[NANOG] OSPF minutia, and, technote publication venues
Steve Gibbard
scg at gibbard.org
Tue May 6 10:09:43 UTC 2008
On Tue, 6 May 2008, Nathan Ward wrote:
> This stuff about customers and things sounds too hard.
>
> Steve, have you actually had to do anycast without having control of
> the routing hop in front of your service providing hosts, or is this
> getting unnecessarily complicated? I'd imagine that the ability to
> install routing equipment would be a pre-requisite for any anycast
> service deployment..
Yes I have. Or rather, I've done the network infrastructure for anycast
services without having administrative control of the anycasted servers.
PCH's anycast platform hosts some blade servers for some other DNS
infrastructure operators (in addition to the name servers PCH operates
itself). Those operators operate their own servers. PCH operates the
routing infrastructure. There is filtering in place to limit the routing
announcements from the servers.
But also, most of the larger organizations I've worked for have had
separate systems and network engineering groups. In general, the network
groups haven't wanted to let the systems engineers configure the routers,
and the systems groups haven't wanted to let network engineers configure
the servers (with good reason). Filtering of routing announcements from
anycast servers would be useful in that environment too.
To address Paul's point about multipath BGP, I never saw Cisco's
implementation of it causing a problem even with full routing tables. I
haven't used any other implementations.
In the Cisco version (and at least for EBGP; I haven't looked at this with
IBGP), it only applies to otherwise identical AS paths. Multiple
directly-connected DNS servers sourcing the same announcement with the
same AS path and other BGP attributes get load balanced between. Paths
learned from different peers had different AS paths and do not get
balanced between. I suppose there probably is load balancing in cases
where there are multiple sessions with the same peer at the same exchange.
That's a relatively rare case in this implementation, and using hash based
rather than per-packet load balancing makes it not really matter.
-Steve
More information about the NANOG
mailing list