18
1/15 Basic Theorems on the Backoff Process in 802.11 Basic Theorems on the Backoff Process in 802.11 JEONG-WOO CHO Q2S, Norwegian University of Science and Technology (NTNU), Norway Joint work with YUMING JIANG Q2S, Norwegian University of Science and Technology (NTNU), Norway A part of this work was done when J. Cho was at EPFL, Switzerland.

Basic Theorems on the Backoff Process in 802.11

Embed Size (px)

DESCRIPTION

Basic Theorems on the Backoff Process in 802.11. JEONG-WOO CHO Q2S, Norwegian University of Science and Technology (NTNU), Norway. Joint work with YUMING JIANG Q2S, Norwegian University of Science and Technology (NTNU), Norway. - PowerPoint PPT Presentation

Citation preview

Page 1: Basic Theorems on the Backoff Process  in 802.11

1/15Basic Theorems on the Backoff Process in 802.11

Basic Theorems on the Backoff Process in 802.11

JEONG-WOO CHOQ2S, Norwegian University of Science and Technology (NTNU), Norway

Joint work withYUMING JIANG

Q2S, Norwegian University of Science and Technology (NTNU), Norway

A part of this work was done when J. Cho was at EPFL, Switzerland.

Page 2: Basic Theorems on the Backoff Process  in 802.11

2/15Basic Theorems on the Backoff Process in 802.11

Understanding 802.11

• Single-cell 802.11 network• Every node interferes with the rest of the nodes.

• CSMA synchronizes all nodes.• User activity is determined by whetherwhether there is a carrier in the medium or

not.

• Sufficiency Sufficiency of the backoff analysis• The kernel lies in backoff analysis

• Backoff process is simple(i) Every node in backoff stage k attempts transmission with probability pk for

every time-slot.

(ii) If it succeeds, k changes to 0; otherwise, k changes to (k+1) mod (K+1) where K is the index of the highest backoff stage.

Page 3: Basic Theorems on the Backoff Process  in 802.11

3/15Basic Theorems on the Backoff Process in 802.11

Why MMean FField TTheory?

Node 1 @ backoff stage iNode 2 @ backoff stage j

Co

llisi

on

Node 1 @ backoff stage i+1Node 2 @ backoff stage j+1

Inve

rse

co

llisi

on

?

• Markov chain models of the backoff process• Due to their irreversibilityirreversibility, mathematically intractable.

• Decoupling approximation • Backoff process at a node is asymptotically independent from

those at other nodes.

[BEN08] M. Benaim and J.-Y. Le Boudec, “A class of mean field limit interaction models for computer and communication systems”, Perf. Eval., Nov. 2008.

[BOR07] C. Bordenave, D. McDonald, and A. Proutiere, “A particle system in interaction with a rapidly varying environment: Mean Field limits and applications”, to appear in NHM.

• Q: Decoupling approximation is valid?• Exactly under which conditions?

• Recent advances in Mean Field Theory [BEN08] [BOR07]Recent advances in Mean Field Theory [BEN08] [BOR07]• If the following nonlinear ODEs are globally stable, it is valid; otherwise, oscillations may occur.

K

k kkKK

kkkkk

tptpttptpttptdt

d

Kktpttptdt

d

0000

11

)()( where)()()()(1)()(

,,1for ,)()()()(

Page 4: Basic Theorems on the Backoff Process  in 802.11

4/15Basic Theorems on the Backoff Process in 802.11

Decoupling Approximation Validated

. :

,:

, stage backoffat :

}.{0,1,...,stages,backoff1 and nodesareThere

rateattempt average

yprobabilit collision

rateattempt

p

kp

KKN

k

0.for holds regime stationary in the

ceindependen fieldmean thesequence, ingnonincreas a is ,0 , If

ODEs) FieldMean ofStability (Global1Theorem

K

Kkpk

• Bianchi’s Formula• Representative formula exploiting decoupling approximation.

• A set of fixed-point equations to compute collision probability.

pNγ,

γp

K

kk

k

K

k

k

1exp1

0

0

fixed. are and , kpKN

Page 5: Basic Theorems on the Backoff Process  in 802.11

5/15Basic Theorems on the Backoff Process in 802.11

Beyond Throughput Analysis

• New Interest in Backoff Distribution

• How much backoff time should a packet wait for transmission?

backoffpacket -percalledpacket, afor generated valuesbackoff of sum the: Ω

[BRE09] M. Bredel and M. Fidler, “Understanding fairness and its impact on quality of service in IEEE 802.11”, IEEE Infocom, Apr. 2009.

[BER04] G. Berger-Sabbatel et al., “Fairness and its impact on delay in 802.11 networks”, IEEE Globecom, Nov. 2004.

• Possible misunderstanding misunderstanding for N=2• Based on extensive simulations, for the case N=2, [BRE09] and [BER04]

concluded that Ω is exponentially and uniformly distributed, resp.

• Possible misunderstanding about the distribution of Ω.

Page 6: Basic Theorems on the Backoff Process  in 802.11

6/15Basic Theorems on the Backoff Process in 802.11

OutlineMean Field Technique Revisited

• Supports us to apply decoupling approximation in the following principles

1. Per-Packet Backoff Principle• One of the two works is incorrect?

2. Power-Tail Principle• What is the distribution type of the delay-related variables?

• Is there long-range dependence inherent in 802.11?

3. Inter-Transmission Principles• Can we develop an analytical model for short-term fairness?

• When does the short-term fairness undergo a dramatic change?

Conclusion

Page 7: Basic Theorems on the Backoff Process  in 802.11

7/15Basic Theorems on the Backoff Process in 802.11

Per-Packet Backoff Principle

2

1

1

002

22

0

1

000

22

121

1

bygiven are of variance theand mean, pdf, Then the ./ variance

and 1/mean with stage backoffat valuesbackoff of pdf thedenote)(Let

Principle)BackoffPacket -(Per2Theorem

Ωpp

γ

p

γvσ,

p

γΩ

γxffγγxff(x)f

Ωpv

pkf

K

k

k

i ik

kK

k k

k

Ω

K

k k

k

K

k

kk

KKΩ

k

kk

• Misunderstandings cleared up: both works [BRE09] [BER04] are correct.

• The contradicting conclusions are due to the different contention window size in 802.11b and 802.11a/g.

• For N=2,

• In the sense that

• 802.11b leads to approx. uniform backoff distribution, while 802.11a/g leads to approx. exponential backoff distribution

802.11a/g,in 1

802.11b,in ,3

1

Ω

σvΩ

Page 8: Basic Theorems on the Backoff Process  in 802.11

8/15Basic Theorems on the Backoff Process in 802.11

Long-range Dependence (LRD)

Self-similarprocesses

Processesw/ finite 2nd moment

LRDProcesses

• There are LRD processes that are – either not self-similarnot self-similar

– or with infinite varianceswith infinite variances.

• Correctly speaking, harmful is LRD.

• Why LRD, termed “ “Joseph EffectJoseph Effect” [MAN68],” [MAN68], is harmful? – [Bible, Genesis 41] “Seven years of great abundance are coming throughout the

land of Egypt, but seven years of famine will follow them.”• long periods of overflow followed by long periods of underflow

• hard to derive efficient bandwidth (envelope) of the traffic and to decide buffer size

[MAN68] B. Mandelbrot and J Wallis, “Noah, Joseph and operational hydrology”, Water Resources Research, 1968.

Page 9: Basic Theorems on the Backoff Process  in 802.11

9/15Basic Theorems on the Backoff Process in 802.11

Bridging between Maths on LRD and 802.11

Black BoxApproach

• Empirical studiesEmpirical studies based on high volume data sets of traffic measurements

Getting to KnowYour Network

Approach

• Qualitative studies Qualitative studies based on rigorous mathematical theories

• “Focuses on understanding of LRD and providing physical explanationsphysical explanations.” [WIL03]

• Developed by Kaj & Taqqu et al. (around 2005)

• A A bridgebridge between this approach and 802.11 is between this approach and 802.11 is required.required.

Theoretical

Gap

The state of the art in 802.11

[WIL03] W. Willinger, V. Paxson, R. Riedi, and M. Taqqu, “Long-Range Dependence and Data Network Traffic”, Theory and Applications of Long-Range Dependence, Birkhäuser Boston, 2003.

Page 10: Basic Theorems on the Backoff Process  in 802.11

10/15Basic Theorems on the Backoff Process in 802.11

Power-Tail Principle

• Per-packet backoff has a truncated form of Pareto-type distributiona truncated form of Pareto-type distribution.

• Sketch of proof:

(1) Discovery of recursive relation in LST of

(2) The quantifier set in regular variation theory is dense.

(3) Application of advanced Karamata Tauberian Theorem

• A bridgebridge between recent mathematical theories on LRD and 802.11

802.11)in 2(.log)log( andinfinity at yingslowly var is)(where

)(~)()(

, as Formally, . a has backoffpacket -per the, If

Principle)Tail(Power5Theorem

c

mm/γ-αx

xxdxxfxF

xΩK

x

tail type-Pareto

)(xf

Page 11: Basic Theorems on the Backoff Process  in 802.11

11/15Basic Theorems on the Backoff Process in 802.11

LRD in 802.11 Identified

[KAJ05] I. Kaj, “Limiting fractal random processes in heavy-tailed systems”, Fractals in Engineering, 2005.

Long-range dependenceLong-range dependence in 802.11 is identified.

• Backoff process of each node can be viewed as a renewal counting process.

2log)log( if tailed-heavy is

,By

m/γ-α

Principle Tail-Power

0 10 20 30 40 50 600

5

10

15

20

25

Time

Co

un

t

Ω

Superpose ? )()( of form thebe what willsevere, is contention If1

N

n

n tAtAprocess ionsuperposit the

motion.Levy and fBm out turns)(

KAJ05],By

between

[

tA

Intermediate Telecom Process

LRD processLRD process that is not self-similar

Page 12: Basic Theorems on the Backoff Process  in 802.11

12/15Basic Theorems on the Backoff Process in 802.11

Short-Term Fairness in 802.11• Long-termLong-term Fairness in 802.11 (without enhanced functionalities)

• the total throughput shared equally.

• Short-termShort-term Fairness in 802.11: not quantified yet.

pkts. ing transmittis node whilepkts transmit 1-,1, nodes:P ζNzNζzN

• Inter-transmission probability Inter-transmission probability

• Node N is the tagged node.

zPN

Page 13: Basic Theorems on the Backoff Process  in 802.11

13/15Basic Theorems on the Backoff Process in 802.11

Inter-Transmission principles

r.v. of pdf the,exp1

and

r.v., of pdf the,2

exp2

1

where

1xP

3Theorem

2

Poissonλλz!

λ,z

normalx

πx

dx,zxvζζNζz

z

ΩN

Ps

Nm

PsNm

yζζyτ

ζcyτζNδzyq

.,r.v stable-αLevy y

process, Telecom teIntermediaτx

dyydxxζz

/αα

α

α

(y)q

(y)q

τ(y)/cN

011

,1

,)0,1,1( of pdf theis

, of pdf theis

where

P

6Theorem

SLv

YTc

LvTc

Doubly stochastic Poisson processDoubly stochastic Poisson process

: : a Poisson process on the line with random intensity

The resultant dist. is approx. Gaussian.The resultant dist. is approx. Gaussian.

General formula forGeneral formula for(i) small K(i) small K

(ii) large K and (ii) large K and αα>2>2

General formula for General formula for (iii) large K and (iii) large K and αα<2<2

The resultant dist. is approx. Lévian The resultant dist. is approx. Lévian entailing skewness.entailing skewness.

Leaning: dist. is leaning to the left

Directional: dist. has heavy-tail on its right part and decays faster than exponentially on its left part.

/ζvN-,N-ζλ Ω2211N

],0(

Page 14: Basic Theorems on the Backoff Process  in 802.11

14/15Basic Theorems on the Backoff Process in 802.11

Collision Dominates Aggregation

ζ

v

ζ

v

ζNZ

σv ΩΩZ

Z

2

1

1

Aggregation EffectAggregation Effect

: Poisson Limit for

Superposition Process

: Decreases with NDecreases with N

Collision EffectCollision Effect

: Gaussian Intensity

: Increases with NIncreases with N

• Gaussian (collision effect) dominates Poisson (aggregation effect).

Given by Per-Packet Backoff

Principle

Page 15: Basic Theorems on the Backoff Process  in 802.11

15/15Basic Theorems on the Backoff Process in 802.11

Conclusion

Decoupling Approximation Revisited

Per-Packet Backoff Principle– Possible misunderstanding removed.

Power-Tail Principle– Backoff distribution formula: truncated Pareto-type.Backoff distribution formula: truncated Pareto-type.

Inter-Transmission Principles– Short-term fairness formulas: approximately Gaussian or LShort-term fairness formulas: approximately Gaussian or Léévianvian

pNγ,

p

γ

γp

K

kk

k

K

k

k

1exp1

0

0

Page 16: Basic Theorems on the Backoff Process  in 802.11

16/15Basic Theorems on the Backoff Process in 802.11

Self-Similarity and Long-Range Dependence

)]1(E[ moments, 2 finitewith 22nd Z

1,

2

1for is )( that implying ,)12(~E)(

bygiven isfunction ation autocorrel the),(-1)( Defining

222 HkrkHHXXkr

kZkZX

H

kii

k

summablenot

0:)(0:)(

,0 allfor if, )1,0(index with is increments stationary with process stochsticA d

ttZatatZ

aH

H

similar-self

• Roughly, a self-similarself-similar process with finite 2nd moment is long-range long-range dependentdependent if H>1/2, in the sense r(k) possesses non-summability.

• Self-similarity doesn’t have negative implications. It is long-range long-range dependencedependence which has a serious impact on the network performance.

Page 17: Basic Theorems on the Backoff Process  in 802.11

17/15Basic Theorems on the Backoff Process in 802.11

NS-2 Simulation Results

– Estimated slopes on log-log scale show a good match with analytical formulae.

802.11bK=6

802.11bK=6

Page 18: Basic Theorems on the Backoff Process  in 802.11

18/15Basic Theorems on the Backoff Process in 802.11

NS-2 Simulation Results

– Leaning tendency and directional unfairness can be observed as predicted by analysis.

802.11bK=6