1. The Internet’s Dynamic Geography Scott Kirkpatrick, School of Engineering, Hebrew University of Jerusalem and EVERGROW Collaborators (thanks, not blame…) Yuval Shavitt, Eran Shir, Shai Carmi, Shlomo Havlin, Avishalom Shalit Bremen, June 11-12, 2007

2. Measuring and monitoring the Internet Has undergone a revolution Traceroute – an old hack  basic tool in wide use Active monitors – hardware intensive  distributed software DIMES (“Dimes@home”) an example, not the only one now Many enhancements under consideration, as the problems in traceroute become very evident Ultimately, we expect every router (or what they become in the future internet) will participate in distributed active monitoring. The payoff comes with interactive and distributed services that can achieve greater performance at greatly decreased overhead

3. History of TraceRoute active measurement Jacobson, “traceroute” from LBL, February 1989 Commonly uses ICMP echo or UDP Variants exist – tcptraceroute, NANOG, “Paris traceroute” And this is something that can be rewritten for special situations, such as cellphones Single machine traces to many destinations – Lucent, 1990s (Burch and Cheswick) Great pictures, but interpretation not clear, demonstrate need for more analytic visualization techniques But excellent for magazine covers, t-shirts… First attempt to determine the time evolution of the Internet First experience in operating under the “network radar” Lumeta, their spinoff, ended up as a network radar supplier.

4. IP address map of August 1998

5. IP address map of Jan 1999

6. IP address map of June 1999

7. Map interpreted: color by ISPs

8. History of Internet Measurement, ctd. Skitter and subsequent projects at CAIDA (SDSC) 15-50 machines (typically <25), at academic sites around world RIPE and NLANR, 1-200 machines, commercial networks and telco backbones, information is proprietary DIMES (>10,000 software agents) represents the next step A complementary approach is available at the coarser level of ISPs (actually “autonomous systems” or ASes) RouteViews (Univ. of Oregon) since 2001 has monitored BGP preferred routes broadcast from a healthy sampling of ASes’ border routers.

9. Traceroute is more than a piece of string A flood of feigned suicide packets (with TTL values t=1 to about 30 hops), each sent more than one time. Ideal situation, each packet dies at step t, router returns echo message, “so sorry, your packet died at ip address I, time T” Non ideal situations must be filtered to avoid data corruption: Errors – router inserts destination address for I Non-response is common Multiple interfaces for a single (complex) router Route flaps, load balancing create false links Route instabilities can be reduced with careful header management (requires guessing router tricks)

10. The Internet is more than a random graph Internet is a federation of subnetworks (ASs or ISPs) It has at least a two-level structure (AS, ip-level) because two different routing strategies and software are used to direct packets. Other coarse grain views – country, city, POP… There are no global databases, many local databases, poor data quality available. Models have evolved steadily Waxman (Random graph with Poisson distribution of ngbrs) “Transit-stub” model with two-level hierarchy Power law pictures, such as preferential attachment, reordering Jellyfish and Medusa

11. What is the quality of today’s measurements? Bias issues – does a superposition of shortest-path trees converge to the actual underlying graph? Concerns about diminishing returns? Filters needed to screen as many false links as possible. Once you have a flood of data, need to address two issues: Has it converged to cover the real graph? Betweenness and visit count help address this How stable are the measurements over time? And finally, how does traceroute discovery compare with online tables of AS-disclosed information (BGP tables)?

12. What do we see with DIMES? New graphical analysis methods reveal considerable structure, apparently related to function. Yes, Virginia, there are power laws! But the initial conditions and some of the patterns of growth reflect distinct roles of subnetworks as well as growth dynamics, and economic incentives. The Internet is a moving target, and we are observing it through a very shaky telescope. How should we characterize its evanescent behavior? How to integrate to see the fainter stars? Discussions of bias and “diminishing returns” may be addressing the wrong hypotheses.

13. Use a new analytical tool – k-pruning Prune by grouping sites in “shells” with a common connectivity further into the Internet: All sites with connectivity 1 are removed (recursively) and placed in the “1-shell,” leaving a “2- core” then 2-shell, 3-core and so forth. The union of shells 1-k is called the “k-crust” At some point, kmax, pruning runs to completion. Identify nucleus as kmax-core This is a natural, robust definition, and should apply to other large networks of interest in economics and biology. Cluster analysis finds interesting structure in the k-crusts

14. Does degree of site relate to k-shell?

15. Numbers of site-distinct paths in the nucleus Conclusion: innermost k-cores are k-connected. But outer k- cores (2,3,4) show exceptions (sites with 1,2,3 paths). kmax (03-06) = 41 kmax (05-06) = 39

16. Distances and Diameters in cores

17. Distances and Diameters

18. K-crusts show percolation threshold Data from 01.04.2005  These are the hanging tentacles of our (Red Sea) Jellyfish For subsequent analysis, we distinguish three components: Core, Connected, Isolated Largest cluster in each shell

19. Michalis Faloutsos ’ Jellyfish Highly connected nodes form the core Each Shell: adjacent nodes of previous shell, except 1- degree nodes Importance decreases as we move away from core 1-degree nodes hanging The denser the 1-degree node population the longer the stem Core Shells 1 2 3

20. Meduza ( מדוזה ) model This picture has been stable from January 2005 (kmax = 30) to present day, with little change in the nucleus composition. The precise definition of the tendrils: those sites and clusters isolated from the largest cluster in all the crusts – they connect only through the core.

21. Non-communication Networks

22. Communication networks

23. Who’s “tier-1” in Medusa? 7012992 70182766 33562665 12392619 1741967 2091387 129561261 12991251 35491219 35611215 2914998 7132951 702923 6730923 6461907 4323772 1273728 3491687 6453644 3303612 3320590 6939584 2828577 4513570 4637544 7911542 8220531 5400522 1221508 1668496 16150460 6395453 3257450 286391 3246389 8342387 5511384 4766367 25462365 8928360 7473359 3292347 3786343 2516330 3209329 12989327 6539317 6320283 10026283 6695277 3352263 8001259 1257258 22773250 6327247 5650245 19151239 13237237 8075226 2497225 15412213 6762208 19029206 4589203 5459202 5089197 852180 5462176 15290174 577156 2856153 8546153 9318145 6079137 13768136 4725133 22822128 293122 4134122 3300117 4355113 6830110 12322108 443698 638996 821095 478893 2335289 1954887 2334280 1031075 81264 1516950 Data from months 10-12, 2005 kmax = 42, 93 nodes All fall within CAIDA’s top 200 ASes, measured by size of “customer input cone.”

24. What about the error bars, the bias, etc.? Need to address the specifics of the “network discoveries” How frequently observed? How sensitive are the observations to the number of observers? How do the measurements depend on the time of observation? The extensive literature on the subject is mostly straw-man counterexamples, that show bias from this class of observation can be serious, in graphs of known structure, but do not address how to estimate structure from actual measurements.

25. Lecture 2 Efforts to model the Internet Waxman (Poisson statistics, single scale) Zegura and co-workers (GaTech) two scales “Transit” and “stub” Preferential attachment Shalit et al (2001) showed exponent in (2,3) possible, and k- shells also give simple power laws Counterattack of the establishment Luddites?

26. The Empire Strikes Back!

27. Willinger et al. analysis of models Is a particular model “descriptive” or “explanatory”? Descriptive models are “evocative “data-driven” But too generic in nature Explanatory models are Structural Can close the loop by validating the explanatory steps with real data “Demystify emergent phenomena”

28. So models  excerpts of actual measurements Power laws occur in the k-shells as well as in degree distrib: But the k-cores are not scale invariant!

29. Where is a pure “emergent phenomenon” happening? Box cover construction shows true fractal only as the shells percolate

30. Back to the actual data Visit count and betweenness Best evidence for reliability of data How much better will it get with 100,000 agents observing? Can’t ask the question. But can ask, how much worse will it be with fewer. Three approaches in prospect. All future work. Study betweenness of present graph with reduced traffic model Reanalyze our raw data with fewer agents included Run retrospective experiments with agents selected specially

31. What sort of coverage is obtained?

32. Agents from entire two years participate

33. Weekly coverage and agent utilization

34. Time dependences – even RouteViews’ BGP speakers vary Study 6 weeks in 2006 (June, July) 50,245 to 51,309 edges found per week In wk 26, 48,221 edges seen all week 335 edges seen for 6 days 192 edges seen only 5 days 294 edges seen only 4 days 354 edges seen only 3 days 260 edges seen only 2 days 175 edges seen only 1 day 451 edges seen only one time. Single observations peak on Sunday (149 edges, other days typically ~40) Edges seen 3 or more days peak at ends of the week Twice as many edges are created on Monday as are deleted on Sunday…

35. Random scale-free graphs produce the same basic structure, different details

36. Percolation “ attacks ” K-core based attack (“by reputation”) is comparable to accurate degree-based attack for random networks, but not for the real AS graph.

37. Preliminary reachability data (using whole graph) Sites reachable

38. Now restrict to the 20-crust Up then downSide step at topThree sidesteps