Server names for query

[Kevin Darcy]

Casey Deccio wrote:
> On Mon, Mar 23, 2009 at 3:20 PM, Kevin Darcy <kcd at chrysler.com> wrote:
>   
>> For the *initial* NS query, I believe BIND will resolve those names down to
>> a flat set of addresses, all of which have equal chance of being tried, so,
>> yes, if a given NS name resolves to more addresses than other names, it is
>> more likely to be tried on the initial NS query.
>>
>> But that's just the *initial* NS query. Once BIND, and/or virtually any
>> other full-resolver implementation, builds up a history of how fast each
>> nameserver responds (based on round-trip-time or RTT), it will start using
>> nameservers which respond faster (although there is some "banding" that
>> occurs, so that nameservers which respond more-or-less at the same speed get
>> tried equally often). So if the point of your question is to try to control
>> the distribution of query load to nameservers, be aware that this will be
>> determined much more by the speed at which they respond, respectively, to
>> clients, than to how the NS names are organized. Clients gravitate to faster
>> servers. If the extra volume causes the fast servers to bog down and be
>> slower, then clients gravitate away from them, and some sort of query-volume
>> equilibrium is achieved between all of the nameservers which are published
>> for the zone. In a sense, it is "auto-tuning" in this regard.
>>
>>     
>
> Thanks, Kevin.  My analysis is more based on collective queries (i.e.,
> from clients in diverse geographical/network locations), rather than
> from any particular client.  So the point of my question is
> determining whether it would be expected that the *total* number of
> queries from different clients for a given domain would be distributed
> roughly proportional to the number of A records corresponding to each
> NS target in the NS RR set for the domain.
>
>   
It all depends on where the clients are and how quickly the nameservers 
can answer queries for the zone. A nameserver that can answer more 
quickly for a given community of clients will attract more queries from 
those clients, over time. It's only the initial query that can really be 
"weighted" in the way you describe. After that, the clients auto-tune 
their nameserver selection choices according to load, network latency, etc.

- Kevin