Overlay networks - broad interest of what we're doing
(1) without the help of the underlying networks
(2) with the help of underlying networks (confederated networks)

(2) is cooperative overlays.  Peer oriented approach.

 |
 |
 |
 V
Redirection ------>
Network A          Network B

Terminology in overlay networks: transport networks, transit networks

- What is known?
  Ingress/Egress routers
- What can I discover?
  Perf characteristics is via measurement
- What to negotiate?
  Ingress/Egress SLAs

* Self-organization:
  placement of services
  connection between service instances
* Topology discovery of service provider network
* Centralized vs Distributed services in an environment of infinite
  and free bandwidth
* Content-level peering

Note: $340K/month to get from SF to NY in current infrastructure

Bottleneck is in the last hop access network
- so why push services close to the access network?

AOL - all clients redirected to Virginia or Wash DC
- has its own cache

Distributed service can have good fault-tolerance

Does service placement matter?
- what applications are latency sensitive?
- what apps don't scale with centralized services?  given latency
  constraints...
- it is much cheaper to buy bandwidth than to setup and manage a
  distributed service

87% of ASs are not transit networks

Internet core:

Highly interconnected transit networks
            |
            |
            |
            V
Regional networks
    |     |
    |     |
    V     V
87% of the ASs


20 ISPs -- clique
     |
     |
     |
     V
Transit networks and big ISPs
  |     |
  |     |
  |     |
  V     V
local disjoint groups (regional)

National backbone
    |
    |
    |
    V
SF metro
    |
    |
    V
Berkeley headend
  |   |
  |   | O(500)
  V   V
 N1  N2

The case for service placement
- What will they do?
- Why do they need to be distributed?
- What is the nature of the costs?  Is bandwidth cheaper than
  management costs?

Need experimental proposal -- detecting the bottleneck.