The validity of IEEE 802.11 MAC modelinghypotheses
David Malone
Joint work with Kaidi Huang and Ken Duffy
Hamilton Institute, National University of Ireland Maynooth
MACOM, Barcelona, September 14th 2010
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Talk outline.
I DCF — the IEEE 802.11 CSMA/CA MAC.
I Mathematical modeling of 802.11 MAC.
I Implicit approximations made to make modeling practical.
I Directly testing these hypotheses with test-bed data.
I Summary, thoughts and conclusions.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
The 802.11 DCF
DataData
SIFS DIFS Decrement counter
Select Random Number in [0,31]
SIFS DIFS
CounterExpires;Transmit
Pause CounterResumeAck Ack
Data
Figure: 802.11 MAC operation (not to scale)
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
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
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Popular mathematical modeling approaches
I 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).
I Mean-field Markov models: seminal work by Bianchi (IEEEComms L. 1998, IEEE JSAC 2000).
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
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.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
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
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
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 0.5p
tau(p)
Figure: Attempt probability τ(p) vs p
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
The self-consistent equation
Network 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-ta
u(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
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Example developments
I Unsaturated 802.11, Small buffer: Ahn, Campbell, Veres andSun, IEEE Trans. Mob. Comp., 2002; Ergen, Varaiya,ACM-Kluwer MONET, 2005; Malone, Duffy, Leith,IEEE/ACM Trans. Network., 2007.
I Unsaturated 802.11, Big buffer: Cantieni, Ni, Barakat andTurletti, Comp. Comm., 2005; Park, Han and Ahn,Telecomm. Sys., 2006; Duffy. and Ganesh, IEEE Comm.Lett., 2007.
I 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.
I 802.11s, unsaturated: Duffy, Leith, Li and Malone, IEEEComm. Lett., 2006.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
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.
Why do these models produce good predictions?Is there a Therom we should know?
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Why is this important?
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
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.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
A first look at the data
0
0.01
0.02
0.03
0.04
0.05
0.06
100 200 300 400 500 600 700 800
Pro
babi
lity
Offered Load (per station, pps, 496B UDP payload)
Average P(col)P(col on 1st tx)
P(col on 2nd tx)1.0/321.0/64
Figure: Collision probability at backoff stages versus load. 2 stations.
Also checked with simulations.David Malone The validity of IEEE 802.11 MAC modeling hypotheses
What are the hypotheses?
Common assumptions to all:• Ck = 1 if kth transmission results in collision.• Ck = 0 if kth transmission results in success.Assumptions:
I (A1) {Ck} is an independent sequence;
I (A2) {Ck} are identically distributed with P(Ck = 1) = p.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Testing (A1): {Ck} independent
0 5 10 15 20−0.1
−0.05
0
0.05
0.1
0.15
0.2
Lag
Aut
oCov
aria
nce
Coe
ffici
ent
Saturated
N=2 λ=750N=5 λ=300N=10 λ=150
Figure: Saturated C1, . . . ,CK normalized auto-covariances. Experimentaldata, N = 2, 5, 10, K = 2500k, 1200k, 711k.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Testing (A1): {Ck} pairwise independent
0 5 10 15 20−0.02
0
0.02
0.04
0.06
0.08
0.1
Lag
Aut
oCov
aria
nce
Coe
ffici
ent
Unsaturated, Big Buffer
N=2 λ=250N=5 λ=100N=10 λ=50
Figure: Unsaturated, big buffer C1, . . . ,CK normalized auto-covariances.Experimental data, N = 2, 5, 10, K = 1800k, 750k, 380k.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
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 .
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
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
Col
lisio
n P
roba
bilit
y
Saturated
N=2 λ=750N=5 λ=300N=10 λ=150Bianchi
Figure: Saturated collision probabilities. Experimental data.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
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
Col
lisio
n P
roba
bilit
y
Unsaturated, Big Buffer
N=2 λ=250N=5 λ=100N=10 λ=50
Figure: Unsaturated, big buffer collision probabilities. Experimental data.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
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:
I (A3) {Qk} is an independent sequence;
I (A4) {Qk} are identically distributed with P(Qk = 1) = q.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Testing (A3): {Qk} pairwise independent
0 5 10 15 200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Lag
Aut
oCov
aria
nce
Coe
ffici
ent
Unsaturated, Big Buffer
N=2 λ=250N=5 λ=100N=10 λ=50
Figure: Unsaturated, big buffer queue-non-empty sequence normalizedauto-covariances. Experimental data. K = 1700k, 720k, 360k.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
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=250N=5 λ=100N=10 λ=50
Figure: Unsaturated, big buffer queue-non-empty probabilities.Experimental data. (Note the large y-range!)
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
What about 802.11e?
DataData
SIFS DIFS Decrement counter
Select Random Number in [0,31]
SIFS DIFS
CounterExpires;Transmit
Pause CounterResumeAck Ack
Data
Figure: 802.11 MAC operation (not to scale)
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
What are the 802.11e hypotheses?
Models with different AIFS values:• Hk is length of kth period we spend in hold-states.Assumptions:
I (A5) {Hk} is an independent sequence;
I (A6) {Hk} are identically distributed and if we know silenceprobability distribution can be determined from Markov chain.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Testing (A5): {Hk} pairwise independent
0 10 20 30 40 50
0
0.05
0.1
0.15
0.2
Lag
Aut
oCov
aria
nce
Coe
ffici
ent
D=2D=4D=8
Figure: Hold state normalized auto-covariances. 5 class 1, 5 class 2stations, D = 2, 4 &8. K = 1700k, 1200k, 850k. ns-2 data
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Testing (A6): {Hk} specific distribution
0 2 4 6 8 10 12 14 160
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
i
P(H
=i)
D=2
SimTheory
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
SimTheory
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.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
What are the 802.11s hypotheses?
Mesh model(s) assume:• Dk is kth inter-departure time.Assumptions:
I (A7) {Dk} is an independent sequence;
I (A8) {Dk} are exponentially distributed.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Testing (A7): {Dk} pairwise independent
0 5 10 15 20−0.1
−0.05
0
0.05
0.1
0.15
0.2
Lag
Aut
oCov
aria
nce
Coe
ffici
ent
Unsaturated, Small Buffer
N=2 λ=400N=5 λ=160N=10 λ=80
Figure: Inter-departure time normalized auto-covariances. Experimentaldata data
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Testing (A8): {Dk} exponentially distributed
0 2 4 6 8 10 12 14x 104
10−6
10−5
10−4
10−3
10−2
10−1
100
Inter−departure Time, t (µs)
P(D
>t)
Unsaturated, Big Buffer, N=5 !=100
Experimental DataTheoretical Data
0 1 2 3 4 5 6x 104
10−6
10−5
10−4
10−3
10−2
10−1
100
Inter−departure Time, t (µs)
P(D
>t)
Saturated, N=5 !=300
Experimental DataTheoretical Data
Figure: Inter-departure time distribution. 5 stations, small buffer. Lowload, Big Biffer and Saturated. Experimental data
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Summary
Assumption Sat. Small buf. Big buf.
(A1) {Ck} indep. X X X(A2) {Ck} i. dist. 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. × light load light load
Table: {Ck} collision sequence; {Qk} queue-occupied sequence; {Hk}hold sequence; {Dk} inter-departure time sequence.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
What to do?
I Collision probability assumption pretty good.I Full Markov chain?
I Modeling variable queue more tractable.I Arrival process structure.I Can also build queue into Markov chain.
R.P. Liu, G.J. Sutton, I.B. Collings, IEEE TWC, To Appear.
I 11e assumptions look OK, for moderate AIFS.I More specialized.
I When network is busy Poisson not that good.I Insensitive to distribution?
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Impact of incorrect hypotheses?
0 200 400 600 800 1000 1200 1400 1600 1800 20000
10
20
30
40
50
60
70
Network Input (Packets/s)
Indi
vidu
al T
hrou
ghpu
t (P
acke
ts/s
)
Sim.Var−qConst−q
0 200 400 600 800 1000 1200 1400 1600 1800 20000
50
100
150
200
250
300
350
400
Network Input (Packets/s)
Indi
vidu
al T
hrou
ghpu
t (P
acke
ts/s
)
λ1/λ
2=30
Sim. Class1Sim. Class2Var−q Class1Var−q Class2Const−q Class1Const−q Class2
Figure: Theory & ns-2 data.
K.D. Huang & K.R. Duffy IEEE Comms Letters 2009.
David Malone The validity of IEEE 802.11 MAC modeling hypotheses
Conclusions
I Some of our assumptions are good,
I Some are not so good,
I Our results are usually good, but not always.
I Possible to provide any analysis?
I Other assumptions: slottedness and channel.
Thanks! Questions?
David Malone The validity of IEEE 802.11 MAC modeling hypotheses