University of Southern California Dissertations and Theses

Adaptive energy conservation protocols for wireless ad hoc routing

(USC Thesis Other)

Adaptive energy conservation protocols for wireless ad hoc routing

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content ADAPTIVE ENERGY CONSERVATION PROTOCOLS FOR WIRELESS AD HOC ROUTING by Ya Xu 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) December 2002 Copyright 2002 Ya Xu Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. UMI Number: 3093948 UMI UMI Microform 3093948 Copyright 2003 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 Ann Arbor, Ml 48106-1346 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. UNIVERSITY OF SOUTHERN CALIFORNIA THE GRADUATE SCHOOL UNIVERSITY PARK LOS ANGELES, CALIFORNIA 90089-1695 This dissertation, written by Y a- under the direction o f h i 3 dissertation committee, and approved by all its members, has been presented to and accepted by the Director o f Graduate and Professional Programs, in partial fulfillment o f the requirements fo r the degree o f DOCTOR OF PHILOSOPHY Director D ecem ber 1 8 , 2002 Date. Dissertation Comrkittee Chair Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Dedication To my parents, my wife and my daughter For their love, understanding and support Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Acknowledgments I am deeply grateful for my advisors John Heidemann and D eborah Estrin. I thank John Heidem ann for his encouragement and support throughout my tim e of PhD research ing. I thank D eborah for her impeccable patience and consistent guidance throughout my tim e at USC, ISI, and after-ISI from 1995 to 2002. Thanks for my comm ittee members, John Heidemann, Deborah Estrin, Ram esh Govin- dan and Ahm ed Abdel-Ghaffar Helmy, for their useful discussions and comments on the dissertation, and for serving on my qualification and defense exams. Thanks also go to M aja M ataric for serving on my qualification comm ittee in addition to the above mem bers. I would like to thank all my colleagues in the VINT and SCADDS projects. Their interaction enriched my experience and inspired me. Especially, I would thank Yutaka Mori and Solomon Bien for their hard work on implem enting G A F/ CEC in the testbeds and evaluating them . Thanks also go to Dawn Johnson at ISI, who helped to bridge comm unications between my advisors and me when I spend m ost of tim e working toward PhD degree from remote. I am especially thankful for my family, my father Wu, chuanlie, my m other Xu, zhangying, my wife Nan, miying, and my daughter Xu, nancy for their love, understanding and support in my 8-year PhD journey, w ith an Outstanding Student Research Award iii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (2001) from USC and publication in A C M /IE E E International Conference on Mobile Computing and Networking(Mobicom 2001), one of the most com petitive conference in the world, while working full-tim e in research institutes and industries m ost of tim e from 1994 to 2002. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Contents D edication ii A cknow ledgm ents iii A bstract xiii 1 Introduction 1 1.1 The Energy dissipation characteristics of ad hoc n e tw o rk s ............................ 2 1.2 The Thesis ................................................................................................................... 6 1.3 C o n trib u tio n s................................................................................................................ 6 1.4 Overview of our energy conservation p ro to co ls.................................................. 8 1.5 Research G e n e ra liz a tio n ........................................................................................... 10 2 R elated W ork 11 2.1 Energy-efficient routing ........................................................................................... 12 2.1.1 S p a n ..................................................................................................................... 12 2.1.2 L E A C H .............................................................................................................. 15 2.2 Energy-efficient MAC P ro to c o ls.............................................................................. 15 2.2.1 IEEE 802.11 16 2.2.2 P A M A S .............................................................................................................. 16 2.2.3 Sparse Topology and Energy M anagem ent (S T E M ).............................. 16 2.2.4 Sensor-MAC (S -M A C ).................................................................................. 17 2.2.5 TDM A .............................................................................................................. 18 2.2.6 L IN T /L IL T ....................................................................................................... 19 2.2.7 LAZY Scheduling............................................................................................. 20 2.3 Application-specific energy c o n s e rv a tio n ............................................................. 20 2.3.1 A S C E N T .......................................................................................................... 20 2.3.2 P ic o N e t.............................................................................................................. 21 2.4 Sensitivity study in ad hoc routing protocols ................................................... 21 2.5 O ther related W o rk ...................................................................................................... 23 2.5.1 Ad hoc routing protocols ............................................................................ 23 2.5.2 Cluster-based ad-hoc routing p ro to c o ls..................................................... 24 2.5.3 Geographic Ad hoc r o u t i n g ........................................................................ 26 2.5.4 O ther examples of adaptive fid e lity ............................................................ 27 v Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3 B E C A /A F E C A fo r e n e rg y c o n s e rv a tio n 28 3.1 Basic Energy-Conserving A lg o r ith m ................................................................... 29 3.2 Adaptive Fidelity Energy-Conserving A lg o r ith m ............................................ 32 3.3 S u m m a r y ...................................................................................................................... 35 4 G e o g ra p h ic -in fo rm e d E n e rg y c o n s e rv a tio n p ro to c o l 37 4.1 Determ ining node equivalence................................................................................. 38 4.2 GAF state t r a n s it io n s ............................................................................................... 40 4.3 Tuning GAF ................................................................................................................ 41 4.4 Load balancing energy usage ................................................................................. 43 4.5 A dapting to high m o b ility ........................................................................................ 44 4.6 GAF interactions w ith ad hoc r o u t i n g ................................................................ 45 4.7 S u m m a r y ....................................................................................................................... 46 5 C lu s te r-b a s e d E n e rg y C o n s e rv a tio n (C E C ) A lg o rith m 47 5.1 D eterm ining network re d u n d a n c y .......................................................................... 50 5.2 D istributed Cluster Form ation .............................................................................. 50 5.3 Controlling the Duty Cycle of CEC N o d e s .......................................................... 54 5.4 A dapting to Network M obility .............................................................................. 55 5.5 S u m m a r y ....................................................................................................................... 57 6 E v a lu a tin g e n e rg y c o n s e rv a tio n p ro to c o ls 58 6.1 Extensions to ns-2 Sim ulator ................................................................................. 59 6.1.1 Outgoing G A F/C E C P a c k e ts .................................................................... 60 6.1.2 Incoming G A F/C E C p a c k e t s ................................................................... 61 6.1.3 Adding power control into mobile node ................................................. 63 6.1.4 Adding Energy Model into mobile n o d e ................................................. 63 6.1.5 Extending propagation m o d e l s ................................................................. 64 6.2 M etrics .......................................................................................................................... 64 6.3 Sensitivity s t u d y ......................................................................................................... 65 6.3.1 Mobility m o d e l............................................................................................... 67 6.3.2 Network c a p a c ity ............................................................................................ 67 6.3.3 Energy m o d e l................................................................................................... 68 6.3.4 Propagation m o d e l......................................................................................... 68 6.3.5 Location Error(G A F only) .................... 70 6.3.6 Node d e n s i t y ................................................................................................... 70 6.4 S u m m a r y ....................................................................................................................... 70 7 E v a lu a tio n o f B E C A /A F E C A 72 7.1 Sim ulation d e s ig n ......................................................................................................... 72 7.2 Experim ental s c e n a r io ............................................................................................... 74 7.3 BECA performance evaluation .............................................................................. 75 7.4 AFECA perform ance evaluation ........................................................................... 79 7.5 Evaluation w ith lim ited e n e r g y .............................................................................. 81 7.6 S u m m a r y ....................................................................................................................... 84 vi Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 8 E v a lu a tin g G A F 87 8.1 GAF Analytic perform ance analysis ..................................................................... 87 8.2 Sim ulation m e th o d o lo g y ............................................................................................ 89 8.3 GAF extension of network lif e tim e ........................................................................ 92 8.4 GAF energy s a v i n g s ................................................................................................... 94 8.5 GAF effects on d ata delivery .................................................................................. 95 8.6 GAF perform ance under high m o b ility ................................................................. 95 8.7 How network density affects G A F ............................................................................ 98 8.8 Sensitivity to shadowing m o d e l ................................................................................... 102 8.9 S u m m a r y ............................................................................................................................ 106 9 C E C P e rfo rm a n c e E v a lu a tio n 107 9.1 CEC Analytic P e rfo rm a n c e .......................................................................................... 107 9.1.1 CEC Protocol Message Analysis .................................................................107 9.1.2 Network Lifetime E x te n sio n ...........................................................................108 9.1.3 A dapting to Network D ynam ics....................................................................109 9.1.4 Sim ulation M ethodology..................................................................................110 9.2 Energy C o n se rv a tio n ........................................................................................................113 9.3 CEC protocol overhead .................................................................................................114 9.4 Extended network life tim e ..............................................................................................115 9.5 Routing F i d e l i t y ...............................................................................................................119 9.6 Sensitivity to network d e n s ity .......................................................................................123 9.7 Sensitivity to tim e-varying shadowing m o d e l............................................................125 9.8 S u m m a r y ............................................................................................................................ 129 10 I m p le m e n ta tio n a n d E v a lu a tio n o f e n e rg y c o n s e rv a tio n p ro to c o ls 131 10.1 GAF im p lem en tatio n .......................................................................................................132 10.1.1 GAF test-bed overview ..................................................................................132 10.1.2 Radio transm ission r a n g e .............................................................................. 133 10.1.3 Network T o p o lo g y ............................................................................................135 10.1.4 Energy d is s ip a tio n ............................................................................................138 10.1.5 Packet re a c h a b ility ............................................................................................138 10.1.6 Packet delivery latency ..................................................................................140 10.1.7 GAF evaluation S u m m a ry .............................................................................. 140 10.2 CEC im p lem en tatio n .......................................................................................................142 10.2.1 CEC test-bed overview ..................................................................................... 142 10.2.2 Extension of Network L if e tim e .................................................................... 143 10.2.3 D ata Delivery R a t i o .........................................................................................145 10.2.4 CEC evaluation S u m m a ry ...............................................................................147 10.3 S u m m a r y ........................................................................................................................... 147 11 C o n c lu sio n s a n d F u tu r e W o rk 149 11.1 Thesis S u m m a r y .............................................................................................................149 11.1.1 Self-configuring mechanisms enable robust p r o to c o ls ............................149 11.1.2 Localized, distributed algorithm s can provide energy-efficient design 150 11.1.3 Sensitivity study is critical for protocol design and evaluation . . . 151 vii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 11.1.4 G A F /C E C is practical and im plem entable in the real world . . . . 152 11.2 Thesis C o n trib u tio n s ......................................................................................................153 11.3 Future Work ............................................................................................................... 155 11.3.1 Coordination of ad hoc routing protocols and energy conservation protocols ..............................................................................................................155 11.3.2 Balance energy conservation w ith protocol r o b u s tn e s s .......................... 156 11.3.3 Joint work between Energy conservation protocols and energy-efficient MAC .....................................................................................................................158 11.3.4 Building tools for protocol developm ents......................................................158 R eference List 158 viii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. List Of Tables 7.1 Com parison of BECA loss rates for different values of Ts................................. 75 7.2 Com parison of loss rates for AODV, BECA (Ts = 10s), and AFECA (k = 10s).......................................................................................................................... 79 ix Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. List Of Figures 1.1 Com parison of energy consumed for different ad hoc routing protocols . . 3 1.2 Exam ple of node redundancy in ad hoc routing.................................................. 5 3.1 States in B E C A ........................................................................................................... 30 3.2 Establishing routing between to an adjacent (but possibly sleeping) node. 31 3.3 “H” topology: An example of extremely non-uniform topologies ............... 35 4.1 The problem of node equivalence.............................................................................. 38 4.2 Exam ple of virtual grid in G A F................................................................................ 39 4.3 State transitions in G A F.............................................................................................. 40 5.1 Exam ple of how network connectivity is broken in GAF under non-uniform ally deployed network........................................................................................................... 47 5.2 Exam ple of CEC cluster fo rm a tio n ........................................................................ 49 6.1 Schematic of a mobile node w ith G A F/C E C extension in n s ......................... 62 7.1 Com parison of BECA and A O D V ......................................................................... 77 7.2 P E comparison of AFECA w ith different values of k for com puting T s a - ■ 79 7.3 Comparisons of AODV, BECA (Ts = 10s), and AFECA (k = 10s)............... 80 7.4 P E comparison of unm odified AODV, BECA, and A FECA under different traffic loads w ith lim ited energy............................................................................... 82 7.5 Network lifetime comparison of AODV, BECA and A F E C A ......................... 83 x Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 7.6 Network lifetime w ith increasing node density................................................... 85 8.1 Network lifetime comparison of AODV and G A F ........................................... 93 8.2 Network lifetime comparison of AODV and GAF under high mobility . . 97 8.3 D ata delivery comparison of AODV and GAF .............................................. 99 8.4 Quantifying network lifetime: GAF over AODV ............................................... 101 8.5 Network lifetime changes under different node d e n s ity ..................................... 102 8.6 Com parison of packet delivery ratio under shadowing m o d el........................... 105 9.1 Percentage of CEC energy use over total system energy u s e ........................... 116 9.2 Com parison of non-zero energy node fraction over t i m e ..................................116 9.3 Com parison of network mobility im pact on network life tim e........................... 118 9.4 R outing fidelity comparison under heavy tra ffic ...................................................120 9.5 D ata Delivery comparison under m obility m o d e l ................................................122 9.6 Network lifetime comparison among CEC, GAF and AODV under differ ent node d e n s ity ...............................................................................................................124 9.7 Packet delivery ratio comparison of CEC and AODV under time-varying shadowing m o d e l..............................................................................................................127 9.8 Packet delivery ratio comparison of revised CEC and AODV under time- varying shadowing m o d e l .............................................................................................129 10.1 GAF im plem entation a r c h ite c tu r e ........................................................................... 132 10.2 Radio Range of node X ............................................................................................... 134 10.3 Radio Range M e a s u re m e n t.........................................................................................135 10.4 Node location and Grid s e tu p ..................................................................................... 136 10.5 Conumed E n e rg y ............................................................................................................. 137 10.6 Conumed E n e rg y ............................................................................................................. 139 xi Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 10.7 Conumed E n e rg y .............................................................................................................141 10.8 Extension of network lifetime (95% confidence i n te r v a l ) ....................................143 10.9 CEC active nodes over tim e (95% confidence in te r v a l) ......................................144 10.10 D ata delivery ratio (95% confidence in te r v a l) ....................................................... 146 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Abstract M ultihop, ad hoc networking has been the focus of many recent research and devel opment efforts. In ad hoc networks such as sensor networks, energy use m aps directly to lifetime and utility, thus energy use becomes a very im portant m etric for wireless ad hoc networks. My thesis is th a t energy conservation protocols can be designed and im plem ented for ad hoc networks to extend network lifetime while m aintaining routing fidelity. Our protocols exploit the network redundancy to achieve the goal of conserving energy to extend network lifetime. Once network redundancy is identified, our protocols power off the redundant p art to conserve energy while adaptively controlling the duty cycle of network nodes to m aintain routing fidelity. In order to maximize performance for energy conservation, we explore different de signs to develop localized, self-configuring protocols to extend network lifetime. Basic Energy-Conserving Algorithm (BECA) and A daptive Fidelity Energy-Conserving Al gorithm (AFECA) take advantage of ad hoc routing protocols to conserve energy and introduce the concept of using network density to find network redundancy. Geographic- informed Adaptive Fidelity (GAF) self-configures redundant nodes into small groups based on their location and uses localized, distributed algorithm s to control node duty cycle to extend network lifetime. Cluster-based Energy Conservation (CEC) continues xiii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. G A F’s effort in developing adaptive, localized, distributed and self-configuring algorithm s while elim inating the dependency on location systems. In order to make our protocols robust, we study the protocol sensitivity to different network dynamics factors such as m obility model, network capacity, propagation model, energy dissipation model, location error model and network density, through analysis, simulations, and im plem entations. We also provide guidance on how to correctly and reasonably use these models in the wireless ad hoc network research for some types of applications. The sensitivity analysis can apply to other protocol developments and studies in ad hoc network area. xiv Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 1 Introduction M ultihop, ad hoc networking has been the focus of many recent research and development efforts. W ireless networks and m ultihop routing have applications in military, commercial, and educational environm ents including wireless office LAN connections, home networks of devices, and sensor networks. A num ber of routing protocols have been proposed to provide m ulti-hop communica tion in wireless, ad hoc networks [40, 9, 41, 38]. Traditionally these protocols are evaluated in term s of packet loss rates, routing message overhead, and route length [8, 29, 15]. Since ad hoc networks will often be deployed using battery-powered nodes, comparison and op tim ization of protocol energy consum ption are also im portant (as suggested for future work by some researchers [29]). Reducing energy consum ption has been a hot research area in ad hoc networks. W hile past research aimed at low power design at the hardw are level, the recent research has dem onstrated th at significant energy conservation can be achieved at the routing level by powering off nodes th a t are redundant from a routing perspective [57, 56, 14]. 1 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In order to develop good schemes for energy conservation, we m ust understand the energy dissipation characteristics in ad hoc networks. This chapter presents our investi gation results of energy dissipation in ad hoc networks, gives an overview of our energy conservation approaches and defines my thesis. 1.1 The Energy dissipation characteristics of ad hoc networks W hen ad hoc networks are deployed using battery-pow ered nodes, the im portant question of how lim ited energy resources affect system lifetime and overall performance becomes critical. For scenarios such as sensor networks where energy use m aps directly to lifetime and utility, energy use is the im portant metric. To understand energy efficiency we examined existing ad hoc routing protocols using models of Lucent WaveLAN direct sequence spread spectrum radio w ith the IE E E 802.11-1997 protocol w ith representative models of energy consum ption [53] and radio propagation [8]. We first only consider energy cost due to packet transm ission or reception. Such costs may also include energy dissipation in MAC-level retransm issions, R T S/C T S etc. We studied energy consum ption of four ad hoc routing protocols (AODV, DSR, DSDV, and TO RA ) w ith a simple traffic model where a few nodes send d ata over a m ulti-hop path [56] (Figure 1.1). W ith this energy model we found th a t on-dem and protocols such as AODV and DSR consume much less energy th an a priori protocols such as DSDV (the left, dark bars in Figure 1.1). A priori protocols are constantly expending energy pre-com puting routes, even though there is no traffic passing on these routes. 2 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 30 AODV DSR DSDV TORA ad hoc routing protocol Figure 1.1: Com parison of energy consumed for four ad hoc routing protocols w ith dif ferent energy models (left, black bars are w ithout considering energy consumed when listening; right, gray bars include this consumption). The sim ulation has 50 nodes in a 1500m*300m area. Nodes move according to the random way-point model. The energy model is based on Stemm and K atz [53]. In other words, on-dem and protocols, by their very nature, are more efficient in the energy consumed by routing overhead packets. As a result, energy use is dom inated by routing protocol overhead. In fact, the m ajor source of extraneous energy consum ption was from overhearing, as previously observed in PAM AS [51]. Radios have a relatively large broadcast range. All nodes in th a t range m ust receive each packet to determ ine if it is to be forwarded or received locally. A lthough most of these packets are immedi ately discarded, they consume energy w ith this simple energy model. This observation m otivates approaches th at avoid overhearing. The PAMAS protocol suggests a MAC- layer approach to minimize this cost [52]; TDM A protocols would also be applicable (for example [42]). 3 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Actual Radios consume power not only when sending and receiving, b u t also when lis tening or idle (the radio electronics m ust be powered and decoding to detect the presence of an incoming packet). Research [53, 30] shows th at idle energy dissipation can not be ignored in comparing to sending and receiving energy dissipation. Stemm and K atz show idle:receive:transm it ratios are 1:1.05:1.4 by m easurem ent [53], while more recent studies show ratios of 1:2:2.5 [30] and 1:1.2:1.7 [14]. In any of these cases, energy dissipation in idle state can not be ignored. W ith such energy model, all ad hoc routing protocols considered consume roughly the same am ount of energy (within a few percent) as shown in the grey bars in Figure 1.1. In the scenario with modest traffic, idle tim e completely dom inates system energy consum ption. The studies based on an energy model th a t considers energy dissipation in sent/ received packets and idle tim e, suggest th a t energy optim ization m ust tu rn off the radio, not sim ply reduce packet transm ission and reception. Powering off radio conserves energy both in overhearing due to d ata transfer, and in idle state energy dissipation when no traffic exists. We therefore explore nodes th a t power off their radios much of the tim e. This approach is similar to the use of TDM A for power conservation [42], or PAMAS [51]. However, unlike these approaches, we employ inform ation from above the MAC-layer to control radio power. (We make use of the power m anagement controls in IEEE 802.11 to control power1.) The application- or routing-layers provide b etter inform ation about when the radio is not needed. 1802.11 supports power saving mode in both infrastructure network and ad hoc network. Note that powering on/off MAC is just like node moving in/out communication range with other nodes. RTS/CTS is still used in unicast communication to address hidden terminal issue. [53] shows that time for 802.11 MAC on/off is in a few milliseconds. In other words, powering on/off MAC does not affect normal 802.11 operation. 4 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. © nominal radio range Figure 1.2: Exam ple of node redundancy in ad hoc routing. On the other hand, we observed th a t when there is significant node redundancy in an ad-hoc network, m ultiple paths exist between nodes. Thus we can power off some interm ediate nodes while still m aintaining connectivity. For example, In Figure 1.2, if node 2 is awake, nodes 3 and 4 are extraneous for comm unication between 1 and 5. We define routing fidelity as uninterrupted connectivity between com m unicating nodes. Thus routing fidelity (that 1 and 5 can comm unicate) can be m aintained as long as any interm ediate node is awake. In summary, our analysis shows two basic facts in ad hoc networks: first, there is sig nificant node redundancy in many ad-hoc networks in th at m ultiple paths exist between nodes. Thus we can power off some interm ediate nodes while still m aintaining connec tivity. Secondly, actual radios consume power not only when sending and receiving, but also when listening or idle (the radio electronics m ust be powered and decoding to detect the presence of an incoming packet). Studies [56, 57, 21, 51, 52] suggest th a t energy optim ization m ust tu rn off the radio, not only reduce packet transm ission and reception. 5 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1.2 The Thesis The above observations suggest th a t we can conserve energy to extend network lifetime by turning off the redundant nodes in the network. This is the general approach in our energy conservation protocols. The challenges are: first, how to identify network redundancy; second, after network redundancy is discovered, how to control the duty cycle of redundant nodes to conserve energy; third, how to m aintain routing fidelity, the uninterrupted connectivity between comm unicating nodes, when redundant nodes are powered off in a dynamic network. This thesis argues th a t energy conservation protocols can be designed and imple m ented for ad hoc networks to extend network lifetime while m aintaining routing fidelity. We focus on two key factors to design energy conservation protocols in order to operate successfully in ad hoc networks. First, the protocol must be self-configuring, meaning th at it m ust actively, m easure the networks in order to work robustly and react to the network dynamics. Second, the protocol m ust find the redundant p art of the networks in a distributed and localized fashion and power it off in order to conserve energy, since it is extremely expensive or impossible to m aintain and distribute a consistent state across all nodes in rapidly changing ad hoc networks. 1.3 Contributions Our prim ary contributions are: 1. The use of application- and system -inform ation to tu rn off node radios for extended periods of time. Node duty cycles are influenced by application end-points and 6 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. node movement patters to preserve comm unication fidelity. Predictions about node lifetimes allow energy-conscious load balancing. This property also makes our protocols independent of ad hoc routing protocols because only the application or system-level inform ation, not the routing informa tion, is used to control node duty cycle. In principle, Our protocols can run over any ad hoc routing protocols. There is no need to add energy m etrics into routing algorithm s to complicate the routing protocols. 2. The use of node deployment density to adaptively adjust routing fidelity. Routing redundancy is correlated w ith denser node deployment (when many nodes can hear each o th er). We show how to use this inform ation to increase node duty cycles and to extend the lifetime of the network as a whole. 3. The use of node local coverage information to lim it protocol overhead locally. Nodes only exchange control messages w ith their direct neighbors which is very im portant in the resource-lim ited (in term s of both energy and bandw idth) ad hoc networks. 4. The m ethodology to evaluate the protocols’ sensitivity to different network dy namics factors in wireless ad hoc networks. Through sensitivity studies, we can identify w hat network dynamic factors have the m ajor im pact on our protocols. We also gain insights into the strengths and weakness in our protocols to help us understand applicability of our protocols and identify further research focus. We study the protocols’ sensitivities to different factors such as m obility model, net work capacity, propagation model, energy dissipation model, location error model 7 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. and network density. The methodology can be applied to other wireless ad hoc protocol developments. We describe and analyze four protocols: Basic Energy-Conserving A lgorithm (BECA) and Adaptive Fidelity Energy-Conserving Algorithm (AFECA) [56] take advantage of ad hoc routing protocols to conserve energy and introduce the concept of using net work density to find network redundancy. Geographical Adaptive Fidelity (GAF) [57] self-configures redundant nodes into small groups based on their locations and uses lo calized, distributed algorithm s to control node duty cycle to extend network lifetime. Cluster-based Energy Conservation (CEC) continues G A F’s effort in adaptive, localized, distributed and self-configuring algorithm s while elim inating the dependency on location systems. 1.4 Overview of our energy conservation protocols Our goal has been developing protocols for conserving energy consum ption of routing and extending network lifetime. We designed energy-conservation protocols such as BECA (Section 3) which work very closely w ith particular ad hoc routing protocols such as AODV, DSR to tu rn off the radio in order to reduce energy consumption. We further developed AFECA [56] which adds the additional use of node deployment density to adaptively adjust routing fidelity to extend network lifetime. Each AFECA node m aintains a count of the num ber of nodes w ithin its radio range, obtained by listening to transm issions on the channel. A node switches between sleeping and listening, w ith random ized sleep tim e proportional to the num ber of nearby nodes. 8 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The net effect is th at the num ber of listening nodes is roughly constant, regardless of node density; as the density increases, more energy can be conserved. A FEC A ’s constants are chosen so th at there is a high probability th a t the listening nodes form a connected graph, so th at ad hoc forwarding works. An AFECA node does not know w hether it is required to listen in order to m aintain connectivity, so to be conservative AFECA tends to make nodes listen even when they could be asleep. The im plem entation of BECA and AFECA depends on the particular ad hoc routing protocols. More accurately, they are designed for reactive ad hoc routing protocols such as AODV and DSR. We further generalize our effort in energy conservation protocols so th at the protocols do not need to depend on what ad hoc routing protocols are used. In addition, we accurately estim ate how many nodes are required to listen in order to m aintain connectivity. This is the basic idea in designing GAF, Geographical Adaptive Fidelity [57]. Each GAF node uses location inform ation to associate itself w ith a “virtual grid” , where all nodes in a particular grid square are equivalent w ith respect to forwarding packets (Section 4.1); Nodes in the same grid then coordinate w ith each other to determ ine who will sleep and how long(Section 4.2); this determ ination is m oderated by application and system information. Nodes then periodically wake up and trade places to accomplish load balancing (Section 4.4 and 4.5). We also consider how GAF interacts w ith the underlying ad hoc routing protocol in Section 4.6. GAF takes advantage of global location inform ation systems such as GPS to achieve its goal of conserving energy. However, in many applications, the location inform ation 9 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. system is not always available, such as indoors or under trees where GPS does not work. The dependency on global location inform ation lim its G A F’s applicability. This m otivates Cluster-based, Energy Conservation (CEC), a cluster-based protocol th a t exploits network redundancy by self-configuring equivalent nodes into clusters w ith independent controls. Different from GAF, CEC does not need support from global ge ographic location systems, routing protocols or other protocols. CEC directly measures network connectivity by itself, m aking it robust to connectivity error due to obstructions. Com pared w ith G A F’s conservative connectivity assum ption, the direct connectivity mea surement can find network redundancy more accurately so th at it could conserve more energy. We discuss the details of CEC algorithm in Section 5. 1.5 Research Generalization Our work is one example of adaptive fidelity [17] and RTCP [49] adaptive frequency techniques. O ther examples include beacon density for localization [10]. More generally, we wish to design self-configuring networks th a t exploit redundancy to conserve energy while preserving the fidelity of network applications. 10 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 2 Related Work Reducing energy consum ption has been a recent focus of wireless ad hoc network research and development. There are many approaches to try to address energy conservation from different research groups. From the view of protocol architecture, these approaches can be classified as MAC- based approaches and routing-based approaches. From the view of applications, these approaches can be classified as application-specific approaches and general-purpose approaches. Three of the protocols presented in this paper, Basic Energy-Conserving Algorithm (BECA) and A daptive Fidelity Energy-Conserving Algorithm (AFECA), Geographical Adaptive Fidelity (GAF) and Cluster-based Energy Conservation (CEC) can be treated as routing-based approaches. The three protocols identify network redundancy from routing perspective. They also m aintain routing fidelity while powering off nodes for energy conservation. Span [14] is another example of routing-based energy conservation protocol. Energy-efficient MAC has been the goal of low power hardware design. Past research includes Tim e Division M ultiple Access (TDM A) [42], and 802.11 (power m anagement) 11 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. etc. Recent research includes Sparse Topology and Energy M anagem ent (STEM ) [50] and Sensor-MAC(S-MAC) [58]. Our energy conservation protocols are all general-purpose protocols. They can be used as infrastructure protocols and can be used by all kinds of applications. O ther protocols, such as Span, 802.11, and TDMA also belong to this category. Some protocols are application-specific. Adaptive Self-Configuring sEnsor Networks Topologies (ASCENT) [11] is an adaptive self-configuration protocol designed for sensor networks. Some MAC protocols, such as STEM and S-MAC, they are all specifically designed for sensor network applications. 2.1 Energy-efficient routing Energy-efficient routing protocols exchange inform ation among nodes so th a t each node can build up a view of the network (or p art of the network). The decisions for energy conservation are based on network redundancy from a routing perspective so th a t routing fidelity can be m aintained. 2.1.1 S p an M IT’s Span [14] is an energy-efficient coordination algorithm for topology m aintenance in ad hoc wireless networks. It has the same goal as GAF. In Span, a lim ited set of nodes forms a m ulti-hop forwarding backbone, which tries to preserve the original capacity of the underlying ad-hoc network. O ther nodes transition to sleep states more frequently, as they no longer carry the burden of forwarding d ata of other nodes. To balance out 12 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. energy consum ption, the backbone functionality is rotated between nodes, and as such there is a strong interaction w ith the routing layer. W hile Span does not need any location inform ation support, it is dependent on in form ation provided by a particular routing protocol, such as geographic forwarding algo rithm . In Span, each node “consults state stored in local routing tables” to decide node’s role in keeping network connectivity. Furtherm ore, Span piggybacks routing packets to carry Span control messages. The dependency on routing protocol also lim its Span’s ap plicability: First, not all routing protocols can provide the inform ation th at Span needs. Second, for some applications like sensor net, there is no traditional routing protocols to support Span. Although CEC and Span share some common properties such as self-configuration, localized algorithm s and balanced energy usage among nodes, CEC differs from Span fundam entally in the following aspects: P rotocol robustness. Span aims to build up a backbone network by powering off redundant nodes. A lthough a backbone network approach seems to achieve m axim um energy conservation, it is vulnerable to network dynamics, especially in ad hoc networks. Such approaches will not work robustly under high m obility or dynamic ratio propagation environment. CEC takes a different approach. It uses node coverage inform ation to form clusters. It does not aim to build up a backbone network. Instead, CEC allows some level of network redundancy for robustness purpose. Simulations and im plem entations have shown th at 13 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CEC works well even under high mobility, high traffic load, dynamic radio propagation conditions, which are not fully investigated by Span. Independent on routing protocols. One of the design ideas of CEC (and GAF) is to keep it independent on routing protocols. The benefit is th at they can work w ith any existing ad hoc routing protocols. CEC (and GAF) have been im plem ented w ith different ad hoc routing protocols, such as AODV and DSR, and have been proved its portability by being integrated w ith STEM . Span still needs inform ation from a geographic forwarding algorithm as we stated above. O verhead control m echanism s. Each CEC node only needs to pass its own state about itself to its adjacent nodes w ithin a hop. Simulations have show th a t CEC control message overhead only uses up to 0.4% of total system energy usage. Each span node needs to pass its own state and its neighbor’s state in two hops. So far, Span only shows its performance by piggybacking Span control message over routing messages. Such piggyback mechanism further lim its Span’s use w ith different routing protocols. M obility prediction m echanism s. CEC (and GAF) use m obility prediction mecha nism to help them work robustly under high m obility environment. Span does not have such mechanism. 14 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2.1.2 LEACH LEACH [23] proposes a cluster-based routing protocol th at reduces energy dissipation from a global perspective and extends system lifetime by distributing the load to all the nodes at different points in time. W hile LEACH shares many properties w ith CEC proposed in this paper, such as self-organizing, adaptive clustering, and load sharing, CEC differs from LEACH in three areas. First, unlike LEACH, CEC does not require radio transm ission in different power levels. LEACH assumes high-energy transm ission at the cluster head to reach the base station in one hop. Second, CEC exploits network redundancy to conserve energy while m aintaining routing fidelity. It can work w ith any ad hoc routing protocols to achieve the goal of energy-efficient routing. LEACH is a routing protocol for the sensor networks. 2.2 Energy-efficient M AC Protocols Energy-efficient MAC is a hardw are approach for energy conservation. MAC protocols conserve energy by powering off nodes or controlling transm it power. Different from the energy-efficient routings, nodes using energy-efficient MAC do not have the view of net work topology. Therefore, they typically trade-off network delay for energy conservation. Energy-efficient MAC and Energy-efficient routing are com plem entary each other. The applications can use both of them together for b etter energy conservation. 15 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2.2.1 IEEE 802.11 IEEE 802.11 [33] supports ad hoc network configuration: mobile nodes are brought to gether to form a network on the fly. Recent work [53] has shown th a t its energy con sum ption is high due to overhearing and idle listening under norm al working condition. IEEE 802.11 also provides power m anagement controls to allow disabling the transceiver to conserve energy. A lthough it specifies how to tu rn off the radio, they do not discuss specific policies. We propose these policies assuming the presence of 802.11-like controls for basic and adaptive cases. 2.2.2 PAMAS PAMAS is a MAC-level protocol where radios power off when not actively transm itting or receiving packets [51, 52] by using out-of-channel signaling. PAMAS requires two independent radio channels in order to sense neighbor node activity. PAMAS avoids the overhearing problem we discuss in Section 1, but it does not address the problem of energy consum ption when nodes are idle. Solutions to overhearing are relevant, but for radios w ith high idle power consum ption work such as we propose will be necessary. 2.2.3 Sparse Topology and Energy Management (STEM) STEM [50] trades off power savings versus path setup latency in sensor networks. It emulates a paging channel by having a separate radio operating at a lower duty cycle. Upon receiving a wakeup message, it turns on the prim ary radio, which takes care of the regular d ata transmissions. 16 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. STEM is a pure MAC layer approach. Each node only controls its own behavior. Different from G A F/C E C , each node does not have the view of the network. STEM reactively establishes capacity on dem and only when it is needed. STEM can operate on top of the topology th a t G A F/C E C create to further reduce energy use for those scenarios where the networks, such as sensor networks, spend m ost of their tim e waiting for events to happen, instead of forwarding traffic. W hile dual channel MAC is required for STEM , its approach to energy conservation is orthogonal to C E C /G A F in general. STEM has been integrated w ith GAF and dem onstrated much b etter perform ance th an each of them used individually. STEM is also aim ed to sensor networks. It does not try to keep the routing fidelity, especially packet transm ission delay. Both G A F/C E C can keep the same packet delivery delay even when redundant nodes are powered off. 2.2.4 Sensor-MAC (S-MAC) S-MAC [58], as indicated by its name, is a MAC protocol explicitly designed for wireless sensor networks. S-MAC sets its prim ary goal as energy conservation while per-node fairness and latency are less im portant which are critical for traditional MACs such as 802.11. S-MAC reduces energy consum ption in idle listening by periodically powering off nodes. Each node goes to sleep for some tim e, and then wakes up and listens to see if any other nodes want to talk to it. S-MAC also sets the radio to sleep during transm issions of other nodes by using in-channel signaling. 17 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Such scheme requires periodic synchronization among neighbor nodes. Different from STEM, which em ulates a paging channel by having a separate radio operating at a lower duty cycle, S-MAC synchronizes tim ers by relative tim estam p technique and significantly long listen period in the same channel. A pparently the latency is increased due to the periodic sleep of each node in the S-MAC (and STEM ). Moreover, the delay can accum ulate on each hop. Although energy-efficient routing periodically powers off nodes too, it does not nec essarily cause longer latency. The m ajor difference is th at energy-efficient routing only powers off the redundant nodes while S-MAC (and STEM) periodically powers off all nodes. Because energy-efficient routing and energy-efficient MAC sense the network in different level, the energy-efficient routing can find network redundancy from routing perspective while energy-efficient MAC does not have such knowledge. This is the fun dam ental difference between energy-efficient routing and energy-efficient MAC. Like STEM , S-MAC should be able to work w ith any energy-efficient routing schemes to provide b etter energy conservation for applications. 2.2.5 TDM A TDMA protocols have been proposed to reduce energy consum ption in sensor networks [42] By reducing the duty cycle these protocols can trade idle-time energy consum ption for latency. We believe TDM A MAC protocols will very im portant for power-constrained networks. A lthough we have not yet exam ined use of our approaches over TDM A proto cols, our use of application-level inform ation and node density can further improve power conservation. 18 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2.2.6 LINT/LILT Ram R am anathan and Regina Rosales-Hain investigated the problem of creating a desired topology by adjusting transm it power in ad hoc networks [44]. Ideally, the topology control can reduce energy use in dense network by controlling the num ber of neighbors of each node, and improve connectivity in a sparse network. Its two mobile protocols, LINT and LILT, use sim ilar approach as AFECA [56] to collect neighbor inform ation from routing protocols and attem pt to keep some level of num ber of neighbors for each node. W hile we believe th a t the transm it power control can adaptively conserve energy (presumably the idle energy cost can be reduced too) if the hardw are provides enough support, we are concerned w ith the im pact on the ad hoc routing protocols. Intuitively, the hop-by-hop transm ission w ith restrained transm it power may lead to fragile route due to mobility. The paper does not give packet drops comparison w ith existing ad hoc routings, nor does the comparison under different m obility model. Although G A F /C E C control network topology for energy conservation, they use to tally different approaches for this purpose. G A F /C E C power off redundant nodes for energy conservation while LIN T/LILT changes transm it power. As we discussed above, lowering transm it power does not help reducing energy usage in idle time. In addition, network connectivity will become more vulnerable to m obility due to changed trans m it power, LIN T/LILT need m obility prediction mechanisms like G A F/C E C to adapt to network mobility. Overall, LIN T/LILT can be treated as com petitive approaches to G A F/C E C in energy conservation. 19 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2.2.7 LAZY Scheduling Prabhakar et al. [43] developed a “lazy” scheduling of packet transm issions for energy conservation based on the observation th a t the energy required to transm it a packet over a wireless link can be significantly reduced by lowering transm ission power and transm itting the packet over a longer period of tim e. Their scheme tries to balance energy conservation and delay constraint of packets. As we pointed out in the beginning of the paper, energy is consumed not only in the transm ission, but also in the idle tim e and receiving. The simple control on the transm ission does not fix the problem of energy dissipation in idle time. Similar to STEM , LAZY scheme is complementary to G A F/C E C and can be used w ith G A F/C E C together for further energy conservation. 2.3 Application-specific energy conservation Some protocols are explicitly designed for applications in wireless sensor networks. They use integrated solutions for applications, not ju st considering approaches from MAC layer or routing level. 2.3.1 ASCENT ASCENT measures local connectivity based on neighbor threshold and loss threshold to decide which nodes should join the routing infrastructure [11]. CEC differs from ASCENT in three areas. First, ASCENT targets for single optim ized application in sensor networks while CEC is a general purpose protocol. CEC can sim ultaneously support different 20 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. types of applications for energy conservation purpose. Second, CEC proactively preserves network capacity at all tim es while ASCENT does not assume th a t the network capacity needs to be preserved at all times. Third, CEC balances all nodes energy use while some ASCENT nodes may drain out of energy much faster than the others. 2.3.2 PicoNet PicoNet proposes an integrated design of radios, small battery powered nodes, MAC and application protocols th a t minimize power consum ption [4]. They reduce power consump tion w ith a very low, application-dependent duty cycle (their paper does not specify, but presentations suggest interm ittent polling with periods of 50 to 100s of seconds). They prim arily use local base stations instead of m ulti-hop wireless routing, and assume fre quent or continuous node movement. Their approaches are promising, but we are not aware of a detailed study of PicoNet power consum ption. O ur work differs from theirs by building on existing ad hoc routing protocols and by m aking use of adaptive fidelity to reduce power in dense node configurations. 2.4 Sensitivity study in ad hoc routing protocols To b etter understand how a protocol performance is affected by network dynamics, many researchers conduct sensitivity studies in the development of ad hoc routing protocols. M obility models have been widely used in sensitivity studies of ad hoc routing proto cols. M obility models attem p t to represent the m otion of network nodes and are used to evaluate the effect of the nodal movement on the perform ance of protocols. A so-called random waypoint m obility model is used by several authors in the ad hoc networking 21 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. comm unity [48, 8, 27, 57, 46]. In the random waypoint m obility model, Nodes alternate between pausing and then move to a random ly chosen location at a fixed speed. We use random waypoint m obility model to study GAF sensitivity to network mobil ity. Broch et al [8] use the same m obility model to evaluate routing protocols performance. Although Sanchez et al. [48] and Hu et al. [27] introduce more random m obility models in addition to the random waypoint model in their ad hoc network research, the impact of all random mobility models on the protocol performance th a t they study is alm ost the same. W hile the simple random m obility models are blam ed for unrealistic mobile behavior because the movement in the new tim e interval has no relation to the past values, there are some proposed m obility models to try to fix the problem. B ettstetter proposes a Sm ooth Random Mobility model [5] to use two stochastic principles for direction and speed control in which the new values for speed and direction are correlated to previous values. Hong et al. [26] develop a Reference Point Group M obility model to organize mobile nodes into groups according to their logical relationship. Their work show th at when an ad hoc network is deployed in a real situation, it is not sufficient to test it with random mobility model since the m otion p attern can interact in a generally positive, but sometimes negative way w ith network protocols. Radio propagation model is also an im portant factor to affect protocol performance in ad hoc networks. A determ inistic propagation model called tworayground model is widely used in ad hoc network research [8]. In reality, radio propagation is strongly affected by m ulti-path effects (fading). The determ inistic propagation model can not catch the radio property in the real world. Different from the previous work, we examine GAF sensitivity 22 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. to both determ inistic propagation model and a non-propagation model called shadowing model [45]. Our result shows th a t the shadowing model will cause the d ata delivery ratio decreased by 10%. Since we focus on energy conservation on densely distributed network, we have to examine protocol sensitivity to radio energy and network density. W hile only radio trans m it/receive energy consum ption is considered by many other researchers [51, 52], we find th at the idle tim e energy consum ption dom inates the network energy usage in 802.11 networks. We show th at an inappropriate model in radio energy dissipation causes wrong conclusions in the energy dissipation comparison research of ad hoc routing protocols in Figure 1.1. Our GAF approach depends on location system. We therefore examine G A F’s sensi tivity to location error model and find th a t GAF is not affected by small location errors. The location error model th at we use is the same as the one used by Ko and Vaidya [31] in their Location-aided Routing (LAR) research. 2.5 Other related Work Our work also builds on related work of ad hoc routing protocols, clustering algorithm s and geographic routing. 2.5.1 Ad hoc routing protocols A num ber of routing protocols have been proposed to provide m ulti-hop comm unication in wireless, ad hoc networks [40, 9, 41, 38]. Traditionally these protocols are evaluated in term s of packet loss rates, routing message overhead, and route length [8, 29, 15]. We 23 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. evaluate our protocols by these metrics for comparison, and we add measures of power consum ption and network lifetime to consider power consum ption as well. Both Chang and Tassiulas [12] and P ottie et al. [42] have recently suggested th a t one might select routes in an ad hoc network based on available energy. The effect of this work would be longer network lifetime. O ur approach is to conserve energy by powering radios off rather th an m anaging a fixed energy consum ption, so our work complements their effort. Heinzelman et al. present a set of protocols for comm unication in sensor networks based on flooding [24]. They examine the energy consum ption of these protocols and show th at suppressing duplicate transm issions of the same d ata can save power as calculated from a simple energy model (not considering energy consum ption while radios are idle). Unlike their work, we consider more accurate power models and ad hoc routing protocols rather th an flooding. These differences result in much different optim ization. We also consider optim ization based on adaptive fidelity th a t are specific to dense networks. 2.5.2 Cluster-based ad-hoc routing protocols Some ad-hoc network routing approaches [35, 13, 32, 3] employ a cluster-based philosophy. They structure the ad-hoc network as a two-level network: in the lower level, nodes in geographical proxim ity create peer-to-peer networks. In each one of these lower-level networks, at least one node is designated to serve as a “gateway” to the higher tier. These gateway nodes create the higher-level network. R outing between nodes th a t belong to different lower-level networks is through the gateway nodes. Clustering schemes are often used w ith a TDM A MAC protocol to reduce the cost of radio listening. 24 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In the Near Term Digital Radio (NTDR) project [47, 37], the radios are designed to self-organize into a dynamic two-tiered network scheme of backbone cluster heads and affiliated cluster members. D ata is routed and relayed autom atically between users on three separate frequency hopping patterns. However, the N TD R protocol is designed for long-range communications w ith large am ount of power consum ption. It does not fit into our model of lim ited energy nodes. In addition, N TD R assumes adaptive power control in transm ission to reduce interference. In the self-configuring LandM ark (LM) hierarchy [55, 39], A “landm ark” node is elected to route packets for all nodes w ithin its radio radius range. The LM nodes organize themselves into an hierarchy such th a t radius increases w ith its level and radius of the radius of the top-level LM include all nodes. Cluster-based routings do not directly address the energy dissipation issues due to the cost of radio listening and overhearing, although a TDMA protocol helps some. In addition, they introduce the new cost of hierarchy formation, and possibly additional cost by forcing comm unication through gateways even for nodes th a t could directly hear each other. The cluster heads also could become bottleneck of the network since all traffic have to go through the cluster heads. A lthough GAF finds redundant nodes w ithin each virtual grid, GAF is not a cluster- based routing algorithm . By using application-level inform ation GAF can have much lower duty cycles even th an clustering w ith a TDM A MAC. GAF also does not force the packets from the source nodes to go through all representative nodes in each grids in order to reach the destination although it is up to the ad hoc routing protocols to find the path. 25 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CEC is not a cluster-based routing algorithm . This makes CEC applicable to other protocols like d ata diffusion where routing protocol based on node addresses is not desired. W hile the selected cluster-heads or landm ark nodes in these hierarchical routings are typically responsible for routing packets for other nodes and could become a vulnerable centers, each node w ithin a CEC cluster is treated equally. CEC can be used with any hierarchical clustering routings to conserve energy. However, CEC does not assume clustering routing support. 2.5.3 Geographic Ad hoc routing Ko and Vaidya presented a Location-aided Routing (LAR) approach to utilize location inform ation to improve performance of routing protocols in ad-hoc networks [31]. They use location inform ation to decrease overhead of routing discovery by lim iting the search space for a desired route. GAF is com patible w ith LAR as it is w ith other ad hoc routing protocols. G A F’s ap proaches to energy savings are also complem entary w ith those in LAR; LAR optim ization does not aim to reduce energy dissipation. Grid location service (GLS) [34] provides location services by duplicating a node’s location in a small subset of other nodes. W ith the help of GLS, a node can geographically forward d ata to the destination w ith the destination location. A node chooses its location servers in a hierarchy of grids w ith increasing size. Such grid-based partition is very close to the grid used in GAF. Since GAF is not a geographical routing protocol, there is no need for GAF to propagate location inform ation to other nodes. Grid system w ith 26 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. geographic inform ation has different use in GLS and GAF: it helps choosing location servers in GLS while it helps finding node coverage in GAF. 2.5.4 Other examples of adaptive fidelity Ahn et al. [1] develop ARM control mechanisms to adapt to the m obility and route- dem and pattern in order to control the routing overhead in proactive routing protocols. GAF uses the similar idea to adapt to high mobility although its goal is to m aintain routing fidelity under high mobility. CEC will borrow some schemes used in ARM to let each node m onitor its neighbors’ m obility in order to improve routing fidelity and control the overhead. Bulusu et al. [10] are investigating the use of adaptive fidelity for localization sys tems. Given a field of energy-constrained beacon nodes, she is examining what duty cycle optimizes network lifetime. 27 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 3 B E C A /A F E C A for energy conservation In this section we present two application-driven energy-conserving ad hoc routing al gorithms. These algorithm s are based on the observations presented in Section 1: first, because radios commonly used for 802.11-like networks consume nearly as much power listening as receiving, the only way to substantially reduce energy consum ption is by tu rn ing the radio off. Second, we can take advantage of information above the MAC-layer to control how long we can keep the radio turned off. Third, it is possible to take advantage of node density to further conserve power. O ur first algorithm is the basic energy-conserving algorithm (BECA). The basic idea is th at nodes do not need to be listening and consuming power when they are not involved in sending, forwarding, or receiving data. The PAMAS protocol applies this result at the MAC level [51], turning off after determ ining packets are addressed elsewhere, b u t it still listens when idle to receive new packets. We improve this result using higher-level inform ation to tu rn off the radio more frequently and for a longer duration, thus reducing the substantial energy dissipated during the idle state. We describe this algorithm in detail in Section 3.1. 28 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Our second algorithm is the adaptive fidelity energy-conserving algorithm (AFECA). This algorithm uses observations about node density to increase the tim e the radio is powered off. W hen m any equivalent nodes are able to forward data, they power off for longer intervals. In a sense, AFECA adapts the num ber of nodes participating in ad hoc routing to keep a constant level of routing fidelity (number of nodes th a t will route packets) to reduce energy consumption. We describe this algorithm in Section 3.2. In principle these algorithm s can be applied as m odifications to any ad hoc routing protocol. We study on-dem and ad hoc routing protocols such as AODV and DSR, for two reasons. First, on-dem and protocols have been shown to perform b etter (in term s of packet loss) th an a priori ad hoc routing protocols [8, 29], Second, a priori routing proto cols depend on periodic message exchanges. Care m ust be taken to avoid synchronization problems if combining our algorithm s with such routing protocols. 3.1 Basic Energy-Conserving Algorithm The goal of BECA is to minimize energy consum ption by keeping radios powered-off as much as possible, trading higher latency for reduced energy use. We will show th at this algorithm will establish routes in all cases where a standard ad hoc routing protocol would, although it m ay introduce longer latency. Preliminary algorithm: In BECA,nodes are in one of three states: sleeping, listening, active. A state diagram is shown in Figure 3.1. Initially nodes sta rt out in the sleeping state. W hen sleeping the radio is off, not consuming power. In this state they keep their radio turned off for tim e Ts, then transition 29 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Ta after last traffic sleeping active after T| with no traffic listening upon first traffic Figure 3.1: States in BECA to listening. If when a node is sleeping, it has d ata to send, it transitions to active and starts sending the data. (Although the radio is off, sensors or other parts of the node may be on.) W hen in state listening, a node turns on its radio and listens for messages. It listens for tim e Ti. During this tim e, if it gets a routing message and participates in the route, or if it decides to send data, it transitions to state active. Otherwise it returns to sleeping after J}. W hen in the state active, a node sends or transits data. If at any tim e it hasn’t sent or transited d ata in tim e Ta, it transitions to state sleeping. We m ust control how this duty cycle interacts w ith ad hoc routing; sometimes the recipient of the routing request (RREQ) will be in sleep mode. We require th at the ad hoc routing protocol retry requests every T0 seconds, and th a t it retry R times. To manage interactions between BECA and the underlying ad hoc routing protocol we set T) to be the same as T0, then we pick Ts as some m ultiple k of T0 and then adjust R to insure th at some request will get through. Since we always listen for periods of T0, 30 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Node A Node B L .M 'Send RREP Data Transfer RREQ R evived Sleeping state Listening state RREQ! arrives but dropped Because of sleeping state Active state T O TO+To TO+To+Ta T1 T2 Tl+Ta T2+Ta Figure 3.2: Establishing routing between to an adjacent (but possibly sleeping) node. we are assured th at if another node is trying to establish a new route, we will hear their RREQ message sometime. B E C A a n d a d h o c ro u tin g : To illustrate this algorithm , in Figure 3.2 we consider routing. Initially we set Ts to 1 x T0, so th at nodes have a 50% duty cycle. W hen node A sends a routing request, node B is either sleeping or listening. If it is listening, the request is honored and the route is established; both nodes become active until d ata exchange is completed. In the figure, we assume the worst case, th a t B starts to sleep ju st as we send the initial RREQ. We are guaranteed th at B will wake up in Ts = T0 seconds and hear our second RREQ. On average, we add T0/2 to latency. We therefore conclude th a t for this value of Ts we can reduce energy usage by half, we will increase latency by at m ost Ts, and we can establish routes in R — 2 retries. We can generalize this argum ent to m ultiple-hop routing. We observe th a t once the first hop hears the R R EQ message, it transitions to active, and it stays in th a t state for at least Ta seconds. Since Ta > To, even RREQs th a t do not reach their final destination will keep interm ediate nodes from sleeping. W ith Ts = 1 x T0, each hop will incur up to T0 additional latency in the worst case (we don’t assume any synchronization between 31 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. node sleep and wakeup patterns). Thus, for a network H hops in diam eter, BECA adds at most H T 0 to latency, on average (HT0)/2. and we require R = 2H retries to insure we succeed in establishing routes. After a route is set up, those nodes th at are not in the route will no longer receive RREQ message or data, so th a t they will retu rn to sleeping (after Ta). Those nodes that are on the route will rem ain in active until d ata exchange ceases. We can generalize this approach for Ts = kT0 for k > 1 to get b etter duty cycles. Larger values of k linearly increase latency b u t improve power savings only by a factor of l/k . In Section 7 we examine the relationship between power savings, latency, and packet loss. A c c o m m o d a tin g p a c k e t loss: The above discussion has assumed th a t no routing requests are lost in transit. In real ad hoc networks packets can be lost because of d ata corruption, collisions, or node movement. To account for loss from the first two sources ' we transm it routing requests twice each listen or sleep interval, setting Tg — Ts = 2 x T0. To account for node movement we need to increase the num ber of retries. 3.2 A daptive Fidelity Energy-Conserving Algorithm In densely-populated ad hoc networks many nodes are interchangeable for routing pur poses. O ur Adaptive Fidelity Energy-Conserving Algorithm (AFECA) takes advantage of this observation to improve energy conservation by estim ating node populating and increasing sleep tim e when other nodes are available. We will show th a t this approach increases network lifetime as node density increases. 32 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. AFECA raises two design issues: how do we estim ate neighborhood density, and, how do we adjust the duty cycle to account for this information. In our prototype, each node estim ates its neighborhood by keeping a list of w hat nodes it hears whenever it is listening. This list is treated as soft state, nodes are autom atically removed from it if at any tim e it hasn’t been updated in tim e Te (we use a fixed Te of 50 seconds in our experim ents). By m aintaining this list from inform ation we happen to hear we avoid any additional message or energy overhead of explicit neighbor-discovery messages. We define the size of the neighborhood list as N. In AFECA, each node increases its sleeping a some factor proportional to the num ber of nodes in its neighborhood. We define Tsa by node’s actual sleeping tim e in order to differentiate it from B EC A ’s Ts. In our prototype im plem entation, we define Tsa = Random (1, N ) x T, A node recom putes T$a before it begins sleeping using it latest estim ate of N . We expect an upper bound on Tsa is appropriate, although we do not currently set one. We can define the aggregate duty cycle of N nodes using AFECA as the collective tim e they spend listening divided by the collective tim e they spend listening and sleeping. Assuming each node has the correct estim ate of IV, they will each listen for T0 and sleep for m ean N x Tsf 2. For Ts a - , therefore the aggregate duty cycle is 2/(2 + N ). We are currently considering other definitions of Ts a • We do not carefully try to insure accurate m easurem ent of neighborhood size, prim arily because there is a feedback effect in neighborhood estim ation. If we underestim ate it, neighboring nodes will spend more tim e awake and will be more likely to hear any neighbors and therefore increase their 33 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. estimates. If we overestim ate it, the converse holds. A secondary reason is th a t th at incorrect estim ates simply alter latency, not correctness. This approach is w hat we use in our simulations. A more complete analysis of alter natives to estim ate neighbor population and adjust Tsa are subjects of future work. We close w ith brief analysis of three cases th at illustrate directions we are examining. First, w ith completely passive neighbor discovery, nodes in a quiet network will forget about all neighbors. In this case AFECA simplifies to BECA w ith Ts = Ta and has 50% duty cycle. W hen sensor networks are used for surveillance applications a quiet network (nothing detected) is the typical condition. In these cases, introduction of hard-state measures of neighborhood may be appropriate. Second, our current definition of neighborhood works best when nodes are relatively evenly distributed. For extremely non-uniform topologies large latencies are possible. For example, consider an “H” topology (see Figure 3.3) where two rows of n nodes are on the vertical bars of the H, while a single node is on the horizontal bar. The center node will hear all nodes and assume a large neighborhood, but it is the only node th a t can hear both sides and so should rarely sleep. W ith our current approach, once the center node awakens it will rem ain awake as long as traffic passes through it (due to Ta), but initial latency will be high. Addressing this problem requires topological understand of the network. Fortunately, dense sensor networks where AFECA is appropriate reduce the likelihood of extrem e non-uniform ity of distribution. Finally, in restricted topologies we can com pute optim al solutions to the problem s of estim ating neighborhood size and sleep duration. Consider an infinite line of nodes, each spaced 1 unit apart, num bered by the integers. W ith perfect radios th a t reach 1 + e, 34 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Node XI I 1 Node Y1 □ □ □ \ Node Xn □ / | _ j Node Yn : Nodes can hear each Figure 3.3: “H” topology: An example of extremely non-uniform topologies each node can hear only its neighbors and so Ts cannot be increased at all. Next assume the radio reaches 2 + e. Each node can hear 4 neighbors, so a duty cycle of one-third is appropriate. Using an argum ent sim ilar to in the basic algorithm , nodes will always be able to get connectivity. In general, for this topology, neighborhood size N = 2R, and duty cycle is 1 / (JV - 1). 3.3 Summary We have dem onstrated two approaches to energy conservation for ad hoc routing. Power consum ption in current wireless networks is idle-time dom inated, so b oth focus on turning the radio off as much as possible. BECA, our basic algorithm , uses routing- and application-layer inform ation to control node duty cycle Our second algorithm , AFECA, dem onstrates adaptive fidelity. It adapts sleep tim es based on node density, scaling back node duty cycles (and so reducing routing “fidelity”) when many interchangeable nodes are present. 35 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Our algorithm s work above existing on-dem and ad hoc routing protocols, such as AODV and DSR, w ithout modification to the underlying routing protocols. O ur m ajor contributions are: algorithm s th a t turn off the radio to reduce energy consum ption w ith the involvement of application-level information, and the additional use of node deploy ment density to adaptively adjust routing fidelity to extend network lifetime. The common thread to these approaches is avoiding unnecessary energy consumption. In BECA we tu rn the radio off because it’s unneeded and we’ll check again later; w ith AFECA, we tu rn an unneeded radio off because our neighbors can check for us. These algorithm s will be im portant to maximize the utility of networks of battery-powered embedded devices. The simple “add more to improve service” behavior of AFECA and adaptive fidelity are particularly im portant as the num bers of em bedded devices and the ratio of devices to hum ans increases. 36 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 4 Geographic-informed Energy conservation protocol BECA and AFECA depend on the particular ad hoc routing protocols. More accurately, they are designed for reactive ad hoc routing protocols such as AODV and DSR. In this chapter, we try to generalize our effort in energy conservation protocols so th at the protocols do not need to depend on w hat ad hoc routing protocols are used. In addition, we try to accurately estim ate how many nodes are required to listen in order to m aintain connectivity. This is the basic idea in designing GAF, Geographical Adaptive Fidelity [57]. Each GAF node uses location inform ation to associate itself w ith a “virtual grid” , where all nodes in a particular grid square are equivalent with respect to forwarding packets (Section 4.1); Nodes in the same grid then coordinate w ith each other to determ ine who will sleep and how long(Section 4.2); this determ ination is m oderated by application and system inform ation. Nodes then periodically wake up and trade places to accomplish load balancing (Section 4.4 and 4.5). We also consider how GAF interacts w ith the underlying ad hoc routing protocol in Section 4.6. 37 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Figure 4.1: The problem of node equivalence. 4.1 Determ ining node equivalence GAF uses location inform ation and virtual grids to determ ine node equivalence. Location inform ation used in GAF may be provided by GPS or other location systems under development (for example [2, 10, 16]). For our initial discussion, we assume th at each node knows its current location exactly relative to other nodes. In Section 6.3.5 we relax this assum ption and show th a t GAF is not affected by m oderate location error or even by large, correlated error. Even w ith location inform ation, it is not trivial to find equivalent nodes in an ad hoc network. Nodes th a t are “equivalent” between some nodes may not be equivalent for comm unication between others. For example, in Figure 4.1 nodes are equidistant 1 unit apart w ith radio range slightly larger th an 2 units. For com m unication between nodes 1 and 4, nodes 2 and 3 are equivalent, while between 1 and 5 only node 3 is acceptable. GAF addresses this problem by dividing the whole area where nodes are distributed into small “virtual grids” . The virtual grid is defined such th at, for two adjacent grids A and B, all nodes in A can communicate w ith all nodes in B and vice versa. Thus all nodes in each grid are equivalent for routing. In GAF, nodes exchange grid IDs to adjust 38 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Figure 4.2: Exam ple of virtual grid in GAF. their duty cycle. For example, Figure 4.2 overlays virtual grids on Figure 1.2, creating three virtual grids A, B, and C. According to our definition of virtual grids, node 1 can reach any of 2, 3, or 4, and 2, 3, and 4 can all reach 5. Therefore nodes 2, 3, and 4 are equivalent and two of them can sleep. In the definition of virtual grid, we require th a t any node in adjacent grid can commu nicate with each other. In reality, a node’s radio com m unication range is not determ inistic or even sym m etric due to radio propagation effects such as m ulti-path reflection. In our initial discussion, we assume th a t the comm unication range is determ inistic (using the tworayground propagation model frequently used in many ad hoc routing studies, for example [8, 29]). In Section 8.8, we compare the effects of non-determ inistic radio prop agation using a shadowing model, finding th a t shadowing propagation models do not change our comparisons between GAF and ad hoc routing protocols. We size our virtual grid based on the nom inal radio range R. Assume virtual grid is a square w ith r units on a side as shown in Figure 4.2. In order to meet the definition of virtual grid, the distance between two possible farthest nodes in any two adjacent grids, such as grid B and C in Figure 4.2, m ust not be larger th an R. For example, node 2 of 39 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Figure 4.3: State transitions in GAF. grid B and node 5 of grid C in Figure 4.2 are at the end of the long diagonal connecting two adjacent grids. Therefore, we get: r 2 + (2r ) 2 < R 2 (4.1) or r < ^ (4.2, 4.2 GAF state transitions In GAF, nodes are in one of three states: sleeping, discovery, active. A state transition diagram is shown in Figure 4.3. Initially nodes start out in the discovery state. W hen in state discovery, a node turns on its radio and exchanges discovery messages to find other nodes w ithin the same grid. The discovery message is a tuple of node id, grid id, estim ated node active tim e (enat), and node state. As described above, a node uses its location and grid size to determ ine the grid id. 40 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. W hen a node enters discovery state, it sets a tim er for seconds. W hen the tim er fires, the node broadcasts its discovery message and enters state active. The tim er can also be suppressed by other discovery messages. This tim er reduces the probability of discovery message collision. W hen a node enters active, it sets a tim eout value Ta to define how long this node can stay in active state. After Ta, the node will return to discovery state. W hile active, the node periodically re-broadcasts its discovery message at intervals Tt j:. A node in discovery or active states can change state to sleeping when it can determ ine some other equivalent node will handle routing. Nodes negotiate which node will handle routing through an application-dependent ranking procedure described in the next sec tion. (Node ranking can be an arbitrary ordering of nodes to decide which nodes should be active, or it can be selected to optimize overall system lifetime.) W hen transitioning to sleeping, a node cancels all pending tim ers and powers down its radio. A node in the sleeping state wakes up after an application-dependent sleep tim e Ts and transitions back to discovery. 4.3 Tuning GAF GAF leaves choices of m any param eters including enat, Tj, Ta, node rank, Ts to appli cations. In this section, we describe how and why these param eters are chosen in the current GAF algorithm . Applications may wish to optim ize these choices, for example, perhaps trading increased packet loss for greater energy savings. 41 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Estimated node active time (enat) can be set to the expected node lifetime (enlt), conservatively set by assum ing the node will constantly consume energy at a m axim um rate until it dies. R ather th an this conservative enat, GAF uses an approach described in Section 4.4 to balance energy usage across nodes. GAF selects the discovery message interval (T^) as a uniform random value between 0 and some constant. This approach avoids contention from synchronized discovery mes sages (as inspired by SRM [20]). The range of I]/ can also be influenced by node rank to encourage highly ranked nodes to suppress low-ranked nodes, allowing them to rapidly go to sleep. Nodes in the active state may wish to chose a larger Ta to avoid bandw idth and energy overhead. Nodes active duration (Ta) can be its expected lifetime (enlt). GAF instead uses Ta to accomplish load balancing as described in Section 4.4. Node ranking in GAF is chosen to maximize network lifetime by selecting which nodes handle routing. Rank is determ ined by several rules. First, A node in the active state has higher rank th an a node in discovery state. This rule tries to quickly reach the state each grid only m aintains one active node. For nodes w ith the same state, GAF gives nodes w ith longer expected lifetime (enat) higher rank. This rule p ut nodes w ith longer expected lifetime into use first. Finally, node ids are used to break ties. It is possible to let applications choose different node rank rules according to their own bias, for example, application may favor the active nodes until they drain out energy in a sensor network. Node sleep duration (Ts) can be set to the enat of the active node since this is the conservative assum ption of its lifetime. Due to node mobility, the active node may move out the grid (of course there is chance th a t other nodes move into this grid). This can 42 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. leave a grid w ithout any active nodes although some nodes are sleeping, reducing routing fidelity. One approach th at statistically reduces this problem is to set Ts as a uniform random tim e between 0 and enat. This large range of Ts may often have nodes wake up quite early. In GAF therefore Ts is uniformly from the range [enat/2 ,enat ]. We also consider an alternate approach using node m obility inform ation in Section 4.5. 4.4 Load balancing energy usage GAF employs a load balancing strategy so th at all nodes rem ain up and running to gether for as long as possible. The idea behind this is th at all nodes in the network are equally im portant and no one node m ust be penalized more than any of the others. (An alternative is to completely exhaust the energy of each node in tu rn while other nodes sleep.) GAF uses the following load balancing strategy. After a node rem ains in the active state for tim e Ta, it changes its state to discovery to give a chance to other nodes w ithin the same grid to become active (figure 5). Recall th at nodes are ranked according to their rem aining energy levels. W hen the active node changes its state to discovery, it is more likely th a t it has less rem aining energy th an its neighbor nodes because presum ably the neighbors were in the sleeping state conserving energy during the node’s active time. Consequently, the node th a t was active is less likely to rem ain active after the discovery phase. The active node sets Ta to the value enat and advertises enat in its discovery messages. The non-active nodes in the neighborhood use enat to determ ine their sleeping period. 43 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The active node sets enat to a value less th an the tim e to use up all rem aining energy (enlt). In our sim ulations we set enat to enlt/2 so th at the node consumes half of its energy before handing off to another node in the neighborhood. To avoid thrashing, when enlt becomes less th an a threshold (say, 30s) GAF sets enat to the full enlt. 4.5 Adapting to high mobility GAF tries to adapt the num ber of nodes participating in ad hoc routing to keep a constant level of nodes th at route data. The ideal scenario would be one active node in each grid at any tim e. However, as nodes move, the active node may leave its grid. This may leave the prior grid w ithout an active node, reducing routing fidelity. In scenarios w ith high m obility this problem can greatly increase packet drop rates. We can accom modate high m obility by considering this system-level behavior explic itly in GAF. Each node estim ates the tim e it expects to leave its grid (the expected node grid time or engt) and includes this inform ation in the discovery message. W hen other node enter sleeping state, they sleep for the smaller of enat and engt to decide how long it can stay in sleeping state. This change does not change the node rank; nodes use the same ranking rules to decide who sleeps, but they sleep for shorter time. A node can estim ate engt based on its current speed s (speed can be obtained or estim ated from most GPS receivers) and the grid size, engt = r/s. This estim ate works well for the m obility model we use (the random way-point model w ith pauses); in systems where movement is less predictable this value may be more difficult to estim ate. 44 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In order to compare the effect of the node m obility adaption, we call the GAF w ithout node m obility adaption as GAF-basic (GAF-b), the GAF w ith node m obility adaption as GAF-mobility adaption (GAF-ma) and use sim ulation to compare their performance. GAF-b and GAF-m a behave the same when m obility is low. W hen mobility is high, GAF-b will tend to have fewer active nodes and therefore lower energy consum ption but higher packet loss rates. 4.6 GAF interactions with ad hoc routing In principle, GAF will run over any ad hoc routing protocols because it only uses application- and system-level inform ation to decide each node’s duty cycle, and since discovery messages are only broadcast to direct neighbors. In later sections we evaluate GAF combined w ith AODV [40] and DSR [8] against unaugm ented AODV and DSR. For brevity, we will sometimes say “we compare GAF and AODV” in place of “we compare AODV w ith GAF w ith unm odified AODV” . GAF decision to tu rn nodes on and off is independent of ad hoc routing protocols. If a node is actively routing packets when it is powered off, GAF depends on the ad hoc routing protocol quickly re-routing traffic. This may cause some packet loss, although most of ad hoc routing protocols react to changes quickly. A n optim ization th a t we have not explored is to have GAF inform the ad hoc routing protocol of im pending suspension, allowing it to preem ptively re-route any traffic. 45 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4.7 Summary Power consum ption in current wireless networks is idle-time dom inated, so GAF focus on turning the radio off as much as possible. GAF adapts sleep tim e based on node location scaling back node duty cycles (and so reducing routing “fidelity” ) when many interchangeable nodes are present. Different from AFECA, GAF quantitatively defines how many nodes are needed to stay actively in order to m aintain routing fidelity. GAF also introduces the design ideas of use of geographic inform ation to group redundant nodes and m obility prediction which will help our protocol adapt to network dynamics. 46 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 5 Cluster-based Energy Conservation (CEC) Algorithm GAF takes advantage of global location inform ation systems such as GPS to achieve its goal of conserving energy. However, in many applications, the location inform ation system is not always available, such as indoors or under trees where GPS does not work. The dependency on global location inform ation lim its G A F’s applicability. On the other hand, the geographic proxim ity does not always lead to network connec tivity. GAF m ust make very conservative connectivity assum ptions because it guesses at connectivity instead of directly m easuring it. The conservative approach requires more nodes to stay active th an necessary, leading less energy conservation. Figure 5.1: Exam ple of how network connectivity is broken in GAF under non-uniform ally deployed network. 47 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. G A F’s connectivity assum ptions based on geographic inform ation are also subject to radio propagation and node deployment. GAF may cause loss of existing network connectivity when network is sparsely deployed. Figure 5.1 shows an example where GAF breaks the existing network connectivity when nodes are non-uniform ally deployed. Node 1 can not communicate w ith node 3 b ut node 2 can com m unicate w ith both node 1 and node 3. In this case, if GAF powers off node 2, thinking it is equivalent to node 1, connectivity to node 3 is lost. This m otivates Cluster-based Energy Conservation (CEC), a cluster-based protocol th at exploits network redundancy by self-configuring equivalent nodes into clusters with independent controls. Different from GAF, CEC does not need support from global ge ographic location systems, routing protocols or other protocols. CEC directly measures network connectivity by itself, m aking it robust to connectivity error due to obstructions. Com pared w ith G A F’s conservative connectivity assum ption, the direct connectivity mea surem ent can find network redundancy more accurately so th a t it could conserve more energy. CEC elects a subset of nodes in a dense network to cover the entire node population. The elected nodes m aintain routing fidelity and network capacity while other nodes can be powered off for energy conservation. T he algorithm to elect nodes is localized, distributed, is independent of other systems or protocols and has negligible overhead. CEC spreads energy usage evenly across nodes. It also predicts mobility to b etter react to network dynamics. CEC organizes the nodes into dusters th a t are interconnected to each other as shown in Figure 5.2. The cluster head and gateway nodes, which can com m unicate w ith m ultiple 48 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. ■ cluster-head node @ gateway nodes ordinary nodes Figure 5.2: Exam ple of CEC cluster formation. The circle around the cluster-head in dicates the radio transm ission range. Clusters are interconnected by the gateways, to provide overall network connectivity. clusters, are elected in a fully distributed fashion (Section 5.2) to provide overall network connectivity. Each node has its own cluster head. As shown in Figure 5.2, a cluster can be viewed as a circle around the cluster head w ith the radius equal to the radio transm ission range of the cluster head. In each cluster, all nodes other th an the cluster heads and the selected gateway nodes are redundant. The redundant nodes in each cluster can be powered off to conserve energy. In order to balance energy use in the network, CEC periodically powers on the redundant nodes, which were powered off for energy conservation, before the cluster heads run out of energy. CEC then executes a cluster-form ation algorithm to re-elect the cluster-heads and gateways (Section 5.3). In order to adapt to network mobility. CEC uses predictions of network m obility to anticipate topology changes and thereby m aintain network connectivity (Section 5.4). 49 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 5.1 Determ ining network redundancy CEC divides network into a num ber of overlapping clusters. In CEC, a cluster is defined as a subset of nodes which are m utually ’ ’reachable” by at m ost 2 hops. Nodes find whether they can reach the others in at m ost 2 hops by exchanging discovery messages (Section 5.2). The cluster-head is the node which can reach all nodes in the cluster within 1 hop. CEC algorithm ensures th a t there is only one cluster-head in each cluster. Each cluster is identified by its cluster-head. A node is a gateway node if it is a m em ber of more than one cluster. The gateway nodes connect all clusters together to ensure overall network connectivity. The connection from each node to every other nodes can be m aintained by a subset of gateway nodes and cluster-heads. A node is ordinary node if it is neither cluster-head nor gateway node. If cluster connectivity is not affected due to removal of a node, the node is redundant. According to our definitions above, apparently all ordinary nodes in each cluster are redundant nodes. In addition, if m ultiple gateway nodes exist between the two same clusters, some of these nodes might be redundant nodes too. CEC finds these redundant nodes by a distributed cluster algorithm (Section 5.2). 5.2 Distributed Cluster Formation The CEC algorithm first divides the system topology into small clusters w ith distributed control. Each cluster elects only one cluster-head. CEC then chooses gateway nodes to interconnect the clusters in order to keep the network connectivity. All nodes other th an 50 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. cluster-heads and gateway nodes become ordinary nodes which can be powered off for energy conservation. 1 In order to elect cluster-heads and gateway nodes, each node periodically b ro ad casts' a discovery message th a t contains the node’s id (we assume th a t each node is assigned w ith a distinct ID), the node’s cluster-id and its estim ated lifetime as basic content. A node’s estim ated lifetime can be conservatively set by assuming the node will constantly consume energy at a m axim um rate until it runs out of energy. W hile forming clusters, CEC first elects cluster-heads, then elects gateways to connect clusters. 1 . Cluster-head Selection A node selects itself as a cluster-head if it has the longest lifetime of all its “un clustered” neighbor nodes, breaking ties by node id. Each node can independently decides w hether it becomes a cluster-head based on the discovery messages ex changed among the neighbor nodes. The neighbor nodes of a cluster-head become “clustered” after they receive discovery message from the cluster-head. Once a node selects its cluster-head, it sets its cluster id as the cluster-head’s node id. In other words, the cluster id is defined as the id of the cluster-head. 2. Gateway Node Selection W ithin a cluster, since the cluster-head is a direct neighbor of every other nodes in the cluster, any two nodes can reach each other, at most 2-hops away. This does not exclude the situation th a t the non-clusterhead nodes could be the direct neighbors, 1 We use the same term of cluster-head, gateway and oridnary as the ones used in [3] 51 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. e.g., the node 5 and node 6 in Figure 5.2. The cluster head therefore guarantees th a t each node w ithin the same cluster are connected. However, the cluster-heads do not guarantee the connectivity between the clusters. In order to interconnect the clusters to provide the entire network connectivity, CEC uses gateway nodes to interconnect the clusters to cover the entire node population. A node which has neighbors in other clusters is a gateway node. Among the gateway nodes, those nodes which can hear more th an one cluster-head are primary gateway nodes. Those gateway nodes which are not prim ary gateway nodes are secondary gateway nodes. W hen m ultiple gateway nodes exist between two adjacent clusters, CEC suppresses some of them in order to conserve energy since these gateway nodes are redundant. Gateway selection is determ ined by several rules. First, prim ary gateway nodes have higher priority th an secondary gateway nodes. This rule tries to conserve more energy because we need at least two secondary gateway nodes to connect adjacent clusters while only one prim ary gateway node is needed for the same purpose. Second, gateway nodes w ith more cluster-head neighbors have higher priority. This is based on the same reason of last rule to keep less nodes on. T hird, the gateway node w ith longer lifetime has higher priority, breaking ties by node id. This rule tries to balance the node energy so th a t nodes can equally dissipate energy. The gateway selection algorithm makes sure th a t at least one or one pair of gateway nodes between clusters exists. It does not try to guarantee only one or one pair of gateway nodes exist between adjacent clusters. CEC uses the scheme to improve 52 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. the robustness of the connectivity between clusters. On the other hand, it keeps the algorithm simple. In order to support gateway selection, CEC extends the basic discovery message to include the cluster’s ids which the gateway nodes connect with. Only the gate way nodes include such inform ation in its discovery message for gateway selection purpose. All other nodes which are neither cluster-heads and gateway nodes are ordinary nodes. Because the network connectivity can be m aintained by the cluster-heads and gateway nodes, all ordinary nodes become redundant in term s of routing. As we have pointed out in the gateway selecting algorithm , not all gateway nodes are necessary in order to m aintain network connectivity. Therefore, there are two type of redundant nodes in CEC: ordinary nodes and suppressed gateway nodes. Figure 5.2 shows an example of CEC cluster formation. This example assumes th a t all nodes have the same estim ated network lifetime, nodes with lower ID have higher priority. As we see in this figure, node 1 and node 10 can directly decide they are the cluster-heads because they have the lowest id in all of their neighbor nodes. Node 7 becomes the cluster- head after nodes 2 and 3 are clustered by receiving cluster-lead l ’s discovery message. Nodes 2 and 3 are prim ary gateway nodes because they are the neighbors of two cluster- heads: node 1 and node 7. Nodes 9 and 11 are secondary gateway nodes. Since both nodes 2 and 3 are the prim ary gateway nodes between clusters 1 and 7. One of them is redundant. Between clusters 7 and 10, since only one pair gateway nodes exist, both nodes 9 and 11 are the elected gateway nodes. 53 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 5.3 Controlling the D uty Cycle of CEC Nodes The Cluster-heads and the selected gateway nodes provides overall network connectivity. O rdinary nodes and those gateway nodes which are suppressed by CEC gateway electing algorithm , become redundant and are powered off to conserve energy. CEC treats all nodes equally. In order to achieve this goal, CEC wakes up the powered- off redundant nodes before the cluster-heads run out energy, and re-forms clusters. W henever a cluster is formed, the redundant node sets a wake-up tim er which will wake itself up in tim e Ts. Ts is set to partial of the estim ated node lifetime (enlt) of the cluster-head, which is included in its discovery message. In our CEC im plem entation, we normally set Ts to be enlt/2. To avoid thrashing, we set Ts to be enlt when it becomes less th an a threshold (say 30s). Since the enlt is the lifetime th a t a cluster-head tells all nodes w ithin its cluster how long the cluster head can survive, all nodes in the same cluster should be powered on to re-form cluster before the cluster-head runs out of energy. W hile re-forming clusters, it is more likely th a t the current cluster-head has less rem aining energy th an other nodes in the cluster because presum ably the other nodes were in sleeping state conserving energy. Consequently, the current cluster-head node is less likely to become cluster-head again based on CEC cluster-head selecting algorithm . CEC therefore achieves the goal of balanced energy use by periodically forming clusters. W hen a node’s radio is powered off, its forwarding role can be replaced by other nodes. An interesting question would be how it handles traffic originated from it or destined to it. For the case of traffic source, if the node has d ata to send, it can simply power on its radio and send out data. For the case of traffic destination, the situation 54 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. can be addressed as below: first, for some applications like sensor nets, the traffic is usually not addressed to a particular node. Instead, the traffic is sent to a group of nodes w ith similar properties [28]. W hen a node is powered off, other nodes could stay alive to pick up the traffic. Second, some MAC protocols, such as 802.11, support power-saving mode where active node (typically base-stations) can tem porarily buffer the d ata for the sleeping nodes. W hen traffic is sent to sleeping node, the energy conservation protocol can be integrated w ith 802.11, like Span [14], to have the neighbor active nodes buffer the d ata for the sleeping destination node first. The sleeping nodes put into power-saving mode periodically exchanges beacons w ith the active node to if there is d a ta destined to it. If there is traffic for it, the sleeping node wakes up and picks up buffered d a ta and following traffic. 5.4 Adapting to Network M obility W hen the selected cluster-heads and gateways provide the overall network connectivity, the network m obility could cause the loss of network connectivity, especially under high mobility. CEC uses mobility prediction as a technique to anticipate movement in order to m aintain network connectivity. As shown in Figure 5.2, a cluster can be viewed as a circle around its cluster-head. If we assume the cluster circle area rem ain static, when the cluster-head leaves this area, it can not represent the other nodes in this cluster any more. The other nodes m ust reconfigure new clusters. By estim ating the tim e how soon the cluster-head m ight leave 55 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. its current cluster and telling all nodes in the cluster the time, these nodes can power on themselves before the cluster-head leaves the current cluster. To estim ate the tim e for the cluster-head to leave the cluster, we can use the cluster- head’s current speed s (we assume internal m easurem ent for each node’s speed, ju st like how a car measures its speed) and its radio transm ission range R to roughly estim ate the tim e as R /s. Note th a t the cluster-leaving tim e as R / s m ight be too long because the connectivity between the moving cluster head and some of nodes in the cluster m ight have lost con nectivity before the R /s time. On the other hand, if we set this cluster-leaving tim e too small, CEC will not be able to conserve any energy. In CEC im plem entation, we set the cluster-leaving tim e to be R /4 s to balance the energy conservation and routing fidelity. In order to tell all other nodes in the cluster about the cluster-leaving tim e of the cluster-head, we extends the basic discovery message to include the leaving-cluster tim e in cluster-head’s discovery messages. All nodes in the same cluster should wake up to reconfigure clusters before the m inim um tim e of Ts and cluster-leaving tim e of its current cluster-head in order to m aintain network connectivity. For the gateway nodes, when a gateway node is suppressed by another gateway node based on the gateway electing rules, the gateway node refers to the another gateway node’s estim ated tim e to leave the cluster to decide when to wake up to avoid possible loss of network connectivity. The gateway nodes roughly estim ate its tim e to leave the cluster by R /2 s. A lthough GAF uses the sim ilar idea, it anticipates handoffs by using global location system while CEC uses only local m easurem ent. W ith global geographic inform ation, 56 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. GAF may have more accurate m obility predication. However, CEC generalizes the tech nique of m obility prediction to different applications. 5.5 Summary We have dem onstrated our adaptive techniques (CEC) to self-configure network in order to take advantage of network redundancy to conserve energy while m aintaining routing fidelity and network capacity. CEC is a localized, distributed algorithm . Each node only needs to communicate w ith its neighbors in order to execute CEC algorithm . CEC is also a self-contained algorithm so th a t it does not need support from other systems such as global geographic system, or a particular routing protocol. CEC can operate w ith ad hoc network protocols such as AODV and DSR and with non-IP-routing based approaches such d ata diffusion, to conserve energy in order to ex tend network lifetime. Com pared w ith GAF, CEC can m easure network connectivity more robustly so th at it does not need to be conservative in powering off nodes. This would help CEC ex tend network lifetime longer th an GAF. In addition, unlike GAF, CEC does not have dependence on location systems. 57 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 6 Evaluating energy conservation protocols This section describes a methodology for evaluating energy conservation protocols in ad hoc network environm ent. We focus on the sim ulation systems, the m etrics and sensitivity studies used for analyzing and com paring protocols later in this thesis. It is difficult to capture all details of energy conservation protocols in an analytical model. For this reason we use sim ulations in our protocol studies. The advantage of sim ulation in protocol studies is th a t it can evaluate protocols under varying network conditions. Studying protocols, b oth individually and interactively, under varied condi tions is critical to understand their behavior and characteristics. Sim ulation also provides the opportunity to study large-scale protocol interactions in a controlled environm ent by sim ulating more nodes when large scale physical devices are not available or not existing. The disadvantage of sim ulation is its possible oversimplification. The abstraction inside a sim ulation could lead to the results th at are misleading or incorrect. Selecting the correct level of detail (or level of abstraction) for a sim ulation is a difficult problem. We im plem ented our energy conservation protocols in the ns-2 sim ulator [7], evaluat ing variations in node movement, traffic pattern, radio propagation and energy model as 58 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. described below. By examining different levels of detail in the sim ulations, we show how our energy conversation protocols work under network dynamics. 6.1 Extensions to ns-2 Simulator Ns is a discrete event sim ulator developed by the VINT project [19]. Ns provides sub stantial support for sim ulation of TC P, routing, and m ulticast protocols over wired and wireless network. CMU M onarch Project group [54] added the support for accurately sim ulating the physical aspects of m ulti-hop wireless networks and MAC protocols needed in such environm ent. CMU also added the support of modeling arbitrary movement pat terns, a variety of ad hoc routing protocols and the Address Resolution Protocol (ARP). Their extension to ns as described above has been integrated into V IN T /ns m ain release from ns-2.1b6. We im plem ented BEC A /A FEC A , GAF and CEC in ns-2.1b6. We use locally modified and extended version of ad hoc routing contributed by CMU [54]1, an improved AODV im plem entation from the AODV designers [15], and a energy model described later. We compared ad hoc routing w ith and w ithout our energy conservation protocol extensions using AODV and DSR. Figure 6.1 shows the basic schem atic layout of our G A F/C E C extension to ns mo bile node (B EC A /A FEC A extension is sim ilar). The ns Notes and D ocum entation [19] contains the detail inform ation of ns mobile node architecture. We will only discuss our extension to ns in this thesis. 1CMU also provided a validated 802.11 (2M) MAC layer simulation with the simulation package 59 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. To extend ns for G A F/C E C support, we im plem ented two agents to execute GAF, CEC protocols individually. On the packet incoming path, we insert a ns connector called G A F/C E C switch to forward the incoming G A F/C E C packets directly to G A F/C E C agent. We im plem ented an energy model so th at whenever the physical interface object sends/receives packets, it needs to check the node’s energy model to see if the node has enough energy to send/receive packets. We also extended the propagation model to sup port more radio propagation models in addition to the existing tworayground propagation model. 6.1.1 Outgoing G A F/C EC Packets Packets sent by the G A F/C E C agents are handled to the mobile node’s entry point, which passes them to the address classifier. Since G A F/C E C only send broadcast packets, the G A F/C E C packets will be delivered to the routing agent, which is the default target of the address classifier. Both G A F/C E C will only do localized broadcast, therefore the routing agent will simply hand the G A F /C E C packets down to the link layer (LL). There is no A R P queries for the broadcast packets. LL will insert the G A F/C E C packets to the interface queue (IFq). The m edia access control object (MAC) takes the G A F/C E C packets from the interface queue and sends them to the networking interface (NetIF) when appropriate. The N etIF first checks the energy model to see if the node still have enough energy to send the packets (we will describe the detail behaviors later when we discuss the energy model extension in Section 6.1.4). If the node has enough energy, the N etIF stam ps the common header of the packet w ith properties such as power and position of the 60 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. transm itting interface, and then passes the packet to the channel, where a copy is made for all other interfaces on the channel. 6.1.2 Incoming G AF/C EC packets A copy of each G A F/C E C packet sent onto a channel is delivered to each of the network interfaces on the channel. The tim e to reach each network interface is the tim e th a t the first bit of the packet would begin arriving at the physical interface, based on distance between the nodes and the speed of light. Each network interface stam ps the packet w ith the receiving interface’s properties. Then it consults the energy model to find out whether the node has enough energy to receive the packet (we will describe the detail behaviors later when we discuss the energy model extension at Section 6.1.4). If the node has enough energy, the N etIF then invokes the propagation model (Prop Model). The propagation model uses the transm it and receive stam ps and the properties of the receiving interface to determ ine the power w ith which the interface will receive the packet. The receiving network interface then uses their properties to determ ine if they actually successfully receive the packet, and hand the packet to their MAC layer if appropriate. If the MAC layer receives the packet w ithout collision and error-free, it passes the packet to the link layer object (LL). Since G A F/C E C packets are broadcast packet, they will be always sent to LL if they are received successfully at MAC. Before the packet reaches the entry point of mobile node, the G A F /C E C switch will filter the incoming packets. If they are G A F/C E C packets, they will be sent directly to G A F/C E C agents. 61 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. demux / Src/Sink demux ip addr default RTagent GAF/r.F,C packets U O ) arptable recvtarget LL O sendtarget OMAC recvtarget sendtarget recvtarget propagation_ channel Channel Figure 6.1: Schematic of a mobile node w ith G A F/C E C extension in ns 62 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 6.1.3 Adding power control into mobile node In order to support sim ulation of energy conservation protocols, we add node state, ON or OFF, as a node attribute. A node could be at ON state, which will consume energy even when it is idle (not receiving or sending packets). A node could be p u t at O FF state, which will consume energy at a very low level based on its sleeping power usage. We also implem ent ns APIs to tu rn O N /O F F nodes. 6.1.4 Adding Energy Model into mobile node Energy model, as im plem ented in ns, is a node attribute. The energy model represents level of energy in a mobile node. The energy model in a node has an initial value which is the level of energy th at the node has at the beginning of the sim ulation. It also has a given energy usage for every packet it transm its and receives, and the energy usage for being idle. A mobile constantly dissipates energy if it is in ON state. Since ns is a discrete event sim ulator, we add a dedicated tim er event to periodically wake up to calculate node’s energy dissipation. W hen sending/receiving packets, the N etIF will consult the energy model. The energy model will decrease the energy level by some units, based on the given energy usage, packet size and sending/receiving rate. If the energy level is still positive, the energy m odel will tell the N etIF to go ahead sending/receiving the packets. Otherwise, the energy level will tell the N etIF to drop the packets and put the node into O FF state because it has run out of energy. 63 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 6.1.5 Extending propagation models The radio propagation models are used to predict the received signal power of each packet. At the physical layer of each wireless node, there is a receiving threshold. W hen a packet is received, if its signal power is below the receiving threshold, it is m arked as error and dropped by the MAC layer. The propagation models from CMU [54] are free space model and tworayground re flection model. The free space model and the tworayground model predict the received power as a determ inistic function of distance. They both represent the comm unication range as an ideal circle. In reality the received power at certain distance is a random variable due to m ultipath propagation effects, which are also known as fading effects. We first extended ns to support a shadow model (Section 8.8). The shadowing model extends the ideal circle model of free space and tworayground to a richer statistical model: node can only probabilistically comm unicate when near the edge of the comm unication range. We further extend ns to support tim e varying shadowing model (Section 9.7). In tim e varying shadowing model, not only do the nodes probabilistically com m unicate near the edge, b u t also the comm unication range varies probabilistically over the time. We will discuss the two shadowing models in details in Section 8.8 and Section 9.7. 6.2 M etrics In order to compare, ad hoc routing w ith and w ithout our energy conservation protocols, we choose to evaluate them according to the following three metrics: 64 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1. N e tw o rk life tim e , the fraction of transit nodes with non-zero energy as a function of time. 2. E n e rg y c o n se rv a tio n . Assume th a t sim ulation starts w ith n nodes w ith initial total energy of n nodes as E q, After tim e t, the rem aining total energy of n nodes is Ef. Then the m ean energy consum ption per node (m ecn ) is: E q - E t (R 1 ^ mecn = ----------- (6.1) n * t 3. R o u tin g p e rfo rm a n c e . I define two metrics to m easure routing performance. The first m etric is the d ata delivery ratio, which is the ratio of the num ber of received packets over the total sent packets. The second m etric is the average d ata transfer delay, which the the m ean delay for those received packets. The three m etrics capture the m ost im portant property of our energy conservation protocols in term s of network lifetime, energy conservation and routing fidelity. 6.3 Sensitivity study W hen we develop protocols to conserve energy for extended network lifetime, we m ust evaluate their sensitivities to different network conditions in order to ensure th at our design goals are accomplished. W hile not all network dynamics have a big im pact on our protocols, we can find which factors have the big im pact through sensitivity studies. Due to the huge space of network dynamic scenarios, it is not feasible to explore ah possible network scenarios in evaluating network protocols. It is im portant to find out 65 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. w hat is necessary and w hat is not necessary to evaluate among all network dynamic scenarios. From such thorough analysis, we gain insights into the strengths and weakness of our protocols to help us understand applicability of our protocols and identify further research focus. We will focus on our protocols sensitivity studies in this section, and we will discuss how our work can benefit other ad hoc network research in the future work(Section 6.2). To b etter understand how our protocols performance is affected by network dynamics, we examine the sensitivity of them to network mobility, traffic, radio propagation, energy dissipation, network density and location error(G A F only) using simulations. In these factors, network mobility and traffic models have been widely used in analysis of protocols in self-organizing wireless ad hoc networks. Energy dissipation model is im portant here because we focus on energy conservation issues. Since our general approach is to exploit network redundancy to conserve energy, we m ust examine the im pact of network density. Radio propagation is im portant because in reality, it is strongly affected by m ulti-path effects (fading). A simple assum ption of determ inistic propagation does not catch the real radio property. GAF depends on location systems so th at its perform ance might be affected by location errors. We have found th at a mis-used energy model can lead to the wrong conclusion in the energy dissipation comparison research of ad hoc routing protocols, if the 802.11 MAC protocol is used(Figure 1.1). W hen much research was only focused on reducing the transm itting power of the radio, our finding pointed out the im portance of the radio idle state energy dissipation, which may dom inate the overall energy dissipation in the 802.11 network. 66 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Our future thesis work will continue our efforts on how to correctly and reasonably use the network dynamic models to study protocols’ sensitivity in ad hoc network research. 6.3.1 M obility model In our current work, nodes in the sim ulation move according to the random way-point model used in CMU [8]. Nodes alternate between pausing and then move to a random ly chosen location at a fixed speed. We consider seven pause times: 0, 30, 60, 120, 300, 600, and 900 seconds. For each pause tim e we generate 10 sets of initial placements and random way-points. In the future work, we will introduce non-random models to find w hether our protocol performance is affected by m obility models w ith different properties. 6.3.2 Network capacity Simulation traffic is generated by continuous bit rate (CBR) sources spreading the traffic random ly among 10 traffic nodes. The packet size is 512 bytes and 1024 bytes. The packet rate is set to three different values: 1 p k t/s, 20 pkts/s, 200 p k ts/s to evaluate CEC sensitivity to traffic load. Note th a t while packet size is 1024 bytes and packet rate is 200 p k ts/s, the traffic reaches the m aximum link bandw idth of 2M b/s. Com pared w ith the light traffic load used in previous work such as GAF [57] and Span [14], heavy traffic can help us truly understand protocol’s ability to preserve network capacity. 67 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 6.3.3 Energy model Our energy consum ption model is based on Stem m and K atz’s m easurem ents of an 1995 AT&T 2M b/s WaveLAN (pre-802.11) wireless LAN [53]. They m easured costs of 1.6W for transm it, 1.2W for receiving, and 1.0W for listening. To this we add a cost of 0.025W when sleeping. We chose their model as representative at the tim e we began this work. Although this hardw are is now somewhat old, newer evaluations of more recent versions of the WaveLAN card and com patible hardw are by other vendors show very sim ilar costs. Specifically, K asten m easured the Digitan 2M b/s 802.11 wireless LAN, observing costs of 1.9W (transm it), 1.5W (receive), 0.75W (listen), 0.025W (sleep) in 2000 [30], and Chen et al. m easured the Lucent 2M b/s WaveLAN 802.11 cards, observing costs of 1.4W (transm it), 1.0W (receive), 0.83W (listen), and 0.043W (sleep) [14]. In all of these studies of 802.11 hardware, transm ission is 50-100% more expensive th an listening, while sleeping is a tiny fraction of this cost. We use the energy model in our current and future work. 6.3.4 Propagation model Deterministic propagation model Uses Friss free-space attenuation ( 1 /r2) at near distances and an approxim ation to Two Ray G round (1 /r4) at far distances [45]. The approxim ation assumes specular reflection off a flat ground plane. Statistical shadowing model The shadowing model consists of two parts. The first one is known as p ath loss model, which predicts the m ean received power at certain dis tance. The path loss is usually m easured in dB. The second part of the shadowing model 68 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. reflects the variation of the received power at certain distance. The shadowing model extends the ideal circle model to a richer statistic model: nodes can only probabilistically communicate when near the edge of the comm unication range. T im e -v a ry in g sh a d o w in g m o d e l The shadowing model described above addresses the problem of the two-ray ground propagation model where the edge of node commu nication range is a circle. The shadowing model forces the nodes to only comm unicate probabilistically when near the edge of the comm unication range. Our observation of radio comm unication in the field shows th at the shadowing model can not completely reflect the characteristics of radio propagation. In a long-run (48 hours) observation of radio com m unication between fixed nodes, zhao et al. [61] found th at the quality of radio com m unication between nodes varied dram atically. A pparently there are tim e-varying interferences affecting radio comm unication in the field. A sim ple shadowing model can not catch effect of tim e-varying interferences in the real radio communication field. We therefore extend our shadowing model to tim e-varying shadowing model in order to reflect the tim e varying interferences on radio propagation in the field. In the time- varying shadowing model, not only do the nodes probabilistically comm unicate near the edge, b u t also the com m unication range varies probabilistically. Recall th at the shadowing model contains two parts: poss loss model and the variation of received power at certain distance. O ur tim e-varying shadowing model adds the statistical factor to the path loss model so th a t the received power at certain distance changes probabilistically as well. 69 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 6.3.5 Location Error(GAF only) Many kinds of localization systems (including GPS) include inherent error. We artificially introduced location error in simulations. We modeled error by random ly recording the location of each node in the range [x — e, x + e] and [y — e,y + e] for e of 5 meters. Our CEC protocol does not depend on the location system so th a t we will not study location error im pact on CEC in our future work. 6.3.6 Node density Energy conservation protocols extend network lifetime by identifying equivalent routing nodes and putting these equivalent routing into different duty cycles. As node density increases we expect the protocols to take advantage of the additional equivalent nodes to extend network lifetime. We address this issue by introducing more nodes into the same sim ulation scenarios used above. We change the node num ber from 50 to 100 and 200 while keeping the area constant. In the future, we will try to reduce the network density to ensure our protocols work properly when network does not become redundant. 6.4 Summary We briefly described the details of our extension to ns sim ulator in order to evaluate our energy conservation protocols through simulations. Our extension includes ns node restructure, G A F/C E C agents, node energy model and extended propagation models. 70 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. We also define the m etrics which we will use in the following chapters to quantita tively evaluate our energy conservation protocols. These metrics include network lifetime metrics and routing fidelity metrics. We outline our sensitivity study methodology to study our protocols sensitivity to network mobility, network capacity, energy model, radio propagation model, location model and node density. 71 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 7 Evaluation of B E C A /A F E C A Since our energy-conserving algorithm s can only be studied analytically in restricted scenarios, we explore their perform ance with a set of sim ulation studies. In this chapter, we first describe our sim ulation designs. We then use them evalu ate BECA and AFECA under the scenarios w ith infinite energy and lim ited energy to compare loss rate and energy consumption. We also use sim ulations to find the best pa ram eters for B E C A /A FE C A protocols. Finally, we change network density to find how B E C A /A FEC A can adapt to node density for extending network lifetime. 7.1 Simulation design Our exploration of energy-conserving algorithm s takes place in the ns-2 sim ulator [19]. We chose ns-2 because of our fam iliarity w ith it and its support for a wide range of ad hoc routing protocols, including DSR, AODV, DSDV, and TORA. Our work takes place in a snapshot of ns-2.1b5, which includes a modified and extended version of ad hoc routing contributed from CMU [54] and extended locally, and an improved AODV im plem entation from the AODV designers [15]. We have verified th a t our integration of 72 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. the CMU code reproduces their results [8], and th a t our sim ulation results of unmodified ad hoc protocols are consistent w ith other published results [8, 15, 29]. To this base we have added a model of node energy consum ption and prototypes of BECA and AFECA. O ur energy model is based on results reported by Stemm et al. [53]. We assume th at a radio consumes 1.15W when listening but idle, 1.2W receiving, and 1.6W sending. These values correspond to a 915MHz WaveLAN im plem entation of 802.11. We have implem ented BECA and AFECA as extensions to the AODV routing pro tocol [40]. The algorithm s could be applied to other on-dem and routing protocols. We plan to evaluate their effect on DSR in the future. We also plan to make our sim ulations available in future releases of ns. In Section 3.1 defined BECA in term s of Tp, Ts, and other constants. For our im plem entation of AODV, the m axim um possible tim eout value is 10s [15]; we therefore set Tp = 10s. The actual value of T0 varies, so it is difficult to com pute the num ber of retries required to insure routes complete. O ur im plem entation actually retries route re quests endlessly, so we will not lose requests due to exhausting retries. We are considering m odifying AODV to lim it T0 to a sm aller range so we can bound the num ber of retries. AODV already tim es out routes autom atically after 50-60s for interm ediate or end nodes. We therefore set Ta at 60s, the larger of these values. 73 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 7.2 Experim ental scenario We placed 50 nodes in a square space (1500m by 1500m). Nodes move random ly using a way-point model. At each way point, a node pauses for a predefined tim e (600 seconds) and then moves to its next way-point at a random ly chosen speed uniformly distributed between 0 and 3m /s. This model does not attem pt to reproduce a particular mobile networking scenario, b ut to provide conditions sim ilar to those used in other studies of ad hoc routing [8, 29]. We generate traffic between these nodes by placing a num ber of constant-bit-rate (CBR) sources on nodes, random ly selecting sources and destinations. Each CBR source sends 512-byte packets for a random duration chosen uniformly between 0 and 1500 seconds. We place 25, 50, 75 or 100 sources (depending on load) and adjust their sending rate between 1-10 packets/s to obtain aggregate traffic loads from 5-20 packets/s. We compute aggregate traffic load by averaging the sending rate of all nodes over the whole simulation. We consider two levels of initial energy. First, we give all nodes an infinite am ount of energy and vary algorithm param eters to compare loss rates and power consumption. Since nodes do not run out of power, these results are not com plicated by node failure resulting network partitions. Using this model we evaluate choices of Ts and Ta for BECA. Second, we select the best choices for these param eters and network lifetime in a scenario where nodes have a lim ited am ount of energy. 74 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. tra ffic lo ad (pkts/s) p ro to c o l 5 10 15 20 AODV 0.04% 0.4% 0.3% 0.5% BECA, Ts — 2s 0.13% 0.64% 0.21% 0.48% BECA, Ts = 3.3s 0.09% 0.44% 0.42% 0.58% BECA, Ts = 5s 0.27% 0.5% 0.22% 0.52% BECA, Ts = 10s 0.12% 0.5% 0.4% 0.3% BECA, Ts = 20 s 4% 2.2% 1.1% 1.2% BECA, Ts = 30s 1.6% 1.3% 1.2% 0.9% BECA, Ts = 40s 2% 1.8% 1.3% 1.4% BECA, Ts = 50s 6.5% 3.9% 3.1% 2.3% Table 7.1: Com parison of BECA loss rates for different values of Ts. Finally, all graphs presented in this section represent the m ean values from 10 sim ulation runs. Sim ulation runs vary traffic placem ent randomly, but all use the same movement patterns. 7.3 BECA performance evaluation We first evaluate how BECA changes loss rates, latency, and energy consum ption com pared to unmodified AODV for cases where nodes do not run out of energy. We chose Tf = 10s, Ta = 60s, and vary Ts for a sim ulation lasting a fixed 1500s. Table 7.1 and Figure 7.1 shows how these metrics vary for a range of Ts values. L oss r a te : We evaluate loss rate by m easuring the difference in the num ber of d ata packets sent vs. received. We calculate this as (Ps — Pr)/P s where Ps is the num ber of data packets generated by all traffic sources and Pr is the num ber of d ata packets delivered to all destinations. Table 7.1 shows packet loss as a function of Ts. Loss rate of AODV and BECA for small values of Ts is very similar. Once a route is established unmodified AODV and 75 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. AODV w ith BECA perform similarly. At large values of Ts, BECA observes higher loss due to packet losses while the route is established. Routing latency: By causing nodes to sleep we add latency when setting up new routes. We m easure routing latency as the tim e from the first routing request message until the routing reply is received. Figure 7.1(a) shows routing latency for BECA as Ts varies. For comparison, unm odi fied AODV has a fixed routing latency of 0.2s. Standard deviations are fairly high because delay of each routing request is between 0 and Ts, a wide range. The m ean is not strictly m onotonic because of statistical variation, (at low loads there are relatively few routing requests). First, we observe th at route latency grows roughly linearly w ith increasing Ts. Second, this growth is slightly lower at higher traffic rates. This effect is because in busier networks nodes are less likely to be sleeping. Finally, we conclude from this d ata th a t applications th at use pass frequent, short messages cannot tolerate high values of Ts. Fortunately, short values of Ts achieve very good energy savings and have reasonable latency. In addition to route latency, we also m easured m ean data packet latency. M ean packet latency for AODV and BECA at Ts = 10s are both about 1%. Loss rates for d ata packets are similar because once the route is established, BECA keeps nodes on the route awake. Packets only suffer sleep-induced delay if the route changes. Energy savings: Loss and latency are the costs of BECA; its benefit is energy savings. We compute the energy consumed in the sim ulation and compute how much lower this is th an the same sim ulation w ith unmodified AODV. We calculate the percentage energy 76 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 45 5pkts/s --- 10pkts/s ;.... 15pkts/s l - 20pkt/s ' 40 35 30 25 I 2 0 « 15 10 5 0 •5 60 20 30 40 50 10 0 Ts (sec) (a) Route setup latency 60 50 40 30 10 5pkt/s — i — 10pkt/s —-x— 15pkt/s ••••*•••• 20pkt/S a.... 0 60 20 30 40 50 0 10 T s (sec) (b) Energy consumption 250 200 150 LL) C L 100 5pkts/s — 10pkts/s —- 15pkts/s 20pkts/s < 20 40 60 10 30 50 0 Ts (sec) (c) delivery:energy ratio (PE) Figure 7.1: Com parison of BECA to AODV for different values of Ts in term s of route setup latency, energy consum ption, and PE. 77 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. saved as (Er - E s)/E r where E r is the total energy consum ption for unmodified AODV and E s is the energy consum ption for BECA. Figure 7.1(b) shows energy savings for various values of Ts. Standard deviations of energy consum ption are very small (less th an 1%) and so are not shown. Analysis predicted a 50% savings at Ta — 10s (Section 3.1). Sim ulation validates shows we reach 10% of this optim um at Ts = 10s in our scenario. We observe th a t there is less energy savings at higher traffic loads. More traffic leaves more nodes in active states, reducing tim e spent sleeping sleep tim e. However, even with the heaviest traffic of 20 packets/s, BECA still reduces energy consum ption by 35% energy. Selecting Ts: We also observe th a t much higher sleep tim es show no energy improve ment. Based on the observations from Figure 7.1 we conclude th a t there is little to be gained from high Ts values. To select an optim ism Ts we need a m etric th a t considers b oth packet loss and energy savings at the same tim e. We introduce the value P E to evaluate this ratio. P E = P / E , where P is the size of d ata delivered to the destinations in bytes, and E is the total energy consumed by all nodes of the network in Joules. Figure 7.1(c) shows this trade off. A lthough P E increases monotonically for these scenarios, we consider a Ts of 10s to be reasonable, capturing the m ajority of efficiency while avoiding high route setup latency. 78 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 250 20pkts/s" 200 15pkts/s 150 L U O - 10pkts/s 1 0 0 5pkts/s 50 60 30 40 0 10 20 Ts (sec) Figure 7.2: P E comparison of AFECA w ith different values of k for com puting T$a - tra ffic lo a d (pkts/ 0 p ro to c o l 5 10 15 20 AODV 0.04% 0.4% 0.3% 0.5% BECA, Ts = 10s 0.12% 0.5% 0.4% 0.3% AFECA, k = 10s 0.45% 1.3% 0.97% 1.7% Table 7.2: Com parison of loss rates for AODV, BECA (Ts = 10s), and AFECA (k = 10s). 7.4 AFECA performance evaluation AFECA defines node sleeping tim e as T s a — R andom (l,lV ) x k. (see Section 3.2 for details). Our first task is to select k. Table 7.2 and Figure 7.2 compares the P E ratio for various values of k. A value of k = 10s is best by this m etric for our workload. We have suggested th a t P E is a m easure of data transfer energy efficiency. Figure 7.2 shows th a t heavier traffic loads are correlated w ith higher P E ratios. At heavier traffic loads packets are delivered more energy-efficiently because interm ediate nodes are able to forward d ata for m ultiple stream s, thus reducing m ean per-packet energy. Figure 7.3 summarizes AFECA performance and compares the three protocols. Ta ble 7.2 and 7.3(a) show th at, as expected, BECA and AFECA are worse th an AODV in 79 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Packet delay (sec) Route request delay (sec) 14 AODV BECA> * AFECA >■-*- 12 10 8 6 2 0 -2 30 10 15 20 25 5 0 Traffic load (pkts/s) (a) Route setup latency 220 Unmodified AODV — > ..-••j S e c a > ' /A FECA 200 180 160 140 120 1 0 0 80 60 40 25 5 10 15 Traffic load (pkts/s) 20 0 BECA AFECA Q ) 40 5 10 15 Traffic load (pkts/s) (b) Energy consumption (c) delivery:energy ratio (PE) Figure 7.3: Comparisons of AODV, BECA (Ts = 10s), and AFECA (k = 10s). 80 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. term s of packet loss rate and route setup latency for these choices of param eters. In ad dition, AFECA delay is higher and shows more larger variance th an BECA, as expected because of longer, random sleep times. However, the cost is bounded and modest. In Figure 7.3(b) we compare BECA and AFECA energy to unmodified AODV. W ith these param eters they reduce energy consum ption by 35-45%, w ith AFECA 2-5% more thrifty th an BECA. In Section 7.5 we will show th a t this savings has a noticeable effect over network lifetime. We also observe th a t higher loads offer less chance to save energy— more nodes m ust stay on to forward data. Finally, Figure 7.3(c) compares the energy efficiency (measured as PE) for the proto cols. AFECA aggressive power savings results in the consistently highest efficiency. 7.5 Evaluation w ith limited energy Previous experim ents have started each node w ith enough energy so th a t none run out during the sim ulation. We next examine sim ulations where nodes have lim ited energy to study the effect of nodes running out of power and leaving the network. We set the initial am ount of energy for each node to 1000J. Nodes th a t send no packets and listen at all times will run out of power in 870s. We run sim ulations until all nodes are out of power. In this section we set Ts = 10s according to the results from Section 7.3, and k = 10s as described in Section 7.4. System efficiency: In Figure 7.4 we evaluate P E for unmodified AODV, BECA and AFECA w ith lim ited energy under different traffic load. Since this scenario runs all nodes 81 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 250 200 150 UJ CL 100 Unmodified AODV — BECA — - x — AFECA 15 20 25 5 10 0 Traffic load (pkts/s) Figure 7.4: P E comparison of unmodified AODV, BECA, and A FECA under different traffic loads w ith lim ited energy. out of power, the energy expended by all nodes is the same and this m etric really m easure the num ber of data packets each protocol is able to successfully deliver. Both of the energy-conserving protocols are able to send more packets th an unmodified AODV. At low loads they are equivalent, sending about 30% more d a ta than AODV. At high loads AFECA rations energy b etter and sends up to 15% more data. In Section 7.3 we argued th at BECA and AFECA will show packet loss sim ilar to AODV over short periods. This figure suggests th at over longer tim e periods nodes will run out of energy. In this case, the energy-conserving nature of BECA and AFECA allows them to successfully deliver more packets th an unomdified AODV. Network lifetime: Our goal is to extend the lifetime of the network as a whole through energy conservation; our energy-conserving algorithm s do this by p u ttin g nodes to sleep. Ultim ately, the application wants to know how long the network can deliver inform ation for it. It is difficult to directly evaluate this quantity directly, though, because application 82 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. unmodified AODV - BECA -— x— AFECA w < D " O O C X J o > '£ U > ‘ o CD O ) s c CD O 40 0 ) CL 200 400 600 800 1000 1200 1400 0 Simulation time(sec) Figure 7.5: Network lifetime comparison of unmodified AODV, BECA, and AFECA under traffic load of 10 packets/s w ith lim ited energy. needs vary widely. We therefore m easure node survival rates as a function of time, running the sim ulation until all nodes expire. Figure 7.5 shows node survival as a function of tim e. (We consider each of our four traffic loads, but report only the 10 packet/s load since the other results are similar.) A first observation is th a t all AODV nodes run out of power at about the same tim e (870s into this sim ulation). This tim e is not affected by the network traffic load, confirming our claim th a t energy consum ption in this scenario is dom inated by idle-time consum ption and independent of load. By powering down radios, b oth BECA and AFECA networks last much longer th an AODV. BECA is about 20% longer and AFECA about about 55% longer. These re sults bolster the argum ent th a t the reason BECA and AFECA show greater efficiency in Figure 7.4 is ionger network lifetime and so more packet delivery. 83 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. N ode density and lifetim e: Our adaptive fidelity is designed to perform b etter in densely deployed networks. To evaluate this claim we considered a denser scenario: we place 50, 100, 150, and 200 nodes in a 1000m square area. Figure 7.6 summ arizes these results for the protocols and a traffic load of 10 packets/s. (Again, we looked at our other traffic loads and found similar results.) From this d ata we conclude th at AFECA is effective at m aking use of additional nodes to extend network lifetime. AODV and BECA performance is identical or about the same as node density increases, but a four-fold increase in density doubles network lifetime w ith AFECA. Our m ajor interest for energy-conserving study is to extend network lifetime by con serving resources. BECA dem onstrates th a t we can avoid needlessly keeping the radio on longer by using inform ation from above the MAC level. AFECA takes conservation a step further: by observing the size of their neighborhood, nodes can avoid needlessly duplicating routing offered by equivalent adjacent nodes. The result is th at as node den sity rises (for example, m any people attend a meeting, or sensors are random ly deployed in an area of interest), the network lasts longer, rather th an unnecessarily exhausting itself through duplicated work. A corollary is th at, w ith AFECA, one can simply “throw down” additional nodes to improve network lifetime. 7.6 Summary In this chapter, we evaluate two algorithm s, B E C A /A FEC A for routing in energy- constrained, ad hoc, wireless networks. Nodes running our algorithm s can trade off 84 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. -1 ■ ■ ■ 1 - i i ' 50 nodes — i — 100 nodes —x— 150 nodes 200 nodes ... e d ..... - i 0 500 1000 1500 2000 Simulation time(sec) (a) Unmodified AODV 1 0 0 50 nodes — i — 100 nodes ~-x— 150 nodes -■ ■ ■ * -- -■ E D ..... 200 nodes X J C D > '> 3 W o C D O ) r o c C D C D C L 2000 1500 1000 0 500 Simulation time(sec) (b) BECA 1 0 0 50 nodes — i — 100 nodes —x— 150 nodes -- 200 nodes C / > 03 •o O c ■ O 03 > £ 3 C O O 03 a > T O c 0 ) o ■ I 2000 1500 0 500 1000 Simulation time(seo) (c) AFECA Figure 7.6: Network lifetime w ith increasing node density. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. energy dissipation and d ata delivery quality according to application requirem ents. Our algorithm s work above existing on-dem and ad hoc routing protocols, such as AODV and DSR, w ithout m odification to the underlying routing protocols. Our sim ulation studies show th at B E C A /A FEC A consume as little as 50% of the energy of an unm odified ad hoc routing protocol. Moreover, sim ulations of adaptive fidelity suggest th a t greater node density can be used to increase network lifetime; in one example a four-fold increase in density doubles network lifetime. 86 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 8 Evaluating GAF To evaluate our GAF schemes, we first use a simple m athem atical analysis to determ ine an idealized level of energy conservation in GAF. Since analysis cannot capture the complexity in a full GAF scenario, we then use sim ulation to study GAF effects on network lifetime, how and why it conserves energy, and w hether or not it increases the num ber of packet drops. Finally, we show th a t network lifetime under GAF is proportional to the density of node deployment, and we examine the sensitivity of our sim ulations due to error in the radio propagation model and quality of location. 8.1 GAF Analytic performance analysis To get an upper bound on how much GAF may extend network lifetime we next consider a very simple analytic model. Assume th a t n nodes are evenly distributed in a area with topography size A. Nom inal radio range for each node is R. According to Equation 4.2, the grid size can be set as ^ which is the m axim um size of a virtual grid. The m inimum total num ber of virtual grid cells, m , would be 87 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. According to our assum ption of evenly distributed nodes, each grid would have at most n /m nodes, or - ( 8.2) 5* A ' nodes. At best, assuming stationary nodes and no GAF overhead, only one node in each grid will be active while the rest sleep. Since Equation 8.2 gives the m axim um num ber in each grid, the network lifetime will be extended at m ost (n * R 2)/(5 * A) times. The form ula basically reflects the fact th a t w ith GAF algorithm , the more nodes, the longer network lifetime, and the fewer num ber of virtual grids, the longer network lifetime. The num ber of virtual grids m ainly depends on the nom inal radio transm ission range and the topography size. E quation 4.2 gives the upper-bound of grid size. T he overhead due to GAF discovery message is small. A lthough GAF periodically sends out discovery message if the node is in the discovery or active state, the frequency will be very low. Since the broadcast is lim ited in one hop around a node, such overhead will not affect the whole system energy dissipation too much. In the following sections we use sim ulation to relax our assum ptions and explore GAF performance in more realistic conditions. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 8.2 Simulation m ethodology It is difficult to capture the details of GAF perform ance in an analytical model. For th at reason we im plem ented GAF in the ns-2 sim ulator [7], evaluating variations in node movement, traffic p attern and energy model as described below. In order to dem onstrate the flexibility of GAF, we im plem ented GAF in a snapshot of ns-2.1b6. We use locally modified and extended version of ad hoc routing contributed by CMU [54]1, an improved AODV im plem entation from the AODV designers [15], and a energy model described below. We attached GAF to AODV to get G A F/AO DV, and GAF to DSR to get G A F/D SR . We then run GA F/A O D V , AODV, G A F/D SR , DSR on the same sim ulated scenarios to compare the perform ance in term s of energy dissipation and d ata delivery quality. We have verified th a t our integration of the CM U’s ad hoc routing reproduces their results [8], and th at our sim ulation results of unmodified ad hoc protocols are consistent w ith other published results [8, 15, 29]. T raffic a n d m o b ility m o d els: Nodes in the sim ulation move according to the random way-point model used in CMU [8]. Nodes alternate between pausing and then move to a random ly chosen location at a fixed speed. We consider seven pause times: 0, 30, 60, 120, 300, 600, and 900 seconds. For each pause tim e we generate 10 sets of initial placements and random way-points. We also evaluate two different node movement speeds: uniform distribution between 0 and 20m /s and uniform distribution between 0 and lm /s. Nodes move in a 1500m by 300m m eter area. ^ M U also provided a validated 802.11 (2M) MAC layer simulation with the simulation package 89 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In most scenario we use 50 tran sit nodes (nodes running ad hoc routing) and 10 traffic nodes acting as sources and sinks. W hen we examine node density, we vary the num ber of nodes from 50 to 100 and 200 while keeping area constant. Simulation traffic was generated by continuous bit rate (CBR) sources spreading the traffic random ly among 10 traffic nodes. The packet size was 512 bytes. The packet rate was set to three different values: 1 p k t/s, 10 p k ts/s and 20 p k ts/s to evaluate GAF sensitivity to traffic load. Even w ith 20 p k ts/s packet rate, the traffic load is still well below the network capacity. We chose to study fight to m oderately loaded systems because nodes in ad hoc networks are expected to be energy constrained, more so th an bandw idth constrained. We model a radio w ith a nom inal range of 250 m eters w ith b oth the two-ray-ground propagation model and a shadowing model. E n e rg y m o d el: O ur energy consum ption model is based on Stem m and K atz’s mea surem ents of an 1995 AT&T 2M b/s WaveLAN (pre-802.11) wireless LAN [53]. They m easured costs of 1.6W for transm it, 1.2W for receiving, and 1.0W for listening. To this we add a cost of 0.025W when sleeping. We chose their model as representative at the tim e we began this work. A lthough this hardw are is now somewhat old, newer evaluations of more recent versions of the WaveLAN card and com patible hardw are by other vendors shows very sim ilar costs. Specifically, K asten m easured the D igitan 2M b/s 802.11 wireless LAN, observing costs of 1.9W (transm it), 1.5W (receive), 0.75W (listen), 0.025W (sleep) in 2000 [30], and Chen et al. m easured the Lucent 2M b/s WaveLAN 802.11 cards, ob serving costs of 1.4 W (transm it), 1.0W (receive), 0.83W (listen), and 0.043W (sleep) [14]. 90 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In all of these studies of 802.11 hardw are, transm ission is 50-100% more expensive than listening, while sleeping is a tiny fraction of this cost. It is impossible to evaluate the behavior of the network as whole if traffic nodes run out energy before the transit nodes. To avoid this artifact we separate the traffic nodes from routing nodes and give traffic nodes infinite energy. Traffic nodes follows the same m obility model as transit nodes, b u t they do not run GAF, nor do they forward traffic. Because we treat traffic nodes specially, we do not count them when reporting the num ber of nodes in the sim ulation (for example, our 50 node sim ulation consists of 60 nodes, 50 transit nodes and 10 traffic nodes). Note th at our scheme does not exclude the possibility of accom m odating lim ited-energy traffic nodes as well. In this case, node in the active state would need to buffer d a ta for sleeping end nodes. We give each transit node enough energy so th a t it can listen for about 450 seconds. Since nodes in AODV and DSR have nodes listening constantly, all nodes expire after 450s even w ithout traffic. Since we depend on GPS, we m ust model GPS energy consumption. We m odel GPS as consuming 0.46W in continuous reporting mode, w ith power conserva tion modes th a t consume 0.165W by reporting every second and 0.033W reporting every 8 seconds [30], Since GAF does not require constant position inform ation, we add a GPS cost of 0.033W to GAF sim ulations (and not to simple AODV or DSR sim ulations). We do not tu rn off GPS when we tu rn off radio to avoid modeling satellite acquisition tim e, and because GPS cost is quite small (about equal to radio sleep cost). S u m m a ry : We compared G A F/A O D V w ith norm al AODV and G A F/D SR w ith nor m al DSR. We conducted our comparison into two phases. In the first phase, we sim ulated 91 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 50 nodes for 900s (twice of AO D V /D SR lifetime) to see how GAF affects network life time, d ata delivery, and energy usage. O ur goal in this phase is to show th a t GAF does not reduce the quality of ad hoc routing and th a t it will conserve energy com pared to an unmodified ad hoc routing protocol. We then repeat the same comparison, varying the num ber of nodes (50, 100, and 200), for 3600s of sim ulation tim e to observe w hat will happen when all nodes run out of energy. In this second phase, we study how long GAF will extend network lifetime. In each phase, We consider 1680 sim ulations, all combinations of 4 protocols (GAF/AODV, AODV, G A F/D SR , DSR), 7 movement patterns, 10 initial placements, 3 traffic loads, and 2 movement speeds (lm /s and 20m /s). The rem ainder of this section presents our sim ulation results. Sections 8.3-8.5 analyze GAF perform ance in term s of network lifetime, energy conservation, and d a ta delivery quality under low m obility w ith various traffic loads. Sections 8.6-6.3.5 investigate GAF sensitivity to various scenario param eters, nam ely high mobility (8.6 ), node density (8.7), shadowing propagation model (8.8), and location error (6.3.5). 8.3 GAF extension of network lifetime The first question we examine is how GAF affects network lifetime. We m easure network lifetime as the fraction of transit nodes w ith non-zero energy as a function of tim e. In this section we fix the num ber of nodes at 50 and consider the range of pause tim es and the two m obility speeds. We present the results for AODV with traffic loads at 20 pkts/s. We also evaluated DSR and other traffic loads; we do not present those results as they were 92 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1 I i h11iiii|h 1 1 1 1 1 1mmmimiiiiiiii|« srrri--------1 --------1 — ..i ] \ \ * ^ w < D 0.8 _ ~ o O c \ X 'X ~ o ».................• — -A 0 > 0.6 - V -:v- - 3 < J > AODV — i — O 0.4 - GAF,0 — -X — c 0 GAF,30 o GAF,60 ....-E3.... 1 C O L L 0.2 GAF,120 ‘ GAF,300 GAF,600 -- 0-- -- #- - - - GAF,900 i . . . -A 0 I 1 \ ii-w-liiiiiiiiiliiiiimiiiiiiiiiiiliiiiiiii 0 100 200 300 400 500 600 700 800 900 Simulation Time(sec) Figure 8.1: Network lifetime comparison: GAF vs. AODV at low node speed (lm /s) under various pause time. Movement: lm /s , traffic: 20 pkts/s. similar. (Traffic load does not substantially affect network lifetime because the energy used in packet forwarding is much less th an the cost of idle-time listening.) In Figure 8.1 we present the results for AODV and GAF-b w ith movement of lm /s and traffic loads at 20 pkts/s. GAF-b and -m a are equivalent at this movement speed. We leave the discussion of high mobility effect to section 8.5. The first thing we observe in Figure 8.1 is th a t w ith AODV all nodes run out of energy at the same tim e at about 450s. This is a function of the “battery” in each node and the fact th a t AODV does nothing to conserve energy. This represents the cost of continuously listening. GAF extends network lifetime considerably; after 900s 30-40% of nodes are still alive depending on m obility pattern. Scenarios w ith shorter pause tim es consistently have b et ter network lifetime th an those w ith longer pause tim e. This is because moving nodes are better at load balancing. Consider the extrem e case of a 900s pause tim e (no movement during the sim ulation). In this case, grid cells w ith a single node m ust rem ain constantly 93 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. active and so will expire at 450s. Grids w ith two nodes will last longer and then all die at the same tim e producing the stair-steps at 550s and 850s. W hen nodes move (with pause tim es less th an 600s), they will sometimes move into these grids and share the load, thus producing more gradual node failure. 8.4 GAF energy savings Node lifetime is a useful measure, but it can be a crude m easure of actual energy con sum ption because node life is binary, so 50 about-to-die nodes are considered as good as 50 fresh nodes. In order to quantify how much energy GAF saves, we instead compute the m ean energy consum ption per node defined in Equation 6.1. We calculate mecn for both AODV, GAF-b and GAF-ma for all different scenarios including node moving speed, movement p attern and different traffic load. As we did in section 8.3, we only discuss the low mobility scenario here. We leave the discussion of high m obility into section 8.6. The calculation of mecn shows th a t at low node move speed (lm /s ), b oth GAF-b and GA F-m a have about 40% lower m ecn th an AODV no m atter w hat movement patterns (different pause time) are. The different traffic load does not affect the average energy consum ption per node too much. The reason is although traffic load is increasing, over the long sim ulation tim e, idle energy dissipation still dom inates the to tal energy dissipation in the whole system. 94 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 8.5 GAF effects on data delivery W hile we have shown th at GAF conserves energy, because GAF causes nodes to sleep it may reduce how m any packets are successfully delivered, reducing routing fidelity. We have designed GAF to m aintain a constant level of routing fidelity, b u t we expect some d ata loss when nodes go to sleep as routes change, and we have identified several cases where node dynamics may cause unintentional data loss. In this section, we compare GAF and AODV during the first 450s of sim ulation when all AODV nodes are alive. After this tim e AODV fails to deliver any packets; considering the entire tim e would therefore skew these measures in G A F’s favor. O ur goal here is to show th a t GAF is not significantly worse th an AODV when AODV is effective. In the section 8.7 we will evaluate these m etrics over the whole lifetime of GAF. We calculate the data delivery ratio and average delay tim e (defined in section 6.2) for both AODV, GAF-b and G A F-m a for all different scenarios including node moving speed, movement pattern and different traffic load. At low node speed (lm /s), bo th GAF and AODV have the same d a ta delivery ratio of 99% and the same mean delay across all pause times. These results are because at low speeds, nodes rarely moves out of their virtual grids. We leave the discussion on high m obility to section 8.6. 8.6 GAF performance under high mobility We summ arize GAF-b and G A F-m a perform ance in term s of network lifetime extension, energy saving, and data delivery quality under high mobility in this section. 95 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Network lifetime extension We evaluated GAF w ith a higher movement speed (20m/s) W ith more movement we expect to see differences in GAF-b and -m a performance. In the scenarios of high node moving speed, GA F-b and GAF-m a should have different behavior in network lifetime because GA F-m a uses high m obility adaptation to keep conservative on powering off nodes while GAF-b aggressively powers off nodes. We therefore plot net work lifetime of GAF-b and GAF-m a in high node moving speed in two different Figures: Figure 8.2(a) is for GAF-b and Figure 8.2(b) is for GAF-ma. Figure 8.2(a) and Figure 8.2(b) share the same characteristics as we discussed in Figure 8.1. First, AODV lifetime is not affected in all scenarios; secondly, GAF extends the network lifetime in different degrees: the less pause time, the longer network lifetime; and thirdly, traffic load does not change the network lifetime too much. The difference between Figure 8.2(a) and Figure 8.2(b) is th a t GAF-b can keep more nodes alive for a longer tim e, especially in the high mobility pattern. For example in Figure 8.2(a), there are still 90% of nodes are still alive at the tim e 900s when the pause tim e is 0. There is only about 40% survived in GAF-ma in the same scenario (See Figure 8.2(b)). The reason, as we discussed before, is th at GA F-b aggressively powers node off for energy conservation while G A F-m a conservatively powers off the nodes in high mobility. Data delivery quality At high node moving speed (20m/s) we see some differences between GA F-b and GAF-ma (see Figure 8.3). At long pause tim es (more th an 300 seconds) GAF-b, GAF-ma, and AODV continue to have the same d a ta delivery ratio and mean delay. At shorter pause tim e (less th an 300 seconds) GA F-m a and AODV rem ain 96 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. < / ) CD ~ o O c X 5 < 1 > > £ 3 C .2 o CO 0.8 0.6 0.4 0.2 1 " " a . "v;. AODV — i — GAF,0 — x— GAF,30 GAF,60 n GAF, 120 - GAF,300 — e — GAF,600 GAF,900 -~-a - I _ _ _ _ _ _ _ 1 _ _ _ _ _ _ _ L - * ‘Q - \ • C l V \ \w Q '' A © - 100 200 300 400 500 600 700 800 900 Simulation Tim e(sec) (a) GAF-b © 0.8 T3 o c T 3 C D > £ 3 C O 0.6 c o " ■ 4 — » o o J LL 0.4 0.2 -t -9z AODV — i — GAF,0 — x— GAF,30 — •*— GAF,60 0 I GAF,120 — GAF,300 — o - GAF,600 ------ GAF,900 ^ 1 1 ® \ \ \ \ \ \ \ \ '^ \ H V'-AV' xV t'y, ' W <3\ I V 100 200 300 400 500 600 700 800 900 Simulation Tim e(sec) (b) GAF-ma Figure 8.2: Network lifetime comparison: G A F-b,GA F-m a vs. AODV at high node speed (20m /s) under various pause tim e. Different traffic load does not affect the result. In the legend, “GA F,x” means running G A F/A O D V w ith pause tim e x. 97 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. similar, but GAF-b has worse data delivery ratio and longer average delay. At the most extrem e case w ith a 0 pause tim e, GAF-b has an 85% d ata delivery ratio, while GAF-m a and AODV can reach 95%. The average delay for GAF-b is 0.35 second, when both G AF-m a and AODV are around 0.16 second. G A F-b’s node sleep policy is the reason for its worse behavior under high or constant mobility. We address this problem in three ways: first, GAF-b may be unsuitable at high mobility under the current node density; Second, by considering m obility in the sleep policy, GA F-m a is more appropriate th an GA F-b in high m obility scenarios. T hird, so far we have only considered comparison w ithin the AODV lifetime. Since typically GAF has much longer lifetime th an AODV, over this extended lifetime GAF can perform much better th an AODV. We consider extended lifetimes in the next section. Finally, we show later (Figure 8.4(a)) th a t at higher node density GA F-b is more robust to packet loss. 8.7 How network density affects GAF Figure 8.4(a) shows our result of m onitoring packet delivery ratio for every 100s for GAF-b and GA F-m a in a scenario of 100 nodes sim ulation w ith pause tim e 0, node speed 20m /s and traffic load lO pkts/s. We also m ark the point where the packet delivery ratio drops more th an 20% of AODV packet delivery ratio. For comparison, we p ut AODV packet delivery ratio inside the figure too. From Figure 8.4(a), we find th a t roughly GAF-b quadruples AODV lifetime and GAF-m a triples the AODV lifetime. Before the drops, the packet delivery ratio rem ains almost constantly compared to AODV packet delivery ratio. This shows th a t both GAF-b 98 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. c © W w © o © Q_ C O c O T D " O © > ' © O © © o c O Q. © ■ O AODV GAF-b GAF-ma 100 200 300 400 500 600 700 800 900 P au se time (sec) (a) Data delivery ratio 0.5 AODV - GAF-b - ^ 0 I 0.4 Q © ^ 0.3 © Q — X — GAF-ma © o c 0 a. 0.2 © D ) C O a > > < 0 100 200 300 400 500 600 700 800 900 P au se time (sec) (b ) A vpr^.^p d elay Figure 8.3: D ata delivery as a function of pause tim e comparison: G A F-b,m a vs. AODV under moving speed 20m /s. Traffic load is lO pkts/s. O ther load does not change the result. 99 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. and GA F-m a are very stable in delivering d ata during their lifetime. Such stability also shows th a t the num ber of active nodes is m aintained at a constant level by GAF to m aintain routing fidelity. Node fraction over tim e is shown in Figure 8.4(b). We can find th a t the total num ber of survived nodes are m aintained in the constant level during most of GAF-b and GAF-m a lifetime. This is m ainly due to the load balance algorithm used in GAF so th a t all nodes try to evenly share their lifetime. It is interesting to notice th a t from Figure 8.4(a), GAF-b has a very good d a ta delivery quality in high node density. This is because the high node density provides more nodes to cover should an active node leaves the grid. Thus at high node densities, GA F-b may be suitable even at higher movement rates. We next perform ed similar analysis for network lifetime w ith 200 nodes. These results are similar in shape to Figure 8.4 and so are not presented here. Instead, we summ arize the scalability of GAF as density increases in Figure 8.5. The y-axis shows network lifetime as defined above in m ultiples of AODV lifetime. To show the results independent of the num ber of nodes and the size of topography and radio, we define the nodes in nominal radio area, or ninra, as tv * R * n mnra — --------- -—- (8.3) w * h , where R is the nom inal radio range, n is the to tal num ber of nodes, and w and h are the w idth and length of the sim ulation area. 100 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. n etw ork lifetim e (a) packet ratio over network lifetime = o .t S > i « j -x_— ^ i '"'X * > \ AODV — i — GAF-b — x— GAF-ma •••*--- \ \ - * h -*.......- <it- 1000 network lifetime (b) fraction of survived nodes over network lifetime Figure 8.4: Quantifying network lifetime over AODV. Simulating 100 nodes, at pause tim e 0, speed 20m /s, traffic load lO pkts/s. In figure, n times means n tim es of AODV lifetime. 101 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 9 8 7 g 6 1 5 o 4 2 g 3 2 1 0 0 20 40 60 80 100 ninra Figure 8.5: Network lifetime changes along w ith node density for GA F-b w ith constant mobility (pause tim e 0), GA F-m a w ith constant mobility, GAF w ith no mobility, and AODV. Figure 8.5 shows th at GAF extends the network lifetime proportionally to the increase of node density regardless the m obility p a tte rn while AODV’s lifetime keeps flat for all scenarios. The greater node density can be used to increase network lifetime for about 4 to 6 tim es (depending on the m obility pattern) w ith a four-fold increase in density. Under high mobility, the saving reaches the upper bound of 6 tim es of network lifetime while it only reaches its lower bound, 4 tim es of AODV network lifetime, when all nodes keep still. (As described in Section 8.3, m obility allows longer lifetimes.) 8.8 Sensitivity to shadowing model The shadowing model consists of two parts. The first one is known as p a th loss model, which predicts the m ean received power at certain distance. The p a th loss is usually m easured in dB. The second part of the shadowing model reflects the variation of the received power at certain distance. It is a log-normal random variable, th a t is, it is of 102 1 r -........ i i AODV — i — - GAF-b p/t=0 — x— - GAF-ma p/t=0 - GAF-b and GAF-ma, p/t = 3600 .....a ..... “ - - - - - .....a - - X '""' - - x '" '* ....■ET"" i i - U J 1 > i l 1 1 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Gaussian distribution if measured in dB, by using a close-in distance d0 as denoted by Pr (d). It uses a close-in distance do as a reference. The path loss is usually m easured in /3 is called the p ath loss exponent, and is usually empirically determ ined by field m easurem ent. We know th a t (3 = 2 is for free space propagation. Larger values correspond to more obstructions and hence faster decrease in average received power as distance becomes larger. So from Eqn. (8.4) we have The second p art of the shadowing model reflects the variation of the received power at where X^b is a Gaussian random variable w ith zero m ean and standard deviation a dB- a dB is called the shadowing deviation, and is also obtained by m easurem ent. Eqn. (8.6) is also known as a log-normal shadowing model. dB. Pr (d) is com puted relative to Pr (do) as follows. (8.4) m 106 { { ) (8.5) certain distance. It is a log-normal random variable, th a t is, it is of Gaussian distribution if m easured in dB. The overall shadowing model is represented by (8.6) 103 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. /3 is called the path loss exponent, and is usually empirically determ ined by field measurem ent. Larger values correspond to more obstructions and hence faster decrease in average received power as distance becomes larger. The shadowing model extends the ideal circle model to a richer statistic model: nodes can only probabilistically com m unicate when near the edge of the com m unication range. We select a set of typical outdoor value w ith path loss exponent 3.0 and the shadowing deviation 4.0 in our sim ulation to find out how GAF works under shadowing model. Due to the p ath loss, we expect to have worse d ata delivery ratio. However, norm al ad hoc routing is also affected by the p a th loss. O ur goal is to m atch GAF packet loss ratio with th a t of AODV under shadowing model when network lifetime can be extended. We repeated our sim ulation by ju st replacing two-ray ground propagation model with shadowing propagation model. We sum m arized the result in Figure 8.6. The figure shows th a t for AODV, compared with the sim ulation w ith tworayground propagation model, the packet delivery ratio is about 20% worse under shadowing model. GAF shows the the same downgrade. However, under shadowing model, GAF still can m atch AODV packet delivery ratio w ithin AODV lifetime. W ith extended lifetime, GAF-b can m aintain less th an 10% packet delivery ratio decrease w ithin 3 tim es of AODV network lifetime while GA F-m a can m atch the result w ithin 2 tim es of AODV network lifetime. The result is the same as Figure 8.5. The result shows th at shadowing model does not change the result on GAF which we showed in the section 6.3.6. The shadowing model does affect the packet delivery ratio but its im pact is the same on AODV and GAF. 104 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. i I I I I 1 1 1 I 1 I i I I Q* 0.2 i 1 ■ AODV BGAF-b,lx □ GAF-ma, lx 0 GAF-b,2x ■ GAF-ma, 2x HGAF-b,3x OGAF-ma,3x Pause time Figure 8.6: Com parison of packet delivery ratio for AODV, GAF-b and G A F-m a under shadowing model. GAF-*,nX means GAF-b or -m a at n tim es of AODV lifetime. 105 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 8.9 Summary GAF conserves energy by identifying nodes th a t are equivalent from a routing perspective and then turning off unnecessary nodes, keeping a constant level of routing fidelity. GAF m oderates this policy using application- and system-level information; nodes th a t source or sink data rem ain on and interm ediate nodes m onitor and balance energy use. GAF is independent of the underlying ad hoc routing protocol. We sim ulate GAF over unmodified AODV and DSR. Analysis and sim ulation studies of GAF show th a t it can consume 40% to 60% less energy th an an unmodified ad hoc routing protocol. Moreover, sim ulations of GAF suggest th at network lifetime increases proportionally to node density; in one example, a four-fold increase in node density leads to network lifetime increase for 3 to 6 tim es (depending on the m obility p a tte rn ). Com pared w ith AFECA, GAF quantitatively decides how many nodes need to be active in order to m aintain routing fidelity based on network redundancy degrees. GAF therefore has b etter routing performance and longer network lifetime while network den sity increases. 106 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 9 CEC Performance Evaluation To evaluate CEC, we first use analytical model to bound CEC overhead, determ ine how long the network lifetime can be extended, examine CEC adaption to both high and low network density, and understand C E C ’s improvement on network capacity. However, it is difficult to capture all details of CEC protocol in analytical models. For this reason, we implem ented CEC in the ns-2 sim ulator [7], evaluating variations in term s of node movement, traffic load, radio propagation and network density. 9.1 CEC Analytic Performance 9.1.1 CEC Protocol Message Analysis CEC is a localized, distributed algorithm . Since each node only needs to tell its direct neighbors about its own states, each node will possess only partial inform ation about the whole network. This distributed nature is very im portant in controlling the communica tion overhead in ad hoc wireless networks. 107 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In the average case, the message size of discovery message is 0 (1 ). In the worst case, the message size is 0 ( n ) where n is the num ber of neighbors of a node. The reason for 0 (n ) in the worst case is th at the gateway nodes need to put the num ber of clusters it can connect into the discovery message for gateway electing algorithm . In the worst case, the gateway nodes could connect to n clusters. Also note th a t the 0 (n ) message only happens in a small portion of gateway nodes which are a small portion of all nodes. Once a gateway node is suppressed, it will power off itself so th a t no more discovery message is sent until it powers back on. Only the gateway nodes which are not suppressed will periodically sends 0 (n ) discovery message. Such gateway nodes presum ably are very small portion of nodes. (In our sim ulations, we m easure th a t the message size is 0 (1 )) In GAF [57], the message size is always 0 (1 ). This is achieved w ith the help of the global geographic inform ation systems. In Span, the message size is always O (n) because node puts its neighbor inform ation into its HELLO messages [14]. 9.1.2 Network Lifetime Extension To get an upper bound on how much CEC could extend network lifetime, we consider a very simple analytic model. Assume th a t nodes are distributed in an area w ith topography size A. Nom inal radio range for each node is R. As we can see from Figure 5.2, a cluster area can be viewed as a circle around cluster head w ith the radius equal to R. In order to cover the whole area, the m inim um num ber of clusters, m, would be m A t tR 2 (9.1) 108 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In each cluster, at least the cluster-head and a few gateway nodes m ust stay alive in order to m aintain network connectivity. Assume the average num ber of adjacent clusters is k, we need at least m (k + 1) nodes to cover the whole area. At best, assuming stationary nodes and no CEC overhead, the network lifetime could be extended at m ost n /(m (k + 1)) tim es which is: n ' R2 (9.2) {k + I) A GAF can extend network lifetime at m ost by (nR2)/ (5A) tim es [57]. These equations show th a t CEC can extend network lifetime longer th an GAF when k < 16. It is reason able to think k is less th a t 16 in m ost of scenarios, for example, k = 1 in Figure 5.2, the result tells us th a t CEC in general can extend network lifetime longer th an GAF. One reason for this difference is th a t GAF keeps network connectivity based on geo graphic proximity. Due to the unreliable geographic proximity in network connectivity, GAF has to be conservative to use smaller grid size to group redundant nodes. CEC uses connectivity measurement to discover network redundancy. It does not need to be conservative while grouping redundant nodes. 9.1.3 Adapting to Network Dynamics CEC forms clusters fully based on the network connectivity which is discovered by each node. Due to its nature of dynamic cluster form ation, C EC ’s connectivity m easurem ent 109 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. ensures th a t network connectivity would not be broken by CEC algorithm , if the connec tivity does exist. This property is very im portant when the network is not dense enough, where GAF may break the existing network connectivity as shown in Figure 5.1. If we apply CEC algorithm to the same scenario as Figure 5.1, we will find th a t nodes 1 and 3 become the cluster-heads when node 2 becomes the only gateway node between cluster 1 and cluster 3. All three nodes will stay alive so th at the network connectivity will not be broken. More generally, the example reflects a requirem ent in the energy conservation algo rithm s where when the network redundancy does not exist, the algorithm s m ust adapt it to stop powering off node in order to m aintain network connectivity. CEC meets the requirem ent. CEC also adapts to asym m etric radio propagations. If a node w ithin the cluster-head radio range can not receive the discovery messages from the cluster-head due to asym m etric radio, the node will form new cluster. This may lead to less energy conservation b u t keep network connectivity. 9.1.4 Simulation M ethodology In order to compare CEC, GAF, and unmodified AODV and DSR on the same sim ulated scenarios in term s of network lifetime, energy dissipation and routing fidelity, we imple m ented CEC in ns-2.1b8, and attached CEC to AODV and DSR to get CEC/A O D V , C E C /D SR because ns has released AODV, DSR, GAF/AODV, and G A F/D SR imple m entation. The AODV and DSR im plem entation in ns is based on modified and extended version of ad hoc routing protocols contributed by CMU [54] and an improved AODV 110 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. im plem entation from the AODV designers [15]. Ns released G A F/A O D V and G A F/D SR at ns-2.1b7. T raffic a n d m o b ility m o d els: Nodes alternate between pausing and then move to a random ly chosen1 location at a fixed speed. We consider seven pause times: 0, 30, 60, 120, 300, 600, and 900 seconds. For each pause tim e we generate 10 sets of initial placements and random way-points. We also evaluate two different node movement speeds: uniform distribution between 0 and 20m /s and uniform distribution between 0 and lm /s. Nodes move in a 1500m by 300m area. We vary the num ber of nodes in the same area in order to examine CEC sensitivity to different node density. We put 25, 50, 100, 200 nodes into the same area to sim ulate low density network and high density network. Simulation traffic is generated by continuous bit rate (CBR) sources spreading the traffic random ly among 10 traffic nodes. The packet size is 512 bytes and 1024 bytes. The packet rate is set to three different values: 1 p k t/s, 20 p k ts/s, 200 p k ts/s to evaluate CEC sensitivity to traffic load. Note th a t while packet size is 1024 bytes and packet rate is 200 p k ts/s, the traffic reaches the m axim um link bandw idth of 2M b/s. Com pared w ith the light traffic load used in previous work such as GAF [57] and Span [14], heavy traffic can help us truly understand protocol’s ability to preserve network capacity. We model a radio w ith a nom inal range of 250 m eters w ith bo th the two-ray-ground propagation model and a shadowing model, which described in [45]. We select a set of typical outdoor value w ith p ath loss exponent 3.0 and the shadowing deviation 4.0. I l l Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The shadowing model extends the ideal circle model to a richer statistic model: nodes can only probabilistically com m unicate when near the edge of the comm unication range. E n e rg y m o d el: Prior work [53, 57] has shown th a t transm ission is 50-100% more expensive th an listening, while sleeping is a tiny fraction of this cost. We use 1.6W for transm it, 1.2W for receiving, 1.0W for listening and a cost of 0.025W when sleep ing. These values are based on the m easurem ent of an 1995 AT&T 2M b/s WaveLAN (pre-802.11) wireless LAN [53]. A lthough this hardw are is now somewhat old, newer evaluations of more recent versions of the 802.11 card and com patible hardw are by other vendors shows very similar costs. It is impossible to evaluate the behavior of the network as whole if traffic nodes run out energy before the transit nodes. To avoid this artifact we separate the traffic nodes from routing nodes and give traffic nodes infinite energy. Traffic nodes follows the same m obility model as transit nodes. Because the traffic nodes are treated specially, we do not count them when reporting the num ber of nodes in the sim ulation. In addition, the traffic nodes do not forward traffic. We give each transit node enough energy so th at it can listen for about 450 seconds. If a node listens constantly, it will expire after 450s even w ithout traffic. Summary: We compared CEC /A O D V (C EC /D SR ) w ith unm odified AODV (DSR) and G A F/A O D V (G A F/D SR). Due to the lim itation of pages, we only present the AODV comparisons in the following paper. T he DSR comparison is very similar. 112 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. For each type of protocol, we run 1680 sim ulations, all com binations of 7 movement patterns, 10 initial placements, 3 traffic loads, 2 movement speeds and 4 different network density. The rem ainder of this section presents our sim ulation results. Sections 9.2-9.5 analyze CEC performance in term s of energy conservation, protocol overhead, network lifetime and routing fidelity. Section 6.3.6 investigates CEC sensitivity to different network den sity. 9.2 Energy Conservation We examine how much energy is conserved compared to AODV and GAF. In order to quantify the comparison, we use the same metric, mean energy consumption per node (mecn), defined in E quation 6.1, to compare energy conservation performance. We calculate mecn for CEC, AODV and GAF for all sim ulated scenarios. The result shows th a t in general, CEC uses alm ost about half energy of GAF except at the scenario where nodes move at high speed (20m /s) and constant movement (zero pause tim e). GAF typically can save 30-40% energy th an AODV. CEC can save about 60-70% energy th an AODV. W hen nodes move at high speed and w ith constant movement (no pause tim e), CEC high m obility adaption will p u t more nodes in duty cycle and lead to more frequent cluster formations. Such overhead causes CEC to use more energy th an GAF although it still uses about 30% less energy th an AODV. We can find th a t w ith the help of global location information, GAF is more efficient to adapt to high mobility. 113 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. W hen nodes move at low speed, the pause tim e does not affect CEC. This is mainly because CEC mobility adaption is sensitivity to node moving speed, not how frequent the node moves. Varied traffic load does not affect the energy conservation resuits. 9.3 CEC protocol overhead We m easured the energy used by CEC control messages (discovery message) and com puted the percentage of energy used by CEC protocol over the to tal system energy usage. Figure 9.1 shows how the CEC control overhead varies w ith network mobility under no traffic. W ith traffic, some nodes may run out of energy faster th an the others, which could cause less CEC overhead. In other words, the CEC overhead under no traffic reflects the worst case of CEC energy consum ption. fn Figure 9.1, X axis is the pause tim e, Y axis is the percentage of CEC overhead, the ratio of energy used by CEC control messages over total system energy usage. Since our total sim ulation tim e is 900 seconds, and we choose 7 different pause tim es for the study, the pause tim e less th an 120 seconds generally represents a high m obility scenarios so th a t the CEC overhead for pause tim e of 0, 30s, 60s and 120s looks sim ilar (0.4% of total system energy). For the other two pause tim es, 300 second and 600 second, they represent slow mobility. Therefore, the CEC overhead is sim ilar in the scenario of slow mobility, which is about 0.3% of system energy use. Finally, the pause tim e of 900 seconds represents a static network, in which case the CEC control overhead is minimum, only 0.2% of total system energy use. 114 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. We find th a t CEC protocol overhead constantly uses 0.2 to 0.4% of total system energy. Under high speed, CEC overhead is affected by pause tim e of the m obility model. The less pause time, the more frequent node moves, the higher overhead. This is because CEC high m obility adaption algorithm turns on node more frequently under high speed and high mobility. We also calculated the CEC overhead over the total system energy use under typical light traffic load. The result is close to th a t in the Figure 9.1. For example, under no traffic, the overhead is 0.4% for pause tim e 0. W ith fight traffic, the overhead is about 0.399% for pause tim e 0. The reason for alm ost no change in overhead w ith and w ithout traffic is th at under fight traffic, the energy usage is still dom inated by the radio idle tim e energy use, each node’s lifetime is barely affected by the traffic load. The overhead w ith traffic is slightly lower th an th a t of no traffic because some nodes th a t send traffic run out of energy faster th an the others (radio consumes more energy in sending than receiving/idle), so th at those nodes w ith shorter lifetime generate less CEC overhead. The sim ulation results show th a t CEC efficiently controls its overhead by localized, distributed algorithms. The scalability of CEC protocol is good in term s of energy use. 9.4 Extended network lifetim e The goal of conserving energy is to extend network lifetime. In Section 9.2, we have found th at CEC consumes about half of the energy of GAF in most scenarios. We next examine how much this savings extend network lifetime, comparing CEC to GAF and AODV. 115 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 0.005 0.0045 0.004 0.0035 * o CO a ) o > O 0.003 LU ° 0.0025 " o < 1 ) a > to c 0 ) o 0.002 0.0015 0.001 to CL 0.0005 600 800 1000 0 200 400 Pause time(sec) Figure 9.1: Percentage of CEC energy use over to tal system energy use under no traffic. The result is not affected by node moving speed. tn a > ■ a o c ■ o tl ) > £ 3 to AODV — i CEC,0 — x- CEC,900 GAF,0 Q GAF,900 — 500 1000 1500 2000 Simulation Time(sec) 2500 Figure 9.2: Com parison of non-zero energy node fraction over time: CEC, GAF and ADOV under different network mobility. Traffic load is 20pkts/s. Different traffic load does not affect the result. In the legend, “CEC,x” means running CEC w ith pause tim e x, so is “GA F,x” . 116 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Figure 9.2 gives the comparison of non-zero energy nodes change over tim e for CEC, GAF, and AODV where nodes move at 20m /s. W hen nodes moves at low speed (lm /s), CEC plot is close to the pause tim e of 900s CEC curve in Figure 9.2, regardless of pause time. All AODV nodes run out of energy at the same tim e at about 430s. Since AODV does nothing to conserve energy, this result reflects the cost of continuously listening. For both CEC and GAF, we only plot the zero pause tim e which represents constant node movement and 900s pause tim e which represents almost no node movement. The results based on other pause tim es sit between the zero pause tim e and 900s pause time. Not only does CEC consume less energy th an GAF as shown in Section 9.2, but also CEC balances energy more evenly in nodes th an GAF. For example, in Figure 9.2, at tim e 900s, at least 80% of CEC nodes are still alive while at most 40% of GAF nodes are alive except the constant movement where pause tim e is 0. Although bo th CEC and GAF try to spread energy equally on all nodes, CEC is more efficient because of its connectivity m easurem ent based approach. CEC also shows a different trend from GAF. In CEC, more nodes survive under low m obility (zero pause tim e). In GAF, more nodes survive under high m obility (pause tim e 900s). This is because high m obility causes more frequent cluster form ations (so th at more energy is used), due to m obility adaption algorithm in CEC. In GAF, high m obility helps balance energy use because changes of node location cause active node re-election w ithin grids. In addition, GAF is more efficient in predicting m obility w ith the help of global location information. 117 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. < D ■ o O 2500 2000 1500 1000 500 0 1 1 ..." " 1 — AODV — i — CEC — x— _ GAF ____________________ 1 _____________ X X ' ' ' .......... _ i l I 1 200 400 600 Pause Time(s) 800 1000 Figure 9.3: Com parison of network m obility im pact on network lifetime: CEC, GAF and ADOV. Traffic load for all scenarios is 20pkts/s. In order to quantitatively compare network lifetime of CEC, GAF and AODV, we measure when only 20% of nodes rem ain alive and plot this tim e against the degree of m obility (pause tim e) in Figure 9.3. This figure reflects how network m obility im pacts on network lifetime. According to Figure 9.3, CEC extends network lifetime at most two tim es longer than GAF, five tim es longer th an AODV. In CEC, network lifetime increases along w ith the pause tim e. As we explained above, this is the effect of CEC high m obility adaption algorithm . At constant movement where pause tim e is zero, CEC has the same network lifetime as GAF (yet it is still twice of the AODV lifetime). This is the worst case for CEC. In Section 9.2, we have found th a t CEC consumes 20% more energy th an GAF because its m obility adaption algorithm is not efficient as GAF since it lacks global location information. The same explanation explains why the network lifetime for CEC under zero pause tim e is not as good as larger pause tim e. 118 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 9.5 Routing Fidelity C E C ’s goal is to m aintain routing fidelity while conserving energy. We have shown th a t energy can be dram atically conserved, we now examine routing fidelity under CEC. In order to help us b etter understand how routing fidelity is affected, we m easure data delivery ratio, which is the ratio of the num ber of received packets over the total sent packets, and the average data transfer delay, which the the m ean delay for those received packets. We compare CEC, AODV, GAF for their d ata delivery ratio and average d ata transfer delay under different traffic loads. Different from the previous work [57, 14], we examine both light and heavy traffic loads. The d ata delivery ratio and average data transfer delay under heavy traffic loads, which reaches the m axim um network bandw idth, can truly reflect the effect on network capacity in addition to routing fidelity. Figure 9.4 compares CEC and AODV’s d ata delivery ratio and average delivery delay under heavy traffic loads which reaches 2M b/s, the m axim um bandw idth of the sim ulated network. For CEC, we plot them in both norm al AODV lifetime and extended lifetime. The extended lifetime is two tim es of AODV lifetime. The node’s m axim um moving speed is 20m /s. O ther movement speeds do not change the result. The result also rem ains the same under light traffic. W ithin the norm al AODV lifetime, we can find CEC performs alm ost the same as AODV (with reasonable standard deviations shown in Figure 9.4). W hen network mo bility is slow, CEC perform s even b etter th an AODV. The reason is th at under heavy traffic, energy dissipation in AODV is unbalanced. Those nodes on the routing p ath 119 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. C O c c o Q C O f f l Q AODV CEC at extended network lifetime 400 600 Pause Time (s) 1000 (a) Data delivery ratio AODV CEC at extended network lifetime 400 600 Pause Time (s) 1000 (b) Average delay Figure 9.4: CEC and AODV routing fidelity comparison under heavy traffic load (200pkts/s). Node maximum moving speed is 20m /s. The slow moving speed of lm /s does not change the result. 120 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. or close to the routing path consume more energy th an other nodes. Such unbalanced energy use cause some nodes to run out of energy faster th an other nodes. Worse data delivery ratio is caused by loss of some nodes in AODV. In CEC, since the redundant nodes are powered off to conserve energy and node energy use is balanced, the system performs well even under low mobility. We also noticed th at CEC can m aintain alm ost the same d a ta delivery quality at extended network lifetime. This means th a t d ata carried by the network is doubled. This observation reflects the fact th a t ad hoc network is energy-constrained, not bandw idth constrained. Both CEC and GAF [57] can extend network lifetime while m aintaining routing fi delity. In order to further understand how CEC and GAF perform at extended network lifetime, we compare CEC w ith GAF at extended network lifetime in Figure 9.5. We also plot the AODV routing fidelity at its norm al network lifetime as ideal value to see how routing fidelity is changed in the extended network lifetime. Under high m obility (pause tim e less th an 120s), both CEC and GAF can m aintain the same d a ta delivery ratio as ideal value (w ith reasonable standard deviation as shown in Figure 9.5(a). However, as the increase of pause tim e (pause tim e larger 120s), GAF data delivery ratio becomes worse. At the worst case (pause tim e 900s), GAF packet delivery ratio dram atically decreases to only 60% of ideal level. However, CEC still follows the trend of ideal d ata delivery ratio: the lower mobility, the b etter d a ta deliver ratio. The difference between CEC d ata delivery ratio and ideal ratio rem ains below 5%. The reason for bad perform ance of GAF at lower mobility is due to its static gridding mechanism. W hen node density is good enough to keep at least one node in each grid, 121 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1 •+. .8 0.6 0.4 0.2 Ideal * —+— h CEC 0 800 1000 0 200 400 600 Pause Time (s) (a) Data Delivery Ratio J 5 t o Q £ ■ CD > t o 0 T O 1 3 Q CD O ) T O T O 1 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 1 1 • \ j i i 1M i ! ---------- * ------------ 1 Ideal u—x—! - CEC ;....x....i GAF ^ V *' f . • i 1 ! ' '---i ill J U- \ l -• * .................. .. 1 I i i i \ ! i " " - i -> ■ J . X X 1 i i 1 T 200 400 600 Pause Time (s) 800 1000 (b) Average Delay Figure 9.5: D ata Delivery Q uality as a function of pause tim e comparison: CEC vs. GAF under moving speed 20m /s at extended network lifetime. Traffic load is 20pkts/s. O ther loads do not change the result. The AODV performance under norm al network lifetime is plotted as ideal value. 122 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. GAF works fine. However, when the node density decreases in the extended lifetime, as shown in Figure 5.1, routing fidelity might be affected. If a node moves in high mobility, the situation is not very severe because the movement can help change the uneven distribution If node never moves or moves at very low mobility, the unbalanced node distribution will have no chance to get changed. It will also cause high standard deviation as shown in Figure 9.5(a). The same trend is reflected in the delay tim e as shown in Figure 9.5(b). CEC follows the trend of the ideal delay time: the lower mobility, the lower delay tim e. GAF performs better th an CEC under high m obility but performs worse at low mobility. In summary, w ith enough network density, both CEC and GAF can keep good routing fidelity. As the network density decreases, CEC can m aintain b etter routing fidelity than GAF, especially under low mobility. 9.6 Sensitivity to network density CEC is designed to take advantage of network redundancy to extend network lifetime. We change the num ber of nodes in the same sim ulated area to observe C E C ’s ability to extend network lifetime under different network density: not only high density, b u t also low density. In order to show the results independent of the num ber of nodes and the size of topography, we use the ninra [57], the num ber of nodes in nominal radio area, to m easure node density. 123 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Ideal — i — AODV — x— CEC,p/t=0 — ■ * — CEC, p/t=3600 e .... GAF, p/t=0 — » — GAF,p/t=3600 — o— 45 C D E C D i— o 5 h — > < 1 ) -a ■a- —-1 ---- x — t — 100 60 80 20 40 0 ninra Figure 9.6: Network lifetime comparison among CEC, GAF and AODV under different node density. The Figure 9.6 shows th at CEC extends the network lifetime under different network density and compares CEC w ith GAF and AODV. Under high network density (ninra larger th an 20) where network becomes redundant, both CEC and GAF can extend network lifetime in proportion to the increase of node density while AODV lifetime keeps flat. Under high mobility, CEC and GAF perform the same. Under the low mobility, CEC extends the network lifetime consistently longer th an GAF. W ith a 4-fold increase in node density (ninra 88) compared w ith non-redundant network, CEC can extends network lifetime for 12-fold over AODV, 3-fold over GAF. Under low density (ninra less th an 20) where network is not redundant, b oth CEC and GAF have the same network lifetime as AODV. Another observation is th at in GAF, the network lifetime is increased more under high m obility while in CEC, the network lifetime is increased more under low mobility. This 124 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. trend complies w ith our previous analysis regarding CEC and G A F’s different behavior under different mobility. We also plot the ideal network lifetime extension for CEC under different network density for comparison. The plot is based on the equation 9.2 where we assum e k = 1, which means only 1 cluster head and 1 gateway node for each cluster. This plot also assumes no node movement. 9.7 Sensitivity to tim e-varying shadowing model The shadowing model we used above addresses the problem of the two-ray ground prop agation model where the edge of node comm unication range is a circle. The shadowing model forces the nodes to only comm unicate probabilistically when near the edge of the com m unication range. Our observation of radio comm unication in the field shows th a t the shadowing model can not completely reflect the characteristics of radio propagation. In a long-run (48 hours) observation of radio comm unication between fixed nodes, zhao et al. [59, 60] found th at the quality of radio comm unication between nodes varied dram atically. A pparently there are tim e-varying interferences, which m eans radio transm ission quality varies over time, affecting radio comm unication in the field. A simple shadowing model can not catch effect of tim e-varying interferences in the real radio comm unication field. We therefore extend our shadowing m odel to time-varying shadowing m odel in order to reflect the tim e varying interferences on radio propagation in the field. Recall th a t the shadowing m odel contains two parts: p ath loss model and the variation of received power 125 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. at certain distance. Our tim e-varying shadowing model adds the statistical factor to the p ath loss model so th a t the received power at certain distance changes probabilistically as well. We still use the equation 8.6 to com pute the received power. In the shadowing model, only X^b varies probabilistically. In the tim e-varying shadowing model, both /3 and X^b vary probabilistically. /3 varies exponentially as a function of tim e while X^b uniformally whenever a packet is received. We repeated our sim ulation by replacing the propagation model w ith the time-shadowing model. We choose the typical outdoor value w ith path loss exponent in a range of 3.0 to 4.0 and shadowing deviation 4.0 in the simulation. Our tim e-varying shadowing model adds one more factor to probabilistically change the path loss exponent param eter and control how often the p a th loss exponent param eter should be changed, following ex ponential distribution. In the sim ulations, we also observed between 10% to 20% of asym m etric links over the sim ulation time. As shown in Figure 9.7, under high m obility (pause tim e less th an 120s), the simula tions show th at both CEC and AODV under tim e-varying propagation model have the same downgraded packet deliver ratio (20% worse) th an th a t under tworayground model. CEC still m atches AODV in term s of packet delivery ratio in the norm al network lifetime. CEC also keeps the same packet delivery ratio in the extended network lifetime while all AODV nodes run out of energy. In other words, tim e-varying propagation model does not change the result of perform ance comparison between AODV and CEC under high mobility. 126 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1 0.8 o c S * 0-6 CD > " a > Q 0.4 0 5 + - • 05 o 0.2 0 0 200 400 600 800 1000 Pause Time (s) Figure 9.7: Packet delivery ratio comparison of CEC and AODV under tim e-varying shadowing model w ith different pause times. Traffic load is lp k t/s. Under low m obility (pause tim e larger th an 300s), CEC shows alm ost 30% worse packet delivery ratio th an th a t of AODV although CEC nodes can still run in the extended network lifetime. W hile the packet delivery ratio under AODV can reach about 80%, CEC only can reach about 50% of packet delivery ratio, in both norm al and extended network lifetime. The reason for bad performance for CEC under low m obility is th a t CEC does not sense network m obility frequently enough. The tim e-varying shadowing m odel leads to frequent change of network topology. The network topology m easured at past tim e m ight change at current tim e due to tim e-varying propagation. It has the sim ilar net effect as high mobility. In order to adapt to network topology changes due to high mobility, we designed m obility prediction algorithm for CEC. The tim e-varying characteristics of ratio propagation suggests th a t we m ust have mechanisms to deal w ith topology changes due to time-varying propagation. 127 i-----------------r AODV • — h - h CEC ; x i CEC at extended network lifetime j_________i _________i ------------- 1 ---------- Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. We have found th at CEC performs well under time-varying propagation model when network m obility is high. The reason is th a t the m obility prediction algorithm forces CEC to sense the network connectivity more frequently under high mobility. The frequent measurem ent ensures th at CEC can adapt to any network topology changes, no m atter the changes are due to mobility, or tim e varying propagations. The above observations suggest th a t we can make CEC work more robustly under tim e varying propagation model by having the nodes measure network connectivity more frequently. We can achieve this by adjusting C E C ’s Ts param eter. We used to set C E C ’s Ts to be enlt/2 (if w ithout m obility prediction). W hen a cluster head has long lifetime, it would have large enlt so th a t all nodes in the cluster would check the network connectivity in very low frequency. To make nodes m easure network connectivity more frequently, we can set Ts to be the m inimum value of enlt/2 and a threshold, 7),. Th can be decided based on m easurem ent. The use of Th will p u t nodes’ m easurem ent for network connectivity in a evenly short period. We change our CEC im plem entation a little bit by forcing each node to sense network connectivity in average every 10 secs (equal to a 20m /s node move speed under high mo bility). We repeated the sim ulations by ju st replacing tworayground propagation model w ith tim e-varying propagation model. As shown in Figure 9.8, we found th a t b oth CEC and ADOV perform ed the same w ithin the AODV’s lifetime while only CEC can keep the same perform ance in the extended network lifetime. The result shows th a t CEC works more robust w ith more frequent network connectivity measurement. In summary, the tim e-varying propagation model introduces the issue of network topology changes even under low mobility. Depending on the applications, CEC might 128 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1 0.8 \7i C O D C < 1 > .> "5 Q B O B Q 0.6 0.4 0.2 AODV !---+— ! CEC : x - i CEC at extended network lifetime 1 — *— ■ 200 4 0 0 600 Pause Time (s) 800 1000 Figure 9.8: Packet delivery ratio comparison of revised CEC and AODV under time- varying shadowing model w ith different pause times. Traffic load is lp k t/s . need to be adapted to have more frequent network connectivity m easurem ent in order to work robust for m aintaining the same d ata delivery quality as AODV. 9.8 Summary We have dem onstrated our adaptive techniques (CEC) to self-configure network in order to take advantage of network redundancy to conserve energy while m aintaining routing fidelity and network capacity. Simulations show th a t CEC consumes 60%-70% less energy th an unmodified ad hoc routing protocols. Energy savings scale w ith increases in node density; w ith CEC a 4-fold increase in node density results in a 12-fold increase in lifetime compared to a non-redundant network. Since CEC directly measures connectivity rather th an m aking conservative connectivity assum ptions, it can extend network lifetime 200% longer than 129 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. previous location-based protocols such as GAF th a t m ust be more conservative. More over, these benefits come w ithout sacrificing network connectivity and network capacity. CEC control overhead is quite small; we observe only 0.4% of energy is consumed by control traffic. The comparisons between CEC, AODV and GAF show CEC uses alm ost about half energy of GAF, 30% energy of AODV. W ith less energy use, network lifetime scales with increases in node density. CEC consistently outperform s GAF when nodes are not in extremely high movement. Under extremely high node movement, the global location inform ation can help GAF predicate node movement more efficiently th an C E C ’s self-discovery technique for hand- off anticipation. GAF can conserve more 20% more energy th an CEC in such scenario. On the other hand, C E C ’s self-configuring technique makes it work b etter in the low network m obility and uneven node deployment where GAF may cause loss of network connectivity. 130 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 10 Im plem entation and Evaluation of energy conservation protocols A lthough we have evaluated our protocols through simulations, the results we got from sim ulations do not necessarily guarantee good performance in the real world, or on the nodes w ith real radios. Ultim ately, the final test for any protocols is to deploy it in the environm ents which it was designed for and m easure it to find out if it can operate successfully. Yutaka Mori [36] at USC im plem ented and evaluated GAF on the PC104 sensor net test-bed. Solomon Bien [6] at UCLA drove the CEC im plem entation and evaluation work on the Motes and iPAQ testbed. I was involved in the two projects to give my opinions on design and evaluation issues since I am the designer for both GAF and CEC protocols. This chapter describes the testbed setup, discusses the results of initial experim ents by Yutaka and Solomon, and analyzes the im pact of real-world radio propagation on the protocols. A dditional inform ation regarding the im plem entation details and can be found at [36, 6]. 131 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Filter Apps i i Diffusion Core Diffusion Message GAF Control GAF Message Communication ^Devices_______ Figure 10.1: GAF im plem entation architecture 10.1 GAF im plem entation In this section, we will discuss GAF test-bed, GAF im plem entation architecture and GAF evaluation results from the test-bed. 10.1.1 GAF test-bed overview GAF is im plem ented on PC104 sensor network test-bed. Each node in the test-bed has a 66MHZ CPU w ith 16MB RAM and a flash disk, running Linux operating system. We use 14 nodes in the test-bed. Since directed diffusion [28] was available on PC104, we use directed diffusion as our routing protocol to forward data, in order to save tim e in im plem enting other routing protocols. Since GAF is protocol-independent, we quickly make the test environm ent ready for GAF evaluation. GAF is im plem ented as a sublayer between the diffusion core and the radio devices as shown in Figure 10.1 (See [22] for details of the diffusion software architecture). Such 132 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. im plem entation avoids any changes on the existing im plem entation of directed diffusion. GAF m odule also does not need to know about the directed diffusion m odule either. GAF m odule only filters GAF discovery packets from incoming packet flows and process them in the GAF control layer, and forwards all other packets to upper layer. GAF layer does not affect the outgoing packets from upper layer. 10.1.2 Radio transmission range GAF grid size is determ ined by node’s radio transm ission range. In order to determ ine a node’s radio transm ission range in the real world, we use an application developed in directed diffusion, called “linkScan” , to measure node connectivity in order to estim ate radio transm ission range. The LinkScan application overhears diffusions and periodically reports which nodes th a t a node can hear from. In order to quantify our m easurem ent, we define connectivity index as how often a node appears in another node’s report w ithin lim ited tim e (30 m inutes in our experim ent). In other words, connectivity index from node A to node B is defined as the ratio of the num ber of A ’s appearance in B ’s reports to to tal num ber of B ’s report. If a node frequently appears in another nodes’ report, the node’s radio range should be longer th an the distance between the two nodes. If a node rarely appears in another node’s report, the node’s radio range should be shorter than the distance between this two nodes. In order to quantify node transm ission range, we define two thresholds, high connectivity threshold and low connectivity threshold to help us to decide the radio transm ission range. In our test-bed, we choose the high connectivity threshold as 95% and low connectivity threshold as 10%. 133 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. © © © 4 ► more than 95% connectivity M ---------------------------------- ► Radio Range -4 ---------------------------------------------------- ► l e s s t h a n 10% c o n n e c t i v i t y Figure 10.2: Radio Range of node X By using the two thresholds, we can define a node’s conservative radio range as the furthest distance where the connectivity index from this node is larger th an high connectivity threshold, and a node’s aggressive radio range as the nearest distance where the connectivity index from this node is less th an the low connectivity threshold. A node’s radio transm ission range can be estim ated to be a value between its conservative radio range and aggressive radio range. Figure 10.2 shows an example of a node’s radio range. In this figure, node A is on the the conservative radio range position of node X, and node B is on the aggressive radio range position of node X. Node X ’s radio range is between the distance of X to A and X to B. Figure 10.3 shows the result of estim ated radio range based on the connectivity mea surem ent. We norm ally think th a t nodes w ithin nearer distance should have b etter con nectivity index th an nodes w ithin further distance, in other words, a node’s aggressive radio range should be longer th an its conservative radio range. However, our m easure m ent shows th a t it is not always true. Some nodes, such as node 15, 16, 26, 32, and 38 in Figure 10.3, have shorter aggressive radio transm ission range but longer conservative range. In other words, we found th a t nearer nodes have less connectivity constantly than further nodes. This example shows the difficulty to model the radio transm ission range 134 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 10 15 20 ra n g e (m em ters) Normal i — i — A nom alous ■ — x ~ Figure 10.3: Radio Range M easurement in the real world. In Figure 10.2, we m ark these nodes, which have shorter aggressive radio range but longer conservative radio range, as anomalous nodes. Since GAF requires conservative estim ated radio transm ission range in order to ensure nodes in any adjacent grid can com m unicate w ith each other, we choose 13.5m as node’s radio transm ission range, based on the m easurem ent results shown in the Figure 10.2. 10.1.3 Network Topology We have determ ined th a t the radio transm ission range is 13.5m. According to the Equa tion 4.2, we choose 5.5m as G A F’s grid size in the testbed. Figure 10.4 shows node location in the test-bed and how they are grouped in GAF grid. In grid 4, there are 5 nodes, 15,16,25,35,38, located alm ost in the same position. 135 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 20 15 10 0 0 5 10 15 20 25 Figure 10.4: Node location and Grid setup. The x and y axis are locations in meters. The energy model used in the test-bed is the same as th at used in GAF simulations. Since the test-bed is not really battery-operated, instead of directly m easuring the con sumed energy, we m easure the tim e th a t each node spends in each state and calculate consumed energy based on our energy model. A diffusion application runs on node 19 and node 37 respectively. A pplication on node 37 injects a packet every 10 seconds and the other application on node 19 collects the packet and m easure latency and packet reachability. The clock of bo th nodes are synchronized through external mechanism. The diffusion applications run over 4000 seconds for GAF and non-GAF configurations and the power consum ption of each node is recorded. 136 . . . . . . . . . . . . . . . . . 3 © Grid 1 ' — T “ ” afi) Grid 2 3 § ) Grid 3 2 © Grid 4 1© Grid 5 © O a o>....... ......................... Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Node 25 node 25 1500 2000 2500 time (sec) 4000 (a) With GAF Node 25 1.6 1.4 1.2 oi 0.8 3 0.6 0.4 0.2 500 1000 1500 2000 time (sec) 2500 3000 3500 4000 (b) Without GAF Figure 10.5: Conumed Energy 137 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 10.1.4 Energy dissipation Figure 10.5 shows the energy consumed by node 25 in grid 4 w ith and w ithout GAF. The graph is plotted as consumed power over time. O ther nodes in grid 4 show sim ilar results for energy consumption. We can find th a t node 25, when running GAF, periodically goes to sleep state to conserve energy. W ithout GAF, its energy is constantly dissipated. Apparently, a node can conserve energy w ith GAF. 10.1.5 Packet reachability We define packet reachability as the ratio of received packets over sent packets. Fig ure 10.6 shows the packet reachability in the testbed w ith and w ithout GAF configuration. The graph is plotted for each packet where if the packet is received, its reachability is 100%, otherwise, 0%, and for every 20 packets. The dot vertical line in the figure 10.6(a) shows the tim e when GAF switches the active node in grid 4. W ith GAF, we can find th a t during several periods, packets are not delivered at all. These periods start alm ost right after GAF switches the active node. This indicates th a t GAF m ight tu rn off the active node th a t is on the current forwarding p ath for load balance purpose. How fast a forwarding p a th can be rebuilt fully depending on the forwarding mech anisms. Since the directed diffusion im plem entation used in the test bed sends out its exploratory d ata every 30 seconds, it may take more th an 30 seconds to rebuild the forwarding path. All packets transm itted during this tim e will be dropped. 138 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Packet Reachability average of 20 packets . 100 > > o j s i o C O C D C D O C O Q . 4 f + ■ « I Hlim u -f- + 350 200 250 300 400 100 150 0 50 packet id (a) With GAF 100 I ««)» I I I ) 80 - = 6 0 .n o C O C D 5 40 o C O 20 - 0 50 100 Packet Reachability 1 ---------- 150 200 packet id 250 average of 20 packets ■ ■ III) II ............. r -mtf + + «-+ + 300 350 (b) Without GAF Figure 10.6: Conumed Energy 400 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In our sim ulation, since A O DV /D SR use MAC link detection mechanism to trigger routing recovery, the tim e to rebuild the routing p ath can be lim ited to milli-seconds. This is why our sim ulation results do not show such big packet loss rate. We will discuss this issue more in the future work (Section 11). O ther th an the packet loss caused by the switch-over, we can find GAF generally can keep alm ost the same packet reachability. 10.1.6 Packet delivery latency The packet delivery latency is m easured as the period from the tim e th a t a packet is sent by the source node to the tim e th a t it is received by the destination node. Figure 10.7 shows the packet delivery latency w ith and w ithout GAF in the test-bed. We also plot latency for each packet and latency for every 20 packets. We can find th a t the packet delivery latency varies considerably. This is due to exploratory d a ta of directed diffusion which is sent every 30 seconds. Ignoring the effect of exploratory data, packet delivery latency is alm ost the same w ith and w ithout GAF configuration. 10.1.T GAF evaluation Summary Our results based on the m easurem ent of test-bed show th at GAF can conserve en ergy while m aintaining routing fidelity. The packet reachability is affected by the rout ing/forw arding protocol’s capability to restore the routing path when GAF turns off a node on the active path. This is one of areas for the future work. 140 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. latency 0 7 average of 20 packets each packet 6 5 ++ + 4 3 2 *++t^ + + + + + + -«-+■ , H 1 0 400 250 300 350 50 1 0 0 150 200 0 packet id (a ) W i t h G A F latency 7 average of 20 packets each packet 6 5 4 + + + + + + + + - w - 3 2 + +, f V A++ ++ + + + + $ -t ++ ++ 1 ++ + 4 o 300 0 50 1 0 0 150 200 250 350 400 packet id (b) W ithout G AF Figure 10.7: Conumed Energy- Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 10.2 CEC im plem entation In this section, we discuss CEC test-bed and evaluation results of the test-bed. 10.2.1 CEC test-bed overview We ran experim ents w ith CEC on a testbed of 21 iPAQs. Each iPAQ is equipped w ith a UCB m ote [25], which we use as a radio interface. In addition, each iPAQ has an 802.11 card for experim ental control and logging purposes. The layout of the nodes forms a square grid structure, where a node is present at most of the vertices of the grid. The nodes have an average of seven neighbors each. In our CEC experiments, d ata packets are routed through the use of flooding: each node keeps a cache of the d ata packets it has received and, w ith each subsequent arrival of a packet, it forwards (broadcasts) th a t packet only if it is not already present in the cache. All nodes in the network generate traffic according to the following rule: every five seconds, each node probabilistically decides w hether to generate a new d a ta packet. The probabilities used generate an expected one new packet per five seconds over the whole network. The energy usage of each node is modeled after a 2Mbps radio for which transm ission requires 1.6W, reception requires 1.2W, idle listening requires 1.0W, and a radio turned off requires 0.025W. We only model the energy usage of the radio, since the energy usage of the other components of the node is negligible in comparison. Each node is given enough energy to rem ain in idle listening mode for 450 seconds. 142 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. >. 1A g> 12 Flooding, with CEC Flooding, without CEC 0 200 400 600 800 1000 1200 1400 Experiment Time (sec) Figure 10.8: Extension of network lifetime (95% confidence interval) 10.2.2 Extension of Network Lifetime We first look at the extension of network lifetime. Figure 10.8 shows the num ber of nodes w ith rem aining energy over tim e. W ithout CEC, there is a sharp drop at tim e 450s to zero, when all nodes run out of energy. W ith CEC, though, we see th a t network lifetime (20nodes remaining) is extended until tim e 1000s. Further, the fact th at the curve is sm ooth indicates th a t CEC successfully alternates awake and sleeping periods evenly between nodes: instead of keeping a small set of nodes awake until they run out of energy then choosing a new small set of nodes, CEC uses energy evenly from all nodes. 143 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Number o f Nodes 18 Number of Nodes With Remaining Energy >-+-—■ Number of Active Nodes ; x... lumber of Active Nodes Receiving Packets (Flooding) 16 14 12 10 8 6 4 2 j -ii 0 800 1000 1200 200 400 600 0 Experiment Time (sec) Figure 10.9: CEC active nodes over tim e (95% confidence interval) 144 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 10.2.3 Data Delivery Ratio Figure 10.9 shows the d ata delivery ratio among CEC active nodes over tim e. In this graph, we can also see how much of the network is active. The first thing to notice is th at the plot of the num ber of active nodes receiving packets and the plot of the num ber of active nodes are about the same. This means th a t about 100% of the active nodes are receiving the d a ta packets. From this, we can conclude th at the set of active nodes is indeed connected. The second thing to notice is the fluctuations in the bottom two plots. These fluctuations in the num ber of active nodes seem to be due to two phenomena. The first reason is th a t nodes in discovery phase are counted as active. The peak ju st after tim e 200s is due to the second discovery phase (all nodes wake up in order to re-choose cluster heads). W hile the periodic discovery phases amplify the fluctuations in this plot, the fluctuations are present even if one does not count nodes in discovery state as active. CEC chooses active nodes based on two factors: a node’s rem aining energy and a node’s num ber of neighboring nodes. In the first iteration of CEC, all nodes have approxim ately equal am ounts of energy. As a result, CEC chooses active nodes prim arily based on the num ber of neighboring nodes of a node. In the first iteration, therefore, the num ber of active nodes is close to the smallest possible for the given network topology. However, in the second iteration these same nodes (which make up something close to the m inim um dom inating set of nodes) cannot be picked (since they have used up half of their energy in the first iteration). Thus, a different and larger set of active nodes m ust be chosen in the second iteration. 145 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Number o f Nodes Number of Nodes With Energy h —t —- j Number of Nodes Receiving Packets (Flooding) ; x .... 16 1200 600 800 1000 200 400 0 Experiment Time (sec) Figure 10.10: D ata delivery ratio (95% confidence interval) 146 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Figure 10.10 shows the second part of the d ata delivery metric. We graph the num ber of nodes in the network th a t receive packets over time. The num ber of nodes depicted in this graph includes nodes th a t are asleep. The num ber of nodes on the y-axis is the sum of the num ber of active nodes and the num ber of nodes th at are both asleep and one hop away from an active node. This m etric shows us how much of the network is covered by (i.e. w ithin one hop of) the active nodes. We can see in Figure 10.10 th a t although CEC turns off a subset of nodes, d ata is still effectively delivered throughout the network. 10.2.4 CEC evaluation Summary O ur CEC im plem entation in the testbed shows th a t CEC works as designed to extend network lifetime. CEC has to build on the m easurem ent of reliable radio links. By adapting to the unreliable links in our testbed, CEC can have the same packet reachability as th at w ithout CEC. 10.3 Summary O ur testbeds successfully dem onstrates th a t G A F/ CEC can be successfully implem ented in the real world. G A F/C E C have been im plem ented in different testbeds, and integrated w ith different data tran sp o rt mechanisms. We m easured the network lifetime and found both G A F/C E C can extend network lifetime. W hile it is not new for the researchers in radio comm unication, the dram atic effect of radio propagation on energy-conservation routing protocols have not been addressed previously. In particular, GAF is affected by the estim ation of radio transm ission range. 147 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CEC is affected by the unreliable radio links. Our experiences w ith the testbed have led us to reexamine the radio propagation models in the sim ulator and extended it w ith more realistic behaviors. The packet reachability is affected by the routing/forw arding protocol’s capability to restore the routing p ath when GAF turns off a node on the active path. A local repair mechanism (such as used in AODV and DSR) can help routing protocols to quickly restore broken routing path. The testbeds help us select proper param eters for G A F/C E C to make them work more robustly in the real world. W ith appropriate tuning of G A F/C E C param eters, G A F/C E C can m aintain routing fidelity while extending network lifetime. 148 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 11 Conclusions and Future Work Conserving energy in ad hoc networks deployed w ith battery-pow erred nodes is a critical area of current research. We have dem onstrated a set of protocols to deal w ith the challenges of powering off redundant nodes to conserve energy while m aintaining routing fidelity by self-configuring mechanisms which adapt to network dynamics in a robust fashion. Ad hoc networks running such protocols w ith the dem and of energy conservation can support useful applications well enough. 11.1 Thesis Summary Self-configuration mechanism and localized distributed algorithm s have been our focus in our protocol design. The m ajor conclusions of this thesis are as follows: 11.1.1 Self-configuring mechanisms enable robust protocols Self-configuration has been the focus of our energy conservation protocol developments. Due to the highly dynam ic nature of ad hoc networks, it becomes even more im portant to use self-configuration mechanisms in protocol design. 149 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. We started to use self-configuration mechanism in AFECA to adapt to network density for energy conservation. AFECA shows substantially extended network lifetime th an BECA. However, since AFECA only passively senses the network conditions, we found th a t there are still many spaces to explore to improve protocol performance. We then design GAF to actively m easure network and self-configure the nodes w ithin the same grid for energy conservation. GAF also introduces the technique of mobility adaption to adapt to network m obility in order to work robustly in high m obility networks. Simulations show th at GAF performs much b etter th an AFECA in energy conservation. GAF assumes Global Position Systems support in order to organize redundant nodes. However, not only is the GPS not available all the tim e, b ut also it leaves GAF vulnerable to radio propagation. We further develop CEC which enables the nodes to form clusters w ith self-discovery. Since CEC discovers network redundancy be the node itself through their discovery messages, it works more robust to adapt to radio propagation. CEC also dem onstrates its capability to extend network lifetime much longer th an GAF since it does not need to be conservative in organizing redundant nodes. O ur results show the im portance of using self-configuration,i.e., autonom ously mea suring and adapting to environm ental and system dynamic, to achieve robust operation in ad hoc networks. 11 .1 .2 L o ca lized , d istr ib u te d a lg o r ith m s ca n p ro v id e en erg y -efficien t d esig n Due to the rapidly changing network topology of ad hoc networks, it is hard or impossible to share the same states among all nodes. In addition, sharing states among all nodes 150 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. leads to high comm unication overhead. Such overhead could fail the effort on energy conservation. All of our protocols use localized, distributed algorithms. AFECA passively overhears neighbor nodes communications. G A F/C E C uses discovery messages which only reaches the adjacent nodes. Among these protocols, CEC has the highest com m unication over head since it has to uses discovery message to form clusters while G A F’s grid can be preset based on the location information, and AFECA does not even send any discovery message at all. Simulations show th a t CEC overhead is only up to 0.4% of to tal energy usage of the systems. Such low overhead ensures th at our protocol’s success in energy conservation. To find a network redundancy, the ideal solution would be have nodes are synchronized w ith the same state and control node duty cycle based the shared states. Unfortunately, this solution does not scale at all, and turns out to be unrealistic in dynam ic ad hoc net work environm ent. Instead, our localized, distributed approach ensures energy-efficient design. Simulations show th at such approach is one of the key factors for the success of our energy conservation protocols. 11.1.3 Sensitivity study is critical for protocol design and evaluation Sensitivity study is critical because it addresses two m ajor issues in protocol design: first, w hat factors have the m ost im portant effect on the protocols, and second, w hat param eters should be chosen in the protocols in order to have the best performance. O ur studies show th a t m obility p attern and radio propagation have the m ajor im pact on the energy conservation protocols in ad hoc networks. In order to adapt to the high 151 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. mobility, we uses m obility prediction algorithm s in both GAF and CEC so th a t they can perform well under high mobility. Sim ulation shows th at even a low m obility network could be affected by radio propa gation. T hrough sensitivity study, we choose appropriate param eters in CEC to control node duty cycles in order to adapt to tim ely-varying radio propagation. W ith the se lected param eters based on sensitivity study, CEC works more robust in dynam ic ad hoc networks. W hen our focus has been developing self-configuration for network protocols, the performance of self-configuring protocols is extrem ely sensitive to choice of param eters. Sensitivity study is critical for choosing appropriate param eters for optim al performance. 11.1.4 G AF/CEC is practical and implementable in the real world A lthough many ad hoc network schemes seem to work well in sim ulations, such results do not necessarily guarantee good perform ance in the real world (e.g., which has obstacles) or on nodes w ith real radios (the propagation of which is very difficult to model). The final test for any protocol is to deploy it in the environm ents for which it was designed and m easure it to determ ine if it can operate successfully. B oth G A F/ CEC have been successfully im plem ented in real sensor networks carrying traffic across m ultiple hops. Most of results in the real im plem entations m atch w hat we observed in sim ulations, which shows th a t G A F /C E C can work successfully to extend network lifetime in the real world. We also find the effect of radio propagation on pro tocol performance. Such findings have helped us to choose appropriate mechanisms and param eters in CEC in order to make it work robustly in the real world. They also helps 152 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. us to develop b etter radio propagation model in our sim ulation studies for future protocol developments. 11.2 Thesis Contributions This thesis dem onstrates th a t practical energy conserving ad hoc networks can be built, and provides m ethodology and tools th a t can be reused in ad hoc network research. The thesis makes three m ain contributions in the area of protocol design, research m ethodol ogy, and sim ulation tools for future research. In the area of protocol design, I have quantitatively dem onstrated th at the value of self-configuring, localized, distributed algorithm s in energy conservation protocols for ad hoc networks. I have also examined in detail how the self-configuring, localized and dis tributed mechanisms should interact to create B EC A /A FEC A , GAF, and CEC protocols. Through the comparisons of ad hoc routing protocols w ith and w ithout B E C A /A FEC A , GAF and CEC, I dem onstrated each protocol’s capability in extending network lifetime. The algorithm s of the self-configuration, localized and distributed approaches should help guide future designers in developing energy conservation protocols and other protocols. In the design of energy conservation protocols, I have proposed and used some design ideas which had not used by other protocols before our protocols are published, and make these ideas success. These ideas include independence of energy conservation on other protocols or systems, im portance of idle-state radio energy dissipation in energy conservation, trading off network density for energy conservation, balancing energy con servation w ith protocol robustness,and m obility prediction. I believe these ideas provide 153 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. new design spaces in future protocol developments. And finally, GAF and CEC, each as a complete protocol has dem onstrated its usefulness in the real world in two different sensor network testbeds. GAF also has been integrated w ith STEM by UCLA researchers to dem onstrate its success to work w ith energy efficient MAC in extending network lifetime even longer th an the MAC itself by powering off the redundant part of network. In the area of research methodology, I have developed a sensitivity study m ethodol ogy through sim ulations to study protocol perform ance under different m obility models, radio propagation models, traffic models, energy models, and location error models. I identified the problem of idle-state radio energy usage in conserving energy in wireless ad hoc networks and implem ented energy models in sim ulator and used it in protocol study which had not previously studied. I also showed the im portance to use such energy model in ad hoc network research. I extended the radio propagation model to capture the tim e-varying behavior of radio propagation in CEC protocol study. I used heavy traffic load to study whether network capacity can be m aintained under energy conser vation protocol which had not studied like this before. I also developed new metrics for m easuring and comparing network lifetime and energy conservation in order to ana lyze energy conservation protocols. The m ethodology I developed for sensitivity study of energy-conservation protocols will help the researchers to easily evaluate their protocols, choose appropriate param eters to make the protocols work robustly, and reduce the cost for deploying protocols. In the area of tool generation, I have been participating in the design, im plem entation and testing of V IN T /ns wireless extension. I integrated and restructured C M U /ns mobile node and extended it to support energy models and new radio propagation models for 154 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. protocol study. I also extended VIN T Network A nim ator (NAM) [18] to for visualize mobile network simulations. The extended sim ulator and anim ator, the B EC A /A FEC A , GAF, CEC im plem entations w ith ns and testbed, and their associated tools, validation test scripts are the concrete contributions of this thesis. They have been widely adopted and used by the researchers in ad hoc network area. 11.3 Future Work We have identified a num ber of areas for future work. 11.3.1 Coordination of ad hoc routing protocols and energy conservation protocols Although GAF and CEC are independent of underlying ad hoc routing protocols, there is a research area of developing a mechanism to coordinate ad hoc routing protocols and energy conservation protocols to achieve better routing fidelity. In order to conserve energy, GAF and CEC control node duty cycle independently. W hen a node is powered off, it is up to the ad hoc routing protocols to figure out new routing p ath in order to m aintain routing fidelity. Depending on the design of ad hoc routing protocols, the recovery tim e of finding a new p ath varies from milli-seconds to a few minutes! For example, A O D V /D SR can use 802.11 MAC support for link failure detection. They can recover routing failures in mini-seconds w ith the help of MAC support. However, when the MAC does not support link failure detection, such as the MAC we developed in-house for GAF im plem entation, it takes a few m inutes for the diffusion algorithm s to figure out the new path. It will cause long packet transfer delay and possible packet loss. 155 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. To get rid of the dependence on MAC link failure detection, a solution is to have the energy conservation protocols notify the underlying routing protocols before they tu rn off the nodes. W hile receiving such notifications, the routing protocols can gracefully switch to the new p ath before the node on the current routing path is turned off. The solution does not necessarily m ean th a t we have to combine the energy conserva tion protocols w ith the ad hoc routing protocols. An idea approach would be developing a (routing) protocol independent notification mechanism so th at the energy conservation protocol can invoke it without knowing the details of underlying ad hoc routing protocols. A (routing) protocol dependent “middle-ware” then takes the notification and adapts it to routing protocols individually. Such m echanism will ensure energy conservation can work w ith any ad hoc routing protocols. In other words, we still adhere to the rules of having the energy conservation protocols independent of ad hoc routing protocols. However, a new research area is to develop a “middle-ware” to glue energy conservation protocols seamlessly w ith ad hoc routing protocols to gain b etter routing fidelity. 11.3.2 Balance energy conservation with protocol robustness Different from other energy conservation approaches such as SPAN, all of our protocols do not pursue a “backbone” architectures. In order words, our protocols will still leave some redundant connections even after powering off redundant nodes. The m ain reason to allow lim ited redundant connections is th a t rapidly changing characteristics of ad hoc networks, especially its dynam ic radio propagation, could lead to flakey network connections when 156 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. network becomes less dense. It will make our protocol work more robustly by allowing some percentage of network redundancy. There always exists trade-off between energy conservation and protocol robustness. To one extreme end, the protocols will work most robustly if we do not power off any part of network. Of course there will be no any energy conservation. To the other extrem e end, we can keep infinite long node duty cycle to gain m aximum energy conservation but worst network connectivity. G A F/C E C uses grids/clusters mechanisms to balance such trade-off. Simulations and sensitivity studies help us to find appropriate param eters to obtain better balance point. However, research can be done in the area of providing more mechanisms for the applications to quantitatively control such trade-off more flexibly and accurately. For example, when G A F /C E C typically tries to get one grid/cluster leader in each grid/cluster, some applications may want to have a backup grid/cluster leaders so th a t in case of the failure of prim ary leaders, the backup leaders can take over quickly for robust purpose. In such applications, more control mechanisms are needed for the selection of backup nodes and switch-over mechanisms. O ther some applications, such as m onitoring in sensor networks, may not want quick takeover so th at they can p u t the node in long duty cycle and only wake them up when needed. The application-driven nature for energy conservation protocol requires flexible con trol mechanisms to balance the trade-off between energy conservation and robustness. 157 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 11.3.3 Joint work between Energy conservation protocols and energy- efficient MAC W hen energy-efficient MAC is more focused on individual node energy conservation, our energy conservation protocols are more focused on the network-wide energy conservation mechanisms. More specifically, our mechanisms find the redundant part of network and power them off where it is impossible for MAC to have such view of the network. Combining the energy conservation protocols and energy-efficient MAC will provide better energy conservation and longer network lifetime. STEM has showed the benefit of combining energy-efficient MAC with GAF so th a t it can gain much longer network lifetime th an ju st using STEM MAC or GAF individually. More work can be done in this area to combine different energy conservation protocols, such as C E C /G A F, w ith energy efficient MAC such as SMAC etc, in order to gain the best network lifetime. 11.3.4 Building tools for protocol developments Increased refinement of energy conservation protocols for ad hoc networks is always re quired. W ith the sim ulation and testbed tools developed as p art of this thesis, we have dem onstrated their benefit in improving the protocol performance. Currently the sim u lation and testbed im plem entation are independent of each other. As a future work, we expect to see a wireless em ulation environm ent to combine b oth sim ulation and testbed under the same unified framework. Such unified framework will benefit future protocol developments. 158 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Reference List [1] S Ahn and A U Snakar. A dapting to route-dem and and mobility (arm) in ad-hoc network routing. In 9th International Conference on Network Protocols, pages 56-66, November 2001. [2] Param vir Bahl and Venkata N. Padm anabhan. Radar: An in-building RF-based user location and tracking system. In Proceedings of the IEEE Infocom, pages 775-784, Tel-Aviv, Israel, M arch 2000. [3] D. J. Baker, A. Ephrem ides, and J. A. Flynn. The design and sim ulation of a mobile radio network w ith distributed control. IEEE Journal on Selected Areas in Communications, 2(1):226— 237, January 1984. . [4] Frazer B ennett, David Clarke, Joseph B. Evans, Andy Hopper, Alan Jones, and David Leask. Piconet: Em bedded mobile networking. IEEE Personal Communica tions Magazine, 4(5):8— 15, October 1997. [5] C B ettstetter. Sm ooth is better th an sharp: A random m obility model for sim ulation of wireless networks. In A C M MSWiM, pages 19-27, July 2001. [6] Solomon Bien. Connectiviy study of a csma based wireless network, Jan., 2002. U npublished m anuscript. [7] Lee Breslau, D eborah Estrin, Kevin Fall, Sally Floyd, John Heidemann, Ahm ed Helmy, Polly Huang, Steven McCanne, K annan Varadhan, Ya Xu, and Haobo Yu. Advances in network sim ulation. IEEE Computer, 33(5):59— 67, May 2000. Ex panded version available as USC T R 99-702b at h ttp ://w w w .is i.e d u /~ jo h n h / PAPERS/Baj a j 9 9 a .htm l. [8] J. Broch, D.A. M altz, D.B. Johnson, Y. Hu, and J. Jetcheva. A perform ance com parison of m ulti-hop wireless ad hoc network routing protocols. In Proceedings of the A C M /IE E E International Conference on Mobile Computing and Networking, pages 85-97, October, 1998. [9] Josh Broch, David B. Johnson, and David A. Maltz. The dynamic source routing protocol for mobile ad hoc networks. IN TERNET-DRAFT, draft-ietf-manet-dsr- 03.txt,, O ctober 1999. Work in progress. [10] N irupam a Bulusu, John Heidemann, and D eborah Estrin. GPS-less low cost out door localization for very small devices. IEEE Personal Communications Magazine, 7(5):28-34, O ctober 2000. 159 with permission of the copyright owner. Further reproduction prohibited without permission. [11] A Cerpa and D Estrin. Ascent: A daptive self-configuring sensor network topolo gies. Technical R eport CSD-TR-01-0009, UCLA Com puter Science D epartm ent, May 2001. Available at w w w .is i.e d u /s c a d d s /p u b lic a tio n s .h tm l. [12] Jae-Hwan Chang and Leandros Tassiulas. Energy conserving routing in wireless ad- hoc networking. In Proceedings of the IEEE Infocom, pages 22-31, Tel Aviv, Israel, M arch 2000. A C M /IEEE. [13] Yu-Liang Chang and Ching-Chi Hsu. Routing in wireless/m obile ad-hoc networks via dynam ic group construction. Mobile Networks and Application, pages 27-37, May 2000. [14] Benjie Chen, Kyle Jamieson, H. Balakrishnan, and R obert Morris. Span : An energy-efficient coordination algorithm for topology m aintenance in ad hoc wireless networks. In Proceedings of the A C M /IE E E International Conference on Mobile Computing and Networking, pages 85-96, July, 2001. [15] Sam ir R. Das, Charles E. Perkins, and Elizabeth M. Royer. Performance comparison of two on-dem and routing protocols for ad hoc networks. In Proceedings of the IEEE Infocom, pages 3-12, Tel Aviv, Israel, M arch 2000. [16] Lance Doherty, Kristofer S. J. Pister, and Laurent El Ghaoui. Convex position estim ation in wireless sensor networks. In Proceedings of the IEEE Infocom, pages 1655-1663, Alaska, April 2001. IEEE. [17] D eborah Estrin, Ram esh Govindan, John Heidemann, and Satish K um ar. Next century challenges: Scalable coordination in sensor networks. In Proceedings of the A C M /IE E E International Conference on Mobile Computing and Networking, pages 263-270, Seattle, W ashington, USA, August 1999. ACM. [18] D eborah Estrin, M ark Handley, John Heidemann, Steven M cCanne, Ya Xu, and Haobo Yu. Network visualization w ith nam, the vint network anim ator. IEEE Computer, 33(11) :63— 68, November 2000. [19] ns notes and docum entation. The VINT Project, UC Berkeley, LBL, U SC /ISI, and Xerox PARC, November 1997. Available from http://w w w - mash. cs. berkeley.edu/ns /. [20] S. Floyd, V. Jacobson, C. Liu, S. McCanne, and L. Zhang. A reliable m ulticast framework for light-weight sessions and application level framing. In Proceedings of the A C M SIGCOMM, pages 342-356, Cambridge, M assachusetts, August 1995. ACM. [21] J Heidemann, N Bulusu, J Elson, C Intanagonwiwat, K Lan, Ya Xu, and Wei Ye. Effects of detail in wireless network simulation. In SCS Multiconference on Dis tributed Simulation, pages 3-11, Phoenix, Arizona, USA, January 2001. Available from http://w w w .isi.edu/scadds/ 160 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [22] John Heidemann, Fabio Silva, Yan Yu, D eborah Estrin, and Padm aparm a Haidar. Diffusion filters as a flexible architecture for event notification in wireless sensor networks. Technical Report ISI-TR-556, U SC /Inform ation Sciences Institute, April 2002. [23] W R Heinzelman, A Chandrakasan, and H Balakrishnan. Energy-efficient com m unication protocols for wireless m icrosensor networks. In Hawaii International Conference on System Sciences, pages 2-12, January 2000. [24] W endi R. Heinzelman, H oanna Kulik, and Hari Balakrishnan. Adaptive protocols for inform ation dissem ination in wireless sensor networks. In Proceedings of the A C M /IE E E International Conference on Mobile Computing and Networking, pages 175-185, August, 1999. [25] J Hill, R Szewczyk, A Woo, S Hollar, D Culler, and K Pister. System architecture directions for networked sensors. In Proceedings of the 9th International Confer ence on Architectural Support for Programming Languages and Operating Systems (ASPLOS IX), pages 93-104, Cambridge, MA, Nov, 2000. [26] X Hong, M Gerla, G Pei, and C Chiang. A group m obility model for ad hoc wireless networks. In A C M /IE E E M S W iM ’99, pages 53-60, August 1999. [27] Y Hu and D Johnson. Caching stratgies in on-dem and routing protocols for wireless networks. In Proceedings of the A C M /IE E E International Conference on Mobile Computing and Networking, pages 231-242, August 2000. [28] C Intanagonwiwat, R Govindan, and D eborah Estrin. Directed diffusion: A scalable and robust comm unication paradigm for sensor networks. In Proceedings of the A C M /IE E E International Conference on Mobile Computing and Networking, pages 70-84, Boston, M assachusetts, USA, Aug, 2000. [29] P. Johansson, T. Larsson, N. Hedm an, B. Mielczarek, and M. Degermark. Scenario- based perform ance analysis of routing protocols for mobile ad-hoc networks. In Proceedings of the A C M /IE E E International Conference on Mobile Computing and Networking, pages 195-206, August, 1999. [30] Oliver Kasten. Energy consumption. ETH-Zurich, Swiss Federal Institute of Technology. Available at h ttp ://w w w .in f .e th z .c h /~ k a s te n /r e s e a r c h /b a th tu b / energy_consum ption.htm iy,, 2001. [31] Y-B. Ko and N.H. Vaidya. Location-aided routing (LAR) in mobile ad hoc networks. In Proceedings of the A C M /IE E E International Conference on Mobile Computing and Networking, pages 66-75. ACM, 1998. [32] P. K rishna, N.H. Vaidya, M. Chatterjee, and D.K. Pradhan. A cluster-based ap proach for routing in dynamic networks. A C M SIGCOM M Computer Communica tions Review, pages 372-378, 1997. 161 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [33] LAN MAN Standards Com m ittee of the IEEE Com puter Society. Wireless L A N Medium Access Control (M AC) and Physical Layer (PHY) Specification, 1999. IEEE Std 802.11, 1999 Edition. [34] J. Li, J. Jannotti, D.S.J. De Couto, D.R. Karger, and R. Morris. A scalable location service for geographic ad hoc routing. In Proceedings of the A C M /IE E E Interna tional Conference on Mobile Computing and Networking, pages 120-130, August, 2000 . [35] A. Bruce M cDonald and T. Znati. A m obility based framework for adaptive cluster ing in wireless ad-hoc networks. IEEE Journal on Selected Areas in Communications, 17(8)=1466-1487, August 1999. [36] Yutaka Mori. Connectiviy study of a csma based wireless network, Jan., 2002. Un published m anuscript. [37] R North, D Bryan, and D Baker. W ireless networked radios com parison of mili tary, commercial and r& d protocols. In 2nd Annual UCSD Conference on Wireless Communications, pages 88-98, February 1999. [38] V. Park and S. Corson. Tem porally-ordered routing algorithm (TORA) version 1 functional specification. IN TERN ET-D RAFT, draft-ietf-manet-tora-spec-02.txt,, O ctober 1999. Work in progress. [39] G. Pei, M. Gerla, and X. Hong. Lanm ar: Landm ark routing for large scale wireless ad hoc networks w ith group mobility. In IE E E /A C M MobiHOC, pages 11-18, Boston, M assachusetts, USA, August 2000. [40] Charles Perkins. Ad hoc on dem and distance vector (AODV) routing. Internet- Draft, draft-ietf-manet-aodv-Of.txt, pages 3-12, October 1999,Work in progress. [41] Charles E. Perkins and Pravin Bhagwat. Highly dynamic destination-sequenced distance-vector routing (DSDV) for mobile computers. In Proceedings of the A C M SIGCOMM, pages 234-244, Auguest 1994. A revised version of the paper is available at h t t p : / /www. cs.u m d . ed u /p ro jec ts/m cm l/p a p e rs/S ig c o m m 9 3 . p s . [42] Gregory J. P ottie and W illiam J. Kaiser. Em bedding the internet: wireless integrated network sensors. Communications of the ACM, 43(5):51-58, May 2000. [43] B Prabhakar, E Uysal-Biyikoglu, and A E Gamal. Energy-efficient transm ission over a wireless link via lazy packet scheduling. In Proceedings of the IEEE Infocom, pages 386-394, April 2001. [44] R. R am anathan and R. Rosales-Hain. Topology control of m ultihop wireless net works using transm it power adjustm ent. In Proceedings of the IEEE Infocom, pages 404-413, M arch 2000. [45] Theodore S. R appaport. W ireless communications: Principles and practice. Prentice Hall, Upper Sadde River, New Jersey 07458, Reprinted 1999. 162 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. [46] E.M. Royer, P.M .M elliar-Smith, and L.E. Moser. An analysis of the optim um node density for ad hoc mobile networks. In IEEE International Conference on Commu nications, pages 36-41, June 2001. [47] R R uppe, S Griswald, P Walsh, and R m artin. Near term digital radio (ntdr) systems. In MILCOM, pages 1282-1287, November 1997. [48] M Sanchez, P Manzoni, and Z. J. Haas. D eterm ination of critical transm ission range in ad-hoc networks. In Multiaccess, Mobility and Teletraffic for Wireless Communi cations (M M T ’99), pages 45-54, O ctober 1999. [49] H. Schulzrinne, S. Casner, R. Rrederick, and V. Jacobson. RTP: A tran sp o rt protocol for real-tim e application. RFC 1889, IE T F, January 1996. [50] C Schurgers, V Tsiatsis, and M Srivastava. Stem: Topology m anagem ent for energy efficient sensor networks. In IEEE Aerospace Conference, pages 78-89, M arch, 2002. [51] S. Singh and C.S. Raghavendra. PAMAS: Power aware multi-access protocol with signalling for ad hoc networks. A C M Computer Communication Review, 28(3):5-26, July 1998. [52] S. Singh, M. Woo, and C.S. Raghavendra. Power-aware routing in mobile ad hoc networks. In Proceedings of the A C M /IE E E International Conference on Mobile Computing and Networking, pages 181-190, October, 1998. [53] M ark Stem m and Randy H. Katz. M easuring and reducing energy consum ption of network interfaces in hand-held devices. IEICE Transactions on Communications, E80-B(8): 1125-1131, August 1997. [54] The CMU M onarch Project. The CMU m onarch projects wireless and m obility extension to ns. h t t p : //w w w .m onarch.cs. edu. [55] P Tsuchiya. The landm ark hierarchy: A new hierarchy for routing in very large networks. In Proceedings of the A C M SIGCOMM, pages 35-42, Stanford, CA, USA, 1988. [56] Ya Xu, John Heidemann, and D eborah Estrin. Adaptive energy-conserving routing for m ultihop ad hoc networks. Technical R eport TR-2000-527, U SC /Inform ation Sci ences Institute, O ctober 2000. Available at f t p : / / f t p . i s i . e d u / i s i - p u b s / t r - 5 2 7 . pdf. [57] Ya Xu, John Heidemann, and Deborah Estrin. Geography-informed energy conserva tion for ad hoc routing. In Proceedings of the A CM /IEEE International Conference on Mobile Computing and Networking, pages 70-84, July, 2001. [58] Wei Ye, John Heidemann, and D eborah Estrin. An energy-efficient mac protocol for wireless sensor networks. In Proceedings of the IEEE Infocom, pages 3-12, New York, USA, June 2002. 163 with permission of the copyright owner. Further reproduction prohibited without permission. [59] Y Zhao, R Govindan, and D Estrin. Aggregate network properties for m onitoring wireless sensor networks, 2002. U npublished M anuscript. [60] Y Zhao, R Govindan, and D Estrin. Residual energy scans for m onitoring wireless sensor networks. In Proceedings of IEEE Wireless Communications and Networking Conference (W C N C ’02), pages 78-89, Florida, USA, March, 2002. [61] Yonggang Jerry Zhao. Connectiviy study of a csma based wireless network, Jan., 2002. U npublished m anuscript. 164 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Applying aggregate-level traffic control algorithms to improve network robustness

PDF

Diagnosis and localization of interdomain routing anomalies

PDF

Adaptive routing services in ad-hoc and sensor networks

PDF

Directed diffusion: An application -specific and data -centric communication paradigm for wireless sensor networks

PDF

Grid workflow: A flexible framework for fault tolerance in the grid

PDF

Characterizing Internet topology, routing and hierarchy

PDF

Energy latency tradeoffs for medium access and sleep scheduling in wireless sensor networks

PDF

Systematic test synthesis for multipoint protocol design

PDF

Design issues in large-scale application -level routing

PDF

Distributed resource allocation in networks for multiple concave objectives

PDF

Boundary estimation and tracking of spatially diffuse phenomena in sensor networks

PDF

Call admission control and resource allocation for quality of service support in wireless multimedia networks

PDF

Analysis of wired short cuts in wireless sensor networks

PDF

Design of wireless sensor network based structural health monitoring systems

PDF

Algorithms for performance and trust in peer -to -peer systems

PDF

A hybrid systems modeling framework for transport protocols

PDF

Architecture -independent programming and software synthesis for networked sensor systems

PDF

Evaluating the dynamics of agent -environment interaction

PDF

Consolidated logic and layout synthesis for interconnect -centric VLSI design

PDF

Congestion control in multi-hop wireless networks

Asset Metadata

Creator Xu, Ya (author)

Core Title Adaptive energy conservation protocols for wireless ad hoc routing

School Graduate School

Degree Doctor of Philosophy

Degree Program Computer Science

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

Tag Computer Science,OAI-PMH Harvest

Language English

Contributor Digitized by ProQuest (provenance)

Advisor Heidemann, John (committee chair), Estrin, Deborah (committee member), Govindan, Ramesh (committee member), Helmy, Ahmed Abdel-Ghaffar (committee member)

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c16-274861

Unique identifier UC11339228

Identifier 3093948.pdf (filename),usctheses-c16-274861 (legacy record id)

Legacy Identifier 3093948.pdf

Dmrecord 274861

Document Type Dissertation

Rights Xu, Ya

Type texts

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

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the au...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus, Los Angeles, California 90089, USA