Re: [Xenomai-help] Perfs and rt alarms questions.

[Jan Kiszka <jan.kiszka@domain.hid>]

[-- Attachment #1: Type: text/plain, Size: 1765 bytes --]

RAKOTOSALAMA, Nirilanto wrote:
> Hello everybody.
> 
> I managed to compile a stable kernel (the Scalability O(1) scheduler options seems to make my system freeze)

That's not good. What precisely did you do? Applied the .config you
posted, just enabling CONFIG_XENO_OPT_SCALABLE_SCHED? And what test
case, latency?

> xenomai 2.3.1 and vanilla 2.6.20.4, gcc 4.1.1 on FC6.
> 
> I'm programming and testing a xenomai support for a timer library using rt_alarm. The test program
> lauches the timer periodically and calculate the precision (difference between theoretical and practical results).
> 
> When I launch the latency test, the max latency I have doesn't exceed 10us. So I supposed the precision shouldn't 

Using what load for the system? Keep in mind that unloaded benchmark
numbers are practically meaningless.

> exceed 10 to 20us as well, No ? But sometimes, the precision exceed 30 to 60us. Is it normal ?

Depending on your hardware (CPU, chipset, peripherals), this can be very
well a worst-case latency of the loaded system. Try to capture such a
result with the tracer (you may put your own xntrace_user_freeze() calls
into your timer lib).

> Something wrong in the way I use rt alarms ? in the kernel configuration ?
> 
> I attach : 
> 	- .config
> 	- tests results with overtime with this actual config
> 	- tests results without overtime I had with an old config (unfortunately I cannot restore it). The precision was less than 
> 		8us, I didn't enable xeno debug option. I haven't managed to get such high precision since I recompiled my kernel.
> 	- test program test_stability_trace.c
> 	- xenomai timer library support implementation support_xenomai.c
> 
> Thanks in advance.
> 
> Niry.
> 

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 250 bytes --]