[e2e] traffic engineering considered harmful

[David G. Andersen]

Ben Black just mooed:
> 
> Perhaps this is naive, but why not have the natural traffic aggregation
> points (read: servers) collect the performance data?  Trying to have
> end points collect and disseminate the information seems nightmarish.
> 
> Hosts wishing to learn the best path to a given destination would have
> to either carry full BGP tables, be able to query an entity with full
> BGP tables, or trust an intermediary (as Randy mentioned).  

   Not really.  It depends how anal you are about "best."

   Routing information is necessary for the functioning of the Internet;
performance data is not.  Therefore, in some sense, you really only
need to maintain performance data for entities with whom you
communicate, and it's soft-state.  It makes some sense to maintain
this information at the host-level, becauses hosts know with
whom they're communicating.  It also makes sense to share
the information between hosts on the same LAN or within the
same AS, of course, but it's not intuitively clear how this
best balances out - the extra overhead from each person maintaining
more state vs. the cost and complexity of sharing the information.

> > and maybe the agents could be near the edge of the site and make the
> > forwarding decision for the hosts at that site.
> > 
> > darn, this sure reminds me of something i have seen before.

   :-)

  I have a pet theory that you can "optimize" (e.g. exhaustively
search the routing and path space) the routes for a small
collection of hosts pretty easily, but as you start trying
to scale this up to all-to-all communication, then you're trying
to re-create BGP, and you'll probably make some of the design
decisions that result in the problem in the first place.

   I'd love to be proven wrong on that.

   -Dave