Re: [PATCH] perf_events: add sampling period randomization support (v2)

[Ingo Molnar <mingo@elte.hu>]

* Stephane Eranian <eranian@google.com> wrote:

> On Thu, Mar 4, 2010 at 1:13 PM, Ingo Molnar <mingo@elte.hu> wrote:
> >
> > * Stephane Eranian <eranian@google.com> wrote:
> >
> >> On Thu, Mar 4, 2010 at 3:32 AM, Ingo Molnar <mingo@elte.hu> wrote:
> >> >
> >> > * eranian@google.com <eranian@google.com> wrote:
> >> >
> >> >> This patch adds support for randomizing the sampling period. ??Randomization
> >> >> is very useful to mitigate the bias that exists with sampling. The random
> >> >> number generator does not need to be sophisticated. This patch uses the
> >> >> builtin random32() generator.
> >> >
> >> >> + ?? ?? if (width > 63 || attr->freq)
> >> >> + ?? ?? ?? ?? ?? ?? return -EINVAL;
> >> >
> >> > Why not for freq counters? Those are semi-randomized already, but it might
> >> > make sense to make them 'more' randomized in special circumstances. That would
> >> > also allow us to enable the randomization in perf top and perf record, by
> >> > default.
> >> >
> >>
> >> What's the goal of freq?
> >> Achieve and maintain the target interrupt/rate.
> >> In doing so, it has to adjust the period (not randomly).
> >
> > No, the goal of auto-freq is to keep a steady average rate of sampling.
>
> rate of samples = rate of interrupts (if period < 32 bits on Intel).

What's your point? I corrected your statement which said that the goal of 
auto-freq was to maintain a target interrupt-rate and as such wouldnt be 
randomizable. So i said that auto-freq is slightly different from that: it 
provides a steady _average_ rate, and as such small amounts of randomization 
'fuzz' could still be injected - the auto-freq system would auto-correct the 
effects of that.

Think of it as a dynamic steady-state equilibrium with noise injected. If the 
noise isnt too brutal and the system can adapt, the average sampling rate 
doesnt change.

> > There is no requirement to keep it 'steady' - each sample comes with a 
> > specific weight.
> >
> >> Randomization may prevent achieving the rate, or it may slow it down. 
> >> What's the value add of that?
> >
> > Why do you assume that the two are incompatible? We can randomize 
> > auto-freq and still have a perfectly stable average rate.
>
> What would that buy you compared to what you already have?

The same goal as randomization in general: to decrease the chance for sampling 
artifacts that can occur due to the sampling frequency oscillating together 
with some internal workload parameter, skewing the sample.

	Ingo