Overlay Network Creation and Maintenance with Selfish Users Georgios Smaragdakis Dissertation committee members: Azer Bestavros, Nikolaos Laoutaris, John.

Overlay Network Creation and Maintenance

with Selfish Users

Georgios Smaragdakis

Dissertation committee members: Azer Bestavros,

Nikolaos Laoutaris, John Byers

2

Overlay applications:overlay routing,p2p file sharing,content distribution..

Access ISP

Access ISP

Transit ISP

Overlays & Neighbor Selection

Internet Overlay links

Transit ISP

Access ISP

Overlay node

Focus on service quality!

3

Challenges

v1v2

v3

v4

v5 v6

v7

v8

v9

p1=[v2v3v4v5v6v7v8v9]




Selfish node

What is the performance gain that can be achieved by a selfish node?

What is the impact of selfish neighbor selection to overlay network performance?

What are the implications of selfish neighbor selection to system design?

4

Selfish Neighbor Selection

Implications

to

Overlay Routing

Implications to File SharingImplications to

Service Provisioning

Outline

5


Implications

to

Overlay Routing



6

Selfish Neighbor Selection (SNS)

Constraints that need to be addressed in a realistic model for overlay networks:

Bounded degree Preference vectors Realistic network distance Link directionality

Fundamentally different from other models that have been proposed for other networks.

[Fabrikant et al.,PODC’03; Chun et al., Infocom’04 …]

7

Optimal Neighbor Selection

vi: choose k neighbors, s.t.

vi

G-i=( V-i , S-i )

u

w

ij Vv

jiSiji vvdpSC ),()(min

over all siSi

vi’s residual network

Set of residual nodes

Set of residual wiring

8

SNS & Facility Location

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

9

k-mediank-median: Find a subset I of F and a function σ:CI

to min ( Σi,j sjcij ) such that |I| ≤ k

F: set of

facilities

C: set of clients,

cij: cost connecting

client jfacility I

sj: demand of node j

10

Uncapacitated Facility LocationUncapacitated Facility Location (UFL): Find a subset I of F and a function σ:CI

to min ( Σi fi + Σi,j sjcij )F: set of facilities

fi: cost to

openfacility

C: set of clients,

cij: cost connecting

client jfacility I

sj: demand of node j

11

Non-uniform link weights, and uniform preference

ILP formulation

SNS & Facility Location

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

u

w

w,u can be obtained from k-median on

reversed distances

w

u

vi

ij Vv


Since the wiring cost is the same

12

Local Search (LS)

vi: choose k neighbors

viu

w

ij Vv


over all siSi

vi’s residual network

[Arya et al,STOC’01]

G-i=( V-i , S-i )

Set of residual nodes

Set of residual wiring

13

SNS : the GameGame <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

ij Vv

jiSiji vvdpSC ),()(

14

SNS : Equilibria

n=15 k=2

k=3

k=8

k=11

Uniform Preference Skewness of preference

k (Link density)

In-degrees are highly skewed even under uniform preference ! Quality-based

“preferential attachment”

15

Performance of ILP & LS is close to Utopian!

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

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

SNS : Efficiency

Link density Skewness of preference Link density

Skewness of preference

16

SNS : Trace-Driven Evaluation 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)

17

Connecting on a k-Random graph

k kk

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

18

Connecting on a k-Closest graph

kkk

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!

“Greed is not good”


19

Connecting on a k-Regular graph

k kk

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”


20

Connecting on a Best Response graph

The BR graph is highly optimized!

kkk

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


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

21

SNS vs. Heuristics: Social 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

22

Real-Time Applications

Min-Max Best Response

Worst delay in the overlay:

k

0 2 3 5 11 22

23

SNS with Variable Degree

Real-time applications Variable degree through LS:

Swap 1 link Add 1 link Drop 1 link

Application requirement

(Performance when k=5, n=50 i.e. 250 links)

100 links

120 links

24


Implications

to

Overlay Routing



25

Basic design of EGOIST:Link state protocolMeasurements of distance to candidate

neighborsWirings according to chosen strategy Re-wirings every T second

A newcomer bootstraps by connecting to arbitrary neighbors

26

EGOIST : Performance

BestResponse

27

EGOIST: Passive Measurements

Passive measurements based on virtual coordinates (pyxida system) with minimal cost

28

EGOIST: Other Metrics

End-to-end available bandwidth (pathchirp) with minimal measurement overhead CPU load (loadavg)

29

EGOIST: Marginal Utility of Rewiring

There exists a performance knee (k=3 or 4) Re-wirings could be reduced with lazy BR

BR Lazy BR (threshold = 10%)

30

EGOIST: Effect of Churn

Connectivity is guaranteed (in T/n time) HybridBR (a connected ring is maintained) delivers much of the efficiency of BR

Effi

ciency

Index

Connect

ivit

y

qualit

y

31

EGOIST: Effect of Churn

BR and Hybrid BR dominate all the other heuristics HybridBR pays off at high churn

Effi

ciency

Index

Connect

ivit

y

qualit

y

32

EGOIST : Other Work CPU and memory load is very low

Robust to cheating

Scalability via topological sampling via layered architecture

Applications including multi-player P2P games, real-time traffic over IP etc.

33


Implications

to

Overlay Routing



34

Access ISP

Access ISP

Transit ISP

Modern File Sharing Systems

Parallel upload/ download- Swarming

Local scheduling - Local Rarest First

Flat connectivity- Choke/unchoke

Internet

Transit ISP

Access ISP

Overlay node

Seeder

Leecher

35

n-way Broadcast

Internet Synchronization- Distributed databases - Backups

Batch parallel processing

- The files have to be received by all nodes before the next step

of processing begins

36

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

37

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

38

Reducing the Average Download Time

Objective: Minimize the average download time

Max-Sum: Neighbor selection strategy of node vi:

max (sum (MaxFlow(vi, vj)), for all vj

39

Reducing the Download Time

Objective: Minimize the total download time

Max-Min: Neighbor selection strategy of node vi:

max (min (MaxFlow(vi, vj)), for all vj

40

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

41

Performance Evaluation

File ID

Nod

e ID

Deliv

ery

Tim

e

Naive Max-Sum Max-Min

File ID File ID

Flattens distribution time! Guarantees synchronization! Comparable average

download time

Selfish Upload:

Protects the uplinkcapacity of the slow

node

Improves the download time in the

system

42

Other Work: File Searching

Best response: max #nodes reached

Bootstrap

Server 1

2

3

4

5

6

TTL of scoped flooding is 2

Maximum Coverage Problem

selfishly

43


Implications

to

Overlay Routing



44

Server Selection

Hardware server

45

Centralized Deployment

Generic Service Host

Software server

Demand changee.g. Flash crowd, time-of-dayeffect

46

Dynamic Service Deployment

Generic Service Host

Software server

Demand changee.g. Flash crowd, time-of-dayeffect

47

r-ball (r=2)

Distributed Service Migration (DSM)

Solve k-median or UFL in an r-ball ..BUT nodes outside the r-ball are totally neglected

“ring” nodes

Iterate until

convergence

48

DSM: Properties

Convergence: Migration only if the cost of facilitating the

demand decreases at least be a%, converges in O(log1+a n) steps We can control the speed of convergence by

tuning a

Limited horizon view requirement: r regulates the trade-off between scalability and

performance

49

Similar results for UFL under different cost functions to open and maintain the server

DSM: Evaluation

50

Dynamic vs. Static Deployment

Static deployment

Dynamic deployment

DSM

DSM

51

Conclusions What is the performance gain that

can be achieved by a selfish node?

Selfish nodes can reap substantial performance gain.

What is the impact of selfish neighbor selection to overlay network performance?

Surprisingly, the evolving graphs have also good performance!

52

Conclusions What are the implications of selfish

neighbor selection to system design?

Selfish wiring strategies are easily realizable

Selfish wiring behavior can be used towards distributed overlay network creation and maintenance

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

53

Thank You!

http://csr.bu.edu/sns

http://csr.bu.edu/dfl

Overlay Network Creation and Maintenance with Selfish Users Georgios Smaragdakis Dissertation committee members: Azer Bestavros, Nikolaos Laoutaris, John.

Documents

s i s i

optimal neighbor selection

game game v

facilities f i

selfish routing slide

j s j c ij f

residual wiring slide

demand of node j slide