[ntp:questions] Query about NTP accuracy
davehart at gmail.com
Fri May 22 10:15:58 UTC 2009
NTP includes error budget calculations at every level. You should be
able to demonstrate the difference between directly using a stratum 1
appliance compared to using higher-strata intermediate NTP servers
simply by observing the root dispersion on the client.
I suspect you are correct that the benefit is not worth the cost, but
convince yourself with measurement. Using FreeBSD instead of Linux
for the intermediate servers will probably give better results, if
that is a possibility, give that a test as well.
I would also encourage you to consider using NTPv4 in preference to
NTPv3 where it is practical, but especially on the intermediate
servers. A lot has improved in NTP since version 3.
Finally, while it is true you can improve scalability using broadcast
or multicast, it also requires extra configuration to provide matching
keys on both sides, or using the "disable auth" option which is not
recommended. NTP uses very few resources even with hundreds or
thousands of unicast clients, as each client polls at most once per
minute and typically, after stabilizing, once per 17 minutes (1024
seconds for default maxpoll 10). Processing each packet and replying
is very fast, so even after a power outage when many clients are
starting at once, I would expect even relatively decrepit hardware to
have no problem keeping up with the traffic. With 10,000 clients
polling a single server every 64 seconds, the average rate is 156
packets per second in and out each 48 bytes long. I bet your existing
servers could handle that rate easily.
More information about the questions