University of Southern California Dissertations and Theses

Understanding and exploiting the acoustic propagation delay in underwater sensor networks

(USC Thesis Other)

Understanding and exploiting the acoustic propagation delay in underwater sensor networks

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content UNDERSTANDING AND EXPLOITING THE ACOUSTIC PROPAGATION DELAY IN UNDERWATER SENSOR NETWORKS by Affan Ahmed Syed A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (COMPUTER SCIENCE) August 2009 Copyright 2009 Affan Ahmed Syed Dedication Dedicated to my parents, especially my father who passed away a month before my defense. They gave me the freedom and confidence to be the best I can be, and instilled within me a sense of morality that I will cherish for my entire life. For all this and much, much more — thank you. ii Acknowledgments The journey for a doctorate in philosophy is a challenging one where many skills need to be acquired. I have been most fortunate to have my advisor, John Heidemann, guide me through this arduous journey. His mentoring has not only provided me with an understanding that shall help me in doing fundamental research but he also inspired me with his passion and great work-ethics. He has always been tough but fair in his criticism of my work. I have great appreciation for his management in maintaining such a difficult balance. Beyond academics, John has also provided me a good example of how to live a balanced life— managing family, work, and fun. I also want to acknowledge and appreciate Wei Ye’s significant contribution to my academic development during the two years he spent as a co-advisor. Wei Ye has always been a bed-rock of support and has provided sincere and helpful advice every time I approached him for help. Similarly, I have great appreciation for the advice and help extended by Bhaskar Krishnamachari, with whom I have collaborated throughout my time at University of Southern California (USC). I extend a special thank you to the two remaining members of my qualifying and dissertation committee: Guarav Sukhatme and Michael Neely. Both of them provided me with valuable feedback that has allowed me to improve the content and presentation of this dissertation. I would also like to thank Robert Sholtz and Hossein Hashemi for iii allowing access to the anechoic chamber in ultraLab and to Ta-Shun Chu for helping me during my experiments there. I owe gratitude to several people who helped and encouraged me during my research at USC and ISI. Brian Tung guided me to a simple solution in my Markov analysis. My friends Waheed Bajwa, Qasim Chaudhri, and Iqbal Shahid have all been a tremendous help both as sounding-boards for my ideas and in helping me out whenever I get entan- gled. I want to thank two other friends made at USC; Jung Jun Hyun and Joon Ahn, both of whom have taught me a great deal during our work together on research papers and projects. Last, but definitely not the least, I would like to thank my family. My wife, Maria, deserves a doctorate herself for the patience and steadfastness shown during the four months preceding my thesis defense. She has provided me with a sense of direction and helped me through some very difficult time of my life when I had to deal with great personal loss. My daughter, Zunaira, is the light of my life— with just her smile she lightens the burdens of my day. My mother has been a bedrock of patience and perse- verance and had it not been for her I might not have made it to the end of this journey. I would like to finish off by thanking my Creator for His munificence in bestowing me with so many of His favors in this world. iv Table of Contents Dedication ii Acknowledgments iii List Of Tables ix Abstract xiv Chapter 1: Introduction 1 1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.2.1 Communication Medium for UWSN . . . . . . . . . . . . . . . 3 1.2.2 The Vision of UWSN . . . . . . . . . . . . . . . . . . . . . . . 4 1.2.3 Acoustic Propagation and Sensornet Protocols . . . . . . . . . 6 1.3 Thesis Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.3.1 Summary of Work . . . . . . . . . . . . . . . . . . . . . . . . 7 1.4 Contribution and Novelty . . . . . . . . . . . . . . . . . . . . . . . . . 10 Chapter 2: Related Work 12 2.1 Underwater Acoustic Communication . . . . . . . . . . . . . . . . . . 12 2.1.1 Underwater Acoustic Channel . . . . . . . . . . . . . . . . . . 12 2.1.2 Multipath in UWA Communication . . . . . . . . . . . . . . . 14 2.2 Time Synchronization in Computer Networks . . . . . . . . . . . . . . 17 2.2.1 Wide-Area-Network Time Synchronization . . . . . . . . . . . 18 2.2.2 Time Synchronization in SensorNets . . . . . . . . . . . . . . . 19 2.3 Shared Medium Access . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.3.1 Terrestrial RF-based MACs . . . . . . . . . . . . . . . . . . . 21 2.3.2 Satellite MACs . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.3.3 Underwater Acoustic MACs . . . . . . . . . . . . . . . . . . . 24 Chapter 3: Space-Time Uncertainty 27 3.1 Space-time Uncertainty and the ALOHA Protocol . . . . . . . . . . . . 29 v 3.1.1 Space Time Uncertainty . . . . . . . . . . . . . . . . . . . . . 29 3.1.2 Analysis of ALOHA with Time Uncertainty . . . . . . . . . . . 30 3.1.3 ALOHA with Space Uncertainty . . . . . . . . . . . . . . . . . 33 3.2 The PDT-ALOHA Protocol . . . . . . . . . . . . . . . . . . . . . . . . 38 3.2.1 Protocol Description of PDT-ALOHA . . . . . . . . . . . . . . 38 3.3 Mathematical Analysis of PDT-ALOHA . . . . . . . . . . . . . . . . . 39 3.3.1 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 3.3.2 Key Analytic Result . . . . . . . . . . . . . . . . . . . . . . . 40 3.4 Analysis and Comparison with Protocol Simulation . . . . . . . . . . . 41 3.4.1 Effect of Guard Time on Throughput . . . . . . . . . . . . . . 42 3.4.2 Effect of Delay Regimes on Throughput . . . . . . . . . . . . . 43 3.4.3 Optimal Guard Time in Unknown Delay Regimes . . . . . . . . 45 3.4.4 Short Hops are Better . . . . . . . . . . . . . . . . . . . . . . . 46 3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 Chapter 4: Time Synchronization for High-Latency, Acoustic Networks 48 4.1 Need for Time Synchronization: Clock offset and skew . . . . . . . . . 50 4.1.1 Sources of Error in Time Synchronization . . . . . . . . . . . 52 4.2 Quantifying the Challenges of High Latency Links . . . . . . . . . . . 54 4.3 Design of Time Synchronization for High Latency Channels (TSHL) . . 57 4.3.1 Overview and Assumptions . . . . . . . . . . . . . . . . . . . 57 4.3.2 Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 4.3.3 Time Synchronization Over Multiple Hops . . . . . . . . . . . 61 4.4 Performance Evaluation of TSHL . . . . . . . . . . . . . . . . . . . . 62 4.4.1 Goals and Methodology . . . . . . . . . . . . . . . . . . . . . 62 4.4.2 Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . 63 4.4.3 Comparative Evaluation . . . . . . . . . . . . . . . . . . . . . 64 4.4.4 TSHL Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 68 4.5 Experimental Evaluation of TSHL . . . . . . . . . . . . . . . . . . . . 71 4.5.1 Experiences with TSHL on the Cricket Platform . . . . . . . . 72 4.5.2 Experimental Result Summary . . . . . . . . . . . . . . . . . . 77 4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 Chapter 5: T-Lohi MAC: Exploiting Spatial Uncertainty 79 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 5.2 Challenges and Opportunities . . . . . . . . . . . . . . . . . . . . . . . 81 5.2.1 The SNUSE Modem . . . . . . . . . . . . . . . . . . . . . . . 82 5.2.2 Spatial Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . 83 5.2.3 Deafness Conditions . . . . . . . . . . . . . . . . . . . . . . . 85 5.3 Tone-Lohi MAC Protocol Design . . . . . . . . . . . . . . . . . . . . . 88 5.3.1 Overview of T-Lohi . . . . . . . . . . . . . . . . . . . . . . . . 88 5.3.2 T-Lohi Flavors . . . . . . . . . . . . . . . . . . . . . . . . . . 90 vi 5.3.3 Discussion on Protocol Correctness . . . . . . . . . . . . . . . 93 5.4 Basic Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . 95 5.4.1 Simulation Methodology . . . . . . . . . . . . . . . . . . . . . 96 5.4.2 Network Throughput . . . . . . . . . . . . . . . . . . . . . . . 98 5.4.3 Energy Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . 103 5.4.4 Protocol Correctness: Impact of Deafness and Aggression . . . 104 5.4.5 Impact of Contender Detection and Counting . . . . . . . . . . 105 5.5 Analyzing T-Lohi Reservation Period . . . . . . . . . . . . . . . . . . 106 5.5.1 Super-Rounds . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 5.5.2 A Simplified Analytical Solution . . . . . . . . . . . . . . . . . 110 5.5.3 Markov Chain Model . . . . . . . . . . . . . . . . . . . . . . . 111 5.5.4 Generalized Reservation Period Duration . . . . . . . . . . . . 114 5.6 Evaluation of Design Alternatives . . . . . . . . . . . . . . . . . . . . 119 5.6.1 Choice of Contention Round Duration . . . . . . . . . . . . . . 119 5.6.2 Comparison of T-Lohi Flavors . . . . . . . . . . . . . . . . . . 121 5.6.3 Comparison with existing MAC protocols . . . . . . . . . . . . 122 5.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 5.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 Chapter 6: Tone Self-Multipath 127 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 6.1.1 Impact of Multipath . . . . . . . . . . . . . . . . . . . . . . . 130 6.2 SRTL: Learning to Ignore Echos . . . . . . . . . . . . . . . . . . . . . 132 6.2.1 Overview of SRTL . . . . . . . . . . . . . . . . . . . . . . . . 132 6.2.2 Bayesian Inference Background . . . . . . . . . . . . . . . . . 135 6.2.3 Sampling in SRTL . . . . . . . . . . . . . . . . . . . . . . . . 136 6.2.4 Modeling Truth and Observations . . . . . . . . . . . . . . . . 137 6.2.5 The SRTL Algorithm . . . . . . . . . . . . . . . . . . . . . . . 138 6.3 Experimental Evaluation of SRTL . . . . . . . . . . . . . . . . . . . . 143 6.3.1 Experimental Methodology . . . . . . . . . . . . . . . . . . . 144 6.3.2 SRTL Correctness . . . . . . . . . . . . . . . . . . . . . . . . 148 6.3.3 Robustness to Noise . . . . . . . . . . . . . . . . . . . . . . . 150 6.3.4 SRTL in a Changing Environment . . . . . . . . . . . . . . . . 153 6.4 SRTL Parameter Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . 154 6.4.1 Estimates of Observation Errors . . . . . . . . . . . . . . . . . 154 6.4.2 Parameter Alignment with Environmental Conditions . . . . . . 157 6.4.3 Impact of Bin Discretization . . . . . . . . . . . . . . . . . . . 160 6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 vii Chapter 7: Future Work and Conclusions 163 7.1 Short-Term Future Work . . . . . . . . . . . . . . . . . . . . . . . . . 163 7.2 Long-Term Future Directions . . . . . . . . . . . . . . . . . . . . . . . 166 7.2.1 Underwater Acoustic Link-level Characterization . . . . . . . . 167 7.2.2 Incorporating Delay Tolerant Networking Ideas . . . . . . . . . 167 7.2.3 Hybrid Underwater Networks . . . . . . . . . . . . . . . . . . 168 7.2.4 Capture-aware Protocol Design . . . . . . . . . . . . . . . . . 169 7.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 Bibliography 174 viii List Of Tables 1.1 Comparison of communication characteristics of several wireless tech- nologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 2.1 Sampling of Acoustic modems . . . . . . . . . . . . . . . . . . . . . . 15 4.1 Transmit to detection delay jitter for fixed node positions. . . . . . . . 75 5.1 Acoustic Modem Power Draws . . . . . . . . . . . . . . . . . . . . . . 102 5.2 Table comparing the performance of T-Lohi Flavors . . . . . . . . . . . 121 6.1 Payoff table used in determining decision threshold that maximizes payoff.142 6.2 Research questions asked about the merits of our Bayesian learning algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 ix List of Figures 1.1 Underwater Acoustic Sensor Network Topology . . . . . . . . . . . . . 4 2.1 The two broad schemes for time-synchronization . . . . . . . . . . . . 17 3.1 Illustration of space-time uncertainty . . . . . . . . . . . . . . . . . . . 30 3.2 Vulnerability intervals for ALOHA and slotted ALOHA. . . . . . . . . 31 3.3 Classical throughput analysis for ALOHA [BG96]. . . . . . . . . . . . 32 3.4 Throughput of pure-ALOHA is not affected by propagation delay (rep- resented by delay parametera). . . . . . . . . . . . . . . . . . . . . . . 35 3.5 Throughput of slotted ALOHA degrades with any propagation latency (all delay curves overlap). . . . . . . . . . . . . . . . . . . . . . . . . . 36 3.6 Slotted transmission results in cross slots overlap at receiver. . . . . . . 37 3.7 Time diagram of packet transmission using PDT-ALOHA;A andB are transmitters andR is the receiver. B locates closer to the receiver thanA. 39 3.8 Throughput of PDT-ALOHA as guard time length¯ is varied. . . . . . 41 3.9 Maximum throughput of PDT-ALOHA as the normalized propagation delaya is varied. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.10 Comparison of throughput capacity for ¯ 0 = 0:69 and the ¯-optimal PDT-ALOHA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 4.1 Effect of clock skew . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 4.2 Sources of error in estimating message latency . . . . . . . . . . . . . . 52 4.3 Phases in TSHL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 4.4 Comparison of Instantaneous Error: Distance Variation . . . . . . . . . 65 x 4.5 Comparison of Error:Receive Jitter . . . . . . . . . . . . . . . . . . . . 66 4.6 Comparison of Error: Time since Sync . . . . . . . . . . . . . . . . . . 67 4.7 Effect of clock skew . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4.8 Effect of Changing the Number of Sync Beacons . . . . . . . . . . . . 70 4.9 Varying the Clock Granularity . . . . . . . . . . . . . . . . . . . . . . 71 4.10 A Blow up of the sources of uncertainty we identified specifically for the Cricket platform used in our hybrid RF-Ultrasound testbed. Note that these are below MAC level, above which we have factored out by timestamping after MAC access. . . . . . . . . . . . . . . . . . . . . . 74 5.1 Spatial Unfairness: (a) Transmitter and close neighbors have channel cleared earlier. (b) In slotted access, close neighbor A can attempt in slot 3 while C and D can not. . . . . . . . . . . . . . . . . . . . . . . . 83 5.2 Spatio-temporal uncertainty in acoustic medium access. . . . . . . . . . 84 5.3 The three cases where deafness can occur. (a) Bidirectional deafness. Unidirectional deafness at B with A’s tone reaching B (b) before B starts transmitting, (c) after B starts transmitting. . . . . . . . . . . . . . . . . 85 5.4 The Tone-Lohi protocol frame . . . . . . . . . . . . . . . . . . . . . . 88 5.5 Overview of (a) ST-Lohi, (b) UT-Lohi . . . . . . . . . . . . . . . . . . 90 5.6 Benefit of (a) Higher contention, (b) Aggression and asynchrony. . . . . 92 5.7 Maximum theoretical utilization for T-Lohi protocols as ¹ is varied, showing the operational points for our simulations. . . . . . . . . . . . 96 5.8 Channel utilization of three T-Lohi flavors. The vertical lines show the channel capacity and the protocol capacities, in packets/s. . . . . . . . 97 5.9 Average number of contention rounds in a reservation period for ST-Lohi.100 5.10 Channel utilization as¹ is varied by changing the packet length. . . . . 101 5.11 Relative energy overhead for T-Lohi for an 8 node network . . . . . . . 102 5.12 Packets lost in a fixed duration as offered load is varied . . . . . . . . . 103 5.13 Jain’s fairness index for T-Lohi that can count contender vs. a MAC that can only detect contenders and uses BEB. . . . . . . . . . . . . . . . . 105 xi 5.14 Super-rounds: Example of a T-Lohi reservation process broken into con- ceptual super-rounds. The number in the contention round represents the number of nodes attempting contention. . . . . . . . . . . . . . . . 107 5.15 Average Length of Super-round (in number of contention rounds) with variable number of nodes . . . . . . . . . . . . . . . . . . . . . . . . . 108 5.16 A special 2-Node model of the T-Lohi reservation process. . . . . . . . 112 5.17 A generalized Markov chain model to analyze the convergence time of a T-Lohi reservation period. . . . . . . . . . . . . . . . . . . . . . . . . 114 5.18 Average Length of Reservation Period (in number of contention rounds) for a varying network density. . . . . . . . . . . . . . . . . . . . . . . 116 5.19 Impact of different contention round length on Unsynchronized T-Lohi 119 5.20 Comparison of different MAC approaches with aUT-Lohi with bursty traffic and infinite packet queuing (the graphs consider slightly different ranges of offered load). . . . . . . . . . . . . . . . . . . . . . . . . . . 123 6.1 Self and receiver multipath shown for a single reflecting surface at a distance h from transmitting node A. Similar multipath will occur for each reflecting surface in the environment. . . . . . . . . . . . . . . . 130 6.2 Key idea behind the sample collection process for SRTL algorithm: A node sends a tone and waits for a Sample period. Bin location of echos repeat but that of non-echo detections do not. . . . . . . . . . . . . . . 133 6.3 Controlled Experiments: The Setup in the anechoic chamber. On the left was the transmitter/receiver setup and the right figure shows a reflecting surface whose location was varied in our experiments. . . . . 144 6.4 Uncontrolled Experiments: Uncontrolled experiments were performed at two locations. The lab/office location provided for in-air experiments, while the test setup off the docks in Marina del Rey harbor provided for underwater experiments. . . . . . . . . . . . . . . . . . . . . . . . . . 144 6.5 Experimental results showing the CDF of 200 samples and SRTL response with objects at known location, adjusted for measurement error (shown as the shaded area with dashed boundary). . . . . . . . . . . . . . . . 148 6.6 Results of a controlled experiment (object at 3.87m) showing the CDF of 200 samples and SRTL response, with varying levels of noise. . . . 149 6.7 Laboratory experiment empirically find two stationary reflections with two levels of artificial noise. . . . . . . . . . . . . . . . . . . . . . . . 150 xii 6.8 SRTL response to change in location of reflecting surface from 3.36m (bin 5) to 4.11m (bin 7) with different input parameters. . . . . . . . . . 152 6.9 Impact of parameter choice for an uncontrolled environment. . . . . . . 154 6.10 Fraction of false positive and false negative as simulated noise and cho- sen noise estimates vary. Error bars show 95% confidence intervals. (They-axis starts below zero to show values along the origin.) . . . . . 157 6.11 SRTL belief for a single bin being SR for a worst case scenario of reflec- tion time alternating between adjacent bins. . . . . . . . . . . . . . . . 160 xiii Abstract An understanding of the key areas of difference in acoustic underwater sensor networks and their impact on network design is essential for a rapid deployment of aquatic sen- sornets. Such an understanding will allow system designers to harvest the vast literature of research present in RF sensornets and focus on just those key aspects that are dif- ferent for acoustic sensornets. Most complexities at the physical layer will eventually be handled either by assuming short ranges or with technology advancements making complex algorithms both cost and power efficient. However, the impact of large latency and the resulting magnification of multipath will remain a great impediment for devel- oping robust sensor networks. This thesis contributes towards an understanding of, and solutions to, the impact of latency on sensornet migration to an underwater acoustic environment. The thesis of this dissertation is that Latency-awareness allows both migration of existing terrestrial sensornet protocols and design of new underwater protocols that can overcome and exploit the large propagation delay inherent to acoustic underwater networks. We present four studies that contribute to this thesis. First, we formalize the impact of large propagation delay on networking protocols in the concept of space-time uncertainty. Second, we use the understanding developed from this con- cept to design the first high-latency aware time synchronization protocol for acoustic xiv sensor networks that is able to overcome an error source unique to the underwater envi- ronment. Third, we exploit the space-time volume during medium access to propose T-Lohi, a new class of energy and throughput efficient medium access control (MAC) protocols. Last, with our protocol implementations we are able to indicate the impor- tance of a different type of multipath which we call self-multipath. This self-multipath adversely affects the throughput of T-Lohi MAC, and to overcome this affect we develop a novel Bayesian learning algorithm that can learn-and-ignore such multipath. xv Chapter 1 Introduction Nothing in the world is more flexible and yielding than water. Yet when it attacks the firm and the strong, none can withstand it, because they have no way to change it. So the flexible overcome the adamant, the yielding overcome the forceful. Everyone knows this, but no one can do it. —Lao Tzu 1.1 Overview Two-thirds of our planet Earth is geographically covered by water. Half the world’s population is within 100 km of the ocean; nearly all of it is encompassed if habitation near rivers and lakes are considered. Water bodies, by absorbing solar energy at the tropics and circulating this energy with a gigantic network of hot and cold water current, are the single biggest factor determining the long-term climate of continents. Moreover water forms the basis for all living processes; from its importance in raising crops, to satisfying the thirst of all living creatures, and to providing a renewable food source with the huge diversity of aquatic wildlife, water—and therefore water bodies— is/are perhaps mankind’s most important natural resource. For all its importance to humanity there is surprisingly little that we know about Earth’s water bodies. Just 10% of all the ocean volume has ever been charted by man. Even this perfunctory knowledge is constrained mostly to the surface and water bod- ies near to coastal habitations. The cause for such limited knowledge is not lack of 1 human inquisitiveness but the lack of means. For underwater is an area that is intrin- sically inhospitable to human exploration. As recently as the last century humankind did not possess technological capabilities to survive or observe most of the underwa- ter phenomena. With technological advancement, and the knowledge that the seas hold vast reserves of natural resources, we have finally devised mechanisms to improve our understanding of the underwater world. 1.2 Motivation With a multitude of observational tools ranging from deep-sea diving equipment and sonar based ranging and visualization, to tethered and robotic submersibles, we finally possess the capability to start exploring the underwater world in much finer detail. Human presence is still mostly needed either due to the remoteness of the oceans or the nature of these observation tools or sensors. Also using a single sensor to observe a physical phenomena provides a single observation point that can provide either a spatial trend (by locomotion) or a temporal trend (by repeated observation). This deficiency leads naturally to a solution by deploying of a group of such sensors that also necessar- ily collaborate and provide a complete 4-D (of time and space) picture of an observable phenomena. The deployment of such a collaborative group, or network, therefore lead to research in underwater communication and networking. Recent research is pointing towards a growing interest in extending terrestrial wire- less sensor networks to an aquatic environment [HLS + 06, APM05, PKL06]. Under- water applications like 4D seismic monitoring of sub-sea oilfields [HLS + 06], tagging of aquatic wildlife [TOP06], coral reef monitoring [Net07], and monitoring of maritime borders currently rely on either complex, high cost deployments or on non-real time, low granularity (in both space and time) solutions. A robust underwater sensor network will 2 Table 1.1: Comparison of communication characteristics of several wire- less technologies Characteristic Satellite IEEE 802.11 Mica2 Radio Short-Range UWA [WYH06] Long-Range UWA [Ben] Bit Rate 155 Mb/s 11 Mb/s 20–50kb/s 1 kb/s 2.4Kb/s Typical BER 10 ¡10 10 ¡5 10 ¡5 10 ¡5 10 ¡7 Propagation Delay »120 ms < 1¹s < 1¹s »300 ms »6s Distance » 42,000km <3km <:5km <:5km 6–10km provide a real-time, high-granularity, and low cost alternative and enable, for example, optimal extraction strategies for sub-sea oil fields. 1.2.1 Communication Medium for UWSN Underwater sensor network communication requirements rule out many communica- tion mediums. High frequency radio propagation (RF) has a very limited range under- water; thus mica-2 [Cro02] transmit range has been measured as less than 1m in fresh water [ZSR04]. Low frequency propagation, though possible, requires a long antenna size with high transmission power. An underwater RF modem has recently been devel- oped, but its technical details are vague [Sys07]. The range depends upon data rate, power requirements and the depth of receiving communicating nodes (as having a por- tion of wireless path over the air increases range). Optical links are another option for communication underwater [VKR + 05]. Although good for point-to-point links, they are not adept for a distributed network because of their very short range (less than 5m) and because narrow beam optical transmitters requires precise positioning. Acoustic communication provides a reliable medium that is most suitable for short range, low-cost, adhoc, and dense underwater sensor network (UWSN). Acoustic com- munication allows omnidirectional transmission with distributed channel access, and 3 Buoys Radio Radio Buoys Platform Acoustic Acoustic Buoys Radio Radio Buoys Platform Acoustic Acoustic Buoys Radio Radio Buoys Buoys Buoys Radio Radio Buoys Buoys Platform Acoustic Acoustic Platform Acoustic Acoustic Figure 1.1: Underwater Acoustic Sensor Network Topology has minimal signal attenuation (relative to RF). The underwater acoustic (UWA) chan- nel, despite being a suitable communication medium, introduces a host of new com- munication challenges. UWA channel magnifies the impact of RF wireless bandwidth limitations, transmit energy costs, multipath spread, and doppler shift. This magnifi- cation is due to temporal and spatial sensitivity of channel characteristics to temper- ature, salinity, pressure, frequency, water currents, wind speed, and type of sea bot- tom [Bur84, Uri91, HLS + 06, APM05]. Stojanovic also identifies a number of sources of signal loss [Sto03]: spreading and absorption loss (a function of distance and fre- quency); multipath reflections from the surface, obstacles, the bottom, and tempera- ture variations in the water; noise due to artificial and natural sources; and scattering from reflections off a potentially rough ocean surface. Table 1.1 compares radio-based networks for general computing, sensor, and satellite networks with short (SNUSE modem [WYH06]) and long range (Benthos ATM-885 [Ben]) acoustic networks. With bitrates on the order of last generation sensornet platforms, and delay on the order of satellite networks, acoustic communications can be thought of as a long-slim pipe, with the worst features of satellite and RF-based wireless links. 1.2.2 The Vision of UWSN Many of these forms of loss are unique to acoustic communication at longer distances. In particular, multipath reflections, temperature variation, and surface scattering are all 4 exaggerated by distance. Inspired by the benefits of short-range RF communication in sensor networks, we seek to exploit short-range underwater acoustics. We propose a multi-hop acoustic network targeting communication distances of 50–500m and com- munication rates of 5kb/s. Using a simple FSK signaling scheme we anticipate sending 5 kb/s over a range of 500m, using a 30mW transmitter output. The primary limitation on performance is set by spreading loss and the background noise of the ocean. As with RF, we expect a combination of software and hardware techniques such as duty cycling can result in energy requirements a fraction of the basic transmit costs. Figure 1.1 shows a conceptual view of a target system using underwater acoustic communication. A relatively dense deployment of sensor nodes (small circles) are con- nected by wireless underwater acoustic links. These communicate, possibly via mul- tiple hops, with each other and with tethered nodes (small squares) at buoys to users on a platform or ship. Where possible we can exploit a hybrid network with high- speed RF communication on buoys (large squares), or with underwater or surface-based robotic or manually operated data mules (not shown). We see this network being used for applications such as 4-D seismic monitoring of oil reservoirs, underwater mining and exploration, perimeter monitoring, marine border enforcement, and ecological system monitoring. There are several other possible applications that require different deployment archi- tectures. Oil and technology companies are envisioning more rapid and frequent under- water structure construction. Such applications range from the monitoring and assis- tance in underwater structure completion. Other such applications include a floating, and therefore mobile, sensornet to track long-term trends in ocean currents, aquatic wildlife, and the impact of global warming on oceans. Another interesting application 5 of underwater sensornets is to randomly deploy sensors in a no-fishing zone and then enforce the policy by detecting trawler pickup. The energy consumption of acoustic hardware is significantly different from ter- restrial sensornets, with transmission often 100 times more expensive than recep- tion [PKL06]. For example the typical receive power for the WHOI micro-modem is 80mW, while the transmit power is 10W (a receive:transmit ratio of 1:125) [FGS + 05], while short-range radios for sensornets generally provide ratios around 1:1.5 [SK97]. Thus collision and data retransmission are expensive for an acoustic environment. 1.2.3 Acoustic Propagation and Sensornet Protocols Large propagation latency in acoustic communications, however, poses the most sig- nificant challenge from a networking perspective. Sound propagates (on average) at 1500m/s, which is five order of magnitude slower than that of wireless RF using elec- tromagnetic waves propagating at nearly the speed of light. While industry standard terrestrial protocols consider propagation delays, the delays are typically small (20¹sec for 802.11). Protocols designed for terrestrial radio networks, either ignore the propa- gation delay [EGE02, MKSL04] or provide solutions that do not perform satisfactorily under a high delay regime [GKS03]. Under such cases it is imperative to incorporate the large delay constraint in designing protocols for UWSN as we do in our time syn- chronization and medium access protocols [SH06, SYH08]. Thus, the central problem addressed in this thesis is that while there is a nearly 1-to-1 mapping of applications between terrestrial and underwater sensor networks, migrating protocols directly to the large delays of acoustic networks significantly degrades their performance. The question therefore is whether we can capture the essence of this degradation and use the developed understanding to migrate and develop new protocols 6 for UWSN that continue to provide similar application level performance as terrestrial sensornets. 1.3 Thesis Statement The thesis of this dissertation is that Latency-awareness allows both migration of existing terrestrial sensornet protocols and design of new underwater protocols that can overcome and exploit the large propagation delay inherent to acoustic underwater networks. 1.3.1 Summary of Work We next present four studies that contribute to this thesis. First, we formalize the impact of large propagation delay on networking protocols in the concept of space-time uncer- tainty. Second, we use the understanding developed from this concept to present the design of a high-latency aware time synchronization protocol. Third, we exploit the space-time volume during medium access to propose a new class of energy and through- put efficient medium access control (MAC) protocols for UWSN. Lastly, we observe high latency induced self-multipath during protocol implementations and propose a bayesian learning algorithm to identify such multipath. First, in Chapter 3 we formalize the impact of high propagation latency on acoustic networks with the concept of space-time uncertainty [SYKH07]. We explore the impact of this uncertainty on wireless protocols in general by using ALOHA, an early wireless MAC protocol, as a case study. By understanding the impact of this uncertainty we show that a simple solution can be proposed. This solution allows us to argue that shorter hops 7 can be throughput efficient, beside being energy efficient [PK00], for an acoustic sensor network. Next, in Chapter 4 we present a Time Synchronization protocol that is aware of the High Latency (TSHL); the first time synchronization protocol for acoustic sensor networks [SH06]. We first show that the best current time synchronization schemes for terrestrial RF sensor networks (such as RBS [EGE02] and TPSN [GKS03]) loose their designed microsecond level accuracy under the spatial uncertainty of orders of magnitude large propagation delays in UWSN. By first identifying the portion of time synchronization affected by the high latency we design a new protocol that removes the source of error. We then show that our protocol maintains accuracy independent of propagation latency and at accuracy comparable to best current terrestrial sensornet time synchronization protocols. We then present, in Chapter 5, an energy and throughput efficient MAC designed for the special constraints of UWSN [SYH08]. Any contention based MAC has to deal with access latency on the order of maximum propagation delay that makes channel arbitration significantly difficult. This delay is significant (330ms for 500m) for UWA channel and leads to significant degradation of throughput. We propose T-Lohi, a class of MAC protocols, that leverages the spatial uncertainty in medium access to allow rapid and throughput-efficient contention resolution. Our acoustic modem provides an ultra low-power wake up circuit that allows energy efficient contention [WYH06]. We demonstrate the efficacy of our protocol using simulations and validate these results using rigorous mathematical analysis. Finally, in Chapter 6 we present an algorithm that learns to identify self-multipath of acoustic tones. An important side affect of large latency in UWA is the magnifica- tion of multipath delays. Multipath rays travel similar distances in both RF and acoustic 8 environment, but the slower propagation delay results in a much larger delay spread for the UWA channel. While equalization techniques can compensate such multipath at a receiver, and a rake receiver can benefit from it, delay spreads in the order of tens of symbol duration increase their complexity [SCP94]. This delay spread poses another problem of self-multipath: reception of own transmission via multi-path echoes. Since this delay spread can be as large as 100ms even for short range (less than 500m) net- works, dealing with the resulting interference demands a new approach. Moreover tone echoes can significantly degrade the throughput of T-Lohi. We develop a Bayesian learn- ing algorithm which constantly updates its belief about the identity of echoes based on samples taken after each tone transmission. We perform an extensive set of experiments using the SNUSE modem to understand the behavioral limits of the algorithm. Our results show that the algorithm is correct, robust to noise, and capable of adapting to a dynamic environment with varying surface locations. Our research efforts across these network layers provides a holistic view of the chal- lenges of point-to-point latency for UWSN. Our efforts to overcome and exploit the latency point that characteristics problems, similar to those in wireless RF networks over mutlihop/wide-area networks, appear in singlehop/local-area of acoustic underwa- ter sensor networks. Thus the issue of skew during message exchange adding to time synchronization error observed for singlehop synchronization in UWSN (Section 4.3) would also appear for multihop protocols such as NTP [Mil89] (it is however negligi- ble, and has therefore been ignored, when compared with other sources of error such as network congestion). Similarly, there appears to be a striking similarity between the hidden terminal and exposed terminal problem of RF multihop networks to the failure of carrier sensing in UWSN. In both cases the cause of failure is a transmitter-centric view; what is actually important is the view at the receiver. Thus our insight shows that 9 the propagation latency adds another layer of issues to consider in wireless single-hop communication, but that these challenges are similar to those considered extensively in literature for multi-hop networks. We document a broad range of research related to our work in Chapter 2. We con- clude our thesis with overall comments about the direction of research in UWSN and possible further line of research in Chapter 7 1.4 Contribution and Novelty The contribution of this thesis are three-fold. First, we were the first to formalize a clear and concise understanding of the impact of acoustic delays on networking protocol with the space-time uncertainty principle. We believe that while many of UWA channel complexities at the physical layer will be eventually handled (either by assuming short ranges [HLS + 06] or with technology advancements), the impact of these large propagation delays on performance of proto- cols is the most significant. While this concept has been present implicitly in previous work, we provided a concise treatment of the fundamental concept relating propagation delays to uncertainty, and hence possible errors, in UWSN protocols. Such an under- standing will allow system designers to harvest the vast literature of research present in terrestrial sensornets and lead to rapid development of UWSN protocols. Second, we develop protocols that overcome and exploit the large propagation latency in acoustic networks. We were the first to identify a source of error that was ignored in existing RF sensornet time-synchronization protocols, but becomes impor- tant during large propagation delays. We overcame this problem and designed the first underwater sensornet time-synchronization protocol. This protocol development indi- cates that migration of protocols from terrestrial networks is entirely feasible, but with a 10 comprehensive understanding of the impact of spatial-uncertainty of protocols. We also use the understanding developed to actually exploit the vacant spatial volume present in medium access and design a new family of MAC protocols that provide some novel features unique to high latency of acoustic UWSN. This new protocol is one of the first MAC protocols for acoustic sensornets, and provides both highly energy efficient medium access that also provides efficient and flexible (to application requirements) throughput. Both these protocol development provide a successful template for devel- opment of future underwater networking protocols. Lastly, our experience with experiments and protocol implementation led to the iden- tification of self-multipath; a unique form of multipath that forces us to reevaluate how multipath needs to be handled. We realize that the problem needs to be solved in the context of the cheap and energy-efficient paradigm of sensornets, ruling out most of the traditional approaches. Thus our efforts to implement UWSN protocols led to identifi- cation of a new and unique problem. We develop a novel Bayesian learning algorithm which helps us in identify self-multipath using our cheap and energy-efficient hardware. Our experiments in-air and underwater demonstrate the workability of this solution in the new and challenging environment of underwater acoustic. 11 Chapter 2 Related Work There are 10 types of people in the world: those who understand binary, and those who don’t. —Anonymous We now present three areas of research related to the work in this thesis. These three distinct areas are: Underwater acoustic communication (Section 2.1), time synchro- nization in computer networks (Section 2.2), and wireless medium access techniques (Section 2.3). We next position our work in each of these related fields. 2.1 Underwater Acoustic Communication We first present background information describing the unique nature of the underwa- ter acoustic (UWA) channel. We then introduce communication techniques that tackle, among other challenges, multipath in the UWA channel to provide robust communica- tion links. 2.1.1 Underwater Acoustic Channel The UWA channel is broadly divided into the vertical and horizontal channel. While the vertical channel is generally assumed to be quite consistent in its characteristics and has little multipath, the horizontal UWA channel presents many challenges above and beyond the RF channel [Sto96]. We next present these challenges which impact underwater communication techniques. 12 Noise in the UWA channel is significantly different from the wireless RF chan- nel. Most RF receiver designs assume an additive, white, and guassian noise (AWGN) channel and are able to correctly handle noise. However, the underwater channel has (except for short-range deep-water communication) non-guassian and frequency depen- dent ambient noise characteristics [CPH04]. Thus, mechanisms to estimate and filter noise in UWA channel need to be different from the well established AWGN model widely used in wireless RF receivers [Wha71]. Attenuation in the acoustic channel is dependent on two factors [Uri91, Bur84]. The first is a distance dependent spreading loss. Spreading loss results from energy decay from propagation and decays as R ¡2 , where R represents the distance from transmit- ter. The other significant factor in attenuation is due to frequency dependent absorption loss. The acoustic signal decays proportional to e ¡®(f)R , where ®(f) is an increas- ing function of frequency. A practical implication of this frequency dependence is that the acoustic channel becomes effectively bandwidth (and not, like RF, power)-limited channel [KB00]. This bandwidth-limitation occurs since in general the low cutoff of the available bandwidth is determined by ambient noise levels and the high point by absorption. As the absorption loss becomes more significant at larger distances, the effective bandwidth available also decreases with range. Thus, due to frequency depen- dent attenuation the UWA channel also exhibits a unique relationship between range and bandwidth [Sto96, Sto06]. Therefore the underwater channel exhibits larger bandwidth for shorter distances [Sto06]. Kilfoyle and Baggeroer mention an effective range-rate envelope of 40km-kbps for shallow and 100km-kbps for deep water communication systems [KB00]. 13 Significant fluctuation in the amplitude and phase of the acoustic signal is an impor- tant characteristic of the UWA channel. These fluctuations can be due to waves, turbu- lence, temperature gradients, wind, and many other related phenomena that can cause local changes in sound speed [Bur84]. The received signal strength varies not only due to time dispersive multipath (which we will discuss next) but also due to local, within a single path, channel variations [Cat90]. Such local variations combined with the large propagation delay of acoustic transmission means that receiver feedback of channel sta- tus is also not useful with the channel state becoming stale by the time feedback is received. UWA channel is thus one of the few communication channels where the chan- nel characteristics do not change slowly compared to symbol rate. All of the above features of the UWA channel make communication system design challenging. We however focus on multipath, a challenge that becomes uniquely impor- tant due to the large acoustic propagation delay. 2.1.2 Multipath in UWA Communication Combating the large multipath spread to achieve robust data communication is consid- ered the most challenging task of an UWA communication system [Sto96]. In deep oceans, these multi-paths can be described as rays or using wave-guide modes, but in shallow water they are generally due to reflection from discrete reflecting surfaces. The relatively slow sound speed in UWA communication also gives rise to extended multi- path structures [Cat90]. These multipath, combined with refraction of sound due to different temperature, pressure, or salinity results in shadow zones where no acoustic signal can reach. With in-situ sound velocity profile and bathymetric data available, locations of such shadow zones can be predicted. However, such prediction requires detailed information about 14 Table 2.1: Sampling of Acoustic modems Name Type Modulation Bitrate Range Price SNUSE academic FSK 512bps 50-500m < $1000 WHOI micro-modem academic FSK or PSK 100 or 4800 bps 2-4km $4000 UCONN MIMO-modem academic QPSK/MIMO OFDM 6-12 kbps n/a $1000 Benthos ATM-885 commercial MFSK, PSK 360bps 2-8km $9000 Tritech AM-300 commercial Spread Spectrum/Q-PSK 8-16 kbps 2-8km $35000 the channel and environmental conditions which is generally unavailable for an ad-hoc communication network. The UWA channel is highly reverberant with channel impulse response having delay spreads of 100s of millisecond. With such large time and frequency spreading of the UWA channel, non-coherent (FSK-like) systems were the initial choice for UWA com- munication [Cat90, Sto96]. Such systems, while robust to the time and frequency selec- tive fading of the channel, inefficiently use the limited bandwidth available in UWA channel. Coherent systems are much better at utilizing the limited bandwidth of UWA chan- nel. However, such coherent schemes are much more sensitive (than non-coherent systems) to the large multipath spread which can result in inter-symbol-interference (ISI) even after 100s of symbol. An impulse response with such large delays must still be estimated at the receiver to correctly decode the signal transmitted despite ISI. With the rapid fluctuation of the channel response and the large delay spread within each response, the resulting optimal maximum-likelihood estimator becomes very com- plex. Stojanovic et al. were the first to propose a suboptimal, and therefore less complex, Decision Feedback Equalizer (DFE) jointly optimized with a Phased Locked Loop (PLL) that enabled coherent underwater communication [SCP94]. Recently Time- Reversal-Mirror (TRM) has also been considered as a mechanism to handle multipath in an underwater environment [SHK07]. 15 Table 2.1 lists a few representative acoustic modems. A trend is quite apparent; while the coherent communication based modems can deliver high data rates (or longer ranges) their complexity increases the cost of deployment. Our group (SNUSE) at USC/ISI opted for the lower complexity FSK-based SNUSE modem to bring the low- cost, high-density paradigm of sensor networks to underwater networks. Our choice was also based on an understanding that for short-range systems ( < 200m, typical for energy-efficient sensornet-style communication) the phase fluctuations in the acoustic channel are mild [WYH06]. In such cases, relatively simple techniques can yield rea- sonable (for sensor-networks) data rates [Cat90]. Additional studies have since shown that the bandwidth of an acoustic channel increases with shorter range [Sto06]. With similar benefits being observed for multihop networks [ZM07], the short-hop paradigm becomes a desired choice from both energy and throughput-efficiency perspective. The SNUSE modem introduces a low-power tone wake-up circuit for energy-savings in UWSN deployments that remain dormant for long durations [WYH06]. We use this tone detection capability in a unique way to perform medium access (Chapter 5). One of the requirements of our MAC is to send tones and then listen for tones from other nodes to detect contention. This requirement introduces the issue of self-multipath, where nodes receive echoes of sent tones that remain indistinguishable from another node’s tone. The existing physical-layer techniques, such as Rake receiver [PG58] or TRM [SHK07], distinguish between several copies of the same signal at a receiver. Thus such techniques are not directly applicable to the above described self-multipath at the transmitter. Detecting copies of reflection at the transmitter is more closely related to SONAR [KPK81]. However, such signal processing solutions require more compli- cated, costly and energy-consuming hardware than our low-power wake-up detection circuit. Our Bayesian learning algorithm (Chapter 6) is able to use this more limited, 16 Phase offset = [(T2-T1) - (T4-T3)]/2 Propagation Delay = [(T2-T1) + (T4-T3)]/2. T I M E Reference Node A Synchronizing node B T1 T2 T3 T4 (a) Sender-Receiver Synchronization Reference Receiver A Receiver B T A T B { T A } { T B } (b) Receiver-Receiver Synchronization Figure 2.1: The two broad schemes for time-synchronization but also low-cost and energy-efficient, detection capability and still accurately learn the location of echoes using multiple samples. 2.2 Time Synchronization in Computer Networks Time-synchronization is a crucial concept in computer networks. An important notion in time is that it has to be relative to a given reference standard. Lamport clarified the relationship between computer events and global reference time [Lam78]. Time- synchronization in computer networks is motivated by the need to relate computer sensed events to the outside world. This need implies an existence of a reference time or clock, and all nodes in the network synchronize to that reference. Any node synchronizing with a reference clock requires message exchanges that include reference timestamps. As we elaborate in Chapter 4, this message exchange introduces several sources of uncertainty, and therefore error. The entire research field of time-synchronization focuses on identifying, and then eliminating, these source of errors. 17 At the most fundamental level we can delineate clock synchronization into just two schemes: Sender-Receiver (Figure 2.1(a)) and Receiver-Receiver (Figure 2.1(b)). Both of these schemes use different mechanisms to remove uncertainty during message exchange. While the Sender-Receiver scheme is able to remove uncertainty regarding the delay during a message exchange, the Receiver-Receiver scheme removes all uncer- tainty on the transmitter side. All schemes operate within these two basic frameworks. In addition, some schemes will add additional mechanisms to reduce uncertainty, for example the choice of MAC layer-time stamping removes access uncertainty for sensor- network time-synchronization protocols. 2.2.1 Wide-Area-Network Time Synchronization Network Time Protocol (NTP) is widely used in the Internet and is an example of sender- receiver synchronization. It is distinguished by working well over paths with high latency and high variability [Mil89]. The NTP protocol has a long-term, bi-directional exchange of time information to estimate both offset and skew. It incrementally adjusts the local clock frequency to align it with the reference time base. Unfortunately, NTP is a poor match for sensor networks for several reasons. First, it assumes communications are relatively inexpensive, while sensor networks are bandwidth and energy constrained. Second, it is designed for constant operation in the background at low rates. (At a max- imum polling rate of 16 seconds, NTP took around an hour to reduce error to about 70¹s [Els03]). Thus there is an inherent trade-off between the accuracy of NTP and the synchronization overhead. Such a trade-off is not possible for energy constrained sensor networks. For example our protocol (TSHL, Chapter 4.3) exchanges number of broad- cast beacons to compute skew and then perform one bidirectional exchange to compute 18 a skew-corrected offset. In some sense, TSHL and NTP possess the same informa- tion, however TSHL reduces energy consumption by replacing long-term bidirectional communication with a smaller number of unidirectional, broadcast beacons. In addi- tion, TSHL is not constrained by portability requirements and thus exploits MAC-level timestamping to reduce uncertainty and increase accuracy. An interesting extension of NTP considers the Interplanetary Internet (IPin) [Mil04]. The protocol iNTP, as proposed [Mil04], assumes very high latencies but very pre- dictable node position and movement (for example, predictable trajectories of satellites). While we expect the approximate locations of underwater nodes to be known with some accuracy, we expect ocean currents and environmental effects to render position infor- mation insufficiently reliable. An alternate Internet based protocol was clock skew compensation for streaming audio in the Internet [FOL02]. Faced with large and varying path delays, Fober demon- strates how to model the drift of between node clocks without modeling the offset. He uses statistical measures to remove the high jitter expected for their application. Although we could apply these techniques in underwater acoustic networks to remove this high jitter, but they can also be removed considerably in our point to point network through MAC layer time stamping. 2.2.2 Time Synchronization in SensorNets The research closest to our work is time synchronization effort in the sensor networks community. Underwater sensor networks share many of the design goals of surface sensor networks. Thus, energy conservation and longevity given a fixed power budget are common goals. 19 Reference Broadcast Synchronization (RBS) introduced receiver-receiver synchro- nization, completely eliminating transmitter side uncertainties as described in Sec- tion 4.1.1 [EGE02]. RBS accounts for clock skew by modeling the clock with linear regression on multiple Reference Broadcasts. Its accuracy is quite good, about 6¹s on IPAQ’s with 802.11 cards. In addition, it introduces the concept of post-facto synchro- nization, allowing correction of clock errors after data collection rather than ahead of time. RBS is not applicable to our high-latency acoustic environment because as its central algorithm is built on the simultaneous reception of reference broadcasts at all nearby nodes. With underwater acoustics there are large variations in propagation time between nodes, resulting in synchronization error proportional to the propagation delay, which can be in 100s of milliseconds or more. Timing-sync Protocol for Sensor Networks (TPSN) exploited cross-layer optimiza- tions, minimizing the sender and receiver side uncertainty by time stamping packets at the MAC [GKS03]. Since TPSN also uses a two-way (Sender-Receiver) exchange for synchronization, it can therefore factor out most major sources of non-determinism (the Send, Receive and Access time) and propagation delay. The TPSN authors report achieving accuracy of 8¹s on Mica-2 motes. Like NTP, TPSN requires formation of a hierarchy, where each node synchronizes to its parents. However, it does not model skew of the local clock against the reference; instead it only computes offset. It therefore requires frequent resynchronization to correct for drift due to variation in clock skew, and it cannot do post-facto synchronization. The primary reason TPSN is not applicable to our environment is that it does not consider the effect of clock skew during message exchange. Although this effect is tiny during radio-based synchronization, it causes inaccuracies proportional to message propagation latency rises. We show in Section 4.4 that at moderate distances»300m this error can be nearly 30% worse (see Figure 4.4). 20 Flooding Time Synchronization Protocol (FTSP) takes an additional step towards avoiding timestamp uncertainty by timestamping in the MAC and radio message layer in multiple places and using the average of these values to account for byte alignment jitter [MKSL04]. It uses a variation of sender-receiver synchronization. A sender broad- casts its global reference time to all receivers. Receivers use linear regression over mul- tiple broadcasts to model their clock skew and offset. Compared to RBS, this approach reduces the number of message exchanges that need to take place for synchronization, since the message exchange is not between every pair of nodes. For underwater acous- tic networks, however, FTSP suffers from the same problem as RBS, in that it assumes near instantaneous message propagation. This delay is large and variable in underwater acoustic networks. 2.3 Shared Medium Access In any shared medium arbitrating access to the medium is an important task. There is a significant volume of research which deals with different aspects of such medium access control (MAC) requirements. We touch upon the following three areas of MAC design related to underwater sensor networks: Terrestrial RF based MACs, Satellite MACs, and underwater acoustic MACs. 2.3.1 Terrestrial RF-based MACs Packet radio networks pioneered two classes of wireless MACs: centralized protocols such as TDMA and distributed contention-based protocols such as ALOHA [Abr70] and CSMA [TK75a]. While these protocols are widely used in terrestrial wireless networks, they have significant performance loss when applied to high-latency acoustic networks. 21 Recent work on sensor networks has raised the importance of energy efficiency. Scheduled contention in S-MAC [YHE02] and low-power listening (LPL) in B- MAC [PHC04] and WiseMAC [EHD04] are two major approaches to conserve energy. However, they also become less effective in underwater networks. S-MAC synchronizes the listen intervals of neighboring nodes to perform CSMA-based contention. In under- water networks, senders cannot start contention at the same time due to the large and location-dependent propagation delay to a receiver. A naive solution requires significant extension to the listen interval, and thus largely reduces its performance. LPL shifts the burden from receivers to transmitters by using long preambles. However, in underwater networks, the cost of transmission can be two orders of magnitude higher than that in short-range radios. Moreover, LPL still uses a simple CCA (Clear Channel Assessment) approach in contention, which does not handle the space-time uncertainty. SCP-MAC combines both LPL with scheduled contention thus providing very-low power duty- cycling operation [YSH06]. Underwater, however, its energy-efficiency and effective- ness will suffer the drawback for both LPL (high transmit cost and inaccurate CCA) and scheduled contention (long listen intervals). In contrast, our T-Lohi protocol (Chapter 5) uses a novel tone-based reservation mechanism that handles space-time uncertainty and efficiently resolves contention. It also exploits an ultra low power tone receiver to achieve excellent energy efficiency. There are other protocols, such as BTMA [TK75b] and DBTMA [HD02], that use busy tones in dealing with the hidden terminal and exposed terminal problems. These protocols, however, assume separate channels for tones and data, but we only assume a single channel. They do not consider large propagation delays, as they are designed for wireless RF networks. 22 2.3.2 Satellite MACs Satellite networks are an area where protocols do consider large propagation delays in the order of what UWSN experiences, for example, 125ms. However, the special, asymmetric topology in satellite networks largely simplifies their MAC design [JBH78, TI86]. Such a network usually consists of a satellite and many small nodes on the ground. The down link is a simple broadcast channel that requires no MAC. Although the uplink may involve many transmitters, there is only a single receiver, effectively removing the uncertainty in space. It therefore allows existing protocols such as ALOHA to handle contention in time [Pey99]. Alternatively, a centralized MAC can be easily implemented. In comparison, both terrestrial and underwater sensornets have several applications where a fully distributed protocol is needed for medium access in an ad-hoc topology. The design of sloppy slotted ALOHA (SSA) added guard time to the ALOHA slots for satellite networks with a single, centralized receiver (the satellite) [CW92]. In such networks variable propagation delay is induced by the imperfection (or “sloppiness”) of each node’s implementation, not by the location of each node. In fact, nodes are located on the ground suffer nearly equal delays to the satellite. Our work in Chapter 3, where we also introduce guard times to ALOHA, focuses on ad-hoc acoustic sensor networks where the relative distance to the receiver can vary greatly from node to node. An important difference between satellite MACs and sensornet MAC protocols in general is that energy efficiency is not a design parameter for satellite networks. A common characteristic in terrestrial sensor and satellite networks is that they usually have abundant bandwidth, so channel utilization may not be a big concern. However, the bandwidth in the underwater acoustic channel is very limited, and thus underwater MAC’s need to have efficient channel utilization. 23 2.3.3 Underwater Acoustic MACs Initial result for UWSN MAC targets the development of basic understanding of how latency affects medium access. ALOHA protocols, being the simplest of shared medium access technique, received significant attention. ALOHA protocols were first ana- lyzed for underwater networks with Vieira et al. performing simple analysis of slotted ALOHA and predicting that slotted ALOHA’s performance degrades to pure ALOHA under high latency [VKLG06]. Xie et al. [XC06] have compared the performance of ALOHA and CSMA with RTS/CTS mechanism for underwater networks. Gibson et al. [GXXC07] have extended this work to analyze the performance of ALOHA in a lin- ear multi-hop topology. These papers, however do not attempt to address the following questions: why does pure ALOHA’s performance in underwater remain the same as in RF? why does slotted ALOHA’s performance degrade to pure ALOHA in the presence of varying propagation delay? How can this degradation be handled and what are the optimal parameters for it? In Chapter 3 we specifically address these questions and provide solutions. Concurrently several MAC protocols have recently been proposed for underwater acoustic networks that also deal with high latency. Early work uses naive CSMA with RTS/CTS (Seaweb 2000 [Ric05]), resulting in low throughput. The other work employs CDMA by developing code distribution techniques [XG00], which has high energy cost. The focus of these work has not been short-range or cost effective sensor networks, and as such these solutions increase the cost and energy expenditure of a sensor network. Rodoplu and Park extend S-MAC’s schedule synchronization to sender-receiver pairs in UWSN [RP05]. It allows energy-efficient operation, but lacks effective mechanism and flexibility for on-demand contention. As a result, the protocol is only suited for applications that have extremely low traffic rates. S-FAMA extends the floor acquisition 24 principal using a slotted RTS/CTS exchange that prevent collisions. There is however an round-trip penalty per packet attempt as the slots equals the round-trip delay [MS06]. Thus, this protocol has low channel utilization and without an effective mechanism save energy during the control slots, remains energy inefficient. After these initial incremental works that attempted to overcome the spatial- uncertainty induced challenge in medium access, several MAC protocols have been propose that attempt to exploit the space-time volume available during a packet’s transit through the network [PS06, GFR07, CsSC08, YS08] Peleato and Stojanovic extend S- FAMA by proposing Distance-Aware Colllision Avoidance Protocol (DACAP) [PS06]. DACAP uses the fact that inter-node distance is seldom the maximum transmission range and sufficiently stronger signal will be captures, thus allowing less than RTT penalty per packet. However, just like S-FAMA it does not optimize for the idle-listen energy. UW-FLASHR is another protocol that provide variable data-transmission slots in which packets to different destinations can be sent simultaneously without ensuing packet collision at different receivers [YS08]. This protocol is, however, designed for isochronous data transmitting applications. In APCAP, the authors suggest utilizing the wait-time for RTS/CTS exchange by sending previously queued packets [GFR07]. However, beyond requiring a synchronized network such pipelining adds a minimum delay between responses to prevent the pipelined packets from colliding. Such addi- tional minimum delay will under-utilize the channel at low-traffic loads. Recently, using receiver-initiated packet trains (RIPT) to coordinate the arrival of packets from multiple nodes at the receiver has been introduced [CsSC08]. While this protocols can effi- ciently use the acoustic medium, the protocol is not able to adapt to a change in network latencies after initialization. Also, it is not clear what sort of application traffic will maximally utilize its receiver-coordinated packet-train concept. 25 Our T-Lohi MAC protocol (Chapter 5) was developed more or less in parallel with these UWSN MAC protocols. We also exploit the space-time volume by sending small packet that reserve the packet transmission. These packets allow us to detect and count other contenders and therefore backoff intelligently based on the instantaneous traf- fic load. With contention based access, our protocol is flexible and can accommodate all sort of traffic patterns with minimal impact on performance. We use a low-power wakeup circuit to assist our protocol in providing highly energy efficient contention res- olution. Thus the T-Lohi protocols offers energy efficiency, good and stable throughput with flexibility for all types of applications. 26 Chapter 3 Space-Time Uncertainty If there is magic on this planet, it is contained in water. — Loran Eisely The key differences, as we mention in the Chapter 1, for acoustic network protocols is the five-orders of magnitude greater than Radio-Frequency (RF) propagation laten- cies, and bandwidths one-thousandth that of RF. Both these aspects have significant impact on the design of networking protocols. Since the basic functionality of network- ing protocols is to transfer packets between nodes in a network we will use medium access in UWSN to develop understanding about space-time uncertainty. Both propaga- tion delays and bandwidth limitation have significant effects on MAC protocol’s control algorithms. We first introduce the space-time uncertainty concept and apply it on ALOHA pro- tocols as a case study. ALOHA protocols have been the basis of many wireless MACs since their invention in the 1970s [Abr70]. They are the first class of contention-based MAC protocols in a shared wireless medium. Later protocols, such as carrier sense mul- tiple access (CSMA), achieve better performance than ALOHA in RF networks, due to their conservative mechanism of “listening before transmitting” [KT75]. However, car- rier sense becomes very expensive in underwater acoustic networks due to the large propagation delay. The effect of the propagation delay on ALOHA protocols has been analyzed by Kleinrock and Tobagi [KT75] showing that the protocols are not sensitive 27 to the propagation delay. However, their analysis does not consider the varying propa- gation delays from different locations of nodes, thus its results do not completely hold for underwater networks. In this chapter we first introduce a concept central to understanding the challenge of underwater acoustic medium access. Thus uncertainty in acoustic medium access is characterized not only due to when data is sent (time uncertainty), something typical for RF based terrestrial networks. Uncertainty about channel state is also characterized by the significantly different propagation latencies from spatially diverse transmitters(space uncertainty), together we call these two characteristics space-time uncertainty. We next show that this concept can help explain the impact of location-dependent propagation latency on the slotted ALOHA (Section 3.1.3). Then, we propose the Propagation Delay Tolerant ALOHA (PDT-ALOHA) protocol to improve the performance of the slotted ALOHA by adding guard times(Section 3.2.1). We explore its performance through extensive simulations and compare them with mathematical results (separately done here [AK08]) with an aim to discover the best operating parameters (Section 3.4). We find that, in the high latency environment of UWSN, throughput capacity (through- put optimized across all loads) of PDT-ALOHA improves by 17–100% compared to slotted ALOHA, depending on the maximum propagation delay of the network. Our analysis also show that across all delay regimes under consideration using a guard time with ¯ = 0:69 keeps throughput within 97% of optimal throughput capacity of PDT- ALOHA. This result indicates that a static parameter value can remain a good choice even with unknown or variable operative delay regime. We find that the throughput capacity decreases as the maximum propagation delay increases reinforcing the benefit of short-range, multi-hop communication in underwater networks besides just energy efficient communication. 28 This chapter is meant to explore intrinsic characteristics of the high latency acous- tic channel in UWSN and provide insight for the exploration and development of new protocols. We will use the understanding of medium access developed here to propose a more sophisticated protocol than PDT-ALOHA in Chapter 5. 3.1 Space-time Uncertainty and the ALOHA Protocol In this section we summarize the concept of space-time uncertainty with regards to medium access. We then explore this concept in terms of the ALOHA protocol in a high- latency environment. This exploration provides us with design guidelines for modifying ALOHA for an underwater MAC protocol which we then present in the next section. 3.1.1 Space Time Uncertainty Channel contention state in short-range RF networks can be estimated quickly since propagation delay is negligible. The large propagation delay of acoustic media makes it essential to also consider the locations of a receiver and potential interferers. Dis- tance between nodes translates into uncertainty of current global channel status: space- time uncertainty. Although prior underwater work implicitly considers this uncer- tainty [MS06, BMP + 06, XC06], we are the first (to our knowledge) to present a sys- tematic description of this principle and its impact on UWSN MAC design and also proposing a general solution (Section 3.2). Consider Figure 3.1: the two concurrent transmissions from A and E are received separately at nodes B and D (Figure 3.1(a)). Conversely, packets transmitted at different times by A and E will still collide at node B(Figure 3.1(b)) These examples are contrary 29 (a) Same transmission time; no collision at B (b) Different transmission time but collision at B Figure 3.1: Illustration of space-time uncertainty to conventional RF understanding that concurrent transmissions will collide but not oth- erwise (due to simple channel sensing). Thus we observe that collision and reception in slow networks depend on both transmitter time and receiver location. This space-time uncertainty can also be viewed as a duality where similar collision scenarios can be con- structed by varying either the transmission times or the locations of nodes. Although, in principle, this uncertainty occurs in all communication, it is only significant where latency is very high. The lesson from understanding this two dimensional uncertainty is that we have to take care of both dimensions to obtain guarantees similar to that in handling transmission time in RF networks. We now explore the performance of pure and slotted ALOHA in an underwater, high-latency networked scenario, in the light of our space-time uncertainty principle. 3.1.2 Analysis of ALOHA with Time Uncertainty We first refresh the classical analysis of the simple, or pure, ALOHA protocol [BG96]. where nodes immediately transmit arriving application packets. This analysis is cen- tered at the transmitter and thus only considers time uncertainty. It makes a simplified 30 Vulnerability Interval Data 2T Time T (a) Aloha Vulnerability Interval Data T Time T Slot 0 Slot 1 (b) Slotted ALOHA Figure 3.2: Vulnerability intervals for ALOHA and slotted ALOHA. assumption that all nodes are equidistant to a single receiver, thus mapping any collision at transmitter to that at the receiver. It further assumes that there are infinite numbers of nodes, which implies that all arriving packets are served at a new node and transmit- ted immediately into the network. The packets that collide are buffered, making nodes backlogged. Such backlogged nodes retransmit after an exponential delay. The total offered load to the network is thus combination of the Poisson arrival and backlogged exponential retransmissions. This results in a combined Poisson packet arrival process to the network with parameterG(n) (expected packet/unit time) wheren represents the number of backlogged nodes in the network. The vulnerability interval (VI) is defined as the time interval relative to a sender’s transmission within which another node’s transmission causes collision [KT75]. Assuming T as the packet transmission time, Figure 3.2(a) shows that the VI is equal to 2T. Without going into the details, if we normalizeT to unit time, the instantaneous throughput of ALOHA becomes [BG96]: TH ALOHA =Ge ¡2G (3.1) Slotted ALOHA extends pure ALOHA to allow transmission only at the start of syn- chronized slots of length T . As Figure 3.2(b) shows, this synchronization ensures that only interfering packets that arrive in slot 0 will result in a collision. It thus reduces the 31 0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.1 0.2 0.3 0.4 0.5 Offered Load G (packets/transmit time) Normalized Throughput Slotted ALOHA ALOHA Figure 3.3: Classical throughput analysis for ALOHA [BG96]. VI from 2T toT by preventing any cross slot overlapping. The instantaneous throughput of slotted ALOHA is thus increased to [BG96]: TH slotted ALOHA =Ge ¡G (3.2) The analytic results for both protocols are shown in Figure 3.3, reproduced from several early works [BG96, KT75]. Slotted ALOHA achieves maximum normalized throughput of 36.8% at aG of 1 packet/transmit time, while simple ALOHA achieves it maximum of 18.4% at 0.5 packets/transmit time. As mentioned above, this analysis is carried out with respect to the transmitter’s time. The assumption of a single receiver that is equidistant to all transmitters will result in a similar vulnerability interval at the receiver—regardless of the propagation delay (as shown by Kleinrock and Tobagi [KT75]). Strictly speaking, these assumptions do not hold for all ad hoc wireless networks, but with short-range RF networks the variation 32 in delay is small enough that it has virtually no effect on performance (for example, 10¹s delay over 25m). In satellite networks delay is long, but there is typically only one sender or receiver. We next show, through extensive simulation, that the performance of ALOHA can be significantly affected in acoustic networks where these assumptions do not hold. 3.1.3 ALOHA with Space Uncertainty In order to understand the impact of location-dependent propagation latency, we now evaluate both simple ALOHA and slotted ALOHA using simulations. In the previ- ous section we presented mathematical analysis of ALOHA protocols with assumptions designed to make the analysis tractable. Modeling the protocol with variable delay using similar mechanisms was beyond the scope of our current work; thus we chose to sim- ulate the impact of propagation delays. We next explain the simulation methodology and its differences from the assumptions in mathematical analysis. Then we present the results of these simulations for both pure and slotted ALOHA and comment on these results in light of the space-time uncertainty concept. Simulation Methodology We run our simulations using a custom-built, packet-level simulator designed for UWSN MAC research [SYH08] 1 . Our simulation scenario is different from the classical math- ematical analysis model. We use a single receiver at the center of circular area with uniformly deployed transmitters. For our simulations (unless explicitly mentioned oth- erwise) we uniformly deploy 32 nodes. We have separately verified that our results hold with packets reception at random nodes in the network as well. All nodes have a single 1 This simulator is available from the authors athttp://www.isi.edu/ilense/software/. 33 packet buffer, and do not retransmit lost packets (no collision feedback). Thus there are no backlogged nodes to cause instability. Each node has a Poisson packet generation source. The mean offered load (as defined for our simulations) is the sum of the mean number of packets generated per packet transmit timeT at each node. We only observe the packets successfully received at our designated receiver. To accommodate different transmission ranges and propagation delays, we utilize a normalizing parametera, defined as the ratio of maximum propagation delay to packet transmission time T (the same definition used by Kleinrock and Tobagi [KT75]). To enable this normalization we use a packet length of 125 bytes, resulting in a transmission time of 1 second (at 1kbit/s). We also assume a constant speed of sound as 1500m/s. We alter the maximum range to simulate different delay regimes (the value of parametera). Each simulation data point is the mean of 25 simulation runs with error bars showing 95% confidence intervals. Our simulation results, as mentioned above are different to the assumptions made for its classical analysis. However, with a 32 node network we observe that the through- put curve (with no propagation delay) exhibits similar characteristics to the analytically predicted results. Thus, we extend our simulation results next to observe ALOHA’s behavior with varying level of propagation delays. Simple ALOHA We first evaluate how the throughput of pure ALOHA is affected by different values of a. 34 0 0.5 1 1.5 2 2.5 3 0 0.1 0.2 0.3 0.4 0.5 Offered Load G (packets/transmit time) Normalized Throughput a=0.01 a=0.5 a=1 Figure 3.4: Throughput of pure-ALOHA is not affected by propagation delay (repre- sented by delay parametera). Figure 3.4 shows the simulation results. We can see that the normalized throughput achieved by simple ALOHA remains the same as when no propagation delay is consid- ered, even under different delay regimes. To understand this result, we need to look at it from the receiver’s perspective in both scenarios: with and without propagation delay. With no propagation delay and packets being transmitted as soon as they arrive, they reach the receiver instantaneously with exactly the same Poisson distribution. Hence the analysis at the transmitter faithfully reflects the situation at the receiver. With differ- ent node locations, the arrival time at the receiver will be offset by a constant delay for any transmitter in a particular topology. However, such arrival at the receiver is still a Poisson process with the same parameter, as when there is no latency. Therefore, from the receiver’s perspective, there is no change in probability of collision, and thus the 35 0 0.5 1 1.5 2 2.5 3 0 0.1 0.2 0.3 0.4 0.5 Offered Load G (packets/transmit time) Normalized Throughput No delay a=0.01 a=0.5 a=1 Figure 3.5: Throughput of slotted ALOHA degrades with any propagation latency (all delay curves overlap). throughput performance is the same as that in Figure 3.3. We should point out that sim- ple ALOHA does not attempt to even compensate for time uncertainty, hence further ignoring space uncertainty has no impact due to their duality. Slotted ALOHA Next we evaluate how slotted ALOHA performs in different delay regimes. Slotted ALOHA does handle one dimension (time) of uncertainty by synchronizing transmis- sion slots. Since it does not take care of the space uncertainty, our initial intuition is that the impact of this uncertainty will increase with larger values ofa. As shown by simulation results in Figure 3.5, the throughput of slotted ALOHA does degrade to a level similar to simple ALOHA when propagation latency is considered. (A similar observation is made by Vieira et al. [VKLG06]). What is more interesting, however, is that such degradation occurs immediately with any propagation delay even 36 Figure 3.6: Slotted transmission results in cross slots overlap at receiver. when it is very small (a=0.01). In order to understand this, we look at how the globally synchronized transmission slots overlap at a receiver. Figure 3.6 shows how the packets transmitted in transmission slots of nodes A and B overlap at the receiver R. Node A’s transmission in slot 1 can collide with any packet transmitted by node B in slot 1 (arrived at B during slot 0) and any one transmitted in slot 2 (arrived at B during slot 1). As long as node A and B are not equidistant to the receiver (or the difference in their propagation delays is an exact multiple of T ), either node’s transmission can collide with the other’s due to a packet arrival during a vulner- ability interval of2T . This is the same vulnerability interval as in simple ALOHA, and thus any propagation latency to the receiver completely losses the benefit of time syn- chronization. If the network always has a single receiver, and nodes knew their relative locations, it is conceivable for slotting to be made relative to the receiver. However this simplification does not match the ad hoc network paradigm where any node can be a potential receiver. Radio networks, although having very small propagation latency, do undergo a sim- ilar performance degradation, as we model any packet overlap as collision. However, most RF systems can usually tolerate an overlap of up to a few bits (depending on cod- ing techniques). As a result for high speed RF networks, if bit rate is 10Mb/s (e.g., IEEE 37 802.11b), the maximum propagation delay that slotted ALOHA can tolerate is 1ns, or 30m in distance. Thus such systems do not exhibit the immediate performance degrada- tion that we have shown for any propagation delay. On the other hand, acoustic systems even with low data rate modems (1Kb/s [WYH06]) can tolerate only 1ms or 1.5m in distance due to much slower speed of propagation (about 1500m/s). Thus the impact of spatial uncertainty for slotted ALOHA will be more evident for any acoustic network than it is for RF networks. 3.2 The PDT-ALOHA Protocol We now postulate that space-time uncertainty can be handled by the addition of extra guard time beyond the transmission time in time slots. These guard times are added to ensure a single slot overlap at the receiver, thus tolerating the large propagation delays. We refer to this modified version as propagation-delay-tolerant ALOHA (PDT- ALOHA). As we argue in Section 3.1.3, while a centralized network can handle large delays by synchronizing slots at the receiver, a similar solution is not feasible for ad-hoc networks where every node can be a potential receiver. We first describe the modified protocol and the intuition on how guard time add tolerance to space-uncertainty. 3.2.1 Protocol Description of PDT-ALOHA In our modification to slotted ALOHA, nodes still transmit only at the start of globally synchronized slots. The slot duration, however, is increased from T to T +¯¢¿ max , where ¯ represents the fraction of maximum propagation delay (¿ max ) that nodes wait after finishing their transmission (Figure 3.7). Choosing¯ = 1 ensures that no overlap 38 Figure 3.7: Time diagram of packet transmission using PDT-ALOHA; A and B are transmitters andR is the receiver. B locates closer to the receiver thanA. at the receiver occurs unless packets are transmitted in the same slot, the guarantee that slotted ALOHA was originally designed to achieve when delay is near zero. However, this value of ¯ results in a long wait time after each packet that will increase packet transmission latency and decreases throughput. With¯ <1 there remains the possibility that some node pairs still have the vulnerability interval of two slot durations (as in Figure 3.6). Therefore, reducing¯ value lowers the bandwidth overhead, but increases collision probability. Based on the intuition that the distance between node pairs is often smaller than the maximum propagation delay, we vary¯ to evaluate the tradeoff between bandwidth overhead and collision probability. 3.3 Mathematical Analysis of PDT-ALOHA Here we recap, for context to the next section, the work by Ahn and Krishnamachari which mathematically analyzes the throughput of PDT-ALOHA for both a finite and infinite node network [AK08]. We first describe the assumptions made in their analysis and how these relate to our simulations. Finally, we briefly introduce the key results of their work that we will use in our next section that compares the analysis with simulation results. 39 3.3.1 Assumptions We now describe the key assumptions intrinsic to the mathematical results obtained by Ahn and Krishnamachari [AK08]. They assume a network withn transmitters uniformly deployed in a two-dimensional disk with a single receiver located at the center of the disk. The propagation speed is a positive finite constant,¿ m . Without loss of generality, the packet transmission time is assumed to be a unit constant. Thus, the normalized maximum propagation delay a = ¿ m . They assume that nodes always have packets for transmission but attempt transmission with probability p in each time slot. This assumption represents a Bernoulli packet arrival to the network with parameterp which for a large node density begins approximating a Poisson packet arrival. Packet collisions are not detected and thus no packet retransmission occur. Transmissions take place in a slot-synchronized, error-less network and transmission time is no less than the maximum propagation time so thata· 1. The assumption thata· 1 ensures collisions confined to consecutive time slots (with this assumption, there is no possibility that a packet sent ini-th time slot collides with another inj-th time slot, wherej = 2fi¡1;i;i+1g). These assumptions are slightly different from the simulation assumptions we describe in Section 3.1.3. However, since our results show that both analysis and simu- lations closely match, these differences are not significant. 3.3.2 Key Analytic Result Ahn and Krishnamachari, based on the above assumptions, model the throughput S (packets per transmission slot) of PDT-ALOHA for a finite number of nodes as a func- tion of the delay regimea, the guard-time parameter¯, and the packet departure proba- bilityp. 40 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Guard band fraction β Normalized throughput Slotted ALOHA (no delay) Pure ALOHA a=0.01 a=0.1 a=0.5 a=1 (a) Analytic result 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Guard band fraction β Normalized Throughput a=0.01 a=0.1 a=0.5 a=1 Slotted ALOHA (no delay) Pure ALOHA (b) Simulation result Figure 3.8: Throughput of PDT-ALOHA as guard time length¯ is varied. S(n;¯;p;a)= npPrfNCjn i g 1+¯a (3.3) Here, PrfNCjng is the probability that there are no collisions for an n node net- work. This probability was derived by splitting the 2-D disk (as described in assump- tions above) into three separate collision regions (for details and exact formulation refer to [AK08]). This equation also shows that the throughput is a monotonically decreasing function of¯. We will use this equation in experimental evaluation of PDT-ALOHA’s performance that compares protocol simulation with analytic results next. 3.4 Analysis and Comparison with Protocol Simulation We now analyze the results of optimal throughput of PDT-ALOHA to observe the effect of guard time and network delay regime. Furthermore, for comparison we simulate PDT-ALOHA to verify the correctness of our analysis in a more realistic network. We end this section by drawing some interesting conclusions from these results. 41 3.4.1 Effect of Guard Time on Throughput We now look at throughput capacity (throughput optimized over offered load) that PDT- ALOHA can achieve as a function of guard time length ¯ which is a fraction of the maximum propagation delay. Figure 3.8(a) shows this function of ¯ as the throughput capacity for a fixed n (32 nodes) using numerical methods for maximizing Equation 3.3 over p. We plot the response for different delay regimes characterized by different values of a. Fig- ure 3.8(b) shows the plot for the exact same parameters. However, here instead of using analysis we derive our results from empirical data collected from simulations. As we see results from both simulation and analysis compliment each other. Both results show that throughput capacity of a network can be increased by using PDT-ALOHA and that the benefit of the guard time is highly correlated to its size and the delay regime in which the network is operating. We also observe two trends as ¯ increases. First, with very smalla= (e.g. 0.01 in simulation results) we see the throughput increases (approaching the optimum) as larger guard time is used due to a decreased inter-timeslot collision probability. Conversely, with largea (e.g. equal to 1 when the propagation delay equals transmission time) the throughput becomes insensitive to the use of guard time. Further- more, simulation results show that for any value ofa beyond 1, the benefit of choosing additional guard time diminishes [AK08]. Thus, choosing a packet length that normal- izes the propagation delay to an appropriate value is essentially to yield the benefits of PDT-ALOHA. 42 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 a (propagation delay/transmit time) Normalized Throughput Slotted ALOHA (no delay) Optimal β β=0.5 β=1 β=0.1 Pure ALOHA (a) Analytical result 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 a (propagation delay/transmit time) Normalized Throughput Slotted ALOHA (no delay) β =0.5 β =1 Pure ALOHA β =0.1 Optimal β (b) Simulation result Figure 3.9: Maximum throughput of PDT-ALOHA as the normalized propagation delay a is varied. 3.4.2 Effect of Delay Regimes on Throughput We next vary a to observe how the throughput capacity is affected by propagation delay in PDT-ALOHA. We generate a figure similar to Kleinrock and Tobagi’s (Fig- ure 10, [KT75]) that shows the impact of propagation delay on throughput capacity for different MAC protocols. However, due to their equidistant and single receiver assump- tion, they showed the capacity of slotted ALOHA not being affected by latency. We have shown this result to be incorrect for general ad-hoc networks in Section 3.1.3. Figure 3.9(a) shows throughput capacity S ¤ (¯;a) as a function of the normalized maximum propagation delay a when the guard time ¯ is given and fixed. They are obtained for n = 32 maximizing Equation 3.3 over p with given ¯ and a. For com- parison, we have the same plot generated from simulation results in Figure 3.9(b). We plot the response using different values of ¯. We also show the ¯-optimal throughput capacity curveS ¤ (a) (choosing¯ that optimizes the throughput capacity PDT-ALOHA over all values of a) using the value of ¯ that maximizes the throughput capacity at a given value ofa (using Equation 3.3)). It can be seen that a fixed value of¯ might lead 43 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 a (propagation delay/transmit time) Normalized throughput Maximum throughput of β=0.69 Throughput capacity S * Figure 3.10: Comparison of throughput capacity for¯ 0 = 0:69 and the¯-optimal PDT- ALOHA. to a suboptimal throughput (this is investigated in more detail in next section). When ¯ = 0:5, PDT-ALOHA is closest to the ¯-optimal curve when a is near 1 but the gap increases asa goes to 0. Conversely, for¯ =1 PDT-ALOHA is closest to the¯-optimal curve for smaller values ofa but becomes inefficient asa approaches 1. Although the throughput decreases monotonically with increasing values of a, we observe very little sensitivity to a with smaller ¯ values. This insensitivity is due to limited collision prevention provided by shorter guard time. Also the monotonically decreasing slope increases with ¯ causing throughput to become more sensitive to a. Figure 3.9 shows that PDT-ALOHA can achieve about 17% (when a = 1) to 100% (whena! 0) improvement on throughput over standard slotted ALOHA in an under- water environment. 44 3.4.3 Optimal Guard Time in Unknown Delay Regimes Our results above show that we can find an optimal value for¯ if a given ratio of maxi- mum propagation delay to packet length is known beforehand. In many practical scenar- ios this ratio will vary, either due to changing delays or different packet lengths. Thus it is interesting to find a single guard time ¯ 0 that will provide good throughput over the delay regimes under consideration. For this purpose we consider the ratio of the throughput capacity for a fixed¯ to the optimum throughput capacityS ¤ obtained over all values of ¯. We then find out the optimal value of ¯ 0 which satisfies the max-min ratio as follows: ¯ 0 =argmax 0·¯·1 µ min a2A S ¤ (¯;a) S ¤ (a) ¶ (3.4) where S ¤ (¯;a) = max p S(¯;p;a), S ¤ (a) = max (¯;p) S(¯;p;a) as defined in Section 3.3, and A is the interested set of delay regimes. In this paper we adopt A=[0:01;1]½R. This optimization formulation provides us with a value of¯ that keeps the through- put capacity closest to the ¯-optimal curve over the set of delay regimes A as shown in Figure 3.9. We find (numerically) that this ¯ 0 is ¼ 0:69 and the value makes S ¤ (¯=0:69;a) S ¤ (a) ¸ 0:97 for all a 2 [0:01;1]. Therefore the throughput of a system with such a fixed guard time can always remain within 97% of the ¯-optimal throughput capacity for PDT-ALOHA. Figure 3.10 shows the throughput capacityS ¤ and the max- imum throughput when¯ = 0:69 and the offered load is optimum. As can be seen, the value of ¯ = 0:69 is a good choice for a practical deployment of PDT-ALOHA where the packet length or range can be expected to change. 45 3.4.4 Short Hops are Better Our analytic and simulation results also show higher throughput can be achieved by using guard time for lower values ofa. For example, assume we use an acoustic modem that has a communication range of 300m and a speed of 1kb/s [WYH06]. If we use packet length of 250 bytes,a will be 0.1, and the modified slotted ALOHA can achieve performance similar to slotted ALOHA in RF networks. Thus, in terms of how much of throughput can be reclaimed, shorter communication hops will provide higher through- put benefit. This conclusion is complimentary to the physical layer argument presented by Sto- janovic that higher throughput in acoustic networks can be obtained using smaller hops [Sto06]. Similar arguments from an information theoretic perspective have also been proposed for bit level [CM06] and multi-hop [ZM07] underwater acoustic net- works. All these results, along with the results in this paper, reinforce the benefit of short-range communication in underwater networks, for reasons beyond energy effi- ciency. 3.5 Summary In this chapter we have explored the impact of spatio-temporal uncertainty on UWSN MAC protocols. For such networks we have shown that location-dependent acoustic propagation delay significantly affects MAC protocols such as slotted ALOHA. Thus it is necessary to consider both space and time uncertainties while designing MAC proto- cols under varying latency environment of an acoustic UWSN. We proposed PDT-ALOHA to deal with the spatio-temporal uncertainty in slotted ALOHA by adding guard times each slots. We have investigated different metrics of its 46 performance — success rate, throughput, and their optimal values— using both mathe- matical analysis and protocol simulations. Our results showed that the throughput capac- ity of PDT-ALOHA is 17-100% better than that of simple slotted ALOHA in an under- water environment. We showed that for the optimal throughput capacity the value of optimal¯ changes based on operating delay regime. However a guard time of¯ =0:69 keeps throughput within 97% of capacity with an optimal ¯ when the traffic load is optimal. Thus, it is a good choice even if the operative delay regime is unknown before deployment. Our results indicate a significant throughput benefit when shorter commu- nication links are used. This result argues for deploying dense, short range, multi-hop networks as opposed to sparse and long range networks currently used in underwater networks. Thus, our work in formalizing the impact of propagation latency has provided a clear understanding which we have shown can lead to development of newer protocols that can overcome the detrimental effects of such delay. 47 Chapter 4 Time Synchronization for High-Latency, Acoustic Networks Roll on, thou deep and dark blue Ocean - roll! Ten thousand fleets sweep over thee in vain; Man marks the Earth with ruin - his control Stops with the shore — Lord Byron While the previous chapter built an understanding of the broader impact of propa- gation delay on underwater sensor networks, we now use that understanding to analyze how delay affects network-wide time-synchronization. Time synchronization is an important part of many distributed applications. In the classical networking and operating system literature it has been widely studied in context of database queries [BGS00] and security applications [NT94]. In the Internet, NTP is the canonical approach to provide distributed time synchronization [Mil89]. Time synchronization is even more important in sensor networks, where applications such as acoustic beamforming [WYM + 02] and target tracking [CHS03] require collabo- ratively processing of time-sensitive data [EGE02]. Sensor networks add the additional requirement that energy consumption of the synchronization protocol be minimized. Many time synchronization protocols for sensor networks have been proposed recently [EGE02, vGR03, MKSL04, GKS03]. These mechanisms provide a high degree of precision while being reasonably energy efficient. The protocols adopt increasingly 48 sophisticated approaches to reduce noise and account for latency in communications, but all assume that propagation latency is negligible and thus can be effectively factored out of design consideration. While assumptions about propagation latency are appropriate for RF-based com- munications (where these protocols were intended), this thesis along with a number of researchers are beginning to explore underwater sensor networks (for example, see [APM05]). Underwater, radio propagation is very limited (Mica-2 transmit range has been measured as less than 1m in fresh water [ZSR04]). Underwater acoustic com- munication provides a viable alternative [Uri91]. A number of recent efforts have begun exploring acoustic communication for underwater sensor networks. In fact, commercial systems exist today, typically focused on long-range communication, but our focus is low-power, low-complexity short-range networks. The high propagation delay of underwater acoustics is especially hazardous for time synchronization algorithms (Chapter 2.2). While NTP tolerates large delay [Mil89], it does not consider energy consumption. Sensor-network-based time synchroniza- tion protocols consider energy consumption, but all current protocols are optimized for RF-based networks and assume nearly instantaneous and simultaneous recep- tion (RBS [EGE02], FTSP [MKSL04]) or ignore clock drift during synchronization (TPSN [GKS03], LTS [vGR03]). This chapter quantifies the inaccuracies the constraints of UWSN impose on current time synchronization protocols (Section 4.2). Finally we introduce a new protocol, Time Synchronization for High Latency (TSHL), that compensates for high-latency commu- nication while also minimizing energy consumption (Section 4.3). The key idea in TSHL is that it splits time synchronization into two phases. In the first phase, nodes estimate clock skew to a centralized time-base. In the second phase 49 , Real Time Clock time Ideal Clock (f(x) =x ) Node Clock ( f(x) = ax +b ) Offset =b Sync Exchange Error Figure 4.1: Effect of clock skew they swap skew compensated synchronization messages to determine the offset. The first phase is impervious to the propagation latency, while the second phase explicitly handles propagation delay induced errors (Section 4.3). We demonstrate the importance of this approach by analysis (Section 4.2) and simu- lations (Section 4.4). Analysis and simulations show that RBS and FTSP are not appli- cable to high-latency networks because they do not consider propagation latency at all. Through simulations we compare TSHL to TPSN, the closest current protocol. An implementation of TSHL on an in-the-air testbed was also performed. We discuss how this in-the-air acoustic network will approximate the latency of an underwater network in Section 4.5. 4.1 Need for Time Synchronization: Clock offset and skew Two challenges face synchronization of distributed clocks. 50 First, they must by synchronized to a single common event in absolute time or offset (shown as b in Figure 4.1). Second, because clocks are imperfect and run at slightly different rates, one must determine the skew of a given clock relative to some reference frequency. Since computers boot at different times, there must be some way for distributed computers to determine a common offset. Offset can be determined by a single message exchange, provided we can compensate for any sources of non-deterministic latency in the path (described in Section 4.1.1). As examples, NTP compensates for jitter in the path by using control theoretic mechanisms [Mil89], while RBS uses a jointly visible broadcast packet to synchronize nodes [EGE02]. Time in most modern, inexpensive computers is derived from oscillating frequency of a quartz crystal. Due to environmental variation (temperature, humidity, etc.) or minor manufacturing differences, variations in crystal oscillation frequency on the order of 15–25 parts per million are common [Atm]. Thus, even nodes that are synchronized to a common offset will drift out of synchronization over time. In Figure 4.1 the dotted y = x line represents an ideally ticking clock where time on the x-axis matches time on the y-axis perfectly. Two clocks ticking at different rates with ratio a will appear as a line with a different slope y = ax where the ratio a represents clock skew. Pro- tocols handle clock skew by estimating and compensating for it. For example, NTP keeps the long term average skew error low by using a phase-locked loop to correct the local clock frequency [Mil89],while RBS explicitly calculates this skew using linear regression [EGE02]. With reasonable quality oscillators, clock skew is generally small enough that it can be ignored over short time intervals. Thus existing protocols ignore skew when synchro- nizing offset. While appropriate for RF-based networks, we show that this assumption 51 Application Transport Network Send Time MAC PHY Access Time Interrupt, Encoding Time Application Transport Network Receive Time MAC PHY Interrupt, Decoding, & Byte alignment Time Propagation Time Time Sync Message Sender Time Sync Message Receiver Sync packet Possible Locations for Time Stamping packet (at both sender or receiver) Transmission Time ReceptionTime Figure 4.2: Sources of error in estimating message latency does not hold for slower acoustic communication (see Section 4.2). For these networks, one must consider skew even during synchronization. 4.1.1 Sources of Error in Time Synchronization The main cause of error in time synchronization schemes is the non-determinism in the latency estimates of the message delivery delay. A description of sources of variability was first described by Koeptz and Schwabl [KS89] and extended recently by Horauer et. al. [HSS + 02] incorporating physical layer jitter that cannot be over looked for high precision time synchronization. We briefly review these sources of error below and in Figure 4.2. 52 1. Send Time: The delay in the packet traversal from the message assembly at the application layer all the way down to MAC layer. Highly non-deterministic. 2. Access Time: Is the channel contention time, that in dense broadcast medium such as ours can be in the order of hundreds of milliseconds. Least deterministic part of the message delivery. 3. Interrupt Handling Time: The delay between the radio chip raising and the microcontroller responding to an interrupt. Can be an issue if interrupts are dis- abled on the microcontroller. 4. Transmission and Reception Time:The delay in sending or receiving the entire length of the packet over the channel. Largely deterministic, is a function of bandwidth and packet size. 5. Propagation Time: The delay, for a particular symbol of the message, in travers- ing all the way to the receiver. The propagation time can be deterministic if the speed of propagation is assumed constant, and endpoint location is known, or if synchronization exchange is performed with assumption of path symmetry. This delay can be significant in the underwater acoustic channel since assuming clocks will not skew over packet exchanges would be incorrect. 6. Encoding and Decoding Time: The time taken by the radio chip to encode/decode and transform a part of the message to/from electromagnetic waves. This time is deterministic and is in the order of hundred microsec- onds [MKSL04]. 53 7. Byte Alignment Time: The delay because of the different byte alignment at the receiver. This time is deterministic and can be computed on the receiver side from the bit offset and the speed of the radio. 8. Receive Time: Time for the incoming message to traverse up till the receiver application. Highly variable and varies for each (stack,OS) pair. Existing time synchronization schemes (reviewed in the next section) focus on elim- inating or accounting for these sources of error. Schemes typically differ due to differing assumptions in which sources of variation are dominant in different domains, and due to different approaches to eliminate the sources of error. 4.2 Quantifying the Challenges of High Latency Links Ideally we could simply use an existing time synchronization protocol in underwater acoustic networks. We next develop a simple analytic model to demonstrate why exist- ing protocols such as RBS, TPSN, and FTSP do not work well for high-latency links. In this analysis we focus only on propagation latency as a source of error in time syn- chronization. Clearly this assumption is not correct in general, however, since existing schemes address other sources of error and are already accounted for by most existing protocols. First we define a simple notation for realistic (inaccurate) clocks. For a nodeS, we model its uncorrected clock asf s (t) and its corrected time as b f S (t): f S (t)=a S t+b S b f S (t)=f S (t)+¯ S (t) (4.1) 54 These are first order linear function of its skewa S and offsetb S , wheret is the global reference time, and¯(t) is the correction factor calculated att. Protocols based on one-way exchanges: We first quantify the proportion of error in both RBS and FTSP to the propagation delay by taking a simple case where a perfectly synchronized beacon node f B (t) = c f B (t) = t sends synchronizing pulses with its sending time to an uncorrected nodeR. Note that this is very similar to what happens in FTSP, and with slight modification can be applied to RBS as well. With no propagation delay, the correction factor calculated by the synchronization protocol is: ¯ R (t)=f B (t)¡f R (t)=(1¡a R )t¡b R which will synchronize nodeR’s clock to the correct time, using Equation 4.1. The clock skew can be computed with multiple exchanges by observing how¯ R (t) changes. However, given a propagation delay ofd between beacon and nodeS, the computed correction factor is: ¯(t+d)=f B (t)¡f S (t+d)=(1¡a S )t¡(a S d+b S ) Based on Equation 4.1 alone, the corrected time is offset incorrectly by d. To cor- rect for this error we need to estimate propagation delay. To stay within the RBS or FTSP models we must estimate this delay without sending additional messages. A prior computation of underwater propagation speed is quite difficult. Node locations may be known, but are likely inexact due to placement error or node movement. Underwa- ter node localization is an open problem, and often requires synchronized clocks for computation. Finally, knowledge of acoustic propagation speed assumes that the nodes 55 have capabilities to measure, at the very least, the temperature and pressure. For these reasons, a priori computation of delay seems quite challenging. Protocols based on two-way exchanges: We next quantify the effect of the prop- agation delay on protocols that allow two-way exchanges, such as NTP and TPSN. The key element of these protocols are that they can factor out propagation delay via a two-way message exchange. However there is one assumption they make about sender’s clock not skewing during the message exchange. Consider the sender S and the receiver B (perfect clock) exchanging timestamps as shown in Figure 2.1(a), with T 1 = f S (t 1 );T 2 = f B (t 1 +d);T 3 = f S (t 3 ); and T 4 = f S (t 3 +d). With that scenario, the clock model for the sender S will be: f S (t)=t+b S ¯ S (t)= (T 2 ¡T 1 )¡(T 4 ¡T 3 ) 2 b f S (t)=f S (t)+¯ S (t) (4.2) With no skew, and with our assumption of none of the sources of error are present, this exchange will perfectly synchronize nodes. We will now show that if we consider skew it does have an effect on the error as either the skew or the duration for the message exchanges is increased (intuitively grasped by looking at the message exchange shown in Figure 4.1). f S (t)=at+b S ¯(t 3 +d)= (1¡a S )(t 1 +t 3 +d)¡2b S 2 (4.3) The delay compensating clock at S as: 56 b f S (t 3 +d)=f S (t 3 +d)+¯(t 3 +d) Using Equation 4.4 below shows that this leads to an Error =(t 3 +d)¡ b f S (t 3 +d), that is proportional to the skew and the packet exchange period. Error = (1¡a S )((t 3 ¡t 1 )+d) 2 (4.4) For Berkeley motes, the upper bound given in the datasheet [Atm] is 40ppm i.e. a clock in mote can loose up to 40¹s in a second. Thus a node can drift around 15-20¹s with nodes»400m apart. This drift significantly affects the accuracy of time synchro- nization, as predicted by Equation 4.4, and our simulation will show the degradation of TPSN precision with increasing distance, or skew, between nodes. 4.3 Design of Time Synchronization for High Latency Channels (TSHL) While prior time synchronization protocols addressed many sources of error in estima- tion, only NTP faced large propagation delay, and it is not appropriate for sensor net- works (see Section 2.2). We now present an alternate time synchronization algorithm for sensor networks that can manage high propagation delays while remaining energy efficient. 4.3.1 Overview and Assumptions TSHL is a two phase protocol. The the core idea is to first model the skew of a node’s clock so that each node is skew synchronized. We compute skew by performing linear 57 regression over multiple beacon values. After skew synchronization nodes may still operate with different offsets, but because all (one hop) nodes share a frequency standard they are now able to maintain an accurate relative timer for any additional events. In the second phase we correct for clock offsets. Prior protocols such as NTP and TPSN did this with a two-way message exchange; we take this approach as well, but TSHL considers a skew-compensated two-way exchange. When both phases have been completed we have a model mapping the local, inaccu- rate clock to the reference timebase. We can then compute a global time for all events, even those before our synchronization, with post-facto correction if necessary, similar to RBS. In phase one, skew is estimated without any knowledge of propagation delay. The quality of our estimate of skew is dependent on the consistency, not the duration, of propagation delay. In our current approach we assume propagation delay is constant over the message exchange. Underwater, changes in temperature and pressure affect the speed of sound and so will change propagation delay over long durations. However, we expect a 20-beacon exchange to take a few seconds; over this period our assumption seems reasonable. Verifying this assumption is an area of future work, however if it does not hold we can fall back on a statistical model of skew as done by Fober et al. [FOL02]. Beacon Nodes can either be specialized nodes with accurate clocks, perhaps con- nected to an external time reference like a tethered GPS receiver on a buoy. Alterna- tively, Beacons Nodes could be elected or externally designated. As noted in Section 4.1.1 and first addressed TPSN [GKS03], a great deal of non- determinism in message exchange can be removed by placing message timestamping in the MAC layer. For highest accuracy we believe low-level timestamping is essential. We expect our acoustic modems to provide this bit- or byte-level radio access, making 58 Beacon Node B Node R f' R (t) represents skew-corrected clock, calculated in Phase 1. Note that events are ordered, in time, based on their index. Sync Request (1) T 1 ‘= f’ R (T 1 ) (2) T 2 = f B (T 1 +D B->R ) = T 1 +D B->R Sync Reply (3) T 3 = f BFS (T 2 ) (4) T 4 = f’ R (T 3 +D B->R ) Phase 2 Beacon Node B Node S D B->R D B->S (3) T R = f R (t B,i + D B->R ) (1) t B,i (2) T S = f S (t B,i + D B->S ) Node R f R (t),f S (t) represent the individual, unsynchronized clocks of nodes R & S. Phase 1 Beacon Messages Figure 4.3: Phases in TSHL MAC-level timestamping easy to implement. Our current prototype implementation includes this feature (as discussed in Section 4.5). A second assumption of TSHL is that clocks are short-term stable. Clock frequency and hence skew must remain constant over a short period of time (typically 5-10 min- utes) . Short term instability occurs mainly due to environmental factors such as sudden variation in temperature, supply voltage or shock [Vig92]. This assumption allows us to model the clock skew using linear regression and use it for predicting the future time accurately as well. One typically accommodates long-term instability by periodic resyn- chronization (as described in RBS [EGE02]); optimization of that interval is not part of our current research. 4.3.2 Details Figure 4.3 shows the message exchange in TSHL’s two phases. In Phase 1 of the pro- tocol, each node in the broadcast range of a Beacon Node models its clock skew. We define the Beacon Nodes clock as the reference timebase (i.e., f B (t) = c f B (t) = t). The Beacon Node then sends out enough Beacon Messages for skew estimation. Our current simulations and implementation send 25 of these beacon messages, needed for reasonable linear regression. In the future we can make this value adaptive based on 59 feedback from receivers; each receiver can estimate error and enter Phase 2 when skew error reaches a target threshold. Each Beacon Message BM i contains the transmit timestamp t B;i noted just before the message leaves the Beacon Node. Each receiver R gets this message at absolute time t B;i + D B!R where D B!R represents the unknown propagation delay between the Beacon Node and the receiver. It then assigns it the local time f R (t B;i +D B!R ). This local time includes error due to clock skew and offset in addition to propagation delay. However, as with Fober et al. [FOL02], we can still model the drift of the local clock with respect to the Beacon’s reference clock by doing a linear regression on the difference between receive timestamp and the timestamp in the message. This difference changes by the same amount as clock skew as they drift apart (provided our assumption that path delay is constant over our estimation interval). Thus, forN messagesM i , we do linear regression over the following data points: (t B;i ¡f R (t B;i +D B!R );f R (t B;i +D B!R ) (4.5) This computation gives us skew correcting conversion of the local time, so from now each node is skew synchronized with its skew corrected local time represented asf 0 R (t) (as opposed to a fully synchronized clock c f R (t)). This skew correction allows us, in phase two, to do correct offset estimation and ultimately map prior or future events to the reference timebase. Phase 2 is similar to the classical two-way synchronization exchange as shown in TPSN and Figure 2.1(a). TSHL’s two-way exchange differs in that we correct for skew when computing the clock offset. When the receiver obtains enough beacons to estimate the skew it sends a Synchronization Request message with T1 = f 0 R (T1), the skew-corrected local timestamp. The Beacon Node records its local version of 60 this T2 = f B (T1 + D R!B ) and returns this value to the receiver in a Synchroniza- tion Reply message timestamped at T3 who computes a skew-corrected receive time T4=f 0 R (T3+B!R). Finally the receiver can compute its clock offset: offset R =[(f B (f 0 R (T1)+D R!B )¡f 0 R (T1)) ¡(f 0 R (T3+B!R)¡T3)]=2 (4.6) When this exchange completes, the nodes are able to factor out the error that occurs because of skew and get the exact offset. Our algorithm builds on prior synchronization algorithms, drawing the concept of skew modeling from RBS, and of MAC-layer time-stamping (to reduce jitter) from TPSN. To this prior work we add skew compensation used during the synchroniza- tion exchange. This addition is essential for high latency environment such as ours, as is demonstrated in the simulation results shown in the next section. 4.3.3 Time Synchronization Over Multiple Hops While these equations and the protocol are specific to synchronization between two hosts, we can generalize this result to multi-hop time synchronization. We briefly con- sider the effects of multiple hops on the protocol and accuracy. Some optimizations are possible to the protocol when one considers multiple nodes synchronization with a single reference time. All nodes can share the same series of Phase 1 packets and independently estimate their skews. Since Phase 2 requires a two-way exchange, a straightforward adaptation of TSHL would employ as many such exchanges as each node has neighbors. Potentially one could optimize Phase 2 by bundling the exchange for several neighbors in a single reply. Such an optimization 61 would require additional synchronization and increase delay, but since we model skew during the Phase 2 exchange, such delay should not greatly affect accuracy. Time synchronization over multiple hops can be done either to a single, com- mon time base, as is typical, or with multiple independent timebases as done in RBP [EGE02]. Accuracy of the protocol over multiple hops should be similar to the accuracy of other multi-hop protocols for time synchronization; additional hops will degrade accuracy. RBS observed that per-hop jitter follows a Gaussian distribution, so multi-hop accuracy there degrades as the square-root of the number of hops [EGE02]. An open area is to verify that our acoustic modems provide similar per-hop jitter. 4.4 Performance Evaluation of TSHL In this section we present some preliminary results evaluating our time synchronization algorithm. We also describe the goals of our simulations, the methodology, comparison against existing time synchronization protocols for sensor networks, and its dynamics under different parameters. 4.4.1 Goals and Methodology Our goals in conducting this evaluation of TSHL were two-fold: ² Compare performance of TSHL with other established sensor network time syn- chronization schemes such as TPSN and RBS. ² Understand the dynamics of TSHL under varying parameters such as node clock skew, granularity, and prediction error. Our early simulation results compared TSHL against TPSN, RBS, and FTSP. We omit RBS and FTSP results because those protocols do not consider propagation delay 62 at all and so exhibit error proportional to propagation distance. This error is significant, even at short distances. RBS, for example, shows 6ms error at distances of 10m, and the error grows to 100ms and more at larger distances. Overall, our simulations show that TSHL maintains low synchronization error, irrespective of the path delays, and its accuracy is directly proportional to the receive jitter (due to bit encoding/decoding or interrupt handling) or the granularity (smallest time increment possible) of the clocks used on the nodes themselves. 4.4.2 Simulation Setup We simulated the protocols in a custom event driven, packet level simulator 1 designed for an acoustic underwater environment with high latency. Each node’s clock was sim- ulated as having some skew and offset relative to the global simulation time. We place a single Beacon Node, with no skew and zero offset in all simulations; all receiving nodes are within a 500m radial distance. We model encoding/decoding and interrupt handling errors referred in [MKSL04] by introducing a Gaussian receive jitter. Gaussian distri- butions have been found to provide reasonable approximations of this error, both by Elson and Estrin [EGE02] above the MAC layer, and more recently by the authors of TPSN [GKS03] even with MAC-layer timestamping. The simulator allowed us to alter the following parameters in the simulations: ² Number of Beacon Messages broadcast by Beacon Nodes in phase 1 of the proto- col. ² Individual clock skew and offset. ² Granularity of the clocks. 1 This simulator is available from the authors athttp://www.isi.edu/ilense/software/ 63 ² Receive Jitter distribution. Although we assume the environment was stable during the protocol design, as a worst case we vary water temperature uniformly between 25 and 35 o C for each packet sent by the Beacon Node. This temperature variation, with randomized receive jitter, accounts for any inaccuracy observed in the TSHL protocol. Comparison with TPSN was made by implementing a “TPSN-like” (TPSN-L) protocol in the simulator. By “TPSN-like” we mean that the protocol captured the essence of TPSN through MAC layer timestamping and offset correction scheme as described in their paper [GKS03] and implemented in their nesC code. In each of the simulation test performed the following parameters were used and are assumed to hold unless specifically mentioned otherwise: ² Skew = 40 ppm. ² Initial Offset = 10¹s. ² Number of Beacon Messages = 25. ² Granularity = 1¹s. ² Mean Receive Jitter = 15¹s. Each data point shown in a graph is the mean of 1000 simulation runs. Error bars show standard deviations. 4.4.3 Comparative Evaluation We now compare TSHL performance with TPSN-L, demonstrating that TPSN-L’s accu- racy deteriorates at high latencies because it did not need to model skew in its environ- ment. In the first experiment we evaluate error as a function of the distance between the 64 0 50 100 150 200 250 300 350 400 450 500 0 5 10 15 20 25 Distance (meters) Error (usec) TPSN TSHL Figure 4.4: Comparison of Instantaneous Error: Distance Variation receiver and the Beacon node. To confirm that TPSN-L error is proportional to clock skew, our second second experiment examines offset error at a fixed distance as clock skew varies. Offset Error: Effect of Distance Variation In this test we measure the absolute offset error as a function of distance. (Offset error is the difference between the global time and the corrected local time of the node.) First we measure it at the instant immediately proceeding the final synchronization exchange (TSHL phase 2). We expect that the variation in distance (and thus the effect of clock skew) should cause deterioration of error for TPSN-L, but not for TSHL, as it specifically accounts for delay. Figure 4.4 confirms that the error increases with receiver distance for TPSN-L, from about the same accuracy as TSHL at distances less than 100m, to about double the 65 0 50 100 150 200 250 300 350 400 450 500 −5 0 5 10 15 20 25 Distance (meters) Error (usec) TimeSync Error − Variation with Distance for TSHL Gaussian Receive Error Mean =5 usec Gaussian Receive Error Mean =15 usec Gaussian Receive Error Mean =30 usec Figure 4.5: Comparison of Error:Receive Jitter error at 500m. Error does increase for TSHL as well, but by a much smaller amount; about 12% over 500m. We also observed that the synchronization error of TSHL varies linearly as the mean of the receive jitter distribution (Figure 4.5). This simulation presents error immediately after the synchronization ends, thus these results represent the best case performance for TPSN-like protocols that do not model the clock and consider skew. We consider that case next. Offset Error: Effect of Time since Synchronization Here we compare the performance of each of the scheme in predicting the time after a particular delay from the time the final synchronization exchange occurred. For this section we consider a receiver placed 400m from the Beacon Node. Since TPSN-L does not model the skew and simply measures the offset at that time, we expect it to show a larger error as the time progresses (and the clock skew causes larger drift from the prior synchronized value). TSHL, on the other hand models the 66 10 0 10 1 10 2 10 3 10 4 10 5 10 6 10 7 0 50 100 150 200 250 Time after sync (usec) Offset Error (usec) TSHL TPSN Figure 4.6: Comparison of Error: Time since Sync clock drift and offset, and hence would be able to correct the local clock more accurately in the future. Figure 4.6 confirms our expectations. Error in both protocols is linear with time, but the slope of TSHL is much less because it models skew. Even after 5 seconds, with TSHL, error in offset is below 50¹s. Note that RBS and FTSP do model skew and would do well here. We do not show them because at 400m, their error due to propagation overwhelms the benefits of mod- eling skew. (Note that these simulations assume a constant, random clock skew. Over these timescales, variation in clock skew is unlikely.) Effect of Variation of Skew In this experiment we vary the node skew with respect to the global (and the Beacon Node) clock, and observe the effect on the accuracy of the synchronization. The nodes 67 are kept at a distance of 400m from the beacon node. We vary skew from 5 to 100ppm; this range is wider than typical for real clocks by a factor of about two. Since TSHL models the skew in phase one, its should be able to cater for what- ever the skew is. (Its error should be dominated by mostly by the non-determinism in the delay path i.e. receive jitter, not skew.) TPSN-L on the other hand, should show increased error as the skew (and thus the error over a single exchange) increases. Figure 4.7 validates this expectation as over a wide range of clock skew Synchro- nization error remains nearly constant for TSHL. Again, this difference is because of TPSN-L’s not modeling skew. 4.4.4 TSHL Parameters Having established the robustness of TSHL in a high latency acoustic channel, we ana- lyze the dynamics of the protocol by altering key parameters of the algorithm. Note that we have not shown results of varying the initial offset, since all protocols are able to factor this out consistently. Number of Beacon Messages In this simulation we analyze the effect of changing the number of Beacon Messages transmitted in phase 1. Since we do linear regression over these points to estimate clock skew we expect additional messages to result in higher accuracy. In this simulation we set the receive jitter to 30¹s and consider a receiver 400m from the Beacon node. We show two separate results: One where the error is calculated 0.5s after the final synchronization exchange, and in the other 5s after the exchange. Figure 4.8 shows the results of these simulations. First, we observe that there is diminishing benefit in increasing the number of beacons after a certain optimal value. 68 0 10 20 30 40 50 60 70 80 90 100 110 0 5 10 15 20 25 30 35 40 skew (ppm) Error (usec) TSHL TPSN Figure 4.7: Effect of clock skew This value seems to be at around 25 beacons. These results corroborate very similar results for the group dispersion as a function of beacons in RBS [EGE02]. A second interesting interpretation of this simulation shows the effect of control packet loss. As the number of beacons received decreases (in this interpretation, due to packet loss), we can observe the resulting deterioration in the synchronization accuracy. The loss of two way synchronization messages can be easily rectified by retransmission at the sender, if it does not get the sync reply. Variation of Clock Granularity In this section we observe the effect clock granularity has on the accuracy of the pro- tocol. Since the clock granularity puts a fundamental limit on how accurate a clock can be, we can expect that the performance will degrade as clock granularity increases. 69 0 5 10 15 20 25 30 35 40 45 50 0 200 400 600 800 1000 1200 Number of Beacons Offset Error (usec) After 5 sec After 0.5 sec Figure 4.8: Effect of Changing the Number of Sync Beacons While clock granularity is limited by hardware constraints, software controlled inter- rupts allow one to select coarser clock granularities to reduce interrupt rate and energy consumption, thus there may be a desire to vary this parameter. For this test we placed the receiver at a distance of 100m from the Beacon Node node. We examine granularity at uniform increments of 5¹s. Figure 4.9 shows that both the mean error and the standard deviation of the error increases as the clock granularity degrades. However this is not a bad reflection on the protocol itself; it just goes to show the fundamental limits imposed by the granularity of a node clock on the level of synchrony that can be achieved. 70 0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35 Precision (usec) Error (usec) Figure 4.9: Varying the Clock Granularity 4.5 Experimental Evaluation of TSHL Analysis and simulation help demonstrate the importance of considering propagation delay in acoustic communication, but these approaches necessarily simplify the detail of the real world. The main challenge in testing TSHL with high-latency acoustic communication is that our underwater acoustic modems are currently under development. A number of important design issues must be solved before short-range acoustic modems are avail- able. We considered but rejected the alternative of substituting off-the-shelf, long-range acoustic modems. Not only do they have very different characteristics than our prelimi- nary short-range design, but prior time synchronization protocols have demonstrated the importance of integrating timestamping with the MAC layer [GKS03]. Such integration is impossible with easily available packages. 71 Instead, we substituted in-the-air acoustic communication for underwater communi- cation. Ultrasound and audible sound have been widely used for localization in surface sensor networks [PCB00, SHS01, GBEE02]; we adopted the Cricket platform due to its commercial availability and good support for low-level hardware access. In-the-air acoustics changes several things. First and most importantly, the ratio of the speed of sound in air to water is about 1:5 (»300m/s vs.»1500m/s). This scaling factor plays to our advantage in that shorter distances appropriate for in- office testing scale to much longer equivalent underwater distances. Second, in-the-air acoustics capture some of the noise and random variation of the real world. Although we make no claims that different air temperatures and multipath effects of an office accu- rately model underwater currents and propagation, it does capture some unpredictable variation. Our initial choice of the Cricket platform was promising since it had low level sup- port for acoustic transmission and detection. However, the accuracy provided by in its detection hardware proved to be insufficient for the microsecond-level accuracy that we wanted in our testbed. We next discuss our experience with the in-air experiments on the cricket platform. 4.5.1 Experiences with TSHL on the Cricket Platform As mentioned earlier in Section 4.5, we want to emulate underwater acoustic transmis- sion with in-the-air ultrasound transmission. We initially chose the Cricket platform due to its commercial availability and good support for low-level hardware access. This section summarizes our experiences with this platform for time synchronization. 72 Sending data via hybrid RF and Ultrasound Our choice of the Cricket platform forced us to approximate sending data over ultra- sound. The Cricket hardware consists of a Mica-2-like mote core with Chipcon radio, and it adds a 40kHz ultrasound-based transmitter and microphone. While this platform supports sending and receiving acoustic pulses, it currently does not support data trans- mission. We therefore approximate the acoustic channel by sending an acoustic pulse and a coupled data packet over the standard Chipcon radio. Although the data packet arrives quickly, we delay reception of it until the acoustic pulse arrives, experimentally accounting for propagation latency, and we add a computed transmission delay, account- ing for bandwidth limitations. While not ideal, we were satisfied with this approxima- tion. Jitter in hybrid (RF-Ultrasound) time stamping. After implementing TSHL in TinyOS and building a hybrid RF-Ultrasound emulation module (using the ultrasound control provided by the MIT Cricket group), we observed significant inaccuracy in synchronization (in 100’s of¹s). This necessitated a review of the hardware capabilities and any jitter that it might accrue. We identify several sources of uncertainty in timestamping due to the hybrid emulation module, as illustrated in Figure 4.10. All of these sources of uncertainty are below the MAC layer, and are endemic to the communication hardware employed. To verify these sources of jitter we test timing at the hardware-level. We attach an oscilloscope to the input of ultrasound transmitter (output of pin C54) and the amplified received signal (input to pin D7) for actual signal timing. For software timing we toggle a GPIO (AD0) pin. These experiments allow us to measure the jitter/variance at each of these delays. The sources of delay and our measurement results are as follows: 73 T ran sm itter T ransmission Delay T ime S tam p Radio P kt Ultrasound Pulse Receiver Received Signal Receiver T ime S tam p R eception notification Propagation/D etection T ime Application T ransport Network S end T im e M AC PH Y A ccess Tim e Interrupt, Encoding Tim e A pplication Transport Netw ork R eceive Tim e M AC PH Y Interrupt, D ecoding, & Byte alignment T im e Propagation Tim e Tim e Sync M essage Sen der Tim e S ync M essage Re ceiver S ync packet Transmission Time ReceptionTime Figure 4.10: A Blow up of the sources of uncertainty we identified specifically for the Cricket platform used in our hybrid RF-Ultrasound testbed. Note that these are below MAC level, above which we have factored out by timestamping after MAC access. Ultrasound Transmission Delay: Since we time stamp an outgoing Radio packet and then initiate a software procedure to pulse the ultrasound transmitter, the accuracy of time stamp depends on the delay till when the pulse actually starts going out. This can vary due to code size, interrupts and the resonance delay of transmit circuitry. Our measurements showed that this was constant (less than6¹s) and this value can always be determined, if needed. Reception Notification: This delay occurs between the start of the received pulse till interrupt handler for pulse detection gets called (where we timestamp the reception time). The delay could vary due to existing interrupt handlers or interrupts between the execution of our code. Typically the interrupt handler was called at regular periods. measurements, how- ever, showed that this infrequently (1 in 50) did change by 20–50¹s. Although this did not account for the variance observed in our experiments, we decided to remove it as a possible source of error by using the value of a hardware counter,T C , at the input cap- ture event for that counter (ICP pin is wired to the ultrasound detection circuitry). The 74 Table 4.1: Transmit to detection delay jitter for fixed node positions. Detection Time delay (ms) Jitter from mean (ms) 2.246 -0.0674 2.364 0.0506 2.280 -0.0334 2.318 0.0046 2.359 0.0456 Mean = 2.3134 ms, Range = 0.113 ms counter was started when the correlated radio packet’s start symbol was received, and at that instant we also took a local timestampT L . We were able to circumvent any interrupt related inaccuracies by using the receive time stamp asT L +T C . This is similar to the mechanism for measuring time-of-flight used in the localization of Crickets at MIT. Detection Time: The above results convinced us that most of the uncertainty exists in the Detection time (see Fig.4.10). This jitter is determined by the point in time when, relative to the start of the transmit pulse, the start of the receive pulse is detected. We measured the jitter by creating a positive edge on a GPIO pin at the transmitter when it sends a command to transmit ultrasound pulse, and used this as an edge trigger. At the same time we hooked the second line to the ICP1 pin (the pin that triggers an input capture on rising edge), which is the absolute best timing of ultrasound pulse detection that the Cricket system can provide. The resulting range of detection times, for the same node locations, with a jitter range of about 130¹s, is shown in Table 4.1. In similar experiments that we conducted, we observe that the detection instant also shows correlation to environmental change. Furthermore, multipath effects were evident even at short ranges with this simple acoustic hardware. Another interesting observation is that at larger inter-node distances, the direct path is usually weaker than the first reflection. One possible reason for this 75 “bad” performance might be that the transducers have very narrow bandwidth (1KHz at 3dB) and thus large (about 1 ms) response time; unless we send pulses for longer dura- tion, the pulse will never reach maximum amplitude and we will have a hard time detect- ing the exact start of the pulse. An open issue is to verify that this bandwidth limitation is actually the core problem at hand and address that in our ongoing hardware imple- mentation. The above arguments also suggest that low-level jitter compensation would be beneficial; we could use mechanisms such as described by the FTSP [MKSL04]. We accept that this is not an exhaustive study of the Cricket ultrasounders. Since they were not designed for data transmission, the problems we encountered are not surprising. Our goal here is to understand the limitations of the hardware for time syn- chronization, shed light on our experiments, and to identify important characteristics for future hardware designs. Short range of ultrasounders We also wanted to vary the internode distance to observe the effect of latency. Our experimental results show benefit at distances larger than 100m, with substantial (100%) improvement at around 500m underwater. In the main text we argued (see 4.5) that slower speed of sound in air allowed us to relax this range requirement by a factor of five. However this still amounts to at least a range of around 20–100m to show substantial benefits. The original ultrasound module, with no tweaking, provides a range of around 10-15m. The Cricket architecture provided two ways of extending the range. One way was through increasing the receiver sensitivity; the other by increasing the number of pulse sent per ultrasound transmission. Both have the intuitive drawback that they will result in less accurate time stamping. Increasing sensitivity results in becoming less tolerant 76 of noise; increasing the number of pulses makes it more difficult to correlate the start of the pulse with when a detection event gets triggered (this effect will possibly be more pronounced at larger distances or with higher multipath). We have been able to extend the detection range of Cricket platform to around 100m in the air (by sending 40 pulses at 40kHz), but the resulting increase in detection jitter belies its utility. 4.5.2 Experimental Result Summary Our simulations showed that TSHL will show a performance improvement in the sub 100¹s and>100m regime. This section showed that the cricket platform neither allow us to time-stamp at the needed accuracy not does it have the desired range to show benefit of TSHL. We now have our own acoustic modem using which we plan to perform in- the-air experiments to validate TSHL as part of our short-term future work(Chapter 7.1). 4.6 Summary In this chapter we have presented Time Synchronization for High Latency, a time syn- chronization protocol for high point-to-point latency environment. As short-range, underwater sensor networks are developed we expect protocols that consider high- latency communication to be increasingly important. TSHL represents a first such pro- tocol in this domain, and potentially has application even to other high-latency domains such as interplanetary networks. Through analysis and simulation we explored TSHL for a wide range of charac- teristics. Prior protocols like RBS and FTSP exploit the rapid propagation of RF; in an high-latency acoustic network they perform quite poorly (about 6ms error at 10m, 77 growing to hundreds of milliseconds at our target maximum range of 500m). By con- sidering propagation latency, TPSN does much better. However, we demonstrated that it is critical to consider skew as well, even during the synchronization exchange. At short ranges TPSN and TSHL are equivalent in terms of accuracy, but at long ranges TSHL shows up to two times better accuracy (compared to TPSN-L). Our work on TSHL has demonstrated unequivocally that being propagation-latency- aware is essential for the migration of protocols from terrestrial to underwater sensor networks. 78 Chapter 5 T-Lohi MAC: Exploiting Spatial Uncertainty Water is the only drink for a wise man. — Henry David Thoreau 5.1 Introduction Networks with shared media require access protocols (MACs) to control access of the shared channel. In underwater sensornets (UWSN), a shared acoustic medium raises challenges absent from traditional RF wireless [HLS + 06, PKL06]. Acoustic communi- cation magnifies wireless bandwidth limitations, transmit energy costs, and variations in channel propagation. Control algorithms of MAC protocols are significantly changed by acoustic propagation latencies that are five orders of magnitude greater than radio. The focus of this chapter is to design an energy and throughput efficient MAC pro- tocol for dense and short range acoustic sensor networks. While a vast majority of exist- ing underwater networking applications demand sparse and long range deployments, our earlier work motivates a host of promising sensornet-like applications [HLS + 06]. Recently several innovative acoustic modems have been proposed [FGS + 05, WYH06], 79 but MAC protocols have not yet been proposed that exploit their unique low-power capa- bilities. We will show that the challenges of high latency also enable new MAC tech- niques and solutions that provide good throughput across varying application require- ments (Section 5.2.2). Many underwater applications will require long-term deployment, making energy- efficient design an important goal. These include static applications such as 4-D seis- mic sensing of oilfields [HLS + 06], gliders and low-energy mobile platforms, or plat- forms with parasitic mobility, such as tagging of aquatic-life [TOP06]. Compared to radio communications, underwater acoustic networking presents different design trade- offs, since transmit energy costs are higher [PKL06], idle times are longer, and battery replacement is harder. We propose a new class of MAC protocols called Tone Lohi (“Lohi” means slow in Hawaiian). Besides being energy and throughput efficient, Tone Lohi (T-Lohi) provides flexible, fair and stable medium access for acoustic networks. T-Lohi is designed for general underwater applications (not a specific application, like very low duty-cycle operation [RP05]), and conserves energy through a tone-based contention algorithm with low-power wake-up hardware. This chapter provides three novel contributions. First, we exploit the space-time uncertainty effect (Section 3 and [SYKH07]) to provide contender counting, and show how contender counting can improve fairness and provide throughput stability under high load (Section 5.2.2). Second, we present T-Lohi, a new class of MAC protocols for underwater acoustic networks that utilizes contender counting and low-power wake-up capability of acoustic modems (Section 5.3). Finally, we validate the design decisions behind T-Lohi flavors (Section 5.6.1) and compare T-Lohi to several canonical medium access mechanisms (Section 5.6.3). 80 To understand T-Lohi performance we perform extensive simulations (Section 5.4). Our channel model is quite simple; ignoring any channel related packet loss and mul- tipath to focus on protocol performance. Our results are a good preliminary indication of the viability of our MAC and considering channel effects is a very promising future research area. We discuss these assumption and their impact on our MAC in Section 5.7. Our simulations show that T-Lohi is energy efficient within 3–9% of optimal, and can achieve utilization within 30% of the theoretical optimal channel capacity. We also show that our protocols are stable and fair under both low and very high offered loads (Section 5.4.2). We deepen our previous evaluations by considering different lengths of the contention duration in T-Lohi and confirming our selected values (Section 5.6.1). Moreover, we evaluate design choices with T-Lohi flavors (Section 5.6.2), and compare T-Lohi with representative MAC protocols, such as TDMA, CSMA and ALOHA. Such comparison improves our understanding of when and where T-Lohi is the best choice for medium access (Section 5.6.3). Overall, our results are promising and suggest further evaluation in multi-hop conditions and field tests as promising future directions. 5.2 Challenges and Opportunities As we have previously argued, among several challenges inherent to underwater acous- tic communications [HLS + 06, PKL06], propagation latency (5 orders of magnitude larger than radio) has the greatest impact on networking protocols. Acoustic modems also have very different energy consumption patterns compared to radios, with transmis- sion often 100 times more expensive than reception [PKL06]. For example the typical receive:transmit power ratio of WHOI micro-modem is 1:125 [FGS + 05], while short- range radios for sensornets generally provide ratios around 1:1.5 [KIM07]. Several acoustic modems have a low-power idle mode that draws considerably lower power 81 than either receive or transmit mode [WYH06, Ben, FGS + 05]. As we will show, our design exploits this capability for energy-efficient medium access. These unique characteristics create behavior in acoustic networks that are not seen in radios. We next describe these characteristics and how we can exploit them to improve MAC design. 5.2.1 The SNUSE Modem The simulation results in this paper and our MAC design assumes the specifications of our SNUSE modem [WYH06], transmitting at 1 kbaud in the 17–19 kHz frequency range with FSK encoding. This modem has an expected range of 50–500m with a sub-mW wake-up receiver (power specifications provided in Table 5.1). We have done in-water tests of data transmission with this modem, and our implementation of T-Lohi over this hardware is in progress. The SNUSE modem has been designed to provide a low-power wake-up circuit. The principal goals for the wakeup receiver are good sensitivity and very low power consumption. The only purpose of the receiver is to monitor the total energy level present in a narrow band of frequencies around 18 khz, and to produce an interrupt signal when the energy exceeds some preset threshold level. The optimum threshold level depends on the bandwidth of the receiver, and the operational system noise floor. The total power consumption of this wake-up receiver is 0.5mW. Thus nodes can be woken up from a sleeping state by sending tones (18 khz acoustic signal). Although designed for this modem, we expect T-Lohi to operate with any modem providing wake- up capabilities. 82 Figure 5.1: Spatial Unfairness: (a) Transmitter and close neighbors have channel cleared earlier. (b) In slotted access, close neighbor A can attempt in slot 3 while C and D can not. 5.2.2 Spatial Uncertainty In chapter 3 we explicitly introduced the concept of space-time uncertainty. We now further explore how this concept can be exploited its impact on fairness while accessing the acoustic channel. Finally, we show how this challenge of additional spatial uncer- tainty due to propagation latency can be exploited for the purpose of contender counting and detection. Spatial Unfairness A significant impact of space-time uncertainty is an inherent bias for medium access that depends on location. We call this bias spatial unfairness; it is conceptually similar to channel capture [BDSZ94], but caused by physical location and propagation latency rather than backoff estimates. Since a packet’s arrival time is proportional to distance from transmitter, the channel becomes clear earlier at nodes closer to the transmitter. In Figure 5.1(a) transmitter A and its close neighbor B have a greater chance to recapture the channel after sending than nodes C and D that are far away. With slotted media 83 (a) Collision uncertainty (b) Contender detection and counting Figure 5.2: Spatio-temporal uncertainty in acoustic medium access. access spatial unfairness becomes more pronounced. In Figure 5.1(b), B’s data ends in slot 2 for nodes A and B, but ends in slot 3 for C and D. Thus, even if the transmitter is prevented from immediately reacquiring the channel, nodes A and B can swap the chan- nel back and forth. We handle spatial unfairness in our protocol design by employing a distributed backoff mechanism (Section 5.3.2). Exploiting Space-Time Uncertainty Although latency increases uncertainty, we next show that it can also be exploited for contender detection (CTD) and contender counting (CTC). Nodes in our protocol detect contenders by listening to the channel after sending short reservation tones that are analogous to RTS messages. Unlike low-latency wireless protocols, large propagation delays allows observation of tones sent concurrently because they may arrive after their own transmissions complete. Contender detection depends on relatively short tones and a long listen period. Nodes can further count the number of contenders, if tones are short enough (we formalize shortness in Section 5.2.3), since tones from different transmitters arrive at different times due to varying propagation latencies. An example is shown in Figure 5.2(b), where nodes A and E send short tones. At nodes where the tones do 84 A B (b) Non Overlap Zone Overlap Zone A B (a) A B (c) A’s tone B’s tone A’s tone B’s tone A’s tone B’s tone t tx,A t tx,B t rx,A? B Figure 5.3: The three cases where deafness can occur. (a) Bidirectional deafness. Uni- directional deafness at B with A’s tone reaching B (b) before B starts transmitting, (c) after B starts transmitting. not collide, such as nodes A, E, B, and D, they can count the number of contenders. Even if the tones collide on some nodes, e.g., node C, they can still detect the presence of contention. The capability of contender counting (CTC) is not generally possible for RF-based networks due to short propagation delays, although concurrent with our work, some researchers have begun to use game theoretic approximations of contender counts [CLD06]. We exploit CTC in our MAC design in Section 5.3. Others have proposed flow-level contention counting for multi-hop 802.11 networks [SCBR04]; our work differs by focusing on single-hop contention as applied to MAC protocols. 5.2.3 Deafness Conditions Wireless transceivers often work in half-duplex mode, and thus on a single channel a node that is transmitting cannot receive another packet arriving at the same time. In the case of transmitting tones, a node will be unable to completely receive another tone with a probability that is proportional to the tone length. Therefore, a node becomes deaf to another transmission in these situations. The following analysis is valid for any packet length that needs to be entirely received for contention detection. We employ the low-power wake-up tone hardware 85 proposed by Willis et al. [WYH06] for sending contention tones. Note that we assume that tones and data share the same channel: one transmitter but two receivers in the same channel. Such tone detection mechanism requires energy accumulation over a minimum duration, denoted asT detect , larger than the symbol detection time for data on the same channel. This detection delay can lead to a node transmitting a tone failing to hear another tone — thus deafness occurs. Our aim is to identify the conditions that will lead to deafness. Refer to three dif- ferent circumstances in Figure 5.3, which can cause deafness at node B. We define the non-overlap zone (NOZ) as the non-overlapping region of two partially overlapping tones at B. Deafness will then occur if: NOZ <T detect (5.1) We take the first case, Figure 5.3(a), where both nodes are transmitting at the same instant. HereNOZ =T tone ¡(t tx;B +T tone ¡t rx;A!B ), A’s transmission starts att tx;A whilet rx;A!B is the global time at which B receives A’s tone. IfT A;B is the propagation delay between A and B, thent rx;A!B =t tx;A +T A;B . With these definitions and noting thatt tx;A =t tx;B we get a deafness condition specified by equation (5.1) as: T A;B <T detect (5.2) If node transmissions are synchronized, then the deafness condition is dependent solely upon the delay between the nodes, and thus with nodes spaced closer than the deaf region,D deaf =T detect ¤v sound , causing bidirectional deafness where neither node can hear the other. 86 In the second case, Figure 5.3(b), t rx;A!B < t tx;B , thus the NOZ = t tx;B ¡ t rx;A!B =t tx;B ¡t tx;A ¡T A;B . Thus the deafness condition becomes: (t tx;B ¡t tx;A )¡T A;B <T detect (5.3) Similarly for the third case, Figure 5.3 (c), we get the following deafness condition (t tx;A ¡t tx;B )+T A;B <T detect (5.4) Observe that in case (b) and (c) the deafness is unidirectional, i.e., only node B is deaf to A’s tone but A can still detect B’s tone. We can simplify all three deafness conditions above into a single generalized con- dition. We make the convention that node A transmits its tone before B. Then we can define the following time differences: Time Difference of Transmission (TDT) =t tx;B ¡t tx;A Time difference of Location (TDL) =T A;B With these definitions it is straightforward to see that all the three deafness condi- tions described above can be coalesced into a single generalized condition. Generalized Deafness Condition (GDC): jTDT ¡TDLj<T detect (5.5) GDC shows that as long as all nodes transmit equal-length tones, the deafness con- dition in equation (5.5) is not affected by the tone length. This is because of the binary 87 Data T-Lohi Frame Contention Round Reservation period Figure 5.4: The Tone-Lohi protocol frame Algorithm 1 Pseudocode for the T-Lohi protocol 1: if you receive a contention tone (CTD) while idle 2: set blocking state to true; unset at end of current frame 3: When application invokes MAC send 4: if blocked; wait for end of frame and attempt in next RP. 5: else transmit contention tone; wait for end of current CR. 6: if (contender count (CTC)> 1) 7: Computew uniformly from [0,CTC]; backoffw CR(s) 8: if CTD; while in backoff 9: set blocking state to true; unset at end of frame 10: wait for end of frame and attempt in next RP. 11: else backoff ends; goto line 5 and repeat contention 12: else contender count= 1; data reservation successful 13: transmit data; when DP ends go to idle state nature of information in a tone. The GDC also reflects the space time uncertainty by the dependence of deafness on both relative location (TDL) and transmit time (TDT). 5.3 Tone-Lohi MAC Protocol Design In this section we describe T-Lohi in detail, including its motivation and the design trade- offs behind different flavors of T-Lohi. We end by identifying the underlying assump- tions and limitations of the current T-Lohi design. 5.3.1 Overview of T-Lohi The primary objective of T-Lohi is to provide a MAC protocol that has efficient channel utilization, stable throughput, and low energy consumption. By its use of contention to reserve the channel, it provides efficient channel utilization and throughput stabil- ity. This reservation prevents data packet collision (or makes them very unlikely), thus 88 avoiding loss of throughput and energy waste. It also exploits our modem’s very-low- power wake-up tone receiver [WYH06]. In T-Lohi, nodes contend to reserve the channel to send data. (Pseudocode for the T-Lohi protocol is shown as Algorithm 1.) Figure 5.4 shows this process: each frame consists of a reservation period (RP) followed by a data transfer. Each RP consists of a series of contention rounds (CRs) until one node successfully reserves the channel. While not shown in the figure, the packet length is provided in the data header, allowing nodes to compute the end-of-frame. Contention requires that nodes first send a short tone and then listen for the dura- tion of the contention round (CR) to decide if reservation is successful. If only one node contends in a CR, it wins, ending the RP and then transmitting its data. When nodes detect contention (Algorithm 1, line 7), they randomly back-off in proportion to the contender count, extending the RP. Random backoff promotes fairness, while the window size equal to contender count allows quick convergence based on current load. The CR is long enough to allow nodes to detect and count contenders (CTD and CTC). T-Lohi uses our custom, low-power, wake-up tone receiver to conserve energy [WYH06]. Wake-up tones share the channel with data transmissions, but detect- ing a tone consumes only 2% the energy of listening for data. Transmitters send a wake- up tone before any data transmission, allowing receivers to keep their CPU and data receiver off. Powering off transmit and receive and using our low-power wake-up cir- cuit are essential to reduce energy consumption, since correct estimate of channel state requires channel awareness for times on the order of propagation delays (large fractions of a second). A larger potential source of savings follows because T-Lohi’s reservation mechanism can prevent data collisions and avoid expensive (re)transmissions. We also suppress successive transmissions from a successful sender to reduce spatial unfairness 89 A B C C’s Contention Round CR aUT A B C A and C attempt synchronously tone packet CR ST A ‘’wins’’ and transmits data Window =2, A & C backoff. A attempts. Data packet Reservation Period Data Period CR cUT CR aUT aUT-Lohi: C assume it won and transmit data. cUT-Lohi: C hears A and both back off.. (a) (b) A and C attempt asynchronously Figure 5.5: Overview of (a) ST-Lohi, (b) UT-Lohi (Section 5.2.2). The exact duration of this quiet time depends on the T-Lohi variants to be discussed in Section 5.3.2. 5.3.2 T-Lohi Flavors The T-Lohi reservation mechanism deals with how nodes contend for the channel and make their decisions on channel acquisition by taking the space-time uncertainty into consideration. The backoff mechanism dictates the reaction to a failed contention round as well as the policy to start contention in a new T-Lohi frame, leveraging information about medium access such as CTC. We next define three flavors of T-Lohi that vary the reservation mechanism with different implementation requirements and performance results. (In Section 5.4.5 we also vary the backoff mechanism.) Algorithm 2 ST-Lohi Backoff(FCC,didCntd,SAI) 1: if didCntd = true then 2: return b(random[0;1]+ SAI)¢ FCCc 3: else 4: return b(random[0;1]+ SAI)¢2 FCC c 5: end if 90 Synchronized T-Lohi (ST-Lohi) We begin by assuming all nodes are time synchronized and present ST-Lohi. Synchro- nizing each contention round simplifies reasoning about protocol correctness, at the cost of requiring distribution of some reference time. ST-Lohi synchronizes all communication (contention and data) into slots. This dura- tion of contention round is CR ST =¿ max +T tone , where¿ max is the worst case one-way propagation time andT tone is the tone detection time. Figure 5.5(a) shows ST-Lohi in action, where two nodes contend in the first CR, one in the second CR, then the winner starts sending data in the third slot. Since tones are sent only at the beginning of each CR, we know that any tones must arrive before the end of the CR and will be detected assuming no bidirectional deafness (Section 5.2.3). Since bidirectional deafness happens deterministically based on node location (and only rarely when nodes are extremely close), ST-Lohi contention will always converge and provide collision-free data transfer. Synchronization also provides information about the approximate number of nodes with data to send. We call this value the first contender count (FCC). FCC is updated if in any CR the CTC is greater than the currentFCC and decremented after each frame. In addition, all nodes can estimate the distance from a transmitter by measuring the propagation delay relative to the start of the current slot (¢T in Figure 5.1(a)). We use ¢T to compute a spatial advantage index, SAI = 1¡ ¢T CR ST . Nodes also maintain a boolean variable didCntd indicating if they attempted contention in a previous frame. This variable is reset every time node wins the frame and sends data. Algorithm 2 shows ST-Lohi’s backoff algorithm using the SAI and the didCntd flag. After contending, nodes prioritize their channel access, thus reducing the medium access latency. Nodes with higher SAI are more likely to wait an extra slot, which also reduces 91 Figure 5.6: Benefit of (a) Higher contention, (b) Aggression and asynchrony. potential unfairness that can result in channel monopoly between spatially nearby nodes (see Section 5.2.2 and Figure 5.1(b)). Conservative Unsynchronized T-Lohi (cUT-Lohi) ST-Lohi is simple to reason about and we can exploit synchronization to estimate con- tender behavior. However, time synchronization is not free, and maintaining time synchronization adds run-time overhead and protocol complexity. We therefore next explore unsynchronized protocols. In unsynchronized T-Lohi, nodes can start contending any time they know the chan- nel is not busy. To provide the same contention detection guarantee as ST-Lohi, cUT- Lohi must observe the channel forCR cUT = 2¿ max +2T tone . Consider Figure 5.5(b), where node C sends a tone at timet C . In the worst case, the second contender A sends its tone att C +¿ max +T tone ¡² because it is as far from C as possible and sends just before hearing C’s tone, and A’s tone will arrive and be detected at C att C +2¿ max +T tone ¡². Unlike ST-Lohi, cUT-Lohi cannot estimate a variable similar to FCC because of an asyn- chronous view of a contention round, it therefore defaults to just the quite period of a single CR duration after each transmission. 92 Aggressive UT-Lohi (aUT-Lohi) Although cUT-Lohi avoids the complexity of synchronization, its long contention time reduces throughput. Aggressive unsynchronized T-Lohi (aUT-Lohi) follows cUT-Lohi, but cuts the duration of its contention round to CR aUT =¿ max +T tone . The purpose of the long listen in cUT was to account for worst-case timing of tones. In aUT-Lohi, worst-case timing results in either a tone detection (as before), or a tone-data collision or data-data collision, depending on the relative distances of the two senders and a receiver. Consider Figure 5.6(b): B’s tone will not be heard by A within CR aUT , so A will assume it has acquired the channel and transmit data at t tx;C +CR aUT . B’s tone and A’s data transmissions will collide at a node located within the shadow region near A (a tone-data collision), but be received separately at other nodes. Also, node B will hear A’s tone and backoff. We describe these scenarios in more detail in Section 5.3.3, arguing that the conditions that result in data collisions are quite unlikely. Simulation results in Section 5.4.4 verify the low probability of such events as there are few packet losses for aUT-Lohi. 5.3.3 Discussion on Protocol Correctness T-Lohi avoids packet collisions through a reservation mechanism. However, deafness and aggressive contention can cause the reservation mechanism to fail and lose packets. We next define conditions that lead to incorrect reservation, protocol incorrectness, and can cause packet loss. These cases include tone-data collision, data-data collision and persistently incorrect reservation. We also discuss how higher contention can lead to partially correcting these problems. 93 Tone-Data Collision As described above in Section 5.3.2, tone-data collision can occur in aUT-Lohi because contenders listen for only¿ max . (It also occurs in very unlikely corner cases with cUT- Lohi and ST-Lohi.) The necessary conditions for tone-data collision in aUT-Lohi are: Tone-Data coexistence conditions: TDT < (TDL+T tone ); TDL¸¿ max =2 (5.6) The left inequality states that the interferer B must transmit before A’s tone is detected by B, as tone detection precludes any contention attempt. (see Figure 5.6(b)). This condition is a superset of the deafness condition, so if deafness occurs, it will be satisfied, but not vice versa. The second equation represents the case that B is located far enough from A so that the CR at A ends before A can detect the tone sent by B. However these conditions are not sufficient for tone-data collision. The overlap of tone-data must occur at the receiver (within the shadow region of A as shown in Fig- ure 5.6(b)) for an actual tone-data collision. This additional condition makes tone-data collision less likely to occur; (also supported by the very small number of tone-data packet losses in simulations in Section 5.4.4). In fact, if the receiver is not in the shadow region, a transmission in aUT-Lohi actually succeeds (because tone and data do not collide) in situations where ST-Lohi and cUT-Lohi would extend the reservation period. Data-Data Collision Data-data collisions can also occur in T-Lohi if two nodes believe they have won the reservation and so transmit simultaneously. 94 In ST-Lohi, data-data collisions occur only as a result of bidirectional deafness when reserving nodes are within D deaf —this condition is necessary and sufficient for data- data collisions. (D deaf is quite small for our T detect ; in simulations with random node placement only 0.14% of node pairs are bidirectionally deaf.) Data-data collisions can also occur in aUT-Lohi when pseudo-bidirectional deafness occurs, that is when both tone-data coexistence conditions (Equation 5.6) and deafness condition (Equation 5.5) are met. This collision occurs as one node of the pair will assume data reserved because of its aggressive round length, while the other will do the same due to deafness. Such collisions need to be handled at a higher layer using back off and retransmission. Benefit of High Contention Finally, although we describe collision scenarios above, the presence of an additional contender can solve these situations by effectively extending the reservation period. Figure 5.6(a) illustrates this effect for ST-Lohi, where contending nodes A and B are within each other’s deaf region. In this case, bidirectional deafness would normally cause both nodes to send data packets that would then collide. However, addition of a third contender C causes both A and B to detect another contender. All nodes backoff and prevent an incorrect data reservation. If this backoff places A and B in separate CRs, then no collision will occur. Similarly additional contenders also “break” the pseudo bidirectional deafness of aUT-Lohi and prevent packet collisions. 5.4 Basic Performance Evaluation We next evaluate basic performance of T-Lohi through simulation. We begin by look- ing at the design tradeoffs between its three flavors. We also evaluate important MAC 95 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 μ − Packet duration in multiples of CR Maximum Throughput/Channel Utilization cUT−Lohi Operation point (μ c =0.96, U c =0.49) aUT and ST−Lohi Operation point (μ a =1.92, U c =0.66) Figure 5.7: Maximum theoretical utilization for T-Lohi protocols as¹ is varied, showing the operational points for our simulations. metrics such as throughput, energy efficiency, and fairness. Simulation results show that T-Lohi achieves better throughput (34–50%) than prior published results (22%) of other throughput-efficient, underwater MAC protocols [PS06, MS06], while maintain- ing energy efficiency comparable to UWAN-MAC [RP05]. Finally, we quantify the impact of the unique characteristics in acoustic medium access, such as deafness and contender counting, on performance. In the next section (Section 5.5) we verify these simulation results using mathematical analysis of the duration of reservation period. 5.4.1 Simulation Methodology We develop a custom acoustic network simulator based on a prior model for under- water time synchronization [SH06]. (The simulator and simulations are available for download from the authors’ website.) We do not currently model packet loss due to 96 0 1 2 3 4 5 6 0 0.1 0.2 0.3 0.4 0.5 0.6 Mean Offered Load (packet/sec) Channel Utilization (8 nodes) Optimal Throughput (U a ) Optimal Throughput (U a ) T−Lohi (conservative) Capacity aUT−Lohi cUT−Lohi ST−Lohi (a) Eight Node Network 0 1 2 3 4 5 6 0 0.1 0.2 0.3 0.4 0.5 0.6 Mean Offered Load (packet/sec) Channel Utilization (2 nodes) Channel Capacity T−Lohi (aggressive) Capacity aUT−Lohi ST−Lohi cUT−Lohi (b) Two node network Figure 5.8: Channel utilization of three T-Lohi flavors. The vertical lines show the channel capacity and the protocol capacities, in packets/s. channel noise and multi-path, and mainly focus on protocol behavior. Exploration of these effects of is an important direction for future work. Our default simulation parameters are randomly deployed nodes in a 300£400m area. The network is fully connected with acoustic transmission range of 500m. Trans- mit data rate is 8kb/s, with 650-byte packets, implying that packet transmission duration P tx is 650ms. Tone detection time is 5ms. Each simulation lasts 100s and repeats 500 times. We show the mean and 95% confidence intervals in each graph, but in most cases confidence intervals are barely visible. All simulations in the current section used a single packet buffer to which packets arrived with exponential inter-arrival delays. We employ no retransmission of dropped or missed packets. Mean offered load is the aggregate packet arrival over all nodes in the network. In Section 5.6.3 we change to a 5 packet transmit bursts. We also re-evaluated selected results with infinitely deep queues and found T-Lohi performance did not change significantly. 97 5.4.2 Network Throughput Our first goal is to understand how throughput is affected by changes in offered load, network density, and protocol choice. Throughput is an important metric in acoustic communication because of the very limited bandwidth. We first define the maximum theoretical throughput for T-Lohi, assuming perfect scheduling, to provide a perfor- mance goal. In this ideal case, there is only one contender per frame, and all T-Lohi RPs will consist of a single contention round (CR). With perfect scheduling, the best possible throughput is the ratio of data to frame length: P tx =(P tx + CR). To divorce achievable throughput from a particular topology or hardware, we normalize by¹=P tx =CR, the packet transmission time in multiples of contention rounds. T-Lohi’s maximum throughput is then: TH max =¹=(¹+1) (5.7) Figure 5.7 shows how the best possible performance varies with¹. In simulation, we send fixed amount of data in each packet (650B), but variation in the duration of the con- tention round means that aUT- and cUT-Lohi have different achievable performances. This figure also shows the operational points we use in our simulations with a fixed data size; other points on this curve could also be used. For these parameters, the best possible utilization,U, isU a = 0.66 for ST- and aUT-Lohi, andU c =0.49 for cUT-Lohi. Throughput as Load Varies We first examine how the throughput of T-Lohi responds to varying offered load. We expect T-Lohi to be throughput stable because it can detect and count contenders. 98 Figure 5.8(a) shows channel utilization as a function of aggregate offered load. The figure also shows two theoretical limits while operating at¹ a and¹ c . First, the vertical lines show limits on the offered load due to channel and protocol capacities. Second, we also plot the optimal utilization curves for T-Lohi as the load varies. We have three observations from this simulation. First, T-Lohi is very efficient at low offered load, where contention rates are low. When the load is less than 0.5 packets/s, T-Lohi is very close to the maximum theoretical utilization. Second, as offered load approaches the practical capacity (0.5–1 packet/s), we see T-Lohi reaches about 50% of maximum utilization. Finally, as offered load exceeds practical capacity (more than 1 packet/s), we observe that T-Lohi throughput remains stable. As Figure 5.9 shows (for ST-Lohi; both aUT and cUT-Lohi exhibit similar curves), the duration of reservation period doubles when the offered load increases between 0.5– 1 packet/s, resulting in the decreased throughput. This figure can also be construed as the MAC delay or latency, which in this case is independent of offered load. Further- more, Figure 5.9 also indicates that the reason for stable throughput is the near constant reservation period duration. Thus the combination of contention detection and load- influenced contention counting allows makes throughput stable and load-independent. Impact of Protocol Choice on Throughput To observe how different protocol design (Section 5.3.2) affects channel utilization, we next compare the three T-Lohi flavors. Figure 5.8 shows the channel utilization of T-Lohi flavors at two different network densities. We first observe that cUT-Lohi has saturation capacity about two-third of aUT- Lohi, primarily because of its longer CR length. Although cUT-Lohi has a contention 99 0 1 2 3 4 5 6 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 Mean Offered Load (packet/sec) Mean contention rounds in a RP 8 nodes 4 nodes Figure 5.9: Average number of contention rounds in a reservation period for ST-Lohi. round that is twice that of aUT-Lohi, its capacity is not halved. This disparity is due to the non-linearity of achievable utilization as predicted by Equation 5.7. More interesting is that aUT-Lohi always achieves higher utilization than ST-Lohi (slightly higher with 8 nodes and much better with only 2 nodes). This result is due (except at low densities, which we explain next) to the slotted access in ST-Lohi that delays all access attempts to the start of the next slot. When both have the same CR (CR aUT =CR ST ), this delay (on average half CR) results in greater reservation latency for ST-Lohi. In summary, all the T-Lohi flavors have similar throughput behavior, but ST-Lohi and aUT-Lohi offer higher throughput than cUT-Lohi due to their smaller CRs. Impact of Network Density and Packet Length on Throughput We next explore how network density and packet length affect T-Lohi’s throughput. The throughput of traditional wireless MACs degrades with density, but we expect T-Lohi to remain stable based on the results from Section 5.4.2. 100 0 2000 4000 6000 8000 10000 12000 14000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Packet length in bytes Saturated Channel Utilization (2 nodes) aUT−Lohi cUT−Lohi ST−Lohi Maximum Theoretical 0 5 10 15 20 25 30 35 40 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Packet duration in multiples of CR (aggressive) (μ) Figure 5.10: Channel utilization as¹ is varied by changing the packet length. Comparing Figures 5.8(a) and 5.8(b), we observe that utilization is significantly lower for aUT-Lohi and cUT-Lohi in denser networks (compared to a 2 node network). In fact the decrease by nearly 15% is evident even at 4 nodes (not shown here) and does not vary significantly for higher densities. Utilization of ST-Lohi, however, does not show such dependency on network density. We have separately evaluated T-Lohi throughput at higher densities (16 and 32 nodes), but we observe no significant differ- ences in throughput curves there. The higher throughput with two nodes is explained by a combination of asynchrony and the mechanism to handle spatial unfairness. With two nodes and asynchronous access, the quiet period after successful transmission (Section 5.3.2), allows the two nodes to repeatedly swap the channel with just a single CR per frame. However, the similar effect does not often occur in ST-Lohi because of slotted transmission times. In Figure 5.1(b), only node A contends in slot 3, since B remains quiet in slot 3 to promote 101 0 1 2 3 4 5 6 0 1 2 3 4 5 6 7 8 9 10 Mean Offered Load (packet/sec) Percentage overhead to optimal energy/Packet aUT−Lohi cUT−Lohi ST−Lohi Figure 5.11: Relative energy overhead for T-Lohi for an 8 node network Table 5.1: Acoustic Modem Power Draws Mode Data Wake-up Tone Transmit (Max) 2W 2W Receive 20mW 0.5mW Idle/Listen 20mW 0.5mW fairness. Nodes that are further away, such as C or D, start contending in slot 4 (not shown in figure) along with B whose quiet period would have ended. With more than two nodes, this channel swapping is not possible with either flavor of Lohi, since more than one CR will be required. We also varied ¹ using longer packet length, and observed (Figure 5.10) that the throughput increases monotonically with packet length or ¹. Furthermore, under all operating regimes the utilization achieved by T-Lohi remains within 35% of the theoret- ical optimal given by Equation 5.7. 102 0 5 10 15 20 25 30 35 40 45 50 −1 0 1 2 3 4 5 6 7 Mean Offered Load (packet/sec) Packets Lost (in 100 sec) aUT−Lohi ST−Lohi cUT−Lohi 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8 Network size Packets Lost (in 100 sec) Load=0.433 Load=3.22 Load=53.59 (a) 2 node network (b) aUT-Lohi with> 2 nodes Figure 5.12: Packets lost in a fixed duration as offered load is varied 5.4.3 Energy Efficiency Since underwater sensornets are often energy constrained, we next consider the energy efficiency of T-Lohi under varying loads. We expect T-Lohi to be energy efficient because wake-up tone detection reduces the energy cost of long data reservation peri- ods. The modem power values used in our simulations are shown in Table 5.1 and roughly match the power consumption of our acoustic modem with wake-up tone sup- port [WYH06]. Figure 5.11 shows the percentage energy overhead of T-Lohi in an eight-node net- work. We define energy overhead as the cost beyond the optimal energy per packet used in transmitting and receiving a single packet. All protocols are very efficient under all loads, with energy overhead at most 9% over the optimal cost. ST-Lohi has a very low and nearly constant energy overhead (just 4% over the optimal) because it prevents any data collision. The overhead is solely due to the cost of sending and receiving tones dur- ing the contention rounds. The energy cost of aUT-Lohi increases marginally at higher loads (9% over optimal at high load versus 4% at low load) due to data collisions caused by its aggressive policy. 103 More interestingly, aUT-Lohi and cUT-Lohi have similar energy overhead. While aUT-Lohi gets more packets through than cUT-Lohi, the latter sleeps for longer periods, so the energy cost per packet becomes similar under the Poisson traffic model. For lower network density (4 nodes) cUT-Lohi is about 40% more energy efficient than aUT-Lohi. The reason can be explained from results in next section where we show that higher density reduces the probability of packet loss for aUT-Lohi. 5.4.4 Protocol Correctness: Impact of Deafness and Aggression We now evaluate the impact of deafness and aggressive contention on T-Lohi. Deaf- ness and aggressive contention can cause protocol incorrectness (Section 5.3.3), where multiple nodes believe they have reserved the channel. We quantify the impact of these factors by measuring packet loss over a fixed interval as offered load varies. Figure 5.12(a) allows us to make several interesting observations for a two node net- work. First, cUT-Lohi experiences practically no collision at any offered load. ST-Lohi has very few packet losses but shows high variability, while packet loss for aUT-Lohi increases proportionally to the network load. Investigation of packet loss in aUT-Lohi reveals that, in most cases, loss is due to data-data collisions, and that such collisions become more likely at higher load due to its aggressiveness. The results of packet loss for both cUT-Lohi and ST-Lohi show very little variation over all network densities (omitted here). In Figure 5.12(b), we see that in aUT-Lohi, the number of losses and its variance decrease as more nodes contend, because more nodes easily break the pseudo-deafness conditions necessary for data-data collision. These results show that under high contention, the impact of both deafness and aggression (in aUT and ST-Lohi) becomes negligible. Meanwhile, cUT-Lohi provides the most reliable data transfer, especially for sparse and low traffic networks. 104 0 1 2 3 4 5 6 7 8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Mean Offered Load (packet/sec) Jain’s Fairness Index aUT−Lohi cUT−Lohi ST−Lohi ST−Lohi with BEB T−Lohi Protocols ST−Lohi with BEB Figure 5.13: Jain’s fairness index for T-Lohi that can count contender vs. a MAC that can only detect contenders and uses BEB. 5.4.5 Impact of Contender Detection and Counting T-Lohi exploits space-time uncertainty to provide contention detection (CTD) and counting (CTC). Here we separate these capabilities to quantify the impact of contender counting on fairness. To evaluate the benefits of channel observation we compare T-Lohi with contender counting to a modified version that can only detect (but not count) contention and thus uses binary exponential backoff (BEB). Since systems with collision detection exhibit throughput stability even at high loads (as in Ethernet [WK99]), here we focus on fairness. Binary backoff provides insufficient information to easily provide fairness. T-Lohi’s contender counting allows for a traffic-adaptive backoff mechanism (Algo- rithm 1). 105 We use Jain’s fairness index defined as (§x i ) 2 (n¢§x 2 i ) where x i is the number of packets successfully sent by a node, andn represents the number of nodes in the network. Fig- ure 5.13 shows the result for an experimental setup consisting of eight nodes run for 500s to strenuously test protocol fairness. We first observe that the T-Lohi protocols exhibit a high fairness index (0.9 and above) that remains nearly constant across all offered loads. In comparison the version employing BEB instead of using contention count for back- off shows an exponential decay in its fairness index. The reason for traffic independent fairness in T-Lohi is again the ability to backoff based on an accurate view of the current congestion level. 5.5 Analyzing T-Lohi Reservation Period We now rigorously analyze the performance of our MAC protocol to mathematical understand its throughput stability under high loads. For this purpose we mathemati- cally estimate the duration of T-Lohi’s reservation period (RP), which we define as its convergence time. This convergence time is essential in determining the bounds on T- Lohi’s throughput and latency, as data transmission immediately follows the end of a RP. To make our mathematical analysis we have made three assumptions. First, we con- sider only a synchronized version of T-Lohi where all nodes make attempts in temporally equal slots. Second, we ignore the backoff introduced in our algorithm for spatial fair- ness. Lastly, we assume a saturated network so that we have all nodes simultaneously attempting in the first slot at the end of each data frame. We first model the reservation process by defining conceptual super-rounds. Next we use this model to describe a simplified Markov chain model of the reservation process 106 Data Reservation period 3 0 0 3 0 1 3 0 2 0 2 1 1 st SR 2 nd SR 1 st SR 2 nd SR Entry contention round 3 rd SR Figure 5.14: Super-rounds: Example of a T-Lohi reservation process broken into con- ceptual super-rounds. The number in the contention round represents the number of nodes attempting contention. for a 2 node network and then extend it to a more general model. Finally, we describe how we can obtain bounds on the convergence time using our Markov model. 5.5.1 Super-Rounds To help model T-Lohi, we conceptually divide a reservation period into super-round(s). These conceptual super-rounds , as we show next, can be tractably solved for their dura- tion which we then use to solve the broader goal of finding the duration of a reservation period. Super-rounds are characterized by the number of nodes at its beginning. Thus an N-node super-round (SR) starts withN contending nodes, each aware of the otherN¡ 1 contenders. The super-round ends when at least one node attempts contention by sending a tone. In between the nodes backoff uniformly in a window of N contention round (CR). If M (· N) nodes simultaneously re-attempt a new SR starts with M nodes and the reservation period continues. Thus a reservation period consists of one or more super-rounds, ending when only a single contender remains at the end of a SR. To complete the length of a reservation period we need to add the first CR after data in which all nodes make an attempt (and were previously not aware of how many contenders exist). 107 0 5 10 15 20 25 30 35 40 1 1.1 1.2 1.3 1.4 1.5 1.6 Number of nodes Average Length of Super−rounds (CP/SR) Figure 5.15: Average Length of Super-round (in number of contention rounds) with variable number of nodes Figure 5.14 explains the concept of a super-round for a saturated network with three nodes. Once a data transmission ends all three saturated nodes contend. After the first contention all nodes backoff with a window of three. As an example scenario consider the first reservation period where all three nodes choose the third contention round. The first super-round thus spans three contention rounds, at the end of which all nodes again have a window of three. The next super-rounds ends within two rounds with only one node attempting in the second round. The reservation period ends as there is just a single contender who then transmits data onto the channel. As our example shows, super-rounds can be of any length up to the size of the contention window (thus a 3 node SR can last 1,2 or 3 contention rounds). Expected Length of a Super-Round We now compute the expected length of any super-round. This expected length will then be used to compute the duration of the complete reservation period. For this purpose we define a random variable X N : the length of super-round with N nodes, measured in 108 contention rounds (CRs). We can define the probability of this random variable being a particular valuei2[1;N] as follows: P(X N =i)=P(A N i \B N i )=P(A N i jB N i )P(B N i ) (5.8) In the above equationA N i represents the event that atleast one node makes an attempt in the i th CR and thus ending the super-round. B N i , on the other hand represents the event that no node as made an event in any of the priori¡1 contention rounds; other- wise the super-round would have ended earlier. Since a super-round ends once the first attempt has been made the eventX N =i occurs when both these events occur together. Since P(B N i ) represents the probability that no one has made an attempt in any previous CR, it is the compliment of the sum of events thatX N =k wherek2[1;i¡1]. ThusP(B N i )= ³ 1¡ P i¡1 k=1 P(X N =k) ´ . Also, given that no attempt has been made in thei¡1 previous contention round, the probability is obtained by considering a uniform distribution overN¡(i+1). ThusP(A N i jB N i )=1¡(1¡1=(N+1¡i)) N . Combining these two definitions into Equation 5.8 gives us a recursive solution for P(X N = i), where the termination condition isP(X N =1)=1¡(1¡1=(N)) N . P(X N =i)= Ã 1¡ i¡1 X k=1 P(X N =k) !Ã 1¡ µ 1¡ 1 N +1¡i ¶ N ! (5.9) Using Equation 5.9 we can now estimate the average duration of a super round. E[X N ]= N X i=1 iP(X N =i) (5.10) Figure 5.15 shows the result of plotting this expectation for different numbers of nodes. We observe that having the exact count of contending nodes (using the CTC 109 capability exposed by T-Lohi) allows the super-round to end within an average of 1.5 CR, even for very high density networks. We next use the result from this section to ana- lytically compute the length of a reservation period for a simplified two node network. 5.5.2 A Simplified Analytical Solution We now present an analytically closed form solution for a saturated 2-node network. We want to evaluate the difficulty of extending this solution to a general network. Here we model the length of a reservation period in multiple of super-rounds. In such a simple case the probability of collision (both nodes attempting simultaneously) or resolution (only one node attempts) is equally distributed. We introduce a new random variable Y i , indicating the event that a RP ends after exactlyi super-rounds. For the simple two node case this event has the following prob- ability: P(Y i )=P(a single contender in thei th SR j all previous i-1 SRs had collisions) =( 1 2 ) i¡1 ( 1 2 )=( 1 2 ) i (5.11) In order to find the average duration of a reservation period we use the standard definition of expectation for the random variable as E[Y]= 1 X i=1 i£P(Y i ) = 1 X i=1 i£(0:5) i =2 (5.12) 110 We solve the above equation by taking a derivative of a standard summation P 1 i=1 (a) i which has a well known closed form. Thus on average two super-rounds will be required before the medium contention is successful. Since the average duration for two node super-round is 1.25 contention rounds (from Equation 5.10), we say that on average 2£ 1:25 = 2:5 contention rounds are spent in just super-round for each reservation period. Adding to this duration the first CR in which nodes become aware of other contenders, and is not part of any super-round (see Figure 5.14), we find the estimated duration of a reservation period to be3:5 contention rounds. The result above agrees with separately (not shown here) performed simulation result for a two node case where the mean reservation period length is found to be 3.474 contention rounds with 95% confidence interval of 0.04 CR. While Equation 5.12 provides a closed-form solution, generalizing the probabilities to a network of more than two nodes becomes intractable. Instead, we next present a Markov-chain model of the reservation period which we then solve for a generalized network. 5.5.3 Markov Chain Model We now present a Markov-chain model of the reservation period which we will next show can be solved for any general network. This solution will provide a mathematical bound for T-Lohi’s convergence time that we compare with our simulation results in Section 5.4. We start by representing each super-round as a state. Thus anN node super-round’s state is S N . In our Markov model of T-Lohi’s reservation process, the network starts from a special entry state S Ne . This entry state represents the first contention round 111 S 2e S 1 P 2e,2 =1 S 2 P 2,1 =0.5 P 1,2e =1 P 2,2 =0.5 Entry Contention Data transmission Figure 5.16: A special 2-Node model of the T-Lohi reservation process. where all the contending nodes necessarily collide. This is necessary to capture the ini- tial contention round which is not part of any SR (Figure 5.14) Thus nodes transition from this entry state to a SR state in exactly one contention round. Each node then either stays in that state or transitions to a state with fewer number of nodes. The transitions between and within the contention states can happen with a variable delay based on the length of the super-round. For example in the first reservation period of Figure 5.14 the transition from a 3 node state (S 3 as per our definition here) can happen in three contention rounds (as shown) or also in two CR if all three nodes make an attempt in the third round. Modeling such intricate transitions makes the model cumbersome; we instead approximate by assigning E[X N ], the estimated length of for S N (from Equa- tion 5.10), as the transition delay from S N to any other state. We show, by comparing with simulation results, that this approximation has no significant affect on the accuracy of our model. The reservation period ends when the transition is made to another special state S 1 that represents selection of one of the original N nodes for data transmission. Finally this special state returns to the original entry state to represent a saturated network where allN nodes start contending again. We purposefully ignore the data transmission time that must occur for this transition since that delay is not part of the convergence time that we are modeling here. 112 Example for a two node network We first provide, for the purpose of clarity, an example of the Markov chain model in a relatively simple two node network. This example will capture the essence of our modeling process and how we solve our model to find the convergence time of T-Lohi. Our Markov chain model for a two node saturated network is shown in Figure 5.16. The network starts in the special entry stateS 2e in which both nodes contend simultane- ously. After a single CR they transition to the super-round (SR) state ofS 2 , where each node chooses to backoff and reattempt within the next two contention rounds. Since its equally probable that the nodes choose the same round or a different round. the net- work both returns toS 2 or transition to the data transmission stateS 1 with 1 2 probability. As noticed before, in order to model a saturated network (and make the Markov chain irreducible) we return to the special entry stateS 2 e . We now find the average duration of the reservation period as T 2 , the time taken in going from state S 2 to S 1 . Since we assume no time is spent in S 1 (where reservation has ended and data transmission starts), T 1 is by identity zero. We can therefore solve for the RP duration as follows: T 2 =P 2;1 E[X 2 ]+P 2;2 E[X 2 ] =(0:5)(1:25)+(0:5)(1:25)=2:5 (5.13) Adding the initial CR for the transition fromS 2 e toS 2 we reach an average reserva- tion period length of 3.5 contention round; this agrees with both our simulation previous sections result. We will now generalize this mechanism used to solve the two node network to solve a generalized Markov chain which can represent any network density. 113 Contention states S Ne S N S 1 P N,1 P Ne,N =1 S N-1 P N,N-1 P N-1,1 S 3 P 3,1 P 3,2 P N,3 P N-1,3 S 2 P N,2 P N-1,2 P 2,1 P 1,I e =1 P 3,3 P 2,2 P N-1,N-1 P N,N Entry Data transmission Figure 5.17: A generalized Markov chain model to analyze the convergence time of a T-Lohi reservation period. 5.5.4 Generalized Reservation Period Duration We now explain our model for a general network topology. Figure 5.17 shows this general Markov chain model with a network of N saturated nodes. The chain starts in the entry stateS N e depictingN saturated nodes at the start of every reservation period. We transition from this entry state to a corresponding super-round state S N in a single contention round (the first of the reservation period) where each node discovers there are N ¡1 other contenders. Each state of the Markov chain either loops-back or goes to a lower value state with a certain transition probability. Thus stateS N can transition to any of theN¡1 states where fewer number of nodes contend or loop-back. We now need to define the transition probability P N;j between states S N and S j . Since the transition to state S j can happen in any of the N possible contention rounds we define this probability in the following manner: 114 P N;j = N X i=1 P(C N i;j \B N i ) = N X i=1 P(C N i;j jB N i )P(B N i ) (5.14) HereC N i;j represents the event that exactlyj nodes collide in thei th ·N contention round of a super-round with N nodes. B N i is the same event defined in Section 5.5.1; the event that none of the N contending node made a contention attempt in any of the prior i¡1 contention rounds. P(C N i;j j B N i ) therefore can be defined as the combined probability of ¡ N j ¢ possible event where j nodes choose the i th round with probability ( 1 N+1¡i ) j while the remaining N ¡j choose not to with probability (1¡ 1 N+1¡i ) N¡j . Using the prior definition ofP(X N = k) andP(B N i ) we get the following state transi- tion probability: P N;j = N X i=1 Ã 1¡ i¡1 X k=1 P(X N =k) !Ã µ N j ¶µ 1 N +1¡i ¶ j µ 1¡ 1 N +1¡i ¶ N¡j ! (5.15) We would like to point out that Equation 5.15 does not cover the transitions from the special state S Ne and S 1 . The transition to the state S 1 results in an immediate (incurring no time-cost in contention rounds) transition to the original entry state S N e . This transition is added to model a fully saturated network as explained previously. Also, we separately define thatS Ne transitions toS N in one contention round to capture the first round not part of any super-round. 115 0 5 10 15 20 25 30 35 40 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Number of nodes Reservation Period Length (CR/RP) Simulation Result Markov Analysis Result Figure 5.18: Average Length of Reservation Period (in number of contention rounds) for a varying network density. Solving for RP Duration We now solve our Markov model of the reservation period to find it average duration in the number of contention rounds 1 . For this purpose we observe that, barring the special entry state, any state S j is reached only from either itself or higher states (fromS i ,i¸j). Hence if we defineT i as the time duration it takes to transition from stateS i toS 1 , we can use the above obser- vation to solve the model as a simple recurrence. To find the length of the reservation period we thus need to find justT N , which is solved as the following recurrence: T N = N X j=1 P N;j (E[X N ]+T j ) T 1 =0 (5.16) 1 We would like to acknowledge and thank Dr. Brian Tung who assisted us in finding a mechanism to solve our markov chain for the reservation period duration. 116 Here we have assigned a length of E[X N ] (which is the average length of super- round from Equation 5.10) to any transition from stateS N to make the problem tractable. Equation 5.16 thus provides us with the duration in terms of contention round(s) of going from the first contention state S N of a generalized N node network to S 1 where the reservation period ends. By adding the initial single contention round transition from S N e to S N we are therefore able to find the average duration of a T-Lohi reservation period for a saturated (and therefore the worst case)N-density network. Equation 5.16 does not have a closed form solution. We numerically solve the equa- tion for different network densities and show the result in Figure 5.18. We also provide a comparison with simulation results of a simplified (saturated, synchronized and with no consideration for spatial fairness) T-Lohi reservation process. We next discuss the results shown in Figure 5.18. Discussion of Results In Figure 5.18 we first observe the initial decline for average duration of reservation period beyond the two node case. This decline is thereafter followed by a gradual and monotonic increase in RP duration. This initial decline — from 3.5 rounds to 3.43 rounds per RP — is explained by observing that the values of P N;1 steadily increase from a lower-bound of 0.5 forN = 2. In other words, the probability of a single node making the first attempt, and therefore winning the right to channel access, increases with density (e.g. it is 0.55 for N = 3 and 0.58 forN = 40). The average reservation length (Equation 5.16), however, depends on the sum of all intermediate transitions that lead to the termination state. Since the two node case has no intermediate state the duration is largely dependent on the direct transition to the termination stateS 1 . Going from two to three nodes, the increased probability of going directly to termination state 117 coupled with a similar transition duration (E[X 2 ] = 1:25 and E[X 3 ] = 1:33) and very little contribution from intermediate transitions, results in a shorter total duration for RP. However as the number of nodes, and the length of the Markov chain, only increases the intermediate transitions start to monotonically increase the duration of a reservation period. The above conclusion initially seem to deviate from the throughput results described earlier (Section 5.4.2 and Figure 5.8) where we explain a two node network having higher throughput (and therefore shorter average RP length) due to a combination of asynchronous access and the spatial fairness imbued into the T-Lohi protocol. Our results are for a simpler version of a synchronized T-Lohi that does not implement spa- tial fairness. We next observe that our numerical results are consistent with our simulation results as they are within its 95% confidence interval. This result shows that our approximation of using expected values (and not the exact value) for each super-round duration, does not greatly impact our models accuracy. Our analytical results allow us to also conclude that the performance of T-Lohi throughput and access latency are generally insensitive to network density. Thus, regard- less of network density, we can get a packet transmission reserved in nearly constant time (3.5-3.6 contention rounds) due to a count of contending nodes (using T-Lohi’s CTC capability). As a corollary we observe that these results are for a saturated net- work; we have shown through our protocol simulations that the performance of T-Lohi is sensitive to load and the reservation period will be shorter at lower load (Section 5.4.2). The insensitivity of T-Lohi to network size makes it especially suited for high density networks where significant amount of data would need to be transmitted simultaneously, 118 0 1 2 3 4 5 6 0 0.1 0.2 0.3 0.4 0.5 0.6 Mean Offered Load (packet/sec) Channel Utilization: Unsynchronized 0.5 * MAX DELAY 2 * MAX DELAY (cUT−Lohi) 1* MAX DELAY (aUT−Lohi) 0.3 * MAX DELAY (a) Throughput 0 1 2 3 4 5 6 5 10 15 20 25 30 35 40 45 Number of Nodes (packet/sec) Relative Energy Efficiency: UTLohi 0.5 * MAX DELAY 2*MAX DELAY (cUT−Lohi) 1*MAX DELAY (aUT−Lohi) (b) Relative Energy Efficiency Figure 5.19: Impact of different contention round length on Unsynchronized T-Lohi for example reporting a seismic event to a gateway node. For such scenarios the perfor- mance of T-Lohi will not be degenerate for any network density under a sudden increase of traffic. 5.6 Evaluation of Design Alternatives After having evaluated the protocol performance using simulation, we now evaluate protocol design alternatives like the duration of contention round and different T-Lohi flavors. We also compare T-lohi’s throughput and energy efficiency with a few canonical protocols. 5.6.1 Choice of Contention Round Duration We first investigate the impact of the duration of contention round (CR) in T-Lohi, as this parameter limits the throughput. While the round-trip time is selected in cUT-Lohi (Section 5.3.2) to guarantee colli- sion avoidance, the choice of maximum propagation delay as CR duration for aUT-Lohi 119 seems arbitrary. Moreover, our research on adapting slotted ALOHA to underwater acoustic environment shows transmission slots with additional guard bands (quiet time after data) achieves higher throughput [SYKH07]. Another recent MAC protocol for UWSN similarly advocates the benefit of ignoring worst case delays since such sce- narios are rare [PS06]. We believe that although the throughput would increase with shorter contention durations, the stability and energy efficiency is likely to decrease (due to more data collisions). Figure 5.19 compares the throughput and energy conservation of the unsynchronized version of T-Lohi with different contention round durations (for reasons of clarity and space we omit similar analysis and results for ST-Lohi). We examine a wide set of contention durations (between 0.1 to 2 times the maximum delay, with a granularity of 0.1); here we show specific cases to represent general trends. The throughput per- formance (Figure 5.19(a)) of the protocol is maximized at a CR duration equal to half of the maximum propagation delay. Any shorter CR duration lowers throughput and reduces throughput stability at high loads. On the other hand, reducing CR to below the maximum propagation delay (the aUT-Lohi case) increases energy overhead (Fig- ure 5.19(b)). We see that the energy overhead nearly quadruples when CR duration is half of the maximum delay, with further increase for shorter CR durations (not shown here). These results are explained by noting that while a shorter CR duration results in less per packet overhead (and therefore higher throughput), it also increases the probabil- ity of data-data collision as a result of incorrect reservations (lower throughput). The energy cost, on the other hand, only increases with a shorter CR duration (due to higher collision probability). The reason for similar energy efficiency of aUT- and cUT-Lohi was discussed in more detail in Section 5.4.4. 120 Table 5.2: Table comparing the performance of T-Lohi Flavors T-Lohi Flavor Throughput Energy Efficiency Fairness Correctness Complexity ST-Lohi Good High High High High cUT-Lohi Low High High High Low aUT-Lohi Good High High Good Low Conclusion: These results show that our choice for the contention round duration faithfully fulfills the design goals of T-Lohi for both energy and throughput efficiency. Other values might improve throughput, but cannot provide correctness guarantees and thus result in unacceptable amount of packet and energy losses. 5.6.2 Comparison of T-Lohi Flavors We next compare the T-Lohi flavors (ST-Lohi, cUT-Lohi, and aUT-Lohi). Table 5.2 encapsulates our evaluation of these protocols for a few important MAC attributes that are part of our initial design. ST-Lohi performs well, but slot synchronization increases implementation complex- ity. While cUT-Lohi has lower implementation complexity, it also has lower through- put. On the other hand, aUT-Lohi shows good performance for all measurements, with a slight degradation on correctness (prevention of any collision) due to its aggressive contention policy. However, such correctness issue becomes less of a concern at higher densities, since more contending nodes break the collision conditions (Section 5.4.4). Therefore, we conclude that the aggressive unsynchronized T-Lohi (aUT-Lohi) is the best choice for general applications. We next use aUT-Lohi to represent the T-Lohi class of MAC protocols when comparing with other protocols. 121 5.6.3 Comparison with existing MAC protocols We next compare T-Lohi to three canonical medium access mechanisms one might con- sider for underwater use: TDMA, CSMA and ALOHA. We focus on how throughput and energy efficiency compare in different scenarios. Comparative Protocols Parameters TDMA works well in some networks, although its synchronization often requires cen- tralized or complex coordination. We allocate TDMA slots to senders, with slot duration equal to packet transmission time plus the maximum propagation delay (similar to an implementation by a group at MIT [VKR + 05]). The additional wait time is required to guarantee collision free reception of a TDMA transmission. ALOHA provides an opposite extreme, with a very simple, fully distributed MAC. For the underwater acoustic environment, we compare with a modified version of slotted ALOHA that increase slot duration beyond packet length (with guard bands) to reduce collisions, an extension that is important for high-latency networks [SYKH07]. We select guard band duration to maximize throughput in our simulations from empirical evaluations. In RF, carrier sensing or CSMA significantly improves throughput compared to ALOHA. With high-latency acoustic networks, space-time uncertainty makes sensed channel state less reliable. Our implementation of CSMA is relatively simple, and is similar to the one used in Seaweb [Ric05]. Nodes transmit if channel is sensed clear at that instant; if not they backoff uniformly within the maximum propagation delay to attempt later. We now compare the throughput and energy overhead of aUT-Lohi with TDMA, CSMA, and ALOHA. We expect that TDMA will provide the best throughput near 122 0 0.5 1 1.5 2 0 0.1 0.2 0.3 0.4 0.5 0.6 Mean Bursty Offered Load (packet/sec) Channel Utilization Channel Capacity UW−Slotted ALOHA CSMA aUT−Lohi TDMA (a) Throughput 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 5 10 15 20 25 30 Mean Bursty Offered Load (packet/sec) Multiple of optimal energy UW−Slotted ALOHA aUT−Lohi TDMA CSMA (b) Relative Energy Efficiency Figure 5.20: Comparison of different MAC approaches with aUT-Lohi with bursty traf- fic and infinite packet queuing (the graphs consider slightly different ranges of offered load). saturation, while CSMA and ALOHA performance will degrade at higher loads. Tradi- tional protocols do not attempt to reduce energy consumption, so we expect T-Lohi to be much more energy frugal. We keep the same simulation parameters as Section 5.6.1, but for two. Here we consider a very dense, 32-node network. We also adopt a traffic model with a 5 packet burst, where the mean offered load is the same as in previous simulations. Comparing Channel Utilization Figure 5.20 shows the channel utilization of the protocols under bursty traffic. As expected utilization at high load is best with TDMA, since when all nodes are satu- rated, round-robin is the best possible policy. However, at low to medium loads (until 1 packet/s), before the channel starts to saturate, aUT-Lohi provides higher channel uti- lization because contention-based channel access provides lower latency. In addition, 123 both aUT-Lohi and TDMA are stable when the channel is overcommitted. By compar- ison, utilization for CSMA and our modified slotted ALOHA drops when offered load exceeds one-third of channel capacity, and asymptotically approaches zero. Comparing Energy Efficiency Figure 5.20(b) shows the energy consumed in sending a single packet as a multiple of the optimal energy. We make several observations for different protocols. First, aUT-Lohi shows a constant, load-independent energy-overhead that is lower than other protocols. Second, at low loads, other protocols show a large energy overhead due to idle listening. Finally, while energy overhead in TDMA decreases as load increases, the energy con- sumed by CSMA and modified slotted ALOHA falls first and then rises significantly. TDMA does well because it provides stable, high throughput at high loads by eliminat- ing collision (Figure 5.20). In contrast, collisions at even moderate loads cause poor performance and unstable energy usage in CSMA and modified slotted ALOHA. TDMA, CSMA and ALOHA are not designed for energy efficiency, and we have not tried to optimize their energy consumption. Improvements to their basic approaches, such as using wake-up tones before data, could greatly improve their energy efficiency. However, while this approach would help TDMA by a constant factor, it cannot address the inefficiency of CSMA and ALOHA at high loads that lead to more data collisions. 5.7 Discussion Although we explore a number of factors affecting T-Lohi above, two remain to be studied: channel effects and multi-hop operation. Our evaluation of these additional factors is ongoing, but we next provide brief, preliminary speculations as to their effects. 124 Of course, underwater channels have many sources of noise. For data packets, noise typically causes packet loss due to corruption; for tone transmission noise can result in false tones or tone loss. We do not consider packet loss in this paper, however one could add an ARQ mechanism to our protocol (for example, see Stojanovic [Sto05]). We expect that loss of data packets will lower effective throughput. We are currently evaluating the effects of noise on tones (both false positive tones and tone loss). We believe that false detection of tones simply prolongs the reservation period as the tone is consider an indication of contention. For low to moderate rates of false detection, T-Lohi will work correctly albeit with lower throughput. Numer- ical analysis of reservation period using the T-Lohi contention resolution mechanism shows that with a 50% chance of a false tone (in each CR), reservations periods require less than two additional rounds to converge (from 3.5 to 4.9 rounds per RP), although with nearly constant false tones convergence becomes unacceptable (for example, 95% chance of false tones requires 32 CR per RP). Tone loss, on the other hand, could result in incorrect reservations. However, data-data collisions will occur only if the tone is lost in the final round of contention with exactly two contenders (in other cases, contention continues in future rounds due to other tones). For these reasons, we expect that the performance degradation due to moderate channel noise will be small. In addition, the SNUSE modem targets relatively short range communication (500m or less), so some forms of channel noise will be less likely than with longer range modems [HLS + 06]. One specific class of channel noise is multi-path reflections. Our simulator does not currently consider them; we need to evaluate their effects. We do anticipate that tone echoes (“self”-multipath) could prolong T-Lohi reservation periods. We explore this effect in more detail in Chapter 6 next. 125 Finally, our simulations here consider only a fully connected network. We wish to fully understand single hop characteristics of our MAC before extending it to a multihop network. In future work we expect to explore multi-hop options and effects, including transmit power control, data and control pipelining, and hidden-terminal effects. 5.8 Summary In this chapter we show how to leverage the opportunities in acoustic medium access along with low power wake-up tone hardware to design T-Lohi, a new class of energy efficient, stable and flexible MAC protocols for UWSN. We describe three flavors of T-Lohi representing different design choices, and compare them to traditional CSMA, TDMA, and ALOHA. We show that T-Lohi provides both good throughput and energy efficiency. Furthermore, we model the T-Lohi reservation period as a Markov chain and analyze duration for a saturated network. Our numerical results from this model validate our simulation results and show the benefit of being about count contender during a reservation process. We have provided two unique capabilities to count and detect contenders based on our exploitation of the propagation delay. The contention counting capability is paramount in providing a load and density independent throughput for the T-Lohi MAC under wide range of offered loads. Thus, our work on T-Lohi MAC protocol has shown that not only can we overcome the impact of spatial uncertainty, we can actually exploit this unique characteristic of UWSN. 126 Chapter 6 Tone Self-Multipath Thousands have lived without love, not one without water. — W. H. Auden 6.1 Introduction The underwater acoustic (UWA) channel amplifies several traditional physical layer problems like, for example, large multipath spread as acoustic propagation is five orders of magnitude slower than RF [Cat90]. Large multipath spreads lead to wider inter- symbol interference (ISI) which is a big issue for coherent, phase-based communication. A significant amount of research has been done to tackle multipath in the RF domain. Techniques like adaptive equalization and error correcting codes can compensate for ISI-related distortion. Other techniques actually benefit from multipath by accumulat- ing energy from multiple paths using RAKE receivers [PG58]. Time reversal mirror (TRM) is another mechanism that cancels multipath related ISI [KHS + 98]. However, until the early 90’s researchers widely believed that coherent communica- tion would not be viable for the UWA channel due to its time variability and dispersive multipath [Cat90]. However, Stojanovic showed that coherent communication in UWA was possible by using a powerful receiver that combined adaptive decision feedback 127 equalization with a phased-locked-loop [SCP94]. Similarly, TRM has been demon- strated as a possible solution to handle multipath for an underwater acoustic environ- ment [SHK07]. Thus, we conclude that significant research exists for efficiently dealing with multipath as the physical layer issue of inter-symbol interference. Commercial underwater acoustic modems today primarily target long-range (multi- km) point-to-point communication (for example, the Benthos modem [Ben]), and recent research efforts have targeted higher throughput using coherent communica- tion [FGS + 05, SS06]. Since energy is a constraint in many underwater networks (for example, stationary networks, or those using battery powered gliders [RARF06])), some modems employ special wake-up tones and low-power wakeup receivers to activate modems [Ben, WYH06], and recent work has shown how to integrate tones with a low- power MAC protocol [SYH08]. Tone-based wakeup is also important for wakeup after long-duration sleep [RYC + 05, LYH06], or triggering more sophisticated data reception algorithm [Ben, FGS + 05, WYH06]. Replacing full packet reception with tone detection (detecting energy on the channel) is a very effective way to save energy since it can be done both very quickly, as seen in low-power listening with radios [EHD04, PHC04]; and very efficiently with dedicated hardware, as seen in radio-based pagers and recent underwater modems [WYH06]. While a problem for data communication, multipath can be crippling for tones. When tones are used for coordination, echos reflecting from stationary objects cause the transmitter to hear self-reflections of their tones. If tones indicate contention and their absence indicates a channel clear for transmission (as it does for T-Lohi MAC protocol [SYH08]), self-reflections will prohibit all communication, since a sender will always mistake the echo for another contender! We explore this scenario in more detail 128 in Section 6.1.1 where we use the T-Lohi MAC protocol to illustrate tone-based self- multipath. We are the first to identify the problem of self-multipath for low-power underwater acoustic communication. The challenge of self-multipath is that a transmitter will always interfere with itself. We show that this challenge provides the means to address the problem: senders can and must learn their self-multipath patterns. This goal is difficult because of the unique challenges of underwater sensornets. First, current data-focused multi-path techniques do not directly apply because we are concerned with self-multipath, not sender-receiver multipath. Also wakeup tones are simple signals and are therefore not amenable to sophisticated coding. Finally we must allow for long acoustic propagation times and reflections to hundreds of milliseconds. In Section 6.2 we describe Self-Reflection Tone Learning (SRTL) where we use Bayesian techniques to learn the channel state. In addi- tion to directly applying to underwater MAC protocols, SRTL’s ability to estimate the number of reflecting surfaces in the environment also indicates the sparsity of the chan- nel. This information may assist optimizations that exploit sparsity in managing the complexity of multipath in data reception [LP07]. Since wireless channels are very difficult to model or simulate, we demonstrate the effectiveness of our approach with experiments using an acoustic modem in three differ- ent physical environments in Section 6.3. We establish that what SRTL learns matches to physical-world predictions through tests in a completely controlled anechoic chamber (Section 6.3.2). We show that SRTL copes with reasonable amounts of noise through tests there and in a less controlled laboratory, and with simulations that let us vary noise precisely (Section 6.3.3). Our controlled tests require in-air acoustics; underwater tests are currently underway. Finally, we show that SRTL can track changes to the envi- ronment (Section 6.3.4), how environmental parameters and their estimates affect the 129 Transmitting node A Receiving node B Self Multipath Receiver Multipath d h Figure 6.1: Self and receiver multipath shown for a single reflecting surface at a distance h from transmitting node A. Similar multipath will occur for each reflecting surface in the environment. accuracy of algorithm (Section 6.4.1 and 6.4.2 ), and finally that SRTL remains effec- tive with low-complexity, low-memory implementations (Section 6.4.3) The contributions of this work are to identify self-reflections as a new problem posed in low-power, high-latency underwater communication; to show how Bayesian learning with SRTL can mitigate this problem; and finally to show that it works experimentally. We believe this approach will be essential to energy conserving underwater media access protocols, and also useful for long-duration sleep, and may apply to the broader problem of multipath in high-latency channels. 6.1.1 Impact of Multipath We now show how tone echos result in self-multipath that cripples the T-Lohi protocol. Figure 6.1 shows multipath occurring between two contenders, A and B. Tone sent by A reaches B directly and, a little later, via a surface reflection. This regular, or receiver-multipath, will cause multiple tones to be detected at B. From the perspective of T-Lohi, the additional tones increase the contention count. However, as our simulation and analysis of reservation period show the duration of reservation period does not increase significantly with network density (Chapter 5.5). 130 Therefore, although the throughput decreases due to slightly longer duration before a packet is sent, T-Lohi will successfully contend the medium to send data. Other than the multi-path reflections received at B there is also the echo that A receives of its own tone; this reflection causes what we call self-multipath as it inter- feres at the transmitter(Figure 6.1) understanding of the contention status. As opposed to traditional multipath that results in data interference at receivers, self-multipath is essentially echo-interference amplified in the acoustic channel due to large propagation delays. Moreover, while the multipath spread (the maximum delay between direct and multipath signal) of receiver-multipath is proportional to the difference in the path taken by a direct and reflected signal (typically 10s of msec for 500m range) self-multipath is proportional to just the delay to a reflecting surface (and thus larger than just the difference, typically 100s of msec for 500m range). This self-multipath breaks T-Lohi MAC in such a way that contending nodes are never able to transmit data. This is because a self-reflection 1 for a contention tone sent in any contention round will result in an echo-induced tone detection. Since contenders transmit data in T-Lohi only when no other tone is detected, even a single reflecting surface, results in a contender always hearing an echo tone that it interprets as another contender. Thus a contender will always backoff and never be able to transmit data. While one could work around this problem, perhaps by timing out after non- convergence, or by sending information with the tone, these approaches would increase energy consumption. Instead, we next show how Bayesian techniques allow contenders to learn about self-reflections, then choose to ignore them. 1 We refer to each echo as self-reflection, distinct from the broader protocol-interference concept of self-multipath. 131 6.2 SRTL: Learning to Ignore Echos We now introduce the Self-Reflection Tone Learning (SRTL) algorithm, our approach to manage self-reflection. We first give an overview of SRTL, then review Bayesian learning, the theory we draw upon (Section 6.2.2). We then cover SRTL details: what it observes about the channel, sources of error in those observations, and how these factors come together. 6.2.1 Overview of SRTL The goal of SRTL is to learn about self-reflection and allow higher-level protocols (such as T-Lohi) to distinguish between echos, random background noise, and tones sent from other contenders. The key intuition behind SRTL is that echos are deterministic and repeatable relative to a transmission, while noise and tones from other senders are independent and uncor- related. An example of this observation is shown in Figure 6.2. When a node transmits a tone that reflects off other surfaces (A and B in Figure 6.2), the echos return to the sender with the same delay after each transmission (Figure 6.2). The tone receiver will also trigger in response to ambient noise (as shown by lightening bolts in Figure 6.2), or other transmitters, but these triggers occur independently of transmission times (com- pare location in Figure 6.2(a) and 6.2(b)). SRTL can therefore learn repeated echos using Bayesian learning techniques (reviewed next in Section 6.2.2). SRTL makes two assumptions: echos are repeatable, and other triggers are indepen- dent. Since echos are dependent on the physical placement of nodes and reflecting sur- faces, they will be repeatable in a static network. We confirm both of these assumptions, showing in Section 6.3.2 that the echos we observe correspond directly with the phys- ical environment. We tolerate minor variation in echo delay by discretizing responses 132 Surface A Surface B 1 2 3 4 5 N Sample period noise (a) First Sample: Two echo and one noise-triggered tone detection. Surface A Surface B 1 2 3 4 5 N Sample period noise (different location) (b) Subsequent Sample: Echo locations repeat, but noise at different loca- tion. Figure 6.2: Key idea behind the sample collection process for SRTL algorithm: A node sends a tone and waits for a Sample period. Bin location of echos repeat but that of non-echo detections do not. into bins; in Section 6.4.3 we show our choice of bin size is reasonable and that we tolerate responses on bin edges. Although SRTL is optimized for a static network, we show in Section 6.3.4 that it can accommodate environmental change due to drift or slow movement. The second assumption is that other (than echos) tone triggers are independent of transmissions. Most observations confirm that underwater noise is uncorrelated and 133 random, although the noise distribution is different for shallow and deep water chan- nels [Cat90, Sto03]. We have confirmed this observation in the experiments at the Marina del Rey harbor. It is possible that higher-level protocols or applications would create synchroniza- tion, either intentionally or accidentally [FJ94]. While some variations of T-Lohi inten- tionally synchronize transmissions similar to slotted ALOHA, MAC protocols in general (and the variant of T-Lohi that we use) can explicitly randomize transmissions to guaran- tee protocol-level synchronization does not occur, and thus concurrent application-level transmissions are desynchronized at the MAC-level. Learning about echos is not the only approach to identify self-reflection. We con- sider related work in detail in Chapter 2, but two alternate approaches are to use signal processing techniques to correlate echos and transmitted signals for disambiguation, or to code sender identification in tones. For example, Rake receivers handle data multi- path [PG58] using chip-correlations, but because its complexity is proportional to delay, it doesn’t easily scale to high-latency acoustic communication. Similarly Girod and Estrin use a similar approach in acoustic localization [GE01]. Coding sender identi- fication in a data packet can also solve the issue of self-multipath. Using data pack- ets, however, precludes the energy benefit of tone wake-up. Alternatively, we could use time-based codes, pulsing the tone on and off. However, to reach low power, the low-cost tone wake-up circuit uses very simple analog electronics and so is unable to use sophisticated codes or provide bit synchronization guarantees. Also, to maximize sensitivity, the wakeup receiver requires long activation times (up to 5ms in the worst case). Thus, the requirement to minimize energy consumption precludes many alternate approaches. We therefore turn to learning, and next provide background about Bayesian inference as employed by SRTL. 134 6.2.2 Bayesian Inference Background Bayesian Inference is a well established approach that measures a probabilistic belief or knowledge regarding an event or hypothesis [Jay03]. (Note that probability here repre- sents belief in a hypothesis, not probability as frequency of occurrence.) We next briefly introduce the Bayes’ theorem and show how it can be extended to perform inference based on empirically collected data. The classical Bayes’ theorem as defined for a bimodal hypothesisH, incorporating current evidenceX is: P(HjX)= P(XjH)P(H) P(X) = P(XjH)P(H) P(XjH)P(H)+P(XjH)P(H) (6.1) We assume, in all cases, either the hypothesis holds (H) or is not true (H). The prior probability, P(H) is the confidence before considering current evidence, while P(HjX) is the confidence after observing X. P(XjH) is the conditional probability, the likelihood that event X implies that hypothesis H is true, while P(XjH) is the likelihoodX occurs even though the hypothesis is not true. P(X) is the probability of witnessing the new evidenceX under the two mutually-exclusive hypotheses. The factor P(XjH) P(X) represents the impact that evidence has on belief in the hypothesis. When evidence strongly indicates the hypothesis, this factor is large, but if evidence is inconclusive, perhaps because the environment is noisy, it will be small. This ability to incorporate evidence lends naturally to using Bayesian inference for interactive sensing. The Bayes theorem describes how a single observation modifies a belief. We build on work in landmark-based localization in robotics [Thr98] to use successive observa- tions to learn about the environment. This work uses robots with sensors that identify 135 landmarks and move or change their environment. Observations of landmarks represent evidence increasing belief in the current location; manipulation and movement change the environment and decrease belief. Changes in belief are scaled by models of accu- racy of sensing and actuation. Next, we describe how SRTL “senses” echos, and how we model its accuracy to scale the belief update, thus learning echo locations. 6.2.3 Sampling in SRTL To apply Bayesian reasoning, and inspired by work in robotics (Section 6.2.2), we must decide how to sense the environment, and how our samples correspond to our hypothe- sis. Each time we send a tone, we follow that transmission with a sampling period (as in Figure 6.2). The duration of this period is defined by the maximum range of our tone hardware. Since transmissions attenuate, we listen until any additional echos would be too faint to detect. To manage observations, we divide the sampling period into fixed-duration bins. For each bini, we track the hypothesisH representing belief that thei th bin corresponds to a self-reflection; we call such bins self-reflection (SR) bins. For each bin the hypothesis H, that a bin does not have echos, is simply complementary toH due to bimodality of the hypothesis space. After a transmission and the entire sampling period, we have an array of evidences, one sample per bin. Each sample can take on two values, either E i , indicating a tone detection (an apparent echo), orE i , indicating absence of any detection in that bin. 136 Bayesian learning has a rich mathematical background for estimating hypothesis by incorporating empirical data. While we currently model static nodes, the incre- mental Bayesian learning can also incorporate motion if an appropriate model is pro- vided [Thr98]. Thus we believe that a Bayesian learning approach is appropriate for learning tone-echoes. 6.2.4 Modeling Truth and Observations A transmitter’s observation corresponds to four possible real-world events: True echo detection, true silence detection, Non-echo Detection (ND), and Tone Cancellation (TC). The first two events are the accurate observations about the world from the point of view of the transmitter. However, we model the next two events because they represent error introduced into our observations. Non-echo detection corresponds to an incorrect observation where our tone detector triggers, but it is not due to an echo of our transmission. Channel noise is a common source of a ND. A ND can also be caused by a valid tone transmission from another node. We cannot distinguish between these sources of incorrect observations, but our approach can filter both out, since both are effectively random sources of noise (as jus- tified in Section 6.2.1). A second source of error is Tone cancellation. While unlikely, it’s possible that channel noise or another node’s transmission can interfere with and cancel reception of a tone. In these cases, we are unable to observe an echo that should be there. Missing tones also occur when a tone lies on the boarder of a bin as we discuss in Section 6.4.3. We model the event ND and EC with parametersp nd andp tc . These parameters are engineering estimates of the probabilities that these events will occur in for a given hard- ware and physical environment. We show in Section 6.4.1 that our algorithm tolerates a 137 wide range of parameters, so they need not be set exactly, but its estimate tracks reality best when the parameters closely approximate the actual probability of these events. 6.2.5 The SRTL Algorithm We now apply Bayesian learning to our system. We apply our algorithm in parallel to the hypothesisH andH for each bin, using the samples array observed in each sample period. For simplicity we stop using the subscript and describe a single bin. We refer to E and E as positive and negative evidence of a tone activation in that bin. We next describe how we initialize our algorithm, then how we substituteE andE intoX from Equation 6.1 to update, as we learn,Bel(H) which is our confidence in the hypothesis, then finally how we make decisions from our observation. Initializing the Algorithm We initialize the algorithm with Bel(H) init , representing our initial belief of the bin being a self-reflection (SR) bin. This value seeds the Bayesian algorithm and is the same for all bins. Subsequent samples update the current belief using the update equations we describe next. From the perspective of Bayesian learning, the initial value should reflect our assumptions about the environment, presumably from prior experiments for a given transmitter. Since in general we do not have such knowledge, we instead start with an arbitrary value of Bel(H) init = 0:3. The algorithm is largely unaffected by choice of initial value because it rapidly aligns with reality using experience; Section 6.3.4 confirms that we can quickly track environmental changes, and we have confirmed in experiments (omitted here due to space) that we are largely unaffected by the initial value. 138 We next explore how negative and positive evidence change our estimate Bel(H) posterior and mark bins as self-reflecting. Update for Negative Evidence We consider absence of a tone as negative evidence (E) that indicates a bin does not receive echos. We therefore update our estimate from Equation 6.1, replacing X with E. Also, we replace the termP(HjX) withBel(H) posterior (belief after incrementally incorporating current evidence) andP(H) withBel(H) prior (current belief incorporat- ing all prior evidence) to reflect the standard Bayesian inference terminology [Thr98]. Bel(H) posterior = P(EjH)Bel(H) prior P(EjH)Bel(H) prior +P(EjH)Bel(H) prior We next explore how this update equation for negative evidence is modeled using our parametersp nd andp tc . P(EjH) is the conditional probability for the event when no tone is detected, given that we know an echo tone should be detected. This event is essentially the failure of our tone detection hardware to be triggered in the presence of tone energy. Assuming that bin duration is small enough to allow only a single detection, such an event can happen only if the tone echo is canceled by noise or interference (the event TCjH, which reduces to TC since TC is independent of the hypothesisH). This assumption simplifies our modeling as we no longer consider the more complicated case when an additional tone detection occurs, in the same bin, after the first. Thus, we can now define: P(EjH) ´ p tc (6.2) 139 P(EjH) is the probability for the event when no tone is detected given that we know that no echo can occur in that bin. However, this knowledge does not rule out non-echo detections caused by noise or other transmitters. Thus the event can be described by the union of two disjoint events: no non-echo noise is detected (the ND event) union with the event that although wake-up-triggering noise could have been detected it was canceled (the ND T TC event). Noting that the above events are disjoint, we formulate the following probability: P(EjH) ´ (1¡p nd )+p nd £p tc (6.3) Finally, using the definitions from Equation 6.2 and 6.3 the update equation for negative evidence becomes: Bel(H) posterior = p tc Bel(H) prior p tc Bel(H) prior +(p nd p tc +1¡p nd )Bel(H) prior (6.4) This update is applied to our current belief Bel(H) prior to realize a new belief Bel(H) posterior when no tone is detected within that bin (negative evidence). Update for Positive Evidence Detection in any bin during a sample period is considered positive evidence (E) of the bin being a SR bin. Similar to the negative update, we update our belief according to Equation 6.1, replacingE for the update eventX. We next define the components of the update equation for positive evidence based on the input parametersp nd andp tc . 140 P(EjH) is the conditional probability for the event that tone detection occurs in a bin with known self-reflection. This probability is simply the complement ofP(EjH) ; Thus: P(EjH) ´ 1¡p tc (6.5) On the other hand,P(EjH) is the conditional probability for a detection occurring in a bin we know has no echos. This detection can, therefore, occur only because of wake- up-triggering noise (ambient noise and other contention tones) that is not canceled; the event ND T TC. Thus this probability becomes: P(EjH) ´ p nd (1¡p tc ) (6.6) Finally, using the definitions from Equation 6.5 and 6.6 the update equation for positive evidence becomes: Bel(H) posterior = (1¡p tc )Bel(H) prior (1¡p tc )Bel(H) prior +(p nd (1¡p tc ))Bel(H) prior (6.7) This update is applied to our current belief Bel(H) prior to realize a new belief Bel(H) posterior when a tone is detected within a bin (positive evidence). While Bayes works as described in theory, in our experiments we observed that long runs of consistent evidence would saturate the bins with perfect positive or negative belief (Bel(H) = 1 or Bel(H) = 0). We expect saturation occurs because of floating point rounding error. These equations are unable to shift from certainty, even in the face 141 Table 6.1: Payoff table used in determining decision threshold that maximizes payoff. Reality Decision SR bin not a SR bin SR bin (Ignore tones) 10 2 not a SR bin (Count tones) 1 9 of later contrary evidence. We therefore cap the belief for each bin at a maximum value of 0.999 and a minimum value of 0.0001 to avoid saturation. Identification of SR Bins We intend for the SRTL algorithm to work continuously in the background to our MAC protocol (or any other networking protocol using tones). In T-Lohi, we automatically get one sensing period for each contention round, so observations about the environment occur automatically as a side effect of MAC operation, incurring no additional overhead. We next evaluate how to translate these continuous observations into decisions about which bins represent self-reflections and therefore should be ignored. To reach a decision, we bias our estimates by the payoffs of correct or incorrect deci- sions, then select the most profitable. We believe this decision threshold is reasonable, and we show later (Section 6.4.3) that our assumptions about the environment (p nd and p tc ) have much stronger influence on correctness. For the values given in Table 6.1, we can derive a fixed threshold of 0.45. Thus if the Bel i (H) is greater than 0.45 we will consider that to be a SR bin. We use this threshold in our experiments. In principle, one could adapt these values to the environment. 142 Table 6.2: Research questions asked about the merits of our Bayesian learning algo- rithm. Environment Questions asked about SRTL Section Controlled Uncontrolled Underwater Correctly ID known surface? 6.3.2 Yes n/a n/a Is robust to Noise? 6.3.3 Yes Yes Planned Can handle dynamic environment? 6.3.4 Yes n/a n/a How sensitive to parameters? 6.4.1 Yes Yes Planned What’s the impact of discrete bins? 6.4.3 Yes n/a n/a Selecting Bin Duration Our algorithm uses fixed size bins; bin granularity is one factor to the sensitivity of our algorithm. In practice, bin size is limited by hardware. As a lower bound, our micro- controller has a millisecond level clock granularity, and interrupt debouncing causes a 2ms delay between successive tone detections. We therefore set bin size conservatively at 3ms in both simulations and experiments. 6.3 Experimental Evaluation of SRTL We now evaluate SRTL through experiments, both in the controlled setting of an ane- choic chamber and a less controlled open lab. Table 6.2 summarizes our research ques- tions, but our overall goal here is to show that SRTL can successfully manage echos. To do that, we first confirm SRTL’s conclusions are justified by the physical environment (Section 6.3.2), and evaluate how to copes with different levels of noise (Section 6.3.3). Finally we verify that the algorithm adapts to changes in environment, either due to movement of the node or other objects (Section 6.3.4). We begin by summarizing our experimental methodology. 143 (a) Transmitter Setup (b) Reflecting Surface Figure 6.3: Controlled Experiments: The Setup in the anechoic chamber. On the left was the transmitter/receiver setup and the right figure shows a reflecting surface whose location was varied in our experiments. (a) Lab/office test setup (b) Underwater test setup Figure 6.4: Uncontrolled Experiments: Uncontrolled experiments were performed at two locations. The lab/office location provided for in-air experiments, while the test setup off the docks in Marina del Rey harbor provided for underwater experiments. 6.3.1 Experimental Methodology We evaluate SRTL using our acoustic modem in different environments: a controlled, anechoic chamber and a less controlled laboratory, both using in-air acoustics. In addi- tion, we are currently carrying out underwater tests. We next describe details in common to the three environments we report here, how they differ, and bounds on our ground truth. 144 Details common to experiments We run experiments using the SNUSE acoustic modem [WYH06], hardware revision 2. We use tweeters for in-air tests, and hydrophones when underwater. The modem is driven by a custom data collection program running on a Mica-2 mote. The microcon- troller directly controls modem operation via I/O through a custom digital interface. Each experiment consists of 200 sample periods of tone transmission followed by echo observation. For each sample, the mica2 configures the modem to transmit a wake- up tone, then switches to tone-sleep where it is quiescent until woken up by a tone. We timestamp each tone reception on the Mica-2 with 1ms resolution, then compute delay between initial transmission and echo. We later map detection delay into a correspond- ing 3ms bin (Section 6.2.5). The SRTL algorithm currently runs in a host PC connected to Mica-2, although in principle it could run on the mote itself. After each transmission we record all tone triggers as positive evidence (E), and assume negative evidence (E) for all other bins. We then update SRTL belief estimates based on Equations 6.4 and 6.7. We set the sensing duration based on the maximum observed in-air range of our modem. We measure in-air range at 20m, so we anticipate reflections from objects up to 10m from the transmitter. We therefore anticipate echos arriving with up to 60ms delay (20m, with speed of sound as 343m/s at 24 ± Celsius). We conservatively extend sensing duration to 100ms after each transmission. We next describe details specific to our three experimental locations. Location-specific experimental details We carried out experiments at three locations, each providing us with different level of complexity in the reflective nature of the environment. 145 The first experiment location is an anechoic chamber at USC’s UltraLab Laboratory. The chamber is designed to absorb all RF radiation for controlled radio experiments, but it also provides a good acoustically neutral environment. When necessary we place a large metallic pan in the chamber to act as a reflecting surface (Figure 6.3). We measure the distance from our transmitter as described in Section 6.3.1. Since the anechoic chamber is designed to be reflection free, this configuration lets us confirm against a strong ground truth: the physics of the measured location of our reflecting surface. Our second environment is an office laboratory (Figure 6.4(a)). This location is much more complex, with multiple possible reflecting surfaces (walls, file cabinets, machine rack doors, etc.). We therefore observe more complicated channel response and so cannot provide firm ground truth. However, this more complex environment provides a richer level of noise and signal response. For both in-air tests (anechoic chamber and laboratory), the modem uses high effi- ciency, Motorola piezoelectric tweeters that were impedance matched for both transmis- sion and reception. Our final experimental environment explores underwater performance. We test at the docks in Marina del Rey harbor (Figure 6.4(b)). Our current experimental modems are not yet packaged for underwater use, so we operate them on dock, connected to a Benthos AT-18AT hydrophone lowered about 1m underwater off the docks. Our ini- tial underwater experiments were not successful because the wake-up receiver was too sensitive, receiving near-continuous activations from background noise. We have recal- ibrated it and are in the process of collecting additional data underwater. 146 Estimating ground truth We estimate ground truth based on the physical distance between transmitter and reflec- tor and compare this distance to measured echo distance (converted from measured echo delay). Both these measurements, however, have potential sources of error due to our modem hardware and the measurement process. The largest source of error in measurement of echo delay is the detection circuit of our modem. We time-stamp the transmit time of tone and detection time of echos to calculate the distance to reflector. Due to transmit side warm-up, the actual transmission time of the tone can vary by about 1ms. Similarly the actual detection time can vary by 2ms based on the strength of echo. We measure the physical distance with a HILTI PD-40 high-precision laser range- finder [HIL]. Accuracy is§1mm, so we believe error in the distance measurement is minimal. However, the most significant source of error is in our measurement process when we approximate the angle for the line-of-sight measurement between the piezo- electric crystal located inside the transducer and the reflecting surface. This measure- ment error is approximately§2cm and results in a corresponding delay error of about §0:5ms (with speed of sound as 343m/s at 24 ± Celsius). When comparing the ground truth to identified echos, we have to reconcile the above independent errors, in both measured distance and echo location. In the figures, we show the tolerance region that accounts for the worst case error in each measurement. The identified location can safely be considered to match the ground truth if these regions overlap. Thus, due to the resulting overlap of the error bounds, the tolerance region varies on a case-to-case basis for each measurement. 147 (a) Reflecting Surface at 3.36m (b) Reflecting Surface at 4.11m Figure 6.5: Experimental results showing the CDF of 200 samples and SRTL response with objects at known location, adjusted for measurement error (shown as the shaded area with dashed boundary). 6.3.2 SRTL Correctness We first seek to confirm that SRTL can correctly identify the location of a known reflect- ing surface: does our algorithm and experimental setup match the physical configuration of the world? Since this experiment needs knowledge of the ground-truth, we use the anechoic chamber to perform controlled experiments. For this experiment we place a reflecting surface perpendicular to the piezoelectric tweeters, measure its distance and compute the expected delay. We take several measurements (as described in Section 6.3.1) with the surface at a particular location. Since there is only one reflecting surface in the room, our algorithm should only identify the bin corresponding to the measured distance. We then compare SRTL’s estimate with our prediction from the physical distance. Figure 6.5 shows the result of our experiments for reflectors at two different dis- tances. Each figure combines three different values: prediction from physics, all obser- vations, and the SRTL belief distribution. The dotted box indicates the prediction from our distance measurements, including estimated error. The solid black line represents 148 (a) Uniform noise source over 500ms (b) Uniform noise source over 100ms Figure 6.6: Results of a controlled experiment (object at 3.87m) showing the CDF of 200 samples and SRTL response, with varying levels of noise. the cumulative distribution function of delay values for all the samples considered by SRTL, measured against the left axis. Finally, solid blue bars represents the belief dis- tribution (Bel i (H)) for each bin in the 100ms sensing period (we show only the first 25ms and omit the remainder since there is no belief there). Bin indices are given on the top axis. Figure 6.5(a) shows the result of our experiment with the surface measured to be at 3.36m from the transmitter. The sample CDF shows that nearly all samples are received at a delay of 17ms, which corresponds to the fifth bin. SRTL is able to identify this bin with complete confidence and we can see that the identified bin lies within the error bounds of the physically measured surface location. Figures 6.5(b) shows results from the same experiment with the surface at 4.11m (the seventh bin). We observe that the bin identified by SRTL matches the location indicated by the CDF and predicted by our range measurements. From these experiments we conclude that we can completely explain SRTL perfor- mance under known conditions; SRTL places known reflections in their correct bin as predicted by the physical setup. 149 (a) Uniform noise source over 500ms (b) Uniform noise source over 100ms Figure 6.7: Laboratory experiment empirically find two stationary reflections with two levels of artificial noise. 6.3.3 Robustness to Noise Although we verified that SRTL works as expected in perfect conditions, we also care about performance in the face of environmental noise. We investigate this question by adding an artificial noise source to our experiments. This noise source is a second modem that transmits tones that trigger non-echo tone detections at our original sender (on which SRTL algorithm is running). Our artificial noise source transmits tones repeatedly, with inter-transmission times chosen uniformly randomly within a fixed interval. We then vary this interval to adjust the degree of noise, with smaller intervals causing greater (more frequent) noise. Since timing of noise is random, we expect SRTL to ignore such noise and still be able learn detections from known surfaces. (We select this noise model to provide simple, controlled tests. Exploration of richer noise sources is an area of future work.) We performed our experiments with different levels of noise in both controlled and uncontrolled environments (anechoic chamber and laboratory). Figure 6.6 shows the result for the controlled environment of the anechoic chamber. 150 The presence of a gradual slope in the sample CDF (the solid black line) indicates the presence of noise. The slope of the CDF indicates the level of noise; a steeper slope indicate greater noise, as can be seen comparing Figures 6.6(a) and 6.6(b). However, we observe that SRTL identifies the correct bin of the reflector even with substantial interference. We looked at several noise levels, Figure 6.6(b) shows the case where there are, on average, two noise triggers in each sample period for the single true echo. However, because noise is randomly distributed, SRTL can suppress it and learn the true echos. We conclude from this experiment that while SRTL will learn its environment, it will not be fooled by competing contenders (in the case of T-Lohi) or some levels of environmental noise. We anticipate that real-world noise and multipath will be more complex, so we next reproduce this experiment in the uncontrolled environment of our lab where bin identi- fication is more challenging for SRTL. Figure 6.7 shows the result of our experiments. Over many experiments with and without noise (not shown) SRTL detects two likely echos in bins 4 and 8. Although we do not know the ground truth surface, these indicate empirically the presence of two reflecting surfaces. Again, we see that the algorithm consistently identifies these as self-reflection bins and is able to see through and suppress random noise. Our underwater experiments provide a final evaluation of noise. Our preliminary tests in the Marina show tone triggers every few milliseconds. Assuming a single reflect- ing surface, this corresponds to 50 false triggers per sensing period, far more noise than regular reflections. SRTL is unable to track known surfaces with this level of noise. We are currently working to better characterize SRTL tolerance, and to improve our wakeup receiver’s ability to filter brief noise in hardware. 151 From the above results conclude that SRTL can tolerate random noise up to at least two false triggers per true echo, and shows the need to further characterize the limits of noise tolerance. We do not characterize further due to hardware limitation, but explore greater noise tolerance using simulations in Section 6.4.2. 0 2 4 6 8 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 0 samples 0 5 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 5 samples 0 5 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 10 samples 0 5 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 15 samples Decision threshold (a) p nd =0.4,p tc =0.4 0 2 4 6 8 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 0 samples 0 5 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 5 samples 0 5 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 10 samples 0 5 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 15 samples (b) p nd =0.4,p tc =0.1 0 2 4 6 8 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 0 samples 0 5 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 5 samples 0 5 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 10 samples 0 5 0 0.2 0.4 0.6 0.8 1 Bin ID Confidence 15 samples (c) p nd =0.1,p tc =0.4 Figure 6.8: SRTL response to change in location of reflecting surface from 3.36m (bin 5) to 4.11m (bin 7) with different input parameters. 152 6.3.4 SRTL in a Changing Environment Most underwater environment change, either due to tides or currents, precession on an anchor, or movement of human artifacts or fish. We therefore next wish to evaluate how well SRTL adapts to a changing environment. Properly configured, Bayesian learning can track changes in belief so we expect SRTL to track changes in the environment successfully. To investigate SRTL response to environmental changes, we return to the anechoic chamber. We place a reflecting surface at a known location (at 3.32m, corresponding to the fifth bin). We then take twenty consecutive samples at that location to train the SRTL algorithm to identify that location with maximum confidence. We then move the reflecting surface to a different location (4.11m, corresponding to bin 7) relative to the transmitter. We then observe how SRTL’s belief distribution (about which bins are SR) evolves as it collect additional observations. We expect SRTL to track the changed environment and identify the new bin after incorporating a few samples. The decision threshold is set using the mechanism described in Section 6.2.5. Figure 6.8(a) shows how SRTL’s belief changes as it takes more observations. Ini- tially SRTL is certain that bin five is the surface, but after 5 samples (the second figure from the left), this believe begins to waver. At 10 samples it has begun recognizing bin 7, the new location, as a self-reflection, although it still remembers the old location. Finally, after about 15 samples, SRTL has nearly completely shifted its understanding of the environment. This experiment demonstrates that SRTL will adapt to changes in environment. In the next section we consider how SRTL parameter settings affect this convergence time. 153 (a) p nd =0.4,p tc =0.5 (b) p nd =0.7,p tc =0.7 Figure 6.9: Impact of parameter choice for an uncontrolled environment. 6.4 SRTL Parameter Sensitivity The previous section describes experiments that show SRTL works well, even with noise and environmental change. SRTL has several parameters that affect its opera- tion, including estimates of ND and TC events and choice of bin size. We next evaluate how sensitive SRTL is to choice of these parameters. 6.4.1 Estimates of Observation Errors SRTL’s learning algorithms takes two parameters,p nd andp tc , that are used to adapt its belief to new observations. In Section 6.2.4 we describe how these parameters model our estimates of observation errors. To understand how the parameters affect SRTL, we continue our experiment with changing locations of reflecting surfaces first for a controlled environment. We then reexamine parameter setting for the more complex and uncontrolled laboratory environment. 154 Parameter Choices: Controlled Dynamic Environments We next repeat the experiments from Section 6.3.4 where we move the reflective surface. Figure 6.8 shows the result for three different estimates of error probability. We make four observations from these results. First, in every case SRTL is able to adapt to the changed environment and identifies the new bin location (after 15 samples). However, parameter choice affects the rate of adaptation and the transient behavior. For moderate and equal parameters (Figure 6.8(a), p nd = 0:4, p tc = 0:4), beliefs change relatively slowly. However, when tone cancellation is lower than non-echo detection (Figure 6.8(b),p nd =0:4,p tc =0:1), confidence in the old bin decays quickly. So after 5 samples, neither bin is identified as self-reflective. We explain this result because a lowp tc means that cancellation is unlikely, so negative evidence (absence of tone triggers in the old bin) is quickly accepted. In Figure 6.8(c) we see the opposite case, with low non-echo detections but likely cancellation (p nd = 0:1, p tc = 0:4). Now new evidence is rapidly accepted (bin 7 is detected after five samples), but old assumptions decay slowly. Here, a low value ofp nd indicates a low-noise environment, so new triggers are quickly taken as echos. Overall, we conclude that highp tc values act to damp the response of our algorithm to missed reflections, while large p nd dampens response to new bins. We believe that keeping a moderate value ofp tc is essential, because our lab experiments that show that sometimes there are several consecutive absent echos in even regular locations. Also, a moderate value of p nd helps suppress short, transient noise. We therefore suggest moderate and equal values of both parameters (like, p nd = 0:4, p tc = 0:4). We expect to review these settings as we gather additional data from underwater experiments. Finally, detection theory allows one to trade certainty of detection for time; we observe this tradeoff in the value of parameters. Thus, a higher value of p tc confirms 155 removal of a reflecting surface slowly, but also increases accuracy as we now do not react to transient changes. Parameter Choices: Uncontrolled Environments We have shown that balanced values are appropriate in a controlled setting. We next evaluate their impact in the uncontrolled laboratory setting. To do so, we return to our laboratory experiments with relatively complex multipath. We performed experiments for a wide range of parameters sets, including the ones shown in Figure 6.8. Figure 6.9, however shows the results for two additional sets of parameters as they depict an inter- esting case where the choice of parameters becomes important. Figure 6.9(a) shows response with moderate parameters (p nd = 0:4, p tc = 0:5). Here the algorithm identifies just a single bin (bin 4) as self-reflective. However, the same observations but with (p nd = 0:7, p tc = 0:7) causes SRTL to also identify an additional bin (bin 11, Figure 6.9(b)). Which parameter set is best? The CDF of the observations suggest that both potential reflections are consistent reflection and above the noise threshold. Thus it appears that the second parameter set with high values is a better choice. However, a more detailed look at the data shows that detections in bin 11 were consistent at the beginning of the experiment, but after 160 of the 200 samples, detections in that bin became infrequent (just five more times in the next 40 sample periods). These different observations suggest that multipath changes even in an apparently static room over a few minutes. (This change does not show up in the CDF because it sums the entire experiment.) We conclude therefore that the moderate parameters are a better choice since they correctly dismiss the second bin by the end of the experiment to reflect the change in environment. 156 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Simulated noise probability for each bin False Positive Ratio p nd =0.1 p nd =0.4 p nd =0.7 p nd =0.9 (a) False Positives 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 −0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Simulated probability for tone cancelation in each bin False Negitive Ratio p tc =0.05 p tc =0.15 p tc =0.3 p tc =0.5 (b) False Negatives Figure 6.10: Fraction of false positive and false negative as simulated noise and chosen noise estimates vary. Error bars show 95% confidence intervals. (They-axis starts below zero to show values along the origin.) 6.4.2 Parameter Alignment with Environmental Conditions Experimental results have a couple of limitations. First, we are limited by our hardware in how much noise we can introduce into our system 2 . Also, it is not possible to vary or control the tone cancellation event in an actual experiment. To study both types of noise, we simulate the algorithm with artificial noise, allowing us to explore arbitrarily high noise levels under controlled conditions. Our goal is to understand what levels of noise cause SRTL to fail, and how SRTL behaves when our noise estimates (p nd and p tc ) differ from the actual amount of noise. We start by describing the simulation methodology used in our experiments. Simulation Methodology In our simulations we keep the environment parameters for the probability of the ND and TC events different from their estimates provided as parameters to the SRTL algorithm. 2 Our modem, used to generate random tonal noise, has a hardware delay/warm-up of about 25ms before a tone can be sent. Thus the rate of noise generation becomes restricted. 157 Thus we vary and control the environment by changing these simulation parameters and evaluating our algorithm with fixed input parameters. We simulate 30 bins, and fix six of them as self reflecting (SR) bins. We measure two major metrics of interest: fraction of false positives (the fraction of non-SR bins classified as SR) and false negatives (the fraction of SR bins classified as non-SR) in our algorithm. Our decision threshold is 0.45 based on the decision payoff scheme shown in Table 6.1. We initializeP(H) init to 0.3 in our experiments and make a decision after 30 samples. False Positives To observe false positives, we vary simulated (wake-up tone triggering) noise in the environment. False positives in SRTL indicate when the algorithm fails by incorrectly identifying a bin as self-reflecting. Our understanding from Section 6.4.1 indicates that higher values of simulated noise will correspond to higher false identifications. In Fig- ure 6.10(a) we fix both the simulated and algorithm parameters of tone cancellation probability at 0.05. We can, therefore, observe the response of the algorithm to just the environment’s (simulated) noise triggering probability. We then vary noise (thex-axis) and observe the fraction of false identifications for different values ofp nd . Figure 6.10(a) shows that the fraction of incorrectly identified bins increases as the simulated environmental noise increases. At greater than 80% noise, nearly all bins are incorrectly identified as self-reflections; SRTL cannot tolerate this level of noise. However, this exact threshold is a function of our noise estimate p nd , because larger estimates make SRTL more skeptical that triggers indicate self-reflections. In general, SRTL performs reasonably as long as the estimate is at least as large as true noise, with some leeway when noise is low. Thus, withp nd = 0:1, SRTL handles up to 40% noise 158 (the leftmost line), with p nd = 0:4, it works reasonably to 70% noise (the second line from the left). In fact, withp nd =0:9, there are only 10% false positives at 90% noise. False Negatives To observe false negatives we vary the simulated probability of a tone cancellation. False negatives in SRTL indicate when the algorithm fails to correctly identify a self- reflection. Our understanding from Section 6.4.1 indicates that a high probability of tone cancellation leads to a higher possibility that SRTL fails to identify a SR bin. In Figure 6.10(b) we fix both the simulated and algorithm input values of noise detection probability at 0.4. We can, therefore, observe the response of the algorithm to just the environment’s (simulated) tone-cancellation probability. We then vary the tone can- cellation probability (the x-axis) and observe the fraction of failed identifications for different values ofp tc . Figure 6.10(b) shows that the fraction of SR bins that are SRTL fail to identify (false negative) increases with the simulated tone cancellation probability. Just as with false positives, performance is best when the estimate (p tc ) is close to actual tone cancel- lation probability, but at low levels SRTL will tolerate noise two or three times that estimate. The main difference is that SRTL is less forgiving of tone cancellation than noise. Because there are few echos, moderate cancellation makes them difficult to iden- tify. Fortunately, as it is rarer for interference to completely suppress channel energy, so cancellation is rarer than non-echo noise. Thus we conclude that SRTL can tolerate a wide range of noise in the environment, provided that corresponding estimates (bothp tc andp nd ) are reasonable approximations. Moreover, over-estimating the actual observation errors has a lesser penalty than under- estimating these parameters. 159 0 5 10 15 20 25 30 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Samples Considered Hypothesis confidence P nd =0.1, P tc =0.4 P nd =0.05, P tc =0.05 P nd =0.4, P tc =0.4 P nd =0.4, P tc =0.1 Decision Threshold Figure 6.11: SRTL belief for a single bin being SR for a worst case scenario of reflection time alternating between adjacent bins. 6.4.3 Impact of Bin Discretization SRTL uses discrete bins to track belief and provide efficient and low-complexity opera- tion even on mote-class devices. Our final question is to explore how bin size affects our algorithm. We have three concerns: bins that are too small may disperse observations, bins too large may cause echos to hide real tones from other transmitters, and even with correctly sized bins, tones might fall on the border between two bins. Our observations about system sensitivity (Section 6.2.5) limit our bins to at least 3ms, and we have not seen drift by more than a few ms, so we believe our bins are neither too large nor too small. We next look at echos that lie on the border to investigate the worst case impact of using such a low-complexity, low-memory implementation. For a reflecting surface exactly on the edge of two bins, minor observation jitter (hardware delays, software locks, clock granularity, or simply very slight motion of either the node or reflecting object) can easily move observations in either direction. 160 To simplify evaluation, we consider the worst case scenario: each sample produces detection at an alternate of two adjacent bins. We emulate this scenario by artificially providing such a sample input to our SRTL algorithm. Since the bins are symmetric, belief in each bin is identical. Figure 6.11 plots the belief in one of the adjacent bins as more samples are incorporated by the algorithm using SRTL for four different parameter sets. We see that combining a low value of tone cancellation probability with a higher value of noise detection prevents both bins from ever being declared as SR bins (p nd = 0:4,p tc = 0:1). This result is because low value of p tc implies any confidence is lost quickly, while a higher value of p nd requires long time for new identification to be made. Flipping these parameters results in an almost immediate identification of both bins as SR bins (p nd = 0:1,p tc = 0:4). This choice is very aggressive and thus inappropriate as a small value of p nd makes for a higher possibility of false positives as discussed in Section 6.4.1. A more appropriate choice of parameters is possible with both values at a moderate and equal value such as 0.4 (as suggested in Section 6.4.1). In this case the algorithm requires just six samples (three positives for each bin) before successfully identifying both as being SR bins. The SRTL input parameters provide us with a fine-tuning-knob to adapt our algo- rithm to the need of our environment. The results in this section and Section 6.4.1 suggest setting these parameters is more critical to the performance than the decisions threshold, since belief increases exponentially with positive reinforcements (so a thresh- old of 0.45 and 0.8 might be reached in just one additional sample). Our results also suggest that using moderate and equal values (for example, 0.4 and 0.4) for both noise- detection and tone-cancellation events provides good performance. This combination 161 damps both the increase in confidence for a new bin as well as the decrease in confi- dence for bin for which reflections come and go. 6.5 Summary In this chapter we first identified the issue of tone self-multipath unique to the large propagation delays of acoustic networks. We then used T-Lohi as an example to show that this delay can significantly reduce the throughput of our MAC protocol. We then introduced a Bayesian learning algorithm, Self-Reflecting Tone Learning (SRTL), that can be used to learn-and-ignore these self-multipath or echo tones. We performed exper- iments to verify that our algorithm is correct, robust to noise, and can adapt to a dynamic environment. The problem of self-multipath is uniquely significant in the UWA channel due to the large propagation delays. We were able to observe its crippling affect while implement- ing the wake-up tone based T-Lohi MAC protocol. This issue was then resolved using the understanding we have developed during the design of latency-aware protocol in the previous chapters. Thus not only can latency-awareness help identify new and unique problems, all such problems can be dealt with provided we have a firm understanding of the impact of spatial uncertainty in the context of underwater networking protocols. 162 Chapter 7 Future Work and Conclusions The shattered water made a misty din. Great waves looked over others coming in, And thought of doing something to the shore That water never did to land before. — Robert Frost We close our dissertation by listing some short-term future work to our efforts in this dissertation and also discuss some open research questions in the broader field of UWSN. We then make some concluding statements about our work. 7.1 Short-Term Future Work We first present some incremental future work that is not considered part of this disser- tation. We first talk about future work regarding implementation of protocols. Next we describe some future directions to consider for our protocols and propose extensions to the T-Lohi protocol for these future research issues. We finally present future research direction for the SRTL algorithm. While we have simulated and verified both our time-synchronization and MAC pro- tocols using mathematical analysis, actual experimental results for both TSHL and T- Lohi have been lacking. One significant reason for why experimental evidence have been absent because a low-power, low-cost, sensornet oriented acoustic underwater modem was not available. Our group has been developing a hardware and software stack to enable to low-power acoustic communication using the SNUSE modem [WYH06]. 163 While this modem currently has demonstrable short-range data reception, a new version is being developed that can be reproduced in significant numbers to setup an underwater sensor test-bed. An immediate opportunity then would be implement our protocols on this acoustic test-bed. Such system-level protocol testing is essentially to truly under- stand the unique challenges of underwater networks and verify if we captured all the challenges involved. If not such testing will allow us to, similar to our identification of self-multipath as an issue, discover new challenges in underwater deployments. Underwater channels have many sources of noise. For data packets, noise typically causes packet loss due to corruption; for tone transmission noise can result in false tones or tone loss. We have not considered packet loss in this dissertation, however one could add an ARQ mechanism to our protocol (for example, see Stojanovic [Sto05]). Explor- ing in detail the impact of such data packet loss on both TSHL and T-Lohi protocols is an important future work direction. Another aspect of future work is to look at multi-hop extensions to our protocols. We expect to explore multi-hop options and effects, including transmit power control, data and control pipelining, and hidden-terminal effects. For T-lohi, specifically, we believe that considering pipelining of data across two hops, while interleaving transmission of each hop to not intersect in space-time will provide an interesting extension to the single- hop exploitation of the space-time uncertainty. We are currently evaluating the effects of noise on tones (both false positive tones and tone loss), which are an important and unique aspect of our T-Lohi protocol. We believe that false detection of tones simply prolongs the reservation period as tone indi- cate of contention (the issue of false detection of echoes was extensively dealt with in Chapter 6). For low to moderate rates of false detection, T-Lohi will work correctly albeit with lower throughput. Numerical analysis of reservation period using the T-Lohi 164 contention resolution mechanism shows that with a 50% chance of a false tone (in each CR), reservations periods require less than two additional rounds to converge (from 3.5 to 4.9 rounds per RP), although with nearly constant false tones convergence becomes unacceptable (for example, 95% chance of false tones requires 32 CR per RP). Tone loss, on the other hand, could result in incorrect reservations. However, data-data colli- sions will occur only if the tone is lost in the final round of contention with exactly two contenders (in other cases, contention continues in future rounds due to other tones). For these reasons, we expect that the performance degradation due to moderate channel noise will be small. In addition, the SNUSE modem targets relatively short range com- munication (500m or less), so some forms of channel noise will be less likely than with longer range modems [HLS + 06]. There can be several incremental improvements to T-Lohi that can potentially mit- igate the impact of issues described above and increase T-Lohi’s throughput in a real- world, multi-hop network. One improvement would be to add additional information in the header of data packets. Since in T-Lohi all nodes have to wakeup and decode data packet header (to verify destination), the additional information can be used to make several improvements. For example adding FEC (forward error correction) information to packets will provide us the benefit of not requiring ACK to handle packet loss (saving costly, in both time and energy, retransmission) while distributing the FEC overhead. Similarly, adding information about your one-hop neighbor-hood can help build an effi- cient pipelining mechanism for multi-hop medium access. Adding the size of contention window for the next reservation period can also help save on crucial contention rounds and get faster data transmission. Another improvement to T-Lohi would be using a sin- gle reservation period to elect two or three contiguous data transmission based on the order in which nodes dropped out of contention. 165 Our SRTL experiments verified the correctness of the algorithm in identifying the noise sources. SRTL is designed to make a T-Lohi implementation robust to the self- reflection of tones. Thus combining SRTL algorithm with an implementation of T-Lohi would allow us to validate the working of both our MAC protocol and the algorithm designed to handle self-multipath with T-Lohi as the medium access protocol. While we performed initial underwater experiments for SRTL, the hardware sensitivity to noise in our underwater environment (a boating Marina) prevented these experiments to be successful. Thus a crucial next step will be to test SRTL in an underwater environment where the channel noise will be much different from our current in-air experiments. We believe that the locations identified by SRTL can be used to suppress receiver multipath by scaling down our count of detected contenders by SRTL’s count of SR bins. The intuition over here is that all nodes see approximately the same number of reflecting surfaces. Our algorithm can also be extended to other scenario where the physical layer’s handling of multipath is either absent (as it was for our low-power tone wake-up circuit) or not at a sufficient level from the protocol’s perspective. Exploring such opportunities is also part of the future work. 7.2 Long-Term Future Directions The field of underwater networks in general and underwater sensor networks specifically is relatively nascent. As such, we believe there are several future directions in which long-term research will be interesting and quite fruitful. Next, we briefly introduce each of these future areas. We specify why research in these areas is interesting and what possible directions within each area remain novel and worth an academic pursuit. 166 7.2.1 Underwater Acoustic Link-level Characterization A very important future area of research is to characterize the link-level packet- delivery performance of acoustic modems in an underwater, networked environment. Several research studies have empirically evaluated RF-wireless link-level reliability [ZG03, ABB + 04, ZK04, SKH04]. These studies tremendously increased the under- standing of the real-world performance of protocols. These studies provided deep insight that the less-than-expected performance of wireless protocols stemmed from a design assumption that the wireless channel has predictable reliability behavior. We, as well as many others in the UWSN field, believe that since the acoustic chan- nel is more complex than the RF channel [Pre06], its link-level reliability will also be difficult to model. Due to an incorrect simplification of link layer characteristics, many protocols designed for UWSN might fail to perform or achieve their full potential. Our understanding based on physical-layer studies indicated that such models will vary sig- nificantly with different environmental parameters such as depth, wind-speed, and shal- low vs deep water, among many others. As such, detailed studies that characterize the behavior of link-level packet reliability will preempt bad design choices and lead to development of robust underwater acoustic protocols. While existing long-range acoustic systems consider the use of FEC essential [Ben], a similar evaluation for short-range, sensornet style communication is currently unavail- able. Link-level measurement, as one example, will provide data to facilitate such an empirical evaluation, comparing the complexity/overhead of FEC with ARQ. 7.2.2 Incorporating Delay Tolerant Networking Ideas Bringing Delay Tolerant Networking (DTN) ideas to underwater networks is a promis- ing area of future research. We believe that the research in DTN will be complimentary 167 to our propagation-latency-aware research for UWSN. While DTN deals with network- level temporal disconnections, our work in UWSN focuses on point-to-point delays that introduces spatial uncertainty for what still remains a connected network. Underwater networks, however, can face temporary network disconnections. Such disconnections can be due to wave-breakages that introduce vertical columns of air bub- bles [Pre06], or node movement into the so-called “shadow-zones” that exist due to sound refraction through varying temperature gradient. Such loss of connectivity might fit in with the DTN architecture of store-and-forward, if the disconnection time-scales are large enough (in minutes). Another way where DTN methodology might fit in is with vastly variable channel reliability perhaps due to temporally continuous sources such passage of ships or a pod of whales that use similar frequencies as modems for communicating. In all these situations, it might be beneficial to store data and forward only during much improved channel conditions. Thus we believe that a synergistic approach combining the idea of DTN with protocols dealing with point-to-point delays is a promising future research direction. 7.2.3 Hybrid Underwater Networks Another open research area is the development of protocols for underwater networks with a hybrid of acoustic and other communication technologies. While acoustic com- munication is arguably the best choice for long-range, multi-access communication, alternatives for short-range and high data communications exist. Both optical and RF communications can provide a much higher data rate than acoustic communication at shorter (less than a meter) range. Having a network with hybrid communication capabilities opens many interesting possibilities (in fact one such network has been shown to work [VKR + 05]) As such, 168 employing acoustic as a long-range coordination or control channel is a viable solution for some network scenarios. Mobile nodes can use the acoustic channel to position aggregating nodes or cluster heads together for large data transfers. Another interesting scenario is the use of data-mules that can coordinate, using the acoustic channel, with a data extraction node to collect data rapidly in a large acoustic network. Finally, considering the recently standardize IEEE 1902.1 (RuBee) long-wavelength magnetic communication appears to be an interesting research area [IEE]. These mag- netic transmission are claimed to operate through water and also provide low-power operation (although at low, near acoustic data rates). While the data-rates and ranges might be similar to acoustic communication, point-to-point latency would no longer exist and the channel noise and complexity might be reduced. 7.2.4 Capture-aware Protocol Design Capture-aware protocol design is important for UWSN, more so than its previously explored benefits for RF wireless networks [DSK07]. As previously mentioned, there is a interesting parallel between the problems and techniques use for RF multihop wireless networks and problems for acoustic single-hops. For example the hidden terminal prob- lem of RF multihop networks is very similar in reason to the failure of carrier sensing in UWSN. In both cases the cause of failure is a transmitter-centric view; what is actually important is the view at the receiver. Using RTS/CTS to avoid hidden terminal has been shown to result in overly con- servative protocols that, because of not being capture-aware waste a lot of band- width [WWJ + 05, SKH06]. For underwater, due to the parallelism described above, this waste becomes more pronounced, as conservative MAC prevent capture-able pack- ets from concurrent transmitting over single hop. Exploring the impact of capture over 169 single-hop is an interesting directions. Thus, for example, while using a synchronized transmission scheme a good choice will be to select the closest next-hop. Such concur- rent transmissions succeed more frequently as nearby nodes receive the strongest signal first, leading to an automatic strongest-first capture for both concurrently transmitted packet. 7.3 Conclusion Development of acoustic underwater sensor networks (UWSN) has opened a new door for fine-grained in-situ observation of what lies beneath nearly 75% of our planet. The acoustic channel poses many challenges for underwater communication. While the physical layer challenges have been extensively researched, the unique challenges for network protocols have not been given a similar treatment. We however believe that the five orders of magnitude large propagation delay is the factor most significantly impact- ing network protocol design. This delay is ignored in existing terrestrial RF protocols and results in them not performing below expectations. This dissertation shows, using protocol design, simulations and experiments that by understanding the impact of propagation-latency induced spatial uncertainty is impor- tant. Such an understanding allows the design of protocols that not only overcome but actually exploit the unique nature of acoustic networks. We were the first to formalize the impact of propagation latency in the concept of space-time uncertainty. This concept essentially reiterates the importance of considering not just the packet transmission time (time uncertainty) but also the unequal delays (spa- tial uncertainty) to a receiver in any point-to-point packet transfer. With large propaga- tion delays this spatial uncertainty becomes as important as the time uncertainty which 170 is traditionally considered in terrestrial RF communications. We next use the under- standing developed from the space-time uncertainty concept to propose new protocols that are able to overcome and even exploit these propagation delays. We first identify a source of error that was ignored in existing RF sensornet time- synchronization protocols, but becomes important because of the large delay induced spatial-uncertainty during message exchanges. The error occurs due to ignoring the skew in clocks during the message exchange that is used to factor out propagation delay from time synchronization. We showed that under the high-delay of underwater net- works, this error becomes significant beyond 100m of range. We proposed a new pro- tocol, TSHL, that overcame this source of error by first estimating the skew using one way message exchanges. With knowledge of skew, the equations used to factor out propagation delay and compensate for the skew during the message exchange and arrive at propagation delay independent synchronization accuracy. This protocol development showed that latency-aware design of protocols is essential for migration from terrestrial to underwater sensor networks. For medium access, we observed that the space-time uncertainty implied that some of the channel is under-used with traditional collision prevention schemes. Thus when- ever a packet is transmitted, traditional time-uncertainty based approach advises block- ing all transmission during its existence in the network. Such conservative blocking, due to the large propagation delay in UWSN, results in a large volume of the channel in space-time that is not used. This unused region becomes larger with shorter packet transmissions. Along with a low-power tone wake-up circuit in our acoustic modem, we use this spatially redundant region to design T-Lohi,a family of medium-access protocol. In T-Lohi nodes send small tones that indicate their desire to contend for the channel. The unused region makes the simultaneous collision of these tones unlikely while also 171 providing the unique capability to detect and count other contenders in the network. Our protocol is therefore able to respond intelligently to the current load on the network. We showed, through simulations and mathematical analysis, that this contender counting ability allows data packet transmission in nearly constant time, independent of network density or load. Since the tones are sent using the low power wakeup-circuit, the entire process remain highly energy efficient. We therefore show that spatial uncertainty can actually be exploited to design a throughput- and energy-efficient MAC. We also identified how large-propagation delay can introduce the issue of self- multipath. We observed this unique aspect of acoustic multipath during implementation of T-Lohi. We observed that echoes from a node’s contention tone prevent the protocol from sending data. In T-Lohi, as an echo is not distinguishable using our low-power wake-up circuit, nodes considering their self-reflected tones as other contenders and are never able to send data. We designed a Bayesian learning algorithm that uses the fact that echoes from stable reflecting surface will have repeatable offsets when compared to ambient noise or unsynchronized contenders. This observation is used by our algo- rithm to repeatedly send tones and use the the response detection times as evidence. This evidence can increase belief about tones being classified as self-reflecting which can then be ignored. We perform detailed experimental evaluation of the algorithm in a wide-range of environments with varying complexity. Our results show that the algo- rithm correctly identifies echoes, robust to noise, and capable of adapting to a dynamic environment with varying surface locations. Our research in developing protocols and algorithm has shown that an understanding of the impact of acoustic propagation delay allows for developing protocols that not only overcome any new challenges but can actually benefit from the unique nature of large- propagation-delay acoustic networks. We foresee the resulting awareness foment a rapid 172 development of protocols dealing with more complex networking issues of scheduling, congestion, routing, end-to-end reliability, and transport layer protocols. Such a com- plete suite of protocols will enable many application of underwater sensornets, some that we envision (in Section 1) and hopefully a many we might not have. 173 Bibliography [ABB + 04] D. Aguayo, J. Bicket, S. Biswas, G. Judd, and R. Morris. Link-level mea- surements from an 802.11b mesh network. In ACM SIGCOMM, pages 121–132, Aug 2004. [Abr70] Norman Abramson. The ALOHA system. In Proceedings of AFIPS 1970 Fall Joint Computer Conference, volume 37, pages 281–285, 1970. [AK08] Joon Ahn and Bhaskar Krishnamachari. Performance of Propagation Delay Tolerant ALOHA Protocol for Underwater Wireless Networks. In The 4th IEEE International Conference on Distributed Computing in Sen- sor Systems (DCOSS ’08), pages 1–16, June 2008. [APM05] Ian F. Akyildiz, Dario Pompili, and Tommaso Melodia. Underwater acous- tic sensor networks: research challenges. Ad Hoc Networks Journal, 3(3):257–279, March 2005. [Atm] Atmel Atmega. Atmel atmega 128l. http://www.atmel.com/dyn/ resources/prod_documents/doc2467.pdf. [BDSZ94] Vaduvur Bharghavan, Alan Demers, Scott Shenker, and Lixia Zhang. MACAW: A media access protocol for wireless LAN’s. In Proceedings of the ACM SIGCOMM Conference, pages 212–225, London, UK, Septem- ber 1994. ACM. [Ben] Benthos, Inc. Fast And Reliable Access To Undersea Data. http:// www.benthos.com/pdf/Modems/ModemBrochure.pdf. [BG96] Dimitri Bertsekas and Robert Gallager. Data Networks. Prentice Hall, second edition, 1996. [BGS00] Philippe Bonnet, Johannes E. Gehrke, and Praveen Seshadri. Querying the physical world. In IEEE Personal Communications, pages 10–15, 2000. 174 [BMP + 06] L. Badia, M. Mastrogiovanni, C. Petrioli, S. Stefanakos, and M. Zorzi. An optimization framework for joint sensor deployment, link scheduling and routing in underwater sensor networks. In WUWNet ’06: Proceedings of the 1st ACM international workshop on Underwater networks, pages 56– 63, New York, NY , USA, 2006. ACM Press. [Bur84] William S. Burdic. Underwater Acoustic Systems Analysis. Prentice-Hall, 1984. [Cat90] J.A. Catipovic. Performance limitations in underwater acoustic telemetry. IEEE Journal of Oceanic Engineering, 15(3):205–216, Jul 1990. [CHS03] W.P. Chen, J.C. Hou, and L. Sha. Dynamic clustering for acoustic target tracking in wireless sensor networks. In Proceedings of the 11th IEEE International Conference on Network Protocols (ICNP’03), pages 258– 271, Atlanta, Georgia, USA, November 4-7 2003. [CLD06] Lijun Chen, Steven H. Low, and John C. Doyle. Random access game and medium access control design. Unpublished manuscript available at http://www.cds.caltech.edu/ ˜ chen/papers/ ramac.pdf, December 2006. [CM06] Cecilia Carbonelli and Urbashi Mitra. Cooperative multihop communi- cation for underwater acoustic networks. In WUWNet ’06: Proceedings of the 1st ACM international workshop on Underwater networks, pages 97–100, New York, NY , USA, 2006. ACM. [CPH04] M. Chitre, J. Potter, and O.S. Heng. Underwater acoustic channel char- acterisation for medium-range shallow water communications. OCEANS ’04. MTTS/IEEE TECHNO-OCEAN ’04, 1:40–45 V ol.1, Nov. 2004. [Cro02] Crossbow Technologies, Inc. Mica2 Data Sheet. http://www.xbow. com/, 2002. [CsSC08] N. Chirdchoo, Wee seng Soh, and Kee Chua. RIPT: A receiver-initiated reservation-based protocol for underwater acoustic networks. volume 26, pages 1744–1753, December 2008. [CW92] S. Crozier and P. Webster. Performance of 2-dimensional sloppy-slotted ALOHA random access signaling. In Conference Proceedings of IEEE International Conference on Selected Topics in Wireless Communications, pages 383–386, Jun 1992. [DSK07] J. Heidemann D. Son and B. Krishnamachari. Towards concurrent com- munication in wireless networks. Technical Report ISI-TR-646, USC/ISI, July 2007. 175 [EGE02] J. Elson, L. Girod, and D. Estrin. Fine-grained network time synchroniza- tion using reference broadcasts. In Proceedings of the Fifth Symposium on Operating Systems Design and Implementation (OSDI 2002), pages 147– 163, Boston, MA, USA, December 2002. [EHD04] A. El-Hoiydi and J.-D. Decotignie. Wisemac: an ultra low power mac protocol for the downlink of infrastructure wireless sensor networks. In Proceedings of ISCC”04, pages 244–251, Washington, DC, USA, 2004. IEEE Computer Society. [Els03] Jeremy Elson. Time Synchronization in Wireless Sensor Networks. PhD thesis, UCLA, 2003. [FGS + 05] L. Freitag, M. Grund, S. Singh, J. Partan, P. Koski, and K. Ball. The WHOI micro-modem: an acoustic communications and navigation system for multiple platforms. OCEANS, 2005. Proceedings of MTS/IEEE, pages 1086–1092 V ol. 2, 2005. [FJ94] S. Floyd and V . Jacobson. The synchronization of periodic routing mes- sages. ACM/IEEE Transactions on Networking, 2(2):122–136, April 1994. [FOL02] D. Fober, Y . Orlarey, and S. Letz. Clock skew compensation over a high latency network. In Proceedings of the ICMC, pages 548–552, G¨ oteborg, Sweden, 2002. ICMA. [GBEE02] Lewis Girod, Vladimir Bychkovskiy, Jeremy Elson, and Deborah Estrin. Locating tiny sensors in time and space: A case study. In ICCD ’02: Pro- ceedings of the 2002 IEEE International Conference on Computer Design: VLSI in Computers and Processors (ICCD’02), pages 214–219, Washing- ton, DC, USA, 2002. IEEE Computer Society. Invited paper. [GE01] L. Girod and D. Estrin. Robust range estimation using acoustic and multi- modal sensing. In Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems, volume 3, pages 1312–1320, 2001. [GFR07] Xiaoxing Guo, M.R. Frater, and M.J. Ryan. An adaptive propagation- delay-tolerant MAC protocol for underwater acoustic sensor networks. OCEANS 2007 - Europe, pages 1–5, June 2007. [GKS03] Saurabh Ganeriwal, Ram Kumar, and Mani B. Srivastava. Timing-sync protocol for sensor networks. In Proceedings of the First International ACM Conference on Embedded Networked Sensor Systems (SenSys), pages 138–149, Los Angeles, California, USA, 2003. ACM Press. 176 [GXXC07] John H. Gibson, Geoffrey G. Xie, Yang Xiao, and Hui Chen. Analyzing the performance of multi-hop underwater acoustic sensor networks. In Proceedings of the MTS/IEEE OCEANS Conference, Aberdeen, Scotland, June 2007. [HD02] Z. J. Haas and Jing Deng. Dual busy tone multiple access (DBTMA)-a multiple access control scheme for ad hoc networks. IEEE Transactions on Communications, 50(6):975–985, 2002. [HIL] HILTI, Inc., http://www.us.hilti.com. PD-40 Laser Range Meter. [HLS + 06] John Heidemann, Yuan Li, Affan Syed, Jack Wills, and Wei Ye. Under- water sensor networking: Research challenges and potential applications. In Proceedings of IEEE WCNC’06, pages 228–235, Las Vegas, Nevada, USA, April 2006. [HSS + 02] Martin Horauer, Ulrich Schmid, Klaus Schossmaier, Roland H¨ oller, and Nikolaus Ker¨ o. PSynUTC—evaluation of a high precision time synchro- nization prototype system for ethernet lans. In Proceedings of 34th Annual Precise Time and Time Interval Meeting (PTTI), pages 263–279, Reston, Virginia, USA, December 2002. [IEE] IEEE. RuBee status report. IEEE 1902.1, http://standards. ieee.org/db/status/index.shtml. [Jay03] E. T. Jaynes. Probability Theory: The Logic of Science. Cambridge Uni- versity Press, April 2003. [JBH78] I.M. Jacobs, R. Binder, and E.V . Hoversten. General purpose packet satel- lite networks. Proceedings of the IEEE, 66(11):1448–1467, Nov. 1978. [KB00] D.B. Kilfoyle and A.B. Baggeroer. The state of the art in underwater acous- tic telemetry. IEEE Journal of Oceanic Engineering, 25(1):4–27, Jan 2000. [KHS + 98] W. A. Kuperman, W. S. Hodgkiss, H. C. Song, T. Akal, C. Ferla, and D. R. Jackson. Phase conjugation in the ocean: Experimental demonstration of an acoustic time-reversal mirror. Acoustical Society of America Journal, 103:25–40, January 1998. [KIM07] Kurtis Kredo II and Prasant Mohapatra. Medium access control in wireless sensor networks. Comput. Networks, 51(4):961–994, 2007. [KPK81] W.C. Knight, R.G. Pridham, and S.M. Kay. Digital signal processing for sonar. Proceedings of the IEEE, 69(11):1451–1506, Nov. 1981. 177 [KS89] H. Kopetz and W. Schwabl. Global time in distributed real-time systems. Technical report 15/89, Technische Universitat Wien, 1989. [KT75] Leonard Kleinrock and Fouad Tobagi. Packet switching in radio chan- nels: Part I - carrier sense multiple access modes and their throughput delay characteristics. IEEE Transactions on Computing, 23(12):1400– 1416, December 1975. [Lam78] L. Lamport. Time, clocks, and the ordering of events in a distributed sys- tem. Communications of the ACM, 21, 7:558–564, 1978. [LP07] Weichang Li and J.C. Preisig. Estimation of rapidly time-varying sparse channels. IEEE Journal of Oceanic Engineering, 32(4):927–939, Oct. 2007. [LYH06] Yuan Li, Wei Ye, and John Heidemann. Energy efficient network recon- figuration for mostly-off sensor networks. In Proceedings of the IEEE SECON, pages 527–535, Reston, Virginia, USA, September 2006. IEEE. [Mil89] D. Mills. Internet time synchronization: the network time protocol; RFC 1129. Internet Request for Comments, (1129), October 1989. [Mil04] David Mills. Inter-planetary internet. http://www.eecis.udel. edu/ ˜ mills/ipin.html, 2004. [MKSL04] Miklos Maroti, Branislav Kusy, Gyula Simon, and Akos Ledeczi. The flooding time synchronization protocol. In Proceedings of the Second International ACM Conference on Embedded Networked Sensor Systems (SenSys), pages 39–49, Baltimore, MD, USA, 2004. ACM Press. [MS06] Marcal Molins and Milica Stojanovic. Slotted FAMA: A MAC protocol for underwater acoustic networks. In Proceedings of the IEEE OCEANS’06 Asia Conference, pages 1–7, Singapore, May 2006. [Net07] Coral Reef Environmental Observatory Network. CREON. http:// www.coralreefeon.org/, 2007. [NT94] B. Clifford Neuman and Theodore Ts’o. Kerberos: An authentication ser- vice for computer networks. IEEE Communications Magazine, 32(9):33– 38, September 1994. [PCB00] Nissanka B. Priyantha, Anit Chakraborty, and Hari Balakrishnan. The cricket location-support system. In Proceedings of the 6th ACM Interna- tional Conference on Mobile Computing and Networking, pages 32–43, Boston, MA, USA, August 2000. ACM. 178 [Pey99] Hassan Peyravi. Medium access control protocols performance in satellite communications. IEEE Communication Magazine, 37(3):62–71, March 1999. [PG58] R. Price and P.E. Green. A communication technique for multipath chan- nels. Proceedings of the Institute of Radio Engineers, 46(3):555–570, March 1958. [PHC04] Joseph Polastre, Jason Hill, and David Culler. Versatile low power media access for wireless sensor networks. In Proceedings of the 2nd ACM Con- ference on Embedded Networked Sensor Systems (SenSys), pages 95–107, Baltimore, MD, USA, November 2004. [PK00] G. J. Pottie and W. J. Kaiser. Embedding the Internet: wireless integrated network sensors. Communications of the ACM, 43(5):51–51, May 2000. [PKL06] Jim Partan, Jim Kurose, and Brian Neil Levine. A survey of practical issues in underwater networks. In Proceedings of the WUWNet ’06, 1st ACM workshop on Underwater networks, pages 17–24, New York, NY , USA, 2006. ACM Press. [Pre06] James Preisig. Acoustic propagation considerations for underwater acous- tic communications network development. In WUWNet ’06: Proceedings of the 1st ACM international workshop on Underwater networks, pages 1–5, New York, NY , USA, 2006. ACM Press. [PS06] Borja Peleato and Milica Stojanovic. A MAC protocol for ad-hoc under- water acoustic sensor networks. In Proceedings of the WUWNet ’06, pages 113–115, New York, NY , USA, 2006. ACM Press. [RARF06] Sumit Roy, Payman Arabshahi, Dan Rouseff, and Warren Fox. Wide area ocean networks: architecture and system design considerations. In WUWNet ’06: Proceedings of the 1st ACM international workshop on Underwater networks, pages 25–32, New York, NY , USA, 2006. ACM Press. [Ric05] Joseph Rice. SeaWeb acoustic communication and navigation network. In Proceedings of the International Conference on Underwater Acoustic Measurements: Technologies and Results, Heraklion, Greece, June 2005. [RP05] V olkan Rodoplu and Min Kyoung Park. An energy-efficient MAC protocol for underwater wireless acoustic networks. In Proceedings of the IEEE OCEANS’05 Conference, pages 1198–1203 V ol. 2, Sept. 2005. 179 [RYC + 05] Nithya Ramanathan, Mark Yarvis, Jasmeet Chhabra, Nandakishore Kushalnagar, Lakshman Krishnamurthy, and Deborah Estrin. A stream- oriented power management protocol for low duty cycle sensor network applications. In Proceedings of the IEEE Workshop on Embedded Net- worked Sensors, pages 53–62, Sydney, Australia, May 2005. IEEE. [SCBR04] Kimaya Sanzgiri, Ian D. Chakeres, and Elizabeth M. Belding-Royer. Determining intra-flow contention along multihop paths in wireless net- works. In Proceedings of BROADNETS’04, pages 611–620. IEEE Com- puter Society, 2004. [SCP94] Milica Stojanovic, Josko Catipovic, and John Proakis. Phase coherent dig- ital communications for underwater acoustic channels. IEEE Journal of Oceanic Engineering, 19(1):100–111, January 1994. [SH06] Affan Syed and John Heidemann. Time synchronization for high latency acoustic networks. In Proceedings of the IEEE Infocom, pages 1–12, Barcelona, Catalunya, Spain, April 2006. [SHK07] Hee-Chun Song, William S. Hodgkiss, and William A. Kuperman. MIMO time reversal communications. In Proceedings of WuWNet ’07, pages 5– 10, New York, NY , USA, 2007. ACM. [SHS01] Andreas Savvides, Chih-Chien Han, and Mani Srivastava. Dynamic fine- grained localization in ad-hoc networks of sensors. In Proceedings of the ACM International Conference on Mobile Computing and Networking, pages 166–179, Rome, Italy, July 2001. ACM. [SK97] Mark Stemm and Randy H Katz. Measuring and reducing energy con- sumption of network interfaces in hand-held devices. IEICE Transactions on Communications, E80-B(8):1125–1131, August 1997. [SKH04] D. Son, B. Krishnamachari, and J. Heidemann. Experimental study of the effects of transmission power control and blacklisting in wireless sensor networks. In IEEE SECON, pages 289–298, Oct 2004. [SKH06] D. Son, B. Krishnamachari, and J. Heidemann. Experimental study of con- current transmission in wireless sensor networks. In ACM Sensys, pages 237–250, Nov 2006. [SS06] Ethem Mutlu S¨ ozer and Milica Stojanovic. Reconfigurable acoustic modem for underwater sensor networks. In Proceedings of WUWNet ’06, pages 101–104, New York, NY , USA, 2006. ACM. 180 [Sto96] M. Stojanovic. Recent advances in high-speed underwater acoustic com- munications. Oceanic Engineering, IEEE Journal of, 21(2):125–136, Apr 1996. [Sto03] Milica Stojanovic. Wiley Encyclopedia of Telecommunications, chapter Acoustic (underwater) Communications. John Wiley & Sons, Inc., 2003. [Sto05] Milica Stojanovic. Optimization of a data link protocol for an underwa- ter acoustic channel. In Proceedings of the IEEE OCEANS’05 - Europe Conference, pages 68–73, June 2005. [Sto06] Milica Stojanovic. On the relationship between capacity and distance in an underwater acoustic communication channel. In Proceedings of the WUWNet ’06, 1st ACM workshop on Underwater networks, pages 41–47, New York, NY , USA, 2006. ACM Press. [SYH08] Affan A. Syed, Wei Ye, and John Heidemann. T-Lohi: A new class of MAC protocols for underwater acoustic sensor networks. In Proceedings of the IEEE Infocom, pages 231–235, Pheonix, AZ, April 2008. [SYKH07] Affan Syed, Wei Ye, Bhaskar Krishnamachari, and John Heidemann. Understanding spatio-temporal uncertainty in medium access with aloha protocols. In Proceedings of WUWNet06, pages 41–48, Montreal, Que- bec, Canada, September 2007. ACM. [Sys07] Wireless Fiber Systems. Underwater Radio Modem. http://www. wirelessfibre.co.uk/datasheets.htm, 2007. [Thr98] S. Thrun. Bayesian landmark learning for mobile robot localization. Machine Learning, 33(1):41–76, 1998. [TI86] S. Tasaka and K. Ishida. The sruc protocol for satellite packet communication–a performance analysis. Communications, IEEE Trans- actions on, 34(9):937–945, Sep 1986. [TK75a] Fouad A. Tobagi and Leonard Kleinrock. Packet switching in radio chan- nels: Part I–carrier sense multiple-access modes and their throughput- delay characteristics. IEEE Trans. on Computing, 23(12):1400–1416, 1975. [TK75b] Fouad A. Tobagi and Leonard Kleinrock. Packet switching in radio chan- nels: Part II–the hidden terminal problem in carrier sense multiple-access and the busy-tone solution. IEEE Trans. on Computing, 23(12):1417– 1433, 1975. 181 [TOP06] Tagging of Pacific Pelagic project.http://www.toppcensus.org/, 2006. [Uri91] R.J. Urick. Principles of Underwater Sound. McGraw-Hill Book Com- pany, 1991. [vGR03] Jana van Greunen and Jan Rabaey. Lightweight time synchronization for sensor networks. In Proceedings of the Second ACM international con- ference on Wireless sensor networks and applications, pages 11–19, San Diego, CA, USA, 2003. ACM Press. [Vig92] John R. Vig. Introduction to quartz frequency standards. Techni- cal Report SLCET-TR-92-1, Army Research Laboratory, Electronics and Power Sources Directorate, 1992. [VKLG06] Luiz Filipe M. Vieira, Jeijun Kong, Uichin Lee, and Mario Gerla. Anal- ysis of ALOHA protocols for underwater acoustic sensor networks. In Extended abstract in WUWNet’06, 2006. [VKR + 05] I. Vasilescu, K. Kotay, D. Rus, M. Dunbabin, and P. Corke. Data collection, storage, and retrieval with an underwater sensor network. In Proceedings of ACM SenSys’05, pages 154–165, 2005. [Wha71] Anthony D. Whalen. Detection of Signals in Noise. Academic Press, 1971. [WK99] Jia Wang and Srinivasan Keshav. Efficient and accurate ethernet simula- tion. In LCN ’99: Proceedings of the 24th Annual IEEE Conference on Local Computer Networks, pages 182–191, Washington, DC, USA, 1999. IEEE Computer Society. [WWJ + 05] K. Whitehouse, A. Woo, F. Jiang, J. Polastre, and D. Culler. Exploiting the capture effect for collision detection and recovery. In IEEE Workshop on Embedded Networked Sensors (EmNetS-II), pages 45–52, May 2005. [WYH06] Jack Wills, Wei Ye, and John Heidemann. Low-power acoustic modem for dense underwater sensor networks. In Proceedings of WUWNet05, pages 79–85, Los Angeles, California, USA, September 2006. ACM. [WYM + 02] H. Wang, L. Yip, D. Maniezzo, J. Chen, R. Hudson, J. Elson, and K. Yao. A wireless time-synchronized COTS sensor platform Part II: Applications to beamforming. In Proceedings of IEEE CAS Workshop on Wireless Com- munications and Networking, pages 2–9, Pasadena, CA, USA, September 2002. 182 [XC06] Peng Xie and Jun-Hong Cui. Exploring random access and handshaking techniques in large-scale underwater wireless acoustic sensor networks. In Proceedings of the MTS/IEEE OCEANS Conference, pages 1–6, Boston, Massachusetts, September 2006. [XG00] Geoffrey G. Xie and John Gibson. A networking protocol for underwater acoustic networks. Technical Report TR-CS-00-02, Department of Com- puter Science, Naval Postgraduate School, December 2000. [YHE02] Wei Ye, J. Heidemann, and D. Estrin. An energy-efficient mac protocol for wireless sensor networks. In Proceedings of IEEE INFOCOM, volume 3, pages 1567–1576, New York, NY , Jun 2002. [YS08] Justin Yackoski and Chien-Chung Shen. Uw-flashr: achieving high chan- nel utilization in a time-based acoustic mac protocol. In WuWNeT ’08: Proceedings of the third ACM international workshop on Wireless network testbeds, experimental evaluation and characterization, pages 59–66, New York, NY , USA, 2008. ACM. [YSH06] Wei Ye, Fabio Silva, and John Heidemann. Ultra-low duty cycle MAC with scheduled channel polling. In Proceedings of the Fourth ACM SenSys Conference, pages 321–333, Boulder, Colorado, USA, November 2006. ACM. [ZG03] J. Zhao and R. Govindan. Understanding packet delivery performance in dense wireless sensor networks. In SenSys ’03: Proceedings of the 1st international conference on Embedded networked sensor systems, pages 1–13, Los Angeles, California, USA, Nov 2003. [ZK04] M. Zuniga and B. Krishnamachari. Analyzing the transitional region in low power wireless links. In First Annual IEEE Communications Society Con- ference on Sensor and Ad Hoc Communications and Networks (SECON), pages 517–526, Oct 2004. [ZM07] Wenyi Zhang and Urbashi Mitra. A delay-reliability analysis for multihop underwater acoustic communication. In WuWNet ’07: Proceedings of the second workshop on Underwater networks, pages 57–64, New York, NY , USA, 2007. ACM. [ZSR04] Bin Zhang, Gaurav S. Sukhatme, and Aristides A. G. Requicha. Adaptive sampling for marine microorganism monitoring. volume 2, pages 1115– 1122, Sendai, Japan, 2004. IEEE. 183

Abstract (if available)

Abstract An understanding of the key areas of difference in acoustic underwater sensor networks and their impact on network design is essential for a rapid deployment of aquatic sensornets. Such an understanding will allow system designers to harvest the vast literature of research present in RF sensornets and focus on just those key aspects that are different for acoustic sensornets. Most complexities at the physical layer will eventually be handled either by assuming short ranges or with technology advancements making complex algorithms both cost and power efficient. However, the impact of large latency and the resulting magnification of multipath will remain a great impediment for developing robust sensor networks. This thesis contributes towards an understanding of, and solutions to, the impact of latency on sensornet migration to an underwater acoustic environment.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Sensing with sound: acoustic tomography and underwater sensor networks

PDF

Time synchronization and scheduling in underwater wireless networks

PDF

Design of cost-efficient multi-sensor collaboration in wireless sensor networks

PDF

Aging analysis in large-scale wireless sensor networks

PDF

Robust routing and energy management in wireless sensor networks

PDF

Gradient-based active query routing in wireless sensor networks

PDF

Efficient and accurate in-network processing for monitoring applications in wireless sensor networks

PDF

Analysis and countermeasures of worm propagations and interactions in wired and wireless networks

PDF

Algorithmic aspects of throughput-delay performance for fast data collection in wireless sensor networks

PDF

Rate adaptation in networks of wireless sensors

PDF

Models and algorithms for energy efficient wireless sensor networks

PDF

Efficient data collection in wireless sensor networks: modeling and algorithms

PDF

Multichannel data collection for throughput maximization in wireless sensor networks

PDF

Realistic modeling of wireless communication graphs for the design of efficient sensor network routing protocols

PDF

Data replication and scheduling for content availability in vehicular networks

PDF

Relative positioning, network formation, and routing in robotic wireless networks

PDF

Anycast stability, security and latency in the Domain Name System (DNS) and Content Deliver Networks (CDNs)

PDF

Crowd-sourced collaborative sensing in highly mobile environments

PDF

Disentangling the network: understanding the interplay of topology and dynamics in network analysis

PDF

Benchmarking interactive social networking actions

Asset Metadata

Creator Syed, Affan Ahmed (author)

Core Title Understanding and exploiting the acoustic propagation delay in underwater sensor networks

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Computer Science (Computer Networks)

Publication Date 06/16/2009

Defense Date 02/22/2009

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag acoustic,latency-aware,networks,OAI-PMH Harvest,sensor networks,underwater

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Heidemann, John (committee chair), Krishnamachari, Bhaskar (committee member), Neely, Michael J. (committee member), Ye, Wei (committee member)

Creator Email affan.syed.usc@gmail.com,asyed@isi.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-m2301

Unique identifier UC1495244

Identifier etd-Syed-2948 (filename),usctheses-m40 (legacy collection record id),usctheses-c127-249132 (legacy record id),usctheses-m2301 (legacy record id)

Legacy Identifier etd-Syed-2948.pdf

Dmrecord 249132

Document Type Dissertation

Rights Syed, Affan Ahmed

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Repository Name Libraries, University of Southern California

Repository Location Los Angeles, California

Repository Email cisadmin@lib.usc.edu