1

A Distributed Transmission Power Control Protocol

for Mobile Ad Hoc Networks
Alaa Muqattash and Marwan Krunz
Department of Electrical and Computer Engineering
The University of Arizona
Tucson, AZ 85721
{alaa,krunz}@ece.arizona.edu

Abstract— In this paper, we propose a comprehensive used in situations where network connectivity is tem-
solution for power control in mobile ad hoc networks porarily needed or where it is infeasible (or expensive)
(MANETs). Our solution emphasizes the interplay between to install a fixed infrastructure network. Power control
the MAC and network layers, whereby the MAC layer in- in MANETs has recently received a lot of attention for
directly influences the selection of the next-hop by properly
two main reasons. First, power control has been shown
adjusting the power of route request packets. This is done
while maintaining network connectivity. Channel-gain in- to increase spatial channel reuse, hence increasing the
formation obtained mainly from overheard RTS and CTS overall (aggregate) channel utilization [15]. This issue
packets is used to dynamically construct the network topol- is particularly critical given the ever-increasing demand
ogy. Unlike the IEEE 802.11 approach and previously pro- for channel bandwidth in wireless environments. Second,
posed schemes, ours does not use the RTS/CTS packets to power control improves the overall energy consumption
silence the neighboring nodes. Instead, collision avoidance in a MANET, consequently prolonging the lifetime of the
information is inserted in the CTS packets and sent over
network. Portable devices are often powered by batter-
an out-of-band control channel. This information is used
to dynamically bound the transmission power of potentially ies with limited weight and lifetime, and energy saving is
interfering nodes in the vicinity of a receiver. By properly a crucial factor that impacts the survivability of such de-
estimating the required transmission power for data pack- vices.
ets, our protocol allows for interference-limited simultane- The Distributed Coordination Function (DCF) of the
ous transmissions to take place in the neighborhood of a re- IEEE 802.11 [1] standard is, by far, the most domi-
ceiving node. Simulation results indicate that compared to nant MAC protocol for ad hoc networks1 . This protocol
the IEEE 802.11 approach, the proposed protocol achieves
generally follows the CSMA/CA paradigm, with exten-
a significant increase in the channel utilization and end-to-
end network throughput, and a significant decrease in the sions to allow for the exchange of RTS/CTS (request-to-
total energy consumption. send/clear-to-send) handshake packets between the trans-
mitter and the receiver. These control packets are needed
Index Terms— Power control, ad hoc networks, energy effi-
to reserve a transmission floor for the subsequent data
cient routes, IEEE 802.11, interference margin.
packets. Nodes transmit their control and data packets at
a common maximum power level, preventing all poten-
tially interfering nodes from starting their own transmis-
I. I NTRODUCTION
sions. Any node that hears the RTS or the CTS message
Mobile ad hoc networks (MANETs) are multi-hop net- defers its transmission until the ongoing transmission is
works in which mobile nodes cooperate to maintain net- over. While such an approach is fundamentally needed to
work connectivity and perform routing functions. These avoid the hidden node problem, it negatively impacts the
fast deployable, self-organizing networks are typically channel utilization by not allowing concurrent transmis-
1
This work was supported in part by the National Science Foundation In addition to the DCF, the 802.11 standard also supports a Point
under grants ANI 9733143, CCR 9979310, and ANI 0095626; and Coordination Function (PCF), which is essentially a polling scheme
by the Center for Low Power Electronics (CLPE) at the University of that is intended for delay-sensitive traffic. The DCF can be deployed in
Arizona. An abridged version of this paper was presented at the IEEE both the Ad Hoc and the Infrastructure modes. The former is assumed
Infocom 2003 Conference, San Francisco, April 1–3,2003. in this paper.
 2

Collision at B
sions to take place over the reserved floor. This situation
is exemplified in Figure 1, where node A uses its max-
imum transmission power to send its packets to node B
(for simplicity, we assume omnidirectional antennas, so
A B C D
a node’s reserved floor is represented by a circle in the
2D space). Nodes C and D hear B’s CTS message and,
therefore, refrain from transmitting. It is easy to see that
both transmissions A → B and C → D can, in principle,
take place at the same time if nodes are able to select their
Fig. 2. Challenge of implementing power control in a dis-
transmission powers in an appropriate manner. In Figure tributed fashion. Node C is unaware of the communication
1, the reserved floors based on the standard (fixed, max- A → B, and hence it starts transmitting to node D at a power
imum power) approach are indicated by dashed circles, that destroys B’ s reception.
while the ones that are based on the minimum required
power for coherent reception are indicated by solid cir-
cles.

tion is the key to our proposal. However, to enable dy-

namic adjustment of the (data packet) transmission power,
D
separate channels are needed for data and control pack-
B A ets. Control packets are transmitted at power level Pmax ,
C
and are received by all potentially interfering nodes, as
in the IEEE 802.11 standard. However, in contrast to the
S
CT

S
RT

IEEE 802.11, interfering nodes may be allowed to trans-

mit concurrently, depending on some criteria that will be
discussed later.
Fig. 1. Inefficiency of the 802.11 approach. Nodes A and B
are allowed to communicate, but nodes C and D are not. Power control for MANETs has been extensively stud-
ied (see Section IV for related work). However, previ-
While the idea of power control is simple, achieving it ously proposed protocols address the issue from a single-
in a distributed manner is challenging. More specifically, layer perspective, by either implementing power control
it is not enough to only adjust the transmission power of with proper MAC functionality in mind (e.g., [24], [34]),
each transmitter according to the minimal power needed or by using it as a means of controlling the connectiv-
for coherent reception. This situation is shown in Figure ity and topological properties of the network (e.g., [32],
2, where node A has just started a transmission to node B [25], [27], [29]). While the two approaches may at first
at a power level that is just enough to ensure correct de- seem orthogonal, integrating them in one framework is, at
coding at B. Suppose that node B uses the same power best, highly inefficient. Consequently, none of these ap-
level to communicate with A. Nodes C and D are out- proaches offers a comprehensive solution to the problem.
side the floors of A and B, so they do not hear the RTS- Our view is that inter-layer dependence plays a critical
CTS exchange between A and B. For nodes C and D role in providing an efficient and comprehensive solution
to communicate, they have to use a power level that is to the power control problem, and this view is a key design
reflected by the transmission floors in Figure 2 (the two principle in our proposed protocol.
circles centered at C and D). However, the transmission
C → D will interfere with A → B transmission, causing The rest of the paper is organized as follows. In Sec-
a collision at B. In essence, the problem is caused by the tion II we present the proposed protocol, emphasizing the
asymmetry in the transmission floors. design considerations that were taken into account. The
From the above example, one can make the following operational details are discussed in Section III. In Sec-
observation: if nodes send their control (RTS-CTS) pack- tion IV we review related work in the area of power con-
ets at a fixed power level (Pmax ), but send their data pack- trol for MANETs. The simulation results are presented
ets at an adjustable (lower) power level, then the collision and discussed in Section V. Finally, our main conclusions
in the previous example can be avoided. This observa- are drawn in Section VI.
 3

II. P OWER C ONTROLLED D UAL C HANNEL (PCDC) spacing between the channels is enough to ensure that the
P ROTOCOL outgoing signal on one channel does not interfere with the
A. Channel Model and Protocol Assumptions incoming signal on the other channel.
In designing our protocol, we assume that: (1) the chan-
nel gain is stationary for the duration of the control and the B. Protocol Overview and Design Considerations
ensuing data packet transmission periods; (2) the gain be- The interaction between the network and MAC layers
tween two nodes is the same in both directions; and (3) is fundamental for power control in MANETs. On the
data and control packets between a pair of nodes observe one hand, the power level determines who can receive the
similar channel gains. The justification of these assump- route request packets, and hence, it directly impacts the
tions follows next. selection of the next hop. Obviously, this is a network-
Radio channels typically exhibit large- and small-scale layer issue. On the other hand, the power level also de-
propagation behaviors [28]. Large-scale propagation termines the floor reserved for the node’s transmission,
characterizes the mean signal strength for an arbitrary which is a MAC-layer issue. Hence, we have to somehow
transmitter-receiver separation. Such propagation behav- introduce power control from the perspectives of both lay-
ior has no impact on the validity of our channel assump- ers .
tions, since the distance and the level of clutter are the A power controlled MAC protocol reserves different
same in both directions and for both data and control floors for different uses of the channel, depending on the
channels; hence, the mean signal strength will also be the node’s transmission power. The selection of the “best”
same. Also, the time needed for the RTS/CTS exchange transmission range has been investigated in the literature,
followed by a data-packet transmission is typically in the but not in the context of collision-avoidance MAC proto-
order of tens of milliseconds. Within this time interval, cols. In [16] the authors have shown that a higher net-
very little change occurs in the locations of the mobile work throughput can be achieved by transmitting packets
nodes, and consequently in the average signal strength. to the nearest neighbor in the forward progress direction.
Small-scale propagation characterizes the fluctuations In [15] the authors have proved that using a smaller trans-
of the received signal strength over very short time dura- mission range increases network throughput. The intu-
tions. These fluctuations result from multiple versions of ition behind these results is that halving the transmission
the signal (i.e., multipath waves) arriving at the receiver at range increases the number of hops by two, but decreases
slightly different times and combining to give a resultant the area of the reserved floor to one forth of its original
signal that can vary widely in amplitude and phase. Small- value, allowing for more transmissions in the neighbor-
scale propagation may affect our protocol assumptions hood.
since signals may combine differently in both directions In addition to improving network throughput, reducing
and for both channels. However, in a spread spectrum the transmission range plays a significant role in reduc-
environment where the system spreads the signal into a ing the energy consumption [12]. The power consumed
relatively wide bandwidth using a pseudo-noise (PN) se- by the radio frequency (RF) power amplifier of the net-
quence, the receiver can exploit the multipath components work interface card (NIC) is directly proportional to the
to improve the performance of the system. This is ac- power of the transmitted signal, and thus, it is of great in-
complished by using several diversity techniques (such as terest to control the signal transmission power to increase
RAKE receivers) that take advantage of the random nature the operational lifetime of mobile nodes. Presently, the
of the signal by finding uncorrelated signal paths. There- RF power amplifier consumes almost half (or more in the
fore, our proposed protocol relies on physical-layer tech- case of small computing devices such as sensors) of the
niques to mitigate the multipath effect, and in modest fad- total energy consumed by the NIC. This ratio is expected
ing channels the assumptions will hold. to increase in future NICs, as the processing components
In addition to the above assumptions, we assume that become more power efficient. Therefore, there is poten-
the radio interface can provide the MAC layer with the av- tial for a significant energy saving by reducing the signal
erage power of a received control signal as well as the av- transmission power. In [29] the authors have showed that
erage interference power. The radio interface is equipped power-efficient routes can be found by considering only
with carrier-sense hardware that senses the control chan- the nodes in the “enclosure region” as potential next hops.
nel for any carrier signal. No carrier-sense is needed for Similar results have been provided in [32]. Another ad-
the data channel. The control channel is further divided vantage of power control that has not received much at-
into two sub-channels: a RTS-CTS channel and an ac- tention in the literature is related to reducing the power
knowledgement (ACK) channel. The carrier frequency consumption at unintended receivers (those who are not
 4

addressed by the transmission). Significant power is con- a distributed mechanism by which a node can dynami-
sumed in receiving a packet. For example, in the 2003 cally compute its connectivity set (CS) (defined below).
model of the Cisco Aironet 350 Series Client Adapter card From this CS, the node can then decide on the set of next-
operating at 5 volts, the reception power amounts to about hop nodes. We now describe a localized algorithm for
67% of the transmission power. Since reducing the trans- constructing the CS of an arbitrary node i (CSi ). This
mission range results in a smaller number of nodes over- algorithm aims at producing power-efficient end-to-end
hearing the transmission, less power will be consumed by routes while simultaneously maintaining network connec-
those unintended receivers. tivity and introducing as little overhead as possible.
The above discussion provides sufficient motivation to The intuition behind the algorithm is that CSi must con-
dynamically adjust the transmission range for data pack- tain only the neighboring nodes with which direct com-
ets. The question is how can a node select the lowest pos- munication requires less power than the indirect (two-hop)
sible power that ensures network connectivity while si- communication via any other node that is already in CSi .
multaneously guaranteeing proper MAC functionality and To construct CSi , node i continuously caches the esti-
introducing little overhead? Section II-C answers this mated channel gain of every signal it receives over the
question and explains how next-hop selection can be re- control channel, regardless of the intended destination of
stricted by MAC-layer considerations. this signal. Note that computing the gain is possible be-
Having decided on varying the transmission range, an cause control packets are transmitted at a fixed, known
access mechanism is required to avoid the type of colli- power, and hence, node i uses the reception power of the
sions in Figure 2. For that, PCDC uses a modified RTS- signal to determine the channel gain. Each node in CSi
CTS reservation mechanism. Unlike the 802.11 approach is associated with a timer that expires T seconds from the
(and others, e.g., [4], [14], [18]), the RTS-CTS control time this node was added to CSi . The value of T will be
signals in our scheme are not used to silence the neigh- discussed later. If the timer expires, then the correspond-
bors of a receiving node. Instead, the control signals in ing node is deleted from CSi .
the proposed scheme are used to dynamically bound the Let Puv be the minimum power required to transmit a
transmission power of interfering nodes in the vicinity of data packet from node u to node v at a given time instant.
a receiver. The details of this mechanism are explained in Upon receiving an RTS/CTS packet from another node,
Section II-D. say j, node i does the following. If j ∈ CSi and the
The third key consideration in PCDC is to provide co- newly computed channel gain matches the already stored
operation among neighboring nodes at the MAC layer. one, then the timer associated with j’s entry in CSi is reset
A node that intends to transmit has to account for po- and no further action is taken. On the other hand, if j 6∈
tential future transmissions in its neighborhood. This is CSi or if j ∈ CSi but the newly computed gain does not
achieved by having an interference margin that allows match the already stored one, then node i checks if Pij <
nodes at some interfering distance to start new transmis- Piu + Puj for every node u ∈ CSi , u 6= j. If so, then
sions. Nodes that are in the neighborhood may commence (i)
node j is added to CSi ; otherwise, it is not. Let Pconn
def
their transmissions if such transmissions will not disturb = maxj∈CSi Pij . If node j is added to CSi and Pij <
the ongoing ones. In Section II-E we develop a distributed (i)
Pconn , then all other elements of CSi must be re-examined.
strategy that dynamically adjusts the interference margin The reason is that a two-hop path between nodes i and u,
to maximize the number of simultaneous transmissions. u ∈ CSi , that goes through node j may now be more
power efficient than the direct path between i and u. In
C. Connectivity Set this case, node u has to be deleted from CSi . However,
(i)
In PCDC, the MAC layer affects the performance of if Pij ≥ Pconn , then Pij + Pju > Piu for any u ∈ CSi
the network layer by controlling the power used to trans- and hence, there is no need to re-examine CSi . Figure 3
(i)
mit the route request (RREQ) packets. These packets depicts the algorithm for updating CSi and Pconn upon the
are broadcasted by a node to inquire about the next hop receipt of an RTS/CTS packet from node j.
to a given destination. By controlling the transmission The computation of Pij is simple since node i es-
power of a RREQ packet, the MAC layer effectively con- timates the channel gain Gij from j’s control signals
trols the set of candidate next-hop nodes. From a power (RTS/CTS/hello). However, computing the power re-
consumption standpoint, a smaller transmission power is quired for indirect communication requires the exchange
preferable, which also means a smaller set of next-hop of additional information between one-hop neighbors, as
nodes. But reducing the size of this set may result in los- we now explain. Each node i keeps a list Ni of all one-
ing network connectivity. Hence, the goal is to provide hop neighbors that are within the maximum transmission
 5

(i) all x ∈ CSi . We now show that Piu + Puv + Pvj > Pij for
UPDATE-CS(CSi , j, Pconn )
any nodes u and v in CSi . The proof is by contradiction,
1 for every node u ∈ CSi do
i.e., suppose that Piu + Puv + Pvj ≤ Pij for some nodes u
2 if Piu + Puj ≤ Pij
and v in CSi . Then the communication i → u → v must
3 terminate UPDATE-CS
require less power than i → v, and hence, node v cannot
4 end-for
(i) be in CSi . This contradicts the assumption that v ∈ CSi .
5 if Pij < Pconn
As mentioned in Section II-B, maintaining network
6 for every node u ∈ CSi do
connectivity is crucial. The following theorem shows that
7 if Pij + Pju ≤ Piu
if the network is connected under the standard maximum-
8 CSi ← CSi − {u}
power approach, then it must also be connected when each
9 end-for
node communicates only with nodes in its connectivity
10 CSi ← CSi ∪ {j}
(i)
set.
11 Pconn ← max {Piu : u ∈ CSi } Theorem 1: Let G = (V, E) be the undirected graph
12 end UPDATE-CS that results from using the power Pmax to reach other
nodes. Let H = (V, E 0 ) be the undirected graph con-
(i) structed based on the CS approach. If G is connected,
Fig. 3. Algorithm for updating CSi and Pconn after receiving a
control packet from node j. then H is also connected.
Proof: See Appendix.
One nice feature of the algorithm is its symmetrical
range (Pmax ) of node i. Note that CSi ⊂ Ni . Node i also property: if i ∈ CSj then j ∈ CSi , and vice versa. The
maintains the value of the minimum transmission power reason for this property is that if the direct path from node
required to communicate directly with each node in Ni . i to node j is more power efficient than any other path,
The set Ni is updated dynamically whenever node i over- then so is the direct path from j to i.
hears a control packet (RTS, CTS, or hello) in its neigh- At high loads, there is enough RTS-CTS activity to al-
borhood. Initially, when node i comes up, it broadcasts its low for the computation of the connectivity set at no ex-
current Ni at the power Pmax . Subsequently, whenever a tra bandwidth overhead. However, at light loads, channels
new node is added to Ni or whenever the minimum power are often idle, and an auxiliary scheme is needed to ensure
required to communicate directly with an existing node accurate computation of the connectivity set. In our pro-
in Ni changes by more than a threshold, say  dB, then tocol, we let each node broadcast a “hello” packet over
only the updated information is piggybacked on the next the data channel at power Pmax every ∆ seconds, where
“hello” message that is broadcasted by node i at power ∆ is a random variable that is uniformly distributed in the
Pmax . All nodes in Ni will receive this broadcast. Let interval [T /2, T ]. Randomization is needed to avoid col-
u ∈ Ni . When u receives i’s update, it uses that informa- lisions between synchronized “hello” transmissions. The
tion along with Nu to update its connectivity set CSu , as value of T is determined according to the overall mobility
described in Figure 3. pattern in the network. For example, for conference room
An advantage of the above approach is that it accounts scenarios, the network topology hardly changes within a
for the effect of shadowing (as part of the large-scale chan- 3-second interval, so T can be set to, say, 4 seconds. The
nel variations), independent of any specific propagation format of the “hello” packet is similar to that of the IEEE
model. The communication overhead is relatively low 802.11 CTS packet, except for two changes. First, the ad-
since changes in channel gains due to shadowing occur dress field in the IEEE 802.11 CTS packet that indicates
on the time scale of seconds (channel gain is a character- the receiver address is now used to indicate the transmit-
istic of large-scale models). Furthermore, these channel ter address. Second, the duration field in the standard CTS
gains are broadcasted locally, and are not flooded beyond packet is now used for a different purpose, which will be
the maximum-power neighborhood of a node. explained in Section II-E. Figure 4 shows the format of
In deciding whether to add node j to CSi or not, we the “hello” packet. Note that initially the CS of a node is
only considered the two-hop indirect paths. The reason empty. However, it takes only T seconds in the worst case
is that if the two-hop path is less power-efficient than the for the node to discover its neighborhood and start using
direct path, then so are the L-hop paths, L ≥ 2. We now a reduced power. The above “hello” approach incurs lit-
prove this claim for the case L = 3, and the general case tle overhead. This is in contrast to the scheme in [32],
follows by induction. Suppose that node i has just heard where periodic or on-demand reconfiguration of the net-
a control signal from node j and that Pix + Pxj > Pij for work topology is always needed if nodes are moving (the
 6

Octets: 2 6 2 4

Frame Control Transmitter Address Reserved Field FCS (32−bit CRC)

Fig. 4. Format of the “hello” packet in PCDC.

authors simulated only a static network). This affects net- period over the channel between nodes that intend to prop-
work resources and increases packet delays, especially at agate the RREQ. This results in many collisions between
peak load times. Our scheme, on the other hand, exploits RREQ packets (the transmissions of which are typically
information freely available through the control channel unacknowledged), which delays the process of finding the
at those peak times. destination and requires retransmitting these packets. In
Now that node i has computed the connectivity power
(i)
Pconn , it uses
this power level to broadcast its RREQ pack-
ets. This results in two significant improvements. First,
any simple min-hop routing protocol, such as AODV or
DSR, can now be used to produce routes that are very A

power efficient and that increase network throughput (i.e., B

reduce the total reserved floor). Hence, no intelligence

is needed at the network layer and no link information
D
(e.g., power) has to be exchanged or included in the RREQ
packets in order to find power-efficient routes. Second,
considering how RREQ packets are flooded throughout
the network, significant improvements in throughput and Fig. 5. Example showing the inefficiency of broadcasting
RREQ packets at the maximum power.
power consumption can be achieved by limiting the broad-
casting of these packets to nodes that are within the con-
(i)
nectivity range Pconn . Take, for example, the network in contrast, in PCDC the RREQ packet is broadcasted to
Figure 5 (this topology approximates a classroom envi- the connectivity set only (and not to the maximum reach-
ronment). Suppose that DSR is used for route discovery. able set) and hence, the number of contenders following
Consider first the “standard” approach, whereby RREQ a RREQ does not vary significantly, making it possible
packets are transmitted at power Pmax . If node A intends to design an efficient contention window for RREQ pack-
to send a packet to node D, it broadcasts a RREQ packet at ets. Therefore, in the process of finding the destination,
Pmax . Upon receiving A’s RREQ packet, node B searches PCDC results in lower overhead, less contention, and less
its route cache for the next hop to the destination node D. consumed power.
If no route is found, node B forwards the RREQ packet
to its neighbors after adding its own address. Note that
D. Channel Access Mechanism
all nodes in A’s maximum transmission range will per-
form the same procedure. The RREQ packet propagates We now describe the admission control and channel ac-
through the network until it reaches the destination or a cess strategy in the PCDC protocol. RTS and CTS pack-
node with a route to the destination. Simulation results ets are used to provide three additional functions. First,
in [17] show that for DSR, the overhead of RREQ pack- they allow nodes to estimate the channel gains between
ets in bytes is approximately 38% of the total received transmitter-receiver pairs. Second, a receiver i uses the
bytes. Obviously, this overhead and the corresponding CTS packet to notify its neighbors of the additional inter-
(i)
power consumption are significant. A close look at DSR ference power (denoted by Pnoise ) that each of the neigh-
reveals that these drawbacks become more significant as bors can add to node i without impacting i’s current re-
the range over which the RREQ packets are broadcasted is ception (hence, allowing for interference-limited concur-
made larger. First, as this range increases, the number of rent transmissions). These neighbors constitute the set of
receivers that receive multiple inquiries for the same des- potentially interfering nodes. Finally, each node keeps lis-
tination also increases. As we pointed out earlier, a sig- tening to the control channel regardless of the signal des-
nificant amount of energy is consumed in just receiving a tination in order to keep track of its connectivity set, as
transmission. More significantly, following the transmis- explained in Section II-C. These functions are now ex-
sion of a RREQ packet, there will be a high contention plained in detail.
 7

If node j has a packet to transmit, it sends an RTS CTS, since the requested transmission power may cause a
packet over the control channel at power Pmax and in- collision in the neighborhood of j. On the other hand, if
(j) (ji)
cludes in this packet the maximum allowable power level Pmap > Prequested , then node j can transmit to node i using
(j)
(Pmap ) that node j is allowed to use without disturbing any (ji)
Prequested without disturbing any of the ongoing receptions
ongoing reception in its neighborhood. The exact compu- (i)
tation of this power will be discussed shortly. The format in j’s vicinity. Scaling up the transmission power by αmin
of the RTS packet is similar to that of the IEEE 802.11 amounts to “inflating” the total interference η (i) in (1) by
except for an additional one-byte field that indicates the the same factor. So the additional interference that node i
(j) can tolerate from future unintended transmissions is given
value of Pmap .
by:
Upon receiving the RTS packet, the intended receiver, (i) (i)
say node i, uses the known Pmax value and the power PMAI-future = η (i) (αmin − 1). (3)
(ji)
of the received signal Preceived to estimate the channel (i)
We refer to PMAI-future as the interference margin at node
(ji)
gain Gji = Preceived /Pmax between nodes i and j at that i.
time (note that we assume channel reciprocity, and so The next step is to equitably distribute this power tol-
Gij = Gji ). Accordingly, node i will be able to correctly erance among future potentially interfering users in the
decode the data packet if this packet was transmitted at vicinity of i. The rational behind this distribution is to pre-
(ji) (i)
power Pmin given by: vent one neighbor from consuming the entire PMAI-future .
(i)
(i) In other words, we think of PMAI-future as a network re-
(ji) SNRth (Pthermal + PMAI-current ) SNRth η (i)
Pmin = = source that should be shared among the neighbors of i.
Gji Gji Let K (i) be the number of nodes in the vicinity of i that
(1) (i)
where SNRth is the minimum SNR ratio that is needed are to share PMAI-future . This number is determined as fol-
to achieve the target bit error rate at that receiver (we as- lows. Node i keeps track of the instantaneous number
sume SNRth is the same for all nodes, i.e., all nodes re- of simultaneously active transmissions (i.e., load) in its
(i)
quire the same QoS), Pthermal is the thermal noise power, neighborhood, which we donate by Kinst . This can be eas-
(i)
PMAI-current is the current multiple access interference ily achieved by monitoring the RTS/CTS exchanges over
(MAI) from all already ongoing (interfering) transmis- the control channel. In addition, i keeps track of a moving
(i) (i)
def (i)
sions, and η (i) = Pthermal + PMAI-current . Note that because average of Kinst , denoted by Kavg . Then, K (i) is calcu-
of the assumed stationarity in the channel gain over small lated as follows:
time intervals, Gji is approximately constant throughout (
(i) (i) (i) (i)
(i) β(Kavg − Kinst ), if Kavg > Kinst
the transmissions of the control packet and the ensuing K = (4)
β, otherwise
data packet.
(ji)
The value of Pmin in (1) is the minimum power that where β > 1 is a safety margin. The rationale behind (4)
node j must use for data transmission in order for node (i)
is as follows. At a given time instant, there are Kinst active
i to correctly decode the data packet at the current level transmissions in the neighborhood of i, the interference of
(ji)
of interference. This Pmin , however, does not allow for which must have been accounted for in the current value
any interference tolerance at node i, thus all neighbors of of η (i) . Note that before the start of their transmissions,
node i will have to defer their transmissions during node (i) (i)
these Kinst interferers were accounted for in PMAI-future ,
i’s ongoing reception (i.e., no simultaneous transmissions but once they have started their transmissions, their in-
can take place in the neighborhood of i). To allow for a (i) (i)
terference is now part of η (i) . This leaves Kavg − Kinst
number of future interfering transmissions to take place
potential future interferers to share the current value of
in its vicinity, receiver i requests that node j scales up (i) (i)
(ji) (i) the interference margin. As Kinst increases beyond Kavg ,
the transmission power Pmin by the factor αmin , where
(i)
there are fewer inactive neighboring nodes that could be-
αmin ≥ 1. Therefore, node j must use a transmission come potential future interferers. We limit the number of
power given by: such interferers to β.
(i)
(ji) (i) (ji) Accordingly, the interference tolerance Pnoise that each
Prequested = αmin Pmin . (2)
future neighbor can add to node i is given by:
(i)
The computation of αmin will be explained in the next sec- (i)
(j) (ji) (i) PMAI-future
tion. If Pmap < Prequested , then node i does not send a Pnoise = . (5)
K (i)
 8

Octets: 2 2 6 2 1 4

Frame Control Duration Receiver Address Interference Margin Requested Power FCS

Additional Fields

Fig. 6. Format of the CTS packet in the proposed protocol.

When responding to j’s RTS, node i indicates in its argument. Part (a) of the figure shows node A com-
(ji)
CTS the power level Prequested that j must use. In addition, municating with node B, and node C (which is just
(i) outside the reception range of node B) communicat-
node i inserts Pnoise in the CTS packet and sends this
ing with node D, all according to the IEEE 802.11
packet back to node j at Pmax over the control channel.
approach. The distance between B and the intended
The format of the CTS packet is shown in Figure 6.
transmitter A is comparable to the distance between
A potentially interfering node, say s, that hears the CTS B and the interfering transmitter C. Since C is trans-
message uses the signal strength of the received CTS to mitting at power Pmax , the chances that this situation
compute the channel gain Gsi between itself and node i. will cause a collision at node B are high.
(i) Now, consider part (b) of Figure 7, where again A
The channel gain along with the broadcasted Pnoise values
(s) communicates with B and C with D, this time ac-
are used to compute the maximum power Pmap that s can cording to the proposed PCDC protocol. In this case,
use in its future transmissions. This is the power that node since A can only communicate with nodes within its
s can use in its future transmissions that will not add more connectivity set, the distance between B and the in-
(i)
than Pnoise to the received noise at node i. Furthermore, tended transmitter A is much larger than the distance
(s)
Pmap is updated dynamically whenever s overhears a new between B and the interfering transmitter C. Fur-
(k)
CTS, and is taken as the minimum of the Pnoise /Gsk val- thermore, the interfering transmitter C is transmit-
ues, for all neighbors k of s. Note that it is possible for ting at a power that is much less than Pmax . So the
more than K (i) nodes to start transmitting during i’s re- interference that C causes at B is much less than in
ception and this may result in MAI at i that is greater than the case of the IEEE 802.11 scheme.
(i) (i)
PMAI-future . We address this issue in Section II-G. (ii) Let PMAI-other denote the part of the multiple access
Obviously, the interference range is larger than the re- interference (MAI) at receiver i that is attributed to
(i)
ception range set by Pmax (in theory, any unintended trans- nodes outside i’s Pmax range. PMAI-other fluctuates as
mission causes some interference), so collisions may still neighboring nodes finish their transmissions or start
(i)
occur because of interferers outside the reception range. new ones. However, since PMAI-other is caused by a
This problem is present in the IEEE 802.11 scheme as large number of interfering nodes (all nodes between
well. But compared to the 802.11 scheme, PCDC sig- Pmax and ∞), we can assume, using the law of large
nificantly reduces the severity of this problem, for the fol- (i)
numbers, that the mean value of PMAI-other is almost
lowing reasons: fixed. Now, this mean has already been accounted
(i) According to the PCDC protocol, a node, say j, can (i)
for in Equation (2) in the paper (in PMAI-current ; the
communicate directly only with neighbors who are current noise plus interference at node i). In other
in j’s connectivity set (CSj ), which is associated with words, the receiver will request that the transmitter
(j) (i)
the transmission power Pconn . This power is typically uses a power to combat PMAI-current , which already
much smaller than Pmax . So when node j communi- (i)
includes the mean of PMAI-other .
cates with another node i (which must be in CSj ),
(iii) Node i receives the signal at a scaled up power level,
it uses transmission power Pji that is typically much
allowing for some interference margin. If the inter-
less than Pmax . Since Pji < Pmax , the interference
ference power goes above the margin, then node i
that node j causes to nodes that are outside j’s Pmax
can respond with a special CTS packet over the con-
range can be considered small, and vice versa. This is
trol channel, preventing the RTS sender from com-
in contrast to the 802.11 approach, where nodes can
mencing its transmission.
communicate with any other node within the Pmax
range. The approach we discussed in this section provides a
We use the example in Figure 7 to illustrate the above distributed mechanism for admission control. In contrast
 9

Connectivity set range for A

A DATA B C DATA D A B C D
Interference Interference

(a) IEEE 802.11 scheme. (b) Proposed PCDC protocol.

Fig. 7. Impact of interference (the maximum power of each node is indicated by a dashed circle centered at that node, while in
part (b), the solid circles indicate the connectivity set ranges of transmitters A and C).

to cellular systems where the base station makes the ad- received at two different nodes consists of the power sig-
mission decision, in here each node, and depending on nals received from two different sets of transmitters. To
previously heard RTS and CTS packets, decides whether account for these differences we treat the problem in a
its transmission can proceed or not. slightly different manner. First, while in the cellular sce-
nario the base station applies the algorithm only to active
users, in our case the notion of “users” is different, as it
E. Interference Margin
(.)
refers to the expected number of future users. Second, in
In this section, we show how the scaling factor αmin is our case, each node i tries to accommodate nodes that are
computed dynamically. In [5] a power-control algorithm within its own maximum transmission range, since those
was proposed for the uplink channel of a DS-CDMA cel- are the nodes which node i may interfere with.
lular system. The purpose of that algorithm is to main- To implement power scaling in a distributed manner,
tain the QoS of ongoing users while simultaneously max- node i uses its dynamically computed connectivity power
imizing the free capacity for new users. We propose a (i)
Pconn to compute the maximum scaling constant α(i) that
distributed algorithm to implement the idea in MANETs.
node i can accommodate:
First, note that the SNR at a receiving node i is given by:
Pmax
(i)
α(i) = (i)
(7)
Pj Pconn
SNR(i) = (i) (i)
(6)
Pthermal + PMAI-current + PMAI-future
This value represents the maximum scaling constant that
(i) node i can be asked to use. A larger value implies that
where Pj is the “desired” power at the receiver i from
one node will have to transmit at power greater than Pmax ,
the intended transmitter j (see the previous section for the
which is not possible. Note here that a more clustered
definition of the other variables). It was proven in [5] that
(i) topology would result in a larger interference margin, and
to increase channel capacity, PMAI-future must be increased
(i)
hence, more simultaneous transmissions.
(and so, from (3), αmin must also be increased). The au-
While the maximum available capacity for prospective
thors proposed an algorithm that scales up the power of
transmitters can be achieved by maximizing α(i) , this has
active links (transmissions in progress) by the largest pos-
a negative effect on the node’s battery life. The reasons is
sible constant α. This constant α is calculated to accom-
because the transmitter scales up the transmission power
modate the user with the maximum ratio of the currently
by α(i) , thus, as α(i) increases, more energy is consumed
used power over the peak power imposed by the hard-
to deliver the packet. To account for these two conflicting
ware. If α is made larger than that, then at least one of the (i)
goals, we use the ratio of the remaining energy (Eremain )
users will be peak-power limited (i.e., reaches its maxi- (i)
mum power) and will be unable to attain its QoS. to the full energy (Efull ) of the battery to scale down the
The authors in [5] presented a centralized algorithm value of α(i) as follows:
that implements the aforementioned power-scaling idea
at the base station. Applying the same algorithm in  $
(i)
%
ξ×E
MANETs is not so straightforward due to the absence of

 remain 

(i)
 
E

 

a centralized control. Moreover, in a MANET, the chan- (i) full
αeff = max 1, α(i) (8)
nel consists of overlapping regions where nodes do not 

 ξ 


 
hear all transmitted signals. This means that the power  
 10

where ξ is a pre-specified positive integer, say 4. Note that duration (≈ 1%). Hence, we propose the use of a second
control channel for sending ACK messages.
 (i) (i)
 0 if 0 ≤ Eremain /Efull < 1/ξ In our scheme, if a node, say i, hears an RTS that is in-
(i) (i)
 
(i)
α(i) /ξ

ξ×E
 if 1/ξ ≤ Eremain /Efull < 2/ξ
tended for some other node, then node i defers from trans-
remain

 (i) (i)
(i)
2α(i) /ξ

E
 if 2/ξ ≤ Eremain /Efull < 3/ξ
α(i) full
= .. mitting over the ACK control channel for the duration of
ξ . an ACK packet. This deference duration starts right af-


 (i) (i)
(ξ − 1)α(i) /ξ if (ξ − 1)/ξ ≤ Eremain /Efull < 1



 (i) (i) ter the end of transmission of the data packet (computed
α(i) if Eremain /Efull = 1
(9) from the information in the RTS). In case of two neigh-
boring nodes that start their data receptions at different
The value of ξ is battery dependent, and can be se- times but complete them at the same time, the one with
lected by the system designer to reflect any given battery the later start-of-reception must wait for the duration of
model (different values of ξ can be used by different nodes an ACK packet before acknowledging the receipt of the
in the same network). Other forms of utility functions data packet.
can be used to control the throughput/battery life tradeoff Although PCDC uses a collision avoidance backoff al-
(i)
(e.g., exponentially decreasing the value of αmin to one as gorithm similar to the IEEE 802.11b standard, more so-
(i) (i) (i) phisticated backoff algorithms such as the one in [7] can
Eremain /Efull approaches zero, etc.). Note also that αeff
also be used.
must be greater than or equal to one, or otherwise coher-
ent reception at node i is not possible. Node i broadcasts
(i)
the value of αeff in the reserved field of the “hello” pack- G. Protocol Recovery
(i)
ets mentioned in Section II-C. The value of αmin is set to In [11] the authors observed that when the transmis-
(.) sion and propagation times of control packets are long,
the minimum of the αeff values that node i receives from
(i) (j)
its neighbors, i.e., αmin = minj∈Ni αeff . The intuition is the likelihood of a collision between a CTS packet and an
(i) RTS packet of another contending node increases dramat-
that if the scaling factor is made larger than αmin , then at
ically; the vulnerable period being twice the transmission
least one of the nodes that is within the maximum range
duration of a control packet. At high loads, such a colli-
of node i will be peak-power limited (or battery limited)
sion can lead to collisions with data packets, as illustrated
and will be unable to attain its QoS while conserving its
in Figure 8. Suppose that node D starts sending a RTS
battery energy if it needs to start a communication with
to node C while C is receiving B’s CTS that is intended
one of its connectivity set neighbors.
to A. A collision happens at C, and hence, C is unaware
of B’s subsequent data reception. Afterwards, if C de-
F. Link Layer Reliability cides to transmit a CTS to D, it will destroy B’s reception.
Another problem that was mentioned earlier is if the inter-
Providing link-layer error control is important not only
because it provides faster recovery than transport-layer Collision at C

error control, but also because the performance of tradi-

tional transport layer protocols (such as TCP) degrades
significantly over wireless links, resulting in a large num-
ber of unnecessary retransmissions [8]. This can reduce
D C B A
throughput, incur unacceptable delays, and consume bat- RTS CTS

tery energy.
The protection of ACK packets was addressed in previ-
ous MAC protocols (e.g., [1], [6]), but in the absence of
power control. For example, in the IEEE 802.11 standard,
Fig. 8. Example of a collision between control packets that
a node that hears an RTS packet must defer its transmis- eventually leads to a collision with a data packet.
sion, since it may destroy the reception of the ACK at the
sender. While such an approach is fundamentally needed
(.)
to protect the ACK, it reserves the floor around the trans- ference power goes above PMAI-future . In PCDC, we solve
mitter for the whole duration of the data and ACK trans- the above two problems as follows. If while receiving a
missions, when, in fact, the floor needs to be reserved for data packet, node i hears over the control channel a RTS
the duration of the ACK packet only. In practice, the ACK message (destined to any node) that contains an allowable
(.)
transmission period is small compared to the data-packet power Pmap value that if used could cause an unacceptable
 11

interference with node i’s ongoing reception, then node i 3) K (i) : The expected number of simultaneous trans-
shall respond immediately with a special CTS packet over missions around node i, as computed in (4).
the control channel, preventing the RTS sender from com- 4) CSi : A table that contains the CS of node i. For
mencing its transmission. The duration field of the CTS each node v ∈ CSi , the table contains v’s address,
packet contains the time left for node i to finish its ongo- the channel gain Gvi , and the associated time stamp.
ing reception. To see how this solution helps in reducing 5) Ni : A table that contains the neighbors of node i.
the likelihood of collisions with data packets, consider the For each node v ∈ Ni , the table contains v’s ad-
situation in Figure 8. Suppose that node A sends a RTS dress, the channel gain Gvi , and the associated time
to node B, and B responds back with a CTS that collides stamp.
at C with a RTS from node D. Now, C does not know Table I lists the parameters used in the rest of this sec-
about B’s ongoing reception. Two scenarios can happen. tion. The actions taken by a node at various stages of its
In the first, node C may later wish to send a packet to,
say, node D. It sends a RTS, which will be heard by node (i)
Precv cont Received power of a control packet at node i
B. Node B responds back to node C with a special CTS. Tdata Duration of a data packet
Note that there is a good chance that B’s special CTS will TCTS Duration of a CTS packet
collide with the CTS reply from D; however, this is de- TACK Duration of an acknowledgement packet
sirable since C will fail to recover D’s CTS packet, and TPROP Maximum propagation delay
will therefore defer its transmission and invoke its backoff TSIFS Short interframe space1
procedure. In essence, B’s special CTS acts as a jamming TDIFS Distributed interframe space1
signal to prevent C from proceedings with its transmis-
TABLE I
sion. The second possible scenario is that D (or any other
PARAMETERS USED IN IMPLEMENTING THE PCDC PROTOCOL .
node that is out of the maximum range of node B) may
send a new RTS to node C. Node C will respond to node
D with a CTS, and D will start sending data to node C. operation are as follows:
Simultaneously, node A may be sending to B, without
Step 1: When node j intends to send a packet to node
any collision. This is possible because in PCDC, DATA
i:
and RTS/CTS packets are sent on separate channels.
• Node j must wait for both the physical carrier-sense
Note that in PCDC we try to avoid highly probable col- (j)
lision scenarios like the one mentioned in [11]. However, mechanism and the NAVcont to indicate that the con-
there will still be few complicated (and definitely much trol channel has been idle for a duration of TDIFS .
less probable) scenarios where data packets may collide; After that, node j generates a random backoff period
recovery from such collisions is left to the upper layers. for an additional deferral time before transmitting
(unless the backoff timer already contains a nonzero
value). If the medium is determined to be busy at
III. PCDC AT W ORK
any time during a backoff slot, then the backoff pro-
In this section, we provide the operational details cedure is frozen and only allowed to resume after the
of PCDC. As is the case in practice, we assume medium is determined to be idle for a TDIFS interval.
that each node i has M transmission power levels At this point, node j sends an RTS packet over the
P (i) [1], . . . , P (i) [M ] (for example, CISCO 350 series control channel at Pmax . In that packet, j includes
Aironet has six levels [3]). Each node i maintains the fol- (j)
its Pmap and the time duration Tdata of the yet to be
lowing variables:
transmitted data packet.
1) NAV(i) [x], x = 0, 1, . . . , M (data channel virtual- • After transmitting the RTS packet, node j sets a timer
sense mechanisms for node i). This is the amount of to a timeout value of 2TPROP + TSIFS + TCTS seconds.
time during which node i is not allowed to use the This value is the sum of the time for the RTS packet
power level P (i) [x]. It is implemented as a counter to reach the destination (TPROP ), the time the receiver
that counts down to zero at a uniform rate. The must wait before sending back the CTS (TSIFS ), the
NAV(i) [x] values constitute the network allocation time it takes the CTS to reach the sender (TPROP ),
vector NAV for node i, which is similar to the one and the CTS transmission duration (TCTS ). If after
used in the IEEE 802.11 standard [1]. this period node j has not received a correct CTS
(i)
2) NAVcont (control channel virtual-sense mechanism packet, it concludes that the transmission of the RTS
for node i). This is the time period during which
1
the control channel will be busy (another counter). As defined in the IEEE 802.11 standard [1].
 12

has failed, and hence, it invokes its backoff proce- as follows:

dure. (m)
Step 2: When a node n receives j’s RTS that is intended NAV(m) [x] = 2TPROP +Tdata ∀x : P (m) [x] > Pmap
for node i: but only if NAV(m) [x] already contains a smaller
(n) (n)
• It updates its NAVcont as follows: NAVcont = value.
2TPROP + TSIFS + TCTS , which is the maximum du- Step 5: When node j receives i’s CTS:
ration it takes receiver i to respond with a CTS. (ji)
(n) • It transmits the data packet at power level Prequested
P
• It computes the channel gain Gnj = recv cont
Pmax , and (included in the CTS packet).
(j)
uses this value to update both Nn and CSn , as ex- P
• It computes the channel gain Gij = recv cont and
Pmax
plained in Section II-C.
uses this value to update both Nj and CSj , as de-
Step 3: When node i receives an intended RTS packet scribed before.
from node j:
(i)
• After transmitting the data packet, node j sets a timer
Precv cont
• It computes the channel gain Gij = Pmax , and to 2TPROP +TACK seconds. If after this period node j
uses this value to compute the
(ji)
power Prequested
that has not received a correct ACK packet, it concludes
the data packet should be sent at using (2). that the transmission of the data packet has failed,
(ji) (j) and hence, it invokes its backoff procedure.
• If Prequested < Pmap (which is provided in the RTS
Step 6: Once the destination node i has successfully re-
packet), then node i sends a CTS packet over the con-
ceived the data packet, it immediately transmits the ACK
trol channel at Pmax . The CTS is transmitted after a
packet over the ACK control channel.
TSIFS period if both the physical carrier-sense mech-
(i) In parallel with the above actions, each node, say i,
anism and NAVcont indicate that the control chan-
must do the following:
nel is idle. The CTS packet includes the interfer- (i)
(i) (ji) • Decide on the value of αeff using (7) and (8).
ence margin Pnoise , the requested power Prequested , (i)
and the length of the yet to be received data packet • Broadcast the computed αeff as part of the “hello”

Tdata (copied from the RTS packet). packet, as explained in Section II-E.
(.)
• After transmitting the CTS packet, node i sets a timer • Cache the minimum of all the αeff values it has heard.
(i)
to 2TPROP + TSIFS seconds. If after this time node i • Compute K (i) using (4), and Pnoise using (5)
has not started receiving a data packet (recognized The IEEE 802.11 specifications [1] state that “the trans-
from its header), it concludes that the transmission mit power-on ramp for 10% to 90% of maximum power
of the CTS has failed. At this point, node i sends a shall be no greater than 2 µsec,” and that “the transmit
control packet announcing the release of the channel. power-down ramp for 90% to 10% maximum power shall
This packet is just another CTS packet with zero in be no greater than 2 µsec.” Given that the RTS (or CTS)
the duration field. Note that although the transmitter transmission duration is in the order of 100s of µsec and
sends the data immediately after receiving the CTS the data transmission duration is in the order of 1000s of
packet, a TSIFS period is needed for the receiver i to µsec, the delay attributed to changing power levels (≤ 2
process the packet header. µsec) can be safely ignored. Note also that this delay is
Step 4: When an irrelevant node m receives i’s CTS: less than the turn-around period (period it takes a node to
(m) (m) switch from a receiving mode to a transmitting mode); the
• It updates its NAVcont as follows: NAVcont = TPROP
which allows the CTS to reach back to sender j. later is approximately 5 µsec. Hence, after node j trans-
P
(m) mits an RTS packet at power Pmax , the data packet can go
• It computes the channel gain Gmi = recv cont and
Pmax out on a different power level with a very small delay that
uses this value to update both Nm and CSm , as de- has negligible effect on the system efficiency.
scribed before.
• It finds the maximum power level that it can use with-
(i) IV. R ELATED W ORK
out adding more than Pnoise to the ongoing reception
at node i. This is calculated as follows: Previous schemes for power control in MANETs have
focused on either throughput enhancement or energy con-
(i)
(m) Pnoise sumption. None of these schemes provide a compre-
Pmap = hensive solution that enables a node to communicate via
Gmi
energy efficient links using different transmission ranges
Accordingly, node m updates its NAV(m) [x] vector while still maintaining exclusive use of the channel (i.e.,
 13

proper MAC functionality). In [27] the authors suggested of the total received bytes [17]).
a protocol that exploits global topological information The problem of adjusting the transmission power of
provided by the routing protocol to reduce the nodes trans- broadcast messages was addressed in [9]. The proposed
mission powers such that the degree of each node is upper- approach relies on using the distance information be-
and lower-bounded. In [32] a cone-based solution that tween nodes to construct a restricted neighborhood graph
guarantees network connectivity was proposed. The au- (RNG). A node adjusts its transmission power to reach
thors in [13] proposed the use of a synchronized global only those nodes that are in its RNG. This approach was
signaling channel to build a global network topology in- improved in [10] by taking into account physical shadow-
formation where each node communicates only with its ing, realistic MAC protocols and the effect of collisions.
nearest N neighbors (N is a design parameter). In [29] the In [4], [33], a single channel was used to send the RTS-
authors proposed a position-based distributed algorithm CTS control packets but at different power levels. This
aided by a GPS system to allow each node to communi- again results in the situation in Figure 2. In [19], [14], [26]
cate only with its enclosure region. One common defi- the authors proposed that communicating nodes exchange
ciency in the above protocols is that they rely solely on their RTS and CTS packets at power Pmax , but send their
CSMA for accessing the wireless channel. It has been DATA/ACK packets at the minimum power Pmin needed
shown in [30], [21] that using CSMA alone for accessing for reliable communication. The energy consumption in
the wireless channel significantly degrades network per- this approach is expectedly less. However, similar to the
formance. 802.11 scheme, control signals are used to silence neigh-
The COMPOW protocol [25] relies completely on boring nodes, preventing concurrent transmissions in the
routing-layer agents to converge to a common lowest neighborhood of a receiver. In fact, as pointed out in
power level for all network nodes. However, for con- [18], such schemes achieve less throughput than the IEEE
stantly moving nodes, the scheme (like any routing- 802.11 since they introduce a new problem; interference
protocol-based scheme) incurs significant overhead, and with the reception of the ACK message at the source node.
convergence to a common power level may not be possi- The authors in [18] proposed a solution to that problem.
ble, leading to a situation like the one described in Figure The proposed protocol in that paper was shown to pre-
2. Moreover, in situations where network density varies serve energy without decreasing the network throughout
widely (i.e., nodes are clustered), restricting all nodes to below that of the IEEE 802.11. However, the protocol
converge to a common power level can be conservative, in [18] does not allow for any concurrent transmissions
and may achieve little gain. to take place over the reserved floor, where the reserved
Clustering as proposed in [22] is another interesting ap- floor is the maximum transmission range (i.e., the control
proach for power control. An elected cluster head (CH) packets transmission range).
performs the function of a base station in a cellular sys- Of the several schemes for power control, the ones in
tem. It uses closed-loop power control to adjust the trans- [24], [34] are the most relevant to our scheme. Our work is
mission powers of nodes in the cluster. Communications in line with [24] in the sense that we use the signal strength
between different clusters occur via GateWays, which are of a received control message to bound the transmission
nodes that belong to more than one cluster. This approach power of neighboring nodes. However, our scheme differs
simplifies the forwarding function for most nodes, but at from [24] in the following ways. First, the protocol in [24]
the expense of reducing network utilization since all com- relies on the network layer to find a power efficient next
munications have to go through the CHs. This can also hop. In dense networks, where power control is supposed
lead to the creation of bottlenecks. A joint clustering and to achieve a higher channel reuse factor, the next hop will
power control protocol was proposed in [20], where each be in the maximum range region, and hence, little gain (if
node runs several routing-layer agents that correspond any) will be achieved in using power control. Even if we
to different power levels. These agents build their own assume that a more intelligent power-aware routing pro-
routing tables by communicating with their peer routing tocol runs on top of the scheme in [24], this incurs the
agents at other nodes (i.e., the protocol is distributed with overhead of exchanging link-power information. In ad-
no CHs). Each node along the packet route determines dition, routing packets will still have to be broadcasted
the lowest-power routing table in which the destination is at maximum power; something we avoid in PCDC. It is
reachable. The routing overhead in this protocol grows worth mentioning that the connectivity set that each node
in proportion to the number of routing agents, and can be builds in PCDC is a result of sending the control packets
significant even for simple mobility patterns (recall that (RTS-CTS) over a separate control channel at fixed power.
for DSR, RREQ packets account for approximately 38% Hence, this set cannot be built with protocols like the one
 14

described in [24]. Finally, while PCDC dynamically ad- no significant changes in topology take place within these
justs the interference margin of the receiver, depending periods. Nodes are assumed to have full energy in their
on the nodes density and battery energy left, in [24] the batteries during the simulation time and thus, (8) was not
authors use a fixed interference margin value that is deter- implemented for simplicity. Other parameters used in the
mined offline. simulations are given in Table II. These paraments corre-
A busy-tone based power control protocol was pro- spond to realistic hardware settings [3].
posed in [34], where the sender transmits the data and the
busy tone at minimum power. The receiver transmits its Data packet size 2 KB
802.11 data rate 2 Mbps
busy tone at maximum power. A neighbor estimates the
PCDC data rate 1.6 Mbps
channel gain from the busy tone and is allowed to trans- Control channel rate 400 Kbps
mit if its transmission is not expected to add more than SNR threshold 6 dB
a fixed “noise” value to the ongoing reception. However, Reception threshold −94 dBm
in the suggested protocol, the receiver does not take into Carrier-sense threshold −108 dBm
account the additional noise that future transmitters add to Thermal+receiver noise -169 dBm/Hz
the ongoing reception. Consequently, the criterion for cor- Pmax 20 dBm
rect reception will simply not be met as soon as neighbors TABLE II
start their transmissions. In addition, a similar argument PARAMETERS USED IN THE SIMULATIONS .
to the one mentioned above concerning next-hop selection
also applies to the protocol in [34].

V. P ROTOCOL E VALUATION B. Simulation Results

A. Simulation Setup We consider two types of topologies: random grid and
We now evaluate the performance of the PCDC pro- clustered. In the random grid topology, 49 mobile hosts
tocol and contrast it with the IEEE 802.11 scheme. Our are placed across a square area of length 3000 meters. The
results are based on simulation experiments conducted us- square is split into 49 smaller squares. The location of
ing CSIM programs (CSIM is a C-based process-oriented a mobile user is selected randomly within each of these
discrete-event simulation package) [2]. In our simula- squares. For each generated packet, the destination node
tions, we investigate both the network throughput as well is randomly selected.
as the energy consumption. We use two measures for Figure 9 depicts two instances of the network topol-
the throughput: channel utilization (U ) and end-to-end ogy as constructed under PCDC and 802.11, respectively.
throughput. Channel utilization refers to the average num- As expected, the topology is much denser in the case of
ber of successfully received packets per packet transmis- 802.11 because Pmax is used to determine node connec-
sion time. Essentially, it is a measure of the one-hop tivity. On average, we found that the node degree is 12.74
goodput. Note that according to this definition, U can be under the 802.11 scheme, compared to 4.81 under PCDC,
greater than 1, since multiple transmissions can occur si- which is a reduction of about 62%.
multaneously. The performance for random grid topologies is demon-
For simplicity, data packets are assumed to be of a fixed strated in Figure 10 as a function of the packet genera-
size. Each node generates data packets according to a tion rate. It can be observed that under PCDC, U is about
Poisson process with rate λ (same for all nodes). The cap- 2.5 times that of the 802.11 standard (on average). This
ture model is similar to the one in [31]. We use a min- significant increase in the utilization is achieved mainly
hop routing policy, but we ignore the routing overhead. because in PCDC, communicating nodes reserve smaller
For the 802.11 scheme, the next-hop candidates are nodes floors to achieve successful communications. This allows
that are within the maximum power range of the sender. for a tighter packing of source-destination pairs within a
For PCDC, these candidates are nodes that are within the network environment, thereby improving channel spectral
(.)
connectivity power range (based on Pconn ). The random reuse.
waypoint model is used for mobility, with a host speed Part (b) of the figure depicts the end-to-end through-
that is uniformly distributed between 0 and 2 meters/sec. put, which is a more significant measure of effectiveness
Note, however, that mobility has a little effect on our pro- than the utilization. It is shown that PCDC achieves up
tocol, since an RTS-CTS exchange precedes every packet to 45% increase in the end-to-end throughput. Further-
transmission. The transmission periods for the RTS, CTS, more, PCDC saturates at about twice the load at which
data, and ACK packets are all in tens of milliseconds, so the 802.11 scheme saturates.
 15

3000 3000

2500 2500

2000 2000
(meters)

(meters)
1500 1500

1000 1000

500 500

0 0
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
(meters) (meters)
(a) PCDC. (b) 802.11.

Fig. 9. Instances of generated network topologies under PCDC and the IEEE 802.11 standard.

Note that PCDC has a more significant impact on uti- be spent on retransmissions. For PCDC, as λ increases
lization than on end-to-end throughput. The reason for the interference increases, and so, more power will be re-
this is that PCDC limits the set of possible next-hop nodes quested by receivers to achieve their SNR thresholds.
(.)
to ones that are within the Pconn range, and thus, it forces Part (b) of Figure 11 depicts the reception energy con-
shorter transmissions and longer routes. On the other sumption versus λ. This is the average energy spent on re-
hand, the 802.11 scheme allows for communication with ceiving a data packet by both intended and unintended (or
any node that is within the maximum range, and hence, irrelevant, see Section II-B) receivers, normalized by the
produces higher progress toward the destination per hop. energy needed to receive a data packet by one node. As
Part (a) of Figure 11 depicts the transmission energy the transmission range of data packets is reduced, fewer
consumption versus λ. This is the total energy used by the irrelevant nodes receive this packet and thus, a significant
network to successfully transmit a packet end-to-end, nor- amount of energy is preserved.
malized by the energy needed to send the data packet one The packet delay performance of the PCDC and the
hop at maximum power. It includes the energy lost in re- 802.11 scheme is shown in Figure 12. For both proto-
transmitting data and control packets in case of collisions. cols, the delay increases as the traffic rate is increased.
For almost all cases, PCDC requires less than 23% of the Note, however, that as soon as λ exceeds a certain value,
energy required under the 802.11 scheme. The reason is PCDC incurs less delay than the 802.11 scheme. At first
that PCDC enforces shorter hops. In fact, our results indi- this may seem counterintuitive. One would think that
cate that, for the random grid topologies, the average num- since packets travel longer routes under PCDC, it would
ber of hops generated by PCDC is about 1.76 that of the take these packets longer times to reach their destinations
802.11 scheme (using min-hop routing). Given that the than it would under the 802.11 scheme. However, a close
signal power decreases as 1/d4 , where d is the transmitter- look at the network reveals that as λ increases, the chan-
receiver separation, PCDC should, on average, consume nel contention period also increases. In fact, as the traf-
about ν = 21% of the energy consumed by the 802.11 fic picks up, this period becomes much longer than the
scheme. However, there are two other factors that con- packet transmission period and dominates the end-to-end
tribute to ν. First, PCDC scales up the transmission power packet delay. The reason why PCDC incurs larger delays
(i) for very low load is that there is not enough packets to
by αmin as explained in Section II-E. This increases ν. On
the other hand, the node density is finite (49 nodes) in the utilize the available space efficiently. If these very low
studied topologies, and thus the transmitter-receiver sepa- traffic situations happen in a non-energy constrained net-
ration distance is not continuous. Therefore, nodes using works, then it may be desirable to disable power control,
the 802.11 scheme do not achieve the maximum range by possibly through the wireless network card interface. The
using Pmax and energy is wasted. This decreases the value disabling of power control in a node, say i, is equivalent
(i)
of ν. to setting Pconn = Pmax .
Note that in both protocols, the required energy in- The authors in [23] argued that traffic locality is the key
creases as the load increases, but for different reasons. factor for determining the feasibility of large ad hoc net-
For the 802.11 scheme, as λ increases the probability of works. This motivates studying the performance of PCDC
collisions also increases, and hence more energy has to under clustered topologies. In such topologies, a node
 16

80

End−to−End Throughput (packets/sec)

5

4 60
Channel Utilization

3
40
PCDC
802.11
2

20
PCDC
1
802.11

0 0
0 10 20 30 40 50 0 10 20 30 40 50

Packet Generation Rate (packets/sec) Packet Generation Rate (packets/sec)

(a) (b)

Fig. 10. Throughput performance of PCDC and 802.11 as a function of λ (random grid topologies).

7 60
PCDC PCDC
802.11 802.11
Transmission Energy Consumption (normalized)

Reception Energy Consumption (normalized)

6
50

5
40

30

20
2

10
1

0 0
5 10 20 5 10 20

Packet Generation Rate (packets/sec) Packet Generation Rate (packets/sec)

(a) (b)

Fig. 11. Energy consumption in PCDC and 802.11 as a function of λ (random grid topologies).

1.8 a clustered topology, we consider an area of dimensions

1.6 1000 × 1000 (in meters). 24 nodes are split into 4 equal
1.4
groups, each occupying a 100 × 100 square in one of the
Delay (seconds/packet)

1.2
corners of the complete area. For a given source node, the
destination is selected from the same cluster with proba-
1
bility 1 − p or from a different cluster with probability p.
0.8
PCDC
802.11
In each case, the selection from within the given cluster(s)
0.6
is done randomly.
0.4

0.2 Figure 13 depicts the channel utilization and network

0
0 10 20 30 40 50
throughput versus λ for p = 0.25. According to the
Packet Generation Rate (packets/sec) 802.11 standard, only one transmission proceeds at a time
since all nodes are within the carrier-sense range of each
Fig. 12. Delay performance of PCDC and 802.11 as a function
of λ (random grid topologies). other. However, according to PCDC, up to four trans-
missions can proceed simultaneously, resulting in signif-
icant improvements in channel utilization and end-to-end
throughput. Notice that for low traffic rates, and since the
communicates mostly with nodes within its own cluster, number of nodes is only 24 nodes, there is not enough
and rarely with neighboring cluster nodes. These topolo- packets to utilize the channel efficiently, which reduces
gies are common in practice (e.g., a historical site where the throughput of PCDC. This means that controlling the
users of wireless devices move in groups). To generate transmission range (and thus the next hop selection) can
 17

5 200
PCDC
802.11
PCDC
802.11

End−to−End Throughput (packets/sec)

4 160
Channel Utilization

3 120

2 80

1 40

0 0
0 10 20 30 40 50 0 10 20 30 40 50

Packet Generation Rate (packets/sec) Packet Generation Rate (packets/sec)

(a) (b)

Fig. 13. Throughput performance of PCDC and 802.11 as a function of λ (clustered topologies with p = 0.25).

2.5
reduce the throughout but only at very low traffic rates.
PCDC
802.11
Figure 14 also shows that PCDC consumes much
2
less energy to successfully deliver a data packet than the

Delay (seconds/packet)
802.11 standard. The figures shows the total energy dissi-
1.5
pated in transmissions and receptions, normalized by the
energy needed to send the data packet one hop at max-
imum power. Note that in the case of clustered topolo- 1

gies, the energy consumption for the 802.11 does not vary
with λ. The reason is that all the nodes are within each 0.5

other’s transmission range, which significantly reduces

collisions. For PCDC, concurrent transmissions in differ- 0
0 10 20 30 40 50

ent clusters add little interference to each other and so, the Traffic Generation Rate (packets/sec)

energy consumption does not vary significantly with λ. Fig. 15. Delay performance of PCDC and 802.11 as a function
of λ (clustered topologies with p = 0.25).
20
PCDC
802.11

16 VI. C ONCLUSION
Energy Consumption (normalized)

In this paper, we proposed a power controlled dual

12
channel (PCDC) MAC protocol for wireless ad hoc net-
works. To produce power-efficient routes, PCDC allows
8
the MAC layer to indirectly influence the routing deci-
sion at the network layer by controlling the power level
4 of the broadcasted RREQ packets. PCDC uses the sig-
nal strength of the overheard control (RTS/CTS) signal
0
20 30
to build a power-efficient network topology. By allow-
Packet Generation Rate (packets/sec) ing for a receiver-specific, dynamically computed inter-
Fig. 14. Total energy consumption (transmission and reception)
ference margin, PCDC enables simultaneous interference-
under PCDC and 802.11 as a function of λ (clustered topologies limited transmissions to take place in the vicinity of a re-
with p = 0.25). ceiver.
We compared the performance of PCDC to that of the
IEEE 802.11 standard. Our simulation results showed that
Finally, the delay performance of the clustered topol- PCDC can improve the channel utilization by up to 250%
ogy is shown in Figure 15. Again, the figure shows that and the end-to-end throughput by over 45% (for random
PCDC incurs much less delay for most values of λ, and grid topologies). At the same time, PCDC provides for
performs worse only for very small values of λ. more than 76% reduction in the energy consumed to suc-
 18

cessfully deliver a packet from the source to the destina- [5] D. Ayyagari and A. Ephramides. Power control for link qual-
tion. It also reduces the end-to-end packet delay. To the ity protection in cellular DS-CDMA networks with integrated
(packet and circuit) services. In Proceedings of the IEEE/ACM
best of our knowledge, PCDC is the first protocol to pro-
MOBICOM Conference, pages 96–101, 1999.
vide a comprehensive and efficient solution to the power [6] V. Bharghavan. Performance evaluation of algorithms for wire-
control problem in MANETs. Our future work will focus less medium access. In Proceedings of the IEEE Performance
on tuning the parameters of PCDC, studying a number of and Dependability Symposium, pages 86–95, Aug. 1998.
design issues, and investigating its performance under var- [7] V. Bharghavan, A. Demers, S. Shenker, and L. Zhang. MACAW:
A media access protocol for wireless LANs. In Proceedings of
ious mobility scenarios. the ACM SIGCOMM Conference, volume 24, pages 212–225,
Oct. 1994.
[8] R. Caceres and L. Iftode. Improving the performance of reli-
A PPENDIX able transport protocols in mobile computing environment. IEEE
Journal on Selected Areas in Communications, 13:850–857, June
P ROOF OF T HEOREM 1 1995.
[9] J. Cartigny, D. Simplot, and I. Stojmenović. Localized minimum-
Let x1 and xn be any pair of nodes in V . Since G is energy broadcasting in ad-hoc networks. In Proceedings of the
connected, then there exist a path in G from x1 to xn . Let IEEE INFOCOM Conference, volume 3, pages 2210–2217, Apr.
this path be {x1 , x2 , . . . , xi , . . . , xn−1 , xn }. To prove that 2003.
[10] X. Chen, M. Faloutsos, and S. V. Krishnamurthy. Power adap-
x1 and xn are connected in H, it is sufficient to prove that
tive broadcasting with local information in ad hoc networks. In
there exists a path in H between every pair of successive Proceedings of the IEEE ICNP Conference, pages 168–178, Nov.
nodes, xi and xi+1 , i = 1, 2, . . . , n − 1. 2003.
Consider any two successive nodes xi and xi+1 on the [11] J. Deng and Z. Haas. Dual busy tone multiple access (DBTMA):
A new medium access control for packet radio networks. In Pro-
G-path from x1 to xn . These two nodes are one-hop apart
ceedings of the IEEE ICUPC, pages 973–977, Oct. 1998.
in G, that is, they are within each other’s maximum trans- [12] S. Doshi, S. Bhandare, and T. X. Brown. An on-demand min-
mission range. Therefore, according to our algorithm, one imum energy routing protocol for a wireless ad hoc network.
of the following two cases must hold: ACM SIGMOBILE Mobile Computing and Communications Re-
view, 6(3):50–66, 2002.
• Case 1: xi+1 ∈ CSi , and hence xi and xi+1 are al-
[13] T. A. ElBatt, S. V. Krishnamurthy, D. Connors, and S. Dao.
ready connected in H. Power management for throughput enhancement in wireless ad-
• Case 2: xi+1 6∈ CSi . Then, there exists a node hoc networks. In Proceedings of the IEEE ICC Conference, vol-
u ∈ CSi for which the indirect communication xi → ume 3, pages 1506–1513, 2000.
[14] J. Gomez, A. T. Campbell, M. Naghshineh, and C. Bisdikian.
u → xi+1 requires less power than the direct one
PARO: Supporting dynamic power controlled routing in wireless
xi → xi+1 . Then, our problem now reduces to ad hoc networks. ACM/Kluwer Wireless Networks, 2003, to ap-
showing that there is a path in H between nodes u pear.
and xi+1 . Considering that the power required for [15] P. Gupta and P. R. Kumar. The capacity of wireless net-
transmission is proportional to d4 , then it must be the works. IEEE Transactions on Information Theory, 46(2):388–
404, March 2000.
case that du,xi+1 < dxi ,xi+1 , i.e., u and xi+1 are also [16] T.-C. Hou and V. O. K. Li. Transmission range control in mul-
within the transmission range of one another. By re- tiple packet radio networks. IEEE Transactions on Communica-
peating the same process, one can prove that either tions, 34(1):38–44, Jan. 1986.
xi+1 ∈ CSu or that u can reach xi+1 through another [17] P. Johansson, T. Larsson, N. Hedman, B. Mielczarek, and
M. Degermark. Scenario-based performance analysis of rout-
(closer) node. This process continues until we reach ing protocols for mobile ad-hoc networks. In Proceedings of the
a node that is in sufficiently close to xi+1 (i.e., in the IEEE/ACM MOBICOM Conference, pages 195–206, 1999.
CS of xi+1 ), which completes the proof. [18] E.-S. Jung and N. H. Vaidya. A power control mac protocol for
ad hoc networks. In Proceedings of the IEEE/ACM MOBICOM
Conference, pages 36–47, 2002.
R EFERENCES [19] P. Karn. MACA - a new channel access method for packet radio.
In Proceedings of the 9th ARRL Computer Networking Confer-
[1] International Standard ISO/IEC 8802-11; ANSI/IEEE Std ence, pages 134–140, 1990.
802.11, 1999 Edn. Part 11: wireless LAN Medium Access Con- [20] V. Kawadia and P. R. Kumar. Power control and clustering in ad
trol (MAC) and Physical Layer (PHY) specifications. hoc networks. In Proceedings of the IEEE INFOCOM Confer-
[2] Mesquite Software Incorporation, http://www.mesquite.com. ence, 2003.
[3] The Cisco Aironet 350 Series of wireless LAN, [21] S. Khurana, A. Kahol, and A. P. Jayasumana. Effect of hidden
http://www.cisco.com/warp/public/cc/pd/witc/ao350ap/prodlit/a350c ds.pdf.
terminals on the performance of IEEE 802.11 MAC protocol. In
[4] S. Agarwal, R. H. Katz, S. V. Krishnamurthy, and S. k. Dao. Proceedings of the IEEE LCN Conference, pages 12–20, 1998.
Distributed power control in ad-hoc wireless networks. In IEEE [22] T. J. Kwon and M. Gerla. Clustering with power control. In
International Symposium on Personal, Indoor and Mobile Radio Proceedings of the IEEE MILCOM Conference, volume 2, pages
Communications, volume 2, pages 59–66, Oct. 2001. 1424–1428, 1999.
 19

[23] J. Li, C. Blake, D. S. Couto, H. I. Lee, and R. Morris. Capacity

of ad hoc wireless networks. In Proceedings of the IEEE/ACM
MOBICOM Conference, pages 61–69, 2001.
[24] J. Monks, V. Bharghavan, and W.-M. Hwu. A power controlled
multiple access protocol for wireless packet networks. In Pro-
ceedings of the IEEE INFOCOM Conference, volume 1, pages
219–228, 2001.
[25] S. Narayanaswamy, V. Kawadia, R. S. Sreenivas, and P. R. Ku-
mar. Power control in ad-hoc networks: Theory, architecture,
algorithm and implementation of the COMPOW protocol. In
Proceedings of the European Wireless Conference, pages 156–
162, Feb. 2002.
[26] M. B. Pursley, H. B. Russell, and J. S. Wysocarski.
Energy-efficient transmission and routing protocols for wireless
multiple-hop networks and spread spectrum radios. In Proceed-
ings of the EUROCOMM Conference, pages 1–5, 2000.
[27] R. Ramanathan and R. Rosales-Hain. Topology control of multi-
hop wireless networks using transmit power adjustment. In Pro-
ceedings of the IEEE INFOCOM Conference, volume 2, pages
404–413, 2000.
[28] T. Rappaport. Wireless Communications: Principles and Prac-
tice. Prentice Hall, 1996.
[29] V. Rodoplu and T. Meng. Minimum energy mobile wireless net-
works. IEEE Journal on Selected Areas in Communications,
17:1333–1344, Aug. 1999.
[30] F. A. Tobagi and L. Kleinrock. Packet switching in radio
channels: Part ii–the hidden terminal problem in carrier sense
multiple-access and the busy-tone solution. IEEE Transactions
on Communications, 23(12):1417–1433, Dec 1975.
[31] C. Ware, J. Chicharo, and T. Wysocki. Simulation of capture
behaviour in IEEE 802.11 radio modems. In Proceedings of the
IEEE Vehicular Tech. Conference, volume 3, pages 1393–1397,
Fall 2001.
[32] R. Wattenhofer, L. Li, P. Bahl, and Y.-M. Wang. Distributed
topology control for power efficient operation in multihop wire-
less ad hoc networks. In Proceedings of the IEEE INFOCOM
Conference, volume 3, pages 1388–1397, 2001.
[33] J. Whitehead. Distributed packet dynamic resourse allocation
(DRA) for wireless networks. In Proceedings of the IEEE Vehic-
ular Tech. Conference, volume 1, pages 111–115, 1996.
[34] S.-L. Wu, Y.-C. Tseng, and J.-P. Sheu. Intelligent medium ac-
cess for mobile ad hoc networks with busy tones and power
control. IEEE Journal on Selected Areas in Communications,
18(9):1647–1657, 2000.