Li, Aubrey wrote:
> Rafael.Vanoni at Sun.COM wrote:
> 
>> Li, Aubrey wrote:
>>> Rafael.Vanoni wrote:
>>>
>>>>> These issues are replicable when you keep pressing "I" per ptop
>>>>> interval. Actually, when you close the cstate statistics, total
>>>>> event number is not valid for the subsequent report, and the whole
>>>>> report will be messed up.
>>>> Ok, I'll have a look at this tomorrow morning. But the problem is
>>>> not persistent, if you wait until all the keystrokes are processed,
>>>> the report goes back to normal. Not sure there's a way to avoid
>>>> this since it's the delay of stopping and restarting the DTrace
>>>> program 
>>>> every time
>>>> you freeze one of the windows.
>>>>
>>>> Rafael
>>> The problem is here, why freezing windows need to stop dtrace?
>>> When you stop cstate dtrace probe, the total event number is not
>>> valid any longer, the subsequent report will be broken.
>>>
>>> Just want to know, for what reason is this feature requested?
>>>
>>> -Aubrey
>> My motivation was/is to allow the user to lower the side effects of
>> having three DTrace programs running at the same time, while
>> still being
>> able to use some of the information ptop provides. So we need to stop
>> the scripts, and not just freeze the display.
>>
>> I've also been able to identify the c-state program as responsible for
>> the vast majority of the xcalls we see on mid-large systems, and I've
>> been trying to tune our DTrace parameters around it.
>>
>> What do you think ?
>>
>> Thanks,
>> Rafael
> 
> Your answer didn't address the problem above. When stop the cstate dtrace
> probe, the total event number is not valid any longer, the subsequent report 
> is broken.
> 
> And, for what reason we need to lower the side effects? When we are monitoring
> the system, the impact was expected.


On a dual socket, quad core Nehalem box:

--
Wakeups-from-idle per second: 3545.9    interval: 5.0s
no ACPI power usage estimate available

Top causes for wakeups:
91.4% (3242.1)                 sched :  <xcalls> unix`dtrace_xcall_func
  2.8% (100.0)               <kernel> :  genunix`clock
  2.2% ( 77.4)               <kernel> :  genunix`cv_wakeup
  0.8% ( 29.1)                  sched :  <xcalls> unix`hati_demap_func
--

~3k xcalls generated by PowerTOP. But if you turn the c-state report off:

--
Wakeups-from-idle per second: 335.0     interval: 5.0s
no ACPI power usage estimate available

Top causes for wakeups:
29.9% (100.0)               <kernel> :  genunix`clock
23.4% ( 78.3)               <kernel> :  genunix`cv_wakeup
21.4% ( 71.7)                  sched :  <xcalls> unix`dtrace_xcall_func
  7.0% ( 23.5)               <kernel> : 
uhci`uhci_handle_root_hub_status_change
--

We no longer have the ~3k xcalls, and the number of wakeups is down by 
that same amount. Is this what you're referring to ?

> Actually, I really like to separate this patch into two, and commit 7951 
> first.
> As for 7953, I think it's a big change for powertop, it would be better to me
> to putback after the new version release.

Ok, maybe we should ask someone else's opinion on this one, since we 
can't seem to agree on it :) It will take me more time to split the 
patch in two than to test and fix whatever bugs we find.

> We have several critical bug-fixes in ptop repo. How's the progress to putback
> into ON? I think we should prioritize the putback.

It's in progress, we just have to stop finding bugs in the fixes. 
Patches to both p and c-state reports should go into the current build, 
I'm not waiting for this patch.

Thanks,
Rafael




Reply via email to