delayacct regression in 2.6.24 git

delayacct regression in 2.6.24 git

[Balbir Singh]

Hi, Ingo,

I missed commit 75d4ef16a6aa84f708188bada182315f80aab6fa
"sched: fix delay accounting performance regression"

Without these changes, I find that delay accounting does
not work as expected, I see zeros in the field for virtual,delay
and count(s).

CPU             count     real total  virtual total    delay total
                    0      535861848              0              0


This breaks existing delay accounting behaviour.

Am I missing something?

-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

Re: delayacct regression in 2.6.24 git

[Ingo Molnar]

* Balbir Singh <balbir@linux.vnet.ibm.com> wrote:

> Hi, Ingo,
> 
> I missed commit 75d4ef16a6aa84f708188bada182315f80aab6fa
> "sched: fix delay accounting performance regression"
> 
> Without these changes, I find that delay accounting does
> not work as expected, I see zeros in the field for virtual,delay
> and count(s).
> 
> CPU             count     real total  virtual total    delay total
>                     0      535861848              0              0
> 
> 
> This breaks existing delay accounting behaviour.
> 
> Am I missing something?

delay accounting was causing context-switch slowdown because it called 
sched_clock() explicitly - but indeed my fix to that is wrong ...

how about reverting the change but also using rq->clock instead of 
sched_clock()?

	Ingo

Re: delayacct regression in 2.6.24 git

[Balbir Singh]

Ingo Molnar wrote:
> * Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
> 
>> Hi, Ingo,
>>
>> I missed commit 75d4ef16a6aa84f708188bada182315f80aab6fa
>> "sched: fix delay accounting performance regression"
>>
>> Without these changes, I find that delay accounting does
>> not work as expected, I see zeros in the field for virtual,delay
>> and count(s).
>>
>> CPU             count     real total  virtual total    delay total
>>                     0      535861848              0              0
>>
>>
>> This breaks existing delay accounting behaviour.
>>
>> Am I missing something?
> 
> delay accounting was causing context-switch slowdown because it called 
> sched_clock() explicitly - but indeed my fix to that is wrong ...
> 
> how about reverting the change but also using rq->clock instead of 
> sched_clock()?

Hmm... I think that might work, I find that schedule() calls
__update_rq_clock() which in turn calls sched_clock(), so it
might be a worthwhile optimization.

-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

Re: delayacct regression in 2.6.24 git

[Ingo Molnar]

* Balbir Singh <balbir@linux.vnet.ibm.com> wrote:

> >> Am I missing something?
> > 
> > delay accounting was causing context-switch slowdown because it 
> > called sched_clock() explicitly - but indeed my fix to that is wrong 
> > ...
> > 
> > how about reverting the change but also using rq->clock instead of 
> > sched_clock()?
> 
> Hmm... I think that might work, I find that schedule() calls 
> __update_rq_clock() which in turn calls sched_clock(), so it might be 
> a worthwhile optimization.

yep, that's the idea - in general rq->clock is updated for almost every 
scheduling event so you can rely on it instead of calling sched_clock() 
directly.

	Ingo