[BUG] perf_event: inst_retired:any_p:c=16:i=1 is not equivalent to cpu_clk_unhalted:thread_p

[BUG] perf_event: inst_retired:any_p:c=16:i=1 is not equivalent to cpu_clk_unhalted:thread_p

[Stephane Eranian]

Peter,

commit 7639dae0ca11038286bbbcda05f2bef601c1eb8d
Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
Date:   Tue Dec 14 21:26:40 2010 +0100

    perf, x86: Provide a PEBS capable cycle event



Even with the transformation cmask=16:invert=1, you are not quite
counting the same thing.

inst_retired:any_p:c=16:i=1  is different from cpu_clk_unhalted:thread_p

Simply because, I think, this form of inst_retired counts in HALTED state.

This automatic transformation would be "okay" in per-thread mode, but
not in system-wide mode. It should not be done under the cover. I would
rather have this option at the user level.

Re: [BUG] perf_event: inst_retired:any_p:c=16:i=1 is not equivalent to cpu_clk_unhalted:thread_p

[Peter Zijlstra]

On Tue, 2011-01-18 at 19:05 +0100, Stephane Eranian wrote:
> Peter,
> 
> commit 7639dae0ca11038286bbbcda05f2bef601c1eb8d
> Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
> Date:   Tue Dec 14 21:26:40 2010 +0100
> 
>     perf, x86: Provide a PEBS capable cycle event
> 
> 
> 
> Even with the transformation cmask=16:invert=1, you are not quite
> counting the same thing.
> 
> inst_retired:any_p:c=16:i=1  is different from cpu_clk_unhalted:thread_p
> 
> Simply because, I think, this form of inst_retired counts in HALTED state.

Drad it does indeed:

# perf stat -a -e cycles:p -e cycles sleep 1

 Performance counter stats for 'sleep 1':

        61,634,870 cycles                  
        26,703,249 cycles                  

        1.000860820  seconds time elapsed


> This automatic transformation would be "okay" in per-thread mode, but
> not in system-wide mode. It should not be done under the cover. I would
> rather have this option at the user level.

Agreed, since its not identical it had best be done in userspace, too
bad :/ I'll revert this patch.

Re: [BUG] perf_event: inst_retired:any_p:c=16:i=1 is not equivalent to cpu_clk_unhalted:thread_p

[Stephane Eranian]

On Tue, Jan 18, 2011 at 7:36 PM, Peter Zijlstra <peterz@infradead.org> wrote:
> On Tue, 2011-01-18 at 19:05 +0100, Stephane Eranian wrote:
>> Peter,
>>
>> commit 7639dae0ca11038286bbbcda05f2bef601c1eb8d
>> Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
>> Date:   Tue Dec 14 21:26:40 2010 +0100
>>
>>     perf, x86: Provide a PEBS capable cycle event
>>
>>
>>
>> Even with the transformation cmask=16:invert=1, you are not quite
>> counting the same thing.
>>
>> inst_retired:any_p:c=16:i=1  is different from cpu_clk_unhalted:thread_p
>>
>> Simply because, I think, this form of inst_retired counts in HALTED state.
>
> Drad it does indeed:
>
> # perf stat -a -e cycles:p -e cycles sleep 1
>
>  Performance counter stats for 'sleep 1':
>
>        61,634,870 cycles
>        26,703,249 cycles
>
>        1.000860820  seconds time elapsed
>
What's interesting here is that this cycles:p is not event equal to 1x CPU freq
given your measuring for 1s. So it must stop counting at some point.

>
>> This automatic transformation would be "okay" in per-thread mode, but
>> not in system-wide mode. It should not be done under the cover. I would
>> rather have this option at the user level.
>
> Agreed, since its not identical it had best be done in userspace, too
> bad :/ I'll revert this patch.
>
Yes, it would have been nice. But if you know what you're doing, this is
still doable from user space. So I think we are okay.