[ntp:questions] drift value very large and very unstable

David L. Mills mills at udel.edu
Thu Mar 6 18:28:57 UTC 2008


1. As per the advice in the documentation, do not use iburst with 
reference clock drivers. The only reason you might want to do this is to 
reduce the time to set the clock on initial start. This is unnecessary 
as the clock is now set on the first reply received. Using iburst anyway 
will screw up the radio protocol in some drivers.

2. The temptation to reduce the poll interval below the default can be 
counterproductive. The driver interface uses a median filter to clean up 
nominal jitter due to serial port and interrupt latencies. Generally, 
the jitter is reduced as the number of stages and the poll interval are 
increased. There are cases, in particular with kernel PPS signals, where 
a smaller poll interval can result in marginally better performance, but 
in other cases it generally not a good idea.

3. The burst mode is designed for use when the poll interval of 
necessity must be very long, like at least 1024 s. The current design 
will rate-limit if burst is used with a poll interval of 512 s or less. 
This is to protect busy servers with a minimum average default headway 
of 16 s. Some operators might set the headway higher.


Kevin Oberman wrote:

>>From: Harlan Stenn <stenn at ntp.org>
>>Date: Thu, 06 Mar 2008 06:33:17 +0000
>>Sender: questions-bounces+oberman=es.net at lists.ntp.org
>>iburst is good.  burst, very likely not.
> In general, I agree, but the context is for a connected reference clock.
> I would set maxpoll to 4 as there is no reason NOT to update from the
> reference clock on a frequent (16 second) basis, but I'm less certain if
> iburst is really appropriate. (Nor am I sure that it's inappropriate.)

More information about the questions mailing list