1. Has the Internet Delay Gotten Better or Worse? Universidad Carlos III de Madrid 2010.6.30. 1 DK Lee, Keon Jang, Changhyun Lee, Gianluca Iannaccone, Kenjiro Cho Sue Moon Associate Professor Department of Computer Science

2. Questions we need to answer first 1.Define Internet delay 2.Random sampling of Internet hosts 3.Estimate accuracy 2

3. #1 Definition of Internet delay Delay distribution of host pairs in the Internet 3

4. #2 Random sampling Issues in random sampling of IP addresses – Not all ASes have the same-size blocks of IP addresses – Not all blocks of IP addresses are in use – Not all IP addresses are in use – Not all IP addreses are always in use => /24 block as a unit of random sampling 4

5. #3 Accuracy of estimates iPlane has shown better performance than landmark-based estimates All known delay estimation methodologies require some form of active in-situ measurement but "path stitching" 5

6. Path: Delay: r A + r AB + r B + r BC + r C Overview of Path Stitching Router-level paths and RTT from a to c ? ac AC Step 1. IP-to-AS mapping AC Step 2. AS-level path inference from A to c B Step 3. Stitching path segments :A: A::B B::C:B::C: rArA r AB rBrB r BC rCrC

7. What If There Are 7 March 15, 2010, dklee@an.kaist.ac.kr A::B ?B::C :A: :C: :B: ?... Too few segments: Too many segments:

8. When path stitching produces no stitched path Case #1: No path segments in source or destination AS Case #2: No segments in the middle of inferred AS path – inter-domain: use reverse segment – intra-domain: no solution Case #3: Segments does not rendezvous at the same address – Use approximation

9. When path stitching produces multiple stitched paths Use preferences rules #1 Same destination-bound prefix #2 Closeness to source and destination #3 Most recent vs median

10. Comparison with iPlane 10 Very promising results: With accurate AS paths inference, errors <= 20ms for 80% of pl-hard pairs

11. Now we ask the question again: Has it gotten better or worse? 11

12. Review of random /24 prefixes 12

13. BGP RIB Entries 13 http://bgp.potaroo.net/

14. # of /24 blocks in the BGP tables 14

15. Graphical distribution of host pairs (AS: Asia, AF: Africa, EU: Europe, OC: Oceania, NA: North America, SA: South America) 15

16. Varying sample sizes 16

17. Response rates (n = 10,000) 17

18. Our data set CAIDA's Skitter/Ark from 2004 RouteView and RIPE BGP tables 18

19. Chronicle of Ark monitors 19

20. Delay distribution between random pairs of hosts in 2004 and 2009 20

21. 2004 vs. 2009 21/21 Delay distribution has gotten worse from 2004 to 2009 (Median delay 164.0 msec  211.6 msec) IP/AS hop counts decreased end-to-end

22. Regional Growth of the Internet 22/21 Fraction of host pairs in NA decreased significantly from 40 % to 20% Fractions of all other regional pairs increased NA: North America SA: South America AS: Asia EU: Europe OC: Oceania AF: Africa

23. Delay Distributions for NA-NA and AF-EU pairs 23/21 Delays distributions for NA pairs in 2004 and 2009 are almost identical Delay performance for AF-EU pairs for most part improved 10% of AF-EU pairs experience delays more than 1 sec in 2009

24. For the same pairs of hosts in 2004 and 2009 24

25. 2004 vs. 2009 25/21 Delay distributions for the same set of sample host pairs remain almost identical of slightly improved IP/AS hop counts decreased

26. Concluding Remarks We present the methodology for the Internet delay history reconstruction and analysis: – Path stitching with existing measurements – Random sampling of the Internet host pairs We demonstrate the our approach is feasible in showing insight about the overall Internet delay distribution. Future work will focus on: – Rigorous statistical analysis about the sources of errors – Trends from 1999 to 2009 Match the trend with the Internet-wide upgrades Find the corroborating evidences for the observations 26/21

27. BACKUP SLIDES 27

28. Internet-wide Coverage: Approximations 28 pl-easy pairs pl-hard pairs we show incremental improvement in the fraction of pairs with stitched paths from 5% to 70% (for pl-hard pairs)

29. Preference Rules – (1) 29 Pair #N 0 Delay (ms) pl-easy pairs pl-hard pairs estimated delay (min) without preference rules real delay (max) real delay (min) estimated delay (max) without preference rules proximity+dst.bound (min) proximity+dst.bound (max) All three rules Preference rules bring the estimated delays close to the real measurements

30. Delay distributions, from 2005 to 2009 in comparison with 2004 (Different pairs) 30

31. Median Delays from 2004 to 2009 31

32. Delay distributions, from 2005 to 2009 in comparison with 2004 (Same pairs) 32

33. End-to-end delay performance for specific pairs 33