[Question] How to sample PERF_TYPE_HARDWARE

[Question] How to sample PERF_TYPE_HARDWARE

[rainkin]

Hello~

My goal is to record the number of occurring PERF_TYPE_HARDWARE events
(e.g., branches) every N events.

In my understanding, the "perf stat -e branches -I N " command is not
sampling, it is actually count model which records branches every N
millisecond. Thus perf stat does not meet my goal.

For the "perf record -e cycles -b" command, it does sampling, which
records a LBR sample every N cycles events. However, it seems that
"perf record" does not support sampling PERF_TYPE_HARDWARE.

I also look at the syscall perf_event_open, but still cannot find any
documentation about how to sample PERF_TYPE_HARDWARE.

Does anyone know how to do that？Any suggestions is helpful for me :)

Rainkin

Re: [Question] How to sample PERF_TYPE_HARDWARE

[Namhyung Kim]

Hello,

On Mon, Feb 8, 2021 at 12:50 PM rainkin <rainkin1993@gmail.com> wrote:
>
> Hello~
>
> My goal is to record the number of occurring PERF_TYPE_HARDWARE events
> (e.g., branches) every N events.
>
> In my understanding, the "perf stat -e branches -I N " command is not
> sampling, it is actually count model which records branches every N
> millisecond. Thus perf stat does not meet my goal.
>
> For the "perf record -e cycles -b" command, it does sampling, which
> records a LBR sample every N cycles events. However, it seems that
> "perf record" does not support sampling PERF_TYPE_HARDWARE.

The 'cycles' is a PERF_TYPE_HARDWARE event.

>
> I also look at the syscall perf_event_open, but still cannot find any
> documentation about how to sample PERF_TYPE_HARDWARE.
>
> Does anyone know how to do that？Any suggestions is helpful for me :)

  perf record -e branches -c N

?

Thanks,
Namhyung

Re: [Question] How to sample PERF_TYPE_HARDWARE

[rainkin]

Hi Namhyung,
Thanks for your reply.
Let me make my goal more clear:
record the number of occurring branches instructions
(PERF_COUNT_HW_BRANCH_INSTRUCTIONS) every N cycles
(PERF_COUNT_HW_CPU_CYCLES).

>   perf record -e branches -c N
I use strace to capture the syscall invoked by the above command:

perf_event_open({type=PERF_TYPE_HARDWARE, size=PERF_ATTR_SIZE_VER6,
config=PERF_COUNT_HW_BRANCH_INSTRUCTIONS, sample_period=100000,
sample_type=PERF_SAMPLE_IP|PERF_SAMPLE_TID|PERF_SAMPLE_TIME|PERF_SAMPLE_ID|PERF_SAMPLE_CPU,
read_format=PERF_FORMAT_ID, disabled=1, inherit=1, pinned=0,
exclusive=0, exclusive_user=0, exclude_kernel=0, exclude_hv=0,
exclude_idle=0, mmap=0, comm=0, freq=0, inherit_stat=0,
enable_on_exec=0, task=0, watermark=0, precise_ip=0 /* arbitrary skid
*/, mmap_data=0, sample_id_all=1, exclude_host=0, exclude_guest=1,
exclude_callchain_kernel=0, exclude_callchain_user=0, mmap2=0,
comm_exec=0, use_clockid=0, context_switch=0, write_backward=0,
namespaces=0, wakeup_events=0, config1=0, config2=0,
sample_regs_user=0, sample_regs_intr=0, aux_watermark=0,
sample_max_stack=0, aux_sample_size=0}, -1, 10, -1,
PERF_FLAG_FD_CLOEXEC) = 15

In my understanding ,WHAT this command actually does is recording
IP,TID, and so on
(sample_type=PERF_SAMPLE_IP|PERF_SAMPLE_TID|PERF_SAMPLE_TIME|PERF_SAMPLE_ID|PERF_SAMPLE_CPU)
every N (here N = 100000) branches
(config=PERF_COUNT_HW_BRANCH_INSTRUCTIONS). Thus this cannot meet my
goal.

Thanks,
Rainkin


Namhyung Kim <namhyung@gmail.com> 于2021年2月16日周二 下午3:50写道：
>
> Hello,
>
> On Mon, Feb 8, 2021 at 12:50 PM rainkin <rainkin1993@gmail.com> wrote:
> >
> > Hello~
> >
> > My goal is to record the number of occurring PERF_TYPE_HARDWARE events
> > (e.g., branches) every N events.
> >
> > In my understanding, the "perf stat -e branches -I N " command is not
> > sampling, it is actually count model which records branches every N
> > millisecond. Thus perf stat does not meet my goal.
> >
> > For the "perf record -e cycles -b" command, it does sampling, which
> > records a LBR sample every N cycles events. However, it seems that
> > "perf record" does not support sampling PERF_TYPE_HARDWARE.
>
> The 'cycles' is a PERF_TYPE_HARDWARE event.
>
> >
> > I also look at the syscall perf_event_open, but still cannot find any
> > documentation about how to sample PERF_TYPE_HARDWARE.
> >
> > Does anyone know how to do that？Any suggestions is helpful for me :)
>
>   perf record -e branches -c N
>
> ?
>
> Thanks,
> Namhyung

Re: [Question] How to sample PERF_TYPE_HARDWARE

[rainkin]

Hi Namhyung,
Thanks for your reply.
Let me make my goal more clear: record the number of occurring branch
instructions (PERF_COUNT_HW_BRANCH_INSTRUCTIONS) every N cycles
(PERF_COUNT_HW_CPU_CYCLES).

For the command you mentioned : perf record -e branches -c N

I use strace to capture the syscall invoked by the above command:

perf_event_open({type=PERF_TYPE_HARDWARE, size=PERF_ATTR_SIZE_VER6,
config=PERF_COUNT_HW_BRANCH_INSTRUCTIONS, sample_period=100000,
sample_type=PERF_SAMPLE_IP|PERF_SAMPLE_TID|PERF_SAMPLE_TIME|PERF_SAMPLE_ID|PERF_SAMPLE_CPU,
read_format=PERF_FORMAT_ID, disabled=1, inherit=1, pinned=0,
exclusive=0, exclusive_user=0, exclude_kernel=0, exclude_hv=0,
exclude_idle=0, mmap=0, comm=0, freq=0, inherit_stat=0,
enable_on_exec=0, task=0, watermark=0, precise_ip=0 /* arbitrary skid
*/, mmap_data=0, sample_id_all=1, exclude_host=0, exclude_guest=1,
exclude_callchain_kernel=0, exclude_callchain_user=0, mmap2=0,
comm_exec=0, use_clockid=0, context_switch=0, write_backward=0,
namespaces=0, wakeup_events=0, config1=0, config2=0,
sample_regs_user=0, sample_regs_intr=0, aux_watermark=0,
sample_max_stack=0, aux_sample_size=0}, -1, 10, -1,
PERF_FLAG_FD_CLOEXEC)

In my understanding ,WHAT this command actually does is recording
IP,TID, and so on
(sample_type=PERF_SAMPLE_IP|PERF_SAMPLE_TID|PERF_SAMPLE_TIME|PERF_SAMPLE_ID|PERF_SAMPLE_CPU)
every N (here N = 100000) branches
(config=PERF_COUNT_HW_BRANCH_INSTRUCTIONS). Thus this cannot meet my
goal.

Thanks,
Rainkin

Re: [Question] How to sample PERF_TYPE_HARDWARE

[Namhyung Kim]

On Thu, Feb 18, 2021 at 11:05 AM rainkin <rainkin1993@gmail.com> wrote:
>
> Hi Namhyung,
> Thanks for your reply.
> Let me make my goal more clear: record the number of occurring branch
> instructions (PERF_COUNT_HW_BRANCH_INSTRUCTIONS) every N cycles
> (PERF_COUNT_HW_CPU_CYCLES).

Please try this:

  perf record -c N -e '{cycles,branches}:S'
  perf script

Thanks,
Namhyung