Paper ID

156 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 42, NO.
1, JANUARY 2017
Throughput-Efficient Super-TDMA MAC

Transmission Schedules in Ad Hoc Linear
Underwater Acoustic Networks
Said Lmai, Mandar Chitre, Senior Member, IEEE, Christophe Laot, Senior Member, IEEE,
and Sebastien Houcke, Member, IEEE
Abstract—Underwater acoustic (UWA) sensor network deploy- I. I NTRODUCTION

ments may be used in many applications for environmental,
scientific, military, and commercial purposes. Several topologies
are in use, but the most widely used topology is linear. Indeed,
placing nodes on a single line offers more opportunities in terms
I N RECENT years, underwater acoustic (UWA) sensor net-
work deployments have been explored for use in many
applications: offshore survey and exploration, costly infras-
of large coverage and high-rate services. The UWA channel is a tructure monitoring, surveillance in sensitive areas, and other
shared medium. Thus, a medium access control (MAC) protocol
is necessary, primarily to regulate and coordinate nodes’ access.
more general scientific, military, and commercial missions. To
MAC protocol design should take into consideration large prop- monitor or explore the largest possible area, a network with
agation delays to favor better network throughput. Performance a linear geometry (wherein the nodes are placed on a single
of most developed protocols in linear topologies does not exceed line) may be desirable. As in all shared-medium networks, a
1 in terms of normalized network throughput, or equivalently, medium-access control (MAC) protocol is necessary, primar-
channel utilization. We explore transmission schedules in three
important contexts. 1) Single collision domain with unicast traffic.
ily to regulate and coordinate medium access, i.e., how and
In an N -node network, we develop transmission schedules achiev- when nodes access the UWA channel to be able to effectively
ing a normalized network throughput of 2 − (2/N ). This is the use this shared medium. Due to the low speed of sound in
best that can be done in such a context, as demonstrated using water (about 1500 m/s), the propagation delays in UWA net-
a general greedy approach combined with an exhaustive search works are inherently large, i.e., comparable to or even greater
for small-size networks. 2) Single collision domain with broad-
cast traffic. We propose a periodic per-node fair schedule with the
than the packet duration. In an environment with such fea-
shortest period. Achievable throughput in such conditions is close tures, and in view of the linear network architecture, our main
to N/2. Likewise, we prove that the throughput is upper bounded objective is to propose transmission schedules that exploit the
by N − 1 under the per-node fairness constraint. 3) Partially large propagation delays and make it possible to obtain a desir-
overlapping collision domains with unicast traffic. We consider a able level of network throughput, or high channel utilization.
simple illustration of such a configuration. The proposed trans-
This can be achieved by allowing simultaneous transmissions
mission schedule depicts a scenario where messages originate at
one end of the network, and are sequentially relayed node by node within the network as often as possible, and concentrating
(i.e., hop by hop) in the direction of the final destination located interferences at unintended nodes as much as possible. Time-
at the other end of the network. Furthermore, for all three dis- division multiple-access (TDMA)-based solutions offer more
cussed contexts, we build up computationally efficient algorithms valuable opportunities in this environment, as compared to
that generate transmission schedules regardless of network size.
handshaking protocols and acknowledgment-based retransmis-
We explore the idea of exploiting nonzero propagation delays for
linear topologies to improve network throughput. In recent UWA sion schemes. Furthermore, contention-based MAC protocols
sensor networks, the linear topology is a fundamental component are usually more energy consuming than their TDMA coun-
that may be used to build more complex network configurations. terparts since they waste energy during collisions and idle
This study would then serve as a base for future research into this listening [1].
area.
In practice, some application scenarios involve unicast deliv-
Index Terms—Ad hoc networks, large propagation delays, lin- ery within a fully connected zone or equivalently within a single
ear topology, throughput bound, time-division multiple-access collision domain. A simple example is the sensor network case
(TDMA)-based protocol, underwater sensor networks.
where each node has to transfer the data it has collected to
Manuscript received July 02, 2014; revised March 06, 2015, October 23, specific destination nodes. Other applications may bring nodes
2015, and December 20, 2015; accepted February 26, 2016. Date of publication
May 02, 2016; date of current version January 11, 2017.
to share navigation information, command and control system
Associate Editor: M. Stojanovic. instructions, or even predefined messages for node safety. This
S. Lmai is with the Royal Moroccan Naval Academy (Ecole Royale Navale), requires the use of broadcast traffic. When the extent of the
Casablanca, Morocco (e-mail: said.lmai.lmai@gmail.com). area to be covered gets larger, it becomes necessary to use mul-
M. Chitre is with the Acoustic Research Laboratory, Tropical Marine Science
Institute, Singapore 119227, Singapore (e-mail: mandar@arl.nus.edu.sg).
tihop relaying where a long distance is divided into multiple
C. Laot and S. Houcke are with the Département Signal & Communications, shorter hops in order to gain more advantageous communi-
Telecom Bretagne, Plouzané 29200, France (e-mail: christophe.laot@telecom- cation conditions. In fact, the attenuation of the UWA signal
bretagne.eu; sebastien.houcke@telecom-bretagne.eu). depends exponentially on propagation distance, and also on
Digital Object Identifier 10.1109/JOE.2016.2537659
0364-9059 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on October 18,2023 at 13:58:48 UTC from IEEE Xplore. Restrictions apply.
LMAI et al.: THROUGHPUT-EFFICIENT SUPER-TDMA MAC TRANSMISSION SCHEDULES IN AD HOC LINEAR UNDERWATER ACOUSTIC NETWORKS 157
frequency. Thus, bandwidth and power are severely limited throughput superior to the optimal performance obtained when
[2]. The arrangement of nodes consequently leads to partially the propagation delay is zero (i.e., normalized throughput of
overlapping collision domains. 1). An ad hoc handshaking-based scheme was also consid-
We consider the scheduling of transmissions between all ered in [4]. Using time information carried by control packets,
nodes within a linear N -node network in three separate con- range between different communicating nodes is estimated.
texts. By exploiting information on propagation delay, we Afterwards, an algorithm employs this delay information for
propose throughput-efficient TDMA transmission schedules. data packet scheduling at each node. According to the node’s
The main contributions of this paper are summarized as follows. current schedule, RTS/CTS exchange is used to plan transmis-
• Context 1: Single collision domain with unicast traffic. sion of future data packets. For all the variants studied, this
– We propose a transmission schedule that achieves protocol achieves normalized throughput less than 1. Kredo
a normalized network throughput of 2 − (2/N ). A et al. [5] use node position information to overlap commu-
greedy search method allows us to achieve this high nications without collisions to increase channel utilization.
level of network performance. Likewise, they allow multiple messages to arrive simultane-
– Once the scheduling problem is formulated with ously at an unintended node. Performance in terms of normal-
respect to throughput, we develop a practical com- ized throughput, as defined by the authors, never reaches 1,
putationally efficient algorithm that makes it possi- irrespective of the network topology adopted. However, if we
ble to carry out the design of the aforementioned assume the final destination node in the third context discussed
transmission schedule. This algorithm describes an to be a sink node, our proposed schedule achieves a throughput
approximate dynamic programming solution that is of 1. The study in [6] essentially highlights the value of adopt-
insensitive to network size. ing bidirectional, concurrent transmissions in UWA networks
• Context 2: Single collision domain with broadcast traffic. as regards throughput and end-to-end packet delay. The authors
– Periodic per-node fair schedule with the short- adopt a packet bursting approach, where a sender–receiver pair
est period is presented. Achievable throughput in can exchange multiple rounds of bidirectional packet transmis-
such conditions is demonstrated through a greedy sions. They propose an asynchronous handshaking-based MAC
approach. We propose a fast, effective algorithm that protocol, but because of limiting factors such as idle backoff
allows generation of the transmission schedule. durations, packet collisions, and control packet exchanges, best
– As compared to the aforementioned broadcast performance in terms of the achieved network throughput nor-
schedules, we introduce illustrations for multi- malized by the theoretical maximum throughput is again less
cast schedules with good connectivity and better than 1. Long propagation delay of acoustic signals is also used
throughput. Having shorter periods, the pro- in [7] to facilitate concurrent transmissions and receptions. In a
posed schedules perform more frequent network fully connected network with one final destination for all nodes,
transmissions. MAC transmission scheduling has been explored for maximum
– We prove that N − 1 is the upper bound for the network throughput. Apart from its complexity, the optimal
throughput in the case of per-node fair access. solution found for the model formulated leads to normalized
Furthermore, we propose per-node fair schedules network throughput values that are upper bounded by 1. The
that may move toward this upper bound as closely as same idea is exploited by Diamant and Lampe [8] to develop a
desired, assuming no constraints on the time extent. broadcast transmission schedule for spatial reuse TDMA. The
Likewise, schedules without fairness constraints are motivation behind this design is still to improve throughput by
also presented. exploiting network topology information. However, the demon-
• Context 3: Partially overlapping collision domains with strated performance in terms of packets per time slot does not
unicast traffic. exceed 1.
– Considering a simple illustration of such a con- A linear topology for multihop networking in UWA sys-
figuration, we propose a scheme where messages tems has been considered to some extent in the literature
originating at one end of the network are sequen- within the last ten years (e.g., [9]–[12]). Most of the inter-
tially relayed node by node (i.e., hop by hop) in the est therein was on physical link capability. Nevertheless, the
direction of the final destination located at the other authors did not use network-oriented metrics to characterize
end of the network. network performance and find limits. Overlapping transmis-
– We present a computationally efficient algorithm sions in linear topology also drew interest in the analysis in
that generates the transmission schedule regardless [2]. Explicitly, in a slotted model for transmissions, each time
of network size. slot is used by certain nodes to transmit while others receive.
The main idea of taking advantage of large propagation With respect to the network model considered therein, traf-
delays has been taken into consideration in many studies. fic originating from one source node at one end is eventually
In [3], Peleato and Stojanovic consider a request-to-send received by a destination node at the other end of the network.
(RTS)/clear-to-send (CTS)-based protocol for ad hoc networks. Taking into account interhop interference, achievable informa-
This distance-aware protocol makes it possible to increase over- tion rates are optimized over the signal power spectral density,
all system efficiency by decreasing the average control packet versus per-node power. Nevertheless, no network-oriented per-
size. In fact, each node uses different control packet sizes for formance has been explored. In a UWA sensor network with
different destination nodes. The gain achieved does not make linear topology, Chen et al. [13] introduce TDMA-based MAC
158 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 42, NO. 1, JANUARY 2017
protocol allowing concurrent transmissions in the same time networks with broadcast traffic. We provide explicit and attain-
slot. This protocol uses heuristic strategies to improve chan- able bounds, and propose a fast algorithm that allows construct-
nel utilization, calculated as the number of transmissions within ing the transmission schedule we suggest first. A representative
the network per time unit. Performance evaluation has focused instance of the third context related to partially overlapping col-
on medium-sized networks. For instance, channel utilization lision domains with unicast traffic is depicted in Section V.
does not exceed 3/2 in the 11-node network. However, in this Finally, Section VI concludes the paper.
paper, using the same network structure, we explore transmis-
sion schedules with achievable network throughput that may
II. C ONTEXT AND S YSTEM M ODEL
increase proportionally with network size. A multihop linear
topology has been studied in [14] under fair access criterion for A. Delay Matrix
all nodes. The sensor network model consists of one base sta- In a nonzero propagation delay environment, we consider a
tion as the final destination of all the network traffic. Sensor regular N -node linear network where each node is identified
nodes are placed along a line with identical distance between by i s.t. i ∈ N∗ and 1 ≤ i ≤ N . Let ri be the position vector of
every two neighboring nodes. The transmission range for each node i in 3-D Euclidean space. The propagation delays between
node is assumed to be just one hop, while the interference range every pair of nodes may be expressed using a delay matrix [19]
is less than two hops. In this paper, message duration is set to denoted by D. The time is assumed to be slotted. Thus
be equal to the one-hop propagation delay for analytical sim-
|ri − rj |
plicity. With this network model formulation, Xiao et al. [14] Dij = , 1 ≤ i, j ≤ N (1)
proved that under a per-node fair access constraint, an asymp- cτ
totic upper bound in terms of overall network utilization is where c is the signal propagation speed and τ is the length of
1/2. However, they do not provide transmission schedules that one time slot. Hence, the entries in the delay matrix are non-
achieve this upper bound. Though we assume the interference negative real numbers. Furthermore, we consider in this paper a
range in the third context discussed to be twice the transmission regular N -node linear network with unit spacing between every
range, we present a 1/2-network utilization (as has been defined pair of neighboring nodes. Indeed, for the sake of analytical
in [14]) transmission schedule. simplicity, we have also assumed that the time slot duration is
The time interference alignment concept, also referred to as set to the propagation delay between two neighboring nodes.
interference alignment by delay, is based on the main idea of As shown next, there is no loss of generality in this assump-
overlapping interferences at selected nodes during a given time, tion. Provided that we seek to improve channel utilization, in
and trying to allow the nodes in question to use these periods for the case where slots are an integer fraction 1/K (K > 1) of
transmission. Among the studies that have explored this con- unit propagation delay, one can consider, for instance, an arbi-
cept, we cite [15]–[18]. However, the analysis has been taken trary transmission cycle where every node i transmits mi times
a step further by Chitre et al. [19]. They prove that N/2 is over the duration of one cycle T cycle (which is not necessarily a
an upper bound on the maximum achievable throughput in an multiple of K). By repeating the cycle K times, we end up with
N -node network with nonzero propagation delay. In addition, a transmission schedule having a time extent of Π = KT cycle ,
they pinpoint some geometries where developed transmission during which every node i transmits mi K times. This can be
schedules achieve this upper bound. Likewise, they prove the seen as node i transmitting during mi units propagation delay.
existence of a throughput maximizing optimal schedule that To make things simpler, an illustration where all slots are well
is periodic, for each network. We capitalize on the invaluable occupied consists in transmitting K successive messages for
findings in [19] to conduct a more sharply focused study. In the each node during one unit propagation delay. This could be
context of a linear topology, we carry out an in-depth analysis to reduced to transmit one message of a duration equal to one unit
explore to what extent we can exploit large propagation delays propagation delay, as assumed in our model for the sake of ana-
in consideration of traffic policy, collision domain extent, and lytical simplicity, while one may, furthermore, make gains on
certain fairness constraints. We can see that almost all the work overheads of K messages. Therefore, no additional insight is
cited involves multihop relaying policies within the network. provided by considering the general case with slots that are an
In contrast, we study various scenarios applicable to differ- integer fraction of unit propagation delay in a linear network
ent UWA applications. The three contexts considered provide with regularity in the distances separating nodes. The delay
explicit segmentation of the analysis where we expect to deliver matrix entries are then natural numbers and Dij becomes
a far deeper understanding and promising results. Note that our Dij = |i − j| ∀i, j. (2)
focus in this paper is on the MAC sublayer. Given the regular
linear topology, we demonstrate to what extent a TDMA-based The linear geometry of the network under consideration is
MAC protocol can achieve high network throughput in the case fully described by the delay matrix. Besides, D is symmet-
of the physical link being reliable. ric, i.e., Dij = Dji , since |i − j| = |j − i|, and has an all-zero
The remainder of the paper is organized as follows. The diagonal, i.e., Dii = 0 (since |i − i| = 0). The delay matrix cor-
general context and system model are described in Section II. responding to the three-node regular linear network is given
The first context is widely discussed in Section III, which by D3
⎡ ⎤
focuses on network throughput in regular linear networks with 0 1 2
unicast traffic. In addition, we develop an efficient problem- D3 = ⎣ 1 0 1 ⎦ . (3)
solving algorithm. Section IV explores achievable throughput 2 1 0
with reference
Authorized to different
licensed use limitedfairness constraints,
to: VIT University- inCampus.
Chennai regularDownloaded
linear on October 18,2023 at 13:58:48 UTC from IEEE Xplore. Restrictions apply.
TABLE I
S UMMARY OF P LANNED T RANSMISSIONS AND C ORRESPONDING R ECEPTIONS IN ACCORDANCE W ITH S(3)
B. Network Environment Features simply throughput) Y is the total number of information bits
All nodes in the network are operating in half-duplex mode, successfully received by all nodes in the network per unit time,
i.e., a node cannot simultaneously transmit and receive. In a sin- normalized by the link data rate ν.
gle collision domain, or equivalently a fully connected network, Provided that the message duration is equal to τ , which
a message transmitted by a node reaches all other nodes in the is also the propagation delay between two adjacent nodes,
network after the corresponding propagation delays. With ref- we define a transmission schedule (or simply schedule) S
erence to the traffic type, unicast transmission identifies with a as the matrix that determines when each node transmits and
message sent from a single source node to a single destination receives messages [19]. As regards the entries of S, we have
node, while multicast transmission is used to refer to a message the following:
sent from a single source node to a set of destination nodes • Sit = Si,t = j > 0 indicates that node i transmits a mes-
within the network. If this set of destination nodes includes all sage to node j at time slot t;
• equivalently, in the case of multicast delivery, Sit =
nodes other than the transmitting node, we deal with broadcast
Si,t = jkl > 0 signifies that node i transmits a message
traffic.
In this paper, we consider three scenarios of interest: single to nodes j, k, and l at time slot t;
• Sjt = Sj,t = −i < 0 implies that node j receives a mes-
collision domain with unicast traffic, single collision domain
sage from node i during the time slot t;
with broadcast traffic, and a simple illustration of partially
overlapping collision domains with unicast traffic. Environment • in all other cases, node i is designated as an idle node
during time slot t, which is represented by Sit = Si,t = 0.
monitoring and tactical surveillance in a specific zone, as well
as many other applications, are carried out through underwater If Si,t+T = Sit , ∀i, t, this indicates that the transmission
sensor networks where the first and second scenarios are quite schedule is repeating with a period T . It is said to be periodic
with period T and may be depicted using an N × T matrix S(T )
appropriate. In view of the numerous communication link chal-
lenges in a doubly selective underwater acoustic channel, the where
(T )
transmit power for each node is generally high enough to reach Sit = Si,t(modT ) . (4)
the whole network. If the extent of the monitored area becomes
wider, the third scenario is more pertinent. By way of illustration, we might consider the following trans-
In the case of the first scenario, excluding the source and des- mission schedule in the three-node network with unicast traffic:
⎡ ⎤
tination nodes, a message is considered as an interference at all 3 2 −3
nodes it reaches. On the other hand, in the second scenario, a S(3) = ⎣ 3 0 −1 ⎦ . (5)
message is considered to be an interference at a node only if it 1 −2 −1
overlaps in time with another message intended for that node.
A simple explanation of the transmissions and the correspond-
However, in the context of the third scenario, it becomes nec-
ing receptions handled in this schedule is reported in Table I. If
essary to cope with two types of interference constraints, both
the columns of matrix S(3) are circularly shifted to the right or
related to the strength of the detected signal, i.e., whether it is
to the left, the resulting matrix describes the same transmission
high enough to decode. A message is not only an interference
schedule.
at all nodes between the source and destination nodes within
From schedule S(T ) with period T , the average network
the collision domain in question. It is also an interference at the
throughput can be calculated considering the number of recep-
surrounding nodes, where the signal is not sufficiently strong to
tions in schedule S(T )
be decoded, while it is still detectable.
1 (T )
T −1 N
Y = 1 Sit < 0 (6)
C. Throughput and Transmission Schedule T t=0 i=1
Assuming the physical link to be reliable (error-free) with where 1(E) is the indicator function of event E, with value 1 if
constant data rate ν, the loss of a message is only due to col- E is true and 0 otherwise.
lision. A collision is said to occur at a certain node if two or
more messages overlap in time, while a successful transmis-
D. Fair Schedule Types
sion is defined as a transmission that results in a successful
reception (or receptions when it comes to multicast traffic) of The set of different access opportunities between nodes
the message at the destination node (or nodes). The normalized should be taken into consideration, such that the access dur-
network throughput (also designated by network throughput or ing one period T is done in a certain fair way. The objective
is always to ensure a high network utilization, i.e., a high

throughput. We use the terminology of fairness as introduced
in [19].
When a schedule S(T ) provides the same number of trans-
mission opportunities to all nodes, S(T ) is said to be per-node
fair. This signifies that
T−1
(T )
1 Sit > 0 = constant > 0 ∀i. (7)
t=0
However, with respect to unicast traffic, we can add the stronger

fairness characteristic of transmitting to all other nodes at the
same rate. This allows us to refer to S(T ) as a per-link fair
schedule. Correspondingly
T
−1
(T )
1 Sit = j = constant > 0 ∀i, j = i. (8)
t=0
In spite of that, we may obtain a lower level of fairness by offer-

ing to all nodes the opportunity to transmit to all other nodes at
least m > 0 times during one period T , which would imply
T
−1
(T )
1 Sit = j ≥ m ∀i, j = i. (9)
t=0
Such a transmission scheme might be referred to as a weakly Fig. 1. Depiction of the network state when node i transmits to node j (i < j).
per-link fair schedule. Likewise, a weakly per-node fair sched-
ule can be used to describe the case where all nodes have the Lemma 1: In a regular N -node linear network with a single
opportunity to transmit at least n > 0 times during one period collision domain and unicast traffic, a transmission from node
T , i.e., i to node j at time slot t generates 2(N − 1 − Dij ) restricted
T
−1 entries in a transmission schedule matrix under development,
(T )
1 Sit > 0 ≥ n ∀i. (10) along with Dij − 1 advantageous entries.
t=0 Proof: Node i transmits a message to node j during time
A fairness criterion as has been introduced will be one of slot t only if node j is able to successfully receive the message
the main components considered in developing algorithms that during time slot t + Dij . Therefore
generate transmission schedules of interest.
Sit = j ⇔ Sj,t+Dij = −i ∀i = j. (11)
III. T HROUGHPUT IN A R EGULAR L INEAR N ETWORK Furthermore, to ensure the successful reception of a transmitted
WITH A S INGLE C OLLISION D OMAIN message at time slot t, it is required that no other nodes transmit
AND U NICAST T RAFFIC messages that arrive at node j during time slot t. Thus
A. Achievable Network Throughput and Illustrative
Sjt = −i ⇒ Sk,t−Djk ≤ 0 ∀k = i (12)
Transmission Schedules
1) Achievable Network Throughput: Let us first clarify Let us assume that i < j.
what is meant by a useful message. Once transmitted from a First category of restrictions generated: Once emitted at t,
source node i to a destination node j, the message is consid- the message of node i will transit within the network in both
ered to be useful while it does not reach yet the destination, directions. Then, it is a negative interference (useless message)
i.e., certainly it is an interference (positive interference) on all on each node k, 1 ≤ k < i, at future time slot t + i − k. Hence,
nodes that are in between i and j, but it carries useful informa- i − 1 nodes will be affected, due to this transmission, after t.
tion intended for j, unlike its zero utility (negative interference) In terms of the transmission schedule matrix under develop-
at all the other nodes outside the segment i ↔ j, where it is ment, each of the corresponding i − 1 entries is impacted by
rather a useless message. In Fig. 1, an effective useful message one useless message, and the best that can be done to increase
is referred to as a confirmed useful message. the network throughput is to allow the transit of a useful mes-
In a transmission schedule matrix under development, an sage (once it is possible). In a similar way, after reaching node
entry is said to be advantageous if it can be subject to two j at t + Dij [according to (11)], the message sent by node i
overlapping useful messages traveling in opposite directions. is a negative interference (useless message) on each node l,
However, an entry is said to be restricted if there is already j < l ≤ N at future time slot t + Dij + l − j. Thus, this nega-
a useless message (negative interference) such that it can tive interference created by the transmission in question will get
accommodate a maximum of one single useful message. at N − j additional nodes. In the transmission schedule matrix
under development, each of N − j related entries is inevitably to all nodes of the network during the first time slot t. In doing
impacted, and it would be desirable to be able to transmit. The so, we will have N useful messages in the network. However,
total number of resulting restrictions due to a useless message according to Lemma 1, any transmission from node i to node
is N − j + i − 1 = N − Dij − 1, since Dij = j − i. A graph- j results in 2(N − Dij − 1) restricted entries in the transmis-
ical illustration of this first category of restrictions is given in sion schedule matrix, as well as Dij − 1 advantageous entries.
Fig. 1, where the second category is also depicted. Therefore, three key elements should be taken into account.
Second category of restrictions generated: Furthermore, let First, for each transmission, selected Dij should be as high
us focus on time slots before t. Based on (12), each node as possible to minimize negative interference and maximize
k, 1 ≤ k < i, should not be transmitting at previous time slot potential upcoming transmissions. Second, generated negative
t − i + k. In the best case scenario, one would then expect interferences are to be managed in such a way that they leave
the reception (or the transit) of a useful message at the corre- the network as rapidly as possible, and ideally, are replaced by
sponding entry. This concerns i − 1 entries in the transmission useful messages. Last, consider condensing receptions to take
schedule matrix. Additionally, the same constraint is applied to the shortest possible time.
each node l, j < l ≤ N , at time slot t + Dij − l + j, before Let us consider an N -node network where N is even.
the scheduled time slot for reception, i.e., t + Dij . N − j sup- At time slot 1, for each node i, 1 ≤ i ≤ (N/2), we select
plementary entries are implied. Thus, the second category of node N as the j destination node to maximize the distance
restrictions involves N − j + i − 1 = N − Dij − 1 entries in Dij . As regards nodes k where (N/2) + 1 ≤ k ≤ N , they have
the transmission schedule matrix. node 1 as their destination. Thus far, we have at 1, (N/2) con-
It should be noted that each of the total 2(N − Dij − 1) tiguous useful transmissions on the right half of the network,
entries subject to restrictions may accommodate a maximum of oriented to the center, as well as N/2 other contiguous useful
one useful message, being transmitted or transiting. In contrast, transmissions on the left half. N simultaneous transmissions
each of the intermediate entries corresponding to j − i − 1 = [(N/2) in one direction and (N/2) in the other] give rise to
Dij − 1 intermediate nodes m, i < m < j, is idle at time slot N − 2 negative interferences, since the nodes at the edges do
t + m − i with one useful message in transit. However, it can not generate negative interference, while the others generate
potentially accommodate a second useful message as described one negative interference each. However, it should be noted
in Fig. 1, as long as the ultimate goal is to increase the network that all ((N − 2)/2) negative interferences on the right half of
throughput. the network are oriented to the right end (i.e., extremity node
In sum, transmission from i to j generates 2(N − Dij − N ), which is the nearest end. In parallel, the situation is the
1) restricted entries in the transmission schedule matrix, in same on the left half, with negative interferences pointing to
addition to Dij − 1 advantageous entries. the left end (i.e., extremity node 1). Therefore, all negative
The result is the same when i > j. interferences will, in pairs, completely clear out the network
Theorem 1: In a regular N -node linear network with a single at (N/2). One pair evacuating at each running time slot 1 + e,
collision domain and unicast traffic, the maximum achievable 1 ≤ e ≤ (N/2) − 1, is the maximum we can achieve in terms
throughput is 2 − (2/N ). of rapid negative interference evacuation. Nevertheless, it is
Proof: The general forms of schedules for N even and N important to replace each evacuating negative interference by
odd are summarized as S(N ) and S(N ) , respectively. However, a new useful message as long as we aim to reach Net opt . To
let us first show how we built these solutions. Given the linear do so, we need to look for transmissions that do not intro-
geometry, the network can be considered as a set of N − 1 adja- duce additional negative interference, and try to schedule the
cent sections. Each section consists of two neighboring nodes. corresponding receptions within a time horizon not exceeding
Let us use exchange state to denote the activity within the net- the last time slot planned for the reception of previous trans-
work in terms of messages being transmitted, received, or still missions. Only nodes 1 and N can transmit without causing
circulating in the network. At a certain time slot, the network negative interference. Thus, the following transmissions follow
exchange state that hosts the maximum traffic of useful mes- the general form of the schedule summarized in (13), shown at
sages is the exchange state where each section accommodates the bottom of the next page, as S(N ) .
two messages simultaneously. It obviously comes to two useful In the case where N is odd, the proposed transmission
messages transiting in opposite directions, or equivalently one scheme is slightly modified compared to that for N even. The
message transiting and the other being received, but still with changes are identified in the general form of schedule S(N ) , as
two flying useful messages on each of the two surrounding sec- shown in (14), at the bottom of the next page. S(N ) and S(N )
tions (if they exist). Actually, this last case concerns only the are weakly per-link fair, since each node transmits at least once
two sections at the ends of the network. Let Net opt denote this during one period N .
optimum exchange state. As we are interested in the presence of useful messages in the
While seeking to maximize the network throughput, the network and their receptions, let us follow up on the evolution
transmission schedule should ensure that Net opt is reached as of the transmissions until the network reaches Net opt , and then
quickly as possible, and then allow the receptions to likewise the receptions until all messages reach their respective destina-
be performed as quickly as possible. 2(N − 1) is the number tions. Table II describes the progress of this process for N even.
of messages corresponding to Net opt , and since the network From S(N ) and S(N ) , we notice that the period duration
includes only N nodes, this state cannot be reached after one is N and the number of transmissions within one period
time slot. It should be achieved gradually. Let us grant access is N + 2((N/2) − 1) = 2(N − 1). Hence, the corresponding
TABLE II needed is prohibitively long. We found that 2 − 2/N is the

OVERVIEW OF THE N ETWORK E XCHANGE -S TATE
best that can be done in terms of the normalized network
E VOLUTION OVER T IME (N E VEN )
throughput.
2) Illustrative Transmission Schemes: To have a clear view
of the proposed transmission schedules, we provide illustrative
examples. For instance, let us consider four-node and five-
node regular linear networks. The corresponding schedules are
represented by S(4) and S(5) , respectively
⎡ ⎤
4 3 −3 −4
⎢ 4 0 0 −4 ⎥
S(4) = ⎢
⎣ 1 0 0 −1 ⎦
⎥ (15)
1 2 −2 −1
⎡ ⎤
5 4 −3 −4 −5
⎢5 0 0 0 −5 ⎥
⎢ ⎥
S(5) =⎢
⎢1 0 0 0 −5 ⎥
⎥. (16)
⎣1 0 0 0 −1 ⎦
1 2 3 −2 −1
It is worth mentioning that the proposed transmission sched-
ule, achieving 2 − (2/N ) network throughput, is not unique.
Several schedules reaching the same performance can be found,
network throughput in such an N -node linear network is 2 − especially for high values of N . To put it in context, let us
2/N . It is also worth pointing out that, for validation and examine the example of seven-node network. According to
strengthening purposes, we performed the exhaustive search the approach discussed above, the corresponding transmission
(7)
over small-size networks (3 ≤ N ≤ 5), although the time schedule is depicted by S<1> . The index < 1 > is used to
1 2 ··· e ··· N/2 (N/2) + 1 (N/2) + 2 ··· (N/2) + f ··· N −1 N

⎡ ⎤
1 N N − 1 · · · N − e + 1 · · · (N/2) + 1 −((N/2) + 1) −((N/2) + 2) · · · −((N/2) + f ) · · · −N + 1 −N
⎢ ⎥
2 ⎢N 0 ··· 0 ··· 0 0 0 ··· 0 ··· 0 −N ⎥
⎢ ⎥
.. ⎢ .. .. .. .. ⎥
. ⎢ . . . . ⎥
⎢ ⎥
⎢ ⎥
(N/2) − 1 ⎢ N 0 0 −N ⎥
S (N )
= N/2 ⎢ ⎥
⎢N 0 0 −N ⎥
⎢ ⎥
(N/2) + 1 ⎢⎢ 1 0 0 −1 ⎥
⎥
⎢ ⎥
(N/2) + 2 ⎢ 1 0 0 −1 ⎥
⎢ ⎥
.. ⎢ . .. .. .. ⎥
. ⎢ .. . . . ⎥
⎢ ⎥
⎢ ⎥
N −1 ⎣ 1 0 ··· 0 ··· 0 0 0 ··· 0 ··· 0 −1 ⎦
N 1 2 ··· e ··· N/2 −(N/2) −(N/2) + 1 · · · −(N/2) + f − 1 · · · −2 −1
(13)
1 2 ··· e · · · (N − 1)/2 (N + 1)/2 (N + 3)/2 ··· (N − 1)/2 + f ··· N −1 N

⎡ ⎤
1 N N − 1 · · · N − e + 1 · · · (N + 3)/2 −(N + 1)/2 −(N + 3)/2 · · · −((N − 1)/2) − f · · · −N + 1 −N
⎢ ⎥
2 ⎢N 0 ··· 0 ··· 0 0 0 ··· 0 ··· 0 −N ⎥
⎢ ⎥
.. ⎢ .. .. .. .. ⎥
. ⎢ . . . . ⎥
⎢ ⎥
⎢ ⎥
(N − 1)/2 ⎢N 0 0 −N ⎥
S(N ) = (N + 1)/2 ⎢ ⎥
⎢ 1 0 0 −N ⎥
⎢ ⎥.
(N + 3)/2 ⎢ 1 0 0 −1 ⎥
⎢ ⎥
⎢ ⎥
(N + 5)/2 ⎢ 1 0 0 −1 ⎥
⎢ ⎥
.. ⎢ . .. .. .. ⎥
. ⎢ .. . . . ⎥
⎢ ⎥
⎢ ⎥
N −1 ⎣ 1 0 ··· 0 ··· 0 0 0 ··· 0 ··· 0 −1 ⎦
N 1 2 ··· e · · · (N − 1)/2 (N + 1)/2 −(N − 1)/2 · · · −((N − 1)/2) + f − 2 · · · −2 −1
(14)
denote the first alternative derived network is given by

⎡ ⎤
7 6 5 −4 −5 −6 −7 A = max Dij . (20)
⎢7 0 0 0 0 0 −7 ⎥ i,j
⎢ ⎥
⎢7 0 0 0 0 0 −7 ⎥
⎢ ⎥ In the case of a regular N -node linear network, the correspond-
S<1> = ⎢ −7 ⎥
(7)
⎢1 0 0 0 0 0 ⎥. (17) ing size is A = N − 1.
⎢1 0 0 0 0 0 −1 ⎥
⎢ ⎥ As in [19], we formulate the problem of finding an opti-
⎣1 0 0 0 0 0 −1 ⎦ mal schedule, denoted by S∗ , as a sequential decision problem.
1 2 3 4 −3 −2 −1 Since every network has an optimal schedule that is periodic
(7) [19], we are interested in schedule S∗ with period T . The cur-
With minor modifications we can come up with S<2> rent state of the system is known and represented by S{t} ,
(7)
and S<3> also performing a network throughput of (12/7). which is the partial schedule given all transmissions occurring
Actually, the changes made in terms of receptions affect idle between time slots t − A and t − 1. It should be remem-
entries in the transmission schedule matrix that are impacted by bered here that a transmission will not remain in the network
a single interference (as opposed to at least two interferences in more than A time slots. Indeed, S{t} contains all transmissions
opposite directions) made at t − 1 and A − 1 time slots earlier. However, result-
⎡ ⎤ ing receptions and interference at later time slots are taken into
7 6 5 −4 −5 0 0
⎢ 7 0 0 0 −6 −7 −7 ⎥ consideration in the current state S{t} . Let S denote the discrete
⎢ ⎥ state space of S{t} . Node i is said to be available at time slot t,
⎢ 7 0 0 0 0 0 −7 ⎥
⎢ ⎥ if i is neither the intended recipient nor subject to paralyzing
S<2> = ⎢ ⎥
(7)
⎢ 1 0 0 0 0 0 −7 ⎥
⎢ 1 0 0 0 0 0 −1 ⎥ interference. By paralyzing interference, we mean two or more
⎢ ⎥ unintended messages in opposite directions, which impede the
⎣ 2 0 0 0 0 0 −1 ⎦
{t}
2 2 3 4 −3 −2 −1 node from transmitting. At time slot t, the decision xi to be
taken at each available node i involves granting access to i,
⎡ ⎤ {t} {t}
6 6 5 −4 −5 −6 −7 i.e., xi > 0, or maintaining i idle, i.e., xi = 0. Let X (S{t} )
⎢7 0 0 0 0 0 0 ⎥ denote the set of all possible decisions (also referred to as
⎢ ⎥
⎢ 7 0 0 0 0 −7 −7 ⎥ actions), whereas X designates the decision space. The effect of
⎢ ⎥
S<3> = ⎢ ⎥
(7) {t} {t} {t}
⎢ 1 0 0 0 0 0 −7 ⎥ . (18) action x{t} = (x1 , x2 , . . . , xN ) is defined in the following
⎢ 1 0 0 0 0 0 −1 ⎥
⎢ ⎥ state transition equation:
⎣ 1 0 0 0 0 −1 −1 ⎦
1 3 3 4 −3 −2 0 S{t+1} = Δ S{t} , x{t} (21)
Moreover, unlike N -node networks with N even, the proposed

which produces the new state S{t+1} . In other words, with
transmission schedule for N odd is unbalanced. Specifically,
the transition function Δ(.) using transmissions in x{t} , the
over N time slots (one period), node 1 transmits ((N − 1)/2)
current state S{t} is updated to the next state S{t+1} . The tran-
times and is involved as destination node ((N + 1)/2) times
sition reward is nothing other than the number of transmissions
at the starting time slot. Node N transmits ((N + 1)/2) times
introduced by action x{t} . It can be expressed as
concurrently while it is intended ((N − 1)/2) times. Hence, a
simple permutation provides a new schedule. After replacing 1 N
{t}
(7)
in row 4 column 1 of S<1> by 7, and −4 in row 1 column 4 H x{t} = 1 xi > 0 ∀x{t} ∈ X (S{t} ). (22)
i=1
by 4, and considering the corresponding receptions, we obtain
(7)
S<4> . If such a permutation is combined with the aforemen-
By taking optimal actions (x∗{t } , x∗{t +1} , . . . , x∗{t +T −1} ),
tioned modifications, it will result in many new transmission we come up with an optimal schedule S∗ . This optimal strategy
schedules with the same network throughput X ∗ , where
⎡ ⎤
7 6 5 4 −5 −6 −7
⎢ 7 0 0 0 0 0 −7 ⎥ x∗{t } = X ∗ (S{t } ) (23)

⎢ ⎥
⎢ 7 0 0 0 0 0 −7 ⎥
⎢ ⎥ provides the maximum throughput designated by Y ∗ . Thus,
S<4> = ⎢ ⎥
(7)
⎢ 7 0 0 0 0 0 −1 ⎥ . (19) using the transition rewards over one period T , the value of Y ∗
⎢ 1 0 0 0 0 0 −1 ⎥
⎢ ⎥ is computed by
⎣ 1 0 0 0 0 0 −1 ⎦
1 ∗{t }
T
1 2 3 −4 −3 −2 −1
Y∗ = H x . (24)
T
t =1
Note that there may be more than one optimal strategy. In

B. Problem Modeling and Solution Algorithm
terms of the dynamic programming approach adopted in [19],
1) Problem Modeling: The geometry of an N -node net- Y (S, x): (S, X ) → R+ can be seen as the objective function
work is depicted by the delay matrix D and the size of the having Y ∗ as the best possible value. Besides, the action value
function denoted by Q(S, x): (S, X ) → R+ can be used to throughput (i.e., maximizing transmissions). We have seen that
describe the optimal strategy X ∗ the time extent of the current state is A. Similarly, any upcom-
ing transmission will act on the network within A time slots,
X ∗ (S) = arg max (H(x) + Q(Δ(S, x), x )) (25) and we seek to maximize the number of future transmissions.
x∈X (S)
Therefore, one must go through a state that allows this. Given
where x is the decision immediately following x. It will be a transmission decision ẋ, the capacity of a state to accommo-
recalled that satisfying the Bellman equation, the action value date future transmissions within A time slots is used thus far to
function can be expressed in its recursive form as measure the state–action pair value, denoted by ϑi (Ṡ{t} , ẋ).
Let us first enumerate the constraints related to an N -node
Q(S, x) = max (H(x) + Q(Δ(S, x), x )) . (26) network in the UWA environment. Knowing the partial sched-
x∈X (S)
ule Ṡ{t} at t, a single transmission from node k to node l at
Since we do not know the true action value function, it should t ≥ t is allowed only if:
be estimated iteratively at each time slot. To this end, stan- • it is not a self-transmission, i.e., l = k;
dard algorithms require performing exhaustive state space and • there is no transmission or reception already planned for
{t}
decision space enumerations. Thus, the complexity grows very node k at t , i.e., Ṡk,t = 0;
fast with the size of the network. The approximation of the • there is no transmission or reception already planned for
{t}
action value function is a more practical technique. Although node l at t + Dkl , i.e., Ṡl,t +Dkl = 0;
it is a suboptimal solution, it surprisingly achieves good per- • there is no interference at node l at t + Dkl , which origi-
formance. This method relies on the concept of approximate nates from any other node i(i = k) at t + Dkl − Dil , i.e.,
{t}
dynamic programming [20]. i s.t. Ṡi,t +Dkl −Dil > 0;
2) Practical Algorithm: Instead of using decision space X • the transmission from node k will not cause inter-
having a cardinality of O(N N ), we use successive sequential ference on any other node j(j = l) at t + Dkj ,
transmission decisions [19] within one time slot. Each decision while j is receiving a message from node i, i.e.,
is represented by a 2-tuple (k, l) for a single transmission from {t}
i, j s.t. Ṡi,t +Dkj −Dij = j.
node k to node l at t. Thus, the computational complexity of To summarize this first category of constraints, let us use the
the decision space enumeration problem is reduced to O(N 3 ). transmission indicator function Cklδ where δ = t − t, as intro-
We will therefore introduce a new numbering scale, designated duced in [19]. The action value function approximation should
by a, within each time slot t. Let H {t} denote the number take into account Cklδ (Ṡ{t} ), which is given by
of transmissions in time slot t. After a − 1 transmission deci-
⎧
sions and using the transition function Δ(.), the partial schedule ⎪ 0, if l = k
⎪
⎪
Ṡ{t,a} ∈ Ṡ is combined with the transmission decision ẋ{t,a} to ⎪
⎪ {t}
⎪ 0, if Ṡk,t
⎪ = 0
find the next partial schedule, as indicated by ⎪
⎪
⎪
⎪
⎨ 0, if Ṡ {t}
l,t +Dkl = 0
Ṡ{t,a+1} = Δ Ṡ{t,a} , ẋ{t,a} ∀a < H {t} (27) Cklδ Ṡ{t} = (30)
⎪
⎪ {t}
0, if ∃i s.t. Ṡi,t +Dkl −Dil > 0
{t} {t} ⎪
⎪
Ṡ{t+1,1} = Δ Ṡ{t,H } , ẋ{t,H } . (28) ⎪
⎪
⎪
⎪ {t}
0, if ∃i, j s.t. Ṡi,t +Dkj −Dij = j
⎪
⎪
⎪
⎩
In agreement with the new formulation of the problem within 1, otherwise.
state space Ṡ and decision space Ẋ , we introduce an action
value function Hence

Q̇(Ṡ, ẋ) = max H(ẋ) + Q̇(Δ(Ṡ, ẋ), ẋ ) N
N
A
ẋ∈Ẋ (Ṡ) ϑi Ṡ{t} , ẋ = Cklδ Δ(Ṡ{t} , ẋ) . (31)
= max Q̇(Δ(Ṡ, ẋ), ẋ ) + 1 (29) k=1 l=1 δ=0
ẋ∈Ẋ (Ṡ)
The particular architecture of the N -node network being con-
since H(ẋ) = 1 is the reward for the single transmission sidered deserves to have an appropriate look. Actually, the
decided within the considered time slot. Note that at the stage linear arrangement of the nodes makes the network reach its
illustrated by (29), only decision ẋ matters in the optimal action maximum size, i.e., A = N − 1. There is no other larger N -
finding process, unlike ẋ . As stated above, the true value of Q̇ node network. While nodes are spread over a wide area, two
may be obtained using computationally expensive techniques important features should be taken into consideration. On the
such as relative value iteration. Nevertheless, their complexity one hand, this configuration should be taken advantage of to
makes them quickly infeasible, especially for networks for a allow simultaneous transmissions as much as possible over
significant size A. As a practical and much simpler alternative, both close and faraway branches of the network. On the other
a good approximation of the value function can be developed hand, the exchanges between distant nodes should be consid-
on the basis of our apprehension of the problem structure and ered. To this end, the distance Dkl between a transmitting node
properties. k and a destination node l is a key factor to consider in the
The key feature for any solution is the component current approximation of the action value function. As opposed to the
state/future states that should evolve toward maximizing the first category of constraints, this second category is related to
the network architecture. Correspondingly, the new expression

used to determine the state–action pair value is

ϑf Ṡ{t} , ẋ = D(ẋ) · ϑi Ṡ{t} , ẋ (32)
where ẋ = (i, j) and D(ẋ) = Dij . Taking into account all

constraints inherent to the network under consideration, this
measure can be used to approximate the action value function Q̇
N
N
A
Q̇ Δ(Ṡ{t} , ẋ), ẋ = D(ẋ) · Cklδ Δ(Ṡ{t} , ẋ) .
k=1 l=1 δ=0
(33)
However, the use of distance as a weighting parameter could

lead to maintaining the communication between the same pairs
of remote nodes. Interferences generated thus prevent other
nodes from transmitting. We add a weak fairness constraint
which helps avoid this issue. Indeed, each transmitting node
Fig. 2. Network throughput for schedules generated using Algorithm 1 for
should diversify its destination nodes over A time slots. In other
various N ≤ 25. The throughput consistently achieves the bound 2 − (2/N ).
words, for a certain node k, if decision space Ẋ contains sev-
eral opportunities, it is updated before making use of the action
value function, i.e., the opportunities to transmit to destination Algorithm 1. Algorithm to determine transmission decisions in
nodes already served by k within the previous A time slots, are time slot t and update transmission schedule (Context 1)
excluded. Therefore, the optimal decision ẋ∗ from the updated
1 [h!] Input Delay matrix D, current transmission schedule
decision space Ẍ is expressed by
S{t}
ẋ∗ = arg max Q̇ Δ(Ṡ{t} , ẋ), ẋ . (34) Output Updated transmission schedule S{t+1}
ẋ∈Ẍ (Ṡ{t} ) 2 Ṡ ← S{t}
3a←0
By collecting all the steps discussed above, we end up with an 4 while true do
efficient practical algorithm that makes good transmission deci- 5 Compute Ckl0 (Ṡ), ∀k, l according to (30)
sions at time slot t and updates the transmission schedule. A 6 Ẋ ← {(k, l), ∀k, l s.t. Ckl0 = 1}
concise description of this algorithm is given as Algorithm 1. 7 if Ẋ is empty then
Given the partial schedule S{t} , the algorithm goes through 8 return S{t+1} ← Ṡ, H {t} ← a
the following operations and keeps successively repeating –
them, until there is no more allowable transmission for that 9 Ẍ ← Ẋ
time slot. First, it explores all possible transmission decisions 10 for j ∈ {i, ∃i s.t. (i, i ) ∈ Ẍ } do
in time slot t. Afterwards, the decision space found is updated 11 Bj ← {j , ∀j s.t. (j, j ) ∈ Ẍ }
using the weak fairness constraint of diversifying destination 12 If card {Bj } > 1 then
nodes within a time horizon of A time slots. Thereafter, the 13 B̄j ← Bj ∩ {Ṡj,t−A+1 → Ṡj,t−1 }
approximated action value function is used to find the opti- 14 Ẍ ← Ẍ \{(j, j ), ∀j ∈ B̄j }
mal transmission decision. This is a transmission that has little – –
15 a←a+1
effect on forthcoming transmissions. This algorithm can be
16 Compute Q̇(Δ(Ṡ, ẋ), ẋ ), ∀ẋ ∈ Ẍ according to (33)
seen as an extension of the algorithm proposed in [19], which
17 ẋ∗ ← arg max Q̇(Δ(Ṡ, ẋ), ẋ )
is well adapted to regular linear networks and even to gen-
18 – Ṡ ← Δ(Ṡ, ẋ∗ )
eral linear networks. It requires a short time to result in the
throughput-effective schedules presented in Section III-A1, for
any considered size of the network. These schedules are weakly
per-node fair, since every node transmits at least once during Algorithm 1 makes it possible to build up a variety of optimal
one period T = N , i.e., schedules. To this end, one can modify some parameters in the
algorithm. For instance, there are generally more than one entry
T
−1 in the updated decision space Ẍ that correspond to “arg max
(T )
1 Sit > 0 ≥ 1 ∀i. (35)
Q̇(Δ(Ṡ, ẋ), ẋ ).” Thus, each one of these entries leads to a new
t=0
optimal solution and we have to define just the criterion used to
We summarize results from multiple simulation runs of the choose one of them.
algorithm for various N in Fig. 2. The throughput consis- Note that Algorithm 1 has been successfully tested with
tently achieves the bound 2 − (2/N ). The action value function three-node equilateral triangle, regular tetrahedron, and four-
appropriately approximated combined with the weak fair- node stretched (edge length doubled) tetrahedron networks, i.e.,
ness constraint makes it possible to reach such performance. it makes it possible to find the optimal throughput.
IV. T HROUGHPUT IN A R EGULAR L INEAR N ETWORK WITH

A S INGLE C OLLISION D OMAIN AND B ROADCAST T RAFFIC
A. Per-Node Fair Schedule with Shortest Period

Lemma 2: In a regular N -node linear network with a single
collision domain and broadcast traffic, a maximum of two nodes
may transmit in the same time slot.
Proof: Consider a regular N -node linear network where i
is a node transmitting (i.e., broadcasting) at time slot t. Thus,
all the other nodes in the network will receive from i according
to
Sit > 0 ⇒ Sl,t+Dil = −i ∀l, l = i. (36) Fig. 3. Depiction of the state of six-node network in the case where node pairs
(1, 5) and (1, 6) are broadcasting at the same time slot.

Equivalently, if node l is receiving from node i at time slot t ,
this demonstrates that i has sent a message intended for the
whole network Case 1) |i − k| is odd, i.e., ∃l ≥ 0 s.t. |i − k| = 2l + 1:
According to (36), the transmission of node k will
Slt = −i ⇒ Si,t −Dil > 0. (37) give rise to interference in time slot t + l + 1 at node
i − (l + 1) if k < i, and node i + (l + 1) if k > i.
If another node j attempts to transmit in the same time slot t, Case 2) |i − k| is even: Since |i − k| is even and knowing
we will face two cases. that |i − j| is odd, |j − k| is odd. This is a situation
Case 1) |i − j| is even. Based on (36), the two messages similar to the first case. Therefore, node k will cause
originating from nodes i and j will overlap at node interference.
((i + j)/2) at time slot t + (|i − j|/2). It follows It is concluded that there is no possible transmission at t + 1.
that the transmission of node j at t leads to interfer- Transmissions at t + 2
ence and consequently is not allowed. To help under- Suppose now that there exists a node k that may transmit at
stand this circumstance, a simple instance where t + 2, where k = i and k = j. Three separate cases will arise.
nodes 1 and 5 attempt to transmit at the same time Case 1) k has received from both nodes i and j: Given the
slot is given in the graph in Fig. 3. interference constraint, k requires two time slots
Case 2) |i − j| is odd: Once more using (36), the two mes- (t + 1 and t + 2) to receive the two messages sent
sages originating from nodes i and j will travel by i and j. Besides, given the half-duplex constraint
in opposite directions between the two neighbor- and based on (36), k is not able to transmit at t + 2
ing nodes ((i + j − 1)/2) and ((i + j + 1)/2) at since it is receiving a message.
time slot t + ((|i − j| − 1)/2). However, they do Case 2) k has received from i but has not yet received from
not overlap at any node in the network. Therefore, j: Since k has already received from i before t + 2,
the transmission of j at t is allowed. In Fig. 3, this signifies that k and i are neighboring nodes,
a graphical illustration is given of the simple case i.e., k = i − 1 or k = i + 1. Hence, |j − k| is even
where nodes 1 and 6 are broadcasting in the six-node as long as |i − j| is odd. Consider l > 0 where
network at the same time slot. |j − k| = 2l. As can be deduced from (36), the inter-
Let us now consider two nodes i and j transmitting at t, ference problem will occur in time slot t + (l + 1)
where |i − j| is odd. Let us further take into consideration a at node j + (l + 1) if k > j, and node j − (l + 1) if
third node k, k = i and k = j. If |i − k| is even, then k is not k < j. Consequently, the transmission of k at t + 2
allowed to transmit at t. If |i − k| is odd, then |j − k| is even is not possible. The case where k has received from j
since |i − j| is odd. Hence, the transmission of k at t is not but not yet from i is exactly the same as the situation
permitted. In sum, two is the maximum number of allowable that has just been discussed.
simultaneous transmissions in a regular N -node linear network Case 3) k has not yet received from both nodes i and j: If
with a single collision domain and broadcast traffic. |i − k| is even, the transmission of k is not allowed,
Lemma 3: In a regular N -node linear network with a single as has been shown in the case before. Otherwise,
collision domain and broadcast traffic, if two nodes i and j are |i − k| is odd. Thus, |j − k| is even and the conse-
transmitting at time slot t, there is no transmission allowed at quence is still the infeasibility of the transmission
time slots t + 1 and t + 2 for any other node k different from i of k.
and j. It is concluded then that there is no possible transmission at
Proof: Consider a regular N -node linear network where i t + 2.
and j are transmitting nodes at time slot t. To sum up, after two simultaneous transmissions at time slot
Transmissions at t + 1 t, the next possible transmission will occur at t + 3.
Suppose that there exists a node k that may transmit at t + 1, Theorem 2: In a regular N -node linear network with a single
where k = i and k = j. Two cases will arise. collision domain and broadcast traffic, while seeking to grant
fair access to all nodes as often as possible (i.e., transmission similar to that where N is even. Actually, it relies on a trans-
schedule with the shortest period T ), the maximum achievable mission scheme allowing two simultaneous transmissions every
throughput is (N/2) if N is even, and (N (N − 1)/(2N − 1)) three time slots, or equivalently two successive transmissions,
if N is odd (except for N = 3). but still every three time slots. The particularity to note is that
Proof: Let us consider an N -node network where N is since N is odd, except for transmissions in pairs every three
even. time slots, there will be one single transmission for the cen-
Since we are seeking to increase the network throughput, we tral node (N + 1)/2. It is the last transmission in the schedule
allow the maximum possible transmissions at the same time and it takes ((N − 1)/2) time slots to simultaneously reach
slot. From Lemma 2, a maximum of two nodes may trans- both ends of the network, i.e., nodes 1 and N . Thus, the cor-
mit at time slot 1. Let us then grant access to nodes 1 and responding period is T = 3((N − 1)/2) + 1 + ((N − 1)/2) =
N . Using Lemma 3, the next possible transmissions will occur 2N − 1, in which each node has broadcasted once. Therefore,
at 1 + 3 = 4 and so on. Consequently, a transmission scheme the throughput achieved is (N (N − 1)/(2N − 1)).
As regards the particular case where N = 3, S3 and S 3 ,
(6) (6)
that accommodates two transmissions every three time slots
will be optimal. Obviously, there will be a need for additional shown in
time at the end to allow the last receptions. However, this time
should be the most advantageous. In (38), shown at the bot- ⎡ ⎤
tom of the page, we introduce the general form of the proposed 23 −2 0 0 0 −3
= ⎣ 13 −1 0 0 −3 0 ⎦
(6)
schedule where we merely highlight the entries correspond- S3
ing to transmissions denoted by T x and the last receptions. 0 −2 −1 12 0 0
Thereby, after nodes 1 and N at time slot 1, it is the turn of ⎡ ⎤
nodes 2 and N − 1 to transmit at time slot 4. Thus, at every time 23 0 0 0 −2 −3
S 3 = ⎣ 0 −1 0 13 −3 0 ⎦
(6)
slot 1 + 3n (n ≥ 0), the pair of nodes selected to broadcast is (39)
[1 + n, N − n], where n ≤ (N/2) − 1. The transmissions of 0 0 −1 12 −2 0
nodes (N/2) and (N/2) + 1 will occur at 3(N/2) − 2, and
they are the last to transmit. Furthermore, these transmissions
require (N/2) time slots to reach the farthest edge of the net- exemplify per-node fair schedules and the corresponding net-
work, which is node N when node (N/2) is transmitting and work throughput is 1. The size of the network is indicated as an
node 1 when it comes to node (N/2) + 1. It is the shortest index.
time that a transmission can take to reach the entire network. To help better understand the schemes designed above, we
With such a scheme where (N/2) − 1 pairs of nodes are trans- provide an example for the case where N is even: six-node net-
mitting every three time slots and the receptions from the last work, and another one for N odd: seven-node network. The
pair take (N/2) time slots, we end up with a period of T = corresponding per-node fair schedules with the shortest periods
(10) (13)
3((N/2) − 1) + 1 + (N/2) = 2(N − 1). During this period, are illustrated by S6 and S7 , respectively, shown in (40)
each node has transmitted once and its message is received and (41), at the bottom of the next page.
by N − 1 nodes in the network. Subsequently, the throughput (10)
The entry 23456 in the first row of S6 represents a message
achieved is (N (N − 1)/2(N − 1)) = (N/2). originating from node 1 toward destination nodes 2, 3, 4, 5, and
In the case where N is odd (except for N = 3), the through- 6. Note that since each transmission made by a node reaches all
put is (N (N − 1)/(2N − 1)) and the demonstration is quite the other nodes in the network, a per-node fair schedule is also
a per-link fair schedule.
1 2 3 4 · · · 1 + 3n · · · 3(N/2) − 2 · · · 2(N − 1)
⎡ ⎤
1 Tx | | | −((N/2) + 1)
2 ⎢ −− −− −− Tx | | ⎥
⎢ ⎥
.. ⎢ .. ⎥
. ⎢ | | . ⎥
⎢ ⎥
1+n ⎢ −− −− −− −− −− Tx | ⎥
⎢ ⎥
.. ⎢ .. ⎥
. ⎢ | . ⎥
(2N −2) ⎢ ⎥
SN = (N/2) ⎢ −− −− −− −− −− −− −− Tx ⎥ (38)
⎢ ⎥
(N/2) + 1 ⎢
⎢ −− −− −− −− −− −− −− Tx ⎥
⎥
.. ⎢ .. ⎥
. ⎢ . ⎥
⎢ ⎥
N −n ⎢ −− −− −− −− −− Tx ⎥
⎢ ⎥
.. ⎢ .. ⎥
⎢ ⎥
. ⎢ . ⎥
N −1 ⎣ −− −− −− Tx ⎦
N Tx ··· ··· ··· −(N/2)
B. Practical Algorithm for Per-Node Fair Schedule With is considered. An approximation of the action value function Q̇
Shortest Period can thus be expressed as
Unlike the unicast case discussed in the Section III, a node N
N
A
k is allowed to transmit at a certain time slot, if and only if its Q̇ Δ(Ṡ{t} , ẋ), ẋ = D(ẋ) · Cklδ Δ(Ṡ{t} , ẋ)
message can be successfully received by all other nodes. k tar- k=1 l=1 δ=0
gets a group of nodes, rather than a single one. Let L denote (44)
this destination group composed of all nodes in the network
excluding k. ∀l ∈ L, if node l is concerned with one of the con- where ẋ = i and D(ẋ) = |i − ((N + 1)/2)|, as long as the
straints expressed by the transmission indicator function Cklδ , network central position is identified by ((N + 1)/2).
each node l ∈ L\{l} is also concerned, i.e., In other respects, since we are interested in per-node fair
schedule, each node should, once it has transmitted, allow
access to other nodes within a certain time horizon. Likewise,
Cklδ Ṡ{t} = 0 ⇒ Ckl δ Ṡ{t} = 0 ∀l ∈ L\{l}. we are interested as well in the shortest period. Thus, we add
(42) the following constraint to update the decision space Ẋ to Ẍ
before making use of the action value function: at a time slot t,
node k is elected to transmit only if it has no transmissions that
As regards this first category of constraints, broadcast traffic have occurred during the previous 2(N − 1) − 1 time slots in
is more restrictive than unicast delivery. the case where N is even, or during the previous (2N − 1) − 1
By adopting the same problem formulation used in time slots in the case where N is odd. Actually, this time
Section III-B, the action value function is given by horizon corresponds to one period duration, as shown in the
demonstration above.

Q̇(Ṡ, ẋ) = max H(ẋ) + Q̇(Δ(Ṡ, ẋ), ẋ ) For N even, a summarization of the algorithm thus designed
ẋ∈Ẋ (Ṡ) is presented as Algorithm 2.
= max Q̇(Δ(Ṡ, ẋ), ẋ ) + N − 1 (43) It makes appropriate transmission decisions at time slot t and
ẋ∈Ẋ (Ṡ) updates the transmission schedule. The algorithm in the case
where N is odd is almost the same. The only difference is that,
at the step corresponding to line 14, for transmission at the same
since H(ẋ) = N − 1 is the reward for the single transmission
time slot, we favor a pair of distant nodes rather than a pair of
decided within the considered time slot. Note here that each
neighboring nodes once these pairs are concurrent. The distance
decision is represented by ẋ, that is, one single node (instead of
is even more essential with the particular architecture of a linear
2-tuple for unicast traffic).
network.
As has been pointed out previously, distance is a key fac-
Results from several simulation runs of the algorithm for
tor in a network with linear architecture. However, since a
various N are summarized in Fig. 4. For all the transmission
transmitting node has not one destination node but almost the
schedules at the output, the throughput consistently achieves
whole network, the distance from the network central position
⎡ ⎤
23456 0 0 0 −2 −6 0 −5 −3 −4
⎢ ⎥
⎢ 0 −1 0 13456 −6 0 ⎥ 0 −5 −3 −4
⎢ ⎥
⎢ ⎥
⎢ 0 0 −1 −6 −2 −5 12456 −4 0 0 ⎥
(10) ⎢ ⎥
S6 =⎢ ⎥ (40)
⎢ 0 0 −6 −1 −5 −2 12356 −3 0 0 ⎥
⎢ ⎥
⎢ ⎥
⎢ 0 −6 0 12346 −1 0 −2 −4 −3 0 ⎥
⎣ ⎦
12345 0 0 0 −5 −1 0 −2 −4 −3
⎡ ⎤
234567 0 0 0 −2 −6 0 0 −3 −7 0 −5 −4
⎢ ⎥
⎢ 0 −1 0 134567 −6 0 0 −3 0 ⎥ −7 0 −5 −4
⎢ ⎥
⎢ ⎥
⎢ 0 0 −1 −6 −2 0 124567 −7 0 −5 −4 0 0 ⎥
⎢ ⎥
⎢ ⎥
=⎢ 0 ⎥
(13)
S7 ⎢ 0 0 −6 −1 0 −2 −7 −3 −5 123567 0 0 ⎥. (41)
⎢ ⎥
⎢ 0 −6 0 0 −1 −7 −2 123467 −3 0 −4 0 0 ⎥
⎢ ⎥
⎢ ⎥
⎢ 123457 0 0 0 −7 −1 0 −2 −5 −3 0 −4 0 ⎥
⎣ ⎦
0 −6 0 123456 0 0 −1 0 −2 −5 −3 0 −4
better throughput, we provide two illustrations for multicast

transmission schedules: five-node and seven-node networks.
Although the proposed schemes do not offer bidirectional full
connectivity as is the case with the broadcast schemes dis-
cussed above, the throughput achieved is better with good
connectivity. Furthermore, the respective periods are shorter
and accordingly more interesting in terms of supporting fre-
quent exchanges in the network. In the notations below, M
index is used to designate multicasting while B denotes broad-
casting as defined in Section IV-A. The size of the network in
question is exceptionally specified as index of S.
1) Five-Node Network: The transmission scheme depicted
(6)
by S5 ensures good network connectivity during a period of
TM = 6, which is shorter than TB = 9. Moreover, the cor-
responding network throughput is YM = (7/2), which is an
improvement of more than 57% compared to YB = (20/9)
⎡ ⎤
2345 −5 0 −4 −2 −3
⎢ −5 −1 −4 1345 −3 345 ⎥
Fig. 4. Network throughput for schedules generated using Algorithm 2 for var- ⎢ ⎥
S5 = ⎢ ⎥
(6)
ious N ≤ 25. The throughput consistently achieves the bounds for both even
⎢ −2 −4 −1 125 −2 −5 ⎥ . (45)
and odd N . ⎣ 1235 −2 0 −1 0 −2 ⎦
−2 −4 −2 123 −1 −3
Algorithm 2. Algorithm to determine transmission decisions in
time slot t and update transmission schedule where N is even
2) Seven-Node Network: The proposed transmission
(context 2) (8)
schedule for the seven-node network is depicted by S7 . Over
Input: Delay matrix D, current transmission schedule S{t} the period TM = 8 (while TB = 13), it performs a network
Output: Updated transmission schedule S{t+1} throughput of YM = 5. With respect to YB = (42/13), this is
1 Ṡ ← S{t} an improvement of almost 55%
2a←0 ⎡ ⎤
234567 −7 0 −4 −2 −6 34567 −5
3 while true do ⎢ −7 −1 −4 134567 −6 0
⎢ −5 0 ⎥
⎥
4 Compute Ckl0 (Ṡ), ∀k, l according to (30) and (42) ⎢ −1 −4 −1 −6 −2 −5 4567 −7 ⎥
5 Ẋ ← {k, ∀k, l s.t. Ckl0 = 1} ⎢ ⎥
S7 = ⎢ ⎥
(8)
6 if Ẋ is empty then ⎢ 1236 −1 −6 −1 −5 −2 −7 −3 ⎥ .
⎢ −3 0 −1 12347 −1 −7 −2 0 ⎥
7 return S{t+1} ← Ṡ, H {t} ← a ⎢ ⎥
– ⎣ 12347 −3 −4 −1 0 −1 0 −2 ⎦
8 Ẍ ← Ẋ −2 −6 −3 12345 −1 −5 −1 0
9 Ẍ ← {k , ∀k ∈ Ẍ s.t.all{Ṡk ,t−2(N −1)+1 → (46)
Ṡk ,t−1 } ≤ 0}
10 if Ẋ is empty then
11 return S{t+1} ← Ṡ, H {t} ← a D. Network Throughput Upper Bound
–
12 a←a+1 We have seen that in a regular N -node linear network with
13 Compute Q̇(Δ(Ṡ, ẋ), ẋ ), ∀ẋ ∈ Ẍ according to (44) a single collision domain and broadcast traffic, the transmis-
14 ẋ∗ ← arg max Q̇(Δ(Ṡ, ẋ), ẋ ) sion schedule achieving the shortest period performs a network
15 Ṡ ← Δ(Ṡ, ẋ∗ ) throughput of (N/2) if N is even, and (N (N − 1)/(2N − 1))
– if N is odd. Now, irrespective of the period duration, we are
interested in exploring further performance.
the bound N/2 for N even and N (N − 1)/(2N − 1) when N Theorem 3: In a regular N -node linear network with a single
is odd. collision domain, broadcast traffic and per-node fair access, the
network throughput upper bound is N − 1.
Proof: Let us consider an N -node network where N is
C. Multicast Transmission Schedules With Good Connectivity even. Since we do not have any restrictions on the period
and Better Throughput duration, we will make changes to the proposed transmission
In general, the connectivity in ad hoc networks can be schedule in Section IV-A. At time slot 1, nodes 1 and N are
defined as the connection probability of any two nodes via transmitting. They will resume transmission every two time
multiple intermediate connections [21]. In our context, since slots until they reach b transmissions each. Thus, their bth
we have one collision domain with reliable physical link, only transmissions occur at time slot 2b − 1. Afterwards, it is the
direct information exchange is considered. Keen to exploit the turn of nodes 2 and N − 1 to transmit at time slot 2b + 2
large propagation delays in the UWA environment to achieve (according to Lemma 3). Nodes 2 and N − 1 will also remain
transmitting every two time slots until they reach b transmis- If, in a different context, the interest is rather to allow two nodes
sions each at time slot 4b. To a large extent, every time slot to transmit simultaneously, we would propose transmission
(2b + 1)n + 1 (n ≥ 0), the pair of nodes selected to broad- schedules where i and j are two transmitting nodes at the same
cast b times is [1 + n, N − n], while n ≤ (N/2) − 1. Hence, time slot t. These transmissions are feasible only if |i − j| is
the transmissions of nodes (N/2) and (N/2) + 1 will occur odd (as has been shown in the demonstration of Lemma 2). By
at b(N − 2) + (N/2), and they are the last to transmit. Their way of illustration, we propose for the six-node network the
bth transmissions are made at time slot bN + (N/2) − 2 and (2)
schedule S6<1>
require (N/2) time slots to reach the entire network. Along ⎡ ⎤
these lines, (N/2) pairs of nodes are transmitting b times. 23456 −6
⎢ −6 −1 ⎥
Transmissions from each of (N/2) − 1 pairs take 2b + 1 time ⎢ ⎥
⎢ −1 −6 ⎥
slots: b − 1 times two time slots in addition to three time slots S6<1> = ⎢ ⎥
(2)
⎢ −6 −1 ⎥ . (51)
to allow the transition to the transmissions of the next pair of ⎢ ⎥
⎣ −1 −6 ⎦
nodes. However, the receptions from the last pair take 2(b −
1) + 1 + (N/2) time slots. Therefore, the corresponding period 12345 −1
is T = ((N/2) − 1)(2b + 1) + 2(b − 1) + 1 + (N/2) = (b + Regarding this second category of schemes where two nodes
1)N − 2. During this period, each node has transmitted b times are transmitting, one may propose a particular design where
and its messages are received by N − 1 nodes in the network. these transmissions are spread over two successive time slots.
As a result, the throughput achieved is For transmitting nodes i and j, this is allowed only if |i − j| is
(2)
bN (N − 1) even. The transmission schedule S6<2> is an example for the
Y = . (47) six-node network
(b + 1)N − 2
⎡ ⎤
23456 −5
Let us take an example where N = 4 and b = 2. The corre- ⎢ −5
(10) ⎢ −1 ⎥ ⎥
sponding transmission schedule is described by S4 and the ⎢ −1 −5 ⎥
resulting network throughput is Y = 2.4, which is greater than
(2)
S6<2> = ⎢ ⎢ ⎥. (52)
⎢ −5 −1 ⎥ ⎥
Y = 2 obtained when the period is T = 6, i.e., b = 1 ⎣ −1 12346 ⎦
⎡ ⎤ −5 −1
234 −4 0 −4 −2 −3 −2 −3 234 0
⎢ −4 −1 −4 134 −3 134 −3 0 0 −1 ⎥
S4 = ⎢ ⎥
(10)
⎣ −1 −4 −1 124 −2 124 −2 0 0 −4 ⎦ .
V. T HROUGHPUT IN A R EGULAR L INEAR N ETWORK WITH
123 −1 0 −1 −3 −2 −3 −2 123 0 PARTIALLY OVERLAPPING C OLLISION D OMAINS
(48) AND U NICAST T RAFFIC
By letting b → ∞, we end up with an upper bound of N − 1. Having partially overlapping collision domains in wireless
Thus communication systems is a common phenomenon, when seek-
ing to extend network coverage. This is particularly the case in
Y ≤N −1 (49) underwater sensor networks deployed over a large area of inter-
est. A single collision domain may be defined as the geograph-
but we can develop per-node fair schedules with network ical area of the network, where transmissions are restricted
throughput that may move as close to this upper bound as between all nodes belonging to that collision domain, due to
desired, assuming no constraints on the time extent. the interference constraint. Actually, the interference range in
Let us now investigate the possibility of having trans- wireless radio networks is often considered to be approxi-
mission schedules that achieve the upper bound regardless mately twice the transmission range [22], [23]. In the UWA
of the constraints on connectivity and fairness within the environment, we equivalently assume the same proportions.
network.
If we grant permanent access to one node, the correspond-
ing transmission schedule performs a network throughput of A. Efficient Relaying Transmission Schedule
N − 1. One simple situation such a schedule can refer to is When a linear geometry is under consideration, the collision
when a node does not cease to broadcast specific instructions domain is defined as the set of neighboring nodes with a shared
in the whole network. All other nodes are constantly receiving medium along which a transmission is propagating, and where
from the broadcasting node. An example to highlight this type resulting interferences terminate. Therefore, a single collision
(1)
of schedule is given as S6 for the six-node network domain is identified with respect to a certain node in the net-
⎡ ⎤ work. We have then N partially overlapping collision domains.
23456 Every collision domain overlaps in part with its adjacent colli-
⎢ −1 ⎥
⎢ ⎥ sion domains. Let g ≥ 1 denote the transmission range of each
⎢ −1 ⎥
(1)
S6 = ⎢ ⎢ ⎥. (50) node, i.e., as regards this constraint, node i can transmit to node
⎥
⎢ −1 ⎥ j only if
⎣ −1 ⎦
−1 max(1, i − g) ≤ j ≤ min(i + g, N ). (53)
(10)
entries elsewhere. Let us take the example where c = 5. S3
represents the resulting transmission schedule
⎡ ⎤
2 2 2 2 2 0 0 0 0 0
S3 = ⎣ 3 −1 −1 −1 −1 −1 3 3 3 3 ⎦ .
(10)
−2 −2 0 0 0 0 0 −2 −2 −2
(57)
Fig. 5. Partially overlapping collision domains for N = 6 with g = 1. Let us now look at larger networks, i.e., N ≥ 4. Having three
adjacent nodes transmitting at the same time slot t is not
In addition, let us use Gi to designate the extent of the colli- allowed, as there will be interference on the middle node at t +
sion domain within the regular linear N -node network, which 1. Therefore, a transmission schedule with period T = 2 (like
corresponds to node i. Thus that for the three-node network) is no longer interesting. In con-
trast, the period T = 4 seems to be more attractive. Indeed, the
2g ≤ Gi ≤ 4g. (54) corresponding transmission schedule does not accommodate
more than two adjacent positive entries in the same column. By
Nodes 1 and N at both ends of the network have collision (4)
domains of extent G1 = GN = 2g, while every node j s.t. way of example, S6 describes the schedule for N = 6 where
2g < j ≤ N − 2g has as collision domain extent Gj = 4g. the throughput is Y = (5/2)
⎡ ⎤
We consider in this paper a simple illustration of such a 2 2 0 0
configuration where g = 1. Regarding packet delivery, unicast ⎢ 3 −1 −1 3 ⎥
⎢ ⎥
traffic is in use. One case of interest may be when each col- ⎢ −2 −2 4 4 ⎥
S6 = ⎢ ⎥
(4)
lision domain has to be able to forward the sum of the traffic ⎢ −3 5 5 −3 ⎥ . (58)
⎢ ⎥
of its nodes to its immediate neighbor, and so on until reaching ⎣ 6 6 −4 −4 ⎦
the opposite network end. This action can be performed alterna- 0 −5 −5 0
tively: in one direction over a specified duration and afterwards
in the other direction. The scheme showing different collision When formulating generally, over four time slots, it would be
domains in the case where N = 6 and g = 1 is given in Fig. 5. node 1 transmitting two messages successively to node 2 and
CD k is used to designate the collision domain relative to node remaining idle during the other two time slots. Node 2 will in
k. turn transmit two messages consecutively to node 3, in addi-
Since g = 1, each node has the opportunity to transmit to its tion to receiving from node 1. In this way, node i transmits
immediate neighbor only, i.e., node i can transmit to node i + 1 two messages successively to node i + 1, and receives from
only. Let us first consider the simple case of the three-node node i − 1. This operation is performed iteratively until reach-
network. The straightforward way to allow aforementioned ing node N − 1, which is going to transmit twice to node
transmissions is as follows: at time slot 1, node 1 transmits to N and receive twice from node N − 2. Still over the four
node 2 whereas node 2 transmits to node 3. Time slot 2 will be time slots in question, node N receives two messages from
reserved for the corresponding receptions. The resulting trans- node N − 1 and is idle otherwise. Apart from the first and
(2) last rows, which include two idle entries each, the resulting
mission schedule S3 contains two transmissions over a period
T = 2. Hence, the network throughput is Y = 1 transmission schedule S(4) consists of nonzero entries. It con-
⎡ ⎤ tains 2(N − 1) transmissions over a period T = 4. Hence, the
2 0 network throughput is Y = ((N − 1)/2). Moreover, excluding
S3 = ⎣ 3 −1 ⎦ .
(2)
(55) node N , while it is the final destination of the network traffic,
0 −2 S(4) can be assumed to be per-node fair schedule since
In other respects, there may exist a need to successively trans-
4
(4)
mit two messages for each node instead of one message, as we 1 Sit > 0 = 2 ∀i = N. (59)
have seen so far. We thus propose the following transmission t=1
(4)
schedule depicted by S3 with a period T = 4: In [2], Zhang et al. consider transmission protocols with spatial
⎡ ⎤
0 2 2 0 reuse factor Q, where each node uses on average the frac-
S3 = ⎣ 3 3 −1 −1 ⎦ .
(4)
(56) tion 1/Q of its time slots for transmission (Q is an integer
0 −2 −2 0 for simplicity). In the transmission schedule they proposed and
studied, each node has one time slot out of two to use to trans-
The corresponding network throughput is still Y = 1. More mit, to receive, or to remain idle. In contrast, if we consider
generally, the same throughput can be achieved even in the case the analogy with our proposed transmission schedule, all time
where there is a requirement for more successive messages to slots are fully used by nodes. In terms of network-oriented met-
be transmitted by nodes 1 and 2. Thus, let c be the number of rics, numerical study in [2] shows that it is optimal to operate at
consecutive messages that every node should transmit. The cor- Q = 2. Our proposed transmission schedule satisfies this con-
(2c)
responding schedule S3 having a period T = 2c consists of dition even with fully occupied time slots. Moreover, in [13],
c idle entries in each of the first and third rows, and nonzero for the 11-node network, channel utilization is less than 3/2,
while it is equal to 5 for our proposed transmission schedule. Algorithm 3. Algorithm to determine transmission decisions in
Additionally, in the linear topology considered in [14], Xiao time slot t and update transmission schedule (context 3)
et al. assume that the transmission range of each node is just
one hop and that the interference range is less than two hops. An Input: Delay matrix D, transmission range g, current
overall network utilization is a good indicator of network per- transmission schedule S{t}
formance. The authors define it as the fraction of time in which Output: Updated transmission schedule S{t+1}
the base station is busy receiving messages. With respect to our 1 Ṡ ← S{t}
network modeling, where message duration is set to propaga- 2a←0
tion delay between two neighboring nodes, it has been proven 3 while true do
in [14] that, under a per-node fair access constraint with mul- 4 Compute Ckl0 (Ṡ), ∀k, l according to (60)
tihop relaying, 1/2 is an asymptotic tight upper bound for the 5 Ẋ ← {(k, l), ∀k, l s.t. Ckl0 = 1}
overall network utilization. However, no transmission schedule 6 if Ẋ is empty then
example achieving this bound is provided. Nevertheless, even 7 return S{t+1} ← Ṡ, H {t} ← a
–
if the interference range we adopt in our network modeling is 8 Ẍ ← {(i, j) ∈ Ẋ s.t. i < j}
twice the transmission range for each node, our proposed trans- 9 a←a+1
mission schedule achieves the upper bound in [14], whatever 10 Compute Q̇(Δ(Ṡ, ẋ), ẋ ), ∀ẋ ∈ Ẍ according to (61)
the network size. 11 ẋ∗ ← arg max Q̇(Δ(Ṡ, ẋ), ẋ )
12 Ṡ ← Δ(Ṡ, ẋ∗ )
–
B. Effective Algorithm for Relaying Scheme
The whole N -node network consists of several collision We use the same problem formulation that has been set in
domains that are partially overlapping. Given this context, let Section III-B. However, the time horizon to take into consid-
us first resume the constraints on transmission under such eration becomes 2g, since the expiration time of a transmission
new conditions. Given the partial schedule Ṡ{t} at t, a single in the network is equal to 2g time slots. This is the time extent of
transmission from node k to node l at t ≥ t is allowed only if: the current state of the network. In addition to Cklδ (Ṡ{t} ), the
• node l is within the transmission range of node k, i.e., action value function approximation Q̇ relies on the distance
Dkl ≤ g; parameter, i.e.,
• it is not a self-transmission, i.e., l = k; N
N
2g
• there is no transmission or reception already planned for Q̇ Δ(Ṡ{t} , ẋ), ẋ = D(ẋ) · Cklδ Δ(Ṡ{t} , ẋ)
{t}
node k at t , i.e., Ṡk,t = 0; k=1 l=1 δ=0
• there is no transmission or reception already planned for (61)
{t}
node l at t + Dkl , i.e., Ṡl,t +Dkl = 0;
where ẋ = (i, j) and D(ẋ) = Dij . Note that throughout the
• there is no interference at node l at t + Dkl , which
transmission schedule examples introduced above for any com-
is originated from any other node i (i = k and
municating nodes pair (i, j), we have Dij = 1. Hence, the
Dil ≤ 2g) at t + Dkl − Dil , i.e., i s.t. [ Dil ≤
{t} distance factor does not have any impact. Nevertheless, we did
2g ] & [ Ṡi,t +Dkl −Dil > 0 ]; identify its utility in the case where g = 2, while implementing
• the transmission from node k will not cause interfer- the same action value function approximation.
ence on any other node j (j = l and Dkj ≤ 2g) at Furthermore, the action of nodes generating and relaying
t + Dkj while j is receiving a message from node messages to the network end should be transposed using an
i (Dij ≤ g), i.e., i, j s.t. [ Dkj ≤ 2g ] & [ Dij ≤ additional constraint. Indeed, we update the decision space Ẋ
{t}
g ] & [ Ṡi,t +Dkj −Dij = j ]. to Ẍ before making use of the action value function, in order to
This first category of constraints can be summarized using orient all transmissions in one direction. In other words, if the
the new transmission indicator function Cklδ where δ = t − t. flow of messages is planned to be in the direction node 1 →
Cklδ (Ṡ{t} ) is expressed as node N

Cklδ Ṡ{t} Ẍ = {(i, j) ∈ Ẋ s.t. i < j}. (62)
⎧
⎪ 0, if Dkl > g Otherwise, selected 2-tuple (i, j) should be such that i > j,
⎪
⎪
⎪
⎪ 0, if l = k when transmissions are oriented in the opposite direction, i.e.,
⎪
⎪
⎪
⎪ node N → node 1 .
⎪
⎪ {t}
⎪ 0, if Ṡk,t = 0
⎪ For N ≥ 4, the algorithm passing through all the steps
⎪
⎪
⎨ 0, if Ṡ {t}
l,t +Dkl = 0
outlined above is summarized as Algorithm 3.
= In addition to delay matrix D and the current transmission
⎪
⎪ {t}
⎪
⎪ 0, if ∃i s.t. [ Dil ≤ 2g ] & Ṡi,t +Dkl −Dil > 0 schedule S{t} , the transmission range g is a required input for
⎪
⎪
⎪
⎪ 0, if ∃i,
⎪
⎪
⎪ j s.t. [ Dkj ≤ 2g ] & [ Dij ≤ g ] & the algorithm.
⎪
⎪ {t}
Ṡi,t +Dkj −Dij = j Results obtained from numerous simulation runs of the algo-
⎪
⎪
⎩ rithm for various N are shown in Fig. 6. (N − 1)/2 is the
1, otherwise. throughput evermore achieved by all the transmission schedules
(60)
Authorized licensed use limited to: VIT University- Chennai Campus. Downloadedat
onthe output.
October 18,2023 at 13:58:48 UTC from IEEE Xplore. Restrictions apply.
[2] W. Zhang, M. Stojanovic, and U. Mitra, “Analysis of a linear multi-

hop underwater acoustic network,” IEEE J. Ocean. Eng., vol. 35, no. 4,
pp. 961–970, Oct. 2010.
[3] B. Peleato and M. Stojanovic, “Distance aware collision avoidance pro-
tocol for ad-hoc underwater acoustic sensor networks,” IEEE Commun.
Lett., vol. 11, no. 12, pp. 1025–1027, Dec. 2007.
[4] X. Guo, M. R. Frater, and M. J. Ryan, “Design of a propagation-delay-
tolerant MAC protocol for underwater acoustic sensor networks,” IEEE
J. Ocean. Eng., vol. 34, no. 2, pp. 170–180, Apr. 2009.
[5] K. Kredo, P. Djukic, and P. Mohapatra, “STUMP: Exploiting position
diversity in the staggered TDMA underwater MAC protocol,” in Proc.
IEEE INFOCOM Conf., Rio de Janeiro, Brazil, Apr. 2009, pp. 2961–
2965.
[6] H.-H. Ng, W.-S. Soh, and M. Motani, “A bidirectional-concurrent MAC
protocol with packet bursting for underwater acoustic networks,” IEEE J.
Ocean. Eng., vol. 38, no. 3, pp. 547–565, Jul. 2013.
[7] Y. Guan, C.-C. Shen, and J. Yackoski, “MAC scheduling for high
throughput underwater acoustic networks,” in Proc. IEEE Wireless
Commun. Netw. Conf., Cancun, Quintana-Roo Mexico, Mar. 2011, DOI:
10.1109/WCNC.2011.5779160.
[8] R. Diamant and L. Lampe, “Spatial reuse time-division multiple access
for broadcast Ad Hoc underwater acoustic communication networks,”
IEEE J. Ocean. Eng., vol. 36, no. 2, pp. 172–185, Apr. 2011.
[9] C. Carbonelli and U. Mitra, “Cooperative multihop communication
Fig. 6. Network throughput for schedules generated using Algorithm 3 for for underwater acoustic networks,” in Proc. 1st ACM Int. Workshop
various N ≤ 25. The throughput consistently achieves the bound (N − 1)/2. Underwater Netw., Los Angeles, CA, USA, Sep. 2006, pp. 97–100.
[10] C. R. Benson, M. J. Ryan, and M. R. Frater, “On the bene-
fits of high absorption in practical multi-hop networks,” in Proc.
OES/IEEE OCEANS Conf., Aberdeen, Scotland, Jun. 2007, DOI:
VI. C ONCLUSION 10.1109/OCEANSE.2007.4302280.
[11] W. Chen and U. Mitra, “Packet scheduling for multihopped under-
We considered achievable throughput in the widely used water acoustic communication networks,” in Proc. MTS/IEEE
linear network topology. We designed TDMA-based MAC pro- OCEANS Conf., Vancouver, BC, Canada, Sep. 2007, DOI:
10.1109/OCEANS.2007.4449267.
tocols taking advantage of large propagation delays to improve [12] M. Vajapeyam, S. Vedantam, U. Mitra, J. C. Preisig, and M. Stojanovic,
network throughput. We explored three concrete contexts: “Distributed space-time cooperative schemes for underwater acoustic
single collision domain with unicast traffic, single collision communications,” IEEE J. Ocean. Eng., vol. 33, no. 4, pp. 489–501, Oct.
2008.
domain with broadcast traffic, and partially overlapping col- [13] Y.-D. Chen, C.-Y. Lien, S.-W. Chuang, and K.-P. Shih, “DSSS: A TDMA-
lision domains with unicast traffic. In an N -node network, based MAC protocol with dynamic slot scheduling strategy for under-
we developed transmission schedules achieving 2 − (2/N ) as water acoustic sensor networks,” in Proc. OES/IEEE OCEANS Conf.,
Santander, Spain, Jun. 2011, DOI: 10.1109/Oceans-Spain.2011.6003632.
normalized network throughput in the first context. Moreover, [14] Y. Xiao et al., “Tight performance bounds of multihop fair access for
we proposed periodic per-node fair schedules with network MAC protocols in wireless sensor networks and underwater sensor net-
throughput upper bounds of N/2 and N − 1 under distinctive works,” IEEE Trans. Mobile Comput., vol. 11, no. 10, pp. 1538–1554,
Oct. 2012.
fairness constraints in the second context. Likewise, a practical [15] V. R. Cadambe and S. A. Jafar, “Degrees of freedom of wireless
instance of the third context was considered. It depicts a mul- networks—What a difference delay makes,” in Proc. 41st Asilomar
tihop scenario where messages originating at one end of the Conf. Signals Syst. Comput., Pacific Grove, CA, USA, Nov. 2007, DOI:
10.1109/ACSSC.2007.4487180.
network are sequentially relayed node by node in the direction [16] R. Mathar and G. Bocherer, “On spatial patterns of transmitter-receiver
of the final destination located at the other end of the network. pairs that allow for interference alignment by delay,” in Proc. 3rd Int.
All contexts have been discussed in depth, and with illustrative Conf. Signal Process. Commun. Syst., Omaha, NE, USA, Sep. 2009, DOI:
10.1109/ICSPCS.2009.5306396.
examples. In addition, computationally efficient algorithms that [17] R. Mathar and M. Zivkovic, “How to position n transmitter-
generate the transmission schedules regardless of the network receiver pairs in n-1 dimensions such that each can use half of
size are developed. the channel with zero interference from the others,” in Proc. IEEE
Global Telecommun. Conf., Honolulu, HI, USA, Nov. 2009, DOI:
The findings of this study help identify the impact of inher- 10.1109/GLOCOM.2009.5425222.
ent UWA large propagation delays on network throughput in [18] F. L. Blasco, F. Rossetto, and G. Bauch, “Time interference align-
linear topology networks. But above all, the study is a step ment via delay offset for long delay networks,” in Proc. IEEE
Global Telecommun. Conf., Houston, TX, USA, Dec. 2011, DOI:
toward more complex network configuration analysis, such as a 10.1109/GLOCOM.2011.6133800.
grid topology with two or more rows. The combination of net- [19] M. Chitre, M. Motani, and S. Shahabudeen, “Throughput of networks
work architecture, traffic policy, collision domain extent, and with large propagation delays,” IEEE J. Ocean. Eng., vol. 37, no. 4,
pp. 645–658, Oct. 2012.
certain fairness constraints should definitely lead to interesting [20] W. B. Powell, Approximate Dynamic Programming: Solving the Curses of
new research problems. Dimensionality, New York, NY, USA: Wiley-Interscience, 2007, ch. 8.
[21] L. Zhang, L. Cai, J. Pan, and F. Tong, “A new approach to the directed
connectivity in two-dimensional lattice networks,” IEEE Trans. Mobile
Comput., vol. 13, no. 11, pp. 2458–2472, Nov. 2014.
R EFERENCES [22] D. Jing, L. Ben, and P. K. Varshney, “Tuning the carrier sensing range of
IEEE 802.11 MAC,” in Proc. IEEE Global Telecommun. Conf., Dallas,
[1] S. C. Ergen and P. Varaiya, “PEDAMACS: Power efficient and delay TX, USA, Nov. 2004, DOI: 10.1109/GLOCOM.2004.1378900.
aware medium access protocol for sensor networks,” IEEE Trans. Mobile [23] S. Boppana and J. M. Shea, “Overlapped carrier-sense multiple access
Comput., vol. 5, no. 7, pp. 920–930, Jul. 2006. (OCSMA) in wireless ad hoc networks,” IEEE Trans. Mobile Comput.,
vol. 8, no. 3, pp. 369–383, Mar. 2009.
Said Lmai received the Engineering degree from Christophe Laot (M’07–SM’12) was born in Brest,
the Royal Moroccan Naval Academy (Ecole France, on March 12, 1967. He received the Eng.
Royale Navale), Casablanca, Morocco, in 2003, the degree from the Ecole FRancaise d’Electronique et
Telecommunications’ Engineering degree from the d’Informatique (EFREI), Paris, France, in 1991 and
National Communications Institute (Institut National the Ph.D. degree from the University of Rennes,
des Postes et Télécommunications), Rabat, Morocco, Rennes, France, in 1997, all in communications and
in 2009, and the Ph.D. degree from TELECOM information systems.
Bretagne, Brest, France, in 2014, all in wireless In 1997, he joined the Signal and Communications
communications and networking. Department, TELECOM Bretagne, Brest, France, as
He served on an Overseas Patrol Vessel as the an Associate Professor. Since 2013, he has been (full)
Operations Officer. From 2010 to 2011, he headed an Professor in the same institution. His research inter-
infrastructure networking unit within the Communications Division at Royal ests lie in the areas of communications and signal processing, including equal-
Moroccan Navy’s HQ. He is currently an Associate Professor at the Royal ization, turbo-equalization, iterative receivers for interference cancellation,
Moroccan Naval Academy. His research interests lie in multiple-access tech- synchronization, and underwater acoustic communications.
niques in ad hoc networks, wireless communications, and signal processing, Dr. Laot is a member of the IEEE Communication Society and Vice-
including underwater communications and multicarrier blind synchronization Chair (Europe) of the Technology Committee “Underwater Communication,
algorithms. Navigation, and Positioning” for the IEEE Oceanic Engineering Society.
Mandar Chitre (M’04–SM’11) received the B.Eng. Sebastien Houcke (M’09) was born in Amiens,
and M.Eng. degrees in electrical engineering from the France, in 1974. He received the Engineering
National University of Singapore (NUS), Singapore, degree from the the Université de Technologie de
in 1997 and 2000, respectively, the M.Sc. degree Compiègne, Compiègne, France, the DEA degree
in bioinformatics from the Nanyang Technological (french equivalent to the M.Sc. degree) in signal
University (NTU), Singapore, in 2004, and the Ph.D. and image processing from the Université de Cergy,
degree from NUS in 2006. Cergy-Pontoise, France, and the Ph.D. degree from
From 1997 to 1998, he worked with the Acoustic Laboratoire Systéme de Communication, Université
Research Laboratory (ARL), NUS. From 1998 to de Marne-la-Vallée, Champs-sur-Marne, France, in
2002, he headed the Technology Division of a 2002. His Ph.D. dissertation focused on the esti-
Regional Telecommunications Solutions Company. mation issue in non-cooperative telecommunication
In 2003, he rejoined ARL, initially as the Deputy Head (Research) and is systems.
now the Head of the Laboratory. He also holds a joint appointment with the At the end of 2002, he joined the Signal and Communications Department,
Department of Electrical and Computer Engineering at NUS as an Assistant Telecom Bretagne, Plouzané, France, as an Associate Professor and became a
Professor. His current research interests include underwater communications, full Professor in 2013. His research areas include statistical and digital signal
autonomous underwater vehicles, and acoustic signal processing. processing and signal processing for digital communications for blind applica-
Dr. Chitre has served on the Technical Program Committees of the IEEE tions (synchronization for coded systems, multiple-access technique, and blind
OCEANS Conference, the International Conference on Underwater Wireless identification of communication parameters).
Networks and Systems (WUWNet), the Defence Technology Asia (DTA)
International Conference, and the Offshore Technology Conference (OTC).
He has served as a reviewer for numerous international journals. He was the
Chairman of the student poster committee for IEEE OCEANS’06 in Singapore
and the Chairman for the IEEE Singapore AUV Challenge 2013. He is currently
the IEEE Ocean Engineering Society Technology Committee Cochair of under-
water communication, navigation, and positioning, and an Associate Editor for
the IEEE J OURNAL OF O CEANIC E NGINEERING.

Paper ID

Uploaded by

Copyright:

Available Formats

Paper ID

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Paper ID

Uploaded by

Copyright:

Available Formats

156 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 42, NO.

Throughput-Efficient Super-TDMA MAC

Abstract—Underwater acoustic (UWA) sensor network deploy- I. I NTRODUCTION

is always to ensure a high network utilization, i.e., a high

However, with respect to unicast traffic, we can add the stronger

In spite of that, we may obtain a lower level of fairness by offer-

TABLE II needed is prohibitively long. We found that 2 − 2/N is the

1 2 ··· e ··· N/2 (N/2) + 1 (N/2) + 2 ··· (N/2) + f ··· N −1 N

1 2 ··· e · · · (N − 1)/2 (N + 1)/2 (N + 3)/2 ··· (N − 1)/2 + f ··· N −1 N

denote the first alternative derived network is given by

Moreover, unlike N -node networks with N even, the proposed

⎢ 7 0 0 0 0 0 −7 ⎥ x∗{t } = X ∗ (S{t } ) (23)

Note that there may be more than one optimal strategy. In

the network architecture. Correspondingly, the new expression

where ẋ = (i, j) and D(ẋ) = Dij . Taking into account all

However, the use of distance as a weighting parameter could

IV. T HROUGHPUT IN A R EGULAR L INEAR N ETWORK WITH

A. Per-Node Fair Schedule with Shortest Period

better throughput, we provide two illustrations for multicast

[2] W. Zhang, M. Stojanovic, and U. Mitra, “Analysis of a linear multi-

You might also like