I'm terribly sorry to hijack this old thread, but as some process (not shown
among processes, huh?) froze and took 100% of one core (and sometimes
skipped from one core to another), I got really weird results on my Phenom
II X6 1055T running at 4GHz. (4GB 1780DDR3, 334MHz HTT)
Notably, the overall minimum latency was -1841ns, the overall maximum
latency 1702ns, with average of -1808ns.
This certainly is due to the cache subsystem running at full speed, namely
3GHz in my system (the CPU northbridge is clocked at 3GHz, that is).. I made
some photos if anybody is interested.
I am going to try this with an intentional hog loop (echo "empty" >
/dev/null), but I have some photos I would like to share.
I had cool&quiet enabled in BIOS, along with C1E state, along with something
else power management related. Trying to run the menu item latency test
resulted in nothing. But when I managed to do that results were really
astonishing.
User test of switch time:
SUSP/RES 232ns
SEM SIG/WAIT 247ns
RPC/RCV-RET 326ns
Kernel switch time (including full FP support):
111ns, 116ns, 147ns
some jitter test:
fast: 985ns, slow 1119ns
I wasn't able to run glxgears, but those results are insane by themselves,
aren't they?
Where do I put the photos (if anybody wants to see), should I try to edit
the wiki if this hardware configuration is exceptional?
With all this computing power, this is one very serious HARD real time
machine! My point is I think I know how to decrease those latencies even
further on this platform...
This is a fresh install from CD Ubuntu 10.04 linuxcnc i386, while
downloading 150MB of updates...
On Tue, Feb 10, 2009 at 2:33 PM, John Kasunich <[email protected]>wrote:
> paul_c wrote:
>
> >
> > Running a minimal script that just burns CPU cycles is of little value -
> An
> > analogy would be to jack up the wheels of a car and run the engine to the
> red
> > line and claim it will do 150MPH.
>
> Of course not, and that isn't what I was saying. Both SWP and I have
> observed that with an SMP RTAI kernel, running a simple do-nothing loop
> IN ADDITION to all the other things that we normally do during a latency
> test results in better overall maximum latency numbers.
>
> I don't know the mechanism, and was only speculating in my previous
> message. It could also be that the busy-loop keeps the Linux idle
> process from running (on at least one core), and the idle process does
> something that hurts latency. Maybe it slows the clock, or puts the CPU
> into a halt or sleep state so that it needs a bit of extra time to resume.
>
> My observations were on a Core2 Duo, with an Intel 975XBX2 motherboard.
> I haven't consistently seen the same effect with the Atom CPU/mobo
> combo. From what I can tell, the Atom doesn't slow the clock during
> idle periods, while the Core 2 does - maybe that is why the busy-loop
> makes a difference on one board and not the other...
>
> > What you _should_ be doing is providing some real-world loads over an
> extended
> > period, i.e:
> >
> > Start the RTAI latency test then:
> >
> > dd if=/dev/urandom of=/dev/null
> > ping -f -s 1600 -l 1000 localhost # also flood-ping from another box.
> > cd <kernel-src>/ && make -j
>
> > For more extreme testing, run cpuburn or (forkbomb). Starting up X
> > while the latency-test is running will also give the system a good
> > short term loading, and if it doesn't lock up starting/stopping X,
> > then you will have a solid setup. With X running, any graphics
> > intensive processes will also provide a good workout - Try some of
> > the OpenGL screensavers.
>
> I typically run glxgears, do a cvs update to get some network activity,
> and do a make clean ; make on my EMC2 source tree. Not quite as severe
> as what you listed above, but the same general idea.
>
> Regards,
>
> John Kasunich
>
>
>
> ------------------------------------------------------------------------------
> Create and Deploy Rich Internet Apps outside the browser with
> Adobe(R)AIR(TM)
> software. With Adobe AIR, Ajax developers can use existing skills and code
> to
> build responsive, highly engaging applications that combine the power of
> local
> resources and data with the reach of the web. Download the Adobe AIR SDK
> and
> Ajax docs to start building applications today-
> http://p.sf.net/sfu/adobe-com
> _______________________________________________
> Emc-developers mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/emc-developers
>
------------------------------------------------------------------------------
Nokia and AT&T present the 2010 Calling All Innovators-North America contest
Create new apps & games for the Nokia N8 for consumers in U.S. and Canada
$10 million total in prizes - $4M cash, 500 devices, nearly $6M in marketing
Develop with Nokia Qt SDK, Web Runtime, or Java and Publish to Ovi Store
http://p.sf.net/sfu/nokia-dev2dev
_______________________________________________
Emc-developers mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/emc-developers
