A Multi-Level Approach for Evaluating Internet Topology Generators - - PowerPoint PPT Presentation

a multi level approach for evaluating internet topology
SMART_READER_LITE
LIVE PREVIEW

A Multi-Level Approach for Evaluating Internet Topology Generators - - PowerPoint PPT Presentation

A Multi-Level Approach for Evaluating Internet Topology Generators Ryan Rossi 1 , Sonia Fahmy 1 , Nilothpal Talukder 1 , 2 1 Purdue University, IN 2 Rensselaer Polytechnic Institute, NY Email: { rrossi,fahmy } @cs.purdue.edu, talukn@cs.rpi.edu May


slide-1
SLIDE 1

A Multi-Level Approach for Evaluating Internet Topology Generators

Ryan Rossi1, Sonia Fahmy1, Nilothpal Talukder1,2

1Purdue University, IN 2Rensselaer Polytechnic Institute, NY

Email: {rrossi,fahmy}@cs.purdue.edu, talukn@cs.rpi.edu May 23, 2013

1 / 24

slide-2
SLIDE 2

Motivation

Why Topology Generators?

◮ Generate representative network topologies of different sizes ◮ Used for experiments to design protocols, predict performance, and

understand robustness and scalability of the future Internet

◮ Unfortunately, many fail to capture static and evolutionary properties

  • f today’s Internet, e.g., assume average path length and clustering

coefficient constant Our goal:

◮ Determine how to quantitatively assess a generator through a

multi-level hierarchy of graph, node and link measures

◮ Focus on 2 popular generators: Orbis [SIGCOMM06] and WIT

[INFOCOM07]

◮ Validate using different views of the Internet: data (traceroute),

control (BGP tables), and management (WHOIS) planes

2 / 24

slide-3
SLIDE 3

Taxonomy of Topology Generators

Generators Process Model Type Topology WIT Random-walks Parametric AS RSurfer Parametric N/A Orbis Optimization Data-driven AS & RL HOT Parametric RL

  • Mod. HOT

Parametric AS AB Preferential Parametric N/A BRITE Attachment Data-driven AS & RL Inet Parametric AS GLP Parametric AS SWT Geometry Parametric AS & RL GT-ITM Parametric AS & RL

3 / 24

slide-4
SLIDE 4

Orbis Topology Generator [SIGCOMM 2006]

◮ Series of measures based on degree correlations ◮ The first few dK distributions are:

◮ 0K (average degree) ◮ 1K (degree distribution: P(k) = n(k)/n) ◮ 2K (joint degree distribution: P(k1, k2) = m(k1, k2)µ(k1, k2)/(2m),

where µ(k1, k2) = 2 if k1 = k2, otherwise 1)

◮ 3K (wedges and triangles), etc.

◮ Fails to capture global characteristics ◮ d must be small in practice due to increasing complexity ◮ Relies on rescaling technique; inaccurate as topology becomes larger

4 / 24

slide-5
SLIDE 5

WIT Topology Generator [INFOCOM 2007]

◮ Captures the “wealth” of ISPs over time ◮ Multiplicative stochastic process, ui(t) = λi(t) ui(t − 1), where ui(t)

is the unscaled wealth and λi(t) is an independent random variable

◮ wi(t) is the normalized wealth for node i, and zi(t) = C · di(t) is the

expense

◮ In each iteration,

◮ If wi(t) − zi(t) > C + T, place a link between the node i and an

arbitrary node by randomly walking l-steps from i

◮ If wi(t) − zi(t) > −T, remove a random link of node i

◮ Threshold T is carefully chosen to avoid oscillation

5 / 24

slide-6
SLIDE 6

Orbis versus WIT

◮ WIT attempts to model the evolution of the AS topology

◮ Fails when the underlying process and growth of the Internet change

◮ Orbis generates topologies that preserve a set of measures

◮ Fails if the set of characteristics is incomplete w.r.t. the actual AS

topology

◮ What is the best representative set of local and global measures?

6 / 24

slide-7
SLIDE 7

Network Properties

Measure Importance in Computer Networks Local Degree Fault tolerance, local robustness Assortativity Clustering coefficient Path diversity, fault tolerance, local ro- bustness Distance Scalability, performance, protocol design Global Betweenness Traffic engineering, potential congestion points Eigenvector Network robustness, performance, clus- ters/hierarchy, traffic engineering

7 / 24

slide-8
SLIDE 8

Measures Used

The order of evaluation measures in terms of the difficulty of preservation Link ≥ Node ≥ Graph Graph Measures

◮ Traditional: Average degree, Assortativity coefficient, Average

clustering and Average distance, etc.

◮ Additional: largest singular value (λ1), Network conductance

(λ1 − λ2), radius, and diameter, etc. Node Measures

◮ Traditional: Degree distribution, Clustering coefficient, distance,

eccentricity, betweenness, etc.

◮ Additional: Network values, Scree Plots, K-walks, K-core, etc.

8 / 24

slide-9
SLIDE 9

Measures Used (cont’d)

Link Measures

◮ Order of the nodes with respect to the magnitude of their coordinates

along the principal direction

◮ The closest k-approximation of the topology

Community measures

◮ Louvain’s modularity

9 / 24

slide-10
SLIDE 10

Measures Used (cont’d)

Quantitative Measures

◮ Graph based:

◮ The normalized root-mean-square error (NRMSE)

DNRMSE( x, ˆ

  • x) =

E[( x − ˆ

  • x)2]

max( x, ˆ

  • x) − min(

x, ˆ

  • x)

.

◮ Node based:

◮ Kolmogorov-Smirnov (KS): KS(F1, F2) = maxx |F1(x) − F2(x)| . ◮ Kullback-Leibler (KL) divergence:

DKL(P||Q) =

  • i

P(i) ln P(i) Q(i).

10 / 24

slide-11
SLIDE 11

Learning Graph Measures [ReFeX, SIGKDD 2011]

Instead of selecting a set of graph measures, we automatically learn a set

  • f graph measures recursively.
  • 1. Base set of measures. The process starts by computing degree

(in/out/total edges) and egonet measures (in/out egonet).

◮ egonet includes the node, its neighbors, and any edges in the

induced subgraph on these nodes.

  • 2. Aggregate measures. The existing measures of a node are aggregated

to create additional measures by taking the sum/mean of the neighbors (done in a recursive fashion). One simple measure is the mean value of the degree among all neighbors of a node.

  • 3. Prune correlated measures. At each iteration, we test for redundant

measures using a simple correlation test, and remove all measures that are highly correlated.

  • 4. Stopping Criteria. Repeat steps 2-3 until no new measures are

retained.

11 / 24

slide-12
SLIDE 12

Evaluation Strategy

  • 1. Given G ⋆

n of size n, generate same size graph Gn s.t. M(Gn) ≈ M(G ⋆ n )

  • 2. Given G ⋆

n of size n, generate Gm of size m where m ≥ n s.t.

M(Gm) ≈ M(G ⋆

n )

  • 3. Given an ordered sequence G ⋆

t for t = 1, 2, ..., m, generate a

corresponding sequence Gt for t = 1, 2, ..., m s.t. Gt is the same size as G ⋆

t and M(Gt) ≈ M(G ⋆ t )

12 / 24

slide-13
SLIDE 13

Datasets for Validation

◮ Skitter traceroute ◮ RouteViews’ BGP tables (RV)1 ◮ RIPE’s WHOIS ◮ HOT ◮ RocketFuel

1AS-level subgraphs for 2004-2012 13 / 24

slide-14
SLIDE 14

Results: Graph Measures

2005 2006 2007 2008 2009 2010 2011 2012 4.15 4.2 4.25 4.3 4.35

(a) Average Degree

2005 2006 2007 2008 2009 2010 2011 2012 0.21 0.22 0.23 0.24 0.25 0.26

(b) Average Clustering

2005 2006 2007 2008 2009 2010 2011 2012 −0.205 −0.2 −0.195 −0.19 −0.185 −0.18

(c) Assortativity

2005 2006 2007 2008 2009 2010 2011 2012 3.75 3.8 3.85 3.9

(d) Characteristic Path Length

2005 2006 2007 2008 2009 2010 2011 2012 9 10 11 12 13 14 15 16 17

(e) Diameter

2005 2006 2007 2008 2009 2010 2011 2012 70 75 80 85

(f) Largest Eigenvalue

14 / 24

slide-15
SLIDE 15

Results: Node Measures

10 10

1

10

2

10

3

10

4

10

−5

10

−4

10

−3

10

−2

10

−1

10 Degree CCDF RV Orbis WIT

(a) Degree

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Clustering Coefficient CCDF RV Orbis WIT

(b) Clustering Coefficient

5 10 15 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Distance CCDF RV Orbis WIT

(c) Distance

5 10 15 20 25 30 10

−3

10

−2

10

−1

10 K−cores CCDF RV Orbis WIT

(d) K-cores

15 / 24

slide-16
SLIDE 16

Results: Node Measures

(a) Scree plot (b) Network values

16 / 24

slide-17
SLIDE 17

Results: Link Measures

(a) WHOIS (b) HOT (c) RocketFuel (d) Orbis (WHOIS) (e) Orbis (HOT) (f) Orbis (RocketFuel)

17 / 24

slide-18
SLIDE 18

Results: Quantitative Measures

Table : Quantitative Evaluation of Orbis using KS Distance.

Deg. CC Ecc. Kcores PR EigDiff Net-Value Hot 0.009 0.000 0.000 0.078 0.067 0.588 0.131 RF 0.013 0.450 0.000 0.088 0.215 0.629 0.680 Whois 0.059 0.480 0.224 0.060 0.536 0.169 0.159 Skitter 0.010 0.211 0.029 0.009 0.342 0.096 0.182

18 / 24

slide-19
SLIDE 19

Results: Community Measures

Table : Evaluating the Community Structure of the Topologies.

Communities Q Nodes Edges Degree CC 2004 RouteViews 24 0.65 3951 13360 3.38 0.45 Orbis 46 0.48 957 2254 2.36 0.10 WIT 57 0.92 755 2653 3.51 0.64 2011 RouteViews 34 0.68 6048 18496 3.06 0.22 Orbis 60 0.48 2347 5640 2.40 0.12 WIT 66 0.94 2095 11727 5.60 0.45 Communities C-path Radius Diameter 2004 RouteViews 24 2.74 3 6 Orbis 46 3.01 4 8 WIT 57 2.75 4 7 2011 RouteViews 34 3.27 5 9 Orbis 60 2.91 4 8 WIT 66 3.44 5 10

19 / 24

slide-20
SLIDE 20

Results: Selected versus Learned Measures

2005 2006 2007 2008 2009 2010 2011 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45

Normalized RMSE

Orbis (vs. RV) WIT (vs. RV)

(a) Selected Measures

2005 2006 2007 2008 2009 2010 2011 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

(b) Learned Measures

20 / 24

slide-21
SLIDE 21

Results: Learned Graph Measures

(a) RV (Internet) (b) Orbis (c) WIT

21 / 24

slide-22
SLIDE 22

Conclusions

◮ We propose a multi-level framework for understanding Internet

topologies, and evaluating generators (focus on Orbis, WIT)

◮ We leverage both macro measures (graph) and micro measures (node

and link measures) to accurately compare topologies

◮ We show that the existing generators fail to capture static and

evolutionary properties of the Internet AS topology

◮ Data-driven generators generate static topologies with little or no

variance

◮ Parametric generators typically cannot accurately model Internet

evolution

22 / 24

slide-23
SLIDE 23

Future Directions

◮ Investigate additional topology generators ◮ Develop a parameter estimation technique for WIT and analyze its

behavior with the refined parameters

◮ Study Internet evolution and investigate causes for the changes we

  • bserved

23 / 24

slide-24
SLIDE 24

Thank you. Questions?

24 / 24