on the validity (or otherwise) of ieee 802.11 mathematical modeling...
TRANSCRIPT
On the validity (or otherwise) of IEEE 802.11mathematical modeling hypotheses
Ken Duffy
Joint work with Kaidi Huang and David Malone
Hamilton Institute, National University of Ireland Maynooth
Swinburne University of Technology, May 14th 2009
Talk outline.
Talk outline.
• The IEEE 802.11 CSMA/CA MAC.
Talk outline.
• The IEEE 802.11 CSMA/CA MAC.
• Recent advances in mathematical modeling.
Talk outline.
• The IEEE 802.11 CSMA/CA MAC.
• Recent advances in mathematical modeling.
• Implicit approximations made to enable analytic tractability.
Talk outline.
• The IEEE 802.11 CSMA/CA MAC.
• Recent advances in mathematical modeling.
• Implicit approximations made to enable analytic tractability.
• Directly testing these hypotheses with test-bed data.
Talk outline.
• The IEEE 802.11 CSMA/CA MAC.
• Recent advances in mathematical modeling.
• Implicit approximations made to enable analytic tractability.
• Directly testing these hypotheses with test-bed data.
• Summary, an epilogue and conclusions.
The 802.11 DCF
DataData
SIFS DIFS Decrement counter
Select Random Number in [0,31]
SIFS DIFS
Counter
Expires;
TransmitPause Counter
ResumeAck AckData
Figure: 802.11 MAC operation
The 802.11 MAC flow diagram
(0, W−2)(0,1)
(1,1)(1,0)
Collision
Collision
(2,0) (2,1)
Collision
(0,0)
No collision
No collision
No collision
(0, W−1)
(1, 2W−1)
(2, 4W−1)(2, 4W−2)
(1, 2W−2)
Figure: Saturated 802.11 MAC operation
Popular mathematical modeling approaches
• P-persistent:approximate the back-off distribution be ageometric with the same mean. E.g. work by Marco Contiand co-authors (F Cali, M Conti, E Gregori, P AlephIEEE/ACM ToN 2000).
Popular mathematical modeling approaches
• P-persistent:approximate the back-off distribution be ageometric with the same mean. E.g. work by Marco Contiand co-authors (F Cali, M Conti, E Gregori, P AlephIEEE/ACM ToN 2000).
• Mean-field Markov models: seminal work by Bianchi (IEEEComms L. 1998, IEEE JSAC 2000).
Bianchi’s approach
Observation: each individual station’s impact on overall networkaccess is small.
Bianchi’s approach
Observation: each individual station’s impact on overall networkaccess is small.Mean field approximation: assume a fixed probability of collision ateach attempted transmission p, irrespective of the past.
Bianchi’s approach
Observation: each individual station’s impact on overall networkaccess is small.Mean field approximation: assume a fixed probability of collision ateach attempted transmission p, irrespective of the past.Each station’s back-off counter then a Markov chain.
Mean-field Markov Model’s Chain
(2, 4W−2)
(0,1)
(1,1)(1,0)
(0,0)
1−p
(2,0) (2,1)
111
1 1 1
11
p
p
p
1−p
1−p
1
(2, 4W−1)
(1, 2W−1)
(0, W−1)(0, W−2)
(1, 2W−2)
Figure: Individual’s Markov Chain if p known
Mean-field Markov OverviewStationary distribution gives the probability the station attemptstransmission in a typical slot
Mean-field Markov OverviewStationary distribution gives the probability the station attemptstransmission in a typical slot
τ(p) =2(1− 2p)
(1− 2p)(W + 1) + pW (1− (2p)m).
Mean-field Markov OverviewStationary distribution gives the probability the station attemptstransmission in a typical slot
τ(p) =2(1− 2p)
(1− 2p)(W + 1) + pW (1− (2p)m).
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0 0.1 0.2 0.3 0.4 0.5
p
tau(p)
Figure: Attempt probability τ(p) vs p
The self-consistent equationNetwork of N stations.
The self-consistent equationNetwork of N stations. Mean field decoupling idea:
The self-consistent equationNetwork of N stations. Mean field decoupling idea: the impact ofevery station on the network access of the others is small,
The self-consistent equationNetwork of N stations. Mean field decoupling idea: the impact ofevery station on the network access of the others is small, so that
1− p = (1− τ(p))N−1. (1)
Solution of equation (1) determines the network’s “real” p∗.
The self-consistent equationNetwork of N stations. Mean field decoupling idea: the impact ofevery station on the network access of the others is small, so that
1− p = (1− τ(p))N−1. (1)
Solution of equation (1) determines the network’s “real” p∗.
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5
(1-t
au(p
))^{
N-1
}
p
1-pN=2N=4N=8
N=16
Figure: 1− p and (1− τ(p))N for N = 2, 4, 8 &16
Example developments
• Unsaturated 802.11, Small buffer: Ahn, Campbell, Veres andSun, IEEE Trans. Mob. Comp., 2002; Ergen, Varaiya,ACM-Kluwer MONET, 2005; Malone, K.D., Leith,IEEE/ACM Trans. Network., 2007.
Example developments
• Unsaturated 802.11, Small buffer: Ahn, Campbell, Veres andSun, IEEE Trans. Mob. Comp., 2002; Ergen, Varaiya,ACM-Kluwer MONET, 2005; Malone, K.D., Leith,IEEE/ACM Trans. Network., 2007.
• Unsaturated 802.11, Big buffer: Cantieni, Ni, Barakat andTurletti, Comp. Comm., 2005; Park, Han and Ahn,Telecomm. Sys., 2006; K.D. and Ganesh, IEEE Comm. Lett.,2007.
Example developments
• Unsaturated 802.11, Small buffer: Ahn, Campbell, Veres andSun, IEEE Trans. Mob. Comp., 2002; Ergen, Varaiya,ACM-Kluwer MONET, 2005; Malone, K.D., Leith,IEEE/ACM Trans. Network., 2007.
• Unsaturated 802.11, Big buffer: Cantieni, Ni, Barakat andTurletti, Comp. Comm., 2005; Park, Han and Ahn,Telecomm. Sys., 2006; K.D. and Ganesh, IEEE Comm. Lett.,2007.
• 802.11e, Saturated: Kong, Tsang, Bensaou and Gao, IEEEJSAC, 2004; Robinson and Randhawa, IEEE JSAC, 2004.Unsaturated: Zhai, Kwon and Fang, WCMC, 2004. Chen,Xhai, Tian and Fang, IEEE Trans. W. Commun., 2006.
Example developments
• Unsaturated 802.11, Small buffer: Ahn, Campbell, Veres andSun, IEEE Trans. Mob. Comp., 2002; Ergen, Varaiya,ACM-Kluwer MONET, 2005; Malone, K.D., Leith,IEEE/ACM Trans. Network., 2007.
• Unsaturated 802.11, Big buffer: Cantieni, Ni, Barakat andTurletti, Comp. Comm., 2005; Park, Han and Ahn,Telecomm. Sys., 2006; K.D. and Ganesh, IEEE Comm. Lett.,2007.
• 802.11e, Saturated: Kong, Tsang, Bensaou and Gao, IEEEJSAC, 2004; Robinson and Randhawa, IEEE JSAC, 2004.Unsaturated: Zhai, Kwon and Fang, WCMC, 2004. Chen,Xhai, Tian and Fang, IEEE Trans. W. Commun., 2006.
• 802.11s, unsaturated: K.D., Leith, Li and Malone, IEEEComm. Lett., 2006.
Standard approach to model verification
ASK: Do the model throughput and delay predictions match wellwith results from simulated system?
Standard approach to model verification
ASK: Do the model throughput and delay predictions match wellwith results from simulated system?NOT: Make the approximations explicit hypotheses and checkthem directly.
A warning from hydrology
”The modelling technology has far outstripped the level of our
understanding of the physical processes being modeled. Making
use of this technology then requires that the gaps in the factual
knowledge be filled with assumptions which, although often
appearing logical, have not been verified and may sometimes be
wrong”.
Vit Klemes, WCP-98, WHO, 1985.
Test bed
Figure: PC as AP, 1 PC and 9 PC-based Soekris Engineering net4801 asclients. All with Atheros AR5215 802.11b/g PCI cards. ModifiedMADWiFi wireless driver for fixed 11 Mbps transmissions and specifiedqueue-size.
What are the hypotheses?
All models:• Ck = 1 if kth transmission results in collision.
What are the hypotheses?
All models:• Ck = 1 if kth transmission results in collision.• Ck = 0 if kth transmission results in success.
What are the hypotheses?
All models:• Ck = 1 if kth transmission results in collision.• Ck = 0 if kth transmission results in success.Assumptions:
• (A1) {Ck} is an independent sequence;
What are the hypotheses?
All models:• Ck = 1 if kth transmission results in collision.• Ck = 0 if kth transmission results in success.Assumptions:
• (A1) {Ck} is an independent sequence;
• (A2) {Ck} are identically distributed with P(Ck = 1) = p.
Testing (A1): {Ck} independent
0 5 10 15 20−0.1
−0.05
0
0.05
0.1
0.15
0.2
Lag
AutoCovariance Coefficient
Saturated
N=2 λ=750
N=5 λ=300
N=10 λ=150
Figure: Saturated C1, . . . ,CK normalized auto-covariances. Experimentaldata, N = 2, 5, 10, K = 2500k, 1200k, 711k.
Testing (A1): {Ck} independent
0 5 10 15 20−0.02
0
0.02
0.04
0.06
0.08
0.1
Lag
AutoCovariance Coefficient
Unsaturated, Big Buffer
N=2 λ=250
N=5 λ=100
N=10 λ=50
Figure: Unsaturated, big buffer C1, . . . ,CK normalized auto-covariances.Experimental data, N = 2, 5, 10, K = 1800k, 750k, 380k.
Testing (A2): {Ck} identically distributed
Record the backoff stage at which the attempt was made.Probability pi of collision given backoff stage i .
Testing (A2): {Ck} identically distributed
Record the backoff stage at which the attempt was made.Probability pi of collision given backoff stage i .Assumption (A2): pi = p for all i .
Testing (A2): {Ck} identically distributed
Record the backoff stage at which the attempt was made.Probability pi of collision given backoff stage i .Assumption (A2): pi = p for all i .MLE
p̂i =#collisions at back-off stage i
#transmissions at back-off stage i.
Testing (A2): {Ck} identically distributed
Record the backoff stage at which the attempt was made.Probability pi of collision given backoff stage i .Assumption (A2): pi = p for all i .MLE
p̂i =#collisions at back-off stage i
#transmissions at back-off stage i.
Hoeffding’s inequality (1963):
P(|p̂i − pi | > x) ≤ 2 exp (−2x(#transmissions at back-off stage i)) .
To have 95% confidence that |p̂i − pi | ≤ 0.01 requires 185attempted transmissions at backoff stage i .
Testing (A2): {Ck} identically distributed
0 2 4 6 8 10 120
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Backoff Stage
Co
llisio
n P
rob
ab
ility
Saturated
N=2 λ=750
N=5 λ=300
N=10 λ=150
Bianchi
Figure: Saturated collision probabilities. Experimental data.
Testing (A2): {Ck} identically distributed
0 2 4 6 8 10 120
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
Backoff Stage
Co
llisio
n P
rob
ab
ility
Unsaturated, Big Buffer
N=2 λ=250
N=5 λ=100
N=10 λ=50
Figure: Unsaturated, big buffer collision probabilities. Experimental data.
What are the big-buffer hypotheses?
Big-buffer models:• Qk = 1 if packet waiting after kth successful transmission.
What are the big-buffer hypotheses?
Big-buffer models:• Qk = 1 if packet waiting after kth successful transmission.• Qk = 0 if no packet waiting after kth successful transmission.
What are the big-buffer hypotheses?
Big-buffer models:• Qk = 1 if packet waiting after kth successful transmission.• Qk = 0 if no packet waiting after kth successful transmission.Assumptions:
• (A3) {Qk} is an independent sequence;
What are the big-buffer hypotheses?
Big-buffer models:• Qk = 1 if packet waiting after kth successful transmission.• Qk = 0 if no packet waiting after kth successful transmission.Assumptions:
• (A3) {Qk} is an independent sequence;
• (A4) {Qk} are identically distributed with P(Qk = 1) = q.
Testing (A3): {Qk} independent
0 5 10 15 20
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Lag
Auto
Covariance C
oeffic
ient
Unsaturated, Big Buffer
N=2 λ=250
N=5 λ=100
N=10 λ=50
Figure: Unsaturated, big buffer queue-non-empty sequence normalizedauto-covariances. Experimental data. K = 1700k, 720k, 360k.
Testing (A4): {Qk} identically distributed
0 2 4 6 8 10 120
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Backoff Stage
P(Q
>0)
Unsaturated, Big Buffer
N=2 λ=250
N=5 λ=100
N=10 λ=50
Figure: Unsaturated, big buffer queue-non-empty probabilities.Experimental data. (Note the large y-range!)
What are the 802.11e hypotheses?
Models with different AIFS values:• Hk is length of kth stuck in a hold-state.
What are the 802.11e hypotheses?
Models with different AIFS values:• Hk is length of kth stuck in a hold-state.Assumptions:
• (A5) {Hk} is an independent sequence;
What are the 802.11e hypotheses?
Models with different AIFS values:• Hk is length of kth stuck in a hold-state.Assumptions:
• (A5) {Hk} is an independent sequence;
• (A6) {Hk} are identically distributed with a distribution thatcan be determined from a stopping time problem.
Testing (A5): {Hk} independent
0 10 20 30 40 50
0
0.05
0.1
0.15
0.2
Lag
Auto
Covariance C
oeffic
ient
D=2
D=4
D=8
Figure: Hold state normalized auto-covariances. 5 class 1, 5 class 2stations, D = 2, 4 &8. K = 1700k, 1200k, 850k. ns-2 data
Testing (A6): {Hk} specific distribution
0 2 4 6 8 10 12 14 16
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
i
P(H
=i)
D=2
Sim
Theory
0 10 20 30 40 50 60 70 80
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
i
P(H
=i)
D=12
Sim
Theory
Figure: Hold state distributions, D = 2, 12. ns-2 data.
Testing (A6): {Hk} specific distribution
0 2 4 6 8 10 12 14 16
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
i
P(H
=i)
D=2
Sim
Theory
0 10 20 30 40 50 60 70 80
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
i
P(H
=i)
D=12
Sim
Theory
Figure: Hold state distributions, D = 2, 12. ns-2 data.
Kolmogorov-Smirnov test accepts fit for K of the order 10, 000;
Testing (A6): {Hk} specific distribution
0 2 4 6 8 10 12 14 16
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
i
P(H
=i)
D=2
Sim
Theory
0 10 20 30 40 50 60 70 80
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
i
P(H
=i)
D=12
Sim
Theory
Figure: Hold state distributions, D = 2, 12. ns-2 data.
Kolmogorov-Smirnov test accepts fit for K of the order 10, 000;rejects it for K of the order 1, 000, 000.
What are the 802.11s hypotheses?
Mesh model(s) assume:• Dk is kth inter-departure time.
What are the 802.11s hypotheses?
Mesh model(s) assume:• Dk is kth inter-departure time.Assumptions:
• (A7) {Dk} is an independent sequence;
What are the 802.11s hypotheses?
Mesh model(s) assume:• Dk is kth inter-departure time.Assumptions:
• (A7) {Dk} is an independent sequence;
• (A8) {Dk} are exponentially distributed.
Summary
Assumption Sat. Small buf. Big buf.
(A1) {Ck} indep. X X X
(A2) {Ck} i. dist. X X X/×
(A3) {Qk} indep. - - X/×
(A4) {Qk} i. dist. - - ×
(A5) {Hk} indep. X/× - -
(A6) {Hk} dist. X - -
(A7) {Dk} indep. X X X
(A8) {Dk} exp. dist. × X X
Table: {Ck} collision sequence; {Qk} queue-occupied sequence; {Hk}hold sequence; {Dk} inter-departure time sequence.
Summary
Assumption Sat. Small buf. Big buf.
(A1) {Ck} indep. X X X
(A2) {Ck} i. dist. X X X/×
(A3) {Qk} indep. - - X/×
(A4) {Qk} i. dist. - - ×
(A5) {Hk} indep. X/× - -
(A6) {Hk} dist. X - -
(A7) {Dk} indep. X X X
(A8) {Dk} exp. dist. × X X
Table: {Ck} collision sequence; {Qk} queue-occupied sequence; {Hk}hold sequence; {Dk} inter-departure time sequence.
K. D. Huang, K.D & D. Malone, Tech. Report.(Preliminary report: K. D. Huang, K.D, D. Malone & D. Leith,IEEE PIMRC 2008.)
Epilogue: Impact of erroneous hypotheses?
0 200 400 600 800 1000 1200 1400 1600 1800 20000
10
20
30
40
50
60
70
Network Input (Packets/s)
Indiv
idual T
hro
ughput (P
ackets
/s)
Sim.
Var−q
Const−q
Epilogue: Impact of erroneous hypotheses?
0 200 400 600 800 1000 1200 1400 1600 1800 20000
10
20
30
40
50
60
70
Network Input (Packets/s)
Indiv
idual T
hro
ughput (P
ackets
/s)
Sim.
Var−q
Const−q
0 200 400 600 800 1000 1200 1400 1600 1800 20000
50
100
150
200
250
300
350
400
Network Input (Packets/s)
Ind
ivid
ua
l T
hro
ug
hp
ut
(Pa
cke
ts/s
)
λ1/λ
2=30
Sim. Class1
Sim. Class2
Var−q Class1
Var−q Class2
Const−q Class1
Const−q Class2
Figure: Theory & ns-2 data.
K. D. Huang & K.D, IEEE Comms Letters 2009.
Conclusions
Assumption Sat. Small buf. Big buf.
(A1) {Ck} indep. X X X
(A2) {Ck} i. dist. X X ×
(A3) {Qk} indep. - - ×
(A4) {Qk} i. dist. - - ×
(A5) {Hk} indep. X/× - -
(A6) {Hk} dist. X - -
(A7) {Dk} indep. X X X
(A8) {Dk} exp. dist. × X X
Conclusions
Assumption Sat. Small buf. Big buf.
(A1) {Ck} indep. X X X
(A2) {Ck} i. dist. X X ×
(A3) {Qk} indep. - - ×
(A4) {Qk} i. dist. - - ×
(A5) {Hk} indep. X/× - -
(A6) {Hk} dist. X - -
(A7) {Dk} indep. X X X
(A8) {Dk} exp. dist. × X X
Reports available at:
http://www.hamilton.ie/ken duffy