TELEHOUSE America & Internet Software Consortium Develop DNS F-root Server in New York & Los Angeles

Tue Feb 11 14:38:13 UTC 2003

On Tuesday, Feb 11, 2003, at 07:50 Canada/Eastern, Robert E. Seastrom 
wrote:

> Charles Sprickman <spork at inch.com> writes:
>
>> On Mon, 10 Feb 2003, Paul Vixie wrote:
>>
>>> Deal Enables ISC to Mirror DNS Root Server in Additional U.S. 
>>> Locations
>>
>> Let's hope Telehouse put them on the "good" generator.  "N+1" is no 
>> fun if
>> the "+1" can't be routed to the 5th floor when "N" chokes up.
>
> All is well if the router that announces the network is plugged into
> the same circuit (or if the announcement comes from a BGP speaker on
> the box itself).  No big deal to lose a single root anyway, but this
> scenario would keep F "working as advertised", so to speak.

[Apologies to Suzanne for pre-empting her discussion about this.]

Each F-root node is carefully designed so that most failures which 
could stop a nameserver answering queries are reflected in the network, 
both within the F-root node, and within the F-root's service area. If a 
nameserver within a node is not available, the node will not send it 
queries; if all nameservers within a node are not available, the node 
will stop advertising 192.5.5.0/24 to its local community of peers, who 
will stop sending queries to the node.

The potential for global instability in (and corresponding dampening 
of) 192.5.5.0/24 due to some oscillatory error condition in a 
particular node is limited by the fact that each non-Palo Alto node 
advertises 192.5.5.0/24 to peers only, and precautions are taken to 
limit the propagation of that prefix through peer networks. Only the 
Palo Alto node advertises 192.5.5.0/24 for global transit.

If a local F-root node withdraws service, resolvers within its 
catchment area will see the BGP path to the global F-root node in Palo 
Alto exposed and selected. The change in relative RTTs will then cause 
resolvers (BIND-like resolvers, anyway) to reorder their ranking of how 
close the 13 root servers are, and referrals to the root from the 
catchment of the dead node will tend towards the new closest server, 
which may or may not be F.

Hence, a failure of a restricted-anycast node restores the usual 
availability of root servers -- it effectively just removes the local 
optimisation that the anycast node was providing.

Joe