Internet Aware Overlays to Support Architectural Change

      	       John Jannotti, Brown

Overlay networks have freed researchers to consider serious redesigns of
the Internet architecture.  Researchers have pursued clean slate designs in
virtual networks that are free of legacy constraints.  While overlays allow
the deployment of new designs in isolated virtual networks, we must now
develop a plan to bring this innovation to the Internet.  The challenge is
to build Internet Aware overlays that interface cleanly and completely with
the existing network.

Internet Aware overlays will allow researchers the ability to engage in
nearly clean slate design while interfacing with the network of today.
Though this interface will likely create some constraints on these designs,
these constraints are valuable.  They are the constraints that will produce
architectures that the current Internet might gracefully evolve toward.

Internet Aware overlays should provide universal service, including service
to unmodified clients, servers, or both.  Work in this direction has begun,
using a variety of interposition techniques on clients.  This work has
focused on getting packets onto the overlay. We must also consider how the
overlay can efficiently return packets to the legacy Internet for legacy
destinations.  Viable exit strategies might use latency probing, network
coordinates, and BGP feeds to find a suitable exit point.

For the special (but common) case of new architectures that do not change
the fundamental network-layer service model, we can do better. An overlay
can be used to implement a Virtual AS that peers with traditional networks
and provides transit for unmodified Internet traffic.  A Virtual AS would
allow for the deployment of new wide-area routing protocols.  Internet
routes that traverse the Virtual AS would take advantage of the change.
Architecturally, the new protocol would be an IGP, though its
Internet-scale requirements will necessitate a design that will resemble a
replacement for BGP, an EGP.

Internet Aware overlays must interact with the Internet as it truly exists
today, not the idealized architecture we might wish existed.  NATs,
firewalls, and other middle-boxes complicate today's architecture, but
their existence is undeniable.  Internet Aware overlays must take advantage
of existing work in middle-box traversal, and extend that work, in order to
span the entire Internet.

In order to support this research, wide-spread network resources will be
required.  PlanetLab offers a fine starting point, but more is required.
First, nodes must be deployed in Internet Exchange Points to allow
pervasive peering with tier-1 ISPs.  Yet peering is not simply a matter of
physical deployment.  More difficult (and expensive) will be the
establishment of institutional relationships to allow peering, and
eventually, transit.  At the other end of the spectrum, we must deploy
nodes in "residential" settings with NATs, asymmetric bandwidth, and
uncooperative firewalls.

The goal of this challenge is to create a step-by-step plan to evolve
today's Internet toward support for future architectures.  A concrete
artifact will be an Overlay Toolkit that allows researchers to build
Internet Aware overlays without revisiting the same problems
(NAT-traversal, interposition, etc) with each system.  A central repository
of useful Internet metrics will also be created, allowing better exit
strategies for overlays (AS maps, IP Address to AS mappings,
pseudocoordinate databases).  Success will mean that future architectures
can be deployed, tested, and used by anyone, anywhere.