On Mon, Jul 11, 2022 at 12:46 PM, heasley <[email protected]> wrote:

> Sat, Jul 09, 2022 at 09:32:51PM +0200, Robert Raszuk:
>
> And here comes I think need for authors to clarify something. They said
> that such marking is going to be used along with NO-EXPORT.
>
> no, you might have misread that; with no-export is possible but not
> necessary.
>
> Eg: several (perhaps all now?) of the root servers lie within anycast
> prefixes.
>

All.

These should not be marked no-export, but *could* be marked ANYCAST.
>

Many of the root server operators have the concept of "local" and "global"
nodes.
If the root server's address is e.g 10.0.0.17 [0],  global nodes will
announce 10.0.0.0/23 and local nodes announce 10.0.0.0/24, but tagged with
NO-EXPORT. This will[1] make it that networks that the node peers with
directly will use this local node, without having the great unwashed
hitting that node - this allows scaling the nodes appropriately.  If the
node falls over, traffic will simply follow the shorter /23.  This is much
better explained in ISC's "Hierarchical Anycast for Global Service
Distribution" -  https://www.isc.org/pubs/tn/isc-tn-2003-1.html ,
https://www.caida.org/catalog/papers/2007_dns_anycast/dns_anycast.pdf ,
etc..

Many nodes might be both global and local - they announce the /23 and /24,
so that networks peering with the node prefer this particular instance over
some other instance. Every now and then some network will ignore or strip
NO-EXPORT and announce the "local" prefix - this suddenly makes that node
the "bet" (longest-match), and hilarity ensues..
BGP does have many knobs that can be twiddled, but unfortunately some of
the inbound ones are still fairly heavy hammers...

W
[0]: Grrrr. The IPv4 Documentation prefixes are all /24s, and I needed
something larger, so just pretend, 'k?
[1]: Ok, should...
_______________________________________________
GROW mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/grow

Reply via email to