64bit + linux-3.8.13/rtai-4.0 + linuxcnc-2.6.0-pre

I've playing around with the software combo in the topic title and have been getting conflicting latency scores depending on which test is run.

The machine is an ASUS P6TD with a core I7 cpu.

All the usual bios/kernel settings have been set, Hyper-threading off, power management disabled and the kernel was booted with isolcups=2,3 and noirqbalance, I can attach kernel/rtai configs if needed.

The results from realtime/testsuite/kern/latency look good and are attached as latency.txt
The results from latency-test vary wildly and just plain don't make any sense, two screen shots attached.

Linuxcnc itself runs fine and never complains about realtime delays.

Any ideas of where I should look to get latency-test to behave ?

Regards.

Steve.

Hi

If you let the realtime terminal latency test run for a bit, I would suspect you will get one or more spikes.
You should be able to work out then what events are causing it

Because the dialog based latency test is cumulative, all you see is the worst historical figure.
The immediate difference between the 2 can be something as simple as a mouse, you are never going to use it for the terminal test, but will instinctively do so for the dialog based test

regards

I ran the terminal based test for several hours, with 8 instances of glxgears, watched a bunch of videos on YouTube and created a few 4.7G DVD iso images, the results are attached. Load average was over 7.00 on cpus 0 and 1, cpus 2 and 3 are isolated.

Another thing to note is the numbers shown on the gui dialogs are what is displayed at launch, clicking reset has no effect, only the least 2 or 3 significant digits change, stranger still, the startup values can vary by as much as 3 orders of magnitude at launch.

Regards.

Steve.

rythmnbls wrote:

The results from realtime/testsuite/kern/latency look good and are attached as latency.txt
The results from latency-test vary wildly and just plain don't make any sense, two screen shots attached.
[/attachment]

These LinuxCNC Hal latency pages look suspciously like the middle column
Max Jitter (ns) is actually the period of a thread, and NOT the jitter, which
would be some form of difference between thread times. Has anybody
changed this code recently? Is it possible somehow a 64 bit bug has
been introduced?

Jon

Another minor problem I've encountered is the Base Information dialog in stepconf. The Step Time, Step Space and Direction Hold values seem to be limited to a max value of 100ns. I've attached a patch that fixed it for me.


Regards,

Steve.

The latency problem is fixed, the problem was the kernel config, the interval numbers now look right and the jitter values look quite good.

Sorry for the noise.

Regards.

Steve.

Can you please post your kernel config for others to try it?

Yes, and the kernel patch used would be good.

I have an old one, but there are currently no published 3.x.x patches in magma or RTAI for x86_64, that I am aware of.

regards

github.com/ShabbyX/RTAI is where the new code starts.

github.com/ShabbyX/RTAI/tree/linuxcnc-old

Check with NTULINUX / memleak (in #linuxcnc-devel)