[ntp:questions] oscillations in ntp clock synchronization

Moe Trin ibuprofin at painkiller.example.tld
Fri Jan 18 00:47:54 UTC 2008


On Thu, 17 Jan 2008, in the Usenet newsgroup comp.protocols.time.ntp, in article
<kbBjj.34002$fj2.20407 at edtnps82>, Unruh wrote:

>Yup ordinary computers. They are between 0-100PPM out-- across about
>8 of them. But the cyclic variation is strange.

A 100 ppm error due to thermal problems ALONE should be unusual, as
that would be suggestive of the device temperature changing a very
large amount (multiple tens of degrees C) or an extremely crappy
crystal cut (I don't think anyone is using X-Cut crystals any more).
But are you discussing the variation of an individual unit changing
up to 100 ppm, or that two (or more) units differ by up to 100 ppm.

>What is strange is the about 1.5 hour timescale. There is nothing I
>know of in any of the rooms that has that time scale.

Are all units in a single room showing similar time/rate/direction?
Room temperature isn't the only thing that can cause problems - it
could be load, vibration, and/or power supply voltage just as well.
But an oscillation with a period that long still suggests a wrong
set of constants in a feedback loop, especially in the absence of
external changes of a similar nature. If the period/span remains the
same 24/7 - I'd be looking at the filter a lot closer.

>Is there any way of getting the temperature from one of those CPU
>thermometers (onboard)  to better than 1C?

If you mean something like "lm_sensors" or reading the "Thermal
Reference Byte" in some Intel processors, no, I don't think so, and
I'd be wary of the data anyway. The "Thermal Reference Byte" is
subject to _calibration_ errors, (zero and scale), while the
"lm_sensors" data is subject to the sensor error as well as the
errors in the circuit that is converting the voltage to a digital
representation (I've discussed this in the past - recall that this
is commodity gear, accurate to 5-10 percent AT BEST, never calibrated,
and tested by half-starved chimpanzees to a "yes it's working/no it's
dead" tolerance).

If you wanted to construct something - a thermistor, a couple of
precision resistors, a supply reference, and a decent op-amp AND THEN
CALIBRATING the darn thing, then 1/20th degree should be obtainable
but I wouldn't look to commodity products for this.

>I can see some major variations in teh temp causing rate changes in
>the clock, but the small oscillations, if they are caused by thermal
>effects, are much smaller temp variations than are visible with a 1C
>accuracy.

Accuracy, or resolution, or both?    In one case up-thread, you talk
about a 2-3 usec amplitude, and that would be trivial - Assume a 100
ppm over 0-70C, then a 1C change _could_ produce a 1ppm error, and
that puts you out 2-3 usec in 2-3 seconds. (I'm talking time for the
drift, not the response time of the change in temperature to the actual
change in frequency.) 

        Old guy




More information about the questions mailing list