An Adaptive Energy-Efficient MAC Protocol

for Wireless Sensor Networks

Tijs van Dam
tijsvd@xs4all.nl

Koen Langendoen
K.G.Langendoen@its.tudelft.nl

Faculty of Information Technology and Systems

Delft University of Technology
The Netherlands

ABSTRACT

1. INTRODUCTION

In this paper we describe T-MAC, a contention-based Medium Access Control protocol for wireless sensor networks.
Applications for these networks have some characteristics
(low message rate, insensitivity to latency) that can be exploited to reduce energy consumption by introducing an
active/sleep duty cycle. To handle load variations in time
and location T-MAC introduces an adaptive duty cycle in
a novel way: by dynamically ending the active part of it.
This reduces the amount of energy wasted on idle listening,
in which nodes wait for potentially incoming messages, while
still maintaining a reasonable throughput.
We discuss the design of T-MAC, and provide a head-tohead comparison with classic CSMA (no duty cycle) and
S-MAC (fixed duty cycle) through extensive simulations.
Under homogeneous load, T-MAC and S-MAC achieve similar reductions in energy consumption (up to 98 %) compared
to CSMA. In a sample scenario with variable load, however,
T-MAC outperforms S-MAC by a factor of 5. Preliminary
energy-consumption measurements provide insight into the
internal workings of the T-MAC protocol.

Communication in wireless sensor networks can, like most

network communication, be divided into several layers. One
of those is the Medium Access Control (MAC) layer. This
layer is described by a MAC protocol, which tries to ensure
that no two nodes are interfering with each others transmissions, and deals with the situation when they do.
Wireless sensor networks have an additional aspect: as
sensor nodes are generally battery-operated, energy consumption is very important. The radio on a sensor node
is usually the component that uses most energy. Not only
transmitting costs energy; receiving, or merely scanning the
ether for communication, can use up to half as much, depending on the type of radio [8].
While traditional MAC protocols are designed to maximize packet throughput, minimize latency and provide fairness, protocol design for wireless sensor networks focuses
on minimizing energy consumption. The application determines the requirements for the (modest) minimum throughput and maximum latency. Fairness is usually not an issue,
since the nodes in a wireless sensor network are typically
part of a single application and work together for a common
purpose.

Categories and Subject Descriptors

C.2.1 [Network Architecture and Design]: Wireless
communicationMAC protocol ; C.2.5 [Local and WideArea Networks]: Access schemes; D.4.4 [Communications Management]: Message sending

General Terms
Design, Experimentation, Measurement, Performance

Keywords
Ad-hoc, sensor networks, MAC protocol, energy-efficiency

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
SenSys03, November 57, 2003, Los Angeles, California, USA.
Copyright 2003 ACM 1-58113-707-9/03/0011 ...$5.00.

1.1 Communication patterns

It is important to design and test the behavior of MAC
protocols based on the kind of traffic they have to handle.
We have identified two main communication patterns in sensor applications:
Local uni-/broadcast When a real-world event in the network occurs, we expect nodes to perform some innetwork processing. This will generally involve local
messages being exchanged between neighbors.
Nodes to sink reporting After processing a local event,
or just periodically, nodes may want to report something. We expect messages to be directed to one or a
few sink nodes, which are hooked up to a fixed network
or a computer. Messages from different nodes may,
or may not, be aggregated along the way. We do
not specify an exact routing protocol, but we expect
some random variation in message pathsmessages
flow roughly in the correct direction. In this communication pattern, we see a more or less unidirectional
flow of messages through the network.

We explicitly exclude routed, multi-hop communication

between random nodes in the network, although this pattern
is frequently used to study MAC protocols. After identifying
several realistic wireless sensor applications, we have determined that this communication pattern does not occur.
The two basic communication patterns imply that the
message rate in the network may vary, both in time and
location: events trigger periods of increased activity, and,
around sink nodes, the message rate will be higher than at
the edge of the network, even when aggregation is used.

1.2 EYES nodes

Wireless sensor hardware is generally cheap and simple.
We must consider this when designing a MAC protocol for
wireless sensor nodes. As a base for our protocol design,
simulations and implementation we use the EYES wireless
sensor nodes [13].
The EYES nodes have a Texas Instruments MSP430F149
processor with 2 KB RAM and 60 KB of Flash memory; the
16 bit processor runs at a variable clock rate, with a maximum of 5 MHz. Nodes communicate using a 115 kbps radio
(RFM TR1001, 868.35 MHz, hybrid transceiver), and are
equipped with a 2 Mb EEPROM memory (AST 25P20V6).
The nodes contain multiple interfaces to interact with the
outside world, including JTAG, RS232, 2 LEDs, and 16
general purpose I/O pins (8 with ADC capability). The
nodes run from 3 V supplied by two AA batteries taking
up most of the nodes volume. Table 1 provides a power
breakdown of the processor and radio for active and sleep
modes.
CPU
active (5 MHz)
sleep
Radio
transmit
receive
sleep

2.1 mA
1.6 A
10 mA
4 mA
20 A

Table 1: Power breakdown of EYES nodes.

The capabilities and power consumption of the EYES
nodes are quite similar to other prototype sensor nodes, for
example, the popular popular Berkeley Motes [14]. We like
to stress that memory (2 KB of RAM), next to energy, is a
scarce resource. Consequently, a MAC protocol should use
as little of it as possible, which limits, for example, the possibility of maintaining elaborate information on neighbors.

1.3 Idle listening problem

Most energy in traditional MAC protocols is wasted by
idle listening: since a node does not know when it will
be the receiver of a message from one of its neighbors, it
must keep its radio in receive mode at all times. Consider,
for example, a sensor application that requires nodes to
exchange messages with their neighbors at an average rate
of one per second. Messages are fairly short: they take less
than 5 milliseconds to transmit. This results in each node
spending on average 5 ms per second on transmitting, 5 ms
on receiving a message from another node, and 990 ms on
listening while nothing happens. The radio is then doing
nothing for 99% of the time.

1.4 Outline
We present T-MACTimeout-MAC, an adaptive energyefficient MAC protocol for wireless sensor networks that
minimizes idle listening, while considering wireless sensor
communication patterns and hardware limitations. In Section 2, we will describe some existing energy-saving solutions. Then, in Section 3, we will elaborate on the design
of the T-MAC protocol, problems we encountered, and the
novel way in which we solved these problems. In Section 4,
we will describe our simulation setup, followed by a detailed
report of our results. In Section 5 we report the measured
energy consumption of a limited T-MAC implementation
running on a pair of Eyes nodes.

2. RELATED WORK
There are several solutions addressing the problem of energy waste due to idle listening. In general, some kind of
duty cycle is involved, which lets each node sleep periodically. For example, TDMA-based protocols are naturally
energy preserving, because they have a duty cycle built-in,
and do not suffer from collisions [2]. However, maintaining
a TDMA schedule in an ad-hoc network is not an easy
task and requires much complexity in the nodes. Keeping a
list of neighbors schedules takes valuable memory capacity.
Allocating TDMA slots is a complex problem that requires
coordination. Furthermore, as TDMA divides time into very
small slots, the effect of clock drift can be disastrous; exact
timing is critical.
Another way of energy saving is to use an extra radio
the so-called wake-up radio, which operates on a different
frequency than the radio used for communication [7]. As the
wake-up radio is only for waking up other nodes, it needs
no data processing and therefore uses much less energy. It
does, however, require an extra component on the node and
most wireless sensor nodes currently used in research only
have a single radio that operates on a single frequency.
Introducing a duty cycle into a contention-based (CSMA)
protocol that only uses a single frequency requires some kind
of in-band signalling. The well known IEEE 802.11 protocol,
for example, has power-saving features, even when working
in ad-hoc mode [4]. However, this protocol was designed
with the presumption that all nodes are located in a single
network cell, while wireless sensor networks will often be
multi-hop. Adaptations for multi-hop networks have been
proposed, but seem to require more complexity and dynamic
state than would generally be available in wireless sensor
networks [10].
The TinyOS project [14] includes a sensor-networks specific optimization of the basic CSMA protocol that tackles the idle-listening problem: by sending out a very long
preamble, receivers only need to weak up periodically to
sense activity. This shifts the cost from the receiver (the frequent case) to the transmitter (the rarer case). The TinyOS
approach, briefly discussed in [3], is basically an optimization at the physical layer and may be applied in combination
with link-layer (MAC) solutions discussed next.
Another protocol specifically designed for sensor networks
is S-MAC [12]. The basic idea of this single-frequency contention-based protocol is that time is divided intofairly
largeframes. Every frame has two parts: an active part
and a sleeping part. During the sleeping part, a node turns
off its radio to preserve energy. During the active part, it

normal

normal

active time

active state

sleep state

S-MAC

Figure 1: The S-MAC duty cycle; the arrows

indicate transmitted and received messages; note
that messages come closer together.
can communicate with its neighbors and send any messages
queued during the sleeping part, as shown in Figure 1. Since
all messages are packed into the active part, instead of being spread out over the whole frame, the time between
messages, and therefore the energy wasted on idle listening,
is reduced.
S-MAC needs some synchronization, but that is not as
critical as in TDMA-based protocols: the time scale is much
larger. Typically, there may be an active part of 200 ms in
a frame of one second. A clock drift of 500 s will not be a
problem.
The S-MAC protocol essentially trades used energy for
throughput and latency. Throughput is reduced because
only the active part of the frame is used for communication.
Latency increases because a message-generating event may
occur during sleep time. In that case, the message will be
queued until the start of the next active part.

3.

T-MAC PROTOCOL DESIGN

Energy consumption is the main criterion for our MAC

protocol design. We have already identified the problem of
idle listening. Other forms of energy waste are:
collisions if two nodes transmit at the same time and interfere with each others transmission, packets are corrupted. Hence, the energy used during transmission
and reception is wasted;
protocol overhead most protocols require control packets
to be exchanged; as these contain no application data,
we may consider any energy used for transmitting and
receiving these packets as overhead;

TA

sleep time

TA

TA

T-MAC

Figure 2: The basic T-MAC protocol scheme, with

adaptive active times.
The problem is that, while latency requirements and buffer
space are generally fixed, the message rate will usually vary
(Section 1.1). If important messages are not to be missed
and unimportant messages should not have been sent in any
case, the nodes must be deployed with an active time that
can handle the highest expected load. Whenever the load is
lower than that, the active time is not optimally used and
energy will be wasted on idle listening.
The novel idea of the T-MAC protocol is to reduce idle
listening by transmitting all messages in bursts of variable
length, and sleeping between bursts. To maintain an optimal
active time under variable load, we dynamically determine
its length. We end the active time in an intuitive way: we
simply time out on hearing nothing.

3.1 Basic protocol

Figure 2 shows the basic scheme of the T-MAC protocol.
Every node periodically wakes up to communicate with its
neighbors, and then goes to sleep again until the next frame.
Meanwhile, new messages are queued. Nodes communicate with each other using a Request-To-Send (RTS), ClearTo-Send (CTS), Data, Acknowledgement (ACK) scheme,
which provides both collision avoidance and reliable transmission [1]. This scheme is well known and used, for example, in the IEEE 802.11 standard [4].
A node will keep listening and potentially transmitting,
as long as it is in an active period. An active period ends
when no activation event has occurred for a time TA. An
activation event is:
the firing of a periodic frame timer;
the reception of any data on the radio;

overhearing since the air is a shared medium, a node may

receive packets that are not destined for it; it could
then as well have turned off its radio.

the sensing of communication1 on the radio, e.g. during a collision;

These other sources of energy consumption are relatively

insignificant when compared to the energy wasted by idle
listening, especially when messages are infrequent. Consider
our example where 99% of the time is spent on idle listening. If, then, the actual transmission and receiving time
increases by a factor twodue to collisions and overhead,
idle listening time decreases only from 99% to 98%.

the knowledge, through overhearing prior RTS and

CTS packets, that a data exchange of a neighbor has
ended.

Although reducing the idle listening time, a solution with a

fixed duty cycle, like the S-MAC protocol [12], is not optimal.
S-MAC has two important parameters: the total frame time,
which is limited by latency requirements and buffer space,
and the active time. The active time depends mainly on
the message rate: it can be so small that nodes are able to
transfer all their messages within the active time.

the end-of-transmission of a nodes own data packet or

acknowledgement;

A node will sleep if it is not in an active period. Consequently, TA determines the minimal amount of idle listening
per frame.
The described timeout scheme moves all communication
to a burst at the beginning of the frame. Since messages
between active times must be buffered, the buffer capacity
determines an upper bound on the maximum frame time.
1
Through a Received Signal Strength Indication (RSSI)
signal from the radio.

3.2 Clustering and synchronization

Frame synchronization is inspired by virtual clustering, as
described by the authors of the S-MAC protocol [12]. When
a node comes to life, it starts by waiting and listening. If
it hears nothing for a certain amount of time, it chooses
a frame schedule and transmits a SYNC packet, which contains the time until the next frame starts. If the node, during
startup, hears a SYNC packet from another node, it follows
the schedule in that SYNC packet and transmits its own
SYNC accordingly.
Nodes retransmit their SYNC once in a while. Nodes
must also listen for a complete frame sporadically, so they
can detect the existence of different schedules. This allows
new and mobile nodes to adapt to an existing group.
If a node has a schedule and hears a SYNC with a different
schedule from another node, it must adopt both schedules.
It must also transmit a SYNC with its own schedule to the
other node, to let the other node know about the presence
of another schedule. Adopting both schedules means that
the node will have an activation event at the start of both
frames.
Nodes must start a data transmission only at the start
of their own active time. At that time, both neighbors
with the same schedule, and neighbors that have adopted
the schedule as extra, are awake. If a node would start
transmission at the start of a neighbors frame, it might be
transmitting to another, sleeping neighbor. Note that this
scheme makes it possible that broadcasts only need to be
transmitted once.
The described synchronization scheme, which is called
virtual clustering [12], urges nodes to form clusters with the
same schedule, without enforcing this schedule to all nodes
in the network. It allows efficient broadcast, and obviates
the need to maintain information on individual neighbors.
The virtual clustering technique is easy to implement.
Keeping multiple schedules with a fixed-length active time
is more complex, since active times overlap.

contend

RTS

CTS

DATA

ACK

B
C

contend

TA

Figure 3: A basic data exchange. Node C overhears

the CTS from node B and will not disturb the
communication between A and B. TA must be long
enough for C to hear the start of the CTS.

RTS retries
When a node sends an RTS, but does not receive a CTS
back, one of three things has happened:
1. the receiving node has not heard the RTS due to collision; or
2. the receiving node is prohibited from replying due to
an overheard RTS or CTS; or
3. the receiving node is asleep.
When the sending node receives no answer within the interval TA, it might go to sleep. However, that would be wrong
in cases 1 and 2: we would then have a situation where the
sending node goes to sleep, while the receiving node is still
awake. Since this situation might occur even at the first
message of the frame, the throughput would (and did, in
our preliminary experiments) dramatically decrease.
Therefore, a node should retry by re-sending the RTS if it
receives no answer. If there is still no reply after two retries,
it should give up and go to sleep.

Determining TA

3.3 RTS operation and choosing TA

We will now discuss the additional features of the T-MAC
protocol that provide optimal tuning.

Fixed contention interval

In contention-based protocols, like IEEE 802.11, nodes wait
for a random time within a contention interval after detecting a collision. Only when the air is clear during that time do
they restart transmission. Usually, a back-off scheme is used:
the contention interval increases when traffic is higher. The
back-off scheme reduces the probability of collisions when
the load is high, while minimizing latency when the load is
low.
In the T-MAC protocol, every node transmits its queued
messages in a burst at the start of the frame. During this
burst, the medium is saturated: messages are transmitted
at maximum rate. A node may expect to be in a fierce
fight for winning the medium every time it sends an RTS.
An increasing contention interval is not useful, since the
load is mostly high and does not change. Therefore, RTS
transmission in T-MAC starts by waiting and listening for a
random time within a fixed contention interval. This interval
is tuned for maximum load. The contention time is always
used, even if no collision has occurred yet.

A node should not go to sleep while its neighbors are still

communicating, since it may be the receiver of a subsequent
message. Receiving the start of the RTS or CTS packet from
a neighbor is enough to trigger a renewed interval TA.
Since a node may not hear, because it is not in range,
the RTS that starts a communication with its neighbor, the
interval TA must be long enough to receive at least the start
of the CTS packet (Figure 3). This observation gives us a
lower limit on the length of the interval TA:
TA > C + R + T
where C is the length of the contention interval, R is the
length of an RTS packet, and T is the turn-around time
(the short time between the end of the RTS packet and the
beginning of the CTS packet). In our experiments, we used
TA = 1.5 (C + R + T ), which proved to be satisfactory. A
larger TA increases the energy used.

3.4 Overhearing avoidance

The S-MAC protocol introduced the idea of sleeping after
overhearing an RTS or CTS destined for another node. Since
a node is prohibited from sending during that time, it can
not take part in any communication and may as well turn
off its radio to save energy.

contend

RTS

CTS

DATA

ACK

B
C
D

contend

RTS

CTS DS

DATA

ACK

B
contend

active

C
sleep
RTS?

TA

contend

active

active
FRTS

RTS

TA

Figure 4: The early sleeping problem. Node D goes

to sleep before C can send an RTS to it.

Figure 5: The future-request-to-send packet exchange keeps Node D awake.

In general, overhearing avoidance is a good idea, and it

is an option in the T-MAC protocol. However, we have
observed in our experiments that, as a side effect, collision overhead becomes higher: a node may miss other RTS
and CTS packets while sleeping and disturb some communication when it wakes up. Consequently, the maximum
throughput decreases; for short packets by as much as 25%.
Thus, although overhearing avoidance saves energy, it must
not be used when maximum throughput is (at times) required.

active part of the network. We believe that the problem

may occur in any asymmetric communication pattern. We
devised two solutions.

3.5 Asymmetric communication

Preliminary simulation experiments revealed a problem
with the T-MAC protocol when traffic through the network
is mostly unidirectional, like in a nodes-to-sink communication pattern. This problem is simplified in Figure 4. Each
of the nodes A though D in the picture forms a cell with
its neighbors. Messages flow from top to bottom, so node A
sends only to B, B only to C, and C only to D. Now consider
node C. Every time it wants to send a message to D, it must
contend for the medium and may loose to either node B
(by receiving an RTS packet) or to node A (indirectly, by
overhearing a CTS packet from node B).
If node C looses contention because of an RTS packet from
node B, it will reply with a CTS packet, which can also be
heard by node D. In that case, node D will be awake when
the communication between C and B ends. However, if node
C looses contention because it overhears a CTS packet from
B to A (see Figure 4), C must remain silent. Since D does
not know of the communication between A and B, its active
time will end, and node D will go to sleep. Only at the start
of the next frame will node C have a new chance to send to
node D.
Thus for every packet that node C wants to send to node
D, it may either succeed or fail (by loosing to node A). Both
of these events have equal probability. Failure implies that
the frame ends and C can send no more packets. We can
therefore calculate that, in this simplified setup, node C has
a 50% probability of sending a single packet to node D,
a 25% probability of sending two packets (it must succeed
twice), etcetera, in each frame.
We call the observed effect the early sleeping problem,
since a node goes to sleep when a neighbor still has messages
for it. In the nodes-to-sink communication pattern, the early
sleeping problem reduced the total possible throughput of
T-MAC to less than half of the maximum throughput of
traditional protocols or S-MAC. In later experiments, we
have also encountered this problem at the border of a highly

Future request-to-send
The first solution is a scheme that we call future request-tosend. The idea is to let another node know that we still
have a message for it, but are ourselves prohibited from
using the medium. It works as follows: if a node overhears a
CTS packet destined for another node, it may immediately
send a future-request-to-send (FRTS) packet, like node C
in Figure 5. The FRTS packet contains the length of the
blocking data communication (this information was in the
CTS packet). A node must not send an FRTS packet if
it senses communication right after the CTS, or if it is
prohibited from sending due to a prior RTS or CTS.
A node that receives an FRTS packet knows it will be
the future target of an RTS packet and must be awake by
that time. The node can determine this from the timing
information in the FRTS packet.
As the FRTS packet would otherwise disturb the data
packet that follows the CTS, the data packet must be postponed for the duration of the FRTS packet. To prevent
any other node from taking the channel during this time,
the node that sent the initial RTS (node A in Figure 5)
transmits a small Data-Send (DS) packet. After the DS
packet, it must immediately send the normal data packet.
Since the FRTS packet has the same size as a DS packet,
it will collide with the DS packet, but not with the following
data packet. The DS packet is lost, but that is no problem:
it contains no useful information.
For the FRTS solution to work, TA must be increased
with the length of a control packet (CTS), as follows from
Figure 5. Implementing the FRTS feature increased the
maximum throughput in unidirectional communication patterns by approximately 75%. However, due to the somewhat
higher overhead of DS and FRTS packets, energy consumption also increases slightly. One may want to use FRTS
packets only if a reasonably high load in a more or less
unidirectional communication pattern is expected. Note,
however, that when the load is low, the number of exchanged
packets, and therefore the increased overhead, is also low.

Taking priority on full buffers

The second solution is a scheme that we call full-buffer priority. When a nodes transmit/routing buffers are almost full,
it may prefer sending to receiving. This means that when a

contend

contend

B
C

contend
RTS

RTS CTS

DATA

ACK

TA

Figure 6: Taking priority upon receiving RTS.

node receives an RTS packet destined for it, it immediately
sends its own RTS packet to another node, instead of replying with a CTS like normal. This is depicted in Figure 6. It
has two effects. First, the node has an even higher chance
of transmitting its own message, since it effectively wins the
medium upon hearing a competing RTS; in Figure 6, node
C may transmit to node D after losing contention to node B.
Thus the probability that the early sleeping problem occurs
is lower. Secondly, the full-buffer priority scheme introduces
a limited form of flow control into the network, which is
advantageous in a nodes-to-sink communication pattern; in
Figure 6, node B is prevented from sending until node C has
enough buffer space.
We must, however, be careful with the full-buffer priority
scheme, since it is dangerous in a high-load situation where
communication is not unidirectional. When all nodes in an
omnidirectional communication pattern start taking priority, chances of collisions increase rapidly. Therefore, T-MAC
uses a threshold: a node may only use this scheme when it
has lost contention at least twice. This threshold guarded,
in our experiments, the performance in an omnidirectional
communication pattern, while still increasing the maximum
throughput in a unidirectional pattern.

4.

EXPERIMENTS

In our experiments, we compared three protocols: CSMA,

S-MAC, and T-MAC. We include CSMA because we consider it as a worst case: it has no energy saving features
at all. When a node is not transmitting, its radio is set to
receive. We include S-MAC because it was, like T-MAC,
designed for wireless sensor networks and uses in-band signalling as well.

4.1 Simulation setup and parameters

Our protocol design and evaluation is, for now, based on
simulation. In the OMNeT++ discrete event simulation
package [11], we have built a realistic model of the EYES
wireless sensor nodes [13]. The model has the same limits
on clock resolution and precision, radio turn-around and
wake-up times, and transmission bit rates as the EYES
nodes do. Energy consumption in the model is based on
the amount of energy the real nodes use: 20 A while sleeping, 4 mA while receiving and 10 mA while transmitting a
DC-balanced signal [5, 6, 9].
Using these modeled nodes, we have built a network of
100 nodes in a 10 by 10 grid. We have chosen a radio
range so that non-edge nodes all have 8 neighbors. We
admit that this perfect world is not a realistic setup for most

sensor networks, but it serves our simulation needs well, for

example, we do not need to know the exact location of each
node to generate a workload.
For the nodes-to-sink communication pattern, we used a
randomized shortest path routing method: for each message, the possible next hops are enumerated. Next hops are
eligible if they have a shorter path to the final destination
than the sending node. From these next hops, a random one
is chosen. Thus messages flow in the correct direction, but
do not use the same path every time. No control messages
are exchanged for this routing scheme: nodes automagically
determine the next hop. Multi-hop messages travel at least
two hops per frame because of T-MACs in-band signaling:
the CTS and ACK messages keep direct neighbors awake.
In our experiments, we tested the S-MAC protocol with
a frame length of one second, and with several lengths of
the active time, varying from 75 ms to 915 ms. For the
T-MAC protocol, we always used a frame length of 610 ms
(20000 ticks of a quartz crystal) and an interval TA with
a length of 15 ms. The T-MAC protocol can optionally
be deployed with overhearing avoidance, full-buffer priority,
and FRTS. We show the best combination of these options
in each experiment.

4.2 Results
We start with a simple benchmark, where network load
is homogeneous. We then introduce experiments based on
communication patterns and end with a complete, realistic
scenario.
Except for the complete scenario, we exercised all experiments with multiple network loads. The load is on the
horizontal axes of the graphs. The vertical axes show the
energy consumption in the network, since this is our main
design criterion.
Since throughput is not completely unimportant, we will
end each graph at the point where less than 90% of the
messages are correctly received. In a multi-hop pattern, like
nodes-to-sink, this means that we want 90% of all messages
to finally reach the sink node. We do not use the 90%restriction for CSMA: since CSMA has no built-in retransmits, it is far less reliable.
The message lengths are data payload sizes, and do not
include the MAC header: 4 bytes for CSMA, 6 bytes for
S-MAC and T-MAC.

Homogeneous local unicast

(Figure 7) In out first experiment, nodes send packets with
some payload (20 or 100 bytes) to one of their neighbors
at random. Although this is not a realistic communication
pattern, it serves as a base case. For the T-MAC protocol,
we used overhearing avoidance, but no FRTS or full-buffer
priority mechanisms.
In Figure 7(left), we can clearly see that CSMA provides
no energy savings. The energy consumption when idle is
4 mA (the current drawn by the radio in receive mode),
slightly going up when more messages are sent.
For the S-MAC protocol, various lines are shown for different lengths of the active time. An additional line is drawn,
which connects the S-MAC graphs that use the least energy
for each load. So on this line, the S-MAC protocol is tuned
to provide at least 90% throughput while using as little
energy as possible for each individual load. From now, we
will only show similarly tuned lines for the S-MAC protocol.

Homogeneous unicast, msglength=100

4.5

3.5

3.5
energy used [avg. mA/node]

energy used [avg. mA/node]

Homogeneous unicast, msglength=20

4.5

3
2.5
2
1.5
1

3
2.5
2
1.5
1

0.5

0.5

CSMA
S-MAC
T-MAC

CSMA
S-MAC
T-MAC

0
0

20

40

60

80

100

20

40

load [byte / node / s]

60

80

100

load [byte / node / s]

Figure 7: Homogeneous local unicast for small (left) and large messages (right). The numbers on the S-MAC
line indicate the length of the active time per 1-second frame.
Nodes-to-sink, msglength=100
4.5

3.5

3.5
energy used [avg. mA/node]

energy used [avg. mA/node]

Nodes-to-sink, msglength=20
4.5

3
2.5
2
1.5
1

3
2.5
2
1.5
1

0.5

0.5

CSMA
S-MAC
T-MAC

CSMA
S-MAC
T-MAC

0
0

6
8
load [byte / node / s]

10

12

6
8
load [byte / node / s]

10

12

Figure 8: Nodes-to-sink performance for small (left) and large messages (right).
We expect that the homogeneous experiment is the best
case for the S-MAC protocol, since the load is constant in
both time and location. We see that the T-MAC protocol
performs as well as the (per load tuned) S-MAC protocol.
The fact that the T-MAC protocol uses even less energy
than S-MAC is due to the fact that we tested the S-MAC
protocol only with a limited number of discrete lengths of
the active time. To get an optimal parameter value for each
load, the S-MAC protocol would require complicated tuning,
as it would in real deployment. The adaptive behavior of
T-MAC, on the other hand, requires no explicit tuning.
Figure 7(right) shows that that the maximum throughput of the T-MAC protocol with large messages (100 byte
payload) is less than that of the S-MAC protocol. This is
mainly due to variations on the early sleeping problem as
described in Section 3.5.

Nodes-to-sink communication
(Figure 8) In this experiment, nodes send messages to a
single sink node at the corner of the network. Messages
are routed from node to node with a (slightly randomized)
shortest path algorithm. No data aggregation is used. For
the T-MAC protocol, we used overhearing avoidance, the
full-buffer priority mechanism, and the FRTS mechanism.

Figure 8 shows that T-MAC uses less energy than S-MAC.

We expected this, because in the nodes-to-sink communication the load varies with the location of nodes: there is
more traffic in the neighborhood of the sink node. This also
explains why the absolute load is much lower (factor of 10)
for nodes-to-sink than for local unicast: the rate at which the
sink can handle incoming messages limits the load individual
nodes can generate without congesting the network (around
the sink).
As with homogeneous local unicast we see that the maximum throughput of T-MAC is less than that of S-MAC.
Our experience with other message sizes and communication
patterns is that the maximum throughput of T-MAC is
at worst about 70% of S-MAC. We do not worry about
too much about T-MACs reduced maximum throughput,
because it only occurs under extreme loads that are best
avoided by sensor applications, for example, by aggregating
messages.

Early sleeping problem

(Figure 9) In this experiment, we show the effectiveness of
the measures addressing the early sleeping problem (Section 3.5). We see that the FRTS mechanism increases maximum throughput by approximately 75% (0.08 vs. 0.14 mes-

T-MAC nodes-to-sink, msglength=20

Event-based unicast, msglength=20

1.2

4.5
4

0.8

energy used [avg. mA / node]

energy used [avg. mA / node]

FRTS + prio
1

FRTS

0.6

prio

0.4

no measures

0.2

3.5
3
2.5
2
1.5
1
CSMA
S-MAC
T-MAC

0.5
0

0
0

20

load [msg / node / s]

sages per second), at the cost of some energy. Adding the

full-buffer priority mechanism adds approximately 30% (0.14
vs. 0.18 messages per second) without the cost of extra
energy.

Event-based local unicast

(Figure 10) We now proceed towards a more realistic scenario. In this experiment, events occur in the network with
a frequency of one per 10 seconds. Events have an average
duration of 5 seconds and affect an area of approximately 9
nodes. These nodes then send local unicast messages to their
neighbors for the duration of the event. A neighbor that
receives one of these messages replies with a probability of
20%. We performed multiple measurements, with different
message frequencies during events. This frequency is on the
horizontal axis of the graph (Figure 10). For T-MAC, we
used overhearing avoidance but no FRTS and no full-buffer
priority.
Figure 10 shows that T-MAC uses much less energy than
either S-MAC or CSMA, especially when the message frequency during events increases. However, the maximum
frequency that T-MAC can handle is lower than that of
S-MAC, like we have seen in the nodes-to-sink communication pattern. Again, T-MAC suffers from the early sleeping
problem, because we have relatively many edge nodes.

Event-based local unicast and node-to-sink reporting

(Figure 11) This is an experiment with a complete scenario.
When no events happen, nodes exchange local messages of
10 bytes with each other every 20 seconds. They also report
to a sink node every 100 seconds. When an event happens
(once every 10 seconds, like in the previous experiment),
nodes in the neighborhood of the event start sending local
unicast messages of 30 bytes, with a rate of 4 per second.
They then also send messages of 50 bytes to the sink, once
per second. These messages are aggregated in the network.
To be able to handle this kind of traffic, we had to tune
S-MAC to listen for at least 715 ms in every second. For
the T-MAC protocol, we used overhearing avoidance, but
no FRTS and no full-buffer priority.
Figure 11 clearly shows the disadvantage of a fixed-length
active part: to be able to handle a short-lived burst of

60

80

100

120

140

160

peak load [bytes / node / s]

Figure 10: An event-based scenario, where active

nodes exchange local unicast messages.
Event triggered reporting
4
3.5

energy used [avg. mA / node]

Figure 9: T-MAC options in a nodes-to-sink communication pattern.

40

3
2.5
2
1.5
1
0.5
0
CSMA

S-MAC

T-MAC

Figure 11: A complete scenario with both periodic

reporting and event-based messages.

trafficeven though that happens infrequentlythe deployer

of S-MAC must choose a long active part of the duty cycle. S-MAC therefore wastes much energy at times when
no events happen. By adaptively changing the duty cycle,
T-MAC can decrease the used energy in this scenario by a
factor of 5.

5. REAL IMPLEMENTATION
After optimizing the T-MAC protocol using simulations,
we implemented most of the protocol on the actual EYES
hardware [13]. We did not implement all features. To test
the effectiveness of the full-buffer-priority and future-RTS
schemes, a large-scale experiment is needed, involving a lot
of nodes. Since there was no time to do such an experiment,
implementation of these features was senseless.
We have also not implemented the possibility to keep
multiple schedules yet. Although this is fairly easy to implement, we have only tested T-MAC with single-cluster
configurations.
During testing, we noted that the nodes schedules would
drift apart relatively fast. Even though the time-keeping on

20

25

18

RTS

16

DATA

RTS

DATA

RTS

DATA

20

12

current [mA]

current [mA]

14

10
8

15

10

6
4

TA

CONTEND
CTS

2
0

CTS
ACK
CONTEND

CTS
ACK
CONTEND

ACK

0
0

10

6.18

6.2

time [s]

6.22

6.24

6.26

6.28

time [s]

Figure 12: Trace of the electrical current, receiving

node, 1 message / second.

Figure 14: Trace of the electrical current, node

successfully transmitting 3 messages.

20
18
16

current [mA]

14
12
10
8
6
4
2
0
0

10

time [s]

Figure 13: Trace of the electrical current, transmitting node, full speed.
all nodes is based on ticks of a quartz crystal, some of the
nodes became unreachable within as little as 10 minutes.
We had not simulated this effect.
To solve the drift problem, we used a simple correction
scheme: when a node receives a SYNC message that contains almost, but not exactly, its own schedule, the node
adjusts its own schedule towards the received schedule. To
allow a converging situation, the schedule is only adjusted
for 50% of the difference between the two schedules.
The drift correction solved the problem: an experiment
showed that nodes were still perfectly synchronized after
more than 10 hours.
The final implementation performs well. When sending a
continuous stream of data messages, the nodes rarely go
to sleep. But when no messages are sent, the indication
LEDs on the nodes blink peacefully. The radio is then in
the receive mode for 15 ms out of every 610 ms, which is
less than 2.5% of the time.

5.1 Energy consumption

After the implementation of the T-MAC protocol, we performed a number of energy consumption experiments. In

these experiments, one node was sending, another receiving.

The message length was 20 bytes. We measured the electrical current through both the sending and the receiving node
using using specialized equipment, which sent the values
to a logging computer. After precise calibration, we could
measure the current through the nodes with a precision of
approximately 15 A. The sample rate was 500 Hz (limited
by the PC software).
In Figure 12 we see a power trace of a node receiving
approximately one message per second. There is a transmission in most frames, but the radio is still in sleep mode
most of the time. The high spikes are caused by transmitting
CTSs. In Figure 13 we see a node that transmits messages
at maximum rate. It is clear that this node never sleeps, as
it should not. Instead, it toggles between receive (3.75 mA)
and transmit (18 mA). Figure 14 shows a closeup of a node
transmitting 3 messages during a single frame.
msg / s

transmit mA

receive mA

0
1
10

0.138
0.400
1.516

0.138
0.246
0.890

max

9.590

7.473

Table 2: Average energy consumption of sending

and receiving EYES nodes (T-MAC protocol).
Table 2 shows the average electrical current during each
experiment. We can see that transmitting nodes use significantly more energy than receiving nodes. This is logical, since transmitting with our radio takes more energy
than receiving. More importantly, we see that the idle average current (0.138 mA) is less than 4% of the current
of a non-energy saving protocol (which would be between
3.75 and 4 mA). This is well above the theoretical limit
(TA/Tf rame = 15ms/610ms = 2.5%). However, the theoretical limit only takes the radio energy consumption into
account, not the CPU, EEPROM, RS232 voltage converter
and other components. We think that 96% overall energy
savings is a very good result.

6.

CONCLUSIONS AND FUTURE WORK

To solve the problem of idle listening in a wireless sensor

network, we have proposed the T-MAC protocol. T-MAC
dynamically adapts a listen/sleep duty cycle in a novel way,
through fine-grained timeouts, while having minimum complexity. Simulations have shown that the T-MAC protocol introduces a way of decreasing energy consumption in
a volatile environment where the message rate fluctuates,
either in time or in location. Implementation of the T-MAC
protocol has shown that, during a high load, nodes communicate without sleeping, but during a very low load, nodes
will use their radios for as little as 2.5% of the time, saving
as much as 96% of the energy compared to a traditional
protocol.
We have also identified a problem with the T-MAC protocol (the early sleeping problem) and proposed original solutions for this problem (FRTS and full-buffer priority). The
trade-off for very low energy consumption is a decreased
maximum throughput. We have encountered the early sleeping problem only in a perfect, simulated world. Future work
is needed to determine the actual seriousness of the problem
and to test the effectiveness of the proposed solutions. Other
solutions may well exist.
We have only experimented with a static, non-mobile network. We expect problems when applying virtual clustering
to groups of mobile nodes. The virtual clustering technique, proposed in the S-MAC protocol and re-used in the
T-MAC protocol, has not been researched thoroughly. Since
synchronization of schedules can have a great impact on
the energy consumption, clustering is important. Virtual
clustering, and multi-hop synchronization in general, is an
interesting research subject in itself.
T-MACs reductions in energy consumption are very promising. This novel protocol is the subject of an ongoing study,
and we expect more detailed results in the future.

7.

ACKNOWLEDGEMENTS

We like to thank Niels Reijers, the anonymous reviewers,

and our shepherd David Culler for providing constructive
comments that improved the quality of the paper.

8.

REFERENCES

[1] V. Bharghavan, A. Demers, S. Shenker, and L. Zhang.

MACAW: a media access protocol for wireless LANs.
In Conf. on Communications Architectures, Protocols
and Applications, pages 212225, London, 1994.
[2] P. Havinga and G. Smit. Energy-efficient TDMA
medium access control protocol scheduling. In Asian
International Mobile Computing Conference (AMOC
2000), pages 19, November 2000.
[3] J. Hill and D. Culler. Mica: a wireless platform for
deeply embedded networks. IEEE Micro, 22(6):1224,
Nov-Dec 2002.
[4] LAN MAN Standards Committee of the IEEE
Computer Society. IEEE Std 802.11-1999, Wireless
LAN Medium Access Control (MAC) and Physical
Layer (PHY) specifications. IEEE, 1999.
[5] RFM. TR1001 868.35 MHz Hybrid Tranceiver.
[6] RFM. ASH Transceiver Designers Guide, October
2002.

[7] S. Singh and C. Raghavendra. PAMAS: Power aware

multi-access protocol with signalling for ad hoc
networks. ACM SIGCOMM Computer
Communication Review, 28(3):526, July 1998.
[8] M. Stemm and R. H. Katz. Measuring and reducing
energy consumption of network interfaces in hand-held
devices. IEICE Transactions on Communications,
E80-B(8):11251131, 1997.
[9] Texas Instruments. MSP430x1xx Family Users Guide.
SLAU049B.
[10] Y. Tseng, C. Hsu, and T. Hsieh. Power-saving
protocols for IEEE 802.11-based multi-hop ad hoc
networks. In 21st Conf. of the IEEE Computer and
Communications Societies (INFOCOM), volume 1,
pages 200209, June 2002.
[11] A. Varga. The OMNeT++ discrete event simulation
system. In European Simulation Multiconference
(ESM2001), Prague, Czech Republic, June 2001.
[12] W. Ye, J. Heidemann, and D. Estrin. An
energy-efficient MAC protocol for wireless sensor
networks. In 21st Conference of the IEEE Computer
and Communications Societies (INFOCOM),
volume 3, pages 15671576, June 2002.
[13] http://eyes.eu.org/sensnet.htm.
[14] http://webs.cs.berkeley.edu/tos/.