[Cbe-oss-dev] oprofile question
Carl E. Love
cel at us.ibm.com
Thu Jul 29 06:11:25 EST 2010
On Tue, 2010-07-27 at 11:31 +0900, Yasuhiko Yokote wrote:
> Hi!
>
> I have a question on oprofile use on Cell/QS22. Here is an example of opcontrol --list-events;
> Group 41 :dual_instrctn_commit: (counter: 0)
> Dual instruction committed. (min count: 10000)
> Unit masks (default 0x113)
> ----------
> 0x01: Count cycles [mandatory]
> 0x00: Negative polarity [optional ]
> 0x02: Positive polarity [default ]
> 0x110: SPU Bus Word 0 [default ]
> 0x140: SPU Bus Word 2 [optional ]
> 0x00: SPU 0 [default ]
> 0x1000: SPU 1 [optional ]
> 0x2000: SPU 2 [optional ]
> 0x3000: SPU 3 [optional ]
> 0x4000: SPU 4 [optional ]
> 0x5000: SPU 5 [optional ]
> 0x6000: SPU 6 [optional ]
> 0x7000: SPU 7 [optional ]
> Group 41 :sngl_instrctn_commit: (counter: 0)
> Single instruction committed. (min count: 10000)
> Unit masks (default 0x113)
> ----------
> 0x01: Count cycles [mandatory]
> 0x00: Negative polarity [optional ]
> 0x02: Positive polarity [default ]
> 0x110: SPU Bus Word 0 [default ]
> 0x140: SPU Bus Word 2 [optional ]
> 0x00: SPU 0 [default ]
> 0x1000: SPU 1 [optional ]
> 0x2000: SPU 2 [optional ]
> 0x3000: SPU 3 [optional ]
> 0x4000: SPU 4 [optional ]
> 0x5000: SPU 5 [optional ]
> 0x6000: SPU 6 [optional ]
> 0x7000: SPU 7 [optional ]
> In this list, there is an info on "min count: 10000." But I cannot set it to the oprofiled because it's value is lower than the system's default... So, what is appropriate value for the count?
The smaller you make the count value the more samples you collect. The
count value is how many events between samples. Each sample is
collected by an generating an interrupt. The idea is that the min count
value should be high enough that the system will not hang up trying to
service too many interrupts. The setting of the min is a bit of a guess
as we are not sure since the event occurrence is workload dependent. I
thought you should be able to set the value to the minimum, i.e.
opcontrol --event=sngl_instrctn_commit:10000. But perhaps there is a
bug and where the test is greater then instead of greater then or equal
to. Did you try a minimum of 10001? If that minimum really is too high
for your application, there is a way to change it to a lower value.
This is not really recommended unless you really know what your overhead
is.
Carl Love
>
> Thank you & regards,
More information about the cbe-oss-dev
mailing list