[ntp:questions] Tighter regulation?

David Woolley david at ex.djwhome.demon.invalid
Sat May 25 12:57:27 UTC 2013

Mischanko, Edward T wrote:

> I would modify the current algorithm with an exception that if offsets
>  exceed 1 millisecond for more than one polling cycle, then, polling will be
>  reduced by one interval, else, continue normal operation.

Whilst I still think the OP is trying to make the statistics look good, 
rather than have good time keeping, I can sort of see two feature 
requests here:

1) a tinker option to specify an assumed maximum jitter, which will 
allow the user to put in a long term jitter statistic for their system, 
rather than having ntpd work it out for itself.

2) a tinker option for a panic level of offset, where, it is assumed to 
be so improbable, given the specific noise statistics of the actual 
system, that it should immediately be considered to be due to a local 
clock frequency error.  Measurement noise is often non-gaussian, so it 
unlikely to be safe to make this a fixed multiple of the value in item 
(1).  If this is to be used without filtering by the minimum delay 
filter, it will need to be a lot higher than the filtered value, as the 
minimum delay filter will take out most one off spikes in offset.

I am not volunteering to write the code for these, and I do have concern 
that an urban folklore will develop in which in become it becomes common 
advice to set them excessively low.

However, I still feel that, if the underlying problem is clock frequency 
drift, the primary fix is to address the cause of that drift, and the 
stop gap solution is to reduce maxpoll.  If it is almost anything else, 
the offset changes do not indicate a timekeeping error and should not be 
acted upon.

More information about the questions mailing list