Installing LinuxCNC 2.9 or Master on Linux Mint 21

More
3 years 9 months ago #261554 by Stanislavz

That helped. Still needed apparmor=0
This is Debian Bookworm with my 6.1 kernel.
Code:
quiet isolcpus=nohz,domain,2-3 apparmor=0   [attachment=50991]debian-11Jan2023-70.png[/attachment] So I think that will do me for now as its way under what a mesa card requires. I really would like to poke around with the kernel parameters to see if there are settings that improve network latency (which you should be able to measure roughly with ping times. PCW has ping command syntax  that lets it run without a screen refresh until you stop it. Then look at averages...
 
Agree, that apparmor is really messing system.

And while perfect is an enemy of good :)

sysctl kernel.sched_rt_runtime_us=-1

this to be run before testing.

Please Log in to join the conversation.

More
3 years 9 months ago #261620 by Stanislavz
One more approach. I was curious - why i have marginally better results without nohz. And why some graphs have trident shape. On same setup/session.Reason - nohz is useable if we have 1 process in isolated core. But while testing with base period - we have two.Moved servo thread to core 0 with taskset - and we are in 6-7k region with glxgears and youtube. Without youtube even less.. 

Please Log in to join the conversation.

More
3 years 9 months ago #261622 by rodw
sometimes a fat latency is becasuse the horizontal axis is clipped (eg <  the 52 max shown) This is controlled by sbins parameter
Always use this:
latency-histogram --nobase --sbins 1000

Please Log in to join the conversation.

More
3 years 9 months ago #261627 by Stanislavz
No no. All here is ok. I am tuning for parport driver. And yes - i was able to see different shapes / wrong results.
So - because i cant run it with noservo parameter i just moved servo to 0 core.

If run with nobase - it is always consistent in shape ant similar in delay.

Please Log in to join the conversation.

More
3 years 9 months ago #261629 by rodw
then use
latency-histogram --sbins 1000

Please Log in to join the conversation.

More
3 years 9 months ago #261630 by Stanislavz
Nope still two processes at core which we optimise to one.

Please Log in to join the conversation.

More
3 years 9 months ago #261632 by Stanislavz
Will let it run in this test for night

Please Log in to join the conversation.

More
3 years 9 months ago #261662 by Stanislavz
Soo. Only test one process on single core.

Please Log in to join the conversation.

More
3 years 9 months ago #261684 by tommylight
Looking good, thank you.

Please Log in to join the conversation.

More
3 years 9 months ago #261691 by Stanislavz

Looking good, thank you.
 
And it have stable result now.  After restart. Around 7k. You should not test it with both threads. All optimization is done for single core only.

Could anybody explain - are there any difference between servo and base thread on testing ? If servo is run at same period as base ?

 

Please Log in to join the conversation.

Time to create page: 0.179 seconds
Powered by Kunena Forum