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

[Stephane Eranian <eranian@google.com>]

On Mon, Apr 25, 2011 at 8:05 PM, Ingo Molnar <mingo@elte.hu> wrote:
>
> * Dehao Chen <danielcdh@gmail.com> wrote:
>
>> > ... and the resulting low level of noise in the average period length is
>> > what matters. The instruction itself will still be one of the hotspot
>> > instructions, statistically.
>>
>> Not true. This skid will lead to some aggregation and shadow effects on some
>> certain instructions. To make things worse, these effects are deterministic
>> and cannot be removed by either sampling for multiple times or by averaging
>> among instructions within a basic block. As a result, some actual "hot spot"
>> are not sampled at all. You can simply try to collect a basic block level
>> CPI, and you'll get a very misleading profile.
>
> This certainly does not match the results i'm seeing on real applications,
> using "-e instructions:pp" PEBS+LBR profiling. How do you explain that? Also,
> can you demonstrate your claim with a real example?
>

LBR removes the off-by-1 IP problem, it does not remove the shadow effect, i.e.,
that blind spot of N cycles caused by the PEBS arming mechanism.