1. Drafting Behind Akamai (Travelocity-Based Detouring) Aleksandar Kuzmanovic Northwestern University Joint work with: A. Su, D. Choffnes, and F. Bustamante To appear in Sigcomm 2006 http://www.cs.northwestern.edu/~akuzma/

2. A. Kuzmanovic Drafting Behind Akamai 2 Drafting Detour

3. A. Kuzmanovic Drafting Behind Akamai 3 Motivation Overlay networks –Can’t change IP, so change the layers above –E.g., end-point multicast Common problem –Build the “view” of the underlying network –Passive and active measurements

4. A. Kuzmanovic Drafting Behind Akamai 4 Problem Independent measurements –Redundant –Non-scalable –Can cause problems E.g., synchronization Common service needed –Knowledge plane - MIT –A routing underlay for overlays - Princeton –Network Weather Service - Berkeley

5. A. Kuzmanovic Drafting Behind Akamai 5 Our Approach Fact –CDNs (e.g., Akamai) perform extensive network and server measurements Publish the results through DNS over short time scales Can overlay networks reuse measurements collected by production CDNs? –Significantly reduce the amount of measurements (a complementary service) –No new infrastructure need to be deployed –Inherit the robustness of DNS –Easy integration with existing systems

6. A. Kuzmanovic Drafting Behind Akamai 6 CDN-Driven One-Hop Source Routing S A1 An A2 …….. D DNS ServerEnE2E1

7. A. Kuzmanovic Drafting Behind Akamai 7 Key Questions How does Akamai work? –DNS translation –How many web replicas does a client see? –Impact of different sites (e.g., Yahoo vs. NY Times)? –DNS redirection dynamics? –Network or server latency? An example application: one-hop source routing –Mapping CDN servers to overlay nodes –Low-overhead protocols for exploiting CDN redirections

8. A. Kuzmanovic Drafting Behind Akamai 8 DNS “Black Magic” Web Client Akamai High-Level DNS Server Akamai Low-Level DNS Server Local DNS Server PCWorld Authoritative DNS Server PCWorld Web Server Akamai Edge Server Root DNS Server Top-Level Domain DNS Server www.pcworld.com http request/response images.pcworld.com CNAME: a1694.g.akamai.net images.pcworld.com a1694.g.akmai.net 2 ip addresses of Akamai Edge Servers 2 ip addresses of Akamai Edge Servers fetch image files

9. A. Kuzmanovic Drafting Behind Akamai 9 Measuring Akamai 2-months long measurement 140 PlanetLab (PL) nodes –50 US and Canada, 35 Europe, 18 Asia, 8 South America, the rest randomly scattered Every 20 sec, each PL node queries an appropriate CNAME for –Yahoo, CNN, Fox News, NY Times, etc. PL Node Akamai Edge Server 2 Akamai Edge Server 3 Akamai Low-Level DNS Server.…… Akamai Edge Server 1

10. A. Kuzmanovic Drafting Behind Akamai 10 Initial Results day night BerkeleyPurdue

11. A. Kuzmanovic Drafting Behind Akamai 11 Server Diversity for Yahoo Good overlay-to-CDN mapping candidates Majority of PL nodes see between 10 and 50 Akamai edge-servers Nodes far away from Akamai hot-spots

12. A. Kuzmanovic Drafting Behind Akamai 12 Multiple Akamai Customers Amazon is not hosted on UMass and LBNL Yahoo and NYTimes are not hosted in U. Oregon FEMA is poor!

13. A. Kuzmanovic Drafting Behind Akamai 13 Redirection Dynamics BerkeleyBrazil Korea

14. A. Kuzmanovic Drafting Behind Akamai 14 Key Questions How does Akamai work? –DNS translation –How many web replicas does a client see? –Impact of different sites (e.g., Yahoo vs. NY Times)? –DNS redirection dynamics? –Network or server latency? Potentials for one-hop source routing? –Mapping CDN servers to overlay nodes –Low-overhead protocols for exploiting CDN redirections

15. A. Kuzmanovic Drafting Behind Akamai 15 Methodology Planet Lab Node Akamai Low-Level DNS Server 10 Best Akamai Edge Servers ………

16. A. Kuzmanovic Drafting Behind Akamai 16 Do CDN redirections correlate to network latencies? (1) Rank = r1+r2-1 –16 means perfect correlation –0 means poor correlation Brazil is poorMIT and Amsterdam are excellent

17. A. Kuzmanovic Drafting Behind Akamai 17 Do CDN redirections correlate to network latencies? (2) CNN is poor! 70% paths are among best 10% 97% better than average

18. A. Kuzmanovic Drafting Behind Akamai 18 Akamai-Driven One-Hop Source Routing S A1 An A2 …….. D DNS Server Redirection dynamics sufficiently small for network control EnE2E1 Redirections driven by network conditions Potential for CDN-to-overlay mapping

19. A. Kuzmanovic Drafting Behind Akamai 19 Methodology Source Akamai Low-Level DNS Server Destination Direct Path Akamai path 3 Akamai path 1 ……..... Akamai path 10 Akamai path 2 Akamai path 1 Akamai path 2 Akamai path 3 Akamai path 10

20. A. Kuzmanovic Drafting Behind Akamai 20 Akamai-driven source routing (1) 80% Taiwan 15% Japan 5 % U.S. 75% U.K. 25% U.S. Taiwan-UKUK-Taiwan

21. A. Kuzmanovic Drafting Behind Akamai 21 Akamai-driven source routing (2) Direct paths better than Akamai paths 25% of Akamai paths better than direct paths Akamai and direct paths have similar performance Experiment: US (6), Europe (3), S. America (2), Asia (3)

22. A. Kuzmanovic Drafting Behind Akamai 22 Path pruning Fact –Not always is Akamai-driven path better than the direct one Practical issues –How frequently to make a decision whether to use the direct or the Akamai path? –Should one use the first (of the 2 paths) returned by Akamai (FAS) the better (of the 2 paths) returned by Akamai (BTAS)

23. A. Kuzmanovic Drafting Behind Akamai 23 Path Pruning Result Direct path accounts for 78% of the gain 2 hours update frequency before the performance declines BTAS better than FAS

24. A. Kuzmanovic Drafting Behind Akamai 24 Conclusions Reuse measurements performed by CDNs Reverse-engineering Akamai –DNS redirections sufficiently small –Strong correlation to network conditions –All clients see a large number of paths CDN-driven one-hop source routing –25% of Akamai paths outperform direct paths –50% of nodes “discovered” by Akamai outperform direct paths –Low-overhead pruning algorithms Global Internet “weather-report” service for little to no cost