“ analysis of the increase and decrease algorithms for congestion avoidance in computer networks...
TRANSCRIPT
“Analysis of the Increase and Decrease Algorithms for Congestion Avoidance in Computer Networks”
by Dah-Ming Chiu and Raj Jain, DEC
Computer Networks and ISDN Systems
17 (1989), pp. 1-14
Prof. Xi Zhang
Motivation (1)
Internet is heterogeneous Different bandwidth of links Different load from users
Congestion control Help improve performance after
congestion has occurred
Congestion avoidance Keep the network operating off the
congestion
Prof. Xi Zhang
Power of a Network The power of the network describes this
relationship of throughput and delay: Power = Goodput/Delay This is based on M/M/1 queues ( 1 server and
a Markov distribution of packet arrival and service).
This assumes infinite queues, but real networks the have finite buffers and occasionally drop packets.
The objective is to maximize this ration, which is a function of the load on the network.
Ideally the resource mechanism operates at the peak of this curve.
Prof. Xi Zhang
Motivation (2)
Fig. 1. Network performance as a function of the load.
Power = {Goodput}/{Response Time}
Prof. Xi Zhang
Relate Works
Centralized algorithm Information flows to the resource managers and
the decision of how to allocate the resource is made at the resource [Sanders86]
Decentralized algorithms Decisions are made by users while the
resources feed information regarding current resource usage [Jaffe81, Gafni82, Mosely84] Binary feedback signal and linear control Synchronized model What are all the possible solutions that converge
to efficient and fair states
Prof. Xi Zhang
Linear Control (1)
4 examples of linear control functions Multiplicative Increase/Multiplicative
Decrease Additive Increase/Additive Decrease Additive Increase/Multiplicative Decrease Additive Increase/ Additive Decrease
Decreasetyiftxba
Increasetyiftxbatx
iDD
iIIi 1)( )(
,0)( )()1(
Prof. Xi Zhang
Linear Control (2) Multiplicative Increase/Multiplicative Decrease
Additive Increase/Additive Decrease
Additive Increase/Multiplicative Decrease
Multiplicative Increase/ Additive Decrease
Decreasetyiftxb
Increasetyiftxbtx
iD
iIi 1)( )(
,0)( )()1(
Decreasetyiftxa
Increasetyiftxatx
iD
iIi 1)( )(
,0)( )()1(
Decreasetyiftxb
Increasetyiftxatx
iD
iIi 1)( )(
,0)( )()1(
Decreasetyiftxa
Increasetyiftxbtx
iD
iIi 1)( )(
,0)( )()1(
Prof. Xi Zhang
Criteria for Selecting Controls Efficiency
Closeness of the total load on the resource to the knee point
Fairness Users have the equal share of bandwidth
Distributedness Knowledge of the state of the system
Convergence The speed with which the system approaches
the goal state from any starting state
)(
)(2
2
i
i
xn
xFairness
Prof. Xi Zhang
Responsiveness and Smoothness of Binary Feedback System
Equlibrium with oscillates around the optimal state
Prof. Xi Zhang
Convergence to Efficiency
Negative feedback
So If y=0: If y=1: Or
).()1(1)(
),()1(0)(
txtxty
txtxty
ii
ii
).( 0)()1(
),( 0)()1(
txandntxbna
txandntxbna
iiDD
iiII
)(1
,)(
1
tx
nab
tx
nab
i
DD
i
II
Prof. Xi Zhang
Convergence to Fairness (2)
c>0 implies:
Furthermore, combined with (3) we have:
(9) 0 0
(8) 0 0
D
D
I
I
D
D
I
I
b
aand
b
a
or
b
aand
b
a
)10( 10 ,0
,0 ,0
DD
II
ba
ba
Prof. Xi Zhang
Distributedness Having no knowledge other than the feedback y(t) Each user tries to satisfy the negative feedback
condition by itself
Implies (10) to be
)11( . )()1(1)(
, )()1(0)(
itxtxty
itxtxty
ii
ii
)12( 10 ,0
,1 ,0
DD
II
ba
ba
.0)( 0)()1(
,0)( 0)()1(
txtxba
txtxba
iiDD
iiII
Prof. Xi Zhang
Important Results
Proposition 1: In order to satisfy the requirements of distributed convergence to efficiency and fairness without truncation, the linear increase policy should always have an additive component, and optionally it may have a multiplicative component with the coefficient no less than one.
Proposition 2: For the linear controls with truncation, the increase and decrease policies can each have both additive and multiplicative components, satisfying the constrains in Equations (16)
Prof. Xi Zhang
Optimizing the Control Schemes
Optimal convergence to Efficiency Tradeoff of time to convergent to
efficiency te, with the oscillation size, se.
Optimal convergence to Fairness
Prof. Xi Zhang
Optimal convergence to Efficiency
Given initial state X(0), the time to reach Xgoal is:
Prof. Xi Zhang
Optimal convergence to Fairness
Equation (7) shows faireness function is monotonically increasing function of c=a/b.
So larger values of a and smaller values b give quicker convergence to fairness.
In strict linear control, aD=0 => fairness remains the same at every decrease step
For increase, smaller bI results in quicker convergence to fairness => bI =1 to get the quickest convergence to fairness
Proposition 3: For both feasibility and optimal convergence to fairness, the increase policy should be additive and the decrease policy should be multiplicative.
Prof. Xi Zhang
Practical Considerations
Non-linear controls Delay feedback Utility of increased bits of feedback Guess the current number of users n Impact of asynchronous operation
Prof. Xi Zhang
Conclusion
We examined the user increase/decrease policies under the constrain of binary signal feedback
We formulated a set of conditions that any increase/decrease policy should satisfy to ensure convergence to efficiency and fair state in a distributed manner We show the decrease must be multiplicative to
ensure that at every step the fairness either increases or stays the same
We explain the conditions using a vector representation
We show that additive increase with multiplicative decrease is the optimal policy for convergence to fairness