Re: hard questions: request routing

[Francis Zane <francis@research.bell-labs.com>]

My new concern is:

* Is this fundamentally a 2-level (content provider CDN -> end CDN) or
a multilevel (ie, chains of CDNs inbetween) problem.

(I'm leaving out the producer to the first CDN, as well as
distribution within the terminal CDN, since those should be easily
loop-free.)

This comes out of Oliver's suggestion of the 3-level mechanism (ie,
one forwarding CDN inbetween).  Which begs the question: if we're
going to limit the hierarchy depth, why one forwarder and not two or
zero?  More specifically, 

Question: What value do intermediate CDNs provide?

The only advantages I see is that they perform some useful
aggregations: collecting the metrics, decentralizing the routing
decisions, and simplifying the business arrangements.  My thinking is
that other than that, they just add complexity to the forwarding.
But maybe there are others?  Load-splitting?

On the flip side, this aggregation could be done just as well by
filtering the announcement traffic, or even an out-of-band system
(think airline travel agents) which CDNs subscribe to which sends them
updates on announcements/prices they care about, thus reducing the
actual RR to a 2-level system.  The point would be that they're not
involved in forwarding at the request level, and resolutions always
lead to terminal CDNs, so there aren't loops.

Suppose we go to a 2-level system.  What does it look like?  The one
I'm envisioning sounds more like the one Abbie is discussing (I think;
I'm still not 100% clear on the picture there, sorry): Each source of
content (I forget the word here) collects announcements it cares
about, builds a forwarding table based on this, and uses it to direct
to terminal CDNs.  Without loops, though, the lack of a consistent
global view doesn't matter.  One could even envision pseudo-terminal
CDNs, which aggregate terminal CDNs and never redirect to outside
CDNs, which would also be loop-free.

In a sense, I'd be sort of sad to go this way, as the forwarding
mechanism is somehow more, I don't know, algorithmically sexy, but I
don't yet see what the advantages are.

Francis