An Energy-Efficient MAC Protocol for Wireless

Sensor Networks
Wei Ye, John Heidemann, Deborah Estrin

Abstract—This paper proposes S-MAC, a medium-access control (MAC) than recharged. Prolonging network lifetime for these nodes is a
protocol designed for wireless sensor networks. Wireless sensor networks critical issue. Another important attribute is the scalability to the
use battery-operated computing and sensing devices. A network of these
devices will collaborate for a common application such as environmental change in network size, node density and topology. Some nodes
monitoring. We expect sensor networks to be deployed in an ad hoc fashion, may die over time; some new nodes may join later; some nodes
with individual nodes remaining largely inactive for long periods of time, may move to different locations. The network topology changes
but then becoming suddenly active when something is detected. These char-
acteristics of sensor networks and applications motivate a MAC that is dif-
over time as well due to many reasons. A good MAC proto-
ferent from traditional wireless MACs such as IEEE 802.11 in almost every col should easily accommodate such network changes. Other
way: energy conservation and self-configuration are primary goals, while important attributes include fairness, latency, throughput and
per-node fairness and latency are less important. S-MAC uses three novel bandwidth utilization. These attributes are generally the primary
techniques to reduce energy consumption and support self-configuration.
To reduce energy consumption in listening to an idle channel, nodes period- concerns in traditional wireless voice and data networks, but in
ically sleep. Neighboring nodes form virtual clusters to auto-synchronize sensor networks they are secondary.
on sleep schedules. Inspired by PAMAS, S-MAC also sets the radio to
sleep during transmissions of other nodes. Unlike PAMAS, it only uses
This paper presents sensor-MAC (S-MAC), a new MAC pro-
in-channel signaling. Finally, S-MAC applies message passing to reduce tocol explicitly designed for wireless sensor networks. While
contention latency for sensor-network applications that require store-and- reducing energy consumption is the primary goal in our design,
forward processing as data move through the network. We evaluate our our protocol also has good scalability and collision avoidance
implementation of S-MAC over a sample sensor node, the Mote, developed
at University of California, Berkeley. The experiment results show that, on capability. It achieves good scalability and collision avoidance
a source node, an 802.11-like MAC consumes 2–6 times more energy than by utilizing a combined scheduling and contention scheme. To
S-MAC for traffic load with messages sent every 1–10s. achieve the primary goal of energy efficiency, we need to iden-
tify what are the main sources that cause inefficient use of en-
I. I NTRODUCTION ergy as well as what trade-offs we can make to reduce energy
consumption.
IRELESS sensor networking is an emerging technology
W that has a wide range of potential applications includ-
ing environment monitoring, smart spaces, medical systems and
We have identified the following major sources of energy
waste. The first one is collision. When a transmitted packet
robotic exploration. Such a network normally consists of a large is corrupted it has to be discarded, and the follow-on re-
number of distributed nodes that organize themselves into a transmissions increase energy consumption. Collision increases
multi-hop wireless network. Each node has one or more sen- latency as well. The second source is overhearing, meaning
sors, embedded processors and low-power radios, and is nor- that a node picks up packets that are destined to other nodes.
mally battery operated. Typically, these nodes coordinate to per- The third source is control packet overhead. Sending and re-
form a common task. ceiving control packets consumes energy too, and less useful
Like in all shared-medium networks, medium access control data packets can be transmitted. The last major source of inef-
(MAC) is an important technique that enables the successful op- ficiency is idle listening, i.e., listening to receive possible traf-
eration of the network. One fundamental task of the MAC pro- fic that is not sent. This is especially true in many sensor net-
tocol is to avoid collisions so that two interfering nodes do not work applications. If nothing is sensed, nodes are in idle mode
transmit at the same time. There are many MAC protocols that for most of the time. However, in many MAC protocols such
have been developed for wireless voice and data communication as IEEE 802.11 or CDMA nodes must listen to the channel
networks. Typical examples include the time division multiple to receive possible traffic. Many measurements have shown
access (TDMA), code division multiple access (CDMA), and that idle listening consumes 50–100% of the energy required
contention-based protocols like IEEE 802.11 [1]. for receiving. For example, Stemm and Katz measure that the
idle:receive:send ratios are 1:1.05:1.4 [2], while the Digitan 2
To design a good MAC protocol for the wireless sensor net-
Mbps Wireless LAN module (IEEE 802.11/2Mbps) specifica-
works, we have considered the following attributes. The first is
tion shows idle:receive:send ratios is 1:2:2.5 [3].
the energy efficiency. As stated above, sensor nodes are likely
to be battery powered, and it is often very difficult to change or S-MAC tries to reduce the waste of energy from all the above
recharge batteries for these nodes. In fact, someday we expect sources. In exchange we accept some reduction in both per-hop
some nodes to be cheap enough that they are discarded rather fairness and latency. Although per-hop fairness and latency are
reduced, we will argue that the reduction does not necessarily
W. Ye (weiye@isi.edu) and J. Heidemann (johnh@isi.edu) are with the In- result in lower end-to-end fairness and latency.
formation Science Institute (ISI), University of Southern California (USC). D.
Estrin (destrin@cs.ucla.edu) is with the Computer Science Department, Univer- In traditional wireless voice or data networks, each user de-
sity of California, Los Angeles and USC/ISI. sires equal opportunity and time to access the medium, i.e.,
sending or receiving packets for their own applications. Per- • The use of in-channel signaling to put each node to sleep
hop MAC level fairness is thus an important issue. However, in when its neighbor is transmitting to another node. This method
sensor networks, all nodes cooperate for a single common task. avoids the overhearing problem and is inspired by PAMAS [10],
Normally there is only one application. At certain time, a node but does not require an additional channel.
may have dramatically more data to send than some other nodes. • Applying message passing to reduce application-perceived la-
In this case fairness is not important as long as application-level tency and control overhead. Per-node fragment-level fairness is
performance is not degraded. In our protocol, we re-introduce reduced since sensor network nodes are often collaborating to-
the concept of message passing to efficiently transmit a very wards a single application.
long message. The basic idea is to divide the long message into • Evaluating an implementation of our new MAC over sensor-
small fragments and transmit them in a burst. The result is that net specific hardware.
a node who has more data to send gets more time to access the
medium. This is unfair from a per-hop, MAC level perspec- II. R ELATED W ORK
tive, for those nodes who only have some short packets to send, The medium access control is a broad research area, and many
since their short packets have to wait a long time for very long researchers have done research work in the new area of low
packets. However, as we will show later, message passing can power and wireless sensor networks [11], [12], [13], [14].
achieve energy savings by reducing control overhead and avoid- Current MAC design for wireless sensor networks can be
ing overhearing. broadly divided into contention-based and TDMA protocols.
Latency can be important or unimportant depending on what The standardized IEEE 802.11 distributed coordination function
application is running and the node state. During a period that (DCF) [1] is an example of the contention-based protocol, and
there is no sensing event, there is normally very little data flow- is mainly built on the research protocol MACAW [15]. It is
ing in the network. Most of the time nodes are in idle state. widely used in ad hoc wireless networks because of its simplic-
Sub-second latency is not important, and we can trade it off for ity and robustness to the hidden terminal problem. However, re-
energy savings. S-MAC therefore lets nodes periodically sleep cent work [2] has shown that the energy consumption using this
if otherwise they are in the idle listening mode. In the sleep MAC is very high when nodes are in idle mode. This is mainly
mode, a node will turn off its radio. The design reduces the en- due to the idle listening. PAMAS [10] made an improvement
ergy consumption due to idle listening. However, the latency is by trying to avoid the overhearings among neighboring nodes.
increased, since a sender must wait for the receiver to wake up Our paper also exploits similar method for energy savings. The
before it can send out data. main difference of our work with PAMAS is that we do not use
An important feature of wireless sensor networks is the in- any out-of-channel signaling. Whereas in PAMAS, it requires
network data processing. It can greatly reduce energy con- two independent radio channels, which in most cases indicates
sumption compared to transmitting all the raw data to the end two independent radio systems on each node. PAMAS does not
node [4], [5], [6]. In-network processing requires store-and- address the issue of reduce idle listening.
forward processing of messages. A message is a meaningful The other class of MAC protocols are based on reservation
unit of data that a node can process (average or filter, etc.). It and scheduling, for example TDMA-based protocols. TDMA
may be long and consists of many small fragments. In this case, protocols have a natural advantage of energy conservation com-
MAC protocols that promote fragment-level fairness actually in- pared to contention protocols, because the duty cycle of the ra-
crease message-level latency for the application. In contrast, dio is reduced and there is no contention-introduced overhead
message passing reduces message-level latency by trading off and collisions. However, using TDMA protocol usually re-
the fragment-level fairness. quires the nodes to form real communication clusters, like Blue-
To demonstrate the effectiveness and measure the perfor- tooth [16], [17] and LEACH [13]. Managing inter-cluster com-
mance of our MAC protocol, we have implemented it on our munication and interference is not an easy task. Moreover, when
testbed wireless sensor nodes, Motes, developed by Univer- the number of nodes within a cluster changes, it is not easy for
sity of California, Berkeley [7]. The mote has a 8-bit Atmel a TDMA protocol to dynamically change its frame length and
AT90LS8535 microcontroller running at 4 MHz. It has a low time slot assignment. So its scalability is normally not as good
power radio transceiver module TR1000 from RF Monolithics, as that of a contention-based protocol. For example, Bluetooth
Inc [8], which operates at 916.5 MHz frequency and provides a may have at most 8 active nodes in a cluster.
transmission rate of 19.2 Kbps. The mote runs on a very small Sohrabi and Pottie [12] proposed a self-organization protocol
event-driven operating system called TinyOS [9]. In order to for wireless sensor networks. Each node maintains a TDMA-
compare the performance of our protocol with some other pro- like frame, called super frame, in which the node schedules dif-
tocols, we also implemented a simplified IEEE 802.11 MAC on ferent time slots to communicate with its known neighbors. At
this platform. each time slot, it only talks to one neighbor. To avoid inter-
The contributions of this work are therefore: ference between adjacent links, the protocol assigns different
• The scheme of periodic listen and sleep reduces energy con- channels, i.e., frequency (FDMA) or spreading code (CDMA),
sumption by avoiding idle listening. The use of synchroniza- to potentially interfering links. Although the super frame struc-
tion to form virtual clusters of nodes on the same sleep sched- ture is similar to a TDMA frame, it does not prevent two inter-
ule. These schedules coordinate nodes to minimize additional fering nodes from accessing the medium at the same time. The
latency. actual multiple access is accomplished by FDMA or CDMA. A
drawback of the scheme is its low bandwidth utilization. For Listen Sleep Listen Sleep
example, if a node only has packets to be sent to one neighbor,
it cannot reuse the time slots scheduled to other neighbors. time
Piconet [11] is an architecture designed for low-power ad hoc Fig. 1. Periodic listen and sleep.
wireless networks. One interesting feature of piconet is that it
also puts nodes into periodic sleep for energy conservation. The
scheme that piconet uses to synchronize neighboring nodes is
traffic, while collaborative signal processing can reduce traffic
to let a node broadcast its address before it starts listening. If
and improve sensing quality. In-network processing implies that
a node wants to talk to a neighboring node, it must wait until it
data will be processed as whole messages at a time in store-and-
receives the neighbor’s broadcast.
forward fashion, so packet or fragment-level interleaving from
Woo and Culler [14] examined different configurations of
multiple sources only increases overall latency.
carrier sense multiple access (CSMA) and proposed an adap-
Finally, we expect that applications will have long idle pe-
tive rate control mechanism, whose main goal is to achieve fair
riods and can tolerate some latency. In sensor networks, the
bandwidth allocation to all nodes in a multi-hop network. They
application such as surveillance or monitoring will be vigilant
have used the motes and TinyOS platform to test and measure
for long periods of time, but largely inactive until something
different MAC schemes. In comparison, our approach does not
is detected. For such applications, network lifetime is critical.
promote per-node fairness, and even trade it off for further en-
These classes of applications can often also tolerate some ad-
ergy savings.
ditional latency. For example, the speed of the sensed object
places a bound on how rapidly the network must detect an ob-
III. S ENSOR -MAC P ROTOCOL D ESIGN
ject. (One application-level approach to manage latency is to
The main goal in our MAC protocol design is to reduce en- deploy a slightly larger sensor network and have edge nodes
ergy consumption, while supporting good scalability and colli- raise the network to heightened awareness when something is
sion avoidance. Our protocol tries to reduce energy consump- detected.)
tion from all the sources that we have identified to cause en- These assumptions about the network and application
ergy waste, i.e., idle listening, collision, overhearing and control strongly influence our MAC design and motivate its differences
overhead. To achieve the design goal, we have developed the S- from existing protocols such as IEEE 802.11.
MAC that consists of three major components: periodic listen
and sleep, collision and overhearing avoidance, and message B. Periodic Listen and Sleep
passing. Before describing them we first discuss our assump- As stated above, in many sensor network applications, nodes
tions about the wireless sensor network and it applications. are in idle for a long time if no sensing event happens. Given
the fact that the data rate during this period is very low, it is
A. Network and Application Assumptions not necessary to keep nodes listening all the time. Our protocol
Since sensor networks are somewhat different than traditional reduces the listen time by letting node go into periodic sleep
IP networks or ad hoc networks of laptop computers, we next mode. For example, if in each second a node sleeps for half
summarize our assumptions about sensor networks and applica- second and listens for the other half, its duty cycle is reduced to
tions. 50%. So we can achieve close to 50% energy savings.
We expect sensor networks to be composed of many small
nodes deployed in an ad hoc fashion. Sensor networks will B.1 Basic Scheme
be composed of many small nodes to take advantage of phys- The basic scheme is shown in Figure 1. Each node goes to
ical proximity to the target to simplify signal processing. The sleep for some time, and then wakes up and listens to see if any
large number of nodes can also take advantage of short-range, other node wants to talk to it. During sleep, the node turns off
multi-hop communication (instead of long-range communica- its radio, and sets a timer to awake itself later.
tion) to conserve energy [4]. Most communication will be be- The duration of time for listening and sleeping can be selected
tween nodes as peers, rather than to a single base-station. Be- according to different application scenarios. For simplicity these
cause there are many nodes, they will be deployed casually in values are the same for all the nodes.
an ad hoc fashion, rather than carefully positioned. Nodes must Our scheme requires periodic synchronization among neigh-
therefore self-configure. boring nodes to remedy their clock drift. We use two techniques
We expect most sensor networks to be dedicated to a single to make it robust to synchronization errors. First, all timestamps
application or a few collaborative applications, thus rather than that are exchanged are relative rather than absolute. Second,
node-level fairness (like in the Internet), we focus on maximiz- the listen period is significantly longer than clock error or drift.
ing system-wide application performance. For example, the listen duration of 0.5s is more than 105 times
In-network processing is critical to sensor network life- longer than typical clock drift rates. Compared with TDMA
time [5], [6]. Since sensor networks are committed to one or schemes with very short time slots, our scheme requires much
a few applications, application-specific code can be distributed looser synchronization among neighboring nodes. All nodes are
through the network and activated when necessary or distributed free to choose their own listen/sleep schedules. However, to
on-demand. Techniques such as data aggregation can reduce reduce control overhead, we prefer neighboring nodes to syn-

  

  

  

rebroadcasting the schedule.

C A B D 3. If a node receives a different schedule after it selects and
broadcasts its own schedule, it adopts both schedules (i.e., it
Fig. 2. Neighboring nodes A and B have different schedules. They synchronize
with nodes C and D respectively. schedules itself to wake up at the times of both is neighbor and
itself). It broadcasts it own schedule before going to sleep.
We expect that nodes only rarely adopt multiple schedules,
chronize together. That is, they listen at the same time and go to since every node tries to follow existing schedules before choos-
sleep at the same time. It should be noticed that not all neigh- ing an independent one. On the other hand, it is possible that
boring nodes can synchronize together in a multi-hop network. some neighboring nodes fail to discover each other at beginning
Two neighboring nodes A and B may have different schedules if due to collisions when broadcasting schedules. They may still
they each in turn must synchronize with different nodes, C and find each other later in their subsequent periodic listening.
D, respectively, as shown in Figure 2. To illustrate this algorithm, consider a network where all
Nodes exchange their schedules by broadcasting it to all its nodes can hear each other. The timer of one node will fire first
immediate neighbors. This ensures that all neighboring nodes and its broadcast will synchronize all of its peers on its sched-
can talk to each other even if they have different schedules. For ule. If instead two nodes independently assign schedules (either
example, in Figure 2 if node A wants to talk to node B, it just because they cannot hear each other, or because they happen
wait until B is listening. If multiple neighbors want to talk to to transmit at nearly the same time), those nodes on the bor-
a node, they need to contend for the medium when the node is der between the two schedules will adopt both. In this way, a
listening. The contention mechanism is the same as that in IEEE node only needs to send once for a broadcast packet. The dis-
802.11, i.e., using RTS (Request To Send) and CTS (Clear To advantage is that these border nodes have less time to sleep and
Send) packets. The node who first sends out the RTS packet consume more energy than others.
wins the medium, and the receiver will reply with a CTS packet. Another option is to let the nodes on the border adopt only
After they start data transmission, they do not follow their sleep one schedule, which is the one it receives first. Since it knows
schedules until they finish transmission. another schedule that some other neighbors follow, it can still
Another characteristic of our scheme is that it forms nodes talk to them. However, for broadcast packets, it needs to send
into a flat topology. Neighboring nodes are free to talk to each twice to the two different schedules. The advantage is that the
other no matter what listen schedules they have. Synchronized border nodes have the same simple pattern of period listen and
nodes from a virtual cluster. But there is no real clustering and sleep as other nodes.
thus no problems of inter-cluster communications and interfer-
ence. This scheme is quite easy to adapt to topology changes. B.3 Maintaining Synchronization
We will talk about this issue later. The listen/sleep scheme requires synchronization among
The downside of the scheme is that the latency is increased neighboring nodes. Although the long listen time can tolerate
due to the periodic sleep of each node. Moreover, the delay fairly large clock drift, neighboring nodes still need to period-
can accumulate on each hop. So the latency requirement of the ically update each other their schedules to prevent long-time
application places a fundamental limit on the sleep time. clock drift. The updating period can be quite long. The mea-
surements on our testbed nodes show that it can be on the order
B.2 Choosing and Maintaining Schedules of tens of seconds.
Before each node starts its periodic listen and sleep, it needs Updating schedules is accomplished by sending a SYNC
to choose a schedule and exchange it with its neighbors. Each packet. The SYNC packet is very short, and includes the ad-
node maintains a schedule table that stores the schedules of all dress of the sender and the time of its next sleep. The next-sleep
its known neighbors. It follow the steps below to choose its time is relative to the moment that the sender finishes transmit-
schedule and establish its schedule table. ting the SYNC packet, which is approximately when receivers
1. The node first listens for a certain amount of time. If it does get the packet (since propagation delays are short). Receivers
not hear a schedule from another node, it randomly chooses a will adjust their timers immediately after they receive the SYNC
time to go to sleep and immediately broadcasts its schedule in packet. A node will go to sleep when the timer fires.
a SYNC message, indicating that it will go to sleep after t sec- In order for a node to receive both SYNC packets and data
onds. We call such a node a synchronizer, since it chooses its packets, we divide its listen interval into two parts. The first
schedule independently and other nodes will synchronize with part is for receiving SYNC packets, and the second one is for
it. receiving RTS packets, as shown in Figure 3. Each part is fur-
2. If the node receives a schedule from a neighbor before choos- ther divided into many time slots for senders to perform carrier
ing its own schedule, it follows that schedule by setting its sense. For example, if a sender wants to send a SYNC packet,
schedule to be the same. We call such a node a follower. It it starts carrier sense when the receiver begins listening. It ran-
then waits for a random delay td and rebroadcasts this schedule, domly selects a time slot to finish its carrier sense. If it has not
indicating that it will sleep in t − td seconds. The random delay detected any transmission by the end of the time slot, it wins the
is for collision avoidance, so that multiple followers triggered medium and starts sending its SYNC packet at that time. The
from the same synchronizer do not systematically collide when same procedure is followed when sending data packets.
 Receiver Listen

Sleep E C A B D F
for SYNC for RTS
Fig. 4. Who should sleep when node A is transmitting to B?

Sender 1 SYNC

CS Sleep ments the NAV value until it reaches zero. When a node has
data to send, it first looks at the NAV. If its value is not zero, the
node determines that the medium is busy. This is called virtual
Sender 2 RTS carrier sense.
Physical carrier sense is performed at the physical layer by
CS Send data if CTS received
listening to the channel for possible transmissions. The pro-
cedure was described in section III-B.3. The randomized car-
Sender 3 rier sense time is very important for collision avoidance. The
SYNC RTS
medium is determined as free if both virtual and physical carrier
CS CS Send data if CTS received sense indicate that it is free.
All senders perform carrier sense before initiating a transmis-
Fig. 3. Timing relationship between a receiver and different senders. CS stands sion. If a node fails to get the medium, it goes to sleep and wakes
for carrier sense. up when the receiver is free and listening again. Broadcast pack-
ets are sent without using RTS/CTS. Unicast packets follow the
sequence of RTS/CTS/DATA/ACK between the sender and the
Figure 3 also shows the timing relationship of three possi- receiver.
ble situations that a sender transmits to a receiver. CS stands
C.2 Overhearing Avoidance
for carrier sense. In the figure, sender 1 only sends a SYNC
packet. Sender 2 only wants to send data. Sender 3 sends a In 802.11 each node keeps listening to all transmissions from
SYNC packet and a RTS packet. its neighbors in order to perform effective virtual carrier sensing.
Each node periodically broadcasts SYNC packets to its neigh- As a result, each node overhears a lot of packets that are not
bors even if it has no followers. This allows new nodes to join an directed to itself. This is a significant waste of energy, especially
existing neighborhood. The new node follows the same proce- when node density is high and traffic load is heavy.
dure in the above subsection to choose its schedule. The initial Our protocol tries to avoid overhearing by letting interfering
listen period should be long enough so that it is able to learn nodes go to sleep after they hear an RTS or CTS packet. Since
and follow an existing schedule before choosing an independent DATA packets are normally much longer than control packets,
one. the approach prevents neighboring nodes from overhearing long
DATA packets and the following ACKs. In next subsection we
C. Collision and Overhearing Avoidance describe how to efficiently transmit a long packet combining
with the overhearing avoidance. Now we look at which nodes
Collision avoidance is a basic task of MAC protocols. S-
should go to sleep when there is an active transmission going
MAC adopts a contention-based scheme. It is common that any
on.
packet transmitted by a node is received by all its neighbors even
As shown in Figure 4, node A, B, C, D, E, and F forms a
though only one of them is the intended receiver. Overhearing
multi-hop network where each node can only hear the transmis-
makes contention-based protocols less efficient in energy than
sions from its immediate neighbors. Suppose node A is cur-
TDMA protocols. So it needs to be avoided.
rently transmitting a data packet to B. The question is, which of
the remaining nodes should go to sleep now.
C.1 Collision Avoidance
Remember that collision happens at the receiver. It is clear
Since multiple senders may want to send to a receiver at the that node D should go to sleep since its transmission interferes
same time, they need to contend for the medium to avoid col- with B’s reception. It is easy to show that node E and F do
lisions. Among contention based protocols, the 802.11 does a not produce interference, so they do not need to go to sleep.
very good job of collision avoidance. Our protocol follows sim- Should node C go to sleep? C is two-hop away from B, and its
ilar procedures, including both virtual and physical carrier sense transmission does not interfere with B’s reception, so it is free
and RTS/CTS exchange. We adopt the RTS/CTS mechanism to to transmit to its other neighbors like E. However, C is unable to
address the hidden terminal problem [15]. get any reply from E, e.g., CTS or data, because E’s transmission
There is a duration field in each transmitted packet that indi- collides with A’s transmission at node C. So C’s transmission is
cates how long the remaining transmission will be. So if a node simply a waste of energy. In summary, all immediate neighbors
receives a packet destined to another node, it knows how long of both the sender and the receiver should sleep after they hear
it has to keep silent. The node records this value in an variable the RTS or CTS packet until the current transmission is over.
called the network allocation vector (NAV) [1] and sets a timer Each node maintains the NAV to indicate the activity in its
for it. Every time when the NAV timer fires, the node decre- neighborhood. When a node receives a packet destined to other
nodes, it updates its NAV by the duration field in the packet. immediately. However, the node will learn it from the extended
A non-zero NAV value indicates that there is an active trans- fragments or ACKs when it wakes up.
mission in its neighborhood. The NAV value decrements every It is worth to note that IEEE 802.11 also has the fragmenta-
time when the NAV timer fires. Thus a node should sleep to tion support. We should point out the difference between that
avoid overhearing if its NAV is not zero. It can wake up when scheme with our message passing.
its NAV becomes zero. In 802.11, the RTS and CTS only reserves the medium for
the first data fragment and the first ACK. The first fragment and
D. Message Passing ACK then reserves the medium for the second fragment and
This subsection describes how to efficiently transmit a long ACK, and so forth. So for each neighboring node, after it re-
message in both energy and latency. A message is the collection ceives a fragment or an ACK, it knows that there is one more
of meaningful, interrelated units of data. It can be a long series fragment to be sent. So it has to keep listening until all the
of packets or a short packet, and usually the receiver needs to fragments are sent. Again, for energy-constrained nodes, over-
obtain all the data units before it can perform in-network data hearing by all neighbors wastes a lot of energy.
processing or aggregation. The reason for 802.11 to do so is to promote fairness. If
The disadvantages of transmitting a long message as a single the sender fails to get an ACK for any fragment, it must give
packet is the high cost of re-transmitting the long packet if only a up the transmission and re-contend for the medium. So other
few bits have been corrupted in the first transmission. However, nodes have a chance to transmit. This causes a long delay if
if we fragment the long message into many independent small the receiver really need the entire message to start processing.
packets, we have to pay the penalty of large control overhead In contrast, message passing extends the transmission time and
and longer delay. It is so because the RTS and CTS packets are re-transmits the current fragment. Thus it has fewer contentions
used in contention for each independent packet. and a small latency. There should be a limit on how many exten-
Our approach is to fragment the long message into many sions can be made for each message in case that the receiver is
small fragments, and transmit them in burst. Only one RTS really dead or lost in connection during the transmission. How-
packet and one CTS packet are used. They reserve the medium ever, for sensor networks, application-level fairness is the goal
for transmitting all the fragments. Every time a data fragment as opposed to per-node fairness.
is transmitted, the sender waits for an ACK from the receiver.
If it fails to receive the ACK, it will extend the reserved trans- E. Energy Savings vs. Increased Latency
mission time for one more fragment, and re-transmit the current
This subsection analyzes the trade-offs between the energy
fragment immediately.
savings and the increased latency due to nodes sleep schedules.
As before, all packets have the duration field, which is now
We compare our protocol with protocols that do not have peri-
the time needed for transmitting all the remaining data frag-
odic sleep such as the IEEE 802.11,
ments and ACK packets. If a neighboring node hears a RTS
For a packet moving through a multi-hop network, it experi-
or CTS packet, it will go to sleep for the time that is needed to
ences the following delays at each hop:
transmit all the fragments.
Switching the radio from sleep to active does not occur in- Carrier sense delay is introduced when the sender performs
stantaneously. For example, the RFM radio on our testbed needs carrier sense. Its value is determined by the contention window
20µs to switch from sleep mode to receive mode [8]. Therefore, size.
it is desirable to reduce the frequency of switching modes. The Backoff delay happens when carrier sense failed, either be-
message passing scheme tries to put nodes into sleep state as cause the node detects another transmission or because collision
long as possible, and hence reduces switching overhead. occurs.
The purpose of using ACK after each data fragment is to pre- Transmission delay is determined by channel bandwidth,
vent the hidden terminal problem. It is possible that a neigh- packet length and the coding scheme adopted.
boring node wakes up or a new node joins in the middle of a Propagation delay is determined by the distance between the
transmission. If the node is only the neighbor of the receiver but sending and receiving nodes. In sensor networks, node distance
not the sender, it will not hear the data fragments being sent by is normally very small, and the propagation delay can normally
the sender. If the receiver does not send ACK frequently, the be ignored.
new node may mistakenly infer from its carrier sense that the Processing delay. The receiver needs to process the packet
medium is clear. If it starts transmitting, the current transmis- before forwarding it to the next hop. This delay mainly depends
sion will be corrupted at the receiver. on the computing power of the node and the efficiency of in-
Each data fragment and ACK packet also has the duration network data processing algorithms.
field. In this way, if a node wakes up or a new node joins in the Queuing delay depends on the traffic load. In the heavy traffic
middle, it can properly go to sleep no matter if it is the neighbor case, queuing delay becomes a dominant factor.
of the sender or the receiver. For example, suppose a neigh- The above delays are inherent to a multi-hop network using
boring node receives an RTS from the sender or a CTS from contention-based MAC protocols. These factors are the same
the receiver, it goes to sleep for the entire message time. If the for both S-MAC and 802.11-like protocols. An extra delay in
sender extends the transmission time due to fragment losses or S-MAC is caused by nodes periodic sleeping. When a sender
errors, the sleeping neighbor will not be aware of the extension gets a packet to transmit, it must wait until the receiver wakes
 Delay vs. energy savings
3.5

Listen time 300ms

3 Listen time 200ms

2.5
Average sleep delay (s)

1.5

1 Fig. 6. The UCB Rene Mote.

0.5

AT90LS8535 microcontroller [18], which has 8K bytes of pro-

0
0 20 40 60
Percentage energy savings (%)
80 100 grammable flash and 512 bytes of data memory.
The radio transceiver on the mote is the model TR1000 from
Fig. 5. Energy savings vs. average sleep delay for the listen time of 30ms. RF Monolithics, Inc [8]. When using the OOK(on-off keyed)
modulation, it provides a transmission rate of 19.2 Kbps. It
has three working modes, i.e., receiving, transmitting and sleep,
up. We call it sleep delay since it is caused by the sleep of the each drawing the input current of 4.5mA, 12mA (peak) and 5µA
receiver. respectively.
We call a complete cycle of the listen and sleep a frame. As- Our motes use TinyOS, an efficient event-driven operating
sume a packet arrives at the sender with equal probability in time system [9], [19]. It provides the basic mechanism for packet
within a frame. So the average sleep delay on the sender is transmitting, receiving and processing. TinyOS promotes mod-
ularity, data sharing and reuse.
Ds = Tframe /2 (1) As of July 2001, the standard release of TinyOS has only one
type of packet, which consists of a header, the payload and a
where cyclic redundancy check (CRC). The length of the header or the
payload can be changed to different values. However, once they
Tframe = Tlisten + Tsleep (2) are defined, all packets have the same length and format. In our
MAC implementation, the header, payload and CRC fields have
Comparing with protocols without periodic sleep, the relative 6B, 30B and 2B respectively.
energy savings in S-MAC is Normally the control packets, such as RTS, CTS and ACK,
are very short and without payload. So we have created an other
Tsleep Tlisten packet type in TinyOS, the control packet, which only has the
Es = =1− (3)
Tframe Tframe 6-byte header and the 2-byte CRC. We have modified several
TinyOS components to accommodate the new packet. This en-
The last item in the above equation is the duty cycle of the ables us to efficiently implement MAC protocols and accurately
node. It is desirable to have the listen time as short as possible measure their performance.
so that for a certain duty cycle, the average sleep delay is short.
In our implementation we set the listen time as 300ms. Figure 5 B. Implementation of MAC Protocols
shows the percentage of energy savings Es vs. average sleep
We have implemented three MAC modules on the mote and
delay Ds on each node for the listen time of 300ms and 200ms.
TinyOS platform, as listed below.
We can see that even if the sleep time is zero (no sleeping) there
1. Simplified IEEE 802.11 DCF
is still a delay. This effect is because contention only starts at
2. Message passing with overhearing avoidance
the beginning of each listen interval.
3. The complete S-MAC
IV. P ROTOCOL I MPLEMENTATION For the purpose of performance comparison, we first imple-
mented a simplified version of IEEE 802.11 DCF. It has the fol-
The purpose of our implementation is to demonstrate the ef- lowing major pieces: physical and virtual carrier sense, back-
fectiveness of our protocol and to compare our protocol with off and retry, RTS/CTS/DATA/ACK packet exchange, and frag-
802.11 through some basic experiments. mentation support.
The duration of each carrier sense is a random time within
A. Testbed
the contention window. The randomization is very important to
We use Rene Motes, developed at UCB [7], as our develop- avoid collisions at the first step. For simplicity, the contention
ment platform and testbed (see Figure 6). A mote is slightly window does not exponentially increase when backoff happens.
larger than a quarter. The heart of the node is the Atmel The fragmentation support follows the same procedure as in
 A E loads.
Source 1 Sink 2 The two sources periodically generate a sensing message,
which is divided into some fragments. In the simplified IEEE
C 802.11 MAC, these fragments are sent in a
burst, i.e., RTS/CTS is not used for each fragment. We did not
measure the 802.11 MAC without fragmentation, which treats
each fragment as an independent packet and uses RTS/CTS for
Source 2 Sink 1 each of them, since it is obvious that this MAC consumes much
more energy than the one with fragmentation. In our protocol,
B D
message passing is used, and fragments of a message are always
Fig. 7. Topology used in experiments: two-hop network with two sources and transmitted in a burst.
two sinks. We change the traffic load by varying the inter-arrival period
of the messages. If the message inter-arrival period is 5 seconds,
a message is generated every 5 seconds by each source node.
IEEE 802.11 standard [1] and is described in Section 3 of this In our following experiments, the message inter-arrival period
paper. varies from 1s to 10s.
With 802.11 the radio of each node does not go into sleep For each traffic pattern, we have done 10 independent tests
mode. It is either in listen/receiving mode or transmitting mode. to measure the energy consumption of each node when using
The second module is the message passing with overhearing different MAC protocols. In each test, each source periodi-
avoidance. It achieves energy savings by avoiding overhearing, cally generates 10 messages, which in turn is fragmented into 10
reducing control overhead and contention times. It does not in- small data packets supported by the TinyOS. Thus in each ex-
clude the period listen and sleep. So there is no additional delay periment, there are 200 TinyOS data packets to be passed from
comparing with the simplified IEEE 802.11. The radio of each their sources to their sinks. For the highest rate with a 1s inter-
node goes into the sleep mode only when its neighbors are in arrival time, the wireless channel is nearly fully utilized due to
transmission. its low bandwidth.
With the message passing module we have incorporated peri- We measure the amount of time that each node has used to
odic listen and sleep, and completed most basic functionalities pass these packets as well as the percentage time its radio has
in S-MAC. Currently, the listen time for each node is 300ms, and spent in each mode (transmitting, receiving, listening or sleep).
sleep time can be changed to different values, such as 300ms, The energy consumption in each node is then calculated by mul-
500ms, 1s, etc., which makes different duty cycles of the radio. tiplying the time with the required power to operate the radio in
We can also specify the frequency that the SYNC packet is sent that mode. We found the power consumption from the data sheet
for schedule update between neighboring nodes. In our follow- of the radio transceiver, which is 13.5mW, 24.75mW and 15µW,
ing experiments, we have chosen the sleep time as 1 second and in receiving, transmitting and sleep respectively. There is no dif-
the frequency for schedule update is 10 listen/sleep period, i.e., ference between listening and receiving in this radio transceiver
13 seconds. model.
It should be noted that the energy savings in the current imple-
mentation is only due to the sleep of the radio. In other words, B. Results and Analysis
the microcontroller does not go to sleep. It actually has a sleep The experiments are carried out on the three MAC modules
mode, which consumes much less energy and can be waked up we have implemented on our testbed nodes. In the result graphs,
by a low-frequency watchdog timer. If we put the microcon- the simplified IEEE 802.11 DCF is denoted as ‘IEEE 802.11’.
troller into the sleep mode as well when the radio is sleeping, The message passing with overhearing avoidance is identified
we are able to save more energy. as ‘Overhearing avoidance’. The complete S-MAC protocol,
which includes all pieces of our new protocol, is denoted as ‘S-
V. E XPERIMENTATION
MAC’.
The main goal of the experimentation described here is to We first look at the experiment results on the source nodes
measure the energy consumption of the radio for using each of A and B. Figure 8 is the measured average energy consumption
the MAC modules we have implemented. from these two nodes. The traffic is heavy when the message
inter-arrival time is less than 4s. In this case, 802.11 MAC uses
A. Experiment Setup more than twice the energy used by S-MAC. Since idle listening
Figure 7 is the topology we used in our experiments. This rarely happens, energy savings from periodic sleeping is very
is a two-hop network with two sources and two sinks. Packets limited. S-MAC achieves energy savings mainly by avoiding
from source A flow through node C and end at sink D, while overhearing and efficiently transmitting a long message.
those from B also pass through C but end at E. The topology is When the message inter-arrival period is larger than 4s, traffic
simple, but it is sufficient to show the basic characteristics of the load becomes light. In this case, the complete S-MAC protocol
MAC protocols. has the best energy property, and far outperforms 802.11 MAC.
We will look at the energy consumption of each node when Message passing with overhearing avoidance also performs bet-
utilizing different MAC protocols and under different traffic ter than 802.11 MAC. However, as shown in the figure, when
 Average energy consumption in the source nodes Percentage time source nodes in sleep
1800 100

IEEE802.11 90 Overhearing avoidance

1600 Overhearing avoidance S−MAC
S−MAC
80
1400
Energy consumption (mJ)

70
1200

Percentage time
60
1000
50
800
40
600
30

400
20

200 10

0
0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11
Message inter−arrival period (second) Message inter−arrival period (second)

Fig. 8. Measured energy consumption in the source nodes. Fig. 9. Measured percentage of time that the source nodes in the sleep mode.

Energy consumption in the intermediate node

idle listening dominates the total energy consumption, the pe- 1800

riodic sleep plays a key role for energy savings. The energy IEEE802.11
1600 Overhearing avoidance
consumption of S-MAC is relatively independent of the traffic S−MAC
pattern. 1400

Compared with 802.11, message passing with overhearing

Energy consumption (mJ)

1200
avoidance saves almost the same amount of energy under all
traffic conditions. This result is due to overhearing avoidance 1000

among neighboring nodes A, B and C. The number of packets

800
to be sent by each of them are the same in all traffic conditions.
Figure 9 shows the percentage of time that the source nodes 600

are in the sleep mode. It is interesting that the S-MAC protocol 400
adjusts the sleep time according to traffic patterns. When there
is little traffic, the node has more sleep time (although there is 200

a limit by the duty cycle of the node). When traffic increases, 0

0 1 2 3 4 5 6 7 8 9 10 11
nodes have fewer chances to go to periodic sleep and thus spend Message inter−arrival period (second)
more time in transmission.
This is a useful feature for sensor network applications, since Fig. 10. Measured energy consumption in the intermediate node.
the traffic load indeed changes over time. When there is no sens-
ing event, the traffic is very light. When some nodes detects an
event, it may trigger a big sensor like a camera, which will gen-
source nodes (Figure 8). But we cannot see similar results on
erate heavy traffic. The S-MAC protocol is able to adapt to the
the intermediate node C, since all packet transmissions involve
traffic changes. In comparison, the module of message passing
this node. In this case, its energy consumption is about the same
with overhearing avoidance does not have periodic sleep, and
as that of using the 802.11 MAC.
nodes spend more and more time in idle listening when traffic
load decreases.
Figure 10 shows the measured energy consumption in the in- VI. C ONCLUSIONS AND F UTURE W ORK
termediate node C. We can see in the light traffic case, it still
outperforms 802.11 MAC. In heave traffic case, it consumes This paper presents a new MAC protocol for wireless sensor
slightly more energy than 802.11. One reason is that S-MAC networks. It has very good energy conserving properties com-
has synchronization overhead of sending and receiving SYNC paring with IEEE 802.11. Another interesting property of the
packets. Another reason is that S-MAC introduces more latency protocol is that it has the ability to make trade-offs between en-
and actually uses more time to pass the same amount of data. ergy and latency according to traffic conditions. The protocol
In fact, if the traffic is extremely heavy and a node does not has been implemented on our testbed nodes, which shows its
have any chance to follow its sleep schedule, the scheme of pe- effectiveness.
riodic listen and sleep does not benefit at all. However, message Future work includes system scaling studies and parameter
passing and overhearing avoidance are still effective means of analysis. More tests will be done on larger testbeds with differ-
saving energy. This has been illustrated in the results of the ent number of nodes and system complexity.
 ACKNOWLEDGMENTS [18] Atmel Corporation, http://www.atmel.com/, AVR Microcontroller
AT90LS8535 Reference Manual.
This work is supported by NSF under grant ANI-9979457 [19] http://tinyos.millennium.berkeley.edu/.
as the SCOWR project (http://robotics.usc.edu/projects/
scowr/), and by DARPA under grant DABT63-99-1-0011 as
the SCADDS project (http://www.isi.edu/scadds/) and un-
der contract N66001-00-C-8066 as the SAMAN project (http:
//www.isi.edu/saman/) via the Space and Naval Warfare Sys-
tems Center San Diego.
The authors would like to acknowledge the discussions and
suggestions from members of the SCOWR, SCADDS and
SAMAN projects.
We would also like to thank the TinyOS group (http://
tinyos.millennium.berkeley.edu/) at UCB for their support
with TinyOS and Motes.

R EFERENCES
[1] LAN MAN Standards Committee of the IEEE Computer Society, Wireless
LAN medium access control (MAC) and physical layer (PHY) specifica-
tion, IEEE, New York, NY, USA, IEEE Std 802.11-1997 edition, 1997.
[2] Mark Stemm and Randy H Katz, “Measuring and reducing energy con-
sumption of network interfaces in hand-held devices,” IEICE Transactions
on Communications, vol. E80-B, no. 8, pp. 1125–1131, Aug. 1997.
[3] Oliver Kasten, Energy Consumption, http://www.inf.ethz.ch/
˜kasten/research/bathtub/energy_consumption.html,
Eldgenossische Technische Hochschule Zurich.
[4] Gregory J. Pottie and William J. Kaiser, “Embedding the internet: wireless
integrated network sensors,” Communications of the ACM, vol. 43, no. 5,
pp. 51–58, May 2000.
[5] Chalermek Intanagonwiwat, Ramesh Govindan, and Deborah Estrin, “Di-
rected diffusion: A scalable and robust communication paradigm for sen-
sor networks,” in Proceedings of the ACM/IEEE International Conference
on Mobile Computing and Networking, Boston, MA, USA, Aug. 2000, pp.
56–67, ACM.
[6] John Heidemann, Fabio Silva, Chalermek Intanagonwiwat, Ramesh
Govindan, Deborah Estrin, and Deepak Ganesan, “Building efficient wire-
less sensor networks with low-level naming,” in Proceedings of the Sym-
posium on Operating Systems Principles, Lake Louise, Banff, Canada,
Oct. 2001.
[7] http://www.cs.berkeley.edu/˜awoo/smartdust/.
[8] RF Monolithics Inc., http://www.rfm.com/, ASH Transceiver
TR1000 Data Sheet.
[9] Jason Hill, Robert Szewczyk, Alec Woo, Seth Hollar, David Culler, and
Kristofer Pister, “System architecture directions for networked sensors,” in
Proceedings of the 9th International Conference on Architectural Support
for Programming Languages and Operating Systems, Cambridge, MA,
USA, Nov. 2000, pp. 93–104, ACM.
[10] S. Singh and C.S. Raghavendra, “PAMAS: Power aware multi-access pro-
tocol with signalling for ad hoc networks,” ACM Computer Communica-
tion Review, vol. 28, no. 3, pp. 5–26, July 1998.
[11] Frazer Bennett, David Clarke, Joseph B. Evans, Andy Hopper, Alan Jones,
and David Leask, “Piconet: Embedded mobile networking,” IEEE Per-
sonal Communications Magazine, vol. 4, no. 5, pp. 8–15, Oct. 1997.
[12] Katayoun Sohrabi and Gregory J. Pottie, “Performance of a novel self-
organization protocol for wireless ad hoc sensor networks,” in Proceedings
of the IEEE 50th Vehicular Technology Conference, 1999, pp. 1222–1226.
[13] Wendi Rabiner Heinzelman, Anantha Chandrakasan, and Hari Balakrish-
nan, “Energy-efficient communication protocols for wireless microsensor
networks,” in Proceedings of the Hawaii International Conference on Sys-
tems Sciences, Jan. 2000.
[14] Alec Woo and David Culler, “A transmission control scheme for media
access in sensor networks,” in Proceedings of the ACM/IEEE Interna-
tional Conference on Mobile Computing and Networking, Rome, Italy,
July 2001, ACM.
[15] V. Bharghavan, A. Demers, S. Shenker, and L. Zhang, “Macaw: A media
access protocol for wireless lans,” in Proceedings of the ACM SIGCOMM
Conference, 1994.
[16] Jaap C. Haartsen, “The Bluetooth radio system,” IEEE Personal Commu-
nications Magazine, pp. 28–36, Feb. 2000.
[17] Bluetooth SIG Inc., “Specification of the Bluetooth system: Core,” http:
//www.bluetooth.org/, 2001.