Re: [PATCH v2 0/3] perf-stat: share hardware PMCs with BPF

[Song Liu]

> On Mar 17, 2021, at 9:32 PM, Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
> 
> On Thu, Mar 18, 2021 at 12:52 PM Song Liu <songliubraving@xxxxxx> wrote:
>> 
>> 
>> 
>>> On Mar 17, 2021, at 6:11 AM, Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:
>>> 
>>> Em Wed, Mar 17, 2021 at 02:29:28PM +0900, Namhyung Kim escreveu:
>>>> Hi Song,
>>>> 
>>>> On Wed, Mar 17, 2021 at 6:18 AM Song Liu <songliubraving@xxxxxx> wrote:
>>>>> 
>>>>> perf uses performance monitoring counters (PMCs) to monitor system
>>>>> performance. The PMCs are limited hardware resources. For example,
>>>>> Intel CPUs have 3x fixed PMCs and 4x programmable PMCs per cpu.
>>>>> 
>>>>> Modern data center systems use these PMCs in many different ways:
>>>>> system level monitoring, (maybe nested) container level monitoring, per
>>>>> process monitoring, profiling (in sample mode), etc. In some cases,
>>>>> there are more active perf_events than available hardware PMCs. To allow
>>>>> all perf_events to have a chance to run, it is necessary to do expensive
>>>>> time multiplexing of events.
>>>>> 
>>>>> On the other hand, many monitoring tools count the common metrics (cycles,
>>>>> instructions). It is a waste to have multiple tools create multiple
>>>>> perf_events of "cycles" and occupy multiple PMCs.
>>>> 
>>>> Right, it'd be really helpful when the PMCs are frequently or mostly shared.
>>>> But it'd also increase the overhead for uncontended cases as BPF programs
>>>> need to run on every context switch.  Depending on the workload, it may
>>>> cause a non-negligible performance impact.  So users should be aware of it.
>>> 
>>> Would be interesting to, humm, measure both cases to have a firm number
>>> of the impact, how many instructions are added when sharing using
>>> --bpf-counters?
>>> 
>>> I.e. compare the "expensive time multiplexing of events" with its
>>> avoidance by using --bpf-counters.
>>> 
>>> Song, have you perfmormed such measurements?
>> 
>> I have got some measurements with perf-bench-sched-messaging:
>> 
>> The system: x86_64 with 23 cores (46 HT)
>> 
>> The perf-stat command:
>> perf stat -e cycles,cycles,instructions,instructions,ref-cycles,ref-cycles <target, etc.>
>> 
>> The benchmark command and output:
>> ./perf bench sched messaging -g 40 -l 50000 -t
>> # Running 'sched/messaging' benchmark:
>> # 20 sender and receiver threads per group
>> # 40 groups == 1600 threads run
>>     Total time: 10X.XXX [sec]
>> 
>> 
>> I use the "Total time" as measurement, so smaller number is better.
>> 
>> For each condition, I run the command 5 times, and took the median of
>> "Total time".
>> 
>> Baseline (no perf-stat)                 104.873 [sec]
>> # global
>> perf stat -a                            107.887 [sec]
>> perf stat -a --bpf-counters             106.071 [sec]
>> # per task
>> perf stat                               106.314 [sec]
>> perf stat --bpf-counters                105.965 [sec]
>> # per cpu
>> perf stat -C 1,3,5                      107.063 [sec]
>> perf stat -C 1,3,5 --bpf-counters       106.406 [sec]
>> 
>> From the data, --bpf-counters is slightly better than the regular event
>> for all targets. I noticed that the results are not very stable. There
>> are a couple 108.xx runs in some of the conditions (w/ and w/o
>> --bpf-counters).
> 
> Hmm.. so this result is when multiplexing happened, right?
> I wondered how/why the regular perf stat is slower..

I should have made this more clear. This is when regular perf-stat time 
multiplexing (2x ref-cycles on Intel). OTOH, bpf-counters does enables 
sharing, so there is no time multiplexing. IOW, this is overhead of BPF 
vs. overhead of time multiplexing. 

Thanks,
Song