Path selection by hosts, shudder [was: Re: The state of...]

[Simon Leinen <simon@limmat.switch.ch>]

Craig,

I absolutely agree with Ohta-san's response to yours - I see this
issue as fundamentally similar to the bandwidth management issue.

Bandwidth is obviously something that IT people want to manage, but
yet this is done on the end systems on the Internet (in TCP).

Of course some may argue that this has failed, because there sometimes
congestion somewhere on the Internet, and IT people are forced to buy
rate shapers and other middleboxes to gain the control back from end
systems, and lots of other people have spent a lot of effort to devise
various ways to improve the QoS of the network, most of them involving
making the network handle "less-important" traffic less well.

But the fact is that for many (most?) people the Internet works or
would work well even without all these things.

I think there are good ("end-to-end") arguments for considering to
leave the path selection decision to end systems too.  For instance,
applications run on end systems, and are in a good position to choose
a path that fits their requirement best (cheapest/quickest
response/highest bulk throughput...).

On Tue, 22 Oct 2002 12:02:54 -0500 (Central Daylight Time), "Craig
A. Huegen" <chuegen@cisco.com> said:
> When the host selects the path by virtue of address selection, such
> that the routing infrastructure can see no alternate paths to the
> destination, then enterprise network operations teams have no
> capability to manage traffic policy.  I believe the majority of
> enterprise network operations teams would have an issue with that.

Yes, but the question is how much we should care about that.  It's
true that IT people are the ones who have to buy our stuff, but in the
end our requirements are driven by business (in the large sense) needs
of "end users".

And you can rest assured that someone will sell those IT people boxes
that give them back some control over their vict^H^H^H^Hcustomers'
machines' path selection decisions, just like rate shapers "fix" TCP's
rate computation in many enterprises today.

> We should keep traffic routing within the network layer, IMO, or at
> least give visibility to the network layer of alternate paths that
> can be taken for a variety of policy reasons.

Or we push the policy information out to the edge, so that end systems
can select paths accordingly.
-- 
Simon.