-shm/vm-scalability/0x2006906
>>
>> commit:
>> dcdf11ee14 ("mm, shmem: add vmstat for hugepage fallback")
>> 85b9f46e8e ("mm, thp: track fallbacks due to failed memcg charges separately")
>>
>> dcdf11ee14413332 85b9f46e8ea451633ccd60a7d8c
>> ---------------- ---------------------------
>> fail:runs %reproduction fail:runs
>> | | |
>> 1:4 24% 2:4 perf-profile.calltrace.cycles-pp.sync_regs.error_entry.do_access
>> 3:4 53% 5:4 perf-profile.calltrace.cycles-pp.error_entry.do_access
>> 9:4 -27% 8:4 perf-profile.children.cycles-pp.error_entry
>> 4:4 -10% 4:4 perf-profile.self.cycles-pp.error_entry
>> %stddev %change %stddev
>> \ | \
>> 477291 -9.1% 434041 vm-scalability.median
>> 49791027 -8.7% 45476799 vm-scalability.throughput
>> 223.67 +1.6% 227.36 vm-scalability.time.elapsed_time
>> 223.67 +1.6% 227.36 vm-scalability.time.elapsed_time.max
>> 50364 ± 6% +24.1% 62482 ± 10% vm-scalability.time.involuntary_context_switches
>> 2237 +7.8% 2412 vm-scalability.time.percent_of_cpu_this_job_got
>> 3084 +18.2% 3646 vm-scalability.time.system_time
>> 1921 -4.2% 1839 vm-scalability.time.user_time
>> 13.68 +2.2 15.86 mpstat.cpu.all.sys%
>> 28535 ± 30% -47.0% 15114 ± 79% numa-numastat.node0.other_node
>> 142734 ± 11% -19.4% 115000 ± 17% numa-meminfo.node0.AnonPages
>> 11168 ± 3% +8.8% 12150 ± 5% numa-meminfo.node1.PageTables
>> 76.00 -1.6% 74.75 vmstat.cpu.id
>> 3626 -1.9% 3555 vmstat.system.cs
>> 2214928 ±166% -96.6% 75321 ± 7% cpuidle.C1.usage
>> 200981 ± 7% -18.0% 164861 ± 7% cpuidle.POLL.time
>> 52675 ± 3% -16.7% 43866 ± 10% cpuidle.POLL.usage
>> 35659 ± 11% -19.4% 28754 ± 17% numa-vmstat.node0.nr_anon_pages
>> 1248014 ± 3% +10.9% 1384236 numa-vmstat.node1.nr_mapped
>> 2722 ± 4% +10.6% 3011 ± 5% numa-vmstat.node1.nr_page_table_pages
>
> I'm not sure that I'm reading this correctly, but I suspect that this just
> happens because of NUMA: memory affinity will obviously impact
> vm-scalability.throughput quite substantially, but I don't think the
> bisected commit can be to be blame. Commit 85b9f46e8ea4 ("mm, thp: track
> fallbacks due to failed memcg charges separately") simply adds new
> count_vm_event() calls in a couple areas to track thp fallback due to
> memcg limits separate from fragmentation.
>
> It's likely a question about the testing methodology in general: for
> memory intensive benchmarks, I suggest it is configured in a manner that
> we can expect consistent memory access latency at the hardware level when
> running on a NUMA system.

So you think it's better to bind processes to NUMA node or CPU? But we
want to use this test case to capture NUMA/CPU placement/balance issue
too.

0day solve the problem in another way. We run the test case
multiple-times and calculate the average and standard deviation, then
compare.

For this specific regression, I found something strange,

10.93 ± 15% +10.8 21.78 ± 10% perf-profile.calltrace.cycles-pp._raw_spin_lock_irqsave.pagevec_lru_move_fn.__lru_cache_add.shmem_getpage_gfp.shmem_fault

It appears the lock contention becomes heavier with the patch. But I
cannot understand why too.

Best Regards,
Huang, Ying