1. GENESIS: An agent-based model of interdomain network formation, traffic flow and economics
Aemen Lodhi (Georgia Tech)
Amogh Dhamdhere (CAIDA)
Constantine Dovrolis (Georgia Tech)
31st Annual IEEE International Conference on Computer Communications (IEEE INFOCOM 2012)

2. Outline GENESIS: Introduction & Motivation The model: Key features
Results
Validation
Analysis of results
Case study
How to use GENESIS in your research
Outline of the presentation:
Introduction/Objective
What is the model all about?
What does it tell? How can it be used?
What the model is not?
Comparison with other approaches
Key features of the model
Physical co-location constraints
Provider selection model
Peering strategies
Traffic matrix
Economic attributes
Describe modularity/flexibility
How does the model work?
Order in which decisions are made/nodes play etc.
Sources of heterogeneity
Some implementation attributes e.g. scalability, computational resources required
Validation attempts
Analysis of the model
Oscillations
Variation in equilibria
Zero sum game
Case study
Extension: Model used in evaluation of peering strategy adoption to be presented at NetEcon
Availability of the model
Questions

3. Introduction

4. Motivations for an interdomain network formation model
Insight into dynamics of interdomain network
Study pricing schemes
Study increasing asymmetry in interdomain traffic matrix
Evaluate peering strategies
Impact of actions on economic fitness
Internet “ecosystem” in the future?
What does it tell? How can it be used?
What if questions

5. What is GENESIS Agent based interdomain network formation model
Autonomous Systems (AS) as independent agents acting in a distributed asynchronous manner
Enterprise customer
Transit Provider
Internet
Content Provider

6. What is GENESIS Actions by ASes Outcome of these actions
Transit provider selection
Peering strategy selection
Peering and Depeering decisions
Outcome of these actions
Formation of an interdomain network starting from a random initial state
Mostly ending in equilibrium

7. What GENESIS is not Not a topology generation model
Not a crystal ball to accurately predict the economic fitness or hierarchical status of a single specific AS in future
Use GENESIS for
computing statistical properties of network topology + economic fitness of different categories of ASes
What does it tell? How can it be used?
What if questions

8. The Model

9. Model features
Geographic co-location constraints in provider/peer selection
Traffic matrix
Public & Private peering
Set of peering strategies
Transit provider selection mechanism
Economic attributes: Peering costs, Transit costs, Transit revenue

10. Model features Objective: Maximize economic fitness
Fitness = Transit Revenue – Transit Cost – Peering cost
Objective: Maximize economic fitness
Optimize connectivity through peer and transit provider selection

11. Geographic presence & constraints
Geographic overlap
Regions corresponding to unique IXPs

12. Traffic Matrix
Traffic for ‘N’’ size network represented through an N * N matrix
Illustration of traffic matrix for a 4 AS network
Intra-domain traffic not captured in the model
Traffic sent by AS 0 to other ASes in the network
Traffic received by AS 0 from other ASes in the network

13. Traffic components Autonomous system
Inbound traffic
Traffic consumed in the AS
Traffic generated within the AS
Traffic transiting through the AS
Autonomous system
Outbound traffic
Transit traffic = Inbound traffic – Consumed traffic
same as
Transit traffic = Outbound traffic – Generated traffic

14. Peering strategies
Restrictive: Peer only to avoid network partitioning
Selective: Peer with ASes of similar size
𝑉 𝑥 𝑉 𝑦 ≤𝜎
𝑉 𝑥 =𝑇𝑟𝑎𝑛𝑠𝑖𝑡+𝐺𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝑑+𝐶𝑜𝑛𝑠𝑢𝑚𝑒𝑑
Open: Every co-located AS except customers

15. Peering strategy selection
Default model
Tier 1 Transit providers: Restrictive
All other transit providers: Selective
Stubs: Open

16. Execution of a sample path
No exogenous changes
Finite states
Depeering
Peering
Transit provider selection
Peering strategy update
Depeering
Peering
Transit provider selection
Peering strategy update
Depeering
Peering
Transit provider selection
Peering strategy update
Iteration
Iteration 1 2 N 1 2 N
Time

17. results

18. Stability of the model
Equilibrium: No topology, peering strategy changes in two consecutive iterations
90% simulations reach equilibrium
Short time scales
Average time to equilibrium: 6 iterations 1 2 N
Iteration
Time

19. Oscillations: An artifact?
10% simulations oscillate
Always involve Tier-1 ASes
Resemble real Tier-1 peering disputes
GENESIS captures that endogenous dynamics cannot always produce stable network
Exogenous intervention required 1 2 N
Iteration
Time

20. Validation Comprehensive validation not possible
Should be viewed as proof of concept
10% ASes end up being transit providers
Average path length 3.7 (500 nodes) vs. Average Internet measured path length 4
Path length does not increase significantly as GENESIS scales from 500 to 1000 nodes

21. Validation Highly skewed degree distribution
Not exactly a power law owing to limited number of nodes
Few links carry several orders of magnitude more traffic

22. Variability across equilibria
Sources of variation in a single population: Initial topology, Playing order
Same population but different initial topology: 85% distinct equilibria
Same population & initial topology but different playing order: 90% distinct equilibria
Distinct equilibria quite similar in terms of topology
Coefficient of variation of fitness close to zero for 90% ASes

23. Variability across equilibria
Most predictable ASes
Stubs: Enterprise customers, Small ISPs
Very large transit providers
Most unpredictable ASes
Midsize (regional) transit providers

24. Case study: Peering Openness
How does peering openness affect the properties of the network?
Optimal fitness in range of peering ratios observed in the real world (1.5 to 5)

25. Case study: Peering Openness
Widespread peering: Saving on costs not the only outcome
Results in loss of transit revenue

26. Summary of GENESIS findings
Individual AS status hard to predict
Regional transit providers most sensitive to network level changes
Overall network characteristics more predictable
Internet a stable network (mostly) in the absence of exogenous factors
Increased peering may result in loss of transit revenue

27. How can I use GENESIS in my research?
Presence at IXPs
Presence at IXPs
Flexible & Modular
Resulting network
Pricing schemes
Traffic matrix
Peering strategies
Peering strategies

28. How can I use GENESIS in my research?
C++ single thread implementation
Fast: average simulation time for 500 nodes: 1.25 hours
Scales up to 1000 nodes
Used in “Analysis of peering strategy adoption by transit providers in the Internet” NetEcon 2012
Available at:

29. Thank YOu