the internet is like a jellyfish michalis faloutsos uc riverside uc riverside joint work with:...
TRANSCRIPT
The Internet Is Like A JellyfishThe Internet Is Like A Jellyfish
Michalis FaloutsosMichalis Faloutsos
UC RiversideUC Riverside
Joint work with:Joint work with:
Leslie Tauro, Georgos Siganos (UCR)Leslie Tauro, Georgos Siganos (UCR)
Chris Palmer(CMU) Chris Palmer(CMU)
Michalis Faloutsos 2
Big Picture: Modeling the InternetBig Picture: Modeling the Internet
Measure and model each componentMeasure and model each component• Identify simple properties and patternsIdentify simple properties and patterns
Model and simulate their interactionsModel and simulate their interactions
Topology
Protocols
Traffic
Routing, Congestion ControlRouting, Congestion Control
Michalis Faloutsos 3
Power-law: Frequency of degree vs. degree
The Goal of Internet ModelingThe Goal of Internet Modeling
Find Find simplesimple fundamental properties fundamental propertiesUnderstand why they appear and their effectsUnderstand why they appear and their effects
A real Internet instance
Michalis Faloutsos 44
Claim: We Need The Right ToolsClaim: We Need The Right Tools
““This is just not effective…This is just not effective…
We need to get some chains”We need to get some chains” The Far Side -- G. Larson
Michalis Faloutsos 5
The World Wide Web is a Bow-TieThe World Wide Web is a Bow-Tie
Captures several properties [Captures several properties [WWW-Tomkins et alWWW-Tomkins et al]]
The components are of comparable sizeThe components are of comparable size
In outStrongly connected
•Core•In•Out•Tendrils
Michalis Faloutsos 6
The Accuracy-Intuition Space Of The Accuracy-Intuition Space Of ModelsModels
More tools…More tools…• Self-similaritySelf-similarity• Power-lawsPower-laws• WaveletsWavelets• EigenvaluesEigenvalues
……less intuitionless intuition• Something a human Something a human
can picture can picture
Is it a real conflict?Is it a real conflict?Accuracy
Intuition
Ideal
Clueless
trend
low
low high
highNaive
Complex
Michalis Faloutsos 7
Why Do We Need an Intuitive Model?Why Do We Need an Intuitive Model?
Human mind is simpleHuman mind is simple
Visualizable: creates a mental pictureVisualizable: creates a mental picture
Memorable: captures the main propertiesMemorable: captures the main properties
MaximizesMaximizes information/effortinformation/effort ratioratio
Makes you thinkMakes you think
Michalis Faloutsos 8
What does the Internet look like?What does the Internet look like?
Can I develop a simple model of the AS Can I develop a simple model of the AS Internet topology that I can Internet topology that I can draw by handdraw by hand??
Can I identify a sense of hierarchy in the Can I identify a sense of hierarchy in the network?network?
Focus: Autonomous Systems topologyFocus: Autonomous Systems topology
and data from NLANRand data from NLANR
Michalis Faloutsos 9
Possible Topological ModelsPossible Topological Models
FurballFurball BroomBroom DonutDonut
(One degree nodes are(One degree nodes are
at the periphery)at the periphery)(Military Hierarchy)(Military Hierarchy) (Circular connectivity:(Circular connectivity:
Around the earth?)Around the earth?)
Michalis Faloutsos 10
An Intuitive Model : The Internet An Intuitive Model : The Internet Topology as JellyfishTopology as Jellyfish
Highly connected nodes Highly connected nodes form the coreform the coreEach Shell: adjacent Each Shell: adjacent nodes of previous shell, nodes of previous shell, except 1-degree nodesexcept 1-degree nodesImportanceImportance decreases as decreases as we move away from corewe move away from core1-degree nodes hanging1-degree nodes hangingThe denser the 1-degree The denser the 1-degree node population the node population the longer the stemlonger the stem
CoreShells 12
3
Michalis Faloutsos 11
Roadmap Roadmap
Identify a HierarchyIdentify a Hierarchy• Define the Importance of a nodeDefine the Importance of a node
Present topological propertiesPresent topological propertiesPresent the jellyfish model Present the jellyfish model Why is the jellyfish a good model?Why is the jellyfish a good model?ConclusionsConclusionsAppendix: Latest News on power-laws for the Appendix: Latest News on power-laws for the Internet topologyInternet topology
Michalis Faloutsos 12
How Can We Develop a Simple How Can We Develop a Simple Model?Model?
We need an anchor and a compassWe need an anchor and a compass
Anchor: Anchor: • We need a starting point in the networkWe need a starting point in the network
Compass:Compass:• We want to classify nodes according to We want to classify nodes according to importanceimportance
Michalis Faloutsos 13
Defining the Importance of a NodeDefining the Importance of a Node
The topological importance has many The topological importance has many aspectsaspectsDegreeDegree: number of adjacent nodes: number of adjacent nodesEccentricityEccentricity: the maximum distance: the maximum distance
to any other nodeto any other node
SignificanceSignificance: Significant nodes are near :: Significant nodes are near :1.1. many nodesmany nodes2.2. significant nodessignificant nodes
Michalis Faloutsos 14
Significance of a NodeSignificance of a Node
The significance of a node is the sum of the The significance of a node is the sum of the significance of its neighbors.significance of its neighbors.The iterative procedure convergesThe iterative procedure converges• At each round, total significance is normalized to 1At each round, total significance is normalized to 1
This is equivalent to This is equivalent to [Kleinberg ‘96][Kleinberg ‘96]::• the eigenvector of the max eigenvalue of the adjacency the eigenvector of the max eigenvalue of the adjacency
matrix matrix
Relative Significance: Signif. times No. NodesRelative Significance: Signif. times No. Nodes• Relative Significance = 1, fair share of significanceRelative Significance = 1, fair share of significance
Michalis Faloutsos 15
Roadmap Roadmap
Identify a HierarchyIdentify a Hierarchy• Defining the Importance of a nodeDefining the Importance of a node
Present topological propertiesPresent topological properties
Present the jellyfish model Present the jellyfish model
Why is the jellyfish a good model?Why is the jellyfish a good model?
ConclusionsConclusions
Michalis Faloutsos 16
Observation 1: Significance and Observation 1: Significance and Eccentricity Are CorrelatedEccentricity Are Correlated
Significant nodes have low eccentricitySignificant nodes have low eccentricityIntuitively, significant nodes are in the middle of the Intuitively, significant nodes are in the middle of the network network [Global Internet ‘01][Global Internet ‘01]
SignificanceSignificance
EccentricityEccentricity
Michalis Faloutsos 17
Observation 2: Many One-Degree Observation 2: Many One-Degree Nodes Connect to High-Degree NodesNodes Connect to High-Degree Nodes
One-degree nodes are scattered everywhereOne-degree nodes are scattered everywhereThe distribution of one-degree nodes follows a powerlawThe distribution of one-degree nodes follows a powerlaw
Order of decreasingOrder of decreasing
neighborsneighbors
No of 1-No of 1-degree degree neighborsneighbors
The failure of The failure of
Furball modelFurball model
Michalis Faloutsos 18
Observation 3: The Internet Premise:Observation 3: The Internet Premise:One Robust Connected NetworkOne Robust Connected Network
Robust to random, sensitive to focused failuresRobust to random, sensitive to focused failures
The network tends to stay as one connected componentThe network tends to stay as one connected component
Size of Size of
LargestLargest
ConnecteConnectedd
ComponeComponentnt
0
1000
2000
3000
4000
0332 664 996 132816601992232426562988
Iterations
Random
Highest Degree firstorderHighest Significancefirst order
#Deleted nodes#Deleted nodes
Michalis Faloutsos 19
Observation 4: The Number of Observation 4: The Number of Alternate Paths Between Two NodesAlternate Paths Between Two Nodes
All alternate paths go through the same directionAll alternate paths go through the same direction
No shortcuts or loop-aroundsNo shortcuts or loop-arounds
Path Length
Number of paths
The Failure of the Donut Model
Michalis Faloutsos 20
Roadmap Roadmap
Identify a HierarchyIdentify a Hierarchy• Defining the Importance of a nodeDefining the Importance of a node
Present topological propertiesPresent topological properties
Present the jellyfish model Present the jellyfish model
Why is the jellyfish a good model?Why is the jellyfish a good model?
ConclusionsConclusions
Michalis Faloutsos 21
Defining a Hierarchy RecursivelyDefining a Hierarchy Recursively
Define the core:Define the core:• Maximal clique of highest Maximal clique of highest
degree nodedegree node
Define the Layers:Define the Layers:• All nodes adjacent to previous All nodes adjacent to previous
layerlayer
Define the Shells:Define the Shells:• A layer without its one-degree A layer without its one-degree
nodesnodes
Michalis Faloutsos 22
The Internet Topology as a JellyfishThe Internet Topology as a Jellyfish
CoreCore: High-degree : High-degree cliqueclique
ShellShell: adjacent nodes of : adjacent nodes of previous shell, except previous shell, except 1-degree nodes1-degree nodes
1-degree nodes1-degree nodes: shown : shown hanginghanging
The denser the 1-The denser the 1-degree node population degree node population the longer the stemthe longer the stem
CoreShells 12
3
Michalis Faloutsos 23
The Hierarchy: The Model Respects The Hierarchy: The Model Respects the Node Importance the Node Importance
-6
-4
-2
0
2
4
6
8
Core Layer-1
Layer-2
Layer-3
Layer-4
Layer-5
Eff EccentricityLog Relative SignificanceLog Degree
The importance of The importance of nodes decreases as we nodes decreases as we move away from the move away from the corecore
The effective The effective eccentricity decreases eccentricity decreases by one in each layerby one in each layer
(see paper for details)(see paper for details)
Michalis Faloutsos 24
The Evolution of the JellyfishThe Evolution of the Jellyfish
The structure of the The structure of the jellyfish has not jellyfish has not changed much in changed much in 1997-20001997-2000
Nodes become more Nodes become more connected:connected:
Small increase in Small increase in shells and decrease shells and decrease in hanging nodesin hanging nodes
0
5
10
15
20
25
30
35
40
CoreHang-0Shell1Hang-1Shell2Hang-2Shell3Hang-3Shell4Hang-4
shell/hang
% of total graph
8/11/1997 4/30/1998 8/1/1998 11/30/1998
4/30/1999 7/15/1999 10/10/1999 Jun-00
Michalis Faloutsos 25
The Diameter Remains ConstantThe Diameter Remains Constant
6 hops reach approximately 98% of the network!6 hops reach approximately 98% of the network!The jellyfish diameter remains the sameThe jellyfish diameter remains the same
PercentagePercentage
Of nodesOf nodes
reachedreached
TimeTime
HopsHops66
55
44
Michalis Faloutsos 26
Theory Supports the Jellyfish!Theory Supports the Jellyfish!
A surprising theoretical result A surprising theoretical result [Reittu Norros 03][Reittu Norros 03]
• A network with powerlaw degree -> jellyfishA network with powerlaw degree -> jellyfish
Assume degree powerlaw and random Assume degree powerlaw and random connectionsconnections• The network will have a clique of high degree nodesThe network will have a clique of high degree nodes• The diameter of the network is O(log logN)!The diameter of the network is O(log logN)!
In total aggrement with our observationsIn total aggrement with our observations
Michalis Faloutsos 27
Why Is The Jellyfish a Good Model?Why Is The Jellyfish a Good Model?
It’s cute, in addition…It’s cute, in addition…
Michalis Faloutsos 28
The Jellyfish Captures Many The Jellyfish Captures Many PropertiesProperties
The network is compact:The network is compact:• 99% of pairs of nodes are within 6 hops99% of pairs of nodes are within 6 hops
There exists a highly connected centerThere exists a highly connected center• Clique of high degree nodesClique of high degree nodes
There exists a loose hierarchy:There exists a loose hierarchy:• Nodes far from the center are less importantNodes far from the center are less important
One-degree nodes are scattered everywhereOne-degree nodes are scattered everywhereThe network has the tendency to be one The network has the tendency to be one large connected componentlarge connected component
Michalis Faloutsos 29
And It Looks Like A Jellyfish…And It Looks Like A Jellyfish…
Independent Independent ObservationObservation
Router Level Router Level TopologyTopology
Produced by CAIDAProduced by CAIDA
Michalis Faloutsos 30
Conclusions Conclusions
We model the Internet as a jellyfishWe model the Internet as a jellyfishThe jellyfish represents graphically several The jellyfish represents graphically several topological propertiestopological properties• Network is compactNetwork is compact• We can identify a centerWe can identify a center• We can define a loose hierarchyWe can define a loose hierarchy• The network tends to be one connected componentThe network tends to be one connected component
Theoretical results support our observationsTheoretical results support our observationshttp://www.cs.ucr.edu/~michalis/http://www.cs.ucr.edu/~michalis/
Michalis Faloutsos 31
My Other Research InterestsMy Other Research Interests
1.1. Characterize and model network behavior:Characterize and model network behavior:• Poisson and Long Range Dependence [GI 02 - Globecom]Poisson and Long Range Dependence [GI 02 - Globecom]
2.2. Model and simulate the Internet topologyModel and simulate the Internet topology• Identify structure and hierarchy [GI 01] [COMNET*]Identify structure and hierarchy [GI 01] [COMNET*]
3.3. Model and simulate BGPModel and simulate BGP• Large scale simulations (10,000 nodes) [GI 02]Large scale simulations (10,000 nodes) [GI 02]
4.4. Wireless networksWireless networks• Improving TCP over ad hoc networks [Globecom 02]Improving TCP over ad hoc networks [Globecom 02]
5.5. Multicast: supporting scalability and QoS Multicast: supporting scalability and QoS (Cui Gerla)(Cui Gerla)• Efficient management through overlay trees [Globecom 02]+Efficient management through overlay trees [Globecom 02]+
6.6. A novel network layer: DART (PeerNet) A novel network layer: DART (PeerNet) [IPTPS 03][IPTPS 03]
Michalis Faloutsos 32
Appendix:Appendix:Latest News on Power-lawsLatest News on Power-laws
The Internet topology can be described by The Internet topology can be described by power-laws power-laws [Faloutsos x 3, SIGCOMM’99][Faloutsos x 3, SIGCOMM’99]
The power-laws are here to stayThe power-laws are here to stay• Appear consistently over five yearsAppear consistently over five years• Even with newer more complete data [Infocom’02]Even with newer more complete data [Infocom’02]
Michalis Faloutsos 33
Powerlaw: Degree Exponent DPowerlaw: Degree Exponent D
Degree distribution of nodes: CCDFDegree distribution of nodes: CCDF
It holds even for the more complete graph: 99%It holds even for the more complete graph: 99%
Newer More Complete AS graphNewer More Complete AS graphRouteViews - NLANR DataRouteViews - NLANR Data
Michalis Faloutsos 34
Thank you!Thank you!
http://www.cs.ucr.edu/~michalis/http://www.cs.ucr.edu/~michalis/
Michalis Faloutsos 35
Eigenvalues of the TopologyEigenvalues of the Topology
Let Let A A be the adjacency matrix of graph be the adjacency matrix of graphThe eigenvalue The eigenvalue is real number s.t.is real number s.t.::• A A vv = = vv, , wherewhere vv some vectorsome vector
Eigenvalues are strongly related to topological propertiesEigenvalues are strongly related to topological propertiesMore details in Part BMore details in Part B
1
3
2 00 11 11
11 00 00
11 00 00
A =
Michalis Faloutsos 36
Power-law: Eigen Exponent Power-law: Eigen Exponent EE
Find the eigenvalues of the adjacency matrixFind the eigenvalues of the adjacency matrixEigenvalues in decreasing order (first 100)Eigenvalues in decreasing order (first 100)
E = -0.48
Exponent = slope
Eigenvalue
Rank of decreasing eigenvalue
May 2001
Michalis Faloutsos 37
Surprising Result!Surprising Result!
Exponent E is half of exponent DExponent E is half of exponent DTheorem: Given a graph with relatively large degrees Theorem: Given a graph with relatively large degrees ddii then with high probability: then with high probability:• Eigenvalue Eigenvalue ii = = d di,i, , where i rank of decreasing order , where i rank of decreasing order
Thus, if we compare the slope of the plot the Thus, if we compare the slope of the plot the eigenvalues and the degrees:eigenvalues and the degrees:
• log log ii = = 0.5 log d 0.5 log dii
[[Fabrikant, Koutsoupias, Papadimiitriou in STOC’01]Fabrikant, Koutsoupias, Papadimiitriou in STOC’01]
[Mihail Papadimitriou Random 02][Mihail Papadimitriou Random 02]
Michalis Faloutsos 38
Time Evolution of The Topology Time Evolution of The Topology
Powerlaws are here to stayPowerlaws are here to stay
Degree distribution slope is invariantDegree distribution slope is invariant
Network becomes denserNetwork becomes denser
The rich get richer phenomenonThe rich get richer phenomenon
Michalis Faloutsos 39
The Number of ASes in TimeThe Number of ASes in Time
The number of AS doubled in two years The number of AS doubled in two years Growth seems to slow down!Growth seems to slow down!
3 K
13 K
1997 2002
Michalis Faloutsos 40
Degree Distribution Did Not Change!Degree Distribution Did Not Change!
Slope is practically constant for over 3 yearsSlope is practically constant for over 3 years
Michalis Faloutsos 41
The Topology Becomes Denser!The Topology Becomes Denser!
6 hops reach approximately 98% of the network!6 hops reach approximately 98% of the network!Denser: 6 hops reach more nodesDenser: 6 hops reach more nodes
Recall six Recall six
degrees ofdegrees of
separationseparation
Michalis Faloutsos 42
The Rich Get RicherThe Rich Get Richer
The increase of the degree versus the initial degreeThe increase of the degree versus the initial degreeNew connections prefer “highly connected nodes”New connections prefer “highly connected nodes”
Michalis Faloutsos 43
That’s it!That’s it!
Thank you!Thank you!
http://www.cs.ucr.edu/~michalis/http://www.cs.ucr.edu/~michalis/
Michalis Faloutsos 44
I. Power-law: rank exponent I. Power-law: rank exponent RR
The plot is a line in log-logThe plot is a line in log-log scalescale
Exponent = slope
R = -0.74
R
degree
Rank: nodes in decreasing degree order
Dec’98
[Faloutsos, Faloutsos and Faloutsos SIGCOMM’99]
Michalis Faloutsos 45
I. Estimations Using With Rank Exponent I. Estimations Using With Rank Exponent RR
Lemma:Lemma:
Given the nodes Given the nodes N, N, and an estimate for the and an estimate for the rank exponent rank exponent R, R, we predict the edges E:we predict the edges E:
NNR R
⋅−⋅+
=Ε + )1
1()1(2
11
Michalis Faloutsos 46
Some Current ResultsSome Current Results
Measuring the performance of real-time applicationsMeasuring the performance of real-time applications• E2e performance is asymmetric (by 10)E2e performance is asymmetric (by 10)
Estimating Long Range DependenceEstimating Long Range Dependence• No definitive estimating method existsNo definitive estimating method exists• SELFIS software tool for performance analysisSELFIS software tool for performance analysis
A study of BGP routing robustness A study of BGP routing robustness • Persistence and prevalence of pathsPersistence and prevalence of paths• Paths are fairly robust, but there is a lot of “noise” tooPaths are fairly robust, but there is a lot of “noise” too• A data repository: 107Gb, 1 billion BGP pathsA data repository: 107Gb, 1 billion BGP paths
Michalis Faloutsos 47
Measuring Real-time PerformanceMeasuring Real-time Performance
““Can the Internet support VoIP Can the Internet support VoIP nownow?”?”
We conduct globe-wide experimentsWe conduct globe-wide experiments• UCR, CMU, Japan, Australia, GreeceUCR, CMU, Japan, Australia, Greece
Experimental set-upExperimental set-up• Approx. 6 4Kbit/sec sending rateApprox. 6 4Kbit/sec sending rate• Small packet sizes every 20, 30, 40, 50 msec, 1 Small packet sizes every 20, 30, 40, 50 msec, 1
secsec
Michalis Faloutsos 48
How Many Distinct Paths Does an IP How Many Distinct Paths Does an IP Prefix Use?Prefix Use?
Almost 70% of the IP prefixes have Almost 70% of the IP prefixes have 2-10 distinct paths2-10 distinct paths30% of IP prefixes have only one path30% of IP prefixes have only one path
30%30%
70%70%
Michalis Faloutsos 49
For More InformationFor More Information
www.cs.ucr.edu/~michalis/www.cs.ucr.edu/~michalis/
Michalis Faloutsos 50
Routing Is PersistentRouting Is Persistent
CDF of the relative duration of CDF of the relative duration of the most persistent paththe most persistent path
70% of prefixes use 70% of prefixes use one path one path continuously for continuously for 50% of their time!50% of their time!
70%70%
Michalis Faloutsos 51
Measurements: The Death of the Measurements: The Death of the Symmetry Assumption Symmetry Assumption
One-way delay:One-way delay:
Forward can be 10 times higher than backward delayForward can be 10 times higher than backward delay
Michalis Faloutsos 52
Characterizing Network Behavior Characterizing Network Behavior with Long Range Dependence with Long Range Dependence
LRD captures the “memory” of the behaviorLRD captures the “memory” of the behavior
It is quantified by a single scalar numberIt is quantified by a single scalar number
LRD appears in many aspects of networksLRD appears in many aspects of networks• Traffic load, arrival times, delays, packet lossTraffic load, arrival times, delays, packet loss
Open Question: what does it really tell us?Open Question: what does it really tell us?
PROBLEM:PROBLEM: We do not know how to calculate LRD! We do not know how to calculate LRD!• Many estimators with conflicting estimatesMany estimators with conflicting estimates• No systematic approachNo systematic approach
Michalis Faloutsos 53
The Intuition Behind LRDThe Intuition Behind LRD
White NoiseWhite Noise
Pink NoisePink Noise
Brownian NoiseBrownian Noise
Capturing the “dependency” of the signal to Capturing the “dependency” of the signal to its previous valuesits previous values
Michalis Faloutsos 54
Idea: Reverse Engineering LRDIdea: Reverse Engineering LRD
Develop a library of behaviors to know dataDevelop a library of behaviors to know data
Three Series of Tests for the EstimatorsThree Series of Tests for the Estimators
1.1. Evaluating the accuracy of the estimatorsEvaluating the accuracy of the estimators• Synthetic Fractional Gaussian Noise (FGN) Synthetic Fractional Gaussian Noise (FGN)
2.2. Deceiving the estimators with non-LRD data Deceiving the estimators with non-LRD data Periodicity, Noise, TrendPeriodicity, Noise, Trend
3.3. Applying the estimators in real data Applying the estimators in real data • Characterizing delay and packet lossCharacterizing delay and packet loss
Michalis Faloutsos 55
BGP Routing AnalysisBGP Routing Analysis
Overarching Goal:Overarching Goal:• Develop a realistic detailed model for large scale Develop a realistic detailed model for large scale
realistic simulationsrealistic simulations
Now: A study of BGP routing robustnessNow: A study of BGP routing robustness• Persistence and prevalence of pathsPersistence and prevalence of paths• Stability of advertisementsStability of advertisements
Next step: Next step: • Study the customer-provider relationshipsStudy the customer-provider relationships
Michalis Faloutsos 56
Using Massive BGP Routing DataUsing Massive BGP Routing Data
We use data from NLANR for almost 3 yearsWe use data from NLANR for almost 3 years• Late 1997 to early 2001Late 1997 to early 2001
Daily snapshots of BGP routing tablesDaily snapshots of BGP routing tables
Created a database to facilitate path queriesCreated a database to facilitate path queries• 107Gb of data, 1 billion BGP paths107Gb of data, 1 billion BGP paths
Michalis Faloutsos 57
Overview of Results for BGP RoutingOverview of Results for BGP Routing
Stable and persistent routing with some “noise”Stable and persistent routing with some “noise”
44% prefixes are advertised for < 30 days44% prefixes are advertised for < 30 days
50% prefixes have a dominant path 84% of time50% prefixes have a dominant path 84% of time
35% of prefixes use one path continuously for 35% of prefixes use one path continuously for 90% of their time!90% of their time!Significant path multiplicity due to traffic engin.Significant path multiplicity due to traffic engin.
Michalis Faloutsos 58
Graph Reduction ToolsGraph Reduction Tools
Reduce: large real graph to small realistic graphReduce: large real graph to small realistic graph• Achieve 70% reduction Achieve 70% reduction
Satisfy degree distribution, but increases diameterSatisfy degree distribution, but increases diameter
Reduce
Large
Small
Michalis Faloutsos 59
The Jellyfish Captures “Direction” The Jellyfish Captures “Direction” of Connectivityof Connectivity
Most edges are Most edges are between layers between layers 80% 80%
Less edges are Less edges are within a layer within a layer 20%20%
0
5
10
15
20
25
30
35
40
45
Layer
8/11/1997 4/30/1998 8/1/1998 11/30/1998
4/30/1999 7/15/1999 10/10/1999 Jun-00
Michalis Faloutsos 60
The Model Respects the Node The Model Respects the Node ImportanceImportance
-3
-2
-1
0
1
2
3
4
5
6
7
Core Shell-1 Shell-2 Shell-3 Shell-4
Effective EccentricityLog Relative SignificanceLog Degree
The importance of each The importance of each layer decreases as we layer decreases as we move away from the move away from the corecore
Michalis Faloutsos 61
Intuitive Models Are UsefulIntuitive Models Are Useful
Cons:Cons:• Danger of oversimplificationDanger of oversimplification
Pros:Pros:• MemorableMemorable• VisualizableVisualizable• Maximizing information/effort ratioMaximizing information/effort ratio
They can be very useful when exploring They can be very useful when exploring unknown territoryunknown territory• Even disproving a wrong model is progress!Even disproving a wrong model is progress!