Selfish Overlay Network Formation Georgios Smaragdakis 1 1 - - PowerPoint PPT Presentation

1 1

Selfish Overlay Network Formation

Georgios Smaragdakis

2 2

Deutsche Telekom Laboratories.

T-Labs US, Stanford University T-Labs, Ben-Gurion University T-Labs, An-Institute of Technische Universität Berlin

3 3

Strategic Research concentrates on long- term technology and applied research.

Intelligent Networks Quality and Usability Lab Security in Telecommunications Service-centric Networking

• Functional groups Networking / Security / Usability
• Network

Measurement and Security

• Routing
• Wireless

Networks

• Virtualization
• Peer to Peer
• Content

Distribution Networks

• Audio

Technology

• Image and

Vision Computing

• Mobile and

Physical Interaction

• Quality
• Speech

Technology

• Usability
• Microkernel

Security

• Vehicular

Security

• Wireless

Security

• Server Security
• Data Security

and Cryptography

• Definition in

2010

4 4

Innovation Development

• Innovation Development develops

innovative solutions, as a basis for the commercial use by the Group‘s business areas.

Strategic Research

Innovation Development and Strategic Research work side by side, to jointly achieve goals.

• Strategic Research concentrates on

the long-term technology research and applied research.

• Strategic Research creates the

foundation for the development of innovative solutions in Innovation Development.

• Results
• Publications
• Patents
• Demonstrations
• Results
• Market studies
• Acceptance tests
• Business models
• Prototypes

5 5

The success of Telekom Laboratories is measured at the transfer to the Group’s business areas or to spin-offs.

6 6

Telekom Laboratories cooperate according to the Open Innovation model with selected research institutes.

Norwegian University of Science and Technology Technische Universität Berlin Fraunhofer-Institut für Nachrichtentechnik Heinrich-Hertz-Institut Fraunhofer-Institut für Offene Kommunikationssystem e Ben-Gurion University Ludwig-Maximilian- Universität München Technische Universität München Rheinische Friedrich- Wilhelms-Universität Bonn Imperial College London École Nationale d’Ingénieurs de Brest Univeridad Carlos III de Madrid Technische Universität Darmstadt Universite Catholique de Louvain École Polytechnique Fédérale de Lausanne Universität St. Gallen Stanford University University of Illinois Boston University Princeton University UC Berkeley/ICSI

7 1

Selfish Overlay Network Formation

Georgios Smaragdakis

Joint work with Nikolaos Laoutaris, Azer Bestavros, John Byers, Pietro Michiardi, Mema Roussopoulos and Vassilis Lekakis

8

O ver l ays

physical plane

O 1 O 2 O 3 R 1 R 2 R 3 R 4

• verlay

plane

process

• verlay

node router, AS

2

9

a c c e s s I SP a c c e s s I SP t r a ns i t I SP

O ver l ay Econom i cs & Nei ghbor Sel ect i

• n

Investment:

Flat Resource Allocation

i nt e r ne t

$ $$ $$

Market:

• verlay links

t r a ns i t I SP a c c e s s I SP $$

• verlay node

3

10 4

Connect i vi t y M anagem ent

• Full mesh architectures for reliability

(e.g. RON)

• Myopic heuristics

random or proximity based neighbor selection

• Tree forest or mesh construction to optimize multicast

(e.g. Bullet, Splitstream)

• Optimization for network delay

(e.g. Detour, QRON)

• Opportunistic choke/unchoke

(e.g. BitTorrent)

• Distributed hashing tables

(e.g. Chord, Pastry, Tapestry)

11

Chal l enges

• Network Heterogeneity:

pair wise delay or available bandwidth, storage, cpu cycles, budget…

• Load Variability:

diurnal variation of traffic, dynamic routing or pricing, node churn…

• Diversity of users:

different prospective, conflicting objectives

Op p

• r

t uni t i e s

5

12

Sel f i sh O ver l ay Net w or k Cr eat i

• n

I m pl i cat i

• ns

t

• Pr
• t
• col

Desi gn EG O I ST

Appl i cat i

• n

t

• sw ar

m i ng syst em s I NFOCOM ’ 7 Tr a ns a c t i

• n
• n

Ne t wo r k i ng Co NEXT 2 8

St r at egi c Resour ce Al l

• cat

i

• n

I nf

• c
• m

2 8 , TPDS

6

13

Sel f i sh O ver l ay Net w or k Cr eat i

• n

I m pl i cat i

• ns

t

• Pr
• t
• col

Desi gn EG O I ST

Appl i cat i

• n

t

• sw ar

m i ng syst em s

St r at egi c Resour ce Al l

• cat

i

• n

7

14

Local Connection Game <V,{si},{Ci}> [Fabrikant et al,PODC’03]

• V: set of n players (nodes)
• {si}: strategies available to vi (wirings)
• {Ci}: set of utilities for vi (cost)
• Outcome: S is the global wiring

Net w or k Cr eat i

• n

∑

−

∈

+ ⋅ =

i j V

v j i S i i

v v d s S c ) , ( | | ) ( α

a a m i n

8

15

O ver l ay Net w or k Cr eat i

• n

Towards a Real i st i c model for Overlay Networks:

• Directed Edges
• Bounded out- and in-degree
• Non-uniform preference vectors
• Realistic models of physical distance

Towards a Ri cher G am e, easi l y r eal i zabl e via a network protocol.

9

16

Sel f i sh Nei ghbor Sel ect i

• n

( SNS)

vi: Choose k neighbors vi

G -

i

=( V-

i

, S-

i

)

u w

∑

−

∈

⋅ =

i j V

v j i S ij i

v v d p S C ) , ( ) (

m i n

• v

e r a l l s

i

∈S

i

vi ’ s r esi dual net w or k

Se t

• f

r e s i d ua l no d e s Se t

• f

r e s i d ua l wi r i ng

10

17

SNS & k- m edi an

Uniform link weights, and uniform preference  k-median on asymmetric distances

11

18

k- m edi an

k- m edi an: Find a subset I of F and a function σ:CI, to: min ( Σi,j sjcij ) such that |I| ≤ k

F: set of facilities C: set of clients, cij: cost connecting client jfacility I sj: demand of node

12

19

k- m edi an

13

20

Non-uniform link weights, and uniform preference  ILP formulation

SNS & k- m edi an

Uniform link weights, and uniform preference  k-median on asymmetric distances

u w

w , u can be

• bt

ai ned f r

• m

k- m edi an

• n

r ever sed di st ances

w u vi

∑

−

∈

⋅ =

i j V

v j i S ij i

v v d p S C ) , ( ) (

m i n

Si nce t he w i r i ng cost i s t he sam e

14

21

Local Sear ch ( LS)

vi: choose k neighbors

vi u w

∑

−

∈

⋅ =

i j V

v j i S ij i

v v d p S C ) , ( ) (

m i n

• v

e r a l l s

i

∈S

i

vi ’ s r esi dual net w or k

[Arya et al,STOC’01]

G -

i

=( V-

i

, S-

i

)

Se t

• f

r e s i d ua l no d e s Se t

• f

r e s i d ua l wi r i ng

15

22

SNS : t he G am e

Game <V,{si},{Ci}>

• V : set of n players (nodes)
• {si}: strategies available to vi (wirings),

choose k out of n to connect

• {Ci}: set of costs for vi

min

Best response of a node: node’s optimal wiring Outcome: S, the global wiring

• A stable wiring is a pure Nash equilibium
• Using iterative best response

 Fundamentally different from selfish routing

∑

−

∈

⋅ =

i j V

v j i S ij i

v v d p S C ) , ( ) (

16

23

SNS : Equi l i br i a

n=15 k=2 k=3 k=8 k=11

Uniform Preference Skewness of preference k (Link density)

I n- degr ees ar e hi ghl y skew ed even under uni f

• r

m pr ef er ence !  Qua l i t y

• b

a s e d “ p r e f e r e nt i a l a t t a c h me nt ”

17

24

Performance of ILP & LS is close to Utopian!

Theoretical results showed in the worst case the social cost can be bad

[Laoutaris, Poplawsi, Rajaraman, Sundaram, Teng,PODC’08]

SNS : Ef f i ci ency

Link density Skewness of preference Link density Skewness of preference

18

25

Sel f i sh O ver l ay Net w or k Cr eat i

• n

I m pl i cat i

• ns

t

• Pr
• t
• col

Desi gn EG O I ST

Appl i cat i

• n

t

• sw ar

m i ng syst em s

St r at egi c Resour ce Al l

• cat

i

• n

19

26

SNS : Tr ace- Dr i ven Eval uat i

• n

How we assign the distance:

• Synthetically using BRITE
• Empirically from PlanetLab
• Empirically from AS-level maps [Routeviews]

Neighbor Selection Strategies:

• k-Random heuristic
• k-Closest heuristic
• k-Regular heuristic
• k-Best Response

Control parameter:

• Bound on out-degree k (link density)

20

27

SNS vs. Heur i st i cs: Soci al Cost

Macroscopic view: Focusing on the social welfare

The network is better off with selfish nodes!

(k=2) k-Random/BR k-Closest/BR k-Regular/BR BRITE 1.44 1.53 3.61 PlanetLab 2.23 1.48 3.84 AS 2.04 1.90 4.78

21

28

Connect i ng

• n

a k- Random gr aph

k k k

AS-Level (n=50) PlanetLab (n=50) BRITE (n=50)

If your neighbors are naïve, it pays to be selfish!

0 2 3 5 11 22 0 2 3 5 11 22 0 2 3 5 11 22

22

29

Connect i ng

• n

a k- Cl

• sest

gr aph

k k k 0 2 3 5 11 22 0 2 3 5 11 22 0 2 3 5 11 22

If your neighbors are greedy, it pays to be selfish!

AS-Level (n=50) PlanetLab (n=50) BRITE (n=50)

23

30

Connect i ng

• n

a k- Regul ar gr aph

k k k 0 2 3 5 11 22 0 2 3 5 11 22 0 2 3 5 11 22

If your neighbors have the same wiring pattern, it pays to be selfish!

“Common pattern is not good”

AS-Level (n=50) PlanetLab (n=50) BRITE (n=50)

24

31

Connect i ng

• n

a Best Response gr aph

The BR graph is highly optimized!

k k k 0 2 3 5 11 22 0 2 3 5 11 22 0 2 3 5 11 22

AS-Level (n=50) PlanetLab (n=50) BRITE (n=50)

If your neighbors are selfish, it is OK to be naï ve!

25

32

Sel f i sh O ver l ay Net w or k Cr eat i

• n

I m pl i cat i

• ns

t

• Pr
• t
• col

Desi gn EG O I ST

Appl i cat i

• n

t

• sw ar

m i ng syst em s

St r at egi c Resour ce Al l

• cat

i

• n

26

33 7

34 28

Key Cont r i but i

• ns

System Architecture

 Link state protocol to support connectivity

information dissemination.

 Overlay monitoring and maintenance mechanism.  Computationally efficient neighbor selection.

Performance Evaluation

 Average performance in real operational scenaria.  Performance under different performance metrics (delay,

system load, available bandwidth)

 Overhead of the implementation.  Performance under churn.  Applications.

35 29

Basi c Ar chi t ect ur e

111. 1. 1. 1 122. 2. 2. 2 133. 3. 3. 3 144. 4. 4. 4

111. 1. 1. 1 122. 2. 2. 2 25m secs 111. 1. 1. 1 133. 3. 3. 3 165m secs

111. 1. 1. 1 122. 2. 2. 2 25m secs 111. 1. 1. 1 133. 3. 3. 3 165m secs 111. 1. 1. 1 122. 2. 2. 2 25m secs 111. 1. 1. 1 133. 3. 3. 3 165m secs 111. 1. 1. 1 122. 2. 2. 2 25m secs 111. 1. 1. 1 133. 3. 3. 3 165m secs 111. 1. 1. 1 122. 2. 2. 2 25m secs 111. 1. 1. 1 133. 3. 3. 3 165m secs 111. 1. 1. 1 122. 2. 2. 2 25m secs 111. 1. 1. 1 133. 3. 3. 3 165m secs 111. 1. 1. 1 122. 2. 2. 2 25m secs 111. 1. 1. 1 133. 3. 3. 3 165m secs 111. 1. 1. 1 122. 2. 2. 2 25m secs 111. 1. 1. 1 133. 3. 3. 3 165m secs

X

36 30

Basi c Ar chi t ect ur e

111. 1. 1. 1 122. 2. 2. 2 133. 3. 3. 3 144. 4. 4. 4

37 31

M oni t

• r

i ng

111. 1. 1. 1 122. 2. 2. 2 133. 3. 3. 3 144. 4. 4. 4

38 32

M oni t

• r

i ng

111. 1. 1. 1 122. 2. 2. 2 133. 3. 3. 3 144. 4. 4. 4

39 33

Rew i r i ng

111. 1. 1. 1 122. 2. 2. 2 133. 3. 3. 3 144. 4. 4. 4

40 34

New com er s

111. 1. 1. 1 122. 2. 2. 2 133. 3. 3. 3 144. 4. 4. 4 99. 9. 9. 9

41 35

Node Dr

• p/

Fai l ur e

111. 1. 1. 1 122. 2. 2. 2 133. 3. 3. 3 144. 4. 4. 4

133. 3. 3. 3 DO W N

42

Objectives

36

Per f

• r

m ance Eval uat i

• n:

Exper i m ent al Set t i ng

Nodes:

• 50 PlanetLab nodes for 2

months. Wiring policies:

• EGOIST
• k-Random, k-Closest, k-

Regular (DHT).

• Wiring frequency: 60

seconds. Metrics of interest:

• Delay (ping, Pyxida).
• CPU load (loadavg).
• Available Bandwidth

(pathChirp). Control variables:

• We vary the number (k) of

30 11 7 1 1

43 37

Act i ve M easur em ent s

EG O I ST

delay/EGOIST delay

44 38

Passi ve M easur em ent s

Wiring delay/EGOIST delay

EG O I ST EG O I ST

45 39

Syst em Load

111. 1. 1. 1

10% ut i l i zat i

• n

10% ut i l i zat i

• n

10% ut i l i zat i

• n

46 40

Syst em Load

EG O I ST

delay/EGOIST delay

47 41

Avai l abl e Bandw i dt h

111. 1. 1. 1

3M bps 1M bps 3M bps 2M bps pat hChi r p

48 42

Avai l abl e Bandw i dt h

EG O I ST

bwth/EGOIST bwth

49 43

Re- w i r i ng Fr equency

EGOIST wiring Approximate EGOIST wiring (e= 10%)

CPU, memory and bandwidth consumption is minimal.

EGOIST delay/optimal delay EGOIST re-wirings

• Appr. EGOIST/optimal delay
• Appr. EGOIST re-wirings

Normalized delay

50 44

Per f

• r

m ance Under Chur n: Hybr i d- EG O I ST

111. 1. 1. 1 122. 2. 2. 2 133. 3. 3. 3 144. 4. 4. 4

51 45

Per f

• r

m ance Under Chur n

Efficiency Index

Co nne c t i v i t y q ua l i t y

EG O I ST

K-Random K-Regular K-Closest Hybrid-EGOIST

Connect ed i n O ( T/ n)

52 46

Per f

• r

m ance Under Chur n

Efficiency Index

Co nne c t i v i t y q ua l i t y

EG O I ST

K-Random K-Regular K-Closest Hybrid-EGOIST

53 47

Appl i cat i

• ns
• Multi-path file transfer
• Real-time VoIP
• Online multiplayer P2P games

[Quake III traces from Donnybrook, SIGCOMM’08 EGOIST k-Closest k-Random k-Regular

54

Sel f i sh O ver l ay Net w or k Cr eat i

• n

I m pl i cat i

• ns

t

• Pr
• t
• col

Desi gn EG O I ST

Appl i cat i

• n

t

• sw ar

m i ng syst em s

St r at egi c Resour ce Al l

• cat

i

• n

48

55

a c c e s s I SP a c c e s s I SP t r a ns i t I SP

P2P Fi l e Shar i ng Syst em s

Par al l el Upl

• ad/

Dow nl

• ad
• Swarming

Local Schedul i ng

• Local Rarest First

Peer Sel ect i

• n
• Choke/Unchoke

Random Graphs

i nt e r ne t

$ $$ $$ t r a ns i t I SP a c c e s s I SP $$

• verlay node

49

56

A Cl

• ser

St udy

i nt e r ne t

Fl

• w

Net w or ks Analysis of 1-way broadcast [Massoulie et al., Infocom’07] Max-Flow abst r act s the behavior of Swarming

50

57

Li m i t at i

• ns

i nt e r ne t

Per f

• r

m ance i s t i ed t

• t

he t

• pol
• gy

The topology is not

• ptimized for Swarming!

Case study: Multiple Files

51

58

n- w ay Br

• adcast

: Focusi ng

• n

Di st r i but ed Dat a Cent er s

i nt e r ne t

Synchr

• ni

zat i

• n
• Distributed Databases
• Backups

Bat ch Par al l el Pr

• cessi

ng

• Distributed Anomaly

Detection

• Cloud Computing

52

59

Pr el i m i nar y Sol ut i

• ns

n co- exi st i ng sw ar m s (-) stress of physical links (-) exchange of multiple chunks in parallel overpartitions the uplink capacity [Tian et al., ICPP’06] End- Syst em m ul t i cast ( m esh) [ Spl i t St r eam , Bul l et ] (-) Creates an overlay for each swarm (-) No coordination among swarms (-) Monitor overhead

53

60

O ur Appr

• ach

Cr eat i

• n
• f

Net w or ks f

• r

Sw ar m i ng! Com m on O ver l ay

• Joint optimization of the entire overlay
• Amortization of monitor cost and available resources

Bounded degr ee Bandw i dt h- Cent r i c/ Dat a- Agnost i c

• Improvement of the end-to-end performance
• local scheduling

Di st r i but ed For m at i

• n

54

61

O pt i m i zed G r aphs f

• r

Sw ar m i ng

Swarming is too complicated to be described with an analytic function Max Flow  abst r act s the behavior of swarming Cr eat i

• n of Optimized Graphs based on bandwidth from Max

Flow Per f

• r

m ance of swarming over optimized graphs with simulation and PlanetLab

55

62

Reduci ng t he Aver age Dow nl

• ad

Ti m e

O bj ect i ve: M i ni m i ze the aver age download time Max-Sum: Wiring strategy of node vi: ma x ( s um ( M a x Fl

• w(

v

i,

v

j)

) , f

• r

a l l v

j

56

63

Reduci ng t he Dow nl

• ad

Ti m e

O bj ect i ve: M i ni m i ze the wor st download time Max-Min: Wiring strategy of node vi: m ax ( m i n ( M axFl

• w (

vi, vj) ) , f

• r

al l vj

57

64

Feasi bi l i t y

Both Max-Sum and Max-Min are NP-hard Max-Min: Choose k  Reduction to the SET-COVER

b

2

b

3

v

i

v

j

b

1

b

1

>> b

2

>> b

3

58

65

Local Sear ch

b

1

b

2

b

3

b

1

>> b

2

>> b

3

v

i

v

j

Wiring {si}, for the residual wiring S-i

60

66

Per f

• r

m ance Eval uat i

• n

File ID Node ID Delivery Time

Naive Max-Sum Max-Min

File ID File ID

• Flattens Distribution Time!
• Guarantees

Synchronization!

• comparable average

download time

61

67

I m pact

• f

Sel f i sh Behavi

• r

Upload-Selfishness

• Selfish-FIFO
• Most Replicated First:
• protect the uplink capacity
• Selfish Fast nodes:
• no improvement of upload time
• Selfish Slow nodes:
• significant improvement of upload time
• significant improvement of download time in all

nodes

62

68

Concl usi

• ns

W hat i s t he per f

• r

m ance gai n t hat can be achi eved by a sel f i sh node?

 Selfish nodes can reap substantial performance gain.

W hat i s t he i m pact

• f

sel f i sh nei ghbor sel ect i

• n

t

• ver

l ay net w or k per f

• r

m ance?  Surprisingly, the evolving graphs have also good

performance!

63

69

Concl usi

• ns

W hat ar e t he i m pl i cat i

• ns
• f

sel f i sh nei ghbor sel ect i

• n

t

• syst

em desi gn?



Selfish wiring strategies are easily realizable



Selfish wiring must be a component of any system to protect it from abuse



Selfish wiring behavior can be used for efficient dynamic service provisioning

64

70 65

Thank you.

ht t p: / / csr . bu. edu/ sns

71

Back- up sl i des

72

Real-Time Applications

Min-Max Best Response

Worst delay in the overlay:

k 0 2 3 5 11 22

73

SNS with Variable Degree

Real-time applications Variable degree through LS:

• Swap 1 link
• Add 1 link
• Drop 1 link

Ap p l i c a t i

• n

r e q ui r e me nt ( Pe r f

• r

ma nc e wh e n k =5 , n=5 i . e . 2 5 l i nk s ) 1 l i nk s 1 2 l i nk s

74 24

Performance Under Cheating

Delay/ Delay with abuse

t r ut hf ul EG O I ST t r ut hf ul EG O I ST

Untruthful Truthful Untruthful Truthful

Many Untruthful nodes Single Untruthful node

75

Ac c e s s I SP Ac c e s s I SP Tr a ns i t I SP

Modern File Sharing Systems

Parallel upload/ download

• Swarming

Local scheduling

• Local Rarest First

Flat connectivity

• Choke/unchoke

I nt e r ne t

Tr a ns i t I SP Ac c e s s I SP

Overlay node Seeder Leecher

76

n-way Broadcast

I nt e r ne t

Synchronization

• Distributed databases
• Backups

Batch parallel processing

• The files have to be received by

all nodes before the next step

• f processing begins

77

Preliminary Solutions

n co-existing swarms (-) Stress of physical links (-) Exchange of multiple chunks in parallel overpartitions

the uplink capacity [Tian et al., ICPP’06]

End-system multicast (mesh) [SplitStream, Bullet] (-) Creates an overlay for each swarm (-) No coordination among swarms

(-) Monitor overhead

78

Design Strategies for n-way Broadcast

Joint optimization of upload/download while participating in many swarms Data Agnostic

• Keeps swarming and local scheduling

Bandwidth-Centric

• Max-flow to approximate swarming behavior

[Massoulie et al., Infocom’07]

Bounded Degree

79

Reducing the Average Download Time

Objective: M i ni m i ze the aver age download time

Max-Sum:

Neighbor selection strategy of node vi: max (sum (MaxFlow(vi, vj)), for all vj

80

Reducing the Download Time

Objective: M i ni m i ze the t

• t

al download time

Max-Min:

Neighbor selection strategy of node vi: max (min (MaxFlow(vi, vj)), for all vj

81

Optimized Graphs and Swarming

Formation of stable graphs Each node strives to improve both the upload and download flow Performance of swarming on optimized graphs

• Max flow might not

be realizable

82

Performance Evaluation

File ID Node ID Delivery Time

Naive Max-Sum Max-Min

File ID File ID

Flattens distribution time! Guarantees synchronization! Comparable average download time

Sel f i sh Upl

• ad:

Pr

• t

e c t s t h e up l i nk c a p a c i t y

• f

t h e s l

• w

no d e  I mp r

• v

e s t h e d

• wnl
• a

d t i me i n t h e s y s t e m

83

Faci l i t y Locat i

• n

Uncapaci t at ed Faci l i t y Locat i

• n

( UFL) : Find a subset I of F and a function σ:CI to min ( Σi fi + Σi,j sjcij )

F: set of facilities fi: cost to

• pen

facility C: set of clients, cij: cost connecting client jfacility I sj: demand of node j

84

Faci l i t y Locat i

• n

f

i

dj

i

sj

13