RE: Fwd: Minutes / Notes

["Tony Li" <Tony.Li@procket.com>]

|    I don't know about most implementations. We definitely 
|    need a solution
|    that is "upward compatible" with the current implementations, so we
|    should assume that there will be co-existence for a long 
|    time. And I
|    could definitely see why some implementations would just decide to
|    forget the benefits of multi-homing, or at least those of TCP
|    survivability. But I am concerned mostly with the tax on 
|    applications
|    that run on a multi-homed system, yet are perfectly happy dealing
|    directly with addresses/locators. I would assert that this 
|    is the bulk
|    of current applications.


Back in Ye Olden CIDR days, we estimated that about 10% of all sites
were multihomed and that the most common ones were large enterprises.
Those same large enterprises typically run a large variety of different
OS's, so to provide this benefit, we clearly need to hit a high 
percentage of the v6 implementations.

The size of the tax is not high.  If we work it so that we exchange
identifiers as a TCP option only on SYNs, then we're talking about
something like 40 bytes of additional overhead.  And this is in
a packet that is normally quite small.

I realize that there are some applications out there that could
be munged to handle addresses and locators.  However, changing them
one at a time is both architecturally ugly and very unlikely to
let us make substantial forward progress.  While this may well be
the bulk of current applications, our job is not to architect for
today's applications.  It's to deal with tomorrow's.

Tony