1

Short-term Fairness for TCP Flows in 802.11b WLANs

M. Bottigliengo, C. Casetti, C.-F. Chiasserini, M. Meo

INFOCOM 2004

2

Outline

Introduction Reference Scenario Scheduling Algorithm Description Simulation Scenario Numerical Results Conclusions My Comments

3

Introduction

WLANs using the 802.11b technology add some obstacles to QoS provisioning Poor channel quality depending on relative po

sition of wireless station (WS) Interference from hidden terminals The anomaly due to the different speeds at wh

ich WSs transmit

4

Introduction (cont’d)

This paper addresses the fairness issue among TCP flows

Short-lived flows, representing the majority of today’s Web traffic, suffer from packet losses occurring during or just past the three-way handshake phase The TCP congestion window size may not be

large enough so as to trigger the Fast Recovery algorithm

5

This paper…

The main contribution of this paper is the proposal of an LLC-layer algorithm that can be implemented on both AP and WSs

The algorithm aims at guaranteeing fair access to the medium to every user, by awarding longer transmission opportunities to WSs that experienced short channel failures

6

This paper… (cont’d)

The award mechanism works by monitoring the successful medium accesses over a measurement window, and allowing short or long bursts depending on the WS’s history

7

Reference Scenario

Network and Traffic Network scenario Given N wireless stations

N/2 TCP connections in the uplink direction N/2 TCP connections in the downlink direction

8

Network scenario

Data rate = 11 Mbps

9

Reference Scenario (cont’d)

Channel Model An independent error model for each

communicating pair of nodes was introduced An error model is represented by a

three-state discrete-time Markov chain

10

Three-state Error Model

Long Bad

Short Bad

Lbg

LBp

L_

1

Sbg

SBp

L_

1

GOOD

11

Goals of the Scheduling Algorithm

Improve the fairness among wireless stations that may experience location-dependent channel capacity and errors

Provide short-term fairness to TCP traffic in order to enhance the performance of short-lived flows

12

Scheduling Algorithm Description

At the AP LLC layer, we introduce a separate queue for each WS associated to the AP, while only one queue is implemented at the WS LLC layer

A channel condition estimator is associated to each queue, and transmission is allowed only for the those queues whose channel is estimated to be good

13

Scheduling Algorithm Description (cont’d) Upon switching to good channel state,

A queue that has just experienced a bad channel is rewarded with the possibility to send to the MAC layer a maximum number of back-to-back frame (TXburst),

proportional to its forced silence period The MAC layer will then use the EDCF bursti

ng capability of the current 802.11e draft, to avoid contentions within the burst

14

Scheduling Algorithm

1) channel state estimation

2) queue selection and service

3) TXburst length setting

15

Channel State Estimation

Channel State – GOOD AP receives

A MAC-layer acknowledgment in response to a data frame

A CTS frame in response to an RTS frame An error-free data frame or RTS

16

Channel State Estimation (cont’d)

Channel State – BAD The AP sets the flag to BAD after a

transmission failure The Long Retry Counter (LRC) is incremented

when the transmission of a frame longer than the RTS threshold fails due to channel errors

When LRC (SRC) reaches the LRL value, the MAC layer abandons the transmission of the frame and it signals the failure to the LLC layer

17

Channel State Estimation (cont’d)

Channel State – BAD (cont’d) We can assume that it is highly like that values

of SRC larger than 4 are due to channel errors

SRL = 4

LRL = 0

18

Channel State Estimation (cont’d)

Channel State – PROBE The AP switches the flag from BAD to PROBE

when a configurable value (PTIMER) expires PTIMER starts to run whenever the channel

state switches to BAD The initial value is doubled (PROBE->BAD) The value is reset (PROBE->GOOD)

A WS whose queue flag has a PROBE value can transmit a single data frame (RTS) to check the new channel state

19

Queue Selection Service

20

Queue Selection Selection

NO

YES

21

TXburst Length Setting

The AP has first to compute a fair target throughput (Thr_Fair), that is the throughput value representing the fair service that each WS must receive at a given time (WinLen)

FairThr

iThrQueneFairThri

CounterWS

ThrTotalFairThr

_

)(__

_

__

22

TXburst Length Setting (cont’d)

In order to allow WSs stations to apply the same “controlled” bursty transmission , they need the Thr_Fair value estimated by the AP Since WSs cannot compute the achieved glob

al throughput

The AP must include the Thr_Fair in control frames that are broadcast on the WLAN (for example, in the Beacon Frame)

23

tstat- http://tstat.tlc.polito.it/

24

Simulation Scenario

Simulations run under the ns-2 simulator Using ON-OFF source

The average duration of the OFF period is a configurable parameter

The ON duration depends on the amount of bytes to sent

Uniformly chosen among the ones reported in Table III

25

Simulation Scenario (cont’d)

Transport-layer settings TCP version is NewReno The Max. Segment Size (MSS) of TCP segment is equ

al to 1000 bytes

Data Link-layer settings Set the RTS threshold at 400 bytes so that TCP ACKs

are never using the RTS/CTS handshaking LLC queue at AP are 400 data frames long at AP

26

Up/Down Flow Throughput Ratio

From : Understanding TCP Fairness over Wireless LAN (INFOCOM 2003)

27

Numerical Results

Factors TCP segment Size WinLen size Fixed or Dynamic TXburst Average Off Duration (Traffic Load) Error Probability

28

Average per-connection Throughput

29

3-Downlink (0-2) vs. 3-Uplink (3-5)

30

Average Throughput of Uplink Connection

Average duration of TCP source OFF period = 1s

31

Average Throughput of Connection with Different Segments Size

32

Fairness Index for Different Segments Size

0.7

0.1

N

i i

N

i iJ

rN

rF

1

2

2

1

33

27-Segment Flow Fairness Index

34

Average Connection Completion Time

35

Average Throughput under Different Traffic Load

N = 20

Heavy Load

37

Average per-connection Throughput

38

Fairness Index for Different Segments Size

40

Conclusions

It silences WSs while the channel is unavailable and uses a burst award mechanism to compensate for the missed transmission opportunities

It is highly beneficial for short-lived TCP flows, that normally suffer from losses on their early windows when competing with long-lived flows on a congested link

41

My Comments

LLC-layer Algorithm, not MAC-layer Algorithm

In 802.11e draft, a TSPEC describes the QoS characteristics of a traffic stream, it can reserve resources within the HC and modify the HC’s scheduling behavior Delay Bound Minimum Burst Size Service Interval