Re: Weird instruction profiling

[Brendan Gregg <brendan.d.gregg@gmail.com>]

On Wed, Feb 18, 2015 at 3:46 PM, David Ahern <dsahern@gmail.com> wrote:
> On 2/18/15 4:38 PM, Andi Kleen wrote:
>>
>> On Wed, Feb 18, 2015 at 04:33:31PM -0700, David Ahern wrote:
>>>
>>> On 2/18/15 11:30 AM, Andi Kleen wrote:
>>>
>>>>> I think Andi mentioned this to me last year -- that instruction
>>>>> profiling was no longer reliable.
>>>>
>>>>
>>>> It never was.
>>>
>>>
>>> Why is that? What about profiling with other hardware counters?
>>
>>
>> Practically all profiling has skid of some form.
>
>
> ok, so you meant that only in the context of skid meaning the hit
> instruction is off a few (ie., squint backwards at the annotate output).
>

Thanks Andi and David,

There's skid, but I was noticing what I thought was a different
artifact in the profile: where only every 4th instruction is seen.
Wouldn't this effect be more due to the IP pointing to the next
instruction while a group of 4 are executed in parallel? Then there's
skid on top of that....

Brendan