In message <200204071533.LAA21149@ginger.lcs.mit.edu>, "J. Noel Chiappa" writes
:
> > From: Tony Li <tli@procket.com>
>
> > To generalize this, the point is that the components of the
> > architecture should be cleanly partitioned into subsystems. These
> > subsystems must be as loosely coupled as possible and that it must
> > be possible to replace subsystems incrementally.
>
> Ah, sometimes that's not really possible, the most difficult case often
> being naming systems. E.g. replacing the current IPv4 addresses...
>
> > An obvious question then is "what are the appropriate subsystems?"
>
> Hah. That's known.
>
> 1) Topology element naming.
> 2) Topology information distribution.
> 3) Path computation.
> 4) Path setup.
> 5) User packet forwarding.
>
> Some of these have to be done in a system wide way (e.g. naming), some
> allow of some local tweaking (e.g. information distribution), and some
> are almost inherently local (e.g. path computation).
>
> Noel
Those are subsystems of a specific architecture, one that exists
mostly on paper and in a few research projects.
Not all architectures use a path setup before path forwarding and it
is arguably a bad idea in many cases. For example, path setup before
a single DNS UDP packet to port 53 is not part of the current
architecture and would be inefficient. I'm not arguing that the
current architecture is better, just that your "list of subsystems"
constrain the architecture (to one you'd like to see adopted).
Curtis
This archive was generated by hypermail 2b29 : Mon Aug 04 2003 - 04:10:04 EDT