662 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO.

2, FEBRUARY 2010

Toward Network Coding-Based Protocols for

Data Broadcasting in Wireless Ad Hoc Networks
Alfred Asterjadhi, Student Member, IEEE, Elena Fasolo, Michele Rossi, Member, IEEE,
Joerg Widmer, Member, IEEE, and Michele Zorzi, Fellow, IEEE

Abstract—In this paper we consider practical dissemination settings. This is very important as linear random coding is
algorithms exploiting network coding for data broadcasting in lightweight and inherently localized and, as such, can be
ad hoc wireless networks. For an efcient design, we analyze exploited by communication protocols at low overhead.
issues related to the use of network coding in realistic network
scenarios. In detail, we quantify the impact of random access In this paper, we consider the random network coding
schemes, as used by IEEE 802.11, on the performance of network schemes of [9] for all-to-all data transmission. According to
coding. In such scenarios, deadlock situations may occur where these algorithms, whenever an innovative packet is received at
the delivery process stops and some of the nodes never gather a given node, it generates with probability 𝜌 (the forwarding
the required packets. To tackle this problem, we propose a factor) a new packet through RLNC and broadcasts it over the
proactive mechanism (called proactive network coding) which
adapts its transmission schedule according to the decoding status channel. For the medium access control (MAC) we consider
of neighboring nodes. This scheme can detect when nodes need several variants of carrier sense multiple access (CSMA). We
additional packets in order to decode and acts accordingly. focus on the interaction between MAC and network coding
We nally investigate the behavior of network coding schemes over different wireless network congurations in order to
in multi-rate environments, where we propose a distributed capture the effects of each protocol component and quantify
heuristic approach for the selection of data rates.
the performance degradation due to packet collisions and
Index Terms—Wireless ad hoc networks, network coding, data random transmission schedules. Subsequently, based on our
dissemination, protocol design. performance evaluation we design a proactive dissemination
mechanism and we show that it outperforms the reactive
I. I NTRODUCTION schemes in [9]. The main contributions of this paper are:
∙ We study the performance of random network coding in

N ETWORK coding is a recently introduced paradigm to

efciently disseminate data in wireless networks, where
data ows coming from multiple sources are combined to
selected topologies (circular, grid, and random networks).
In particular, we assess the impact of 1) packet collisions
and 2) the random mixing of ows as dictated by random
increase throughput, reduce delay, and enhance robustness. access at the MAC layer.
In contrast to the traditional store and forward approach [1], ∙ We identify deadlock situations where the delivery pro-
it implements a store, code, and forward technique, where cess prematurely stops due to the lack of further inno-
each node stores incoming packets in its own buffer and trans- vative packets and, as a result, some of the nodes never
mits their combinations, where combining is performed over fully decode.
some nite Galois eld. This technique allows for increased ∙ We propose a lightweight and decentralized combination
throughput efciency as well as scalability and robustness [2]. and transmission technique, called proactive network
These benets arise in the case of multicasting [3], [4] as well coding, to mitigate the deadlock problem in multi-hop
as for other network congurations, such as multiple unicast networks.
communications [5], [6]. Moreover, they are not restricted to ∙ We evaluate the impact of the forwarding factor 𝜌 as well
error-free communication networks, but can also be obtained as the selection of the transmission rate in multi-rate ad
in ad hoc networks [7]–[9], peer-to-peer systems [10], and hoc networks.
optical networks. Important theoretical results are known, see,
Related work on communication protocols for wireless
e.g., [3], [11]–[14], and research is now moving towards the
networks can be found in [5], [6], [8], [9], [15]–[17]. [15]
exploitation of network coding in practical communication
was the rst contribution to present a practical and distributed
protocols. To this end, the work in [3] is of considerable
solution exploiting RLNC. The authors focused on how the
importance as it demonstrates that random linear network
coding matrix as well as the information related to the random
coding (RLNC) is able to reach network capacity in practical
combination of packets in some nite Galois eld 𝐺𝐹 (𝑞)
Manuscript received August 8, 2008; revised June 13, 2009, November 24 can be shared by different nodes at low overhead. This is a
and 27, 2009; accepted November 27, 2009. The associate editor coordinating crucial aspect for network coding algorithms to work in multi-
the review of this paper and approving it for publication was S. Hanly. hop radio networks. The scheme of [16] jointly considers
A. Asterjadhi (corresponding author), M. Rossi, and M. Zorzi are with
the Department of Information Engineering, DEI, University of Padova, via packet combinations with ARQ strategies for wireless sensor
Gradenigo 6/B – 35131, Padova, Italy (e-mail: [email protected]). networks; the transmission paradigm as well as the node
E. Fasolo is with IKS s.r.l., Corso Stati Uniti 14 bis – 35127, Padova, Italy. constraints are however different from those in wireless ad
J. Widmer is with DOCOMO Euro-Labs, Landsberger Strasse 312 – 80687,
Munich, Germany. hoc networks, which is the focus of this paper. COPE [5]
Digital Object Identier 10.1109/TWC.2010.02.081057 applies network coding to unicast ows in wireless networks.
c 2010 IEEE
1536-1276/10$25.00 ⃝

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.
 ASTERJADHI et al.: TOWARD NETWORK CODING-BASED PROTOCOLS FOR DATA BROADCASTING IN WIRELESS AD HOC NETWORKS 663

The authors of the paper experimentally show that signicant II. T HE I MPACT OF MAC L AYER ON DATA
gains, in terms of maximum throughput, are possible even D ISSEMINATION VIA N ETWORK C ODING
in the case of unicast transmissions and even when network
coding is implemented through simple XORing of packets In this section we study the behavior of Network Coding
within a single-hop neighborhood (rather than forwarding in realistic environments. Our aim is to address open issues
encoded packets over multiple hops). [6] presents BFLY, a for its application in wireless networks with a realistic MAC
localized network coding protocol which recognizes buttery Layer.
structures in the network to exploit the coding opportunities Problem description: wireless ad hoc networks are severely
they represent. This protocol builds on COPE and also encodes constrained by interference and channel impairments, espe-
packets through XOR operations, but it additionally allows cially in the case of broadcast communication. The use of
the transmission of encoded packets over multiple hops. The traditional access mechanisms such as CSMA-like protocols,
work in [5], [6] present practical communication schemes, when multiple nodes transmit, may suffer from a high number
where [5] quanties the achievable gains from coding when of collisions and dropped packets. Two main factors are to be
an actual MAC layer is used. Our focus is different in terms of taken into account when using network coding in conjunction
network scenario, as we consider an all-to-all communication with an actual MAC, namely 1) collisions and 2) packet
paradigm and encode packets in 𝐺𝐹 (𝑞) with 𝑞 = 256, scheduling. Both collisions and scheduling are the direct
considering more general coding rules. In [8], the authors consequence of the random (CSMA-like) channel access that
investigate the interaction between MAC and network cod- we adopt in this study. Collisions impact the performance as
ing in wireless multi-hop networks, and propose distributed fewer packets are collected; as a consequence it takes longer
and opportunistic scheduling rules for the combination of to obtain full rank decoding matrices at the receivers. Packet
packets in the presence of time-varying fading links. They scheduling refers to the way in which different nodes take
also look at the impact of MAC schedules. However, this turns in transmitting, which is dictated by the MAC rules. The
topic is treated differently from what we do here as packets transmission order is important when network coding is used
at relay nodes are XORed and possible deadlocks in the at higher layers as it inuences the way encoded packets are
data dissemination are not investigated. The authors of [17] created, i.e., which packets are mixed together. In this paper
study the interaction of network coding and MAC, devis- we focus on the analysis of random access schemes as used
ing suitable conict-free transmission schedules (for a given by IEEE 802.11.
connectivity graph) and related off- and on-line algorithms Network topologies: we start our investigation with circu-
for wireless multi-hop networks. However, their strategies lar and grid reference scenarios and then consider random
entail some coordination among nodes which incurs additional topologies. We ensure that all random topologies used in
communication overhead with respect to RLNC. Reference [9] the simulations are connected. To this end, we do a simple
studies all-to-all communication scenarios and introduces a breadth-rst-search of the underlying connectivity graph and
class of lightweight reactive and distributed network coding check if all nodes are visited (a standard procedure to check
protocols based on RLNC, proving the superiority of these for connectivity). The topology is valid (i.e., it is used in
schemes over ooding [18] and epidemic routing [1]. Our the simulations) if a single connected cluster exists and is
present work is the natural continuation of [9], which we discarded otherwise.
complement investigating the performance degradation due to Physical layer: for the results in this paper we have im-
actual MAC schemes and proposing a solution to deadlocks plemented an extended version of the ns2 physical layer
in the dissemination of the information, which may occur for IEEE 802.11b/g which includes packet error rate (PER)
in certain topologies. The performance evaluation that we calculations accounting for modulation, channel effects, and
carry out in this paper is based on simulation results obtained multi-user interference. In detail, the Signal to Interference
using ns2; the relevant simulation code can be downloaded plus Noise Ratio (SINR) is evaluated for each receiving node
from [19]. and for each packet taking into consideration the interference
The remainder of this paper is organized as follows. In generated by nearby transmitters. PERs are obtained from pre-
Section II we quantify the impact of several IEEE 802.11 calculated packet error rate curves. For the channel, we use the
variants on the performance of network coding. In Section III standard ns2 two ray ground propagation model. More details
we design a lightweight and distributed mechanism to perform can be found in [20].
network coding more efciently and show via simulations that MAC protocols: we consider four different MAC protocols
it can signicantly outperform the data dissemination schemes based on CSMA, which is currently the most widely used
of [9]. Other results are given for multi-rate environments medium access mechanism in wireless ad hoc networks.
in Section IV: on the one hand, high data rates are good 1) IEEE 802.11b: is considered as the baseline MAC. We
as they shorten packet transmission times, thereby reducing adopt the basic access provided by IEEE 802.11b that, in the
the collision probability; on the other hand, however, in a broadcast mode, does not use any acknowledgment mecha-
multi-hop scenario high data rates mean that packets have to nism. In case of collision, no retransmission occurs and the
travel more hops to reach the destination. Thus, determining packet is lost.
suitable data rates for each node, so as to obtain good 2) IEEE 802.11b with pseudo broadcast [5]: this scheme
tradeoffs between delay and delivery ratio, is a challenging and is an improvement of the basic IEEE 802.11b, where an
interesting problem for which we propose a heuristic solution acknowledgment mechanism is implemented. A given node
in Section IV. Section V concludes the paper. randomly picks a neighbor and sends a packet to it via

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.
 664 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 2, FEBRUARY 2010

unicast by including its address in the packet header. All no collisions) assure a packet delivery ratio of 1 when the
other neighbors are in promiscuous mode and can thus over- number of neighbors is 2 (the packet delivery ratio is dened
hear/decode the transmission. However, only the neighbor in Section II-A).
which is the intended receiver of the unicast sends back an 2) Semi-deterministic network coding: in this case, for a given
acknowledgment. This is done according to the basic IEEE forwarding factor 𝜌, each node sends out a new combination
802.11b unicast communication mode (without RTS/CTS). after having received exactly ⌈1/𝜌⌉ innovative packets. As
The packet is retransmitted, after a backoff period, in case an example, 𝜌 = 0.5 means that each node deterministically
there is no acknowledgment from the intended receiver. Using transmits a new combination every two received innovative
this mechanism, only collisions at the addressed receiver can packets. The forwarding factor, in this case, is not related to a
be detected, while collisions occurring at any of the other probability, but it is rather used as a threshold on the number
neighbors are ignored. Also, this strategy does not solve the of incoming messages.
hidden terminal problem. 3) Timed network coding: The two previous schemes have
3) IEEE 802.11 with pseudo broadcast and RTS/CTS hand- two major drawbacks. The rst is that they are particularly
shake: To further improve the performance we consider the sensitive to packet losses, e.g., due to collisions. In fact, if
previous scheme with additional RTS/CTS handshake. These one of the transmitted packets is lost, the propagation of
control messages are introduced to alleviate the hidden node the information through the network could be interrupted.
problem. The CTS is only transmitted by the node addressed The second drawback is that both probabilistic and semi-
in the packet header. As for the previous schemes, this strategy deterministic network coding suffer from some inefciencies
can not detect collisions at all overhearing nodes. when there is a small number of packets to combine. In
4) Ideal MAC: is a simple mechanism where transmitted such cases, new combinations are created from a small set
packets are only affected by the delay, Δ = ℓ + Δprop ≃ ℓ, of packets and, for this reason, are often not innovative. To
where Δprop is the propagation delay and ℓ is the packet alleviate these problems, we introduce a timing strategy into
transmission time. That is, we assume to have an omniscient the rst scheme. For each received innovative packet, a timer
entity which regulates the transmissions in order to completely is activated. When the timer has expired, the node decides
avoid interference and collisions. Hence, as a node sends a to send out a new random combination with probability 𝜌.
packet, all its neighbors successfully receive the message after The timer, 𝜏 , is a uniform random variable in [0, 𝜏max ]. This
the (xed) delay Δ. Note that this MAC still schedules packet timing approach has two advantages. With the introduction
transmissions according to CSMA and is thus non-ideal from of a waiting interval before transmission, nodes have the
a packet scheduling point of view when network coding is chance of collecting other innovative packets and send out
used at the upper layers. This idealized scheme is exploited richer combinations. Moreover, the reduction of the number
to pinpoint the effect of packet errors on the performance of of transmissions and the random characteristic of the timer
network coding. help in decreasing the collision probability at the MAC layer.
Network coding: we consider the RLNC paradigm of [3], The drawback of the timed scheme is the introduction of a
which is implemented as presented in [15]. At any given short delay due to the timer. Hence, the timer value shall be
node, every new encoded packet is obtained as the linear chosen so as to achieve a good trade-off between extra delay
combination of all packets in its receiving buffer, where and performance improvements. In IEEE 802.11b, this value
the combination coefcients are scalars randomly picked in has to be large enough to allow for the collection of more than
𝐺𝐹 (𝑞). These scalars are then stored in an encoding vector one packet, which translates to selecting 𝜏max ≈ 10 − 30 ms.
which is sent along with the encoded packet, see [15] for We picked 𝜏max = 20 ms. Note that in general 𝜏max depends
further details. In addition, we implement three different on network density and ow demands.
packet combination techniques. The rst two are inspired by Trafc pattern: for the trafc pattern, each node 𝑖 inserts into
the work in [9], while the last one is a new proposal. In the network a single original packet 𝑥𝑖 and wants to collect all
the following, we detail the packet combination strategies the other inserted packets. 𝑥𝑖 s are generated either randomly
considered in this section. All the presented schemes are or deterministically. In the former case, each node inserts its
characterized by a design parameter, called forwarding factor original packet by independently picking the insertion time
𝜌, which is dened as the ratio between the average number uniformly in a xed length interval of Δ1 = 100 ms. In the
of packets transmitted and the average number of innovative latter case, we can assume to have a simple application that
packets received, per node. A received packet is innovative inserts original packets sequentially in each node. Subsequent
whenever it increases the rank of the decoding matrix [9]. For insertions, at different nodes, are separated by xed time
the schemes below 𝜌, is decided a priori and equal for all intervals of Δ2 = 1 s. For this value of Δ2 , with the consid-
nodes. ered scenario (e.g., transmission times and network size), the
1) Probabilistic network coding: this approach uses RLNC. collision probability is negligible for both original packets and
Each node sends a random linear combination of the packets subsequent transmissions elicited by network coding. This is
in its buffer. When receiving an innovative message a new useful to assess the performance of our dissemination schemes
combination is transmitted with probability 𝜌 (the forwarding when used with an ideal MAC.
factor) whereas nothing is transmitted with probability 1 − 𝜌.
For 𝜌 = 0.5, a node on average sends a new message every two A. MAC and Combination Strategies: Simulation results
innovative packets received. From [21] we know that 𝜌 = 0.5 In this section, we discuss the most relevant results we
would theoretically (circular topology, ideal scheduling, and obtained via ns2 simulations. All presented schemes are eval-

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.
 ASTERJADHI et al.: TOWARD NETWORK CODING-BASED PROTOCOLS FOR DATA BROADCASTING IN WIRELESS AD HOC NETWORKS 665

NC ! IEEE 802.11b
NC ! IEEE 802.11b pb RTS/CTS
1 NC ! IEEE 802.11b pb
1
NC ! Ideal Mac
Flooding ! IEEE 802.11b

Packet Delivery Ratio, PDR

Packet Delivery Ratio, PDR

Flooding ! IEEE 802.11b pb RTS/CTS

0.8 Flooding ! IEEE 802.11b pb 0.8
Flooding ! Ideal Mac

0.6 0.6

0.4 0.4
NC ! IEEE 802.11b
NC ! IEEE 802.11b pb
NC ! IEEE 802.11b pb RTS/CTS
NC ! Ideal Mac
0.2 0.2 Flooding ! IEEE 802.11b
Flooding ! IEEE 802.11b pb
Flooding ! IEEE 802.11b pb RTS/CTS
Flooding ! Ideal Mac
0 0
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
Forwarding Factor, ! Forwarding Factor, !

Fig. 1. Packet Delivery Ratio: Performance comparison of Probabilistic Fig. 2. Packet Delivery Ratio: Performance comparison of Probabilistic
network coding and Probabilistic Flooding for different MAC protocols in network coding and Probabilistic Flooding for different MAC protocols in
circular networks with 𝑛 = 16. grid networks with 𝑛 = 16.

uated taking into account the random and the deterministic 6

NC ! IEEE 802.11b
trafc patterns. We test the algorithms varying 𝜌 from 0.1 to NC ! IEEE 802.11b pb
NC ! IEEE 802.11b pb RTS/CTS
1 and the number of nodes in the network, 𝑛, from 4 to 64, 5 NC ! Ideal Mac

Protocol Overhead, Pkttx/Pktdcd

Flooding ! IEEE 802.11b
and we consider a data rate of 𝑅 = 1 Mbps. For all MACs we Flooding ! IEEE 802.11b pb
Flooding ! IEEE 802.11b pb RTS/CTS
consider packet lengths of 𝐿 = 𝑃 𝐻𝑌 + 𝑀 𝐴𝐶 + 𝑁 𝐶(𝑛) + 𝑃 , Flooding ! Ideal Mac
4
where 𝑃 𝐻𝑌 and 𝑀 𝐴𝐶 are the physical and the MAC head-
ers, respectively, with 𝑃 𝐻𝑌 = 192 bits, 𝑀 𝐴𝐶 = 224 bits.
𝑁 𝐶(𝑛) is the network coding header that depends on the 3
number of nodes 𝑛 and is of size 𝑁 𝐶(𝑛) = 8(18 + 𝑛) bits.
𝑃 is the payload size that in our case is 64 bits.1 We organize 2
our performance analysis in two parts: we rst focus on
the impact of different MAC protocols on network coding 1
and we subsequently evaluate the effect of different packet
combination strategies. 0
Our performance metrics are: 1) the Packet Delivery Ratio, 0.2 0.4 0.6 0.8 1
Forwarding Factor, !
𝑃 𝐷𝑅, which is dened as the ratio between the number of
successfully received (and decoded) packets and the number Fig. 3. Protocol Overhead: Performance comparison of Probabilistic network
of packets a node is interested in, averaged over all nodes; coding and Probabilistic Flooding for different MAC protocols in grid
2) Packet Delivery Delay: is the average time between the networks with 𝑛 = 16.
rst transmission of a packet and its reception and successful
decoding at the destination nodes. This metric is only com- This is obtained through a proper centralized coordination
puted for correctly received packets; 3) Protocol Overhead: of the nodes’ transmissions which maximizes the probability
is the ratio between the number of transmitted packets at the of sending innovative packets at each transmission attempt.
MAC layer and the number of successfully decoded packets. However, this performance level is never reached in practice
This value depends on the adopted MAC protocol and on the and the actual 𝑃 𝐷𝑅 depends on the number of nodes. Looking
efciency of the network coding strategy. at Fig. 1 for 𝜌 = 0.6, IEEE 802.11b achieves 𝑃 𝐷𝑅 ≈ 0.6,
Impact of MAC protocols: in Fig. 1 we compare probabilistic whereas an ideal MAC achieves 𝑃 𝐷𝑅 ≈ 0.8, which cor-
network coding (solid lines) against probabilistic ooding responds to a decrease in performance of about 25%. Note
(dotted lines) in terms of delivery ratio in a circular net- that our ideal MAC does not provide full reliability as it still
work topology for 𝑛 = 16. Network coding outperforms schedules transmissions according to CSMA and does not use
probabilistic ooding for all values of 𝜌. Gains are more the optimal coordination strategy of [21]. The effectiveness of
pronounced when 𝜌 is close to one and 𝑛 is large (results pseudo broadcast (IEEE 802.11 pb in the gure) and pseudo
for different values of 𝑛 are not shown here due to space broadcast with RTS/CTS (IEEE 802.11 pb RTS/CTS) is also
constraints, the results are however similar to those in [9]). clear, though the improvements are not as large as expected.
As observed in [21], for this topology a delivery ratio of The observed decrease in performance is due to the use of
one is theoretically achievable with 𝜌 slightly larger than 0.5. an actual MAC layer (IEEE 802.11b in this case) and to
the sub–optimality of random scheduling, which indicates the
1 Due to the inefciencies of 802.11 medium access and the additional
importance of these issues for the design of practical schemes.
network coding overhead, using a small packet size represents a lower bound
on throughput performance. The relative performance differences between the Fig. 2 shows results for a different setting where 𝑛 = 16
protocols remain unchanged when using larger packet sizes. nodes are placed over a grid. As expected, the achieved

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.
 666 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 2, FEBRUARY 2010

0
10
NC ! IEEE 802.11b 1
NC ! IEEE 802.11b pb
NC ! IEEE 802.11b pb RTS/CTS
NC ! Ideal Mac

Packet Delivery Ratio, PDR

Flooding ! IEEE 802.11b
Packet Delivery Delay, D [s]

Flooding ! IEEE 802.11b pb 0.8

Flooding ! IEEE 802.11b pb RTS/CTS
!1
10 Flooding ! Ideal Mac

0.6

!2 0.4
10

0.2 Probabilistic NC ! Random traffic pattern

Semi!deterministic NC ! Random traffic pattern
Timed NC ! Random traffic pattern
!3 Semi!deterministic NC ! Deterministic traffic pattern
10 0
0.2 0.4 0.6 0.8 1
0.2 0.4 0.6 0.8 1 Forwarding Factor, !
Forwarding Factor, !

Fig. 5. Performance comparison of different combination strategies in

Fig. 4. Packet Delivery Delay: Performance comparison of Probabilistic
circular networks with 𝑛 = 16 and IEEE 802.11 MAC.
network coding and Probabilistic Flooding for different MAC protocols in
grid networks with 𝑛 = 16.
constant up to 𝜌∗ = 0.4 and then suddenly increases for
performance is better than in the circular case due to the higher forwarding factors. This does not occur for probabilistic
higher number of neighbors per node (4 instead of 2), which network coding (solid lines) whose curves are smooth. This
favors packet mixing and dissemination. Also in this scenario, reects the threshold based transmission policy of semi-
the presence of realistic MAC layers signicantly reduces deterministic network coding. The exact value of the shifting
the 𝑃 𝐷𝑅 metric for a given 𝜌. As expected, the schemes point 𝜌∗ depends on the number of neighbors. For circular
implementing collision avoidance policies (i.e., IEEE 802.11b networks, where each node has exactly two neighbors, 𝜌 < 0.5
with pseudo broadcast and IEEE 802.11 with pseudo broadcast (⌈1/𝜌⌉ > 2) never sufces to trigger the transmission of a
and RTS/CTS handshake) improve 𝑃 𝐷𝑅 but also increase the new combination, as the initial number of innovative packets
protocol overhead. This is due to the MAC retransmissions is equal to two. This aw is not present in probabilistic
in case of collisions and to the control trafc (i.e., ACK, and timed network coding, where sending rules are based on
RTS and CTS packets). In addition, we note that when we probabilities rather than on hard thresholds. Timed network
compare the performance of probabilistic network coding and coding outperforms the semi-deterministic scheme with deter-
ooding against 𝜌, we have a fair comparison as, given a ministic trafc pattern for 𝜌 ≤ 𝜌∗ and performs very close
specic 𝜌 and a xed MAC protocol, both network coding and to this method for larger forwarding factors. In addition, the
ooding lead to very similar protocol overhead (see Fig. 3). timed strategy performs better than both semi-deterministic
Pseudo broadcast and pseudo broadcast with RTS/CTS are and probabilistic network coding with random scheduling. For
effective in decreasing the number of collisions. However, 𝜌 = 0.5, probabilistic network coding with random scheduling
using these additional techniques to recover from packet loss achieves 𝑃 𝐷𝑅 ≈ 0.35, whereas timed network coding leads
leads to longer delays, as can be seen from Fig. 4. The average to 𝑃 𝐷𝑅 ≈ 0.55, which corresponds to an improvement of
delay increase is about one order of magnitude in the worst about 57%. We observed that the timed strategy introduces
case (pseudo broadcast with RTS/CTS). We also note that an additional delay. Also, there are some expected differences
the delay of network coding stabilizes for increasing 𝜌 while between ideal and actual MAC. For IEEE 802.11b, the delay
it continues to increase for ooding. The reason for this is increase is reasonably small (approximately equal to the aver-
that with ooding, a higher number of redundant packets is age value of the timer) and is similar to that introduced by the
received early on, delaying the reception of innovative packets. pseudo broadcast algorithms. Hence, the timed combination
For network coding, the combination of packets prevents this provides higher benets than pseudo broadcast in terms of
from happening and most packets received are innovative even packet delivery ratio, leading to similar extra delays. For this
for high 𝜌. reason, the timed scheme may make sense when the goal
To sum up, we observe that the presence of actual MAC is to maximize the packet delivery ratio (throughput) while
protocols reduces the performance in terms of packet delivery accepting some delay degradation. This extra delay appears,
ratio; however, this reduction is not as high as expected. In however, well tolerable (less than 5% increase over the delay
addition, collision avoidance policies give little improvement without the timed strategy). The results of probabilistic and
in terms of 𝑃 𝐷𝑅, while leading to poor overhead and delay timed network coding for random topologies are plotted in
performance. Fig. 6 (similar results hold for grid networks). Note that in
Impact of packet combination strategies: Fig. 5 shows the these networks the gain becomes larger for decreasing 𝜌; in
packet delivery ratio performance for a circular network with fact, when fewer packets are transmitted timed network coding
𝑛 = 16 for various packet combination strategies for an more effectively exploits coding opportunities.
IEEE 802.11b MAC protocol. The semi-deterministic schemes From the above results we see that there exists a gap
(dotted lines) show a phase change, where 𝑃 𝐷𝑅 remains between the predicted theoretical performance and the results

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.
 ASTERJADHI et al.: TOWARD NETWORK CODING-BASED PROTOCOLS FOR DATA BROADCASTING IN WIRELESS AD HOC NETWORKS 667

an example, think of the case where a given node 𝑥 has a large

1
number of neighbors and one of them, say node 𝑦, has only
𝑥 as its neighbor. Due to its high number of neighbors (small
Packet Delivery Ratio, PDR

0.8 𝜌), 𝑥 sends out a small number of packets and, in turn, 𝑦 is

unlikely to be able to decode all the wanted information (as it
did not receive enough independent combinations from 𝑥). In
0.6
contrast, our proactive scheme does not require the reception
of innovative information to continue data dissemination (so
0.4 it is more robust to interference and collisions), and its
15 Neighbors ! Timed NC ! IEEE 802.11b
15 Neighbors ! Timed NC ! Ideal MAC performance does not depend on the forwarding factor 𝜌. It is
15 Neighbors ! Probabilistic NC ! IEEE 802.11b
0.2 15 Neighbors ! Probabilistic NC ! Ideal MAC based on two important components: 1) a set of conditions
7 Neighbors ! Timed NC ! IEEE 802.11b
7 Neighbors ! Timed NC! Ideal MAC to stop transmissions when all original packets have been
7 Neighbors ! Probabilistic NC ! IEEE 802.11b
7 Neighbors ! Probabilistic NC! Ideal MA delivered to all nodes, i.e., Stopping Conditions (SC) and 2)
0 a strategy to set the frequency at which new random packet
0.2 0.4 0.6 0.8 1
Forwarding Factor, ! combinations are to be sent so as to avoid network congestion.
In the rest of the section we refer to this strategy as Rate
Fig. 6. Random Networks: Performance comparison of Packet Delivery Adaptation mechanism.
Ratio for Probabilistic and Timed Network Coding for ideal and IEEE 802.11b
MAC.
Basic rules for ProNC: each node can be in one of two
different states: active and inactive. The basic idea of the
obtained in realistic environments. This gap is due to the proactive approach is that an active node periodically sends out
random access mechanism of IEEE 802.11, which leads to a new packet combination to its neighbors, while an inactive
collisions, and also to its packet scheduling, which does not node does not transmit. To switch from one state to the other,
allow the combination of packets in the optimal order. a node considers the following set of rules:
R1 A node becomes active upon receiving the rst innovative
packet. This means that a data dissemination phase is
III. P ROACTIVE N ETWORK C ODING (P RO NC)
started and the node has to contribute to it.
The schemes considered up to now are reactive protocols, R2 A node becomes inactive when the Stopping Condition
i.e., nodes participate in the dissemination of data only when is veried. In this case, further transmissions from this
they receive innovative information. If this does not occur, node are no longer useful for its neighbors and should
the dissemination is interrupted even though nodes may still be suppressed to avoid unnecessary overhead.
have innovative information to send. This fact is an inherent R3 A node becomes active again when the Stopping Con-
characteristic of the reactive approach. In this section we dition no longer holds. This last rule is particularly
propose a network coding data dissemination scheme based important as it allows propagation of new information
on a proactive approach (referred to in the following as into an area where all nodes are currently inactive.
ProNC) to address this problem. Even though our focus Note that while a node is inactive, it can still receive packets
here is on scenarios where data is to be exchanged among from its neighbors. This information is used to assess whether
all the users of a wireless ad hoc network, the rationale the stopping condition still holds.
behind ProNC also applies to different settings. Our scheme Stopping conditions: there are different ways to dene the
is completely distributed and self-adaptable and requires very Stopping Conditions for proactive network coding. They de-
limited network knowledge, which can be easily acquired by pend, in general, on the amount of information that each
overhearing the exchanged data. node has to collect in order to decide whether to suspend
In the previous section we have seen that reactive schemes its transmissions. Our main aim is to keep the transmission
are likely to suffer from the presence of interference and overhead as low as possible. We identify two simple cases
collisions in realistic radio environments. The main problem in which a node has to suspend its transmission: 1) In the
of reactive schemes is that new random combinations are rst case, all neighbors of a node 𝑥 have decoded all the
generated and transmitted only when innovative (i.e., linearly packets they require and thus no further transmissions by 𝑥
independent) information is received. Innovative packets may are necessary. 2) The second is when the subspace spanned by
however be lost due to packet collisions, thus interrupting the information vectors (i.e., packets) available at node 𝑥 is
data propagation. Even worse, the insertion of innovative contained in the subspace spanned by the information vectors
information into a given network area often causes all nodes at each of the node’s neighbors. In this case, 𝑥’s packets will
in the area to attempt their new transmissions simultaneously not be innovative for any of its neighbors and the node should
and this further increases the collision probability. suspend its transmission.
In reactive probabilistic network coding, nodes send out Based on these observations, we propose two different
new combinations based on a forwarding factor 𝜌, which conditions which are referred to as Strong and Weak Stopping
depends on their number of neighbors [9]. Setting 𝜌 inversely Conditions (SSC and WSC, respectively). They dene two
proportional to the number of neighbors has the desirable different proactive schemes. According to the SSC, nodes
effect that the number of innovative packets per area is send out beacons (Strong Stopping Messages, SSM) to their
independent of the local node density. We observe that there neighbors when they have decoded all the packets they are
are particular topologies where this strategy does not work. As interested in. Each node collects SSMs from its neighborhood.

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.
 668 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 2, FEBRUARY 2010

When a node receives an SSM from each of its known q/W q/W ... q/W q/W
neighbors, the SSC is veried and transmissions are stopped.
We refer to this scheme as Strong ProNC as it requires strong p p p p
assumptions on the data trafc. In order to send out SSMs, q/W
each node needs to know in advance how many packets it 0 1 2 ... W-2 W-1
wants to collect. This fact implies that each node has full 1-p 1-p 1-p 1-p
knowledge about the amount (and type) of data owing over q/W q/W
the network. Note that the collection of this information may 1-q
q/W ... q/W
not be feasible in practice.
The second strategy we propose is based on the WSC.
-1
During data propagation, each node sends out beacons (Weak
Stopping Messages, WSM) containing a decoding eld which
is set to 1 if it can decode all packets in its buffer and to 0 1-q
otherwise. In addition, beacons contain a rank eld specifying
the rank of the nodes’ decoding matrices. According to the Fig. 7. Markov chain tracking the evolution of the IEEE 802.11 broadcast
Weak Stopping Conditions each node suspends its transmis- transmission process. For 𝑛𝑣 nodes in a neighborhood, 𝑝 is the probability
that at least one of the remaining 𝑛𝑣 nodes transmits when the target node
sions when all its neighbors can decode all the packets in their is in slot 𝑖 = 1, . . . , 𝑊 − 1. 1 − 𝑞 is the probability that the transmission
buffers and their matrices have the same (full) rank. We refer buffer is empty after the transmission of a given packet. 𝑊 is the backoff
to this second strategy as Weak ProNC because it does not window size of IEEE 802.11.
require any knowledge about the data trafc and has a limited
overhead. However, Weak ProNC is suboptimal as there are dimensional Markov chain of Fig. 7 (a two dimensional chain
some situations in which the rank alone does not capture the was used in [22] for the IEEE 802.11 unicast case to take
exact decoding status at different nodes. For instance, it might retransmissions into account). In addition, in order to model
happen that all neighbors of a node can decode all the packets the transmission process in nonsaturated trafc conditions, we
in their buffers and they all have the same rank but the decoded adopt the technique of [23] where idle transmission times
information is different. (due to empty transmission queues) are modeled through the
addition of the further state −1. With reference to Fig. 7
A. Rate adaptation heuristics our model works as follows. When a target node has a
packet to transmit, it starts the backoff process by randomly
We dene 𝜏 as the time elapsed between the completion of selecting a backoff counter value from 0 to 𝑊 − 1 and then
the transmission of a packet by the PHY and the instant when starting to decrement the counter until state 0 is reached. State
the next packet is made available for transmission at the MAC, 𝑖 = 0, 1, . . . , 𝑊 − 1 represents the current backoff counter
i.e., the idle time of the node. Note that 𝜏 is (roughly) inversely value. The transition from state 𝑖 to state 𝑖 − 1 occurs with
proportional to the transmission rate of the nodes. In what probability 1 − 𝑝 after a backoff slot time (of xed duration
follows, we present an approximate model to nd the value of 𝜎), while with probability 𝑝 the process remains in state 𝑖.
𝜏 that maximizes the amount of innovative information that 𝑝 is the probability that at least one of the remaining 𝑛𝑣
is transferred over the channel as a function of the system nodes transmit when the target node is in state 𝑖. If this
parameters. Note that the physical (PHY) layer data rate is kept occurs, the node momentarily stops counting down its backoff
constant. Methods to change the PHY rate are investigated in timer. The current packet is nally transmitted when the
Section IV. backoff process reaches state 0 (transmission state). Upon the
Impact of MAC layer dynamics: in what follows we de- completion of the packet transmission two events can occur:
rive the relationship between the value of 𝜏avg = 𝐸[𝜏 ] E1) with probability 𝑞 the transmission queue is non-empty
that maximizes the throughput, referred to as 𝜏avg ★
, and the and a new backoff timer is uniformly selected at random in
number of neighbors at any given node, 𝑛𝑣 . We consider {0, 1, . . . , 𝑊 − 1}: the probability that the system moves from
the packet transmission process in a given neighborhood of state 0 to any state 𝑖 = 0, 1, . . . , 𝑊 − 1 is thus 𝑞/𝑊 . E2) With
𝑛 = 𝑛𝑣 + 1 nodes making the following assumptions: A1) we probability 1 − 𝑞 the transmission queue is empty and in this
neglect the channel propagation delay as for the considered case the process moves to the idle state −1, where it remains
system parameters it has a negligible impact on the throughput until a new packet arrives (at which point the chain evolution
performance, A2) we assume that all packets involved in a is the same as in E1).
collision are lost and A3) we assume that any transmitted
Let 𝜋𝑖 , 𝑖 = −1, 0, 1, . . . , 𝑊 − 1 be the steady-state proba-
packet is always successfully received by all nodes in the bilities of the above Markov chain. Our goal is to nd 𝜋0 (the
neighborhood unless it collides with another transmission. To
transmission probability) and relate it to 𝜏avg . From the chain
obtain a rate adaptation heuristic we model the IEEE 802.11 regularities, and by computing recursively through the chain
broadcast communication process. As in [22], the evolution from right to left, we obtain:
of the transmission backoff counter is tracked using a suitable
Markov chain. However, in our case the backoff window size
𝑊 is always constant as packets are neither retransmitted nor 𝑖𝑞(𝜋0 + 𝜋−1 )
𝜋𝑊 −𝑖 = , 𝑖 = 1, 2, . . . , 𝑊 − 1 ,
acknowledged. This implies that the broadcast backoff process 𝑊 (1 − 𝑝)
of any of the 𝑛𝑣 + 1 nodes can be modeled through the one 𝜋0 = (𝜋0 + 𝜋−1 )𝑞 . (1)

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.
 ASTERJADHI et al.: TOWARD NETWORK CODING-BASED PROTOCOLS FOR DATA BROADCASTING IN WIRELESS AD HOC NETWORKS 669

∑𝑊 −1
From the normalization condition 𝑖=−1 𝜋𝑖 = 1 we nd The optimal transmission probability 𝜋0★ can be found as (see
1 calculations in Section VI of [22]):
𝜋0 + 𝜋−1 = (𝑊 −1)𝑞
, (2)
1+ 2(1−𝑝)
𝜋0★ = argmax 𝑔(𝜋0 )
𝜋
√ 0
from which we nally obtain 𝜋0 as: [(𝑛𝑣 + 1) + 2𝑛𝑣 (𝑇𝑠 /𝜎 − 1)]/(𝑛𝑣 + 1) − 1
𝑞 = (10)
.
𝜋0 = . (3) 𝑛𝑣 (𝑇𝑠 /𝜎 − 1)
−1)𝑞
1 + (𝑊
2(1−𝑝) The optimal 𝜏avg
★
, is obtained from 𝜋0★ as:
The probability 𝑝 that at least one of the remaining 𝑛𝑣 nodes 1. Express 𝑞 as 𝑞 = 1 − 𝑒−𝐸[slot length]/𝜏avg , which as shown
transmits when the target node is in slot 𝑖 = 1, . . . , 𝑊 − 1 in [23] provides a good approximation of the queue
𝑑𝑒𝑓
is found as: 𝑝 = 1 − (1 − 𝜋0 )𝑛𝑣 = 𝑓1 (𝜋0 ). We additionally behavior in the unsaturated case. Inverting this relation
dene the probability 𝑃𝑡 that at least one node is transmitting gives:
in a given slot, 𝐸[slot length]
𝑛𝑣 +1 𝜏avg = −
𝑃𝑡 = 1 − (1 − 𝜋0 ) , (4) log(1 − 𝑞)
and 𝑃𝑠 as the probability that only one node is in the 𝜎(1 − 𝑃𝑡 ) + 𝑃𝑡 𝑇𝑠
= − . (11)
transmission state, conditioned on the fact that at least one log(1 − 𝑞)
node is transmitting. 𝑃𝑠 is obtained as: 2. Obtain 𝜋0★ from (10).
(𝑛𝑣 + 1)𝜋0 (1 − 𝜋0 )𝑛𝑣 (𝑛𝑣 + 1)𝜋0 (1 − 𝜋0 )𝑛𝑣 3. Invert (3) to nd 𝑝 as a function of 𝜋0 as:
𝑃𝑠 = = . (5)
𝑃𝑡 1 − (1 − 𝜋0 )𝑛𝑣 +1 𝜋0 (𝑊 − 1) 𝑑𝑒𝑓
𝑝=1− = 𝑓2 (𝜋0 , 𝑞) (12)
We are now ready to calculate the normalized throughput 𝑆 2(1 − 𝜋0 /𝑞)
as (see [22]): and nd 𝑞 ★ as the solution of 𝑓1 (𝜋0★ ) − 𝑓2 (𝜋0★ , 𝑞) = 0,
𝐸[payload bits successfully transmitted in a slot] which leads to:
𝑆= . (6)
𝐸[slot length] 2𝜋0★ (1 − 𝜋0★ )𝑛𝑣
Since a successful transmission occurs in a slot with proba- 𝑞★ = . (13)
2(1 − 𝜋0★ )𝑛𝑣 + 𝜋0★ (1 − 𝑊 )
bility 𝑃𝑡 𝑃𝑠 , the average number of payload bits successfully
4. Obtain 𝜏avg
★
from (11) setting 𝑞 ← 𝑞 ★ and expanding
transmitted in a slot time is 𝑃𝑡 𝑃𝑠 𝑃 , where 𝑃 is the payload
𝐸[slot length] using (4):
size. The average length of a slot can be obtained by consid-
ering the following three mutually exclusive cases: C1) with ★ 𝑇𝑠 − (1 − 𝜋0★ )𝑛𝑣 +1 (𝑇𝑠 − 𝜎)
𝜏avg = ( ) (14)
probability 1 − 𝑃𝑡 , none of the nodes transmit in the slot. The 2(1−𝜋 ★ )𝑛𝑣 +𝜋 ★ (1−𝑊 )
log 2(1−𝜋★0)𝑛𝑣 +1 +𝜋0 ★ (1−𝑊 )
duration of an empty backoff slot is 𝜎 = 20 𝜇s. C2) with 0 0

probability 𝑃𝑡 𝑃𝑠 , only one node transmits in the slot. The slot We observe that our model is accurate for sufciently large
duration in this case is 𝑇𝑠 , which is the transmission time 𝑛𝑣 , i.e., 𝑛𝑣 ≥ 4. For smaller values it is however inaccurate
of a packet, given by the sum of the time spent transmitting because of several approximations made in the analysis, i.e.,
the physical header (𝑇PHY ), the MAC header (𝑇MAC ), the the expression of 𝑞 [23], the independence of the busy channel
network coding header (𝑇NC (𝑛) = 𝑁 𝐶(𝑛)/𝑅), the payload probability 𝑝 among subsequent access slots [22] and to the
(𝑇𝑃 = 𝑃/𝑅), plus the distributed inter frame space (DIFS) fact that in our derivation of the steady state probabilities
time, 𝑇DIFS : we neglect the semi-Markov character of the process, i.e.,
𝑇𝑠 = 𝑇PHY + 𝑇MAC + 𝑇NC (𝑛) + 𝑇𝑃 + 𝑇DIFS . (7) that states 0 and 𝑖 ∕= 0 have different durations (a common
simplication for the analysis of IEEE 802.11 throughput [22],
C3) With probability 𝑃𝑡 (1 − 𝑃𝑠 ), multiple nodes transmit in [23]).
a slot time, leading to a collision event. The duration of this Implementation notes: in our implementation we pick
slot is also 𝑇𝑠 because packets are not acknowledged and, 𝜏avg = 𝜏avg★
selecting 𝜏 uniformly in [0, 2𝜏avg ], which gives
therefore, the transmission period for successful and collided 𝐸[𝜏 ] = 𝜏avg . We obtain 𝜏avg★
as a function of 𝑛𝑣 from (14)
packets is the same. Hence, we have 𝐸[slot length] = 𝜎(1 − using 𝑇PHY = 192 𝜇s, 𝑇MAC = 224 𝜇s, 𝑇DIFS = 50 𝜇s,
𝑃𝑡 ) + 𝑃𝑡 𝑃𝑠 𝑇𝑠 + 𝑃𝑡 (1 − 𝑃𝑠 )𝑇𝑠 . These facts together with (6) 𝜎 = 20 𝜇s, 𝑅 = 1 Mbps and 𝑊 = 32 slots, which are used
give: for IEEE 802.11b broadcast with a rate of 1 Mbps. We nd
𝑃𝑡 𝑃𝑠 𝑃 that the relationship between the two is well approximated
𝑆 =
𝜎(1 − 𝑃𝑡 ) + 𝑃𝑡 𝑃𝑠 𝑇𝑠 + 𝑃𝑡 (1 − 𝑃𝑠 )𝑇𝑠 by a linear function, as predicted by the simulation results
𝑃𝑠 (𝑃/𝜎) of [9], [24]. A good approximation is in fact given by the
= . (8) following heuristic: 𝜏avg
★
≃ 𝜅𝑛𝑣 𝑇𝑠 , where 𝜅 = 0.7. We note
(1 − 𝑃𝑡 )/𝑃𝑡 + 𝑇𝑠 /𝜎
that 𝜏avg corresponds to the average amount of time spent
Note that the maximum throughput is achieved when the
in state −1, i.e., to the time elapsed between the completion
following function is maximized: of the transmission of a packet by the PHY and the instant
𝑃𝑠 when a new packet is made available by ProNC. The inter-
𝑔(𝜋0 ) =
(1 − 𝑃𝑡 )/𝑃𝑡 + 𝑇𝑠 /𝜎 packet transmission time is greater than 𝜏avg as it also includes
(𝑛𝑣 + 1)𝜋0 (1 − 𝜋0 )𝑛𝑣 the time spent in backoff. Moreover, ProNC requires the
= . (9)
𝑇𝑠 /𝜎 − (1 − 𝜋0 )𝑛𝑣 +1 (𝑇𝑠 /𝜎 − 1) estimation of the number of neighbors at each node which

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.
 670 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 2, FEBRUARY 2010

0 0
10 10

increasing !

!1 !1
10 10
increasing !

Failure Probability, 1!PDR

Failure Probability, 1!PDR

n v =4 n v =2
!2 Weak ProNC !2 n v =4 n v =2 Weak ProNC
10 10
Strong ProNC Strong ProNC
Reactive NC, nv=2 Reactive NC, n =2
v
Reactive NC, n =4 Reactive NC, n =4
v n v =8 v
Reactive NC, nv=8
Reactive NC, n =8
v n v =8
Reactive NC, n =16 Reactive NC, nv=16
v !3
!3 Reactive NC, nv=32
10 Reactive NC, n =32 10
v
Adaptive NC, n =2 Adaptive NC, nv=2
v
Adaptive NC, n =4 Adaptive NC, nv=4
v
Adaptive NC, nv=8 Adaptive NC, n =8
v
n v =32 n v =16 n v =8 n v =4
Adaptive NC, nv=16 n v =16 Adaptive NC, nv=16
!4
!4
Adaptive NC, nv=32 n v =32 10 Adaptive NC, nv=32
10

!3 !2 !1 0 1 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8

10 10 10 10 10
Packet Delivery Delay [s] Protocol Overhead

Fig. 8. Failure probability 1−𝑃 𝐷𝑅 vs delay: comparison between proactive Fig. 9. Failure probability 1 − 𝑃 𝐷𝑅 vs overhead: comparison between
and reactive schemes. The curves shown for reactive schemes are obtained proactive and reactive schemes. The curves shown for reactive schemes are
for different values of 𝑛𝑣 ∈ {2, 4, 8, 16, 32}, varying 𝜌 as the independent obtained for different values of 𝑛𝑣 ∈ {2, 4, 8, 16, 32}, varying 𝜌 as the
parameter. independent parameter.

can be simply achieved by monitoring the source addresses average node densities by varying the average number of
of incoming packets. Note that both the stopping conditions neighbors, 𝑛𝑣 ∈ {2, 4, 8, 16, 32}. For the MAC, we adopt the
and the packet rate adaptation mechanism depend on the node basic IEEE802.11b broadcast mode, accounting for channel
density. Most importantly, the number of nodes estimated in errors and collisions. In the following graphs we consider
this way is smaller than that of the previous analysis. In the probabilistic network coding of Section II (referred to as
fact, this number of neighbors only accounts for the nodes ”Reactive NC” in the plots) as well as the adaptive network
within transmission range, whereas 𝑛𝑣 should include all coding (referred to as ”Adaptive NC”) scheme of [9] where
nodes in the collision domain, whose range is always greater. 𝜌 is picked independently at each node as 𝜌 = 𝑐/(𝑛𝑣 + 1),
′
Hence, the actual inter-packet transmission time 𝜏avg , which where 𝑐 is a suitable constant equal for all nodes [9]. In Fig. 8
accounts for all these facts, is 𝜏avg = 𝜅 𝑛𝑣 𝑇𝑠 where 𝜅′ = 6.
′ ′
we show the tradeoff between failure probability, 1 − 𝑃 𝐷𝑅,
We thus use this linear heuristic which gives good results and the packet delivery delay. We note that ProNC performs
across all simulations. In addition, Stopping Messages are better in terms of data recovery; 1 − 𝑃 𝐷𝑅 is at least one
included within data packets at the cost of a few extra bits. order of magnitude smaller for ProNC when 𝑛𝑣 ∈ {8, 16}.
For SSM, we need one additional bit, whereas for WSM we For small 𝑛𝑣 , i.e., 𝑛𝑣 ∈ {2, 4}, we often obtain pathological
need a bit to represent the decoding status and a byte to topologies leading to deadlocks of the data dissemination
communicate the rank of the local decoding matrix2 . In both when reactive protocols are used. ProNC efciently deals
cases, the additional overhead is acceptable. On the downside, with these topologies and alleviates the deadlock problem by
when a node becomes inactive it must send out at least one offering better performance in terms of 1 − 𝑃 𝐷𝑅. On the
Stopping Message to communicate its change of status and downside, in these cases the dissemination of data in ProNC
this packet may be useless for coding purposes. We note takes slightly longer due to the waiting periods of proactive
that piggybacking control information within data packets has schemes (see variable 𝜏 ). The tradeoff concerning the protocol
the benecial effect of keeping channel congestion low. In overhead is shown in Fig. 9: similarly to reactive schemes,
addition, the added control information (SSMs and WSMs, the overhead of ProNC increases with decreasing 𝑛𝑣 . This
rank, decoding status) is used to increase the efciency of is because network coding is more efcient when the node
network coding schemes which, in turn, can further reduce density is high. Also, the overhead of ProNC is usually smaller
the number of transmissions for a target performance level. than that of reactive schemes, while it always outperforms
These benets are quantitatively veried below. reactive solutions in terms of data recovery performance. For
both graphs Weak ProNC performs slightly worse than Strong
B. ProNC: Simulation Results ProNC in terms of packet delivery ratio, whereas it performs
Next, we compare the ProNC scheme against the reactive better in terms of overhead performance for the same 𝑛𝑣 .
probabilistic schemes proposed in [9]. The results that follow The difference in performance is more signicant at small
are for topologies where nodes are randomly placed within densities, i.e., where deadlocks are more likely to occur. As
a xed area in such a way that the topology is always con- demonstrated in [25], the overhead performance of all schemes
nected, possibly through multi-hop paths. We consider several at high densities approaches that of an idealized scheme,
having the minimum possible overhead: this reects the fact
2 A single byte often sufces in practice, i.e., the number of nodes in the
that network coding works better when there are more coding
network that generate original packets is lower than or equal to 256. Coding opportunities.
over more original packets would imply the inversion of large matrices at the
receiver which is impractical and difcult to obtain in realtime. To summarize, both Strong and Weak ProNC show satisfac-

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.
 ASTERJADHI et al.: TOWARD NETWORK CODING-BASED PROTOCOLS FOR DATA BROADCASTING IN WIRELESS AD HOC NETWORKS 671

tory performance in actual network settings. Weak ProNC is 1

a distributed and self-adaptable dissemination protocol which 0.9

does not require any knowledge about the trafc and only 0.8

Packet Delivery Ratio, PDR

requires a few local interactions among nodes to work prop-
0.7
erly. Further improvements of Weak ProNC in terms of 𝑃 𝐷𝑅
are possible through the extension of the communication of 0.6
control messages over multiple hops. These issues are left for 0.5
future research.
0.4
n = 10, R = 6 Mbps
v
IV. N ETWORK C ODING IN M ULTI -R ATE A D H OC 0.3 nv = 10, R = 54 Mbps
N ETWORKS 0.2 n = 10, c = 20 %
v f
nv = 10, cf = 40 %
When nodes have multi-rate transmission capabilities, it is 0.1
n = 10, c = 60 %
necessary to pick a suitable PHY layer data rate, besides the v f
0
selection of the forwarding factor. This is not an easy task. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Forwarding Factor, !
When nodes transmit at low data rates their coverage area is
larger and in this case packets travel long hops. Conversely,
Fig. 10. Packet delivery ratio for a multi-rate ad hoc network as a function
transmitting with higher data rates leads to shorter transmis- of 𝜌.
sion delays and shorter hops. Whichever is best depends on
several factors such as network density and topology. The
analysis in, e.g., [26] can be used to calculate the involved nv = 10, R = 6 Mbps
nv = 10, R = 54 Mbps
delay-throughput tradeoffs for different rates. The objective of
nv = 10, cf = 20 %
the following paragraphs is to obtain good tradeoffs between
nv = 10, cf = 40 %
𝑃 𝐷𝑅 and average delay in IEEE 802.11g-based multi-rate Packet Delivery Delay [s] nv = 10, cf = 60 %
scenarios in the presence of RLNC.
As in Section III-B, we consider a random wireless network
!2
where nodes want to disseminate and retrieve information 10

through multi-rate network coding. We investigate the cooper-

ation between IEEE 802.11g PHY/MAC and network coding
in this random scenario considering the all-to-all transmission
paradigm. We present results obtained through ns2 simulations
with varying forwarding factor 𝜌 for reactive network coding.
As per our discussion above, 𝜌 is the probability of sending
a new combination whenever a node receives an innovative 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Forwarding Factor, !
packet and directly determines the amount of trafc which
ows through the wireless network. Note that ProNC can be Fig. 11. Packet delivery delay for a multi-rate ad hoc network as a function
applied to a multi-rate transmission scenario as well and, for of 𝜌.
this case, we found similar advantages as those presented in
Section III-B. 1) id: the address of the sending node, 2) 𝐿: the size of the
received packet, 3) 𝛾𝑗 : the instantaneous 𝑆𝐼𝑁 𝑅 associated
A. Rate Adaptation Heuristics and Simulation Results with this packet reception. Upon reception, an average 𝑆𝐼𝑁 𝑅,
In the following, we present a data rate adaptation heuristic 𝛾 𝑗 , is updated for each neighbor 𝑗 according to a discrete time
which tries to achieve, at the same time, a short delay and a rst order low pass lter as: 𝛾 𝑗 = 𝛼𝛾𝑗 + (1 − 𝛼)𝛾 𝑗 , where
high 𝑃 𝐷𝑅. We assume that wireless nodes initially have no 𝛼 ∈ (0, 1) is the smoothing factor. For our simulations we
knowledge about the network status, i.e., they are completely picked 𝛼 = 0.5. In this way we take into account the variations
unaware of location and number of neighbors as well as the in the received SINRs, while trying to capture its average
transmission opportunities in their neighborhood. In order to value. The packet transmission length 𝐿 and 𝛾 𝑗 are thus used
acquire this knowledge the algorithm uses internal variables, to estimate the PER for neighbor 𝑗 for all transmission rates.
at the MAC layer, to store auxiliary pieces of information In this way, each node estimates the status of its neighbors
such as node addresses, signal to interference plus noise ratios, in terms of associated 𝑆𝐼𝑁 𝑅. Once node 𝑖 has this informa-
𝑆𝐼𝑁 𝑅, as well as data rates. tion it updates its internal variables. Subsequently, it considers
Initially, all nodes begin their transmissions with the lowest the fraction 𝑐𝑓 of the nodes in its neighboring set that have the
available rate (which is 6 Mbps for IEEE 802.11g). This is the highest 𝛾s. 𝑐𝑓 is referred to here as coverage factor. The data
best choice in terms of neighbor discovery as it allows nodes rate at a given node 𝑖 is thus selected such that all of these
to collect information from a larger area. The subsequent re- nodes will receive packets from node 𝑖 with a small packet
ception of packets permits the gathering of useful information error probability, i.e., smaller than a given threshold 𝑃𝑡ℎ . In
which will determine the data rate at this node, as we explain the following results we selected 𝑃𝑡ℎ = 0.03 as it gave good
next. Each time a node, say node 𝑖, receives a new packet from results across all our experiments.
one of its neighbors 𝑗, it extracts the following information: Fig. 10 shows 𝑃 𝐷𝑅 as a function of 𝜌 for different coverage

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.
 672 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 2, FEBRUARY 2010

n = 7, ! = 0.4
v
network sizes. As usual, various tradeoffs can be obtained for
!1
nv = 7, ! = 0.6 different forwarding factors 𝜌: high 𝜌 values always lead to
10 nv = 7, ! = 0.8
n = 10, ! = 0.4
v
good 𝑃 𝐷𝑅 performance at the cost of additional delay, while
a small 𝜌 is a good choice in terms of delay performance at
Failure Probability, 1!PDR

n = 10, ! = 0.6
v
n = 10, ! = 0.8
v
nv = 15, ! = 0.4
the cost of an increased failure probability.
n = 15, ! = 0.6
!2
v
10 n = 15, ! = 0.8
v V. C ONCLUSIONS
In this paper we focused on the design of practical broad-
casting schemes based on network coding for wireless ad
10
!3 increasing c f hoc networks. First we discussed the impact of IEEE 802.11-
like random access on the performance of reactive network
coding [9]. We identied potential problems due to premature
termination of the data dissemination. To solve these, we pro-
10
!4 posed an original proactive network coding (ProNC) strategy.
0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02 0.022 0.024
We nally focused on broadcasting protocols for multi-rate ad
Packet Delivery Delay [s]
hoc networks, where we devised a lightweight rate adaptation
Fig. 12. Tradeoff between failure probability 1 − 𝑃 𝐷𝑅 and packet delivery heuristic. The designed protocols use estimates of the local
delay for a multi-rate ad hoc network with different network densities. node density to reduce the collision probability and keep a
Different curves correspond to different forwarding factors 𝜌. Each curve sufcient amount of innovative packets owing across the
is plotted for varying 𝑐𝑓 .
network. The effectiveness of our schemes was demonstrated
by simulation in a range of different scenarios.
factors 𝑐𝑓 . For this gure, nodes have on average 𝑛𝑣 = 10
neighbors, solid curves represent reactive network coding with ACKNOWLEDGMENT
xed data rate (only the extreme cases of 6 and 54 Mbps are
This paper is partly based on research presented at IEEE
plotted), whereas dotted curves represent the reliability when
ICC 2007 and IEEE GLOBECOM 2007. This work was
nodes are allowed to adapt their data rate based on the above
partially supported by DoCoMo Euro-Labs.
heuristic. First of all, we observe that transmitting with the
highest rate of 54 Mbps performs the worst. This is because
R EFERENCES
in such a case the error probability is quite high and this affects
the overall performance. As expected, a xed rate of 6 Mbps [1] P. Eugster, R. Guerraoui, A.-M. Kermarrec, and L. Massoulie, “Epidemic
information dissemination in distributed systems,” Computer, vol. 37,
leads to the best reliability performance. However, setting no. 5, pp. 60–67, May 2004.
𝑐𝑓 = 40% gets very close to this performance while leading [2] D. Tuninetti and C. Fragouli, “Processing along the way: forwarding vs.
to a better delay metric, as we can see from Fig. 11. From coding,” in ISITA, Parma, Italy, Oct. 2004.
[3] T. Ho, M. Medard, R. Koetter, D. R. Karger, M. Effros, J. Shi, and
this plot we can further appreciate the benets of adapting B. Leong, “A random linear network coding approach to multicast,”
the data rate: overall, for a coverage factor of 𝑐𝑓 = 40% we IEEE Trans. Inf. Theory, vol. 52, no. 10, pp. 4413–4430, Oct. 2006.
get the shortest delays by performing, in terms of reliability, [4] J. Yuan, Z. Li, W. Yu, and B. Li, “A cross-layer optimization framework
for multihop multicast in wireless mesh networks,” IEEE J. Sel. Areas
almost as well as the xed rate scheme with 6 Mbps. For Commun., vol. 24, no. 11, pp. 2092–2103, Nov. 2006.
the non monotonic behavior of the curves of Fig. 11, note [5] S. Katti, H. Rahul, W. Hu, D. Katabi, M. Médard, and J. Crowcroft,
that when 𝜌 is low, e.g., 𝜌 = 0.1, the dissemination of “Xors in the air: practical wireless network coding,” ACM SIGCOMM
Comput. Commun. Rev., vol. 36, no. 4, pp. 243–254, Oct. 2006.
innovative information terminates early (in this case deadlocks [6] S. Omiwade, R. Zheng, and C. Hua, “Practical localized network coding
are frequent) and a large number of nodes are unable to in wireless mesh networks,” in Proc. IEEE SECON, San Francisco,
decode all packets (𝑃 𝐷𝑅 is low, see Fig. 10). An increasing 𝜌 USA, June 2008, pp. 332–340.
[7] Y. Wu, P. A. Chou, Q. Zhang, K. Jain, W. Zhu, and S.-Y. Kung,
allows the dissemination process to continue longer, leading to “Network planning in wireless ad hoc networks: a cross-layer approach,”
fewer deadlocks, thus the delay increases with a corresponding IEEE J. Sel. Areas Commun., vol. 23, no. 1, pp. 136–150, 2005.
increase in 𝑃 𝐷𝑅. A further increase of 𝜌 at rst allows [8] H. Yomo and P. Popovski, “Opportunistic scheduling for wireless
network coding,” IEEE Trans. Wireless Commun., vol. 8, no. 6, pp.
a quicker dissemination of innovative information (shorter 2766–2770, June 2009.
delays), and afterwards the delay increases again due to [9] C. Fragouli, J. Widmer, and J.-Y. L. Boudec, “Efcient broadcasting
the increased trafc load (collisions). Fig. 12 shows tradeoff using network coding,” IEEE/ACM Trans. Networking, vol. 16, no. 2,
pp. 450–463, Apr. 2008.
results representing the failure probability 1 − 𝑃 𝐷𝑅 as a [10] C. Gkantsidis and P. R. Rodriguez, “Network coding for large scale
function of the packet delivery delay for different number of content distribution,” in Proc. IEEE INFOCOM, Miami, FL, USA, Mar.
neighbors 𝑛𝑣 ∈ {7, 10, 15}. Simulation curves are plotted for 2005, pp. 2235–2245.
[11] R. Ahlswede, N. Cai, S.-Y. R. Li, and R. Yeung, “Network information
different values of 𝜌 and each curve is obtained by varying ow,” IEEE Trans. Inf. Theory, vol. 46, no. 4, pp. 1204–1216, July
𝑐𝑓 . Good schemes are those lying on the bottom left of the 2000.
plot (i.e., having short delay and high reliability). From this [12] R. Koetter and M. Médard, “An algebraic approach to network coding,”
IEEE/ACM Trans. Networking, vol. 11, no. 5, pp. 782–795, Oct. 2003.
graph we see that for each (𝑛𝑣 , 𝜌) pair there exists a suitable [13] C. Fragouli and E. Soljanin, “Information ow decomposition for
coverage factor 𝑐𝑓 which minimizes the delay while achieving network coding,” IEEE Trans. Inf. Theory, vol. 52, no. 3, pp. 829–848,
good performance in term of 1 − 𝑃 𝐷𝑅. In practice, 𝑐𝑓 in the Mar. 2006.
[14] C. Chekuri, C. Fragouli, and E. Soljanin, “On average throughput and
range [30, 50]% are good choices as they give good reliability alphabet size in network coding,” IEEE Trans. Inf. Theory, vol. 52, no. 6,
performance while ensuring short delays for all reasonable pp. 2410–2424, June 2006.

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.
 ASTERJADHI et al.: TOWARD NETWORK CODING-BASED PROTOCOLS FOR DATA BROADCASTING IN WIRELESS AD HOC NETWORKS 673

[15] P. A. Chou, Y. Wu, and K. Jain, “Practical network coding,” in Michele Rossi received the Laurea degree in Electri-
41st Allerton Conference on Communication Control and Computing, cal Engineering (with honors) and the Ph.D. degree
Allerton, IL, USA, Oct. 2003. in Information Engineering from the University of
[16] H. Dubois-Ferriére, D. Estrin, and M. Vetterli, “Packet combining in Ferrara in 2000 and 2004, respectively. From March
sensor networks,” in Proc. ACM SenSys, San Diego, CA, USA, Nov. 2000 to October 2005 he has been a Research Fellow
2005, pp. 102–115. at the Department of Engineering of the University
[17] Y. E. Sagduyu and A. Ephremides, “Crosslayer design for distributed of Ferrara. During 2003 he was on leave at the
MAC and network coding in wireless ad hoc networks,” in Proc. IEEE Center for Wireless Communications (CWC) at the
ISIT, Adelaide, Australia, Sep. 2005, pp. 1863–1867. University of California San Diego (UCSD), where
[18] Y.-C. Tseng, S.-Y. Ni, Y.-S. Chen, and J.-P. Sheu, “The broadcast storm he performed research on wireless sensor networks.
problem in a mobile ad hoc network,” Wireless Networks, vol. 8, no. In November 2005 he joined the Department of
2/3, pp. 153–167, Mar. 2002. Information Engineering of the University of Padova, Italy, where he is an As-
[19] “ns2 simulation code for wireless network coding.” [Online]. Available: sistant Professor. Dr. Rossi is currently part of the EU-funded SENSEI project
http://telecom.dei.unipd.it/simulation and of the WISE-WAI project, both on wireless sensor networks. His research
[20] N. Baldo, F. Maguolo, M. Miozzo, M. Rossi, and M. Zorzi, “ns2- interests are centered around the dissemination of data in distributed ad hoc
MIRACLE: a modular framework for multi-technology and cross-layer and wireless sensor networks, including integrated MAC/routing schemes,
support in network simulator 2,” in Proc. ACM ValueTools, Nantes, data dissemination via network coding, the application of compressive sensing
France, Oct. 2007, pp. 1–8. techniques for the reconstruction of signals in wireless sensor networks and
[21] J. Widmer, C. Fragouli, and J.-Y. L. Boudec, “Low-complexity energy- cooperative routing policies for ad hoc networks.
efcient broadcasting in wireless ad hoc networks using network cod-
ing,” in NetCod, Riva del Garda, Italy, Apr. 2005.
[22] G. Bianchi, “Performance analysis of the IEEE 802.11 distributed Joerg Widmer is manager of the Ubiquitous Net-
coordination function,” IEEE J. Sel. Areas Commun., vol. 18, no. 3, working Research Group at DOCOMO Euro-Labs,
pp. 535–547, Mar. 2000. Munich, Germany. His research expertise covers
[23] F. Daneshgaran, M. Laddomada, F. Mesiti, and M. Mondin, “Unsat- wired and wireless networks, ranging from MAC
urated throughput analysis of IEEE 802.11 in the presence of non layer design for wireless communication, sensor
ideal transmission channel and capture effects,” IEEE Trans. Wireless networking, and network coding to transport proto-
Commun., vol. 7, no. 4, pp. 1276–1286, Apr. 2008. cols and future Internet architectures. Before joining
[24] E. Fasolo, M. Rossi, J. Widmer, and M. Zorzi, “On MAC scheduling DOCOMO, Joerg Widmer was senior researcher
and packet combination strategies for random network coding,” in Proc. at EPFL, Switzerland, working on ultra-wide band
IEEE ICC, Glasgow, UK, June 2007, pp. 3582–3589. communication and network coding. Joerg Widmer
[25] ——, “A proactive network coding strategy for pervasive wireless obtained his M.S. degree and Ph.D. degree in com-
networking,” in Proc. IEEE GLOBECOM, Washington DC, USA, Nov. puter science from the University of Mannheim, Germany in 2000 and 2003,
2007, pp. 5271–5276. respectively. In 1999 and 2000 he was a visiting researcher at the International
[26] A. Gkelias, M. Dohler, and H. Aghvami, “Throughput analysis for Computer Science Institute in Berkeley, CA, USA. He authored more than 60
wireless multi-hop CSMA,” in Proc. IEEE PIMRC, Barcelona, Spain, conference and journal papers, holds several patents, and regularly participates
Sep. 2004, pp. 979–983. in program committees of major conferences.

Alfred Asterjadhi was born in Elbasan, Albania,

in 1982. He received the Bachelor’s and Master’s Michele Zorzi was born in Venice, Italy, in 1966.
degree (summa cum laude) in Telecommunications He received the Laurea degree and the Ph.D. degree
Engineering in 2004 and 2007, respectively, from in Electrical Engineering from the University of
the University of Padova, Italy. During the Aca- Padova, Italy, in 1990 and 1994, respectively. During
demic Year 2004/2005 he attended several graduate the Academic Year 1992/93, he was on leave at the
courses of Computer Science at the Rey Juan Carlos University of California, San Diego (UCSD), attend-
University (URJC), Madrid (Spain). Since 2008 he ing graduate courses and doing research on multiple
has been a Ph.D. Student at the Department of access in mobile radio networks. In 1993, he joined
Information Engineering, University of Padova and the faculty of the Dipartimento di Elettronica e
is actively participating in national and international Informazione, Politecnico di Milano, Italy. After
research projects. His current research is focused on the design and evaluation spending three years with the Center for Wireless
of protocols for cognitive wireless networks. Active research is carried out in Communications at UCSD, in 1998 he joined the School of Engineering of the
several sub-elds such as network coding, cross-layer techniques and machine University of Ferrara, Italy, and in 2003 joined the Department of Information
learning. Engineering of the University of Padova, Italy, where he is currently a
Professor. His present research interests include performance evaluation in
Elena Fasolo received the Laurea degree in mobile communications systems, random access in mobile radio networks,
Telecommunication Engineering from the University ad hoc and sensor networks, energy constrained communications protocols,
of Padova in 2004. From January 2005 to December and cognitive radio and networks.
2007 she was a Ph.D. Student at the School of Infor- Dr. Zorzi was the Editor-In-Chief of the IEEE W IRELESS C OMMUNICA -
mation Engineering at the Information Engineering TIONS M AGAZINE from 2003 to 2005, is currently the Editor-In-Chief of the
Department, University of Padova, Italy. During this IEEE T RANSACTIONS ON C OMMUNICATIONS, and serves on the Steering
period, she spent six months at DoCoMo Euro-Labs Committee of the IEEE T RANSACTIONS ON M OBILE C OMPUTING, and
in Munich, Germany, working on network coding on the Editorial Boards of the the W ILEY J OURNAL OF W IRELESS C OM -
techniques for wireless networks. On March 2008 MUNICATIONS AND M OBILE C OMPUTING and the ACM/URSI/K LUWER
she defended her Ph.D. thesis on network coding J OURNAL OF W IRELESS N ETWORKS . He was also guest editor for special
strategies for wireless ad-hoc networks. She now issues in the IEEE P ERSONAL C OMMUNICATIONS M AGAZINE (Energy Man-
works at IKS, Padova, as Application and Network Performance Optimization agement in Personal Communications Systems) and the IEEE J OURNAL ON
Specialist. S ELECTED A REAS IN C OMMUNICATIONS (Multi-media Network Radios).

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on March 03,2010 at 15:45:05 EST from IEEE Xplore. Restrictions apply.