Adrian von Bidder wrote:
On Monday 09 January 2006 21.34, Charles Swiger wrote:
[ ... ]
New connections to the NTP pools might not be load-balanced properly
if the outage persists for a long time, but it would probably need to
be down for a period of several days before it would make much
difference.
No, the load spikes during the time when a timeserver is in the DNS rotation
is quite remarkable. No zone file update for only a few hours already
means that some pool servers on weak lines might actually have problems.
Ah. I believe I understand why; the DNS responses for requests to the pool only
return 10 or so responses (what fits into a 512-byte UDP response?), rather than
returning potentially hundreds, so the 10 or so listed get all of the new
traffic until the rotation is fixed.
What problem are we trying to solve?
In general: one easy to remember DNS name for time services, usable by
anybody without special requirements wrt time quality.
Specifically: distribute the resulting load on as many servers as we can
get.
The second part is where we're currently doing it by more or less abusing
DNS. DNS is a mapping from names to numerical addresses, but it was never
designed for the kind of load-balancing we're doing here. What we do
works, kind of, but IMHO it's worth keeping in mind that tweaking DNS is a
kludge I've started because I am too lazy to design a proper solution,
which would be integrating support for server pools into ntpd directly.
One might be able to improve this by having "pool.ntp.org" and other relatively
full domains (Europe, NA) be CNAMES which point to 0.pool.ntp.org,
1.pool.ntp.org, and so forth, which would let you spread the load-balancing
between 30+ boxes rather than just 10...?
--
-Chuck
_______________________________________________
timekeepers mailing list
[email protected]
https://fortytwo.ch/mailman/cgi-bin/listinfo/timekeepers
