Re: [LKP] [SUNRPC] 0472e47660: fsmark.app_overhead 16.0% regression

[Xing Zhengjun <zhengjun.xing@linux.intel.com>]

On 5/31/2019 3:10 AM, Trond Myklebust wrote:
> On Thu, 2019-05-30 at 15:20 +0800, Xing Zhengjun wrote:
>>
>> On 5/30/2019 10:00 AM, Trond Myklebust wrote:
>>> Hi Xing,
>>>
>>> On Thu, 2019-05-30 at 09:35 +0800, Xing Zhengjun wrote:
>>>> Hi Trond,
>>>>
>>>> On 5/20/2019 1:54 PM, kernel test robot wrote:
>>>>> Greeting,
>>>>>
>>>>> FYI, we noticed a 16.0% improvement of fsmark.app_overhead due
>>>>> to
>>>>> commit:
>>>>>
>>>>>
>>>>> commit: 0472e476604998c127f3c80d291113e77c5676ac ("SUNRPC:
>>>>> Convert
>>>>> socket page send code to use iov_iter()")
>>>>> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git
>>>>> master
>>>>>
>>>>> in testcase: fsmark
>>>>> on test machine: 40 threads Intel(R) Xeon(R) CPU E5-2690 v2 @
>>>>> 3.00GHz with 384G memory
>>>>> with following parameters:
>>>>>
>>>>> 	iterations: 1x
>>>>> 	nr_threads: 64t
>>>>> 	disk: 1BRD_48G
>>>>> 	fs: xfs
>>>>> 	fs2: nfsv4
>>>>> 	filesize: 4M
>>>>> 	test_size: 40G
>>>>> 	sync_method: fsyncBeforeClose
>>>>> 	cpufreq_governor: performance
>>>>>
>>>>> test-description: The fsmark is a file system benchmark to test
>>>>> synchronous write workloads, for example, mail servers
>>>>> workload.
>>>>> test-url: https://sourceforge.net/projects/fsmark/
>>>>>
>>>>>
>>>>>
>>>>> 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/disk/filesize/fs2/fs/iterations/kconf
>>>>> ig/n
>>>>> r_threads/rootfs/sync_method/tbox_group/test_size/testcase:
>>>>>      gcc-7/performance/1BRD_48G/4M/nfsv4/xfs/1x/x86_64-rhel-
>>>>> 7.6/64t/debian-x86_64-2018-04-03.cgz/fsyncBeforeClose/lkp-ivb-
>>>>> ep01/40G/fsmark
>>>>>
>>>>> commit:
>>>>>      e791f8e938 ("SUNRPC: Convert xs_send_kvec() to use
>>>>> iov_iter_kvec()")
>>>>>      0472e47660 ("SUNRPC: Convert socket page send code to use
>>>>> iov_iter()")
>>>>>
>>>>> e791f8e9380d945e 0472e476604998c127f3c80d291
>>>>> ---------------- ---------------------------
>>>>>           fail:runs  %reproduction    fail:runs
>>>>>               |             |             |
>>>>>               :4           50%           2:4     dmesg.WARNING:a
>>>>> t#for
>>>>> _ip_interrupt_entry/0x
>>>>>             %stddev     %change         %stddev
>>>>>                 \          |                \
>>>>>      15118573
>>>>> ±  2%     +16.0%   17538083        fsmark.app_overhead
>>>>>        510.93           -
>>>>> 22.7%     395.12        fsmark.files_per_sec
>>>>>         24.90           +22.8%      30.57        fsmark.time.ela
>>>>> psed_
>>>>> time
>>>>>         24.90           +22.8%      30.57        fsmark.time.ela
>>>>> psed_
>>>>> time.max
>>>>>        288.00 ±  2%     -
>>>>> 27.8%     208.00        fsmark.time.percent_of_cpu_this_job_got
>>>>>         70.03 ±  2%     -
>>>>> 11.3%      62.14        fsmark.time.system_time
>>>>>
>>>>
>>>> Do you have time to take a look at this regression?
>>>
>>>   From your stats, it looks to me as if the problem is increased
>>> NUMA
>>> overhead. Pretty much everything else appears to be the same or
>>> actually performing better than previously. Am I interpreting that
>>> correctly?
>> The real regression is the throughput(fsmark.files_per_sec) is
>> decreased
>> by 22.7%.
> 
> Understood, but I'm trying to make sense of why. I'm not able to
> reproduce this, so I have to rely on your performance stats to
> understand where the 22.7% regression is coming from. As far as I can
> see, the only numbers in the stats you published that are showing a
> performance regression (other than the fsmark number itself), are the
> NUMA numbers. Is that a correct interpretation?
> 
We re-test the case yesterday, the test result almost is the same.
we will do more test and also check the test case itself, if you need
more information, please let me know, thanks.

>>> If my interpretation above is correct, then I'm not seeing where
>>> this
>>> patch would be introducing new NUMA regressions. It is just
>>> converting
>>> from using one method of doing socket I/O to another. Could it
>>> perhaps
>>> be a memory artefact due to your running the NFS client and server
>>> on
>>> the same machine?
>>>
>>> Apologies for pushing back a little, but I just don't have the
>>> hardware available to test NUMA configurations, so I'm relying on
>>> external testing for the above kind of scenario.
>>>
>> Thanks for looking at this.  If you need more information, please let
>> me
>> know.
>>> Thanks
>>>     Trond
>>>

-- 
Zhengjun Xing