On Sunday, February 24, 2013 at 17:30, Hal Murray wrote: > Harlan is getting ready to release the next version of the standard NTP > package.
Great! > The pool command seems to work. How many of you are testing it? > > I think clients using the pool should occasionally check to see if an address > > they picked is still in the pool and pick another one if it has dropped out > of the pool. (or at least stop using the old one) Yes, that would be very nice. There are many reasons a server isn't appropriate to use anymore even if it's still working. For example if a server operator wants to stop providing the service but wants to do it gracefully, he can keep running the service but (unlike now) know that the traffic will stop over a reasonable timeframe (weeks to a few months). > Is there a way to do that? I assume it would use DNS. There isn't. I could setup a "DNSBL-style" zone to check "is this server still ok to use", but I don't think that's really what we need. For example then a server operator might now be asking for less queries per second; or maybe the server used to be in the NL zone, but now is in the DE zone so a server in the Netherlands shouldn't get it anymore, but a server in Germany should still. I think the simplest and for the pool most effective strategy would be if ntpd just replaces the "oldest" server every 7-10 days. With 4 servers that will mean they all rotate every 4-6 weeks which seems reasonable in terms of instability caused and would be a huge improvement over the current state. This would be in addition to replacing "bad" servers already. A more sophisticated algorithm would be to replace the worst server (even if it's okay) every week, but also consider "age" to be bad so all servers are replaced no less than every 60 days. I don't think it's a good idea though; it will be much less predictable and it'll also skew traffic harder to servers that are working best which isn't really helpful in terms of preserving the common good. > How are we going to test this? Would it be reasonable to setup a batch of > systems that are already in the main pool and put them in something like > test.pool.ntp.org (http://test.pool.ntp.org), and rotate them in/out fast > enough so it would be easy to > test? Say in for an hour, then out for a few days. In most countries pool.ntp.org "rotates" every few minutes already. :-) (Because it gives different answers to each query). Ask -- Ask Bjørn Hansen - http://askask.com/ _______________________________________________ pool mailing list pool@lists.ntp.org http://lists.ntp.org/listinfo/pool
