HY-483 Presentation

29
HY-483 Presentation On power law relationships of the internet topology A First Principles Approach to Understanding the Internet’s Router-level Topology On natural mobility models

description

HY-483 Presentation. On power law relationships of the internet topology A First Principles Approach to Understanding the Internet’s Router-level Topology On natural mobility models. On power law relationships of the internet topology. Michalis Faloutsos - PowerPoint PPT Presentation

Transcript of HY-483 Presentation

Page 1: HY-483 Presentation

HY-483 Presentation On power law relationships of

the internet topology

A First Principles Approach to Understanding the Internet’s Router-level Topology

On natural mobility models

Page 2: HY-483 Presentation

On power law relationships of the internet topology

Michalis Faloutsos U.C. Riverside Dept. of Comp. Science

[email protected] Petros Faloutsos

U. of Toronto Dept. of Comp. Science [email protected]

Christos Faloutsos Carnegie Mellon Univ. Dept. of Comp.

Science [email protected]

Page 3: HY-483 Presentation

Previous work

Heavy tailed distributions used to describe LAN and WAN traffic

Power laws describe WWW traffic

There hasn't been any work on power laws with respect to topology.

Page 4: HY-483 Presentation

Dataset & Methodology Three inter-domain level instances of

the internet (97-98), in which the topology grew by 45%.

Router-level instance of the internet in 1995

Min,Max and Means fail to describe skewed distributions

Linear Regression & correlation coeficients, to fit a plot to a line

Page 5: HY-483 Presentation

First power law: the rank exponent R

Lemma1:

Lemma2:

Page 6: HY-483 Presentation

The rank exponent in AS and router level

Page 7: HY-483 Presentation

Second power law: the outdegree exponent

O

Test of the realism of a graph metric follows a power law exponent is close to realistic numbers

Page 8: HY-483 Presentation

The outdegree exponent O

Page 9: HY-483 Presentation

Approximation: the hop-plot exponent H

Lemma 3:

Definition deff:

Lemma 4: O (d·h^H)

Previous definition O(d^h)

Page 10: HY-483 Presentation

The hop-plot exponent H

Page 11: HY-483 Presentation

Average Neighborhood size

Page 12: HY-483 Presentation

Third power law: the eigen exponent ε

The eigen value λ of a graph is related with the graph's adjacency matrix A (Ax = λx) diameter the number of edges the number of spanning trees the number of CCs the number of walks of a certain length

between vertices

Page 13: HY-483 Presentation

The eigenvalues exponent ε

Page 14: HY-483 Presentation

Contributions-Speculations

Exponents describe different families of graphs

Deff improved calculation complexity from previous O(d^h) to O(d·h^H)

What about 9-20% error in the computation of E?

Page 15: HY-483 Presentation

A First Principles Approach to Understanding the Internet’s

Router-level Topology

Lun Li California Institute of Technology [email protected]

David Alderson California Institute of Technology

[email protected] Walter Willinger

AT&T Labs Research [email protected]

John Doyle California Institute of Technology

[email protected]

Page 16: HY-483 Presentation

Previous work

Random graphs Hierarchical structural models Degree-based topology generators.

Preferential attachment General model of random graphs

(GRG) Power Law Random Graph (PLRG)

Page 17: HY-483 Presentation

A First Principles Approach

Technology constraints Feasible region

Economic considerations End user demands

Heuristically optimal networks Abilene and CENIC

Page 18: HY-483 Presentation

Evaluation of a topology Current metrics are inadequate and lack

a direct networking interpretation Node degree distribution Expansion Resilience Distortion Hierarchy

Proposals Performance related Likelihood-related metrics

Page 19: HY-483 Presentation

Abilene-CENIC

Page 20: HY-483 Presentation

Comparison of simulated topologies with power law degree distributions

and different features

Page 21: HY-483 Presentation

Performane-Likelihood Comparison

Page 22: HY-483 Presentation

Contributions-Speculations Different graphs generated by degree-based

models, with average likelihood, are Difficult to be distinguished with macroscopic

statistic metrics Yield low performance

Simple heuristically design topologies High performance Efficiency

Robustness not incorporated in the analysis Validation with real data

Page 23: HY-483 Presentation

On natural mobility models

Vincent Borrel Marcelo Dias de Amorim Serge Fdida

LIP6/CNRS – Université Pierre et Marie Curie 8, rue du Capitaine Scott – 75015 – Paris – France {borrel,amorim,sf}@rp.lip6.fr

Page 24: HY-483 Presentation

Previous work Individual mobility models

Random Walk Random Waypoint Random Direction model Boundless Simulation Gauss-Markov model, City Section model,

Group mobility models Reference Point model, Exponential Correlated Pursue model

Page 25: HY-483 Presentation

Aspects of real-life networks Scale free property and high clustering

coefficient Biology Computer networks Sociology

Page 26: HY-483 Presentation

Proposal: Gathering Mobility (1/2)Why?

Current group mobility models Rigid Unrealistic

Match reality using scale free distributions Human behavior Research on Ad-hoc inter-contacts

Page 27: HY-483 Presentation

Proposal: Gathering Mobility (2/2)

The model

Individuals Cycle behavior

Attractors Appear-dissapear

Probability an individual to choose an attractor

Attractiveness of an attractor

Page 28: HY-483 Presentation

ExperimentScale-free spatial distribution

Scale-free Population growth

Page 29: HY-483 Presentation

Contributions-Speculations A succesive merge of individual and group

behavior: Individual movement No explicit grouping

Vs Strong collective behaviour Influence by other individuals Gathering around centers of interest of varying

popularity levels

Determination of maintenance of this distribution in case of population decrease and renewal