Something Wicked (with my hdwr clock) This Way Comes

Dean S. Messing deanm at sharplabs.com
Sat Dec 29 20:09:10 UTC 2007


Louis Lagendijk writes:
: On Fri, 2007-12-28 at 15:23 -0800, Dean S. Messing wrote:
: <summary>
:   adjtimex --utc --compare=20 --interval=10
:   returns values of "system-cmos" that make it look like the RTC
:   is "jumping"
: 
: Could this be caused by some funny resolution of the system time
: delivered by the RTC to the OS? If it delivers time only in seconds, by
: uses some strange divider internally, one could see this
: behaviour.....  
: The RTC might still be accurate, but might vary upto one second from the
: system clock.

Hi Louis.

This is an interesting idea.  Do you know of an independent way to
verify it?  I guess I'm going to have to find some time (limited!) to
read the kernel docs on /dev/rtc.

If /dev/rtc only delivers time with 1 second granularity that would be
mightly useless, I would think.

One thing that may argue against your idea: each time I invoke

adjtimex --utc --compare=20 --interval=X

X can be _anything_ (I've tried 10, 13, 15, 20, 23, 60),
but the delta between (system-cmos) from line to line
is always almost exactly 0.1 seconds, until the "correction"
occurs.

By the way, I ruled out the possibility of the CMOS clock
running grossly slow.  Last evening I turned off
`ntpd', hand set my system clock to freerun using

adjtimex --tick 10001   --freq 3831728

which are the equivalent parameters that ntpd was
using to drive the sytem clock when it was well-tuned.

I also turned off the 11 minute kernel update to the RTC
(with adjtimex -S 64 )

So both clock are free-running.

In 12 hours or so my system clock has drifted from
the "montpelier.ilan.caltech.edu" stratum 1 clock
by only 23 ms (from -0.005 to 0.018) which is less
than 1 ppm. On the other hand


[root at medulla ~]# hwclock --show; date
Sat 29 Dec 2007 12:04:10 PM PST  -0.080610 seconds
Sat Dec 29 12:04:09 PST 2007

so my RTC has deviated from system time by no more than
1 second.   Indeed it appears be running slightly _fast_
not grossly slow.   

Dean




More information about the fedora-list mailing list