Re: [generalized cache events] Re: [PATCH 1/1] perf tools: Addmissing user space support for config1/config2

[Stephane Eranian]

On Sat, Apr 23, 2011 at 3:16 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Sat, 2011-04-23 at 14:06 +0200, Stephane Eranian wrote:
>> On Sat, Apr 23, 2011 at 9:50 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>> > On Fri, 2011-04-22 at 17:03 -0700, Andi Kleen wrote:
>> >> > > Yes, and note that with instructions events we even have skid-less PEBS
>> >> > > profiling so seeing the precise .
>> >> > Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â - location of instructions is possible.
>> >>
>> >> It was better when it was eaten. PEBS does not actually eliminated
>> >> skid unfortunately. The interrupt still occurs later, so the
>> >> instruction location is off.
>> >>
>> >> PEBS merely gives you more information.
>> >
>> > You're so skilled at not actually saying anything useful. Are you
>> > perchance referring to the fact that the IP reported in the PEBS data is
>> > exactly _one_ instruction off? Something that is demonstrated to be
>> > fixable?
>> >
>> > Or are you defining skid differently and not telling us your definition?
>> >
>>
>> PEBS is guaranteed to return an IP that is just after AN instruction that
>> caused the event. However, that instruction is NOT the one at the end
>> of your period. Let's take an example with INST_RETIRED, period=100000.
>> Then, the IP you get is NOT after the 100,000th retired instruction. It's an
>> instruction that is N cycles after that one. There is internal skid due to the
>> way PEBS is implemented.
>>
>> That is what Andi is referring to. The issue causes bias and thus impacts
>> the quality of the samples. On SNB, there is a new INST_RETIRED:PREC_DIST
>> event. PREC_DIST=precise distribution. It tries to correct for the skid
>> on this event on INST_RETIRED with PEBS (look at Vol3b).
>
> Sure, but who cares? So your period isn't exactly what you specified,
> but the effective period will have an average and a fairly small stdev
> (assuming the initial period is much larger than the relatively few
> cycles it takes to arm the PEBS assist), therefore you still get a
> fairly uniform spread.
>
> I don't much get the obsession with precision here, its all a statistics
> game anyway.
>

The particular example I am thinking about came from compiler people
I work with who would like to use PEBS to do statistical basic block profiling.
They do care about correctness of the profile. Otherwise, it may cause wrong
attribution of "hotness" of basic blocks and mislead the compiler when it tries
to reorder blocks on the critical path. Compiler people can validate a
statistical
profile because they have a reference profile obtained via instrumentation of
each basic block.

> And while you keep saying the examples are too trivial and Andi keep
> sprouting vague non-statements, neither of you actually provide anything
> sensible to the discussion.
>
> So stop f*cking whining and start talking sense or stop talking all
> together.
>
> I mean, you were in the room where Intel presented their research on
> event correlations based on pathological micro-benches. That clearly
> shows that exact event definitions simply don't matter.
>
Yes, and I don't get the same reading of the presentation. He never mentioned
generic events. He never even used them, I mean the Intel generic events.
Instead he used very focused Atom-specific events.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/