Le 21/02/2014 18:16, Franz Engel a écrit :
> Hi,
> 
> I made a little benchmark of the communication performance of my system
> (with Rtnet). Therefore I use an external Data Analyzer (Anritsu) and
> measure the delay between each telegram. I tried to send every 4ms. Pseudo
> code:
> 
> while(true):
> 
> ret=sendto(...)
> 
> sleep(4ms)
> 
> endwhile

If you really want to send a message every 4ms, this is not what you
should do. What you should do is either use an absolute delay, or make
the task periodic. Otherwise, the errors will add-up.

> 
> I make this test two times. One time with a totally unstressed system. That
> means I isolated the specific core, where my real-time task is working, and
> there are no other running programs. The result is that the mean delay of my
> communication was 0,004003518.
> 
> For the second test I stressed my system with "taskset -c 0-11 stress --cpu
> 12" (that means each core is working with 100%). Now I have the funny
> effect, that the result of the second test is better than the first one. In
> the second test I have an mean delay of 0,004000951. That means, the mean
> delay of the stressed system is 2.567 micro seconds faster than the
> unstressed system.
> 
> Does somebody have an idea what could be the reason for this phenomenon?

Sounds like a problem with power management. Maybe C1E mode on recent
intel processor?


-- 
Gilles.

_______________________________________________
Xenomai mailing list
[email protected]
http://www.xenomai.org/mailman/listinfo/xenomai

Reply via email to