Intradomain DNS Anycast revisited

Wed Mar 30 18:10:30 UTC 2005

      On Fri, 25 Mar 2005, Joe Shen wrote:
    > Do you mean Quagga's OSPF route has higher priority
    > than static route?  or even there is static default
    > route configured, once Quagga detects link to default
    > router is down it will replace  0.0.0.0/0.0.0.0  in
    > host routing table?

If you're using dynamic routing (whether it be OSPF or iBGP) to distribute 
default routes for fail-over, yes, you need to make sure that any statics 
you also have are at lower priority.  One way of playing it safe would be 
to not have static _defaults_, but to only have static routes to your 
internal management networks.

    > > Nope, no problem, particularly so long as the two
    > > routers are iBGP peers, 
    > > so they'll both (for the most part) have the same
    > > idea of what selected 
    > > paths are.
    > 
    > I don't understand why should both routers be iBGP
    > peers.  In fact, iBGP does not run on that two
    > routers; the two routers are only members of  OSPF
    > backbone area who only run OSPF; only  border router (
    > at the edge of our network) runs BGP and enject
    > default route into OSPF backbone area. 

Ah, you're correct, there's no requirement for the routers to be iBGP 
peers or to run BGP at all...  If you're doing this principally as an 
intranet thing, you might not have any BGP speakers nearby, or any need 
for BGP.  I've usually done it as a service provider, which meant that the 
point was to have the servers as close to the rest of the world as 
possible, which means directly adjacent to an AS-edge BGP speaker.  But 
you're quite right.

    > if that possible that router3 or router-1 dispers
    > packets of the same TCP connection to different path? 

Depends upon the equal-cost-path load-balancing algorithm that the routers 
are using.  You want to select a source-destination-hash one, to avoid 
that issue.

    > Is there possibility that a DNS requests are divided
    > into multiple UDP packets?

No.  Not unless they hit an undetected MTU below 576 bytes, or whatever it 
is...  Any DNS packet which can't fit inside a single UDP packet is 
supposed to be sent via TCP instead.  Note that I'm a network guy, not a 
DNS guy, so this is possibly an oversimplification.

    > Do I only need to configure BIND to origin request
    > from administration IP address ( configured on NIC and
    > different from DNS service address)?

You mean for requests that the anycast server is making of other servers?  
If it needs to originate a zone transfer, or perform recursive lookups?  
Yes, those need to originate from the unique/administration address, as 
opposed to the shared/service address.

                                -Bill