[ntp:questions] Re: More peering theory - refclocks in the mix

Richard B. Gilbert rgilbert88 at comcast.net
Thu Mar 16 17:59:56 UTC 2006


Joshua Coombs wrote:

> My test server aided with rrdtool has narrowed down my pool of potential 
> remote ntp servers to 17 healthy candidates with consistent latency, low 
> jitter, and similar theories on the one true tick.  That gives me 4 to 
> hand to each of my peered internal servers, plus an extra to toss on at 
> random.
> 
> One of these servers is dialing NIST periodically.  With the server set 
> it has, it never raises the polling interval from min (12) and never 
> chooses it as a time source.  The jitter seems rather high as well.  I 
> was under the impression that if I used the two NIST numbers, the 
> refclock driver could deduce the line latency automatically (always 
> reads 0 on the scoreboard) and took multiple samples per call to reduce 
> jitter.  Should I be using 'burst' on that server statement (is that 
> even valid?).  I'd like to see the server stretch the poll interval up 
> to the max if it can, seems like calling NIST every hour is a bit extreme.
> 
> Also, as that server has access to a reference clock, would I be better 
> served configuring just my peers on it and no external ip servers, 
> instead putting them on my 3 other internal servers?
> 
> I've also seen mention of using burst with extended intervals for remote 
> time sources mentioned in a few places.  If I only have a given remote 
> time source configured on one server, is using burst with a maxpoll of 
> 11 considered abusive?  Should I bump my minpoll if I do?
> 
> Joshua Coombs

Please be very careful about using burst.  Burst sends eight requests at 
every poll interval and creates an undue burden on the server.  It is a 
special purpose hack not intended for general use.  The documentation is 
not very clear about exactly what circumstances justify its use.

Burst should only be used with the knowledge and consent of the victim. 
    Do not confuse burst and iburst.  Iburst is good, it sends an intial 
burst of eight requests before beginning to poll at normal intervals; it 
gets you a fast startup and, provided you do not restart ntpd every few 
minutes does not place an undue burden on the server.




More information about the questions mailing list