Re: [PATCH v2 12/15] perf record: introduce thread local variable for trace streaming

[Alexey Budankov]

On 27.10.2020 15:01, Jiri Olsa wrote:
> On Mon, Oct 26, 2020 at 05:11:30PM +0300, Alexei Budankov wrote:
>>
>> On 26.10.2020 13:34, Jiri Olsa wrote:
>>> On Mon, Oct 26, 2020 at 11:21:28AM +0300, Alexei Budankov wrote:
>>>>
>>>> On 24.10.2020 18:43, Jiri Olsa wrote:
>>>>> On Wed, Oct 21, 2020 at 07:07:00PM +0300, Alexey Budankov wrote:
>>>>>>
>>>>>> Introduce thread local variable and use it for threaded trace streaming.
>>>>>>
>>>>>> Signed-off-by: Alexey Budankov <alexey.budankov@xxxxxxxxxxxxxxx>
>>>>>> ---
>>>>>>  tools/perf/builtin-record.c | 71 ++++++++++++++++++++++++++++++++-----
>>>>>>  1 file changed, 62 insertions(+), 9 deletions(-)
>>>>>>
>>>>>> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
>>>>>> index 89cb8e913fb3..3b7e9026f25b 100644
>>>>>> --- a/tools/perf/builtin-record.c
>>>>>> +++ b/tools/perf/builtin-record.c
>>>>>> @@ -101,6 +101,8 @@ struct thread_data {
>>>>>>  	u64		   bytes_written;
>>>>>>  };
>>>>>>  
>>>>>> +static __thread struct thread_data *thread;
>>>>>> +
>>>>>>  struct record {
>>>>>>  	struct perf_tool	tool;
>>>>>>  	struct record_opts	opts;
>>>>>> @@ -587,7 +589,11 @@ static int record__pushfn(struct mmap *map, void *to, void *bf, size_t size)
>>>>>>  		}
>>>>>>  	}
>>>>>>  
>>>>>> -	rec->samples++;
>>>>>> +	if (thread)
>>>>>> +		thread->samples++;
>>>>>> +	else
>>>>>> +		rec->samples++;
>>>>>
>>>>> this is really wrong, let's keep just single samples counter
>>>>> ditto for all the other places in this patch
>>>>
>>>> This does look like data parallelism [1] which is very true for
>>>> threaded trace streaming so your prototype design looks optimal.
>>>>
>>>> For this specific place incrementing global counter in memory is
>>>> less performant and faces scalability limitations as a number of
>>>> cores grow.
>>>>
>>>> Not sure why you have changed your mind.
>>>
>>> I'm not sure I follow.. what I'm complaining about is to have
>>> 'samples' stat variable in separate locations for --threads
>>> and --no-threads mode
>>
>> It is optimal to have samples variable as per thread one
>> and then sum up the total in the end of data collection.
>>
>> Single global variable design has scalability and performance
>> drawbacks.
>>
>> Why do you complain about per thread variable in this case?
>> It looks like ideally fits these specific needs.
> 
> I think there's misunderstanding.. I think we should move
> samples to per thread 'thread' object and have just one
> copy of that.. and do not increase separate variables for
> thread and non-thread cases

Aw, I see. Using the same __thread object by main thread in
serial and threaded modes. That makes sense.
I will try in v3.

Alexei

> 
> jirka
>