Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Message-ID: <8b7f911abe2cb38665727b0f8ddfc4ea4a4a1623.camel@gmail.com>
Subject: Re: [RFC PATCH v2 10/10] perf synthetic-events: use workqueue
 parallel_for
From:   Riccardo Mancini <rickyman7@gmail.com>
To:     Jiri Olsa <jolsa@redhat.com>
Cc:     Arnaldo Carvalho de Melo <acme@kernel.org>,
        Ian Rogers <irogers@google.com>,
        Namhyung Kim <namhyung@kernel.org>,
        Peter Zijlstra <peterz@infradead.org>,
        Ingo Molnar <mingo@redhat.com>,
        Mark Rutland <mark.rutland@arm.com>,
        linux-kernel@vger.kernel.org, linux-perf-users@vger.kernel.org,
        Alexey Bayduraev <alexey.v.bayduraev@linux.intel.com>
Date:   Mon, 09 Aug 2021 15:24:40 +0200
In-Reply-To: <YREZ4G1xzncpdsVk@krava>
References: <cover.1627657061.git.rickyman7@gmail.com>
         <0e9bdbcb903b24b95841e09bbae180841b6311ca.1627657061.git.rickyman7@gmail.com>
         <YREZ4G1xzncpdsVk@krava>
Content-Type: text/plain; charset="UTF-8"
User-Agent: Evolution 3.40.3 (3.40.3-1.fc34) 
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Hi Jiri,
thanks for looking into this patchset.

On Mon, 2021-08-09 at 14:04 +0200, Jiri Olsa wrote:
> On Fri, Jul 30, 2021 at 05:34:17PM +0200, Riccardo Mancini wrote:
> > To generate synthetic events, perf has the option to use multiple
> > threads. These threads are created manually using pthread_created.
> > 
> > This patch replaces the manual pthread_create with a workqueue,
> > using the parallel_for utility.
> 
> hi,
> I really like this new interface
> 
> > 
> > Experimental results show that workqueue has a slightly higher overhead,
> > but this is repayed by the improved work balancing among threads.
> 
> how did you measure that balancing improvement?
> is there less kernel cycles spent?

I meant that the workqueue with the shared queue is able to balance work among
threads. This is particulary important in synthesize since low pids are
associated to kthreads, which require less processing (as far as I understand),
therefore the current work assignment is not great.

I think the goal of the workqueue is not to be faster than the current
implementation (which it is only due to the aforementioned issue), but to have a
better abstraction with a contained overhead.

> 
> I ran the benchmark and if I'm reading the results correctly I see
> performance drop for high cpu numbers (full list attached below).

The implementation with the shared queue suffers from this problem, but I hoped
it would hold up to more threads.
I'm working on having one queue per thread, in order to be able to scale better
on more cpus. I do not have any workstealing at the moment so there is no
rebalancing of work, but in our usecases it is not that important, at the
moment.
You can find it on https://github.com/Manciukic/linux.git in perf/workqueue/dev
branch. If you can run the same benchmark there, it would be really helpful for
me.

Thanks,
Riccardo

> 
> 
> old perf:                                                                 new
> perf:
> 
> [jolsa@dell-r440-01 perf]$ ./perf.old bench internals synthesize -t      
> [jolsa@dell-r440-01 perf]$ ./perf bench internals synthesize -t
> ...
>   Number of synthesis threads: 40                                          
> Number of synthesis threads: 40
>     Average synthesis took: 2489.400 usec (+- 49.832 usec)                   
> Average synthesis took: 4576.500 usec (+- 75.278 usec)
>     Average num. events: 956.800 (+- 6.721)                                  
> Average num. events: 1020.000 (+- 0.000)
>     Average time per event 2.602 usec                                        
> Average time per event 4.487 usec
> 
> maybe profiling will show what's going on?
> 
> thanks,
> jirka
> 
> 
> ---
> [jolsa@dell-r440-01 perf]$ ./perf.old bench internals synthesize -t      
> [jolsa@dell-r440-01 perf]$ ./perf bench internals synthesize -t
> # Running 'internals/synthesize' benchmark:                               #
> Running 'internals/synthesize' benchmark:
> Computing performance of multi threaded perf event synthesis by          
> Computing performance of multi threaded perf event synthesis by
> synthesizing events on CPU 0:                                            
> synthesizing events on CPU 0:
>   Number of synthesis threads: 1                                           
> Number of synthesis threads: 1 
>     Average synthesis took: 7907.100 usec (+- 197.363 usec)                  
> Average synthesis took: 7972.900 usec (+- 198.158 usec)
>     Average num. events: 956.000 (+- 0.000)                                  
> Average num. events: 936.000 (+- 0.000)
>     Average time per event 8.271 usec                                        
> Average time per event 8.518 usec
>   Number of synthesis threads: 2                                           
> Number of synthesis threads: 2 
>     Average synthesis took: 5616.800 usec (+- 61.253 usec)                   
> Average synthesis took: 5844.700 usec (+- 87.219 usec)
>     Average num. events: 958.800 (+- 0.327)                                  
> Average num. events: 940.000 (+- 0.000)
>     Average time per event 5.858 usec                                        
> Average time per event 6.218 usec
>   Number of synthesis threads: 3                                           
> Number of synthesis threads: 3 
>     Average synthesis took: 4274.000 usec (+- 93.293 usec)                   
> Average synthesis took: 4019.700 usec (+- 67.354 usec)
>     Average num. events: 962.000 (+- 0.000)                                  
> Average num. events: 942.000 (+- 0.000)
>     Average time per event 4.443 usec                                        
> Average time per event 4.267 usec
>   Number of synthesis threads: 4                                           
> Number of synthesis threads: 4 
>     Average synthesis took: 3425.700 usec (+- 43.044 usec)                   
> Average synthesis took: 3382.200 usec (+- 74.652 usec)
>     Average num. events: 959.600 (+- 0.933)                                  
> Average num. events: 944.000 (+- 0.000)
>     Average time per event 3.570 usec                                        
> Average time per event 3.583 usec
>   Number of synthesis threads: 5                                           
> Number of synthesis threads: 5 
>     Average synthesis took: 2958.000 usec (+- 82.951 usec)                   
> Average synthesis took: 3086.500 usec (+- 48.213 usec)
>     Average num. events: 966.000 (+- 0.000)                                  
> Average num. events: 946.000 (+- 0.000)
>     Average time per event 3.062 usec                                        
> Average time per event 3.263 usec
>   Number of synthesis threads: 6                                           
> Number of synthesis threads: 6 
>     Average synthesis took: 2808.400 usec (+- 66.868 usec)                   
> Average synthesis took: 2752.200 usec (+- 56.411 usec)
>     Average num. events: 956.800 (+- 0.327)                                  
> Average num. events: 948.000 (+- 0.000)
>     Average time per event 2.935 usec                                        
> Average time per event 2.903 usec
>   Number of synthesis threads: 7                                           
> Number of synthesis threads: 7 
>     Average synthesis took: 2622.900 usec (+- 83.524 usec)                   
> Average synthesis took: 2548.200 usec (+- 48.042 usec)
>     Average num. events: 958.400 (+- 0.267)                                  
> Average num. events: 950.000 (+- 0.000)
>     Average time per event 2.737 usec                                        
> Average time per event 2.682 usec
>   Number of synthesis threads: 8                                           
> Number of synthesis threads: 8 
>     Average synthesis took: 2271.600 usec (+- 29.181 usec)                   
> Average synthesis took: 2486.600 usec (+- 47.862 usec)
>     Average num. events: 972.000 (+- 0.000)                                  
> Average num. events: 952.000 (+- 0.000)
>     Average time per event 2.337 usec                                        
> Average time per event 2.612 usec
>   Number of synthesis threads: 9                                           
> Number of synthesis threads: 9 
>     Average synthesis took: 2372.000 usec (+- 95.495 usec)                   
> Average synthesis took: 2347.300 usec (+- 23.959 usec)
>     Average num. events: 959.200 (+- 0.952)                                  
> Average num. events: 954.000 (+- 0.000)
>     Average time per event 2.473 usec                                        
> Average time per event 2.460 usec
>   Number of synthesis threads: 10                                          
> Number of synthesis threads: 10
>     Average synthesis took: 2544.600 usec (+- 107.569 usec)                  
> Average synthesis took: 2328.800 usec (+- 14.234 usec)
>     Average num. events: 968.400 (+- 3.124)                                  
> Average num. events: 957.400 (+- 0.306)
>     Average time per event 2.628 usec                                        
> Average time per event 2.432 usec
>   Number of synthesis threads: 11                                          
> Number of synthesis threads: 11
>     Average synthesis took: 2299.300 usec (+- 57.597 usec)                   
> Average synthesis took: 2340.300 usec (+- 34.638 usec)
>     Average num. events: 956.000 (+- 0.000)                                  
> Average num. events: 960.000 (+- 0.000)
>     Average time per event 2.405 usec                                        
> Average time per event 2.438 usec
>   Number of synthesis threads: 12                                          
> Number of synthesis threads: 12
>     Average synthesis took: 2545.500 usec (+- 69.557 usec)                   
> Average synthesis took: 2318.700 usec (+- 15.803 usec)
>     Average num. events: 974.800 (+- 0.611)                                  
> Average num. events: 963.800 (+- 0.200)
>     Average time per event 2.611 usec                                        
> Average time per event 2.406 usec
>   Number of synthesis threads: 13                                          
> Number of synthesis threads: 13
>     Average synthesis took: 2386.400 usec (+- 79.244 usec)                   
> Average synthesis took: 2408.700 usec (+- 27.071 usec)
>     Average num. events: 950.500 (+- 5.726)                                  
> Average num. events: 966.000 (+- 0.000)
>     Average time per event 2.511 usec                                        
> Average time per event 2.493 usec
>   Number of synthesis threads: 14                                          
> Number of synthesis threads: 14 
>     Average synthesis took: 2466.600 usec (+- 57.893 usec)                   
> Average synthesis took: 2547.200 usec (+- 53.445 usec)
>     Average num. events: 957.600 (+- 0.718)                                  
> Average num. events: 968.000 (+- 0.000)
>     Average time per event 2.576 usec                                        
> Average time per event 2.631 usec
>   Number of synthesis threads: 15                                          
> Number of synthesis threads: 15 
>     Average synthesis took: 2249.700 usec (+- 64.026 usec)                   
> Average synthesis took: 2647.900 usec (+- 79.014 usec)
>     Average num. events: 956.000 (+- 0.000)                                  
> Average num. events: 970.000 (+- 0.000)
>     Average time per event 2.353 usec                                        
> Average time per event 2.730 usec
>   Number of synthesis threads: 16                                          
> Number of synthesis threads: 16 
>     Average synthesis took: 2311.700 usec (+- 64.304 usec)                   
> Average synthesis took: 2676.200 usec (+- 34.824 usec)
>     Average num. events: 955.000 (+- 0.907)                                  
> Average num. events: 972.000 (+- 0.000)
>     Average time per event 2.421 usec                                        
> Average time per event 2.753 usec
>   Number of synthesis threads: 17                                          
> Number of synthesis threads: 17 
>     Average synthesis took: 2174.100 usec (+- 36.673 usec)                   
> Average synthesis took: 2580.100 usec (+- 45.414 usec)
>     Average num. events: 971.600 (+- 3.124)                                  
> Average num. events: 974.000 (+- 0.000)
>     Average time per event 2.238 usec                                        
> Average time per event 2.649 usec
>   Number of synthesis threads: 18                                          
> Number of synthesis threads: 18 
>     Average synthesis took: 2294.200 usec (+- 63.657 usec)                   
> Average synthesis took: 2810.200 usec (+- 49.113 usec)
>     Average num. events: 953.200 (+- 0.611)                                  
> Average num. events: 976.000 (+- 0.000)
>     Average time per event 2.407 usec                                        
> Average time per event 2.879 usec
>   Number of synthesis threads: 19                                          
> Number of synthesis threads: 19 
>     Average synthesis took: 2410.700 usec (+- 120.169 usec)                  
> Average synthesis took: 2862.400 usec (+- 36.982 usec)
>     Average num. events: 953.400 (+- 0.306)                                  
> Average num. events: 978.000 (+- 0.000)
>     Average time per event 2.529 usec                                        
> Average time per event 2.927 usec
>   Number of synthesis threads: 20                                          
> Number of synthesis threads: 20 
>     Average synthesis took: 2387.000 usec (+- 91.051 usec)                   
> Average synthesis took: 2908.800 usec (+- 36.404 usec)
>     Average num. events: 952.800 (+- 0.800)                                  
> Average num. events: 978.600 (+- 0.306)
>     Average time per event 2.505 usec                                        
> Average time per event 2.972 usec
>   Number of synthesis threads: 21                                          
> Number of synthesis threads: 21 
>     Average synthesis took: 2275.700 usec (+- 39.815 usec)                   
> Average synthesis took: 3141.100 usec (+- 30.896 usec)
>     Average num. events: 954.600 (+- 0.306)                                  
> Average num. events: 980.000 (+- 0.000)
>     Average time per event 2.384 usec                                        
> Average time per event 3.205 usec
>   Number of synthesis threads: 22                                          
> Number of synthesis threads: 22 
>     Average synthesis took: 2373.200 usec (+- 89.528 usec)                   
> Average synthesis took: 3342.400 usec (+- 112.115 usec)
>     Average num. events: 949.100 (+- 5.843)                                  
> Average num. events: 982.000 (+- 0.000)
>     Average time per event 2.500 usec                                        
> Average time per event 3.404 usec
>   Number of synthesis threads: 23                                          
> Number of synthesis threads: 23 
>     Average synthesis took: 2318.300 usec (+- 39.395 usec)                   
> Average synthesis took: 3269.700 usec (+- 55.215 usec)
>     Average num. events: 954.600 (+- 0.427)                                  
> Average num. events: 984.000 (+- 0.000)
>     Average time per event 2.429 usec                                        
> Average time per event 3.323 usec
>   Number of synthesis threads: 24                                          
> Number of synthesis threads: 24
>     Average synthesis took: 2241.900 usec (+- 52.577 usec)                   
> Average synthesis took: 3379.500 usec (+- 56.380 usec)
>     Average num. events: 954.000 (+- 0.000)                                  
> Average num. events: 986.000 (+- 0.000)
>     Average time per event 2.350 usec                                        
> Average time per event 3.427 usec
>   Number of synthesis threads: 25                                          
> Number of synthesis threads: 25
>     Average synthesis took: 2343.400 usec (+- 101.611 usec)                  
> Average synthesis took: 3382.500 usec (+- 51.535 usec)
>     Average num. events: 956.200 (+- 1.009)                                  
> Average num. events: 988.000 (+- 0.000)
>     Average time per event 2.451 usec                                        
> Average time per event 3.424 usec
>   Number of synthesis threads: 26                                          
> Number of synthesis threads: 26
>     Average synthesis took: 2260.700 usec (+- 18.863 usec)                   
> Average synthesis took: 3391.600 usec (+- 44.053 usec)
>     Average num. events: 954.000 (+- 0.000)                                  
> Average num. events: 990.000 (+- 0.000)
>     Average time per event 2.370 usec                                        
> Average time per event 3.426 usec
>   Number of synthesis threads: 27                                          
> Number of synthesis threads: 27
>     Average synthesis took: 2373.800 usec (+- 74.213 usec)                   
> Average synthesis took: 3659.200 usec (+- 113.176 usec)
>     Average num. events: 955.000 (+- 0.803)                                  
> Average num. events: 992.000 (+- 0.000)
>     Average time per event 2.486 usec                                        
> Average time per event 3.689 usec
>   Number of synthesis threads: 28                                          
> Number of synthesis threads: 28
>     Average synthesis took: 2335.500 usec (+- 49.480 usec)                   
> Average synthesis took: 3625.000 usec (+- 90.131 usec)
>     Average num. events: 954.000 (+- 0.000)                                  
> Average num. events: 994.000 (+- 0.000)
>     Average time per event 2.448 usec                                        
> Average time per event 3.647 usec
>   Number of synthesis threads: 29                                          
> Number of synthesis threads: 29
>     Average synthesis took: 2182.100 usec (+- 41.649 usec)                   
> Average synthesis took: 3708.400 usec (+- 103.717 usec)
>     Average num. events: 954.000 (+- 0.000)                                  
> Average num. events: 996.000 (+- 0.000)
>     Average time per event 2.287 usec                                        
> Average time per event 3.723 usec
>   Number of synthesis threads: 30                                          
> Number of synthesis threads: 30
>     Average synthesis took: 2246.100 usec (+- 58.252 usec)                   
> Average synthesis took: 3820.500 usec (+- 95.282 usec)
>     Average num. events: 954.000 (+- 0.000)                                  
> Average num. events: 998.000 (+- 0.000)
>     Average time per event 2.354 usec                                        
> Average time per event 3.828 usec
>   Number of synthesis threads: 31                                          
> Number of synthesis threads: 31
>     Average synthesis took: 2156.900 usec (+- 26.141 usec)                   
> Average synthesis took: 3881.400 usec (+- 36.277 usec)
>     Average num. events: 948.300 (+- 5.700)                                  
> Average num. events: 1000.000 (+- 0.000)
>     Average time per event 2.274 usec                                        
> Average time per event 3.881 usec
>   Number of synthesis threads: 32                                          
> Number of synthesis threads: 32
>     Average synthesis took: 2295.300 usec (+- 41.538 usec)                   
> Average synthesis took: 4191.700 usec (+- 149.780 usec)
>     Average num. events: 954.000 (+- 0.000)                                  
> Average num. events: 1002.000 (+- 0.000)
>     Average time per event 2.406 usec                                        
> Average time per event 4.183 usec
>   Number of synthesis threads: 33                                          
> Number of synthesis threads: 33
>     Average synthesis took: 2249.100 usec (+- 59.135 usec)                   
> Average synthesis took: 3988.200 usec (+- 25.015 usec)
>     Average num. events: 948.500 (+- 5.726)                                  
> Average num. events: 1004.000 (+- 0.000)
>     Average time per event 2.371 usec                                        
> Average time per event 3.972 usec
>   Number of synthesis threads: 34                                          
> Number of synthesis threads: 34
>     Average synthesis took: 2270.400 usec (+- 65.011 usec)                   
> Average synthesis took: 4064.600 usec (+- 44.158 usec)
>     Average num. events: 954.200 (+- 0.200)                                  
> Average num. events: 1006.000 (+- 0.000)
>     Average time per event 2.379 usec                                        
> Average time per event 4.040 usec
>   Number of synthesis threads: 35                                          
> Number of synthesis threads: 35
>     Average synthesis took: 2259.200 usec (+- 44.287 usec)                   
> Average synthesis took: 4145.700 usec (+- 37.297 usec)
>     Average num. events: 954.000 (+- 0.000)                                  
> Average num. events: 1008.000 (+- 0.000)
>     Average time per event 2.368 usec                                        
> Average time per event 4.113 usec
>   Number of synthesis threads: 36                                          
> Number of synthesis threads: 36
>     Average synthesis took: 2294.100 usec (+- 38.693 usec)                   
> Average synthesis took: 4234.900 usec (+- 81.904 usec)
>     Average num. events: 954.000 (+- 0.000)                                  
> Average num. events: 1010.400 (+- 0.267)
>     Average time per event 2.405 usec                                        
> Average time per event 4.191 usec
>   Number of synthesis threads: 37                                          
> Number of synthesis threads: 37
>     Average synthesis took: 2338.900 usec (+- 80.346 usec)                   
> Average synthesis took: 4337.900 usec (+- 30.071 usec)
>     Average num. events: 954.400 (+- 0.267)                                  
> Average num. events: 1014.000 (+- 0.000)
>     Average time per event 2.451 usec                                        
> Average time per event 4.278 usec
>   Number of synthesis threads: 38                                          
> Number of synthesis threads: 38
>     Average synthesis took: 2406.300 usec (+- 57.140 usec)                   
> Average synthesis took: 4426.600 usec (+- 27.035 usec)
>     Average num. events: 938.400 (+- 7.730)                                  
> Average num. events: 1016.000 (+- 0.000)
>     Average time per event 2.564 usec                                        
> Average time per event 4.357 usec
>   Number of synthesis threads: 39                                          
> Number of synthesis threads: 39
>     Average synthesis took: 2371.000 usec (+- 35.676 usec)                   
> Average synthesis took: 5979.000 usec (+- 1518.855 usec)
>     Average num. events: 963.000 (+- 0.000)                                  
> Average num. events: 1018.000 (+- 0.000)
>     Average time per event 2.462 usec                                        
> Average time per event 5.873 usec
>   Number of synthesis threads: 40                                          
> Number of synthesis threads: 40
>     Average synthesis took: 2489.400 usec (+- 49.832 usec)                   
> Average synthesis took: 4576.500 usec (+- 75.278 usec)
>     Average num. events: 956.800 (+- 6.721)                                  
> Average num. events: 1020.000 (+- 0.000)
>     Average time per event 2.602 usec                                        
> Average time per event 4.487 usec
>