SCTP for multihoming

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

Lode and others,

I've been reading the SCTP for multihoming draft. Here's my feedback.

First a more general issue that comes up in the SCTP-related drafts: 
there is no way for two hosts to find out whether the paths that exist 
between them share any infrastructure. Would it be useful to see if we 
can come up with something that allows hosts to find out? I'm thinking 
something in the way of the record route option.

I'm worried about the much more complex packet format (compared to TCP) 
that SCTP has, such as different chunks that can be in a single packet. 
Are there any figures available about SCTP implementation complexity 
and performance?

I've been opposed to a strict transport layer solution because then all 
transport protocols need to be changed. Still, SCTP is here today so 
why not use it for multihoming in the situations where it suits the job 
at hand. Another problem with SCTP is that moving applications from TCP 
to SCTP means both protocols will have to run side by side for a long 
time. Would it make sense to modify SCTP such that it is possible for 
applications that expect to run TCP to be multihomed without changes, 
and also to be backward compatible? The latter could be done by 
exchanging some "I can do SCTP" options in TCP and then (try to) set up 
an SCTP session if the other also supports it, or even better, switch 
to SCTP in mid-flight.

And I still think the heartbeats are both evil and most of the time 
unnecessary.

BTW, it would be awesome to have HTTP over SCTP where each of the 
elements that make up a web page is loaded over a different stream so 
there is no head of line blocking but the session establishment 
overhead (both in unnecessary repeating data and in round trip times) 
that exists today is mitigated.

Now that we're firmly in science fiction territory (Tony Hain once 
suggest we form an "Internet Fantasy Task Force", maybe not such a bad 
idea...) it occurs to me that SCTP could be wrapped around UDP for 
long-lived UDP associations and subsequently add multihoming to UDP and 
UDP-based transport protocols such as RTP. At least, if the stream 
setup mechanism can be changed to support "one packet streams" 
efficiently. What do you think?