Re: A tunneling proposal

[Ramakrishna Gummadi <ramki@cs.berkeley.edu>]

> Tunnels can fix some types of failures, but they can't fix every type of
> failure, like routing problems that affect the entire ISP. This kind of
> multihoming is clearly inferior to the regular IPv4 way.
>

One of the  primary goals of tunneling, I think, is to make *existing*
connections that use transports such as UDP and TCP survive a partial
outage within the ISP.
I hypothesize that the only scenario where tunneling for TCP
and UDP fails is when the entire ISP is affected in a major way.
If one uses SCTP, or some other transport with
ability to support multiple addresses per connection, one can deal with
this problem completely.


True, regular IPv4  multihoming can provide this facility only because
end-hosts use a
single non-aggregatable address that is announced in the DFZ.
If hosts were to use multiple addresses, regular IPv4 multihoming can not
provide connection survivability for TCP and UDP even in the face of a
partial outage.

It all depends on what the first priority is---if it is scalability, then,
I think, to make TCP/UDP survive all failures, one needs an approach like
GxSE, which affects the IP layer. If you are willing to be less than
perfect, and don't want to change the host stacks, tunnels may be a good
choice, especially when applications can use SCTP to get the
required reliability.

-ramki