Use case for PERF_COUNT_HW_REF_CPU_CYCLES generic PMU event

Use case for PERF_COUNT_HW_REF_CPU_CYCLES generic PMU event

[Anshuman Khandual]

Hello Stephane, 

I was going through the following discussion where we added the
new HW generic event PERF_COUNT_HW_REF_CPU_CYCLES.

https://lkml.org/lkml/2011/12/10/103 

(Sorry, for asking this question bit late)

I am trying to understand the use case for this. Would this new event
help us in generating (during a perf session) a CPU frequency invariant
time metric against which we would plot our other perf event's measurements ?
CPU frequency independent time measurement is it's primary purpose ? or we were
finding a way to expose the fixed counter 2 which was not getting used before 
for not having an event encoding. I guess this would help us in finding equivalent
PMU events or mechanisms in other architecture / platforms.

-- 
Anshuman Khandual
Linux Technology Centre
IBM Systems and Technology Group

Re: Use case for PERF_COUNT_HW_REF_CPU_CYCLES generic PMU event

[Stephane Eranian]

On Wed, Feb 8, 2012 at 10:43 AM, Anshuman Khandual
<khandual@linux.vnet.ibm.com> wrote:
> Hello Stephane,
>
> I was going through the following discussion where we added the
> new HW generic event PERF_COUNT_HW_REF_CPU_CYCLES.
>
> https://lkml.org/lkml/2011/12/10/103
>
> (Sorry, for asking this question bit late)
>
> I am trying to understand the use case for this. Would this new event
> help us in generating (during a perf session) a CPU frequency invariant
> time metric against which we would plot our other perf event's measurements ?
> CPU frequency independent time measurement is it's primary purpose ? or we were
> finding a way to expose the fixed counter 2 which was not getting used before
> for not having an event encoding. I guess this would help us in finding equivalent
> PMU events or mechanisms in other architecture / platforms.
>
The goal was to expose a cycle event that is not subject to frequency scaling
nor turbo boost of any sort. An event that could be used to correlate with time.
An event that could also be used to compute idle time by comparing its value
with wall-clock time.

The fact that on Intel X86 this event is on fixed counter 2 is an
implementation
detail.

> --
> Anshuman Khandual
> Linux Technology Centre
> IBM Systems and Technology Group
>

Re: Use case for PERF_COUNT_HW_REF_CPU_CYCLES generic PMU event

[Anshuman Khandual]

On Wednesday 08 February 2012 03:21 PM, Stephane Eranian wrote:
> On Wed, Feb 8, 2012 at 10:43 AM, Anshuman Khandual
> <khandual@linux.vnet.ibm.com> wrote:
>> Hello Stephane,
>>
>> I was going through the following discussion where we added the
>> new HW generic event PERF_COUNT_HW_REF_CPU_CYCLES.
>>
>> https://lkml.org/lkml/2011/12/10/103
>>
>> (Sorry, for asking this question bit late)
>>
>> I am trying to understand the use case for this. Would this new event
>> help us in generating (during a perf session) a CPU frequency invariant
>> time metric against which we would plot our other perf event's measurements ?
>> CPU frequency independent time measurement is it's primary purpose ? or we were
>> finding a way to expose the fixed counter 2 which was not getting used before
>> for not having an event encoding. I guess this would help us in finding equivalent
>> PMU events or mechanisms in other architecture / platforms.
>>
> The goal was to expose a cycle event that is not subject to frequency scaling
> nor turbo boost of any sort. An event that could be used to correlate with time.

> An event that could also be used to compute idle time by comparing its value
> with wall-clock time.

Why kernel computed idle time is not sufficient ? How much accuracy would it
improve in using PMU event computed idle time over kernel computed idle time. 


> 
> The fact that on Intel X86 this event is on fixed counter 2 is an
> implementation
> detail.
> 
>> --
>> Anshuman Khandual
>> Linux Technology Centre
>> IBM Systems and Technology Group
>>
> 

Re: Use case for PERF_COUNT_HW_REF_CPU_CYCLES generic PMU event

[Stephane Eranian]

On Wed, Feb 8, 2012 at 11:51 AM, Anshuman Khandual
<khandual@linux.vnet.ibm.com> wrote:
> On Wednesday 08 February 2012 03:21 PM, Stephane Eranian wrote:
>> On Wed, Feb 8, 2012 at 10:43 AM, Anshuman Khandual
>> <khandual@linux.vnet.ibm.com> wrote:
>>> Hello Stephane,
>>>
>>> I was going through the following discussion where we added the
>>> new HW generic event PERF_COUNT_HW_REF_CPU_CYCLES.
>>>
>>> https://lkml.org/lkml/2011/12/10/103
>>>
>>> (Sorry, for asking this question bit late)
>>>
>>> I am trying to understand the use case for this. Would this new event
>>> help us in generating (during a perf session) a CPU frequency invariant
>>> time metric against which we would plot our other perf event's measurements ?
>>> CPU frequency independent time measurement is it's primary purpose ? or we were
>>> finding a way to expose the fixed counter 2 which was not getting used before
>>> for not having an event encoding. I guess this would help us in finding equivalent
>>> PMU events or mechanisms in other architecture / platforms.
>>>
>> The goal was to expose a cycle event that is not subject to frequency scaling
>> nor turbo boost of any sort. An event that could be used to correlate with time.
>
>> An event that could also be used to compute idle time by comparing its value
>> with wall-clock time.
>
> Why kernel computed idle time is not sufficient ? How much accuracy would it
> improve in using PMU event computed idle time over kernel computed idle time.
>
I was just listing the what you could do with the event. I have not
yet tried to measure
idle with ref-cycles.

>
>>
>> The fact that on Intel X86 this event is on fixed counter 2 is an
>> implementation
>> detail.
>>
>>> --
>>> Anshuman Khandual
>>> Linux Technology Centre
>>> IBM Systems and Technology Group
>>>
>>
>