You are on page 1of 8

A Study of Spatial Packet Loss Correlation in

802.11 Wireless Networks


Zhe Wang, Mahbub Hassan Tim Moors
School of Computer Science and Engineering School of Electrical Engineering and Telecommunications
University of New South Wales, Sydney 2052, Australia University of New South Wales, Sydney 2052, Australia
Email: {zhewang, mahbub}@cse.unsw.edu.au Email: t.moors@unsw.edu.au

Abstract—This paper examines the spatial correlation of loss between receivers is “roughly independent” but did not
packet loss events in IEEE 802.11 wireless networks for broad- present quantitative data about the loss correlation. Salyers
cast communications. We discuss limitations of previously used et al [3] presented the results of a small range (distances of
metrics to measure spatial loss correlation and show that the
entropy correlation coefficient, which is based on the mutual 3m∼12m and packet loss rate of 2%∼10%) experiment, and
information concept of Information Theory, can overcome these found, using conditional loss probability as a measure of loss
limitations. In our experiments, we find that the packet losses correlation, that the loss correlation increases with distance.
among closely located receivers are highly correlated when signal However, as we will demonstrate later in this paper, condi-
strength is high, but the correlation decreases with decreasing tional loss probability alone cannot truly capture the statistical
signal strength. The losses become totally independent once the
signal strength falls below a threshold. As a first step to model the dependence among the losses observed at different receivers,
relationship between spatial loss correlation and signal strength, hence may invalidate some of the conclusions reported in [3].
we find that the relationship can be approximated as a Gaussian Clearly, further studies are needed to explore the issue of
cumulative distribution function. spatial packet loss correlation in wireless networks. Given
that in many applications, e.g., 802.11-based vehicular safety
I. I NTRODUCTION
communication, the packet loss rate is up to 40% due to
Packet loss is an important issue for IEEE 802.11 wireless distance between the packet sender and receivers can be
networks. A great amount of research has been done to study hundreds of meters ( [6], [7]), it is important to characterise
the average packet loss rate, or the packet loss probability. spatial loss correlation for a wider range of distances than
However, a more fine-grained study of the issue of packet loss reported in [3]. In this paper, we study the effect of distance
correlation also attracts the interest of the many researchers on packet loss correlation1 through a series of experiments.
[1]–[3]. We make the following key contributions:
There are two kinds of packet loss correlation, namely • We discuss the limitations of the metrics used to measure
temporal correlation and spatial correlation. The temporal loss correlation in previous works and propose a new
loss correlation measures the packet loss pattern over time. metric to overcome these limitations.
It is often related to the issue of consecutive packet losses or • We conduct a series of experiments representing a wide
the burstiness of packet loss by a single receiver. However, for range of distances between the transmitter and the re-
multicast/broadcast communication, in which a single packet ceivers.
has multiple receivers, it is also useful to measure the correla- • Applying the proposed metric, we find that the spatial
tion of losses by different receivers. We focus on such spatial packet loss correlation in our 802.11 network experiments
loss correlation between receivers of a single transmission. decreases with the distance between sender and receivers.
The spatial correlation between receivers is important for • We take a first step to analytically model the loss corre-
its effect on the performance of error recovery protocols. If lation. We find that with signal strength as the indepen-
there is low correlation in loss between receivers that are near dent variable, the proposed loss correlation metric can
each other, then when one receiver loses a packet, it may be approximated as a Gaussian cumulative distribution
be able to recover the lost packet from a nearby receiver. function.
Thus, local retransmission is more effective when losses are
The rest of this paper is organized as follows: We review the
less spatially correlated. The extent to which losses among
related work in Section II. The limitation of previously used
multiple receivers are disjoint also influences error recovery
metrics to measure spatial loss correlation are discussed and
when network coding is used to combine (e.g., XOR) multiple
the proposed metric is presented in Section III. We describe
lost packets into a single retransmission [4], [5].
our experiments and discuss the findings in Section IV. We
Despite the importance of spatial loss correlation in wireless
conclude in Section V.
networks, not much is known about this issue. In open
literature, the 802.11 spatial loss correlation has been reported 1 In the rest of this paper, the term loss correlation specifically refers to the
in [2] and [3]. Reis et al [2] stated qualitatively that the spatial packet loss correlation
II. R ELATED W ORK conditional loss probability P (A|B) and unconditional loss
Miu et al. [1] have investigated whether separate 802.11 probability P (A), satisfies this property. On one hand, if
transmitters have correlated packet losses when observed at the ratio PP(A|B)
(A) = 1, we can get P (AB)=P (A)P (B),
a single receiver. The metric used in [1] to measure the which means the packet loss is independent (uncorrelated).
correlation is based on the conditional and unconditional loss On the other hand, if the ratio PP(A|B) 1
(A) = P (A) , we can get
probability. If A is the event of the receiver losing a packet P (A|B) = 1, which means A ⊆ B, i.e., the packet loss
from transmitter 1 and B is the event of the receiver losing is 100% correlated. The ratio of the conditional probability
a packet from transmitter 2, then the level of correlation to the unconditional probability, PP(A|B)
(A) , which is bounded
can be measured by comparing the probability P (A) and 1
by [1, P (A) ], could be used to measure the loss correlation.
P (A|B). If P (A) equals P (A|B), then the packet loss of is The higher this ratio, the higher the packet loss is correlated
totally uncorrelated (probabilistic independent). At the other between the two receivers.
extreme, if P (A|B) equals one, then the packet loss is totally Instead of using the conditional loss probability, [9] uses
correlated. The metric used to measure the loss correlation the Pearson correlation coefficient to measure the spatial loss
is the difference between P (A|B) and P (A). The level of correlation in wired multicast networks. The Pearson corre-
correlation is decided by the position of P (A|B) in the range lation coefficient is a measure of linear dependence between
[P (A), 1]. The further P (A|B) is from P (A) and closer to two random variables [8], which ranges from -1 to 1. If the
1, the higher the correlation is. Although the loss correlation coefficient is 1 or -1, the two random variables have perfect
issue we are investigating is different from [1], the metric linear positive or negative relationship (100% correlated).
used in this work can also be applied to the case of one If the coefficient is 0, the two random variables have no
transmitter and two receivers. In our case, for any packet from linear dependence. The coefficient being 0, i.e., no linear
the transmitter, A is the event of receiver 1 losing the packet dependence, does not guarantee the statistical independence.
and B is the event of receiver 2 losing the packet. In some cases, the two random variables (X and Y ) are
Reis et al [2] observed that the spatial loss correlation dependent, e.g., Y = X 2 while the coefficient is 0. Therefore,
of 802.11 network in a office building is low. However, the Pearson correlation coefficient is inadequate to measure the
neither the metric used to measure the correlation nor the loss statistical dependence of the packet loss between two receivers.
correlation data is shown in the paper. The second property that the loss correlation metric should
The most related work to our study is [3], in which have is that given the packet loss events (A and B) of two
the authors study the loss correlation for different 802.11 receivers, the metric (ρ) should have a unique value no matter
wireless cards at different distances between the sender and which receiver is chosen as A or B, i.e., ρAB =ρBA . The ratio
receivers. They measure the correlation using the conditional of conditional loss probability over unconditional loss proba-
loss probability alone, instead of using the difference between bility does not satisfy this property when PP(A|B) P (B|A)
(A) = P (B) .
the conditional and unconditional loss probabilities as in [1]. By using conditional/unconditional loss probability ratio as
Using this metric, the authors report that the loss correlation the metric, we might have two possible measures for the loss
increases when distance increased from 3m to 12m. correlation between the same pair of receivers.
In summary, previous works have measured the 802.11 The third property the loss correlation metric should have
packet loss correlation using different metrics. Before we start is that the metric should have fixed upper and lower bounds.
our study on how the loss correlation changes with the distance Without this property, it is hard to compare the level of
between sender and receivers, it is necessary to choose an loss correlation between two different pairs of receivers. For
appropriate metric to numerically measure the loss correlation. example, assume that we have two pairs of receivers (R1,
III. METRIC FOR SPATIAL LOSS CORRELATION R2), and (R1 , R2 ). For (R1, R2), the packet loss events
are A and B respectively, and for (R1 , R2 ), the packet
In this section, we propose a metric to measure the loss
loss events are A and B  respectively. By using the ratio of
correlation. We start by discussing the properties that the
unconditional/conditional loss probabilities as the metric, we
metric should have.
cannot compare the level of loss correlation of these two pairs
The first property that the loss correlation metric should P (A |B  )
if P (A) = P (A ) because the ratios: PP(A|B)
(A) and P (A )
have is that it should indicate the statistical dependence of 1 1
the packet loss between two receivers. In probability theory, have different upper bounds ( P (A) and P (A ) ). Suppose that
the definition of two events A and B being independent the packet loss of both these two pairs are 100% correlated
is: P (AB)=P (A)P (B) [8]. Using the conditional proba- (they have the same correlation level), comparing PP(A|B) (A)
 
bility P (A|B) alone to measure the loss correlation as in 1
(equals P (A) in this case) and P P(A(A|B
)
) 1
(equals P (A  ) in this
[3] is problematic because only when P (A|B)=P (A), i.e., case) will lead to the wrong conclusion that the loss correlation
P (AB)
P (B) =P (A), can we get P (AB)=P (A)P (B). The con- of these two pairs are different.
ditional probability P (A|B) does not indicate whether the The metrics used in previous works do not have all the three
packet losses of the two receivers are independent without con- properties discussed above. In the following, we propose a new
sidering its relationship with unconditional probability P (A). metric that exhibits all these three properties.
The metric used in [1], which is the relationship between The entropy correlation coefficient [10] is a metric to
Fig. 1: The experiment setup (to observe a large range of packet loss rates, the external antenna of sender node is not mounted).

measure the dependency of two random variables, which is


based on the entropy concept of information theory. If we
25
have two random variables X and Y , the entropy correlation
coefficient (ρXY ) is defined as:

2IXY 20
ρXY = (1)
HX + H Y

RSSI
where HX and HY are the entropies of X and Y respectively,
15
and IXY is the mutual information of X and Y . If we define
the random variables X and Y as: X = 0 if a packet is
received by receiver R1 , X = 1 if the packet is lost by R1 ;
Y = 0 if the packet is received by receiver R2 , Y = 1 if the 10
Receiver 1
Receiver 2
packet is lost by R2 , then we have:
4 6 8 10 12 14 16 18
 Transmission power (dBm)
HX = − P (X = x) log2 P (X = x)
x=0,1 Fig. 2: The average RSSI of the two receivers


HY = − P (Y = y) log2 P (Y = y) IV. E XPERIMENT
y=0,1
A. Experiment Setup
Our experiment setup consists of 3 Yawarra ALIX 3 [11]
1 
 1
P (X = x, Y = y) single board computers. Each computer is equipped with a
IXY = P (X = x, Y = y) log2 wireless card with the Atheros AR5112 chipset. The wireless
x=0 y=0
P (X = x)P (Y = y)
cards are driven by the Madwifi driver [12] with ath hal
The entropy correlation coefficient (ρXY ) in (1) is the measure version 0.9.18.0 and ath pci version 0.9.4.5. Because in the
of dependency of the packet loss of R1 and R2 and it has the experiment environment there are WLAN networks operating
following properties: in 802.11 b/g mode in 2.4GHz band, the wireless cards are
set to 802.11a mode in 5.18GHz to avoid interference from
• ρXY =ρY X , it has the same value no matter which re- the existing systems.
ceiver is chosen as X and which is chosen as Y . Placement of the three computers are shown in Fig. 1.
• 0 ≤ ρXY ≤ 1, the upper and lower bounds are constants. The transmitting node is placed 4 meters away from the two
• ρXY =1 if and only if X and Y are completely dependent, receivers, which are themselves separated by 5cm 2 . There is
i.e., if we know the value of X, we can predict the value a clear line of sight between the transmitter and receivers.
of Y with certainty. We set the packet length to 100 bytes and the bit rate to
• ρXY =0 if and only if X and Y are independent, i.e., 6Mbps, which is same as a typical vehicular safety application
knowing the value of X does not help at all to predict of 802.11 broadcast [13]. In our experiments, we do not change
the value of Y . the placements of the nodes. To study the effect of transmitter-
In the following section, we conduct a series of experiments receiver distance, which affects the received signal strength at
to study the loss correlation between two closely located
2 In [3], the two receivers are separated by less than the wavelength of the
receivers. By measuring the loss correlation using different
radio signal. To compare with their results, we follow similar setup, where the
metrics, we show that the entropy correlation coefficient is a inter-receiver distance of 5cm is less than the wavelength of 5.18GHz radio,
better metric than the ones used in previous works. which is 5.8cm.
1 1

0.1 0.1
Packet loss probability

Packet loss probability


0.01 0.01

P(A) P(B)
P(A|B) P(B|A)
0.001 0.001
4 6 8 10 12 14 16 18 4 6 8 10 12 14 16 18
Transmission power (dBm) Transmission power (dBm)

(a) Packet loss probability of receiver 1 (in log-scale) (b) Packet loss probability of receiver 2 (in log-scale)

Fig. 3: The conditional and unconditional loss probabilities of the two receivers. P (A) and P (B) are the unconditional loss
probabilities of receiver 1 and 2 respectively. P (A|B) and P (B|A) are the conditional probabilities of a receiver losing
a packet given that the other receiver loses the packet.

1000
the receivers, we repeat the experiment with different trans- P(A|B)/P(A)
P(B|A)/P(B)
mission power levels. Starting from the maximum transmission
power (17 dBm) allowed by the wireless card, the transmission
Conditional/unconditional loss probability ratio

power is reduced by 1dBm for each experiment until 5 dBm.


For each transmission power, the transmitting node broadcasts 100

20,000 packets with 20 ms gap between two consecutive


packets. In order to observe a large range of packet loss rates,
the external antenna is not mounted on the transmitting node
(as shown in Fig. 1). The receiving nodes are set to work in 10
monitor mode so that the Receiving Signal Strength Indicator
(RSSI) is included in the captured packet traces.
Although the two receivers are very close to each other,
which makes the path loss and channel fading have the same
1
effect on them, they have different loss rates. By analyzing the 6 8 10 12 14 16
Transmission power (dBm)
RSSI of received packets as shown in Fig. 2, we find that the
two receivers have slightly different signal to noise ratios (In Fig. 4: The ratio of conditional/unconditional loss probability
Madwifi, the reported RSSI for each packet is equivalent to (in log-scale).
the signal to noise ratio [14]). The average RSSI of receiver 1
is always lower than that of receiver 2. The difference remains
the same even when we swap the locations and the external abilities are “V” shape curves. When the transmission power
antennas. This might be caused by the variation of analog part is low, the conditional loss probability is close to one and
of the hardware. decreases with the transmission power until it reaches the
lowest point of the “V” shape and starts increasing with
B. Conditional/Unconditional Loss Probability
the transmission power until it reaches one. Fig. 3 clearly
The conditional and unconditional packet loss probabilities shows that the conditional loss probability alone cannot
at different transmission powers are shown in Fig. 3. Note that, capture the statistical dependence among the losses of the
the transmission power range we applied increased uncondi- two receivers. For example, when the transmission power is
tional loss probability by about 1000 folds from 0.001 at 17 8dBm, P (A)=0.93 and P (A|B)=0.93; when the transmission
dBm to 0.997 at 5 dBm. This would reflect a large variation in power is 14dBm, P (A)=0.009 and P (A|B)=0.93. Although
the transmitter-receiver distance for a fixed transmission power the conditional loss probabilities are the same, the statistical
(In [3], the unconditional probability changes from 0.02 to dependence of packet loss of these two cases is fairly different.
0.10, an increase of about 5 fold, when the distance increases For 8dBm, we have P (A|B)=P (A), suggesting that there is
from 3m to 12m). little dependency, whereas for the case of 14dBm, P (A|B) is
While the unconditional loss probabilities monotonically much larger than P (A), suggesting that the packet losses at
decrease with transmission power, the conditional loss prob- two receivers are highly dependent.
0.6 0.6
txpower=14dBm
txpower=16dBm
txpower=10dBm
txpower=12dBm μ0=21.2 txpower=10dBm txpower=12dBm txpower=16dBm
μ0=17.4 μ0=22.5 μ0=18.4 μ0=20.9 μ0=24.4
μ0=19.2 σ=0.68
0.5 σ=0.77 σ=0.70 0.5 σ=0.82 σ=0.79 σ=0.79
σ=0.79 txpower=14dBm
μ0=23.0
σ=0.90
Probability density function

Probability density function


0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
16 18 20 22 24 26 16 18 20 22 24 26
RSSI RSSI

(a) Receiver 1 (b) Receiver 2

Fig. 6: The probability density of RSSI

(1). The results are shown in Fig. 5 (along with the packet
1 reception probabilities at the receivers). We can see that
the loss correlation decreases monotonically with decreasing
0.8 transmission power. When the transmission power is high, i.e.,
above 16dBm, the entropy correlation coefficient is one, which
Reception Probability (R1)
Reception Probability (R2)
means that the two receivers lose the same set of packets and
0.6
if one receiver loses (or receives) a packet, we know the other
also loses (or receives) it. Loss correlation starts to decrease as
0.4 transmission power is reduced. When the transmission power
Entropy Correlation Coefficient
is less than 8dBm, the entropy correlation coefficient is almost
0.2
zero, which means that the two receivers independently lose
packets and the knowledge of one receiver losing (receiving) a
packet is not useful to predict the loss (reception) at the other
0
4 6 8 10 12 14 16 18 receiver.
Transmission power (dBm)
Another interesting observation we make from Fig. 5 is
Fig. 5: The entropy correlation coefficient and packet recep- that the curves of both the entropy correlation coefficient and
tion probabilities (1-P (A) and 1-P (B)) the packet reception probabilities have the “S” shape, albeit
the centre points and steepness are different. This observation
suggests that we may be able to model loss correlation in a
Fig. 3 also shows that the difference between the con- similar fashion the packet reception probability is modelled.
ditional loss probability and unconditional loss probabil- We first discuss the possible reasons for the packet reception
ity increases with transmission power. The ratio of condi- probabilities to exhibit the “S” shape.
tional/unconditional loss probability is plotted in Fig. 4. Using
It is known that, if the signal strength attenuates according
the ratio as the loss correlation metric, we can see from
to log-normal shadowing model, then the “S” shaped Gaussian
Fig. 4 that the loss correlation monotonically increases with
cumulative distribution function (CDF) can be used to deter-
transmission power. Fig. 4 also shows that the two ratios:
P (A|B) P (B|A) mine the probability that the received signal strength (RSS) is
P (A) and P (B) are not always equal. For example, when greater than a threshold R (Rth ) [15]. If our experimental radio
P (B|A)
the transmission power is 15dBm, P (B) is 1.5 times of environment is characterised by log-normal shadowing model,
P (A|B) and if we assume that packets are received correctly only when
P (A) . Using the conditional/unconditional loss probability
ratio as a metric to measure dependency, there would be two the RSS exceeds a threshold (this assumption is used widely
different measures for the loss correlation between the same in the literature and simulation tools [16]), the packet recep-
pair of receivers. tion probability can be approximated by Gaussian CDF. To
investigated the radio model of our experimental environment,
C. Entropy Correlation Coefficient we plot the RSSI distribution for different transmission power
From the packet trace collected by the two receivers, we in Fig. 6. Log-normal shadowing implies that RSS at specific
derive the joint distribution of X and Y defined in Section transmitter-receiver separation is Gaussian (normal) distributed
III and calculate the entropy correlation coefficient using with the distance dependent mean and a constant standard
1
deviation [15]. In Fig. 6, we find that the probability density μ=0, σ=0.5
μ=0, σ=1
of RSSI is indeed Gaussian shaped, the mean is changing μ=0, σ=2
μ=-2, σ=1
with transmission power (distance) and the standard deviation 0.8
remains relatively stable. Such results indicate that the log-
normal shadowing model is a good approximation of the
received signal strength in our experiments (note that the unit 0.6

Φμ,σ(X)
of μ0 and σ in Fig. 6 is not dBm since RSSI is defined as
an arbitrary integer value the 802.11 standard [17], and the
0.4
precision of the integer is vendor dependent [18]).
According to the log-normal shadowing model [15], the
received signal strength (R) can be expressed as: 0.2

R =T −L+N (2)
0
where T is the transmission signal strength, L is the path loss -4 -2 0 2 4
X
and N is a Gaussian random variable with zero mean and
standard deviation σ (R, T and σ are measured in dBm). Given Fig. 7: The Φμ,σ (X) function with different parameters
that T and L are static variables, R is Gaussian distributed
with mean μ0 = T − L and standard deviation σ.
Based on (2), if we assume that there is a threshold of the curve of entropy correlation coefficient. Again using the
R(Rth ) above which the packet will be successfully received, curve fitting tool [19], we fit the entropy correlation coefficient
the packet reception probability (r) can be derived as: to function Φμ,σ (X). Fig. 9 shows that, like the packet
reception probability, the entropy correlation coefficient also
r = P (R > Rth )
has a good fit to Gaussian CDF except the fact that the values
= P (N > Rth − T + L) of both its parameters, μ and σ, are larger than those of the
= P (N < T − Rth − L) packet reception probability. How to derive the parameters μ
and σ of the Gaussian CDF that approximates spatial loss
Since N is Gaussian distributed with mean 0 and standard correlation remains a topic of ongoing research.
deviation σ, we get:
V. C ONCLUSION AND F UTURE W ORKS
1 T −μ
r= (1 + erf ( √ )) (3) Using indoor experiments, we have studied the spatial cor-
2 2σ 2
relation of packet loss in IEEE 802.11 wireless networks when
where a packet transmitted by a single transmitter is to be received
μ = Rth + L by multiple receivers. We have found that the metric used
 z to measure the spatial loss correlation is an important factor.
2
erf (z) = √ exp(−x2 ) dx. Specifically, we have shown that simply using conditional loss
π 0
probabilities to measure loss correlation can lead to misleading
Equation (3) has the same form as the CDF (Φ-function) of a conclusions. We have proposed and demonstrated that the
Gaussian random variable with mean μ and standard deviation entropy correlation coefficient, which is based on the mutual
σ: information concept of Information Theory, is a good metric
1 X −μ
Φμ,σ (X) = (1 + erf ( √ )) (4) to measure spatial packet loss correlation in wireless networks.
2 2σ 2 Using this metric, we have discovered that the received signal
The Φμ,σ (X) function with different parameters values are strength (distance between the transmitter and the receivers)
plotted in Fig. 7. It can be seen that parameter μ controls the is a key factor that affects the spatial loss correlation. In our
position of the center point of the curve (increasing μ will experiments, the packet losses among closely located receivers
shift the center point rightwards). Parameter σ controls the are highly correlated when the signal is strong, but the corre-
steepness of the curve (increasing σ will make the curve less lation falls with decreasing signal strength. The losses become
steep). totally independent once the signal strength falls below a
Using the curve fitting tool [19], we fit the packet reception threshold. Our finding suggests that the decay of entropy
probability from the experiment to the function Φμ,σ (X) with correlation coefficient can be approximated as a Gaussian
unknown parameters μ and σ. The curve fitting results in Fig. CDF, the same function that models the packet reception
8 shows that the packet loss rate at different transmission probability at the receivers in a log-normal shadowing radio
powers can be accurately modeled by Φμ,σ (X) function with propagation environment. This finding serves as a first step to
parameters μ and σ being slightly different for different analytically model the spatial packet loss correlation.
receivers. In future work, it will be useful to conduct more experiments
Now that we have reasonable clues why packet reception to investigate the impact of 802.11 mode (a/b/g), data rate,
probability follows a Gaussian CDF, we proceed to examine packet size, and the distance between the two receivers, on the
1 Experiment data 1
Experiment Data
0.9 Φμ,σ(X) 0.9 Φμ,σ(X)

0.8 0.8
Packet Reception Probability

Packet Reception Probability


0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0

6 8 10 12 14 16 6 8 10 12 14 16
Transmission Power (dBm) Transmission Power (dBm)

(a) Receiver 1: (b) Receiver 2:


μ= 9.65 ( 95% confidence interval: (9.526, 9.777)) μ= 8.94 ( 95% confidence interval: (8.812, 9.067))
σ=1.15 ( 95% confidence interval: (0.973, 1.328)) σ=0.957 ( 95% confidence interval: (0.776, 1.138))
SSE: 0.0088 SSE: 0.0109
R-square: 0.9963 R-square: 0.9953
RMSE: 0.02825 RMSE: 0.03149

Fig. 8: Φμ,σ (X) approximation of packet reception probability.

R EFERENCES
1
Experiment data
[1] A. Miu, G. Tan, H. Balakrishnan, and J. Apostolopoulos, “Divert: fine-
0.9 Φμ,σ(X)
grained path selection for wireless lans,” in Proceedings of the 2nd
0.8 international conference on Mobile systems, applications, and services.
Entrophy Correlation Coefficient

Boston, MA, USA: ACM, 2004.


0.7
[2] C. Reis, R. Mahajan, M. Rodrig, D. Wetherall, and J. Zahorjan,
0.6 “Measurement-based models of delivery and interference in static wire-
less networks,” in Proceedings of the 2006 conference on Applications,
0.5
technologies, architectures, and protocols for computer communications.
0.4 Pisa, Italy: ACM, 2006.
0.3
[3] D. Salyers, A. Striegel, and C. Poellabauer, “Wireless reliability: Re-
thinking 802.11 packet loss,” in International Symposium on a World
0.2 of Wireless, Mobile and Multimedia Networks, WoWMoM 2008., June
0.1
2008, pp. 1–4.
[4] D. Nguyen, T. Tuan, N. Thinh, and B. Bose, “Wireless broadcast using
0 network coding,” IEEE Transactions on Vehicular Technology, vol. 58,
6 8 10 12 14 16 no. 2, pp. 914–925, 2009.
Transmission Power (dBm) [5] Z. Wang, M. Hassan, and T. Moors, “Efficient loss recovery using
network coding in vehicular safety communication,” in IEEE Wireless
Fig. 9: Φμ,σ (X) curve fitting of entropy correlation coefficient Communications and Networking Conference, WCNC 2010, Sydney,
μ=13.75 ( 95% confidence interval: [13.48, 14.02]) Australia, 2010.
[6] D. Jiang, T. Vikas, M. Andreas, H. Wieland, and H. Ralf, “Design of
σ=1.629, ( 95% confidence interval: [1.25, 2.01]) 5.9Ghz DSRC-based vehicular safety communication,” IEEE Wireless
SSE: 0.028 Communications, vol. 13, no. 5, pp. 36–43, 2006.
R-square: 0.985 [7] B. Fan and H. Krishnan, “Reliability Analysis of DSRC Wireless
Communication for Vehicle Safety Applications,” in Intelligent Trans-
RMSE: 0.0504 portation Systems Conference, 2006. ITSC ’06. IEEE, 2006, pp. 355–
362.
[8] J. L. Devore, Probability and statistics for engineering and the sciences,
7th ed. Belmont, 2008.
[9] M. Yajnik, J. Kurose, and D. Towsley, “Packet loss correlation in the
MBone multicast network,” in Global Telecommunications Conference,
spatial packet loss correlation. These experiments may provide 1996. GLOBECOM ’96., 1996, pp. 94–99.
useful insight into the factors that affect the key features of [10] I. Virtanen and J. Astola, “On entropy-based dependence measures
the Gaussian CDF, i.e., the centre point and the steepness of for two and three dimensional categorical variable distributions,” in
Proceedings of The First World Congress of Bernoulli Society for
the curve. In addition to studying the loss correlation in the Mathematical Statistics and Probability, Tashkent, USSR, 1986.
“interference free” scenario, it may be worth considering the [11] Yawarra ALIX 3. [Online]. Available: http://www.yawarra.com.au/hw-
case when packet losses are also caused by collisions, which alix3.php
[12] The MadWifi project. [Online]. Available: http://madwifi-project.org/
is common for applications where every node is a potential [13] T. ElBatt, K. G. Siddhartha, H. Gavin, K. Hariharan, and P. Jayendra,
broadcaster (such as vehicular safety communication). “Cooperative collision warning using dedicated short range wireless
communications,” in Proceedings of the 3rd international workshop on [17] Part 11: Wireless LAN Medium Access Control (MAC) and Physical
Vehicular ad hoc networks, VANET06. ACM Press, 2006. Layer (PHY) Specifications, IEEE Std. 802.11, 1999.
[14] RSSI in Madwifi. [Online]. Available: http://madwifi- [18] The truth about 802.11 signal and noise metrics. [Online]. Available:
project.org/wiki/UserDocs/RSSI http://www.connect802.com/download/techpubs/2004/you believe D100201.pdf
[15] T. Rappaport, Wireless communications : principles and practice, [19] Curve fitting toolbox 2.2. The MathWorks, Inc. [Online]. Available:
2nd ed. Prentice Hall, 2002. http://www.mathworks.com/products/curvefitting/
[16] The official NS-2 webpage. [Online]. Available:
http://nsnam.isi.edu/nsnam/index.php/Main Page