1 atomic routing games on maximum congestion costas busch department of computer science louisiana...

50
1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU Malik Magdon-Ismail, RPI

Upload: zechariah-grissom

Post on 31-Mar-2015

218 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

1

Atomic Routing Games on Maximum Congestion

Costas BuschDepartment of Computer Science

Louisiana State University

Collaborators: Rajgopal Kannan, LSUMalik Magdon-Ismail, RPI

Page 2: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

2

Introduction

Price of Stability

Price of Anarchy

Outline of Talk

Bicriteria Game

Page 3: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

3

Network Routing

Each player corresponds to a pair of source-destination

Objective is to select paths with small cost

Page 4: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

4

Main objective of each player is to minimize congestion: minimize maximum utilized edge

3 congestion C

iplayer

Page 5: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

5

A player may selfishly choose an alternativepath that minimizes congestion

CC 31 congestion

Congestion Games:

Page 6: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

6

Player cost function for routing :i

ii Cppc )(

p

Congestionof selected path

Social cost function for routing :

CpSC )(

p

Largest player cost

Page 7: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

We are interested in Nash Equilibriumswhere every player is locally optimal

Metrics of equilibrium quality:

p

Price of Stability

)(

)(min

*pSC

pSCp

Price of Anarchy

)(

)(max

*pSC

pSCp

*p is optimal coordinated routing with smallest social cost

Page 8: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

8

Results:

• Price of Stability is 1

• Price of Anarchy is )log( nLO

Maximum allowed path length

Page 9: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

9

Introduction

Price of Stability

Price of Anarchy

Outline of Talk

Bicriteria Game

Page 10: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

10

We show:

• QoR games have Nash Equilibriums

(we define a potential function)

• The price of stability is 1

Page 11: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

11

],,,,,[)( 21 Nk mmmmpM

number of players with cost km k

Routing Vector

Page 12: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

12

Routing Vectors are ordered lexicographically

],,,[)( 21 NmmmpM

],,,[)( 21 NmmmpM

= = = =

],,,,,[)( 11 Nkk mmmmpM

],,,,,[)( 11 Nkk mmmmpM

< < = =

)()( pMpM

)()( pMpM )( pp

)( pp

Page 13: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

13

If player performs a greedy movetransforming routing to then:p p pp

i

Lemma:

Proof Idea:

Show that the greedy move gives a lower order routing vector

Page 14: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

14

kk

ii Cppck )(

ii Cppck )(

Player Costi

Before greedy move:

After greedy move:

Since player cost decreases:

Page 15: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

15

],,,,,,,[)( 11 Nkkk mmmmmpM

Before greedy move player was counted herei

],,,,,,,[)( 11 Nkkk mmmmmpM

After greedy moveplayer is counted herei

Page 16: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

16

],,,,,,,[)( 11 rkkk mmmmmpM

],,,,,,,[)( 11 rkkk mmmmmpM

> ==No change

Definite Decrease

possibledecrease

possibleincreaseor decrease

Possible increase

>

END OF PROOF IDEA

Page 17: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

17

Existence of Nash Equilibriums

Greedy moves give lower order routings

Eventually a local minimum for every playeris reached which is a Nash Equilibrium

Page 18: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

18

minp

Price of Stability

Lowest order routing :

*min )( SCpSC

• Is a Nash Equilibrium

• Achieves optimal social cost

1)(

Stability of Price*min

SC

pSC

Page 19: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

19

Introduction

Price of Stability

Price of Anarchy

Outline of Talk

Bicriteria Game

Page 20: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

20

We show for any restricted QoR game:

Price of Anarchy = )log( nLO

Page 21: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

21

Path of player

Consider an arbitrary Nash Equilibriump

i

iCedgemaximum congestionin path

Page 22: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

22

must have an edge with congestion

Optimal path of player

In optimal routing :*p

i

iC

1 iCC

)(111*)( ppcCCCppc iiii

**)( CpSC

Since otherwise:

Page 23: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

23

C

00

0

edges use that Paths:

Congestion of Edges :

E

CE

In Nash Equilibrium social cost is: CSC

0 0

Page 24: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

24

C 1C1C

0 0

Edges in optimal paths of 0

Page 25: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

25

C 1C1C

0 01 1

11

1

edges use that Players:

1 least at Congestion of Edges :

E

CE

Page 26: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

26

C 1C1C 2C 2C2C2C

0 01 1

Edges in optimal paths of 1

Page 27: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

27

C *SC *SC 2C 2C2C

0 01 1

2C

2 2

22

2

edges use that Players:

2 least at Congestion of Edges :

E

CE

Page 28: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

28

In a similar way we can define:

jj

j

E

jCE

edges use that Players:

least at Congestion of Edges :

Page 29: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

29

,,,,

,,,,

3210

3210

EEEEWe obtain sequences:

There exist subsequence:

110

110

,,,

,,,,

s

ss EEEE

||2|| 1 jj EEWhere: ||2|| 1 ss EEand1sj

ns log

Page 30: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

30

||))1((|| 11 ss EsCL

|||| 1*

s

s

EC

Maximum edge utilization

Minimum edge utilization

LMaximum path length

)log(* nLOCC

ns log ||2|| 1 ss EE

Known relations

Page 31: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

31

Worst Case Scenario:

Page 32: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

32

Introduction

Price of Stability

Price of Anarchy

Outline of Talk

Bicriteria Game

Page 33: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

33

We consider Quality of Routing (QoR) congestion games where the pathsare partitioned into routing classes:

QQQ ,,, 21

)()()( 21 QSQSQS

With service costs:

Only paths in same routing class can causecongestion to each other

Page 34: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

34

An example:

•We can have routing classes)(lognO

•Each routing class contains paths with length in range

jQ

]2,2( 1jj

12)( jjQS•Service cost:

•Each routing class uses a different wireless frequency channel

Page 35: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

35

Player cost function for routing :i

iii SCppc )(

p

Congestionof selected path

Cost of respectiverouting class

Page 36: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

36

Social cost function for routing :

SCpSC )(

p

Largest player cost

Page 37: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

37

Results:

• Price of Stability is 1

• Price of Anarchy is

)log),(min( ** nSCO

Page 38: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

38

We consider restricted QoR games

For any path :p )(|| pSp

Path length Service Cost of path

Page 39: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

39

We show for any restricted QoR game:

Price of Anarchy = )log),(min( ** nSCO

Page 40: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

40

Path of player

Consider an arbitrary Nash Equilibriump

i

iCedgemaximum congestionin path

Page 41: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

41

must have an edge with congestion

Optimal path of player

In optimal routing :*p

i

iC

*SCC i

)(111 *** ppcSCCSSCSCcp iiiiiiii

***)( SCpSC

Since otherwise:

Page 42: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

42

C

00

0

edges use that Paths:

Congestion of Edges :

E

CE

In Nash Equilibrium: SCSC

0 0

Page 43: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

43

C *SC *SC

0 0

Edges in optimal paths of 0

Page 44: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

44

C *SC *SC

0 01 1

11

*1

edges use that Players:

least at Congestion of Edges :

E

SCE

Page 45: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

45

C *SC *SC *2SC *2SC *2SC *2SC

0 01 1

Edges in optimal paths of 1

Page 46: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

46

C *SC *SC *2SC *2SC *2SC

0 01 1

*2SC

2 2

22

*2

edges use that Players:

2least at Congestion of Edges :

E

SCE

Page 47: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

47

In a similar way we can define:

jj

j

E

jSCE

edges use that Players:

least at Congestion of Edges : *

Page 48: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

48

,,,,

,,,,

3210

3210

EEEEWe obtain sequences:

There exist subsequence:

110

110

,,,

,,,,

s

ss EEEE

||2|| 1 jj EEWhere: ||2|| 1 ss EEand1sj

ns log

Page 49: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

49

||))1((|| 1*

1 ss ESsCL

||

|| 1*

s

s

EC

Maximum edge utilization

Minimum edge utilization

*SLMaximum path length

)log( **

nSOC

C

ns log ||2|| 1 ss EE

Known relations

Page 50: 1 Atomic Routing Games on Maximum Congestion Costas Busch Department of Computer Science Louisiana State University Collaborators: Rajgopal Kannan, LSU

50

)log( **

nSOC

C

)log),(min( Anarchy of Price ****

nSCOSC

SC

We have:

By considering class service costs, we obtain: