RE: Identifier/locator recap

[Iljitsch van Beijnum <iljitsch@muada.com>]

On Wed, 19 Mar 2003, Erik Nordmark wrote:

> At some level one could argue that host multi-addressing is scalable
> (at the expense of moving complexity to the hosts and applications).but
> my concern is that host multi-addressing will more or less have
> the hosts track with source/destination combinations (which approximate
> routing paths at some rough level) work vs. doesn't.
> To get fast failover this essentially turns into doing end-to-end "hello"
> traffic instead of relying on the local hello traffic performed by the routing
> protocols. So I think there are severe limitations in making this scale.

There is one big fat advantage to having this in the hosts: those
already track the end-to-end status of sessions.

Doing it in middleboxes is more complex, but it should be doable without
spending too much bandwidth on reachability tests: a middlebox typically
communicates with many other middleboxes (or hosts) and many of these
sessions share a significant amount of infrastructure. So rather than do
a reachability check for each session every 90 seconds (yes, this is
very long but it's the same as the BGP hold time listed in RFC 1771 and
Cisco even uses 180 seconds by default) the middlebox can simply do a
reachability check for say 1/300th of all sessions each second. That
means each sessions will be checked every 5 minutes. But as checks start
to fail, their number is increased, starting with ones that fall inside
the same 32. The effect is that the closer the failure is to the
middlebox, the faster it will be detected. Only when the path between
two middleboxes shares no infrastructure with other paths, and this
non-shared infrastructure fails, it may be necessary to wait the full
300 seconds.

Iljitsch