Re: 2.6.14-rt21: slow-running clock
From: Jonathan Woithe
Date: Tue Dec 13 2005 - 20:19:51 EST
Hi John
> On Fri, 2005-12-09 at 12:49 +1030, Jonathan Woithe wrote:
> > > Ok, I went digging further and found the c3tsc selection is correct on
> > > your hardware. I'm just too used to my own laptop where the TSC varies
> > > with cpu speed and we lower the rating value. So that should be ok.
> >
> > Ok, good. That leaves the c3tsc slowdown as the only outstanding issue at
> > this stage.
> >
> > > I'm now working on why we mis-compensate the c3tsc clocksource in the
> > > -RT tree.
> >
> > No problem. Let me know when you have something to test or need further
> > info.
>
> Attached is a test patch to see if it doesn't resolve the issue for
> you. I get a maximum change in drift of 30ppm when idling between C3
> states by being more careful with the C3 TSC compensation and I also
> force timekeeping updates when cpufreq events occur.
Unfortunately there's still an issue.
I applied this patch to a 2.6.14-rt21 kernel. It applied with some offsets,
and a few chunks needed a fuzz of 1. If you think this is important let me
know what the patch is against and I'll try that.
Anyway, to the results. Firstly, under 2.6.14-rt21 I find today that if
the computer is left to idle with the network interface enabled, the
system time actually runs *fast* compared to the CMOS clock. Previous
observations (which have been with the NIC module not loaded) have seen
the system clock run *slow* compared to the CMOS clock. If I run
jackd using
/usr/local/bin/jackd -R -d alsa -d hw:0 -r 48000 -n 4 -p 512
the previous situation (system clock slower than CMOS clock) prevails: after
less than a minute of running jack we went from
~/tsc> /sbin/clock -r && date
Wed Dec 14 10:49:54 2005 -0.254472 seconds
Wed Dec 14 10:49:55 CST 2005
to
~/tsc> /sbin/clock -r && date
Wed Dec 14 10:52:46 2005 -0.116721 seconds
Wed Dec 14 10:52:40 CST 2005
Running your drift script under 2.6.14-rt21 without jack running produces
the following:
~/tsc> ./drift-test.py -s 192.168.0.1
14 Dec 10:18:59 offset: 17.504476 drift: 56445.0 ppm
14 Dec 10:19:59 offset: 17.33345 drift: -1878.37704918 ppm
14 Dec 10:20:59 offset: 17.090538 drift: -2954.48760331 ppm
14 Dec 10:22:00 offset: 16.847759 drift: -3298.1978022 ppm
14 Dec 10:23:00 offset: 16.6384 drift: -3345.58264463 ppm
14 Dec 10:24:00 offset: 16.432669 drift: -3362.12582781 ppm
14 Dec 10:25:01 offset: 16.201893 drift: -3432.88705234 ppm
14 Dec 10:26:01 offset: 15.926439 drift: -3597.14420804 ppm
14 Dec 10:27:01 offset: 15.749873 drift: -3515.85507246 ppm
14 Dec 10:28:02 offset: 15.47964 drift: -3618.36580882 ppm
14 Dec 10:29:02 offset: 15.302595 drift: -3552.04635762 ppm
14 Dec 10:30:03 offset: 15.063286 drift: -3586.08270677 ppm
With jackd running on a 2.6.14-rt21 kernel we see this:
~/tsc> ./drift-test.py -s 192.168.0.1
14 Dec 10:55:07 offset: 30.4774 drift: 147306.0 ppm
14 Dec 10:55:58 offset: 39.063024 drift: 167940.961538 ppm
14 Dec 10:56:50 offset: 47.496911 drift: 165065.548077 ppm
14 Dec 10:57:42 offset: 56.076335 drift: 165040.00641 ppm
14 Dec 10:58:33 offset: 64.517798 drift: 165157.990338 ppm
14 Dec 10:59:25 offset: 73.139451 drift: 165287.092664 ppm
14 Dec 11:00:17 offset: 81.534864 drift: 164645.562701 ppm
14 Dec 11:01:08 offset: 90.110464 drift: 165139.143646 ppm
14 Dec 11:02:00 offset: 98.619667 drift: 164950.65942 ppm
14 Dec 11:02:51 offset: 107.154181 drift: 165213.090323 ppm
14 Dec 11:03:43 offset: 115.666054 drift: 165059.883946 ppm
14 Dec 11:04:35 offset: 124.174761 drift: 164929.115993 ppm
14 Dec 11:05:26 offset: 132.740329 drift: 165177.798387 ppm
Moving to a 2.6.14-rt21 with your patch from yesterday, the idling machine
with the tg3 module loaded had the following situation immediately
after booting:
~> /sbin/clock -r && date
Wed Dec 14 11:09:54 2005 -0.662011 seconds
Wed Dec 14 11:09:47 CST 2005
After sitting idle for a short while we had:
jwoithe@halite:~> /sbin/clock -r && date
Wed Dec 14 11:12:05 2005 -0.709709 seconds
Wed Dec 14 11:11:57 CST 2005
So in this case the system clock is definitely running slower than the CMOS
clock.
The drift stats with the sc patch (ie: 2.6.14-rt21 with your patch from
yesterday), without jack running:
~/tsc> ./drift-test.py -s 192.168.0.1
14 Dec 09:53:26 offset: 0.508609 drift: 62263.0 ppm
14 Dec 09:54:26 offset: 0.677778 drift: 3793.96721311 ppm
14 Dec 09:55:26 offset: 0.885301 drift: 3627.72727273 ppm
14 Dec 09:56:26 offset: 1.012511 drift: 3127.98342541 ppm
14 Dec 09:57:26 offset: 1.179381 drift: 3041.63900415 ppm
14 Dec 09:58:26 offset: 1.275088 drift: 2753.29568106 ppm
14 Dec 09:59:25 offset: 1.53823 drift: 3033.01111111 ppm
14 Dec 10:00:25 offset: 1.682737 drift: 2943.78809524 ppm
14 Dec 10:01:25 offset: 1.810167 drift: 2841.29375 ppm
14 Dec 10:02:25 offset: 1.908942 drift: 2708.51111111 ppm
14 Dec 10:03:25 offset: 2.101228 drift: 2758.13666667 ppm
14 Dec 10:04:25 offset: 2.199901 drift: 2656.90151515 ppm
14 Dec 10:05:25 offset: 2.564818 drift: 2942.32222222 ppm
14 Dec 10:06:25 offset: 2.695642 drift: 2883.71282051 ppm
14 Dec 10:07:25 offset: 2.859662 drift: 2872.9952381 ppm
14 Dec 10:08:25 offset: 2.990623 drift: 2826.97444444 ppm
14 Dec 10:09:25 offset: 3.221054 drift: 2890.32083333 ppm
14 Dec 10:10:25 offset: 3.364036 drift: 2860.48039216 ppm
14 Dec 10:11:25 offset: 3.526485 drift: 2851.98055556 ppm
14 Dec 10:12:25 offset: 3.661997 drift: 2820.74649123 ppm
14 Dec 10:13:25 offset: 3.853566 drift: 2839.35 ppm
With the tsc patch (ie: 2.6.14-rt21 + your patch) with jackd running:
~/tsc> ./drift-test.py -s 192.168.0.1
14 Dec 11:14:06 offset: 158.954383 drift: 171300.0 ppm
14 Dec 11:14:55 offset: 169.858831 drift: 221514.96 ppm
14 Dec 11:15:44 offset: 180.577996 drift: 220150.636364 ppm
14 Dec 11:16:33 offset: 191.516626 drift: 221172.587838 ppm
14 Dec 11:17:23 offset: 202.295611 drift: 219760.242424 ppm
14 Dec 11:18:12 offset: 213.217457 drift: 220382.080972 ppm
14 Dec 11:19:02 offset: 223.900228 drift: 219249.646465 ppm
14 Dec 11:19:51 offset: 234.819493 drift: 219758.410405 ppm
14 Dec 11:20:40 offset: 245.524901 drift: 219599.539241 ppm
14 Dec 11:21:29 offset: 256.4786 drift: 220034.948198 ppm
14 Dec 11:22:19 offset: 267.234911 drift: 219538.11336 ppm
With a plain 2.6.14 kernel (for reference) without jackd running:
~/tsc> ./drift-test.py -s 192.168.0.1
14 Dec 11:25:24 offset: 275.996533 drift: -18.0000000114 ppm
14 Dec 11:26:24 offset: 275.995463 drift: -17.8360655744 ppm
14 Dec 11:27:24 offset: 275.994522 drift: -16.7685950413 ppm
14 Dec 11:28:25 offset: 275.993527 drift: -16.6153846156 ppm
14 Dec 11:29:25 offset: 275.992539 drift: -16.5785123966 ppm
14 Dec 11:30:25 offset: 275.991556 drift: -16.5397350994 ppm
14 Dec 11:31:25 offset: 275.990544 drift: -16.593922652 ppm
14 Dec 11:32:25 offset: 275.989575 drift: -16.5308056872 ppm
14 Dec 11:33:25 offset: 275.988622 drift: -16.4502074689 ppm
Finally, 2.6.14 with jackd running:
~/tsc> ./drift-test.py -s 192.168.0.1
14 Dec 11:34:06 offset: 275.987919 drift: -20.0000000063 ppm
14 Dec 11:35:07 offset: 275.986933 drift: -16.225806451 ppm
14 Dec 11:36:07 offset: 275.985932 drift: -16.4508196721 ppm
14 Dec 11:37:07 offset: 275.984939 drift: -16.4835164834 ppm
14 Dec 11:38:07 offset: 275.983928 drift: -16.5743801653 ppm
14 Dec 11:39:07 offset: 275.982924 drift: -16.6059602648 ppm
14 Dec 11:40:07 offset: 275.981923 drift: -16.6187845304 ppm
14 Dec 11:41:07 offset: 275.980915 drift: -16.644549763 ppm
14 Dec 11:42:07 offset: 275.97991 drift: -16.6576763485 ppm
14 Dec 11:43:07 offset: 275.978898 drift: -16.6808118081 ppm
Regards
jonathan
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/