Building an AStopology model that captures route diversity

1
Building an AS-topology model that captures route
diversity
Steve UhligUniversité catholique de Louvain
Wolfgang Mühlbauer Anja Feldmann Technical
University Munich
Olaf Maennel Matthew RoughanUniversity of
Adelaide
2
Agenda

• Background
• Why another model of the Internet?
• A model of the Internet
• Some results
• Conclusion

3
Agenda

• Background
• Why another model of the Internet?
• A model of the Internet
• Some results
• Conclusion

4
A simple Internet
AS 3
AS 2
AS 1
AS 4
AS 6
AS 5
Inter-domain link Intra-domain link
5
Advertising a reachable prefix
AS 3
AS 2
p
p
AS 1
AS 4
AS 6
AS 5
Inter-domain link Intra-domain link
6
Choice of paths towards p
AS 3
AS 2
AS 1
AS 4
Traffic path Inter-domain link Intra-domain
link
AS 6
AS 5
7
AS-paths towards p
AS 3
AS 2
Effect of policy
AS 1
AS 4
AS 5
AS 6
AS path Inter-AS edge
8
Agenda

• Background
• Why another model of the Internet?
• A model of the Internet
• Some results
• Conclusion

9
Observed AS-paths
AS2914 (Verio)
AS4716 (POWEREDCOM)
AS3549 (Global Crossing)
AS5511 (FranceTelecom)
AS7911 (Wiltel/Level 3)
AS24249 (JWAY)
AS3356 (Level 3)
AS4694 (IDC)
AS3561 (CW)
10
Required routers per AS
11
One router per AS

• In favour
• Large fraction of the observable AS-paths can
  still be matched without having multiple routers
  Mao05
• Some policies seem to be defined on a per
  neighboring AS basis Gao00, Subramanian02
• Against
• ASes do contain multiple routers and propagate
  multiple paths
• With one router per AS one cannot explain 100 of
  the observed paths

12
Why another model of the Internet?

• Why would one want a more realistic model of the
• Internet
• How does the traffic flow from one AS to another?
  
• What if topological change happens?
• What if an AS changes its routing policies?
• For that we need to know
• How routes propagate across the network?
• Which policies are applied between ASes ?
• Long-term goal Infer how the real Internet works
  and be
• able to predict it to some extent (what-if)
• Shorter-term goal Learn how to build models that
  capture
• some aspects of the macroscopic reality of the
  Internet

13
Agenda

• Background
• Why another model of the Internet?
• A model of the Internet
• Some results
• Conclusion

14
Data

• Snapshot of BGP data from more than 1,300
  observation points (700 ASes) widest coverage of
  the Internet ever used!
• 300,000 prefixes (more specifics)
• 4,730,222 unique AS paths
• 21,178 ASes
• 58,903 AS-level edges
• Partitioning of observation points into training
  and validation
• Training randomly select 2/3 of observation
  points
• Validation take the remaining 1/3 of observation
  points
• AS-level topology built from the union of all
  known AS-paths of the data (both training and
  validation)

15
Terminology

• Best match simulation selects a path that was
  observed in reality
• Pot. best match simulation learns a path that
  was observed in reality, but due to random
  tie-break did not select this path
• RIB-In match simulation learns a path that was
  observed in reality, but did not pick that
  path as best
• Not found No router at the considered AS in the
  simulation learns about the path that was
  observed in reality

16
Approach

• Build a model of the data based on training
  dataset
• This model must be 100 consistent with observed
  AS paths from the training dataset (100 best
  matches)
• Look at how this model performs for validation
  dataset in terms of the matches
• Note removed paths from validation redundant
  with those of training

17
Reproducing observable paths

• Premises
• Without policies, shortest paths are propagated
• If a non-shortest path is observed, it means some
  policy has been applied somewhere
• Only observable paths give us usable
  informationabout the AS-level topology and
  potential routing policies
• Goal Reproduce perfectly observed paths
  (training
• set) in the simulation model

18
Simulation principles

• Split AS, if multiple paths must be propagated
• Filter shortest paths, if longer paths must be
  propagated
• Get rid of random decisions (lowest router-ID),
  when supporting information is available

19
Initial state of model
prefix p

• One router per AS and shortest paths are chosen

20
Splitting ASes
prefix p

• Split AS into several quasi-routers when
  several
• paths must be propagated

21
How to propagate longer paths?
longer path router
prefix p
shorterpath router

• Filter shorter paths when a longer path must be
• observed

22
How to propagate longer path?
longer path router
prefix p
shorterpath router

• Filter also on egress-part of shorter path
  router

23
Lowest Neighbor ID
lowest neighbor ID decision
sim.
obs.
prefix p

• Fix arbitrary decisions when several equal length
• simulated paths occur

24
Agenda

• Background
• Why another model of the Internet?
• A model of the Internet
• Some results
• Conclusion

25
Training dataset
RIB-In Best match
Iteration
Training achieves 100 matches
26
Validation dataset
RIB-In Best match
Iteration
Accuracy 63 best matches - 94 RIB-In matches
27
Discussion

• Our model achieves what it was supposed to on
  training
• Literature cannot match (RIB-In) more than 87 of
  the paths because of one router assumption and
  simplistic policies
• Our model performs quite well on validation
• 93 of the paths are propagated correctly, 63
  correctly predicted
• Only a single case of validation of AS-topology
  model in Mao05 on 3 observation points
• We used more than 400 observation points
• We cannot compare our results to literature (we
  are the reference now!)

28
Whats next?

• 63 of the paths in the validation dataset were
  correctly predicted
• Reasons
• Trained paths limit the choice we have to do to
  predict validation paths
• We do not reverse engineer the actual Internet!
• We do not know the real policies!
• We build simplest policies which are consistent
  with our observations.
• gt Agnosticism leads to better results than
  incorrect
• assumptions
• gt To go further we need to add assumptions
• about how the real Internet might be working

29
Further work

• Reverse-engineering actual policies
• What is a policy?
• How do each AS define its policies?
• Adding the time dimension
• Current model relies on a snapshot of the BGP
  data
• In practice BGP is converging for a subset of the
  prefixes almost all the time
• Our view of the topology has to include time
  dynamics
• Actual diversity of the real Internet
• How do observations sample the actual routing
  diversity of the real Internet?
• Our current model can check for type 2 errors
  (when model is wrong), not type 1 ones
  (mistakes in observed paths)

30
Outline

• Background
• Why another model of the Internet?
• A model of the Internet
• Some results
• Conclusion

31
Conclusion

• One router per AS is not good enough to model the
  Internet seen from multiple vantage points
• Proposed a first agnostic model that perfectly
  reproduces observations
• Model also predicts well data from validation
  dataset
• Answering what-if questions requires going beyond
  agnosticism

32
Some references

• Gao00 L. Gao. On Inferring Autonomous System
  Relationships in the Internet. Proc. of IEEE
  Global Internet Symposium, 2000.
• Feldmann04 A. Feldmann, O. Maennel, Z. Mao, A.
  Berger, and B. Maggs. Locating Internet routing
  instabilities. Proc. of ACM SIGCOMM, 2004.
• Mao05 Z. Mao, L. Qiu, J. Wang, and R. Katz.
  Towards an accurate AS-level traceroute tool.
  Proc. of ACM SIGMETRICS, 2005.
• Quoitin05 B. Quoitin and S. Uhlig.Modeling the
  Routing of an Autonomous System. IEEE Network
  Magazine, 19(6), 2005.
• Subramanian02 L. Subramanian, S. Agarwal, J.
  Rexford, and R. Katz. Inferring and
  characterizing the Internet hierarchy from
  multiple vantage points. Proc. of IEEE INFOCOM,
  2002.