[ntp:questions] Re: Question about min/max poll and an interesting plot showing maxpoll's effect
jpp at cloudview.com
Thu Apr 20 22:11:21 UTC 2006
Richard B. Gilbert wrote:
> Ntpd is supposed to set it's own polling interval; something to do with
> the "Allen Variance". My feeble understanding translates into English
> as "short poll intervals allow ntpd to correct large errors quickly and
> long poll intervals allow ntpd to correct small errors accurately".
> Anyway, I'm a little surprised that shorter is better. I have two
> machines as clients of my stratum 1 (GPS) server. All three are Sun
> workstations; the server is an Ultra 10 440 MHz, one of the clients is
> also an Ultra 10 440 MHz the other is an Ultra 5 360 MHz. All have been
> running for several days. The Ultra 5 is polling at 64 second intervals
> with an estimated error of 11056 us while the Ultra 10 is polling at 256
> second intervals with an estimated error of 500 us.
> The proper poll interval appears to be highly dependent on the quality
> of the local clock.
My experience is that shorter is better with commodity hardware that is subject to rapid thermal
cycling - my home machines are in the basement which also houses the furnace - I can see the furnace
12 min on/off cycle in the frequency offset graphs. Equally when periodic compute intensive jobs
run (one of my S1's is also a file/backup server) the resultant thermal shock to the machine pushes
it way off - if I let ntpd poll long (or omit flag 3 on refclocks) it spends all it's time chasing
the thermal cycle and never quite makes it. Will a poll of 64 (or less on my own servers) it
catches it before it gets 250us off and corrects. My co-lo machines which have a much more even
load and are in an cabinet in a machine room show a dinural cycle that ntp copes with pretty well
even at a poll interval of 1024.
More information about the questions