Aah, my bad. I didn't realize that I need to set both intx and intxcp flags
to get events within transactions - events within aborted transactions,
which is what I was aiming for. I should have used
UNHALTED_CORE_CYCLES:intx:intxcp instead of
UNHALTED_CORE_CYCLES:intxcp. Thanks
for pointing that out.
On Tue, Aug 20, 2013 at 6:51 PM, Stephane Eranian <eran...@googlemail.com>wrote:
>
>
>
> On Tue, Aug 20, 2013 at 6:34 PM, Bhavishya Goel
> <bhavishya.g...@gmail.com>wrote:
>
>> With the kernel version 3.11-rc6, I am able to get the intx attribute
>> (events within transactions) working. But there seems to be some problem
>> with intxcp attribute (events within transactions - events within aborted
>> transactions). This is what I get now:
>>
>> task -i -e
>> UNHALTED_CORE_CYCLES,UNHALTED_CORE_CYCLES:intx,UNHALTED_CORE_CYCLES:intxcp
>> ./cmd
>>
>> 22 020 143 037 UNHALTED_CORE_CYCLES (0,00% scaling, ena=9 670 910
>> 328, run=9 670 910 328)
>> 14 974 191 218 UNHALTED_CORE_CYCLES:intx (0,00% scaling, ena=9 670
>> 910 328, run=9 670 910 328)
>> 19 240 174 502 UNHALTED_CORE_CYCLES:intxcp (0,00% scaling, ena=9
>> 670 910 328, run=9 670 910 328)
>>
>> Note that the value for UNHALTED_CORE_CYCLES:intxcp is greater than for
>> UNHALTED_CORE_CYCLES:intx which as per my understanding is not possible.
>> The UNHALTED_CORE_CYCLES and UNHALTED_CORE_CYCLES:intx values match what I
>> get from Intel's PCM-TSX tool.
>>
>>
>> intxcp means: do not count cycles from aborted transactions. So what
> happens on an aborted transaction
> is that the count is rolled back to what it was when entering the
> transaction. As such, it cannot be
> greater than unhalted_core_cycles but it can be greater than
> unhalted_core_cycles:intx.
>
>
>
>>
>>
>> On Tue, Aug 20, 2013 at 11:04 AM, Stephane Eranian <
>> eran...@googlemail.com> wrote:
>>
>>>
>>>
>>>
>>> On Tue, Aug 20, 2013 at 11:02 AM, Bhavishya Goel <
>>> bhavishya.g...@gmail.com> wrote:
>>>
>>>> Oh, I am using version 3.10.0. I will try with version 3.11-rc6.
>>>> Thanks. :)
>>>>
>>>>
>>>> 3.10 is not sufficient. HSW support went in 3.11
>>>
>>>
>>>> On Tue, Aug 20, 2013 at 10:41 AM, Stephane Eranian <
>>>> eran...@googlemail.com> wrote:
>>>>
>>>>>
>>>>>
>>>>>
>>>>> On Tue, Aug 20, 2013 at 9:56 AM, Bhavishya Goel <
>>>>> bhavishya.g...@gmail.com> wrote:
>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> I am trying to use the TSX event modifiers for libpfm4 on Intel
>>>>>> Haswell microarchitecture. But it doesn't seem to be working for me. The
>>>>>> counts that I see for performance events are the same with or without the
>>>>>> intx/intxcp attibute. Following is an example of command that I run:
>>>>>>
>>>>>> task -i -e UNHALTED_CORE_CYCLES:intx ./tsx_example
>>>>>>
>>>>>> The count for UNHALTED_CORE_CYCLES remains the same with and without
>>>>>> intx attribute. The intxcp attribute also doesn't make a difference.
>>>>>> Intel's PCM-TSX tool for the same code shows that cycles spent within
>>>>>> transaction are only 56% of the total cycles.
>>>>>>
>>>>>> And what kernel version are you using?
>>>>> You need kernel support to get Haswell support. In fact, you need
>>>>> 3.11.0-rc5 or so.
>>>>> With any older kernels, the tsx modifiers will be ignored.
>>>>>
>>>>>
>>>>
>>>>
>>>> --
>>>> ಠ_ಠ
>>>>
>>>>
>>>> ------------------------------------------------------------------------------
>>>> Introducing Performance Central, a new site from SourceForge and
>>>> AppDynamics. Performance Central is your source for news, insights,
>>>> analysis and resources for efficient Application Performance Management.
>>>> Visit us today!
>>>>
>>>> http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
>>>> _______________________________________________
>>>> perfmon2-devel mailing list
>>>> perfmon2-devel@lists.sourceforge.net
>>>> https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
>>>>
>>>>
>>>
>>
>>
>> --
>> ಠ_ಠ
>>
>>
>> ------------------------------------------------------------------------------
>> Introducing Performance Central, a new site from SourceForge and
>> AppDynamics. Performance Central is your source for news, insights,
>> analysis and resources for efficient Application Performance Management.
>> Visit us today!
>>
>> http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
>> _______________________________________________
>> perfmon2-devel mailing list
>> perfmon2-devel@lists.sourceforge.net
>> https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
>>
>>
>
--
ಠ_ಠ
------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and
AppDynamics. Performance Central is your source for news, insights,
analysis and resources for efficient Application Performance Management.
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
perfmon2-devel mailing list
perfmon2-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
