Ugh, another machine with rotten latency

Rick

Thanks for the reply.

Although it compiled, the previous kernel / patch combination was obviously not right

I tried a different kernel ( 2.6.32 from kernel.org ) and used the 2.6.32.20-x86-2.7-03.patch from magna and compiled that under Ubuntu 10.04 and it worked perfectly ( I am running it currently)
I used the config listed in the neo-technical wiki to start.

Now I have it running, I will play with specifying multi-processor support, getting rid of multi-processor support etc. etc. (both in the kernel and rtai build) and see what gives the best latency figures.

Used to do this sort of thing all the time in the kernel 2.2.x days just to get my hardware working, took a while to remember how!

regards

ArcEye,

Glad to hear you got it going. I too feel there are gains to be made by fine tunning the config for different computers.

Rick G

Hi

The new kernel, rtai and re-compiled emc2 are working well under Ubuntu 10.4.

Because I use 2.6~pre on a RIP compile most of the time I compiled that too

This raised a factor that I had not considered.

With specific 4 core processor support configured, C99 maths on etc.
Emc 2.4.5 gave a max jitter of Base 26342 Servo 20209

Doing the same test under 2.6~pre with same progs to load it,
Emc 2.6~pre gave max jitter of Base 20343 Servo 18003

I was not expecting a 25%+ difference just changing emc versions and don't quite know how that comes about.

I think I am going to have to do 3 vanilla installs of 10.4 from the live cd on separate partitions, then compile a different custom rtai kernel on 2 of them ( 1 full 4 core support, 1 total single core, everything off that will go off) keeping 1 untouched as a reference.
I can then compile different versions of emc against them and compare all the figures.

This could get interesting ( in a geeky sort of way!)

regards

( Test machine is Intel DG451D, Quad core2 Intel processor with 6GB RAM, using on board graphics (Intel G45 Express chipset)
way overkill for emc, but this is my house computer only used for simulation.
However despite its power always returned latency figures 3x that of my P4 Fujitsu Siemens computers running 8.4 in the workshop)

RESULTS

The interim results are in from my multiple kernel, RTAI and EMC re-builds,
got quite reasonable figures in the end, but there are some interesting anomolies.

(System loading during latency test was 5 x glxgears open plus large file copies, deletes etc * however see later)

1) Vanilla install from Live CD Ubuntu 10.4
BIOS enabled multi-cores

Base max jitter 32083 Servo 166390 (not as bad as Jons servo of 320000 - but pretty c**p)

2) kernel 2.6.32 patched with magma 2.6.32.20-x86-2.7.03.patch running EMC 2.4.5 compiled against it
Kernel and magma configured for core2 specific chip, quad core
(with CPUHotplug, SFI, Thermal sysinfo, watchdog, virtualisation and accessability all off)

Base max jitter 21503 Servo 20865

NB. * When I tried to overload completely during latency-test by transcoding a video and streaming live video from an IP cam as well as glxgears
I think I saw the cpu-hog effect Rick outlined because figures now
Base max jitter 21032 Servo 18338

3)kernel 2.6.32 patched with magma 2.6.32.20-x86-2.7.03.patch running EMC 2.4.5 compiled against it
Kernel and magma configured for single core only (but with BIOS enabled multi-cores)

Base max jitter 20516 Servo 17114

4)kernel 2.6.32 patched with magma 2.6.32.20-x86-2.7.03.patch running EMC 2.4.5 compiled against it
Kernel and magma configured for single core only (BIOS DISABLING multi-cores)

Base max jitter 18187 Servo 15367

NB. This gave the best raw figures but trying to open new processes was like swimming through glue

Conclusions:-
You certainly don't need a quad core beast to run EMC, none of this came close to my consistant base jitter of 10500 on my workshop Fujitsu-Seimens P4s running 8.04

However shows that with a processor balanced kernel / RTAI setup, some of the newer machines should be well usable.
Quite reasonable figures retaining 4 cores with kernel and magma built to suit, retaining functionality for other tasks.

I have lost my realtime errors, running a hacked custom simulator that EMC thinks it is attached to a real machine.
I only use this computer for simulation anyway, so this is an academic exercise to an extent

regards

ArcEye wrote:

Conclusions:-
You certainly don't need a quad core beast to run EMC, none of this came close to my consistant base jitter of 10500 on my workshop Fujitsu-Seimens P4s running 8.04

True. It seems that older machines are much better in jitter results. I once tested a >10years PII IBM machine and (if I remember right) it had a max jitter of <10000.

ArcEye,

Looks like you spent some time on this and the information is quite interesting.

It seems to me that the EMC builds shipped with 10.04 and 8.04 do a great job of working with a large number of the computers out there, no small feat.

Also possible that:

Computers that don't work with the stock build a custom build may allow there use.

And a marginal computer's performance may also be increased by a custom build.

The computers I use for EMC do not do anything else so their performance in others areas is not important to me, but if it was I could just boot to a different kennel.

Of course the latency only has to be good enough to work with your hardware.

Again great info.

Rick G

Jon,

Did you get your Dell working or did you move onto another computer?

Rick G

Rick G wrote:

Jon,

Did you get your Dell working or did you move onto another computer?

Rick G

That specific computer is mainly to be used as a web server and firewall
for the Pico Systems network, and only for board testing and such, so
latency isn't really as important as on a real system. Also, it will be
used with servo interface boards that don't use a base thread, so
it may perform better than the latency test might indicate.

So, I am probably going to just leave it as is. But, I wouldn't
use it to control a real machine.

Jon