e. altman, c. touati, r. el-azouzi inria, univ avignon networking games ens january 2010

75
E. Altman, C. Touati, R. El- Azouzi INRIA, Univ Avignon Networking Games ENS ENS January 2010

Upload: shannon-hankinson

Post on 02-Apr-2015

217 views

Category:

Documents


3 download

TRANSCRIPT

Page 1: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon

Networking Games

ENSENSJanuary 2010

Page 2: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

A Trip to Matrix Gameland

Chapter 1

Page 3: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Overview of Chap 1

1. TCP, competition between protocols: motivation for the game theoretic definition of equilibrium

2. Matrix Games and Nash Equilibrium, properties

3. Correlated equilibrium

4. Multi-access 2x2 matrix game

5. Coordinating games

6. Zero sum games, maple

Page 4: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Background: Early TCP

TCP – Transport Control Protocol, used for reliable data transfer and for flow control

Packets have serial numbers. The destination acknowledges received packets

A non-acknowledged packet is retransmitted Initially, data transfers over the Internet used flow

control with fixed window size K: transmission possible only when the number of packets not acknowledged is smaller than K

Problem: NETWORK COLLAPSE

Page 5: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Background: Modern TCP (Van Jacobson)

Adaptive window: keeps increasing linearly in time

When K acks arrive we transmit the window to K+1 and K+1 packets are sent

When a loss occurs we halve K

AIMD: Additive Increase Multiplicative Decrease

Other details are not modelled

Page 6: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Background: Contemporary TCP

Aggressive versions have been proposed to adapt faster

Scalable (Tom Kelly): when K acks are received we multiply K by a constant. MIMD – Multiplicative Increase Multiplicative Decrease

HSTCP (Sally Floyd) like AIMD but with increase and decrease parameters that increase with K

TCP versions are mostly open source (IETF standards) but also patents.

Page 7: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

EVOLUTION OF TCP

First version aggressive Second version (Tahoe) the most gentle,

disappeared Third version Reno and its refinement are the

mostly deployed versions Vegas version was shown to perform better but

was not much deployed due to vulnerability against Vegas. Used in satellite communications

New aggressive versions appear (for grid and storage networks): Scalable, HSTCP…

Page 8: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

How will future Internet look like?

Researchers have tried to determine which version of TCP will dominate

We can pose a more abstract question: will the Internet move towards an aggressive behavior of TCP, a friendly behavior? Or coexistence?

If coexistence, what proportions?

If there is a convergence to one of the above, we call this an Equilibrium.

Page 9: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Definitions of Equilibrium

A1(u) The isolation test: See how well the protocol performs if everyone uses the friendly protocol only. Then imagine the worlds with the aggressive TCP only. Compare the two worlds. The version u for which users are happier is the candidate for the future Internet.

A2(u) The Confrontation test Consider interactions between aggressive and piecefull sessions that share a common congested link. The future Internet is declared to belong to the transport protocol of version u if u performs better in the interaction with v.

A3. Game Theoretic Approach: We shall combine the approaches. If a version u does better than v under both then it will dominate the future Internet. It is called "dominating strategy". Otherwize both versions will co-exist. The fraction of each type will be such that the average performance of a protocol is the same under both u and v

Page 10: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Competition between MIMD

Page 11: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Competition between MIMD

Symmetric MIMD with synch losses: ratio of throughputs remains as the initial ratio since the rate of increase and decrease are the same UNFAIR!UNFAIR!

Asymmetric MIMD with synch losses: connection with lowest RTT gets all the bandwidth VERY UNFAIR!VERY UNFAIR!

Non synchronized losses, Asym: connection with lowest RTT gets all bandwidth VERY UNFAIR!VERY UNFAIR!

Sym: Sym: null recurrent MC UNFAIR!UNFAIR!

[EA, KA, B. Prabhu 2005]

Page 12: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Intra and Inter-version competition

MIMD-AIMD competition, [EA, KA, BP 2005] there is a threshold on the BDP below which AIMD has better throughput

AIMD-AIMD competitionAIMD-AIMD competition: “fair” sharing.

Page 13: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Vegas – Reno interaction

Reno is more aggressive than Vegas. Does better in the confrontation test but worse in the isolation

“The last issue, which was not addressed in this paper, concerns the deploying of TCP Vegas in the Internet. It may be argued that due to its conservative strategy, a TCP Vegas user will be severely disadvantaged compared to TCP Reno users, …. it is likely that TCP Vegas, which improves both the individual utility of the users and the global utility of the network, will gradually replace TCP” (Bonald)

Page 14: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Summary in a Symmetric matrix game

Matrix Game:Sc-NR:NR better in Isolation, Vegas- Better in confr.NR-Sc:Sc better in Isolation, NR- In confr.

Page 15: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 16: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Mixed Strategies

Assume that neither actions is dominant in the TPC game. The game approach predicts that both versions will coexist. The fraction of each is computed so that the average performance of a protocol is the same under both actions

We call this the Indifference Principle

Let this fraction of actions be p and 1-p. Take a=0

Page 17: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Applying the indifference principle

Equating these gives:

Page 18: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Nash Equilibrium

In both the cases of dominating strategy as well as the case of mixed strategy, we predicted the use of BEST RESPONSES at equilibrium – At equilibrium, each player uses an action that is best for him for the given actions of the others

Equilibrium is formally defined through this property.

Page 19: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Pure equilibrium

Page 20: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Equilibrium in Mixed strategies

Page 21: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Characterizing equilibrium

Page 22: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Computing mixed equilibrium

Page 23: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Multiple Access Game

Two mobiles, Collision Channel Only possibility for successful transmission:

only one transmitms Equilibrium:

Always transmit.

Page 24: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Pricing

Adding price E per transmission. 0<E<1. Mixed equilib

r=1-E Thp at equilib:

E(1-E)

maximized at E=1/2 Yields Thp=1/4 Utility at equilib: 0

Page 25: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Capture

Often collision does

not result in loss

Packet Err Prob:

Packet Loss Prob:

Page 26: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Coordination

A non symmetric equilibrium can achieve a global throughput of 1

Is this a correlated equilibrium?

Page 27: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Coordination Games

Page 28: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 29: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 30: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 31: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 32: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 33: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 34: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 35: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Coordination Games

Page 36: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 37: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Zero-sum Game

Page 38: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Lower Value

Page 39: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Upper Value

Page 40: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Saddle Point

Page 41: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Best response to q

Page 42: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Linear Program

Page 43: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Solution in Maple

Page 44: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Equivalent Games

Page 45: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Example: transformation into an optimization problem

Page 46: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Example: Transforming into a zero-sum game

Page 47: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

A vous de jouer!

Page 48: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 49: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Concave Games and Constraintns

Chapter 2

Page 50: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Overview of Chap 2

1. Constrained Game

2. Concave Games

Page 51: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 52: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 53: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 54: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 55: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010

Exampl of general constraints

Page 56: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 57: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 58: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 59: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 60: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 61: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 62: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 63: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 64: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 65: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 66: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 67: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 68: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 69: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 70: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 71: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 72: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 73: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 74: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010
Page 75: E. Altman, C. Touati, R. El-Azouzi INRIA, Univ Avignon Networking Games ENS January 2010