understanding tcp fairness over wireless lan ieee infocom 2003 saar pilosof, ramachandran ramjee,...

Understanding TCP fairness over Wireless LAN

IEEE INFOCOM 2003Saar Pilosof, Ramachandran Ramjee, Danny Raz, Yuval Shavitt, Prasun

Sinha

Presented by Yixin Hua

04/22/23WPI CS577 2

Agenda Introduction Problem Overview Simulation Study Modeling TCP Access Our Solution Related Work Conclusion & Discussion

04/22/23WPI CS577 3

Introduction A typical 802.11 installation

CISCOSYSTEMS

802.11 Base Station

802.11 Mobile host

802.11 Mobile host

802.11 Mobile host

04/22/23WPI CS577 4

IntroductionScenarios (Assumption)

All senders or receivers: Share bandwidth equally.

One sender and two receivers: Sender get half BW, and receivers share other half.

04/22/23WPI CS577 5

Problem OverviewReal Experiment Setup

CISCOSYSTEMS

802.11 Base Station

802.11 Mobile host

802.11 Mobile host

802.11 Mobile host

04/22/23WPI CS577 6

Problem OverviewReal Experiment Setup

Ru: Average TCP uplink throughput Rd: Average TCP downlink throughput Ru/Rd: Ratio MTU: Maximum Transmission Unit – Varied Background UDP: to reduce buffer

available to TCP flows – varied by packet size and arrival interval

04/22/23WPI CS577 7

Problem OverviewResult

04/22/23WPI CS577 8

Problem OverviewResult

In basic case, Ru/Rd = 1.44 Does commercial system give high

priority to downstream? Since most applications involve download rather than upload.

With UDP flows, ratio Ru/Rd increase With smaller MTU, ratio reaches 8

04/22/23WPI CS577 9

Problem OverviewFurther investigation with sniffers

Upstream TCP window size reaches its maximum in all cases

Downstream TCP window size changes

04/22/23WPI CS577 10

Problem OverviewFurther investigation with sniffers

04/22/23WPI CS577 11

Problem OverviewConcerns

Wireless link interference Base station buffer size Implementation details of 802.11

MAC layer Difficult to vary and isolate

parameters, and trace their impacts

04/22/23WPI CS577 12

Simulation StudyExperiment 1: 1 TCP sender and 1 TCP receiver

TCP receiver window size w = 42 MTU = 1500 Base station buffer size B = 6 ~ 85 packets Number of ACK per data packet = 1 Data packet size 1024 bytes 5 runs, each lasts 100 second Nodes don’t move

04/22/23WPI CS577 13

Simulation Study Experiment 1: Observation

04/22/23WPI CS577 14

Simulation Study Experiment 1: Observation

Region 1: over 84, ratio is 1 Region 2: 42 to 84, ratio decreases

from 10 to 1 Region 3: 6 to 42, ratio varies

between 9 and 12 Region 4: below 6, data points wide

spread (too noisy)

04/22/23WPI CS577 15

Simulation Study Experiment 1: More observations

04/22/23WPI CS577 16

Simulation Study Experiment 1: More observations

RTT increases monotonically with base station buffer size w/o significant rate changes

Data packet loss rate is always higher than ACK loss rate, not linear with base station buffer size

04/22/23WPI CS577 17

Simulation Study Experiment 1: Further investigation

04/22/23WPI CS577 18

Simulation Study Experiment 1: Further investigation

When buffer size is smaller than 42, sharing result is 1:10

When buffer size becomes larger, sharing ratio increases

When buffer size is larger than 84, Base Packet is equal to the difference between Down ACK and the Up Packet

04/22/23WPI CS577 19

Simulation StudyExperiment 2: Multiple flows

Case 1: One upstream and multiple downstream flows

Case 2: Equal number of multiple upstream and downstream flows

Base station buffer size is 100 packets

5 runs, each lasts 100 seconds

04/22/23WPI CS577 20

Simulation Study Experiment 2: Observation

04/22/23WPI CS577 21

Simulation Study Experiment 2: Observation

Case 1: Ratio is linear All downstream flows share bandwidth equally Total throughput stay stable

Case 2: Average ratio goes up to 800, since upstream

flows’ ACK clutter base station buffer Upstream flows maintain maximum window size Downstream flows struggle with a window of 0-2

packets

04/22/23WPI CS577 22

Modeling TCP Access Scenario 1: One upstream and one downstream flow

Base station buffer size: B TCP receiver window size: w All packet loss due to buffer

overflows at base station

04/22/23WPI CS577 23

Modeling TCP Access Scenario 1: Analysis

Upstream flow window behavior When sender window is large, a loss of

an ACK has no effect on the window size due to TCP cumulative acknowledgement nature.

Sender window will reach w.

04/22/23WPI CS577 24

Modeling TCP Access Scenario 1: Analysis

Downstream flow window behavior It changes depending on B and w, since loss of

data packet will cause sender half window size. If B ( + 1)w, all packets have room in base

station buffer, no drops. Assume BS buffer is full of w ACKs.(?) If B

(+1)w, B - w buffer available for downstream. Sender window will vary between (B - w)/2 and B - w, average window size is 3(B - w)/4. (Simplified)

Ratio: R = 4w/(3(B - w))

04/22/23WPI CS577 25

Modeling TCP Access Scenario 1: Further Analysis

Using bounded size queuing system (M/M/1/K) Arrival rate Rd + Ru , = 1. The probability of K packets in a buffer in a stable

state, Pk = (1-) k/(1- k+1) (1) is ratio between arrival rate and service rate,

= (Rd + Ru)/ Ru= 1+R, where R = Rd/Ru (2) Drop rate approximate to p = (1+BR)/(B+1) (3) Using Rd= sqrt(3/(2p))/RTTd, and Ru=w/RTTu R = RTTu/RTTd*sqrt(3/(2w2p)) (4)

04/22/23WPI CS577 26

Modeling TCP Access Scenario 1: Further Analysis

Using (3) and (4), We get (1+BR)/(B+1) = 3/(2w2R2) Finally

Using 1+B B and 1+BR BR, R = 1/10.56, it gives an approximation for region 6 to 42.

04/22/23WPI CS577 27

Modeling TCP Access Scenario 1: Further Analysis

04/22/23WPI CS577 28

Modeling TCP Access Scenario 1: Further Analysis

When B > w and only loss is due to buffer overflow, window size is composed from a fixed part B - w, and a part of interaction with acknowledgements in the BS buffer.

Effective average window size is sqrt(3/(2p)) + 3(B - w)/4

R=RTTu/(w*RTTd)*sqrt(3/(2p))+3(B-w)/4 (6) It gives an approximation to region 42 to 85. When w=42, it matches with Eq. 4

04/22/23WPI CS577 29

Modeling TCP Access Scenario 1: Validation

TCP doesn’t provide a nice arrival behavior like M/M/1/K

04/22/23WPI CS577 30

Modeling TCP Access Scenario 2: Small Buffer Upstream flow with small buffer size, using discrete

time Markov chain

State i represents a state where TCP window size is 2i On state i, go to state 0 if a timeout occurs with

probability p2i

Otherwise double the window size and move into state i+1 with probability 1- p2i

With ns2, the upstream flow always end up with maximum window size

04/22/23WPI CS577 31

Modeling TCP Access Scenario 3: Multiple Flows

From eq. 1, = (Rd + nRu)/ Ru= 1+nR (7)

Drop rate, p = (1+nBR)/(B+1)(8)

R=sqrt(3/(2nw2p))+3(B-w)/(4nw)(9)

It fits figure 6 very well!

04/22/23WPI CS577 32

Modeling TCP Access Scenario 3: Multiple Flows

04/22/23WPI CS577 33

Our solution

1. Separately queue for TCP data and ACK packets at base station - Doesn’t wok

2. Fake duplicate ACK packets or discard data packets to force TCP to reduce the upstream window size – Waste BW

3. Using advertised receiver window field in the ACK packets towards TCP sender, BS manipulates the receiver window

04/22/23WPI CS577 34

Our solution Solution 3

Keep a counter for numbering current TCP flow in the system

If n flows in system, BS set receiver window to B/n

? Web traffic, bursty flow, UDP ! An XCP way

04/22/23WPI CS577 35

Our solutionA simulation for solution 3

04/22/23WPI CS577 36

Related Work Lu et al. [2] first identified the problem under a UDP

model. They proposed a centralized scheduling algorithm performed at BS.

Nandagopal et al. [3] suggests a fairness model that identify the different node fairness and flow fairness.

Research [9] suggests employ BW reservation over MA channels to support QoS.

Sobrinho and Krishnakumar [10] suggests blackburst to find the the real-time sender with longest waiting time( and thus the highest priority).

04/22/23WPI CS577 37

Related Work Deng and Chang [1] suggested to change the backoff

period according to a station priority. The lower the priority the higher is the maximum backoff period a station can draw.

Berry et al. [11] follower the line and use two distinct backoff periods for two priority classes.

Vaidya et al. [4] suggests a distributed algorithm that calculates the backoff period for the stations that resulted access to the channel will closely match the Self-Clocked Fair Queueing scheduling.

Ada and Castelluccia suggests three differential mechanism based on scaling of the congestion window, modifying the IFSs, and changing the maximum frame length.

04/22/23WPI CS577 38

Conclusion & Discussion Buffer size at base station plays a key role in the

observed unfairness. Based on simulation, the unfairness in TCP throughput

ration could be as high as 800. Using bounded size queuing system (M/M/1/K), authors

explained TCP’s behavior and interaction with MAC layer. The analysis identified four regions of unfairness that

depend on the buffer availability at base station. Proposed solution using advertised window manipulated

by base station alleviates the problem in simulation and testbed.

04/22/23WPI CS577 39

Conclusion & Discussion Open discussion:

Channel losses TCP with different RTT Providing higher share of the media to

the base station Interaction with IPSec