A transmission control scheme for media access in sensor networks

TL;DR: This work proposes an adaptive rate control mechanism aiming to support media access control in sensor networks and finds that such a scheme is most effective in achieving the authors' fairness goal while being energy efficient for both low and high duty cycle of network traffic.

Abstract: We study the problem of media access control in the novel regime of sensor networks, where unique application behavior and tight constraints in computation power, storage, energy resources, and radio technology have shaped this design space to be very different from that found in traditional mobile computing regime. Media access control in sensor networks must not only be energy efficient but should also allow fair bandwidth allocation to the infrastructure for all nodes in a multihop network. We propose an adaptive rate control mechanism aiming to support these two goals and find that such a scheme is most effective in achieving our fairness goal while being energy efficient for both low and high duty cycle of network traffic.

Summary

1. INTRODUCTION

• Sensor networks are an important emerging area of mobile computing that presents novel wireless networking issues because of their unusual application requirements, highly constrained resources and functionality, small packet size, and deep multihop dynamic topologies.
• Often, applications will arrange periodic rendezvous so that data can be communicated over many hops while allowing nodes to turn off their radios for lengthy periods.
• Much of the traffic moves through the network over several hops, perhaps with some intermediate processing, to points that are connected to a larger processing infrastructure.
• At the very least, fairness is at odds with both energy efficiency and high channel utilization.
• In short, the characteristics and goals of MAC in sensor networks differ strongly from conventional computer networks.

2. SENSOR NETWORK DESIGN POINT

• The authors study is grounded in the small, low-power networked sensor device shown in Figure 1 [7].
• The packet-level component is responsible for spooling incoming bytes and delivering the packet receive event.
• The other component is responsible for building the dynamic multihop network and routing traffic.
• Originating sensor packets are marked for the parent.

2.2 Metric for Evaluation

• The metrics for evaluation of a sensor network MAC protocol stress both fairness and energy efficiency.
• High aggregate bandwidth of packets only from nodes around the base station is not desirable.
• The energy consumed is the total energy that the network has invested in propagating data to the base station.
• The authors experiment is designed such that all nodes are in receive mode even during idle period.
• Therefore, the authors only account useful work as energy spent in channel listening and packet transmission under this metric.

2.3 Simulation Environment

• Given the difficulty in performing actual measurements in wireless networking, the authors first evaluate their system through simulation.
• Each UNIX process represents a networked sensor, and a master process is responsible for synchronizing them to perform bit time simulation.
• There is a simple radio propagation model in the simulation and the authors assume bit error rate to be zero, since their main focus is media access control.
• The simulator doesn’t simulate the actual hardware operating in the TinyOS environment.
• It preserves the event driven semantics and the dynamics of traffic flow shown in Figure 2.

3. RELATED WORK

• In designing the MAC protocol, the authors first examine existing mechanisms and determine whether or not they apply in the regime of sensor networks.
• Other CSMA schemes rely on time synchronized slotted channel, such as Slotted ALOHA [11].
• As a result, re-exploring CSMA strategies with a different fundamental assumption is extremely relevant.
• Given the spontaneous nature of sensor networks where the receiver or the next hop to the base station may change at any time, such a scheme may be very inefficient in publishing different schedules.
• The proposed adaptive rate control scheme builds upon this work, but their goal is to have media access control assist in achieving fair bandwidth delivery to the base station for nodes in a multihop network.

4. DESIGN

• The authors discuss how media access control for sensor network should be done differently in this section.
• First, the authors explore what type of listening mechanism is appropriate for the case where all nodes can hear each other.
• Third, the authors present two mechanisms which they will study their effectiveness in the context of a multihop network.
• The first scheme is a conventional RTS/CTS contention control scheme, and the second one is their proposed adaptive transmission control scheme, and finally, a mechanism that all schemes can leverage off for avoiding some cases of hidden node problems in multihop network.

4.1 Listening Mechanism

• Carrier Sense Multiple Access (CSMA) and the Collision Detection (CD) scheme found in Ethernet are examples of listening mechanisms.
• Unfortunately, collision detection is not possible in wireless network technology without additional circuitry.
• Many protocols such as IEEE 802.11 require sensing the channel even during backoff.
• The highly synchronized nature of the traffic imposes a new criteria for CSMA.
• Detection of one common physical event will synchronize these nodes and lead them to send at the same time, which repeats periodically.

4.2 Backoff Mechanism

• The idea of backoff is to restrain a node from accessing the channel for a period of time and hopefully, the channel will become free after the backoff period.
• In the case of sensor networks where the traffic is a superposition of different periodic streams, backoff should not just restrain a node from sending for the backoff period.

4.3 Contention Based Mechanism

• Explicit contention control schemes, which are widely used in many MAC protocols, e.g., IEEE 802.11 [2] and MACAW [4], require the use of control packets, such as Request to Send (RTS) and Clear to Send (CTS).
• For sensor networks where packet size is small, they can constitute a large overhead.
• One advantage of a bidirectional multihop network is that acknowledgments are free when the receiving node (your parent in the multihop topology) routes the packet to its parent.
• This eliminates an explicit ACK control packet.
• Furthermore, if a node hears a CTS before any of its own transmission, it will defer transmission for one packet time to avoid corrupting the traffic.

4.4 Rate Control Mechanism

• The tension between originating traffic and route-thru traffic has a direct impact in achieving their fairness goal.
• Similarly, some kind of progressive signalling mechanism should exist for route-thru traffic, such that back pressure can propagate deep down into the network for those nodes to lower their rate of originating data.
• The adaptive rate control idea is very simple and can be explained with an analogy of metering traffic onto a freeway where the route-thru traffic is like traffic on the freeway and each node originating data is like cars trying to enter.
• The amount of computation for this adaptive scheme is small and within networked sensor’s computation capability.
• It requires a simple pseudo random number generator and a few addition and divide operations.

4.5 Multihop Hidden Node Problem

• The adaptive transmission control scheme attempts to avoid hidden node problem without explicit control packets by constantly tuning the transmission rate and performing phase changes, so that the aggregate periodic streams of traffic will not repeatedly collide with each other.
• The contention control RTS/CTS scheme can solve the hidden problem to some degree.
• As discussed in Section 3, MACAW has suggested a scenario where multihop hidden node problem cannot be solved due to lack of synchronized information of knowing when is the contention period between a child node and its grandparent’s node.
• If the authors assume that packets will be routed after some processing time x, a child node is able to avoid a potential hidden node problem with its grandparent.
• In fact, if the child node detects such situation it should perform a backoff to change its phase such that it will not encounter the same situation the next time it transmits.

5. ANALYSIS OF CSMA SCHEMES

• The authors use both simulated and empirical measurements to explore the various ways of performing CSMA, and study their performance based on their energy efficiency metric, as well as traditional metrics, such as channel utilization and fairness.
• Traditional CSMA schemes have two basic design parameters: the carrier sense (or listening) mechanism and the backoff mechanism.
• The listening period can be random over a fixed interval or constant.
• Backoff time is random drawn from a fixed window, binary exponentially increasing window, or binary exponentially decreasing window.
• All nodes in the cell are able to hear each other, with node 0 being the base station.

5.1 Simulation Settings

• To make the simulation approximate their real platform, the packet size is set to 30 bytes, which is the actual packet size used in many networked sensor applications on their prototype.
• With 30 byte packets in Manchester Encoding, the 10kbps channel capacity can deliver at most 20.8 packet/s.
• The authors use a 16-bit CRC error detection mechanism to check for corrupted packets.
• The specific the values of all the necessary parameters for the CSMA schemes in Table 1 are given in Table 2.

5.2 Delivered Bandwidth under Simulation

• Since the channel capacity is 20.8 packet/s, the traffic load will exceed capacity when more than 4 nodes are sending.
• Their aggregate bandwidth is not very robust, as indicated by the two dips in the figure.
• The remaining schemes, with random delay or random listening intervals achieve slightly less bandwidth, but are more robust.
• The randomness introduced by the backoff mechanism may seem to be sufficient to avoid repeated collisions, however, without collision detection hardware greater attention must be paid to the listen phase.
• The performance is almost insensitive to backoff mechanism.

5.3 Energy Usage

• In examining the energy consumed in communication, the authors separate the portion spent in actually transmitting and receiving packets from that spent listening.
• The former is determined primarily by the traffic load; differences result from the happenstance of packets being dropped.
• The authors CSMA schemes with constant listen period are the most energy efficient, at approximately 10uJ/packet independent of network size.
• 4 Fairness Figure 7 and Figure 8 show the deviation of mean throughput per node among the three CSMA schemes and 802.11 as an indication of fairness.
• From the 802.11 data that are not shown in the figures, the authors found that nodes, which have an earlier transmission start time, end up capturing the channel and result in this unfairness.

5.5 Sensor Phase Shifting

• The half duplex nature of the networking stack makes CSMA vulnerable to the capturing effect.
• That is, during reception of neighboring node’s transmission, the network stack will not start any transmissions issued by the application.
• Instead, it will fail the transmission back to the application.
• If the two nodes remain in synchrony with one starting its.

6. ANALYSIS OF MULTIHOP SCENARIO

• This section extends their analysis to multihop networks where two essential challenges are present.
• First, if nodes near the base station originate too much traffic, little will bandwidth will be available for more distant nodes.
• The CSMA scheme developed for media access control is augmented with a transmission control protocol so that nodes adapt their data origination rate to give a fair share to downstream nodes and to match available upstream bandwidth.
• Like TCP, it adjusts its rate based on observed packet loss.
• For comparison, the base CSMA and 802.11 schemes are carried forward to multihop networks, as well.

6.1 Reference Topology

• Nodes are hidden from each other if they are not linked by an edge.
• Constructing this topology by physical placement of the nodes is challenging, especially with automatic route discovery, given the variability in cell shapes.
• Still there is significant connectivity and interference not present in the simulation.
• For the multihop scenario shown in Figure 14, the channel capacity of node 2 will determine the ideal transmission rate of the overall network.

6.2 Simulation Measurements

• The simulation runs with each node sending packets to the base station at rate of 4 packet/s with the same start time.
• This is an example of the unfairness resulting from schemes that fail to accommodate the collective behavior.
• “Gateway” nodes, which are close to the base station, dominate the channel and use up most of the channel capacity for delivering their own packets.
• A large β will impose a smaller penalty, especially for route-thru traffic, such that the aggregate bandwidth is higher and is more fair.
• They do so by heavily favoring nodes near the base station.

6.3 Empirical Measurements

• As discussed in Section 1, an ad hoc multihop network topology is dynamically determined by how node placement and physical environment influence radio propagation, and by the choice of route discovery algorithms.
• Cell boundaries are not sharp, so interference effects may overreach useful communication cells and complicate the problem.
• In term of fairness, all three settings of ARC is more fair than D CONST FIX alone.
• This is expected because node 1b’s α is much smaller than 1a’s α since 1b has to deliver traffic from its nine children.
• In fact, Figure 22 shows the yield of each node, which suggests that a lower beta achieves higher yield for each packet sent, and therefore, result in higher aggregate bandwidth.

7. CONCLUSION

• The paper has shown how the application scenario, resource limitation, and network traffic characteristics in sensor networks differ from conventional computer networks and explained why existing MAC protocols are not suitable in this regime.
• A comprehensive study has been performed in understanding the appropriate carrier sensing mechanism.
• The conclusion is that random delay should be introduced prior to any transmission, with backoff acting as a phase shift for the periodicity of the application.
• A new, simple adaptive rate control scheme for achieving the desired metrics in a multihop network has been proposed and compared with conventional contention based schemes.
• The authors adaptive scheme is extremely efficient in energy for low traffic situation which is the common case in sensor networks.

Figures

Figure 14: Five level deep reference multihop network with 11 nodes. Edges represent bidirectional connectivity.

Table 1: Different CSMA schemes with variations on the amount of delay period, listening period, and ways of backoff. All exponential backoffs are binary exponential.

Figure 3: Network topology for the evaluating the CSMA scheme.

Table 4: Transmission rate breakdown of each node in the cell of node 2 in the multihop scenario with each node sending at the ideal transmission rate of X packet/s

Figure 1: Our low-power networked sensor device prototype.

Table 2: Values for the parameters of different CSMA scheme used in simulation. Since the radio’s raw bandwidth is 10kbps, one bit time is 100us.

Figure 4: Aggregate delivered bandwidth for CSMA schemes with typical start conditions and offered load of 5 packets per second per node.

Figure 5: Aggregate Bandwidth for the different CSMA schemes with nodes that perform no random delay to begin transmission at different time.

Figure 6: Average energy spent in listening the channel per unit of successful packet received by the base station.

Figure 12: Actual energy spent in CSMA listening over our network sensor prototypes for the different CSMA schemes.(Empirical)

Figure 13: Fairness comparison in terms of deviation of mean throughput per node. (Empirical)

Figure 23: Dynamics of transmission probabilities of node 1a, 1b, and 3c over time. (Empirical)

Table 7: Summary of aggregate bandwidth and energy efficiency measurement. The energy efficiency metric is the average amount of energy spent in CSMA listening and packet transmission per packet received by the base station.

Figure 2: Complete TinyOS application component graph.

Figure 21: Average transmission rate of each node. (Empirical)

Table 6: Measuring the effectiveness of the hidden node inference scheme over ARC with α = 0.08 and β = 0.5.

Figure 22: Percent yield (packet received by base station/packet transmitted) of each node. (Empirical)

Table 5: Summary of the effects in changing β in the simulation.

Figure 20: Delivered bandwidth received by the base station from each node. (Empirical)

Figure 8: Mean throughput per node delivered to the base station and the corresponding standard deviation as a measure of fairness for the different CSMA schemes.

Figure 7: Mean throughput per node delivered to the base station and standard deviation as a measure of fairness for the different CSMA schemes.

Full text

A Transmission Control Scheme for Media Access in
Sensor Networks
Alec Woo
Department of EECS
Computer Science Division
University of California, Berkeley
awoo@cs.berkeley.edu
David E. Culler
Computer Science Division
University of California, Berkeley
and Intel XIS Lab
culler@cs.berkeley.edu
ABSTRACT
We study the problem of media access control in the novel
regime of sensor networks, where unique application behav-
ior and tight constraints in computation power, storage, en-
ergy resources, and radio technology have shaped this design
space to be very different from that found in traditional mo-
bile computing regime. Media access control in sensor net-
works must not only be energy efficient but should also allow
fair bandwidth allocation to the infrastructure for all nodes
in a multihop network. We propose an adaptive rate control
mechanism aiming to support these two goals and find that
such a scheme is most effective in achieving our fairness goal
while being energy efficient for both low and high duty cycle
of network traffic.
1. INTRODUCTION
Sensor networks are an important emerging area of mobile
computing that presents novel wireless networking issues be-
cause of their unusual application requirements, highly con-
strained resources and functionality, small packet size, and
deep multihop dynamic topologies. Although many high-
level architectural and programming aspects of this area
are still being resolved, the underlying media access con-
trol (MAC) and transmission control protocols are critical
enabling technology for many sensor network applications.
These problems are well-studied for traditional computer
networks, however, the different wireless technologies, appli-
cation characteristics, and usage scenarios create a complex
mix of issues that have led to the existence of many distinct
solutions. It is natural to expect the low-level protocols to
evolve again for this new era.
Application behavior in sensor networks leads to very dif-
ferent traffic characteristics from that found in conventional
computer networks. The primary function of a sensor net-
work application is to sample the environment for sensory
information, such as temperature, and propagate this data
back to the infrastructure, while perhaps performing some
in-network processing, such as aggregation or compression.
The network tends to operate as a collective structure, rather
than supporting many independent point-to-point flows. Traf-
fic tends to be variable and highly correlated. Over lengthy
periods there may be little activity or traffic, but for short
periods the traffic may be very intense. For example, when
an abnormal event, such as a fire, is detected, many de-
vices will initiate communication at once. Often, applica-
tions will arrange periodic rendezvous so that data can be
communicated over many hops while allowing nodes to turn
off their radios for lengthy periods. Even in the simplest
case, roughly periodic sampling of the sensor field yields
correlated bursts, even when the duty cycle is low.
The data that a networked sensor generates for each sam-
ple, such as a temperature value, is relatively small and,
given the low bandwidth of the radio, data packets are kept
small with a typical size around tens of bytes. Multi-path
interference and short, irregular transmission range result in
unpredictable cell structure, but bidirectional connectivity
can generally be achieved.
Much of the traffic moves through the network over several
hops, perhaps with some intermediate processing, to points
that are connected to a larger processing infrastructure. The
network takes on an ad hoc multihop topology comprising
many levels, where the connectivity is determined dynami-
cally by how placement and physical environment influence
radio propagation and by the discovery algorithm. Interfer-
ence effects may overreach useful communication cells. At
each hop, traffic originating from the local sensor must be
merged with route-thru traffic. Often this merging is appli-
cation specific, but at the very least every node is both a
data source and a router. Generally, the amount of route-
thru traffic exceeds that of originating traffic.
The capabilities of sensor devices are also very different from
traditional nodes in a computer network. These devices have
a very limited amount of storage, processing power, and
most importantly, energy resources. These limitations cer-
tainly impose constraints in the design of the MAC protocol.
On these platforms, a typical low power RF radio delivers
moderate bandwidth in a single channel at the ISM band.
There is little or no dedicated support for carrier sensing,
collision detection, and no specific framing or encoding en-
forced by the hardware, other than basic DC-balance. Fur-
Permission to make digital or hard copies of part or all of this work or
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.
ACM SIGMOBILE 7/01 Rome, Italy
© 2001 ACM ISBN 1-58113-422-3/01/07…$5.OO
221

thermore, there are no specific protocol stacks in place to
dictate the MAC protocol design. It is roughly the same
cost per unit time to listen as to transmit or receive. Every
moment the radio is on, it consumes precious power. Thus,
a key requirement is to turn the radio off whenever possible.
Also, it is important to sense contention or frame packets
with a minimal number of bits and to minimize the number
of protocol control packets. Furthermore, there is very little
buffering available on the node. Typically, a single packet is
moving with only a few bits of buffering.
These application and platform characteristics give rise to a
new set of metrics. Not only are we interested in high chan-
nel utilization, we are interested in communication efficiency
in terms of energy consumed per unit of successful commu-
nication. Furthermore, fairness is highly desirable. For ex-
ample, we may want to collect roughly the same amount of
temperature data from each deployed sensor in a field to in-
fer the temperature gradient during a fire. Therefore, a fair
allocation of bandwidth delivered to the base station from
each node over multiple hops is desired. It is not sufficient
to share the channel fairly in an individual cell, we would
like to achieve a crude level of end-to-end fairness even in
a deep and self-organized multihop networks, which may
change dynamically and originate data at each intermediate
node.
The multihop nature of the network poses four interrelated
challenges. First, the originating traffic and the route-thru
traffic compete for the same upstream bandwidth. Trans-
mission rate control can potentially be applied to either.
Rate control is particularly important around the base sta-
tion, because traffic from nodes deep in the topology primar-
ily flows to a few “gateway” nodes. Second, a hidden node
problem exists, by definition, between every other pair of
levels in the network. Thus, it may not be possible to detect
contention at the upstream node and a significant loss rate
is to be expected. Third, the routing distance and degree of
intermediate competition varies widely across the network.
Nodes residing farther from the infrastructure face a much
higher probability of corruption and possible contention at
each hop. Finally, energy is invested in a packet at each
hop, so the cost of dropping a packet varies with packet and
place. All of these factors make it difficult to achieve fair-
ness through the simple, local algorithms that the platform
naturally support. At the very least, fairness is at odds with
both energy efficiency and high channel utilization. The eas-
iest way to reduce energy and fill the channel is to only take
traffic from the nodes adjacent to the base station, but this
hardly provides a valid sample of the overall sensor field.
In short, the characteristics and goals of MAC in sensor net-
works differ strongly from conventional computer networks.
They are dominantly periodic and highly correlated traf-
fic, comprising small packets flowing to base stations in a
deep, irregular multihop network with each node seeking to
achieve a fair bandwidth allocation to the base station in an
energy efficient way over single channel radios.
In this paper we make progress on addressing the array of
design tradeoffs for sensor networks by developing an inno-
vative MAC protocol and adaptive transmission rate control
scheme for multihop networks in the context of a simplified
Figure 1: Our low-power networked sensor device
prototype.
application scenario on a real low-power networked sensor,
as well as in simulation. Section 2 describes the hardware
and software platform of our networked sensor development
and MAC design. It describes the application scenario, lays
out our metrics of fairness and efficiency, and introduces
the simulation environment that we use in evaluating our
design. Section 3 examines related work on existing MAC
protocols and identifies shortcomings relative to the sensor
network challenges. Section 4 outlines our proposed MAC
and transmission control scheme along with a set of open
engineering issues, including the mechanism for carrier sens-
ing, desynchronizing periodic behavior, and backoff scheme
for the MAC. For the multihop scenario, we study both the
conventional RTS/CTS contention based scheme and a sim-
ple adaptive rate control algorithm. Section 5 presents our
analysis and evaluation of the different carrier sense mul-
tiple access (CSMA) techniques that we study. We con-
clude that limiting the length of listening, the introduction
of random delay in addition to backoff, and phase shift at
the application level are necessary. Section 6 compares our
adaptive rate control scheme with a conventional contention
control scheme in a multihop network scenario. We find that
the adaptive rate control mechanism is the most effective in
achieving our fairness goal while being energy efficient for
both low and high duty cycle of network traffic.
2. SENSOR NETWORK DESIGN POINT
Our study is grounded in the small, low-power networked
sensor device shown in Figure 1 [7]. We believe it be rep-
resentative of the constraints of limited computation power,
storage, and energy supply of the tiny devices that will
be deployed into the future [10]. The processor is an AT-
MEL [3] 4MHz, 8 bit micro-controller with 8K bytes of pro-
gram memory and 512 bytes of data memory. The radio is
a single channel RF transceiver operating at 916MHz and
capable of transmitting at 10kbps using on-off-keying en-
coding. Each radio transition and bit sampling is performed
in software. There is no facility for collision detection. A
heterogeneous set of sensors such as light, temperature, hu-
midity, pressure, acceleration, and magnetic field can be in-
tegrated into these prototypes.
2.1 Networking Component Stack
222

Figure 2: Complete TinyOS application component
graph.
TinyOS [7] is an event-based operating system for these de-
vices that provides fine-grained interleaving of event pro-
cessing and tasks from multiple system components. The
complete TinyOS application for our study is shown in Fig-
ure 2. There is a component providing an asynchronous
interface to each sensor and a stack of components to imple-
ment networking over the radio. The lowest layer transmits
or receives bytes bit-by-bit over the radio. It provides phase
and rate controls to lock on to the packet start symbol and
then to spool bits. At this level, the interface is half-duplex
- the radio is receiving except during packet transmission.
The packet-level component is responsible for spooling in-
coming bytes and delivering the packet receive event. It is
where the media access control mechanisms for transmit re-
side. (It also performs the encoding and decoding of the
byte stream onto the link and error checking: Manchester
encoding with an 16-bit CRC.) Packets are short and of a
fixed size, typically 30 bytes including an one byte destina-
tion field, an one byte handler field, and an application data
unit.
The Active Message component delivers tagged packet events
to application level components. Here we have two such
components. The sensor component periodically receives a
clock event, acquires sensor data, and transmits the data
toward a base station over the multihop network. The other
component is responsible for building the dynamic multi-
hop network and routing traffic. A simple beacon-based
discovery protocol maintains a breadth-first spanning tree,
such that each node knows a “parent node” closer to the
base station. Originating sensor packets are marked for the
parent. (All other nodes discard them.) At each hop, the
multihop component receives a packet and retransmits it
to the upstream level. In general, this component might
perform aggregation or statistical analysis. However, we re-
strict ourselves to the case where it forwards all data to the
infrastructure for analysis, as this focuses the work on the
media access and transmission control aspects. This compo-
nent does collect statistics on the number of nodes routing
through it. The only buffering in the system is a fixed num-
ber of small packet buffers at the application level, one of
which is used for the asynchronous receive. Thus, if the
radio is busy transmitting or receiving when a packet send
is requested, the request will fail back up to the applica-
tion component. Once the packet component has accepted
a packet for transmission, it will work on it until it acquires
the channel and transmits it. Thus, the transmission rate
control is implemented within the two application compo-
nents.
2.2 Metric for Evaluation
The metrics for evaluation of a sensor network MAC pro-
tocol stress both fairness and energy efficiency. A fair al-
location of bandwidth delivered to the base station from
each node over multiple hops is desirable. Although aggre-
gate bandwidth is an important metric in evaluating MAC
protocols, it can be misleading. High aggregate bandwidth
of packets only from nodes around the base station is not
desirable.
We evaluate energy efficiency in terms of energy consumed
per unit of successful communication or packets received by
the base station. The energy consumed is the total energy
that the network has invested in propagating data to the
base station. A scenario with nodes near the base station
filling the channel will perform poorly under this metric.
The total energy invested by the network includes energy
spent in listening for the channel and all packet transmis-
sions and forwarding. Our experiment is designed such that
all nodes are in receive mode even during idle period. There-
fore, we only account useful work as energy spent in channel
listening and packet transmission under this metric.
2.3 Simulation Environment
Given the difficulty in performing actual measurements in
wireless networking, we first evaluate our system through
simulation. We have created a simple simulator capable of
creating an arbitrary multihop network topology of a group
of networked sensors. Each UNIX process represents a net-
worked sensor, and a master process is responsible for syn-
chronizing them to perform bit time simulation. There is a
simple radio propagation model in the simulation and we as-
sume bit error rate to be zero, since our main focus is media
access control. We use a simple reachability table that spec-
ifies whether bidirectional connectivity exists between nodes
in the simulated network. The simulator doesn’t simulate
the actual hardware operating in the TinyOS environment.
However, it preserves the event driven semantics and the dy-
namics of traffic flow shown in Figure 2. All the simulations
presented in Section 5 and 6 are collected in this simulation
environment.
3. RELATED WORK
In designing the MAC protocol, we first examine existing
mechanisms and determine whether or not they apply in
the regime of sensor networks.
Many variations of the CSMA [8, 17] strategy can be found
in the literature. Listening to the channel before transmis-
sion to exploit information about other users is a very com-
mon approach found in almost all CSMA schemes, except
pure ALOHA [15]. Another approach is to use explicit posi-
tive or negative acknowledgments to signal collision and per-
form necessary random delay before retransmit. ALOHA
takes this approach. IEEE 802.11 [2] uses it in addition
to listening. Other CSMA schemes rely on time synchro-
nized slotted channel, such as Slotted ALOHA [11]. Finally,
223

CSMA with collision detection [9], which is widely used,
including wired Ethernet. Though many CSMA schemes
exist, they all lean toward a fundamental assumption that
packet transmissions occur with a stochastic distribution,
that is very different from the correlated traffic found in
sensor networks. Furthermore, they aim to support many
independent point-to-point flows while the network in this
new regime tends to operate as a collective structure. As a
result, re-exploring CSMA strategies with a different funda-
mental assumption is extremely relevant.
Energy consumption is an important metric in the design
of media access control. PAMAS [13, 14] is a power aware
media access protocol which powers off radio when not ac-
tively transmitting or receiving packets. Our work does not
explore this power scheduling aspect. Instead, we focus on
the energy efficiency in basic media access control schemes,
and the overall bandwidth, energy, and fairness tradeoff in
a multihop network.
The IEEE 802.11 [2] standard aims to provide a wireless
Ethernet illusion. The design is based on an assumption of
a single cell scenario, with mobile stations always in range
of at least one base station, with a hand-off when migrating
from one cell to another. As a result, there is no multihop
scenario. Furthermore, the large transmission range and the
spread spectrum capability of the radio make it feasible to
create a grid of overlapping cells to cover a field without need
for taking several hops to reach the base station. The ad hoc
aspect of the protocol assumes peer-to-peer communications
rather than many-to-one data propagation scenario as found
in sensor network. Nevertheless, the primary mechanisms of
carrier sensing and contention control scheme are fundamen-
tals that we will study and evaluate.
Bluetooth [1] is an emerging standard for many future wire-
less devices. Its usage model is to create a “wireless cable”
illusion for applications like connecting a cellular phone or
speaker to a notebook computer with voice or data stream-
ing among these devices. The primary media access control
is a centralized Time Division Multiple Access (TDMA) pro-
tocol within a piconet which is a relatively static ad hoc
network supporting a small number of nodes within a sin-
gle cell. Overlapping cells is also feasible due to the spread
spectrum radios. Bluetooth assumes no multihop scenario.
The centralized TDMA protocol and the tight requirement
of time synchronization between each node in the piconet
make it inappropriate for sensor networks.
MACAW [4] shares a single channel radio similar to ours,
but its main focus is for single hop base station interaction
within a cell. In fact, in one of its communication scenar-
ios resembling a hidden node problem in a multihop net-
work, it explicitly states that the scenario cannot be solved
by contention control protocols unless time synchronization
information is present for contention period to be known.
Nevertheless, we will study the most primitive contention
control protocol in the context of multihop network, and in
Section 4, we will discuss how hidden node problem can be
addressed.
In large, dense packet radio networks [12], a collision free
channel access scheme is accomplished using locally gener-
ated and published transmit and receive schedules. A sta-
tion can send if its sending slot overlaps with the receiver’s
receiving slot and the amount of overlap is long enough
for data transmission. To prevent synchronization between
neighboring stations, they produce random or pseudo-random
schedules. This is a distributed TDMA scheme with the
assumption that the network topology is relatively static.
Given the spontaneous nature of sensor networks where the
receiver or the next hop to the base station may change at
any time, such a scheme may be very inefficient in publishing
different schedules.
Prior work shows that an adaptive rate control algorithm
is very effective in achieving proportional fairness of media
access for packet radio [16]. Their design goal is to have a
fair sharing of the channel among local competing neighbors.
Our proposed adaptive rate control scheme builds upon this
work, but our goal is to have media access control assist
in achieving fair bandwidth delivery to the base station for
nodes in a multihop network.
Our adaptive rate control uses loss as collision signal to ad-
just transmission rate in a manner similar to the congestion
control used in TCP [5, 6]. While TCP’s congestion control
is end-to-end over a network with many independent flows,
our proposed adaptive rate control works collectively at ev-
ery node in the network, since each node is both a router
and a sender, and routing is done at the application level.
4. DESIGN
We discuss how media access control for sensor network
should be done differently in this section. First, we ex-
plore what type of listening mechanism is appropriate for
the case where all nodes can hear each other. Second, we
discuss how backoff should be implemented in a sensor net-
work. Third, we present two mechanisms which we will
study their effectiveness in the context of a multihop net-
work. The first scheme is a conventional RTS/CTS con-
tention control scheme, and the second one is our proposed
adaptive transmission control scheme, and finally, a mech-
anism that all schemes can leverage off for avoiding some
cases of hidden node problems in multihop network.
4.1 Listening Mechanism
Carrier Sense Multiple Access (CSMA) and the Collision
Detection (CD) scheme found in Ethernet are examples of
listening mechanisms. Listening is very effective when all
nodes can hear each other, (i.e. without hidden nodes).
Unfortunately, collision detection is not possible in wireless
network technology without additional circuitry. Though
listening is simple, it does come with an energy cost, because
the radio must be on to listen. To conserve energy, it is
important to shorten the length of carrier sensing. Many
protocols such as IEEE 802.11 require sensing the channel
even during backoff. However, CSMA for sensor networks
should take this opportunity to turn the radio off.
The highly synchronized nature of the traffic imposes a new
criteria for CSMA. Given there are no hardware mechanism
for detecting collisions, nodes that happen to send at the
same time will corrupt each other. If the traffic pattern of
each node is independent, this situation is not likely to re-
peat. However, detection of one common physical event will
224

synchronize these nodes and lead them to send at the same
time, which repeats periodically. The result is no packet
transfer at all. The solution is to introduce random delay
for transmission to unsynchronize the nodes. Section 5 will
discuss and evaluate various ways in introducing random-
ness for CSMA.
4.2 Backoff Mechanism
Backoff is a widely used mechanism in media access control
to reduce contention. The idea of backoff is to restrain a
node from accessing the channel for a period of time and
hopefully, the channel will become free after the backoff pe-
riod. In the case of sensor networks where the traffic is a
superposition of different periodic streams, backoff should
not just restrain a node from sending for the backoff period.
In fact, the backoff period should be applied as a phase shift
to the periodicity of the application so that synchronization
among periodic streams of traffic can be broken.
4.3 Contention Based Mechanism
Explicit contention control schemes, which are widely used
in many MAC protocols, e.g., IEEE 802.11 [2] and MACAW
[4], require the use of control packets, such as Request to
Send (RTS) and Clear to Send (CTS). Acknowledgments
(ACKs) serve a different purpose in IEEE 802.11; they in-
dicate lack of collision. For computer networks where pack-
ets are large, these small control packets impose very little
overhead. However, for sensor networks where packet size
is small, they can constitute a large overhead. A RTS-CTS-
DATA-ACK handshake series in transmitting a packet can
constitute up to 40% overhead in our platform. (Each con-
trol packet is 3 bytes long (type,destination,source) and the
packet is 30 bytes long.) This can be extremely costly, since
energy has to be spent in CSMA, transmitting, and receiv-
ing each control packet. One advantage of a bidirectional
multihop network is that acknowledgments are free when
the receiving node (your parent in the multihop topology)
routes the packet to its parent. This eliminates an explicit
ACK control packet. If the receiver performs some kind of
application specific aggregation before routing the packet,
the originator of the packet may still be capable of detect-
ing the success of the transmission.
A contention control scheme for sensor networks should use
a minimum number of control packets. The most basic types
are RTS and CTS. Though it may be effective in solving the
hidden node problem in a multihop network, such a scheme
should only be used if the amount of traffic is high while
a simple CSMA scheme is actually adequate for low traffic
since the probability of corruption due to collision is very
small.
For the contention scheme that we study in Section 6, only
RTS and CTS packets are used for handshakes. A node
wishing to transmit first sends a RTS packet to its parent
and waits for a CTS reply. If no CTS is received for a
timeout period (2 CTS packet times), the node will enter
backoff with a binary exponential increasing backoff window.
Similarly, if it receives a CTS not destined to it, it will also go
into backoff. If no CTS has been received after five retries,
the transmission will be dropped. Furthermore, if a node
hears a CTS before any of its own transmission, it will defer
transmission for one packet time to avoid corrupting the
traffic.
4.4 Rate Control Mechanism
The tension between originating traffic and route-thru traf-
fic has a direct impact in achieving our fairness goal. Media
access control must assist in balancing this tension for the
channel. Specifically, the MAC should control the rate of
originating data of a node in order to allow route-thru traffic
to access the channel and reach the base station. Similarly,
some kind of progressive signalling mechanism should exist
for route-thru traffic, such that back pressure can propagate
deep down into the network for those nodes to lower their
rate of originating data. This in turn will decrease the ag-
gregate route-thru traffic and open up the channel for nodes
closer to the base station to originate data. Such a rate con-
trol mechanism should only use distributed local algorithms.
We propose an implicit mechanism which passively adapts
the rate of transmission of both original and route-thru traf-
fic without the use of any MAC control packets.
The adaptive rate control idea is very simple and can be
explained with an analogy of metering traffic onto a freeway
where the route-thru traffic is like traffic on the freeway and
each node originating data is like cars trying to enter. Peri-
odically, a node attempts to inject a packet. If the packet is
successfully injected, it becomes part of the route-thru traf-
fic. As it is routed by the node’s parent, it signals that the
road still has capacity for more traffic and thus, the node
can increase its transmission rate. However, if the injection
of the packet wasn’t successful, it signals that the road is
jammed and the node decreases its rate of originating data
and backoff to achieve a phase change effect.
The above explains how the originating data rate adapts to
the route-thru traffic. Route-thru traffic will adapt to the
traffic of original data using a similar mechanism. If a node
injects lots of original traffic into the freeway, the route-thru
traffic will be hindered and thus, the rate of transmitting
route-thru traffic will decrease (cars at the back have to slow
down for cars in front once they are on the bridge). It has a
domino effect in propagating this back pressure deep down
into the network which ultimately decreases the amount of
aggregate route-thru traffic.
The metering effect discussed in the analogy above can be
set by a global schedule. Given that each node has an
omniscient knowledge of the total number of nodes N in
the entire network, each node can meter its own rate by
ChannelCapacity/N to achieve our fairness goal. However,
the spontaneous ad hoc nature of sensor networks make such
a global knowledge impractical. Therefore, we propose an
adaptive scheme attempting to approximate it.
Our rate control mechanism uses a linear increase and multi-
plicative decrease approach to control the transmission rate
of the application. Given the application transmission rate is
S, the actual rate of originating data is S ∗ p where p ∈ [0, 1].
This rate control is probabilistic, where p is the probability
of transmission. To linearly increase the rate, simply incre-
ment p by a constant α. To multiplicatively decrease the
rate, multiply p by a factor β where 0 < β < 1. This proba-
bilistic mechanism of rate control is based on the work done
225

Frequently asked questions

Q1. What have the authors contributed in "A transmission control scheme for media access in sensor networks" ?

The authors study the problem of media access control in the novel regime of sensor networks, where unique application behavior and tight constraints in computation power, storage, energy resources, and radio technology have shaped this design space to be very different from that found in traditional mobile computing regime. The authors propose an adaptive rate control mechanism aiming to support these two goals and find that such a scheme is most effective in achieving their fairness goal while being energy efficient for both low and high duty cycle of network traffic. 

Q2. What are the future works in "A transmission control scheme for media access in sensor networks" ?

These simulation results are further supported by real implementation on their tiny network sensor platform. 

Q3. What is the only buffering in the system?

The only buffering in the system is a fixed number of small packet buffers at the application level, one of which is used for the asynchronous receive. 

Q4. What is the main cause of the discrepancy between simulation and empirical measurements?

The reality of cell overlaps and interference in the multihop scenario is probably the main cause of the discrepancy between simulation and empirical measurements. 

Q5. What is the effect of backoff on the fairness of the network?

Their results suggest that backoff with a fixed windowsize or binary exponential decrease in window size are effective in maintaining proportional fairness. 

Q6. What is the ad hoc aspect of the protocol?

The ad hoc aspect of the protocol assumes peer-to-peer communications rather than many-to-one data propagation scenario as found in sensor network. 

Q7. What are the common MAC protocols?

Explicit contention control schemes, which are widely used in many MAC protocols, e.g., IEEE 802.11 [2] and MACAW [4], require the use of control packets, such as Request to Send (RTS) and Clear to Send (CTS). 

Q8. What is the power aware media access protocol?

PAMAS [13, 14] is a power aware media access protocol which powers off radio when not actively transmitting or receiving packets. 

Q9. What makes it inappropriate for sensor networks?

The centralized TDMA protocol and the tight requirement of time synchronization between each node in the piconet make it inappropriate for sensor networks. 

Q10. How does the proposed mechanism achieve fairness?

Simulation have shown that their proposed mechanism is effective in achieving fairness while maintaining good aggregate bandwidth with reasonable energy efficiency. 

Q11. What is the protocol for a simple beacon-based discovery?

A simple beacon-based discovery protocol maintains a breadth-first spanning tree, such that each node knows a “parent node” closer to the base station. 

Q12. What is the effect of backoff mechanism on proportional fairness?

The result suggests that backoff mechanism has an effect on proportional fairness, with binary exponential increasing backoff being the worst. 

Q13. What is the goal of the proposed adaptive rate control scheme?

Their proposed adaptive rate control scheme builds upon this work, but their goal is to have media access control assist in achieving fair bandwidth delivery to the base station for nodes in a multihop network. 

Q14. What is the reason why MACAW has suggested a scenario where multihop hidden node?

as discussed in Section 3, MACAW has suggested a scenario where multihop hidden node problem cannot be solved due to lack of synchronized information of knowing when is the contention period between a child node and its grandparent’s node. 

Q15. How does the graph show the average rate of traffic delivered?

Figure 15 shows that nodes below first level of the tree achieve about 0.2 packet/s of delivered bandwidth, or about 25% of the ideal rate that saturates the bottleneck. 

Q16. What is the easiest way to reduce energy and fill the channel?

The easiest way to reduce energy and fill the channel is to only take traffic from the nodes adjacent to the base station, but this hardly provides a valid sample of the overall sensor field. 

Q17. What is the key requirement to share the channel fairly in an individual cell?

It is not sufficient to share the channel fairly in an individual cell, the authors would like to achieve a crude level of end-to-end fairness even in a deep and self-organized multihop networks, which may change dynamically and originate data at each intermediate node.