Back
[01:38:20] <KGB-linuxcnc> 03Chris Morley 05pncconf_GTK_builder 5f0f83b 06linuxcnc 10(32 files in 3 dirs) pncconf -GTK BUILDER refactor * 14http://git.linuxcnc.org/?p=linuxcnc.git;a=commitdiff;h=5f0f83b
[01:38:20] <KGB-linuxcnc> 03Chris S Morley 05pncconf_GTK_builder 8b153f0 06linuxcnc 10src/emc/usr_intf/pncconf/dialogs.glade 10src/emc/usr_intf/pncconf/pages.py 10src/emc/usr_intf/pncconf/pncconf.py 10src/emc/usr_intf/pncconf/tests.py pncconf -fix live test bugs - openloop and stepper tune test * 14http://git.linuxcnc.org/?p=linuxcnc.git;a=commitdiff;h=8b153f0
[10:38:01] <skunkworks> pcw_home, do you think there could be some math that says - huge step change in 1 period and oposite in the next - throw them out (sum them or whatever(
[10:48:14] <CaptHindsight> that's the fun of feedback loops, response time vs bandwidth, in a mechanical system at least you can rule out some noise
[10:59:50] <KGB-linuxcnc> 03Sebastian Kuzminsky 05v2.5_branch 97780e3 06linuxcnc 10docs/man/man1/axis-remote.1 10src/emc/usr_intf/axis/scripts/axis-remote.py docs: update --help & manpage for axis-remote * 14http://git.linuxcnc.org/?p=linuxcnc.git;a=commitdiff;h=97780e3
[11:08:27] <pcw_home> other than just averaging them I dont know. If you just average over a larger time
[11:08:29] <pcw_home> period you risk sampling only 1/2 of a peak pair. One other possibility is just dropping
[11:08:30] <pcw_home> single (or pairs) of samples that are out of line
[11:09:12] <pcw_home> (or fix linuxcncs stepwise velocity profiles)
[11:11:03] <pcw_home> (not that that will help with measuring other systems acceleration profiles)
[11:23:35] <skunkworks> yes - I think I am still getting 1/2 of the spikes - as I still get acc > limit
[11:23:47] <skunkworks> some times
[11:25:37] <pcw_home> Might be interesting to test Mach and see if it has velocity steps
[11:29:47] <CaptHindsight> well mach is good at missing steps so that you can cut right into your table if you don't double check :)
[11:30:37] <pcw_home> missing step will sure show up an a accel plot :-)
[11:30:42] <pcw_home> steps
[11:31:57] <KGB-linuxcnc> 03Dewey Garrett 05master 555c0fc 06linuxcnc 10tcl/ngcgui.tcl ngcgui: make .gcmc files visible in open dialogs * 14http://git.linuxcnc.org/?p=linuxcnc.git;a=commitdiff;h=555c0fc
[12:09:34] <CaptHindsight> after 2 weeks the preempt-rt latency just jumped from 48uS to 68uS while running videos in chrome
[12:11:40] <pcw_home> I get somewhere 80 usec on my desktop (Preemt_RT) running all sorts of normal stuff
[12:11:54] <pcw_home> somewhere near
[12:12:04] <CaptHindsight> I was out of the room when it jumped
[12:12:17] <pcw_home> NSA access
[12:12:41] <CaptHindsight> but week one was <42uS, week 2 48uS, now week 3 68uS
[12:12:56] <CaptHindsight> might be
[12:13:19] <CaptHindsight> was playing a series of youtube videos, maybe it didn't like MGMT?
[12:13:32] <pcw_home> maybe memory leak also
[12:14:16] <pcw_home> or maybe just statistics
[12:15:39] <pcw_home> if the jitter is peaky (mostly good) there are some things that can be done to lessen their impact even on high speed/precise systems
[12:15:45] <CaptHindsight> but it's safe to say 42uS for preempt-rt
[12:16:16] <CaptHindsight> next is on a real machine and no youtube
[12:16:43] <CaptHindsight> I use a different PC to play videos while I'm working anyway
[12:17:00] <pcw_home> Not bad. If we can get our Ethernet stuff running reliably at 3-4 KHz on Preemt RT I would be pretty happy
[12:23:53] <CaptHindsight> I'm getting the devs in China to start on the A20 support
[12:24:51] <CaptHindsight> so preemptRT and RTAI for A20 along with SPI drivers for FPGA
[12:25:21] <CaptHindsight> we can take a look at ethernet for arm as well
[12:35:32] <CaptHindsight> it ends up that Tom Cubie and other were the only devs building the kernels and ISO's
[12:35:39] <CaptHindsight> 1 other
[12:50:42] <pcw_home> yeah not expert kernel builders
[12:55:02] <CaptHindsight> Tom was allwinners kernel guy
[12:55:11] <CaptHindsight> he left to start cubieboard
[12:55:33] <CaptHindsight> he's not a hwardware guy
[12:55:37] <pcw_home> Is Rockchip more closed the AllWinner?
[12:55:38] <CaptHindsight> hardware even
[12:55:52] <CaptHindsight> actually they just started opening up everything
[12:56:01] <pcw_home> (I keep wanting to say AllWeiner)
[12:56:09] <CaptHindsight> let me the the link
[12:56:51] <CaptHindsight> they aren't foreign invaders, they are more like misdirected kids with lots of backing
[12:57:56] <CaptHindsight> https://gitorious.org/rockchip-android
[12:58:28] <pcw_home> Toms newest looks interesting ( Radxa Rock)
[12:59:09] <CaptHindsight> they come from farms with the bucket, the chair and the roof and get sent to school for engineering since they behaved and tested well
[12:59:29] <CaptHindsight> they get their first hands on a PC at 19
[13:00:24] <pcw_home> yeah pretty quick rate of change
[13:00:27] <CaptHindsight> at 23 they have an MS in EE or CS and get jobs at Oracle, Google, Intel AMd as engineers making $10K usd/year
[13:01:12] <CaptHindsight> so no real experience with tech, no tinkering, no taking apart the microwave or toaster ovens at age 8
[13:02:00] <CaptHindsight> and no real exposure to the west except for 90210 and Baywatch circa '95
[13:03:16] <CaptHindsight> smart kids with no experience
[13:04:45] <pcw_home> Thats an interesting perspective
[13:05:01] <CaptHindsight> that is what I see there
[13:05:14] <CaptHindsight> but the official story is different
[13:05:43] <pcw_home> The official story is always different :-)
[13:06:26] <CaptHindsight> but the leadership there is mostly scientists and engineers
[13:06:29] <CaptHindsight> so they get it
[13:07:00] <CaptHindsight> at the same time there is corruption with so much wealth and power
[13:08:41] <CaptHindsight> that is why they government sets up investments in start-ups
[13:09:28] <CaptHindsight> last year 1000 start-ups were given $100K-3M usd along with offices and factories
[13:11:31] <CaptHindsight> the successful ones get 10x or 1% that in funding or loans after a couple of years
[13:11:44] <CaptHindsight> 1% interest loans
[13:12:39] <CaptHindsight> last year 6000 applied and 600 were chosen
[13:13:00] <CaptHindsight> they tried to find 1000
[13:13:51] <CaptHindsight> Tom just went off on his own and started making boards out of his pocket and kickstarter
[13:15:34] <CaptHindsight> MediaTek is the other big ARM soc in China. They are slowly going GPL as well
[13:22:52] <pcw_home> A good thing overall, I suspect part of TIs lowering the price if Sitaras was done in response to Chinese SOCs
[13:33:37] <CaptHindsight> I have quad core 1.5Ghz smart phones that were $150 from Lenovo
[13:33:53] <CaptHindsight> the chipset is ~$20
[13:35:09] <CaptHindsight> no plan. full price
[13:52:42] <skunkworks> linuxcnc....
http://imagebin.org/292611
[13:53:55] <skunkworks> mach3....
http://imagebin.org/292612
[13:55:16] <skunkworks> sampling at 36ms
[13:55:20] <mozmck> skunkworks: both pictures look like linuxcnc?
[13:55:48] <skunkworks> heh
[13:55:52] <skunkworks> let me explain....
[13:56:01] <pcw_home> and the other shoe: what acceleration constraints
[13:56:21] <skunkworks> both are set to 30in/sec^2 and 500ipm
[13:57:10] <skunkworks> mozmck,
http://youtu.be/qo74moJ30H4
[13:58:13] <mozmck> what is that supposed to explain?
[13:58:38] <mozmck> Looks like 2 computers running linuxcnc
[13:58:39] <pcw_home> a detailed plot of the accel might be needed to see whats going on (or maybe Mach simply doesnt obey accel contraints)
[14:00:20] <skunkworks> mozmck, I am using a mesa 7i80 card to read step/dir signals back into linuxcnc and trying to calculate acceleration
[14:00:33] <skunkworks> it isn't as easy as I thought it would be..
[14:01:10] <skunkworks> the first one is running step/dir from linuxcnc into linuxcnc
[14:01:24] <skunkworks> the second is running step/dir from mach into linuxcnc
[14:01:41] <skunkworks> So linuxcnc is backploting and sampling the step/dir
[14:02:01] <mozmck> Ah, I see. So I guess the point is that Mach is not following accel constraints?
[14:03:20] <skunkworks> well - as well as I can sample it. I am averaging the accelleration over 31ms. That makes it so when sampling linuxcnc - I don't get anything over 30in/sec^2 in the sampling. when I switch to mach - using the same sampling rate - I get much higer peaks
[14:03:42] <CaptHindsight> I didn't think it was their design philosophy, it was more point and pray to support USB
[14:04:03] <skunkworks> CaptHindsight, usb?
[14:04:27] <CaptHindsight> mach + usb
[14:04:46] <skunkworks> oh - I see - motion cards..
[14:05:12] <CaptHindsight> since mach users mostly use steppers anyway and it's open loop
[14:06:20] <CaptHindsight> they don't use a win RT kernel either
[14:08:49] <pcw_home> halscope plots of accel would be nice (or halsample it to a file for analysis )
[14:14:13] <skunkworks> pcw_home, does this make sense?
[14:14:14] <skunkworks> http://imagebin.org/292616
[14:15:16] <skunkworks> you can see the acc sample length :)
[14:24:47] <pcw_home> Yeah that makes it easy to interpret
[14:25:16] <skunkworks> I do have counts also..
[14:25:20] <pcw_home> Sure the Mach scaling is the same?
[14:25:21] <skunkworks> (it was turned off)
[14:25:46] <skunkworks> yes - it follows the path pretty well
[14:26:32] <skunkworks> if that is what you mean
[14:28:55] <skunkworks> this is with counts
[14:28:57] <skunkworks> http://imagebin.org/292619
[14:34:18] <skunkworks> it has peaked to almost 100in/sec^2 - have been running it over and over - it keeps following the path
[14:36:21] <mozmck> You are sampling every 31ms? Maybe there is more jitter in the Mach3 output - that would look like spikes in accel I think.
[14:37:33] <skunkworks> sure
[14:37:52] <micges> skunkworks: any communication errors with 7i80?
[14:40:39] <skunkworks> sometimes it doesn't start... but I have not looked at the logs to see what is going on.. (takes 2 tries for linuxcnc to load)
[14:40:59] <skunkworks> but if it is running - I don't see any issues - been running 1khz
[14:41:48] <micges> good
[14:41:56] <skunkworks> have you seen that?
[14:42:12] <pcw_home> I have not
[14:42:24] <micges> not starting not, not closing yes
[14:43:14] <skunkworks> well - I will get a log output
[14:46:31] <skunkworks> http://imagebin.org/292622
[14:47:08] <skunkworks> looks like quite a correction
[14:51:27] <skunkworks> micges, any updates I should pull?
[14:51:43] <micges> not yet
[14:56:14] <skunkworks> bbl
[15:28:45] <CaptHindsight> huh, firefox jolts latency right when it opens, chrome takes several minutes
[15:29:01] <CaptHindsight> all while playing flash
[15:30:11] <CaptHindsight> nearly 3 weeks of preempt-rt to jump to 68uS with firefox, 1 hour on Chrome and it's at 62uS
[16:10:47] <skunkworks> logger[mah]:
[16:10:47] <logger[mah]> skunkworks: Log stored at
http://linuxcnc.mah.priv.at/irc/%23linuxcnc-devel/2014-02-10.html
[16:12:44] <skunkworks> PCW: did you see
http://imagebin.org/292622
[16:17:27] <skunkworks> seems like quite a correction - it is outputting steps.
[16:21:45] <PCW> Yeah i wonder if you can capture the same section generated by linuxcnc (though the sampling windows will be different)
[16:27:45] <PCW> one thing that might make interpretation easier is to delay the other traces by the (31 ms?) sample interval
[16:27:47] <PCW> (the accel trace is for the previous sample period so always looks a sample period late)
[16:29:29] <skunkworks> oh - so that is from the \/ part?
[16:29:49] <skunkworks> the first one..
[16:35:37] <skunkworks> I probably could - I would just have to log xyz position. Although - this was the luck of the draw where the acc sample timed out perfectly with the violation
[16:36:03] <PCW> yeah because your differences are always after he "integration time" so they are 31 (or whatever) ms late
[16:37:49] <PCW> you can see by looking at the slope of velocity (which should match the accel value) that the accel plot is always late
[16:39:21] <skunkworks> makes sense
[16:45:29] <skunkworks> could be that this machine doesn't run mach very well also...
[16:47:05] <PCW> yes but the velocity slopes dont lie (assuming scaling is correct)
[16:53:45] <PCW> just looking at the deccel near the cursor, I get ~ 2.7 IPS change in 40 ms or about 67 IPS/S
[16:56:53] <skunkworks> yikes
[16:57:01] <PCW> but I would try and duplicate on another machine
[16:58:46] <skunkworks> PCW: it follows path -
http://imagebin.org/292612
[16:59:02] <skunkworks> 1500 step/inch
[16:59:48] <PCW> everyone 1s 1500 steps /inch?
[17:00:22] <skunkworks> yes - (linuxcnc and mach setup the same.)
[17:00:48] <skunkworks> 1500 steps/inch - 30in/sec^2 500ipm
[17:03:32] <PCW> not quite a smoking gun but does suggest that Mach does not respect accel constraints
[17:06:28] <PCW> (or does some trickyness like allowing higher accel when moving slow which would wring better performance out of step motors)
[17:07:43] <skunkworks> heh
[17:08:23] <skunkworks> I think they would be touting that then...
[17:08:31] <PCW> what is the maximum velocity?
[17:08:57] <PCW> Oh you just listed it
[17:10:27] <PCW> maybe its just bad...
[17:34:05] <CaptHindsight> what are the advantages of splitting the functions of EMC2 up into a headless RT/HAL/motion environment from the UI vs just using something like VNC?
[17:34:59] <skunkworks> I think as far as the bbb - it makes it so it can run the hal/realtime parts while a faster video computer can run the gui
[17:35:43] <skunkworks> vnc requires the realtime computer to have decent x performance
[17:35:52] <skunkworks> (from what I understand)
[17:36:58] <CaptHindsight> then why even use the AM335x?
[17:37:30] <CaptHindsight> I don't follow this logic
[17:38:05] <CaptHindsight> everything that gets touched by reprap turns into a situation like this
[17:38:58] <CaptHindsight> that project that used EMC on the old mini2440 worked
[17:39:09] <CaptHindsight> it didn't have a GPU worth a damn
[17:39:34] <CaptHindsight> and that was arm9
[17:40:57] <CaptHindsight> http://code.google.com/p/miniemc2/downloads/list
[17:41:59] <CaptHindsight> http://www.youtube.com/watch?v=YSnNtgZrfyU Pico-SAM9G45 board was the other project
[17:43:16] <CaptHindsight> what am I missing here?
[20:33:12] <cmorley> your missing the coolness factor
[20:33:25] <CaptHindsight> ah
[20:34:13] <cmorley> coolness often waste money and is not practical lol