[e2e] traffic engineering considered harmful

[Erik Nordmark]

>  >>In addition to Randy's point about navigation there is also
>  >>the issue that exploring the available paths to find a good or better
>  >>path isn't free - it requires sending some number of packets.
>  >>And each host will do this exploration independently.
> 
> funny, isn that what TCP does on a single path right now

But those TCP packets are also used to carry data.
If you want to measure N differents paths in parallel a per-host solution
would need to send extra packets to measure the other paths. And to
keep the information current this needs to be done every RTT.
Of course, if you're thinking about having different hosts use different paths
and share information then you don't have this issue. Merely a small
issue of scalable information sharing of information that will be stale
in about one RTT.

>  >>I guess one could envision agents (e.g. in each site) that would do
>  >>this type of exploration and be willing to share their information
>  >>with hosts in the site. But that raises issues about stale information.
> 
> there are a lot more hosts than routers and a lot more flows than
> routing neighbour state entries - ergo, we already haev a lot of this
> data and we jjust need to think a bit better about coordinating it

And aggregating it.
If the hosts have data about www.example.com (really about the IP address
and not the name) how can the hosts tell whether this also applies
to www2.example.com?

   Erik