Sampling hardware counter values with timers

Sampling hardware counter values with timers

[Keno Fischer]

Hello everyone,

I was wondering if it is possible (either using the `perf` user space
tool or using the perf_events API) to periodically sample the value of
a hardware counter (in my particular case I'd like to sample the
retired-branch-counter). I know it is possible to do the opposite
(i.e. sample a timestamp every specified number of branches), but I
wasn't able to see an easy way to sample the hardware counter at a
specified frequency. Is this possible?

Thanks,
Keno

Re: Sampling hardware counter values with timers

[Andi Kleen]

Keno Fischer <kfischer@college.harvard.edu> writes:

> Hello everyone,
>
> I was wondering if it is possible (either using the `perf` user space
> tool or using the perf_events API) to periodically sample the value of
> a hardware counter (in my particular case I'd like to sample the
> retired-branch-counter). I know it is possible to do the opposite
> (i.e. sample a timestamp every specified number of branches), but I
> wasn't able to see an easy way to sample the hardware counter at a
> specified frequency. Is this possible?

perf record works this way by default. The counter period is dynamically
adjusted for a given frequency. The default frequency is 1000.
You can set it with -F.

Another alternative is to use perf stat -I ... if you only need the counts.

-Andi

Re: Sampling hardware counter values with timers

[Keno Fischer]

Hi Andi,

Thanks. I know that perf sample works this way, but I didn't think it
was possible to have it sample a specified counter. Is it? `perf stat
-I` looks exactly like what I need, but it looks like the maximum
frequency is 10Hz. I was hoping for a higher frequency (& ideally
something that
doesn't require switching back into user space and calling `read`).

Thanks,
Keno

On Fri, May 13, 2016 at 9:40 AM, Andi Kleen <andi@firstfloor.org> wrote:
> Keno Fischer <kfischer@college.harvard.edu> writes:
>
>> Hello everyone,
>>
>> I was wondering if it is possible (either using the `perf` user space
>> tool or using the perf_events API) to periodically sample the value of
>> a hardware counter (in my particular case I'd like to sample the
>> retired-branch-counter). I know it is possible to do the opposite
>> (i.e. sample a timestamp every specified number of branches), but I
>> wasn't able to see an easy way to sample the hardware counter at a
>> specified frequency. Is this possible?
>
> perf record works this way by default. The counter period is dynamically
> adjusted for a given frequency. The default frequency is 1000.
> You can set it with -F.
>
> Another alternative is to use perf stat -I ... if you only need the counts.
>
> -Andi

Re: Sampling hardware counter values with timers

[Vince Weaver]

On Fri, 13 May 2016, Keno Fischer wrote:

> Thanks. I know that perf sample works this way, but I didn't think it
> was possible to have it sample a specified counter. Is it? `perf stat
> -I` looks exactly like what I need, but it looks like the maximum
> frequency is 10Hz. I was hoping for a higher frequency (& ideally
> something that
> doesn't require switching back into user space and calling `read`).

you can easily hack perf -I to sample faster than 10Hz, but if you sample 
much more often than that you're going to start ending up with a lot of 
perf_event measurement overhead in your data.

Vince

Re: Sampling hardware counter values with timers

[Keno Fischer]

Thanks Vince,

I figured I could do something like that, but I was indeed concerned
about the performance overhead, esp. since it seems like the way I'd
have to structure it is to have a user space timer interrupt which
then `read`s the perf counter. Since perf can sample e.g. the ip at
sufficiently high frequencies, I was hoping there was some mechanism
that would allow me to do the same for the retired branch counter.

For some background to this question, I'm hoping to use this in
combination with mozilla's record/replay framework (BTW, your work
on deterministic perf counters has come in super handy several times
here, Vince). rr identifies program points by combinations of
(retired-branch-counter, ip), so I was hoping to be able to correlate
those program points with actual time elapsed.

Keno

On Fri, May 13, 2016 at 1:42 PM, Vince Weaver <vincent.weaver@maine.edu> wrote:
> On Fri, 13 May 2016, Keno Fischer wrote:
>
>> Thanks. I know that perf sample works this way, but I didn't think it
>> was possible to have it sample a specified counter. Is it? `perf stat
>> -I` looks exactly like what I need, but it looks like the maximum
>> frequency is 10Hz. I was hoping for a higher frequency (& ideally
>> something that
>> doesn't require switching back into user space and calling `read`).
>
> you can easily hack perf -I to sample faster than 10Hz, but if you sample
> much more often than that you're going to start ending up with a lot of
> perf_event measurement overhead in your data.
>
> Vince