OMB: IPv6 by June 2008

Thu Jul 7 15:11:20 UTC 2005

On 2005-07-07, at 10:23, Andre Oppermann wrote:

> It was about a spot in the global routing table.  No matter if one  
> gets
> PA or PI they get a routing table entry in the DFZ.  There is no  
> way around
> it other than to make the routing protocols more scaleable.

With the hole-punching/CIDR abuse multihoming that is widely used in  
IPv4, a slot in the DFZ gets burned each time an end site adds a  
provider, regardless of whether they are using PA or PI addresses.  
This slot represents state information for the multi-homed site which  
answers the question "how else can this set of addresses be reached?"

The shim6 approach shifts this state from the DFZ to the endpoints  
which are exchanging unicast traffic. The endpoints exchange a set of  
possible locators through a protocol element within the IP layer and  
handle locator migration transparently to the transport layer above.  
Hence the question "how else can this particular remote address be  
reached" is answered using information on the host, not information  
in the network.

With shim6 an end site can multi-home using one PA prefix per  
provider, without taking up additional slots in the DFZ. Hosts within  
the site are given multiple addresses (locators), and the layer-3  
shim handles any change of locator needed for traffic exchanged  
between any two hosts.

If one (or both) of the hosts exchanging traffic don't support shim6,  
then the traffic is exchanged without transport-layer stability  
across re-homing events (and, potentially, without any optimisation  
as to the choice of endpoint addresses for the session).

So, the shim6 future of multihoming looks like this:

1. ISPs multi-home exactly as people are used to doing today, using  
PI prefixes, and taking up a slot in the DFZ per transit provider.  
Everybody is familiar with this already. There is no change for ISPs  
in this picture.

2. Multi-homed end sites obtain one PA prefix per upstream ISP, and  
hosts within those end-sites are assigned multiple addresses (in some  
automated, secure and controllable fashion). There are no additional  
slots burned in the DFZ by end site multi-homing. Hosts obtain  
transport-layer reliability across re-homing events using shim6,  
rather than relying on the network to take care of it.

Joe