Enhancing TCP Fairness in Ad Hoc

Wireless Networks Using Neighborhood

RED

Kaixin Xu, Mario Gerla

University of California, Los Angeles

{xkx, gerla}@cs.ucla.edu

Lantao Qi, Yantai Shu

Tianjin University, Tianjin, China

{ltqi, ytshu}@tju.edu.cn

Motivation

TCP is important in ad hoc network apps Reliable transfer of data/image files

and multimedia streaming Congestion protection Efficient utilization and fair share of

the resources However, TCP has shown unfair

behavior in ad hoc nets

TCP Unfairness in Ad Hoc Networks 3 TCP flows contending with each other Flow 2 is nearly starved Original RED fails to improve the fairness

F T P 1

F T P 2

F T P 3

(1 0 0 , 1 0 0 ) (6 0 0 , 1 0 0 )

(3 5 0 , 3 5 0 )

(3 5 0 , 7 0 0 )(0 , 7 0 0 )

(3 5 0 , 1 0 5 0 )

(1 0 0 , 1 3 0 0 ) (6 0 0 , 1 3 0 0 )

(7 0 0 , 7 0 0 )

Why RED Does Not Work?

Why RED does not work in ad hoc networks? Congestion simultaneously affects multiple

queues Queue at a single node cannot completely

reflect the state Extend RED to the entire congested area –

Neighborhood of the node

Random Early Detection (RED) to improve fairness Active queue management scheme qwavgwavg qq 1 Average queue size:

Drop probability: , proportional to buffer occupancy

thth

thp avg

bp minmax

)min(max

Neighborhood and Its Distributed Queue A node’s

neighborhood consists of the node itself and the nodes which can interfere with this node’s signal 1-hop neighbors

directly interfere 2-hop neighbors

may interfere

Queue size of the neighborhood reflects the degree of local network congestion

A

1

5

3

4

2

6

7

9

1 0

1 2

1 1

8

Simplified Neighborhood Queue Model 2-hop neighborhood queue

model is not easy to operate Too much overhead to

propagate queue values 2 hops away

Simplified model Only include 1-hop

neighbors Two Qs at each

neighbor: Outgoing queue “Incoming queue”=

# CTS packets OH by A Distributed neighborhood

queue – the aggregate of these local queues

A

1

5

3

4

2

6

O u tg o in g Q u eu e

In co m in g Q u eu e

Neighborhood Random Early Detection (NRED) Extending RED to the distributed

neighborhood queue (Virtual) Key Problems

Counting the size of the distributed neighborhood queue

Calculating proper packet drop probability at each node

Components of Neighborhood RED Neighborhood Congestion Detection (NCD) Neighborhood Congestion Notification (NCN) Distributed Neighborhood Packet Drop

(DNPD)

Direct way: Announce queue size upon changes Too much overhead, queue location

problem? Our method: Indirectly estimate an index of

instant queue size by monitoring wireless channel

Neighborhood Congestion Detection

Average queue size is calculated using RED’s alg.

Congestion: queue size exceeds the minimal threshold

ervalsampling

timebusychannelUbusy int

Channel utilization ratio

Queue size index K is a constant

busyUKq *

A

1

5

3

4

2

6

O u tg o in g Q u eu e

In co m in g Q u eu e

Channel busy

CTS

Neighborhood Congestion Notification & Distributed Neighborhood Packet Drop Neighborhood Congestion Notification

Congested node computes drop probability following RED’s alg.

It broadcasts the drop probability to all neighbors

Distributed Neighborhood Packet Drop Neighborhood Drop Prob = Max of all

drop probabilities heard from neighbors

Performance Evaluation: Simple Scenario Both long-term and short-term

fairness is achieved Loss of aggregated throughput

Tradeoff between fairness and throughput

Channel is not fully utilized

Overall Throughput Instantaneous Throughput

F T P 1

F T P 2

F T P 3

(1 0 0 , 1 0 0 ) (6 0 0 , 1 0 0 )

(3 5 0 , 3 5 0 )

(3 5 0 , 7 0 0 )(0 , 7 0 0 )

(3 5 0 , 1 0 5 0 )

(1 0 0 , 1 3 0 0 ) (6 0 0 , 1 3 0 0 )

(7 0 0 , 7 0 0 )

Performance Evaluation: Multiple Congested Neighborhood Multiple congested neighborhoods FTP2 & FTP 5 have more competing flows,

are more likely to be starved

Overall Throughput

F T P 1 F T P 2 F T P 3

F T P 4

F T P 5

F T P 6

Lessons learned TCP unfairness has been found in ad hoc

networks Notion of “wireless neighborhood” key to solve

the “capture” problem -- > NRED for enhancing TCP fairness in ad hoc nets

Good example of “cross layer” interaction (network drops packets to let TCP know..)

Network layer solution ??? Easy to implement ???

Work in progress: Implementation of proposed schemes in the

Linux testbed Measuring channel utilization rate is an over-

layer solution and hardly to implement in practice

Performance measurement

MAC Delay Based NRED Scheme

Node with the maximum MAC delay (Q size) is not the congestion center node, but the previous hop of the congestion center

2

5

1 3

4 6

7

8

9

TCP1

TCP2

TCP3

Node with the Maximum

MAC Delay

Congestion CenterCongestion

Notification

Fig. 6. Congestion detection and notification

Packet drop probability for MAC delay based NRED

minth

1

MAC Delay

Drop Probability

maxth

maxp

pb

Simulation resultsM inth=0.07, M axth=0.12, M axP=?

0

100

200

300

400

500

600

MaxP=0.1 MaxP=0.3 MaxP=0.5 MaxP=0.7 MaxP=0.9

Ove

rall

Thro

ughp

ut (k

bps)

0.0

0.2

0.4

0.6

0.8

1.0

Fairn

ess I

ndex

TCP 1

TCP 2

TCP 3

Fairness

Total

Minth=?, Maxth=0.12, MaxP=0.5

0

100

200

300

400

500

600

Minth=0.03 Minth=0.05 Minth=0.07 Minth=0.09 Minth=0.11

Ove

rall

Thro

ughp

ut (k

bps)

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Fairn

ess I

ndex

T CP 1

T CP 2

T CP 3

Fairness

T otal

Minth=0.07, Maxth=?, MaxP=0.5

0

100

200

300

400

500

600

Maxth=0.07 Maxth=0.09 Maxth=0.12 Maxth=0.15 Maxth=0.18

Ove

rall

Thro

ughp

ut(k

bps)

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Fairn

ess I

ndex

T CP 1

T CP 2

T CP 3

Fairness

T otal

Fig. 8. Performances at different parameters

optimal parameters:

maxp = 0.5 maxth = 0.12 minth = 0.07

Adaptive Pause scheme

Queue length based NRED and MAC Delay Based NRED schemes can not directly find congestion

center congestion notification - hard RED: optimal parameters setting ?

被动：等待受压制节点发出通知才开始控制发包，对网络中的不公平现象反应慢，而且发送消息造成额外的带宽开销。

主动：获得竞争无线信道优势的节点要自律 Adaptive Pause

Adaptive Pause scheme

Each node monitors its bandwidth occupation and the number of its active neighbor nodes

Instead of detecting congestion, the nodes which occupy too much bandwidth, will reduce their sending rate by themselves

If the bandwidth exceeds the threshold, the node will pause an interval to stop sending

0, 0 0, 1 0, W0-2 0, W0-1

1, 0 1, W1-1

m’ , 0

m-1, 0

m’ , 1

1, 1

m-1,Wm-1-1

m, 0 m,Wm-1

,Wm’ -1m’

m,Wm-2

1111

1/W01/W0

p/W1p/W1

p/W2

p/W1

p/Wm’p/Wm’

p/Wm-1

p/Wmp/Wm

...

...

111...

111... ...

...

11... ...

11... ...

1

...

1

1- p

1- p

1- p

1- p

rPause

1 - r

...

r

Markov model for Adaptive_Pause scheme

Simulation result

Throughput of Adaptive_Pause and D_NRED

(Adaptive_Pause parameters: Tp= 1.5 s ，R= 0.5)

0

50

100

150

200

D_NRED Adaptive_Pause

kbps

吞吐

率（

） TCP1

TCP2

TCP3

Simulation result (Adaptive_Pause)

Flow 2 throughput ~ parameters: Tp ， R

Simulation result (Adaptive_Pause)

Flow 2 throughput ~ parameters: Tp ， R

Experiment result on TJU testbed

f

Overal l throughput of 3 TCP fl ows

0

500

1000

15002000

2500

3000

3500

Drop Tai l RED NRED Pause

Thro

ughp

ut(k

bps)

0

0. 2

0. 4

0. 6

0. 8

1

Fair

ness

Ind

ex

fl ow1 fl ow2 fl ow3 total fai rness