Towards an Evolvable Internet Architecture

1. change IP (routers, headers, addressing, …) Towards an Evolvable Internet Architecture IP layer Sylvia Ratnasamy (Intel Research), Scott Shenker (U.C. Berkeley/ICSI), Steven McCanne (Riverbed Tech.) Presenter: Kai Chen and Alex Kiaie

2. hhFolkloreff • The Internet Architecture needs fixing • IPNL, Triad, IP Multicast, Pushback, GIA, Traceback, IPv6, SIFF, FQ, CSFQ, XCP, Capabilities, DTN, HLP, RCP, AIF, i3, LFN, … • But, ISPs don’t deploy these fixes • Exception: IP Multicast, IPv6 are the successful stories! • This contradiction produces two reactions.

3. Overlays to the Rescue (v1) Use overlays to augment IP • Overlays have been proposed for many services • Multicast: ESM (CMU), commercial CDNs • Routing: InterNAP, RON (MIT), SOSR (UW) • Quality-of-Service: OverQoS (UCB/MIT) • DoS: Mayday (MIT), SOS (Columbia), i3 (UCB/CMU)

4. Overlays to the Rescue (v1) Use overlays to augment IP • Overlays have been proposed for many services • Overlay is practical • bypass CISCO and the ISPs

5. Overlays to the Rescue (v1) Use overlays to augment IP • Overlays have been proposed for many services • Overlay is practical • Overlay is often appropriate • keep complexity out of IP

6. Overlays to the Rescue (v1) Use overlays to augment IP • Overlays have been proposed for many services • Overlay is practical • Overlay is often appropriate Although these overlay technologies would not lead to fundamental changes in the underlying Internet architecture, they would only mask some of its most obvious deficiencies.

7. Overlays (v2) Use overlays to undermine ISPs[Peterson, etc., 04] • Next-Generation Service Provider (NGSP) • Use overlays for a new architecture atop existing ISPs • Legacy ISPs only serve to access NGSP

8. Overlays (v2) Use overlays to undermine ISPs[Peterson, etc., 04] • Next-Generation Service Provider (NGSP) • Eventually, NGSP replaces ISPs • By leasing dedicated lines

9. Overlays (v2) Use overlays to undermine ISPs[Peterson, etc., 04] • Next-Generation Service Provider (NGSP) • Eventually, NGSP replaces ISPs • Technically, practical and broad • (and invaluable as an experimental platform)

10. Overlays (v2) Use overlays to undermine ISPs[Peterson, Shenker, Turner 04] • Next-Generation Service Provider (NGSP) • Eventually, NGSP replaces ISPs • Technically, practical and broad But, requires disrupting the existing market structure • Evolution through (repeated) revolution Are there other (more conservative) options?

11. This Paper • Can we enable evolution that • can retain the existing market structure • yet, allows non-incremental change (revolution through evolution ) • Approach: • design for evolution (vs. causing evolution)

12. Design for Evolution The Internet will always be • multi-provider • decentralized in control Common complaint • Internet service providers have little incentive to innovate Many possible reasons for ISP reluctance • architectural barriers to innovation • economic barriers (pricing models, etc.) • disconnect between research and reality • maybe the Internet is doing just fine • maybe the fixes we propose aren’t the right ones

13. This Paper Focus on architectural barriers When a new version of IP, call it IPvN, is defined, what conditions would lead ISPs to deploy it? Outline • Toy example: deploying IPvN • Universal Access • Implementing Universal Access • Conclusion

14. Toy Example IPvN supports comprehensive security • requires router support • new IP headers • Software vendor puts out an IPvN stack • Router vendors support IPvN • Content Provider (CP)is interested in using IPvN • ISPs consider deploying IPvN

15. Deploying IPvN Scalable, flexible  partial deployment a necessity partial deployment  partial usability CP IPv4 ISP A

16. partial usability global usability partial deployment  partial usability development of applications/services stalled on global usability global deployment Proposal: separate deployment from usability • require global usability under partial deployment any ISP can gate usability low usage, user demand high risk, yet offers no competitive advantage for ISPs to deploy IPvN. no incentive for ISPs to deploy IPvN

17. partial deployment  global usability X IPv4 ISP A

18. Universal Access If even a single ISP deploys IPvN, any endhost can use IPvN • enables customer choice, demand • encourages application development • no ISP can gate adoption • independent innovation; others follow to compete Note assumption: UA leads to increased revenue flow

19. Outline • Toy Example: deploying IPvN • Universal Access • Implementing Universal Access • constraints • two components • putting it all together • Conclusion

20. Achieving UA Constraints: • partial deployment • partial ISP participation • allow participating ISPs control • existing players • existing contractual agreements

21. Achieving UA: Two components (1) partial deployment  multi-provider overlays* IPv4 ISP A

22. Achieving UA: Two components (2) universal access  need redirection IPv4 ISP A

23. Redirection for UA Involves knowing: • where IPvN routers are located • which IPvN router is the best choice for a source (And the answer to both changes as deployment spreads!) Mechanism is ~tunneling++ Key is who effects redirection

24. Redirection: Options Who Recall Constraints • partial deployment • partial ISP participation • participant ISP control • no new players • existing contracts

25. Redirection: Options Who • user: unwieldy Recall Constraints • partial deployment • partial ISP participation • participant ISP control • no new players • existing contracts

26. Redirection: Options Who • user: unwieldy • user’s ISP Recall Constraints • partial deployment • partial ISP participation • participant ISP control • no new players • existing contracts

27. Redirection: Options Who • user: unwieldy • user’s ISP • participant ISPs Recall Constraints • partial deployment • partial ISP participation • participant ISP control • no new players • existing contracts

28. Redirection: Options Who • user: unwieldy • user’s ISP • participant ISPs • application-layer Recall Constraints • partial deployment • partial ISP participation • participant ISP control • no new players • existing contracts

29. Redirection: Options Who • user: unwieldy • user’s ISP • participant ISPs • application-layer • network-layer Recall Constraints • partial deployment • partial ISP participation • participant ISP control • no new players • existing contracts

30. Network-Layer Redirection Routers perform redirection

31. Network-Layer Redirection Routers perform redirection • Challenge: no explicit participation from ‘ ’

32. Proposal: Use IP Anycast • ‘A’is the IPv(N-1) address used to deploy IPvN • IPvN routers advertise ‘A’ into the IPv(N-1) routing protocol • adiscovers IPvN routers via IPv(N-1) routing protocol IPv4 DST = A A A A A A A

33. Redirection: Options Who • user: unwieldy • user’s ISP • participant ISPs • application-layer • network-layer* Recall Constraints • partial deployment • partial ISP participation • participant ISP control • no new players • existing contracts      *Caveat: less flexible redirection

34. But, Isn’t Anycast a Non-Starter? Short answer: no. • Scales just fine • restricted service model vis-à-vis RFC 1546 • deployed/used only by ISPs • a new IP needs one anycast address • And is deployable (see paper) • Intra-domain: minor change by participating ISPs • (+) Inter-domain v1 : simple policy change by all ISPs • (~) Inter-domain v2: no change by non-participant ISPs

35. Outline • Toy Example: deploying IPvN • Universal Access • Implementing Universal Access • constraints • two pieces • putting it all together • Conclusion

36. Putting It All Together Case 1: Destination’s ISP supports IPvN IPvN DST =Dn IPvN DST =Dn IPv4 DST = A IPv4 DST = R A A A source A A R Dn

37. Case 2: Destination’s ISP does not supports IPvN • Two issues: • Addressing hosts in non-participant ISP domains IPvN DST = ? IPv4 DST = A A A A source A ?

38. Case 2: Destination’s ISP does not supports IPvN • Two issues: • Addressing hosts in non-participant ISP domains • proposal: interim addressing à la RFC 3056 IPvN DST = D4-to-n IPv4 DST = A A A A source A D4-to-n from D4

39. Case 2: Destination’s ISP does not supports IPvN • Two issues: • Addressing hosts in non-participant ISP domains • Routing to hosts in non-participant ISP domains (paper) • one proposal: advertises D4’s prefix into IPvN routing R D4-to-n ? A A A source A R D4-to-n from D4

40. Case 2: Destination’s ISP does not supports IPvN • Two issues: • Addressing hosts in non-participant ISP domains • Routing to hosts in non-participant ISP domains (paper) A A IPv4 DST = D4 A source A D4-to-n =from D4

41. Putting It All Together Summary: Technical requirements for UA • Redirection • best achieved at the network-level • anycast: works under partial participation • Multi-provider virtual backbones • similar to the MBone, etc. • but, details of addressing and routing to destinations in non-IPvN domains requires some attention

42. Open Questions • End-host software architecture • dual-stack, NAT-PT, BIS, OCALA[UCB] • Exploring revenue flow: • ongoing work at SIMS (UCB)[Laskowski, Chuang] • Architectural limitations due to partial deployment, overlays • Clean-slate design for evolvability

43. Conclusion Proposal: A conservative approach to evolution [Floyd] • a preference for incremental strategies (that lead in the fundamentally right direction?) • value to understanding the compromises possible with existing network vs. brave new solutions

44. Conclusion Proposal: A conservative approach to evolution [Floyd] Conjecture: UA could enable ISP innovation • achievable with no change to the current architecture • a bit of synthesis, but no new mechanisms

45. Conclusion Proposal: A conservative approach to evolution [Floyd] Conjecture: UA could enable ISP innovation Maybe the Internet is evolvable Maybe the problem is not a technical one • worth exploring to avoid repeating the same mistake Or, maybe there is no problem

46. Thank you!