0% found this document useful (0 votes)

42 views

Programming Memory - Constrained Networked Embedded Systems

This document provides a summary of Adam Dunkels' 2007 doctoral thesis on programming memory-constrained networked embedded systems. The thesis focused on three topics: 1) using the TCP/IP protocol suite in memory-constrained networked embedded systems; 2) simplifying event-driven programming for such systems; and 3) dynamic loading of program modules in an operating system for memory-constrained devices. Dunkels showed that TCP/IP can be used in small embedded systems, developed a novel programming mechanism called protothreads to simplify event-driven programming, and implemented dynamic linking of ELF files in the Contiki operating system. The software developed has been widely used in industry and academia.

Uploaded by

amina.kodecta

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

42 views

Programming Memory - Constrained Networked Embedded Systems

Uploaded by

amina.kodecta

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 222

Swedish Institute of Computer Science

Doctoral Thesis
SICS Dissertation Series 47

Programming
Memory-Constrained
Networked Embedded Systems
Adam Dunkels

February 2007

Swedish Institute of Computer Science

Stockholm, Sweden
Copyright c Adam Dunkels, 2007
ISRN SICS-D–47–SE
SICS Dissertation Series 47
ISSN 1101-1335
Printed by Arkitektkopia, Västerås, Sweden
3
Abstract

Ten years after the Internet revolution are we standing on the brink of another
revolution: networked embedded systems that connect the physical world with
the computers, enabling new applications ranging from environmental moni-
toring and wildlife tracking to improvements in health care and medicine.
Only 2% of all microprocessors that are sold today are used in PCs; the re-
maining 98% of all microprocessors are used in embedded systems. The micro-
processors used in embedded systems have much smaller amounts of memory
than PC computers. An embedded system may have as little has a few hundred
bytes of memory, which is thousands of millions times less than the memory in
a modern PC. The memory constraints make programming embedded systems
a challenge.
This thesis focus on three topics pertaining to programming memory-
constrained networked embedded systems: the use of the TCP/IP protocol suite
even in memory-constrained networked embedded systems; simplifying event-
driven programming of memory-constrained systems; and dynamic loading of
program modules in an operating system for memory-constrained devices. I
show that the TCP/IP protocol stack can, contrary to previous belief, be used
in memory-constrained embedded systems but that a small implementation has
a lower network throughput. I present a novel programming mechanism called
protothreads that is intended to replace state machine-based event-driven pro-
grams. Protothreads provide a conditional blocked wait mechanism on top of
event-driven systems with a much smaller memory overhead than full multi-
threading; each protothread requires only two bytes of memory. I show that
protothreads significantly reduce the complexity of event-driven programming
for memory-constrained systems. Of seven state machine-based programs
rewritten with protothreads, almost all explicit states and state transitions could
be removed. Protothreads also reduced the number of lines of code with 31%
on the average. The execution time overhead of protothreads is on the order

i
ii

of a few processor cycles which is small enough to make protothreads usable

even in time-critical programs. Finally, I show that dynamic linking of native
code in standard ELF object code format is doable and feasible for wireless
sensor networks by implementing a dynamic loading and linking mechanism
for my Contiki operating system. I measure and quantify the energy consump-
tion of the dynamic linker and compare the energy consumption of native code
with that of virtual machine code for two virtual machines, including the Java
virtual machine. The results show that the energy overhead of dynamic linking
of ELF files mainly is due to the ELF file format and not due to the dynamic
linking mechanism as such. The results also suggest that combinations of na-
tive code and virtual machine code are more energy efficient than pure native
code or pure virtual machine code.
The impact of the research in this thesis has been and continues to be large.
The software I have developed as part of this thesis, lwIP, uIP, protothreads, and
Contiki, is currently used by hundreds of companies in embedded devices in
such diverse systems as car engines, oil boring equipment, satellites, and con-
tainer security systems. The software is also used both in academic research
projects and in university project courses on embedded systems throughout the
world. Articles have been written, by others, in professional embedded soft-
ware developer magazines about the software developed as part of this thesis.
The papers in this thesis are used as required reading in advanced university
courses on networked embedded systems and wireless sensor networks.
Sammanfattning

Tio år efter Internet-revolutionen står vi nu inför nästa revolution: kommu-

nicerande inbyggda system som kopplas ihop med varandra och därigenom
möjliggör helt nya tillämpningar inom ett stort antal områden, bland annat
sjukvård, miljöövervakning och energimätning.
Endast 2% av alla mikroprocessorer som säljs idag används för att bygga
PC-datorer; resterande 98% går till inbyggda system. Det stora flertalet av
dessa system har avsevärt mycket mindre minne än en modern PC. Inbygga
system har ofta endast ett par hundra bytes minne, att jämföra med de tusen-
tals miljoner bytes minne en modern PC har. Minnesbegränsningarna hos de
inbyggda systemen gör dem till en utmaning att programmera.
Avhandlingen behandlar programmering av minnesbegränsade kommu-
nicerande inbyggda system ur tre synvinklar: möjligheten för mycket
små minnesbegränsade inbyggda system att kommunicera med hjälp av
Internet-protokollen; förenkling av händelsestyrd programmering för minnes-
begränsade system; och dynamisk programladdning i ett operativsystem för
kommunicerande inbyggda system.
För att ett inbyggt system ska kunna kommnunicera i ett nätverk krävs att
systemet kan prata nätverkets språk, nätverksprotokollet. Avhandlingen visar
att det är möjligt även för mycket små system att använda Internet-protokollen,
TCP/IP, utan att behöva göra avsteg från gällande Internet-standarder. Dock
innebär minnesbegränsningarna en avsevärd prestandaminskning. Min pro-
gramvara lwIP och uIP visar att det är möjligt att koppla ihop mycket enkla
inbyggda system, väsentligt mycket mindre än man tidigare trott, med nätverk
som använder TCP/IP-protokollen.
Många program för minnesbegränsade inbyggda system bygger på en
programmeringsmetod som kallas händelsestyrd programmering. Med
händelsestyrd programmering kan man skriva program som kräver mycket lite
minne, men programmen blir ofta svåra att både utveckla och underhålla. För

iii
iv

att underlätta programmeringen av sådana system har jag utvecklat en pro-

grammeringsteknik som jag kallar protothreads, prototrådar. Protothreads gör
händelsestyrda program mindre komplexa utan att nämnvärt öka minnesutnyt-
tjandet och med mycket små prestandaförluster.
Operativsystemet Contiki, som jag har utvecklat under avhandlingsarbetet,
kan under drift ladda nya programmoduler, något som inte stöds av andra oper-
ativsystem för små inbyggda system. Att kunna ladda programvara under drift
underlättar både utveckling av ny programvara och korrigering av felaktig pro-
gramvara. Contiki visar att det trots resursbegränsningarna är möjligt att ladda
programmoduler i standardformatet ELF. Jag kvantifierar energiåtgången både
för att ladda program med dynamisk länkning och för att exekvera de laddade
programmen, samt jämför energiåtgången med den för motsvarande program
skrivet för två virtuell maskiner, bland annat en Java-maskin.
Programvaran har fått mycket stor spridning och används idag av hun-
dratals företag i ett stort antal produkter såsom bilmotorer, satellitsystem, olje-
borrar, TV-utrustning och låssystem från företag såsom BMW, NASA och HP.
Programvaran används i projektkurser på universitetsnivå världen över. Ar-
tiklar i branchtidskrifter har skrivits, av utomstående, om hur man anpassar
programvaran för nya mikroprocessorer. Ledande experter inom programvaru-
utveckling för inbyggda system har ett flertal gånger rekommenderat program-
varan i nyhetsbrev. Både forskningsartiklarna i avhandlingen och program-
varan används i undervisningen vid universitet världen över.
To Castor, Morgan, Maria, and the little one we have not yet met
Preface

I have always loved programming. When I was a kid, my father sometimes

brought home a computer that he used in teaching computer programming to
mathematic teacher students at the university. The computer was an ABC80,
a Swedish Z80-based computer with a BASIC interpreter and 16 kilobytes of
RAM. I learned programming BASIC from modifying my father’s programs
and by typing in BASIC programs from the manual. The programs were very
small and I was never limited by the small amount of memory. A few years
later I got my first own computer, a Commodore 64 with 64 kilobytes of RAM.
I was so eager to start programming that I had already learned programming
the assembly language of its 6510 processor by reading books on the subject
before I got the actual computer. Over the years, I wrote a large number of
assembly language programs for it and frequently felt that its memory was a
limitation. Some six or seven years later I bought my first PC, with a 486
microprocessor and 16 megabytes of RAM. I quickly learned x86 assembly
language but never came anywhere near writing a program that used the entire
memory of the machine.
In 2000 I did my master’s thesis at the Swedish Institute of Computer Sci-
ence in Kista, Sweden. As part of my thesis I developed a TCP/IP stack,
which I named lwIP, for transmitting vital statistics from wireless sensors on
ice hockey-players to people in the audience with laptop computers. The wire-
less sensors were equipped with Mitsubishi M16c CPUs with 20 kilobytes of
RAM and 100 kilobytes of ROM. I was almost back where I started!
While this thesis officially was done over the past four years at the Swedish
Institute of Computer Science, the real work started almost 20 years earlier.

Adam Dunkels
Stockholm, January 8 2007

vii
Acknowledgements

I first and foremost thank my colleague Thiemo Voigt, who has also been the
co-adviser for this thesis, for all the moral support over the past few years, for
being the committed person that he is, and for being genuinely fun to work
with. Working with this thesis would have been considerably less enjoyable if
it had not been for Thiemo. Thiemo and I have been working during the final
hours before paper submission deadlines, sometimes as late/early as 6 AM in
the morning. Thiemo has also gone out of his way to take care of distracting
duties, thus allowing me to focus on doing the research for this thesis.
I am grateful to Mats Björkman, my university adviser for this thesis, for
being a stimulating person and for the smooth PhD process. Seemingly big
problems have always turned into non-problems after a discussion with Mats.
I am also very grateful to the inspiring Juan Alonso. Juan started the
DTN/SN project at SICS within which most of the work in this thesis was
done. I also would like to thank Henrik Abrahamsson for being a good friend
and stimulating discussion partner on subjects ranging from the craft of sci-
ence and research to cars and culinary culture. I am also very happy to work
with the great members of our Networked Embedded Systems Group at SICS:
Joakim Eriksson, Niclas Finne, and Fredrik Österlind. An equally skilled and
dedicated group of people is very hard to find. Many thanks also to Björn
Grönvall for taking a lot of the work of writing project deliverables as well
as porting Contiki to new platforms. Thanks also to Sverker Janson, labora-
tory manager of the Intelligent Systems Laboratory at SICS, for his inspiring
leadership and for his support. Many thanks to all the people at SICS for creat-
ing a stimulating work environment; Lars Albertsson, Frej Drejhammar, Karl-
Filip Faxén, Anders Gunnar, Ali Ghodsi, Kersti Hedman, Janusz Launberg,
Ian Marsh, Mikael Nehlsen, Martin Nilsson, L-H Orc Lönn, Tony Nordström,
Carlo Pompili, Babak Sadighi, and Karl-Petter Åkesson, just to name a few.
Many thanks to Oliver Schmidt for our cooperation on protothreads and his

ix
x

porting and maintaining of Contiki, for always being a very sharp discussion
partner, and for being a good person to work with.
Thanks also to the great master thesis students with whom I have been in-
volved during this work: Max Loubser, Shujuan Chen, Zhitao He, and Nicolas
Tsiftes. Thanks also to Muneeb Ali for his fruitful research visit at SICS.
My thanks also go out to the hundreds of people I have been in contact with
regarding my software over the past few years. I have gotten many warming
words, good comments on my software, bugfixes and patches, as well as new
modules and ports to new architectures. I have gotten so many e-mails that I
unfortunately have only been able to answer a fraction of them.
I am also deeply grateful to the people at Luleå University of Technology
for teaching me the basic aspects of computer science. Lars-Gunnar Taube for
introducing me to the secrets of computing many years ago; Håkan Jonsson for
his introduction to the interesting world of functional programming; Leif Ku-
soffsky for his imperative programming laboratory assignments that taught me
how to write virtual machines and how to develop compilers for object-oriented
languages; Lennart Andersson for giving me the extremely important insight
that external and internal data representation need not be the same, when we
were instructed to not use a two-dimensional array to represent the spreadsheet
data in the VisiCalc-clone we developed as a laboratory assignment; Mikael
Degermark and Lars-Åke Larzon for sparkling my interest in computer com-
munications; and Olov Schelén and Mathias Engan for teaching me how to
read and review scientific papers.
Thanks also go to my mother Kerstin for being supportive throughout my
education and research career, for taking interest in my research work, and for
reading and commenting on this thesis. I will also forever be in debt to my late
father Andrejs, who taught me the skills of everything from living and laughing
to mathematics and music.
Finally, I am extremely fortunate to have been blessed with such a loving
family: my wife Maria, our sons Morgan and Castor, and one who we look
forward to meet in a few months from now. Maria has supported me throughout
the work with this thesis, taken interest in my work, listened to and helped
improve my research presentations, and endured all my research ramblings at
home.

This work in this thesis is in part supported by VINNOVA, Ericsson, SITI, SSF,
the European Commission under the Information Society Technology priority
within the 6th Framework Programme, the European Commission’s 6th Frame-
xi

work Programme under contract number IST-004536, and the Swedish Energy
Agency. Special thanks to Bo Dahlbom and the Swedish Energy Agency for
funding the final writing up of this thesis. The Swedish Institute of Computer
Science is sponsored by TeliaSonera, Ericsson, SaabTech, FMV, Green Cargo,
ABB, and Bombardier Transportation AB.
Included Papers

This thesis consists of a thesis summary and five papers that are all published
in peer-reviewed conference and workshop proceedings. Throughout the thesis
summary the papers are referred to as Paper A, B, C, D, and E.
Paper A Adam Dunkels. Full TCP/IP for 8-bit architectures. In Proceedings
of The First International Conference on Mobile Systems, Applications,
and Services (ACM MobiSys 2003), San Francisco, USA, May 2003.
Paper B Adam Dunkels, Björn Grönvall, and Thiemo Voigt. Contiki - a
Lightweight and Flexible Operating System for Tiny Networked Sen-
sors. In Proceedings of the First IEEE Workshop on Embedded Net-
worked Sensors (IEEE Emnets 2004), Tampa, Florida, USA, November
2004.
Paper C Adam Dunkels, Oliver Schmidt, and Thiemo Voigt. Using pro-
tothreads for sensor node programming. In Proceedings of the Workshop
on Real-World Wireless Sensor Networks (REALWSN 2005), Stockholm,
Sweden, June 2005.
Paper D Adam Dunkels, Oliver Schmidt, Thiemo Voigt, and Muneeb Ali.
Protothreads: Simplifying event-driven programming of memory-
constrained embedded systems. In Proceedings of the 4th Interna-
tional Conference on Embedded Networked Sensor Systems (ACM Sen-
Sys 2006), Boulder, Colorado, USA, November 2006.
Paper E Adam Dunkels, Niclas Finne, Joakim Eriksson, and Thiemo Voigt.
Run-time dynamic linking for reprogramming wireless sensor networks.
In Proceedings of the 4th International Conference on Embedded Net-
worked Sensor Systems (ACM SenSys 2006), Boulder, Colorado, USA,
November 2006.

xiii
Contents

I Thesis Summary 1
1 Introduction 3
1.1 Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . 4
1.2 Programming Memory-Constrained Embedded Systems . . . 5
1.3 Research Approach and Method . . . . . . . . . . . . . . . . 5
1.4 Research Issues . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.1 TCP/IP for Memory-Constrained Systems . . . . . . . 7
1.4.2 Protothreads and Event-Driven Programming . . . . . 8
1.4.3 Dynamic Module Loading . . . . . . . . . . . . . . . 8
1.5 Thesis Structure . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Scientific Contributions and Impact 11

2.1 Scientific Contributions . . . . . . . . . . . . . . . . . . . . . 11
2.2 Impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3 Summary of the Papers and Their Contributions 15

3.1 Paper A: Full TCP/IP for 8-Bit Architectures . . . . . . . . . 16
3.2 Paper B: Contiki - a Lightweight and Flexible Operating Sys-
tem for Tiny Networked Sensors . . . . . . . . . . . . . . . . 17
3.3 Paper C: Using Protothreads for Sensor Node Programming . 18
3.4 Paper D: Protothreads: Simplifying Event-Driven Program-
ming of Memory-Constrained Embedded Systems . . . . . . . 19
3.5 Paper E: Run-time Dynamic Linking for Reprogramming
Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . 19

4 Related Work 21
4.1 Small TCP/IP Implementations . . . . . . . . . . . . . . . . . 21

xv
xvi Contents

4.2 Operating Systems for Wireless Sensor Networks . . . . . . . 22

4.3 Programming Models for Wireless Sensor Networks . . . . . 24
4.3.1 Macro-programming and New Programming Languages 24
4.3.2 Virtual Machines . . . . . . . . . . . . . . . . . . . . 25

5 Conclusions and Future Work 27

5.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

6 Other Software and Publications 31

6.1 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6.2 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

Bibliography 35

II Papers 43
7 Paper A:
Full TCP/IP for 8-Bit Architectures 45
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
7.2 TCP/IP overview . . . . . . . . . . . . . . . . . . . . . . . . 48
7.3 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . 50
7.4 RFC-compliance . . . . . . . . . . . . . . . . . . . . . . . . 52
7.5 Memory and buffer management . . . . . . . . . . . . . . . . 53
7.6 Application program interface . . . . . . . . . . . . . . . . . 55
7.7 Protocol implementations . . . . . . . . . . . . . . . . . . . . 56
7.7.1 Main control loop . . . . . . . . . . . . . . . . . . . . 56
7.7.2 IP — Internet Protocol . . . . . . . . . . . . . . . . . 57
7.7.3 ICMP — Internet Control Message Protocol . . . . . . 58
7.7.4 TCP — Transmission Control Protocol . . . . . . . . 59
7.8 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.8.1 Performance limits . . . . . . . . . . . . . . . . . . . 62
7.8.2 The impact of delayed acknowledgments . . . . . . . 63
7.8.3 Measurements . . . . . . . . . . . . . . . . . . . . . 64
7.8.4 Code size . . . . . . . . . . . . . . . . . . . . . . . . 66
7.9 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.10 Summary and conclusions . . . . . . . . . . . . . . . . . . . 70
7.11 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . 70
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Contents xvii

8 Paper B:
Contiki - a Lightweight and Flexible Operating System for Tiny
Networked Sensors 75
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
8.1.1 Downloading code at run-time . . . . . . . . . . . . . 78
8.1.2 Portability . . . . . . . . . . . . . . . . . . . . . . . . 78
8.1.3 Event-driven systems . . . . . . . . . . . . . . . . . . 79
8.2 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . 80
8.3 System overview . . . . . . . . . . . . . . . . . . . . . . . . 81
8.4 Kernel architecture . . . . . . . . . . . . . . . . . . . . . . . 82
8.4.1 Two level scheduling hierarchy . . . . . . . . . . . . 83
8.4.2 Loadable programs . . . . . . . . . . . . . . . . . . . 83
8.4.3 Power save mode . . . . . . . . . . . . . . . . . . . . 84
8.5 Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
8.5.1 Service replacement . . . . . . . . . . . . . . . . . . 85
8.6 Libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
8.7 Communication support . . . . . . . . . . . . . . . . . . . . . 87
8.8 Preemptive multi-threading . . . . . . . . . . . . . . . . . . . 88
8.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
8.9.1 Over-the-air programming . . . . . . . . . . . . . . . 89
8.9.2 Code size . . . . . . . . . . . . . . . . . . . . . . . . 89
8.9.3 Preemption . . . . . . . . . . . . . . . . . . . . . . . 91
8.9.4 Portability . . . . . . . . . . . . . . . . . . . . . . . . 92
8.10 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

9 Paper C:
Using Protothreads for Sensor Node Programming 97
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
9.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
9.3 Protothreads . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
9.3.1 Protothreads versus events . . . . . . . . . . . . . . . 102
9.3.2 Protothreads versus threads . . . . . . . . . . . . . . . 103
9.3.3 Comparison . . . . . . . . . . . . . . . . . . . . . . . 103
9.3.4 Limitations . . . . . . . . . . . . . . . . . . . . . . . 105
9.3.5 Implementation . . . . . . . . . . . . . . . . . . . . . 106
9.4 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . 107
9.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
xviii Contents

10 Paper D:
Protothreads: Simplifying Event-Driven Programming of
Memory-Constrained Embedded Systems 111
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
10.2 Protothreads . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
10.2.1 Scheduling . . . . . . . . . . . . . . . . . . . . . . . 116
10.2.2 Protothreads as Blocking Event Handlers . . . . . . . 117
10.2.3 Example: Hypothetical MAC Protocol . . . . . . . . . 117
10.2.4 Yielding Protothreads . . . . . . . . . . . . . . . . . 120
10.2.5 Hierarchical Protothreads . . . . . . . . . . . . . . . . 120
10.2.6 Local Continuations . . . . . . . . . . . . . . . . . . 121
10.3 Memory Requirements . . . . . . . . . . . . . . . . . . . . . 122
10.4 Replacing State Machines with
Protothreads . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
10.5 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . 125
10.5.1 Prototype C Preprocessor
Implementations . . . . . . . . . . . . . . . . . . . . 126
10.5.2 Memory Overhead . . . . . . . . . . . . . . . . . . . 130
10.5.3 Limitations of the Prototype Implementations . . . . . 130
10.5.4 Alternative Approaches . . . . . . . . . . . . . . . . . 131
10.6 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
10.6.1 Code Complexity Reduction . . . . . . . . . . . . . . 133
10.6.2 Memory Overhead . . . . . . . . . . . . . . . . . . . 140
10.6.3 Run-time Overhead . . . . . . . . . . . . . . . . . . . 141
10.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
10.8 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . 143
10.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

11 Paper E: Run-time Dynamic Linking for Reprogramming Wireless

Sensor Networks 151
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
11.2 Scenarios for Software Updates . . . . . . . . . . . . . . . . . 154
11.2.1 Software Development . . . . . . . . . . . . . . . . . 154
11.2.2 Sensor Network Testbeds . . . . . . . . . . . . . . . . 154
11.2.3 Correction of Software Bugs . . . . . . . . . . . . . . 155
11.2.4 Application Reconfiguration . . . . . . . . . . . . . . 155
11.2.5 Dynamic Applications . . . . . . . . . . . . . . . . . 155
11.2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . 156
Contents xix

11.3 Code Execution Models and Reprogramming . . . . . . . . . 156

11.3.1 Script Languages . . . . . . . . . . . . . . . . . . . . 157
11.3.2 Virtual Machines . . . . . . . . . . . . . . . . . . . . 157
11.3.3 Native Code . . . . . . . . . . . . . . . . . . . . . . . 157
11.4 Loadable Modules . . . . . . . . . . . . . . . . . . . . . . . 158
11.4.1 Pre-linked Modules . . . . . . . . . . . . . . . . . . . 160
11.4.2 Dynamic Linking . . . . . . . . . . . . . . . . . . . . 161
11.4.3 Position Independent Code . . . . . . . . . . . . . . . 162
11.5 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . 162
11.5.1 The Contiki Operating System . . . . . . . . . . . . . 163
11.5.2 The Symbol Table . . . . . . . . . . . . . . . . . . . 164
11.5.3 The Dynamic Linker . . . . . . . . . . . . . . . . . . 164
11.5.4 The Java Virtual Machine . . . . . . . . . . . . . . . 167
11.5.5 CVM - the Contiki Virtual Machine . . . . . . . . . . 167
11.6 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
11.6.1 Energy Consumption . . . . . . . . . . . . . . . . . . 170
11.6.2 Memory Consumption . . . . . . . . . . . . . . . . . 176
11.6.3 Execution Overhead . . . . . . . . . . . . . . . . . . 176
11.6.4 Quantitative Comparison . . . . . . . . . . . . . . . . 178
11.6.5 Scenario Suitability . . . . . . . . . . . . . . . . . . . 181
11.6.6 Portability . . . . . . . . . . . . . . . . . . . . . . . . 183
11.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
11.8 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . 184
11.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

SICS Dissertation Series 193

Thesis Summary

1
Chapter 1

Introduction

Twenty years ago the computer revolution put PCs in offices and homes
throughout large parts of the western world. Ten years later the Internet revolu-
tion connected the computers together in a world-spanning communication net-
work. Today, we stand on the brink of the next revolution: networked embed-
ded systems that connect the physical world together with computers, enabling
a large variety of applications such as health and heart resuscitation monitor-
ing [8, 69], wildlife tracking and volcano monitoring [35, 48, 70], building
structure monitoring [39], building automation [63], and carbon dioxide mon-
itoring in relation to global warming [41].
It is difficult to estimate the total number of embedded systems in the world
today, but it is possible to get a grasp of the magnitude of the area by looking
at sales figures for microprocessors. We might expect that PCs account for
the bulk of microprocessors because of their widespread use. However, PCs
account for only a very small part of the microprocessor market. In 2002, only
2% of all microprocessors sold were used in PCs [66]. The remaining 98% of
all microprocessors were sold for use in various types of embedded systems.
Embedded systems typically have much less memory than general-purpose
PCs. A normal PC sold in late 2006 had thousands of millions bytes of random
access memory (RAM). This is many million times larger than the RAM size
in many embedded systems; the microprocessors in many embedded systems
have as little as a few hundred or a few thousand bytes of RAM.
It is difficult to estimate the typical or average memory size in embedded
systems today, but again we can get a grasp of the magnitude by looking at
the microprocessor sales figures. Of the total number of microprocessors sold

3
4 Chapter 1. Introduction

in 2002, over 90% were significantly smaller in terms of memory size than a
modern PC [66]. In fact, over 50% of all microprocessors were so-called 8-
bit processors, which typically only can handle a maximum of 65536 bytes of
memory. Because of cost constraints, most of those microprocessors are likely
to have considerably less memory than the maximum amount. For example,
in 2004 the price of the popular Texas Instruments’ MSP430 FE423 micropro-
cessor was nearly 20% higher with 1024 bytes of RAM ($5.95) than the same
microprocessor with 256 bytes of RAM ($4.85) [31]. While the price of on-
chip memory is likely to decrease in the future, the number of applications for
microprocessors will increase. As these applications will require microproces-
sors with an even lower per-unit cost than today, future microprocessor models
are likely to have similar memory configurations as today’s microprocessors
but at lower per-unit prices.
Most microprocessors in embedded systems are programmed in the C pro-
gramming language [22]. Their memory constraints make programming them
a challenge. Programmers that program general-purpose computers or PCs sel-
dom have to take memory limitations into consideration because of the large
amounts of memory available. Moreover, PC microprocessors hardware make
techniques such as virtual memory possible, making the memory accessible
to the programmer almost limitless. In contrast, the small amounts of mem-
ory requires embedded systems programmers to always be aware of memory
limitations when programming their systems. Also, most microprocessors for
embedded systems do not have the ability to extend the physical memory with
virtual memory. In this thesis I use the term memory constrained for systems
where the programmer explicitly must take memory limitations into consider-
ation when programming the system.
Many embedded systems today communicate with each other. Examples
include base stations for mobile telephony, wireless car keys, point of sale
terminals, and data logging equipment in trucks. The embedded systems com-
municate both with other embedded systems and general-purpose computers
using a network. In this thesis I call such systems networked embedded sys-
tems. Networked embedded systems are, just like ordinary embedded systems,
often memory constrained.

1.1 Wireless Sensor Networks

A special kind of networked embedded systems are wireless sensor networks.
Wireless sensor networks consist of many small wireless networked embedded
1.2 Programming Memory-Constrained Embedded Systems 5

systems, equipped with sensors, that form a wireless network through which
sensor readings are transmitted [61]. Each sensor node is a networked embed-
ded system. Sensor data is relayed from sensor node to sensor node towards
a base station. If a node should break, sensor network routing protocols may
reroute the data around the broken node.
A wireless sensor network may consist of up to thousands of sensor nodes.
Because of the potential large scale of the sensor network the individual sensors
must be small, low cost, and expendable. For this reason, the sensor nodes used
in wireless sensor networks typically have memory-constrained microproces-
sors. Commercially available sensor nodes have between 2 and 10 kilobytes of
RAM [1, 59, 62]. Moreover, for sensor networks to run for extended periods
of time, the energy consumption of both individual sensor nodes and of the
network as a whole is of primary importance. Thus energy consumption is an
important performance metric for sensor networks.

1.2 Programming Memory-Constrained Embed-

ded Systems
This thesis is about programming memory-constrained networked embedded
systems, such as sensor nodes in wireless sensor networks. I use the word
programming in the sense of programming-in-the-small [18], not in the sense
of programming-in-the-large, and not in the sense of software engineering.
Programming-in-the-small is the process of connecting individual program
language statements to form program modules, whereas programming-in-the-
large is the process of connecting modules into programs or systems. Software
engineering is the process of gathering requirements, analyzing the require-
ments, designing the software system, and implementing it, possibly with a
large team of developers. Programming-in-the-small is an important part of
the software engineering process as it is what constitutes the final implementa-
tion phase. It is also in this phase that memory constraints are most evidently
visible to the system developers.

1.3 Research Approach and Method

Throughout the work with this thesis I have taken a pragmatic approach to pro-
gramming memory-constrained networked embedded systems: I have used the
C programming language, the most commonly used programming language
6 Chapter 1. Introduction

for embedded systems [22], and I have made it a point to make my software
work on a wide range of embedded systems platforms. An alternative approach
would have been to use uncommon programming languages, develop new pro-
gramming languages, or develop new hardware platforms. However, this prag-
matic approach has enabled me to interact with embedded systems develop-
ers working with actual embedded systems and products which would have
been very difficult if I had not used the C programming language. Interact-
ing with embedded systems developers has given me insights into many of the
actual problems that exist in embedded systems programming and has forced
me to build systems that actually work. Moreover, this pragmatic approach
also makes the research directly accessible to practitioners. This is one of the
reasons behind the large impact of this thesis.
The research method employed in this thesis has been the method of com-
puter systems research [16]: I have built computer systems and conducted ex-
periments with them in order to evaluate a specific method, tool, mechanism,
or implementation technique. The systems I have built are software systems:
two TCP/IP stacks for memory-constrained systems, lwIP and uIP, and an op-
erating system for memory-constrained systems, Contiki.
My research work has typically gone through two phases, one exploratory
phase and one confirmatory phase. In the exploratory phase I have been writing
computer programs, either as part of another research project or for personal
enjoyment. When programming I have come up with an interesting idea and
have become interested in finding out whether the idea is good or not. To test
the idea I have formulated an initial hypothesis to verify or falsify. The work
has then entered the confirmatory phase. In the confirmatory phase I test the
hypothesis that I have developed during the exploratory phase. For the purpose
of testing the hypothesis I have built a software system to carry out experi-
ments that either verify or falsify my hypothesis. I have then conducted the
experiments and evaluated the results. If the experiments have not supported
my hypothesis I have either revised the hypothesis and rebuilt my system, or
have abandoned the hypothesis and continued the exploratory phase.

1.4 Research Issues

In this thesis I focus on three topics pertaining to programming memory-
constrained networked embedded systems: the use of the TCP/IP protocol suite
for memory-constrained embedded systems; a novel programming abstraction
that I have named protothreads, which is intended to simplify event-driven
1.4 Research Issues 7

programming of memory-constrained systems; and dynamic loading of native

code modules in my Contiki operating system for memory-constrained system.
A general theme throughout this research is the use of standard or general-
purpose mechanisms. In academic research we are not restricted to standards.
Rather, we can freely choose to investigate our own protocols, programming
languages, file formats, and mechanisms. This is partly due to that we often
work in areas in which no standards have been created. However, even in
areas where standards exist or are emerging we often develop and use our own
protocols, programming languages, file formats, and mechanisms. Because of
this we cannot be sure why we choose our own solutions over the standard
solutions. Do we go our own way because we want to? Or are we compelled
to do it because of the overheads of the standard solutions?
Part of the research in this thesis is about answering the question of how
far we can push general-purpose or standard mechanisms before we need to
invent our own mechanisms. Or, conversely, how well does general-purpose
and standard mechanisms work for memory-constrained systems?

1.4.1 TCP/IP for Memory-Constrained Systems

All systems connected to the global Internet, wireless networks such as WLAN
and GPRS, and many local area networks communicate using the standard
TCP/IP protocol suite. Due to the prevalence of TCP/IP networks many net-
worked embedded systems are connected to such networks and therefore must
be able to communicate using the TCP/IP protocols. However, the TCP/IP pro-
tocol suite is often perceived to be “heavy-weight” in that an implementation of
the protocols requires large amounts of resources in terms of memory and pro-
cessing power. This perception can be corroborated by measuring the memory
requirements of popular TCP/IP implementations, such as the one in the Linux
kernel [28] or in the BSD operating system [52]. The TCP/IP implementations
in these systems require many hundreds of kilobytes of RAM and have a code
footprint of approximately hundred kilobytes.
The perception that the TCP/IP protocols would require large amounts of
memory leads to system designers equipping their embedded systems with
large microprocessors. If it was possible to implement the TCP/IP protocol
suite in radically less memory, system designers could choose smaller micro-
processors for their embedded systems. This would not only make the resulting
systems less costly to produce, but would also enable an entire class of smaller
embedded systems to communicate using the TCP/IP protocol suite. This is
the focus of Paper A.
8 Chapter 1. Introduction

1.4.2 Protothreads and Event-Driven Programming

Networked embedded systems must handle multiple events that occur concur-
rently; they must handle both interactions with the physical world and interac-
tions with other systems that communication over the network. To handle con-
currency in a small amount of memory, many memory-constrained networked
embedded systems are built on an event-driven programming model rather than
a multithreaded model [29, 53].
In the multithreaded model concurrency is implemented by designing the
system as a set of threads. Each thread is a program that runs concurrently
with the other threads in the system. Threads can wait for events to occur. The
thread then blocks its execution until it is woken up by the operating system.
Each thread requires its own processor stack. In a memory-constrained system,
multiple stacks may require a large part of the total memory.
In the event-driven model, the system is not made up of threads but of
event handlers that all run on the same stack. Since only one stack is used,
the memory requirements are reduced. In the event-driven model all program
execution is triggered by internal or external events. When an event occurs, an
event handler is invoked by the event dispatcher. Event handlers cannot wait for
events to occur but must explicitly return control to the event dispatcher. The
fact that event handlers cannot do a blocked wait complicates implementation
of high-level logic that cannot be expressed as a single event handler. Such
logic must be divided into multiple event handlers where the flow control of the
logic is implemented as explicit state machines. Such state machines typically
are difficult to write, read, and debug [6, 23, 42, 44, 67].
If it was possible to somehow combine the multithreaded and the event-
driven model, perhaps programs could be written in a sequential style without
explicit state machines and still have a low memory overhead. Furthermore,
it would be advantageous if it was possible to do this in a standard, general-
purpose programming language. The study of this problem is initiated in Paper
B and continued in Papers C and D.

1.4.3 Dynamic Module Loading

Most operating systems for memory-constrained embedded systems are de-
signed in a monolithic fashion where the operating system, hardware device
drivers, and all application programs are compiled into a single monolithic bi-
nary that is installed in the read-only memory of the microprocessor in the
embedded system. This monolithic approach has its benefits, such as memory
1.5 Thesis Structure 9

footprint predictability and the opportunity for cross-module optimizations, but

makes run-time replacement of applications difficult. The monolithic binary
has cross dependencies between the operating system and the applications. It
is not possible to replace an application without having to alter the compiled
operating system code. To replace an application, the entire monolithic binary
must be replaced with a new monolithic binary that includes the new appli-
cation and an updated version of the compiled operating system code. When
the new binary has been installed in read-only memory, the system must be
rebooted.
Radio communication is the primary energy consumer in many wireless
sensor network hardware platforms. By reducing the amount of data that needs
to be transmitted across the network when performing a software update, the
energy consumption is reduced. Early methods for updating software in wire-
less sensor networks required the entire monolithic binary to be transmitted
over the network [33]. Even if only a single application needs to be updated,
the entire operating system binary must be transfered, incurring a large energy
overhead.
If it was possible to transmit only the parts of the system that needed to be
updated the energy consumption of software updates would be lower. Further-
more, it would be advantageous if it was possible to do this by using standard
or general-purpose mechanisms. This is investigated in Papers B and E.

1.5 Thesis Structure

This thesis consists of a thesis summary followed by five papers that are pub-
lished in peer-reviewed conference and workshop proceedings. Conference
proceedings are the primary venue for scientific publishing in the area of this
thesis.
The rest of this thesis is structured as follows. Chapter 2 summarizes the
scientific contributions of this thesis and reviews the impact that the research
in the thesis has. Chapter 3 summarizes the papers of the thesis and their
individual contributions. Related work is reviewed in Chapter 4. I conclude
the thesis and present ideas for future work in Chapter 5. Chapter 6 presents
relevant software I have developed but that is not strictly part of the thesis and
a list of papers authored or co-authored by me during the thesis work but not
included in this thesis. Chapters 7, 8, 9, 10, and 11 contain the five papers of
the thesis.
Chapter 2

Scientific Contributions and

Impact

2.1 Scientific Contributions

In this thesis I make three main scientific contributions. My first contribution
is showing that the TCP/IP protocols can be implemented efficiently enough
to run even in memory-constrained embedded systems, but that it leads to sig-
nificantly lower network throughput. My second contribution is protothreads,
an intentionally simple yet very powerful programming abstraction. I show
that protothreads significantly reduce the complexity of event-driven programs
for memory-constrained systems. My third contribution is that I show that
dynamic linking and loading of native code modules is feasible for memory-
constrained operating systems. I quantify the energy consumption and evaluate
dynamic loading, linking, and execution of native code against loading and ex-
ecution of virtual machine code.
Prior to the work presented in this thesis, the general consensus was that
the TCP/IP protocol suite was inherently unsuited for memory-constrained sys-
tems because an RFC standards-compliant implementation of the TCP/IP pro-
tocols would require far too much resources in terms of memory to even be
possible to fit in a memory-constrained embedded system. In this thesis I show
that RFC-compliant TCP/IP can be used in systems an order of magnitude
smaller than previously believed possible.
My protothreads programming abstraction shows that it is possible to do se-

11
12 Chapter 2. Scientific Contributions and Impact

quential programming on event-driven systems without the memory overhead

of full multi-threading. Protothreads are an intentionally simple but powerful
mechanism that significantly reduces the complexity of state machine-based
event-driven programs for memory-constrained systems. Out of seven state
machine-based event driven programs rewritten with protothreads, the state
machines could be almost completely removed for all of them and the num-
ber of lines of code was reduced with 31% on average, at the price of a few
hundred bytes increase in size of the compiled code.
The memory overhead of protothreads is small. A program written with
protothreads uses only one byte more memory than the equivalent program
implemented as a state machine-based event-driven program. The run-time
overhead of a protothread is small; the run-time of a protothread is only a few
processor cycles higher than that of the equivalent state machine-based event-
driven program. This makes protothreads usable even in time-critical code such
as hardware interrupt handlers.
My Contiki operating system is the first operating system in the wireless
sensor network community to support loadable modules. Previous operating
systems for sensor networks require the entire system and applications to be
compiled into a single monolithic binary that is loaded onto each sensor node.
In such systems the entire system binary must be replaced when reprogram-
ming a single module. With loadable modules, only the module that is changed
needs to be replaced. Contiki is also the first operating system for sensor net-
works to support both event-driven and threaded programming.
In this thesis I quantify the energy cost of run-time dynamic loading and
linking of native code modules in the de facto standard ELF object code for-
mat. My results show that the overhead of run-time dynamic linking is fairly
small and that the overhead mainly is due to the ELF file format and not due
to the dynamic linking mechanism as such. Furthermore, when comparing dy-
namic linking and loading of native code modules with virtual machine mod-
ules the results suggest that combinations of virtual machine code and dynami-
cally loaded native code provide better energy efficiency than pure native code
modules and pure virtual machine code modules.

2.2 Impact
The impact of the research in this thesis has been, and continues to be, large.
The lwIP and uIP software developed as part of this thesis has been adopted
by well over one hundred companies world-wide in a wide variety of embed-
2.2 Impact 13

ded devices. Examples include satellite systems, oil boring and oil pipeline
equipment, TV transmission devices, equipment for color post-processing of
movies, world-wide container monitoring and security systems, switches and
routers, network cameras, and BMW racing car engines. The software is also
used in system development kits from hardware manufacturers including Ana-
log Devices, Altera, and Xilinx, which greatly increases the dissemination of
the software. Articles in professional embedded systems developer magazines
have been written, by others, on porting the uIP software for new hardware
platforms [9]. The software is also covered in books on embedded systems and
networking [34, 50] and is included in embedded operating systems [2, 14, 65].
The lwIP and uIP homepages have for a few years been among the top five hits
for Google searches such as TCP/IP stack and embedded TCP/IP.
The Contiki operating system has become a well-known operating sys-
tem in the wireless sensor network community and is used by several research
projects in the area. Many ideas from Contiki such as dynamic module load-
ing and the optional multi-threading library have also been adopted by other
operating systems for wireless sensor networks [26, 51]. The dynamic loader
mechanism in Contiki has also been investigated for use by Ericsson Mobile
Platforms as part of the hardware platform used in many of today’s 3G mobile
telephones.
Protothreads are currently used by numerous embedded software devel-
opers and have been recommended twice in acclaimed embedded developer
Jack Ganssle’s Embedded Muse newsletter [20]. Protothreads have also
been ported, by others, to other programming languages and operating sys-
tems [40, 58].
The papers and the software in this thesis are used in advanced courses on
embedded systems and sensor networks at many universities and institutions
throughout the world. The papers are cited by many scientific papers in the
area of wireless sensor networks. In early 2007 the Google Scholar citation
index [3] reports a total of 121 citations of the first three papers in this thesis.
The last two papers had not yet been indexed by Google Scholar.
Chapter 3

Summary of the Papers and

Their Contributions

The thesis is a collection of five papers, Paper A, B, C, D, and E. All papers are
published in peer-reviewed conference and workshop proceedings. Conference
proceedings are the primary venue for scientific publishing in the area of this
thesis. I have presented all papers at the conferences and workshops at which
they appeared.
Papers A, D, and E are published at top-tier conferences, ACM MobiSys
2003 and ACM SenSys 2006. ACM MobiSys is a high-quality single track
conference. ACM SenSys is regarded as the most prestigious conference in the
area of wireless sensor networks. Paper B was published in the first instance
of a now established high quality workshop, IEEE EmNets 2004. Paper C
was published at the REALWSN 2005 workshop on real-world wireless sensor
networks.
Paper A presents and evaluates the lwIP and uIP TCP/IP stacks. The event-
driven nature of uIP forms the basis of the Contiki operating system introduced
in Paper B. Paper B presents Contiki and the dynamic module loading mech-
anism and the per-process optional multi-threading in Contiki. The dynamic
loading mechanism in Contiki is further developed and evaluated in Paper E.
The multi-threading mechanism in Contiki, presented in Paper B, is the first
step towards the protothreads mechanism that I introduce in Paper C. Paper
D refines, extends, and evaluates the protothreads mechanism. Paper C also
includes a qualitative comparison between protothreads, events, and threads
that is not included in Paper D. Papers B, C, D, and E show how the research

15
16 Chapter 3. Summary of the Papers and Their Contributions

has progressed from development of the underlying mechanisms in Contiki in

Paper B and the novel protothreads mechanism in Paper C, to more refined
mechanisms and their evaluations in Paper D and E.

3.1 Paper A: Full TCP/IP for 8-Bit Architectures

Adam Dunkels. Full TCP/IP for 8-bit architectures. In Proceedings of The
First International Conference on Mobile Systems, Applications, and Services
(ACM MobiSys 2003), San Francisco, USA, May 2003.
Summary. The TCP/IP protocol suite is the family of protocols used for
communication over the global Internet, and is often used in private networks
such as local-area networks and corporate intranets. In order to attach a de-
vice to the network, the device must be able to use the TCP/IP protocols for
communication.
This paper presents two small implementations of the TCP/IP protocol
stack with slightly different designs; lwIP, which is designed in a modular and
generic fashion, similar to how large-scale protocol implementations are de-
signed, and uIP which is designed in a minimalistic fashion and only contains
the absolute minimum set of features required to fulfill the protocol standards.
In order to reduce the code size and the memory requirements, the uIP imple-
mentation uses an event-based API which is fundamentally different from the
most common TCP/IP API, the BSD socket API.
As expected, measurements from an actual system running the implementa-
tions show that the smaller uIP implementation provides a very low throughput,
particularly when sending data to a PC host. However, small systems running
uIP usually do not produce enough data for the performance degradation to
become a serious problem.
Contribution. The main contribution of this paper is that it refutes the
common conception that the TCP/IP protocol suite is too “heavy-weight” to be
possible to fully implement on tiny devices. At the time this paper was written,
most TCP/IP protocol stack implementations were designed for workstations
and server-class systems, where communication performance was the primary
concern. This caused a wide-spread belief that tiny devices such as sensor net-
work nodes would be too constrained to be able to fully implement the TCP/IP
stack. There were also a number of commercial implementations of the TCP/IP
stack intended for embedded devices, where the protocols in the TCP/IP suite
had been modified in order to reduce the code size and memory usage of their
implementation. Such implementations are problematic as they may cause in-
3.2 Paper B: Contiki - a Lightweight and Flexible Operating System for
Tiny Networked Sensors 17

teroperability problems with other TCP/IP implementations. This paper shows

that it is possible to implement the TCP/IP stack in a very small code size and
with a very small memory usage, without modifying the protocols. However,
a small TCP/IP stack will not be able to achieve as high throughput as a larger
implementation.

3.2 Paper B: Contiki - a Lightweight and Flexible

Operating System for Tiny Networked Sen-
sors
Adam Dunkels, Björn Grönvall, and Thiemo Voigt. Contiki - a Lightweight
and Flexible Operating System for Tiny Networked Sensors. In Proceedings
of the First IEEE Workshop on Embedded Networked Sensors (IEEE Emnets
2004), Tampa, Florida, USA, November 2004.
Summary. The ability to reprogram wireless sensor networks using the
radio reduces the time for the software development cycle. However, operat-
ing systems for sensor networks require the full system image to be replaced
when reprogramming a sensor network over the radio. This paper presents
Contiki, a lightweight and flexible operating system for tiny networked em-
bedded devices. Unlike other operating systems for sensor networks, Contiki
has the ability to selectively load and unload individual programs. This re-
duces the amount of data that needs to be transmitted when reprogramming a
sensor network over the radio, since the entire system binary does not need to
be transfered. We show that the dynamic loading can be done while keeping
the memory footprint down.
Contiki supports two kinds of concurrency mechanisms: an event-driven
interface and a preemptive multi-threading interface. The advantages of an
event-driven model is that it is possible to implement using very small amounts
of memory. Preemptive multi-threading, on the other hand, requires compara-
tively large amounts of memory to hold per-thread stacks. Furthermore, there
are types of programs that are unsuited for the event-driven model but work
well with preemptive multi-threading. Computationally intensive programs
such as encryption algorithms are typical examples of this.
Unlike other operating systems, Contiki leverages both models by basing
the system on an event-driven kernel and implementing preemptive multi-
threading as an optional application library. This allows using preemptive
multi-threading on a per-program basis. Experiments show that a Contiki sys-
18 Chapter 3. Summary of the Papers and Their Contributions

tem is able to continue to respond to events in a timely manner while perform-

ing a long-running computation as a preemptible thread.
Contribution. The main contribution of this paper is introducing the idea
of loadable modules for a sensor network operating system. Furthermore, the
combination of multi-threaded and event-driven operation in Contiki was not
previously used in sensor network operating systems.
Per-author contributions. I developed the Contiki system and I wrote the
paper. Contiki was developed and the paper was written on my initiative. The
idea of providing preemptive multi-threading as an application library on top of
the event-driven Contiki kernel was formed in cooperation with Björn Grönvall
who also took part in writing the paper. All other ideas in the paper are mine.
Thiemo Voigt was involved in the writing of the paper.

3.3 Paper C: Using Protothreads for Sensor Node

Programming
Adam Dunkels, Oliver Schmidt, and Thiemo Voigt. Using protothreads for
sensor node programming. In Proceedings of the Workshop on Real-World
Wireless Sensor Networks (REALWSN 2005), Stockholm, Sweden, June 2005.
Summary. Memory-constrained systems are often programmed using the
event-driven model rather than a multi-threaded model because multi-threading
has a higher memory overhead than the event-driven model. The problem with
the event-driven model is that sequencing high-level operations must be done
in a state machine-style. State machine-programs written in C are difficult to
understand, debug, and maintain.
Protothreads is a lightweight mechanism developed to make it possible to
write high-level sequences as a sequence of source code statements rather than
as explicit state machines.
Contribution. This paper introduces the idea of protothreads as a
lightweight mechanism to provide sequential programming on top of an event-
driven system. The protothreads mechanism is intentionally simple; it requires
only two bytes of memory per protothread and it is possible to implement pro-
tothreads using only 6 lines of C code.
Per-author contributions. I invented protothreads, came up with their
implementation in C, and wrote the paper. The ideas in the paper are mine
and the work was done and the paper was written on my initiative. Oliver
Schmidt and I discussed and further developed the ways protothreads could be
implemented. Thiemo Voigt was involved in the writing of the paper.
3.4 Paper D: Protothreads: Simplifying Event-Driven Programming of
Memory-Constrained Embedded Systems 19

3.4 Paper D: Protothreads: Simplifying

Event-Driven Programming of Memory-
Constrained Embedded Systems
Adam Dunkels, Oliver Schmidt, Thiemo Voigt, and Muneeb Ali. Protothreads:
Simplifying event-driven programming of memory-constrained embedded sys-
tems. In Proceedings of the 4th International Conference on Embedded Net-
worked Sensor Systems (ACM SenSys 2006), Boulder, Colorado, USA, Novem-
ber 2006.
Summary. In this paper we significantly extend the results of Paper C in
that we measure the usefulness of protothreads by reimplementing a set of state
machine-based programs with protothreads. We show that protothreads can
significantly reduce the complexity in terms of number of lines of code for the
rewritten programs. Furthermore, we show that the execution time overhead of
protothreads is small: only a few processor cycles per invocation. This is small
enough to make protothreads usable even in time-critical code such as interrupt
handlers. However, protothreads incur a code size overhead of a few hundred
bytes per program.
Contribution. The contribution of this paper is that we show that pro-
tothreads, a very simple mechanism, significantly reduce the complexity of
state machine-based event-driven programs without increasing memory re-
quirements or execution time.
Per-author contributions. I invented protothreads, came up with their im-
plementation in C, designed and carried out the evaluation, and wrote the paper.
The ideas in the paper are mine and work was done and the paper was written
on my initiative. Oliver Schmidt and I discussed and further developed some
of the ways protothreads could be implemented. Thiemo Voigt was involved in
the writing of the paper. Muneeb Ali wrote two paragraphs in the Related Work
section. Measuring the code size of the rewritten programs was suggested by
SenSys paper shepherd Philip Levis.

3.5 Paper E: Run-time Dynamic Linking for Re-

programming Wireless Sensor Networks
Adam Dunkels, Niclas Finne, Joakim Eriksson, and Thiemo Voigt. Run-time
dynamic linking for reprogramming wireless sensor networks. In Proceedings
of the 4th International Conference on Embedded Networked Sensor Systems
20 Chapter 3. Summary of the Papers and Their Contributions

(ACM SenSys 2006), Boulder, Colorado, USA, November 2006.

Summary. Reprogramming wireless sensor networks is a very useful fea-
ture in a large number of situations. Many different methods for reprogram-
ming wireless sensor networks have been investigated. In this paper we intro-
duce a new reprogramming method: dynamic linking of native code using the
standard ELF object code file format.
We compare the energy efficiency of dynamic linking with three other re-
programming mechanisms: full image replacement, an application-specific vir-
tual machine, and a standard Java virtual machine. We conclude that the ELF
format has a high overhead but that the overhead is due to the file format and
not due to the dynamic linking mechanism. Furthermore, we show that, from
an energy consumption perspective, combinations of virtual machine code and
native code may prove to be the most energy efficient alternative.
Contribution. This paper shows for the first time that a standard mecha-
nism for loading code, dynamic linking of ELF object code files, is doable and
feasible even in resource-constrained wireless sensor network nodes. Further-
more, the quantification of the energy cost of dynamic linking and the quantifi-
cation of the energy costs of native code versus virtual machine code and Java
code for sensor networks is a contribution on its own.
Per-author contributions. I designed and implemented the Contiki ELF
loader, the Contiki virtual machine and its compiler, did the initial porting of
the leJOS Java virtual machine, designed and carried out the experimental eval-
uation, and wrote the paper. The ideas in the paper are mine and the work was
done and the paper was written on my initiative. Niclas Finne and Joakim
Eriksson did the largest part of the porting of the leJOS Java virtual machine
to Contiki. Niclas also took part in performing the energy measurements in the
paper and Joakim wrote the initial version of the subsection on the Java virtual
machine. Thiemo Voigt was involved in the writing of the paper.
Chapter 4

Related Work

4.1 Small TCP/IP Implementations

There are several small TCP/IP implementations for memory-constrained em-
bedded systems. However, most of those implementations refrain from imple-
menting required protocol mechanisms in order to reduce the complexity of the
implementation. The resulting implementation may therefore not be fully com-
patible with other TCP/IP implementations. Hence communication between
the memory-constrained system and another system may not be possible.
Many small TCP/IP implementations are tailored for a specific application,
such as running a web server. Tailoring the TCP/IP implementation for a spe-
cific application makes it possible to significantly reduce the implementation
complexity. However, such an implementation does not provide a general com-
munications mechanism that can be used for other applications. The PICmicro
stack [11] is an example of such a TCP/IP implementation. Unlike such imple-
mentations, the uIP and lwIP implementations are not designed for a specific
application.
Other implementations rely on the assumption that the small embedded
device will always be communicating with a full-scale TCP/IP implementation
running on a PC or work-station class device. Under this assumption it is
possible to remove certain mechanisms that are required for full compatibility.
Specifically, support for IP fragment reassembly and for TCP segment size
variation are two mechanisms that are required by the standard but are often
left out. Examples of such implementations are Texas Instrument’s MSP430
TCP/IP stack [17] and the TinyTCP code [15]. Unlike those implementations,

21
22 Chapter 4. Related Work

the uIP or the lwIP stack support both IP fragment reassembly and a variable
maximum TCP segment size.
In addition to the TCP/IP implementation for small embedded systems,
there is a large class of TCP/IP implementations for embedded systems with
less constraining limitations. Typically, such implementations are based on the
TCP/IP implementation from the BSD operating system [52]. These imple-
mentations do not suffer from the same problems as the tailored implementa-
tions. However, such implementations require too large amounts of resources
to be feasible for memory-constrained embedded systems. Such implementa-
tions are typically orders of magnitude larger than uIP.

4.2 Operating Systems for Wireless Sensor Net-

works
TinyOS [29] is probably the earliest operating system that directly targets the
specific applications and limitations of sensor devices. TinyOS is built around a
lightweight event scheduler where all program execution is performed in tasks
that run to completion. TinyOS uses a special description language for com-
posing a system of smaller components [21] that are statically linked with the
kernel to a complete monolithic binary image of the system. After linking,
modifying the system is not possible [42]. The Contiki system is also designed
around a lightweight event-scheduler, but is designed to allow loading, unload-
ing, and replacing modules at run-time.
To load new applications in an operating system where the entire system
is compiled into a monolithic binary, the entire system binary must be up-
dated. Since the energy consumption of distributing code in sensor networks
increases with the size of the code to be distributed several attempts have been
made to reduce the size of the code to be distributed. Reijers and Langen-
doen [60] produce an edit script based on the difference between the modified
and original binary system image. After a number of optimizations including
architecture-dependent ones, the script is distributed throughout the network.
A similar approach has been developed by Jeong and Culler [32] who use the
rsync algorithm to generate the difference between modified and original exe-
cutable. Koshy and Pandey’s diff-based approach [37] reduces the amount of
flash rewriting by modifying the linking procedure so that functions that are
not changed are not shifted.
FlexCup [49] uses another approach to enable run-time installation of
TinyOS software components: loadable applications include the symbolic
4.2 Operating Systems for Wireless Sensor Networks 23

names of functions in the loaded application. When an application is loaded,

the symbolic names are resolved and the loaded binary is updated. FlexCup
uses a complete duplicate image of the system’s monolithic binary image that
is stored in external flash ROM. The copy of the system image is used for
constructing a new system image when a new program has been loaded. How-
ever, FlexCup uses a non-standard format and is less portable than the loader in
Contiki. Further, FlexCup requires a reboot after a program has been installed,
requiring an external mechanism to save and restore the state of all other ap-
plications as well as the state of running network protocols across the reboot.
Contiki does not need to be rebooted after a program has been installed. Also
the Contiki dynamic linker does not alter the core image when programs are
loaded and therefore no external copy of the core image needs to be stored.
In order to provide run-time reprogramming for TinyOS, Levis and Culler
have developed Maté [42], a virtual machine for TinyOS devices. Code for the
virtual machine can be downloaded into the system at run-time. The virtual
machine is specifically designed for the needs of typical sensor network appli-
cations. Similarly, the MagnetOS [10] system uses a virtual Java machine to
distribute applications across the sensor network. The advantages of using a
virtual machine instead of native machine code is that the virtual machine code
can be made smaller, thus reducing the energy consumption of transporting the
code over the network. One of the drawbacks is the increased energy spent
in interpreting the code—for long running programs the energy saved during
the transport of the binary code is instead spent in the overhead of executing
the code. Contiki does not suffer from the executional overhead since modules
loaded into Contiki are compiled to native machine code.
The Mantis system [5] uses a traditional preemptive multi-threaded model
of operation. Mantis enables reprogramming of both the entire operating sys-
tem and parts of the program memory by downloading a program image onto
EEPROM, from where it can be burned into flash ROM. Due to the multi-
threaded semantics, every Mantis program must have stack space allocated
from the system heap, and locking mechanisms must be used to achieve mu-
tual exclusion of shared variables. In Contiki, only programs that explicitly
require multi-threading need to allocate an extra stack.
Systems that offer loadable modules besides Contiki include SOS [26] and
Impala [46]. Impala features an application updater that enables software up-
dates to be performed by linking in updated modules. Updates in Impala are
coarse-grained since cross-references between different modules are not pos-
sible. Also, the software updater in Impala was only implemented for much
more resource-rich hardware than the memory-constrained systems considered
24 Chapter 4. Related Work

in this thesis. SOS [26], which was published after Contiki, is similar in design
to Contiki: SOS consists of a small kernel and dynamically-loaded modules.
However, unlike Contiki SOS uses position independent code to achieve relo-
cation and jump tables for application programs to access the operating system
kernel. Application programs can register function pointers with the operat-
ing system for performing inter-process communication. Position independent
code is not available for all platforms, however, which limits the applicability
of this approach.

4.3 Programming Models for Wireless Sensor

Networks
While I use the C programming language throughout this thesis, others have
investigated the use of new programming languages and programming models
for programming memory-constrained networked embedded systems.

4.3.1 Macro-programming and New Programming Lan-

guages
Research in the area of software development for sensor networks has led to a
number of new abstractions that aim at simplifying the programming of sensor
networks [4, 13]. Approaches with the same goal include macro-programming
of aggregates of sensor nodes [19, 24, 56, 68], high-level database abstractions
of sensor networks [47], and network neighborhood abstractions [54, 68, 71].
The work in this thesis differs from these sensor network programming abstrac-
tions in that I target the difficulty of low-level event-driven programming of
individual sensor nodes, programming-in-the-small, rather than the difficulty
of developing application software for sensor networks, programming-in-the-
large.
Kasten and Römer [36] have also identified the need for new abstractions
for managing the complexity of event-triggered state machine programming.
They introduce OSM, a state machine programming model based on Harel’s
StateCharts[27] and use the Esterel language. The model reduces both the
complexity of the implementations and the memory usage. Their work is dif-
ferent from protothreads in that they help programmers manage explicit state
machines, whereas protothreads are designed to reduce the number of explicit
state machines. Furthermore, OSM requires support from an external OSM
4.3 Programming Models for Wireless Sensor Networks 25

compiler to produce the resulting C code, whereas the prototype implementa-

tions of protothreads only make use of the regular C preprocessor.
Script languages are another alternative for programming embedded sys-
tems. Script languages do not need special compilers as the script code is
interpreted by the embedded systems. However, existing structured script lan-
guages are typically too large for the memory-constraints considered in this
thesis. The SensorWare system [12] uses a reduced version of Tcl to provide
a script-based programming environment for sensor networks. However, the
system is designed for sensor nodes with an order of magnitude more memory
resources than the systems considered in this thesis; for example, the Sensor-
Ware system occupies 180 kilobytes of memory.
Rappit [25] is a development framework for scripting languages for em-
bedded systems. Rappit uses a host environment running Python that sends
commands to the embedded systems. The host environment runs on a resource-
rich server system outside of the sensor network that translates commands into
simpler messages that are executed by the embedded system. Tapper [72] is a
command language and lightweight stack-based script engine for sensor nodes
built with Rappit. The Tapper language provides primitives for accessing hard-
ware devices such as analog to digital converters and for sending and receiving
radio packets. Both Rappit and Tapper require assistance from a host system
for interpreting the commands.

4.3.2 Virtual Machines

Virtual machines have been investigated for sensor networks as an approach
to reduce the distribution energy costs for software updates. The code size of
the programs running on top of the virtual machine can be kept to a minimum
since the virtual machine can be tailored to the needs of an application for a
specific domain such as sensor networks. The drawback of virtual machines
is the increased execution overhead over native code. The virtual machine
approach is different from the approach taken in this thesis as virtual machines
need both special compilers that can produce code for the virtual machine as
well as a virtual machine running on the embedded system.
Maté [42] was the first virtual machine specifically targeted to wireless sen-
sor networks. Maté is a stack-based virtual machine that runs on top of TinyOS.
Maté instructions are 8 bits wide. Each virtual machine instruction is executed
in a separate TinyOS run-to-completion task. Levis et al. [43] have further in-
vestigated the use of application specific virtual machines (ASVMs) that are
compiled for the needs of a particular application or set of applications. Other
26 Chapter 4. Related Work

examples of stack-based virtual machines for sensor networks are DVM [7]
and my CVM Contiki virtual machine (Paper E). There are also Java-based
virtual machines for sensor networks apart from the Java VM in Paper E, such
as the VM⋆ system [38].
Chapter 5

Conclusions and Future

Work

5.1 Conclusions
In this thesis I investigate three aspects of programming memory-constrained
networked embedded systems: the use of the TCP/IP protocol stack for
memory-constrained systems; the novel protothread programming abstraction
for state machine-based programs for event-driven systems; and the use of dy-
namic loading and linking of native code in an operating system for memory-
constrained embedded systems. A general theme throughout this work is
how applicable standard or general-purpose mechanisms and methods are to
memory-constrained systems. I have identified and quantified trade-offs in
both the use of general-purpose mechanisms for memory-constrained systems
and the use of a general-purpose programming language for reducing the com-
plexity of memory-efficient programming for memory-constrained systems.
In this thesis I show that the standard TCP/IP protocol stack can be im-
plemented efficiently enough to be usable even in memory-constrained sys-
tems, but that such an implementation leads to a significant decrease in net-
work throughput. Furthermore, the results show that protothreads reduces the
complexity of state machine-based event-driven programs, while having a very
small memory and run-time overhead. Finally, the Contiki operating system
shows that dynamic loading and run-time dynamic linking of native code is a
feasible mechanism for memory-constrained networked embedded systems.

27
28 Chapter 5. Conclusions and Future Work

There are at least two conclusions that can be drawn from my research.
First, in many cases it is possible to use standard protocols and mechanisms
developed for general-purpose computers even in memory-constrained embed-
ded systems. However, there are trade-offs in terms of both memory footprint
and energy. Second, protothreads show that it is possible to combine features
from multi-threading and event-driven programming; sequential programming
from the multi-threaded model and a small memory overhead from the event-
driven model. Since a slightly limited version of protothreads can be imple-
mented in the general-purpose C programming language, it is possible to do
memory-efficient sequential programming in C without requiring the use of a
special-purpose programming language. However, there are trade-off due to
the limitations of the C-based implementation of protothreads.

5.2 Future Work

The research in this thesis can be continued in two ways: extensions to or con-
tinuations of the research presented in this thesis or application of the methods
and approaches used in this research to new problems or problem domains.
There is one immediate extension to the work in this thesis; the current im-
plementation of protothreads does not implement the full protothreads model
because of the limitations of the C preprocessor. By implementing a special
pre-compiler that would translate protothread-based code into regular C code
it is possible to implement the full model. This also makes it possible to mini-
mize the memory overhead of protothreads.
The work on energy-efficient loading of program modules can be continued
in at least two ways. Firstly, the energy consumption for transferring loadable
modules in Contiki can be reduced by compressing the modules with loss-less
compression techniques before the modules are transmitted across the network.
Investigating the use of standard compression techniques and the trade-offs in
terms of computational energy consumption is an interesting continuation of
the work on the dynamic loader in Contiki. Secondly, it would be interesting to
compare the use of dynamic module loading in Contiki with differential-based
mechanisms for loading modules in monolithic operating system. Investigating
the trade-offs and how they vary with different applications and network types
would be interesting.
Finally, the approaches and methods used in this thesis can be applied to
nearby problem areas. The memory management scheme used in uIP, where a
single buffer is used for both incoming and outgoing packets, might be possible
5.2 Future Work 29

to apply to other types of network protocols, such as power cycling MAC pro-
tocols for sensor networks. By using a single buffer, rather than a multi-buffer
management scheme, it might be possible to reduce both the data memory re-
quirements and the code footprint. However, additional mechanisms might be
needed to ensure that no packet queuing is required.
The research in this thesis suggests that there is a relation between imple-
mentation complexity in terms of memory footprint and the achievable appli-
cation performance; the smaller uIP implementation achieves much lower net-
work throughput than the larger lwIP implementation. A system designer that
requires high throughput can choose the larger lwIP stack, whereas a system
designer that requires a low memory footprint, but does not have any through-
put requirements, can choose the smaller uIP stack. It would be interesting to
investigate if this relation exists in other types of protocols as well. For exam-
ple, many protocols for distributing software updates through sensor networks
have been developed [30, 33, 45, 55, 64], each having different properties in
terms of both performance and footprint. A system designer would want to be
able to choose how much system resources to commit to the software update
feature and what kind of performance to expect from the feature, depending on
the requirements of particular deployments. By implementing a set of software
dissemination protocols and measuring their memory footprint and achievable
throughput it might be possible to quantify trade-offs between memory foot-
print and performance.
The systems in this thesis are built using a “bottom-up” approach rather
than a “top-down” approach. Instead of breaking down the system into smaller
components that are independently implemented and composed into a com-
plete system, I have started from the bottom with a set of simple building
blocks from which I have built a system that is similar to, but not always
equivalent to, a system built from the top down. It would be interesting to use
this approach for building new programming mechanisms such as abstractions
for programming of aggregates of memory-constrained networked embedded
systems, so-called macro-programming. Other approaches towards building
macro-programming systems have started by defining the programming inter-
face and decomposed this into smaller modules that are executed by each indi-
vidual node, e.g. the work by Newton et al. [56, 57] and Gummadi et al. [24].
By instead building upwards from a set of small and simple node-level abstrac-
tions to an abstract interface for network-level programming it may be possible
to build macro-programming systems that are more efficient in terms of both
implementation complexity and network communication than systems built in
a top-down fashion.
Chapter 6

Other Software and

Publications

6.1 Software
In addition to the software I have developed as part of this thesis I have written
a number of other programs that are relevant to the research in this thesis.

Miniweb A proof-of-concept implementation of selected parts of the TCP/IP

protocol stack together with a simple web server. The code con-
sists of 400 lines of C and uses only 30 bytes of RAM. Although
the implementation sacrifices a large number of protocol mechanisms,
it does include TCP’s congestion control mechanisms. The code
is not intended for actual use. The source code is available at
http://www.sics.se/˜adam/miniweb/.

phpstack A proof-of-concept implementation of an application-specific

TCP/IP stack together with a web server written in 600 lines of PHP
code. The TCP/IP stack is more simplified than the Miniweb stack and
can only participate in single-segment sessions. Thus pages served by
the web server are limited to 1460 bytes. If the program had been imple-
mented in a low-level language such as C, it would only have required
a handful bytes of RAM. The code is not intended for actual use. The
source code is available at http://www.sics.se/˜adam/phpstack/.

31
32 Chapter 6. Other Software and Publications

VNC client, SMTP client, HTTP client, IRC client, and web browser
Proof-of-concept implementations of a number of application-layer
protocols built on top of uIP and Contiki. The purpose was to investigate
if the event-driven interface of uIP was expressive enough to implement
complex application layer protocols. However, it turned out that many of
the protocols required complex state machines on top of the event-driven
interface of uIP. This insight lead me to develop protothreads. The
source code is available at http://www.sics.se/contiki/.

6.2 Publications
The following peer-reviewed publications were authored or co-authored by me
during the work with this thesis but are not included in the thesis.

• Fredrik Österlind, Adam Dunkels, Joakim Eriksson, Niclas Finne, and

Thiemo Voigt. Cross-level sensor network simulation with cooja. In
Proceedings of the First IEEE International Workshop on Practical Is-
sues in Building Sensor Network Applications (SenseApp 2006), Tampa,
Florida, USA, November 2006.

• Helena Rivas, Thiemo Voigt, and Adam Dunkels. A simple and efficient
method to mitigate the hot spot problem in wireless sensor networks. In
Workshop on Performance Control in Wireless Sensor Networks, Coim-
bra, Portugal, May 2006.

• Muneeb Ali, Umar Saif, Adam Dunkels, Thiemo Voigt, Kay Römer,
Koen Langendoen, Joseph Polastre, and Zartash Afzal Uzmi. Medium
access control issues in sensor networks. ACM SIGCOMM Computer
Communication Review, April 2006.

• Thiemo Voigt, Adam Dunkels, and Torsten Braun. On-demand con-

struction of non-interfering multiple paths in wireless sensor networks.
In Proceedings of the 2nd Workshop on Sensor Networks at Informatik
2005, Bonn, Germany, September 2005.

• Adam Dunkels, Richard Gold, Sergio Angel Marti, Arnold Pears, and
Mats Uddenfeldt. Janus: An architecture for flexible access to sensor
networks. In First International ACM Workshop on Dynamic Intercon-
nection of Networks (DIN’05), Cologne, Germany, September 2005.
6.2 Publications 33

• Thiemo Voigt and Adam Dunkels. The impact of knowledge about

neighbors on the efficiency of geographic routing. In Proceedings of
Radio Sciences and Communication RVK’05, June 2005.

• Torsten Braun, Thiemo Voigt, and Adam Dunkels. Energy-efficient TCP

operation in wireless sensor networks. PIK Journal Special Issue on
Sensor Networks, 2005.

• Hartmut Ritter, Jochen Schiller, Thiemo Voigt, Adam Dunkels, and Juan
Alonso. Experimental Evaluation of Lifetime Bounds for Wireless Sen-
sor Networks. In Proceedings of the Second European Workshop on
Sensor Networks (EWSN2005), Istanbul, Turkey, January 2005.

• Adam Dunkels, Thiemo Voigt, Niclas Bergman, and Mats Jönsson. The
Design and Implementation of an IP-based Sensor Network for Intru-
sion Monitoring. In Swedish National Computer Networking Workshop,
Karlstad, Sweden, November 2004.

• Thiemo Voigt, Adam Dunkels, and Juan Alonso. Reliability in dis-

tributed TCP caching. In Workshop on Sensor Networks Workshop at
Informatik 2004, Ulm, Germany, September 2004.

• Adam Dunkels, Thiemo Voigt, Juan Alonso, and Hartmut Ritter. Dis-
tributed TCP caching for wireless sensor networks. In Proceedings of
the Third Annual Mediterranean Ad Hoc Networking Workshop (Med-
HocNet 2004), June 2004.

• Thiemo Voigt, Hartmut Ritter, Jochen Schiller, Adam Dunkels, and Juan
Alonso. Solar-aware Clustering in Wireless Sensor Networks. In Pro-
ceedings of the Ninth IEEE Symposium on Computers and Communica-
tions, June 2004.

• Juan Alonso, Adam Dunkels, and Thiemo Voigt. Bounds on the energy
consumption of routings in wireless sensor networks. In Proceedings
of the 2nd WiOpt, Modeling and Optimization in Mobile, Ad Hoc and
Wireless Networks, Cambridge, UK, March 2004.

• Adam Dunkels, Thiemo Voigt, and Juan Alonso. Making TCP/IP Viable
for Wireless Sensor Networks. In Proceedings of the First European
Workshop on Wireless Sensor Networks (EWSN 2004), work-in-progress
session, Berlin, Germany, January 2004.
34 Chapter 6. Other Software and Publications

• Adam Dunkels, Thiemo Voigt, Juan Alonso, Hartmut Ritter, and Jochen
Schiller. Connecting Wireless Sensornets with TCP/IP Networks. In
Proceedings of the Second International Conference on Wired/Wireless
Internet Communications (WWIC2004), Frankfurt (Oder), Germany,
February 2004.
• Laura Marie Feeney, Bengt Ahlgren, Assar Westerlund, and Adam
Dunkels. Spontnet: Experiences in configuring and securing small ad
hoc networks. In Proceedings of The Fifth International Workshop on
Network Appliances (IWNA5), Liverpool, UK, October 2002.
Bibliography

[1] Crossbow mica motes. Web page.

URL: http://www.xbow.com/

[2] eCos Embedded Configurable Operating System. Web page.

URL: http://sources.redhat.com/ecos/

[3] Google Scholar. Web page.

URL: http://scholar.google.com/

[4] T. Abdelzaher, J. Stankovic, S. Son, B. Blum, T. He, A. Wood, and C. Lu.

A communication architecture and programming abstractions for real-
time embedded sensor networks. In Workshop on Data Distribution for
Real-Time Systems, Providence, RI, USA, May 2003.

[5] H. Abrach, S. Bhatti, J. Carlson, H. Dai, J. Rose, A. Sheth, B. Shucker,

J. Deng, and R. Han. Mantis: system support for multimodal networks of
in-situ sensors. In Proceedings of the 2nd ACM international conference
on Wireless sensor networks and applications, pages 50–59, San Diego,
CA, USA, September 2003.

[6] A. Adya, J. Howell, M. Theimer, W. J. Bolosky, and J. R. Douceur. Co-

operative Task Management Without Manual Stack Management. In Pro-
ceedings of the USENIX Annual Technical Conference, pages 289–302,
2002.

[7] R. Balani, S. Han, R. Rengaswamy, I. Tsigkogiannis, and M. B Srivas-

tava. Multi-level software reconfiguration for sensor networks. In Pro-
ceedings of the 6th ACM/IEEE Conference on Embedded Software (EM-
SOFT ’06), Seoul, Korea, October 2006.

35
36 Bibliography

[8] H. Baldus, K. Klabunde, and G. Muesch. Reliable set-up of medical

body-sensor networks. In Proceedings of the First European Workshop on
Wireless Sensor Networks (EWSN’04), Berlin, Germany, January 2004.

[9] D. Barnett and A. J. Massa. Inside the uIP Stack. Dr. Dobb’s Journal,
February 2005.

[10] R. Barr, J. C. Bicket, D. S. Dantas, B. Du, T. W. D. Kim, B. Zhou, and

E. Sirer. On the need for system-level support for ad hoc and sensor
networks. SIGOPS Opererating Systems Review, 36(2):1–5, 2002.

[11] J. Bentham. TCP/IP Lean: Web servers for embedded systems. CMP
Books, October 2000.

[12] A. Boulis, C. Han, and M. B. Srivastava. Design and implementation of a

[13] E. Cheong, J. Liebman, J. Liu, and F. Zhao. TinyGALS: A programming

model for event-driven embedded systems. In Proc. of the 18th Annual
ACM Symposium on Applied Computing (SAC’03), Melbourne, Florida,
USA, March 2003.

[14] A. Christian and J. Healey. Gathering motion data using featherweight

sensors and TCP/IP over 802.15.4. In Proceedings of the IEEE Interna-
tional Symposium on Wearable Computers, On-Body Sensing Workshop,
Osaka, Japan, October 2005.

[15] G. H. Cooper. TinyTCP. Web page. 2002-10-14.

URL: http://www.csonline.net/bpaddock/tinytcp/

[16] National Research Council. Academic Careers for Experimental Com-

puter Scientists and Engineers. National Academy Press, 1994.

[17] A. Dannenberg. MSP430 internet connectivity. SLAA 137, November

2001. Avalible from www.ti.com.

[18] F. DeRemer and H. Kron. Programming-in-the large versus

programming-in-the-small. In Proceedings of the international confer-
ence on Reliable software, pages 114–121, Los Angeles, California,
1975.
Bibliography 37

[19] C. Frank and K. Römer. Algorithms for generic role assignment in wire-
less sensor networks. In Proceedings of the 3rd international conference
on Embedded networked sensor systems (SenSys ’05), San Diego, Cali-
fornia, USA, November 2005.

[20] J. Ganssle. The embedded muse. Monthly newsletter.

URL: http://www.ganssle.com/tem-back.htm

[21] D. Gay, P. Levis, R. von Behren, M. Welsh, E. Brewer, and D. Culler.

The nesC language: A holistic approach to networked embedded systems.
In Proceedings of the ACM SIGPLAN 2003 conference on Programming
language design and implementation, pages 1–11, San Diego, California,
USA, June 2003.

[22] J. Grenning. Why are you still using C? Embedded Systems Design, April
2003.

[23] R. Grimm, J. Davis, E. Lemar, A. MacBeth, S. Swanson, T. Anderson,

B. Bershad, G. Borriello, S. Gribble, and D. Wetherall. Programming
for pervasive computing environments. ACM Transactions on Computer
Systems, January 2002.

[24] R. Gummadi, O. Gnawali, and R. Govindan. Macro-programming wire-

less sensor networks using kairos. In Proc. of Distributed Computing in
Sensor Systems (DCOSS)’05, Marina del Rey, CA, USA, June 2005.

[25] J. Hahn, Q. Xie, and P. H. Chou. Rappit: framework for synthesis of

host-assisted scripting engines for adaptive embedded systems. In Pro-
ceedings of the 3rd IEEE/ACM/IFIP international conference on Hard-
ware/software codesign and system synthesis (CODES+ISSS ’05), Jersey
City, NJ, USA, September 2005.

[26] C. Han, R. K. Rengaswamy, R. Shea, E. Kohler, and M. Srivastava. Sos:

A dynamic operating system for sensor networks. In Proceedings of the
3rd international conference on Mobile systems, applications, and ser-
vices (MobiSys ’05), Seattle, WA, USA, June 2005.

[27] D. Harel. Statecharts: A visual formalism for complex systems. Science

of Computer Programming, 8(3):231–274, 1987.

[28] T. F. Herbert. The Linux TCP/IP Stack: Networking For Embedded Sys-
tems. Charles River Media, 2004.
38 Bibliography

[29] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler, and K. Pister. Sys-

tem architecture directions for networked sensors. In Proceedings of the
9th International Conference on Architectural Support for Programming
Languages and Operating Systems, Cambridge, Massachusets, USA,
November 2000.
[30] J. W. Hui and D. Culler. The dynamic behavior of a data dissemination
protocol for network programming at scale. In Proceedings of the 2nd
international conference on Embedded networked sensor systems (SenSys
’04), Baltimore, Maryland, USA, November 2004.
[31] Texas Instruments. Analog enewsletter. Web page, May 2004. Visited
2006-12-21.
URL: http://focus.ti.com/en/download/aap/enewsletters/0405enewsletter.htm

[32] J. Jeong and D. Culler. Incremental network programming for wire-

less sensors. In Proceedings of the First IEEE Communications Society
Conference on Sensor and Ad Hoc Communications and Networks IEEE
SECON (2004), Santa Clara, California, USA, October 2004.
[33] J. Jeong, S. Kim, and A. Broad. Network reprogramming. TinyOS docu-
mentation, 2003. Visited 2006-04-06.
URL: http://www.tinyos.net/tinyos-1.x/doc/NetworkReprogramming.pdf

[34] M. T. Jones. TCP/IP Application Layer Protocols for Embedded Systems.

Charles River Media, June 2002.
[35] P. Juang, H. Oki, Y. Wang, M. Martonosi, L. S. Peh, and D. Rubenstein.
Energy-efficient computing for wildlife tracking: design tradeoffs an d
early experiences with ZebraNet. In Proceedings of the 10th interna-
tional conference on architectural support for programming languages
and operating systems (ASPLOS-X), pages 96–107, San Jose, California,
2002.
[36] O. Kasten and K. Römer. Beyond event handlers: Programming wire-
less sensors with attributed state machines. In The Fourth International
Conference on Information Processing in Sensor Networks (IPSN), Los
Angeles, USA, April 2005.
[37] J. Koshy and R. Pandey. Remote incremental linking for energy-efficient
reprogramming of sensor networks. In Proceedings of the second Euro-
pean Workshop on Wireless Sensor Networks, 2005.
Bibliography 39

[38] J. Koshy and R. Pandey. Vm*: Synthesizing scalable runtime environ-

ments for sensor networks. In Proc. SenSys’05, San Diego, CA, USA,
November 2005.
[39] V. A. Kottapalli, A. S. Kiremidjian, J. P. Lynch, E. Carryer an d
T. W. Kenny, K. H. Law, and Y. Lei. Two-tiered wireless sensor net-
work architecture for structural health monitoring. In Proceedings of the
SPIE 10th Annual International Symposium on Smart Structures and Ma-
terials, San Diego, CA, March 2000.
[40] Framework Labs. Protothreads for Objective-C/Cocoa. Visited 2006-04-
06.
URL: http://www.frameworklabs.de/protothreads.html

[41] L. Laffea, R. Monson, R. Han, R. Manning, A. Glasser, S. Oncley, J. Sun,

S. Burns, S. Semmer, and J. Militzer. Comprehensive monitoring of CO2
sequestration in subalpine forest ecosystems and its relation to global
warming. In Proceedings of the 4th international conference on Em-
bedded networked sensor systems (SenSys ’06), pages 423–424, Boulder,
Colorado, USA, 2006.
[42] P. Levis and D. Culler. Maté: A tiny virtual machine for sensor networks.
In Proceedings of ASPLOS-X, San Jose, CA, USA, October 2002.
[43] P. Levis, D. Gay, and D Culler. Active sensor networks. In Proceedings of
ACM/Usenix Networked Systems Design and Implementation (NSDI’05),
Boston, MA, USA, May 2005.
[44] P. Levis, S. Madden, D. Gay, J. Polastre, R. Szewczyk, A. Woo,
E. Brewer, and D. Culler. The Emergence of Networking Abstractions
and Techniques in TinyOS. In Proceedings of ACM/Usenix Networked
Systems Design and Implementation (NSDI’04), San Francisco, Califor-
nia, USA, March 2004.
[45] P. Levis, N. Patel, D. Culler, and S. Shenker. Trickle: A self-regulating
algorithm for code propagation and maintenance in wireless sensor net-
works. In Proceedings of ACM/Usenix Networked Systems Design and
Implementation (NSDI’04), March 2004.
[46] T. Liu, C. Sadler, P. Zhang, and M. Martonosi. Implementing software
on resource-constrained mobile sensors: Experiences with Impala and
ZebraNet. In Proc. Second Intl. Conference on Mobile Systems, Applica-
tions and Services (MOBISYS 2004), June 2004.
40 Bibliography

[47] S. R. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong. Tinydb: an

acquisitional query processing system for sensor networks. ACM Trans-
actions on Database Systems, 30(1):122–173, 2005.

[48] A. Mainwaring, J. Polastre, R. Szewczyk, D. Culler, and J. Anderson.

Wireless sensor networks for habitat monitoring. In First ACM Workshop
on Wireless Sensor Networks and Applications (WSNA 2002), Atlanta,
GA, USA, September 2002.

[49] P. José Marrón, M. Gauger, A. Lachenmann, D. Minder, O. Saukh, and

K. Rothermel. Flexcup: A flexible and efficient code update mechanism
for sensor networks. In European Workshop on Wireless Sensor Net-
works, Zurich, Switzerland, January 2006.

[50] A. J. Massa. Embedded Software Development with eCos. Prentice Hall,

November 2002.

[51] W. McCartney and N. Sridhar. Abstractions for safe concurrent program-

ming in networked embedded systems. In Proceedings of the 4th interna-
tional conference on Embedded networked sensor systems (SenSys ’06),
pages 167–180, Boulder, Colorado, USA, 2006.

[52] M. K. McKusick, K. Bostic, M. J. Karels, and J. S. Quarterman. The

Design and Implementation of the 4.4 BSD Operating System. Addison-
Wesley, 1996.

[53] M. Melkonian. Get by Without an RTOS. Embedded Systems Program-

ming, 13(10), September 2000.

[54] L. Mottola and G. Picco. Programming wireless sensor networks with

logical neighborhoods. In Proceedings of the first international confer-
ence on Integrated internet ad hoc and sensor networks (InterSense ’06),
page 8, Nice, France, May 2006.

[55] V. Naik, A. Arora, P. Sinha, and H. Zhang. Sprinkler: A reliable and

scalable data dissemination service for wireless embedded devices. In
Proceedings of the 26th IEEE Real-Time Systems Symposium, Miami,
Florida, USA, December 2005.

[56] R. Newton, Arvind, and M. Welsh. Building up to macroprogramming:

An intermediate language for sensor networks. In Proc. IPSN’05, Los
Angeles, CA, USA, April 2005.
Bibliography 41

[57] R. Newton and M. Welsh. Region streams: functional macroprogram-

ming for sensor networks. In Proceeedings of the 1st international work-
shop on Data management for sensor networks (DMSN ’04), pages 78–
87, Toronto, Canada, 2004.

[58] J. Paisley and J. Sventek. Real-time detection of grid bulk transfer traffic.
In Proceedings of the 10th IEEE/IFIP Network Operations Management
Symposium, Vancouver, Canada, April 2006.

[59] J. Polastre, R. Szewczyk, and D. Culler. Telos: Enabling ultra-low power

wireless research. In Proc. IPSN/SPOTS’05, Los Angeles, CA, USA,
April 2005.

[60] N. Reijers and K. Langendoen. Efficient code distribution in wireless sen-

sor networks. In Proceedings of the 2nd ACM international conference
on Wireless sensor networks and applications, pages 60–67, San Diego,
CA, USA, September 2003.

[61] K. Römer and F. Mattern. The design space of wireless sensor networks.
IEEE Wireless Communications, 11(6):54–61, December 2004.

[62] J. Schiller, H. Ritter, A. Liers, and T. Voigt. Scatterweb - low power

nodes and energy aware routing. In Proceedings of Hawaii International
Conference on System Sciences, Hawaii, USA, January 2005.

[63] V. Singhvi, A. Krause, C. Guestrin, Jr. J. Garrett, and S. Matthews. In-

telligent light control using sensor networks. In Proceedings of the 3rd
international conference on Embedded networked sensor systems (SenSys
’05), pages 218–229, San Diego, California, USA, November 2005.

[64] T. Stathopoulos, J. Heidemann, and D. Estrin. A remote code update

mechanism for wireless sensor networks. Technical Report CENS-TR-
30, University of California, Los Angeles, Center for Embedded Net-
worked Computing, November 2003.

[65] J. T. Taylor. eXtreme Minimal Kernel. Shift-Right Technologies LLC.

URL: http://www.shift-right.com/xmk/

[66] J. Turley. The Two Percent Solution. Embedded Systems Design, Decem-
ber 2002.
[67] R. von Behren, J. Condit, and E. Brewer. Why events are a bad idea (for
high-concurrency servers). In Proceedings of the 9th Workshop on Hot
Topics in Operating Systems, Lihue (Kauai), Hawaii, USA, May 2003.
[68] M. Welsh and G. Mainland. Programming Sensor Networks Using Ab-
stract Regions. In Proceedings of ACM/Usenix Networked Systems De-
sign and Implementation (NSDI’04), San Francisco, California, USA,
March 2004.
[69] M. Welsh, D. Myung, M. Gaynor, and S. Moulton. Resuscitation moni-
toring with a wireless sensor network. In Circulation 108:1037: Journal
of the American Heart Association, Resuscitation Science Symposium.,
October 2003.
[70] G. Werner-Allen, K. Lorincz, J. Johnson, J. Lees, and M. Welsh. Fidelity
and yield in a volcano monitoring sensor network. In Proceedings of the
7th USENIX Symposium on Operating Systems Design and Implementa-
tion 2006, Seattle, November 2006.
[71] K. Whitehouse, C. Sharp, E. Brewer, and D. Culler. Hood: a neighbor-
hood abstraction for sensor networks. In Proceedings of the 2nd interna-
tional conference on Mobile systems, applications, and services (MobiSys
’04), Boston, MA, USA, June 2004.
[72] Q. Xie, J. Liu, and P. H. Chou. Tapper: a lightweight scripting engine
for highly constrained wireless sensor nodes. In Proceedings of the fifth
international conference on Information processing in sensor networks
(IPSN ’06), Poster session, Nashville, Tennessee, USA, 2006.
II

Papers

43
Chapter 7

Paper A:
Full TCP/IP for 8-Bit
Architectures

Adam Dunkels. Full TCP/IP for 8-bit architectures. In Proceedings of The

First International Conference on Mobile Systems, Applications, and Services
(ACM MobiSys 2003), San Francisco, USA, May 2003.
c 2003 Usenix Association.

45
Abstract

We describe two small and portable TCP/IP implementations fulfilling the

subset of RFC1122 requirements needed for full host-to-host interoperability.
Our TCP/IP implementations do not sacrifice any of TCP’s mechanisms such
as urgent data or congestion control. They support IP fragment reassembly
and the number of multiple simultaneous connections is limited only by the
available RAM. Despite being small and simple, our implementations do not
require their peers to have complex, full-size stacks, but can communicate with
peers running a similarly light-weight stack. The code size is on the order of
10 kilobytes and RAM usage can be configured to be as low as a few hundred
bytes.
7.1 Introduction 47

7.1 Introduction
With the success of the Internet, the TCP/IP protocol suite has become a global
standard for communication. TCP/IP is the underlying protocol used for web
page transfers, e-mail transmissions, file transfers, and peer-to-peer network-
ing over the Internet. For embedded systems, being able to run native TCP/IP
makes it possible to connect the system directly to an intranet or even the global
Internet. Embedded devices with full TCP/IP support will be first-class net-
work citizens, thus being able to fully communicate with other hosts in the
network.
Traditional TCP/IP implementations have required far too much resources
both in terms of code size and memory usage to be useful in small 8 or 16-
bit systems. Code size of a few hundred kilobytes and RAM requirements of
several hundreds of kilobytes have made it impossible to fit the full TCP/IP
stack into systems with a few tens of kilobytes of RAM and room for less than
100 kilobytes of code.
TCP [21] is both the most complex and the most widely used of the trans-
port protocols in the TCP/IP stack. TCP provides reliable full-duplex byte
stream transmission on top of the best-effort IP [20] layer. Because IP may
reorder or drop packets between the sender and the receiver, TCP has to im-
plement sequence numbering and retransmissions in order to achieve reliable,
ordered data transfer.
We have implemented two small generic and portable TCP/IP implemen-
tations, lwIP (lightweight IP) and uIP (micro IP), with slightly different design
goals. The lwIP implementation is a full-scale but simplified TCP/IP imple-
mentation that includes implementations of IP, ICMP, UDP and TCP and is
modular enough to be easily extended with additional protocols. lwIP has sup-
port for multiple local network interfaces and has flexible configuration options
which makes it suitable for a wide variety of devices.
The uIP implementation is designed to have only the absolute minimal set
of features needed for a full TCP/IP stack. It can only handle a single network
interface and does not implement UDP, but focuses on the IP, ICMP and TCP
protocols.
Both implementations are fully written in the C programming language.
We have made the source code available for both lwIP [7] and uIP [8]. Our
implementations have been ported to numerous 8- and 16-bit platforms such
as the AVR, H8S/300, 8051, Z80, ARM, M16c, and the x86 CPUs. Devices
running our implementations have been used in numerous places throughout
the Internet.
48 Paper A

We have studied how the code size and RAM usage of a TCP/IP implemen-
tation affect the features of the TCP/IP implementation and the performance of
the communication. We have limited our work to studying the implementa-
tion of TCP and IP protocols and the interaction between the TCP/IP stack and
the application programs. Aspects such as address configuration, security, and
energy consumption are out of the scope of this work.
The main contribution of our work is that we have shown that is it possible
to implement a full TCP/IP stack that is small enough in terms of code size and
memory usage to be useful even in limited 8-bit systems.
Recently, other small implementations of the TCP/IP stack have made it
possible to run TCP/IP in small 8-bit systems. Those implementations are of-
ten heavily specialized for a particular application, usually an embedded web
server, and are not suited for handling generic TCP/IP protocols. Future em-
bedded networking applications such as peer-to-peer networking require that
the embedded devices are able to act as first-class network citizens and run a
TCP/IP implementation that is not tailored for any specific application.
Furthermore, existing TCP/IP implementations for small systems assume
that the embedded device always will communicate with a full-scale TCP/IP
implementation running on a workstation-class machine. Under this assump-
tion, it is possible to remove certain TCP/IP mechanisms that are very rarely
used in such situations. Many of those mechanisms are essential, however, if
the embedded device is to communicate with another equally limited device,
e.g., when running distributed peer-to-peer services and protocols.
This paper is organized as follows. After a short introduction to TCP/IP
in Section 7.2, related work is presented in Section 7.3. Section 7.4 discusses
RFC standards compliance. How memory and buffer management is done in
our implementations is presented in Section 7.5 and the application program
interface is discussed in Section 7.6. Details of the protocol implementations
is given in Section 7.7 and Section 7.8 comments on the performance and max-
imum throughput of our implementations, presents throughput measurements
from experiments and reports on the code size of our implementations. Section
7.9 gives ideas for future work. Finally, the paper is summarized and concluded
in Section 7.10.

7.2 TCP/IP overview

From a high level viewpoint, the TCP/IP stack can be seen as a black box that
takes incoming packets, and demultiplexes them between the currently active
7.2 TCP/IP overview 49

Application
data
Web server application

Web server application

Incoming
Network TCP/IP Web server application
interface packets stack

Mail sender application

Data logger application

Figure 7.1: TCP/IP input processing.

connections. Before the data is delivered to the application, TCP sorts the
packets so that they appear in the order they were sent. The TCP/IP stack will
also send acknowledgments for the received packets.
Figure 7.1 shows how packets come from the network device, pass through
the TCP/IP stack, and are delivered to the actual applications. In this example
there are five active connections, three that are handled by a web server ap-
plication, one that is handled by the e-mail sender application, and one that is
handled by a data logger application.

Web server application Application

data

Web server application

Outgoing
Web server application TCP/IP Network
stack packets interface

Mail sender application

Data logger application

Figure 7.2: TCP/IP output processing.

A high level view of the output processing can be seen in Figure 7.2. The
TCP/IP stack collects the data sent by the applications before it is actually
sent onto the network. TCP has mechanisms for limiting the amount of data
that is sent over the network, and each connection has a queue on which the
data is held while waiting to be transmitted. The data is not removed from
the queue until the receiver has acknowledged the reception of the data. If no
acknowledgment is received within a specific time, the data is retransmitted.
Data arrives asynchronously from both the network and the application,
50 Paper A

and the TCP/IP stack maintains queues in which packets are kept waiting for
service. Because packets might be dropped or reordered by the network, in-
coming packets may arrive out of order. Such packets have to be queued by the
TCP/IP stack until a packet that fills the gap arrives. Furthermore, because TCP
limits the rate at which data that can be transmitted over each TCP connection,
application data might not be immediately sent out onto the network.
The full TCP/IP suite consists of numerous protocols, ranging from low
level protocols such as ARP which translates IP addresses to MAC addresses,
to application level protocols such as SMTP that is used to transfer e-mail.
We have concentrated our work on the TCP and IP protocols and will refer
to upper layer protocols as “the application”. Lower layer protocols are often
implemented in hardware or firmware and will be referred to as “the network
device” that are controlled by the network device driver.
TCP provides a reliable byte stream to the upper layer protocols. It breaks
the byte stream into appropriately sized segments and each segment is sent in
its own IP packet. The IP packets are sent out on the network by the network
device driver. If the destination is not on the physically connected network,
the IP packet is forwarded onto another network by a router that is situated
between the two networks. If the maximum packet size of the other network
is smaller than the size of the IP packet, the packet is fragmented into smaller
packets by the router. If possible, the size of the TCP segments are chosen so
that fragmentation is minimized. The final recipient of the packet will have
to reassemble any fragmented IP packets before they can be passed to higher
layers.

7.3 Related work

There are numerous small TCP/IP implementations for embedded systems.
The target architectures range from small 8-bit microcontrollers to 32-bit RISC
architectures. Code size varies from a few kilobytes to hundreds of kilobytes.
RAM requirements can be as low as 10 bytes up to several megabytes.
Existing TCP/IP implementations can roughly be divided into two cate-
gories; those that are adaptations of the Berkeley BSD TCP/IP implementa-
tion [18], and those that are written independently from the BSD code. The
BSD implementation was originally written for workstation-class machines
and was not designed for the limitations of small embedded systems. Be-
cause of that, implementations that are derived from the BSD code base are
usually suited for larger architectures than our target. An example of a BSD-
7.3 Related work 51

derived implementation is the InterNiche NicheStack [11], which needs around

50 kilobytes of code space on a 32-bit ARM system.
Many of the independent TCP/IP implementations for embedded proces-
sors use a simplified model of the TCP/IP stack which makes several assump-
tions about the communication environment. The most common assumption is
that the embedded system always will communicate with a system such as a PC
that runs a full scale, standards compliant TCP/IP implementation. By relying
on the standards compliance of the remote host, even an extremely simplified,
uncompliant, TCP/IP implementation will be able to communicate. The com-
munication may very well fail, however, once the system is to communicate
with another simplified TCP/IP implementation such as another embedded sys-
tem of the same kind. We will briefly cover a number of such simplifications
that are used by existing implementations.
One usual simplification is to tailor the TCP/IP stack for a specific applica-
tion such as a web server. By doing this, only the parts of the TCP/IP protocols
that are required by the application need to be implemented. For instance, a
web server application does not need support for urgent data and does not need
to actively open TCP connections to other hosts. By removing those mecha-
nisms from the implementation, the complexity is reduced.
The smallest TCP/IP implementations in terms of RAM and code space re-
quirements are heavily specialized for serving web pages and use an approach
where the web server does not hold any connection state at all. For example, the
iPic match-head sized server [26] and Jeremy Bentham’s PICmicro stack [1]
require only a few tens of bytes of RAM to serve simple web pages. In such
an implementation, retransmissions cannot be made by the TCP module in the
embedded system because nothing is known about the active connections. In
order to achieve reliable transfers, the system has to rely on the remote host
to perform retransmissions. It is possible to run a very simple web server with
such an implementation, but there are serious limitations such as not being able
to serve web pages that are larger than the size of a single TCP segment, which
typically is about one kilobyte.
Other TCP/IP implementations such as the Atmel TCP/IP stack [5] save
code space by leaving out certain vital TCP mechanisms. In particular, they
often leave out TCP’s congestion control mechanisms, which are used to re-
duce the sending rate when the network is overloaded. While an implementa-
tion with no congestion control might work well when connected to a single
Ethernet segment, problems can arise when communication spans several net-
works. In such cases, the intermediate nodes such as switches and routers may
be overloaded. Because congestion primarily is caused by the amount of pack-
52 Paper A

ets in the network, and not the size of these packets, even small 8-bit systems
are able to produce enough traffic to cause congestion. A TCP/IP implementa-
tion lacking congestion control mechanisms should not be used over the global
Internet as it might contribute to congestion collapse [9].
Texas Instrument’s MSP430 TCP/IP stack [6] and the TinyTCP code [4]
use another common simplification in that they can handle only one TCP con-
nection at a time. While this is a sensible simplification for many applications,
it seriously limits the usefulness of the TCP/IP implementation. For example,
it is not possible to communicate with two simultaneous peers with such an
implementation. The CMX Micronet stack [27] uses a similar simplification in
that it sets a hard limit of 16 on the maximum number of connections.
Yet another simplification that is used by LiveDevices Embedinet imple-
mentation [12] and others is to disregard the maximum segment size that a
receiver is prepared to handle. Instead, the implementation will send segments
that fit into an Ethernet frame of 1500 bytes. This works in a lot of cases
due to the fact that many hosts are able to receive packets that are 1500 bytes
or larger. Communication will fail, however, if the receiver is a system with
limited memory resources that is not able to handle packets of that size.
Finally, the most common simplification is to leave out support for re-
assembling fragmented IP packets. Even though fragmented IP packets are
quite infrequent [25], there are situations in which they may occur. If packets
travel over a path which fragments the packets, communication is impossible
if the TCP/IP implementation is unable to correctly reassemble them. TCP/IP
implementations that are able to correctly reassemble fragmented IP packets,
such as the Kadak KwikNET stack [22], are usually too large in terms of code
size and RAM requirements to be practical for 8-bit systems.

7.4 RFC-compliance
The formal requirements for the protocols in the TCP/IP stack is specified in a
number of RFC documents published by the Internet Engineering Task Force,
IETF. Each of the protocols in the stack is defined in one more RFC documents
and RFC1122 [2] collects all requirements and updates the previous RFCs.
The RFC1122 requirements can be divided into two categories; those that
deal with the host to host communication and those that deal with communica-
tion between the application and the networking stack. An example of the first
kind is “A TCP MUST be able to receive a TCP option in any segment” and an
example of the second kind is “There MUST be a mechanism for reporting soft
7.5 Memory and buffer management 53

Table 7.1: TCP/IP features implemented by uIP and lwIP

Feature uIP lwIP
IP and TCP checksums x x
IP fragment reassembly x x
IP options
Multiple interfaces x
UDP x
Multiple TCP connections x x
TCP options x x
Variable TCP MSS x x
RTT estimation x x
TCP flow control x x
Sliding TCP window x
TCP congestion control Not needed x
Out-of-sequence TCP data x
TCP urgent data x x
Data buffered for rexmit x

TCP error conditions to the application.” A TCP/IP implementation that vio-

lates requirements of the first kind may not be able to communicate with other
TCP/IP implementations and may even lead to network failures. Violation of
the second kind of requirements will only affect the communication within the
system and will not affect host-to-host communication.
In our implementations, we have implemented all RFC requirements that
affect host-to-host communication. However, in order to reduce code size,
we have removed certain mechanisms in the interface between the application
and the stack, such as the soft error reporting mechanism and dynamically
configurable type-of-service bits for TCP connections. Since there are only
very few applications that make use of those features, we believe that they can
be removed without loss of generality. Table 7.1 lists the features that uIP and
lwIP implements.

7.5 Memory and buffer management

In our target architecture, RAM is the most scarce resource. With only a few
kilobytes of RAM available for the TCP/IP stack to use, mechanisms used in
54 Paper A

traditional TCP/IP cannot be directly applied.

Because of the different design goals for the lwIP and the uIP implemen-
tations, we have chosen two different memory management solutions. The
lwIP implementation has dynamic buffer and memory allocation mechanisms
where memory for holding connection state and packets is dynamically allo-
cated from a global pool of available memory blocks. Packets are contained in
one or more dynamically allocated buffers of fixed size. The size of the packet
buffers is determined by a configuration option at compile time. Buffers are
allocated by the network device driver when an incoming packet arrives. If the
packet is larger than one buffer, more buffers are allocated and the packet is
split into the buffers. If the incoming packet is queued by higher layers of the
stack or the application, a reference counter in the buffer is incremented. The
buffer will not be deallocated until the reference count is zero.
The uIP stack does not use explicit dynamic memory allocation. Instead, it
uses a single global buffer for holding packets and has a fixed table for hold-
ing connection state. The global packet buffer is large enough to contain one
packet of maximum size. When a packet arrives from the network, the device
driver places it in the global buffer and calls the TCP/IP stack. If the packet
contains data, the TCP/IP stack will notify the corresponding application. Be-
cause the data in the buffer will be overwritten by the next incoming packet, the
application will either have to act immediately on the data or copy the data into
a secondary buffer for later processing. The packet buffer will not be overwrit-
ten by new packets before the application has processed the data. Packets that
arrive when the application is processing the data must be queued, either by the
network device or by the device driver. Most single-chip Ethernet controllers
have on-chip buffers that are large enough to contain at least 4 maximum sized
Ethernet frames. Devices that are handled by the processor, such as RS-232
ports, can copy incoming bytes to a separate buffer during application process-
ing. If the buffers are full, the incoming packet is dropped. This will cause
performance degradation, but only when multiple connections are running in
parallel. This is because uIP advertises a very small receiver window, which
means that only a single TCP segment will be in the network per connection.
Outgoing data is also handled differently because of the different buffer
schemes. In lwIP, an application that wishes to send data passes the length
and a pointer to the data to the TCP/IP stack as well as a flag which indi-
cates whether the data is volatile or not. The TCP/IP stack allocates buffers
of suitable size and, depending on the volatile flag, either copies the data into
the buffers or references the data through pointers. The allocated buffers con-
tain space for the TCP/IP stack to prepend the TCP/IP and link layer headers.
7.6 Application program interface 55

After the headers are written, the stack passes the buffers to the network de-
vice driver. The buffers are not deallocated when the device driver is finished
sending the data, but held on a retransmission queue. If the data is lost in the
network and have to be retransmitted, the buffers on retransmission queue will
be retransmitted. The buffers are not deallocated until the data is known to be
received by the peer. If the connection is aborted because of an explicit request
from the local application or a reset segment from the peer, the connection’s
buffers are deallocated.
In uIP, the same global packet buffer that is used for incoming packets is
also used for the TCP/IP headers of outgoing data. If the application sends
dynamic data, it may use the parts of the global packet buffer that are not used
for headers as a temporary storage buffer. To send the data, the application
passes a pointer to the data as well as the length of the data to the stack. The
TCP/IP headers are written into the global buffer and once the headers have
been produced, the device driver sends the headers and the application data
out on the network. The data is not queued for retransmissions. Instead, the
application will have to reproduce the data if a retransmission is necessary.
The total amount of memory usage for our implementations depends heav-
ily on the applications of the particular device in which the implementations
are to be run. The memory configuration determines both the amount of traffic
the system should be able to handle and the maximum amount of simultaneous
connections. A device that will be sending large e-mails while at the same time
running a web server with highly dynamic web pages and multiple simultane-
ous clients, will require more RAM than a simple Telnet server. It is possible
to run the uIP implementation with as little as 200 bytes of RAM, but such
a configuration will provide extremely low throughput and will only allow a
small number of simultaneous connections.

7.6 Application program interface

The Application Program Interface (API) defines the way the application pro-
gram interacts with the TCP/IP stack. The most commonly used API for
TCP/IP is the BSD socket API which is used in most Unix systems and has
heavily influenced the Microsoft Windows WinSock API. Because the socket
API uses stop-and-wait semantics, it requires support from an underlying mul-
titasking operating system. Since the overhead of task management, context
switching and allocation of stack space for the tasks might be too high in our
target architecture, the BSD socket interface is not suitable for our purposes.
56 Paper A

Instead, we have chosen an event driven interface where the application is

invoked in response to certain events. Examples of such events are data arriving
on a connection, an incoming connection request, or a poll request from the
stack. The event based interface fits well in the event based structure used by
operating systems such as TinyOS [10]. Furthermore, because the application
is able to act on incoming data and connection requests as soon as the TCP/IP
stack receives the packet, low response times can be achieved even in low-end
systems.

7.7 Protocol implementations

The protocols in the TCP/IP protocol suite are designed in a layered fashion
where each protocol performs a specific function and the interactions between
the protocol layers are strictly defined. While the layered approach is a good
way to design protocols, it is not always the best way to implement them. For
the lwIP implementation, we have chosen a fully modular approach where each
protocol implementation is kept fairly separate from the others. In the smaller
uIP implementation, the protocol implementations are tightly coupled in order
to save code space.

7.7.1 Main control loop

The lwIP and uIP stacks can be run either as a task in a multitasking system, or
as the main program in a singletasking system. In both cases, the main control
loop (Figure 7.3) does two things repeatedly:

1. Check if a packet has arrived from the network.

2. Check if a periodic timeout has occurred.

If a packet has arrived, the input handler of the TCP/IP stack is invoked. The
input handler function will never block, but will return at once. When it returns,
the stack or the application for which the incoming packet was intended may
have produced one or more reply packets which should be sent out. If so, the
network device driver is called to send out these packets.
Periodic timeouts are used to drive TCP mechanisms that depend on timers,
such as delayed acknowledgments, retransmissions and round-trip time estima-
tions. When the main control loop infers that the periodic timer should fire, it
invokes the timer handler of the TCP/IP stack. Because the TCP/IP stack may
7.7 Protocol implementations 57

Check for packet Process packet

Application events

Output packets

Check for timeout Process timeout

Application events

Output packets

Figure 7.3: The main control loop.

perform retransmissions when dealing with a timer event, the network device
driver is called to send out the packets that may have been produced.
This is similar to how the BSD implementations drive the TCP/IP stack,
but BSD uses software interrupts and a task scheduler to initiate input handlers
and timers. In our limited system, we do not depend on such mechanisms being
available.

7.7.2 IP — Internet Protocol

When incoming packets are processed by lwIP and uIP, the IP layer is the first
protocol that examines the packet. The IP layer does a few simple checks such
as if the destination IP address of the incoming packet matches any of the local
IP address and verifies the IP header checksum. Since there are no IP options
that are strictly required and because they are very uncommon, both lwIP and
uIP drop any IP options in received packets.
58 Paper A

IP fragment reassembly
In both lwIP and uIP, IP fragment reassembly is implemented using a sepa-
rate buffer that holds the packet to be reassembled. An incoming fragment is
copied into the right place in the buffer and a bit map is used to keep track
of which fragments have been received. Because the first byte of an IP frag-
ment is aligned on an 8-byte boundary, the bit map requires a small amount of
memory. When all fragments have been reassembled, the resulting IP packet
is passed to the transport layer. If all fragments have not been received within
a specified time frame, the packet is dropped.
The current implementation only has a single buffer for holding packets
to be reassembled, and therefore does not support simultaneous reassembly of
more than one packet. Since fragmented packets are uncommon, we belive this
to be a reasonable decision. Extending our implementation to support multiple
buffers would be straightforward, however.

Broadcasts and multicasts

IP has the ability to broadcast and multicast packets on the local network. Such
packets are addressed to special broadcast and multicast addresses. Broadcast
is used heavily in many UDP based protocols such as the Microsoft Windows
file-sharing SMB protocol. Multicast is primarily used in protocols used for
multimedia distribution such as RTP. TCP is a point-to-point protocol and does
not use broadcast or multicast packets.
Because lwIP supports applications using UDP, it has support for both
sending and receiving broadcast and multicast packets. In contrast, uIP does
not have UDP support and therefore handling of such packets has not been
implemented.

7.7.3 ICMP — Internet Control Message Protocol

The ICMP protocol is used for reporting soft error conditions and for querying
host parameters. Its main use is, however, the echo mechanism which is used
by the ping program.
The ICMP implementations in lwIP and uIP are very simple as we have
restricted them to only implement ICMP echo messages. Replies to echo mes-
sages are constructed by simply swapping the source and destination IP ad-
dresses of incoming echo requests and rewriting the ICMP header with the
Echo-Reply message type. The ICMP checksum is adjusted using standard
techniques [23].
7.7 Protocol implementations 59

Since only the ICMP echo message is implemented, there is no support for
Path MTU discovery or ICMP redirect messages. Neither of these is strictly
required for interoperability; they are performance enhancement mechanisms.

7.7.4 TCP — Transmission Control Protocol

The TCP implementations in lwIP and uIP are driven by incoming packets and
timer events. IP calls TCP when a TCP packet arrives and the main control
loop calls TCP periodically.
Incoming packets are parsed by TCP and if the packet contains data that is
to be delivered to the application, the application is invoked by the means of
a function call. If the incoming packet acknowledges previously sent data, the
connection state is updated and the application is informed, allowing it to send
out new data.

Listening connections
TCP allows a connection to listen for incoming connection requests. In our
implementations, a listening connection is identified by the 16-bit port number
and incoming connection requests are checked against the list of listening con-
nections. This list of listening connections is dynamic and can be altered by
the applications in the system.

Sending data
When sending data, an application will have to check the number of available
bytes in the send window and adjust the number of bytes to send accordingly.
The size of the send window is dictated by the memory configuration as well
as the buffer space announced by the receiver of the data. If no buffer space is
available, the application has to defer the send and wait until later.
Buffer space becomes available when an acknowledgment from the re-
ceiver of the data has been received. The stack informs the application of this
event, and the application may then repeat the sending procedure.

Sliding window
Most TCP implementations use a sliding window mechanism for sending data.
Multiple data segments are sent in succession without waiting for an acknowl-
edgment for each segment.
60 Paper A

The sliding window algorithm uses a lot of 32-bit operations and because
32-bit arithmetic is fairly expensive on most 8-bit CPUs, uIP does not im-
plement it. Also, uIP does not buffer sent packets and a sliding window im-
plementation that does not buffer sent packets will have to be supported by a
complex application layer. Instead, uIP allows only a single TCP segment per
connection to be unacknowledged at any given time. lwIP, on the other hand,
implements TCP’s sliding window mechanism using output buffer queues and
therefore does not add additional complexity to the application layer.
It is important to note that even though most TCP implementations use
the sliding window algorithm, it is not required by the TCP specifications.
Removing the sliding window mechanism does not affect interoperability in
any way.

Round-trip time estimation

TCP continuously estimates the current Round-Trip Time (RTT) of every active
connection in order to find a suitable value for the retransmission time-out.
We have implemented the RTT estimation using TCP’s periodic timer.
Each time the periodic timer fires, it increments a counter for each connec-
tion that has unacknowledged data in the network. When an acknowledgment
is received, the current value of the counter is used as a sample of the RTT. The
sample is used together with the standard TCP RTT estimation function [13] to
calculate an estimate of the RTT. Karn’s algorithm [14] is used to ensure that
retransmissions do not skew the estimates.

Retransmissions
Retransmissions are driven by the periodic TCP timer. Every time the periodic
timer is invoked, the retransmission timer for each connection is decremented.
If the timer reaches zero, a retransmission should be made.
The actual retransmission operation is handled differently in uIP and in
lwIP. lwIP maintains two output queues: one holds segments that have not
yet been sent, the other holds segments that have been sent but not yet been
acknowledged by the peer. When a retransmission is required, the first segment
on the queue of segments that has not been acknowledged is sent. All other
segments in the queue are moved to the queue with unsent segments.
As uIP does not keep track of packet contents after they have been sent
by the device driver, uIP requires that the application takes an active part in
performing the retransmission. When uIP decides that a segment should be re-
7.7 Protocol implementations 61

transmitted, it calls the application with a flag set indicating that a retransmis-
sion is required. The application checks the retransmission flag and produces
the same data that was previously sent. From the application’s standpoint, per-
forming a retransmission is not different from how the data originally was sent.
Therefore the application can be written in such a way that the same code is
used both for sending data and retransmitting data. Also, it is important to note
that even though the actual retransmission operation is carried out by the ap-
plication, it is the responsibility of the stack to know when the retransmission
should be made. Thus the complexity of the application does not necessarily
increase because it takes an active part in doing retransmissions.

Flow control

The purpose of TCP’s flow control mechanisms is to allow communication

between hosts with wildly varying memory dimensions. In each TCP segment,
the sender of the segment indicates its available buffer space. A TCP sender
must not send more data than the buffer space indicated by the receiver.
In our implementations, the application cannot send more data than the
receiving host can buffer. Before sending data, the application checks how
many bytes it is allowed to send and does not send more data than the other
host can accept. If the remote host cannot accept any data at all, the stack
initiates the zero window probing mechanism.
The application is responsible for controlling the size of the window size
indicated in sent segments. If the application must wait or buffer data, it can
explicitly close the window so that the sender will not send data until the ap-
plication is able to handle it.

Congestion control

The congestion control mechanisms limit the number of simultaneous TCP

segments in the network. The algorithms used for congestion control [13] are
designed to be simple to implement and require only a few lines of code.
Since uIP only handles one in-flight TCP segment per connection, the
amount of simultaneous segments cannot be further limited, thus the conges-
tion control mechanisms are not needed. lwIP has the ability to have multiple
in-flight segments and therefore implements all of TCP’s congestion control
mechanisms.
62 Paper A

Urgent data
TCP’s urgent data mechanism provides an application-to-application notifica-
tion mechanism, which can be used by an application to mark parts of the data
stream as being more urgent than the normal stream. It is up to the receiving
application to interpret the meaning of the urgent data.
In many TCP implementations, including the BSD implementation, the ur-
gent data feature increases the complexity of the implementation because it
requires an asynchronous notification mechanism in an otherwise synchronous
API. As our implementations already use an asynchronous event based API,
the implementation of the urgent data feature does not lead to increased com-
plexity.

Connection state
Each TCP connection requires a certain amount of state information in the
embedded device. Because the state information uses RAM, we have aimed
towards minimizing the amount of state needed for each connection in our
implementations.
The uIP implementation, which does not use the sliding window mech-
anism, requires far less state information than the lwIP implementation. The
sliding window implementation requires that the connection state includes sev-
eral 32-bit sequence numbers, not only for keeping track of the current se-
quence numbers of the connection, but also for remembering the sequence
numbers of the last window updates. Furthermore, because lwIP is able to
handle multiple local IP addresses, the connection state must include the lo-
cal IP address. Finally, as lwIP maintains queues for outgoing segments, the
memory for the queues is included in the connection state. This makes the
state information needed for lwIP nearly 60 bytes larger than that of uIP which
requires 30 bytes per connection.

7.8 Results
7.8.1 Performance limits
In TCP/IP implementations for high-end systems, processing time is domi-
nated by the checksum calculation loop, the operation of copying packet data
and context switching [15]. Operating systems for high-end systems often have
multiple protection domains for protecting kernel data from user processes and
7.8 Results 63

user processes from each other. Because the TCP/IP stack is run in the ker-
nel, data has to be copied between the kernel space and the address space of
the user processes and a context switch has to be performed once the data has
been copied. Performance can be enhanced by combining the copy operation
with the checksum calculation [19]. Because high-end systems usually have
numerous active connections, packet demultiplexing is also an expensive oper-
ation [17].
A small embedded device does not have the necessary processing power to
have multiple protection domains and the power to run a multitasking operating
system. Therefore there is no need to copy data between the TCP/IP stack and
the application program. With an event based API there is no context switch
between the TCP/IP stack and the applications.
In such limited systems, the TCP/IP processing overhead is dominated by
the copying of packet data from the network device to host memory, and check-
sum calculation. Apart from the checksum calculation and copying, the TCP
processing done for an incoming packet involves only updating a few counters
and flags before handing the data over to the application. Thus an estimate
of the CPU overhead of our TCP/IP implementations can be obtained by cal-
culating the amount of CPU cycles needed for the checksum calculation and
copying of a maximum sized packet.

7.8.2 The impact of delayed acknowledgments

Most TCP receivers implement the delayed acknowledgment algorithm [3] for
reducing the number of pure acknowledgment packets sent. A TCP receiver
using this algorithm will only send acknowledgments for every other received
segment. If no segment is received within a specific time-frame, an acknowl-
edgment is sent. The time-frame can be as high as 500 ms but typically is 200
ms.
A TCP sender such as uIP that only handles a single outstanding TCP seg-
ment will interact poorly with the delayed acknowledgment algorithm. Be-
cause the receiver only receives a single segment at a time, it will wait as much
as 500 ms before an acknowledgment is sent. This means that the maximum
possible throughput is severely limited by the 500 ms idle time.
Thus the maximum throughput equation when sending data from uIP will
be p = s/(t+td) where s is the segment size and td is the delayed acknowledg-
ment timeout, which typically is between 200 and 500 ms. With a segment size
of 1000 bytes, a round-trip time of 40 ms and a delayed acknowledgment time-
out of 200 ms, the maximum throughput will be 4166 bytes per second. With
64 Paper A

the delayed acknowledgment algorithm disabled at the receiver, the maximum

throughput would be 25000 bytes per second.
It should be noted, however, that since small systems running uIP are not
very likely to have large amounts of data to send, the delayed acknowledgment
throughput degradation of uIP need not be very severe. Small amounts of data
sent by such a system will not span more than a single TCP segment, and would
therefore not be affected by the throughput degradation anyway.
The maximum throughput when uIP acts as a receiver is not affected by the
delayed acknowledgment throughput degradation.

7.8.3 Measurements
For our experiments we connected a 450 MHz Pentium III PC running
FreeBSD 4.7 to an Ethernut board [16] through a dedicated 10 megabit/second
Ethernet network. The Ethernut board is a commercially available embedded
system equipped with a RealTek RTL8019AS Ethernet controller, an Atmel
Atmega128 AVR microcontroller running at 14.7456 MHz with 128 kilobytes
of flash ROM for code storage and 32 kilobytes of RAM. The FreeBSD host
was configured to run the Dummynet delay emulator software [24] in order
to facilitate controlled delays for the communication between the PC and the
embedded system.
In the embedded system, a simple web server was run on top of the uIP and
lwIP stacks. Using the fetch file retrieval utility, a file consisting of null bytes
was downloaded ten times from the embedded system. The reported through-
put was logged, and the mean throughput of the ten downloads was calculated.
By redirecting file output to /dev/null, the file was immediately discarded
by the FreeBSD host. The file size was 200 kilobytes for the uIP tests, and 200
megabytes for the lwIP tests. The size of the file made it impossible to keep it
all in the memory of the embedded system. Instead, the file was generated by
the web server as it was sent out on the network.
The total TCP/IP memory consumption in the embedded system was var-
ied by changing the send window size. For uIP, the send window was varied
between 50 bytes and the maximum possible value of 1450 bytes in steps of
50 bytes. The send window configuration translates into a total RAM usage of
between 400 bytes and 3 kilobytes. The lwIP send window was varied between
500 and 11000 bytes in steps of 500 bytes, leading to a total RAM consumption
of between 5 and 16 kilobytes.
Figure 7.4 shows the mean throughput of the ten file downloads from the
web server running on top of uIP, with an additional 10 ms delay created by
7.8 Results 65

500000
Throughput with delayed ACKs disabled
Throughput with delayed ACKs enabled
450000

400000

350000
Throughput (bytes/second)

300000

250000

200000

150000

100000

50000

0
0 200 400 600 800 1000 1200 1400 1600
Send window (bytes)

Figure 7.4: uIP sending data with 10 ms emulated delay.

the Dummynet delay emulator. The two curves show the measured through-
put with the delayed acknowledgment algorithm disabled and enabled at the
receiving FreeBSD host, respectively. The performance degradation caused by
the delayed acknowledgments is evident.
Figure 7.5 shows the same setup, but without the 10 ms emulated delay.
The lower curve, showing the throughput with delayed acknowledgments en-
abled, is very similar to the lower one in Figure 7.4. The upper curve, however,
does not show the same linear relation as the previous figure, but shows an
increasing throughput where the increase declines with increasing send win-
dow size. One explanation for the declining increase of throughput is that the
round-trip time increases with the send window size because of the increased
per-packet processing time. Figure 7.6 shows the round-trip time as a func-
tion of packet size. These measurements were taken using the ping program
and therefore include the cost for the packet copying operation twice; once for
packet input and once for packet output.
The throughput of lwIP shows slightly different characteristics. Figure 7.7
shows three measured throughput curves, without emulated delay, and with
emulated delays of 10 ms and 20 ms. For all measurements, the delayed ac-
knowledgment algorithm is enabled at the FreeBSD receiver. We see that for
small send window sizes, lwIP also suffers from the delayed acknowledgment
throughput degradation. With a send window larger than two maximum TCP
66 Paper A

500000
Throughput with delayed ACKs disabled
Throughput with delayed ACKs enabled
450000

400000

350000
Throughput (bytes/second)

300000

250000

200000

150000

100000

50000

0
0 200 400 600 800 1000 1200 1400 1600
Send window (bytes)

Figure 7.5: uIP sending data without emulated delay.

segment sizes (3000 bytes), lwIP is able to send out two TCP segments per
round-trip time and thereby avoids the delayed acknowledgments throughput
degradation. Without emulated delay, the throughput quickly reaches a maxi-
mum of about 415 kilobytes per second. This limit is likely to be the processing
limit of the lwIP code in the embedded system and therefore is the maximum
possible throughput for lwIP in this particular system.
The maximum throughput with emulated delays is lower than without delay
emulation, and the similarity of the two curves suggests that the throughput
degradation could be caused by interaction with the Dummynet software.

7.8.4 Code size

The code was compiled for the 32-bit Intel x86 and the 8-bit Atmel AVR plat-
forms using gcc [28] versions 2.95.3 and 3.3 respectively, with code size op-
timization turned on. The resulting size of the compiled code can be seen in
Tables 7.2 to 7.5. Even though both implementations support ARP and SLIP
and lwIP includes UDP, only the protocols discussed in this paper are pre-
sented. Because the protocol implementations in uIP are tightly coupled, the
individual sizes of the implementations are not reported.
There are several reasons for the dramatic difference in code size between
lwIP and uIP. In order to support the more complex and configurable TCP im-
7.8 Results 67

7
Round-trip time

4
RTT (ms)

0
0 200 400 600 800 1000 1200 1400 1600
Packet size (bytes)

Figure 7.6: Round-trip time as a function of packet size.

Table 7.2: Code size for uIP (x86)

Function Code size (bytes)
Checksumming 464
IP, ICMP and TCP 4724
Total 5188

plementation, lwIP has significantly more complex buffer and memory man-
agement than uIP. Since lwIP can handle packets that span several buffers, the
checksum calculation functions in lwIP are more complex than those in uIP.
The support for dynamically changing network interfaces in lwIP also con-
tributes to the size increase of the IP layer because the IP layer has to manage
multiple local IP addresses. The IP layer in lwIP is further made larger by
the fact that lwIP has support for UDP, which requires that the IP layer is able
handle broadcast and multicast packets. Likewise, the ICMP implementation
in lwIP has support for UDP error messages which have not been implemented
in uIP.
The TCP implementation is lwIP is nearly twice as large as the full IP,
ICMP and TCP implementation in uIP. The main reason for this is that lwIP
implements the sliding window mechanism which requires a large amount of
68 Paper A

500000
Throughput without emulated delay
Throughput with 10 ms emulated delay
450000 Throughput with 20 ms emulated delay

400000

350000
Throughput (bytes/second)

300000

250000

200000

150000

100000

50000

0
2000 4000 6000 8000 10000
Send window (bytes)

Figure 7.7: lwIP sending data with and without emulated delays.

Table 7.3: Code size for uIP (AVR)

Function Code size (bytes)
Checksumming 712
IP, ICMP and TCP 4452
Total 5164

buffer and queue management functionality that is not required in uIP.

The different memory and buffer management schemes used by lwIP and
uIP have implications on code size, mainly in 8-bit systems. Because uIP uses
a global buffer for all incoming packets, the absolute memory addresses of the
protocol header fields are known at compile time. Using this information, the
compiler is able to generate code that uses absolute addressing, which on many
8-bit processors requires less code than indirect addressing.
Is it interesting to note that the size of the compiled lwIP code is larger on
the AVR than on the x86, while the uIP code is of about the same size on the
two platforms. The main reason for this is that lwIP uses 32-bit arithmetic to a
much larger degree than uIP and each 32-bit operation is compiled into a large
number of machine code instructions.
7.9 Future work 69

Table 7.4: Code size for lwIP (x86)

Function Code size (bytes)
Memory management 2512
Checksumming 504
Network interfaces 364
IP 1624
ICMP 392
TCP 9192
Total 14588

Table 7.5: Code size for lwIP (AVR)

Function Code size (bytes)
Memory management 3142
Checksumming 1116
Network interfaces 458
IP 2216
ICMP 594
TCP 14230
Total 21756

7.9 Future work

Prioritized connections. It is advantageous to be able to prioritize certain con-
nections such as Telnet connections for manual configuration of the device.
Even in a system that is under heavy load from numerous clients, it should
be possible to remotely control and configure the device. In order to do pro-
vide this, different connection types could be given different priority. For effi-
ciency, such differentiation should be done as far down in the system as possi-
ble, preferably in the device driver.
Security aspects. When connecting systems to a network, or even to the
global Internet, the security of the system is very important. Identifying levels
of security and mechanisms for implementing security for embedded devices
is crucial for connecting systems to the global Internet.
Address auto-configuration. If hundreds or even thousands of small em-
70 Paper A

bedded devices should be deployed, auto-configuration of IP addresses is ad-

vantageous. Such mechanisms already exist in IPv6, the next version of the
Internet Protocol, and are currently being standardized for IPv4.
Improving throughput. The throughput degradation problem caused by the
poor interaction with the delayed acknowledgment algorithm should be fixed.
By increasing the maximum number of in-flight segments from one to two, the
problem will not appear. When increasing the amount of in-flight segments,
congestion control mechanisms will have to be employed. Those mechanisms
are trivial, however, when the upper limit is two simultaneous segments.
Performance enhancing proxy. It might be possible to increase the per-
formance of communication with the embedded devices through the use of a
proxy situated near the devices. Such a proxy would have more memory than
the devices and could assume responsibility for buffering data.

7.10 Summary and conclusions

We have shown that it is possible to fit a full scale TCP/IP implementation
well within the limits of an 8-bit microcontroller, but that the throughput of
such a small implementation will suffer. We have not removed any TCP/IP
mechanisms in our implementations, but have full support for reassembly of
IP fragments and urgent TCP data. Instead, we have minimized the interface
between the TCP/IP stack and the application.
The maximum achievable throughput for our implementations is deter-
mined by the send window size that the TCP/IP stack has been configured to
use. When sending data with uIP, the delayed ACK mechanism at the receiver
lowers the maximum achievable throughput considerably. In many situations
however, a limited system running uIP will not produce so much data that this
will cause problems. lwIP is not affected by the delayed ACK throughput
degradation when using a large enough send window.

7.11 Acknowledgments
Many thanks go to Martin Nilsson, who has provided encouragement and been
a source of inspiration throughout the preparation of this paper. Thanks also go
to Deborah Wallach for comments and suggestions, the anonymous reviewers
whose comments were highly appreciated, and to all who have contributed
bugfixes, patches and suggestions to the lwIP and uIP implementations.
Bibliography

[1] J. Bentham. TCP/IP Lean: Web servers for embedded systems. CMP
Books, October 2000.

[2] R. Braden. Requirements for internet hosts – communication layers. RFC

1122, Internet Engineering Task Force, October 1989.

[3] D. D. Clark. Window and acknowledgement strategy in TCP. RFC 813,

Internet Engineering Task Force, July 1982.

[4] G. H. Cooper. TinyTCP. Web page. 2002-10-14.

URL: http://www.csonline.net/bpaddock/tinytcp/

[5] Atmel Corporation. Embedded web server. AVR 460, January 2001.
Avalible from www.atmel.com.

[6] A. Dannenberg. MSP430 internet connectivity. SLAA 137, November

2001. Avalible from www.ti.com.

[7] A. Dunkels. lwIP - a lightweight TCP/IP stack. Web page. 2002-10-14.

URL: http://www.sics.se/˜adam/lwip/

[8] A. Dunkels. uIP - a TCP/IP stack for 8- and 16-bit microcontrollers. Web
page. 2003-10-21.
URL: http://dunkels.com/adam/uip/

[9] S. Floyd and K. Fall. Promoting the use of end-to-end congestion control
in the internet. IEEE/ACM Transactions on Networking, August 1999.

[10] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler, and K. Pister. Sys-

tem architecture directions for networked sensors. In Proceedings of the
9th International Conference on Architectural Support for Programming

71
72 Bibliography

Languages and Operating Systems, Cambridge, Massachusets, USA,

November 2000.
[11] InterNiche Technologies Inc. NicheStack portable TCP/IP stack. Web
page. 2002-10-14.
URL: http://www.iniche.com/products/tcpip.htm

[12] LiveDevices Inc. Embedinet - embedded internet software products. Web

page. 2002-10-14.
URL: http://www.livedevices.com/net products/embedinet.shtml

[13] V. Jacobson. Congestion avoidance and control. In Proceedings of the

SIGCOMM ’88 Conference, Stanford, California, August 1988.
[14] P. Karn and C. Partridge. Improving round-trip time estimates in reliablie
transport protocols. In Proceedings of the SIGCOMM ’87 Conference,
Stowe, Vermont, August 1987.
[15] J. Kay and J. Pasquale. The importance of non-data touching process-
ing overheads in TCP/IP. In Proceedings of the ACM SIGCOMM ’93
Symposium, pages 259–268, September 1993.
[16] H. Kipp. Ethernut embedded ethernet. Web page. 2002-10-14.
URL: http://www.ethernut.de/en/

[17] P. E. McKenney and K. F. Dove. Efficient demultiplexing of incoming

TCP packets. In Proceedings of the SIGCOMM ’92 Conference, pages
269–279, Baltimore, Maryland, August 1992.
[18] M. K. McKusick, K. Bostic, M. J. Karels, and J. S. Quarterman. The
Design and Implementation of the 4.4 BSD Operating System. Addison-
Wesley, 1996.
[19] C. Partridge and S. Pink. A faster UDP. IEEE/ACM Transactions in
Networking, 1(4):429–439, August 1993.
[20] J. Postel. Internet protocol. RFC 791, Internet Engineering Task Force,
September 1981.
[21] J. Postel. Transmission control protocol. RFC 793, Internet Engineering
Task Force, September 1981.
[22] Kadak Products. Kadak KwikNET TCP/IP stack. Web page. 2002-10-14.
URL: http://www.kadak.com/html/kdkp1030.htm
[23] A. Rijsinghani. Computation of the internet checksum via incremental
update. RFC 1624, Internet Engineering Task Force, May 1994.
[24] Luigi Rizzo. Dummynet: a simple approach to the evaluation of network
protocols. ACM Computer Communication Review, 27(1):31–41, 1997.
[25] C. Shannon, D. Moore, and K. Claffy. Beyond folklore: Observations
on fragmented traffic. IEEE/ACM Transactions on Networking, 10(6),
December 2002.
[26] H. Shrikumar. IPic - a match head sized web-server. Web page. 2002-
10-14.
URL: http://www-ccs.cs.umass.edu/˜shri/iPic.html

[27] CMX Systems. CMX-MicroNet true TCP/IP networking. Web page.

2002-10-14.
URL: http://www.cmx.com/micronet.htm

[28] The GCC Team. The GNU compiler collection. Web page. 2002-10-14.
URL: http://gcc.gnu.org/
Chapter 8

Paper B:
Contiki - a Lightweight and
Flexible Operating System
for Tiny Networked Sensors

Adam Dunkels, Björn Grönvall, and Thiemo Voigt. Contiki - a Lightweight

and Flexible Operating System for Tiny Networked Sensors. In Proceedings
of the First IEEE Workshop on Embedded Networked Sensors (IEEE Emnets
2004), Tampa, Florida, USA,
c 2004 Institute of Electrical and Electronics Engineers.

75
Abstract

Wireless sensor networks are composed of large numbers of tiny networked

devices that communicate untethered. For large scale networks it is impor-
tant to be able to dynamically download code into the network. In this pa-
per we present Contiki, a lightweight operating system with support for dy-
namic loading and replacement of individual programs and services. Contiki
is built around an event-driven kernel but provides optional preemptive multi-
threading that can be applied to individual processes. We show that dynamic
loading and unloading is feasible in a resource constrained environment, while
keeping the base system lightweight and compact.
8.1 Introduction 77

8.1 Introduction

Wireless sensor networks are composed of large numbers of tiny sensor devices
with wireless communication capabilities. The sensor devices autonomously
form networks through which sensor data is transported. The sensor devices
are often severely resource constrained. An on-board battery or solar panel
can only supply limited amounts of power. Moreover, the small physical size
and low per-device cost limit the complexity of the system. Typical sensor
devices [1, 2, 5] are equipped with 8-bit microcontrollers, code memory on
the order of 100 kilobytes, and less than 20 kilobytes of RAM. Moore’s law
predicts that these devices can be made significantly smaller and less expensive
in the future. While this means that sensor networks can be deployed to greater
extents, it does not necessarily imply that the resources will be less constrained.
For the designer of an operating system for sensor nodes, the challenge lies
in finding lightweight mechanisms and abstractions that provide a rich enough
execution environment while staying within the limitations of the constrained
devices. We have developed Contiki, an operating system developed for such
constrained environments. Contiki provides dynamic loading and unloading of
individual programs and services. The kernel is event-driven, but the system
supports preemptive multi-threading that can be applied on a per-process basis.
Preemptive multi-threading is implemented as a library that is linked only with
programs that explicitly require multi-threading.
Contiki is implemented in the C language and has been ported to a number
of microcontroller architectures, including the Texas Instruments MSP430 and
the Atmel AVR. We are currently running it on the ESB platform [5]. The ESB
uses the MSP430 microcontroller with 2 kilobytes of RAM and 60 kilobytes
of ROM running at 1 MHz. The microcontroller has the ability to selectively
reprogram parts of the on-chip flash memory.
The contributions of this paper are twofold. Our first contribution is that
we show the feasibility of loadable programs and services even in a constrained
sensor device. The possibility to dynamically load individual programs leads
to a very flexible architecture, which still is compact enough for resource con-
strained sensor nodes. Our second contribution is more general in that we show
that preemptive multi-threading does not have to be implemented at the lowest
level of the kernel but that it can be built as an application library on top of an
event-driven kernel. This allows for thread-based programs running on top of
an event-based kernel, without the overhead of reentrancy or multiple stacks in
all parts of the system.
78 Paper B

8.1.1 Downloading code at run-time

Wireless sensor networks are envisioned to be large scale, with hundreds or

even thousands of nodes per network. When developing software for such
a large sensor network, being able to dynamically download program code
into the network is of great importance. Furthermore, bugs may have to be
patched in an operational network [9]. In general, it is not feasible to physically
collect and reprogram all sensor devices and in-situ mechanisms are required.
A number of methods for distributing code in wireless sensor networks have
been developed [21, 8, 17]. For such methods it is important to reduce the
number of bytes sent over the network, as communication requires a large parts
of the available node energy.
Most operating systems for embedded systems require that a complete bi-
nary image of the entire system is built and downloaded into each device. The
binary includes the operating system, system libraries, and the actual applica-
tions running on top of the system. In contrast, Contiki has the ability to load
and unload individual applications or services at run-time. In most cases, an
individual application is much smaller than the entire system binary and there-
fore requires less energy when transmitted through a network. Additionally,
the transfer time of an application binary is less than that of an entire system
image.

8.1.2 Portability

As the number of different sensor device platforms increases (e.g. [1, 2, 5]),
it is desirable to have a common software infrastructure that is portable across
hardware platforms. The currently available sensor platforms carry completely
different sets of sensors and communication devices. Due to the application
specific nature of sensor networks, we do not expect that this will change in
the future. The single unifying characteristic of today’s platforms is the CPU
architecture which uses a memory model without segmentation or memory
protection mechanisms. Program code is stored in reprogrammable ROM and
data in RAM. We have designed Contiki so that the only abstraction provided
by the base system is CPU multiplexing and support for loadable programs
and services. As a consequence of the application specific nature of sensor
networks, we believe that other abstractions are better implemented as libraries
or services and provide mechanisms for dynamic service management.
8.1 Introduction 79

8.1.3 Event-driven systems

In severely memory constrained environments, a multi-threaded model of op-
eration often consumes large parts of the memory resources. Each thread must
have its own stack and because it in general is hard to know in advance how
much stack space a thread needs, the stack typically has to be over provisioned.
Furthermore, the memory for each stack must be allocated when the thread is
created. The memory contained in a stack can not be shared between many
concurrent threads, but can only be used by the thread to which is was allo-
cated. Moreover, a threaded concurrency model requires locking mechanisms
to prevent concurrent threads from modifying shared resources.
To provide concurrency without the need for per-thread stacks or locking
mechanisms, event-driven systems have been proposed [15]. In event-driven
systems, processes are implemented as event handlers that run to completion.
Because an event handler cannot block, all processes can use the same stack,
effectively sharing the scarce memory resources between all processes. Also,
locking mechanisms are generally not needed because two event handlers never
run concurrently with respect to each other.
While event-driven system designs have been found to work well for many
kinds of sensor network applications [18] they are not without problems. The
state driven programming model can be hard to manage for programmers [17].
Also, not all programs are easily expressed as state machines. One exam-
ple is the lengthy computation required for cryptographic operations. Typi-
cally, such operations take several seconds to complete on CPU constrained
platforms [22]. In a purely event-driven operating system a lengthy computa-
tion completely monopolizes the CPU, making the system unable to respond
to external events. If the operating system instead was based on preemptive
multi-threading this would not be a problem as a lengthy computation could be
preempted.
To combine the benefits of both event-driven systems and preemptible
threads, Contiki uses a hybrid model: the system is based on an event-driven
kernel where preemptive multi-threading is implemented as an application li-
brary that is optionally linked with programs that explicitly require it.
The rest of this paper is structured as follows. Section 8.2 reviews related
work and Section 8.3 presents an overview of the Contiki system. We describe
the design of the Contiki kernel in Section 8.4. The Contiki service concept
is presented in Section 8.5. In the following section, we describe how Contiki
handles libraries and communication support is discussed in Section 8.7. We
present the implementation of preemptive multi-threading in Section 8.8 and
80 Paper B

our experiences with using the system is discussed in Section 8.9. Finally, the
paper is concluded in Section 8.10.

8.2 Related work

TinyOS [15] is probably the earliest operating system that directly targets the
specific applications and limitations of sensor devices. TinyOS is also built
around a lightweight event scheduler where all program execution is performed
in tasks that run to completion. TinyOS uses a special description language
for composing a system of smaller components [12] which are statically linked
with the kernel to a complete image of the system. After linking, modifying the
system is not possible [17]. In contrast, Contiki provides a dynamic structure
which allows programs and drivers to be replaced during run-time and without
relinking.
In order to provide run-time reprogramming for TinyOS, Levis and Culler
have developed Maté [17], a virtual machine for TinyOS devices. Code for
the virtual machine can be downloaded into the system at run-time. The vir-
tual machine is specifically designed for the needs of typical sensor network
applications. Similarly, the MagnetOS [7] system uses a virtual Java machine
to distribute applications across the sensor network. The advantages of using a
virtual machine instead of native machine code is that the virtual machine code
can be made smaller, thus reducing the energy consumption of transporting the
code over the network. One of the drawbacks is the increased energy spent
in interpreting the code—for long running programs the energy saved during
the transport of the binary code is instead spent in the overhead of executing
the code. Contiki programs use native code and can therefore be used for all
types of programs, including low level device drivers without loss of execution
efficiency.
SensorWare [8] provides an abstract scripting language for programming
sensors, but their target platforms are not as resource constrained as ours. Sim-
ilarly, the EmStar environment [13] is designed for less resource constrained
systems. Reijers and Langendoen [21] use a patch language to modify parts of
the binary image of a running system. This works well for networks where all
nodes run the exact same binary code but soon gets complicated if sensors run
slightly different programs or different versions of the same software.
The Mantis system [3] uses a traditional preemptive multi-threaded model
of operation. Mantis enables reprogramming of both the entire operating sys-
tem and parts of the program memory by downloading a program image onto
8.3 System overview 81

EEPROM, from where it can be burned into flash ROM. Due to the multi-
threaded semantics, every Mantis program must have stack space allocated
from the system heap, and locking mechanisms must be used to achieve mu-
tual exclusion of shared variables. In contrast, Contiki uses an event based
scheduler without preemption, thus avoiding allocation of multiple stacks and
locking mechanisms. Preemptive multi-threading is provided by a library that
can be linked with programs that explicitly require it.
The preemptive multi-threading in Contiki is similar to fibers [4] and
the lightweight fibers approach by Welsh and Mainland [23]. Unlike the
lightweight fibers, Contiki does not limit the number of concurrent threads to
two. Furthermore, unlike fibers, threads in Contiki support preemption.
As Exokernel [11] and Nemesis [16], Contiki tries to reduce the number
of abstractions that the kernel provides to a minimum [10]. Abstractions are
instead provided by libraries that have nearly full access to the underlying hard-
ware. While Exokernel strived for performance and Nemesis aimed at quality
of service, the purpose of the Contiki design is to reduce size and complexity,
as well as to preserve flexibility. Unlike Exokernel, Contiki do not support any
protection mechanisms since the hardware for which Contiki is designed do
not support memory protection.

8.3 System overview

A running Contiki system consists of the kernel, libraries, the program loader,
and a set of processes. A process may be either an application program or a
service. A service implements functionality used by more than one applica-
tion process. All processes, both application programs and services, can be
dynamically replaced at run-time. Communication between processes always
goes through the kernel. The kernel does not provide a hardware abstraction
layer, but lets device drivers and applications communicate directly with the
hardware.
A process is defined by an event handler function and an optional poll han-
dler function. The process state is held in the process’ private memory and the
kernel only keeps a pointer to the process state. On the ESB platform [5], the
process state consists of 23 bytes. All processes share the same address space
and do not run in different protection domains. Interprocess communication is
done by posting events.
A Contiki system is partitioned into two parts: the core and the loaded
programs as shown in Figure 8.1. The partitioning is made at compile time
82 Paper B

11111111111
00000000000
ROM
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
Loaded program

00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
Communication service
11111111111
00000000000
RAM
Language run−time 00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
Loaded program

00000000000
11111111111
Program loader

Communication service
Kernel
Kernel

Core Core

Figure 8.1: Partitioning into core and loaded programs.

and is specific to the deployment in which Contiki is used. Typically, the core
consists of the Contiki kernel, the program loader, the most commonly used
parts of the language run-time and support libraries, and a communication stack
with device drivers for the communication hardware. The core is compiled into
a single binary image that is stored in the devices prior to deployment. The core
is generally not modified after deployment, even though it should be noted that
it is possible to use a special boot loader to overwrite or patch the core.
Programs are loaded into the system by the program loader. The program
loader may obtain the program binaries either by using the communication
stack or by using directly attached storage such as EEPROM. Typically, pro-
grams to be loaded into the system are first stored in EEPROM before they are
programmed into the code memory.

8.4 Kernel architecture

The Contiki kernel consists of a lightweight event scheduler that dispatches
events to running processes and periodically calls processes’ polling handlers.
All program execution is triggered either by events dispatched by the kernel or
through the polling mechanism. The kernel does not preempt an event handler
once it has been scheduled. Therefore, event handlers must run to completion.
8.4 Kernel architecture 83

As shown in Section 8.8, however, event handlers may use internal mechanisms
to achieve preemption.
The kernel supports two kind of events: asynchronous and synchronous
events. Asynchronous events are a form of deferred procedure call: asyn-
chronous events are enqueued by the kernel and are dispatched to the target
process some time later. Synchronous events are similar to asynchronous but
immediately causes the target process to be scheduled. Control returns to the
posting process only after the target has finished processing the event. This can
be seen as an inter-process procedure call and is similar to the door abstraction
used in the Spring operating system [14].
In addition to the events, the kernel provides a polling mechanism. Polling
can be seen as high priority events that are scheduled in-between each asyn-
chronous event. Polling is used by processes that operate near the hardware
to check for status updates of hardware devices. When a poll is scheduled all
processes that implement a poll handler are called, in order of their priority.
The Contiki kernel uses a single shared stack for all process execution.
The use of asynchronous events reduce stack space requirements as the stack
is rewound between each invocation of event handlers.

8.4.1 Two level scheduling hierarchy

All event scheduling in Contiki is done at a single level and events cannot pre-
empt each other. Events can only be preempted by interrupts. Normally, inter-
rupts are implemented using hardware interrupts but may also be implemented
using an underlying real-time executive. The latter technique has previously
been used to provide real-time guarantees for the Linux kernel [6].
In order to be able to support an underlying real-time executive, Contiki
never disables interrupts. Because of this, Contiki does not allow events to be
posted by interrupt handlers as that would lead to race-conditions in the event
handler. Instead, the kernel provides a polling flag that it used to request a poll
event. The flag provides interrupt handlers with a way to request immediate
polling.

8.4.2 Loadable programs

Loadable programs are implemented using a run-time relocation function and a
binary format that contains relocation information. When a program is loaded
into the system, the loader first tries to allocate sufficient memory space based
84 Paper B

Kernel

Service layer

Service interface
Application process Service process
Version number
Service Function 1 ptr
interface
Function 1(); stub Function 2 ptr
Function 1 implementation
Function 3 ptr
Function 2(); Function 3 implementation

Function 3();

Function 2 implementation

Figure 8.2: An application function calling a service.

on information provided by the binary. If memory allocation fails, program

loading is aborted.
After the program is loaded into memory, the loader calls the program’s
initialization function. The initialization function may start or replace one or
more processes.

8.4.3 Power save mode

In sensor networks, being able to power down the node when the network is
inactive is an often required way to reduce energy consumption. Power conser-
vation mechanisms depend on both the applications [18] and the network pro-
tocols [20]. The Contiki kernel contains no explicit power save abstractions,
but lets the the application specific parts of the system implement such mech-
anisms. To help the application decide when to power down the system, the
event scheduler exposes the size of the event queue. This information can be
used to power down the processor when there are no events scheduled. When
the processor wakes up in response to an interrupt, the poll handlers are run to
handle the external event.
8.5 Services 85

8.5 Services
In Contiki, a service is a process that implements functionality that can be used
by other processes. A service can be seen as a form of a shared library. Ser-
vices can be dynamically replaced at run-time and must therefore be dynam-
ically linked. Typical examples of services includes communication protocol
stacks, sensor device drivers, and higher level functionality such as sensor data
handling algorithms.
Services are managed by a service layer conceptually situated directly next
to the kernel. The service layer keeps track of running services and provides a
way to find installed services. A service is identified by a textual string that de-
scribes the service. The service layer uses ordinary string matching to querying
installed services.
A service consists of a service interface and a process that implements the
interface. The service interface consists of a version number and a function
table with pointers to the functions that implement the interface.
Application programs using the service use a stub library to communicate
with the service. The stub library is linked with the application and uses the
service layer to find the service process. Once a service has been located, the
service stub caches the process ID of the service process and uses this ID for
all future requests.
Programs call services through the service interface stub and need not be
aware of the fact that a particular function is implemented as a service. The first
time the service is called, the service interface stub performs a service lookup in
the service layer. If the specified service exists in the system, the lookup returns
a pointer to the service interface. The version number in the service interface
is checked with the version of the interface stub. In addition to the version
number, the service interface contains pointers to the implementation of all
service functions. The function implementations are contained in the service
process. If the version number of the service stub match the number in the
service interface, the interface stub calls the implementation of the requested
function.

8.5.1 Service replacement

Like all processes, services may be dynamically loaded and replaced in a run-
ning Contiki system. Because the process ID of the service process is used as
a service identifier, it is crucial that the process ID is retained if the service
process is replaced. For this reason, the kernel provides special mechanism for
86 Paper B

replacing a process and retaining the process ID.

When a service is to be replaced, the kernel informs the running version of
the service by posting a special event to the service process. In response to this
event, the service must remove itself from the system.
Many services have an internal state that may need to be transfered to the
new process. The kernel provides a way to pass a pointer to the new ser-
vice process, and the service can produce a state description that is passed to
the new process. The memory for holding the state must be allocated from a
shared source, since the process memory is deallocated when the old process
is removed.
The service state description is tagged with the version number of the ser-
vice, so that an incompatible version of the same service will not try to load
the service description.

8.6 Libraries
The Contiki kernel only provides the most basic CPU multiplexing and event
handling features. The rest of the system is implemented as system libraries
that are optionally linked with programs. Programs can be linked with libraries
in three different ways. First, programs can be statically linked with libraries
that are part of the core. Second, programs can be statically linked with li-
braries that are part of the loadable program. Third, programs can call services
implementing a specific library. Libraries that are implemented as services can
be dynamically replaced at run-time.
Typically, run-time libraries such as often-used parts of the language run-
time libraries are best placed in the Contiki core. Rarely used or application
specific libraries, however, are more appropriately linked with loadable pro-
grams. Libraries that are part of the core are always present in the system and
do not have to be included in loadable program binaries.
As an example, consider a program that uses the memcpy() and atoi()
functions to copy memory and to convert strings to integers, respectively. The
memcpy() function is a frequently used C library function, whereas atoi()
is used less often. Therefore, in this particular example, memcpy() has been
included in the system core but not atoi(). When the program is linked to
produce a binary, the memcpy() function will be linked against its static ad-
dress in the core. The object code for the part of the C library that implements
the atoi() function must, however, be included in the program binary.
8.7 Communication support 87

Communication
Application stack

Routing protocol 1 Routing protocol 2

Device driver 1 Device driver 2

Hardware

Figure 8.3: Loosely coupled communication stack.

8.7 Communication support

Communication is a fundamental concept in sensor networks. In Contiki, com-
munication is implemented as a service in order to enable run-time replace-
ment. Implementing communication as a service also provides for multiple
communication stacks to be loaded simultaneously. In experimental research,
this can be used to evaluate and compare different communication protocols.
Furthermore, the communication stack may be split into different services as
shown in Figure 8.3. This enables run-time replacement of individual parts of
the communication stack.
Communication services use the service mechanism to call each other and
synchronous events to communicate with application programs. Because syn-
chronous event handlers are required to be run to completion, it is possible to
use a single buffer for all communication processing. With this approach, no
data copying has to be performed. A device driver reads an incoming packet
into the communication buffer and then calls the upper layer communication
service using the service mechanisms. The communication stack processes the
headers of the packet and posts a synchronous event to the application program
for which the packet was destined. The application program acts on the packet
contents and optionally puts a reply in the buffer before it returns control to the
communication stack. The communication stack prepends its headers to the
outgoing packet and returns control to the device driver so that the packet can
be transmitted.
88 Paper B

mt yield();
Yield from the running thread.

mt post(id, event, dataptr);

Post an event from the running thread.

mt wait(event, dataptr);
Wait for an event to be posted to the running thread.

mt exit();
Exit the running thread.

mt start(thread, functionptr, dataptr);

Start a thread with a specified function call.

mt exec(thread);
Execute the specified thread until it yields or is preempted.

Figure 8.4: The multi-threading library API.

8.8 Preemptive multi-threading

In Contiki, preemptive multi-threading is implemented as a library on top of
the event-based kernel. The library is optionally linked with applications that
explicitly require a multi-threaded model of operation. The library is divided
into two parts: a platform independent part that interfaces to the event kernel,
and a platform specific part implementing the stack switching and preemption
primitives. Usually, the preemption is implemented using a timer interrupt
that saves the processor registers onto the stack and switches back to the ker-
nel stack. In practice very little code needs to be rewritten when porting the
platform specific part of the library. For reference, the implementation for the
MSP430 consists of 25 lines of C code.
Unlike normal Contiki processes each thread requires a separate stack. The
library provides the necessary stack management functions. Threads execute
on their own stack until they either explicitly yield or are preempted.
The API of the multi-threading library is shown in Figure 8.4. It con-
sists of four functions that can be called from a running thread (mt yield(),
mt post(), mt wait(), and mt exit()) and two functions that are called
to setup and run a thread (mt start() and mt exec()). The mt exec()
8.9 Discussion 89

function performs the actual scheduling of a thread and is called from an event
handler.

8.9 Discussion
We have used the Contiki operating system to implement a number of sen-
sor network applications such as multi-hop routing, motion detection with dis-
tributed sensor data logging and replication, and presence detection and notifi-
cation.

8.9.1 Over-the-air programming

We have implemented a simple protocol for over-the-air programming of en-
tire networks of sensors. The protocol transmits a single program binary to
selected concentrator nodes using point-to-point communication. The binary
is stored in EEPROM and when the entire program has been received, it is
broadcasted to neighboring nodes. Packet loss is signaled by neighbors using
negative acknowledgments. Repairs are made by the concentrator node. We
intend to implement better protocols, such as the Trickle algorithm [19], in the
future.
During the development of one network application, a 40-node dynamic
distributed alarm system, we used both over-the-air reprogramming and man-
ual wired reprogramming of the sensor nodes. At first, the program loading
mechanism was not fully functional and we could not use it during our devel-
opment. The object code size of our application was approximately 6 kilobytes.
Together with the Contiki core and the C library, the complete system image
was nearly 30 kilobytes. Reprogramming of an individual sensor node took just
over 30 seconds. With 40 nodes, reprogramming the entire network required
at least 30 minutes of work and was therefore not feasible to do often. In con-
trast, over-the-air reprogramming of a single component of the application was
done in about two minutes—a reduction in an order of magnitude—and could
be done with the sensor nodes placed in the actual test environment.

8.9.2 Code size

An operating system for constrained devices must be compact in terms of both
code size and RAM usage in order to leave room for applications running on
90 Paper B

Module Code size Code size RAM

(AVR) (MSP430) usage
10 +
Kernel 1044 810 + 4e + 2p
Service layer 128 110 0
Program loader - 658 8
Multi-threading 678 582 8+s
Timer library 90 60 0
Replicator stub 182 98 4
Replicator 1752 1558 200
230 + 4e +
Total 3874 3876 + 2p + s

Table 8.1: Size of the compiled code, in bytes.

top of the system. Table 8.1 shows the compiled code size and the RAM us-
age of the Contiki system compiled for two architectures: the Texas Instru-
ments MSP430 and the Atmel AVR. The numbers report the size of both core
components and an example application: a sensor data replicator service. The
replicator service consists of the service interface stub for the service as well as
the implementation of the service itself. The program loader is currently only
implemented on the MSP430 platform.

The code size of Contiki is larger than that of TinyOS [15], but smaller than
that of the Mantis system [3]. Contiki’s event kernel is significantly larger than
that of TinyOS because of the different services provided. While the TinyOS
event kernel only provides a FIFO event queue scheduler, the Contiki kernel
supports both FIFO events and poll handlers with priorities. Furthermore,
the flexibility in Contiki requires more run-time code than for a system like
TinyOS, where compile time optimization can be done to a larger extent.

The RAM requirement depends on the maximum number of processes that

the system is configured to have (p), the maximum size of the asynchronous
event queue (e) and, in the case of multi-threaded operation, the size of the
thread stacks (s).
8.9 Discussion 91

17
Round-trip time

16.5
Milliseconds

15.5

15
0 2 4 6 8 10 12 14 16
Seconds

Figure 8.5: A slight increase in response time during a preemptible computa-

tion.

8.9.3 Preemption
The purpose of preemption is to facilitate long running computations while be-
ing able to react on incoming events such as sensor input or incoming commu-
nication packets. Figure 8.5 shows how Contiki responds to incoming packets
during an 8 second computation running in a preemptible thread. The curve
is the measured round-trip time of 200 “ping” packets of 40 bytes each. The
computation starts after approximately 5 seconds and runs until 13 seconds
have passed. During the computation, the round-trip time increases slightly
but the system is still able to produce replies to the ping packets.
The packets are sent over a 57600 kbit/s serial line with a spacing of 200
ms from a 1.4 GHz PC to an ESB node running Contiki. The packets are trans-
mitted over a serial line rather than over the wireless link in order to avoid radio
effects such as bit errors and MAC collisions. The computation consists of an
arbitrarily chosen sequence of multiplications and additions that are repeated
for about 8 seconds. The cause for the increase in round-trip time during the
computation is the cost of preempting the computation and restoring the kernel
context before the incoming packet can be handled. The jitter and the spikes of
about 0.3 milliseconds seen in the curve can be contributed to activity in other
92 Paper B

poll handlers, mostly the radio packet driver.

8.9.4 Portability
We have ported Contiki to a number of architectures, including the Texas In-
struments MSP430 and the Atmel AVR. Others have ported the system to the
Hitachi SH3 and the Zilog Z80. The porting process consists of writing the
boot up code, device drivers, the architecture specific parts of the program
loader, and the stack switching code of the multi-threading library. The kernel
and the service layer does not require any changes.
Since the kernel and service layer does not require any changes, an opera-
tional port can be tested after the first I/O device driver has been written. The
Atmel AVR port was made by ourselves in a couple of hours, with help of pub-
licly available device drivers. The Zilog Z80 port was made by a third party, in
a single day.

8.10 Conclusions
We have presented the Contiki operating system, designed for memory con-
strained systems. In order to reduce the size of the system, Contiki is based
on an event-driven kernel. The state-machine driven programming of event-
driven systems can be hard to use and has problems with handling long running
computations. Contiki provides preemptive multi-threading as an application
library that runs on top of the event-driven kernel. The library is optionally
linked with applications that explicitly require a multi-threaded model of com-
putation.
A running Contiki system is divided into two parts: a core and loaded pro-
grams. The core consists of the kernel, a set of base services, and parts of the
language run-time and support libraries. The loaded programs can be loading
and unloading individually, at run-time. Shared functionality is implemented
as services, a form of shared libraries. Services can be updated or replaced
individually, which leads to a very flexible structure.
We have shown that dynamic loading and unloading of programs and ser-
vices is feasible in a resource constrained system, while keeping the base sys-
tem lightweight and compact. Even though our kernel is event-based, preemp-
tive multi-threading can be provided at the application layer on a per-process
basis.
8.10 Conclusions 93

Because of its dynamic nature, Contiki can be used to multiplex the hard-
ware of a sensor network across multiple applications or even multiple users.
This does, however, require ways to control access to the reprogramming facil-
ities. We plan to continue our work in the direction of operating system support
for secure code updates.
Bibliography

Bibliography

[1] Crossbow mica motes. Web page.

URL: http://www.xbow.com/

[2] Eyes prototype sensor node. Web page. Visited 2004-06-22.

URL: http://eyes.eu.org/sensnet.htm

[3] H. Abrach, S. Bhatti, J. Carlson, H. Dai, J. Rose, A. Sheth, B. Shucker,

J. Deng, and R. Han. MANTIS: system support for MultimodAl NeT-
works of In-Situ sensors. In Proc. WSNA’03, 2003.
[4] A. Adya, J. Howell, M. Theimer, W. J. Bolosky, and J. R. Douceur. Coop-
erative Task Management Without Manual Stack Management. In Proc.
USENIX, 2002.
[5] CST Group at FU Berlin. Scatterweb Embedded Sensor Board. Web
page. 2003-10-21.
URL: http://www.scatterweb.com/

[6] M. Barabanov. A Linux-based RealTime Operating System. Master’s

thesis, New Mexico Institute of Mining and Technology, 1997.
[7] R. Barr, J. C. Bicket, D. S. Dantas, B. Du, T. W. D. Kim, B. Zhou, and
E. Sirer. On the need for system-level support for ad hoc and sensor
networks. SIGOPS Oper. Syst. Rev., 36(2), 2002.
[8] A. Boulis, C. Han, and M. B. Srivastava. Design and implementation of
a framework for efficient and programmable sensor networks. In Proc.
MOBISYS‘03, May 2003.
[9] D. Estrin (editor). Embedded everywhere: A research agenda for net-
worked systems of embedded computers. National Academy Press, 2001.

94
Bibliography 95

[10] D. R. Engler and M. F. Kaashoek. Exterminate all operating system ab-

stractions. In Proc. HotOS-V, May 1995.

[11] D. R. Engler, M. F. Kaashoek, and J. O’Toole Jr. Exokernel: an operating

system architecture for application-level resource management. In Proc
SOSP ’95, December 1995.

[12] D. Gay, P. Levis, R. von Behren, M. Welsh, E. Brewer, and D. Culler. The
nesC language: A holistic approach to networked embedded systems. In
Proc. SIGPLAN’03, 2003.

[13] L. Girod, J. Elson, A. Cerpa, T. Stathopoulos, N. Ramanathan, and D. Es-

trin. EmStar: A Software Environment for Developing and Deploying
Wireless Sensor Networks. In Proc. USENIX, 2004.

[14] G. Hamilton and P. Kougiouris. The spring nucleus: A microkernel for

objects. In Proc. Usenix Summer Conf., 1993.

[15] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler, and K. Pister. Sys-

tem architecture directions for networked sensors. In Proc. ASPLOS-IX,
November 2000.

[16] I. M. Leslie, D. McAuley, R. Black, T. Roscoe, P. T. Barham, D. Evers,

R. Fairbairns, and E. Hyden. The design and implementation of an oper-
ating system to support distributed multimedia applications. IEEE JSAC,
14(7):1280–1297, 1996.

[17] P. Levis and D. Culler. Maté: A tiny virtual machine for sensor networks.
In Proc. ASPLOS-X, October 2002.

[18] P. Levis, S. Madden, D. Gay, J. Polastre, R. Szewczyk, A. Woo,

E. Brewer, and D. Culler. The Emergence of Networking Abstractions
and Techniques in TinyOS. In Proc. NSDI, 2004.

[19] P. Levis, N. Patel, D. Culler, and S. Shenker. Trickle: A self-regulating

algorithm for code propagation and maintenance in wireless sensor net-
works. In Proc. NSDI, 2004.

[20] V. Raghunathan, C. Schurgers, S. Park, and M. Srivastava. Energy

aware wireless microsensor networks. IEEE Signal Processing Maga-
zine, 19(2):40–50, 2002.
[21] N. Reijers and K. Langendoen. Efficient code distribution in wireless
sensor networks. In Proc. WSNA’03, 2003.
[22] F. Stajano. Security for Ubiquitous Computing. Wiley, 2002.
[23] M. Welsh and G. Mainland. Programming Sensor Networks Using Ab-
stract Regions. In Proc. NSDI, 2004.
Chapter 9

Paper C:
Using Protothreads for
Sensor Node Programming

Adam Dunkels, Oliver Schmidt, and Thiemo Voigt. Using protothreads for
sensor node programming. In Proceedings of the Workshop on Real-World
Wireless Sensor Networks (REALWSN 2005), Stockholm, Sweden, June 2005.
c 2005 Swedish Institute of Computer Science.

97
Abstract

Wireless sensor networks consist of tiny devices that usually have severe re-
source constraints in terms of energy, processing power and memory. In order
to work efficiently within the constrained memory, many operating systems
for such devices are based on an event-driven model rather than on traditional
multi-threading. While event-driven systems allow for reduced memory us-
age, they require programs to be developed as explicit state machines. Since
implementing programs using explicit state machines is hard, developing and
maintaining programs for event-driven systems is typically more difficult than
for multi-threaded ones.
In this paper, we introduce protothreads, a programming abstraction for
event-driven sensor network systems. Protothreads simplify implementation of
high-level functionality on top of event-driven systems, compared to traditional
methods.
9.1 Introduction 99

9.1 Introduction
Wireless sensor networks consist of tiny devices that usually have severe re-
source constraints in terms of energy, processing power and memory. Most
programming environments for wireless sensor network nodes today are
based on an event-triggered programming model rather than traditional multi-
threading. In TinyOS [7], the event-triggered model was chosen over a multi-
threaded model due to the memory overhead of threads. According to Hill et
al. [7]:
“In TinyOS, we have chosen an event model so that high levels
of concurrency can be handled in a very small amount of space. A
stack-based threaded approach would require that stack space be
reserved for each execution context.”
While the event-driven model and the threaded model can be shown to be
equivalent [9], programs written in the two models typically display differ-
ing characteristics [1]. The advantages and disadvantages of the two models
are a debated topic [11, 14].
In event-triggered systems, programs are implemented as event handlers.
Event handlers are invoked in response to external or internal events, and run to
completion. An event handler typically is a programming language procedure
or function that performs an explicit return to the caller. Because of the run-to-
completion semantics, an event-handler cannot execute a blocking wait. With
run-to-completion semantics, the system can utilize a single, shared stack. This
reduces the memory overhead over a multi-threaded system, where memory
must be allocated for a stack for each running program.
The run-to-completion semantics of event-triggered systems makes imple-
menting certain high-level operations a complex task. When an operation can-
not complete immediately, the operation must be split across multiple invo-
cations of the event handler. Levis et al. [10] refer to this as a split-phase
operation. In the words of Levis et al.:
“This approach is natural for reactive processing and for in-
terfacing with hardware, but complicates sequencing high-level
operations, as a logically blocking sequence must be written in
a state-machine style.”
In this paper, we introduce the notion of using protothreads [3, 6] as a
method to reduce the complexity of high-level programs in event-triggered sen-
sor node systems. We argue that protothreads can reduce the number of explicit
100 Paper C

state machines required to implement typical high-level sensor node programs.

We believe this reduction leads to programs that are easier to develop, debug,
and maintain, based on extensive experience with developing software for the
event-driven uIP TCP/IP stack [4] and Contiki operating system [5].
The main contribution of this paper is the protothread programming ab-
straction. We show that protothreads reduce the complexity of programming
sensor nodes. Further, we demonstrate that protothreads can be implemented
in the C programming language, using only standard C language constructs
and without any architecture-specific machine code.
The rest of this paper is structured as follows. Section 9.2 presents a mo-
tivating example and Section 9.3 introduces the notion of protothreads. Sec-
tion 9.4 discusses related work, and the paper is concluded in Section 9.5.

9.2 Motivation
To illustrate how high-level functionality is implemented using state machines,
we consider a hypothetical energy-conservation mechanism for wireless sensor
nodes. The mechanism switches the radio on and off at regular intervals. The
mechanism works as follows:

1. Turn radio on.

2. Wait for tawake milliseconds.

3. Turn radio off, but only if all communication has completed.

4. If communication has not completed, wait until it has completed. Then

turn off the radio.

5. Wait for tsleep milliseconds. If the radio could not be turned off before
tsleep milliseconds because of remaining communication, do not turn the
radio off at all.

6. Repeat from step 1.

To implement this protocol in an event-driven model, we first need to iden-

tify a set of states around which the state machine can be designed. For this
protocol, we can see three states: on – the radio is turned on, waiting – wait-
ing for any remaining communication to complete, and off – the radio is off.
Figure 9.2 shows the resulting state machine, including the state transitions.
9.2 Motivation 101

enum {
PT_THREAD(radio_wake_thread
ON,
(struct pt *pt)) {
WAITING,
PT_BEGIN(pt);
OFF
} state;
while(1) {
radio_on();
void radio_wake_eventhandler() {
timer_set(&timer, T_AWAKE);
switch(state) {
PT_WAIT_UNTIL(pt,
timer_expired(&timer));
case OFF:
if(timer_expired(&timer)) {
timer_set(&timer, T_SLEEP);
radio_on();
if(!communication_complete()) {
state = ON;
PT_WAIT_UNTIL(pt,
timer_set(&timer, T_AWAKE);
communication_complete()
}
|| timer_expired(&timer));
break;
}
case ON:
if(!timer_expired(&timer)) {
if(timer_expired(&timer)) {
radio_off();
timer_set(&timer, T_SLEEP);
PT_WAIT_UNTIL(pt,
if(!communication_complete())
timer_expired(&timer));
{
}
state = WAITING;
}
} else {
radio_off();
PT_END(pt);
state = OFF;
}
}
}
break;

case WAITING:
if(communication_complete()
|| timer_expired(&timer)) {
state = ON;
timer_set(&timer, T_AWAKE);
} else {
radio_off();
state = OFF;
}
break;
}
}

Figure 9.1: The radio sleep cycle implemented with events (left) and with pro-
tothreads (right).

To implement this state machine in C, we use an explicit state variable,

state, that can take on the values OFF, ON, and WAITING. We use a
switch() statement to perform different actions depending on the state
variable. The code is placed in an event handler function that is called when-
ever an event occurs. Possible events in this case are that a timer expires and
that communication completes. The resulting C code is shown in the left part
of Figure 9.1.

We note that this simple mechanism results in a fairly large amount of C

code. The structure of the mechanism, as it is described by the six steps above,
is not evident from the C code.
102 Paper C

9.3 Protothreads
Protothreads [6] are an extremely lightweight stackless type of threads, de-
signed for severely memory constrained systems. Protothreads provide condi-
tional blocking waits on top of an event-driven system, without the overhead
of per-thread stacks. The purpose of protothreads is to implement sequential
flow of control without complex state machines or full multi-threading.
We developed protothreads in order to deal with the complexity of explicit
state machines in the event-driven uIP TCP/IP stack [4]. For uIP, we were
able to substantially reduce the number of state machines and explicit states
used in the implementations of a number of application level communication
protocols. For example, the uIP FTP client could be simplified by completely
removing the explicit state machine, and thereby reducing the number of ex-
plicit states from 20 to one.

9.3.1 Protothreads versus events

Programs written for an event-driven model typically have to be implemented
as explicit state machines. To illustrate how protothreads solve this problem,
we return to the radio sleep cycle example from the previous section.
The right part of Figure 9.1 shows how the radio sleep cycle mechanism
is implemented with protothreads. Comparing the left and right part of Fig-
ure 9.1, we see that the protothreads-based implementation not only is shorter,
but also more closely follows the specification of the radio sleep mechanism.
Due to the linear code flow of this implementation, the overall logic of the sleep
cycle mechanism is visible in the C code. Also, in the protothreads-based im-
plementation we are able to make use of regular C control flow mechanisms
such as while() loops and if() statements.

t_awake timer expired

OFF ON

t_sleep timer expired

communicaion
completed communication
active
WAITING

Figure 9.2: State machine realization of the radio sleep cycle protocol.
9.3 Protothreads 103

9.3.2 Protothreads versus threads

The main advantage of protothreads over traditional threads is that protothreads
are very lightweight: a protothread does not require its own stack. Rather,
all protothreads run on the same stack and context switching is done by stack
rewinding. This is advantageous in memory constrained systems, where a stack
for a thread might use a large part of the available memory. In comparison, a
protothread requires only two bytes of memory per protothread and no addi-
tional stack.
A protothread runs within a single C function and cannot span over other
functions. A protothread may call normal C functions, but cannot block inside
a called function. Blocking inside nested function calls is instead implemented
by spawning a separate protothread for each potentially blocking function. Un-
like threads, this makes protothreads blocking explicit: the programmer knows
exactly which functions that potentially may yield.

9.3.3 Comparison
Table 9.1 summarizes the features of protothreads and compares them with the
features of events and threads. The names of the features are from [1].

Proto-
Feature Events Threads threads
Control structures No Yes Yes
Debug stack retained No Yes Yes
Implicit locking Yes No Yes
Preemption No Yes No
Automatic variables No Yes No

Table 9.1: Qualitative comparison between events, threads and protothreads

Control structures. One of the advantages of threads over events is that

threads allow programs to make full use of the control structures (e.g., if
conditionals and while loops) provided by the programming language. In
the event-driven model, control structures must be broken down into two
or more pieces in order to implement continuations [1]. Like threads,
protothreads allow blocking statements to be used together with control
structures.
104 Paper C

void radio_wake_thread(struct pt *pt) {

switch(pt->lc) {
case 0:

while(1) {
radio_on();
timer_set(&timer, T_AWAKE);

pt->lc = 8;
case 8:
if(!timer_expired(&timer)) {
return;
}

timer_set(&timer, T_SLEEP);
if(!communication_complete()) {

pt->lc = 13;
case 13:
if(!(communication_complete() ||
timer_expired(&timer))) {
return;
}

if(!timer_expired(&timer)) {
radio_off();

pt->lc = 18;
case 18:
if(!timer_expired(&timer)) {
return;
}

}
}
}
}

Figure 9.3: C switch statement expansion of the protothreads code in Figure 9.1

Debug stack retained. Because the manual stack management and the free
flow of control in the event-driven model, debugging is difficult as the
sequence of calls is not saved on the stack [1]. With both threads and
protothreads, the full call stack is available for debugging.

Implicit locking. With manual stack management, as in the event-driven

model, all yield points are immediately visible in the code. This makes it
evident to the programmer whether or not a structure needs to be locked.
In the threaded model, it is not as evident that a particular function call
yields. Using protothreads, however, potentially blocking statements are
explicitly implemented with a PT WAIT statement. Program code be-
tween such statements never yields.

Preemption. The semantics of the threaded model allows for preemption of a

running thread, save its stack, and continue execution of another thread.
9.3 Protothreads 105

Because both the event-driven model and protothreads use a single stack,
preemption is not possible within these models.

Automatic variables. Since the threaded model allocates a stack for each
thread, automatic variables—variables with function local scope auto-
matically allocated on the stack—are retained even when the thread
blocks. Both the event-driven model and protothreads use a single shared
stack for all active programs, and rewind the stack every time a program
blocks. Therefore, with protothreads, automatic variables are not saved
across a blocking wait. This is discussed in more detail below.

9.3.4 Limitations
While protothreads allow programs to take advantage of some of the bene-
fits of a threaded programming model, protothreads also impose some of the
limitation from the event-driven model. The most evident limitation from the
event-driven model is that automatic variables—variables with function-local
scope that are automatically allocated on the stack—are not saved across a
blocking wait. While automatic variables can still be used inside a protothread,
the contents of the variables must be explicitly stored before executing a wait
statement. The reason for this is that protothreads rewind the stack at every
blocking statement, and therefore potentially destroy the contents of variables
on the stack.
If an automatic variable is erroneously used after a blocking statement, the
C compiler is able to detect the problem. Typically a warning is produced, stat-
ing that the variable in question “might be used uninitialized in this function”.
While it may not be immediately apparent for the programmer that this warn-
ing is related to the use of automatic variables across a blocking protothreads
statement, it does provide an indication that there is a problem with the pro-
gram. Also, the warning indicates the line number of the problem which assists
the programmer in identifying the problem.
The limitation on the use of automatic variables can be handled by using
an explicit state object, much in the same way as is done in the event-driven
model. The state object is a chunk of memory that holds the contents of all
automatic variables that need to be saved across a blocking statement. It is,
however, the responsibility of the programmer to allocate and maintain such a
state object.
It should also be noted that protothreads do not limit the use of static lo-
cal variables. Static local variables are variables that are local in scope but
106 Paper C

allocated in the data section. Since these are not placed on the stack, they are
not affected by the use of blocking protothreads statements. For functions that
do not need to be re-entrant, using static local variables instead of automatic
variables can be an acceptable solution to the problem.

9.3.5 Implementation
Protothreads are based on a low-level mechanism that we call local continu-
ations [6]. A local continuation is similar to ordinary continuations [12], but
does not capture the program stack. Local continuations can be implemented
in a variety of ways, including using architecture specific machine code, C-
compiler extensions, and a non-obvious use of the C switch statement. In this
paper, we concentrate on the method based on the C switch statement.
A local continuation supports two operations; it can be either set or re-
sumed. When a local continuation is set, the state of the function—all CPU
registers including the program counter but excluding the stack—is captured.
When the same local continuation is resumed, the state of the function is reset
to what it was when the local continuation was set.
A protothread consists of a single local continuation. The protothread’s lo-
cal continuation is set before each conditional blocking wait. If the condition
is true and the wait is to be performed, the protothread executes an explicit
return statement, thus returning to the caller. The next time the protothread is
called, it resumes the local continuation that was previously set. This will ef-
fectively cause the program to jump to the conditional blocking wait statement.
The condition is re-evaluated and, once the condition is false, the protothread
continues down through the function.

#define RESUME(lc) switch(lc) { case 0:

#define SET(lc) lc = LINE; case LINE:

Figure 9.4: The local continuation resume and set operations implemented us-
ing the C switch statement.

Local continuations can be implemented using standard C language con-

structs and a non-obvious use of the switch statement. With this technique,
the local continuation is represented by an unsigned integer. The resume op-
eration is implemented as an open switch statement, and the set operation is
9.4 Related Work 107

implemented as an assignment of the local continuation and a case statement,

as shown in Figure 9.4. Each set operation sets the local continuation to a
value that is unique within each function, and the resume operation’s switch
statement jumps to the corresponding case statement. We note that the case
0: statement in the implementation of the resume operation ensures that the
resume statement does nothing if is the local continuation is zero.
Figure 9.3 shows the example radio sleep cycle mechanism from Sec-
tion 9.2 with the protothreads statements expanded using the C switch imple-
mentation of local continuations. We see how each PT WAIT statement has
been replaced with a case statement, and how the PT BEGIN statement has
been replaced with a switch statement. Additionally, the PT END statement
has been replaced with a single right curly bracket, which closes the switch
block that was opened by the PT BEGIN statement.
The non-obviousness of the C switch implementation of local continuations
is that the technique appears to cause problems when a conditional blocking
statement is used inside a nested C control statement. For example, the case
13: statement in Figure 9.3 appears inside an if block, while the corresponding
switch statement is located at a higher block. However, this is a valid use of the
C switch statement: case statements may be located anywhere inside a switch
block. They do not need to be in the same level of nesting, but can be located
anywhere, even inside nested if or for blocks. This use of the switch statement
is likely to first have been publicly described by Duff [2]. The same technique
has later been used by Tatham to implement coroutines in C [13].
The implementation of protothreads using the C switch statements imposes
a restriction on programs using protothreads: programs cannot utilize switch
statements together with protothreads. If a switch statement is used by the pro-
gram using protothreads, the C compiler will is some cases emit an error, but
in most cases the error not be detected. This is troublesome as it may lead to
unexpected run-time behavior which is hard to trace back to an erroneous mix-
ture of one particular implementation of protothreads and switch statements.
We have not yet found a suitable solution for this problem.

9.4 Related Work

Kasten and Römer [8] have also identified the need for new abstractions for
managing the complexity of event-triggered programming. They introduce
OSM, a state machine programming model based on Harel’s StateCharts. The
model reduces both the complexity of the implementations and the memory
108 Paper C

usage. Their work is different from protothreads in that OSM requires sup-
port from an external OSM compiler to produce the resulting C code, whereas
protothreads only make use of the regular C preprocessor.

9.5 Conclusions
Many operating systems for wireless sensor network nodes are based on an
event-triggered programming model. In order to implement high-level opera-
tions under this model, programs have to be written as explicit state machines.
Software implemented using explicit state machines is often hard to under-
stand, debug, and maintain.
We have presented protothreads as a programming abstraction that re-
duces the complexity of implementations of high-level functionality for event-
triggered systems. With protothreads, programs can perform blocking waits on
top of event-triggered systems with run-to-completion semantics.

Acknowledgments
This work was partly financed by VINNOVA, the Swedish Agency for Inno-
vation Systems, and the European Commission under contract IST-004536-
RUNES.
Bibliography

[1] A. Adya, J. Howell, M. Theimer, W. J. Bolosky, and J. R. Douceur. Co-

operative Task Management Without Manual Stack Management. In Pro-
ceedings of the USENIX Annual Technical Conference, pages 289–302,
2002.

[2] T. Duff. Re: Explanation please! Usenet news article, Message-ID:

<[email protected]>, August 1988.

[3] A. Dunkels. Protothreads web site. Web page. Visited 2006-12-15.

URL: http://www.sics.se/˜adam/pt/

[4] A. Dunkels. Full TCP/IP for 8-bit architectures. In Proceedings of The

First International Conference on Mobile Systems, Applications, and Ser-
vices (MOBISYS ‘03), San Francisco, California, May 2003.

[5] A. Dunkels, B. Grönvall, and T. Voigt. Contiki - a lightweight and flexible

operating system for tiny networked sensors. In Proceedings of the First
IEEE Workshop on Embedded Networked Sensors (IEEE Emnets ’04),
Tampa, Florida, USA, November 2004.

[6] A. Dunkels and O. Schmidt. Protothreads – Lightweight Stackless

Threads in C. Technical Report T2005:05, Swedish Institute of Com-
puter Science.

[7] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler, and K. Pister. Sys-

109
[8] O. Kasten and K. Römer. Beyond event handlers: Programming wire-
less sensors with attributed state machines. In The Fourth International
Conference on Information Processing in Sensor Networks (IPSN), Los
Angeles, USA, April 2005.
[9] H. C. Lauer and R. M. Needham. On the duality of operating systems
structures. In Proc. Second International Symposium on Operating Sys-
tems, October 1978.
[10] P. Levis, S. Madden, D. Gay, J. Polastre, R. Szewczyk, A. Woo,
E. Brewer, and D. Culler. The Emergence of Networking Abstractions
and Techniques in TinyOS. In Proceedings of ACM/Usenix Networked
Systems Design and Implementation (NSDI’04), San Francisco, Califor-
nia, USA, March 2004.
[11] J. K. Ousterhout. Why threads are a bad idea (for most purposes). Invited
Talk at the 1996 USENIX Technical Conference, 1996.

[12] J. C. Reynolds. The discoveries of continuations. Lisp and Symbolic

Computation, 6(3):233–247, 1993.
[13] S. Tatham. Coroutines in C. Web page, 2000.
URL: http://www.chiark.greenend.org.uk/˜sgtatham/coroutines.html

[14] R. von Behren, J. Condit, and E. Brewer. Why events are a bad idea (for
high-concurrency servers). In Proceedings of the 9th Workshop on Hot
Topics in Operating Systems, Lihue (Kauai), Hawaii, USA, May 2003.
Chapter 10

Paper D:
Protothreads: Simplifying
Event-Driven Programming
of Memory-Constrained
Embedded Systems

Adam Dunkels, Oliver Schmidt, Thiemo Voigt, and Muneeb Ali. Protothreads:
Simplifying event-driven programming of memory-constrained embedded sys-
tems. In Proceedings of the 4th International Conference on Embedded Net-
worked Sensor Systems (ACM SenSys 2006), Boulder, Colorado, USA, Novem-
ber 2006.
c 2006 Association for Computing Machinery.

111
Abstract

Event-driven programming is a popular model for writing programs for tiny

embedded systems and sensor network nodes. While event-driven program-
ming can keep the memory overhead down, it enforces a state machine pro-
gramming style which makes many programs difficult to write, maintain, and
debug. We present a novel programming abstraction called protothreads that
makes it possible to write event-driven programs in a thread-like style, with
a memory overhead of only two bytes per protothread. We show that pro-
tothreads significantly reduce the complexity of a number of widely used pro-
grams previously written with event-driven state machines. For the examined
programs the majority of the state machines could be entirely removed. In the
other cases the number of states and transitions was drastically decreased. With
protothreads the number of lines of code was reduced by one third. The execu-
tion time overhead of protothreads is on the order of a few processor cycles.
10.1 Introduction 113

10.1 Introduction
Event-driven programming is a common programming model for memory-
constrained embedded systems, including sensor networks. Compared to
multi-threaded systems, event-driven systems do not need to allocate mem-
ory for per-thread stacks, which leads to lower memory requirements. For this
reason, many operating systems for sensor networks, including TinyOS [19],
SOS [17], and Contiki [12] are based on an event-driven model. According
to Hill et al. [19]: “In TinyOS, we have chosen an event model so that high
levels of concurrency can be handled in a very small amount of space. A stack-
based threaded approach would require that stack space be reserved for each
execution context.” Event-driven programming is also often used in systems
that are too memory-constrained to fit a general-purpose embedded operating
system [28].
An event-driven model does not support a blocking wait abstraction. There-
fore, programmers of such systems frequently need to use state machines to im-
plement control flow for high-level logic that cannot be expressed as a single
event handler. Unlike state machines that are part of a system specification, the
control-flow state machines typically have no formal specification, but are cre-
ated on-the-fly by the programmer. Experience has shown that the need for ex-
plicit state machines to manage control flow makes event-driven programming
difficult [3, 25, 26, 35]. With the words of Levis et al. [26]: “This approach is
natural for reactive processing and for interfacing with hardware, but compli-
cates sequencing high-level operations, as a logically blocking sequence must
be written in a state-machine style.” In addition, popular programming lan-
guages for tiny embedded systems such as the C programming language and
nesC [15] do not provide any tools to help the programmer manage the imple-
mentation of explicit state machines.
In this paper we study how protothreads, a novel programming abstrac-
tion that provides a conditional blocking wait operation, can be used to reduce
the number of explicit state machines in event-driven programs for memory-
constrained embedded systems.
The contribution of this paper is that we show that protothreads simplify
event-driven programming by reducing the need for explicit state machines.
We show that the protothreads mechanism is simple enough that a prototype
implementation of the protothreads mechanism can be done using only C lan-
guage constructs, without any architecture-specific machine code. We have
previously presented the ideas behind protothreads in a position paper [13].
In this paper we significantly extend our previous work by refining the pro-
114 Paper D

tothreads mechanism as well as quantifying and evaluating the utility of pro-

tothreads.
To evaluate protothreads, we analyze a number of widely used event-driven
programs by rewriting them using protothreads. We use three metrics to quan-
tify the effect of protothreads: the number of explicit state machines, the num-
ber of explicit state transitions, and lines of code. Our measurements show
that protothreads reduce all three metrics for all the rewritten programs. For
most programs the explicit state machines can be entirely removed. For the
other programs protothreads significantly reduce the number of states. Com-
pared to a state machine, the memory overhead of protothreads is a single byte.
The memory overhead of protothreads is significantly lower than for traditional
multi-threading. The execution time overhead of protothreads over a state ma-
chine is a few processor cycles.
We do not advocate protothreads as a general replacement for state ma-
chines. State machines are a powerful tool for designing, modeling, and ana-
lyzing embedded systems. They provide a well-founded formalism that allows
reasoning about systems and in some cases can provide proofs of the behavior
of the system. There are, however, many cases where protothreads can greatly
simplify the program without introducing any appreciable memory overhead.
Specifically, we have seen many programs for event-driven systems that are
based on informally specified state machines. The state machines for those
programs are in many cases only visible in the program code and are difficult
to extract from the code.
We originally developed protothreads for managing the complexity of ex-
plicit state machines in the event-driven uIP embedded TCP/IP stack [10]. The
prototype implementations of protothreads presented in this paper are also used
in the Contiki operating system [12] and have been used by at least ten differ-
ent third-party embedded developers for a range of different embedded devices.
Examples include an MPEG decoding module for Internet TV-boxes, wireless
sensors, and embedded devices collecting data from charge-coupled devices.
The implementations have also been ported by others to C++ [30] and Objec-
tive C [23].
The rest of the paper is structured as follows. Section 10.2 describes pro-
tothreads and shows a motivating example. In Section 10.3 we discuss the
memory requirements of protothreads. Section 10.4 shows how state machines
can be replaced with protothreads. Section 10.5 describes how protothreads are
implemented and presents a prototype implementation in the C programming
language. In Section 10.6 we evaluate protothreads, followed by a discussion
in Section 10.7. We review of related work in Section 10.8. Finally, the paper
10.2 Protothreads 115

is concluded in Section 10.9.

10.2 Protothreads
Protothreads are a novel programming abstraction that provides a conditional
blocking wait statement, PT WAIT UNTIL(), that is intended to simplify
event-driven programming for memory-constrained embedded systems. The
operation takes a conditional statement and blocks the protothread until the
statement evaluates to true. If the conditional statement is true the first time
the protothread reaches the PT WAIT UNTIL() the protothread continues to
execute without interruption. The PT WAIT UNTIL() condition is evaluated
each time the protothread is invoked. The PT WAIT UNTIL() condition can
be any conditional statement, including complex Boolean expressions.
A protothread is stackless: it does not have a history of function invoca-
tions. Instead, all protothreads in a system run on the same stack, which is
rewound every time a protothread blocks.
A protothread is driven by repeated calls to the function in which the pro-
tothread runs. Because they are stackless, protothreads can only block at the
top level of the function. This means that it is not possible for a regular func-
tion called from a protothread to block inside the called function - only explicit
PT WAIT UNTIL() statements can block. The advantage of this is that the
programmer always is aware of which statements that potentially may block.
Nevertheless, it is possible to perform nested blocking by using hierarchical
protothreads as described in Section 10.2.5.
The beginning and the end of a protothread are declared with
PT BEGIN and PT END statements. Protothread statements, such as the
PT WAIT UNTIL() statement, must be placed between the PT BEGIN and
PT END statements. A protothread can exit prematurely with a PT EXIT state-
ment. Statements outside of the PT BEGIN and PT END statements are not
part of the protothread and the behavior of such statements are undefined.
Protothreads can be seen as a combination of events and threads. From
threads, protothreads have inherited the blocking wait semantics. From events,
protothreads have inherited the stacklessness and the low memory overhead.
The blocking wait semantics allow linear sequencing of statements in event-
driven programs. The main advantage of protothreads over traditional threads
is that protothreads are very lightweight: a protothread does not require its
own stack. Rather, all protothreads run on the same stack and context switch-
ing is done by stack rewinding. This is advantageous in memory constrained
116 Paper D

systems, where a thread’s stack might use a large part of the available mem-
ory. For example, a thread with a 200 byte stack running on an MS430F149
microcontroller uses almost 10% of the entire RAM. In contrast, the memory
overhead of a protothread is as low as two bytes per protothread and no addi-
tional stack is needed.

10.2.1 Scheduling

The protothreads mechanism does not specify any specific method to invoke or
schedule a protothread; this is defined by the system using protothreads. If a
protothread is run on top of an underlying event-driven system, the protothread
is scheduled whenever the event handler containing the protothread is invoked
by the event scheduler. For example, application programs running on top
of the event-driven uIP TCP/IP stack are invoked both when a TCP/IP event
occurs and when the application is periodically polled by the TCP/IP stack. If
the application program is implemented as a protothread, this protothread is
scheduled every time uIP calls the application program.
In the Contiki operating system, processes are implemented as protothreads
running on top of the event-driven Contiki kernel. A process’ protothread is
invoked whenever the process receives an event. The event may be a message
from another process, a timer event, a notification of sensor input, or any other
type of event in the system. Processes may wait for incoming events using the
protothread conditional blocking statements.
The protothreads mechanism does not specify how memory for holding the
state of a protothread is managed. As with the scheduling, the system using
protothreads decides how memory should be allocated. If the system will run a
predetermined amount of protothreads, memory for the state of all protothreads
can be statically allocated in advance. Memory for the state of a protothread
can also be dynamically allocated if the number of protothreads is not known in
advance. In Contiki, the memory for the state of a process’ protothread is held
in the process control block. Typically, a Contiki program statically allocates
memory for its process control blocks.
In general, protothreads are reentrant. Multiple protothreads can be running
the same piece of code as long as each protothread has its own memory for
keeping state.
10.2 Protothreads 117

state: {ON, WAITING, OFF }

radio wake eventhandler:

if (state = ON)
if (expired(timer))
timer ← tsleep
if (not communication complete())
state ← WAITING
wait timer ← twait max
else
radio off()
state ← OFF
elseif (state = WAITING)
if (communication complete() or
expired(wait timer))
state ← OFF
radio off()
elseif (state = OFF )
if (expired(timer))
radio on()
state ← ON
timer ← tawake

Figure 10.1: The radio sleep cycle implemented with events, in pseudocode.

10.2.2 Protothreads as Blocking Event Handlers

Protothreads can be seen as blocking event handlers in that protothreads can
run on top of an existing event-based kernel, without modifications to the un-
derlying event-driven system. Protothreads running on top of an event-driven
system can use the PT WAIT UNTIL() statement to block conditionally. The
underlying event dispatching system does not need to know whether the event
handler is a protothread or a regular event handler.
In general, a protothread-based implementation of a program can act as a
drop-in replacement a state machine-based implementation without any modi-
fications to the underlying event dispatching system.

10.2.3 Example: Hypothetical MAC Protocol

To illustrate how protothreads can be used to replace state machines for event-
driven programming, we consider a hypothetical energy-conserving sensor net-
work MAC protocol. One of the tasks for a sensor network MAC protocol is to
118 Paper D

radio wake protothread:

PT BEGIN
while (true)
radio on()
timer ← tawake
PT WAIT UNTIL(expired(timer))
timer ← tsleep
if (not communication complete())
wait timer ← twait max
PT WAIT UNTIL(communication complete() or
expired(wait timer))
radio off()
PT WAIT UNTIL(expired(timer))
PT END

Figure 10.2: The radio sleep cycle implemented with protothreads, in pseu-
docode.

allow the radio to be turned off as often as possible in order to reduce the over-
all energy consumption of the device. Many MAC protocols therefore have
scheduled sleep cycles when the radio is turned off completely.
The hypothetical MAC protocol used here is similar to the T-MAC proto-
col [34] and switches the radio on and off at scheduled intervals. The mecha-
nism is depicted in Figure 10.3 and can be specified as follows:

1. Turn radio on.

2. Wait until t = t0 + tawake .

t wait_max

t awake t sleep

keep on if communication

Radio ON
off if
no
comm.
Radio OFF
t t + t awake + t sleep
0 0

Figure 10.3: Hypothetical sensor network MAC protocol.

10.2 Protothreads 119

ON
Remaining communication

Timer expired Timer expired WAITING

Timer expired
OFF

Figure 10.4: State machine realization of the radio sleep cycle of the example
MAC protocol.

3. Turn radio off, but only if all communication has completed.

4. If communication has not completed, wait until it has completed or t =
t0 + tawake + twait max .
5. Turn the radio off. Wait until t = t0 + tawake + tsleep .
6. Repeat from step 1.

To implement this protocol in an event-driven model, we first need to iden-

tify a set of states around which the state machine can be designed. For this
protocol, we can quickly identify three states: ON – the radio is on, WAITING
– waiting for remaining communication to complete, and OFF – the radio is off.
Figure 10.4 shows the resulting state machine, including the state transitions.
To implement this state machine, we use an explicit state variable, state,
that can take on the values ON, WAITING, and OFF . We use an if statement
to perform different actions depending on the value of the state variable. The
code is placed in an event handler function that is called whenever an event
occurs. Possible events in this case are an expiration of a timer and the com-
pletion of communication. To simplify the code, we use two separate timers,
timer and wait timer, to keep track of the elapsed time. The resulting pseu-
docode is shown in Figure 10.1.
We note that this simple mechanism results in a fairly large amount of
code. The code that controls the state machine constitutes more than one third
of the total lines of code. Also, the six-step structure of the mechanism is not
immediately evident from the code.
When implementing the radio sleep cycle mechanism with protothreads
we can use the PT WAIT UNTIL() statement to wait for the timers to expire.
120 Paper D

Figure 10.2 shows the resulting pseudocode code. We see that the code is
shorter than the event-driven version from Figure 10.1 and that the code more
closely follows the specification of the mechanism.

10.2.4 Yielding Protothreads

Experience with rewriting event-driven state machines to protothreads revealed
the importance of an unconditional blocking wait, PT YIELD(). PT YIELD()
performs an single unconditional blocking wait that temporarily blocks the pro-
tothread until the next time the protothread is invoked. At the next invocation
the protothread continues executing the code following the PT YIELD() state-
ment.
With the addition of the PT YIELD() operation, protothreads are similar to
stackless coroutines, much like cooperative multi-threading is similar to stack-
ful coroutines.

10.2.5 Hierarchical Protothreads

While many programs can be readily expressed with a single protothread, more
complex operations may need to be decomposed in a hierarchical fashion. Pro-
tothreads support this through an operation, PT SPAWN(), that initializes a
child protothread and blocks the current protothread until the child protothread
has either ended with PT END or exited with PT EXIT. The child protothread
is scheduled by the parent protothread; each time the parent protothread is in-
voked by the underlying system, the child protothread is invoked through the
PT SPAWN() statement. The memory for the state of the child protothread
typically is allocated in a local variable of the parent protothread.
As a simple example of how hierarchical protothreads work, we consider
a hypothetical data collection protocol that runs in two steps. The protocol
first propagates data interest messages through the network. It then continues
to propagate data messages back to where the interest messages came from.
Both interest messages and data messages are transmitted in a reliable way:
messages are retransmitted until an acknowledgment message is received.
Figure 10.5 shows this protocol implemented using hierarchi-
cal protothreads. The program consists of a main protothread,
data collection protocol, that invokes a child protothread, reliable send,
to do transmission of the data.
10.2 Protothreads 121

reliable send(message):
rxtimer: timer
PT BEGIN
do
rxtimer ← tretransmission
send(message)
PT WAIT UNTIL(ack received() or expired(rxtimer))
until (ack received())
PT END

data collection protocol

child state: protothread state
PT BEGIN
while (running)
while (interests left to relay())
PT WAIT UNTIL(interest message received())
send ack()
PT SPAWN(reliable send(interest), child state)
while (data left to relay())
PT WAIT UNTIL(data message received())
send ack()
PT SPAWN(reliable send(data), child state)
PT END

Figure 10.5: Hypothetical data collection protocol implemented with hierar-

chical protothreads, in pseudocode.

10.2.6 Local Continuations

Local continuations are the low-level mechanism that underpins protothreads.
When a protothread blocks, the state of the protothread is stored in a local con-
tinuation. A local continuation is similar to ordinary continuations [31] but,
unlike a continuation, a local continuation does not capture the program stack.
Rather, a local continuation only captures the state of execution inside a single
function. The state of execution is defined by the continuation point in the func-
tion where the program is currently executing and the values of the function’s
local variables. The protothreads mechanism only requires that those variables
that are actually used across a blocking wait to be stored. However, the current
C-based prototype implementations of local continuations depart from this and
do not store any local variables.
A local continuation has two operations: set and resume. When a local
continuation is set, the state of execution is stored in the local continuation.
122 Paper D

This state can then later be restored with the resume operation. The state cap-
tured by a local continuation does not include the history of functions that have
called the function in which the local continuation was set. That is, the lo-
cal continuation does not contain the stack, but only the state of the current
function.
A protothread consists of a function and a single local continuation. The
protothread’s local continuation is set before each PT WAIT UNTIL() state-
ment. If the condition is false and the wait is to be performed, the protothread
is suspended by returning control to the function that invoked the protothread’s
function. The next time the protothread function is invoked, the protothread
resumes the local continuation. This effectively causes the program to execute
a jump to the conditional blocking wait statement. The condition is reevaluated
and either blocks or continues its execution.

10.3 Memory Requirements

Programs written with an event-driven state machine need to store the state
of the state machine in a variable in memory. The state can be stored in a
single byte unless the state machine has more than 256 states. While the ac-
tual program typically stores additional state as program variables, the single
byte needed for storing the explicit state constitutes the memory overhead of
the state machine. The same program written with protothreads also needs to
store the same program variables, and will therefore require exactly the same
amount memory as the state machine implementation. The only additional
memory overhead is the size of the continuation point. For the prototype C-
based implementations, the size of the continuation point is two bytes on the
MSP430 and three bytes for the AVR.
In a multi-threading system each thread requires its own stack. Typically,
in memory-constrained systems this memory must be statically reserved for
the thread and cannot be used for other purposes, even when the thread is not
currently executing. Even for systems with dynamic stack memory allocation,
thread stacks usually are over-provisioned because of the difficulties of pre-
dicting the maximum stack usage of a program, For example, the default stack
size for one thread in the Mantis system [2] is 128 bytes, which is a large part
of the memory in a system with a few kilobytes of RAM.
In contrast to multi-threading, for event-driven state machines and pro-
tothreads all programs run on the same stack. The minimum stack memory
requirement is therefore the same as the maximum stack usage of all pro-
10.4 Replacing State Machines with
Protothreads 123

Events, Threads
protothreads
Stack size
1 2 3 1 2 3

Figure 10.6: The stack memory requirements for three event handlers, the three
event handlers rewritten with protothreads, and the equivalent functions run-
ning in three threads. Event handlers and protothreads run on the same stack,
whereas each thread runs on a stack of its own.

grams. The minimum memory requirement for stacks in a multi-threaded sys-

tem, however, is the sum of the maximum stack usage of all threads. This is
illustrated in Figure 10.6.

10.4 Replacing State Machines with

Protothreads
We analyzed a number of existing event-driven programs and found that most
control-flow state machines could be decomposed to three primitive patterns:
sequences, iterations, and selections. While our findings hold for a number of

a) b) c)

cond1 cond1 condition

cond2
cond2a cond2b

Figure 10.7: Two three primitive state machines: a) sequence, b) iteration, c)

selection.
124 Paper D

an iteration:
a sequence: PT BEGIN
PT BEGIN (* ... *)
(* ... *) while (cond1)
PT WAIT UNTIL(cond1) PT WAIT UNTIL(cond1 or
(* ... *) cond2)
PT END (* ... *)
PT END

Figure 10.8: Pseudocode implementation of the sequence and iteration patterns

with protothreads.

a selection:
PT BEGIN
(* ... *)
if (condition)
PT WAIT UNTIL(cond2a)
else
PT WAIT UNTIL(cond2b)
(* ... *)
PT END

Figure 10.9: Pseudocode implementation of the selection pattern with a pro-

tothread.

memory-constrained sensor network and embedded programs, our findings are

not new in general; Behren et al. [35] found similar results when examining
several event-driven systems. Figure 10.7 shows the three primitives. In this
section, we show how these state machine primitives map onto protothread
constructs and how those can be used to replace state machines.
Figures 10.8 and 10.9 show how to implement the state machine patterns
with protothreads. Protothreads allow the programmer to make use of the con-
trol structures provided by the programming language: the selection and itera-
tion patterns map onto if and while statements.
To rewrite an event-driven state machine with protothreads, we first analyse
the program to find its state machine. We then map the state machine patterns
from Figure 10.7 onto the state machine from the event-driven program. When
the state machine patterns have been identified, the program can be rewritten
using the code patterns in Figures 10.8 and 10.9.
As an illustration, Figure 10.10 shows the state machine from the radio
10.5 Implementation 125

Iteration
Selection

Remaining communication

Timer expired Timer expired

Timer expired

Sequence

Figure 10.10: The state machine from the example radio sleep cycle mecha-
nism with the iteration and sequence patterns identified.

sleep cycle of the example MAC protocol in Section 10.2.3, with the itera-
tion and sequence state machine patterns identified. From this analysis the
protothreads-based code in Figure 10.2 can be written.

10.5 Implementation

We have developed two prototype implementations of protothreads that use

only the C preprocessor. The fact that the implementations only depend on the
C preprocessor adds the benefit of full portability across all C compilers and
of not requiring extra tools in the compilation tool chain. However, the im-
plementations depart from the protothreads mechanism in two important ways:
automatic local variables are not saved across a blocking wait statement and
C switch and case statements cannot be freely intermixed with protothread-
based code. These problems can be solved by implementing protothreads as a
special precompiler or by integrating protothreads into existing preprocessor-
based languages and C language extensions such as nesC [15].
126 Paper D

struct pt { lc_t lc };
#define PT_WAITING 0
#define PT_EXITED 1
#define PT_ENDED 2
#define PT_INIT(pt) LC_INIT(pt->lc)
#define PT_BEGIN(pt) LC_RESUME(pt->lc)
#define PT_END(pt) LC_END(pt->lc); \
return PT_ENDED
#define PT_WAIT_UNTIL(pt, c) LC_SET(pt->lc); \
if(!(c)) \
return PT_WAITING
#define PT_EXIT(pt) return PT_EXITED

Figure 10.11: C preprocessor implementation of the main protothread opera-

tions.

10.5.1 Prototype C Preprocessor

Implementations
In the prototype C preprocessor implementation of protothreads the pro-
tothread statements are implemented as C preprocessor macros that are shown
in Figure 10.11. The protothread operations are a very thin layer of code on
top of the local continuation mechanism. The set and resume operations of the
local continuation are implemented as an LC SET() and the an LC RESUME()
macro. The prototype implementations of LC SET() and LC RESUME() de-
part from the mechanism specified in Section 10.2.6 in that automatic variables
are not saved, but only the continuation point of the function.
The PT BEGIN() statement, which marks the start of a protothread, is im-
plemented with a single LC RESUME() statement. When a protothread func-
tion is invoked, the LC RESUME() statement will resume the local continu-
ation stored in the protothread’s state structure, thus performing an uncondi-
tional jump to the last place where the local continuation was set. The resume
operation will not perform the jump the first time the protothread function is
invoked.
The PT WAIT UNTIL() statement is implemented with a LC SET() op-
eration followed by an if statement that performs an explicit return if the
conditional statement evaluates to false. The returned value lets the caller know
that the protothread blocked on a PT WAIT UNTIL() statement. PT END()
and PT EXIT() immediately return to the caller.
To implement yielding protothreads, we need to change the implementa-
10.5 Implementation 127

#define PT_BEGIN(pt) { int yielded = 1; \

LC_RESUME(pt->lc)
#define PT_YIELD(pt) yielded = 0; \
PT_WAIT_UNTIL(pt, yielded)
#define PT_END(pt) LC_END(pt->lc); \
return PT_ENDED; }

Figure 10.12: Implementation of the PT YIELD() operation and the updated

PT BEGIN() and PT END() statements.

#define PT_SPAWN(pt, child, thread) \

PT_INIT(child); \
PT_WAIT_UNTIL(pt, thread != PT_WAITING)

Figure 10.13: Implementation of the PT SPAWN() operation

tion of PT BEGIN() and PT END() in addition to implementing PT YIELD().

The implementation of PT YIELD() needs to test whether the protothread has
yielded or not. If the protothread has yielded once, then the protothread should
continue executing after the PT YIELD() statement. If the protothread has not
yet yielded, it should perform a blocking wait. To implement this, we add an
automatic variable, which we call yielded for the purpose of this discus-
sion, to the protothread. The yielded variable is initialized to one in the
PT BEGIN() statement. This ensures that the variable will be initialized every
time the protothread is invoked. In the implementation of PT YIELD(), we set
the variable to zero, and perform a PT WAIT UNTIL() that blocks until the
variable is non-zero. The next time the protothread is invoked, the conditional
statement in the PT WAIT UNTIL() is reevaluated. Since the yielded vari-
able now has been reinitialized to one, the PT WAIT UNTIL() statement will
not block. Figure 10.12 shows this implementation of PT YIELD() and the
updated PT BEGIN() and PT END() statements.
The implementation of PT SPAWN(), which is used to implement hier-
archical protothreads, is shown in Figure 10.13. It initializes the child pro-
tothread and invokes it every time the current protothread is invoked. The
PT WAIT UNTIL() blocks until the child protothread has exited or ended.
We now discuss how the local continuation functions LC SET() and
LC RESUME() are implemented.
128 Paper D

typedef void * lc_t;

#define LC_INIT(c) c = NULL
#define LC_RESUME(c) if(c) goto *c
#define LC_SET(c) { __label__ r; r: c = &&r; }
#define LC_END(c)

Figure 10.14: Local continuations implemented with the GCC labels-as-values

C extension.

GCC C Language Extensions

The widely used GCC C compiler provides a special C language extension that
makes the implementation of the local continuation operations straightforward.
The C extension, called labels-as-values, makes it possible to save the address
of a C label in a pointer. The C goto statement can then be used to jump to the
previously captured label. This use of the goto operation is very similar to the
unconditional jump most machine code instruction sets provide.
With the labels-as-values C extension, a local continuation simply is a
pointer. The set operation takes the address of the code executing the oper-
ation by creating a C label and capturing its address. The resume operation
resumes the local continuation with the C goto statement, but only if the local
continuation previously has been set. The implementation of local continua-
tions with C macros and the labels-as-values C language extension is shown
in Figure 10.14. The LC SET() operation uses the GCC label extension
to declare a C label that is local in scope. It then defines the label and stores
the address of the label in the local continuation by using the GCC double-
ampersand extension.

C Switch Statement
The main problem with the GCC C extension-based implementation of local
continuations is that it only works with a single C compiler: GCC. We next
show an implementation using only standard ANSI C constructs which uses
the C switch statement in a non-obvious way.
Figure 10.15 shows local continuations implemented using the C switch
statement. LC RESUME() is an open switch statement, with a case 0:
immediately following it. The case 0: makes sure that the code after
the LC RESUME() statement is always executed when the local continua-
tion has been initialized with LC INIT(). The implementation of LC SET()
10.5 Implementation 129

typedef unsigned short lc_t;

#define LC_INIT(c) c = 0
#define LC_RESUME(c) switch(c) { case 0:
#define LC_SET(c) c = __LINE__; case __LINE__:
#define LC_END(c) }

Figure 10.15: Local continuations implemented with the C switch statement.

1 int sender(pt) { int sender(pt) {

2 PT_BEGIN(pt); switch(pt->lc) {
3 case 0:
4 /* ... */ /* ... */
5 do { do {
6 pt->lc = 8;
7 PT_WAIT_UNTIL(pt, case 8:
8 cond1); if(!cond1)
9 return PT_WAITING;
10 } while(cond); } while(cond);
11 /* ... */ /* ... */
12 PT_END(pt); }
13 return PT_ENDED;
14 } }

Figure 10.16: Expanded C code with local continuations implemented with the
C switch statement.

uses the standard LINE macro. This macro expands to the line number in
the source code at which the LC SET() macro is used. The line number is
used as a unique identifier for each LC SET() statement. The implementation
of LC END() is a single right curly bracket that closes the switch statement
opened by LC RESUME().
To better illustrate how the C switch-based implementation works, Fig-
ure 10.16 shows how a short protothreads-based program is expanded by the C
preprocessor. We see that the resulting code is fairly similar to how the explicit
state machine was implemented in Figure 10.1. However, when looking closer
at the expanded C code, we see that the case 8: statement on line 7 appears
inside the do-while loop, even though the switch statement appears outside of
the do-while loop. This does seem surprising at first, but is in fact valid ANSI
C code. This use of the switch statement is likely to first have been publicly
described by Duff as part of Duff’s Device [8]. The same technique has later
130 Paper D

been used by Tatham to implement coroutines in C [33].

10.5.2 Memory Overhead

The memory required for storing the state of a protothread, implemented ei-
ther with the GCC C extension or the C switch statement, is two bytes; the C
switch statement-based implementation requires two bytes to store the 16-bit
line number identifier of the local continuation. The C extension-based im-
plementation needs to store a pointer to the address of the local continuation.
The size of a pointer is processor-dependent but on the MSP430 a pointer is
16 bits, resulting in a two byte memory overhead. A pointer on the AVR is
24 bits, resulting in three bytes of memory overhead. However, the memory
overhead is an artifact of the prototype implementations; a precompiler-based
implementation would reduce the overhead to one byte.

10.5.3 Limitations of the Prototype Implementations

The two implementations of the local continuation mechanism described above
introduce the limitation that automatic variables are not saved across a blocking
wait. The C switch-based implementation also limits the use of the C switch
statement together with protothread statements.

Automatic Variables

In the C-based prototype implementations, automatic variables—variables

with function-local scope that are automatically allocated on the stack—are
not saved in the local continuation across a blocking wait. While automatic
variables can still be used inside a protothread, the contents of the variables
must be explicitly saved before executing a wait statement. Many C compilers,
including GCC, detect if automatic local variables are used across a blocking
protothreads statement and issues a warning message.
While automatic variables are not preserved across a blocking wait, static
local variables are preserved. Static local variables are variables that are local
in scope but allocated in the data section of the memory rather than on the stack.
Since static local variables are not placed on the stack, they are not affected by
the use of blocking protothreads statements. For functions that do not need to
be reentrant, static local variables allow the programmer to use local variables
inside the protothread.
10.5 Implementation 131

For reentrant protothreads, the limitation on the use of automatic variables

can be handled by using an explicit state object, much in the same way as is
commonly done in purely event-driven programs. It is, however, the responsi-
bility of the programmer to allocate and maintain such a state object.

Constraints on Switch Constructs

The implementation of protothreads using the C switch statements imposes
a restriction on programs using protothreads: programs cannot utilize switch
statements together with protothreads. If a switch statement is used by the
program using protothreads, the C compiler will in some cases emit an error,
but in most cases the error is not detected by the compiler. This is troublesome
as it may lead to unexpected run-time behavior which is hard to trace back
to an erroneous mixture of one particular implementation of protothreads and
switch statements. We have not yet found a suitable solution for this problem
other than using the GCC C extension-based implementation of protothreads.

Possible C Compiler Problems

It could be argued that the use of a non-obvious, though standards-compliant,
C construct can cause problems with the C compiler because the nested switch
statement may not be properly tested. We have, however, tested protothreads
on a wide range of C compilers and have only found one compiler that was
not able to correctly parse the nested C construct. In this case, we contacted
the vendor who was already aware of the problem and immediately sent us
an updated version of the compiler. We have also been in touch with other C
compiler vendors, who have all assured us that protothreads work with their
product.

10.5.4 Alternative Approaches

In addition to the implementation techniques described above, we examine two
alternative implementation approaches: implementation with assembly lan-
guage and with the C language functions setjmp and longjmp.

Assembly Language
We have found that for some combinations of processors and C compilers it
is possible to implement protothreads and local continuations by using assem-
bly language. The set of the local continuations is then implemented as a C
132 Paper D

function that captures the return address from the stack and stores it in the lo-
cal continuation, along with any callee save registers. Conversely, the resume
operation would restore the saved registers from the local continuation and
perform an unconditional jump to the address stored in the local continuation.
The obvious problem with this approach is that it requires a porting effort for
every new processor and C compiler. Also, since both a return address and
a set of registers need to be stored in the local continuation, its size grows.
However, we found that the largest problem with this approach is that some C
compiler optimizations will make the implementation difficult. For example,
we were not able to produce a working implementation with this method for
the Microsoft Visual C++ compiler.

With C setjmp and longjmp Functions

While it at first seems possible to implement the local continuation operations
with the setjmp and longjmp functions from the standard C library, we have
seen that such an implementation causes subtle problems. The problem is be-
cause the setjmp and longjmp function store and restore the stack pointer, and
not only the program counter. This causes problems when the protothread is
invoked through different call paths since the stack pointer is different with
different call paths. The resume operation would not correctly resume a local
continuation that was set from a different call path.
We first noticed this when using protothreads with the uIP TCP/IP stack.
In uIP application protothreads are invoked from different places in the TCP/IP
code depending on whether or not a TCP retransmission is to take place.

Stackful Approaches
By letting each protothread run on its own stack it would be possible to im-
plement the full protothread mechanism, including storage of automatic vari-
ables across a blocking wait. With such an implementation the stack would
be switched to the protothread’s own stack by the PT BEGIN operation and
switched back when the protothread blocks or exits. This approach could be
implemented with a coroutine library or the multi-threading library of Contiki.
However, this implementation would result in a memory overhead similar to
that of multi-threading because each invocation of a protothread would require
the same amount of stack memory as the equivalent protothread running in
a thread of its own due to the stack space required by functions called from
within the protothread.
10.6 Evaluation 133

Finally, a promising alternative method is to store a copy the stack frame of

the protothread function in the local continuation when the protothread blocks.
This saves all automatic variables of the protothread function across a block-
ing wait, including variables that are not used after the blocking wait. Since
all automatic variables are saved, this approach have a higher memory over-
head. Furthermore, this approach requires both C compiler-specific and CPU
architecture-specific code, thus reducing the portability of the implementation.
However, the extra porting effort may be outweighed by the benefits of storing
automatic variables across blocking waits. We will continue to pursue this as
future work.

10.6 Evaluation
To evaluate protothreads we first measure the reduction in code complexity
that protothreads provide by reimplementing a set of event-driven programs
with protothreads and measure the complexity of the resulting code. Second,
we measure the memory overhead of protothreads compared to the memory
overhead of an event-driven state machine. Third, we compare the execution
time overhead of protothreads with that of event-driven state machines.

10.6.1 Code Complexity Reduction

To measure the code complexity reduction of protothreads we reimplement
parts of a number of event-driven applications with protothreads: XNP [20],
the previous default over-the-air programming program from TinyOS; the
buffer management module of TinyDB [27], a database engine for TinyOS;
radio protocol drivers for the Chipcon CC1000 and RF Monolithics TR1001
radio chips; the SMTP client in the uIP embedded TCP/IP stack and a code
propagation program from the Contiki operating system. The state machines in
XNP, TinyDB, and the CC1000 drivers were rewritten by applying the method
for replacing state machines with protothreads from Section 10.4 whereas the
TR1001 driver, the uIP SMTP client and the Contiki code propagation were
rewritten from scratch.
We use three metrics to measure the complexity of the programs we reim-
plemented with protothreads: the number of explicit states, the number of ex-
plicit state transitions, as well as the lines of code of the reimplemented func-
tions.
All reimplemented programs consist of complex state machines. Using
134 Paper D

protothreads, we were able to entirely remove the explicit state machines for
most programs. For all programs, protothreads significantly reduce the number
of state transitions and lines of code.
The reimplemented programs have undergone varying amounts of testing.
The Contiki code propagation, the TR1001 low-level radio driver, and the uIP
SMTP client are well tested and are currently used on a daily basis in live
systems, XNP and TinyDB have been verified to be working but not heavily
tested, and the CC1000 drivers have been tested and run in simulation.
Furthermore, we have anecdotal evidence to support our hypothesis that
protothreads are an alternative to state machines for embedded software devel-
opment. The protothreads implementations have for some time been available
as open source on our web page [9]. We know that at least ten embedded sys-
tems developers have successfully used protothreads to replace state machines
for embedded software development. Also, our protothreads code have twice
been recommended by experienced embedded developers in Jack Ganssle’s
embedded development newsletter [14].

XNP
XNP [20] is one of the in-network programming protocols used in
TinyOS [19]. XNP downloads a new system image to a sensor node and writes
the system image to the flash memory of the device. XNP is implemented on
top of the event-driven TinyOS. Therefore, any operations in XNP that would
be blocking in a threaded system have to be implemented as state machines.
We chose XNP because it is a relatively complex program implemented on top
of an event-driven system. The implementation of XNP has previously been
analyzed by Jeong [20], which assisted us in our analysis. The implementation
of XNP consists of a large switch statement with 25 explicit states, encoded
as defined constants, and 20 state transitions. To analyze the code, we identi-
fied the state transitions from manual inspection of the code inside the switch
statement.
Since the XNP state machine is implemented as one large switch statement,
we expected it to be a single, complex state machine. But, when drawing the
state machine from analysis of the code, it turned out that the switch statement
in fact implements five different state machines. The entry points of the state
machines are not immediately evident from the code, as the state of the state
machine was changed in several places throughout the code.
The state machines we found during the analysis of the XNP program are
shown in Figure 10.17. For reasons of presentation, the figure does not show
10.6 Evaluation 135

DL_START REQ_CIDMISSING

DL_SRECWRITE UP_SRECWRITE
DL_START0 GET_CIDMISSING

DL_START1 ISP_REQ GETNEXTCID

DL_END EEFLASH_WRITE

DL_START2
DL_END_SIGNAL EEFLASH_WRITEDONE ISP_REQ1 GET_DONE

DL_FAIL

DL_FAIL_SIGNAL

Figure 10.17: XNP state machines. The names of the states are from the code.
The IDLE and ACK states are not shown.

the IDLE and ACK states. Almost all states have transitions to one of these
states. If an XNP operation completes successfully, the state machine goes
into the ACK state to transmit an acknowledgment over the network. The IDLE
state is entered if an operation ends with an error, and when the acknowledg-
ment from the ACK state has been transmitted.
In the figure we clearly see many of the state machine patterns from Fig-
ure 10.7. In particular, the sequence pattern is evident in all state machines. By
using the techniques described in Section 10.4 we were able to rewrite all state
machines into protothreads. Each state machine was implemented as its own
protothread.
The IDLE and ACK states are handled in a hierarchical protothread. A
separate protothread is created for sending the acknowledgment signal. This
protothread is spawned from the main protothread every time the program logic
dictates that an acknowledgment should be sent.

TinyDB

TinyDB [27] is a small database engine for the TinyOS system. With TinyDB,
a user can query a wireless sensor network with a database query language
similar to SQL. TinyDB is one of the largest TinyOS programs available.
In TinyOS long-latency operations are split-phase [15]. Split-phase opera-
tions consist of two parts: a request and a completion event. The request com-
pletes immediately, and the completion event is posted when the operation has
completed. TinyDB contains a large number of split-phase operations. Since
programs written for TinyOS cannot perform a blocking wait, many complex
136 Paper D

appendBufferTask loadBufferTask readEEPROMRow

WRITING_LENGTHS READ_OPEN READ_ROW

WRITING_NAME READ_LENGTHS READING_LENGTH

WRITING_QUERY ALLOC_NAME ALLOC_FOR_READ

WRITING_BUFFER READ_NAME READING_DATA

SKIP_BYTES

WRITE_FIELD_DATA ALLOC_QUERY READING_DATA

WRITE_FIELD_LEN

WRITE_NEXT_BUFFER READ_QUERY

READ_BUFFER

READ_FIELD_LEN

ALLOC_FIELD_DATA
READ_FIELD_DATA

Figure 10.18: Three state machines from TinyDB.

operations in TinyDB are encoded as state machines.

To the state machines in TinyDB we analyze the TinyDB buffer manage-
ment module, DBBufferC. DBBufferC uses the MemAlloc module to allocate
memory. Memory allocation requests are performed from inside a function
that drives the state machine. However, when the request is completed, the al-
locComplete event is handled by a different function. This event handler must
handle the event different depending on the state of the state machine. In fact,
the event handler itself implements a small piece of the entire state machine.
The fact that the implementation of the state machine is distributed across dif-
ferent functions makes the analysis of the state machine difficult.
From inspection of the DBBufferC code we found the three state machines
in Figure 10.18. We also found that there are more state machines in the code,
but we were not able to adequately trace them because the state transitions were
scattered around the code. By rewriting the discovered state machines with
protothreads, we were able to completely remove the explicit state machines.

Low Level Radio Protocol Drivers

The Chipcon CC1000 and RF Monolithics TR1001 radio chips are used in
many wireless sensor network devices. Both chips provide a very low-level
interface to the radio. The chips do not perform any protocol processing them-
selves but interrupt the CPU for every incoming byte. All protocol functional-
10.6 Evaluation 137

ity, such as packet framing, header parsing, and MAC protocol must be imple-
mented in software.
We analyze and rewrite CC1000 drivers from the Mantis OS [2] and from
SOS [17], as well as the TR1001 driver from Contiki [12]. All drivers are
implemented as explicit state machines. The state machines run in the interrupt
handlers of the radio interrupts.
The CC1000 driver in Mantis has two explicit state machines: one for han-
dling and parsing incoming bytes and one for handling outgoing bytes. In
contrast, both the SOS CC1000 driver and the Contiki TR1001 drivers have
only one state machine that parses incoming bytes. The state machine that
handles transmissions in the SOS CC1000 driver is shown in Figure 10.19.
The structures of the SOS CC1000 driver and the Contiki TR1001 driver are
very similar.

TXSTATE_PREAMBLE

TXSTATE_SYNC

TXSTATE_PREHEADER

TXSTATE_HEADER

TXSTATE_DATA

TXSTATE_CRC

TXSTATE_FLUSH

TXSTATE_WAIT_FOR_ACK

TXSTATE_READ_ACK

TXSTATE_DONE

Figure 10.19: Transmission state machine from the SOS CC1000 driver.

With protothreads we could replace most parts of the state machines. How-
ever, for both the SOS CC1000 driver and the Contiki TR1001 drivers, we kept
a top-level state machine. The reason for this is that those state machines were
138 Paper D

not used to implement control flow. The top-level state machine in the SOS
CC1000 driver controlled if the driver was currently transmitting or receiving
a packet, or if it was finding a synchronization byte.

uIP TCP/IP Stack

The uIP TCP/IP stack [10] is designed for memory-constrained embedded sys-
tems and therefore has a very low memory overhead. It is used in embedded
devices from well over 30 companies, with applications ranging from pico-
satellites to car traffic monitoring systems. To reduce the memory overhead
uIP follows the event-driven model. Application programs are implemented as
event handlers and to achieve blocking waits, application programs need to be
written as explicit state machines. We have rewritten the SMTP client in uIP
with protothreads and were able to completely remove the state machines.

Contiki
The Contiki operating system [12] for wireless sensor networks is based on an
event-driven kernel, on top of which protothreads provide a thread-like pro-
gramming style. The first version of Contiki was developed before we intro-
duced protothreads. After developing protothreads, we found that they reduced
the complexity of writing software for Contiki.
For the purpose of this paper, we measure the implementation of a dis-
tribution program for distributing and receiving binary code modules for the
Contiki dynamic loader [11]. The program was initially implemented with-
out protothreads but was later rewritten when protothreads were introduced to
Contiki. The program can be in one of three modes: (1) receiving a binary
module from a TCP connection and loads it into the system, (2) broadcasting
the binary module over the wireless network, and (3) receiving broadcasts of a
binary module from a nearby node and loading it into memory.
When rewriting the program with protothreads, we removed most of the
explicit state machines, but kept four states. These states keep track in which
mode the program is: if it is receiving or broadcasting a binary module.

Results
The results of reimplementing the programs with protothreads are presented in
Table 10.1. The lines of code reported in the table are those of the rewritten
functions only. We see that in all cases the number of states, state transitions,
and lines of code were reduced by rewriting the programs with protothreads. In
10.6 Evaluation 139

States, States, Transitions, Transitions,

Program before after before after
XNP 25 - 20 -
TinyDB 23 - 24 -
Mantis CC1000 driver 15 - 19 -
SOS CC1000 driver 26 9 32 14
Contiki TR1001 driver 12 3 32 3
uIP SMTP client 10 - 10 -
Contiki code propagation 6 4 11 3
Lines of Lines of Reduction,
Program code, before code, after percentage
XNP 222 152 32%
TinyDB 374 285 24%
Mantis CC1000 driver 164 127 23%
SOS CC1000 driver 413 348 16%
Contiki TR1001 driver 152 77 49%
uIP SMTP client 223 122 45%
Contiki code propagation 204 144 29%

Table 10.1: The number of explicit states, explicit state transitions, and lines
of code before and after rewriting with protothreads.

most cases the rewrite completely removed the state machine. The total average
reduction in lines of code is 31%. For the programs rewritten by applying
the replacement method from Section 10.4 (XNP, TinyDB, and the CC1000
drivers) the average reduction is 23% and for the programs that were rewritten
from scratch (the TR1001 driver, the uIP SMTP client, and the Contiki code
propagation program) the average reduction is 41%.
Table 10.2 shows the compiled code size of the rewritten functions when
written as a state machine and with protothreads. We see that the code size
increases in most cases, except for the Contiki code propagation program. The
average increase for the programs where the state machines were replaced with
protothreads by applying the method from Section 10.4 is 14%. The Con-
tiki TR1001 driver is only marginally larger when written with protothreads.
The uIP SMTP client, on the other hand, is significantly larger when written
with protothreads rather than with a state machine. The reason for this is that
the code for creating SMTP message strings could be optimized through code
reuse in the state machine-based implementation, something which was not
140 Paper D

Code size, Code size,

before after
Program (bytes) (bytes) Increase
XNP 931 1051 13%
TinyDB DBBufferC 2361 2663 13%
Mantis CC1000 994 1170 18%
SOS CC1000 1912 2165 13%
Contiki TR1001 823 836 2%
uIP SMTP 1106 1901 72%
Contiki code prop. 1848 1426 -23%

Table 10.2: Code size before and after rewriting with protothreads.

possible in the protothreads-based implementation without significantly sac-

rificing readability. In contrast with the uIP SMTP client, the Contiki code
propagation program is significantly smaller when written with protothreads.
Here, it was possible to optimize the protothreads-based code by code reuse,
which was not readily possible in the state machine-based implementation.
We conclude that protothreads reduce the number of states and state tran-
sitions and the lines of code, at the price of an increase in code size. The size
of the increase, however, depends on the properties of the particular program
rewritten with protothreads. It is therefore not possible to draw any general
conclusions from measured increase in code size.

10.6.2 Memory Overhead

We measured the stack size required for the Contiki TR1001 driver and the
Contiki code propagation mechanism running on the MSP430 microcontroller.
We measure the stack usage by filling the stack with a known byte pattern,
running the TR1001 driver, and inspecting the stack to see how much of the
byte pattern that was overwritten.
Neither the Contiki code propagation mechanism nor the Contiki TR1001
driver use the stack for keeping state in program variables. Therefore the entire
stack usage of the two programs when running as threads is overhead.
We compare the memory overhead of the Contiki TR1001 driver and the
Contiki code propagation running as threads with the memory overhead of a
state machine and protothreads in Table 10.3.
10.6 Evaluation 141

State Proto-
machine thread Thread
Contiki TR1001 driver 1 2 18
Contiki code propagation 1 2 34

Table 10.3: Memory overhead in bytes for the Contiki TR1001 driver and the
Contiki code propagation on the MSP430, implemented with a state machine,
a protothread, and a thread.

State Proto- Yielding

machine thread protothread
MSP430 9 12 17
AVR 23 34 45

Table 10.4: Machine code instructions overhead for a state machine, a pro-
tothread, and a yielding protothread.

10.6.3 Run-time Overhead

To evaluate the run-time overhead of protothreads we counted the machine
code instruction overhead of protothreads, compared to a state machine, for
the MSP430 and the AVR microcontrollers. Furthermore, we measured the
execution time for the driver for the TR1001 radio chip from Contiki, imple-
mented both as a state machine and as a protothread. We also compare the
numbers with measurements on a cooperative multi-threading implementation
of said program. We used the cooperative user space multi-threading library
from Contiki [12] which is a standard stack-switching multi-threading library.

Machine Code Instruction Overhead

To analyze the number of additional machine code instructions in the execution
path for protothreads compared to state machines, we compiled our program
using the GCC C compiler version 3.2.3 for the MSP430 microcontroller and
GCC version 3.4.3 for the AVR microcontroller.
With manual inspection of the generated object machine code we counted
the number of machine code instructions needed for the switch and case state-
ments in a state machine function and for the protothread operations in a pro-
tothread function. The results are given in Table 10.4. The absolute overhead
142 Paper D

State Protothreads, Protothreads, C

Compiler machine GCC C extension switch statement
optimization (ms) (ms) (ms)
Size (-Os) 0.0373 0.0397 0.0434
Speed (-O1) 0.0369 0.0383 0.0415
State Protothreads, Protothreads, C
Compiler machine GCC C extension switch statement
optimization (cycles) (cycles) (cycles)
Size (-Os) 91.67 97.56 106.7
Speed (-O1) 90.69 94.12 102.0

Table 10.5: Mean execution time in milliseconds and processor cycles for a
single invocation of the TR1001 input driver under Contiki on the MSP430
platform.

for protothreads over the state machine is very small: three machine code in-
structions for the MSP430 and 11 for the AVR. In comparison, the number of
instructions required to perform a context switch in the Contiki implementation
of cooperative multi-threading is 51 for the MSP430 and 80 for the AVR.
The additional instructions for the protothread in Table 10.4 are caused
by the extra case statement that is included in the implementation of the
PT BEGIN operation.

Execution Time Overhead

To measure the execution time overhead of protothreads over that of an event-
driven state machine, we implemented the low-level radio protocol driver for
the RFM TR1001 radio chip from Contiki using both an event-driven state ma-
chine and a yielding protothread. We measured the execution time by feeding
the driver with data from 1000 valid radio packets and measured the average
execution time of the driver’s function. The time was measured using a pe-
riodic timer interrupt driven by the MSP430 timer A1 at a rate of 1000 Hz.
We set the clock speed of the MSP430 digitally controlled oscillator to 2.4576
MHz. For the protothreads-based implementation we measured both the GCC
C extension-based and the C switch statement-based implementations of the
local continuations operations.
The measurement results are presented in Table 10.5. We see that the av-
erage execution time overhead of protothreads is low: only about five cycles
10.7 Discussion 143

per invocation for the GCC C extension-based implementation and just over
ten cycles per invocation for the C switch-based implementation. The results
are consistent with the machine code overhead in Table 10.4. We also mea-
sured the execution time of the radio driver rewritten with cooperative multi-
threading and found it to be approximately three times larger than that of the
protothread-based implementation because of the overhead of the stack switch-
ing code in the multi-threading library.
Because of the low execution time overhead of protothreads we conclude
that protothreads are usable even for interrupt handlers with tight execution
time constraints.

10.7 Discussion
We have been using the prototype implementations of protothreads described
in this paper in Contiki for two years and have found that the biggest problem
with the prototype implementations is that automatic variables are not pre-
served across a blocking wait. Our workaround is to use static local variables
rather than automatic variables inside protothreads. While the use of static lo-
cal variables may be problematic in the general case, we have found it to work
well for Contiki because of the small scale of most Contiki programs. Also,
as many Contiki programs do not need to be reentrant, the use of static local
variables work well.
Code organization is different for programs written with state machines
and with protothreads. State machine-based programs tend to consist either of
a single large function containing a large state machine or of many small func-
tions where the state machine is difficult to find. On the contrary, protothreads-
based programs tend to be based around a single protothread function that con-
tains the high-level logic of the program. If the underlying event system calls
different functions for every incoming event, a protothreads-based program
typically consists of a single protothread function and a number of small event
handlers that invoke the protothread when an event occurs.

10.8 Related Work

Research in the area of software development for sensor networks has led to a
number of new abstractions that aim at simplifying sensor network program-
ming [1, 4]. Approaches with the same goal include virtual machines [25] and
macro-programming of sensors [16, 29, 37]. Protothreads differ from these
144 Paper D

sensor network programming abstractions in that we target the difficulty of

low-level event-driven programming rather than the difficulty of developing
application software for sensor networks.
Kasten and Römer [21] have also identified the need for new abstractions
for managing the complexity of event-triggered state machine programming.
They introduce OSM, a state machine programming model based on Harel’s
StateCharts[18] and use the Esterel language. The model reduces both the
complexity of the implementations and the memory usage. Their work is dif-
ferent from protothreads in that they help programmers manage their state ma-
chines, whereas protothreads are designed to reduce the number of state ma-
chines. Furthermore, OSM requires support from an external OSM compiler to
produce the resulting C code, whereas the prototype implementations of pro-
tothreads only make use of the regular C preprocessor.
Simpson [32] describes a cooperative mini-kernel mechanism for C++
which is very similar to our protothreads in that it was designed to replace
state machines. Simpson’s mechanism also uses a single stack. However, it
needs to implement its own stack to track the location of a yield point. In con-
trast, protothreads do not use a stack, but hold all their state in a 16-bit integer
value.
Lauer and Needham [24] proved, essentially, that events and threads are
duals of each other and that the same program can be written for either of the
two systems. With protothreads, we put this into practice by actually rewriting
event-driven programs with blocking wait semantics.
Protothreads are similar to coroutines [22] in the sense that a protothread
continues to execute at the point of the last return from the function. In partic-
ular, protothreads are similar to asymmetric coroutines [7], just as cooperative
multi-threading is similar to asymmetric coroutines. However, unlike corou-
tines and cooperative multi-threading, protothreads are stackless and can only
block at the top level of the protothread function.
Dabek et al. [6] present libasynch, a C++ library that assists the program-
mer in writing event-driven programs. The library provides garbage collection
of allocated memory as well as a type-safe way to pass state between callback
functions. However, the libasynch library does not provide sequential execu-
tion and software written with the library cannot use language statements to
control program flow across blocking calls.
The Capriccio system by von Behren et al. [36] shows that in a memory-
rich environment, threads can be made as memory efficient as events. This
requires modification to the C compiler so that it performs dynamic memory
allocation of stack space during run-time. However, as dynamic memory al-
10.9 Conclusions 145

location quickly causes memory fragmentation in the memory-constrained de-

vices for which the protothreads mechanism is designed, dynamic allocation
of stack memory is not feasible.
Cunningham and Kohler [5] develop a library to assist programmers of
event-driven systems in program analysis, along with a tool for visualizing
callback chains, as well as a tool for verifying properties of programs imple-
mented using the library. Their work is different from ours in that they help the
programmer manage the state machines for event-driven programs, whereas
protothreads are designed to replace such state machines.
Adya et al. [3] discuss the respective merits of event-driven and threaded
programming and present a hybrid approach that shows that the event-driven
and multi-threaded code can coexist in a unified concurrency model. The au-
thors have developed a set of adaptor functions that allows event-driven code
to call threaded code, and threaded code to call event-driven code, without
requiring that the caller has knowledge about the callee’s approach.
State machines are a powerful tool for developing real-time systems. Ab-
stractions such as Harel’s StateCharts [18] are designed to help developers to
develop and manage state machines and are very valuable for systems that are
designed as state machines. Protothreads, in contrast, are intended to replace
state machines with sequential C code. Also, protothreads do not provide any
mechanisms for assisting development of hard real-time systems.

10.9 Conclusions
We present protothreads, a novel abstraction for memory-constrained embed-
ded systems. Due to memory-constraints, such systems are often based on an
event-driven model. Experience has shown that event-driven programming is
difficult because the lack of a blocking wait abstraction forces programmers to
implement control flow with state machines.
Protothreads simplify programming by providing a conditional blocking
wait operation, thereby reducing the need for explicit state machines. Pro-
tothreads are inexpensive: the memory overhead is only two bytes per pro-
tothread.
We develop two prototype protothreads implementations using only C pre-
processor and evaluate the usefulness of protothreads by reimplementing some
widely used event-driven programs using protothreads. Our results show that
for most programs the explicit state machines could be entirely removed. Fur-
thermore, protothreads significantly reduce the number of state transitions and
146 Paper D

lines of code. The execution time overhead of protothreads is on the order of

a few processor cycles. We find the code size of a program written with pro-
tothreads to be slightly larger than the equivalent program written as a state
machine.

Acknowledgments
This work was partly financed by VINNOVA, the Swedish Agency for Inno-
vation Systems, and the European Commission under contract IST-004536-
RUNES. Thanks go to Kay Römer and Umar Saif for reading and suggesting
improvements on drafts of this paper, and to our paper shepherd Philip Levis
for his many insightful comments that significantly helped to improve the pa-
per.
Bibliography

[1] T. Abdelzaher, J. Stankovic, S. Son, B. Blum, T. He, A. Wood, and C. Lu.

A communication architecture and programming abstractions for real-
time embedded sensor networks. In Workshop on Data Distribution for
Real-Time Systems, Providence, RI, USA, May 2003.
[2] H. Abrach, S. Bhatti, J. Carlson, H. Dai, J. Rose, A. Sheth, B. Shucker,
J. Deng, and R. Han. Mantis: system support for multimodal networks of
in-situ sensors. In Proceedings of the 2nd ACM international conference
on Wireless sensor networks and applications, pages 50–59, San Diego,
CA, USA, September 2003.
[3] A. Adya, J. Howell, M. Theimer, W. J. Bolosky, and J. R. Douceur. Co-
operative Task Management Without Manual Stack Management. In Pro-
ceedings of the USENIX Annual Technical Conference, pages 289–302,
2002.
[4] E. Cheong, J. Liebman, J. Liu, and F. Zhao. TinyGALS: A programming
model for event-driven embedded systems. In Proc. of the 18th Annual
ACM Symposium on Applied Computing (SAC’03), Melbourne, Florida,
USA, March 2003.
[5] R. Cunningham and E. Kohler. Making events less slippery with eel. In
Proceedings of the 10th Workshop on Hot Topics in Operating Systems
(HotOS-X), Santa Fee, New Mexico, June 2005. IEEE Computer Society.
[6] F. Dabek, N. Zeldovich, F. Kaashoek, D. Mazières, and R. Morris. Event-
driven programming for robust software. In Proceedings of the 2002
SIGOPS European Workshop, Saint-Emilion, France, September 2002.
[7] A. L. de Moura and R. Ierusalimschy. Revisiting coroutines. MCC 15/04,
PUC-Rio, Rio de Janeiro, RJ, June 2004.

147
148 Bibliography

[8] T. Duff. Unwinding loops. Usenet news article, net.lang.c, Message-ID:

<[email protected]>, May 1984.

[9] A. Dunkels. Protothreads web site. Web page. Visited 2006-04-06.

URL: http://www.sics.se/˜adam/pt/

[10] A. Dunkels. Full TCP/IP for 8-bit architectures. In Proceedings of The

First International Conference on Mobile Systems, Applications, and Ser-
vices (MOBISYS ‘03), San Francisco, California, May 2003.

[11] A. Dunkels, N. Finne, J. Eriksson, and T. Voigt. Run-time dynamic link-

ing for reprogramming wireless sensor networks. In Proceedings of the
4th International Conference on Embedded Networked Sensor Systems,
SenSys 2006, Boulder, Colorado, USA, 2006.

[12] A. Dunkels, B. Grönvall, and T. Voigt. Contiki - a lightweight and flexible

operating system for tiny networked sensors. In Proceedings of the First
IEEE Workshop on Embedded Networked Sensors (IEEE Emnets ’04),
Tampa, Florida, USA, November 2004.

[13] A. Dunkels, O. Schmidt, and T. Voigt. Using protothreads for sensor node
programming. In Proc. of the Workshop on Real-World Wireless Sensor
Networks (REALWSN’05), Stockholm, Sweden, June 2005.

[14] J. Ganssle. The embedded muse. Monthly newsletter.

URL: http://www.ganssle.com/tem-back.htm

[15] D. Gay, P. Levis, R. von Behren, M. Welsh, E. Brewer, and D. Culler.

[16] R. Gummadi, O. Gnawali, and R. Govindan. Macro-programming wire-

less sensor networks using kairos. In Proc. of Distributed Computing in
Sensor Systems (DCOSS)’05, Marina del Rey, CA, USA, June 2005.

[17] C. Han, R. K. Rengaswamy, R. Shea, E. Kohler, and M. Srivastava. Sos:

A dynamic operating system for sensor networks. In Proceedings of the
3rd international conference on Mobile systems, applications, and ser-
vices (MobiSys ’05), Seattle, WA, USA, June 2005.
Bibliography 149

[18] D. Harel. Statecharts: A visual formalism for complex systems. Science

of Computer Programming, 8(3):231–274, 1987.

[19] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler, and K. Pister. Sys-

[20] J. Jeong. Analysis of xnp network reprogramming module. Web page,

October 2003. Visited 2006-04-06.
URL: http://www.cs.berkeley.edu/˜jaein/cs294 1/xnp anal.htm

[21] O. Kasten and K. Römer. Beyond event handlers: Programming wire-

less sensors with attributed state machines. In The Fourth International
Conference on Information Processing in Sensor Networks (IPSN), Los
Angeles, USA, April 2005.

[22] D. E. Knuth. The art of computer programming, volume 1: fundamental

algorithms (2nd edition). Addison-Wesley Longman Publishing Co., Inc.,
Boston, MA, USA, 1978.

[23] Framework Labs. Protothreads for Objective-C/Cocoa. Visited 2006-04-

06.
URL: http://www.frameworklabs.de/protothreads.html

[24] H. C. Lauer and R. M. Needham. On the duality of operating systems

structures. In Proc. Second International Symposium on Operating Sys-
tems, October 1978.

[25] P. Levis and D. Culler. Maté: A tiny virtual machine for sensor networks.
In Proceedings of ASPLOS-X, San Jose, CA, USA, October 2002.

[26] P. Levis, S. Madden, D. Gay, J. Polastre, R. Szewczyk, A. Woo,

E. Brewer, and D. Culler. The Emergence of Networking Abstractions
and Techniques in TinyOS. In Proceedings of ACM/Usenix Networked
Systems Design and Implementation (NSDI’04), San Francisco, Califor-
nia, USA, March 2004.

[27] S. R. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong. Tinydb: an

acquisitional query processing system for sensor networks. ACM Trans-
actions on Database Systems, 30(1):122–173, 2005.
[28] M. Melkonian. Get by Without an RTOS. Embedded Systems Program-
ming, 13(10), September 2000.
[29] R. Newton, Arvind, and M. Welsh. Building up to macroprogramming:
An intermediate language for sensor networks. In Proc. IPSN’05, Los
Angeles, CA, USA, April 2005.
[30] J. Paisley and J. Sventek. Real-time detection of grid bulk transfer traffic.
In Proceedings of the 10th IEEE/IFIP Network Operations Management
Symposium, Vancouver, Canada, April 2006.
[31] J. C. Reynolds. The discoveries of continuations. Lisp and Symbolic
Computation, 6(3):233–247, 1993.
[32] Z. B. Simpson. State machines: Cooperative mini-kernels with yield-
ing. In Computer Game Developer’s Conference, Austin, TX, November
1999.
[33] S. Tatham. Coroutines in C. Web page, 2000.
URL: http://www.chiark.greenend.org.uk/˜sgtatham/coroutines.html

[34] T. van Dam and K. Langendoen. An adaptive energy-efficient MAC pro-

tocol for wireless sensor networks. In Proceedings of the first interna-
tional conference on Embedded networked sensor systems, pages 171–
180, Los Angeles, California, USA, 2003.
[35] R. von Behren, J. Condit, and E. Brewer. Why events are a bad idea (for
high-concurrency servers). In Proceedings of the 9th Workshop on Hot
Topics in Operating Systems, Lihue (Kauai), Hawaii, USA, May 2003.
[36] R. von Behren, J. Condit, F. Zhou, G. C. Necula, and E. Brewer. Capric-
cio: scalable threads for internet services. In Proc. SOSP ’03, pages
268–281, 2003.
[37] M. Welsh and G. Mainland. Programming Sensor Networks Using Ab-
stract Regions. In Proceedings of ACM/Usenix Networked Systems De-
sign and Implementation (NSDI’04), San Francisco, California, USA,
March 2004.
Chapter 11

Paper E: Run-time Dynamic

Linking for Reprogramming
Wireless Sensor Networks

Adam Dunkels, Niclas Finne, Joakim Eriksson, and Thiemo Voigt. Run-time
dynamic linking for reprogramming wireless sensor networks. In Proceedings
of the 4th International Conference on Embedded Networked Sensor Systems
(ACM SenSys 2006), Boulder, Colorado, USA, November 2006.
c 2006 Association for Computing Machinery.

151
Abstract

From experience with wireless sensor networks it has become apparent that
dynamic reprogramming of the sensor nodes is a useful feature. The resource
constraints in terms of energy, memory, and processing power make sensor
network reprogramming a challenging task. Many different mechanisms for
reprogramming sensor nodes have been developed ranging from full image
replacement to virtual machines.
We have implemented an in-situ run-time dynamic linker and loader that
use the standard ELF object file format. We show that run-time dynamic link-
ing is an effective method for reprogramming even resource constrained wire-
less sensor nodes. To evaluate our dynamic linking mechanism we have im-
plemented an application-specific virtual machine and a Java virtual machine
and compare the energy cost of the different linking and execution models. We
measure the energy consumption and execution time overhead on real hardware
to quantify the energy costs for dynamic linking.
Our results suggest that while in general the overhead of a virtual machine
is high, a combination of native code and virtual machine code provide good
energy efficiency. Dynamic run-time linking can be used to update the native
code, even in heterogeneous networks.
11.1 Introduction 153

11.1 Introduction
Wireless sensor networks consist of a collection of programmable radio-
equipped embedded systems. The behavior of a wireless sensor network is
encoded in software running on the wireless sensor network nodes. The soft-
ware in deployed wireless sensor network systems often needs to be changed,
both to update the system with new functionality and to correct software bugs.
For this reason dynamically reprogramming of wireless sensor network is an
important feature. Furthermore, when developing software for wireless sen-
sor networks, being able to update the software of a running sensor network
greatly helps to shorten the development time.
The limitations of communication bandwidth, the limited energy of the
sensor nodes, the limited sensor node memory which typically is on the order
of a few thousand bytes large, the absence of memory mapping hardware, and
the limited processing power make reprogramming of sensor network nodes
challenging.
Many different methods for reprogramming sensor nodes have been de-
veloped, including full system image replacement [14, 16], approaches based
on binary differences [15, 17, 31], virtual machines [18, 19, 20], and loadable
native code modules in the first versions of Contiki [5] and SOS [12]. These
methods are either inefficient in terms of energy or require non-standard data
formats and tools.
The primary contribution of this paper is that we investigate the use of stan-
dard mechanisms and file formats for reprogramming sensor network nodes.
We show that in-situ dynamic run-time linking and loading of native code us-
ing the ELF file format, which is a standard feature on many operating systems
for PC computers and workstations, is feasible even for resource-constrained
sensor nodes. Our secondary contribution is that we measure and quantify the
energy costs of dynamic linking and execution of native code and compare it to
the energy cost of transmission and execution of code for two virtual machines:
an application-specific virtual machine and the Java virtual machine.
We have implemented a dynamic linker in the Contiki operating system that
can link, relocate, and load standard ELF object code files. Our mechanism is
independent of the particular microprocessor architecture on the sensor nodes
and we have ported the linker to two different sensor node platforms with only
minor modifications to the architecture dependent module of the code.
To evaluate the energy costs of the dynamic linker we implement an ap-
plication specific virtual machine for Contiki together with a compiler for a
subset of Java. We also adapt the Java virtual machine from the lejOS sys-
154 Paper E

tem [8] to run under Contiki. We measure the energy cost of reprogramming
and executing a set of program using dynamic linking of native code and the
two virtual machines. Using the measurements and a simple energy consump-
tion model we calculate break-even points for the energy consumption of the
different mechanisms. Our results suggest that while the execution time over-
head of a virtual machine is high, a combination of native code and virtual
machine code may give good energy efficiency.
The remainder of this paper is structured as follows. In Section 11.2 we
discuss different scenarios in which reprogramming is useful. Section 11.3
presents a set of mechanisms for executing code inside a sensor node and in
Section 11.4 we discuss loadable modules and the process of linking, relocat-
ing, and loading native code. Section 11.5 describes our implementation of
dynamic linking and our virtual machines. Our experiments and the results are
presented in Section 11.6 and discuss the results in Section 11.7. Related work
is reviewed in Section 11.8. Finally, we conclude the paper in Section 11.9.

11.2 Scenarios for Software Updates

Software updates for sensor networks are necessary for a variety of reasons
ranging from implementation and testing of new features of an existing pro-
gram to complete reprogramming of sensor nodes when installing new appli-
cations. In this section we review a set of typical reprogramming scenarios and
compare their qualitative properties.

11.2.1 Software Development

Software development is an iterative process where code is written, installed,
tested, and debugged in a cyclic fashion. Being able to dynamically reprogram
parts of the sensor network system helps shorten the time of the development
cycle. During the development cycle developers typically change only one
part of the system, possibly only a single algorithm or a function. A sensor
network used for software development may therefore see large amounts of
small changes to its code.

11.2.2 Sensor Network Testbeds

Sensor network testbeds are an important tool for development and experimen-
tation with sensor network applications. New applications can be tested in a
11.2 Scenarios for Software Updates 155

realistic setting and important measurements can be obtained [36]. When a

new application is to be tested in a testbed the application typically is installed
in the entire network. The application is then run for a specified time, while
measurements are collected both from the sensors on the sensor nodes, and
from network traffic.
For testbeds that are powered from a continuous energy source, the energy
consumption of software updates is only of secondary importance. Instead,
qualitative properties such as ease of use and flexibility of the software update
mechanism are more important. Since the time required to make an update is
important, the throughput of a network-wide software update is of importance.
As the size of the transmitted binaries impact the throughput, the binary size
still can be used as an evaluation metric for systems where throughput is more
important than energy consumption.

11.2.3 Correction of Software Bugs

The need for correcting software bugs in sensor networks was early identi-
fied [7]. Even after careful testing, new bugs can occur in deployed sensor
networks caused by, for example, an unexpected combination of inputs or vari-
able link connectivity that stimulate untested control paths in the communica-
tion software [30].
Software bugs can occur at any level of the system. To correct bugs it must
therefore be possible to reprogram all parts of the system.

11.2.4 Application Reconfiguration

In an already installed sensor network, the application may need to be recon-
figured. This includes change of parameters, or small changes in the applica-
tion such as changing from absolute temperature readings to notification when
thresholds are exceeded [26]. Even though reconfiguration not necessarily in-
clude software updates [25], application reconfiguration can be done by repro-
gramming the application software. Hence software updates can be used in an
application reconfiguration scenario.

11.2.5 Dynamic Applications

There are many situations where it is useful to replace the application software
of an already deployed sensor network. One example is the forest fire detection
scenario presented by Fok et al. [9] where a sensor network is used to detect
156 Paper E

Update Update Update Program

Scenario frequency fraction level longevity
Development Often Small All Short
Testbeds Seldom Large All Long
Bug fixes Seldom Small All Long
Reconfig. Seldom Small App Long
Dynamic
Application Often Small App Long

Table 11.1: Qualitative comparison between different reprogramming scenar-

ios.

a fire. When the fire detection application has detected a fire, the fire fighters
might want to run a search and rescue application as well as a fire tracking
application. While it may possible to host these particular applications on each
node despite the limited memory of the sensor nodes, this approach is not scal-
able [9]. In this scenario, replacing the application on the sensor nodes leads
to a more scalable system.

11.2.6 Summary
Table 11.1 compares the different scenarios and their properties. Update frac-
tion refers to what amount of the system that needs to be updated for every
update, update level to at what levels of the system updates are likely to occur,
and program longevity to how long an installed program will be expected to
reside on the sensor node.

11.3 Code Execution Models and Reprogram-

ming
Many different execution models and environments have been developed or
adapted to run on wireless sensor nodes. Some with the notion of facilitat-
ing programming [1], others motivated by the potential of saving energy costs
for reprogramming enabled by the compact code representation of virtual ma-
chines [19]. The choice of the execution model directly impacts the data format
and size of the data that needs to be transported to a node. In this section we
discuss three different mechanisms for executing program code inside each
sensor node: script languages, virtual machines, and native code.
11.3 Code Execution Models and Reprogramming 157

11.3.1 Script Languages

There are many examples of script languages for embedded systems, including
BASIC variants, Python interpreters [22], and TCL machines [1]. However,
most script interpreters target platforms with much more resources than our
target platforms and we have therefore not included them in our comparison.

11.3.2 Virtual Machines

Virtual machines are a common approach to reduce the cost of transmitting
program code in situations where the cost of distributing a program is high.
Typically, program code for a virtual machine can be made more compact than
the program code for the physical machine. For this reason virtual machines
are often used for programming sensor networks [18, 19, 20, 23].
While many virtual machines such as the Java virtual machine are generic
enough to perform well for a variety of different types of programs, most virtual
machines for sensor networks are designed to be highly configurable in order
to allow the virtual machine to be tailored for specific applications. In effect,
this means that parts of the application code is implemented as virtual machine
code running on the virtual machine, and other parts of the application code is
implemented in native code that can be used from the programs running on the
virtual machine.

11.3.3 Native Code

The most straightforward way to execute code on sensor nodes is by running
native code that is executed directly by the microcontroller of the sensor node.
Installing new native code on a sensor node is more complex than installing
code for a virtual machine because the native code uses physical addresses
which typically need to be updated before the program can be executed. In this
section we discuss two widely used mechanisms for reprogramming sensor
nodes that execute native code: full image replacement and approaches based
on binary differences.

Full Image Replacement

The most common way to update software in embedded systems and sensor
networks is to compile a complete new binary image of the software together
158 Paper E

with the operating system and overwrite the existing system image of the sen-
sor node. This is the default method used by the XNP and Deluge network
reprogramming software in TinyOS [13].
The full image replacement does not require any additional processing of
the loaded system image before it is loaded into the system, since the loaded
image resides at the same, known, physical memory address as the previous
system image. For some systems, such as the Scatterweb system code [33], the
system contains both an operating system image and a small set of functions
that provide functionality for loading new operating system images. A new
operating system image can overwrite the existing image without overwriting
the loading functions. The addresses of the loading functions are hard-coded
in the operating system image.

Diff-based Approaches
Often a small update in the code of the system, such as a bugfix, will cause
only minor differences between in the new and old system image. Instead of
distributing a new full system image the binary differences, deltas, between
the modified and original binary can be distributed. This reduces the amount
of data that needs to be transferred. Several types of diff-based approaches
have been developed [15, 17, 31] and it has been shown that the size of the
deltas produced by the diff-based approaches is very small compared to the
full binary image.

11.4 Loadable Modules

A less common alternative to full image replacement and diff-based approaches
is to use loadable modules to perform reprogramming. With loadable mod-
ules, only parts of the system need to be modified when a single program is
changed. Typically, loadable modules require support from the operating sys-
tem. Contiki and SOS are examples of systems that support loadable modules
and TinyOS is an example of an operating system without loadable module
support.
A loadable module contains the native machine code of the program that is
to be loaded into the system. The machine code in the module usually contains
references to functions or variables in the system. These references must be
resolved to the physical address of the functions or variables before the ma-
chine code can be executed. The process of resolving those references is called
11.4 Loadable Modules 159

Core

0x0237
int memcpy() {
/* ... */
}
0x1720
void radio_send() {
/* ... */
}

Pre−linked module Module with dynamic linking information

memcpy(); call 0x0237 memcpy(); call 0x0000

radio_send(); call 0x1720 radio_send(); call 0x0000

radio_send call instruction

memcpy call instruction

Figure 11.1: The difference between a pre-linked module and a module with
dynamic linking information: the pre-linked module contains physical ad-
dresses whereas the dynamically linked module contains symbolic names.

linking. Linking can be done either when the module is compiled or when the
module is loaded. We call the former approach pre-linking and the latter dy-
namic linking. A pre-linked module contains the absolute physical addresses
of the referenced functions or variables whereas a dynamically linked module
contains the symbolic names of all system core functions or variables that are
referenced in the module. This information increases the size of the dynam-
ically linked module compared to the pre-linked module. The difference is
shown in Figure 11.1. Dynamic linking has not previously been considered for
wireless sensor networks because of the perceived run-time overhead, both in
terms of execution time, energy consumption, and memory requirements.
The machine code in the module usually contains references not only to
functions or variables in the system, but also to functions or variables within the
module itself. The physical address of those functions will change depending
on the memory address at which the module is loaded in the system. The
addresses of the references must therefore be updated to the physical address
that the function or variable will have when the module is loaded. The process
160 Paper E

of updating these references is known as relocation. Like linking, relocation

can be done either at compile-time or at run-time.
When a module has been linked and relocated the program loader loads the
module into the system by copying the linked and relocated native code into a
place in memory from where the program can be executed.

11.4.1 Pre-linked Modules

The machine code of a pre-linked module contains absolute addresses of all
functions and variables in the system code that are referenced by the module.
Linking of the module is done at compile time and only relocation is performed
at run-time. To link a pre-linked module, information about the physical ad-
dresses of all functions and variables in the system into which the module is to
be loaded must be available at compile time.
There are two benefits of pre-linked modules over dynamically linked mod-
ules. First, pre-linked modules are smaller than dynamically linked modules
which results in less information to be transmitted. Second, the process of
loading a pre-linked module into the system is less complex than the process
of linking a dynamically linked module. However, the fact that all physical
addresses of the system core are hard-coded in the pre-linked module is a se-
vere drawback as a pre-linked module can only be loaded into a system with
the exact same physical addresses as the system that was to generate the list of
addresses that was used for linking the module.
In the original Contiki system [5] we used pre-linked binary modules for
dynamic loading. When compiling the Contiki system core, the compiler gen-
erated a map file containing the mapping between all globally visible functions
and variables in the system core and their addresses. This list of addresses was
used to pre-link Contiki modules.
We quickly noticed that while pre-linked binary modules worked well for
small projects with a homogeneous set of sensor nodes, the system quickly
became unmanageable when the number of sensor nodes grew. Even a small
change to the system core of one of the sensor nodes would make it impossible
to load binary a module into the system bedcase the addresses of variables and
functions in the core were different from when the program was linked. We
used version numbers to guard against this situation. Version numbers did help
against system crashes, but did not solve the general problem: new modules
could not be loaded into the system.
11.4 Loadable Modules 161

11.4.2 Dynamic Linking

With dynamic linking, the object files do not only contain code and data, but
also names of functions are variables of the system core that are referenced by
the module. The code in the object file cannot be executed before the physical
addresses of the referenced variables and functions have been filled in. This
process is done at run time by a dynamic linker.
In the Contiki dynamic linker we use two file formats for the dynamically
linked modules, ELF and Compact ELF.

ELF - Executable and Linkable Format

One of the most common object code format for dynamic linking is the Exe-
cutable and Linkable Format (ELF) [3]. It is a standard format for object files
and executables that is used for most modern Unix-like systems. An ELF ob-
ject file include both program code and data and additional information such
as a symbol table, the names of all external unresolved symbols, and relocation
tables. The relocation tables are used to locate the program code and data at
other places in memory than for which the object code originally was assem-
bled. Additionally, ELF files can hold debugging information such as the line
numbers corresponding to specific machine code instructions, and file names
of the source files used when producing the ELF object.
ELF is also the default object file format produced by the GCC utilities
and for this reason there are a number of standard software utilities for manip-
ulating ELF files available. Examples include debuggers, linkers, converters,
and programs for calculating program code and data memory sizes. These
utilities exist for a wide variety of platforms, including MS Windows, Linux,
Solaris, and FreeBSD. This is a clear advantage over other solutions such as
FlexCup [27], which require specialized utilities and tools.
Our dynamic linker in Contiki understands the ELF format and is able to
perform dynamic linking, relocation, and loading of ELF object code files. The
debugging features of the ELF format are not used.

CELF - Compact ELF

One problem with the ELF format is the overhead in terms of bytes to be
transmitted across the network, compared to pre-linked modules. There are
a number of reasons for the extra overhead. First, ELF, as any dynamically re-
locatable file format, includes the symbolic names of all referenced functions
or variables that need to be linked at run-time. Second, and more important,
162 Paper E

the ELF format is designed to work on 32-bit and 64-bit architectures. This
causes all ELF data structures to be defined with 32-bit data types. For 8-bit or
16-bit targets the high 16 bits of these fields are unused.
To quantify the overhead of the ELF format we devise an alternative to
the ELF object code format that we call CELF - Compact ELF. A CELF file
contains the same information as an ELF file, but represented with 8 and 16-bit
datatypes. CELF files typically are half the size of the corresponding ELF file.
The Contiki dynamic loader is able to load CELF files and a utility program is
used to convert ELF files to CELF files.
It is possible to further compress CELF files using lossless data compres-
sion. However, we leave the investigation of the energy-efficiency of this ap-
proach to future work.
The drawback of the CELF format is that it requires a special compres-
sor utility is for creating the CELF files. This makes the CELF format less
attractive for use in many real-world situations.

11.4.3 Position Independent Code

To avoid performing the relocation step when loading a module, it is in some
cases possible to compile the module into position independent code. Position
independent code is a type of machine code which does not contain any abso-
lute addresses to itself, but only relative references. This is the approach taken
by the SOS system.
To generate position independent code compiler support is needed. Further-
more, not all CPU architectures support position independent code and even
when supported, programs compiled to position independent code typically
are subject to size restrictions. For example, the AVR microcontroller supports
position independent code but restricts the size of programs to 4 kilobytes. For
the MSP430 no compiler is known to fully support position independent code.

11.5 Implementation
We have implemented run-time dynamic linking of ELF and CELF files in
the Contiki operating system [5]. To evaluate dynamic linking we have im-
plemented an application specific virtual machine for Contiki together with a
compiler for a subset of Java, and have ported a Java virtual machine to Contiki.
11.5 Implementation 163

ROM
11111111111
00000000000
00000000000
11111111111
00000000000
11111111111
Loaded program

00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
Device drivers
RAM
00000000000
11111111111
00000000000
11111111111
Language run−time

00000000000
11111111111
Loaded program Symbol table

00000000000
11111111111
Device drivers
Dynamic linker

Contiki kernel Contiki kernel

Core Core

Figure 11.2: Partitioning in Contiki: the core and loadable programs in RAM
and ROM.

11.5.1 The Contiki Operating System

The Contiki operating system was the first operating system for memory-
constrained sensor nodes to support dynamic run-time loading of native code
modules. Contiki is built around an event-driven kernel and has very low
memory requirements. Contiki applications run as extremely lightweight pro-
tothreads [6] that provide blocking operations on top of the event-driven kernel
at a very small memory cost. Contiki is designed to be highly portable and has
been ported to over ten different platforms with different CPU architectures
and using different C compilers.
A Contiki system is divided into two parts: the core and the loadable pro-
grams as shown in Figure 11.2. The core consists of the Contiki kernel, device
drivers, a set of standard applications, parts of the C language library, and a
symbol table. Loadable programs are loaded on top of the core and do not
modify the core.
The core has no information about the loadable programs, except for infor-
mation that the loadable programs explicitly register with the core. Loadable
programs, on the other hand, have full knowledge of the core and may freely
call functions and access variables that reside in the core. Loadable programs
can call each other by going through the kernel. The kernel dispatches calls
from one loaded program to another by looking up the target program in an
in-kernel list of active processes. This one-way dependency makes it possible
to load and unload programs at run-time without needing to patch the core and
164 Paper E

without the need for a reboot when a module has been loaded or unloaded.
While it is possible to replace the core at run-time by running a special
loadable program that overwrites the current core and reboots the system, ex-
perience has shown that this feature is not often used in practice.

11.5.2 The Symbol Table

The Contiki core contains a table of the symbolic names of all externally vis-
ible variable and function names in the Contiki core and their corresponding
addresses. The table includes not only the Contiki system, but also the C lan-
guage run-time library. The symbol table is used by the dynamic linker when
linking loaded programs.
The symbol table is created when the Contiki core binary image is com-
piled. Since the core must contain a correct symbol table, and a correct symbol
table cannot be created before the core exists, a three-step process is required to
compile a core with a correct symbol table. First, an intermediary core image
with an empty symbol table is compiled. From the intermediary core image an
intermediary symbol table is created. The intermediary symbol table contains
the correct symbols of the final core image, but the addresses of the symbols are
incorrect. Second, a second intermediary core image that includes the interme-
diary symbol table is created. This core image now contains a symbol table of
the same size as the one in the final core image so the addresses of all symbols
in the core are now as they will be in the final core image. The final symbol
table is then created from the second intermediary core image. This symbol
table contains both the correct symbols and their correct addresses. Third, the
final core image with the correct symbol table is compiled.
The process of creating a core image is automated through a simple make
script. The symbol table is created using a combination of standard ELF tools.
For a typical Contiki system the symbol table contains around 300 entries
which amounts to approximately 4 kilobytes of data stored in flash ROM.

11.5.3 The Dynamic Linker

We implemented a dynamic linker for Contiki that is designed to link, relocate,
and load either standard ELF files [3] and CELF, Compact ELF, files. The
dynamic linker reads ELF/CELF files through the Contiki virtual filesystem
interface, CFS, which makes the dynamic linker unaware of the physical loca-
tion of the ELF/CELF file. Thus the linker can operate on files stored either in
11.5 Implementation 165

RAM, on-chip flash ROM, external EEPROM, or external ROM without mod-
ification. Since all file access to the ELF/CELF file is made through the CFS,
the dynamic linker does not need to concern itself with low-level filesystem
details such as wear-leveling or fragmentation [4] as this is better handled by
the CFS.
The dynamic linker performs four steps to link, relocate and load an
ELF/CELF file. The dynamic linker first parses the ELF/CELF file and ex-
tracts relevant information about where in the ELF/CELF file the code, data,
symbol table, and relocation entries are stored. Second, memory for the code
and data is allocated from flash ROM and RAM, respectively. Third, the code
and data segments are linked and relocated to their respective memory loca-
tions, and fourth, the code is written to flash ROM and the data to RAM.
Currently, memory allocation for the loaded program is done using a sim-
ple block allocation scheme. More sophisticated allocation schemes will be
investigated in the future.

Linking and Relocating

The relocation information in an ELF/CELF file consists of a list of relocation
entries. Each relocation entry corresponds to an instruction or address in the
code or data in the module that needs to be updated with a new address. A relo-
cation entry contains a pointer to a symbol, such as a variable name or a func-
tion name, a pointer to a place in the code or data contained in the ELF/CELF
file that needs to be updated with the address of the symbol, and a relocation
type which specifies how the data or code should be updated. The relocation
types are different depending on the CPU architecture. For the MSP430 there
is only one single relocation type, whereas the AVR has 19 different relocation
types.
The dynamic linker processes a relocation entry at a time. For each relo-
cation entry, its symbol is looked up in the symbol table in the core. If the
symbol is found in the core’s symbol table, the address of the symbol is used
to patch the code or data to which the relocation entry points. The code or data
is patched in different ways depending on the relocation type and on the CPU
architecture.
If the symbol in the relocation entry was not found in the symbol table
of the core, the symbol table of the ELF/CELF file itself is searched. If the
symbol is found, the address that the symbol will have when the program has
been loaded is calculated, and the code or data is patched in the same way as if
the symbol was found in the core symbol table.
166 Paper E

Relocation entries may also be relative to the data, BSS, or code segment
in the ELF/CELF file. In that case no symbol is associated with the reloca-
tion entry. For such entries the dynamic linker calculates the address that the
segment will have when the program has been loaded, and uses that address to
patch the code or data.

When the linking and relocating is completed, the text and data have been re-
located to their final memory position. The text segment is then written to flash
ROM, at the location that was previously allocated. The memory allocated for
the data and BSS segments are used as an intermediate storage for transferring
text segment data from the ELF/CELF file before it is written to flash ROM.
Finally, the memory allocated for the BSS segment is cleared, and the contents
of the data segment is copied from the ELF/CELF file.

Executing the Loaded Program

When the dynamic linker has successfully loaded the code and data segments,
Contiki starts executing the program.
The loaded program may replace an already running Contiki service. If the
service that is to be replaced needs to pass state to the newly loaded service,
Contiki supports the allocation of an external memory buffer for this purpose.
However, experience has shown that this mechanism has been very scarcely
used in practice and the mechanism is likely to be removed in future versions
of Contiki.

Portability

Since the ELF/CELF format is the same across different platforms, we de-
signed the Contiki dynamic linker to be easily portable to new platforms. The
loader is split into one architecture specific part and one generic part. The
generic part parses the ELF/CELF file, finds the relevant sections of the file,
looks up symbols from the symbol table, and performs the generic relocation
logic. The architecture specific part does only three things: allocates ROM and
RAM, writes the linked and relocated binary to flash ROM, and understands
the relocation types in order to modify machine code instructions that need
adjustment because of relocation.
11.5 Implementation 167

Alternative Designs
The Contiki core symbol table contains all externally visible symbols in the
Contiki core. Many of the symbols may never need to be accessed by loadable
programs, thus causing ROM overhead. An alternative design would be to let
the symbol table include only a handful of symbols, entry points, that define
the only ways for an application program to interact with the core. This would
lead to a smaller symbol table, but would also require a detailed specification
of which entry points that should be included in the symbol table. The main
reason why we did not chose this design, however, is that we wish to be able to
replace modules at any level of the system. For this reason, we chose to provide
the same amount of symbols to an application program as it would have, would
it have been compiled directly into the core. However, we are continuing to
investigate this alternative design for future versions of the system.

11.5.4 The Java Virtual Machine

We ported the Java virtual machine (JVM) from lejOS [8], a small operating
system originally developed for the Lego Mindstorms. The Lego Mindstorms
are equipped with an Hitachi H8 microcontroller with 32 kilobytes of RAM
available for user programs such as the JVM. The lejOS JVM works within
this constrained memory while featuring preemptive threads, recursion, syn-
chronization and exceptions. The Contiki port required changes to the RAM-
only model of the lejOS JVM. To be able to run Java programs within the 2
kilobytes of RAM available on our hardware platform, Java classes needs to
be stored in flash ROM rather than in RAM. The Contiki port stores the class
descriptions including bytecode in flash ROM memory. Static class data and
class flags that denote if classes have been initialized are stored in RAM as
well as object instances and execution stacks. The RAM requirements for the
Java part of typical sensor applications are a few hundred bytes.
Java programs can call native code methods by declaring native Java meth-
ods. The Java virtual machine dispatches calls to native methods to native code.
Any native function in Contiki may be called, including services that are part
of a loaded Contiki program.

11.5.5 CVM - the Contiki Virtual Machine

We designed the Contiki Virtual Machine, CVS, to be a compromise between
an application-specific and a generic virtual machine. CVM can be configured
168 Paper E

for the application running on top of the machine by allowing functions to be

either implemented as native code or as CVM code. To be able to run the
same programs for the Java VM and for CVM, we developed a compiler that
compiles a subset of the Java language to CVM bytecode.
The design of CVM is intentionally similar to other virtual machines, in-
cluding Maté [19], VM⋆ [18], and the Java virtual machine. CVM is a stack-
based machine with separated code and data areas. The CVM instruction
set contains integer arithmetic, unconditional and conditional branches, and
method invocation instructions. Method invocation can be done in two ways,
either by invocation of CVM bytecode functions, or by invocation of functions
implemented in native code. Invocation of native functions is done through a
special instruction for calling native code. This instruction takes one parame-
ter, which identifies the native function that is to be called. The native function
identifiers are defined at compile time by the user that compiles a list of native
functions that the CVM program should be able to call. With the native func-
tion interface, it is possible for a CVM program to call any native functions
provided by the underlying system, including services provided by loadable
programs.
Native functions in a CVM program are invoked like any other function.
The CVM compiler uses the list of native functions to translate calls to such
functions into the special instruction for calling native code. Parameters are
passed to native functions through the CVM stack.

11.6 Evaluation
To evaluate dynamic linking of native code we compare the energy costs of
transferring, linking, relocating, loading, and executing a native code module
in ELF format using dynamic linking with the energy costs of transferring,
loading, and executing the same program compiled for the CVM and the Java
virtual machine. We devise a simple model of the energy consumption of the
reprogramming process. Thereafter we experimentally quantify the energy and
memory consumption as well as the execution overhead for the reprogram-
ming, the execution methods and the applications. We use the results of the
measurements as input into the model which enables us to perform a quantita-
tive comparison of the energy-efficiency of the reprogramming methods.
We use the ESB board [33] and the Telos Sky board [29] as our experi-
mental platforms. The ESB is equipped with an MSP430 microcontroller with
2 kilobytes of RAM and 60 kilobytes of flash ROM, an external 64 kilobyte
11.6 Evaluation 169

PROCESS_THREAD(test_blink, ev, data)

{
static struct etimer t;
PROCESS_BEGIN();

etimer_set(&t, CLOCK_SECOND);

while(1) {
leds_on(LEDS_GREEN);
PROCESS_WAIT_UNTIL(etimer_expired(&t));
etimer_reset(&t);

leds_off(LEDS_GREEN);
PROCESS_WAIT_UNTIL(etimer_expired(&t));
etimer_reset(&t);
}

PROCESS_END();
}

Figure 11.3: Example Contiki program that toggles the LEDs every second.

EEPROM, as well as a set of sensors and a TR1001 radio transceiver. The

Telos Sky is equipped with an MSP430 microcontroller with 10 kilobytes of
RAM and 48 kilobytes of flash ROM together with a CC2420 radio transceiver.
We use the ESB to measure the energy of receiving, storing, linking, relocat-
ing, loading and executing loadable modules and the Telos Sky to measure the
energy of receiving loadable modules.

We use three Contiki programs to measure the energy efficiency and exe-
cution overhead of our different approaches. Blinker, the first of the two pro-
grams, is shown in Figure 11.3. It is a simple program that toggles the LEDs
every second. The second program, Object Tracker, is an object tracking ap-
plication based on abstract regions [35]. To allow running the programs both
as native code, as CVM code, and as Java code we have implemented these
programs both in C and Java. A schematic illustration of the C implementation
is in Figure 11.4. To support the object tracker program, we implemented a
subset of the abstract regions mechanism in Contiki. The Java and CVM ver-
sions of the program call native code versions of the abstract regions functions.
The third program is a simple 8 by 8 vector convolution calculation.
170 Paper E

PROCESS_THREAD(use_regions_process, ev, data)

{
PROCESS_BEGIN();

while(1) {

value = pir_sensor.value();

region_put(reading_key, value);
region_put(reg_x_key, value * loc_x());
region_put(reg_y_key, value * loc_y());
if(value > threshold) {
max = region_max(reading_key);

if(max == value) {
sum = region_sum(reading_key);
sum_x = region_sum(reg_x_key);
sum_y = region_sum(reg_y_key);
centroid_x = sum_x / sum;
centroid_y = sum_y / sum;
send(centroid_x, centroid_y);
}
}

etimer_set(&t, PERIODIC_DELAY);
PROCESS_WAIT_UNTIL(etimer_expired(&t));

PROCESS_END();
}

Figure 11.4: Schematic implementation of an object tracker based on abstract

regions.

11.6.1 Energy Consumption

We model the energy consumption E of the reprogramming process with

E = Ep + Es + El + Ef

where Ep is the energy spent in transferring the object over the network, Es
the energy cost of storing the object on the device, El the energy consumed
by linking and relocating the object, and Ef the required energy for of storing
the linked program in flash ROM. We use a simplified model of the network
propagation energy where we assume a propagation protocol where the energy
consumption Ep is proportional to the size of the object to be transferred. For-
mally,
Ep = Pp so
11.6 Evaluation 171

20
Current

15
Current (mA)

0
0 0.2 0.4 0.6 0.8 1
Time (s)

Figure 11.5: Current draw for receiving 1000 bytes with the TR1001.

where so is the size of the object file to be transfered and Pp is a constant scale
factor that depends on the network protocol used to transfer the object. We
use similar equations for Es (energy for storing the binary) and El (energy for
linking and relocating). The equation for Ef (the energy for loading the binary
to ROM) contains the size of the compiled code size of the program instead of
the size of the object file. This model is intentionally simple and we consider
it good enough for our purpose of comparing the energy-efficiency of different
reprogramming schemes.

Lower Bounds on Radio Reception Energy

We measured the energy consumption of receiving data over the radio for two
different radio transceivers: the TR1001 [32], that is used on the ESB board,
and the CC2420 [2], that conforms to the IEEE 802.15.4 standard [11] and is
used on the Telos Sky board. The TR1001 provides a very low-level inter-
face to the radio medium. The transceiver decodes data at the bit level and
transmits the bits in real-time to the CPU. Start bit detection, framing, MAC
layer, checksums, and all protocol processing must be done in software running
on the CPU. In contrast, the interface provided by the CC2420 is at a higher
level. Start bits, framing, and parts of the MAC protocol are handled by the
transceiver. The software driver handles incoming and outgoing data on the
172 Paper E

20
Current

15
Current (mA)

0
0 0.2 0.4 0.6 0.8 1
Time (s)

Figure 11.6: Current draw for receiving 1000 bytes with the CC2420.

packet level.
Since the TR1001 operates at the bit-level, the communication speed of the
TR1001 is determined by the CPU. We use a data rate of 9600 bits per second.
The CC2420 has a data rate of 250 kilobits per second, but also incurs some
protocol overhead as it provides a more high-level interface.
Figures 11.5 and 11.6 show the current draw from receiving 1000 bytes
of data with the TR1001 and CC2420 radio transceivers. These measurements
constitute a lower bound on the energy consumption for receiving data over the
radio, as they do not include any control overhead caused by a code propagation
protocol. Nor do they include any packet headers. An actual propagation pro-
tocol would incur overhead because of both packet headers and control traffic.
For example, the Deluge protocol has a control packet overhead of approxi-
mately 20% [14]. This overhead is derived from the total number of control
packets and the total number of data packets in a sensor network. The average
overhead in terms of number of excessive data packets received is 3.35 [14]. In
addition to the actual code propagation protocol overhead, there is also over-
head from the MAC layer, both in terms of packet headers and control traffic.
The TR1001 provides a low-level interface to the CPU, which enabled us
to measure only the current draw of the receiver. We first measured the time
required for receiving one byte of data from the radio. To produce the graph
11.6 Evaluation 173

Time Energy Time per Energy per

Transceiver (s) (mJ) byte (s) byte (mJ)
TR1001 0.83 21 0.0008 0.021
CC2420 0.060 4.8 0.00006 0.0048

Table 11.2: Lower bounds on the time and energy consumption for receiving
1000 bytes with the TR1001 and CC2420 transceivers. All values are rounded
to two significant digits.

in the figure, we measured the current draw of an ESB board which we had
programmed to turn on receive mode and busy-wait for the time corresponding
to the reception time of 1000 bytes.
When measuring the reception current draw of the CC2420, we could not
measure the time required for receiving one byte because the CC2420 does not
provide an interface at the bit level. Instead, we used two Telos Sky boards and
programmed one to continuously send back-to-back packets with 100 bytes
of data. We programmed the other board to turn on receive mode when the
on-board button was pressed. The receiver would receive 1000 bytes of data,
corresponding to 10 packets, before turning the receiver off. We placed the two
boards next to each other on a table to avoid packet drops. We produced the
graph in Figure 11.6 by measuring the current draw of the receiver Telos Sky
board. To ensure that we did not get spurious packet drops, we repeated the
measurement five times without obtaining differing results.
Table 11.2 shows the lower bounds on the time and energy consumption
for receiving data with the TR1001 and CC2420 transceivers. The results show
that while the current draw of the CC2420 is higher than that of the TR1001, the
energy efficiency in terms of energy per byte of the CC2420 is better because
of the shorter time required to receive the data.

Energy Consumption of Dynamic Linking

To evaluate the energy consumption of dynamic linking, we measure the energy
required for the Contiki dynamic linker to link and load two Contiki programs.
Normally, Contiki loads programs from the radio network but to avoid measur-
ing any unrelated radio or network effects, we stored the loadable object files in
flash ROM before running the experiments. The loadable objects were stored
as ELF files from which all debugging information and symbols that were not
needed for run-time linking was removed. At boot-up, one ELF file was copied
174 Paper E

20
Current

15
Current (mA)

Writing to EEPROM Linking and relocating Writing Executing binary

10 to flash
ROM

0
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (s)

Figure 11.7: Current draw for writing the Blinker ELF file to EEPROM (0 -
0.166 s), linking and relocating the program (0.166 - 0.418 s), writing the re-
sulting code to flash ROM (0.418 - 0.488 s), and executing the binary (0.488
s and onward). The current spikes delimit the three steps and are intention-
ally caused by blinking on-board LEDs. The high energy consumption when
executing the binary is caused by the green LED.

into an on-board EEPROM from where the Contiki dynamic linker linked and
relocated the ELF file before it loaded the program into flash ROM.
Figure 11.7 shows the current draw when loading the Blinker program, and
Figure 11.8 shows the current draw when loading the Object Tracker program.
The current spikes seen in both graphs are intentionally caused by blinking the
on-board LEDs. The spikes delimit the four different steps that the loader is
going through: copying the ELF object file to EEPROM, linking and relocating
the object code, copying the linked code to flash ROM, and finally executing
the loaded program. The current draw of the green LED is slightly above 8
mA, which causes the high current draw when executing the blinker program
(Figure 11.7). Similarly, when the object tracking application starts, it turns
on the radio for neighbor discovery. This causes the current draw to rise to
around 6 mA in Figure 11.8, and matches the radio current measurements in
Figure 11.5.
Table 11.3 shows the energy consumption of loading and linking the
11.6 Evaluation 175

20
Current

15
Current (mA)

Writing to EEPROM Linking and relocating Writing Executing

10 to flash binary
ROM

0
0 0.2 0.4 0.6 0.8 1 1.2
Time (s)

Figure 11.8: Current draw for writing the Object Tracker ELF file to EEPROM
(0 - 0.282 s), linking and relocating the program (0.282 - 0.882 s), writing the
resulting code to flash ROM (0.882 - 0.988 s), and executing the binary (0.988
s and onward). The current spikes delimit the three steps and are intentionally
caused by blinking on-board LEDs. The high current draw when executing the
binary comes from the radio being turned on.

Blinker program. The energy was obtained from integration of the curve from
Figure 11.7 and multiplying it by the voltage used in our experiments (4.5 V).
We see that the linking and relocation step is the most expensive in terms of
energy. It is also the longest step.
To evaluate the energy overhead of the ELF file format, we compare the
energy consumption for receiving four different Contiki programs using the
ELF and CELF formats. In addition to the two programs from Figures 11.3
and 11.4 we include the code for the Contiki code propagation mechanism and
a network publish/subscribe program that performs periodic flooding and con-
verging of information. The two latter programs are significantly larger. We
calculate an estimate of the required energy for receiving the files by using the
measured energy consumption of the CC2420 radio transceiver and multiply it
by the average overhead by the Deluge code propagation protocol, 3.35 [14].
The results are listed in Table 11.4 and show that radio reception is more energy
consuming than linking and loading a program, even for a small program. Fur-
176 Paper E

Blinker Energy Obj. Tr. Energy

Step time (s) (mJ) time (s) (mJ)
Wrt. EEPROM 0.164 1.1 0.282 1.9
Link & reloc 0.252 1.2 0.600 2.9
Wrt. flash ROM 0.070 0.62 0.106 0.76
Total 0.486 2.9 0.988 5.5

Table 11.3: Measured energy consumption of the storing, linking and loading
of the 1056 bytes large Blinker binary and the 1824 bytes large Object Tracker
binary. The size of the Blinker code is 130 bytes and the size of the Object
Tracker code is 344 bytes.

thermore, the results show that the relative average size and energy overhead
for ELF files compared to the code and data contained in the files is approxi-
mately 4 whereas the relative CELF overhead is just under 2.

11.6.2 Memory Consumption

Memory consumption is an important metric for sensor nodes since memory is
a scarce resource on most sensor node platforms. The ESB nodes feature only
2 KB RAM and 60 KB ROM while Mica2 motes provide 128 KB of program
memory and 4 KB of RAM. The less memory required for reprogramming,
the more is left for applications and support for other important tasks such as
security which may also require a large part of the available memory [28].
Table 11.5 lists the memory requirements of the static linker, the dynamic
linker and loader, the CVM and the Java VM. The dynamic linker needs to
keep a table of all core symbols in the system. For a complete Contiki system
with process management, networking, the dynamic loader, memory alloca-
tion, Contiki libraries, and parts of the standard C library, the symbol table
requires about 4 kilobytes of ROM. This is included in the ROM size for the
dynamic linker.

11.6.3 Execution Overhead

To measure the execution overhead of the application specific virtual machine
and the Java virtual machine, we implemented the object tracking program in
Figure 11.4 in C and Java. We compiled the Java code to CVM code and Java
bytecode. We ran the compiled code on the MSP430-equipped ESB board.
11.6 Evaluation 177

ELF ELF ELF radio

Code Data file file size reception
Program size size size overhead energy (mJ)
Blinker 130 14 1056 7.3 17
Object tracker 344 22 1668 5.0 29
Code propagator 2184 10 5696 2.6 92
Flood/converge 4298 42 8456 1.9 136
CELF CELF CELF radio
Code Data file file size reception
Program size size size overhead energy (mJ)
Blinker 130 14 361 2.5 5.9
Object tracker 344 22 758 2.0 12
Code propagator 2184 10 3686 1.7 59
Flood/converge 4298 42 5399 1.2 87

Table 11.4: The overhead of the ELF and CELF file formats in terms of bytes
and estimated reception energy for four Contiki programs. The reception en-
ergy is the lower bound of the radio reception energy with the CC2420 chip,
multiplied by the average Deluge overhead (3.35).

The native C code was compiled with the MSP430 port of GCC version 3.2.3.
The MSP430 digitally-controlled oscillator was set to clock the CPU at a speed
of 2.4576 MHz. We measured the execution time of the three implementations
using the on-chip timer A1 that was set to generate a timer interrupt 1000 times
per second. The execution times are averaged over 5000 iterations of the object
tracking program.
The results in Table 11.6 show the execution time of one run of the object
tracking application from Figure 11.4. The execution time measurements are
averaged over 5000 runs of the object tracking program. The energy consump-
tion is calculated by multiplying the execution time with the average energy
consumption when a program is running with the radio turned off. The table
shows that the overhead of the Java virtual machine is higher than that of the
CVM, which is turn is higher than the execution overhead of the native C code.
All three implementations of the tracker program use the same abstract
regions library which is compiled as native code. Thus much of the execution
time in the Java VM and CVM implementations of the object tracking program
is spent executing the native code in the abstract regions library. Essentially,
the virtual machine simply acts as a dispatcher of calls to various native func-
178 Paper E

Module ROM RAM

Static loader 670 0
Dynamic linker, loader 5694 18
CVM 1344 8
Java VM 13284 59

Table 11.5: Memory requirements, in bytes. The ROM size for the dynamic
linker includes the symbol table. The RAM figures do not include memory for
programs running on top of the virtual machines.

Execution type Execution time (ms) Energy (mJ)

Native 0.479 0.00054
CVM 0.845 0.00095
Java VM 1.79 0.0020

Table 11.6: Execution times and energy consumption of one iteration of the
tracking program.

tions. For programs that spend a significant part of their time executing virtual
machine code the relative execution times are significantly higher for the vir-
tual machine programs. To illustrate this, Table 11.7 lists the execution times
of a convolution operation of two vectors of length 8. Convolution is a com-
mon operation in digital signal processing where it is used for algorithms such
as filtering or edge detection. We see that the execution time of the program
running on the virtual machines is close to ten times that of the native program.

11.6.4 Quantitative Comparison

Using our model from Section 11.6.1 and the results from the above measure-
ments, we can calculate approximations of the energy consumption for distri-
bution, reprogramming, and execution of native and virtual machine programs
in order to compare the methods with each other. We set Pp , the scale factor
of the energy consumption for receiving an object file, to the average Deluge
overhead of 3.35.
11.6 Evaluation 179

Execution type Execution time (ms) Energy (mJ)

Native 0.67 0.00075
CVM 58.52 0.065
Java VM 65.6 0.073

Table 11.7: Execution times and energy consumption of the 8 by 8 vector

convolution.

Dynamic Full image

Step linking (mJ) replacement (mJ)
Receiving 17 330
Wrt. EEPROM 1.1 22
Link & reloc 1.4 -
Wrt. flash ROM 0.45 72
Total 20 424

Table 11.8: Comparison of energy-consumption of reprogramming the blinker

application using dynamic linking with an ELF file and full image replacement
methods.

Dynamic Linking vs Full Image Replacement

We first compare the energy costs for the two native code reprogramming mod-
els: dynamic linking and full image replacement. Table 11.8 shows the results
for the energy consumption of reprogramming the blinker application. The size
of blinker application including the operating system is 20 KB which is about
20 times the size of the blinker application itself. Even though no linking
needs to be performed during the full image replacement, this method is about
20 times more expensive to perform a whole image replacement compared to
a modular update using the dynamic linker.

Dynamic Linking vs Virtual Machines

We use the tracking application to compare reprogramming using the Contiki
dynamic linker with code updates for the CVM and the Java virtual machine.
CVM programs are typically very small and are not stored in EEPROM, nor
are they linked or written to flash. Java uncompressed class files are loaded
into flash ROM before they are executed. Table 11.9 shows the sizes of the
180 Paper E

Step ELF CELF CVM Java

Size (bytes) 1824 968 123 1356
Receiving 29 12 2.0 22
Wrt. EEPROM 1.9 0.80 - -
Link & reloc 2.5 2.5 - -
Wrt. flash ROM 1.2 1.2 - 4.7
Total 35 16.5 2.0 26.7

Table 11.9: Comparison of energy-consumption in mJ of reprogramming for

the object tracking application using the four different methods.

corresponding binaries and the energy consumption of each reprogramming

step.
As expected, the process of updating sensor nodes with native code is less
energy-efficient than updating with a virtual machine. Also, as shown in Ta-
ble 11.6, executing native code is more energy-efficient than executing code
for the virtual machines.
By combining the results in Table 11.6 and Table 11.9, we can compute
break-even points for how often we can execute native code as opposed to
virtual machine code for the same energy consumption. That is, after how
many program iterations do the cheaper execution costs outweigh the more
expensive code updates.
Figure 11.9 shows the modeled energy consumption for executing the Ob-
ject Tracking program using native code loaded with an ELF object file, native
code loaded with an CELF object file, CVM code, and Java code. We see that
the Java virtual machine is expensive in terms of energy and will always re-
quire more energy than native code loaded with a CELF file. For native code
loaded with an ELF file the energy overhead due to receiving the file makes
the Java virtual machine more energy efficient until the program is repeated a
few thousand times. Due to the small size of the CVM code it is very energy
efficient for small numbers of program iterations. It takes about 40000 itera-
tions of the program before the interpretation overhead outweigh the linking
and loading overhead of same program running as native code and loaded as
a CELF file. If the native program was loaded with an ELF file, however, the
CVM program needs to be run approximately 80000 iterations before the en-
ergy costs are the same. At the break-even point, the energy consumption is
only about one fifth of the energy consumption for loading the blink program
11.6 Evaluation 181

140
Java VM
ELF
120 CVM
Consumed energy (mJ)

CELF
100

0
0 20000 40000 60000 80000 100000
Number of program iterations

Figure 11.9: Break-even points for the object tracking program implemented
with four different linking and execution methods.

using full image replacement as shown in Table 11.8.

In contrast with Figure 11.9, Figure 11.10 contains the break-even points
from the vector convolution in Table 11.7. We assume that the convolution
algorithm is part of a program with the same size as in Figure 11.9 so that the
energy consumption for reprogramming is the same. In this case the break-
even points are drastically lower than in Figure 11.9. Here the native code
loaded with an ELF file outperforms the Java implementation already at 100
iterations. The CVM implementation has spent as much energy as the native
ELF implementation after 500 iterations.

11.6.5 Scenario Suitability

We can now apply our results to the software update scenarios discussed in
Section 11.2. In a scenario with frequent code updates, such as the dynamic
application scenario or during software development, a low loading overhead
is to prefer. From Figure 11.9 we see that both an application-specific virtual
machine and a Java machine may be good choices. Depending on the type of
182 Paper E

140
Java VM
ELF
120 CVM
Consumed energy (mJ)

CELF
100

0
0 200 400 600 800 1000
Number of program iterations

Figure 11.10: Break-even points for the vector convolution implemented with
four different linking and execution methods.

application it may be beneficial to decide to run the program on top of a more

flexible virtual machine such as the Java machine. The price for such a decision
is higher energy overhead.
In scenarios where the update frequency is low, e.g. when fixing bugs in in-
stalled software or when reconfiguring an installed application, the higher price
for dynamic linking may be worth paying. If the program is continuously run
for a long time, the energy savings of being able to use native code outweigh
the energy cost of the linking process. Furthermore, with a virtual machine it
may not be possible to make changes to all levels of the system. For example,
a bug in a low-level driver can usually only be fixed by installing new native
code. Moreover, programs that are computationally heavy benefit from being
implemented as native code as native code has lower energy consumption than
virtual machine code.
The results from Figures 11.9 and 11.10 suggest that a combination of vir-
tual machine code and native code can be energy efficient. For many situations
this may be a viable alternative to running only native code or only virtual
machine code.
11.7 Discussion 183

Lines of code, Lines of code,

Module total relocation function
Generic linker 292
MSP430-specific 45 8
AVR-specific 143 104

Table 11.10: Number of lines of code for the dynamic linker and the
microcontroller-specific parts.

11.6.6 Portability
Because of the diversity of sensor network platforms, the Contiki dynamic
linker is designed to be portable between different microcontrollers. The dy-
namic linker is divided into two modules: a generic part that parses and ana-
lyzes the ELF/CELF that is to be loaded, and a microcontroller-specific part
that allocates memory for the program to be loaded, performs code and data
relocation, and writes the linked program into memory.
To evaluate the portability of our design we have ported the dynamic linker
to two different microcontrollers: the TI MSP430 and the Atmel AVR. The TI
MSP430 is used in several sensor network platforms, including the Telos Sky
and the ESB. The Atmel AVR is used in the Mica2 motes.
Table 11.10 shows the number of lines of code needed to implement each
module. The dramatic difference between the MSP430-specific module and
the AVR-specific module is due to the different addressing modes used by the
machine code of the two microcontrollers. While the MSP430 has only one
addressing mode, the AVR has 19 different addressing modes. Each addressing
mode must be handled differently by the relocation function, which leads to a
larger amount of code for the AVR-specific module.

11.7 Discussion
Standard file formats. Our main motivation behind choosing the ELF for-
mat for dynamic linking in Contiki was that the ELF format is a standard file
format. Many compilers and utilities, including all GCC utilities, are able to
produce and handle ELF files. Hence no special software is needed to com-
pile and upload new programs into a network of Contiki nodes. In contrast,
FlexCup [27] or diff-based approaches require the usage of specially crafted
184 Paper E

utilities to produce meta data or diff scripts required for uploading software.
These special utilities also need to be maintained and ported to the full range
of development platforms used for software development for the system.
Operating system support. Dynamic linking of ELF files requires sup-
port from the underlying operating system and cannot be done on monolithic
operating systems such as TinyOS. This is a disadvantage of our approach. For
monolithic operating systems, an approach such as FlexCup is better suited.
Heterogeneity. With diff-based approaches a binary diff is created either
at a base station or by an outside server. The server must have knowledge of
the exact software configuration of the sensor nodes on which the diff script
is to be run. If sensor nodes are running different versions of their software,
diff-based approaches do not scale.
Specifically, in many of our development networks we have witnessed a
form of micro heterogeneity in the software configuration. Many sensor nodes,
which have been running the exact same version of the Contiki operating sys-
tem, have had small differences in the address of functions and variables in
the core. This micro heterogeneity comes from the different core images being
compiled by different developers, each having slightly different versions of the
C compiler, the C library and the linker utilities. This results in small varia-
tions of the operating system image depending on which developer compiled
the operating system image. With diff-based approaches micro heterogeneity
poses a big problem, as the base station would have to be aware of all the small
differences between each node.
Combination of native and virtual machine code. Our results suggest
that a combination of native and virtual machine code is an energy efficient
alternative to pure native code or pure virtual machine code approaches. The
dynamic linking mechanism can be used to load the native code that is used by
the virtual machine code by the native code interfaces in the virtual machines.

11.8 Related Work

Because of the importance of dynamic reprogramming of wireless sensor net-
works there has been a lot of effort in the area of software updates for sensor
nodes both in the form of system support for software updates and execution
environments that directly impact the type and size of updates as well as distri-
bution protocols for software updates.
Mainwaring et al. [26] also identified the trade-off between using virtual
machine code that is more expensive to run but enables more energy-efficient
11.8 Related Work 185

updates and running native code that executes more efficiently but requires
more costly updates. This trade-off has been further discussed by Levis and
Culler [19] who implemented the Maté virtual machine designed to both sim-
plify programming and to leverage energy-efficient large-scale software up-
dates in sensor networks. Maté is implemented on top of TinyOS.
Levis and Culler later enhanced Maté by application specific virtual ma-
chines (ASVMs) [20]. They address the main limitations of Maté: flexibility,
concurrency and propagation. Whereas Maté was designed for a single ap-
plication domain only, ASVM supports a wide range of application domains.
Further, instead of relying on broadcasts for code propagation as Maté, ASVM
uses the trickle algorithm [21].
The MagnetOS [23] system uses the Java virtual machine to distribute ap-
plications across an ad hoc network of laptops. In MagnetOS, Java applica-
tions are partitioned into distributed components. The components transpar-
ently communicate by raising events. Unlike Maté and Contiki, MagnetOS
targets larger platforms than sensor nodes such as PocketPC devices. Sensor-
Ware [1] is another script-based proposal for programming nodes that targets
larger platforms. VM* is a framework for runtime environments for sensor
networks [18]. Using this framework Koshy and Pandey have implemented a
subset of the Java Virtual Machine that enables programmers to write applica-
tions in Java, and access sensing devices and I/O through native interfaces.
Mobile agent-based approaches extend the notion of injected scripts by de-
ploying dynamic, localized and intelligent mobile agents. Using mobile agents,
Fok et al. have built the Agilla platform that enables continuous reprogram-
ming by injecting new agents into the network [9].
TinyOS uses a special description language for composing a system of
smaller components [10] which are statically linked with the kernel to a com-
plete image of the system. After linking, modifying the system is not possi-
ble [19] and hence TinyOS requires the whole image to be updated even for
small code changes.
Systems that offer loadable modules besides Contiki include SOS [12] and
Impala [24]. Impala features an application updater that enables software up-
dates to be performed by linking in updated modules. Updates in Impala are
coarse-grained since cross-references between different modules are not possi-
ble. Also, the software updater in Impala was only implemented for much more
resource-rich hardware than our target devices. The design of SOS [12] is very
similar to the Contiki system: SOS consists of a small kernel and dynamically-
loaded modules. However, SOS uses position independent code to achieve
relocation and jump tables for application programs to access the operating
186 Paper E

system kernel. Application programs can register function pointers with the
operating system for performing inter-process communication. Position in-
dependent code is not available for all platforms, however, which limits the
applicability of this approach.

FlexCup [27] enables run-time installation of software components in

TinyOS and thus solves the problem that a full image replacement is required
for reprogramming TinyOS applications. In contrast to our ELF-based solu-
tion, FlexCup uses a non-standard format and is less portable. Further, Flex-
Cup requires a reboot after a program has been installed, requiring an external
mechanism to save and restore the state of all other applications as well as the
state of running network protocols across the reboot. Contiki does not need to
be rebooted after a program has been installed.

FlexCup also requires a complete duplicate image of the binary image of

the system to be stored in external flash ROM. The copy of the system image
is used for constructing a new system image when a new program has been
loaded. In contrast, the Contiki dynamic linker does not alter the core image
when programs are loaded and therefore no external copy of the core image is
needed.

Since the energy consumption of distributing code in sensor networks in-

creases with the size of the code to be distributed several attempts have been
made to reduce the size of the code to be distributed. Reijers and Langen-
doen [31] produce an edit script based on the difference between the modified
and original executable. After various optimizations including architecture-
dependent ones, the script is distributed. A similar approach has been de-
veloped by Jeong and Culler [15] who use the rsync algorithm to generate
the difference between modified and original executable. Koshy and Pandey’s
diff-based approach [17] reduces the amount of flash rewriting by modifying
the linking procedure so that functions that are not changed are not shifted.

XNP [16] was the previous default reprogramming mechanism in TinyOS

which is used by the multi-hop reprogramming scheme MOAP (Multi-
hop Over-the-Air Programming) developed to distribute node images in the
sensor network. MOAP distributes data to a selective number of nodes
on a neighbourhood-by-neighbourhood basis that avoids flooding [34]. In
Trickle [21] virtual machine code is distributed to a network of nodes. While
Trickle is restricted to single packet dissemination, Deluge adds support for the
dissemination of large data objects [14].
11.9 Conclusions 187

11.9 Conclusions
We have presented a highly portable dynamic linker and loader that uses the
standard ELF file format and compared the energy-efficiency of run-time dy-
namic linking with an application specific virtual machine and a Java virtual
machine. We show that dynamic linking is feasible even for constrained sensor
nodes.
Our results also suggest that a combination of native and virtual machine
code provide an energy efficient alternative to pure native code or pure virtual
machine approaches. The native code that is called from the virtual machine
code can be updated using the dynamic linker, even in heterogeneous systems.

Acknowledgments
This work was partly financed by VINNOVA, the Swedish Agency for Inno-
vation Systems, and the European Commission under contract IST-004536-
RUNES. Thanks to our paper shepherd Feng Zhao for reading and commenting
on the paper.
Bibliography

Bibliography

[1] A. Boulis, C. Han, and M. B. Srivastava. Design and implementation of a

framework for efficient and programmable sensor networks. In Proceed-
ings of The First International Conference on Mobile Systems, Applica-
tions, and Services (MOBISYS ‘03), San Francisco, USA, May 2003.
[2] Chipcon AS. CC2420 Datasheet (rev. 1.3), 2005.
URL: http://www.chipcon.com/

[3] TIS Committee. Tool Interface Standard (TIS) Executable and Linking
Format (ELF) Specification Version 1.2, May 1995.
[4] H. Dai, M. Neufeld, and R. Han. Elf: an efficient log-structured flash file
system for micro sensor nodes. In SenSys, pages 176–187, 2004.
[5] A. Dunkels, B. Grönvall, and T. Voigt. Contiki - a lightweight and flexible
operating system for tiny networked sensors. In Proceedings of the First
IEEE Workshop on Embedded Networked Sensors (IEEE Emnets ’04),
Tampa, Florida, USA, November 2004.
[6] A. Dunkels, O. Schmidt, T. Voigt, and M. Ali. Protothreads: Simpli-
fying event-driven programming of memory-constrained embedded sys-
tems. In Proceedings of the 4th International Conference on Embedded
Networked Sensor Systems, SenSys 2006, Boulder, Colorado, USA, 2006.
[7] D. Estrin (editor). Embedded everywhere: A research agenda for net-
worked systems of embedded computers. National Academy Press, 1st
edition, October 2001. ISBN: 0309075688
[8] G. Ferrari, J. Stuber, A. Gombos, and D. Laverde, editors. Programming
Lego Mindstorms with Java with CD-ROM. Syngress Publishing, 2002.
ISBN: 1928994555

188
Bibliography 189

[9] C. Fok, G. Roman, and C. Lu. Rapid development and flexible deploy-
ment of adaptive wireless sensor network applications. In Proceedings
of the 24th International Conference on Distributed Computing Systems,
Tokyo, Japan, June 2005.

[10] D. Gay, P. Levis, R. von Behren, M. Welsh, E. Brewer, and D. Culler.

[11] J. A. Gutierrez, M. Naeve, E. Callaway, M. Bourgeois, V. Mitter, and

B. Heile. IEEE 802.15.4: A developing standard for low-power low-cost
wireless personal area networks. IEEE Network, 15(5):12–19, Septem-
ber/October 2001.

[12] C. Han, R. K. Rengaswamy, R. Shea, E. Kohler, and M. Srivastava. Sos:

A dynamic operating system for sensor networks. In Proceedings of the
3rd international conference on Mobile systems, applications, and ser-
vices (MobiSys ’05), Seattle, WA, USA, June 2005.

[13] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. Culler, and K. Pister. Sys-

[14] J. W. Hui and D. Culler. The dynamic behavior of a data dissemination

protocol for network programming at scale. In Proceedings of the 2nd
international conference on Embedded networked sensor systems (SenSys
’04), Baltimore, Maryland, USA, November 2004.

[15] J. Jeong and D. Culler. Incremental network programming for wire-

less sensors. In Proceedings of the First IEEE Communications Society
Conference on Sensor and Ad Hoc Communications and Networks IEEE
SECON (2004), Santa Clara, California, USA, October 2004.

[16] J. Jeong, S. Kim, and A. Broad. Network reprogramming. TinyOS docu-

mentation, 2003. Visited 2006-04-06.
URL: http://www.tinyos.net/tinyos-1.x/doc/NetworkReprogramming.pdf
190 Bibliography

[17] J. Koshy and R. Pandey. Remote incremental linking for energy-efficient

reprogramming of sensor networks. In Proceedings of the second Euro-
pean Workshop on Wireless Sensor Networks, 2005.

[18] J. Koshy and R. Pandey. Vm*: Synthesizing scalable runtime environ-

ments for sensor networks. In Proc. SenSys’05, San Diego, CA, USA,
November 2005.

[19] P. Levis and D. Culler. Maté: A tiny virtual machine for sensor networks.
In Proceedings of ASPLOS-X, San Jose, CA, USA, October 2002.

[20] P. Levis, D. Gay, and D Culler. Active sensor networks. In Proceedings of

ACM/Usenix Networked Systems Design and Implementation (NSDI’05),
Boston, MA, USA, May 2005.

[21] P. Levis, N. Patel, D. Culler, and S. Shenker. Trickle: A self-regulating

algorithm for code propagation and maintenance in wireless sensor net-
works. In Proceedings of ACM/Usenix Networked Systems Design and
Implementation (NSDI’04), March 2004.

[22] J. Lilius and I. Paltor. Deeply embedded python, a virtual machine for
embedded systems. Web page. 2006-04-06.
URL: http://www.tucs.fi/magazin/output.php?ID=2000.N2.LilDeEmPy

[23] H. Liu, T. Roeder, K. Walsh, R. Barr, and E. Gün Sirer. Design and
implementation of a single system image operating system for ad hoc
networks. In MobiSys, pages 149–162, 2005.

[24] T. Liu, C. Sadler, P. Zhang, and M. Martonosi. Implementing software

on resource-constrained mobile sensors: Experiences with Impala and
ZebraNet. In Proc. Second Intl. Conference on Mobile Systems, Applica-
tions and Services (MOBISYS 2004), June 2004.

[25] G. Mainland, L. Kang, S. Lahaie, D. C. Parkes, and M. Welsh. Using

virtual markets to program global behavior in sensor networks. In Pro-
ceedings of the 2004 SIGOPS European Workshop, Leuven, Belgium,
September 2004.

[26] A. Mainwaring, J. Polastre, R. Szewczyk, D. Culler, and J. Anderson.

Wireless sensor networks for habitat monitoring. In First ACM Workshop
on Wireless Sensor Networks and Applications (WSNA 2002), Atlanta,
GA, USA, September 2002.
[27] P. José Marrón, M. Gauger, A. Lachenmann, D. Minder, O. Saukh, and
K. Rothermel. Flexcup: A flexible and efficient code update mechanism
for sensor networks. In European Workshop on Wireless Sensor Net-
works, Zurich, Switzerland, January 2006.
[28] A. Perrig, R. Szewczyk, V. Wen, D. E. Culler, and J. D. Tygar. SPINS:
security protocols for sensor netowrks. In Mobile Computing and Net-
working, pages 189–199, 2001.
[29] J. Polastre, R. Szewczyk, and D. Culler. Telos: Enabling ultra-low power
wireless research. In Proc. IPSN/SPOTS’05, Los Angeles, CA, USA,
April 2005.
[30] N. Ramanathan, E. Kohler, and D. Estrin. Towards a debugging system
for sensor networks. International Journal for Network Management,
3(5), 2005.
[31] N. Reijers and K. Langendoen. Efficient code distribution in wireless sen-
sor networks. In Proceedings of the 2nd ACM international conference
on Wireless sensor networks and applications, pages 60–67, San Diego,
CA, USA, September 2003.
[32] RF Monolithics. 868.35 MHz Hybrid Transceiver TR1001, 1999.
URL: http://www.rfm.com

[33] J. Schiller, H. Ritter, A. Liers, and T. Voigt. Scatterweb - low power

nodes and energy aware routing. In Proceedings of Hawaii International
Conference on System Sciences, Hawaii, USA, January 2005.
[34] T. Stathopoulos, J. Heidemann, and D. Estrin. A remote code update
mechanism for wireless sensor networks. Technical Report CENS-TR-
30, University of California, Los Angeles, Center for Embedded Net-
worked Computing, November 2003.
[35] M. Welsh and G. Mainland. Programming Sensor Networks Using Ab-
stract Regions. In Proceedings of ACM/Usenix Networked Systems De-
sign and Implementation (NSDI’04), San Francisco, California, USA,
March 2004.
[36] G. Werner-Allen, P. Swieskowski, and M. Welsh. Motelab: A wireless
sensor network testbed. In Proc. IPSN/SPOTS’05, Los Angeles, CA,
USA, April 2005.
SICS Dissertation Series

01: Bogumil Hausman, Pruning and Speculative Work in OR-Parallel PRO-

LOG, 1990.

02: Mats Carlsson, Design and Implementation of an OR-Parallel Pro-

log Engine, 1990.

03: Nabiel A. Elshiewy, Robust Coordinated Reactive Computing in

SANDRA, 1990.

04: Dan Sahlin, An Automatic Partial Evaluator for Full Prolog, 1991.

05: Hans A. Hansson, Time and Probability in Formal Design of Dis-

tributed Systems, 1991.

06: Peter Sjödin, From LOTOS Specifications to Distributed Implemen-

tations, 1991.

07: Roland Karlsson, A High Performance OR-parallel Prolog System,

1992.

08: Erik Hagersten, Toward Scalable Cache Only Memory Architec-

tures, 1992.

09: Lars-Henrik Eriksson, Finitary Partial Inductive Definitions and General

Logic, 1993.

10: Mats Björkman, Architectures for High Performance Communica-

tion, 1993.

193
194 Bibliography

11: Stephen Pink, Measurement, Implementation, and Optimization of

Internet Protocols, 1993.

12: Martin Aronsson, GCLA. The Design, Use, and Implementation of

a Program Development System, 1993.

13: Christer Samuelsson, Fast Natural-Language Parsing Using Explanation-

Based Learning, 1994.

14: Sverker Jansson, AKL - - A Multiparadigm Programming Language, 1994.

15: Fredrik Orava, On the Formal Analysis of Telecommunication Pro-

tocols, 1994.

16: Torbjörn Keisu, Tree Constraints, 1994.

17: Olof Hagsand, Computer and Communication Support for Interac-

tive Distributed Applications, 1995.

18: Björn Carlsson, Compiling and Executing Finite Domain Constraints,

1995.

19: Per Kreuger, Computational Issues in Calculi of Partial Inductive

Definitions, 1995.

20: Annika Waern, Recognising Human Plans: Issues for Plan Recogni-
tion in Human-Computer Interaction, 1996.

21: Björn Gambäck, Processing Swedish Sentences: A Unification-Based

Grammar and Some Applications, 1997.

22: Klas Orsvärn, Knowledge Modelling with Libraries of Task Decom-

position Methods, 1996.

23: Kia Höök, A Glass Box Approach to Adaptive Hypermedia, 1996.

24: Bengt Ahlgren, Improving Computer Communication Performance

by Reducing Memory Bandwidth Consumption, 1997.
Bibliography 195

25: Johan Montelius, Exploiting Fine-grain Parallelism in Concurrent

Constraint Languages, 1997.

26: Jussi Karlgren, Stylistic experiments in information retrieval, 2000.

27: Ashley Saulsbury, Attacking Latency Bottlenecks in Distributed

Shared Memory Systems, 1999.

28: Kristian Simsarian, Toward Human Robot Collaboration, 2000.

29: Lars-åke Fredlund, A Framework for Reasoning about Erlang Code,

2001.

30: Thiemo Voigt, Architectures for Service Differentiation in Over-

loaded Internet Servers, 2002.

31: Fredrik Espinoza, Individual Service Provisioning, 2003.

32: Lars Rasmusson, Network capacity sharing with QoS as a financial

derivative pricing problem: algorithms and network design, 2002.

33: Martin Svensson, Defining, Designing and Evaluating Social Navi-

gation, 2003.

34: Joe Armstrong, Making reliable distributed systems in the presence

of software errors, 2003.

35: Emmanuel Frécon, DIVE on the Internet, 2004.

36: Rickard Cöster, Algorithms and Representations for Personalised

Information Access, 2005.

37: Per Brand, The Design Philosophy of Distributed Programming

Systems: the Mozart Experience, 2005.

38: Sameh El-Ansary, Designs and Analyses in Structured Peer-to-Peer

Systems, 2005.
196 Bibliography

39: Erik Klintskog, Generic Distribution Support for Programming Sys-

tems, 2005.

40: Markus Bylund, A Design Rationale for Pervasive Computing User

Experience, Contextual Change, and Technical Requirements, 2005.

41: Åsa Rudström, Co-Construction of hybrid spaces, 2005.

42: Babak Sadighi Firozabadi, Decentralised Privilege Management for

Access Control, 2005.

43: Marie Sjölinder, Age-related Cognitive Decline and Navigation in

Electronic Environments, 2006.

44: Magnus Sahlgren, The Word-Space Model: Using Distributional

Analysis to Represent Syntagmatic and Paradigmatic Relations between
Words in High-dimensional Vector Spaces, 2006.

45: Ali Ghodsi, Distributed k-ary System: Algorithms for Distributed

Hash Tables, 2006.

46: Stina Nylander, Design and Implementation of Multi-Device Ser-

vices, 2007

47: Adam Dunkels, Programming Memory-Constrained Networked Em-

bedded Systems, 2007.

USKT Documentation
No ratings yet
USKT Documentation
22 pages
Application of Real Time Operating System in The Internet of Things
No ratings yet
Application of Real Time Operating System in The Internet of Things
6 pages
CMPS290 Class Notes Chap 01
No ratings yet
CMPS290 Class Notes Chap 01
24 pages
Chapter # 1 COAL
No ratings yet
Chapter # 1 COAL
24 pages
Introduction To Parallel Computing
0% (1)
Introduction To Parallel Computing
34 pages
Advanced Electronic Protection System To Prevent Exam Paper Leakages
No ratings yet
Advanced Electronic Protection System To Prevent Exam Paper Leakages
135 pages
Debugging Real-Time Multiprocessor Systems: Class #264, Embedded Systems Conference, Silicon Valley 2006
No ratings yet
Debugging Real-Time Multiprocessor Systems: Class #264, Embedded Systems Conference, Silicon Valley 2006
15 pages
under ground cable doc
No ratings yet
under ground cable doc
52 pages
Chapter 1: Introduction: 1.1 What Is A Microprocessor?
No ratings yet
Chapter 1: Introduction: 1.1 What Is A Microprocessor?
69 pages
Internet of Things OS
No ratings yet
Internet of Things OS
9 pages
The Laboratory Use of Computers: Merck Research Laboratories, Merck & Co., Inc., Rahway, NJ, USA
No ratings yet
The Laboratory Use of Computers: Merck Research Laboratories, Merck & Co., Inc., Rahway, NJ, USA
20 pages
Sri Venkateswara Engineering College
No ratings yet
Sri Venkateswara Engineering College
12 pages
Face Recognization Based Car Security System
No ratings yet
Face Recognization Based Car Security System
86 pages
Exposicion de Jimenez
No ratings yet
Exposicion de Jimenez
19 pages
COME6102 Chapter 1 Introduction 1 of 2
No ratings yet
COME6102 Chapter 1 Introduction 1 of 2
18 pages
CS3099 Software Team Project: Group Report: A. Bell, M. Brock, A. Buckland, C. Garrigan, and T. Liang
No ratings yet
CS3099 Software Team Project: Group Report: A. Bell, M. Brock, A. Buckland, C. Garrigan, and T. Liang
11 pages
Rush Collision Prevention in Railways Major Document
No ratings yet
Rush Collision Prevention in Railways Major Document
77 pages
03 Introduction PDF
No ratings yet
03 Introduction PDF
0 pages
Ijettcs 2017 04 16 67 PDF
No ratings yet
Ijettcs 2017 04 16 67 PDF
5 pages
MIPSIT
No ratings yet
MIPSIT
8 pages
Pocket PC Compiler: Mini Project Report
No ratings yet
Pocket PC Compiler: Mini Project Report
22 pages
An Embedded Systems Laboratory To Support Rapid Prototyping of Robotics and The Internet of Things
No ratings yet
An Embedded Systems Laboratory To Support Rapid Prototyping of Robotics and The Internet of Things
3 pages
Cs501 Glossary
No ratings yet
Cs501 Glossary
14 pages
Computer03 Softeninghw
No ratings yet
Computer03 Softeninghw
7 pages
MIPS
No ratings yet
MIPS
6 pages
RICS
No ratings yet
RICS
4 pages
Survey On Minimal Implementation of Clustering Using Virtual Machines
No ratings yet
Survey On Minimal Implementation of Clustering Using Virtual Machines
12 pages
CH 1 Intro
No ratings yet
CH 1 Intro
7 pages
Advanced Computer Architectures
No ratings yet
Advanced Computer Architectures
194 pages
CA - 2marks
No ratings yet
CA - 2marks
80 pages
The Laboratory Use of Computers: Merck Research Laboratories, Merck & Co., Inc., Rahway, NJ, USA
No ratings yet
The Laboratory Use of Computers: Merck Research Laboratories, Merck & Co., Inc., Rahway, NJ, USA
20 pages
ITE 111 Module 1-4
No ratings yet
ITE 111 Module 1-4
5 pages
Cs1304-Computer Architecture Department of Cse & It
No ratings yet
Cs1304-Computer Architecture Department of Cse & It
105 pages
Pic Report
No ratings yet
Pic Report
79 pages
Comp Org and Desgn Notes
No ratings yet
Comp Org and Desgn Notes
5 pages
Bcs 011 PDF
No ratings yet
Bcs 011 PDF
33 pages
ES Notes
No ratings yet
ES Notes
112 pages
Re Engineering Traditional Learning Model With Outcome Based Learning Curve Using Embedded Training Laboratory
No ratings yet
Re Engineering Traditional Learning Model With Outcome Based Learning Curve Using Embedded Training Laboratory
11 pages
Computer Programming Lecture Notes PDF
No ratings yet
Computer Programming Lecture Notes PDF
134 pages
Parallel Computing Main
No ratings yet
Parallel Computing Main
47 pages
Sms Based Digital Notice Board
No ratings yet
Sms Based Digital Notice Board
107 pages
Chapter 1 To 5 Training File1
No ratings yet
Chapter 1 To 5 Training File1
35 pages
Module-1: Chapter-1 Parallel Computer Models
No ratings yet
Module-1: Chapter-1 Parallel Computer Models
42 pages
Phill 12 PT
No ratings yet
Phill 12 PT
6 pages
Pdc Digital Notes 6 17
No ratings yet
Pdc Digital Notes 6 17
12 pages
Documento_completo
No ratings yet
Documento_completo
5 pages
Cloud Computing Unit 1 Ppt
No ratings yet
Cloud Computing Unit 1 Ppt
29 pages
Contiki, Iot Os &: Cooja Network Simulator
No ratings yet
Contiki, Iot Os &: Cooja Network Simulator
37 pages
73493
No ratings yet
73493
51 pages
Advanced Computer Architecture
No ratings yet
Advanced Computer Architecture
28 pages
Parallel and Cluster Computing
No ratings yet
Parallel and Cluster Computing
31 pages
Microprocessor: From Wikipedia, The Free Encyclopedia
No ratings yet
Microprocessor: From Wikipedia, The Free Encyclopedia
33 pages
Advanced Computer Architecture: 1.0 Objective
No ratings yet
Advanced Computer Architecture: 1.0 Objective
27 pages
CC Unit 1
No ratings yet
CC Unit 1
18 pages
Unit 2 - Advanced Computer Architecture - Www.rgpvnotes.in
No ratings yet
Unit 2 - Advanced Computer Architecture - Www.rgpvnotes.in
15 pages
Embedded Systems Paper Presentation
No ratings yet
Embedded Systems Paper Presentation
13 pages
Crusoe Processor
No ratings yet
Crusoe Processor
15 pages
Assembly
No ratings yet
Assembly
3 pages
Iological Equipments
No ratings yet
Iological Equipments
58 pages
Learn Computer Science
From Everand
Learn Computer Science
Knowledge Flow
No ratings yet
Virtual Report Processing: The Mapper Story
From Everand
Virtual Report Processing: The Mapper Story
Louis Schlueter
No ratings yet
Assesment of Learning
33% (6)
Assesment of Learning
8 pages
JD TCS
No ratings yet
JD TCS
3 pages
Quiz Management System Documentation
36% (11)
Quiz Management System Documentation
74 pages
Quill A Letter-Writing Roleplaying Game For A Single Player PDF
No ratings yet
Quill A Letter-Writing Roleplaying Game For A Single Player PDF
16 pages
PCI_24-25_st1_qb1_MRB
No ratings yet
PCI_24-25_st1_qb1_MRB
2 pages
Attributes &entities
No ratings yet
Attributes &entities
15 pages
B2B Consulting
No ratings yet
B2B Consulting
30 pages
Verbs Followed by Gerund or Infinitive
50% (2)
Verbs Followed by Gerund or Infinitive
7 pages
Federico Mompou
0% (1)
Federico Mompou
3 pages
24 Samantamukhaparivarta : Technical Details
No ratings yet
24 Samantamukhaparivarta : Technical Details
13 pages
Byrne Refuge
No ratings yet
Byrne Refuge
31 pages
Matlab Guide Instalation Student Account
No ratings yet
Matlab Guide Instalation Student Account
14 pages
Altivar 61: Communication Parameters User's Manual
No ratings yet
Altivar 61: Communication Parameters User's Manual
132 pages
A Semiotic Study On War Posters
100% (1)
A Semiotic Study On War Posters
6 pages
Lesson Plan 1
100% (1)
Lesson Plan 1
3 pages
C Programming (Unit 2 - Unit 5 Programs)
No ratings yet
C Programming (Unit 2 - Unit 5 Programs)
55 pages
NCERT Exemplar Class 8 Maths Solutions Chapter 4 Linear Equations in One Variable
No ratings yet
NCERT Exemplar Class 8 Maths Solutions Chapter 4 Linear Equations in One Variable
18 pages
Rhys Davids - 1900 - A Buddhist Manual of Psychological Ethics of The Fourth Century B.C.
100% (4)
Rhys Davids - 1900 - A Buddhist Manual of Psychological Ethics of The Fourth Century B.C.
500 pages
Passive Voice
No ratings yet
Passive Voice
6 pages
FEB 1 Reading-Intervention-Activities ENSCIMA
No ratings yet
FEB 1 Reading-Intervention-Activities ENSCIMA
4 pages
ENG8-Wk 6
No ratings yet
ENG8-Wk 6
2 pages
Lstdrvrs
No ratings yet
Lstdrvrs
130 pages
Quotes Crist + Islam
No ratings yet
Quotes Crist + Islam
2 pages
Final Test For Level 3 A Unit 1-4
No ratings yet
Final Test For Level 3 A Unit 1-4
4 pages
Program Studi Ilmu Komunikasi Fakultas Ilmu Sosial Dan Ilmu Politik Universitas Islam Balitar Blitar Email: Andiwimeifilina@unisbablitar - Ac.id
No ratings yet
Program Studi Ilmu Komunikasi Fakultas Ilmu Sosial Dan Ilmu Politik Universitas Islam Balitar Blitar Email: Andiwimeifilina@unisbablitar - Ac.id
10 pages
'C' Programming by NIKHIL
No ratings yet
'C' Programming by NIKHIL
50 pages
PDF
No ratings yet
PDF
63 pages
Test Project: It Software Solutions For Business
No ratings yet
Test Project: It Software Solutions For Business
8 pages
Soal B.inggris Kelas 4
No ratings yet
Soal B.inggris Kelas 4
4 pages