RPI4 Raspbian 64 bit & LinuxCNC

More
5 years 6 months ago #205014 by cakeslob
Ill test it on the 8gb today/tomorrow, the 32 bit version for 5.4 works very well. bonus with the 5.10, i need that for a separate project.

How do these compare with the "official" images for latency?

They were pretty much the same,like underwhelmingly the same. pretty much the same as the previous 4.19.70something 4gb ram. I thought maybe between all the images i had saved, older or newers would be faster, but with vanilla settings they are all pretty close to the same. I tested elovalvo's 32bit 5.4 kernel and the 3gb 4.19 kernel official image side by side with 4gb and 8gb units. The only real difference I noticed was the ram, but i didnt dive to deep into the kernel config or anything. I have it loaded on my test unit and it preforms as expected aside for some reason I can only ever use 25% cpu usage, like when using make -j4. I forgot about that until just now.

Please Log in or Create an account to join the conversation.

More
5 years 6 months ago #205041 by elovalvo

andypugh wrote: I can try it on the Pi400. (if I remember)

In this case, you must verify that a hal_pi_gpio module works because the Pi400 (6) version is not included in the source code.

andypugh wrote: Would you expect it to work on the Pi3?

I don't think so, because I compiled the patch for the runtime with the bcm2711_defconfig option.
Following the link lemariva.com/blog/2019/09/raspberry-pi-4...19y-performance-test
you should use the option bcm2709_defconfig

andypugh wrote: How did you make the images?

I followed the steps in the first post of this thread.
In this case, however, I took into account the suggestion to disable virtualization (KVM)
github.com/kdoren/linux/issues/5

andypugh wrote: How do these compare with the "official" images for latency?

I have not installed the official releases at this time.
These days I try to make them

andypugh wrote: I make the "official" images. I would be more than happy to pass that task on.

Just tell me which version you need and I'll be happy to prepare it
The following user(s) said Thank You: tjtr33

Please Log in or Create an account to join the conversation.

More
5 years 6 months ago #205597 by cakeslob
works on 8gb, the 5.10 atleast , but im getting constant low voltage warnings which are (presumable) causing intense latency spikes, works pretty good on cm4 2gb
The following user(s) said Thank You: elovalvo

Please Log in or Create an account to join the conversation.

More
5 years 6 months ago #206501 by tjtr33
i think there are too many closely related threads here

and this hread is about rpi 64 bit so i post again...

elovalvo's 64bit runs well on pi4 4g
aarch64 5.4.61-rt37-v8+ 2.8.1
30.2 max base latency after 14 hrs
latency-histogram --base 80000 --servo 1000000
with 4 glxgears
hth
tomp

Please Log in or Create an account to join the conversation.

More
5 years 6 months ago #206615 by tjtr33
Another 14 hr run gave me 25.8 max base latency.
I use
latency-histogram --base 80000 --servo 1000000
with boot params of
isolcpus=1,2,3 idle=poll
and disable the screen blanking & screensaver

I dont know why 80 uS base period is such a sweet spot,
but 50,60,70,90 and 100uS were all tested. ( many tmes on may pi's )
and all (non-80) were not as good.
Why is 80 'sweet'?

I did get pretty good resuts from 50uS base period,
seeing 34uS latency after just a couple hours.

I was ready to discard the 50uS as a possible based thread period,
but it leaves 16 uS.

I say this ("leaves 16uS") because I believe the period - latency = reliable time available.
Is that true?

Then...
how good is 16uS as a time to get everything done?

When people say that 20uS latency is good in the context of a 50uS thread period,
are they implying they need 30uS to accomplish all tasks?

Is there a profiler that can determine how much of the period is used/unused?

thx
TomP
tjtr33

Please Log in or Create an account to join the conversation.

More
5 years 6 months ago #206729 by andypugh

tjtr33 wrote: Why is 80 'sweet'?


I suspect that you are closer to knowing that than anyone on the planet. At least you know it.
The number of realtime users on the Pi in the world who have analysed the system is measured in the "severals" I suspect.

Please Log in or Create an account to join the conversation.

Time to create page: 0.506 seconds
Powered by Kunena Forum