Source address selection in IPv6 multihomed multi-addressed sites

[Cedric de Launois <delaunois@info.ucl.ac.be>]

These are my random thoughts about source address
selection in IPv6 multihomed multi-addressed sites.

I believe this problem needs to be resolved to get
a complete IPv6 multi-addressing multihoming solution.
We assume here that the providers perform ingress
filtering.
I'm here only looking to solutions where the source
address selection is not made by applications, but it
is done "automagically". However, for all these solutions,
it is possible to imagine some mechanism where the
applications can overload the automatic behaviour.

Here is my understanding of the different solutions
proposed.

* A solution is to perform source address based routing
  inside the multihomed site. The packets are always
  routed to the site exit router corresponding to the
  source address used.

  Pros: - no new option/protocol
        - no modification to hosts
        - no tunnel
        - routing efficiency (dumb routing) ?

  Cons: - possibly very bad load-balancing if all the
          hosts choose the same source address (and this
          is what they'll do with RFC3484 !) +
          no way to engineer the traffic if no additionnal
          mechanism is used
        - source address based routing has probably
          much more implications than what it is
          expected. This needs to be studied further.

  My point of view : maybe source address based routing
  is not something desirable. Anyway, source address based
  routing alone is not sufficient.

* Another solution is to establish tunnels between the
  site exit routers. When a packet with a wrong source
  address is received by a site exit router, it is
  tunneled to the right exit router.

  Pro : no modification to hosts
  Cons: - a tunnel, reducing the MTU
        - no way to engineer the traffic

  My POV: I think this solution is too rigid and
  that a solution based on an ICMP source redirect can
  lead to a better solution without adding much more
  cost.

* A third solution is that a host first try a source
  address (at random?). If the source address is not
  correctly chosen (e.g. the link is down), then a
  new source ICMP redirect is sent by the site exit
  router to the host.
  This ICMP message tells the host which source address
  to use.

  Pro : no tunnel
  Cons: - modifications needed on the hosts
        - at first look, difficult to engineer the
          traffic with this mechanism

  My POV: I think this solution is the most credible
  when the site is small and doesn't need complex
  policies. However I think the ICMP source redirect
  message should also contain a prefix associated to the
  source, like in the NAROS response messages. This would
  allow the host to fill its policy table step by step.
  This approach has to be reserved for small sites with
  few requirements.

* A fourth solution is that a service (NAROS) tells the
  hosts which source address to use.

  Pro : - the most complete solution since it allows any
          kind of traffic engineering
        - no tunnel
  Cons: - also the most complex solution since it
          requires changes in the hosts and to set up
          a new service
        - implies source address based routing

  My POV: isn't that too complex for small sites ? Maybe.
  Is there a need to manage complex requirements is
  host-centric multihomed sites ? Probably. Anyway, the
  solution has the advantage of being rather complete. At
  the price of complexity.

* A fifth solution is that the hosts know which source
  address to use thanks to DHCP. When they boot, they
  receive a full policy table from their DHCP server.
  This policy table contains informations about which
  destination is accessible through which site exit router
  etc.

  Pros: - no tunnel
        - easy configuration
  Cons: - requires a new DHCP option. Modifications to
          host are needed
        - complex traffic engineering not supported
        - no dynamic, how to react to a failure ?
        - implies source address based routing

  My POV: seems to rigid for me. At almost the same price,
  we can get the same features + dynamic with the ICMP
  solution.

* A sixth solution is that a host tries to detect
  by itself which source address to use. A proposition
  is that a host maintain by itself one or more full BGP
  tables so that it can deduce the source address to use.

  Pro : no tunnel
  Cons: - waste of computer resources (how much ?)
        - no easy configuration


Any comment/suggestion/correction ?

Cédric