Re: [PATCH 3/5] perf core: Prepare writing into ring buffer from end

["Wangnan (F)" <wangnan0@huawei.com>]

On 2016/3/28 9:58, Wangnan (F) wrote:
>
>
> On 2016/3/28 9:07, Wangnan (F) wrote:
>>
>>
>> On 2016/3/27 23:30, pi3orama wrote:
>>>
>>> 发自我的 iPhone
>>>
>>>> 在 2016年3月27日，下午11:20，Peter Zijlstra <peterz@infradead.org> 
>>>> 写道：
>>>>
>>>> On Fri, Mar 25, 2016 at 10:14:36PM +0800, Wangnan (F) wrote:
>>>>>>> I think you enabled some unusual config options?
>>>> x86_64-defconfig
>>>>
>>>>>> You must enabled CONFIG_OPTIMIZE_INLINING. Now I get similar result:
>>>> It has that indeed.
>>>>
>>>>> After enabling CONFIG_OPTIMIZE_INLINING:
>>>>>
>>>>> Test its performance by calling 'close(-1)' for 3000000 times and
>>>>> use 'perf record -o /dev/null -e raw_syscalls:* test-ring-buffer' to
>>>>> capture system calls:
>>>>>
>>>>>                   MEAN         STDVAR
>>>>> BASE            800077.1     23448.13
>>>>> RAWPERF.PRE    2465858.0    603473.70
>>>>> RAWPERF.POST   2471925.0    609437.60
>>>>>
>>>>> Considering the high stdvar, after applying this patch the 
>>>>> performance
>>>>> is not change.
>>>> Why is your variance so immense? And doesn't that render the
>>>> measurements pointless?
>>>>
>>> For some unknown reason, about
>>> 10% of these results raises 2 times of normal
>>> results. Say, "normal results" are about
>>> 2200000, but those "outliers" are about
>>> 4400000 (I can't access raw data now).
>>> Variance becomes much smaller if I remove
>>> those outliers.
>>
>
> Find the reason of these outliners.
>
> If perf and 'test-ring-buffer' are scheduled on different processors,
> the performance is bad. I think cache is the main reason.
>
> I will redo the test, bind them to cores on same CPU.
>
> Thank you.

Test method improvements:

1. Set CPU freq:

  # for f in /sys/devices/system/cpu/cpufreq/policy*/scaling_governor  ; 
do echo performance > $f ; done

2. Bind core:
Add following code into head of test-ring-buffer:

         CPU_ZERO(&mask);
         CPU_SET(6, &mask);
         pthread_setaffinity_np(pthread_self(), sizeof(mask), &mask);
         pthread_yield();

3. Bind core (perf):

Use following command to start perf:

  # taskset -c 7 ./perf record -o /dev/null --no-buildid-cache -e 
raw_syscalls:* test-ring-buffer

New result of 100 test data in both cases:

                   MEAN         STDVAR
BASE            800214.950    2853.083
RAWPERF.PRE    2253846.700    9997.014
RAWPERF.POST   2257495.540    8516.293

Thank you.