On 12/4/2012 2:47 AM, Anders Wallin wrote: >> I'm trying to place your 30us latency number in context. What >> motherboard/bios/cpu? >> > This is a new Intel DH61AG mini-ITX motheboard with an Intel Core i3 2120T > 2.6 GHz LGA1155 CPU. It has 4G of ram and a 60G SSD disk.
On specs alone, this sounds like a sweet combination. I easily could have chosen it for myself. I feel like we are alchemists---if you prefer, Tom Edison---trying every possible combination in our search for the solution. The recent explorations of Michael, Charles, John, Bas, and others of alternative realtime frameworks are fantastic but a proper science hasn't yet emerged. Engineering without underlying scientific principles is really tough but sometimes one is forced to it. > With HyperThreading turned off from the BIOS, and specifying isolcpus=1 in > grub I seem to get around 14us mostly, but still the occasional 30us > latency. The latency-histograms Michael had measured would really tell the > whole story as I feel the 30 or 35us occasional spikes are quite rare. > Any other BIOS/xenomai/linuxcnc tuning that I should try? For my In addition to latency-test there is a script called "latencyplot" that possibly could give some insight. I have spent time staring at the rolling plot trying to discern regularities in the jitter (like when I suspect a BIOS and SMI are doing some fancy footwork). I had conjectured on emc-users about how one might go about making an external pulse-train analyzer which, like Michael's digital logic analyzer, would be able to generate histograms but over a longer time scale. If one is satisfied that the internal, latency-test approach provides a reasonable metric, then it would be dead-simple to take latency-test/latencyplot a step further, bin the results, and derive interesting measures from it. Like latency-test, one could provide a running tally of key measures or like the OSADL does for its RT-Preempt, one could draw histograms and analyze exhaustively on demand. > application a 1ms thread will be enough, so I can probably live with 35us, > but it would be interesting to see if 15us is attainable easily. > > >> cpus, real and/or virtual). I recently reported latency measurements for >> a variety of configurations on my Quad-core AMD Athlon II box >> ( >> https://sites.google.com/site/manisbutareed/linuxcnc-2-5/xenomai-user-threads >> ). >> > nice page, more of that please :) > As I said before collecting this kind of data automatically into a > web-browsable database would be nice. Thanks. I have some other boards and RT-approaches to try but your remark about automation is exactly on target. After kvetching privately to Michael about the tedium of testing all possible combinations (I feel like one of Skinner's pigeons pecking obsessively at the keys) I got to thinking about this too. Fiddling with BIOS options is probably out of scope for automation, but everything else could be done programmatically. I've been idly sketching out what to try on my next board. Regards, Kent ------------------------------------------------------------------------------ LogMeIn Rescue: Anywhere, Anytime Remote support for IT. Free Trial Remotely access PCs and mobile devices and provide instant support Improve your efficiency, and focus on delivering more value-add services Discover what IT Professionals Know. Rescue delivers http://p.sf.net/sfu/logmein_12329d2d _______________________________________________ Emc-developers mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/emc-developers
