DNS Query Behavior with Global Forwarders Statement
Mark Andrews
Mark_Andrews at isc.org
Thu Aug 14 01:00:27 UTC 2008
> On 12 Aug 2008, at 23:12:18, Mark Andrews wrote:
>
> >
> >>> Is this an artifact of the -P2 changes or was the use of RTT dropped
> >>> for some other reason?
> >>
> >> You didn't say which version you were running.
>
>
> Our NMS systems tend to be running BIND 9.3.5-P1. The -P2 rollout is
> in progress. There is a smattering of Engineering name servers in the
> environment that are running BIND 9.4.2-P2.
>
>
> >> I'd be quite surprised if this were an artifact of the -P1 and -P2
> >> changes. I'd be less surprised if it were a bug introduced in 9.5.0.
> >>
> >> --
> >> Evan Hunt -- evan_hunt at isc.org
> >> Internet Systems Consortium, Inc.
> >
> > RTT estimates work well with authoritative servers.
> >
> > The best I expect to get out of rtt and forwarders is whether
> > the forwarder is up or not. If there was some way to
> > accurately work out which answers come from the cache and
> > which first required recursion then rtt times would home
> > in the closest forwarder.
> >
> > It takes 110ms to get a answer from A via F1 and F2.
> >
> > R 10ms F1 100ms A
> > R 100ms F2 10ms A
> >
> > Cache talking to caches don't always get enough second level
> > cache hits to make a statistical difference to the rtt estimate.
>
>
> If I understand you correctly, BIND's RTT is based on how long it
> takes to get an answer to the DNS query. The fact that a NMS and a
> regional name server are in the same rack has little bearing on the
> selection process if the majority of the time is spent querying the
> Internet.
>
> This leads to another question. Is the selection of forwarder in
> anyway affected by what is being asked for in the query? If one
> regional name server could resolve queries related to CNN.COM
> significantly faster than the others, would it be favored for
> resolving queries related to that zone?
In general no. The rtt estimates for all nameservers for
a zone are updated as the result of performing a lookup on
the zone. The server that responed is updated using the
rtt for the query as a component of its new rtt estimate.
Servers that were queried but failed to answer are penalised.
Servers which were not queried have their rtt estimated
improved.
If you have overlapping sets of servers then queries for
one zone can impact on which server is choosen to another
zone which uses some of the servers involved in the first
zone.
> Also, should a regional name server be inaccessible due to a network
> outage how much of a delay might there be in using it after service is
> restored?
I havn't done the caculation. The delay would depend on
lots of things. The simplest way to find out would be to
do a experiment where you make a server unavailable to a
client for a amount of time then make it available again
and look at the query profile.
Note the time to flip would shorten if the working server
stopped working when the other server was brought on line.
Non-prefered servers are tried relatively quickly when other
previously prefered servers die.
Mark
> Merton Campbell Crockett
> m.c.crockett at roadrunner.com
--
Mark Andrews, ISC
1 Seymour St., Dundas Valley, NSW 2117, Australia
PHONE: +61 2 9871 4742 INTERNET: Mark_Andrews at isc.org
More information about the bind-users
mailing list