Re: multihomed host

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

On Wed, 9 Oct 2002, Yoshihiro Ohba wrote:

> > This is possible as of today; a host that has multiple PA addresses is
> > certainly considered a multihomed host. However, this alone is not a
> > multihoming solution.

> What additional things need to be considered for this usage to become
> a multihoming solution?

A good multihoming solution keeps working if there are failures. So the
applications must be smart enough to try all the addresses rather than
just one. Telnet and FTP do this, WWW clients typically don't.
Unfortunately, this must be done at the remote end and not on the
multi(homed|addressed) host itself.

When there is a session, and the path over one ISP becomes unavailable,
the multihomed host must switch to the path over the other ISP. There are
many problems with this. First of all, the failure must be detected. Then
the host can reroute outgoing packets. But if the host still uses the same
source address (from the address space from the failed ISP), it is likely
the packets won't be accepted by the alternate ISP because the source
addresses are "spoofed". And even if they are, the communications partner
keeps sending packets back to th source address tied to the first ISP,
which isn't reachable any more.

One solution for the address problem would be the adoption of new or
improved transport protocols that can handle this. There have been
experiments with modifications to TCP to support this, and the new SCTP
protocol can also handle this. However, modifying all transport protocols
might be problematic.

Another solution is tunnelling/aliasing the packet to hide the original
address while the packet is rerouted over the alternative path. The
destination host or a box very close to the destination host strips of the
tunnel header or replaces the original address so TCP and other transport
protocols don't see the change in address.

Some of us are working on this stuff on a non-IETF mailinglist. The latest
idea is to come up with a common tunnelling/aliasing framework so
different solutions in this area may interoperate.