Luận án tiến sĩ: Wireless MAC layer flexibility for extending effective system lifetime

Offloading protocol overhead from highly loaded nodes to lightly loaded With a shared medium and limited transmission ranges, nodes may To avoid collisions, nodes exchange RTS and CTS pa

WIRELESS MAC LAYER FLEXIBILITY FOR EXTENDING EFFECTIVE SYSTEM LIFETIME

by

Rebecca Lynn Braynard Silberstein

Department of Computer Science

Dissertation submitted in partial fulfillment of the

requirements for the degree of Doctor of Philosophy

in the Department of Computer Science

in the Graduate School of Duke University 2006

UMI Number: 3244842

Copyright 2006 by Silberstein, Rebecca Lynn Braynard

All rights reserved

INFORMATION TO USERS

The quality of this reproduction is dependent upon the quality of the copy submitted Broken or indistinct print, colored or poor quality illustrations and photographs, print bleed-through, substandard margins, and improper alignment can adversely affect reproduction

In the unlikely event that the author did not send a complete manuscript and there are missing pages, these will be noted Also, if unauthorized copyright material had to be removed, a note will indicate the deletion

®

UMI

UMI Microform 3244842 Copyright 2007 by ProQuest Information and Learning Company

All rights reserved This microform edition is protected against unauthorized copying under Title 17, United States Code

ProQuest Information and Learning Company

300 North Zeeb Road P.O Box 1346

Copyright © 2006 by Rebecca Lynn Braynard Silberstein

All rights reserved

ABSTRACT (Computer Science) WIRELESS MAC LAYER FLEXIBILITY FOR EXTENDING EFFECTIVE SYSTEM LIFETIME

by Rebecca Lynn Braynard Silberstein

Department of Computer Science

Dr Maria Papadopouli

An abstract of a dissertation submitted in partial fulfillment of the

requirements for the degree of Doctor of Philosophy

in the Department of Computer Science

in the Graduate School of Duke University 2006

Abstract

The scenario of wireless devices deployed in an ad hoc network is a compelling one For the system to support its assigned tasks, nodes must cooperate to forward data packets through the network The act of forwarding, however, is directly linked to node lifetime The more a node forwards data packets, the sooner it depletes its bat- tery through use of its high energy-consuming radio Once forwarding nodes deplete their batteries, the network may partition and fail to provide its designated services

To reduce the burden on nodes maintaining network connectivity, we propose to bal- ance energy consumption among all participating networked nodes, so they all fail

at approximately the same time We do this through our Medium Access Control

(MAC) layer protocol, SEESAW, which offloads communication control overhead from

heavily loaded nodes to more lightly loaded neighbors SEESAW’s unique combina- tion of asynchrony, asymmetry and batching allow probabilistic communication with reliable delivery, while both reducing and balancing energy consumption

We present SEESAW and an accompanying online, distributed, self-tuning algo- rithm We evaluate SEESAW through simulation using our detailed energy modeling simulator, SENSIM, and implementation on Mica2 Motes in TinyOS Our results show SEESAW is competitive in terms of energy consumption and delivery performance with related protocols, despite additional overhead needed for an asynchronous and asym- metric protocol Additionally, we show SEESAW successfully balances consumption among networked nodes while providing acceptable delivery performance with static random topologies and static, random, and bursty application workloads

iv

Acknowledgements

Without the support of family and friends, completing this thesis would not have been possible I would especially like to thank my parents, Jerry and Gerry Braynard, who help with everything I try to do in whatever way they can My husband, Adam Silberstein, who is not only my best friend but also happens to be a wonderful collaborator, has helped every way possible To Dr Carla Ellis, thank you for guiding me with this project and all of your help, whether or not it was related to this thesis Thank you Dr Alvy Lebeck, Dr Jun Yang and Dr Maria Papadopouli, your help with research and graduate school advice made it possible for me to finish Diane Riggs, thank you for your guidance through every step along the way and all the reminders to make sure I could actually graduate I also could not miss the chance to thank Adolfo Rodriguez, Steve Kacsmark, Andrea Klostermann, Richenda Petersen and Billy and Alexa Corbett who all provide encouragement, help and a good laugh whenever needed

1.2 Wireles and Ad Hoc Network Challengs

1.23 Radio Range and Link Asymmetry

2 Background and Related Work

2.2.3 Asynchronous Protocols 0.0.0 + eee

Evaluating Static Seesaw

5.1 Asymmetry and Asynchrony Overhead

5.2 Parameter Influence 0.00 eee ee ee

Seesaw in TinyOS on Mica2 Motes

7.2 Comparison of TinyOS Implementation and Simulation Results

82 Puture WOrk - c c Q es

A Objective Function Variable Derivations

B Random Test Topology Set

C Individual Test Data Tables

C.1 Constant Bit Rate (CBR) Experiments C.2 Changing Data Rate (CDR) Experimens C.3 Hot Spot Burst (HSB) Experiments

and Vaidya [15], 800useconds and 2*IDLE power [11, 32] 5

Sample neighbor parameter tables for the 4 nodes in the topology of

Node settings after processing the neighbor table entries (action booleans

Node settings after processing action booleans and accounting for the

SEESAW parameter settings for both sets of parameter comparison tests In the base tests, batching is disabled and the enhanced version,

Cross topology parameter values for the three scenarios, Symmetric,

Packets delivered and average delay for SENSIM tests using offline

Node parameter settings after 50000 seconds of simulation without

(left) and with (right) batch size tuning enabled 73

7.1 Parameter settings for interval skew tests -.-, 95

A.1 Variables used in detailed definitions of the components of the model introduced in Table3.1 2.0 20.0.0 2.00 eee ees 104 C1 8 Nodes CBR: 1 packet per 2seconds 108

C2 8Nodes CBR: 1 packet per 4seconds 109

C.3 8 Nodes CBR: 1 packet per 8seconds 109

C.4 20 Nodes CBR: 1 packet per 8seconds 110

C.5 20 Nodes CBR 1 packet per l6 seconds 110

C.6 CDR: rate change every 60-120 seconds 111

C.7 CDR: rate change every 120-240 seconds ., 112

C.8 CDR: rate change every 1000-2000 seconds 112

C.9 HSB: about 120 seconds between hot spots 004 113 C.10 HSB: about 240 seconds between hot spots 114 C.11 HSB: about 480 seconds between hot spots 0.0, 114

Offloading protocol overhead from highly loaded nodes to lightly loaded

With a shared medium and limited transmission ranges, nodes may

To avoid collisions, nodes exchange RTS and CTS packets to estab- lish communication and follow Data packets with acknowledgments

Nodes send a evenly spaced Ads each interval when they have data

Nodes use the same interval length, but the intervals can start at

Accept (establish communication) packet components 25

Decoupling sending and listening allows nodes to build asymmetric

Simple three node topology with a source generating data, a forwarder passing data along the path and a destination acting as the data packet

3.11 State information from a portion of a SEESAW simulation run

3.12 Simple four node topology with nodes 1, 2 and 3 producing data pack- ets for the sink, node 0 To maintain connectivity, node 1 forwards

4.2 SENSIM provides three layers for implementing functionality, applica- tion, routing and medium access The Channel provides communica-

4.3 Reproduced energy consumption results from S-MAC [43]

4.4 Validation of our simulator results for S-MAC and B-MAC against

published results from S-MAC [43] and B-MAC [25),

5.3 Base: Delivered packets © 0 0 0 pee ee ee eee

35

5.12 Varying listen time per interVaÌl,Ì.Ô cuc ee ee

5.15 Using the cross traffic topology (”X” Topology), we compare three

5.16 Node energy consumption for three data rates using CPLEX determined

5.17 Total node consumption with static settings: a=12and/=0.1 5.18 Total node consumption using offline algorithm settings 5.19 Total node consumption with a and / tuning and static B 5.20 Total node consumption with a, / and B tuning enabled

6.1 Sample topology that is difficult to balance Node 1 is forwarding for many nodes and node 7 simply does not need to, nor possibly can, consume enough to balance consumption with nodel

6.3 Simple network where all sources can directly communicate with the

6.4 Simple network with maximum packet forwarding to the sink, node 0 6.5 Box plot for 8-node topologies with nodes generating a packet approx-

6.6 Box plot for 8-node topologies with nodes generating a packet approx-

6.7 Box plot for 8-node topologies with nodes generating a packet approx- imately every 8 seconds 2 ee

71

Box plot for 20 node topologies with nodes generating a packet ap-

Box plot for 20 node topologies with nodes generating a packet ap-

Box plot for 8 node topologies with nodes generating packets with random intervals for random durations 04

Box plot for 8 node topologies with nodes generating packets with random intervals for random durations 0.000

Box plot for 8 node topologies with nodes generating packets with random intervals for random durations 00 Box plot for 8 node topologies with 120 second hot spot bursts spaced

Box plot for 8 node topologies with 120 second hot spot bursts spaced approximately 240 seconds apart

Box plot for 8 node topologies with 120 second hot spot bursts spaced approximately 480 seconds apart cu

Validation of our simulation results for SEESAW against measurements

Comparison of Mica2 Mote and simulator energy consumption results for a node sending 12 Ads per interval and a receiver listening 11% of

Comparison of Mica2 Mote and simulator energy consumption results for two nodes without and with tuning enabled

Comparison of Mica2 Mote and simulator parameter values for two

Delivery performance for varying degrees of clock instability for ran-

Chapter 1

Introduction

Sensor networks are comprised of small wireless devices, or nodes The fundamental

function of such networks are for nodes to take periodic environmental readings (i.e

temperature), write them to data packets, and send those packets to a sink, or collection node Nodes are deployed as an ad hoc network such that nodes forward data through each other toward the sink, and therefore depend on one another to function properly Based on their locations within the network, nodes may have vastly different traffic workloads In turn, nodes may have vastly different rates of energy consumption; the nodes with the highest workloads consume at the highest rates Once these fast-consuming nodes deplete their batteries, they are no longer available to forward packets toward the sink When these forwarding nodes die, the network is partitioned until either their depleted batteries are replaced, or other nodes, whether new or existing, are deployed in the gap Meanwhile, lower workload nodes have energy available to take measurements, receive and transmit packets The loss of forwarders and resulting partition, however, means the sink will not receive the data produced by this work Therefore, when a partition occurs and some nodes have energy remaining, that energy cannot be utilized, and is wasted by the network Delving into the traffic imbalance issue, it is the nodes close to the sink we expect

to have higher forwarding burdens than those at the fringes of the network Fringe nodes have small to no forwarding burdens, since by definition, they do not lie between many other nodes and the sink Nodes near the sink, in contrast, are likely to be used to transmit many packets originating at those fringe nodes to the sink

We observe a double-edged problem Forwarding nodes have the highest burden

Source Forwarder Destination

or fringe Further, if this is done such that this balanced energy consumption rate among all nodes is lower than the otherwise maximal rates of the forwarders, then balancing extends the network lifetime This improvement is possible because, intu- itively, balanced consumption is a more effective utilization of the system’s available energy

In this work, we address the specific goal of extending the effective network lifetime, specifically in sensor networks, defined as time prior to the first node dying The nodes in position to play a critical role in maintaining connectivity and delivering the sensor data are those at greatest risk for battery depletion

2

Figure 1.1 illustrates the problem using a three node topology with source, s, forwarder, f, and destination, d The energy consumption of multihop communica- tion is given in the bar charts for two different protocols In each plot, the lower bar components represent the energy to transmit and receive data, while the upper components represent energy consumed by protocol overhead Data cost for each node is necessarily equal between the two plots because the same data packets are exchanged A horizontal line marks the most heavily consuming node, the first to die from battery depletion In this example that is f, by virtue of its sending and receiving Minimizing the maximum energy consumption across all nodes maximizes the time until some node dies

A solution is illustrated by the right plot in Figure 1.1 Since data costs are fixed, the protocol overhead must be the focus of energy management efforts On the left, each node has equal overhead The key idea is to shift some of the protocol burden onto s and d to reduce the protocol cost at f This decreases total energy consumption at f and increases consumption at s and d, leading to balanced energy consumption among them This better utilization of network energy results in in- creased time until the first node dies, thus extending network lifetime In fact, with complete balance, all nodes die together, leaving no unused energy We have devel- oped SEESAW, an asynchronous and asymmetric MAC layer protocol that balances energy consumption among nodes to prolong network lifetime and strives for low en- ergy consumption In some cases, due to the workload and topology, perfect energy consumption balance cannot be achieved In these, SEESAW balances the heaviest consumers, while avoiding excess consumption at lightly loaded nodes There are two extremes to balancing consumption One way to balance is for all nodes to consume energy at the maximum rate This leaves all nodes dying at the same time, but the network lifetime is cut short On the other extreme, nodes can remain off, never

expending energy This has a very long network lifetime, but the network does not deliver any data We strive for a balance between these extremes Our goal is to balance energy consumption, without expending extra energy, while the network still provides its intended functionality

1.2 Wireless and Ad Hoc Network Challenges

Balancing consumption is not an easy task The target device type and network virtues make the problem challenging The nodes are small devices with limited battery supplies Since the radio is a large energy consumer, it is a primary target for new protocols and research due in part to the available states with varied power

values (more to follow) Wireless devices all use the same medium, air, to transmit

their messages Unlike wired networks, the nodes cannot detect collisions until packet transmissions are complete Additionally, nodes have a limited transmission range,

so they cannot know if a transmission is successful at a neighbor without some form

of an acknowledgment Along with limited transmission ranges, wireless nodes must also deal with link asymmetry Not only are radios for each node different (ranges are not perfectly shaped spheres centered at the node) they also suffer from obstructions and interference These are only a few examples of the interesting complications in sensor network research, and we further explore their consequences in the remainder

of this chapter

1.2.1 Power

The MAC layer provides two main avenues for controlling energy consumption These are the node radio states and communication overhead The radio states trade off energy efficiency and functionality Although radios can be equipped with many possible states, the main examples are send, receive, idle, sleep and off Send and

4

Transmit (ø;) 36mW Receive (p,.) 14.4mW Idle (p;) 14.4mW Sleep (ps) 0.015mW

receive are the highest-consuming states, but also provide the most functionality The next state, idle, is actually a primary culprit for wasted energy consumption When a radio is in idle, it is ready and available to receive a new transmission Idle, however, consumes energy at the same rate as receive Additionally, to receive a packet, the radio must first be in idle This is the primary dilemma for MAC layer protocols A node’s radio must be in idle to receive a transmission, but while in this state when there are no packets to receive, the node wastes valuable energy The remaining two states, sleep and off, consume the least amount of energy and provide the lowest functionality When in either, a node cannot send or receive packets To do so, the node must first transition to a higher consuming active state These transitions come

at a cost, in terms of both energy and time

The MAC layer provides another opportunity to manipulate energy consumption among nodes To communicate, nodes must coordinate using control overhead A protocol may actively synchronize clocks to schedule communication, or it may re- actively coordinate only when nodes have data to send In either case, to establish communication, nodes must transmit and receive control overhead The amount of overhead is protocol-dependent, and provides an opportunity to meet our goal of reduced and balanced consumption

Many protocols have nodes equally share the burden for establishing

communi-cation Alternatively, we propose to skew this burden such that lowly-loaded nodes

(e.g fringe nodes) are given a higher burden than highly-loaded nodes (e.g for-

warders near sink) The intuition is if the data transmission workloads of a pair

of nodes are fixed (once communication is established), we still have the ability to redistribute energy spent on establishing communication between the pair, with the goal of balancing their total consumptions Consider a pair of nodes where the re- ceiver is the higher energy consumer To balance consumption, we must reduce the receiver’s consumption We do this by decreasing the receiver’s time spent in the idle state As mentioned, this reduces its ability to receive messages, hurting func- tionality In response, we increase the frequency with which the sender attempts

to contact the receiver Intuitively, when these adjustments are made, we want to maintain the likelihood that the sender quickly finds the receiver in the idle state In doing so, we shift the overhead burden between sender and receiver for more balanced consumption, without any degradation in connectivity

1.2.2 Shared Medium

As mentioned earlier, wireless devices share their medium, the air When one node

is transmitting a packet, the transmission occupies the medium (or channel) and to prevent collisions, neighboring nodes should refrain from sending Collisions occur when multiple transmissions are heard by a receiving node The transmissions inter- fere with one another and corrupt the packets at the receiving node When collisions occur in a reliable protocol, data must be retransmitted Retransmissions require the sending node to transmit and the receiving node to receive multiple times to transfer

a single data packet one hop along its path to its destination This not only wastes energy at the nodes, but also lengthens delivery time and increases contention for the channel With multiple nodes vying for the same medium it becomes, in a sense, a

competition to acquire the channel and transmit without collisions Nodes need to employ protocols to reduce collisions and effectively share the medium

1.2.3 Radio Range and Link Asymmetry

Node radios have limited range Nodes can alter their transmission range using differ- ent transmission powers, but the signal strength also changes In our work, we utilize the maximum transmission power for all nodes and we consider adjusting this power orthogonal research to our own This line of work is further discussed in Chapter 2 Having limited radio ranges introduces complications for wireless networks Nodes with limited transmission range may not be able to reach their destination directly, forcing neighbors to forward their data Having a limited range does have a subtle benefit When one node transmits, it occupies the medium in the area With a limited range, more than one node can potentially transmit at a time without inter- fering with one another, potentially increasing the network throughput Effectively

“allocating” the medium to avoid collisions is a difficult wireless research problem Asymmetric links in a wireless network introduce an important complication A node may be able to transmit to a neighbor, but due to obstacles, radio power or other issues, the sending node cannot receive the neighbor’s packet acknowledgments This prevents a node from knowing if its packets were received, and therefore, inhibits reliable protocols that acknowledge a packet’s reception Link asymmetry generally precludes its use in many protocols; thereby increasing the forwarding load on well- connected nodes

Nodes cannot detect ongoing transmissions when they are out of the sender’s range or obstructed, introducing the Hidden Terminal Problem The hidden terminal problem even affects nodes not trying to directly communicate with one another In Figure 1.2, we see four nodes Node z is transmitting a packet to node y In addition,

by SEESAW to avoid such collisions

This dissertation presents a MAC-layer protocol, SEESAW, that balances energy con- sumption in a wireless ad hoc network We argue that balancing consumption bet- ter utilizes network energy and extends system lifetime To achieve balance, nodes performing varied tasks and supporting different traffic workloads need individual

schedules not only to fulfill their roles, but also to conserve energy SEESAW ac- complishes this by utilizing asynchrony and asymmetry Asynchrony allows nodes

to avoid regular communication for maintaining individual clocks or a system-wide time Instead, nodes maintain their own notion of time and do not coordinate to agree on a specific time Asymmetry empowers nodes to schedule activities based on their own workload Asynchrony and asymmetry allow the MAC layer flexibility to achieve energy goals such as balancing consumption

The contributions of this thesis include:

e SEESAW: We present our novel asynchronous and asymmetric MAC protocol, the first to decouple the actions of sending and receiving and reduce control overhead through batching data packets

e We employ probabilistic communication to effectively maintain communication while providing comparable performance to related protocols

e We investigate SEESAW’s ability to extend network lifetime by increasing the time until the first node fails

e We demonstrate SEESAW’s flexibility for meeting system-wide energy goals, including that of balancing consumption, through simulation

e SENSIM: We develop a sophisticated network simulator for modeling radio energy consumption, validated by reproducing published results from our and related protocol implementations on Mica2 Motes

e We provide algorithms for automatically tuning SEESAW for real-world deploy- ment in random topologies and changing workloads

e We implement SEESAW on current devices, demonstrating its simplicity and flexibility in practice

In this thesis we present SEESAW We provide background and related work in Chapter 2 and describe SEESAW in Chapter 3 Our simulator, SENSIM is detailed in Chapter 4 We present simulator results and analysis in Chapters 5 and 6 with the Mica2 Mote implementation and evaluation in Chapter 7 We conclude and describe future work in Chapter 8

10

Chapter 2

Background and Related Work

There are two extremes in sending data through an ad hoc network when dealing with energy and performance Ôn one hand, we could keep all nodes active and listening 100% of the time This extreme provides the best connectivity and highest delivery performance as long as energy resources are available, but it forces nodes to consume energy at the highest rate resulting in the shortest node and system lifetime Conversely, nodes could buffer all packets until data collection is no longer needed All remaining resources are used to transmit the stored data to the destination with the caveat of long delivery delays and potential data losses due to limited memory and storage Neither of these extremes will work for most applications Instead, applications are willing to suffer some decrease in performance in exchange for a longer-lived system, able to provide services for an increased period of time

2.1 Energy Savings in Wireless Protocols

Identifying the factors that contribute to energy consumption of wireless sensor proto- cols is useful both to categorize related work and to focus attention on open opportu- nities to reduce wasted energy The amount of traffic passing through a node and the power needed for sending, receiving and listening determine the energy requirements for transmitting application data Routing plays a role in affecting the traffic load through any particular node Energy-aware routing protocols (e.g., [10, 8, 4, 41, 6]) attempt to choose paths that will use the least amount of energy to send the packet

or paths with the most remaining energy, attempting to increase forwarding fairness among nodes Transmission power can be varied in some radios, affecting trans-

mission range and connectivity [15, 1, 4, 26] Additional traffic may be generated

by any necessary maintenance operations such as performing route discovery and maintenance, duty cycle schedule negotiation, or time synchronization (avoided with asynchronous protocols) In this work, we focus on the transmission phase during periods of stable topology where node discovery and routing tables have been com- pleted

2.1.1 Overhearing Avoidance

Overhearing transmissions intended for other nodes while in idle listening mode rep- resents one form of wasted energy The idea of overhearing avoidance, presented in PAMAS [83], allows nodes to determine if a packet is for them or not If not, they disable their radio for the duration of the packet to reduce the energy wasted by receiving a packet intended for another node

2.1.2 Collisions

Collisions are one source of extra traffic and wasted energy A node in a contention- based wireless network may not be able to directly detect another transmission oc- curring within its own transmission range [2] If this node were to send a packet, the transmission could be corrupted and have to be retransmitted, thus wasting valuable energy resources The classic solution to prevent this situation is in the IEEE 802.11 protocol with its distributed coordination function and the RTS-CTS

(Request-To-Send - Clear-To-Send) handshake [16], shown in Figure 2.1 To establish

communication, nodes check for an existing transmission by performing carrier sense

to determine if the medium is idle or busy If idle, they can exchange Request-To-

Send (RTS) and Clear-To-Send (CTS) packets This alerts neighboring nodes to the

upcoming Data packet transmission, allowing the neighbors to avoid transmitting

12

2.2 Utilizing Radio States

Focusing on radio transmission power and/or collisions allows some energy to be saved; however, there exists a potentially greater opportunity to save the energy expended by radios using power to listen for possible traffic, even when there are no active data flows In fact, the long periods of time typically spent in the idle state can result in idle power being a dominate factor in energy costs It is important to more effectively exploit all of the available radio power states (Table 1.1), including sleep and off, to conserve energy Energy-aware protocols employing regular, low power duty cycles must have a mechanism to avoid missing incoming transmissions when the radio is not active

Coordinating between nodes by predictive schemes [44] or scheduling through

advertising data paths [45] can enable nodes to sleep more often and reduce the number of transmission attempts when coordinating communication Another simple idea is to have senders page their destinations and coordinate data transfers using an

always on, alternative and lower power device [5, 31, 30] If nodes are on a heavily

used path to the destination, they may not have the opportunity to enter a lower power state

For the remainder of this discussion, we categorize protocols by their defin- ing characteristic Duty cycle protocols can be divided into slotted or contention- based protocols Slotted protocols divide time into slots and schedule transmissions Contention-based protocols determine if the channel is busy or idle and generally perform the RTS-CTS handshake to avoid collisions Furthermore, protocols can be

(a)synchronous and/or (a)symmetric Synchronous protocols have an idea of a global

time or clock while asynchronous protocols do not have time-based node coordina- tion Symmetric protocols utilize one duty cycle schedule, which may or may not be synchronized Asymmetric protocol nodes may have different schedules This means that one node may be listening 10% of the time while another node listens 25% of the time

2.2.1 Slotted Protocols

To avoid the overhead associated with a contention-based protocol, a TDMA (Time Division Multiple Access) alternative allows nodes to communicate with a reduced

chance of collisions by designating time slots TRAMA [27] is a collision-free MAC

protocol that uses a distributed election scheme to fill time slots based on current network traffic Nodes unable to use their slots are not forced to listen for trans- missions that are not coming and the system does not have to waste time slots on nodes without data to transmit Since TRAMA is a scheduled protocol, it is able

14

to avoid collisions without having to be contention aware, i.e., nodes do not have to perform carrier sense prior to transmitting data Like all TDMA protocols, however, nodes need to be synchronized to schedule their time slots In addition to TRAMA, Sohrabi, et al [34] present an example suite of protocols using TDMA slots

Like a TDMA protocol, Sift [13, 14] utilizes a slotted contention window to reduce

spatially-correlated contention (sensors near a detected event will report and other- wise transmit little or no data) by only having a subset of the reporting nodes actually transmit their event This not only reduces contention, but also energy consumption and latency Nodes use non-uniform probability distributions to choose transmission slots in the contention window If a subset size of nodes have already reported the event, the node will suppress the packet The authors include the idea of combining Sift with other MAC protocols, however, this would require all nodes to receive their neighbors’ transmissions to make sure only the subset size number of reports are transmitted Also, nodes near the base station will still be needed to forward data

to the sink, preventing balanced energy consumption without an additional protocol focused on controlling consumption In our work, we assume the MAC layer does not

do in-network processing, Redundant packet suppression, aggregation and similar techniques are complimentary, but orthogonal to our approach

2.2.2 Synchronous Protocols

Flexible Power Scheduling (FPS) [12] uses a two-level architecture for saving energy

while supporting varying demand At the network layer, the distributed protocol provides coarse-grained scheduling to plan communication of entire data flows, al- lowing nodes to turn the radio off during idle times At the MAC layer, finer-grained scheduling is used to handle channel access FPS uses the combination of slots, cycles and reservations in a time division protocol to coordinate communication between

fo

Figure 2.2: S-MAC nodes participating in multiple virtual clusters take on multiple schedules

neighbors in a tree to forward data towards a sink Node schedules are adaptive, based on supply and demand of data flows in the neighborhood Nodes are only able

to transmit data packets to the next hop in reserved time slots (which can limit the throughput and increase delay) and neighbors are required to be active for the same amount of time in the reserved slots While the protocol provides opportunities to save energy in idle slots, nodes with higher demand (e.g., forwarding nodes) will still have higher duty cycles

S-MAC [42, 43] incorporates virtual clusters to coordinate node sleep schedules, and message passing to reduce latency from contention This involves fragmenting packets into smaller packets to form a burst, with one RTS-CTS exchange, much like fragmentation in 802.11 S-MAC does not attempt to reduce the energy consumption

of vital forwarding nodes In addition, nodes on the edges of the virtual clusters must listen for all cluster times they participate in To assist the border nodes that must

take on multiple schedules and reduce latencies, Li, et al [19] present two algorithms

The first, global schedule algorithm (GSA), is an algorithm developed to allow nodes

to agree upon a single schedule for the entire network While this solves the increased listening burden for border nodes, it is still not able to assist those nodes with a high

forwarding burden The second algorithm, fast path algorithm (FPA), reduces data

16

delivery latencies by waking up nodes on the path from source to sink, If these wake-up times are in addition to the already scheduled listening times, the nodes providing forwarding functionality to the network will again have an additional load placed upon them

T-MAC [89], a protocol built upon S-MAC, adds the concept of an idle time threshold during an active cycle If an event does not occur within a threshold, the radio transitions back to the sleep state earlier, thereby decreasing power consump- tion While this does vary node schedules, nodes are still expected to be active based upon a set schedule so that neighbors can predict their cycle This makes the protocol not truly asymmetric in the sense that we describe

The authors of Clique-Based Randomized Multiple Access (CRMA) [9] present a hybrid proactive-reactive protocol CRMA nodes form 1-hop cliques with neighbors

to perform proactive scheduling Communication between these cliques is reactive with randomized access to avoid collisions Additionally, using local information, nodes can predict medium access conflicts and work to resolve them before collisions occur Since the protocol is synchronous with slotted communication scheduling within cliques, the nodes must coordinate slot start times to remain on schedule In addition, since cliques all agree on communication slots, they will all have the same

duty cycle lengths (symmetric scheduling) Again, nodes with heavier forwarding

burdens cannot offload overhead to balance energy consumption

Despite being a synchronized protocol, PMAC [46] (Pattern-MAC) utilizes local

traffic information to adapt node schedules to their load by adjusting the number of active slots If a node is not participating in nearby transmissions, it uses a TCP-like slow-start algorithm to increase its sleeping time Instead of splitting time into inter- vals, the authors introduce periods During each period, a node follows a repeating slot activity pattern (e.g., [off-off-off-on, off-off-off-on]) for the period If the load

changes, the node can update the pattern to account for increased traffic At the end

of each period, a node resynchronizes with its neighbors by broadcasting a logical time if it does not receive a neighbor’s broadcast first Nodes also broadcast their next interval’s pattern, allowing the neighbors to know when the node is available for communication While this protocol does adjust the node schedules based on their load, it does not assist nodes that have a high forwarding load Additionally, the au- thors claim that loose synchronization, like in S-MAC, is all that is needed If a node hears a synchronization packet, it adjusts its clock to that schedule Nodes broadcast their time if no synchronization packets are received If multiple synchronization broadcasts are received, the node remembers and follows all of them, potentially further increasing the burden on forwarding nodes who must be awake when their neighbors are active

Introducing duty cycles to reduce idle listening impacts delivery performance Nodes are unable to send and receive packets when sleeping, thus data packets must

be buffered until the nodes are active To reduce buffering time imposed by sleeping nodes, Lu et al developed DMAC [20] to reduce delivery delay for a static network with many sources sending to a single sink Nodes in the network have staggered active schedules depending on their depth from the sink in the tree Nodes can also adapt their schedule to the local load through traffic prediction and the use of More

To Send (MTS) packets MTS packets alert later nodes in the data path of needed transmissions The focus is to reduce data packet delay while conserving energy DMAC does not take the additional load into account for the nodes forwarding data packets to the sink For example, if a node receives a packet to forward in its active period, it predicts its other neighbors will also have data for it to send to the sink Thus it adds another slot to listen for data This increases the listening time for the forwarding nodes

18

2.2.3 Asynchronous Protocols

Asynchronous protocols [37, 21, 25] are motivated by clock synchronization overhead Tseng, et al [37] present three power management protocols for the power-save mode

of IEEE 802.11 mobile ad hoc networks (MANETs) The first is dominating-awake-

interval where nodes are active at least 50% of each interval to guarantee overlapping active cycles between neighbors This protocol is not energy efficient, especially with light traffic The second is periodically-fully-awake-interval In this protocol, nodes

interleave intervals where they are active the entire interval between intervals where

they are active a minimum amount of time This reduces consumed energy, but may increase delay and node discovery The final protocol presented in this paper

is gquorum-based Nodes use a quorum to design their wakeup patterns so two nodes can be guaranteed to have overlapping active intervals The quorum-based protocol reduces energy consumption, but may consume additional energy deciding on and sending node schedules and increase delays if the intervals are long

Polastre, et al [25] introduce B-MAC, an asynchronous-symmetric protocol for low power listening B-MAC nodes periodically check for transmissions using a de- fined interval To ensure nodes receive transmissions, B-MAC senders transmit a preamble that takes longer than the check interval These preambles alert neighbors

of an upcoming data transmission Each time a node needs to send data it must first transmit a preamble For nodes that have to send data, sources and forwarders, this requires an increased burden when compared to the sinks that do not transmit preambles Additionally, all data packets are received by all nodes who can hear the preamble; nodes located in a busy area of the network will receive extra packets Z-MAC [29] is a hybrid protocol that dynamically adapts to contention levels at nodes It is built upon B-MAC to behave like a CSMA protocol under low contention and incorporates a loosely synchronized TDMA schedule to prioritize transmissions

under higher contention The primary efect on energy comes from reducing the overhead of collisions at high contention At low contention, it offers no energy benefits over B-MAC

While SEESAw is not the first asymmetric and asynchronous MAC protocol, we do not know of any prior work demonstrating that it is possible to leverage these features

to balance network-wide energy consumption, provide acceptable data delivery, and address the problems of heavily loaded forwarding nodes

20

Chapter 3

Seesaw

SEESAW is a MAC layer protocol that balances energy consumption among wireless nodes in an ad hoc network SEESAW is an asynchronous and asymmetric proto- col, meaning nodes do not have to have coordinated clocks, nor do they all follow the same schedule The actions of sending and receiving are decoupled, allowing flexibility when determining node actions SEESAW nodes communicate probabilisti- cally; there is no guarantee packets are sent further along their path within a certain amount of time This allows nodes to conserve energy and avoid sending coordina- tion packets to coincide active cycles Although nodes communicate probabilistically,

we show SEESAW performs comparably to existing, scheduled protocols despite its asynchrony and asymmetry (Chapter 5) Finally, SEESAW nodes batch packets when possible Batching may increase delay slightly, but the main benefit is reduced control overhead For every packet added to a batch, a sending node will have to establish communication one less time Batching reduces the amount of transmitted control overhead and thereby, energy consumption Batching also decreases how often nodes vie for the shared medium, thus increasing performance

To describe SEESAW, we first present its components, further discuss its charac- teristics and walk through an overview and example We conclude this chapter with the online, distributed tuning algorithm

We implement SEESAW’s flexibility with five simple parameters: three node-dependent values and two system-wide parameters These five values determine system perfor-

3.1.1 Node Parameters

In this subsection, the three node-tunable parameters are described Please note that these descriptions include the use of the term node interval or interval (and its denotation 7) As a short preview, intervals are a subdivision of time and are fully described in the following subsection in the Node Interval bullet

e Listening Percent, /: In each interval, nodes listen for incoming transmissions for ! percent of the interval This means each node is awake for L continuous

seconds each interval, where L = 1 *% I(x) denotes the listening percent for

node z An example of a node’s listening schedule is shown in Figure 3.1

e Advertisements per Interval, a: Ignoring contention for now, when a node has data packets to transmit, it sends a advertisements (Ads) to the next hop during

an interval The only timing requirements are that all a Ads must happen within a single interval and they must be evenly spaced In other words, a node cannot push Ads into the next interval and also cannot bunch them together

in a short burst After a node empties its buffer, it can cancel Ads since it does not have any data packets to send a(x) denotes the number of Ads per interval for node x Figure 3.2 shows a sample advertisement schedule

22

Transmit

e Service Level, S: This value is the inverse of the expected time to establish communication between two nodes For a pair of communicating nodes, the service level determines the product of the nodes’ a and / values If node z is sending data to node y, a(x)*I(y) > S Increasing service levels lead to higher a

23