| Re: [chrony-dev] Poll adjust after long time unreachable? |
[ Thread Index |
Date Index
| More chrony.tuxfamily.org/chrony-dev Archives
]
- To: chrony-dev@xxxxxxxxxxxxxxxxxxxx
- Subject: Re: [chrony-dev] Poll adjust after long time unreachable?
- From: Richard Curnow <rc@xxxxxxxxxx>
- Date: Wed, 16 Oct 2013 22:54:06 +0100
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20120113; h=sender:date:from:to:subject:message-id:references:mime-version :content-type:content-disposition:in-reply-to:user-agent; bh=P5qgFlXHipDkkkI1sXPnNTPr6lE+SPrSpUEqEoLpeJg=; b=JXLoNMmqRdI5h9zKZUkCSQ4LIVQ1IQci9pp+EnxpyJJWsuzZeAZjIILfGjtorcp7Y8 2nohXunWH5pgjp35NGye5sg90369vEdNODb0OAu/pkkHF/U+cvToISfxF9uYqfvSGWVM +UXeUrYt/homaCK36yUEMLRakfaLy0Dh856NV5RugBIRVtCBwiMX/tUuMK//gRcb5+7w c6z8sGd5R14JvqSvglA5LXioKyCVOgpZyJRH9Ww4NI2hj+K1msSLC8vHAweGajVVMP3E vNvqdEl9GafcEmIPUcX3Bhb8gC9sbtS05K2Z4PqjcTB/7o5WvjAl+9lmTgWFjib6RrgK Ydew==
On Mon, Aug 19, 2013 at 11:56:27AM -0700, Bill Unruh wrote:
>
> For short term random, no. But for long term changes, it could. For example
> the crystal ages, and the frequency goes down. over a long time ( hours, days)
> Or the temperature changes over hours. Both would be expected to have a
> relatively show change in frequency which a quadaratic term should capture.
>
> Clearly short terms frequency fluctuations are not captured.
I found last year during another project that using a higher order model
to approximate noisy data got nowhere. The coefficients of the
higher-power terms grew out of control as the curve tried too much to
fit the noise.
However, if the rate of time drift is slowing down or speeding up
(perhaps the temperature is changing?), perhaps a quadratic model could
be worth looking into.
> >
> >From the tests I have done in the past, the weakest link in the
> >chrony loop seems to be the code which determines how many samples
> >should be used in the linear regression.
>
> Yes, that is the non-linearity in the process. At what point does
> linear regression fail. One could fit a high order polynomial and
> demand that the higher order terms be small. The counting of changes
> in sign of the deviations is a crude way of determining when the fit
> gets bad. I do not knw if there are well developed procedures in the
> statistics literature for this kind of decision procedure.
Going back many years, I don't recall much rigour going into the design
of the "number of runs" logic. The code arose from a heuristic that
seemed to work well enough; I didn't find anything in the literature
that led to it being replaced with something else.
If someone can find some applicable "number of runs" theory that could
apply, it would be worth looking into.
--
Richard Curnow | http://about.me/rc0
Weston-super-Mare |
United Kingdom |
--
To unsubscribe email chrony-dev-request@xxxxxxxxxxxxxxxxxxxx with "unsubscribe" in the subject.
For help email chrony-dev-request@xxxxxxxxxxxxxxxxxxxx with "help" in the subject.
Trouble? Email listmaster@xxxxxxxxxxxxxxxxxxxx.