Title: TCP Connection Game: A Study on the Selfish Behavior of TCP Users
1TCP Connection GameA Study on the Selfish
Behavior of TCP Users
Honggang Zhang, Don Towsley, Weibo Gong Univ. of
Massachusetts Amherst
- Presented by
- Honggang Zhang
2Outline
- General TCP connection game formulation
- System optimization problem
- Different games capturing different user
behaviors - Conclusions and future work
3TCP Connection Game
bottleneck link
- Assumptions
- all connections experience the same loss rate
- aggregate goodput bottleneck link capacity
4TCP Connection Game
bottleneck link
- m users compete for congested bottleneck link
- User i strategy number of connections (ni)
- User i adjusts number of connections to search
for maximum goodput or other utilities
5Problems to be studied
- Existence, uniqueness of Nash equilibrium
- Nash equilibrium
- a vector n(n1, n2, , nm) such that no
user can benefit by deviating from it
6What is cost ?
bottleneck link
7System Optimization
minimize aggregate offered rate from all
users while keeping bottleneck link fully utilized
8Results of System Optimization
- Loss rate p is increasing function of number of
connections
Example. Assume all users have same RTT.
- System optimal cost uniquely achieved when each
user opens one connection
9TCP Connection Game
- Different games to capture different user
behaviors - Two most important games
- Game 1 aggressive users
- Game 2 resource constrained
- Strategy space of user i
10Game 1 Aggressive Users
- Utility of user i goodput
Number of connections of user i
Per connection goodout of user i
Round Trip Time of user i
Bottleneck link capacity
- Nash equilibrium nNE(n1max, , nmmax)
- Reason users not resource constrained
11Game 1 Aggressive Users
- Efficiency loss unbounded as nimax increase
Example. Total goodput of user i given other
users having 5 connections.
As nimax increases, loss rate approaches 1
congestion collapse !
12Game 2 Resource Constrained Users
13Game 2 Resource Constrained Users
- Theorem there exists a unique Nash equilibrium
(NE) in m (mgt2) user TCP connection game.
Two-player game example
- Illustrate Nash equilibrium using best response
curves - Best Response of player i
- The optimal number of connections given the
number of connections of all other players.
2
1
3
14Game 2 Resource Constrained Users
System cost of Nash equilibrium
Loss of Efficiency
System cost of optimal point
- Loss of efficiency is bounded
- The upper bound is function of network
parameters and
NO congestion collapse!
15How users play TCP connection game?
- Best-response dynamics or algorithm
- Users take moves alternatively, each move is the
best response to the number of connections of all
other players
4
2
Number of connections of player 2
3
1
Number of connections of player 1
16Stability of Nash equilibrium in Game 2
- Nash equilibrium (NE) of two-player game locally
stable for best response dynamics - Within some neighborhood of NE, if there is
deviation from Nash equilibrium and each player
takes best response alternatively, then this
interaction process converges back to Nash
equilibrium - Result follows from Banach contraction mapping
theorem
17Local Stability
Player 1 deviates from Nash equilibrium
Nash equilibrium
1
4
Number of connections of player 2
3
2
Local Stability Region
Number of connections of player 1
18Conjecture Global Stability
Nash equilibrium
4
2
Number of connections of player 2
3
1
Number of connections of player 1
19Conclusions
- Nash equilibrium exists in all variants of TCP
connection games - Efficiency loss bounded if users resource limited
- More powerful users have higher utility at Nash
equilibrium - Nash equilibrium locally stable for two-player
version of game 2 - Selfish behavior of TCP users will NOT lead to
congestion collapse in TCP connection games
20Thank you!Questions ?