System Architecture for Wireless Sensor Networks

by
Jason Lester Hill
B.S. (University of California, Berkeley) 1998
M.S. (University of California, Berkeley) 2000
A dissertation submitted in partial satisfaction of the
requirements for the degree of
Doctor of Philosophy
in
Computer Science
in the
GRADUATE DIVISION
of the
UNIVERISY OF CALIFORNIA, BERKELEY
Committee in charge:
Professor David E. Culler, Chair
Professor Kris Pister
Professor Paul Wright
Spring 2003
The dissertation of Jason Lester Hill is approved:
Chair Date
Date
Date
University of California, Berkeley
Spring 2003
System Architecture for Wireless Sensor Networks
Copyright 2003
by
Jason Lester Hill
1
Abstract
System Architecture for Wireless Sensor Networks
by
Jason Lester Hill
Doctor of Philosophy in Computer Science
University of California at Berkeley
Professor David Culler, Chair
In this thesis we present and operating system and three generations of a hardware
platform designed to address the needs of wireless sensor networks. Our operating
system, called TinyOS uses an event based execution model to provide support for finegrained
concurrency and incorporates a highly efficient component model. TinyOS
enables us to use a hardware architecture that has a single processor time shared between
both application and protocol processing. We show how a virtual partitioning of
computational resources not only leads to efficient resource utilization but allows for a
rich interface between application and protocol processing. This rich interface, in turn,
allows developers to exploit application specific communication protocols that
significantly improve system performance.
The hardware platforms we develop are used to validate a generalized architecture
that is technology independent. Our general architecture contains a single central
controller that performs both application and protocol-level processing. For flexibility,
this controller is directly connected to the RF transceiver. For efficiency, the controller is
2
supported by a collection of hardware accelerators that provide basic communication
primitives that can be flexibility composed into application specific protocols.
The three hardware platforms we present are instances of this general architecture
with varying degrees of hardware sophistication. The Rene platform serves as a baseline
and does not contain any hardware accelerators. It allows us to develop the TinyOS
operating system concepts and refine its concurrency mechanisms. The Mica node
incorporates hardware accelerators that improve communication rates and
synchronization accuracy within the constraints of current microcontrollers. As an
approximation of our general architecture, we use Mica to validate the underlying
architectural principles. The Mica platform has become the foundation for hundreds of
wireless sensor network research efforts around the world. It has been sold to more than
250 organizations.
Spec is the most advanced node presented and represents the full realization of
our general architecture. It is a 2.5 mm x 2.5 mm CMOS chip that includes processing,
storage, wireless communications and hardware accelerators. We show how the careful
selection of the correct accelerators can lead to orders-of-magnitude improvements in
efficiency without sacrificing flexibility. In addition to performing a theoretical analysis
on the strengths of our architecture, we demonstrate its capabilities through a collection
of real-world application deployments.
_______________________________________
Professor David Culler
Dissertation Committee Chair
i
Table of Contents
Table of Contents................................................................................................................. i
List of Figures:................................................................................................................... iv
Chapter 1: Introduction................................................................................................. 1
Chapter 2: Wireless Sensor Networks ........................................................................ 10
2.1 Sensor network application classes....................................................................... 11
2.1.1 Environmental Data Collection................................................................. 11
2.1.2 Security Monitoring.................................................................................. 14
2.1.3 Node tracking scenarios............................................................................ 16
2.1.4 Hybrid networks........................................................................................ 17
2.2 System Evaluation Metrics ................................................................................... 17
2.2.1 Lifetime..................................................................................................... 18
2.2.2 Coverage ................................................................................................... 19
2.2.3 Cost and ease of deployment .................................................................... 20
2.2.4 Response Time.......................................................................................... 22
2.2.5 Temporal Accuracy................................................................................... 22
2.2.6 Security ..................................................................................................... 23
2.2.7 Effective Sample Rate............................................................................... 24
2.3 Individual node evaluation metrics....................................................................... 25
2.3.1 Power ........................................................................................................ 26
2.3.2 Flexibility.................................................................................................. 26
2.3.3 Robustness ................................................................................................ 27
2.3.4 Security ..................................................................................................... 28
2.3.5 Communication......................................................................................... 28
2.3.6 Computation.............................................................................................. 29
2.3.7 Time Synchronization............................................................................... 30
2.3.8 Size & Cost ............................................................................................... 31
2.4 Hardware Capabilities........................................................................................... 31
2.4.1 Energy....................................................................................................... 32
2.4.2 Radios ....................................................................................................... 37
2.4.3 Processor ................................................................................................... 43
2.4.4 Sensors ...................................................................................................... 48
2.5 Rene Design Point................................................................................................. 50
2.5.2 Baseline Power Characteristics................................................................. 53
2.6 Refined Problem Statement .................................................................................. 54
Chapter 3: Software Architecture for Wireless Sensors ............................................. 56
3.1 Tiny Microthreading Operating System (TinyOS) ............................................... 57
3.2 TinyOS Execution Model ..................................................................................... 58
3.2.1 Event Based Programming ....................................................................... 58
3.2.2 Tasks ......................................................................................................... 59
3.2.3 Atomicity .................................................................................................. 60
ii
3.3 TinyOS Component Model................................................................................... 60
3.3.1 Component Types ..................................................................................... 63
3.3.2 Enabling the migration of hardware/software boundary .......................... 65
3.3.3 Example Components ............................................................................... 66
3.3.4 Component Composition .......................................................................... 67
3.3.5 Application Walk Through ....................................................................... 70
3.4 AM Communication Paradigm............................................................................. 72
3.4.1 Active Messages Overview....................................................................... 72
3.4.2 Tiny Active Messages Implementation .................................................... 73
3.4.3 Buffer swapping memory management .................................................... 75
3.4.4 Explicit Acknowledgement....................................................................... 76
3.5 Components Contained in TinyOS ....................................................................... 77
3.6 TinyOS Model Evaluation .................................................................................... 78
3.7 TinyOS Summary ................................................................................................. 80
Chapter 4: Generalized Wireless Sensor Node Architecture...................................... 82
4.1 Wireless communication requirements................................................................. 82
4.2 Key issues architecture must address.................................................................... 85
4.2.1 Concurrency.............................................................................................. 86
4.2.2 Flexibility.................................................................................................. 86
4.2.3 Synchronization ........................................................................................ 87
4.2.4 Decoupling between RF and processing speed......................................... 88
4.3 Traditional Wireless Design ................................................................................. 91
4.4 Generalized architecture for a wireless sensor node............................................. 93
4.5 Architectural Benefits ........................................................................................... 96
4.5.1 Concurrency.............................................................................................. 96
4.5.2 Protocol Flexibility and Timing Accuracy ............................................. 100
4.6 Summary............................................................................................................. 100
Chapter 5: Approximation of General Architecture: Mica...................................... 102
5.1 Mica design......................................................................................................... 103
5.1.1 Block diagram overview......................................................................... 103
5.1.2 Raw radio interface ................................................................................. 106
5.2 Communication accelerators............................................................................... 107
5.3 Evaluation ........................................................................................................... 109
5.3.1 Concurrency Management ...................................................................... 110
5.3.2 Interplay between RF and Data path speed............................................. 111
5.3.3 Interface flexibility.................................................................................. 112
5.4 Blue: a follow-on to Mica ................................................................................... 120
5.4.1 CPU......................................................................................................... 120
5.4.2 Radio....................................................................................................... 121
5.5 Summary............................................................................................................. 122
Chapter 6: Integrated Architecture for Wireless Sensor Nodes – Spec.................... 124
6.1.1 High Level Overview.............................................................................. 125
6.1.2 General Block Diagram .......................................................................... 126
6.1.3 Radio Back End & ADC Sub blocks ...................................................... 134
6.1.4 Digital frequency lock registers .............................................................. 135
6.1.5 Physical Characteristics .......................................................................... 136
iii
6.2 Performance ........................................................................................................ 138
6.2.1 Start Symbol Detection........................................................................... 139
6.2.2 Interrupt Handling Overhead .................................................................. 140
6.2.3 Program Memory Management .............................................................. 141
6.2.4 Timing Extraction ................................................................................... 142
6.2.5 Encryption Support ................................................................................. 143
6.2.6 Memory I/O and serialization ................................................................. 144
6.2.7 Primitives not included ........................................................................... 145
6.3 Cost of flexibility................................................................................................ 146
6.4 Summary............................................................................................................. 147
Chapter 7: Demonstration Applications and Performance ....................................... 149
7.1 Environmental Data Monitoring......................................................................... 149
7.2 Empirical analysis of performance improvements ............................................. 150
7.2.1 Hardware Trial ........................................................................................ 153
7.3 29 Palms.............................................................................................................. 154
7.3.1 Application Description.......................................................................... 154
7.3.2 Key sub components/application blocks................................................. 156
7.3.3 Importance of in-network processing ..................................................... 159
7.4 Z-Car Tracking.................................................................................................... 159
7.4.1 Application Description.......................................................................... 160
7.5 Conclusion .......................................................................................................... 165
Chapter 8: Related Work .......................................................................................... 166
8.1 TinyOS Supported Work .................................................................................... 166
8.1.1 Wireless Robots ...................................................................................... 166
8.1.2 Media Access Control and Routing ........................................................ 168
8.1.3 Time Synchronization............................................................................. 168
8.1.4 Multi-Hop Routing optimization ............................................................ 169
8.1.5 TinyDB ................................................................................................... 170
8.2 Wireless Platforms .............................................................................................. 171
8.2.1 Smart Dust .............................................................................................. 171
8.2.2 Bluetooth................................................................................................. 172
8.2.3 Zigbee (802.15.4).................................................................................... 173
8.2.4 Pico Radio............................................................................................... 174
8.2.5 Chipcon CC1010..................................................................................... 174
8.3 Embedded Operating Systems ............................................................................ 175
Chapter 9: Conclusions............................................................................................. 177
Bibliography ................................................................................................................... 182
iv
List of Figures:
Figure 1-1: DOT – Wireless sensor network device designed to be the approximate size of a
quarter. Future devices will continue to be smaller, cheaper and longer lasting. 2
Figure 1-2: Possible deployment of ac-hoc wireless embedded network for precision
agriculture. Sensors detect temperature, light levels and soil moisture at hundreds of
points across a field and communicate their data over a multi-hop network for analysis. 4
Figure 1-3: Design lineage of Mote Technology. COTS (Common off the shelf) prototypes
lead to the weC platform. Rene then evolved to allow sensor expansion and enabled
hundreds of compelling applications. The Dot node was architecturally the same as
Rene but shrunk into a quarter-sized device. Mica – discussed in depth in this thesis –
made significant architectural improvements in order to increase performance and
efficiency. Spec represents the complete integrated CMOS vision. 7
Figure 2-1: Battery characteristics for Lithium, Alkaline and NiMH batteries. The discharge
characteristics of alkaline batteries make it essential to design a system to tolerate a
wide range of input voltages. 33
Figure 2-2: Power consumption and capabilities of commonly available. 48
Figure 2-3: Pictures of the Rene wireless sensor network platform. 52
Figure 3-1: Component graph for a multi-hop sensing application. 61
Figure 3-2: AM Messaging component graphical representation. 64
Figure 3-3: NESC component file describing the external interface to the AMStandard
messaging component. 66
Figure 3-4: BlinkM.nc application component that blinks the system LED's once per second. 68
Figure 3-5: NESC application configuration file that wires together the Blink application. 69
Figure 3-6: Listing of components contained in TinyOS 1.0. Application developers select
from this collection of system-level components to create the desired application
attributes. 77
Figure 3-7: Detailed breakdown of work distribution and energy consumption across each
layer on the Rene node. 80
Figure 4-1: Phases of wireless communication for transmission and reception. 83
Figure 4-2: Energy cost per bit when transmitting at 10Kbps and 50Kbps broken down into
controller and RF energy. 89
Figure 4-3: Generalized Architecture for embedded wireless device. 94
Figure 4-4: Breakdown of CPU usage for start symbol detection overhead broken down into
search time, overhead, and CPU time for applications. For the dedicated controller case
the breakdown is included with a realistic 10% chip-to-chip overhead estimate and with
a best-case 0% overhead estimate. 98
Figure 5-1: Block diagram of Mica architecture. The direct connection between application
controller and transceiver allows the Mica node to be highly flexible to application
demands. Hardware accelerations optionally assist in communication protocols. 104
Figure 5-2: Breakdown of active and idle power consumption for Mica hardware components
at 3V. 105
Figure 5-3: CPU utilization of start symbol detection. Optimizations enabled by the Mica
node significantly increase CPU efficiency. 110
Figure 5-4: Equations for determining the power consumption of a sleeping network. 113
Figure 5-5: Follow-on to the Mica node, blue increase range and reduces power consumption.
Blue nodes have a 4-6x longer battery life over Mica on alkaline batteries. 121
Figure 6-1: The single chip Spec node pictured next to a ballpoint pen. 124
Figure 6-2: Block Diagram of Spec, the single chip wireless mote. 127
Figure 6-3: Design layout for single chip mote. Large central block contains CPU, timers and
hardware accelerators. Across the top are 6 memory banks. The ADC is center left.
The radio is in the bottom right corner. 136
Figure 6-4: Area breakdown of digital logic modules included on Spec. Total area of wiring
and gates using National Semiconductor standard cells in .25 um CMOS. 137
v
Figure 6-5: Overview of the performance improvements achieved by the Spec node’s
hardware accelerators. 146
Figure 7-1: Theoretical node energy consumption when performing environmental data
monitoring where data is collected every 4 seconds, averaged and transmitted once
every 5 minutes. Each node has a maximum of 5 children nodes and one parent node.
The table shows results for Rene, Mica, Blue, Spec. The ratio between Rene power
consumption and blue power consumption is also shown. The alarm check cost of blue
is larger than that of Mica due to the CC1000 having a longer turn-on time. The
Sensing cost of blue is reduced by a low power ADC in the microcontroller. 151
Figure 7-2: 29 palms application. Motes dropped out of an airplane self-assemble onto an adhoc
network in order to monitor for vehicle activity at a remote desert location. 154
Figure 7-3: Data filtering performed locally on modes to amplify and extract vehicle signal.
Time stamp assigned to the highest peak of the red activity line. 157
Figure 7-4: Position estimates of each of the sensing node are updated in response to the result
of the linear regression that estimates the passing vehicle’s velocity. 158
Figure 7-5: Picture of z-car tracking application deployment. Remote control car tracked as it
passed through the field of sensors. 160
Figure 7-6: Multi-hip network routing used in z-car tracking application. Nodes detecting the
event (in green) transmit data to a single leader that transits a single report packet. The
report packet travels across the network (dotted line) until it reaches the base node. The
packet contains the estimate of the evader’s (red) position. 164
Figure 8-1: Characteristics and requirements of commercial embedded real-time operating
systems. 175
1
Chapter 1: Introduction
The emerging field of wireless sensor networks combines sensing, computation,
and communication into a single tiny device. Through advanced mesh networking
protocols, these devices form a sea of connectivity that extends the reach of cyberspace
out into the physical world. As water flows to fill every room of a submerged ship, the
mesh networking connectivity will seek out and exploit any possible communication path
by hopping data from node to node in search of its destination. While the capabilities of
any single device are minimal, the composition of hundreds of devices offers radical new
technological possibilities.
The power of wireless sensor networks lies in the ability to deploy large numbers
of tiny nodes that assemble and configure themselves. Usage scenarios for these devices
range from real-time tracking, to monitoring of environmental conditions, to ubiquitous
computing environments, to in situ monitoring of the health of structures or equipment.
While often referred to as wireless sensor networks, they can also control actuators that
extend control from cyberspace into the physical world.
The most straightforward application of wireless sensor network technology is to
monitor remote environments for low frequency data trends. For example, a chemical
plant could be easily monitored for leaks by hundreds of sensors that automatically form
a wireless interconnection network and immediately report the detection of any chemical
leaks. Unlike traditional wired systems, deployment costs would be minimal. Instead of
having to deploy thousands of feet of wire routed through protective conduit, installers
2
simply have to place quarter-sized device, such as the one pictured in Figure 1-1, at each
sensing point. The network could be incrementally extended by simply adding more
devices – no rework or complex configuration. With the devices presented in this thesis,
the system would be capable of monitoring for anomalies for several years on a single set
of batteries.
In addition to drastically reducing the installation costs, wireless sensor networks
have the ability to dynamically adapt to changing environments. Adaptation
mechanisms can respond to changes in network topologies or can cause the network to
shift between drastically different modes of operation. For example, the same embedded
network performing leak monitoring in a chemical factory might be reconfigured into a
network designed to localize the source of a leak and track the diffusion of poisonous
gases. The network could then direct workers to the safest path for emergency
evacuation.
Current wireless systems only scratch the surface of possibilities emerging from
the integration of low-power communication, sensing, energy storage, and computation.
Figure 1-1: DOT – Wireless sensor network device
designed to be the approximate size of a quarter.
Future devices will continue to be smaller, cheaper and
longer lasting.
3
Generally, when people consider wireless devices they think of items such as cell phones,
personal digital assistants, or laptops with 802.11. These items costs hundreds of dollars,
target specialized applications, and rely on the pre-deployment of extensive infrastructure
support. In contrast, wireless sensor networks use small, low-cost embedded devices for
a wide range of applications and do not rely on any pre-existing infrastructure. The
vision is that these devise will cost less that $1 by 2005.
Unlike traditional wireless devices, wireless sensor nodes do not need to
communicate directly with the nearest high-power control tower or base station, but only
with their local peers. Instead, of relying on a pre-deployed infrastructure, each
individual sensor or actuator becomes part of the overall infrastructure. Peer-to-peer
networking protocols provide a mesh-like interconnect to shuttle data between the
thousands of tiny embedded devices in a multi-hop fashion. The flexible mesh
architectures envisioned dynamically adapt to support introduction of new nodes or
expand to cover a larger geographic region. Additionally, the system can automatically
adapt to compensate for node failures.
The vision of mesh networking is based on strength in numbers. Unlike cell
phone systems that deny service when too many phones are active in a small area, the
interconnection of a wireless sensor network only grows stronger as nodes are added. As
long as there is sufficient density, a single network of nodes can grow to cover limitless
area. With each node having a communication range of 50 meters and costing less that
$1 a sensor network that encircled the equator of the earth will cost less that $1M.
4
An example network is shown in Figure 1-2. It depicts a precision agriculture
deployment—an active area of application research. Hundreds of nodes scattered
throughout a field assemble together, establish a routing topology, and transmit data back
to a collection point. The application demands for robust, scalable, low-cost and easy to
deploy networks are perfectly met by a wireless sensor network. If one of the nodes
should fail, a new topology would be selected and the overall network would continue to
deliver data. If more nodes are placed in the field, they only create more potential routing
opportunities.
There is extensive research in the development of new algorithms for data
aggregation [1], ad hoc routing [2-4], and distributed signal processing in the context of
wireless sensor networks [5, 6]. As the algorithms and protocols for wireless sensor
network are developed, they must be supported by a low-power, efficient and flexible
hardware platform.
This thesis focuses on developing the system architecture required to meet the
needs of wireless sensor networks. A core design challenge in wireless sensor networks
is coping with the harsh resource constraints placed on the individual devices. Embedded
Figure 1-2: Possible deployment of ac-hoc wireless
embedded network for precision agriculture. Sensors
detect temperature, light levels and soil moisture at
hundreds of points across a field and communicate
their data over a multi-hop network for analysis.
5
processors with kilobytes of memory must implement complex, distributed, ad-hoc
networking protocols. Many constraints derive from the vision that these devices will
be produced in vast quantities and must be small and inexpensive. As Moore’s law
marches on, each device will get smaller, not just grow more powerful at a given size.
Size reduction is essential in order to allow devices to be produced as inexpensively as
possible, as well as to be able to allow devices to be used in a wide range of application
scenarios.
The most difficult resource constraint to meet is power consumption. As physical
size decreases, so does energy capacity. Underlying energy constraints end up creating
computational and storage limitations that lead to a new set of architectural issues.
Many devices, such as cell phones and pagers, reduce their power consumption through
the use specialized communication hardware in ASICs that provide low-power
implementations of the necessary communication protocols [7] and by relying on highpower
infrastructure. However, the strength of wireless sensor networks is their
flexibility and universality. The wide range of applications being targeted makes it
difficult to develop a single protocol, and in turn, an ASIC, that is efficient for all
applications. A wireless sensor network platform must provide support for a suite of
application-specific protocols that drastically reduce node size, cost, and power
consumption for their target application.
The wireless sensor network architecture we present here includes both a
hardware platform and an operating system designed specifically to address the needs of
wireless sensor networks. TinyOS is a component based operating system designed to
run in resource constrained wireless devices. It provides highly efficient communication
6
primitives and fine-grained concurrency mechanisms to application and protocol
developers. A key concept in TinyOS is the use of event based programming in
conjunction with a highly efficient component model. TinyOS enables system-wide
optimization by providing a tight coupling between hardware and software, as well as
flexible mechanisms for building application specific modules.
TinyOS has been designed to run on a generalized architecture where a single
CPU is shared between application and protocol processing. We detail three generations
of wireless nodes and a host of application deployments that have proven the capabilities
of our general system architecture. Figure 1-3 lays a picture and timeline of several
“mote” generations. The Mica platform has been produced in the largest quantities –
over 5000 Mica nodes have been produced and distributed to over 250 companies and
research organizations from around the country. The Mica platform includes a low
power transceiver, a power management subsystem, extended storage and an embedded
microcontroller.
The most advanced hardware platform we present is a single-chip CMOS device
that integrates the processing, storage and communication capabilities to form a complete
system node. This single chip node – called Spec – measures just 2.5 mm x 2.5 mm,
contains a microcontroller, transmitter, ADC, general purpose I/O ports, UART, memory
and encryption engine. The tiny chip only needs to be supported by a 32 KHz watch
crystal, an off-chip inductor and a power supply, a battery and a 4 MHz clock. The Spec
node represents the coming generation of wireless sensor nodes that will be manufactured
for pennies and deployed in the millions.
7
Both the Mica and Spec node are used to substantiate our claim that optimal
system architecture for wireless sensor networks is to have a single central controller
directly connected to a low-power radio transceiver through a rich interface that supports
hardware assistance for communication primitives. In contrast to having a hierarchical
partition of hardware resources dedicated to specific functions, a single shared controller
performs all processing. This allows for the dynamic allocation of computation
resources to the set of computational tasks demanded by the system. The layers of
abstraction typically achieved through hardware partitioning can instead be achieved
through the use of a highly efficient software-based component model. Software
abstractions allow for a wider scope of cross-layer optimizations that can achieve orders
of magnitude improvements is system performance. The power and viability of this
architecture is demonstrated through a collection benchmarks performed on real-world
hardware and in application level deployments.
􀀀










 
Figure 1-3: Design lineage of Mote Technology. COTS (Common off the shelf) prototypes lead to
the weC platform. Rene then evolved to allow sensor expansion and enabled hundreds of
compelling applications. The Dot node was architecturally the same as Rene but shrunk into a
quarter-sized device. Mica – discussed in depth in this thesis – made significant architectural
improvements in order to increase performance and efficiency. Spec represents the complete
integrated CMOS vision.
8
The main contributions of this work are (1) a general architecture that meets the
strict efficiency and flexibility requirements of wireless sensor networks, (2) an
implementation of the architecture using current microcontroller and low-power radio
technology, (3) an operating system that compliments the hardware capabilities and
provides support for sensor network applications, (4) an integrated single chip CMOS
hardware platform for use in wireless sensor networks and (5) a demonstration of the
flexibility of this architecture on several novel demonstration applications.
This thesis is organized in 9 chapters. Chapter 2 presents three key application
scenarios and an overview of the requirements for wireless sensor networks. It is
intended to provide the background necessary for a general understanding of the issues
discussed in later chapters. Additionally, Chapter 2 provides a general outline of the
capabilities of modern hardware building blocks. It culminates in describing a firstgeneration
sensor node used as a baseline for comparison. Chapter 3 presents the overall
architecture for the TinyOS operating system and a list of included components. TinyOS
is an event based operating system with a highly efficient component model. It plays a
critical role in exposing the capabilities of the hardware architectures we present.
Chapter 4 discusses the critical design issues that must be addressed by a wireless sensor
network platform. We discuss the shortcomings of traditional wireless architectures and
present a generalized architecture that addresses these issues. Our generalized
architecture is designed to allow for flexibility without sacrificing efficiency.
In Chapter 5 we present the Mica platform – an off-the-shelf approximation of our
generalized architecture. Mica represents what is possible with commercial hardware
and allows us to validate the basic principles of our generalized architecture. Mica has
9
bee successfully used in hundreds of real-world sensor network deployments. Chapter 6
presents the single chip Spec node where communication, computation, and storage are
combined onto a 2.5 mm x 2.5 mm die. Spec is a full realization of the generalized
architecture presented. Spec represents the future of wireless sensor network hardware.
Chapter 7 presets an overview and analysis of several demonstration applications. This
analysis underscores the performance impact of the architecture presented. Additionally
the applications presented demonstrate the flexibility and validate the platform.
Chapter 8 provides a survey of related work. This includes a summary of
research efforts that have been layered on top of the system we present as well as other
wireless activities. Chapter 9 summarizes the thesis and concludes with a prediction of
future technological trends.
10
Chapter 2: Wireless Sensor Networks
The concept of wireless sensor networks is based on a simple equation:
Sensing + CPU + Radio = Thousands of potential applications
As soon as people understand the capabilities of a wireless sensor network,
hundreds of applications spring to mind. It seems like a straightforward combination of
modern technology.
However, actually combining sensors, radios, and CPU’s into an effective
wireless sensor network requires a detailed understanding of the both capabilities and
limitations of each of the underlying hardware components, as well as a detailed
understanding of modern networking technologies and distributed systems theory. Each
individual node must be designed to provide the set of primitives necessary to synthesize
the interconnected web that will emerge as they are deployed, while meeting strict
requirements of size, cost and power consumption. A core challenge is to map the
overall system requirements down to individual device capabilities, requirements and
actions. To make the wireless sensor network vision a reality, an architecture must be
developed that synthesizes the envisioned applications out of the underlying hardware
capabilities.
To develop this system architecture we work from the high level application
requirements down through the low-level hardware requirements. In this process we first
attempt to understand the set of target applications. To limit the number of applications
that we must consider, we focus on a set of application classes that we believe are
representative of a large fraction of the potential usage scenarios. We use this set of
11
application classes to explore the system-level requirements that are placed on the overall
architecture. From these system-level requirements we can then drill down into the
individual node-level requirements. Additionally, we must provide a detailed background
into the capabilities of modern hardware.
After we present the raw hardware capabilities, we present a basic wireless sensor
node. The Rene node represents a first cut at a system architecture, and is used for
comparison against the system architectures presented in later chapters.
2.1 Sensor network application classes
The three application classes we have selected are: environmental data collection,
security monitoring, and sensor node tracking. We believe that the majority of wireless
sensor network deployments will fall into one of these class templates.
2.1.1 Environmental Data Collection
A canonical environmental data collection application is one where a research
scientist wants to collect several sensor readings from a set of points in an environment
over a period of time in order to detect trends and interdependencies. This scientist
would want to collect data from hundreds of points spread throughout the area and then
analyze the data offline [8, 9]. The scientist would be interested in collecting data over
several months or years in order to look for long-term and seasonal trends. For the data
to be meaningful it would have to be collected at regular intervals and the nodes would
remain at known locations.
At the network level, the environmental data collection application is
characterized by having a large number of nodes continually sensing and transmitting
12
data back to a set of base stations that store the data using traditional methods. These
networks generally require very low data rates and extremely long lifetimes. In typical
usage scenario, the nodes will be evenly distributed over an outdoor environment. This
distance between adjacent nodes will be minimal yet the distance across the entire
network will be significant.
After deployment, the nodes must first discover the topology of the network and
estimate optimal routing strategies [10]. The routing strategy can then be used to route
data to a central collection points. In environmental monitoring applications, it is not
essential that the nodes develop the optimal routing strategies on their own. Instead, it
may be possible to calculate the optimal routing topology outside of the network and then
communicate the necessary information to the nodes as required. This is possible
because the physical topology of the network is relatively constant. While the time
variant nature of RF communication may cause connectivity between two nodes to be
intermittent, the overall topology of the network will be relatively stable.
Environmental data collection applications typically use tree-based routing
topologies where each routing tree is rooted at high-capability nodes that sink data. Data
is periodically transmitted from child node to parent node up the tree-structure until it
reaches the sink. With tree-based data collection each node is responsible for forwarding
the data of all its descendants. Nodes with a large number of descendants transmit
significantly more data than leaf nodes. These nodes can quickly become energy
bottlenecks [11, 12].
Once the network is configured, each node periodically samples its sensors and
transmits its data up the routing tree and back to the base station. For many scenarios, the
13
interval between these transmissions can be on the order of minutes. Typical reporting
periods are expected to be between 1 and 15 minutes; while it is possible for networks to
have significantly higher reporting rates. The typical environment parameters being
monitored, such as temperature, light intensity, and humidity, do not change quickly
enough to require higher reporting rates.
In addition to large sample intervals, environmental monitoring applications do
not have strict latency requirements. Data samples can be delayed inside the network for
moderate periods of time without significantly affecting application performance. In
general the data is collected for future analysis, not for real-time operation.
In order to meet lifetime requirements, each communication event must be
precisely scheduled. The senor nodes will remain dormant a majority of the time; they
will only wake to transmit or receive data. If the precise schedule is not met, the
communication events will fail.
As the network ages, it is expected that nodes will fail over time. Periodically the
network will have to reconfigure to handle node/link failure or to redistribute network
load. Additionally, as the researchers learn more about the environment they study, they
may want to go in and insert additional sensing points. In both cases, the
reconfigurations are relatively infrequent and will not represent a significant amount of
the overall system energy usage.
The most important characteristics of the environmental monitoring requirements
are long lifetime, precise synchronization, low data rates and relatively static topologies.
Additionally it is not essential that the data be transmitted in real-time back to the central
14
collection point. The data transmissions can be delayed inside the network as necessary in
order to improve network efficiency.
2.1.2 Security Monitoring
Our second class of sensor network application is security monitoring. Security
monitoring networks are composed of nodes that are placed at fixed locations throughout
an environment that continually monitor one or more sensors to detect an anomaly. A
key difference between security monitoring and environmental monitoring is that security
networks are not actually collecting any data. This has a significant impact on the
optimal network architecture. Each node has to frequently check the status of its sensors
but it only has to transmit a data report when there is a security violation. The
immediate and reliable communication of alarm messages is the primary system
requirement. These are “report by exception” networks.
Additionally, it is essential that it is confirmed that each node is still present and
functioning. If a node were to be disabled or fail, it would represent a security violation
that should be reported. For security monitoring applications, the network must be
configured so that nodes are responsible for confirming the status of each other. One
approach is to have each node be assigned to peer that will report if a node is not
functioning. The optimal topology of a security monitoring network will look quite
different from that of a data collection network.
In a collection tree, each node must transmit the data of all of its decedents.
Because of this, it is optimal to have a short, wide tree. In contrast, with a security
network the optimal configuration would be to have a linear topology that forms a
Hamiltonian cycle of the network. The power consumption of each node is only
15
proportional to the number of children it has. In a linear network, each node would have
only one child. This would evenly distribute the energy consumption of the network.
The accepted norm for security systems today is that each sensor should be
checked approximately once per hour. Combined with the ability to evenly distribute the
load of checking nodes, the energy cost of performing this check becomes minimal. A
majority of the energy consumption in a security network is spent on meeting the strict
latency requirements associated with the signaling the alarm when a security violation
occurs.
Once detected, a security violation must be communicated to the base station
immediately. The latency of the data communication across the network to the base
station has a critical impact on application performance. Users demand that alarm
situations be reported within seconds of detection. This means that network nodes must
be able to respond quickly to requests from their neighbors to forward data.
In security networks reducing the latency of an alarm transmission is significantly
more important than reducing the energy cost of the transmissions. This is because
alarm events are expected to be rare. In a fire security system alarms would almost
never be signaled. In the event that one does occur a significant amount of energy could
be dedicated to the transmission. Reducing the transmission latency leads to higher
energy consumption because routing nodes must monitor the radio channel more
frequently.
In security networks, a vast majority of the energy will be spend on confirming
the functionality of neighboring nodes and in being prepared to instantly forward alarm
16
announcements. Actual data transmission will consume a small fraction of the network
energy.
2.1.3 Node tracking scenarios
A third usage scenario commonly discussed for sensor networks is the tracking of
a tagged object through a region of space monitored by a sensor network. There are
many situations where one would like to track the location of valuable assets or
personnel. Current inventory control systems attempt to track objects by recording the
last checkpoint that an object passed through. However, with these systems it is not
possible to determine the current location of an object. For example, UPS tracks every
shipment by scanning it with a barcode whenever it passes through a routing center. The
system breaks down when objects do not flow from checkpoint to checkpoint. In typical
work environments it is impractical to expect objects to be continually passed through
checkpoints.
With wireless sensor networks, objects can be tracked by simply tagging them
with a small sensor node. The sensor node will be tracked as it moves through a field of
sensor nodes that are deployed in the environment at known locations. Instead of
sensing environmental data, these nodes will be deployed to sense the RF messages of the
nodes attached to various objects. The nodes can be used as active tags that announce
the presence of a device. A database can be used to record the location of tracked objects
relative to the set of nodes at known locations. With this system, it becomes possible to
ask where an object is currently, not simply where it was last scanned [13].
Unlike sensing or security networks, node tracking applications will continually
have topology changes as nodes move through the network. While the connectivity
17
between the nodes at fixed locations will remain relatively stable, the connectivity to
mobile nodes will be continually changing. Additionally the set of nodes being tracked
will continually change as objects enter and leave the system. It is essential that the
network be able to efficiently detect the presence of new nodes that enter the network.
2.1.4 Hybrid networks
In general, complete application scenarios contain aspects of all three categories.
For example, in a network designed to track vehicles that pass through it, the network
may switch between being an alarm monitoring network and a data collection network.
During the long periods of inactivity when no vehicles are present, the network will
simply perform an alarm monitoring function. Each node will monitor its sensors
waiting to detect a vehicle. Once an alarm event is detected, all or part of the network,
will switch into a data collection network and periodically report sensor readings up to a
base station that track the vehicles progress. Because of this multi-modal network
behavior, it is important to develop a single architecture that and handle all three of these
application scenarios.
2.2 System Evaluation Metrics
Now that we have established the set of application scenarios that we are
addressing, we explore the evaluation metrics that will be used to evaluate a wireless
sensor network. To do this we keep in mind the high-level objectives of the network
deployment, the intended usage of the network, and the key advantages of wireless sensor
networks over existing technologies. The key evaluation metrics for wireless sensor
18
networks are lifetime, coverage, cost and ease of deployment, response time, temporal
accuracy, security, and effective sample rate. Their importance is discussed below.
One result is that many of these evaluation metrics are interrelated. Often it may
be necessary to decrease performance in one metric, such as sample rate, in order to
increase another, such as lifetime. Taken together, this set of metrics form a
multidimensional space that can be used to describe the capabilities of a wireless sensor
network. The capabilities of a platform are represented by a volume in this
multidimensional space that contains all of the valid operating points. In turn, a specific
application deployment is represented by a single point. A system platform can
successfully perform the application if and only if the application requirements point lies
inside the capability hyperspace.
One goal of this chapter is to present an understanding of the tradeoffs that link
each axis of this space and an understanding of current capabilities. The architectural
improvements and optimizations we present in later chapters are then motivated by
increasing the ability to deliver these capabilities and increasing the volume of the
capability hypercube.
2.2.1 Lifetime
Critical to any wireless sensor network deployment is the expected lifetime. The
goal of both the environmental monitoring and security application scenarios is to have
nodes placed out in the field, unattended, for months or years.
The primary limiting factor for the lifetime of a sensor network is the energy
supply. Each node must be designed to manage its local supply of energy in order to
maximize total network lifetime. In many deployments it is not the average node lifetime
19
that is important, but rather the minimum node lifetime. In the case of wireless security
systems, every node must last for multiple years. A single node failure would create a
vulnerability in the security systems.
In some situations it may be possible to exploit external power, perhaps by
tapping into building power with some or all nodes. However, one of the major benefits
to wireless systems is the ease of installation. Requiring power to be supplied externally
to all nodes largely negates this advantage. A compromise is to have a handful of special
nodes that are wired into the building’s power infrastructure.
In most application scenarios, a majority of the nodes will have to be selfpowered.
They will either have to contain enough stored energy to last for years, or they
will have to be able to scavenge energy from the environment through devices, such as
solar cells or piezoelectric generators [14, 15]. Both of these options demand that that
the average energy consumption of the nodes be as low as possible.
The most significant factor in determining lifetime of a given energy supply is
radio power consumption. In a wireless sensor node the radio consumes a vast majority
of the system energy. This power consumption can be reduced through decreasing the
transmission output power or through decreasing the radio duty cycle. Both of these
alternatives involve sacrificing other system metrics.
2.2.2 Coverage
Next to lifetime, coverage is the primary evaluation metric for a wireless network.
It is always advantageous to have the ability to deploy a network over a larger physical
area. This can significantly increase a system’s value to the end user. It is important to
keep in mind that the coverage of the network is not equal to the range of the wireless
20
communication links being used. Multi-hop communication techniques can extend the
coverage of the network well beyond the range of the radio technology alone. In theory
they have the ability to extend network range indefinitely. However, for a given
transmission range, multi-hop networking protocols increase the power consumption of
the nodes, which may decrease the network lifetime. Additionally, they require a
minimal node density, which may increase the deployment cost.
Tied to range is a network’s ability to scale to a large number of nodes.
Scalability is a key component of the wireless sensor network value proposition. A user
can deploy a small trial network at first and then can continually add sense points to
collect more and different information. A user must be confident that the network
technology being used is capable of scaling to meet his eventual need. Increasing the
number of nodes in the system will impact either the lifetime or effective sample rate.
More sensing points will cause more data to be transmitted which will increase the power
consumption of the network. This can be offset by sampling less often.
2.2.3 Cost and ease of deployment
A key advantage of wireless sensor networks is their ease of deployment.
Biologists and construction workers installing networks cannot be expected to understand
the underlying networking and communication mechanisms at work inside the wireless
network. For system deployments to be successful, the wireless sensor network must
configure itself. It must be possible for nodes to be placed throughout the environment
by an untrained person and have the system simply work.
Ideally, the system would automatically configure itself for any possible physical
node placement. However, real systems must place constraints on actual node
21
placements – it is not possible to have nodes with infinite range. The wireless sensor
network must be capable of providing feedback as to when these constraints are violated.
The network should be able to assess quality of the network deployment and indicate any
potential problems. This translates to requiring that each device be capable of performing
link discovery and determining link quality.
In addition to an initial configuration phase, the system must also adapt to
changing environmental conditions. Throughout the lifetime of a deployment, nodes may
be relocated or large physical objects may be placed so that they interfere with the
communication between two nodes. The network should be able to automatically
reconfigure on demand in order to tolerate these occurrences.
The initial deployment and configuration is only the first step in the network
lifecycle. In the long term, the total cost of ownership for a system may have more to do
with the maintenance cost than the initial deployment cost. The security application
scenario in particular requires that the system be extremely robust. In addition to
extensive hardware and software testing prior to deployment, the sensor system must be
constructed so that it is capable of performing continual self-maintenance. When
necessary, it should also be able to generate requests when external maintenance is
required.
In a real deployment, a fraction of the total energy budget must be dedicated to
system maintenance and verification. The generation of diagnostic and reconfiguration
traffic reduces the network lifetime. It can also decrease the effective sample rate.
22
2.2.4 Response Time
Particularly in our alarm application scenario, system response time is a critical
performance metric. An alarm must be signaled immediately when an intrusion is
detected. Despite low power operation, nodes must be capable of having immediate,
high-priority messages communicated across the network as quickly as possible. While
these events will be infrequent, they may occur at any time without notice. Response
time is also critical when environmental monitoring is used to control factory machines
and equipment. Many users envision wireless sensor networks as useful tools for
industrial process control. These systems would only be practical if response time
guarantees could be met.
The ability to have low response time conflicts with many of the techniques used
to increase network lifetime. Network lifetime can be increased by having nodes only
operate their radios for brief periods of time. If a node only turns on its radio once per
minute to transmit and receive data, it would be impossible to meet the application
requirements for response time of a security system.
Response time can be improved by including nodes that are powered all the time.
These nodes can listen for the alarm messages and forward them down a routing
backbone when necessary. This, however, reduces the ease of deployment for the
system.
2.2.5 Temporal Accuracy
In environmental and tracking applications, samples from multiple nodes must be
cross-correlated in time in order to determine the nature of phenomenon being measured.
The necessary accuracy of this correlation mechanism will depend on the rate of
23
propagation of the phenomenon being measured. In the case of determining the average
temperature of a building, samples must only be correlated to within seconds. However,
to determine how a building reacts to a seismic event, millisecond accuracy is required.
To achieve temporal accuracy, a network must be capable of constructing and
maintaining a global time base that can be used to chronologically order samples and
events. In a distributed system, energy must be expended to maintain this distributed
clock. Time synchronization information must be continually communicated between
nodes. The frequency of the synchronization messages is dependent on the desired
accuracy of the time clock. The bottom line is maintenance of a distributed time base
requires both power and bandwidth.
2.2.6 Security
Despite the seemingly harmless nature of simple temperature and light
information from an environmental monitoring application, keeping this information
secure can be extremely important. Significant patterns of building use and activity can
be easily extracted from a trace of temperature and light activity in an office building. In
the wrong hands, this information can be exploited to plan a strategic or physical attack
on a company. Wireless sensor networks must be capable of keeping the information
they are collecting private from eavesdropping.
As we consider security oriented applications, data security becomes even more
significant. Not only must the system maintain privacy, it must also be able to
authenticate data communication. It should not be possible to introduce a false alarm
message or to replay an old alarm message as a current one. A combination of privacy
24
and authentication is required to address the needs of all three scenarios. Additionally, it
should not be possible to prevent proper operation by interfering with transmitted signals.
Use of encryption and cryptographic authentication costs both power and network
bandwidth [16, 17]. Extra computation must be performed to encrypt and decrypt data
and extra authentication bits must be transmitted with each packet. This impacts
application performance by decreasing the number of samples than can be extracted from
a given network and the expected network lifetime.
2.2.7 Effective Sample Rate
In a data collection network, effective sample rate is a primary application
performance metric. We define the effective sample rate as the sample rate that sensor
data can be taken at each individual sensor and communicated to a collection point in a
data collection network. Fortunately, environmental data collection applications typically
only demand sampling rates of 1-2 samples per minute. However, in addition to the
sample rate of a single sensor, we must also consider the impact of the multi-hop
networking architectures on a nodes ability to effectively relay the data of surrounding
nodes.
In a data collection tree, a node must handle the data of all of its descendents. If
each child transmits a single sensor reading and a node has a total of 60 descendants, then
it will be forced to transmit 60 times as much data. Additionally, it must be capable of
receiving those 60 readings in a single sample period. This multiplicative increase in
data communication has a significant effect on system requirements. Network bit rates
combined with maximum network size end up impacting the effective per-node sample
rate of the complete system [5].
25
One mechanism for increasing the effective sample rate beyond the raw
communication capabilities of the network is to exploit in-network processing. Various
forms of spatial and temporal compression can be used to reduce the communication
bandwidth required while maintaining the same effective sampling rate. Additionally
local storage can be used to collect and store data at a high sample rate for short periods
of time. In-network data processing can be used to determine when an “interesting”
event has occurred and automatically trigger data storage. The data can then be
downloaded over the multi-hop network as bandwidth allows.
Triggering is the simplest form of in-network processing. It is commonly used in
security systems. Effectively, each individual sensor is sampled continuously, processed,
and only when a security breach has occurred is data transmitted to the base station. If
there were no local computation, a continuous stream of redundant sensor readings would
have to be transmitted. We show how this same process can be extended to complex
detection events.
2.3 Individual node evaluation metrics
Now that we have established the set of metrics that will be used to evaluate the
performance of the sensor network as a whole, we can attempt to link the system
performance metrics down to the individual node characteristics that support them. The
end goal is to understand how changes to the low-level system architecture impact
application performance. Just as application metrics are often interrelated, we will see
that an improvement in one node-level evaluation metric (e.g., range) often comes at the
expense of another (e.g., power).
26
2.3.1 Power
To meet the multi-year application requirements individual sensor nodes must be
incredibly low-power. Unlike cell phones, with average power consumption measured
in hundreds of milliamps and multi-day lifetimes, the average power consumption of
wireless sensor network nodes must be measured in micro amps. This ultra-low-power
operation can only be achieved by combining both low-power hardware components and
low duty-cycle operation techniques.
During active operation, radio communication will constitute a significant fraction
of the node’s total energy budget. Algorithms and protocols must be developed to
reduce radio activity whenever possible. This can be achieved by using localized
computation to reduce the streams of data being generated by sensors and through
application specific protocols. For example, events from multiple sensor nodes can be
combined together by a local group of nodes before transmitting a single result across the
sensor network.
Our discussion on available energy sources will show that a node must consume
less that 200 uA on average to last for one year on a pair of AA batteries. In contrast the
average power consumption of a cell phone is typically more than 4000 uA, a 20 fold
difference.
2.3.2 Flexibility
The wide range of usage scenarios being considered means that the node
architecture must be flexible and adaptive. Each application scenario will demand a
slightly different mix of lifetime, sample rate, response time and in-network processing.
A wireless sensor network architecture must be flexible enough to accommodate a wide
27
range of application behaviors. Additionally, for cost reasons each device will have only
the hardware and software it actually needs for a given the application. The architecture
must make it easy to assemble just the right set of software and hardware components.
Thus, these devices require an unusual degree of hardware and software modularity while
simultaneously maintaining efficiency.
2.3.3 Robustness
In order to support the lifetime requirements demanded, each node must be
constructed to be as robust as possible. In a typical deployment, hundreds of nodes will
have to work in harmony for years. To achieve this, the system must be constructed so
that it can tolerate and adapt to individual node failure. Additionally, each node must be
designed to be as robust as possible.
System modularity is a powerful tool that can be used to develop a robust system.
By dividing system functionality into isolated sub-pieces, each function can be fully
tested in isolation prior to combining them into a complete application. To facilitate this,
system components should be as independent as possible and have interfaces that are
narrow, in order to prevent unexpected interactions.
In addition to increasing the system’s robustness to node failure, a wireless sensor
network must also be robust to external interference. As these networks will often coexist
with other wireless systems, they need the ability to adapt their behavior
accordingly. The robustness of wireless links to external interference can be greatly
increased through the use of multi-channel and spread spectrum radios. It is common for
facilities to have existing wireless devices that operate on one or more frequencies. The
28
ability to avoid congested frequencies is essential in order to guarantee a successful
deployment.
2.3.4 Security
In order to meet the application level security requirements, the individual nodes
must be capable of performing complex encrypting and authentication algorithms.
Wireless data communication is easily susceptible to interception. The only way to keep
data carried by these networks private and authentic is to encrypt all data transmissions.
The CPU must be capable of performing the required cryptographic operations itself or
with the help of included cryptographic accelerators [16].
In addition to securing all data transmission, the nodes themselves must secure the
data that they contain. While they will not have large amounts of application data stored
internally, they will have to store secret encryption keys used in the network. If these
keys are revealed, the security of the network could crumble. To provide true security, it
must be difficult to extract the encryption keys of from any node.
2.3.5 Communication
A key evaluation metric for any wireless sensor network is its communication
rate, power consumption, and range. While we have made the argument that the
coverage of the network is not limited by the transmission range of the individual nodes,
the transmission range does have a significant impact on the minimal acceptable node
density. If nodes are placed too far apart it may not be possible to create an
interconnected network or one with enough redundancy to maintain a high level of
reliability. Most application scenarios have natural node densities that correspond to the
29
granularity of sensing that is desired. If the radio communications range demands a
higher node density, additional nodes must be added to the system in to increase node
density to a tolerable level.
The communication rate also has a significant impact on node performance.
Higher communication rates translate into the ability to achieve higher effective sampling
rates and lower network power consumption. As bit rates increase, transmissions take
less time and therefore potentially require less energy. However, an increase in radio bit
rate is often accompanied by an increase in radio power consumption. All things being
equal, a higher transmission bit rate will result in higher system performance. However,
we show later that an increase in the communication bit rate has a significant impact on
the power consumption and computational requirement of the node. In total, the benefits
of an increase in bit rate can be offset by several other factors.
2.3.6 Computation
The two most computationally intensive operations for a wireless sensor node are
the in-network data processing and the management of the low-level wireless
communication protocols. As we discuss later, there are strict real-time requirements
associated with both communication and sensing. As data is arriving over the network,
the CPU must simultaneously control the radio and record/decode the incoming data.
Higher communication rates required faster computation.
The same is true for processing being performed on sensor data. Analog sensors
can generate thousands of samples per second. Common sensor processing operations
include digital filtering, averaging, threshold detection, correlation and spectral analysis.
30
It may even be necessary to perform a real-time FFT on incoming data in order to detect
a high-level event.
In addition to being able to locally process, refine and discard sensor readings, it
can be beneficial to combine data with neighboring sensors before transmission across a
network. Just as complex sensor waveforms can be reduced to key events, the results
from multiple nodes can be synthesized together. This in-network processing requires
additional computational resources.
In our experience, 2-4 MIPS of processing are required to implement the radio
communication protocols used in wireless sensor networks. Beyond that, the application
data processing can consume an arbitrary amount of computation depending on the
calculations being performed.
2.3.7 Time Synchronization
In order to support time correlated sensor readings and low-duty cycle operation
of our data collection application scenario, nodes must be able to maintain precise time
synchronization with other members of the network. Nodes need to sleep and awake
together so that they can periodically communicate. Errors in the timing mechanism will
create inefficiencies that result in increased duty cycles.
In distributed systems, clocks drift apart over time due to inaccuracies in
timekeeping mechanisms. Depending on temperature, voltage, humidity, time keeping
oscillators operate at slightly different frequencies. High-precision synchronization
mechanisms must be provided to continually compensate for these inaccuracies.
31
2.3.8 Size & Cost
The physical size and cost of each individual sensor node has a significant and
direct impact on the ease and cost of deployment. Total cost of ownership and initial
deployment cost are two key factors that will drive the adoption of wireless sensor
network technologies. In data collection networks, researchers will often be operating off
of a fixed budget. Their primary goal will be to collect data from as many locations as
possible without exceeding their fixed budget. A reduction in per-node cost will result in
the ability to purchase more nodes, deploy a collection network with higher density, and
collect more data.
Physical size also impacts the ease of network deployment. Smaller nodes can be
placed in more locations and used in more scenarios. In the node tracking scenario,
smaller, lower cost nodes will result in the ability to track more objects.
2.4 Hardware Capabilities
Now that we have identified the key characteristics of a wireless sensor node we
can look at the capabilities of modern hardware. This allows us to understand what bit
rate, power consumption, memory and cost we can expect to achieve. A balance must be
maintained between capability, power consumption and size in order to best address
application needs. This section gives a quick overview of modern technology and the
tradeoffs between different technologies. We start with a background of energy storage
technologies and continue through the radio, CPU, and sensors.
32
2.4.1 Energy
Just as power consumption of system components are often expressed in
milliamps, batteries are generally rated in milliamp-hours (mAh). In theory a 1000 mAh
battery could support a processor consuming 10 mA for 100 hours. In practice this in
not always true. Due to battery chemistry, voltage and current levels vary depending on
how the energy is extracted from a battery. Additionally, as batteries discharge their
voltage drops. If the system is not tolerant to a decrease in voltage it may not be possible
to use the full rated capacity of a battery. For example, a 1.5 V alkaline battery is not
considered empty by the manufacturer until it is outputting only .8 V [18].
2.4.1.1 Battery technologies
There are three common battery technologies that are applicable for wireless
sensor networks – Alkaline, Lithium, and Nickel Metal Hydride. An AA Alkaline battery
is rated at 1.5 V, but during operation it ranges from 1.65 to .8 V as shown in Figure 2-1
and is rated at 2850 mAh. With a volume of just 8.5 cm3, it has an energy density of
approx 1500 Joules/cm3. While providing a cheap, high capacity, energy source, the
major drawbacks of alkaline batteries are the wide voltage range that must be tolerated
and their large physical size. Additionally, lifetimes beyond 5 years cannot be achieved
because of battery self-discharge. The shelf-life of an alkaline battery is approximately 5
years.
Lithium batteries provide an incredibly compact power source. The smallest
versions are just a few millimeters across. Additionally, they provide a constant voltage
33
supply that decays little as the battery is drained. Devices that operate off of lithium
batteries do not have to be as tolerant to voltage changes as devices that operate off of
alkaline batteries. Additionally, unlike alkaline batteries, lithium batteries are able to
operate at temperatures down to -40 C. The most common lithium battery is the CR2032
[19]. It is rated at 3V, 255 mAh and sells for just 16 cents. With a volume of 1 cm3, it
has and energy density of 2400 J/cm3. In addition to traditional lithium batteries, there
are also specialized Tadiran lithium batteries that have densities as high as 4000 J/cm3
and tolerate a wide temperature range. One of the drawbacks of lithium batteries is that
they often have very low nominal discharge currents. A D-cell size Tadiran battery has a
Figure 2-1: Battery characteristics for Lithium, Alkaline and NiMH batteries. The discharge
characteristics of alkaline batteries make it essential to design a system to tolerate a wide range of
input voltages.
Battery Characteristics
Alkaline discharge curve
Battery Energy Desnsity
0
500
1000
1500
2000
2500
3000
3500
4000
4500
1
J/cm
CR2032 Lithium
Tadiran Lithium
AA Alkaline
AA NiMH
Lithium discharge curve
34
nominal discharge current of just 3 mA. This is compared to an alkaline AA battery’s
nominal discharge rate of 25 mA.
Nickel Metal Hydride batteries are the third major battery type. They have the
benefit of being easily rechargeable. The downside to rechargeable batteries is a
significant decrease in energy density. An AA size NiMH battery has approximately half
the energy density of an alkaline battery at approximately 5 times the cost. Before
considering the use of NiMH batteries it is important to note that they only produce 1.2
V. Because many system components require 2.7 volts or more, they it may not be
possible to operate directly off of rechargeable batteries.
2.4.1.2 Expected lifetime calculation
While it appears easy to quickly look at a battery’s rated capacity, compare it to
the systems energy consumption and calculate the system lifetime, there are several
additional factors to consider.
In looking at real systems, it is important to look at how power supplies decay
over time. Figure 2-1 shows the voltage versus time plot of an AA battery drained at a
500 mW. The graph shows that the battery quickly falls from the 1.5 V starting voltage
and ends at just .8 V.
If a theoretical system was built with components requiring 2.7 V and consumed
250 mW, it would only last for 100 minutes off of 2 AA batteries. However, if the
system components were selected to operate off of voltages down to 2.0 volts, it would
last approximately 5 times as long off of the same power source. A seemingly
unimportant CPU parameter results in a 5x difference in system lifetime.
35
2.4.1.3 Voltage Regulation
In order to avoid issues associated with varying battery voltages, voltage
regulation techniques can be used. In general, voltage regulators take in varying input
voltages and produce a stable, constant output voltage. Standard regulators require the
input voltage be greater than the desired output voltage. Boost converters can output
voltages that are higher than the input voltage.
There are two major drawbacks to using voltage regulation. The most obvious is
the inefficient conversion process. High efficiency voltage converters can boast of
efficiencies up to 90%. However, these are generally tuned for systems consuming
hundreds of milliamps. The efficiencies fall drastically as the current consumption of
the system is reduced.
Secondly, the quiescent current consumption – or consumption when no current is
being output – of regulators can be relatively high. The current consumption of a
standard regulator is between 15 and 50 uA. While irrelevant in a system consuming
hundreds of milliamps, for ultra-low power devices, this small current drain can dominate
overall energy usage.
The inefficiencies associated with voltage conversion and regulation combined
with the high variation in output voltage of alkaline batteries makes it highly
advantageous to build sensor nodes out of components that are tolerant to a wide voltage
range. Selection of components that can operate over a range of (2.1-3.3V) can
significantly improve system performance.
36
2.4.1.4 Renewable Energy
An alternative to relying on batteries with enough energy to last for years is to use
renewable energy. Modern solar cells can produce up to 10 mW per square inch in
direct sunlight. If stored properly, the energy collected during the day can be enough
energy to last through the night. In indoor lighting environments between 10 and 100
uW per square inch can be produced depending on the type of lighting. For solar
powered application scenarios, the key to successfully harnessing solar energy lies in the
ability to store the energy.
Modern ultra-capacitors represent an option for energy storage. Ultra capacitors
are low-voltage capacitors ranging between 1 and 6 Farads. When charged to 3 V they
contain enough energy to send several hundred radio packets. While they can be
charged and drained easily, one of their biggest drawbacks is that they have internal
leakage rates of 20 to 50 uA when charged to 3 V. In low duty cycle operations this will
quickly dominate the energy consumption of the nodes.
Rechargeable batteries can also be used to store solar energy. NiMH batteries can
be trickle charged directly off of a solar cell. During periods of bright sunlight, energy
could be slowly transferred back into the battery. 66% of the energy put into a NiMH
battery turns into usable stored energy [20]. While the initial recharging process is less
efficient than storing energy in an ultra-cap, rechargeable batteries have a significantly
lower self discharge rate. If the energy is going to be used over several hours, the reduced
self-discharge rate will outweigh the charging inefficiency.
37
2.4.1.5 Idealized AA Battery
For the remainder of this thesis we use a pair of idealized AA batteries in order to
estimate system lifetime. For the idealized battery we assume that the battery discharges
from 1.6 V down to .8 V in a linear fashion while delivering 2250 mA hours of charge.
We use 20% less than the rated capacity of AA to ensure that our estimates are not too
high.
We use this idealized battery to generate an energy budget and analyze the impact
of architectural alternatives. If we assume that the network must last for one year and
that the system is able to operate down to 2.0V, it can have an average power
consumption of just 192 uA (approximately 500 uW). This current must be divided
across all system components and functions. We can keep this figure in mind as we
continue to evaluate system components.
2.4.2 Radios
The radio subsystem is the most important system on a wireless sensor node since
it is the primary energy consumer in all three of our application scenarios. Modern lowpower,
short range transceivers consume between 15 and 300 milliwatts of power when
sending and receiving. A key hardware observation is that low power radios consume
approximately the same amount of energy when in receive or transmit mode. This
energy is consumed if the radio is on, weather or not it is receiving actual data. The
actual power emitted out of the antenna only accounts for a small fraction of the
transceiver’s energy consumption. A significant fraction goes to internal operation.
Because of this, the overall cost of radio communication can easily be dominated by the
receiver power consumption – a metric that is often ignored in wireless studies.
38
Additionally, if the receiver is left on 100% of the time during periods of intermittent
communication, the receiver will be the single largest energy consumer in the system.
People commonly make the mistake of assuming that reception is free.
2.4.2.1 Transmission Range
The transmission range of a wireless system is controlled by several key factors.
The most intuitive factor is that of transmission power. The more energy put into a
signal, the farther it should travel. The relationship between power output and distance
traveled is a polynomial with an exponent of between 3 and 4 (non-line of sight
propagation) [21]. So to transmit twice as far through an indoor environment, 8 to 16
times as much energy must be emitted.
Other factors in determining range include the sensitivity of the receiver, the gain
and efficiency of the antenna and the channel encoding mechanism. In general, wireless
sensor network nodes cannot exploit high gain, directional antennas because they require
special alignment and prevent ad-hoc network topologies. Omni-directional antennas
are preferred in ad-hoc networks because they allow nodes to effectively communicate in
all directions.
Both transmission strength and receiver sensitivity are measured in dBm. Typical
receiver sensitivities are between -85 and -110 dBm. (The dB scale is a logarithmic
scale where a 10 dB increase represents a 10x increase in power. The baseline of 0 dBm
represents 1 milliwatt, so 1 watt is 30 dBm.) Transmission range increases can be
achieved by either increasing sensitivity or by increasing transmission power. When
transmitting at 0 dBm, a receiver sensitivity of -85 dBm will result in an outdoor free
space range of 25-50 meters, while a sensitivity of -110 dBm will result in a range of 100
39
to 200 meters. The use of a radio with a sensitivity of -100 dBm instead of a radio with -
85 dBm will allow you to decrease the transmission power by a factor of 30 and achieve
the same range.
2.4.2.2 Modulation Type
A key characteristic of any RF device is the modulation mechanism. Most radios
communicate information by modulating a RF carrier signal. The standard modulation
mechanisms are amplitude modulation and frequency modulation. Amplitude
modulation is the simplest to encode and decode, but it is the most susceptible to noise.
Because the data is encoded in the strength of the transmission, external noise is added to
the signal. In contrast, frequency based modulation is less susceptible to noise because
all data is transmitted at the same power level.
2.4.2.2.1 Frequency Agile Radios
Most transceivers on the market today use a VCO (Voltage Controlled Oscillator)
based radio architecture and have the ability to communicate at a variety of carrier
frequencies. Not only does this give them the potential to be robust to interfering
signals, it also allows the radios to comply with FCC part 15 regulations. In the 902-928
MHz band, radios must channel-hop across at least 25 channels in order to be allowed to
transmit at up to 250 mW or 24 dBm. If only a single channel is used, the transmission
must remain slightly below 1 milliwatt.
An additional reason to use a frequency agile radio is that it allows for the use of
FM or frequency modulation based data communication. FM is more robust to
40
interference and fading than the simpler AM based modulation. Fixed frequency radios
are limited to the use of AM based data encoding schemes.
2.4.2.2.2 Frequency Hopping Spread Spectrum
To increase a system’s ability to tolerate external interference, spread spectrum
transmission techniques can be used. Spread spectrum communication is a technique that
increases channel reliability and tolerance to noise by spreading the signal to be
communicated over a wide range of frequencies. An interference source may corrupt a
small portion of the data, but a majority of the information transmitted will be received.
Frequency hopping (FH) is a common form of spread spectrum. The carrier frequency is
continually changed based on a pseudorandom algorithm. On each hop, the data is
transmitted on a single carrier frequency but over time the data will be spread across the
frequency spectrum. Dwell times – the duration each channel is used – range from
hundreds of microseconds to tens of milliseconds. (.0001s->.1s)
The major drawback of frequency hopping systems when applied to wireless
sensor networks is the overhead of maintaining channel synchronization and the effort to
discover the current hopping sequence. As nodes attempt to discover each other they
must attempt to search all possible channel locations. This is an expensive, time
consuming operation that is not compatible with low duty cycle networks. It can be seen
how this leads to high power consumption in Bluetooth devices.
2.4.2.2.3 Direct Sequencing Spread Spectrum
In contrast to frequency hopping systems, direct sequencing spread spectrum
radios communicate on a single carrier frequency. Instead of hopping carriers, the signal
41
to be communicated is directly spread over a wide frequency band by multiplying the
signal by a higher rate pseudorandom sequence. Typical DSSS radios have spreading
ratios between 15 and 100. This means that the radio must transmit between 15 and 100
times more data than is actually contained in the signal.
During reception, the received signal is passed through a correlator that
reconstructs the original input signal. The main downside to the use of DSSS is the
overhead associated with the spreading codes and the cost of the decorrelation. To
achieve the same effective bit rate as a standard radio, the transmission rate must be
increased by 15x-100x. This bit rate increase translates into an increase in radio power
consumption and a decrease in receiver sensitivity. The decrease in sensitivity is then
offset by the processing gain associated with the correlation, ideally resulting in a
comparable communication range.
2.4.2.3 Power Consumption
To illustrate the power consumption of modern low power transceivers, we
simply outline the characteristics of two commercial radios, the RF Monolithics TR1000
[22] and the Chipcon CC1000 [23].
The TR1000 radio consumes 21 mW of energy when transmitting at 0.75 mW.
During reception, the TR1000 consumes 15 mW and has a receive sensitivity of -85
dBm. In contrast, at optimum input supply voltage, the CC1000 consumes 50 mW to
transmit at 3 mW and consumes 20 mW to have a receive sensitivity of -105 dBm. When
transmitting at the same .75 mW as the TR1000, the CC1000 consumes 31.6 mW. In
practice TR1000 provides an outdoor, line-of-site communication range of up to 300 feet
compared to 900 feet for the CC1000.
42
To put these numbers in perspective we can refer back to our idealized AA
battery. On our idealized power source, the CC1000 can transmit for approximately 4
days straight or remain in receive mode for 9 days straight. In order to last for one year,
the CC1000 must operate at a duty cycle of approximately 2%.
2.4.2.4 Bit Rate
Unlike many high performance data networks, wireless sensor networks do not
require high bit rates. 10-100 Kbps of raw network bandwidth is sufficient for many
applications. Radio bandwidth has a more significant impact on node power
consumption and lifetime than it does on application capabilities.
As bit rates increase, transmission times decrease. As the highest instantaneous
energy consumer it is essential that the radio remain off as much as possible. By
increasing the bit rate without increasing the amount of data being transmitted, the radio
duty cycle is decreased. The CC1000 and the RFM TR1000 both offer bit rates of
approximately 100 Kbps.
2.4.2.5 Turn-on Time
In performing an analysis of a radio’s power consumption there are several factors
to consider in addition to the peak power consumption. A radio’s ability to quickly enter
and exit low power sleep modes is extremely important for efficient operation in wireless
sensor networks. All usage scenarios assume that the radio will only be on occasionally.
Before and after a transmission, time and energy must be spent configuring and powering
up the radio. Ideally this should be a quick, cheap, operation. If a system has to react to
emergency events within seconds, the radio must be powered at least once per second. If
43
a radio’s turn-on-to-receive time is more than a few tens of milliseconds, it quickly
becomes impossible to achieve the required duty cycles of less that 1 percent. Anything
beyond a 5 ms turn-on time would be unacceptable for most applications.
One of the drawbacks to multi-channel radios that use a VCO based frequency
synthesizer is that the VCO must stabilize prior to transmission or reception.
Additionally the VCO is often locked to a high frequency crystal that must also stabilize.
The CC1000 radio from Chipcon requires 2 ms for the primary crystal to stabilize and
200 us for the high frequency VCO. In contrast, the RFM radio that uses a fixed
frequency SAW filter can be turned on and be ready to receive in just a few hundred
microseconds – more than ten times faster.
2.4.3 Processor
Modern microcontrollers integrate flash storage, RAM, analog-to-digital
converters and digital I/O onto a single integrated circuit that costs between $1 and $5.
Their tight integration makes them ideal for use in deeply embedded systems like
wireless sensor networks. When selecting a microcontroller family, some of the key
requirements are energy consumption, voltage requirements, cost, support for peripherals
and the number of external components required.
2.4.3.1 Power
Power consumption of microcontrollers varies greatly from family to family.
Standard 8 or 16 bit microcontrollers consume between .250 to 2.5 mA per MHz. This
10x difference between low-power and standard microcontrollers can have a significant
44
impact on system performance. While it is obvious that selecting the lowest power
controller is essential, peak power consumption is not the most important metric.
A second significant power measurement for a microcontroller is the sleep mode
power consumption. The CPU is a major contributor to the node’s standby power
consumption. During idle periods, the CPU will stop execution and enter a low-power
sleep state. The processor only needs to maintain its memory and maintain time
synchronization so it can properly wake-up when necessary. Sleep mode current
consumption varies between 1 uA [24] and 50 uA across controller families. Despite
being a difference of just a few uA, this 50x variation can have a more significant impact
on node performance than milliamp differences in peak power consumption. This is
because the CPU is expected to be idle 99.9% of the time. A microcontroller that burns
50 uA in sleep mode will have a maximum lifetime of 4 years on our idealized power
source even if it never leaves the sleep mode. In contrast the lifetime of a controller that
consumes 1 uA in sleep mode will be completely dominated by battery self discharge and
active power consumption.
2.4.3.2 Wake-up Time
In addition to the current consumption of the sleep mode, the transition
requirements for entering and exiting sleep modes are also significant. Some
microcontrollers require as much as 10 ms to enter and resume from the sleep state. In
contrast, some controllers have wake-up times of just 6 us. The wake-up delay is used to
start and stabilize system clocks. The faster a controller can enter and leave the lowpower
state, the more often the sleep state can be used and the more responsive the node
will be.
45
With a quick wake-up time, it is not necessary to remain awake through short
periods of inactivity. For example, during a message transmission, the controller can
enter the sleep mode between each bit of the transmission. This can reduce the effective
active current consumption of the controller significantly. Even at a 9600 bps
communication rate, a 1 ms sleep time makes entering the sleep mode between each bit
impossible.
2.4.3.3 Voltage Requirements
The operating voltage range of a low-power microcontroller can also have
significant impact on system performance. Traditional low-voltage microcontrollers
operate between 2.7V and 3.3 V. However, new generations of low-power controllers
are appearing that operate down to 1.8 V. As discussed in our analysis of energy sources,
a wide voltage tolerance allows devices to be more efficiently operated off of Alkaline
batteries.
2.4.3.4 Speed
The main duties of the microcontroller are to execute the communication
protocols, control the radio, perform data encryption, interact with sensors and perform
data processing. Most of these operations place relatively low demands on the speed of
the controller. Current designs demand CPU speeds between 1 and 4 MHz. This range
is consistent with most microcontrollers on the market today. Most controllers are fully
static with the ability to operate between 0 and 8 MHz.
One feature of several controller families is the ability to dynamically change the
operating frequency. However, CMOS power consumption obeys the equation
46
“P=CV2F”. This relation states that the power consumption of the controller scales
linearly with frequency. However, the execution time is inversely proportional to
frequency. The impact of changing the frequency is negligible because it is offset by the
change in execution time resulting in no change in energy usage. If proper techniques
are used to effectively enter low power sleep states there is not a significant advantage to
reducing or increasing the CPU speed.
The critical variable in determining the necessary computational speed for a
wireless sensor node is the amount of data analysis and in-network processing that must
be performed. The CPU must be capable of meeting the real-time deadlines demanded
by the data processing.
2.4.3.5 Memory
In general, sensor nodes only require small amounts of storage and program
memory. Data is only stored long enough for it to be analyzed and then transmitted
through the network to the base station. In general, modern flash-based microcontrollers
contain between 1 and 128 KB of on-chip program storage. This can be used as both
program memory and as temporary data storage. Additionally they contain between 128
and 32KB of data ram that can be used for program execution.
Modern flash technology currently produces memory storage densities upwards of
150 KB per square millimeter in a .25 micron process [25]. This is significantly higher
than Intel’s recent SRAM density record of 60 KB per square millimeter set using a 90
nm process [26]. While it is currently only possible to achieve these storage densities on
chips that are dedicated to performing one specific function, we can conclude that there is
47
a significant memory density advantage for flash over SRAM. In order to achieve the
smallest, lowest cost device, SRAM should be used only when necessary.
In addition to considering the size impact of flash versus SRAM, we should also
consider the power impact. SRAM requires more energy to retain data over time, but
does not require as much energy for the initial storage operation. Flash, on the other
hand, is a persistent storage technology that requires no energy to maintain data. In
contrast, AMD SRAM chips require 2188 nA per megabit for data retention [25].
However, a flash write operation requires 4 us to complete compared to .07 us for SRAM
– both consuming 15 mA. If data is to be stored for long periods of time it is more
efficient to use flash instead of SRAM. However, for storage operation to overcome the
cost of performing the flash write operation it must be stored for at least 7.6 hours. Any
shorter and the initial cost of storage is not outweighed by the decreased cost of retention.
2.4.3.6 Peripheral support
In addition to storage and computation, microcontrollers are specifically designed
to interact with external devices. Microcontrollers generally contain a collection of
specialized external interfaces. These range from general purpose digital I/O pins to
digital communication interfaced to analog inputs and outputs.
Standard digital I/O lines are included on all controllers as the base line interface
mechanism. However, to interface directly with analog sensors many controllers have
an internal analog-to-digital converter. An internal converter allows for precise control
of sample timing and easy access to sample results. If not present, an external converter
must be used.
48
The digital communication interfaces come into play when interacting with digital
sensors, other peripheral chips, or for serial communications over a radio or RS-232
transceiver. There are 3 standard communication protocols supported by digital
communication primitives: UART, I2C and SPI. Both I2C and SPI use synchronous
protocols with explicit clock signals while UART provides an asynchronous mechanism.
2.4.4 Sensors
The last decade has seen an explosion in senor technology. There are currently
thousands of potential sensors ready to be attached to a wireless sensing platform.
Additionally, advances in MEMS and carbon nano-tubes technology are promising to
create a wide array of new sensors. They range from simple light and temperature
monitoring sensors to complex digital noses. Figure 2-2 outlines a collection of common
micro-sensors and their key characteristics.
2.4.4.1 Interfaces
There are two general ways to interface with sensors that can be used in sensor
Current
Discrete Sample
Time
Voltage
Requiremet Manufacturer
Photo 1.9 mA 330 uS 2.7-5.5V Taos
Temperature 1 mA 400 mS 2.5-5.5V Dallas Semiconductor
Humitidy 550 uA 300 mS 2.4-5.5V Sensirion
Pressure 1 mA 35 mS 2.2V-3.6V Intersema
Magnetic Fields 4 mA 30 uS Any Honeywell
Acceleration 2 mA 10 mS 2.5-3.3V Analog Devices
Acoustic .5 mA 1 mS 2-10 V Panasonic
Smoke 5 uA -- 6-12 V Motorola
Passive IR (Motion) 0 mA 1 mS Any Melixis
Photosynthetic Light 0 mA 1 mS Any Li-Cor
Soil Moisture 2 mA 10 mS 2-5 V Ech2o
Commonly avilable sensors
Figure 2-2: Power consumption and capabilities of commonly available.
49
networks: analog and digital. Analog sensors generally provide a raw analog voltage that
corresponds to the physical phenomena that they are measuring. Generally these
produce a continual waveform that must be digitized and then analyzed. While
seemingly straightforward to integrate, raw analog sensors often require external
calibration and linearization. It is common for the sensor to have non-linear response to
stimuli. The host controller must then compensate in order to produce a reading in
meaningful units. For example, a raw reading of 1.873 Volts from an accelerometer
must be translated into a corresponding acceleration. Depending on the characteristics of
the sensor this can be a complex process. In many cases the translation may depend on
other external factors such as temperature, pressure, or input voltage.
A second difficulty in interfacing with raw analog sensors is that of scale. Each
sensor will have different timing and voltage scales. The output voltage will generally
contain a DC offset combined with a time-varying signal. Depending on the ratio of
signal to DC component, an array of amplifiers and filters may be required to match the
output of the sensor to the range and fidelity of the ADC being used to capture it.
Digital sensors have been developed to remove many of these difficulties. They
internally contain a raw analog sensor but provide a clean digital interface to it. All
required compensation and linearization is performed internally. The output is a digital
reading on an appropriate scale. The interface to these sensors is via one of a handful of
standard chip-to-chip communication protocols.
2.4.4.2 Sample Rates and Turn-on times
In looking at the power consumption of a sensor, one of the most important
factors is how quickly a sensor can be enabled, sampled, and disabled. In most cases
50
sensors are capable of producing thousands of samples per second. However, we are
generally only interested in a few samples per minute. It is essential that the sensor be
able to enter and exit a low-power state quickly.
The power consumption of a sensor is equally dependent on the amount of time it
takes to read the sensor as it is to the current consumption. For example, if a sensor takes
100 ms to turn on and generate a reading and consumes just 1 mA at 3 V, it will cost 300
uJ per sample. This is the same amount of energy as a sensor that consumes 1000 mA of
current at 3 V but takes just 100 us to turn on and sample.
2.4.4.3 Voltage requirements
A seemingly low-power, easy to interface sensor can quickly become a major
hassle if its voltage requirements do not match the capabilities of the system. Some
sensors require +/- 6 V. Special voltage converters and regulators have to be added to
systems operating off of AA or lithium batteries in order to use this sensor. The power
consumption and turn-on times of the regulation and conversion circuitry must be
included in the total energy budget for the sensor.
2.5 Rene Design Point
Now that we have provided a background of hardware capabilities, we present a
baseline networked sensor platform. The Rene design point is used as a comparison
point in order to evaluate the more advanced architecture developed over the subsequent
chapters. A photograph for the Rene hardware including sensor and programming
boards appear in Figure 2-3. It consists of a microcontroller with internal flash program
memory, data SRAM and data EEPROM, connected to a set of actuator and sensor
51
devices, including LEDs, a low-power radio transceiver, a digital temperature sensor, a
serial port, and a small coprocessor unit and an I/O connector that allows a suite of
sensors to be attached. The basic design is based on the weC mote, but was redesigned
to serve as an experimental and developments platform. The modifications included a
rich connector interfaced that allowed for stackable sensor and power boards. This
allowes for application specific sensor and power boards.
2.5.1.1 CPU
The processor on the Rene platform is the ATMEL 90LS8535 [27]. It is an 8-bit
Harvard architecture with 16-bit addresses. It provides 32 8-bit general purpose registers
and runs at 4 MHz and 3.0 V. The system is very memory constrained: it has 8 KB of
flash as the program memory, and 512 bytes of SRAM as the data memory. The MCU is
designed such that a processor cannot write to instruction memory; our prototype uses a
coprocessor to perform that function. Additionally, the processor integrates a set of
timers and counters which can be configured to generate interrupts at regular time
intervals. In conjunction with a 32 KHz external crystal, the counters can be used to
keep precise time. Additionally, it also contains an internal analog to digital converter
for reading analog sensor inputs.
2.5.1.2 Storage
52
Besides the storage contained inside the main microcontroller, the Rene hardware
platform also contains an external EEPROM capable of holding 32 kilobytes of data.
This can be used to temporarily record sensor readings or to act as storage for additional
programs.
The CPU interacts with the EPROM via a two wire I2C [28] data communication
bus. The I2C bus is a standard communication mechanism in embedded systems. The
bus architecture allows for a collection of 8 peripherals to be controlled by a single
master via two communication lines and a ground connection.
2.5.1.3 RF Subsystem
The radio is the most important component. It consists of an RF Monolithics
916.50 MHz transceiver (TR1000) [22], antenna, and collection of discrete components
to configure the physical layer characteristics such as signal strength and sensitivity. As
configured, it operates in an ON-OFF key mode at speeds up to 19.2 Kbps but is only
driven at 10 Kbps. Control signals configure the radio to operate in either transmit,
receive, or power-off mode. The radio contains no buffering so each bit must be
Figure 2-3: Pictures of the Rene wireless sensor network platform.
53
serviced by the controller on time. Additionally, the transmitted value is not latched by
the radio, so jitter at the radio input is propagated into the transmitted signal.
2.5.1.4 Coprocessor
A coprocessor is included on the platform to assist in reprogramming. In this
case, it is a very limited controller (AT90LS2343 [29], with 2 KB flash instruction
memory, 128 bytes of SRAM and EEPROM) that uses I/O pins connected to an SPI
controller. SPI is a synchronous serial data link, providing high speed full-duplex
connections (up to 1 Mbit) between various peripherals. The coprocessor is connected in
a way that allows it to reprogram the main microcontroller. The sensor node can be
reprogrammed by transferring data from the network into external EEPROM. Then the
coprocessor can reset the main controller and reprogram it with the new program. This
is required because the main CPU cannot program its own memory.
2.5.1.5 I/O Interface
Three LEDs represent analog outputs connected through a general I/O port; they
may be used to display digital values or status. A photo-sensor represents an analog
input device with simple control lines. In this case, a control line enables and disables
the sensor by cutting the power drain through the photo resistor when not in use. The
input signal can be directed to the internal ADC in continuous or sampled modes.
2.5.2 Baseline Power Characteristics
The AT908535 microcontroller consumes just 5 mA when operating at 4 MHz.
Additionally, it has two low-power states – idle and sleep. When in the idle state, the
54
controller consumes just 3 mA. The idle state can be entered and exited with just one
instruction penalty, making it useful for brief periods of inactivity between computations.
In contrast, the sleep sate consumes just 10 uA of current, but requires 2 milliseconds to
enter and exit. The sleep state can only be used in conjunction with disabling the radio.
The minimum pulse width for the RFM radio is 52 us. Thus, at 3V and 7 mA, it
takes 1.1 uJ of energy to transmit a single bit. Alternatively, it costs approximately 0.5
uJ to receive a bit. During this time, the processor can execute 208 cycles (roughly 100
instructions) and can consume up to .8 uJ. A fraction of this instruction count is devoted
to bit level processing. The remainder can go to higher level processing (byte-level,
packet level, application level) amortized over several bit times. Unused time can be
spent in power-down mode or on application processing.
While listening for incoming communications, the radio will continually consume
5 mA regardless of what is being received. In addition, the CPU must actively search the
incoming bit stream for a valid transmission. The CPU will continually be switching
between its 3 mA idle state and its 5 mA active. Because the CPU must inspect each
arriving bit, it is not possible to enter a low-power sleep mode while scanning for the start
of a packet. Even if we ignore the active power consumption, the node will be
consuming approximately 8 mA to search for valid data.
2.6 Refined Problem Statement
The potential usage scenarios for wireless sensor networks are countless. In order
to reduce the problem space, we have chosen to focus on three distinct application
scenarios that we believe are representative of a large fraction of the potential
applications. These three scenarios – environmental data collection, security monitoring,
55
and node tracking – allow us to extract the evaluation metrics that can be used to measure
the performance of a wireless sensor network. From these system-level evaluation
metrics we have drilled down into the individual node capabilities that will support them.
Our objective is to develop a wireless sensor node architecture that addresses the
node-level requirements that we have set forth. In looking at several of the system-level
metrics, it is apparent that a balance must be obtained between range, bit rate, duty cycle
and power consumption in order to create a system that meets both lifetime and
performance requirements. We develop a generalized node architecture and evaluate it in
depth over a series of instantiations at technology points ranging from off-the-shelf
integrations to a full custom ASIC.
In the next chapter we will look into the software architecture that plays a critical
role in realizing the system capabilities. While often ignored, the software subsystem
plays as important a roll as any radio, processor or battery. The TinyOS operating system
presented will provide a concurrency and flexibility mechanism that allow the
development of a generalized wireless sensor node architecture that is substantially
different from traditional wireless devices.
56
Chapter 3: Software Architecture for Wireless
Sensors
A critical step towards achieving the vision behind wireless sensor networks is the
design of a software architecture that bridges the gap between raw hardware capabilities
and a complete system. The demands placed on the software of wireless sensor
networks are numerous. It must be efficient in terms of memory, processor, and power
so that it meets strict application requirements. It must also be agile enough to allow
multiple applications to simultaneously use system resources such as communication,
computation and memory. The extreme constraints of these devices make it impractical
to use legacy systems. TinyOS is an operating designed explicitly for networked
sensors.
TinyOS draws strongly from previous architectural work on lightweight thread
support and efficient network interfaces. Included in the TinyOS system architecture is
an Active Messages communication system. We believe that there is a fundamental fit
between the event based nature of network sensor applications and the event based
primitives of the Active Messages communication model.
In working with wireless sensor networks, two issues emerge strongly: these
devices are concurrency intensive - several different flows of data must be kept moving
simultaneously, and the system must provide efficient modularity - hardware specific and
application specific components must snap together with little processing and storage
overhead. We address these two problems in the context of current network sensor
57
technology and the tiny micro threaded OS. Analysis of this solution provides valuable
initial directions for architectural innovation.
3.1 Tiny Microthreading Operating System (TinyOS)
Small physical size, modest active power load, and micro standby power load are
must be provided by the hardware design. However, an operating system framework is
needed that will retain these characteristics by managing the hardware capabilities
effectively, while supporting concurrency-intensive operation in a manner that achieves
efficient modularity and robustness. Existing embedded device operating systems do
not meet the size, power and efficiency requirements of this regime. These
requirements are strikingly similar to that of building efficient network interfaces, which
also must maintain a large number of concurrent flows and juggle numerous outstanding
events [30]. In network interface cards, these issues have been tackled through physical
parallelism [31] and virtual machines [32]. We tackle it by building an extremely
efficient multithreading engine. As in TAM [32] and CILK [33], TinyOS maintains a
two-level scheduling structure, so a small amount of processing associated with hardware
events can be performed immediately while long running tasks are interrupted. The
execution model is similar to finite state machine models, but considerably more
programmable.
TinyOS is designed to scale with the current technology trends supporting both
smaller, tightly integrated designs, as well as the crossover of software components into
hardware. This is in contrast to traditional notions of scalability that are centered on
scaling up total power/resources/work for a given computing paradigm. It is essential
58
that software architecture plans for the eventual integration of sensors, processing and
communication.
In order to enable the vision of single-chip, a low cost sensor node, TinyOS
combines a highly efficient execution model, component model and communication
mechanisms.
3.2 TinyOS Execution Model
To provide the extreme levels of operating efficiency required in wireless sensor
networks, TinyOS uses event based execution. The event model allows for high levels
of concurrency to be handled in a very small amount of space. In contrast, a thread
based approach requires that stack space be pre-allocated for each execution context.
Additionally, the context switch overhead of threaded systems is significantly higher than
those of an event-base system. Context switch rates as high as 40,000 switches per
second are required for the base-band processing of a 19.2 Kbps communication rate.
This is twice every 50 us, once to service the radio and once to perform all other work.
The efficiency of an event-based regime lends itself to these requirements.
3.2.1 Event Based Programming
In an event based system, a single execution context is shared between unrelated
processing tasks. In TinyOS, each system module is designed to operate by continually
responding to incoming events. When an event arrives, it brings the required execution
context with it. When the event processing is completed, it is returned back to the
system. Researchers in the area of high performance computing have also seen that event
59
based programming must be used to achieve high performance in concurrency intensive
applications [34, 35].
In addition to efficient CPU allocation, event-based design results in low power
operation. A key to limiting power consumption is to identify when there is no useful
work to be performed and to enter an ultra-low power state. Event-based systems force
applications to implicitly declare when they are finished using the CPU. In TinyOS all
tasks associated with an event are handled rapidly after each event is signaled. When an
event and all tasks are fully processed, unused CPU cycles are spent in the sleep state as
opposed to actively looking for the next interesting event. Eliminating blocking and
polling prevents unnecessary CPU activity.
3.2.2 Tasks
A limiting factor of an event based program is that long-running calculations can
disrupt the execution of other time critical subsystems. If an event were to never
complete, all other system functions would halt. To allow for long running computation,
TinyOS provides an execution mechanism called tasks. A task is an execution context
that runs to completion in the background without interfering with other system events.
Tasks can be scheduled at any time but will not execute until current pending events are
completed. Additionally, tasks can be interrupted by low-level system events. Tasks
allow long running computation to occur in the background while system event
processing continues.
Currently task scheduling is performed using a simple FIFO scheduling queue.
While it is possible to efficiently implement priority scheduling for tasks, it is unusual to
60
have multiple outstanding tasks. A FIFO queue has proven adequate for all application
scenarios we have attempted to date.
3.2.3 Atomicity
In addition to providing a mechanism for long-running computation, the TinyOS
task primitive also provides an elegant mechanism for creating mutually exclusive
sections of code. In interrupt-based programming, data race conditions create bugs that
are difficult to detect. In TinyOS, code that is executed inside of a task is guaranteed to
run to completion without being interrupted by other tasks. This guarantee means that
all tasks are atomic with respect to other tasks and eliminates the possibility of data race
conditions between tasks.
Low-level system components must deal with the complexities associated with
reentrant, interrupt based code in order to meet their strict real-time requirements.
Normally, only simple operations are performed at the interrupt level to integrate data
with ongoing computation. Applications can use tasks to guarantee that all data
modification occurs atomically when viewed from the context of other tasks.
3.3 TinyOS Component Model
In addition to using the highly efficient event-based execution, TinyOS also
includes a specially designed component model targeting highly efficient modularity and
easy composition. An efficient component model is essential for embedded systems to
increase reliability without sacrificing performance. The component model allows an
application developer to be able to easily combine independent components into an
application specific configuration.
61
In TinyOS, each module is defined by the set of commands and events that makes
up its interface. In turn, a complete system specification is a listing of the components to
include plus a specification for the interconnection between components. The TinyOS
component has four interrelated parts: a set of command handlers, a set of event handlers,
an encapsulated private data frame, and a bundle of simple tasks. Tasks, commands, and
event handlers execute in the context of the frame and operate on its state. To facilitate
modularity, each component also declares the commands it uses and the events it signals.
These declarations are used to facilitate the composition process. As shown in Figure
3-1, composition creates a graph of components where high level components issue
commands to lower level components and lower level components signal events to the
higher level components. The lowest layer of components interacts directly with the
underlying hardware. A specialized language, NESC, has been developed to express the
component graph and the command/event interfaces between components. In NESC,
RFM
Radio byte
Radio Packet
UART
Serial Packet
photo
Active Messages
clocks
bit
byte
packet
Ad hocapplication Routing Application
HW
SW
message
Figure 3-1: Component graph for a multi-hop sensing
application.
62
multiple command and events can be grouped together into interfaces. Interfaces
simplify the interconnection between components.
In TinyOS, storage frames are statically allocated to allow the memory
requirements of the complete application to be determined at compile time. The frame is
a specialized C Structure that is statically allocated and directly accessible only to the
component. While TinyOS does not have memory protection, variables cannot be
directly accessed from outside of a component. In addition to allowing the calculation of
maximum memory requirements, pre-allocation of frames prevents the overhead
associated with dynamic allocation and avoids pointer related errors. This savings
manifests itself in many ways, including execution time savings because variable
locations can be statically compiled into the program instead of accessing state via
pointers.
In TinyOS, commands are non-blocking requests made to lower level
components. Typically, a command will deposit request parameters into its local frame
and conditionally post a task for later execution. It may also invoke lower commands,
but it must not wait for long or indeterminate latency actions to take place. A command
must provide feedback to its caller by returning status indicating whether it was
successful or not, e.g., buffer overrun. Event handlers are invoked to deal with hardware
events, either directly or indirectly. The lowest level components have handlers that are
connected directly to hardware interrupts, which may be external interrupts, timer events,
or counter events. An event handler can deposit information into its frame, post tasks,
signal higher level events or call lower level commands. A hardware event triggers a
fountain of processing that goes upward through events and can bend downward through
63
commands. In order to avoid cycles in the command/event chain, commands cannot
signal events. Both commands and events are intended to perform a small, fixed amount
of work, which occurs within the context of their component’s state.
Tasks perform the primary work in a TinyOS application. They are atomic with
respect to other tasks and run to completion, though they can be preempted by events.
Tasks can call lower level commands, signal higher level events, and schedule other tasks
within a component. The run-to-completion semantics of tasks make it possible to
allocate a single stack that is assigned to the currently executing task. This is essential in
memory constrained systems.
Tasks simulate concurrency within each component, since they execute
asynchronously with respect to events. However, tasks cannot block or spin wait or they
will prevent progress in other components. While events and commands approximate
instantaneous state transitions, task bundles provide a way to incorporate arbitrary
computation into the event driven model.
3.3.1 Component Types
In general, components fall into one of three categories: hardware abstractions,
synthetic hardware, and high level software components. Hardware abstraction
components map physical hardware into the TinyOS component model. The RFM radio
component (shown in lower left corner of Figure 3-1) is representative of this class. This
component exports commands to manipulate the individual I/O pins connected to the
RFM transceiver and signals events informing other components about the transmission
and reception of bits. Its frame contains information about the current state of the
64
component, (the transceiver is in sending or receiving mode, the current bit rate, etc.).
The RFM consumes hardware interrupts that are transformed into either the RX bit
events or into the TX bit events, depending on the mode of operation. There are no tasks
within the RFM because the hardware itself provides the concurrency. This model of
abstracting over the hardware resources can scale from very simple resources, like
individual I/O pins, to quite complex ones, like encryption accelerators.
Synthetic hardware components simulate the behavior of advanced hardware. A
good example of such component is the Radio Byte component (see Figure 3-1). It
shifts data into or out of the underlying RFM module and signals when an entire byte has
completed. The internal tasks perform simple encoding and decoding of the data.
Conceptually, this component is an enhanced state machine that could be directly cast
into hardware. From the point of view of the higher levels, this component provides an
(addr,type,data)
send_msg
send_msg_thread internal
state
RX_packet_done
(buffer)
TX_packet_done
(success)
Messaging
Component
msg_send_done(success)
msg_rec(type, data)
init
power(mode)
init
power(mode)
TX_packet(buf)
Figure 3-2: AM Messaging component
graphical representation.
65
interface and functionality very similar to the UART hardware abstraction component:
they provide the same commands and signal the same events, deal with data of the same
granularity, and internally perform similar tasks (looking for a start bit or symbol,
perform simple encoding, etc.).
The high level software components perform control, routing and all data
transfers. A representative of this class is the messaging module presented in Figure 3-2.
It performs the function of filling in a packet buffer prior to transmission and dispatches
received messages to their appropriate place. Additionally, components that perform
calculations on data or data aggregation fall into this category.
3.3.2 Enabling the migration of hardware/software boundary
The TinyOS component model allows for easy migration of the
hardware/software boundary because the event based software model it uses is
complementary to the underlying hardware structure. Additionally, the use of fixed size,
pre-allocated storage is a requirement for hardware based implementations. Migration
from software to hardware is particularly important for networked sensors where the
system designers will want to explore the tradeoffs between the scale of integration,
power requirements, and the cost of the system. Chapter 6 presents a customized design
that pushes several system components from the Rene node across the hardware/software
boundary.
66
3.3.3 Example Components
A typical component including a frame, event handlers, commands and tasks for a
message handling component is pictured in Figure 3-2. Like most components, it
exports commands for initialization and power management. Additionally, it has a
command for initiating a message transmission, and signals events on the completion of a
transmission or the arrival of a message. In order to perform its function, the messaging
component issues commands to a packet level component and handles two types of
events: one that indicates a message has been transmitted and one that signals that a
message has been received.
module AMStandard
{
provides {
interface StdControl as Control;
interface CommControl;
// The interface are as parameterised by the active message id
interface SendMsg[uint8_t id];
interface ReceiveMsg[uint8_t id];
// How many packets were received in the past second
command uint16_t activity();
}
uses {
// signaled after every send completion for components which wish to
// retry failed sends
event result_t sendDone();
interface StdControl as UARTControl;
interface BareSendMsg as UARTSend;
interface ReceiveMsg as UARTReceive;
interface StdControl as RadioControl;
interface BareSendMsg as RadioSend;
interface ReceiveMsg as RadioReceive;
interface Leds;
//interface Timer as ActivityTimer;
}
}
Figure 3-3: NESC component file describing the external interface to the
AMStandard messaging component.
67
Pictorially, we represent the component as a bundle of tasks, a block of state
(component frame), a set of commands (upside-down triangles), a set of handlers
(triangles), solid downward arcs for commands they use, and dashed upward arcs for
events they signal. Figure 3-3 shows how all of these elements are explicit in the
component code. Since the components describe both the resources they provide and the
resources they require, connecting them together is simple.
The programmer matches the signatures of events and commands required by one
component with the signatures of events and commands provided by another component.
Communications across the components takes the form of a function call, which has low
overhead and provides compile time type checking.
3.3.4 Component Composition
In order to support the modular nature of TinyOS, we have developed a set of
tools that assist the developer to link components together. In TinyOS, components are
linked together at compile time to eliminate unnecessary runtime overhead. To facilitate
composition, each component’s external interface is described at the start of each
component file. In these files, a component lists the set of commands it accepts and the
events it handles as well as the set of events it signals and the commands it uses.
Logically we view each of the inputs and outputs of a component as an I/O pin as if the
component were a physical piece of hardware. This complete description of a
component’s upper and lower interface is used by our compile time tools to automatically
generate component header files. Figure 3-4 contains an example component file for a
simple application that blinks the LEDs. In order actually compose individual
components into a complete application; TinyOS originally used description (.desc) files
68
/**
* Implementation for Blink application. Toggle the red LED when the
* clock fires.
**/
module BlinkM {
provides {
interface StdControl;
}
uses {
interface Clock;
interface Leds;
}
}
implementation {
/**
* the state of the red LED (on or off)
*/
bool state;
/**
* Initialize the component.
*
* @return Always returns SUCCESS
**/
command result_t StdControl.init() {
state = FALSE;
call Leds.init();
return SUCCESS;
}
/**
* Start things up. This just sets the rate for the clock component.
*
* @return Always returns SUCCESS
**/
command result_t StdControl.start() {
return call Clock.setRate(TOS_I1PS, TOS_S1PS);
}
/**
* Halt execution of the application.
* This just disables the clock component.
*
* @return Always returns SUCCESS
**/
command result_t StdControl.stop() {
return call Clock.setRate(TOS_I0PS, TOS_S0PS);
}
/**
* Toggle the red LED in response to the Clock.fire event.
*
* @return Always returns SUCCESS
**/
event result_t Clock.fire()
{
state = !state;
if (state) {
call Leds.redOn();
} else {
call Leds.redOff();
}
return SUCCESS;
}
}
Figure 3-4: BlinkM.nc application component that blinks the system LED's once per second.
69
that contained a listing of components to be used as well as the logical connections
between the components. The success of the TinyOS component model led to the
development a specialized language called NESC that helps facilitate the expression of
these logical connections [36].
The application description files can be thought of as a parts lists and a wiring
diagram. These description files are divided into two sections. The first section directly
lists the modules to include in the application. The second section of the file lists the
connections between each component. This wiring diagram takes the form of lists of
ports that are logically connected. Each line of the file contains lists a pair of module
interfaces that are to be connected together with a wiring network. Figure 3-5 contains
and example description file for a simple NESC application.
At compile time, the component description files and the application description
files are preprocessed in order to create a net list that links the components together. Just
as in hardware, a net in our system can have multiple sources and sinks. This is
automatically handled by the NESC compiler. For example, a single event can be
handled by multiple components. At compile time, code will be automatically generated
configuration Blink {
}
implementation {
components Main, BlinkM, ClockC, LedsC,
DustKernelWrapper;
Main.StdControl -> BlinkM.StdControl;
BlinkM.Clock -> ClockC;
BlinkM.Leds -> LedsC;
}
Figure 3-5: NESC application configuration file that wires
together the Blink application.
70
to send the event to as many places as necessary. The output of the NESC compiler is a
standard C file that contains all of the components in the application as well as all of the
necessary linkage information.
3.3.5 Application Walk Through
In order to show what applications look like in TinyOS, we walk through a simple
application that is included in the TinyOS 1.0 release. The application, BLINK, simply
turns on and off the LED’s of the system in a binary counting fashion. While this
application could easily be written without the help of the TinyOS models, it
demonstrates some of the core concepts in TinyOS.
Figure 3-4 shows the module description for the Blink application’s main
component (from BlinkM.nc). From the module declaration, you can see that the
BlinkM component provides the TinyOS StdControl interface and uses the Clock
interface and the LEDs interface. This interface file completely describes what
functionality needs to be provided by other parts of the system for this component to
function correctly. In particular, this component needs to be controlled with the
StdControl interface, and it needs to have access to the clock interface and the LEDs.
The actual code can then be written to fill in this component without knowledge of what
will be providing the necessary support.
All variables declared inside the boundaries of the module are locally scoped
private variables. They are placed inside the component’s private frame. In this
application the module has a single variable called state. This variable contains the
current value of the binary counter being displayed on the LED’s. The second section of
71
the code contains the implementation of the StdControl interface. This interface is
designed to handle initialization. Functions are declared by specifying the type of
function (command/event), the return type, the name of the interface they belong to, the
individual function name, and the argument list. The init function is a command that
returns a result_t and accepts no arguments and is part of the StdControl interface. It
performs all of the necessary initialization for this component. Additionally, it must
initialize all sub-components that are used. In this example, the init function issues a
single command to initialize the LED. The invocation of commands are performed
using the ‘call’ keyword. Similarly, events are signaled using the signal keyword. The
keywords call and signal explicitly state when component boundaries are crossed. This
information is used by the NESC compiler in full-program analysis to optimize away
component boundaries.
Note that the application never holds onto an execution context. Both the
initialization routine and the event handlers release the execution context quickly. This
is an essential property of a TinyOS application. By not holding onto an execution
context, a single context can be shared across multiple concurrent activities. If this
application were to be implemented in a threaded execution model, as opposed to the
event based TinyOS model, this simple function would require a dedicated execution
context. It would block until it was time to update the counter. This would consume a
significant amount of resources when idle. In contrast, this application simply consumes
a single byte of memory while idle.
In addition to writing the individual application level components, the application
developer must also assemble a collection of TinyOS components into a complete
72
application. Figure 3-5 shows the application configuration file for this application. The
first section of the file contains the list of components that need to be included. In this
case, the complete application consists of the Main component, the BlinkM component,
the ClockC component and the LedC component. The second section of the file then
contains the connections between the components. The first line of this section states
that the Main component’s StdControl interface is connected to the BlinkM component’s
StdControl interface. The use of this component model allows applications to be easily
reconfigured.
3.4 AM Communication Paradigm
A key piece of TinyOS is the communication model it supports. Instead of
scaling a PC based communication models down, we believe that it is beneficial to tailor
the communication system to meet the needs of these devices. To accomplish this, we
have used the Active Messages communication model as primary building blocks for
networking in TinyOS.
3.4.1 Active Messages Overview
Active Messages (AM) is a simple, extensible paradigm for message-based
communication widely used in parallel and distributed computing systems [35]. Each
Active Message contains the name of an application-level handler to be invoked on a
target node upon arrival and a data payload to pass in as arguments. The handler function
serves the dual purpose of extracting the message from the network and either integrating
the data into the computation or sending a response message. The network is modeled as
a pipeline with minimal buffering for messages. This eliminates many of the buffering
73
difficulties faced by communication schemes that use blocking protocols or special
send/receive buffers. To prevent network congestion and ensure adequate performance,
message handlers must be able to execute quickly and asynchronously.
Although Active Messages has its roots in large parallel processing machines and
computing clusters, the same basic concepts can be used to the meet the constraints of
networked sensors. Specifically, the lightweight architecture of Active Messages can be
leveraged to balance the need for an extensible communication framework while
maintaining efficiency and agility. More importantly, the event based handler invocation
model allows application developers to avoid busy-waiting for data to arrive and allows
the system to overlap communication with other activities such as interacting with
sensors or executing other applications. It is this event centric nature of Active
Messages which makes it a natural fit for TinyOS.
3.4.2 Tiny Active Messages Implementation
In bringing Active Messages out of the high performance parallel computing
world and down into this low power design regime, we have attempted to preserve the
basic concepts of integrating communication with computation and matching
communication primitives to hardware capabilities. The overlap of computational work
with application level communication is essential. Execution contexts and stack space
must never be wasted because applications are blocked, waiting for communication. The
Active Messages communication model can be viewed as a distributed event model
where networked nodes send each other events. While quite basic, we believe that all
applications can be built on top of this primitive model.
74
In order to make the Active Messages communication model a reality, certain
primitives must be provided by the system. We believe that the three required primitives
are: best effort message transmission with acknowledgements, addressing, and dispatch.
More demanding applications may need to build more functionality on top of these
primitives, but that is left for the applications developer to decide. By creating the
minimal kernel of a communication system, all applications will be able to build on top
of it. Additionally, it is likely that there will be a large variety of devices with different
physical communication capabilities and needs. By building the communication kernel
as separate components in the TinyOS component model, developers can pick and choose
which implementations of the basic components they need. This can take the form of
selecting from a collection of delivery components that perform different levels of error
correction and detection. However, by providing a consistent interface to
communication primitives, application developers can easily transfer their applications to
different hardware platforms.
Just as there will be various implementations of the core components for the
developer to choose from, various other extensions will be available, such as reliable
delivery. This is similar to the design of Horus [37], which attempted to have modular
PC-based communication protocols where application developers could choose from a
variety of building blocks including encryption, flow control, and packet fragmentation.
It is extremely advantageous to be able to customize protocols when dealing with
network sensors due to their extreme performance constraints and their large diversity of
physical hardware.
75
3.4.3 Buffer swapping memory management
Fundamental to the TinyOS communication system is its storage model. As data
arrives over the radio, it must be stored into a memory buffer. The active messages
dispatch layer then delivers the buffer up to the applications. In many cases, applications
will wish to keep the message buffers that are delivered to them. In the case of the
multi-hop communication, applications would need to keep the buffer long enough for it
to be transmitted to the next node.
This situation is problematic in systems that do not support dynamic memory
allocation because the radio subsystem must continually acquire new buffers to deliver to
the applications. To handle this, TinyOS requires that each application return an unused
message buffer to the radio subsystem each time a message is delivered. The radio
simply maintains one extra buffer to receive the next message into. After reception the
buffer is transferred up to the application. Then a free buffer is handed back by the
application component to the radio system. This free buffer is then filled by the next
message.
Through this buffer swapping scheme, a fixed set of pre-allocated messaging
buffers can be used. If an application needs the ability to store several messages
simultaneously, it simply allocates extra message buffers inside its own private frame.
As messages arrive that need to be preserved, the application can return the pre-allocated
buffers. It must maintain which buffers are empty and which are storing useful data. At
all times, the application maintains control of a fixed number of buffers. The actual
buffers are traded throughout the system as communication occurs.
76
3.4.4 Explicit Acknowledgement
While TinyOS only provides best-effort message delivery, we have found that it
is important to receive feedback as to whether a transmission was successfully received
by the next intended recipient. This simple piece of feedback greatly simplifies
application level routing and reliable delivery algorithms.
In TinyOS 1.0, each messages transmission is followed by an immediate,
synchronous acknowledgement that is transmitted from the receiver back to the
transmitter. This acknowledgement is transmitted by an addressed recipient if a packet
successfully passes a CRC check.
This form of packet acknowledgement is included at the physical layer of the
TinyOS communication stack because it is significantly less expensive than providing the
same functionality at the application level. Instead of transmitting a complete packet
that contains an acknowledgement, as well as the information required to match the
acknowledgement with the packet it is acknowledged, the TinyOS acknowledgement
mechanism simply transmits a special sequence immediately after the packet is received.
In addition to reducing the amount of energy it takes to acknowledge a packet,
this mechanism also simplifies the task of processing the acknowledgement. Because
the timing of the acknowledgement is synchronous with respect to the packet
transmission time, applications do not have to wait for a timeout to occur prior to
determining that a transmission failed. The transmitting application knows the fate of
the packet immediately after transmission. This allows the application to immediately
initiate a retransmission or to search for a new recipient.
77
3.5 Components Contained in TinyOS
There are several system level components that are included in TinyOS.
Additionally, there are a collection of example applications that demonstrate the usages
of the TinyOS system. Figure 3-6 contains a listing of the components contained in the
BapM.nc Beaconless ad-hoc routing protocol
Counter.nc Continually incrementing counter
IdentC.nc Component to retrieve and set node identity
IntToLedsM.nc Display an integer value on the LEDs
IntToRfmM.nc Report an integer value over the radio
OscopeM.nc
Component that continually samples and
reports and ADC channel
ResetC.nc System reset component
RfmToIntM.nc Receive and integer over the radio
Route Multi-hop routing suite
SenseToInt.nc
Generate and iteger valued sensor reading
from the ADC
TinyAlloc.nc Memory allocator
TinyDB Database application for TinyOS
ADCM.nc Interface to the ADC
AMStandard.nc Active Messages implemetnation
ByteEEPROM.nc EEProm access componet
CRCPacket.nc CRC packet calculator
ClockC.nc Timing componet
CrcFilter.nc CRC packet filter
GenericComm.nc Generic communication stack for general use
I2CPacketC.nc I2C protocol implementation
LedsC.nc LED interface
LoggerM.nc Interface to log data to the off-chip flash
NoCRCPacket.nc Packet componet without CRC calculation
PotM.nc Signal strength control
RFM.nc TR1000 Radio interface
RadioCRCPacket.nc TR1000 Packet interface with CRC
RadioNoCRCPacket.nc TR1000 Packet interface without CRC
RandomLFSR.nc Random Number Generator
SecDedRadioByteSignal.nc Forward error correction component
TimerM.nc Multi-application timer module
UARTComm.nc UART communicaitons stack
UART.nc UART interface module
UARTNoCRCPacket.nc UART communication stack without CRC
VoltageM.nc Battery voltage measurement component
TinyOS 1.0 System Components
Figure 3-6: Listing of components contained in TinyOS 1.0. Application
developers select from this collection of system-level components to create the
desired application attributes.
78
current TinyOS release.
3.6 TinyOS Model Evaluation
To evaluate the performance of the TinyOS system model we return to our set of
key software requirements: Small physical size, concurrency intensive, efficient
modularity, limited physical parallelism, and robust.
Small physical size: Application code size is our primary metric of performance
for small physical size. TinyOS enables incredibly small applications. In particular, our
scheduler only occupies 178 bytes and our complete network sensor application requires
only about 3KB of instruction memory. Furthermore, the data size of our scheduler is
only 16 bytes, which utilizes only 3% of the available data memory on the baseline Rene
platform. An entire multi-hop data collection application comes in at 226 bytes of data,
still under 50% of the 512 bytes available on the Rene platform.
Concurrency-intensive operations: It is essential that the software architecture be
able to provide efficient concurrency; network sensors need to handle multiple flows of
information simultaneously. In this context, an important baseline characteristic of a
network sensor is its context switch speed. In the TinyOS event based system, the cost
of propagating an event is roughly equivalent to that of copying one byte of data. This
low overhead is essential for achieving modular efficiency. Posting a thread and
switching context costs about as much as moving 6 bytes of memory. Our most
expensive operation involves the low-level aspects of interrupt handling. Though the
hardware operations for handling interrupts are fast, the software operations that save and
restore registers in memory impose a significant overhead. Several techniques can be
79
used to reduce that overhead such as partitioning the register set [32] or use of register
windows [38].
Efficient modularity: One of the key characteristics of our systems is that events
and commands can propagate through components quickly. Projects such as paths, in
Scout [39], and stackable systems [40] have had similar goals in other regimes.
The static compile-time optimization contained in the NESC compiler effectively
eliminates the overhead associated with the TinyOS component model. At compile time,
the component boundaries are automatically eliminated. Full program analysis is used to
create eliminate system partitions and balance code size with execution time. In most
cases all of the overhead associated with the component model is optimized out of the
final compiled binary.
Limited physical parallelism and controller hierarchy: We have successfully
demonstrated a system managing multiple flows of data through a single microcontroller.
Figure 3-7 shows the work and energy distribution among each of our software
components while engaged in active data transmission. It shows that 66.48% of the
work of receiving packets is done in the RFM bit-level component which utilizes 30.08%
of the CPU time during the entire period of receiving a packet. This consumes 451.17nJ
per bit. Even during this highly active period, the processor is idle approximately 50%
of the time. The remaining time can be used to access other sensors, like the photo
sensor, or the I2C temperature controller. In Chapter 5 we investigate the tradeoffs
associated with physical versus virtual parallelism in depth.
80
Diversity in usage and robust operation: Finally, we have been able to test the
versatility of this architecture by creating sample applications that exploit the modular
structure of our system. These include source based multi-hop routing applications,
active badge-like [41] location detection applications and sensor network monitoring
applications. Additionally by developing our system in C, we have the ability to target
multiple CPU architectures in future systems.
3.7 TinyOS Summary
TinyOS is an operating system specifically designed to address the needs of
wireless sensor networks. It is based on an event driven execution engine that
simultaneously provides efficiency and fine-grained concurrency. TinyOS’s event based
Componets
Packet reception
work breakdown
Percent CPU
utilization
Energy
(nJ/bit)
AM 0.05% 0.02% 0.33
Packet 1.12% 0.51% 4.58
Radio Handler 26.87% 12.16% 182.38
Radio decode thread 5.48% 2.48% 37.2
RFM 66.48% 30.08% 451.17
Radio Reception - - 1350
Idle - 54.75% -
Toal 100.00% 100.00% 2028.66
Componets
Packet transmission
work breakdown
Percent CPU
utilization
Energy
(nJ/bit)
AM 0.03% 0.01% 0.18
Packet 3.33% 1.59% 23.89
Radio Handler 35.32% 16.90% 253.55
Radio decode thread 4.53% 2.17% 32.52
RFM 56.80% 27.18% 407.17
Radio Transmission - - 1800
Idle - 52.14% -
Toal 100.00% 100.00% 4317.89
Figure 3-7: Detailed breakdown of work distribution and energy
consumption across each layer on the Rene node.
81
communication is encapsulated in a component model that allows state-machine based
components to be efficiently composed together into application-specific configurations.
The TinyOS communication system is based on the Active Messages
communication paradigm taken from high-performance computing environments. When
data is received from the network, message-specific handlers are automatically invoked.
TinyOS includes a special-purpose memory model which allows a fixed set of message
buffers to be pre-allocated and shared amongst several applications.
Now that we have an understanding of the software that will be running on these
devices, we can begin to dive into the key issues associated with their networking
protocols. In the next chapter we discuss a family of system-level tradeoffs and their
impact on node performance. We use this analysis to develop a generalized node
architecture that is the cornerstone of this thesis.
82
Chapter 4: Generalized Wireless Sensor Node
Architecture
A key step in exploring the challenges of building a low-power wireless
communication platform is in understanding the system requirements for the
communication mechanisms necessary to construct the multi-hop networks that are
envisioned. The focus of this chapter is to develop a general architecture that addresses
the needs of wireless sensor networks. The architecture is not tied to any particular radio
or processing technology but rather details how the computation and communication
should brought together. It is designed to have communication subsystem that allows for
flexible, application specific optimization of communication protocols while
simultaneously obtaining high bandwidth and efficiency. To set the stage for the design,
we first describe the basic operations associated with communication across a radio link.
This exposes several of the core design challenges including the concurrency intensive
nature of wireless sensor networks. To address the set of core challenges presented we
develop a generalized architecture for a wireless sensor node. In the following chapters
we analyze specific implementations of this general architecture.
4.1 Wireless communication requirements
To communicate over a wireless link, protocols must be built up from the raw
electro-magnetic signaling primitives. A transmitter must carefully modulate the RF
carrier while receiver performs demodulation and signal analysis. Figure 4-1 illustrates
the key phases of a packet-based wireless communication protocol. It is important to
83
note that many of the operations must be performed in parallel with each other. This can
be seen in the distinct layers that overlap in time.
The first step in the communication process is to encode the data for transmission.
The coding schemes are designed to increase the probability of a successful transmission
by preventing and correcting slight errors. For efficiency reasons the encoding process is
pipelined with the actual transmission process. Once the first byte is encoded,
transmission may begin. The remaining bytes can be encoded as preceding bytes are
transmitted.
Coding schemes can range from simple DC-balancing schemes, such as 4b-6b or
Manchester encoding, to complex CDMA schemes. In either, a collection of one or
more data bits, called a symbol, are coded into a collection of radio transmission bits
called chips. Manchester encoding has two chips per symbol which represents 1 bit of
Start Symbol Search Receiving individual bits
Start Symbol Detection
Synchronization
Radio Samples
MAC Delay Transmitting encoded bits
Start Symbol
Transmission
Bit Modulations
Transmission
Reception
Encoded data received
Data Received
… Encoded data to be Transmitted
Data to be Transmitted
Encode processing
Decode processing
Wireless Communication Phases
Time
Figure 4-1: Phases of wireless communication for transmission and reception.
84
data. Direct sequence spread spectrum and CDMA schemes often have 15 to 50 chips
per symbol with each symbol containing 1 to 4 data bits.
The actual transmission begins with the execution of a media access control
protocol (MAC). MAC protocols are designed to allow multiple transmitters to share a
single communication channel. One of the simplest MAC protocols is carrier sense
media access (CSMA) where each transmitter first checks for an idle channel prior to
each transmission. If the channel is busy, it waits for a short, random, delay after which
it reinitiates the transmission.
The first piece of data to be actually transmitted over the radio link is a
synchronization symbol or start symbol. The start symbol signals to the receiver that a
packet is coming and is used by the receiver to determine the timing of the arriving
transmission. The start symbol is immediately followed by the encoded data transmitted
as a serial stream. As the transmission proceeds, the transmitter must precisely control
the timing of each bit transition so that the receiver can maintain synchronization.
Skewed bit transitions can cause the sender and receiver to get out of synch, resulting in
an unsuccessful transmission or corrupted data.
For a receiver, the first part of data reception is to detect that a transmission has
begun. The channel is monitored continually in order to filter out background noise and
detect the special start symbol. The length and format of the start symbol can be
optimized for the expected noise levels. In order to properly detect the start symbol, the
receiver must sample the channel at least twice the radio chip rate. Otherwise, the
relative phase of the sampling and the transmission may result in the receiver missing the
start symbol.
85
Once detected, the receiver must then synchronize itself to the exact phase of the
incoming transmission. This synchronization step is critical in allowing the receiver to
determine the start and end of each bit window being used by the transmitter.
Synchronization requires the incoming transmission to be sampled higher than twice the
bit rate so the timing of the bit transitions can be determined.
Once synchronized, the receiver then samples the value of the incoming signal at
the center of each bit. Precise care must be taken to minimize skew in the sampling rate
and timing. As the individual bits are extracted from the radio, they are assembled into
blocks that are the encoded version of actual data messages. Finally, the blocks are
decoded back into the original data and assembled into a packet. The decoding process
can often correct bit errors in the received signal and reproduce the original data.
4.2 Key issues architecture must address
There are four key issues that must be addressed by the system architecture. First,
from our basic description of a wireless communication protocol we can see that several
operations must occur in parallel. The channel must be continually monitored, data must
be encoded, and bits must be transferred to the radio. The ability to deal with finegrained
concurrency is required in order to perform these operations in parallel.
Secondly, the system must be flexible to meet the wide range of target application
scenarios. Third, the architecture must provide precise control over radio transmission
timing. This requirement is driven by the need for ultra-low power communication in the
data collection application scenario. Finally, the system must be able to decouple the
data path speed and the radio transmission rate. We show how a direct coupling between
86
processing speed and communication bit rates can lead to sub-optimal energy
performance.
4.2.1 Concurrency
The architecture must provide an efficient mechanism to handle fine-grained
concurrency. In both the transmission and reception case, the computations associated
with wireless communication must occur in parallel with application-level processing and
potentially with intermediate level packet processing. Sensor events and data
calculations must continue to proceed while communication is in progress. This is
particularly true when considering overlapping start-symbol detection and application
processing.
For example, start symbol detection must be performed continually when waiting
for incoming communication. If any other processing is to be performed in must be done
in parallel with the start symbol search. In addition to the parallel operation of protocol
processing and application-level processing, we it is also required that several different
steps of protocol processing must be performed in parallel. In Figure 4-1 shows how data
encoding and data transmission are pipelined together in order to reduce transmission
latency.
The system architecture must provide a fine-grained concurrency mechanism that
allows for protocol processing and application-level processing to proceed concurrently.
4.2.2 Flexibility
The architecture must provide the flexibility to support a wide range of
application scenarios. For deeply embedded systems, it is important that a system is
87
flexible enough to support a wide rage of application-specific protocols. Unlike cell
phones, wireless local area networks, or Bluetooth devices, wireless senor networks do
not have a fixed set of communication protocols that they must adhere to. Deeply
embedded wireless networks can exploit tradeoffs between bandwidth, latency, and innetwork
processing to reduce power consumption by orders of magnitude.
For example, if sensor data is sampled only once per minute, it might be
acceptable to delay transferring the data for several seconds, allowing the network to
coordinate many such flows efficiently while operating at a low duty cycle. Such`
optimizations can extend battery life from weeks to years.
In addition to supporting flexible communication protocols, it is also important to
support flexible interfaces between protocols and applications. A rich interface between
application processing and communication processing allows programmers to create
arbitrary protocol decompositions. Internal protocol state can be exposed up to
applications and applications can have fine-grained control over the protocols.
A simple example of the need for flexible protocols that allow applications to
view the underlying channel characteristics can be seen in [42]. They discovered that
many 802.11 cards abstract away signal quality from application level processing which
eliminates their ability to support RF localization applications. Had the protocols for
interfacing between applications and the communications hardware been flexible, that
information could have been easily exposed.
4.2.3 Synchronization
The architecture must provide the ability to achieve precise node-to-node
synchronization and give the application direct control of radio timing. This plays a
88
critical role in many wireless sensor network applications. As described in Chapter 2, our
data collection application scenario relies on time synchronization to schedule the childto-
parent communication windows and the temporal accuracy of data collection. The
short duration and frequency of the transmission windows makes it important for the
application to have precise temporal control over the radio. This is in contrast to the
communication patterns of a cell phone where communication occurs infrequently (a few
calls per day) and for long durations (for minutes per call). A 10 ms delay before a cell
phone call starts or stops has a negligible impact on the user experience. However if a
packet is just 2 ms late in our data collection application scenario, it will be lost.
To allow applications to achieve precise time synchronization, radio
communication protocols must provide exact timing information as to when data packets
are received. Additionally, they must allow the application to precisely control when
radio communication begins as well as when the radio is switched on or off for reception.
4.2.4 Decoupling between RF and processing speed
Finally, the system architecture must allow a decoupling between RF transmission
rates and processing speed. A radio is the most efficient when data transmissions occur
at its maximum transmission rate. When transmitting with a fixed power, reducing the
transmission time reduces the energy used. However, modern studies in low power
processor design and dynamic voltage scaling have shown that processors are the most
efficient when they spread computation out in time as much as possible so that they can
run at the lowest possible voltage [43, 44]. Ideally, for power consumption reasons, we
would want the controller to perform all calculations as slowly as possible and just as the
computation is complete, the radio would burst out the data as quickly as possible.
89
Coupling the speed of the microcontroller to the data transmission rate forces both pieces
of the system to operate at non-optimal points. This coupling can be seen in the design
of a wireless sensor node where a balance must be formed between the speed of the
controller and the speed of the radio.
To dig into this requirement further we can analyze a performance shortcoming of
the Rene platform. On the Rene platform, only 10% of the radio’s 115Kbps capability is
used. Increasing the clock speed of the microcontroller could easily increase the radio
transmission speed because the speed of Programmed I/O is directly proportional to the
data path speed. If the clock speed of the microcontroller were doubled, the resulting
transmission rate would be 20Kbps and the energy consumed per bit by the radio would
decrease by 50%. However, the speed increase from 4MHz to 8MHz increases the
power consumption of the microcontroller from 15mW to 50 mW according to the
AT90LS8535 data sheet [27]. Additionally, the idle mode power consumption would
increase from 9mW to 28 mW. The savings in radio transmission power is quickly
Interdependence of CPU and RF Power
0 0.02 0.04 0.06
10Kbps
20Kbps
Average Power Consumption (uJ/bit)
CPU Energy Consumption
RF Energy Consumption
Figure 4-2: Energy cost per bit when transmitting at 10Kbps and
50Kbps broken down into controller and RF energy.
90
offset by the increase in the controller’s power consumption.
To make this concrete, we consider an application where a device is transmitting
100% of the time, doubling the transmission rate would allow the device to spend 50% of
its time in the idle mode. With a radio transmission consuming 21mW, the average
power consumption of the improved system would be approximately
.5*21mW+.5*50mW+.5*28mW or 49.5 mW. On the other hand, our original system
only consumed an average of 36 mW. While expecting a decrease in power
consumption, we see a net increase.
To address this issue, the system architecture must provide a mechanism for
decoupling the transmission rate from the datapath speed. This observation provides an
indication that there needs to be dedicated special-purpose resources for interacting with
the radio. Special purpose hardware can operate at frequencies that are optimal for the
radio, while the general-purpose controller is tuned to the optimal rate for computation.
Through buffering the special-purpose hardware can perform a rate matching between the
general-purpose data path and the radio. The amount of buffering sets the level of
decoupling. If the entire packet were buffered the processor speed could be scaled to
just meet the packet processing requirements set by the application scenario.
At first the need to decouple the transmission speed from the processing speed
seems at odds with the goal of allowing applications precise, direct and flexible control
over the radio communications. However, a processor supported by specialized hardware
accelerators has the ability to do both.
91
4.3 Traditional Wireless Design
Traditional wireless designs such as cell phones, 802.11 wireless cards and
Bluetooth enabled devices all choose to address the concurrency and decoupling issues
by including a dedicated protocol processor. Application and protocol-level processing
can proceed in parallel because they operate on separate physical hardware.
The protocol processor handles the real-time requirements of modulating and
demodulating the radio channel, encoding the data for transmission, and power cycling
the radio. It refines the raw data stream coming from the radio into a formatted packet
interface that the application processor can easily handle. For example, the host channel
interface (HCI) of Bluetooth provides a high-level packet interface over a UART. The
intricacies of packet synchronization, channel encoding and media access control (MAC)
protocols are all hidden from the application. The speed of the protocol processor is then
set to meet the requirements of the communication protocols.
While simple, there are several disadvantages to including a dedicated protocol
processor. The inclusion of the additional processor does not reduce the amount of
computation that must be performed; it simply confines it to a secondary processor. It
only provides a concurrency mechanism that allows application and protocol processing
to occur simultaneously. Unfortunately, the packet processor is a system resource
permanently dedicated to radio processing. This strict partitioning of resources leads to
non-optimal resource utilization. Additionally, inefficient chip-to-chip communication
mechanisms must be used to transfer data between the protocol processor and the
application processor. In the end, the use of dedicated protocol processors leads to
inefficient resource utilization.
92
An additional drawback to using a dedicated protocol processor, is that the
application interface to the communication protocols must occur over a narrow chip-tochip
interface. A fixed set of control messages that travel between the application and
protocol processor must be defined. This physical barrier between the application and
protocol processing makes it difficult to perform cross-layer optimizations. We show
how cross-layer optimizations that are prevented by this artificial barrier can lead to order
of magnitude performance increases. Traditional wireless designs sacrifice flexibility
and create strict system partitions when they include dedicated protocol processors.
An alternative to the protocol processor based approach is represented by the
Rene design discussed in Chapter 2. Instead of using a dedicated packet processor, we
time-share a single execution engine across application and protocol processing. The
concurrency requirements of the system are met virtually by fine-grained interleaving of
event processing in TinyOS instead of physically. This allows the computational
resources of the system to be dynamically partitioned between application and protocol
processing and avoids the overhead of shuttling packets between the processors, as well
as the need to define a fixed interface. We show that the benefit of dynamic resource
allocation can outweigh the overhead of using virtual parallelism. Additionally, in a
virtually partitioned system, the interface between protocol and application processing is
purely in software. This allows for rich interfaces to be constructed that can give
applications tight control over the communication protocols. A rich interface between
applications and their protocols can enable a host of system level optimizations.
93
4.4 Generalized architecture for a wireless sensor node
We have developed a generalized architecture for wireless sensor nodes that
builds from the single controller design used in Rene. The architecture is based on the
premise that shared pools of resources should be used when possible to exploit the
benefits of dynamic allocation, that buffering needs be used to decouple the generalpurpose
data path and the radio, and that protocol flexibility is essential. It addresses
performance and efficiency issues by including special-purpose hardware accelerators for
handling the real-time, high-speed requirements of the radio. Accelerators provide
general building blocks, not complete solutions.
Figure 4-3 depicts our generalized architecture. The core of the architecture is a
central computational engine that is timeshared across application and protocol
processing. Only a single computation engine is included because it allows the
allocation of all processing resources to a single task when necessary. We show that this
allows for an efficient use of processing capabilities.
Ideally, this single processor includes extra hardware support for the fine-grained
concurrency that it must provide. As this processor is intended to be allocated to multiple
concurrent operations, it must be designed so that context switching is as efficient as
possible. A traditional mechanism for decreasing context switch overhead is through the
inclusion of register windows. Multiple register sets can be included in the CPU so that
each context switch does not require the registers to be written out to memory. Instead,
the operating system simply switches to a free register set.
As is typical in microcontroller designs, the data path is connected to the rest of
the system components through a shared interconnect. Memory, I/O ports, analog-to94
digital converters, system timers and hardware accelerators are attached to this
interconnect. By utilizing a high-speed, low latency interconnect, data can be moved
easily between the processor, memory and peripheral devices.
In addition to allowing the CPU to interact with its peripheral devices, this shared
interconnect also allows the individual peripheral devices to interact with each other. A
peripheral placed on this bus has the ability to pull data directly out of the memory
subsystem or to push data into a UART peripheral. This creates a highly flexible system
where a data encoding peripheral can pull data directly from memory and push it into a
data transmission accelerator, such as modulation of an RF communication channel. In
ADC
Shared storage
Program & Data
Register File
I/O Ports
ALU
Instruction Fetch
& Decode
Protocol
Accelerators
System Timers
Radio Control
System
Register File
RF Transceiver
Memory & I/O bus
Processor Core
Protocol
AccePleroratotocrosl
Accelerators
Programming
Interface
Radio Timing
Signals
Data path Timing
Signals
Figure 4-3: Generalized Architecture for embedded wireless device.
95
such a system, the CPU is simply orchestrating the data transmission; it does not have to
directly handle the data.
All devices on this shared interconnect operate through a shared memory
interface. Each device has control structures that are mapped into a shared address space.
This allows components that were not originally intended to function together to be
combined in new and interesting ways. A data encoder designed to read from memory,
transform data, and write to memory many not even know that it is actually pulling data
from a radio receiver block’s memory interface and pushing it into a UART’s portion of
the communication memory. In this architecture, the size of the shared address space
dedicated to each operation can be set dynamically to meet application requirements.
The true power of this system is in the special-purpose hardware accelerators that
it enables. These accelerators provide efficient implementations of low-level operations
that are inefficient on a general-purpose data path. Each accelerator is designed to
provide support for operations that are critical to sensor network communication. By
increasing the efficiency of these operations, the overall power consumption of the
system can be greatly reduced. It is important that these accelerators are communication
primitives instead of complete protocol implementations so that the system can support a
wide range of communication protocols simultaneously simply through software
reconfiguration.
The hardware accelerators also support operations that need to be performed as
fast as possible to optimize the radio power consumption. This includes support for start
symbol detection as well as the low-level bit modulation. The goal is to include the
96
minimum hardware functionality that is necessary to efficiently support the needs of
applications and decouple communication rates from processing rates.
While the hardware accelerators are designed to provide primitives that can be
used to construct communication protocols, these primitives are not necessarily simple.
For example, a hardware accelerator for encryption would be considered a primitive.
This component would take the data it is given and encrypt or decrypt it as necessary. As
a hardware accelerator it could be used for encrypted communications, data
authentication, or to ensure that data stored in an off-chip flash is kept secure. The
hardware accelerator would be implemented so that the core cryptographic
transformation was exposed to the CPU and other peripherals. This is in contrast to a
system where an entire secure communication subsystem would be encapsulated without
exposing any of the internal primitives. Once again this design choice allows for
flexibility without sacrificing efficiency.
4.5 Architectural Benefits
To fully understand the tradeoffs and benefits of our generalized architecture we
can return to the set of issues presented earlier in this chapter. While the hardware
accelerators directly provide the ability to decouple the processing and communication
rates, we must explore the concurrency and flexibility mechanisms of this architecture.
4.5.1 Concurrency
To evaluate the performance of the concurrency mechanisms, we analyze the
performance of the most concurrency intensive aspect of communication, start symbol
detection, with respect to two system architectures. We compare our shared processor
97
architecture to one that has a pair of processors, one dedicated to protocol processing and
one dedicated to application processing. We break the CPU work down into three
categories: search time, which is the cycles dedicated to performing the start symbol
search, application time, which is cycles that the application can decide how to use, and
overhead. In the case of our shared CPU, system overhead is the cost of performing
context switches. In the dedicated protocol processor case, system overhead is the cost
of the chip-to-chip communication and the over-provisioning required to handle worstcase
deadlines.
To compare these two design points we fix the start symbol transmission rate at
10 Kbps and analyze the CPU usage of both systems assuming the same microcontroller
is used. As a baseline for our shared CPU system, we use a single 4 MHz
microcontroller. In the dedicated protocol processor based system, we size the protocol
processor to just meet the worst-case timing requirements of the start symbol search and
then size the application processor so that the system has the total equivalent processing
of the 4 MHz system.
The start symbol search used in TinyOS [45] requires 49 cycles per sample on
average, but has a worst-case overhead of 62 cycles. To detect a start symbol arriving at
10Kbps, a sample must be taken every 50 us. Assuming optimistically that the cost of the
chip-to-chip interface is zero, the protocol processor must be able to perform the 62 cycle
worst case sample every 50 us, so it operates at 1.25 MHz. Therefore, the application
processor would be left with the remaining 2.75 MHz (line 2 of Figure 4-4). In actuality,
we expect significant additional overhead caused by communication between the
98
application and protocol processor. To give the benefit of the doubt to the protocol
processor approach, we ignore this for now.
On the shared CPU system, the context switch overhead must be included for
each sample. Without any optimizations, 116 cycles are used per sample on average.
This leaves just 42% of the 4 MHz processor (1.68 MHz) available for the application as
shown in the top line of Figure 4-4. 33.5% of the CPU resources are spent on context
switch overhead. In contrast, 6.5% of the CPU resources of the dedicated processor
system are spent on the over-provisioning overhead. Figure 4-4 shows that there is more
processing available to the application in the dedicated protocol processor case (line 2
and 3) than in the case of the non-optimized shared controller (line 1). With a brute
force, naive, implementation, our concurrency model is significantly less efficient than
the dedicated processor case.
Breakdown of CPU Resources During Start Symbol Detection
0% 20% 40% 60% 80% 100%
Optimized Shared Controller
Dedicated Packet Controller
(Optimistic 0% chip-to-chip overhead)
Dedicated Packet Controller (Realistic
10% chip-to-chip overhead)
Unoptimized Shared Controller Start Symbol Search
Overhead of Physical Parallelism,
COM CPU
Over Provisioning on COM CPU
Overhead of Physical Parallelism,
App CPU
Overhead of Virtual Parallelism
Application Resources
Figure 4-4: Breakdown of CPU usage for start symbol detection overhead broken down into search time,
overhead, and CPU time for applications. For the dedicated controller case the breakdown is included with a
realistic 10% chip-to-chip overhead estimate and with a best-case 0% overhead estimate.
99
However, while the over-provisioning is a fundamental part of the partitioned
system, context switch overhead is something we can optimize. We can exploit dynamic
resource allocation to reduce the average context switch overhead in the shared case. So
far, our analysis is based on having both implementations process every sample in real
time. While it is essential that a design be able to processes the important bits in realtime,
not all bits have to be processed in that fashion. The first half of the start symbol
can be detected lazily and then switch to reading each sample in real-time. If just 8
samples could be processed in batch, the average context switch overhead would be
reduced by a factor of 8. This would result in a 4.19% context switch overhead, as
shown at the bottom of Figure 4-4 (line 4), which is less that the 6.5% over provisioning
overhead required in the case of the dedicated controller. Because processing the
samples in batch does not change the worst-case requirements, the dedicated controller
does not benefit from this optimization.
This analysis suggests that the widespread use of dedicated protocol processors is
difficult to justify on performance or utilization grounds, especially when the natural
dynamic partitioning of resources in the shared case is used to improve overall system
performance. While the fixed overhead of context switching can be significant,
techniques can be used to properly amortize it. The intermittent nature of the real-time
requirements of wireless protocols makes it beneficial to have dynamic resources
partitioning.
100
4.5.2 Protocol Flexibility and Timing Accuracy
We also must confirm that our architecture allows for protocol flexibility and
direct control over communication timing. A key strength of this general architecture
over traditional wireless systems is that it allows application developers direct access to
each step of the communication process. Instead of abstracting the timing of bit
sampling behind several system partitions, an application can directly calculate the arrival
time of each radio chip. In turn, in can also control the exact transmission time. This
creates the basic primitives for extremely precise communication scheduling. This tight
coupling between applications and their communication protocols can lead to ultra
precise time synchronization.
One of the key phases in wireless protocols is a media access control (MAC)
phase. Arbitrating for the channel introduces random delay in packet transmission times.
An application cannot know precisely when a packet will be transmitted when it initiates
the send operation. When performing time synchronization over the radio link, this
random delay degrades the accuracy of the synchronization algorithms. However, if
there is a tight coupling between the application and the protocol, the MAC layer can
inform the application what delay a packet experiences before it is transmitted.
Additionally, this information can be used to modify the packet once the transmission
begins so that it properly reflects the time that it was sent. This can lead to microsecond
accurate timestamps on packet transmission.
4.6 Summary
In this chapter we have set the background for, and explored the underlying
challenges of wireless node design. We then presented a generalized architecture that
101
addresses these issues. Now, we must evaluate this architecture on real-world hardware
implementations. In the next chapter we present a first approximation of this generalized
architecture in the form of the Mica node. This design serves to validate the architecture
principles we have chosen. With it we will demonstrate the flexibility and efficiency of
this architecture. It is the first step towards the final single-chip implementation
presented in Chapter 6.
102
Chapter 5: Approximation of General Architecture:
Mica
The Rene node served as a starting point for our study by allowing us to develop
the core TinyOS concepts. Only after a complete system was developed around Rene,
was it possible to analyze the hardware tradeoffs of the platform. The mica platform
builds on Rene by adding key hardware accelerators in order to validate the generalized
architecture.
Mica is the first test of our generalized architecture. Like Rene, it continues to
have a central microcontroller that performs all computation but it supplements the
controller with hardware accelerators, as dictated by our general architecture. The
accelerators we have chosen to include support to increase the transmission bit rates and
timing accuracy.
Unfortunately, Mica is only an approximation of our generalized architecture.
Existing interfaces available on commercial microcontrollers severely limit how fully the
architecture can be implemented, but nonetheless substantial performance and efficiency
gains are possible. The design provides the baseline for more aggressive
implementations and is useful in bringing to light the interactions between the various
types of system components and their impact on performance. It is built using a single
CPU that provides multithreading in software using the TinyOS concurrency mechanisms
[45].
103
5.1 Mica design
Mica combines an Atmega103 processor with a RFM TR1000 radio, external
storage and communication acceleration. A direct connection between the application
controller and transceiver provides the flexibility to meet application designs. The
hardware accelerators optionally assist to increase the performance of key phases of the
wireless communication.
The form factor (1.25´ 2.25 inch) is a similar size as a pair of AA batteries,
although, we have compressed a variant of the design to about the size of a 2.5-
centimeter coin .5 cm thick. The standby current for Mica’s components is just a few
micro amps, enabling applications that last for years on a single set of batteries. Mica
also improves experimental flexibility by including an expansion bus that connects to a
wide array of sensor boards. Currently sensor boards include support for monitoring
thermal temperature, barometric pressure, magnetic fields, light, passive infrared,
acceleration, vibration, and acoustics.
5.1.1 Block diagram overview
Figure 5-1 shows the Mica architecture, consisting of five major modules:
processing, radio frequency (RF) communication, power management, I/O expansion,
and secondary storage. A quick survey of the major modules provides a general
overview for the system as a whole, hence, a detailed bill of materials, device schematic
and datasheet for all components mentioned can be found at http://www.tinyos.net.
The main microcontroller is an Atmel ATMEGA103L or ATMEGA128 running
at 4 MHz and delivering about four million instructions per second (MIPS) [46]. This 8-
bit microcontroller has: 128-Kbyte flash program memory, 4-Kbyte static RAM, internal
104
8-channel 10-bit analog-to-digital converter, three hardware timers, 48 general-purpose
I/O lines, one external universal asynchronous receiver transmitter (UART), and one
serial peripheral interface (SPI) port.
Normally, programming of these embedded microcontrollers occurs during
manufacture with a firmware upload or during field maintenance. In our design,
however, the embedded network can be dynamically reprogrammed during routine use.
The coprocessor handling wireless reprogramming is the same controller used on Rene.
Additionally, to provide each node with a unique identification, we include a Maxim
DS2401 silicon serial number—a low-cost ROM device with a minimal electronic
interface and no power requirements [47]. The Mica radio module consists of an RF
Monolithics TR1000 transceiver. It is configured in a manner similar to that of the Rene
node.
A 4-Mbit Atmel AT45DB041B serial flash chip provides persistent data storage
[48]. We chose this chip because of its interface and small footprint, 8-pin small-outline
integrated circuit. It stores sensor data logs and temporarily holds program images that
Atmega103 Microcontroller
TR 1000 Radio Transceiver
4Mbit External Flash
51-Pin I/O Expansion Connector
Power Regulation MAX1678 (3V)
DS2401 Unique ID
8 Analog I/O
8 Programming
Lines
Coprocessor
Transmission Power
Control
Hardware
Accelerators
Digital I/O
Figure 5-1: Block diagram of Mica architecture. The direct connection
between application controller and transceiver allows the Mica node to be
highly flexible to application demands. Hardware accelerations optionally
assist in communication protocols.
105
received over the network interface. To hold a complete program the flash must be
larger than the 128-Kbyte program memory. This prevented the project from
considering use of the lower power, electronically-erasable-programmable-ROM-based
solutions used on Rene because they are generally smaller than 32 Kbytes.
We designed the power subsystem to regulate the system’s supply voltage; a
Maxim1678 DC-DC converter provides a constant 3.3-V supply. Mica operates with
inexpensive alkaline batteries and boosts their output voltage. We chose the Maxim chip
because of its small form factor and high efficiency. The converter takes input voltage
as low as 1.1 V. Input voltage significantly affects the TR1000’s transmission strength
and receive sensitivity. For ultra low-power sleep mode, disabling the power system lets
the system run directly off the unregulated input voltage, reducing power consumption by
the boost converter and the microcontroller. The radio will not operate, however,
without the boost converter enabled. Figure 5-2 summarizes the Mica node’s component
power consumption.
The I/O subsystem interface consists of a 51-pin expansion connector that we
designed to interface with a variety of sensing and programming boards. We divided the
connector into the following sections: eight analog lines, eight power control lines, three
Active Idle
CPU 16.5 mW 30 uW
Radio
21 mW (TX),
15 mW (RX) 0 uW
Silicon ID .015 mW 0 uW
External Flash 45 mW 30 uW
LED's 10 mW 0 uW
Mica Node Component Power Consumption
Figure 5-2: Breakdown of active and idle power consumption for Mica
hardware components at 3V.
106
pulse width modulated lines, two analog compare lines, four external interrupt lines, an
I2C-bus from Philips Semiconductor, an SPI bus, a serial port, and, a collection of lines
dedicated to programming the microcontrollers.
The expansion connector can also be used to program the device and to
communicate with other devices, such as a PC serving as a gateway node. Additionally,
it contains a standard UART interface to control or provide data to any RS-232 protocol
based device. Dozens of sensor boards with a variety of sensors have been developed
that use this expansion connector. It has even been used to allow the Mica node to
control a handful of inch-sized micro-robotic platforms.
5.1.2 Raw radio interface
The radio subsystem is a prime example of the rich simple interface approach.
From this subsystem’s low-level interface, designers can build customized arbitrary
signaling protocols to meet application requirements. The radio dictates certain
parameters, that is, signaling should occur over a half-duplex bit-serial link using
amplitude shift keying with a minimum physical bit time of 10 microseconds (ms), and
the need for a rough direct current (DC) balance (for example, there should not be more
than four consecutive high bits or four low bits). Beyond that, system software
components determine all aspects of the communication protocol.
The TR1000 radio used on the Mica allows the controller direct access to the
signal strength of the incoming RF transmission as well as a sampling the level of
background noise during periods when there is no active transmission. Using this
information in multi-hop networking applications can dramatically improve efficiency by
allowing applications to use links with good signal to noise ratios. Additionally,
107
interactions between transmitter and receiver are very predictable, as is the delay through
the radio interface. Because the radio can be powered on and off from software quickly
and predictably, low duty cycle operation can occur without global coordination or
complex time slotting. This direct, low-level interface to the radio provides flexibility
for application developers. Researchers from the University of California, Los Angeles
have exploited this flexibility to create energy aware MAC protocols [50].
5.2 Communication accelerators
The drawback of a low-level interface to the radio is that it places a significant
overhead on the main controller. The frequency of programmed I/O operations and the
inefficiency of conventional instruction sets for expressing these operations make it
difficult to achieve high bit rates. It is inefficient to handle the bit-serial sliding window
correlation operation required for start symbol detection on a general-purpose data path.
However, this operation takes just a few gates to implement in hardware. To compensate
for this, the Mica architecture includes hardware accelerators for the most demanding
primitives used in protocol construction and allows these primitives to be composed in
software, making rich interfaces available. On the Mica platform, we built the critical
hardware accelerators using conventional serializers in unconventional ways. While
they provide a significant efficiency boost, they represent just the beginning of what is
possible as we progress toward custom devices that fully realize our general architecture.
The first accelerator we include provides a simple shared memory buffer between
the bit-parallel data path and the bit-serial radio channel. It lets the processor deal with
communication data in efficient chunks and overlap data manipulation with the low-level
spooling of bits to or from the radio. The mechanism does not dictate channel coding or
108
signaling and can be bypassed to let the microcontroller directly interact with the radio
using programmed I/O. This option meets real-time requirements when necessary, for
example, during start symbol detection, but does not force real-time processing of all
transfer options. This accelerator lets the Atmel controller drive the radio at 50 kbps
using a small fraction of the processor, whereas programmed I/O peaks at 20 kbps using
the entire processor. (Other microcontrollers with a double-buffered SPI transmit port can
achieve the full 115 kbps of the radio.) The byte streaming performance of more
sophisticated versions of this accelerator would approach conventional direct memory
access channel performance, yet with precise control over timing.
The Mica node also incorporates a synchronization accelerator that captures the
exact timing of an incoming packet to within one clock cycle (250 nanoseconds) at the
start of packet reception. Shared memory stores this packet time stamp for the data path
to access. This hardware accelerator is critical to our design because at a high bit rate,
accurately determining transmission timing is difficult. The synchronization accelerator
forwards this time stamp to the buffering accelerator so that automatic channel sampling
can be performed by the buffering accelerator at the center of each bit transmission.
We built both of these hardware accelerators out of standard microcontroller
functional units. SPI is a synchronous chip-to-chip interconnect with a serial data line
and a separate clock signal that an external SPI master will provide. We drive the
asynchronous radio by combining the functionality of input capture registers, timer
controlled output pins, and the SPI communication hardware. The input capture register
automatically captures a timing pulse contained in the packet. This value is used to
configure the timing register that controls an output pin. One control option lets
109
hardware automatically toggle the output pin each time the counter expires; this output
pin becomes a clocking signal that times each arriving bit. Finally, SPI hardware
captures the value of the radio signal each time the counter triggers the clock line. We
accomplish this by connecting the counter controlled output pin to the synchronous clock
line of the SPI port.
During reception, we combine the radio’s incoming receive data with an
artificially generated clock signal to create an SPI master. The clock signal is the output
of the timing register fed back into the controller’s SPI port. This results in the SPI port
automatically latching and buffering the incoming transmission with precise timing, as
long as the internal timing register is configured correctly at the start of reception.
As in the general architecture presented in Chapter 4, these hardware accelerators
are primitive building blocks that can be flexibly used to create communication protocols.
They do not abstract away the low-level information. Rather they attempt to expose as
must information as possible up to the microcontroller’s data path so that flexible new
compositions are possible.
5.3 Evaluation
To evaluate the Mica node we will return to the set of requirements we used to
develop our general architecture. In particular we focus on the Mica node’s ability to
support the concurrency requirements of sensor networks, the ability to decouple
communication and data path speed, the ability to provide precise temporal control over
the radio and the ability to support a flexible set of communication protocols.
110
5.3.1 Concurrency Management
To analyze the efficiency of our concurrency mechanism, we return to our start
symbol detection analysis. To optimally exploit the dynamic CPU allocation, we
implemented a two-phased start symbol search mechanism. In the first stage, samples
are buffered into 8-sample blocks and analyzed in batch. If the first half of the start
symbol is detected, the next set of samples is analyzed in real-time. If the remainder of
the symbol is found, then the packet reception is started. If the symbol is not found, then
the CPU returns to batch processing mode.
This exploits the buffering accelerator to amortize the context switch cost across
multiple bits of data. It reduces the average context switch overhead from 33.5% down
to just 4.2%. This reduces the average CPU usage of the start symbol search and the
context switch overhead to 22.1% for the 10Kbps start symbol search. Not only did we
successfully amortize the context switch costs, but the batch mode processing of the
search itself is more efficient. The special purpose hardware has increased the CPU
resources available to the application processing by 85% and outperforms our optimally
Breakdown of CPU resources
0% 20% 40% 60% 80% 100%
Optimized Code
Original Code
Start Symbol Search
Overhead of Virtual Parallelism
Application Resources
Figure 5-3: CPU utilization of start symbol detection. Optimizations
enabled by the Mica node significantly increase CPU efficiency.
111
partitioned dedicated packet controller based system.
With this new start symbol search efficiency we can either spend more time on
application processing, slow down the CPU, or increase the start symbol transmission
rate. Depending on application specific demands all three of these options could lead to
a more efficient design point system. Figure 5-3 outlines these results.
5.3.2 Interplay between RF and Data path speed
In addition to using the hardware accelerators to decrease overhead, we have also
used it to decouple the data path from the transmission rate. During transmission, the
real-time, low-level bit sampling is now being performed in hardware. Instead of using
programmed I/O for each bit, we interface the data path at the byte level. Ideally, we
would have included more buffering than a single byte, but it was not possible when
using commodity Atmel microcontrollers. However, including this limited amount of
decoupling allows us to significantly increase transmission rates.
The implementation demonstrated on Rene used programmed I/O exclusively and
recorded a maximum bandwidth of 10Kbps. This was limited by the CPU overhead
associated with taking an interrupt with each bit. With our exploitation of the SPI based
hardware accelerator, we have been able to reach speeds of up to 50 Kbps on the same
processor. Our 5x speedup is limited by the tiny amount of buffering present on our
device. We have been able to demonstrate 115 Kbps transmissions on a TI MSP430 [24]
microcontroller that that also runs at 4 MHz, but has two bytes of buffering in the SPI
port. In addition to increasing the bandwidth, we also have decreased the CPU overhead
by dealing with data in byte-sized chunks. Eliminating the bit manipulation operations
from the data path significantly reduced the computational overhead on the system.
112
5.3.3 Interface flexibility
The third architectural issue is the efficient support of flexible protocol interfaces.
The increase in performance that we have shown demonstrates our designs ability to be
efficient. To demonstrate its flexibility we have implemented a pair of protocols that
exploit application specific optimizations. They are a high accuracy time
synchronization mechanism and an ultra-low power RF sleep protocol.
5.3.3.1 Cross-layer optimizations
Many deeply embedded applications spend the vast majority of the time in a very
low-power state, slowly draining their available energy supply by processing sensor
readings to confirm that no particular action need be taken. Infrequently, these
applications communicate monitoring results and, rarely, portions of the network become
extremely active upon detection of an important event. Performance improvements in an
application’s active work, such as processing and messaging, enhance the application’s
overall effectiveness and let it return to a low-power state more quickly. However, the
larger impact of the richly integrated approach to wireless design is that enables
optimization of the rest of the application. Often these optimizations reach across
traditional layers of abstraction, using low-level information to achieve a high-level goal.
113
In addition to delivering high bandwidth data communication using standard
protocols, we have also been able to demonstrate how the flexibility we provided plays a
critical role in enabling systems-level optimizations. We show this by implementing
four application specific protocols that improve overall system performance for their
intended application. The first of these is a protocol that explores the tradeoff between
transmitter and receiver overhead.
5.3.3.1.1.1 RF Wakeup
The raw interface to the radio can be exploited to implement an ultra low power
radio-based network wakeup signal. As seen in our alarm system application, it
important to be able to initiate unscheduled communication while the network is in a lowpower
state. The first step in this process is to wake-up a sleeping network.
For any RF based wake-up protocol, each node must periodically turn on the radio
and check for wakeup signal. Figure 5-4 contains the equations necessary to determine
the power consumption of a sleeping network. Each time a node checks for the wakeup
signal, it will consume energy equal to the power consumption of the radio times the time
the radios is on (1). The power consumed by the sleeping node will be the energy used
each time it checks for the signal times the frequency of the check (2).
1) Cost of checking = (radio on time) * (radio power consumption)
2) Power consumption = (checking frequency) * (cost of checking)
3) Average wakeup time = ½ (checking period) = 1/(2 * checking frequency)
Figure 5-4: Equations for determining the power consumption of a
sleeping network.
114
Minimizing the time a radio must be turned on each time a node checks for the
wakeup signal and minimizing the check frequency are the two mechanisms for reducing
the energy consumption of the system. However, the frequency that nodes check for a
wake-up signal determines the amount of time that it takes for the network to wakeup. If
each node were to check for a wake-up signal every minute, the average expected wakeup
time for a single node would be 30 seconds (3). This means higher frequency
checking yields faster wake-up times and better application performance. Because of
this, we focus on minimizing the time it takes to check for a wakeup signal.
Using a packet-based radio protocol, each node has to turn on the radio for at least
two packet times1. In our system, a packet transmission time is approximately 50ms, so
each node would have to be awake for at 100ms each time it checks for a wake-up
message. If a node needs to wake-up every minute, this yields a best-case radio duty
cycle of .2%. This time window gets even larger when considering the effects of a
multi-hop network and the contention associated with having a large number of nodes
retransmit the wake-up signal simultaneously.
Instead of interacting with the radio over a high-level packet interface, our low
power sleep mode implementation interacts directly with the analog base-band output.
The wake-up signal is nothing more than a long RF pulse. Each time a node checks for
the wake-up signal, it can determine that the wake-up signal is not present in 50us. The
50us signal detection time is a 2000x improvement over a packet-based detection time.
In our implementation, we choose to sample every 4 seconds which results in a .00125 %
1 If a node were sending a continual stream of wake-up packets, listening for two packet times
would ensure that a complete packet was transmitted while the node was awake.
115
radio duty cycle, a 160x improvement over the packet-based protocol that was sampling
once per minute. This ultra-low duty cycle implementation has been used to consistently
wake-up multi-hop sensor network of more than 800 nodes. This radically different
signaling scheme would not have been possible if a protocol processor was constraining
applications use of the radio.
The system-level protocol for managing the transition between the ultra-low
power sleep mode and the active mode is quite simple. Assuming that nodes start in the
sleep mode, each node checks for the wake-up signal. Once a wake-up signal is detected,
it listens for a secondary message that contains information about how long it is to wakeup
and if it should wake-up its neighbors. If neighbors are to be woken up, then the node
will retransmit the pair of wake-up messages. Supplementing the ultra-low power wakeup
message with a control message prevents unwanted network wake-up. The control
message can exploit error checking and message authentication techniques that are not
possible with the low-power wake-up message. If a secondary message is not received
that the node can quickly return to the low-power state.
Once a network is awake, it can be placed back into a low-power state by flooding
out a command message. The only requirement is that the nodes delay for a short period
of time after receiving the sleep command so that they ensure all of their neighbors have
heard the message also.
5.3.3.1.1.2 Low Power Listening
In sensor networks, multi-hop routing topologies are built by having intermediate
nodes act as relays for remote nodes. These routing nodes must listen for
communication and propagate messages towards their destination. While listening, the
116
radio consumes almost as much energy as when transmitting. Even when no
communication is taking place, considerable amounts of energy is spent searching for the
next packet. In the data collection network the energy spent while waiting for a
transmission can represent more than 50% of a node’s total energy budget. Embedded
sensor researchers have claimed that high-level protocols must be used to reduce energy
consumed of nodes when not actively transmitting [2]. As discussed in Chapter 2, one
solution is to have windows of communication periods and windows of sleep periods[51].
However, it has been shown that in some situations poor interaction between high-level
power saving techniques and low-level communication protocols can actually lead to an
increase in energy consumption when using windowing mechanisms [52].
We have been able to show that alternative is to modify the lowest levels of the
communication stack. Unlike windowing, which layers itself on top of standard lowlevel
protocols, we change the underlying communication protocol to optimize for the
receiver power consumption. Instead of having the sender simply transmit a start symbol
followed by a packet, we can require that the sender first transmit a preamble to get the
attention of any possible receiver. It can then follow with the start symbol and packet.
The receiver then only has to listen often enough to pick up any portion of the attention
signal. Precise control over the power state of the radio allows the protocol to turn off
the radio between each sample. The duty cycle of the receiver becomes proportional to
the length of this preamble. Once detected, the preamble will cause the receiver to
search for the pending start symbol.
This protocol optimization trades power consumption on the sender for power
consumption on the receiver. The sender must transmit longer, but the receiver can
117
sample the radio channel less frequently. The optimal ratio is dependent on the
communication patterns of the application. Specifically, it depends on the transmission
rate, duty cycle and the maximum allowable delay.
We have implemented a version of this scheme where the receiver has a 10%
duty cycle and the sender must transmit a preamble of 5 bits. The 10% receiver duty
cycle not only results in a 90% reduction of radio power consumption, it also produced a
33% reduction in the CPU overhead associated with monitoring the radio. On the other
hand, the sender only incurs a slight increase in overhead – less than 1% overhead on a
30-byte packet. This slight overhead increase is because the sample rate of the receiver
is 3000 times per second. Fine-grained control is used to power-on, sample, and poweroff
the radio in 30us. Application level windowing protocols generally have window
sizes of seconds or more.
Depending on application specific goals, the receiver overhead can be reduced
arbitrarily at the expense of bandwidth, latency, and transmission overhead. Moreover,
an application can change the protocol at runtime based on network activity. This simple
optimization demonstrates the benefits that are enabled by our architecture for flexible
communication protocols by exploiting the ability to tailor protocols to application
specific criteria.
5.3.3.1.1.3 Time synchronization
Our data collection application requires time-correlated sensor readings and
precise communication scheduling. The accuracy of distributed synchronization
protocols is bounded by the unpredictable jitter on communication times [53]. Aside
from variations in the actual network latency, the variation in delay going through
118
sophisticated protocol processors can be hundreds of milliseconds due to buffering, MAC
protocols, and back off. Unlike wide-area time synchronization protocols—such as
network time protocol8—we can determine all sources of communication delay [54].
Our rich interfaces allow us to expose the sources of delay to the application, reducing
the unknown jitter. Additionally, by exploiting shared system timers, we can assign
precise timestamps to incoming and outgoing packets beneath the sources of variation.
We designed the Mica platform to use an internal 16-bit high-frequency counter
to act as the lower 16 bits of a 32-bit continually running system clock. This highly
accurate system clock is directly linked to the synchronization accelerator used to capture
the exact timing of an incoming packet. To synchronize a pair of nodes, a packet can be
time-stamped with a sender’s clock value as it is transmitted and after all MAC delays
have occurred (prior to transmission, a node might have to wait for the radio channel to
be clear). The receiver’s synchronization accelerator can then tag the packet with the
receiver’s clock value. The application can use the difference between the two time
stamps to determine how to adjust its system clock.
Our implementation synchronizes a pair of nodes to within 2 ms. We can directly
attribute the skew of +/- 2 ms to several sources of jitter. When sending, there is a jitter
of +/- 1ms in the transmission propagation due to the internal circuit dynamics of the
radio. Hardware then captures the arriving pulse with an accuracy of +/- .25ms. Finally,
to synchronize the clock based on the captured value introduces an additional +/- .625 ms
of jitter. This implementation is only possible because of the shared access to a highaccuracy
system timer between the bottom of the network stack and the top of the
119
application. Partitioned wireless devices—such as Bluetooth chipsets—hide the exact
timing information from applications.
Generally the receive packet path has less jitter than the transmit path, so an
alternative approach uses the broadcast nature of the radio channel to synchronize
multiple receivers, even if they are poorly synchronized with the transmitter. Mica’s
time synchronization primitives provide highly accurate pair wise synchronization
between any two nodes. This primitive can aid construction of network-wide time
synchronization to within tens of microseconds.
5.3.3.1.1.4 Localization
Node localization is a key part of our node tracking application scenario. Direct
access to the base-band signal being output by the TR1000’s receiver chain can be
exploited to assist with localization. Several groups have attempted to perform
localization in a sensor network by using RF signal strength[42, 55, 56]. The radio is
used as an analog sensor to detect the strength of an incoming signal. RF propagation
models are then applied to infer distance from a collection of strength readings.
By providing the processor direct access to the raw base-band signal, our platform
gives the application developer as much information as possible about the incoming
signal. The central controller can look at the signal strength of each individual bit as
well as the level of the background noise. Additionally, because the sender has direct
control over the base band signal being transmitted, it can intentionally transmit long
duration pulses of variable strength to help the receiver determine the reception strength
more accurately.
120
An alternative to RF localization is to use acoustic localization[57]. The
propagation delay of acoustic pulses is measured and used to infer distance. For accurate
results, it is essential to use the radio to achieve precise time synchronization between
two nodes in order to correlate their sensor readings in time. For accurate results, the
transmission time of the acoustic pulse must be accurately communicated to the receiver
and correlated with the receive time. The precise communication synchronization
provided by the radio layer allows this to be done incredible accurately.
5.4 Blue: a follow-on to Mica
The design point represented by Mica has been taken one step further by Dust Inc.
The Dust Inc. Blue mote replaces the individual system components while maintaining
the same basic architecture. The Blue mote continues to run TinyOS, yet it exploits a
new generation of microcontroller and radio.
5.4.1 CPU
The first system component to be changed is the central microcontroller. Blue
abandons the AVR based microcontroller in favor of a 16-bit Texas Instruments MSP430
controller [24]. The TI controller has been optimized for low-power, battery operated
systems. As noted in Chapter 2, a key microcontroller characteristic is the ability to
quickly enter and exit sleep modes. The MPS430 has the ability to enter and exit a 2 uA
sleep state in under 100 us. Additionally, the core has been designed to let subsystems
operate while the system clock is disabled. ADC operations, UART communication and
radio data transfers can proceed while the main CPU core is dormant. These features
greatly reduce the duty cycle and in turn energy consumption of the CPU core.
121
The TI CPU has additional benefits. The overall power consumption of the CPU
in active mode is reduced from the 5 mA of the AVR core to just 1 mA when operating at
3 V. Additionally, the TI CPU is capable of operating over a voltage range down to 1.8
volts. This allows it to operate directly off of alkaline batteries without requiring a
costly boost converter.
Additionally, the ADC resolution is increased from 10-bits to 12-bits and the
ADC sample rate is increased to 100 Ksps. Finally, the part cost of the TI CPU is
significantly less that that of the Atmel version. Overall the blue mote is expected to
yield a 50% cost reduction and a 4-6x life improvement.
5.4.2 Radio
The second significant upgrade on the Blue node is the use of an FM radio instead
of an AM radio. The Blue node employs a 900 MHz FSK radio – the Chipcon CC1000
Figure 5-5: Follow-on to the Mica
node, blue increase range and reduces
power consumption. Blue nodes have a 4-
6x longer battery life over Mica on alkaline
batteries.
122
that we outlined in Chapter 2[23]. While consuming more power than the RFM
TR1000 radio used on the Mica node, the CC1000 has increased sensitivity and increased
transmission power. The use of FM modulation also allows the CC1000 to be more
tolerant to external interference.
Additionally, the CC1000 radio is frequency agile. The carrier frequency of the
blue mote can be tuned anywhere in the 902-928 frequency range. By default, the Blue
node provides 25 evenly spaced channels. These channels can be used to avoid external
interference or to divide network bandwidth across the individual channels. On blue, the
CC1000 provides 76.8 Kbps of bandwidth on each channel, up from the 50Kbps possible
on the Mica node single channel.
5.5 Summary
The Mica node is the first approximation of our generalized architecture. Built
from off-the-shelf components, it has proven invaluable in verifying the validity of our
general architecture. However, it is constrained by the available interfaces and internal
behavior of commercially available components. Despite its limitations, it has been used
by over a hundred research groups as a platform for experimenting in the area of wireless
sensor networks.
Mica is a careful balance of efficiency and flexibility. Efficiency is essential for
all sensor network systems while flexibility allows the Mica platform to address a wide
range of application scenarios. The success of the Mica platform enables us to take the
next evolutionary step – single chip integration.
The Mica node has demonstrated how the generalize architecture presented in
Chapter 4 has the potential to produce highly efficient and flexible nodes. When
123
compared to the Rene platform, the mica hardware accelerators significantly increase
communication bit rates and timing accuracy while reducing CPU overhead.
The Blue platform follow-on to the Mica node maintains the same basic
architecture of the Mica node but selects hardware components that significantly improve
node lifetime and communication range. As described in Chapter 2 the ability to tolerate
a wide range of input voltages significantly increases node lifetime when operating off of
alkaline batteries.
In the next chapter we make the leap from constructing system out of off-the-shelf
components to creating custom silicon. This allows us to fully realize the generalize
architecture presented in Chapter 4. We show the inclusion of key hardware accelerators
can lead to dramatic improvements in performance without sacrificing efficiency. The
design process results in a single-chip CMOS design that integrates a radio, computation,
and storage into a single millimeter-sized node.
124
Chapter 6: Integrated Architecture for Wireless
Sensor Nodes – Spec
The next leap towards demonstrating the power of our generalized node
architecture is to move beyond what is possible with off the shelf components and design
a custom integrated solution. While approximating our general architecture, the Mica
and Blue nodes were constrained by existing inter-chip interfaces. Development of a
custom ASIC allows us to tear down the artificial constraints imposed by commercial
components. Through the use of custom silicon we are able to achieve orders-ofmagnitude
efficiency improvements on key communication primitives. Additionally, we
demonstrate that our generalized architecture allows us to achieve these performance
improvements without sacrificing flexibility.
Figure 6-1: The single chip Spec node
pictured next to a ballpoint pen.
125
In addition to the architectural benefits, single-chip integration also has significant
size and efficiency improvements. Our latest design is an ASIC based node called Spec.
Despite including a microcontroller, SRAM, communication accelerators and a 900 MHz
multi-channel transmitter, the Spec node measures just 2.5 mm on a side in a .25 um
CMOS process. Its architecture blends the efficiency of a dedicated hardware
implementation with the flexibility of software by providing low-level primitive
accelerators that are dynamically composed together. The integration results in a device
which only requires a handful of low-cost external components, including a crystal,
battery, inductor and an antenna to be a complete wireless sensor network node.
6.1.1 High Level Overview
In order to support a wide range of application scenarios and the ability to
perform application specific optimizations, the Spec node is designed to be a collection of
basic primitives that can be flexibility composed. This is in contrast to systems that are
built as a hierarchy of specialized sub-systems. Instead of having protocol engines and
dedicated radio controllers, Spec is built out of low-level pattern matchers, timing
extractors and serializers. The simple primitives can be composed together by the
central controller to implement a wide array of application specific communication
protocols.
The true benefits of our generalized architecture are only realizable when
combined with high speed on-chip interconnects. With Spec we are able to design
specialized accelerators targeting specific communication needs and connect them to a
126
shared, high-speed interconnect. This high speed, low-latency interconnect is essential
in providing an efficient way of composing and configuring the discrete primitives.
The accelerators in the Spec node were selected so that the CPU intensive
portions of RF communication could be offloaded. As we considered each hardware
accelerator, we compared the potential in energy saved against the loss of flexibility. If
we only wanted to implement a single protocol, we could implement it directly in
hardware. However, such a node could only support a handful of applications. By
carefully selecting the correct set of primitives, we can implement a large set of protocols
with a minimal impact on efficiency.
The Spec node consists of SRAM, a RISC CPU microcontroller core, a 900 MHz
RF transmitter [58], a collection of communication hardware accelerators, and an analogto-
digital converter [59] all packed together. In addition to providing primitives to assist
in implementing RF communication, the Spec node also has support for the highly
efficient interrupt handling. Profiling of TinyOS revealed the significance of efficient
interrupt handling. Sensor networks must be continually monitoring their environment.
Each sample triggers several interrupts that perform small amounts of computation.
Interrupt overhead accounts for a significant amount of the processor’s energy
consumption; this makes it an ideal candidate for optimization.
6.1.2 General Block Diagram
Figure 6-2 depicts the general block diagram of the Spec node. The CPU core is a
basic 8-bit RISC core with 16-bit instructions. To remain compatible with the Mica
instruction set, Spec has a Harvard architecture with separate address spaces for data and
127
program memory. The CPU can execute ALU operations in a single cycle but memory
operations take two cycles. The instruction set architecture contains 32 general purpose
registers some of which can be combined together to form 16-bit addresses for use in
load and store operations. Additionally there is a dedicated stack pointer that is
automatically manipulated by jump and return instructions. On Spec, architected
registers are supplemented by an additional set of 32 general purpose registers to allow
for optimization of interrupt handling routines.
The CPU core is connected to a bank of 6 memory blocks each containing 512
bytes of data. These blocks can be individually mapped into either the data or program
address space of the CPU. Only 3K of memory is included in our current design.
Clearly the total amount of ram could be increased at the expense of die area. It is
important that the memory be divided into banks so that it can be migrated between
ADC
RAM
Blocks
Address Translation
Register File
I/O Ports
ALU
Instruction Fetch
& Decode
Encryption Engine
System Timers
SPI
Programming
Module
Radio Subsystem
Register File
Analog
RF
Memory & I/O bus
RISC Core
Figure 6-2: Block Diagram of Spec, the single chip
wireless mote.
128
instruction and data memory. This mapping is handled by an address translation unit
that is interposed between the memory and the processing core. The memory mapping
unit contains space for 16 physical frames. The Spec chip uses 7 of them – 6 for the
memory banks and one for the interface to the radio subsystem.
In addition to the memory controller, the CPU is also connected to an ultra-low
power analog to digital converter, an encryption accelerator, general purpose I/O ports,
system timers, a chip programming module, and a RF sub-system. For illustrative
purposes we have grouped the communication accelerators into a subsystem. However,
it is actually composed of a collection of accelerators that we describe in detail. Each of
the communication accelerators connect directly to the microcontroller’s data bus. The
accelerators can be linked together or used in isolation.
The communication subsystem contains primitives that: extract and generate bit
timing, perform pattern matching for start symbol detection, stream received and
transmitted data into and out of memory, encrypt and decrypt data automatically and
provide memory-mapped control interfaces to the radio circuitry.
6.1.2.1 CPU Core
The standard 8-bit RISC microcontroller on the Spec node is a traditional single
cycle core. It has a 16-bit instruction bus and an 8-bit data bus. The only form of
pipelining is a pre-fetch of the next instruction while executing the current instruction.
Decode, execution, and write back are performed in a single cycle. In addition to
traditional RISC instructions, Spec also includes a set of instructions for interacting with
129
64 specialized I/O registers. I/O operations to these registers can be performed in a single
cycle. However, memory operations require two cycles.
The actual instruction set architecture is not of particular importance. Key
constraints are that the instruction representation must be compact (as is the case with
most microcontrollers), and that I/O operations can be performed efficiently.
More important than the actual ISA is the ability to efficiently handle interrupts.
The event based programming of TinyOS relies heavily on the ability to process
interrupts quickly. We include specialized support for interrupt handling in the form of
register windows. The CPU core includes two set of registers. One set is used for
general processing and one for use in the context of interrupts. Background processing
such as TinyOS tasks can easily share one set of registers. When an interrupt arrives, the
interrupt handler can quickly switch register windows by executing a single specialized
instruction. It can then proceed to execute the handler code and return. The interrupt
handler can only use the secondary register set if it does not re-enable interrupts during
its execution.
To access the secondary register set, a specialized “switch registers” instruction
has been included in the ISA. This single instruction replaces the multiple memory
accesses that are generally required to write the saved registers out to memory.
6.1.2.2 Paged RAM
In addition to a specialized data path, the Spec chip has a specially designed
memory system. Traditional embedded systems have a strict partition between program
and data memory. This is done, in part, because embedded systems are generally
designed with a ‘write once, run forever’ model in mind. Embedded software is usually
130
written with a single application in mind and then burned into a device where it will
remain for the life of the device. The dynamic nature of sensor networks makes it
impractical to maintain that programming model.
While a strictly partitioned memory system is practical in traditional application
scenarios, wireless sensor networks demand more flexibility. The wide range of target
applications makes it desirable for nodes to be field programmable. Users want to be
able to download new programs over the radio and into the application space.
To support this style of operation, the Spec chip includes a specialized paged
memory system where physical page frames can be dynamically mapped into any part of
program or data memory. With this mechanism applications can be loaded over the
radio and into data memory. Once the application is complete and verified, the memory
frames used to store the program can be re-mapped into the instruction memory of the
CPU. At the same time, the original application is automatically mapped out of the
instruction memory. In just a few cycles, the entire program memory can be swapped
out. In the event that the new program is determined to be faulty, the original program
can be quickly restored.
The Spec memory mapped system contains 16 page frames each containing 512
bytes of memory. The page frame mapping is controlled by writing to special addresses
located at 0xffc0. During operation, the low 9 bits of the memory address are passed
unchanged to the memory system. The remaining 7 bits are passed through a remapping
table that dynamically performs the translation. Both the instruction and data memory
addresses are translated simultaneously by separate control blocks.
131
When a physical page is mapped into a new location that is already filled, the old
page is removed from the address space all together. The old page must be mapped
elsewhere in the address space before it can be accessed again. When mapped into data
memory, each page is byte addressed over an 8-bit data bus. When mapped into data
memory, the ram provides a 16-bit, word addressed, read-only memory.
6.1.2.3 Communication Acceleration
One of the primary concepts contained in the Spec design is the inclusion of
hardware acceleration primitives designed to reduce the cost of radio communication
without sacrificing flexibility.
The communication accelerators on the Spec node include support for
synchronization, timing extraction, encryption and data serialization. During
communication, the first of the accelerators to come into play is the synchronization
accelerator.
6.1.2.3.1 Synchronization
The synchronization accelerator is designed to process the incoming bit stream
and detect the presence of a pre-determined start sequence. First, a configuration register
is used to specify the value and length of the target sequence. Then the incoming data
stream is automatically analyzed in search of the sequence. When detected, the
accelerator signals the CPU via an interrupt mechanism. It also is interconnected to a
data serialization accelerator with automatically records the incoming transmission.
6.1.2.3.2 Memory Mapped I/O & Data serialization
132
Data serialization acceleration has been included to automatically stream data
between the radio and memory system. It is a simplified DMA engine. During radio
transmissions, each bit must be recorded or transmitted serially. The accelerator is
enabled by specifying a memory address for the start of the packet and a maximum
packet length. In TX mode you set the packet length to a non-zero number to initiate a
transmission. The transmission pointer is automatically incremented with each byte
transmitted. Accordingly, the remaining packet length is decremented to represent the
number of bytes left in the transmission. Both of these fields can be accessed by the
processor to monitor progress.
During reception, either a signal from the CPU or a signal from the
synchronization accelerators causes the incoming transmission to be recorded to the
address of the receive memory pointer. Data is recorded based on an receive data length
field.
Memory operations for the I/O system are automatically interleaved with the
memory operations of the CPU. Most CPU instructions do not access the data memory,
only instruction memory. Spec has been designed so the program and data memory
access can occur in parallel. The RF system can claim access to the memory bus when it
is not in use by the CPU. The RF system has enough internal buffering to handle
potential delays. The memory addresses used by the RF system are virtual memory
address prior to translation by the memory map logic.
6.1.2.3.3 Transmission Timing
The serialization and synchronization accelerator are both supported by a timing
accelerator that provides RF transmission timing information. This accelerator is
133
connected to the high frequency oscillator used by the CPU and can be programmed to
generate timing pulses at integer multiples of the clock period.
The transmission bit rate is programmed by specifying the clock divider
corresponding to 5x the transmission rate. This allows the timing accelerator to generate
both the individual bit timing pulses and pulses for 5x over sampling. The 5x over
sampling is used during radio reception to detect edge transition in the data stream. This
allows the transmitter and receiver to remain synchronized during communication.
6.1.2.3.4 Encryption
Many sensor applications require that data be both confidential and authentic.
Our security application scenario demands that all alarm conditions are genuine.
However, encryption is a costly operation on a general purpose CPU. Spec provides a
hardware accelerator for encrypted communication.
When enabled, all communication over the radio link is encrypted prior to
transmission and decrypted upon reception. An application must begin each
communication in the clear so that receivers can lock-on to the incomming signal and
select appropriate encryption settings. At that point, the encryption accelerator can be
engaged to perform the data convolution.
The encryption mechanism on Spec is a LFSR based stream cipher similar to that
used in Bluetooth. Prior to engaging the encryption accelerator the application must
generate a unique key sequence. Both the transmitter and the receiver must share the
same initialization key.
134
This encryption accelerator provides data security but does not provide data
authenticity. It is possible for an attacker to modify transmitted messages. Data
integrity must still be provided by the application.
6.1.3 Radio Back End & ADC Sub blocks
The only analog circuitry contained on the Spec design is inside the radio [58] and
ADC (analog-to-digital converter)[59] sub-componnts. Both the radio and the ADC are
included as pre-designed blocks. The ADC was designed to be as low power as possible.
It performs an 8-bit conversion with just 27 pJ of energy. Its ultra-low power
consumption is ideal for use in wireless sensor networks. It is capable of operating at
over 10,000 samples per second while consuming less that a microwatt.
The transmitter was also designed with power consumption in mind. In isolation
it consumes less that 1 mW while transmitting at -7 dBm. It is a frequency shift keyed
(FSK) radio that is capable of transmitting at a wide range of frequency in the 900 MHz
band. It can communicate to the Chipcon CC1000 radio used on the Blue node at up to
40 feet in a lab environment.
The transmitter uses a 32.768 KHz oscillator as its primary frequency reference.
It only requires support form a couple of off-chip components. In addition to the 32
KHz oscillator, it requires an inductor and an antenna. Advances in CMOS technology
may eventually allow the inductor to be constructed on-chip. The radio is configured
through a set of registers that allow the CPU to control transmission strength,
transmission frequency and frequency separation. The radio is capable of transmitting
up to approximately 100 Kbps.
135
On the current implementation, only a transmitter is included on-chip. The Spec
design can also use an off-chip 900 MHz transceiver built by RF Monolithics. This is
the same radio used on the Mica node. A low-power on-chip receiver is currently under
development and has been simulated to consume just 300 uA at 3V.
6.1.4 Digital frequency lock registers
The analog RF block contains the primitive functions necessary for RF
transmission. However, it is not capable of self-tuning. A digital frequency control
block is included for frequency tuning. The frequency control block implements a
frequency lock loop between the 32.768 KHz crystal oscillator and the RF transmission
frequency.
The frequency lock is performed by counting the number of oscillations of the RF
transmission frequency that occur between each oscillation of the 32.768 KHz clock.
The 900 MHz carrier signal is fed into a specialized block of low-power dynamic logic
that divides the signal by a factor of 8. Once the signal is in the range of 100 MHz, a
standard digital counter is used as a cycle counter.
The frequency lock block is configured by specifying a target count that
represents the desired frequency. The block then adjusts the control lines of the register
to achieve the programmed frequency. If a 902.5 MHz signal is desired, then 27542 can
be fed into the frequency control register. (27542 * 32.768KHz = 902.5 MHz). This
results in the ability to adjust the carrier frequency in 32 KHz steps. To increase the
fidelity of the adjustment, the 32 KHz reference can be divided down by up to a factor of
32 in order to produce a minimum frequency step of 1 KHz resulting in an adjustment
accuracy of 1 KHz.
136
The frequency control block uses a successive approximation algorithm to quickly
narrow in on the desired frequency. Once locked, the control block continually tracks
the desired frequency by incrementing and decrementing the radio control words as
necessary. In addition to providing frequency tuning, the frequency control block also
controls the frequency separation of the FSK modulation. Frequency separation can
range between 10 KHz and 10 MHz. For debugging purposes the frequency control
block can be bypassed to directly access the RF control lines.
The interface to the analog RF blocks consist of 16 frequency control lines, 24
power, bias current, and amplitude control lines.
6.1.5 Physical Characteristics
The Spec design was fabricated by National Semiconductor using a .25 um
CMOS process. Figure 6-3 gives a general feel for the area required by the various sub-
CPU and Accelerators
ADC
Memory Banks
Crystal Driver
900 MHZ transmitter
Frequency Synthesizer
I/O Pins
Figure 6-3: Design layout for single chip mote. Large central block contains CPU, timers
and hardware accelerators. Across the top are 6 memory banks. The ADC is center left. The
radio is in the bottom right corner.
137
components of the design. A large fraction of the area is consumed by the ram blocks –
approx 20 – 30%. Additionally, the analog RF circuitry requires approximately the same
area as 1 kilobyte of SRAM.
The digital logic contained in the design was synthesized from behavioral VHDL.
After simulation, the high-level VHDL code was mapped into standard cells provided by
National Semiconductor using Ambit Build Gates and layout was performed with Silicon
Ensemble – both tools from Cadence Design Systems. In addition to simulation, the
functionality of the digital core was also verified by downloading it onto a Xilinx FPGA.
The actual properties of the National Semiconductor cell libraries are confidential
however, for analysis purpose we assume that the registers that we are using have a delay
of approx 1 ns and consume approximately 1 pJ per transition. These approximations
Component (mm2) %
CPU Core 0.3806 47%
RF System Interface 0.1886 23%
Central Memory Controller 0.0769 10%
UART 0.0407 5%
Timer 0.0406 5%
SPI Programming 0.0283 4%
Service Module 0.0172 2%
Output Ports 0.0120 1%
ALU 0.0114 1%
ADC Interface 0.0100 1%
External Mux 0.0033 0%
RAM Access Controller 0.0030 0%
Input Port 0.0007 0%
Complete Logic Core 0.8065 100%
Logic Core Area Breakdown
Figure 6-4: Area breakdown of digital logic
modules included on Spec. Total area of wiring and gates
using National Semiconductor standard cells in .25 um
CMOS.
138
are well below the actual performance and therefore represent a upper bound on the
energy consumption of our basic blocks. Even with these sub-optimal approximations,
the central core of our device can operate in excess of 50 MHz – well beyond the target
of 4Mhz.
Figure 6-4 gives a breakdown of the area used by various system blocks. This
information can be used to evaluate the area/power tradeoff associated with each system.
The RISC CPU core requires just .381 mm2. The RF subsystem is approximately half
that size at .188 mm2. It is important to note that the areas reported are the total area of
the logic blocks. It is not always possible to achieve logic 100% density. The .806 mm2
of logic is placed inside the 1mm2 block shown in Figure 6-3.
6.2 Performance
In our evaluation, we compare the Spec design against the MICA platform. To
recap, the key characteristics of the Mica platform are that the CPU consumes 15 mW
when active at 4 MHz and the radio consumes 21 mW while transmitting and 15 mW
when receiving. The maximum transmission speed is 50 Kbps. Our goal is to focus on
the architectural improvements of the Spec design over the Mica design. For a fair
comparison we assume that our core consumes the same 15 mW when active as the Mica
core. While in actuality we consume only approx 1.5 mW, the reduction in power not
due to an architectural difference, but simply the CMOS process used. Our comparison
focuses on the duty cycle of the CPU.
139
6.2.1 Start Symbol Detection
In analyzing the power consumption of the wireless networking protocols
implemented on the MICA platform, the power consumption of start symbol detection
was the single largest component of total CPU power consumption. On Mica, prior to
receiving a packet, the channel is sampled at 20 Kbps in order to search for the arrival of
a 10 Kbps start symbol. The CPU contains a buffer of the past 30 samples and performs
a pattern matching operation on the arriving stream. Each sample requires individual
processing from the CPU. Additionally, because the data portion of the packet
immediately follows the start symbol, each sample must be processed in real time so that
the symbol is detected before the data arrives.
This software based approach to start symbol detection takes approximately 100
cycles per sample. At 20 K samples per second, this results in 2 M cycles per second.
On the Mica node this requires 7.5 mW of power (50 % of the CPU). Not only does the
software start symbol detection consume a significant amount of energy but the available
CPU resources, it also limits the maximum sample rate. 50 Kbps starts symbol detection
would be impossible at any energy budget.
As discussed, the Spec node solves this problem by including a hardware pattern
matching primitive. This primitive block is configured with a pattern up to 24 bits long
and signals when it is matched. It works in conjunction with a timing accelerator that
automatically sets the sampling rate. While sampling the channel, the pattern matching
primitive actually samples at 5 times the bit rate. This allows it to average several
samples to reduce the impact of noise.
140
The hardware implementation of this primitive requires .1 mm2 of chip area. A
majority of that goes to the 120 register bits required to hold the 24 bits over sampled by
5x (24*5 = 120). Using the 1 pJ/transition estimate of the energy consumption of a
register described above, the total power consumption of the accelerator when searching
for the same 10 Kbps start symbol is just 5 uW. Despite the 5x over sampling, the power
consumption is reduced by a factor of 1,500.
6.2.2 Interrupt Handling Overhead
Another frequent and expensive operation on the Mica node is the saving and
restoration of registers when handling an interrupt. The real-time nature of wireless
sensor networks makes handling interrupts a critical and frequent operation. The
standard interrupt routine used in TinyOS requires that 14 of the 32 registers be saved
prior to the execution of the interrupt handler. With a 2 cycle memory latency, this
saving and restoring of the registers requires 56 cycles per operation. When sampling an
audio waveform at 5K samples per second to detect an acoustic pulse, this would
represent an overhead of 280,000 cycles per second. When broken down into energy,
the saving and restoring of registers for interrupt handling requires 210 nJ/interrupt.
The Spec node adds specialized hardware support for accelerating the execution
of interrupts. Instead of having only a single register set that must be saved for each
interrupt, the Spec node introduces an additional register set and an instruction that can
be used to switch between them. During operation, one register set is used for generalpurpose
execution and the second is used for short, non-preemptable, interrupts. Thus,
when executing a non-preemptable interrupt, the system executes a single instruction to
141
swap register sets. In this design, the stack pointer is preserved when register sets are
switched. This allows the interrupt routine to use continue to use the system stack after
switching register sets.
The inclusion of this additional register set reduces the cost of interrupt overhead
to just 7.5 nJ/interrupt – 2 instructions. The overall cost of saving and restoring the
registers is reduced from 210 nJ to just 7.5 nJ, a factor of 28. We also considered
including additional register sets for other system functions. However, the area of each
additional register set – .113 mm2 – made this prohibitively expensive. Each register set
represents 30 % of the CPU core’s area.
6.2.3 Program Memory Management
A key requirement for wireless sensor networks is that the nodes need to be
reprogrammed in the field. As nodes are deployed in large numbers it becomes
impractical to collect them for reprogramming. It is important that they be able to accept
a program over the radio and then start executing it. Additionally, as the cost of the
individual nodes drop it will become impractical to interface directly with each node
produced. The nodes will be exclusively programmed over the radio.
With the Mica architecture, a device must be equipped with enough storage to
hold 3 distinct program images. During operation, the node has a currently executing
image stored in its program memory. To reprogram itself it downloads a second code
image and stores it in external memory. Once complete the downloaded image is copied
over the running image and booted. However, in case the new image is damaged or non142
functional, a backup image must also be stored. In total, three copies are required, one
running and two in memory.
In order to reduce this overhead, the Spec node has the ability to dynamically map
pages into the program and data memories of the microcontroller. This allows the device
to download its new code image into ram and then remap the ram pages into the
instruction memory. With this method the code images can be swapped in place. This
results in a 33% reduction in memory requirements.
This is implemented by performing an address translation on all addresses leaving
the data path. The top 6 bits of both the instruction and data address are passed through
a translation table that maps the address to one of 16 physical page frames. The Spec
node currently had 6 of these frames filled with memory. With this setup, arbitrary
allocations of data and instructions memories can be achieved. Additionally the small
512 byte page size keeps fragmentation to a minimum. The translation tables can be
modified by accessing special memory address locations.
6.2.4 Timing Extraction
During transmission precise synchronization between transmitter and receiver
must be maintained. Slight skew in timing mechanisms may cause bits to be lost if
periodic re-synchronization is not performed. While it is not strictly necessary for short
transmissions, it becomes required as transmission length increases.
Currently, the Mica platform does not support long packet lengths. For it to
support long transmission, the receiver would have to periodically measure edge
143
transitions in order to re-synchronize. With the Mica’s CPU already heavily loaded
during transmissions, this is a difficult operation.
In order to facilitate long transmissions as well as to improve the reliability of
short transmissions, the Spec node includes a timing extractor primitive that
automatically signals the center of each bit transition. The timing extractor samples the
channel at 5x the bit rate and searches for perfect edge transitions. If a clean transition
occurs within 1/5 of a bit time of the expected transition, it updates the expectations. A
clean transition is when 4 consecutive samples of ‘0’ are followed by four consecutive
samples of ‘1’ or vice-versa. This allows transmitter/receiver pairs to successfully
communication even if their time references differ by up to 10%. Had the Mica node
had this ability it could have used a low-cost 4 MHz resonator instead of the high
accuracy 4 MHz crystal.
6.2.5 Encryption Support
In addition to providing primitives for low-level operations, the Spec node also
includes primitives to support complex operations such as encryption. The Spec node
includes an encryption accelerator that can be used to automatically encrypt and decrypt
messages for transmission. The primitive includes four separate 40-bit LFSRs that are
xor-ed together generate a single random sequence. This parallels the encryption
methods used in the Bluetooth wireless standard. This collection of 4 LFSR offloads a
majority of the overhead associated with encryption. Unlike in Bluetooth, the processing
for seeding the LFSRs is done in software.
144
In hardware, the LFSR is simply a shift register augmented with feedback through
a handful of xor gates. With the 4 LFSRs containing 160 registers, the power
consumption to perform the encryption is just 160 pJ/bit. In contrast, a software
implementation of the same algorithm on the MICA CPU core would require 34 (15
load/stores, 19 ALU) instructions per bit for each shift register. This would result in a
total of 196 cycles per bit for all 4 LFSRs, or 735 nJ per bit – a 4500x increase.
Additionally, at 196 cycles per bit, the maximum encryption/decryption speed would be
just over 20 Kbps – well below the current transmission rate of 50 Kbps. The area of the
encryption accelerator is just .06 mm2, roughly the same size as the memory address
translator.
6.2.6 Memory I/O and serialization
While the Mica node uses clever optimizations to allow the CPU to interact with
the radio in byte chucks, CPU involvement still accounts for a significant amount of
overhead. Periodic per byte interactions with the radio also prevents the CPU from
entering its lowest power sleep state. On Mica node, the byte level interaction combined
with the overhead of handling interrupts actually sets the maximum transmission rate it is
able to achieve.
The Mica node is only able to achieve a maximum bit rate 50 Kbps despite having
a radio capable of 115 Kbps. Additionally, their CPU must be tasked to service the radio
every 160 us when transmitting at 50 Kbps. With each byte transmitted, the controller
has to execute for approximately 100 cycles. This results in the CPU consuming 375 nJ
per byte.
145
To improve this, the Spec node incorporates a memory-mapped data serializer to
reduce the energy consumption while sending and receiving data. Upon detection of
start symbol or initiation of transmission, base and length registers are used to
automatically read/write data directly to/from ram. After initiating the operation, the
CPU doesn’t need to be involved until the operation signals its completion. Instead of
continually passing each byte of data to the radio every 160 us, the CPU can shutdown
and remain in its lowest power state until the transmission is complete. With this
primitive, the core energy required per byte transferred is reduced to the cost of a single
memory operation or just 3.75 nJ/byte – and improvement of 100x.
6.2.7 Primitives not included
In addition to the primitives that we included, we also evaluated several others
that we choose not to include. In general, the energy savings did not outweigh the loss
of flexibility and complexity that they would introduce. The first of these primitives is a
forward error correction primitive.
The Mica node uses SEC/DED error correction. Each data byte that is
transmitted is supplemented by a set of parity bits that can be used to detect up to two
errors and fix up to one error. It has been shown that even this primitive form of error
correct can significantly improve communication performance [10]. However, the benefit
of error correction is dependent on application specific requirements. If channel quality
is sufficiently high, simple error detection and retransmission may result in lower energy
expenditure.
146
The cost of performing error correction used by the Mica node in software is just
55 instructions per byte encoded. Their encoding process is a rate 3 code with each data
byte translated into 3 data bytes. In total the encoding costs just 8.6 nJ per bit. When
compared to the 735 nJ required for encryption this is quite small. Additionally it can be
performed in software at a maximum rate of 218 Kbps.
CRC calculation hardware was also considered for inclusion. On the Mica node,
a 16 bit CRC is attached to each 30-byte packet transmitted. The overhead of that
calculation result in 29 instructions per byte or just 13.6 nJ per bit. Because it costs just
a few nJ per bit, leaving it in software will not have a significant impact on system
performance. Figure 6-5 gives an overview of the performance improvements achieved
by the hardware accelerators that were included on the Spec node.
6.3 Cost of flexibility
Instead of providing the various building blocks described above, we could have
simply provided a hardware implementation of the protocol stack implemented by the
Mica node. Had we done so, the interconnections between the primitives we have listed
would have been performed in hardware instead of software.
Acceleration Type
Performance
without accelerator
Performance with
accelerator
Improvement
Factor
Register Windows 210 nJ/Interrupt 7.5 nJ/interrupt 28
Start Symbol Detection 7.5 mW 5 uW 1500
Encryption Acceleration 735 nJ/Bit 160 pJ/Bit 4594
Data Serilization 375 nJ/Byte 3.75 nJ/Byte 100
Memory Mapping 3 copies required 2 copies required 1.5
Spec Hardware Accelerators
Figure 6-5: Overview of the performance improvements achieved by the Spec node’s
hardware accelerators.
147
In addition to the channel encoding and CRC checking, the TinyOS messaging
stack also performs address and group ID checking on each message and then parses the
message header for dynamic dispatch. These operations are quite inexpensive to
perform in software when compared to a hardware implementation. However, the
dispatch and header parsing only requires approximately 50 instructions. This translates
into just .260 nJ per bit transmitted because the cost can be amortized over an entire
packet. While this could be reduced by using dedicated hardware, the flexibility gained
by allowing for variable packet formats and application-specific addressing schemes is
more significant than a nJ per bit.
6.4 Summary
The Spec chip demonstrates both the value and feasibility of single chip
integration of a wireless sensor network platform. 1000x performance improvements
were achieved on system functions that were dominating CPU energy consumption on
the Mica platform. Additionally, by providing a set of communication primitives instead
of a hardware protocol implementation, spec is able to support a highly flexible set of
protocols. These components included support for start symbol detection, data
serialization, data encryption and timing extraction. Spec also includes architectural
support for the TinyOS operating system by including a paged memory system and
register windows. Its architecture is focused improving performance without sacrificing
flexibility.
Through integration, the Spec chip also drastically reduces device cost allowing
for a wide range of application scenarios. The Mica generation of wireless sensor nodes
has a manufacturing cost around $60. The spec mote requires approximately $0.25 of
148
silicon to be combined with a $0.16 battery, a $0.02 inductor and a $0.10 watch crystal.
It proves that sub-dollar cost for a fully functional wireless sensor node is possible.
149
Chapter 7: Demonstration Applications and
Performance
The ultimate goal of wireless sensor network research is to enable novel
applications that change the way we interact with the world around us. These
applications have requirements that differ drastically from those of traditional wireless
scenarios. In our exploration of these new scenarios, two distinct classes of applications
have emerged. The first is categorized by a low duty-cycle, low data-rate, long latency,
static topology and a long expected lifetime. For these networks, lifetime is the main
evaluation criterion. A second class of applications is that of highly dynamic sense-andcontrol
networks with higher data rates, latency limits, and highly mobile nodes. Instead
of passively monitoring a relatively static environment, these networks attempt to control
the environment in real time. This class places strict latency and throughput
requirements on the network, which are the main measures of system performance for
this class of device. We have evaluated or architecture with respect to both application
classes.
7.1 Environmental Data Monitoring
As presented in Chapter 2, environmental data monitoring is a key application
area for wireless sensor networks. These scenarios include heating, ventilation, and air
condition monitoring of large buildings or climactic monitoring of outdoor habitats. The
architectural optimization opportunities come together to determine the application’s
capability and lifetime at a given energy budget. We use this scenario to track the
150
performance improvements of the four node architectures we have presented (Rene,
Mica, Blue, Spec).
For illustrative purposes we consider a scenario where the network forms itself
into a tree and each node samples environmental data every 4 seconds, aggregates a
statistical summary over 5 minutes, and transmits the summary to its parent node. Each
parent combines the readings from its children and sends a single summary to its parent.
In the event of a drastic environmental change, any node can send an emergency
notification to its parent at any time. This network is simultaneously acting as a data
collection network and an alarm monitoring network. In alarm mode, the parent checks
for emergency messages every 4 seconds. Figure 7-1 summarizes the theoretical
performance of each of the platforms we presented in this thesis. Computing rates are
fixed; the key performance metric is energy efficiency, which we measure as mj/day to
accomplish the task. The performance improvements of the mica node over the Rene
node can largely be attributed to the hardware accelerators. Overall, MICA’s hardware
accelerators and exploitation of application specific protocols results in a 10x increase in
overall application performance over the baseline Rene node. This comparison is
analyzed in more detail below.
7.2 Empirical analysis of performance improvements
To estimate the energy consumption for each of the architectures we have
presented, we sum the energy usage of each of the application components: Data
Transmission, Rendezvous Overhead, Alarm Checking, Data Sensing, and sleep energy.
The energy cost of data transmission assumes that the worst-case node is responsible for
five children.
151
In this application scenario, a node must receive a summary from each of the
children, combine them, and transmit a new summary to its parent. The cost of each
operation is dependent on the power consumption of the radio system and the bit rate that
data can be communicated at. The Blue node shows an 11x improvement over the
original Rene and outperforms the Mica node. This is due to an increase in transmission
rate enabled by the synchronization and buffering accelerator – blue node has twice the
buffering in its serialization accelerator.
As outlined in Chapter 2, because communication is infrequent and the radio
consumes energy whenever it is powered on, nodes must leave their radios powered off
whenever possible. In the data collection application scenario, nodes agree to
communicate at a future time and turn off their radios until that time. In these protocols,
synchronization precision translates into lower power consumption. If a scheduled
application period of communication is only accurate to approximately 100 ms, then the
receiver must wake up 100 ms early to ensure that it is not late; this could result in the
Platform performance on Environmental Data Collection Application
Implementation
Data
communication
(mJ/day)
Rendevous (mJ/day)
Alarm check
(mJ/day)
Sensing (mJ/day)
Total active
(mJ/day)
Sleep (mJ/day)
Total (mJ/day)
Expected Lifetime
on Idealized AA
(years)
Rene 2,289 691.2 52,462 16.2 54,768 4,230 58,998 0.35
Mica 458 34.56 78 16.2 552 4,319 4,871 4.27
Blue 209 72.576 178 1.701 388 259 647 > 10 yr
Spec 2 0.864 6 5.832E-07 9 259 268 > 10 yr
Improvement factor
Blue vs Rene
11 10 295 10 141 16 91 > 40
Figure 7-1: Theoretical node energy consumption when performing environmental data
monitoring where data is collected every 4 seconds, averaged and transmitted once every 5 minutes.
Each node has a maximum of 5 children nodes and one parent node. The table shows results for
Rene, Mica, Blue, Spec. The ratio between Rene power consumption and blue power consumption is
also shown. The alarm check cost of blue is larger than that of Mica due to the CC1000 having a
longer turn-on time. The Sensing cost of blue is reduced by a low power ADC in the microcontroller.
152
receiver waiting 200 ms before the transmission actually begins. For the short
transmissions common in sensor networks, this can result in a 4x to 5x increase in cost
over basic packet transmission and reception. The use of bit-level time synchronization
on the mica and blue nodes reduces the energy consumed while waiting for the scheduled
communication by a factor of 10. You will also note that the Mica node outperforms the
Blue node on this operation. This is due to the increased receive power consumption of
the CC1000 radio.
In order to meet the strict alarm propagation requirements, each node must also be
continually ready to forward alarm messages. This means that they must be frequently
checking the RF channel for activity. The relatively high frequency of the wake-up
check makes it the single largest consumer of energy in the Rene design. However,
exploiting the low power wake-up functionality on Mica and Blue reduces this cost by as
much as a factor of 600 on the Blue node. Once again the Mica node performs better
than the Blue node due to the characteristics of the CC1000. Spec performs significantly
better because of its fast radio turn-on time.
Taken together, the three optimizations (RF wake up, time synchronization, and
high-speed communication) reduce energy consumption of the active part of the
application by two orders of magnitude. In Figure 7-1 we also show the impact of these
optimizations on the overall impact on system lifetime. In the Rene base case, the power
drain in sleep state was a reasonable 10% of the overall budget, but it dominates the
optimized application. When looking at the application as a whole, battery life is
improved by a factor of more that 90, resulting in an expected lifetime of greater than 10
153
years for a pair of AA batteries. We do not predict a longer lifetime because the shelf life
of an alkaline battery is less than 10 years.
It is interesting to note that the added performance improvements of the spec node
do not have a significant impact on the application lifetime. For analysis purposes we
have assumed that the sleep mode power consumption of the spec node is the same as the
blue node. Because the sleep mode power consumption dominates this application
scenario once the other factors have been addressed, the improvements in radio power
consumption are insignificant. This, however, will not be the case in a node tracking
scenario.
7.2.1 Hardware Trial
To verify the empirical analysis performed, we demonstrated this application
scenario on the blue platform. Once again, a node collects data periodically from a local
sensor, averages its readings and transmits data up once per second. The hardware trial
was performed to confirm the theoretical results.
The data collection application used precise time synchronization to maintain
communication schedules. The hardware trial showed that transmitter and receiver pairs
are indeed able to maintain synchronization of better than 2 ms over the course of the 5
minute communication intervals. At the end of each interval, nodes exchange pair-wise
time synchronization information. This information is used to maintain the <2ms
synchronization indefinitely. The resulting duty cycle of active communication is just
.02% or 60 ms every 5 minutes.
Additionally, when measured experimentally, the blue mote’s periodic wake-up
checks require 8.8 uJ per check or 190 mJ per day. The experimental sampling cost is
154
just 42 nJ per sample or .907 mJ per day. The net result of the hardware trial is that the
Blue implementation performs roughly equivalent to the theoretical predictions.
7.3 29 Palms
Another application scenario that has been implemented on the Berkeley motes is
that of an isolated network designed to monitor activity at a remote location. The “29
Palms” application used the Rene mote platform in conjunction with magnetometers to
detect vehicle activity on a remote desert road.
7.3.1 Application Description
As part of an experiment with the US marines, the motes were deployed to detect
vehicle activity at an isolated intersection in the desert near Palm Springs, California. A
collection of nodes were dropped in a line along the side of a road. They were dropped
Figure 7-2: 29 palms application. Motes dropped out of an
airplane self-assemble onto an ad-hoc network in order to monitor for
vehicle activity at a remote desert location.
155
from a small UAV that was flying autonomously based on a GPS flight plan. The plane
was pre-configured with a multi-point flight plan which included a low-altitude
“bombing run”. The motes were released from approximately 100 feet along a track
parallel to the road.
Once deployed, the nodes configured themselves into a multi-hop network and
synchronized internal clocks so discrete sensor readings could be correlated across
multiple nodes. As vehicles passed the network, individual nodes used magnetometers
to detect deviations in the magnetic field caused by metal contained in the vehicles.
Each node determined the closest point of approach for the vehicle and then assigned a
time stamp for the vehicle event. Events from each sensor were then communicated to
neighboring nodes. Once a node collected 5 readings, it performed regression analysis to
determine the velocity and direction of the vehicle.
The high-level vehicle track information was then logged in persistent storage.
Over time, each node collected a set of vehicle track logs in its memory. These logs
could then be downloaded by a querying node. In the case of the demonstration
application, the vehicle tracks were downloaded back to the UAV that originally released
the nodes. After deployment, the UAV departed the area and returned later to retrieve
vehicle detection events. Upon request, each node uploaded its data to a node contained
in the plane. The plane then returned to base so that the data could be downloaded to a
PC for display. The plane itself acted as a physical data shuttle.
156
7.3.2 Key sub components/application blocks
The 29 Palms application was divided into several distinct sub-pieces: sensor
calibration, data analysis, time synchronization, data communication, data regression,
position estimation and data logging.
7.3.2.1 Sensor Calibration
Upon deployment, each node must re-calibrate its highly sensitive magnetometer.
Any change in node orientation would cause a relative change in the earth’s magnetic
field. This change must be calibrated out of a sensor once the node has reached its final
resting point. This was performed by having a control loop that observed the readings
coming from the magnetometer and detecting when the plane flight ended and the node
was stationary on the ground.
Once the node determined that it reached its final resting point, the node would
begin to tune its physical hardware to compensate out the earth’s magnetic field.
Essentially they attempted to remove the DC component coming from the magnetic
sensor. This was done through the use of digitally controlled potentiometers [60].
7.3.2.2 Data Analysis
Once calibrated, each node continually analyzed the magnetometer sensor
readings in an attempt to detect a vehicle. The analog sensor was sampled at 32 Hz and
then passed through a pair of IIR filters. The filters were carefully constructed to
perform a band-pass operation on the incoming data stream. Each node searched for a
transient signal occurring at .2-2 Hz. When a signal arrived in this frequency band, the
node assigned a time stamp to the peak of the arriving signal. Figure 7-3 provides a
157
graph of the incoming signal from a passing car. The red line represents the filtered
signal.
7.3.2.3 Time Synchronization
In order to calculate vehicle velocity, each node must record when a vehicle
passes with respect to a global time base. This allows events from multiple nodes to be
compared to each other. For this application we targeted time synchronization accuracy
of +/- 30 ms. This level of synchronization is sufficient to successfully determine the
closest point of approach of a vehicle traveling 50 mph to approx 2 feet.
Each node maintained synchronization using a clock advancing algorithm. Each
node would periodically broadcast out its current clock reading. Upon reception of such
a broadcast, the node would compare the reading to its local clock and set its clock value
to the larger of the two readings. This allows the entire network to automatically lock to
the node with the fastest running clock.
Figure 7-3: Data filtering performed locally on modes to
amplify and extract vehicle signal. Time stamp assigned to the
highest peak of the red activity line.
Filter output
Raw signal
Detection
event
158
7.3.2.4 Data regression and position estimation
Nodes used linear regression to determine the vehicle velocity form a collection
of readings. Once a node had collected at least 4 vehicle detection events, it would wait
for one second – to collect any remaining readings – and then perform a linear regression.
The result of the linear regression was an estimation of the vehicle velocity and an
estimate of the time it entered the network.
Any error between the estimated track and the actual sensor readings was
attributed to the inaccuracy of the sensor readings and to the inaccuracy of the position
estimation of each node. The result of the regression was then fed back to help refine the
position estimation for each node.
After the initial deployment, each node only had a rough estimate of its location.
The nodes were released approximately 10 feet apart. However, after falling for 100 feet
and bouncing off the ground, the position of each node will be slightly skewed. Over
time, each vehicle track is used to update the position of the sensor nodes through
􀀀


 

 
 
 
 
 
 

!
"
#
$
%
%
%
%
%
%
&('*),+.- /103245+565/765812 8
9 '7: /;81<=2>5?5@581A /;<
B C 22 +
6D/12 /;A E7C 45/7FG/;H5C 0,-/7FG/-+50C 2 I
J '.KLD65812 /745+565/7L5+5<=C 2 C +54M/;<=2 C E7812 /;<
Figure 7-4: Position estimates of each of the sensing node are
updated in response to the result of the linear regression that
estimates the passing vehicle’s velocity.
159
feedback. The network first calculates the vehicle speed and then asks itself, “If the
vehicle was actually going 15 mph, where are the sensors?”
7.3.3 Importance of in-network processing
The 29 palms application scenario demonstrates the importance of in network
processing. Each individual sensor node processes over 400KB of data per hour.
Collecting the data from hundreds of sensor nodes at a central analysis point would
quickly overwhelm the communication capabilities of the network. Additionally, the
power consumption would be unacceptable.
By including the data processing at the individual sensor, the communication
requirements are drastically reduced. Additionally, the power consumption of the
network is tied to the number of vehicles that are detected, not to the amount of time that
has elapsed. This is ideal for a network that needs to remain in place for several months
in order to confirm that there is no activity in the area. When there are vehicles to detect,
the power consumption increases locally to share the vehicle events with neighbors and
then declines once the vehicles have passed. Only important readings consume the
valuable communication time and energy. Central data analysis requires that all data be
communicated regardless of the information value.
7.4 Z-Car Tracking
A more controlled follow-on application to the 29 Palms vehicle tracking demo
has been done called Z-car tracking. In this applications, a regular grid of sensor nodes
form a mesh of sensing that can automatically direct “high-level” resources towards
interesting events.
160
7.4.1 Application Description
The Z-car tracking application has a network of nodes designed to track a vehicle
once it enters the sensing grid. Unlike the 29 Palms application, the vehicle detection
events are immediately communicated to a high-level resource for prompt action. The
small-scale demo involves tracking a remote-controlled car with a video camera through
a network of 25 nodes. The vehicle detection algorithm used on each node is the same
as that of the 29-Palms tracking demo.
When the vehicle is detected inside the network, local nodes exchange sensor
readings with their neighbors and synthesize a single report message that contains an
estimated position for the vehicle. This message is then handed to a multi-hop
communication mechanism that delivers the report to a camera located at the end of the
network. For demonstration purposes, the network is only 20 x 20 feet and the vehicle
Figure 7-5: Picture of z-car tracking application deployment.
Remote control car tracked as it passed through the field of sensors.
161
being tracked is a small remote controlled toy. However, this demo could be easily
scaled to detect larger vehicles by using larger grid spacing.
The multi-hop data communication is based on geographic routing where the
packet is addressed to nodes that are physically closer to the data’s destination. The data
eventually arrives and a special mote that is in control of the video camera. The readings
are then used to calculate the most likely position of the vehicle and the video camera is
pointed accordingly.
7.4.1.1 Bi-modal operation
The Z-car tracking application has a Bi-modal activity profile. During a real
deployment, days or weeks may pass before any activity is detected. During this time,
the nodes must remain in a dormant, low-power state. However, they cannot be
completely turned off. At any time, they could either detect a vehicle or be called upon
to be router node in the multi-hop network.
When activity is required, the network snaps into a high-duty cycle state.
Messages are transmitted from the sensing cluster to the camera once per second. All
nodes along the routing path from the vehicle to the camera must also remain active
because they may be called upon at any time to take over as a primary routing node.
The Z-car application demonstrates the ability of TinyOS and the Mote
architecture to flexibility meet a wide array of protocol requirements.
7.4.1.2 Leader Election
While tracking a vehicle, sensor readings from several nodes are synthesized into
a single report packet in order to conserve network bandwidth. A vehicle may trigger
several sensor nodes as it passes through a network. The information coming from
162
nodes that are further from the vehicle is less valuable that the data coming from sensor
nodes that are closer to the vehicle.
In order to conserve network bandwidth, a local leader is elected to review the
data coming from several sensors and to select the most significant data. The filtered
data is then handed to the multi-hop layer for the expensive trip across the network. In
actuality there are two levels of filtering being performed simultaneously. Each node is
continually filtering the raw sensor readings and transmitting its neighbors only what may
be significant. Then, the leader then takes the data and determines which of data
readings are the most significant.
The leader is elected by being the node with the strongest sensor reading in a
given time epoch. Each node collects the reports coming from all of its surrounding
neighbors. It then compares the reports to its own local report. The node that has the
strongest sensor reading is the leader and selects data from the next three best nodes.
This election mechanism was selected because of its low overhead and its tolerance to
lost messages.
In the event that a message is lost, our leader election algorithm guarantees that
the best data will still be reported. Regardless of what packets are received by the true
leader, by definition he will still have the highest quality data. If the leader’s data is not
properly received by other nodes, then a second leader will be accidentally selected and
two packets will be transmitted to the base. In either case the most valuable data is
always transmitted.
163
7.4.1.3 Reliable Message Delivery
Once the data from several sensor nodes has been aggregated together, it becomes
essential that the data is delivered reliably across the network. In many sensor
applications, lost data is acceptable because the data transmitted is highly redundant.
However, we are intentionally transmitting only highly valuable data.
To implement a reliable delivery mechanism, we augmented our communication
primitives to include link-level acknowledgements. Each time the application initiates a
message transmission, it automatically waits to receive an acknowledgement. When the
TinyOS system returns the SendDone event to the application, it includes information
about whether an acknowledgement was received or not. A positive acknowledgement
signifies that the messages was received by the node that it was addressed to and that the
receiving node verified the integrity of the packet using a CRC mechanism.
In the event that an acknowledgement was not received, our multi-hop routing
layer automatically initiates a retransmission. This is repeated up to three times before it
is assumed that the communication link has been broken. In contrast, the initial sensor
report messages are not retransmitted at all. This is because they are simply broadcast
messages to surrounding neighbors. There is not a single intended recipient.
7.4.1.4 Geographic Based Routing
The multi-hop routing mechanism used was based on geographic information.
The physical location of each node was used to determine a routing structure. Each node
was told its location on in the plane. A node would route data through any node
that is physically closer to the destination.
164
With a grid based network layout, it is easy for each node to calculate the location
of nodes closer to the base station. Each node attempts to transmit to any node that is up
to 2 grid units away from itself. Greedily, each node first tries to contact the nodes that
are both one step in the X direction and one step in they Y direction towards the base
station. If a transmission to that node is not successful after three attempts, another node
is selected and the process is repeated. Figure 7-6 displays the network behavior while
tracking a node.
Once a successful path has been established, the nodes continue to use the same
routing path. Only after a local retransmission has failed three times does the network
select a new path.
One of the problems faced by geographic routing is that local minima can be
formed from which data cannot escape. Sometimes you have to send data away from the
destination in order to avoid an obstacle. Unfortunately, we did not address this issue.
Route packet
traveled
Evader Track
Leader Node
Evader
Aggregated Nodes
Base Node
Figure 7-6: Multi-hip network routing used in z-car tracking
application. Nodes detecting the event (in green) transmit data to a
single leader that transits a single report packet. The report packet
travels across the network (dotted line) until it reaches the base node.
The packet contains the estimate of the evader’s (red) position.
165
Our grid based deployment minimizes the presence of local minima. Multiple node
failures are required to create such an obstacle in our grid. Future work would be to
include algorithms to avoid such phenomenon.
7.5 Conclusion
Hundreds of applications have been constructed that demonstrate the efficiency
and capability of the Mica, Blue and Spec platforms. We have presented three key
applications that demonstrate several of the capabilities of the overall system. Low-duty
cycle networks built with the Blue node have demonstrated the ability for this
architecture to support precise synchronization which leads to very low network duty
cycle and multi-year lifetime for the environmental data monitoring scenario.
Additionally, complex vehicle tracking deployments have shown TinyOS’s ability to
meet application specific needs for data communication, data aggregation, sensor analysis
and multi-hop routing.
166
Chapter 8: Related Work
In looking at related work for wireless sensor networks, there are three key areas
of active research. The success and widespread adoption of TinyOS has lead to a body
of work that is directly built on-top of the TinyOS/Mica platform. Secondly, there is a
body of work related to other wireless platforms that are applicable to wireless sensor
networks, and finally there is a body of work related to embedded operating systems that
may be used in wireless sensor networks.
8.1 TinyOS Supported Work
In this section, we will quickly provide an overview of work that has been done
on top of the TinyOS platform. The Mica node has been distributed to over 250 research
groups around the world as a platform to support wireless sensor network research.
Additionally, it has been the primary research platform for the NEST research program
[61]. The Network Embedded Software Technology (NEST) research program is a
DARPA funded program with dozens of researchers at more that ten locations around the
nation that are investigating the software technologies that must be developed to support
distributed, embedded systems including those of sensor networks.
8.1.1 Wireless Robots
One of the research areas inside of the NEST program is that of wirelessly
controlled robots. Traditionally research into autonomous robotics has been performed
on robotics platforms that cost tens of thousands of dollars. These devices include and
embedded PCs running windows or Linux and communicate with 802.11. As an
167
alternative, several research groups have developed low-cost robots that are controlled by
a Mica node running TinyOS.
The COTS bots project at UC Berkeley has built their robots out of low-cost
remote control cars [62]. By replacing the standard control system with a wireless
sensor/control node, a child’s toy is transformed into a research platform. The speed,
direction and turning radius are controlled by the wireless sensor node. The COTS bots
project created a robotic layer of abstraction out of TinyOS components that refines the
basic motor control primitives into TinyOS commands and events
The COTS-bots project is also investigating the integration of sensors to help with
navigation. Obvious sensors to include are acceleration sensors, rate gyros, and tilt
sensors. Early results from analysis of acceleration data from an ADXL202 suggests
that acceleration data alone is not sufficient to perform inertial navigation even when the
system is constrained to a single dimension [63]. Low frequency drift in sensor readings
leads to unbounded error. However, the accelerometer has been successfully used to
detect impacts with obstacles. Additionally, it looks promising to use multiple sensors
together to perform inertial navigation.
One of the goals of the COTS-bots project is to develop algorithms for
coordination between multiple robots. Instead of deploying a single high-cost robot,
they want to deploy a fleet of interconnected robots. This is enabled by using the multihop
communication algorithms and in-network processing being developed in wireless
sensor networks.
168
8.1.2 Media Access Control and Routing
Media Access Control and transmission scheduling is a key area of wireless
sensor network research. Media Access Control, the mechanism for determining who
can transmit and when, must be used to optimize for device power consumption.
Researchers at UCLA have demonstrated the benefit of exploiting application specific
protocols by creating customized MAC layers adapted to sensor networks. Their
customized sensor network communication protocols attempt to reduce energy
consumption and have shown 2-6x energy improvement when compared to standard
801.11-like wireless networking protocols [51]. Their work demonstrates the advantage
of using a flexible, modular system that facilitates customization.
Researchers at UCLA have also generated optimized routing protocols designed
to conserve energy in deployed networks[2, 11]. In data collection networks, schemes
must be used to evenly distribute load across potential routing nodes. As described in
Chapter 2, nodes that are forwarding data for other nodes will consume more energy than
other nodes. Over time, their batteries will be depleted and they will be unable to
continue forwarding data. This may result in degraded network performance. Evenly
dividing the load across all possible routing nodes ensures the maximum possible
network lifetime.
8.1.3 Time Synchronization
Time synchronization is a classic distributed systems problem that has been readdressed
in the context of wireless sensor networks. Researchers at UCLA have pushed
the limits of distributed time synchronization over wireless sensor networks. The driving
169
application behind their time synchronization research is to achieve precise localization
through acoustic ranging.
The UCLA approach is to use the broadcast nature of wireless communication to
send timing signals to multiple nodes [50]. Any node in the network can send out a
timing beacon. Any node that hears the beacon records a precise, local time stamp for
the signal. It then exchanges its time stamp with all nodes that have also heard the same
timing beacon. This mechanism allows each node to determine the relative difference
between its local clock and a neighbor clock that also heard the beacon. Each node
stores the relative differences between its clock and all neighbors. Additionally it stores
an estimate rate of the relative drift. UCLA has been able to demonstrate microsecond
accurate time synchronization using this mechanism on the Mica platform.
When a sensing event is detected inside the network, all nodes apply a local time
stamp to the event. The nodes then transmit the time stamped detection event through
the wireless network. Each time the data is transmitted the receiving node updated the
time stamp to account for the difference between its local time base and the time base of
the transmitter. The time reading is dynamically updated to be relative to the local clock
as it is transmitted across the network. As the data propagates through the network, the
time stamp will be updated multiple times. All data that reaches a central collection
point, will have time stamps that are relative to the same local time base.
8.1.4 Multi-Hop Routing optimization
One of the most important areas of research being supported by the TinyOS
platform is that of multi-hop routing optimizations. The wide range of application
scenarios being considered for wireless sensor networks leads to a wide range of potential
170
multi-hop routing protocols. Tradeoffs in latency, bandwidth and power can be made in
each application scenario.
Researchers at UC Berkeley have implemented several generations of multi-hop
data collection algorithms. Reference [64] presents a data collection routing protocol
designed to construct an optimum spanning tree topology. Instead of searching for a
minimum depth spanning tree, reference [10]searches for a tree that maximizes the
probability that each node’s data will reach the base station on the first try. Each node is
assigned a success probability metric that starts from the base station and propagates out.
The base station has 100% probability of success and each of its neighbors has a success
probability that is proportional to the link probability to the base station. Nodes that can
barely communicate with the base station will choose to send data over an extra hop in
order to increase the success probability.
8.1.5 TinyDB
One of the most exciting systems to be constructed on top of the TinyOS platform
is one call TinyDB [65]. TinyDB is designed to transform a wireless sensor network
into streaming database. SQL queries are entered into a user interface and propagated
onto the sensor network. The network executes the queries over incoming sensor
readings and returns the results. Just like in standard SQL, these queries can include
aggregation operations to refine the data into averages, max, or min values [1]. The
TinyDB query optimizer automatically distributes the aggregation operations into the
network.
TinyDB provides a mechanism for reasoning about how to access data coming
from wireless sensor networks. It provides functionality beyond that of standard SQL in
171
that it supports both long-running queries and event triggers. Unlike in traditional data
bases, the data set being queried in wireless sensor networks is an infinite stream of data
extending into the future. Every second or perhaps millisecond, each sensor has the
potential to generate a new sensor reading. Long-running queries can be issued to the
network requesting data to be reported if it exceeds a threshold or fits a special criteria.
In TinyDB the parameters of the query will be propagated to each individual node
allowing for local data filtering. In addition to automatically transmitting data that meets
special criteria, TinyDB also allows for triggers to be included that cause other local
actions to occur when a specific sensing condition occurs.
8.2 Wireless Platforms
In addition to the collection of TinyOS related platforms discussed earlier, there
are a collection of other wireless networking platforms that are relevant.
8.2.1 Smart Dust
The Smart Dust[66, 67] project at UC Berkeley represents the inspiration for
much of the work contained in this thesis. The project has continued to develop
miniature scale hardware structures that are advancing the state-of-the-art in wireless
sensor network technology. In targeting extreme miniaturization and low-power
consumption they are developing many ultra-low power primitives including the radio
and ADC components contained on the Spec node.
Additionally, the Smart dust project has pioneered new optical communication
technologies through their use of MEMS mirror-based optical communication. By
modulating reflected light, their corner cube reflector communication structure is able to
172
transmit data across mile long links while consuming microwatts [68]. In addition to
supporting the development of Spec, they have developed showcase devices that include
sensing, computation, solar power sources, and optical communication. They have
demonstrated autonomous sensor node that are less that 10 mm3.
8.2.2 Bluetooth
Bluetooth is a wireless system designed to be the industry standard for lowpower,
mobile devices [69]. Highly integrated chipsets are being developed that provide
RF circuitry and protocol processing on a single chip. While these designs will provide
highly efficient implementations of the Bluetooth protocol, they will not provide the
flexibility demanded by wireless sensor networks.
Bluetooth has been designed as a wire replacement and includes support for very
low-latency communication. To support low-latency and high-throughput operation,
Bluetooth uses a channel hopping period of just 600 us. This requires all devices to
remain synchronized to within a few microseconds. In wireless sensor networks devices
may only send data once per hour. It would be highly inefficient to force these nodes to
remain synchronized for the entire hour. However, in Bluetooth it is expensive to enter
and leave a network, it would be impractical for low duty cycle nodes to continually enter
and leave a Bluetooth piconet. In a standard configuration, it takes over 2.4 seconds for
connection establishment. During this time, the master node must be in a high-power
scanning mode – typical Bluetooth radios consume hundreds of milliwatts wile
monitoring the channel.
173
Another incompatibility between wireless sensor networks and Bluetooth is that
the master-slave topology of Bluetooth requires each slave to report to the master every
few seconds. Again, this is not efficient for wireless sensor networks. While the
Bluetooth chipsets are highly efficient and well integrated, their inability to provide lowpower
operation modes makes it inefficient for wireless sensor networks.
8.2.3 Zigbee (802.15.4)
Zigbee is an industrial consortium designed to build a standard data link
communication layer for used in ultra-low power wireless application [70]. The Zigbee
alliance was formed because its members felt that existing standard technologies were not
applicable to ultra-low power application scenarios.
The Zigbee data link layer is designed to operate on top of the IEEE 802.15.4
physical layer [71]. IEEE 802.15.4 is a direct sequence spread spectrum physical layer
including transmission bands at 868 MHz, 902-928 MHz and 2.4 GHz. First generation
chipsets will be available in Q4 of 2003 [72]. Direct Sequence Spread Spectrum has a
distinct advantage over channel hopping mechanisms because hop-sequence
synchronization does not have to occur prior to initiating communication. This provides
a large power advantage for low-duty cycle devices and addresses a major shortcoming
of Bluetooth.
Once completed, the Zigbee standard will specify a communication mechanism
that is comparable to the AM communication layer provided in TinyOS. Primitive data
framing mechanisms, error detection, and addressing will be standardized to allow
compatibility. Zigbee is primarily focused on star topology networks where low-wireless
devices communicated to powered collection point. A canonical Zigbee application is a
174
wireless light switch. This would involve a low-cost battery operated switch
communicating to a powered light fixture.
8.2.4 Pico Radio
There are several other groups looking into the architecture of wireless embedded
devices. The Berkeley Wireless Research Center’s PicoRadio project has identified the
importance of application specific protocols, and have built a flexible platform by
exploiting reconfigurable hardware [73]. The design incorporates reconfigurable
building blocks that connect to their PicoRadio protocol processor, to give it the
flexibility to implement a large number of underlying protocols. However, their overall
node architecture does include a dedicated protocol processor and maintains a partition
between protocol and application processing. The flexibility in protocol structure will
still be limited by the protocol processors external interface.
8.2.5 Chipcon CC1010
Several radio manufactures have identified the benefits of integrating general
purpose microcontroller onto the same CMOS die with the radio. One commercial part
that has just become available is the Chipcon CC1010. It is the integration of a CC1000
transceiver and an Intel 8051 microprocessor. While they were able to bring the
configuration registers of the radio onto the I/O bus of the CPU, they did not take the step
of providing any additional integration to aid in the communication process. The data
interface to the radio remains largely unchanged from what was possible when the
microcontroller and radio are two separate chips.
175
We have demonstrated how the integration of CPU and radio on to a single chip
creates new opportunities for interfaces between the radio and the CPU. However the
simple form integration done by Chipcon in their CC1010 has simply reduced physical
size and cost. The inclusion of communication accelerators are required to allow
integration to yield significant performance improvements. The CC1010 does not
realize the full potential that is allowed by our generalized architecture.
8.3 Embedded Operating Systems
Traditional real time embedded operating systems include VxWorks [75], WinCE
[76], PalmOS [77], and QNX [77, 78] and many others [79-81]. Figure 8-1 shows the
characteristics for a handful of these systems. Many are based on microkernels that
allow for capabilities to be added or removed based on system needs. They provide an
execution environment that is similar to traditional desktop systems. The POSIX [82]
compatible thread packages allow system programmers to reuse existing code and
multiprogramming techniques.
The largest Real time operating systems provide memory protection given the
appropriate hardware support. This becomes increasingly important as the size of the
Figure 8-1: Characteristics and requirements of commercial embedded real-time
operating systems.
176
embedded applications grow. In addition to providing fault isolation, memory protection
prevents corrupt pointers from causing seemingly unrelated errors in other parts of the
program allowing for easier software development. These systems are a popular choice
for PDAs, cell phones and set-top-boxes. However, they do not come close to meeting
the requirements of wireless sensor networks; they are more suited to the world of
embedded PCs. For example, a QNX context switch requires over 2400 cycles on a
33MHz 386EX processor. Additionally, the memory footprint of VxWorks is in the
hundreds of kilobytes. Both of these statistics are more than an order of magnitude
beyond the capabilities of the Rene platform.
There is also a collection of smaller real time executives including Creem [83],
pOSEK [84], and Ariel [85], which are minimal operating systems designed for deeply
embedded systems, such as motor controllers or microwave ovens. While providing
support for preemptive tasks, they have severely constrained execution and storage
models. pOSEK, for example, provides a task-based execution model that is statically
configured to meet the requirements of a specific application. Generally, these systems
approach meet space requirements of Rene and represent designs closest to ours.
However, they tend to be control centric (controlling access to hardware resources) and
do not provide the necessary levels of concurrency. Even pOSEK, which meets our
memory requirements, exceeds the limitations we have on context switch times. It is
unfortunate to note that while there is a large amount of information on code size of
embedded OSes, there are very few hard performance numbers published.
177
Chapter 9: Conclusions
This thesis has presented a system architecture for wireless sensor nodes that is
capable of addressing the strict requirements of wireless sensor networks. By utilizing a
single shared controller that is augmented by a collection of specialized hardware
accelerators, the architecture is able to support flexible, application-specific
communications protocols without sacrificing efficiency. This architecture has been
validated through the development of three hardware platforms and a software operating
system.
We have developed the TinyOS operating system which provides the fine-grained
concurrency mechanisms required to implement wireless sensor network protocols and
applications. TinyOS leverages an event based execution model to efficiently share a
single processor across multiple independent functional operations. Additionally,
TinyOS provides a highly-efficient component model that has almost no runtime
overhead yet allows application developers to partition applications into easy-to-manage
modules. This allows for the component modules to be verified independently before
composing them together into a complete application.
We have presented a generalized architecture that addresses key issues that arise
when building a wireless sensor network device that must meet strict power consumption
and size requirements. They include: flexibility, fine-grained concurrency, precise
synchronization, and decoupling between RF and data path speed. We argue that these
properties must be present in the system architecture in order to support wireless sensor
network applications. The platform must be flexible enough to meet the wide range of
178
application requirements that sensor networks are addressing. We have identified core
application scenarios that range from environmental data collection, to security networks,
to node tracking networks. Each scenario has substantially different communication
patters and protocols that must be supported by a single hardware architecture.
To validate our general architecture we first presented the Mica node. Composed
from off-the- shelf components, it only approximated our general architecture. It
included specialized hardware accelerators that help decouple the RF and data path speed
and increase the synchronization accuracy of communication protocols. In evaluation of
this platform we demonstrated how these simple accelerators result in significant
performance improvements without loosing flexibility. The Mica platform has proven
itself both in theory and through deployment in long-term, battery operated application
scenarios.
We have made the leap beyond the capabilities of off-the-shelf hardware by fully
realizing our general architecture in the form a single-chip CMOS device called Spec.
Unconstrained by the capabilities of commercially available components, we have
integrated a suite of hardware accelerators and a custom radio in order to realize order-ofmagnitude
improvements on key evaluation metrics. These include crucial metrics such
as communication rates, processing overhead for start symbol detection and timing
accuracy. Spec is representative of the future of wireless sensor network devises.
In addition to micro-benchmarks and theoretical analysis, we have also presented
real-world application deployments. We have grounded our study by going as far as
taking our nodes to the desert to track military vehicle movement during a live-fire
training exercise. We have also tracked scale size vehicles in mock-up scenarios, and
179
deployed countless-other experimental networks. The Mica platform combined with
TinyOS has been delivered to over 250 organizations to be the foundation of a
nationwide effort into wireless sensor network research and development.
For several years, people have been imagining possible application scenarios for
wireless sensor networks. The system architecture presented in this thesis is ready to
make those visions a reality. The first commercial usage of wireless sensor networks will
likely be in non-mission critical application scenarios. Systems designed to help
improve efficiency or provide insight into how a business or ecosystem operates can be
deployed without risk of impeding existing activities. Additionally, military applications
for wireless sensor networks will become a reality in the near future. Wireless sensor
networks have the ability to provide a radically new perspective on the state of the
battlefield. Military commanders welcome technologies that integrate their information
gathering capabilities and can incorporate the data coming from senor networks without
abandoning existing information sources.
The last stage of adoption for wireless sensor networks will be when mission
critical information is trusted to them. Industrial process control and security systems
demand extremely high levels of reliability and predictability. These systems must be
tested and validated for years before they can be trusted for deployment. Failures in an
industrial process control system can lead to millions of dollars in lost productivity,
injury to personnel, and damage to equipment. While the adoption rate will be slow in
these areas, the technological advantages of wireless sensor networks will outweigh the
risks.
180
While the platforms presented here are ready to meet the demands of real-world
commercial applications, the technology enabling wireless sensor networks will continue
to evolve. As advances in CMOS processes and RF radio technology are incorporated
into next-generation wireless senor nodes, the power consumption and lifetime will
continually improve. Currently, technology allows for multi-year operation off of a
single pair of AA batteries. The upcoming technological advances will most likely be
applied to decreasing the power consumption of the device. In turn, this will enable a
reduction of physical size of the energy storage required for any given application. As
for tighter levels of integration, the cost/size point represented by the Spec platform has
reached the point of diminishing returns. Further reduction in the physical size of the
radio, processing, and storage is no longer necessary. Only a select few applications have
the need for a device that is smaller that 2.5 mm x 2.5 mm. However, all application
scenarios can benefit from reduced power consumption which is translated into longer
network lifetime and/or increased sample rates.
Fifteen years from now, wireless sensors will be a “behind-the-scenes”
technology that has grown to impact every aspect of our lives. All factory and machine
command and control systems will have switched over to relying on wireless sensing and
control points. The millions of miles of cumbersome wiring you hear about in building
control and automation systems will be replaced by an invisible wireless mesh. The
devices themselves will become as common place as light switches and thermostats are
today. They will be tiny, cheap, commodity pieces of silicon that interact with the
physical world. Today we give little thought to the modern electrical grid and what our
181
lives would be like without it. Tomorrow we will give little thought to wireless sensor
network technology and the systems that have grown to impact every aspect of our lives.
182
Bibliography
1. Madden, S., et al., TAG: A Tiny AGgregation Service for Ad-Hoc Sensor
Networks. 2002: OSDI.
2. Intanagonwiwat, C., R. Govindan, and D. Estrin, Directed Diffusion: A Scalable
and Robust Communication Paradigm for Sensor Networks. 2000: Mobile
Computing and Networking.
3. Perkins, C., Ad-hoc on-demand distance vector routing. MILCOM, 1997.
4. Berkeley, University of California, 800 node self-organized wireless sensor
network. 2001: http://today.cs.berkeley.edu/800demo/.
5. Doherty, L., Algorithms for Position and Data Recovery in Wireless Sensor
Networks. UC Berkeley EECS Masters Report, 2000.
6. Mclurkin, J., Algorithms for distributed sensor networks. 1999: Masters Thesis
for Electrical Engineering at the University of California, Berkeley.
7. McMahan, M.L., Evolving Cellular Handset Architectures but a Continuing,
Insatiable Desire for DSP MIPS. 2000, Texas Instruments Incorporated.
8. Cerpa, A., et al., Habitat monitoring: Application driver for wireless
communications technology. ACM SIGCOMM Workshop on Data
Communications in Latin America and the Caribbean, 2001.
9. Mainwaring, A., et al., Wireless Sensor Networks for Habitat Monitoring, in
ACM International Workshop on Wireless Sensor Networks and Applications
(WSNA'02). 2002.
10. Woo, A. and D. Culler, Evaluation of Efficient Link Reliability Estimators for
Low-Power Wireless Networks. 2002: Technical Report, UC Berkeley.
11. Xu, Y., J. Heidemann, and D. Estrin, Geography-informed energy conservation
for Ad Hoc routing. 2001, ACM Press: SIGMOBILE : ACM Special Interest
Group on Mobility of Systems, Users, Data and Computing. p. 70 - 84.
12. Yarvis, M.D., et al., Real-World Experiences with an Interactive Ad Hoc Sensor
Network. 2002: International Conference on Parallel Processing Workshops.
13. Whitehouse, K., The design of calamari: an ad-hoc localization system for sensor
networks. 2003: Masters Report, University of California at Berkeley.
14. Kymissis, J., et al., Parasitic Power Harvesting in Shoes. ISWC, 1998.
15. Roundy, S., P. Wright, and J. Rabaey, A Study of Low Level Vibrations as a
Power Source for Wireless Sensor Nodes. 2002: Computer Communications.
16. Perrig, A., et al., SPINS: Security protocols for sensor networks. Proceedings of
MOBICOM, 2001, 2002.
17. Rivest, R., The RC5 Encryption Algorithm. 1994: Fast Software Encryption. p.
86-96.
18. Energizer Battery Company, Energizer AA, engineering datasheet:
http://data.energizer.com/datasheets/library/primary/alkaline/energizer/consumer_
oem/e91.pdf.
183
19. Energizer Battery Company, Energizer CR2032, engineering datasheet:
http://data.energizer.com/datasheets/library/primary/lithium_coin/cr2032.pdf.
20. Power Stream, NiMH Battery Charging Basics:
http://www.powerstream.com/NiMH.htm.
21. McLarnon, B., VHF/UHF/Microwave Radio Propagation:
A Primer for Digital Experimenters:
http://www.tapr.org/tapr/html/ve3jf.dcc97/ve3jf.dcc97.html.
22. RF Monolithics Inc, TR1000 Data Sheet. 1999:
http://www.rfm.com/products/data/tr1000.pdf.
23. Chipcon, CC1000 Single Chip Very Low Power RF Transceiver. 2002:
http://www.chipcon.com/files/CC1000_Data_Sheet_2_1.pdf.
24. Texas Instruments, MSP430X13x, MSP430x14x Mixed Signal Microcontroller
User Guide. 2003: http://www-s.ti.com/sc/ds/msp430f149.pdf.
25. AMD, AM49DL640BG Stacked Multi-Chip Package (MCP) Flash Memory and
SRAM. 2003: http://www.amd.com/usen/
assets/content_type/white_papers_and_tech_docs/26090a.pdf.
26. Intel Corp, Intel Press Release: Intel Builds World's First One Square Micron
SRAM Cell. 2002:
http://www.intel.com/pressroom/archive/releases/20020312tech.htm.
27. Atmel Corporation, AT90S/LS8535 Datasheet. 2001:
http://www.atmel.com/atmel/acrobat/doc1041.pdf.
28. Philips Semiconductor, The I2C-Bus Specification. 2000:
http://www.semiconductors.philips.com/acrobat/various/I2C_BUS_SPECIFICAT
ION_3.pdf.
29. Atmel Corporation, AT90S2323/LS2323/S2343/LS2343 Datasheet. 2001:
http://www.atmel.com/atmel/acrobat/doc1004.pdf.
30. Culler, D.E., J. Singh, and A. Gupta, Parallel Computer architecture a
hardware/software approach. 1999.
31. Esser, R. and R. Knecht, Intel Paragon XP/S - architecture and software
environment. 1993: Technical Report KFA-ZAM-IB-9305.
32. Culler, D.E., et al. Fine-grain parallelism with minimal hardware support: a
compiler-controlled threaded abstract machine. 1991.
33. Blumofe, R., et al., Cilk: An Efficient Multithreaded Runtime System.
Proceedings of the 5th Symposium on Principles and Practice of Parallel
Programming, 1995.
34. Hu, J., I. Pyarali, and D. Schmidt, Measuring the Impact of Event Dispatching and
Concurrency Models on Web Server Performance Over High-speed Networks.
Proceedings of the 2 nd Global Internet Conference, IEEE, 1997.
35. von Eicken, T., et al. Active messages: a mechanism for integrated
communication and computation. in 19th Annual International Symposium on
Computer Architecture. 1992.
36. Gay, D., et al., The nesC Language: A Holistic Approach to Networked
Embedded Systems. 2003: Programming Language Design and Implementation
(PLDI).
37. Renesse, R.V., et al., A framework for protocol composition in horus. 1995:
Proceedings of the ACM Symposium on Principles of Distributed Computing.
184
38. Agarwal, A., et al., The MIT alewife machine: A large-scale distributed-memory
multi-processor. 1991: Proceedings of Workshop on Scalable Shared Memory
Multiprocessors.
39. Montz, A.B., et al., Scout: A communications-oriented operating system. 1995:
HOT OS.
40. Hutchinson, N.C. and L.L. Peterson, The x-kernel: An architecture for
implementing network protocols. 1991: IEEE Transactions on Software
Engineering. p. 17(1):64-76.
41. Want, R., et al., The Active Badge Location System. ORL, 24a Trumpington
Street, Cambridge CB2 1QA, 1992.
42. Bahl, P. and V. Padmanabhan, RADAR: An in-building RF-based user location
and tracking system. IEEE Infocom, 2000. 2: p. 775-784.
43. Chandrakasan, A.P., S. Sheng, and R.W. Brodersen, Low Power CMOS Digital
Design. 1992, University of California Berkeley.
44. Pering, T., T. Burd, and R. Brodersen, The Simulation and Evaluation of
Dynamic Voltage Scaling Algorithms. 1998.
45. Hill, J., et al. System architecture directions for networked sensors. in ASPLOS
2000. 2000.
46. Atmel Corporation, Atmega103(L) Datasheet. 2001, Atmel Corporation:
http://www.atmel.com/atmel/acrobat/doc0945.pdf.
47. Dallas Semiconductor, DS2401 Silicon Serial Number: http://pdfserv.maximic.
com/arpdf/DS2401.pdf.
48. Atmel Corporation, AT45DB041B Datasheet. 2001:
http://www.atmel.com/atmel/acrobat/doc1938.pdf.
49. Maxim, Maxim 1-Cell to 2-Cell, Low-Noise, High-Efficiency, Step-Up DC-DC
Converter, MAXIM1678. 1998: http://pdfserv.maximic.
com/arpdf/MAX1678.pdf.
50. Elson, J. and D. Estrin, Time Synchronization for Wireless Sensor Networks:
http://www.circlemud.org/~jelson/writings/timesync/timesync.html.
51. Ye, W., J. Heidemenn, and D. Estrin, An Energy-Efficient MAC Protocol for
Wireless Sensor Networks. 2001: Submitted for review, July 2001.
52. Stemm, M., et al. Reducing power consumption of network interfaces in handheld
devices. in International Workshop on Mobile Multimedia Communications
(MoMuc-3). 1996. Princeton, NJ.
53. Lamport, L., Time, clocks, and the ordering of events in a distributed system.
Comm., 1978. ACM 21(7): p. 558-565.
54. Mills, D.L., Internet time synchronization: the Network Time Protocol. IEEE
Trans. Communications, 1991. COM-39(10): p. 1482-1493.
55. Doherty, L., K.S.J. Pister, and L.E. Ghaoui. Convex position estimation in
wireless sensor networks. in IEEE Infocom. 2001: IEEE Computer Society Press.
56. Borriello, J.H.a.G., Location Systems of Ubiuitous Computing. Computer, 2001.
34(8): p. 57-66.
57. Priyantha, N.B., A. Chakraborty, and H. Balakrishnan. The Circket Location-
Support System. in MobiCom 2000. 2000. Boston, Massachusetts.
58. Molnar, A., Personal Communication. 2004: To be submitted to IEEE
International Solid-State Circuits Conference 2004.
185
59. Scott, M.D., An Ultra-Low Power ADC for Distributed Sensor Networks. 2002:
Master of Science Theses, UC Berkeley.
60. Dallas Semiconductor, DS1804 NV Trimmer Potentiometer:
http://pdfserv.maxim-ic.com/arpdf/DS1804.pdf.
61. DARPA, Network Embedded Software Technology:
http://www.darpa.mil/ipto/Solicitations/CBD_02-12.html.
62. Bergbreiter, S. and K. Pister, COTS-BOTS: Specifications. 2002: http://wwwbsac.
eecs.berkeley.edu/~sbergbre/CotsBots/specs.pdf.
63. Analog Devices, ADXL202/ADXL210: Low Cost ±2 g/±10 g Dual Axis
iMEMS® Accelerometers with Digital Output Data:
http://www.analog.com/UploadedFiles/Datasheets/70885338ADXL202_10_b.pdf
.
64. Hill, J., System Architecture for Wireless Sensor Networks. 2000: Masters
Report, University of California Berkeley.
65. Madden, S., et al., TinyDB web page. 2002:
http://www.telegraph.cs.berkeley.edu/tinydb.
66. Atwood, B., B. Waraneke, and K.S.J. Pister. Preliminary circuits for smart dust. in
Southwest Symposioum on Mixed-Signal Design. 2000. San Diego, Ca.
67. Pister, K.S.J., J.M. Kahn, and B.E. Boser, Smart Dust: Wireless Networks of
Millimeter-Scale Sensor Nodes. Electronics Research Laboratory Research
Summary, 1999.
68. Chu, P.B., et al., Optical communication link using micro-macined corner cube
reflectors. 1997: Proceeding of the SPIE vol. 3008-20.
69. The Official Bluetooth Website: http://www.bluetooth.com/.
70. ZigBee Alliance, The official website of the Zigbee Alliance:
http://www.zigbee.org.
71. IEEE, IEEE 802.15 WPAN™ Task Group 4: http://ieee802.org/15/pub/TG4.html.
72. Semiconductor, A., AMI Semiconductor First to Demonstrate Working ZigBee
Wireless Device: http://www.amis.com/news/030414_zigbee.cfm.
73. J.L. Da Silva Jr., J.S., M. J. Ammer, C. Guo, S. Li, R. Shah, T. Tuan, M. Sheets,
J.M. Ragaey, B. Nikolic, A Sangiovanni-Vincentelli, P. Wright., Design
Methodology for Pico Radio Networks. 2001, Berkeley Wireless Research
Center.
74. Pottie, G. and W. Kaiser, Wireless Integrated Network Sensors (WINS):
Principles and Approach. Communications of the ACM, 2000. 43.
75. Windriver Systems, VxWorks 5.4 Datasheet:
http://www.windriver.com/products/html/vxwks55_ds.html.
76. Microsoft Corp., Microsoft Windows CE:
http://www.microsoft.com/windowsce/embedded.
77. PalmOS Software 3.5 Overview:
http://www.palm.com/devzone/docs/palmos35.html.
78. Hildebrand, D., An architectural Overview of QNX:
http://www.qnx.com/literature/whitepapers/archoverview.html.
79. Windriver Systems, pSOSystem Datasheet:
http://www.windriver.com/products/html/psosystem_ds.html.
186
80. QNX Software Systems Ltd., QNX Neutrino Realtime OS:
http://www.qnx.com/products/os/neutrino.html.
81. Microware, Microware OS-9:
http://www.microware.com/ProductsServices/Technologies/os-91.html.
82. IEEE Std. 1003.0-1995: IEEE Guide to the POSIX® Open System Environment.
1995, IEEE: http://standards.ieee.org/catalog/olis/arch_posix.html.
83. Kauler, B., CREEM: Concurrent Realtime Embedded Executive for
Microcontrollers: http://members.dodo.net.au/~void/old/creem.htm.
84. pOSEK: A super-small scalable real-time operating system for high-volume,
deeply embedded applications.: http://www.isi.com/products/posek/index.htm.
85. Microware, Microware Ariel Technical Overview:
http://www.microware.com/ProductsServices/Technologies/ariel_technology_brie
f.html.
86. Liu, L.T. and D.E. Culler, Measurements of active messages perfromance on the
CM-5. 94: Technical Report UCP//CSD-94-087, University of California at
Berkeley, Department of Computer Science.
87. Druschel, P., M. Aron, and W. Zwaenepoel, A scalable and explicit event delivery
mechanism of UNIX. 1999: Proceeding of the USENIX 1999 Annual Technical
Conference.
88. Homewood, M. and M. McLaren. Meiko CS-2 interconnect. Elan-Elite design.
1993.
89. Muller, M., Zero-Copy TCP/IP with Gigabit Ethernet. 1999: PhD thesis, Institute
for Computer Systems, ETH Zurich.
90. Cooper, G.H., Tiny TCP/IP:
91. Seiko Instruments, S-7600A Series iChip TCP/IP Protocol: http://www.seiko-usaecd.
com/intcir/products/rtc_assp/s7600a.html.