nanosleep() accuracy

nanosleep() accuracy

[GolovaSteek]

Hello!
I need use sleep with accurat timing.
I use 2.6.21 with rt-prempt patch.
with enabled rt_preempt, dyn_ticks, and local_apic
But

req.tv_nsec = 300000;
req.tv_sec = 0;
nanosleep(&req,NULL)

make pause around 310-330 microseconds.

I tried to understend how work nanosleep(), but it not depends from
jiffies and from smp_apic_timer_interrupt.

When can accuracy be lost?
And how are process waked up?


GolovaSteek

Re: nanosleep() accuracy

[Michal Schmidt]

GolovaSteek skrev:
> Hello!
> I need use sleep with accurat timing.
> I use 2.6.21 with rt-prempt patch.
> with enabled rt_preempt, dyn_ticks, and local_apic
> But
> 
> req.tv_nsec = 300000;
> req.tv_sec = 0;
> nanosleep(&req,NULL)
> 
> make pause around 310-330 microseconds.

How do you measure this?
If you want to have something done every 300 microseconds, you must not 
sleep for 300 microseconds in each iteration, because you'd accumulate 
errors. Use a periodic timer or use the current time to compute how long 
to sleep in each iteration. Take a look how cyclictest does it.

> I tried to understend how work nanosleep(), but it not depends from
> jiffies and from smp_apic_timer_interrupt.
> 
> When can accuracy be lost?
> And how are process waked up?
> 
> 
> GolovaSteek

Don't forget the process will always have non-zero wakeup latency. It 
takes some time to process an interrupt, wakeup the process and schedule 
it to run on the CPU. 10-30 microseconds is not unreasonable.

Michal

Re: nanosleep() accuracy

[GolovaSteek]

2007/8/17, Michal Schmidt <xschmi00@stud.feec.vutbr.cz>:
> GolovaSteek skrev:
> > Hello!
> > I need use sleep with accurat timing.
> > I use 2.6.21 with rt-prempt patch.
> > with enabled rt_preempt, dyn_ticks, and local_apic
> > But
> >
> > req.tv_nsec = 300000;
> > req.tv_sec = 0;
> > nanosleep(&req,NULL)
> >
> > make pause around 310-330 microseconds.
>
> How do you measure this?
> If you want to have something done every 300 microseconds, you must not
> sleep for 300 microseconds in each iteration, because you'd accumulate
> errors. Use a periodic timer or use the current time to compute how long
> to sleep in each iteration. Take a look how cyclictest does it.

no. I just want my programm go to sleep sometimes and wake up in correct time.

> > I tried to understend how work nanosleep(), but it not depends from
> > jiffies and from smp_apic_timer_interrupt.
> >
> > When can accuracy be lost?
> > And how are process waked up?
> >
> >
> > GolovaSteek
>
> Don't forget the process will always have non-zero wakeup latency. It
> takes some time to process an interrupt, wakeup the process and schedule
> it to run on the CPU. 10-30 microseconds is not unreasonable.

But 20000 operations can be done in 10 microseconds?
and why is there that inconstancy? Why sametimes 10 and sometimes 30?
In which points of implementation it happens?

GolovaSteek

Re: nanosleep() accuracy

[Jan Engelhardt]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 738 bytes --]


On Aug 17 2007 11:44, GolovaSteek wrote:
>> How do you measure this?
>> If you want to have something done every 300 microseconds, you must not
>> sleep for 300 microseconds in each iteration, because you'd accumulate
>> errors. Use a periodic timer or use the current time to compute how long
>> to sleep in each iteration. Take a look how cyclictest does it.
>
>no. I just want my programm go to sleep sometimes and wake up in correct time.

Would it be acceptable to use an optimistic strategy, like the one below?

Let's say that the following tasks happen at each time: A at 0, B at 300, C at
600, D at 900, E at 1200, F at 1500. Assume sleeping takes 500 µs.
Then B and C could be run at 500, D at 1000 and E,F at 1500.


	Jan
-- 

Re: nanosleep() accuracy

[Michal Schmidt]

GolovaSteek wrote:
> 2007/8/17, Michal Schmidt <xschmi00@stud.feec.vutbr.cz>:
>   
>> GolovaSteek skrev:
>>     
>>> Hello!
>>> I need use sleep with accurat timing.
>>> I use 2.6.21 with rt-prempt patch.
>>> with enabled rt_preempt, dyn_ticks, and local_apic
>>> But
>>>
>>> req.tv_nsec = 300000;
>>> req.tv_sec = 0;
>>> nanosleep(&req,NULL)
>>>
>>> make pause around 310-330 microseconds.
>>>       
>> How do you measure this?
>> If you want to have something done every 300 microseconds, you must not
>> sleep for 300 microseconds in each iteration, because you'd accumulate
>> errors. Use a periodic timer or use the current time to compute how long
>> to sleep in each iteration. Take a look how cyclictest does it.
>>     
>
> no. I just want my programm go to sleep sometimes and wake up in correct time.
>   

What does your program do that it has such a strict requirement on the
exact length of sleeping?

>>> I tried to understend how work nanosleep(), but it not depends from
>>> jiffies and from smp_apic_timer_interrupt.
>>>
>>> When can accuracy be lost?
>>> And how are process waked up?
>>>
>>>
>>> GolovaSteek
>>>       
>> Don't forget the process will always have non-zero wakeup latency. It
>> takes some time to process an interrupt, wakeup the process and schedule
>> it to run on the CPU. 10-30 microseconds is not unreasonable.
>>     
>
> But 20000 operations can be done in 10 microseconds?
> and why is there that inconstancy? Why sametimes 10 and sometimes 30?
> In which points of implementation it happens?
>
> GolovaSteek
>   

If a jitter of 20 microseconds is unacceptable for your application,
don't use PC hardware. Consider using a microcontroller.

Michal