1 Seminar on “Clean Slate Design for the Internet” Nick McKeown

2 22 High level  “Given what we know today, if we were to start over with a Clean Slate, how would we design a global communications network?”  “Ideally, how will the network look in years, and how will we get there from here?”

3 33  What’s wrong with the Internet…?  Why is the research and business community not already solving it?  What are other groups doing?  What we plan to do at Stanford  An example of “Clean Slate” design Prelims

4 44 Original Architecture  A dumb connectionless packet-forwarding packet- switched infrastructure, with high-level functionality at the edge  Single, simple lowest-common denominator data delivery service (IP), with reliable stream service built on top  Fixed-size numerical addresses with {network, host} hierarchy; one per physical network interface  Later  Separation of IP and TCP (including congestion control using packet loss as congestion signal)  Subnetting, autonomous systems (EGPs and IGPs), DNS, CIDR

5 55 What is needed  Wouldn’t we like a network that we can trust to be always there, always on, easy to use, universally accessible, secure, and economically viable.  David Cheriton’s example: If the FAA carried all of its traffic over the public Internet, you'd be nuts to fly.  Some obvious desirable characteristics  Robustness and Availability  Security  Naming and Addressing: accountability vs anonymity  Predictability  Mobility  Economic Viability  What else?

6 66  What’s wrong with the Internet…?  Why is the research and business community not already solving it?  What are other groups doing?  What we plan to do at Stanford  An example of “Clean Slate” design Prelims

7 77  What’s wrong with the Internet…?  Why is the research and business community not already solving it?  What are other groups doing?  What we plan to do at Stanford  An example of “Clean Slate” design Prelims

8 88 What are others doing?  Background  Incrementalism and “victim of success” of Internet  New era of more radical and fundamental thinking about the future of networks and communications  New-arch (MIT)  100x100 (CMU)  Geni (NSF/Gov)

9 99 New-arch (2000)  Requirements for new network  Mobility: Highly dynamic and efficient  Policy-driven auto-configuration  Highly time-variable resources  Allocation of capacity 

10 100x100 (CMU/Stanford/Rice)  NSF Large ITR ( )  Questions:  Can structure be used to make networks more robust, predictable and manageable?  What economic principles drive the operation of access and backbone networks?  What security primitives must be built into the network?  Can/should network and protocol architectures be designed to take advantage of long-term technology trends? 

11 NSF Geni Initiative (2005)  CISE major effort, seeking congressional funding of approx $300M starting 2008  Two parts: Research program; Global experimental facility to explore new architectures  Areas of interest:  Creating new core functionality, including naming, addressing, identity, management.  Developing enhanced capabilities: building security intot he architecture; design for high availability; privacy/accountability; design for regional differences and local values  Deploying and validating new architectures  Building higher-level service abstractions  Building new services and applications  Developing new network architecture theories

12  What’s wrong with the Internet…?  Why is the research and business community not already solving it?  What are other groups doing?  What we plan to do at Stanford  An example of “Clean Slate” design Prelims

13  What’s wrong with the Internet…?  Why is the research and business community not already solving it?  What are other groups doing?  What we plan to do at Stanford  An example of “Clean Slate” design Prelims

14 What we plan to do at Stanford  Weekly Seminar in Fall and Winter  Fall: Talk by professor followed by discussion  Goals  To get thinking about the problem  To learn from each other  To identify some collaborative research projects

15  What’s wrong with the Internet…?  Why is the research and business community not already solving it?  What are other groups doing?  What we plan to do at Stanford  An example of “Clean Slate” design How to design backbone networks from a clean slate? Prelims

16 Backbone Networks: Emerging Structure  routing centers interconnected by long-haul optical links  Increasingly rich topology for robustness and load- balancing  Typical utilization < 25%, because  Uncertainty of traffic matrix network is designed for  Headroom for future growth  Headroom to carry traffic when links and routers fail  Minimize congestion and delay variation  Efficiency sacrificed for robustness and low queueing delay

17 How flexible are networks today? Abilene Verio AT&TSprint 25% Over Prov: 0.025% 50% Over Prov: 0.66% What fraction of allowable traffic matrices can they support? 25% Over Prov: % 50% Over Prov: 1.15% 25% Over Prov: % 50% Over Prov: 0.15% 25% Over Prov: % 50% Over Prov: 0.06% Verio, AT&T and Sprint topologies are from RocketFuel

18 Desired Characteristics  Robust Recovers quickly; continues to operate under failure  Flexible Will support broad class of applications, new customers, and traffic patterns  Predictable Can predict how it will perform, with and without failures  Efficient Does not sacrifice cost for robustness

19 Backbone Design  Assume underlying reliable mesh of physical circuits 1. Dynamic circuit switching over underlying mesh, or 2. Load-balanced logical network. Describing today

20 Approach  Assume we know/estimate traffic entering and leaving each Regional Network  Requires only local knowledge of users and market estimates  Use Valiant Load Balancing (VLB) over whole network  Enables support of all traffic matrices

21 Valiant Load-Balancing N … 4 r1r1 2r 1 r 2 /rN r2r2 r3r3 r4r4 rNrN Capacity provisioned over existing robust mesh of physical circuits

22 A Predictable Backbone Network  Performance: 100% throughput for any valid traffic matrix.  Only need to know aggregate node traffic.  Under low load, no need to spread traffic.  Robustness  Upon failure, spread over working paths  Small cost to recover from k failures: Provision approx 2r i r j /r(N-k)  Simple routing algorithm  Efficient  VLB is lowest cost method to support all traffic matrices  Similar cost, while supporting significantly more traffic matrices.

23 How expensive would VLB be? Abilene Verio AT&TSprint 25% Over Prov: 0.026% Cost: % Over Prov: 0.66% Cost: 1.04 Cost normalized to VLB routing. Cost of switching = cost of transmission for 370miles 25% Over Prov: % Cost: % Over Prov: 1.08% Cost: % Over Prov: % Cost: % Over Prov: 0.14% Cost: % Over Prov: % Cost: % Over Prov: 0.04% Cost: 1.04

24 Open questions  Worst case propagation delay doubled  Low variance in delay  There are “express paths”  (How) are multiple VLB networks connected, and how does performance change?  Economics and policy: how do operators compete?