Tom Anderson
University of Washington
1

The Internet’s current evolutionary path will not address its fundamental challenges:
– Security

Security costs of connecting to the Internet dwarf bandwidth costs; no end in sight to viruses, worms, DoS, spam
– Robustness

End to end reliability orders of magnitude lower than phone service
– Manageability
 State of the art: “tweak and pray”
– Performance
 End to end performance orders of magnitude slower than the raw hw
– Evolvability

QoS, ad hoc networks, mobility, etc.

Do we know how to fix these problems?
If we did know, could we implement the fix?
Answer is no to both, for today’s Internet
– Little to no conceptual understanding of how to address these issues
– Little to no ability to implement changes to the Internet architecture, except via point solutions that often make matters worse in the long run
Analogy with programming languages in 80’s/90’s
– dominant standards in industry crowd out academic innovation; eventually more radical approaches succeed

Simple, universal end to end packet delivery service, implemented by multiple, cooperating service providers
IP

Any architectural change requires global agreement
– ISPs have little incentive or ability to evolve architecture
– result: ossification with feature creep
IP IP IP
NAT, VPN, firewalls,
IDS, …
NAT, VPN, firewalls,
IDS, …
NAT, VPN, firewalls,
IDS, …

Services can only have two out of three among: multiprovider (e.g., planetary scale) high level interface (e.g., IP) evolvability
Examples: IP, email, telephony, CDNs, multicast, …

RIP: horizontal, planetary-scale service providers
– At base, a virtual hardware abstraction (cycles and bit pipes to neighbors); layer planetary-scale services on top
…
IP as a service
Routing
Resource Management
Information plane
HWP HWP HWP HWP

Rate of increase in cycles/$ >> internet bw/$
Jim Gray (2003): Cost of sending TCP ack = 500K instructions
What about future?
Moore’s Law vastly understates potential for CPU improvement:
60% squared (density) * 30% (cycle time) * 20% (volume)
Raw optics improving at a similar rate => captive backbones
Cumulative improvement in the engineering of distributed systems
We understand how to engineer secure, reliable, efficient distributed systems, if we aren't constrained by legacy systems

• Add a new layer to the network architecture
– overlay networks overlay
 purpose-built virtual networks that use the existing
Internet for transmission
 the Internet was once deployed as an overlay on top of the telephony network
• Challenges
– isolate services from each other and the Internet
– exploit planetary-scale cooperation/vantage points
– become the intermediary for WAN packets
Internet

Any new architecture needs to be:
Incentive compatible for end users
Opt-in at a fine-grain (hijack packets via name xlation)
Overlay routing for reliability, bandwidth, latency
PCP to manage legacy Internet paths
Incentive compatible for hardware providers
Avoid 95% charging intervals
Win-win bilateral barter
Self-managing, secure, evolvable as an engineered solution

anarchy barter
%
5
1
50
10 distance inflation anarchy
1.0
1.4
2.0
5.9
barter
1.0
1.1
1.2
1.5
path length inflation relative to socially optimal
• Bilateral barter closely approximates socially optimal
– Reduces need for manual route tweaks

Make security, manageability, efficiency, etc. engineering problems, not political ones