On the Performance of Ad Hoc Networks with
Beamforming Antennas
Ram Ramanathan
Internetwork Research Department
BBN Technologies (A part of Verizon)
Cambridge, Massachusetts, USA
ABSTRACT
Beamformingantennashavethepotentialtoprovidea funda-mentalbreakthroughinadhocnetworkcapacity. Wepresent abroad-basedexaminationofthispotential,focusingon ex-ploitingthelongerrangesaswellasthereducedinterference thatbeamformingantennascanprovide.Weconsidera num-berofenhancementsto aconventional adhoc network sys-tem,andevaluatetheimpactofeachenhancementusing sim-ulation. Suchenhancementsinclude\aggressive"and \con-servative"channelaccessmodelsforbeamformingantennas, linkpowercontrol, and directionalneighbor discovery. Our simulationsarebasedondetailedmodelingofsteeredaswell asswitchedbeamsusingantenna patternsofvaryinggains, andarealisticradioandpropagationmodel. Forthe scenar-ios studied,our results showthat beamforming can yielda 28%to118% (dependinguponthe density)improvementin throughput,anduptoafactor-of-28reductionindelay. Our studyalso tells us which mechanismsare likely to be more eectiveandunderwhatconditions, whichinturnidenties areaswherefutureresearchisneeded.
1.
INTRODUCTION
Sincetheearlydaysof adhoc networking, researchershave strivedtoincreasethecapacityofadhocnetworksthrougha variety ofinnovativetechniques. Alongtheway,ithasbeen realizedthat therearefundamentallimitations tohowhigh onecanpushthecapacity. Underlyingthislimitation isthe inherent tension between the number of end-to-end packet hopsand the spatialreuse per hop {that is, if one wishes to decrease the number of multihop transmissions end-to-end(andtherebyreducethedemandsonthenetwork),then onehastoincreasetherangepertransmission,whichresults inareductioninthe numberofsimultaneoustransmissions. Thishasbeenthoroughlyexaminedin[1],whoshowthatthe throughputobtainableby eachnodeis (
W p
(nlogn)
), where
W isthedatarate,andnisthenumberofnodes. This lim-itationon capacity exists no matterwhat routing protocol orchannelaccessschemeisused. Splittingthe channelinto subchannelsdoesnotchangethisresult[1].
Oneof the chief contributors to this capacity limitation is the omni-directional nature of transmissions. Specically, the distribution of energy in directions other than the in-tendeddirectionnotonlycausesunnecessaryinterferenceto othernodes,butalsoreducesthepotentialrangeofthe trans-mission(duetolowersignalstrengthandmultipath compo-nents). Ontheotherhand,withdirectionalcommunications, bothspatial reuse and range { the two contributors to the capacitylimitation{wouldbesimultaneouslyenhanced. In-deed,thecapacitylimitin[1]assumesomni-directional trans-missionsandacknowledges,withoutdetails,that beamform-ing will be \advantageous". In recent years, beamforming technologyhasmadegreatstrides,andoersanunique and timelyopportunityunshackle the capacity limitations ofad hocnetworks.
Whiletheuseofdirectionalcommunicationsinadhoc newt-works promises to open new doors, a number of questions arise. Howcanweachievedirectionalcommunications?What kindsofantennasareavailableforthepurpose? Aretheybe too expensive and/or bulky to make sense? What implica-tions does it have for networking { is it just a question of replacingtheomni-directionalantennawithasmartantenna orwill networking protocols have to be changedto harness thepotential? Whichaspectsofnetworkingareimportantin harnessingthispotential,andwhicharenot? Doesitdepend upon the kind of antennas used? How much performance improvementcanweget? Isitworththetrouble?
(connectivity)andlatencyofanadhocnetwork. Weemploy adetailedandthoroughmodelingofbeamformed communi-cations,usingspecicantennapatternsofdierentgains,and modelingofsideas wellas mainlobes. Finally,weconsider andcomparesteered(e.g. adaptivearrays)andswitched(e.g. diversity)beamforming. Theresultsofourstudyprovidethe motivationandmethodologyforusingbeamformingantennas asa\rstclasscitizen"inadhocnetworking.
Manyofthebenetsofabeamformingantennasuchashigher signal-to-noiseratio, multipathmitigation, etc. arerelevant toandcanbeexploitedatthephysicallayer. However,thisis notthesubjectofthispaper. Ratherweask: what network-ing and medium access advantages does beamforming pro-vide,andhowcantheybeexploited? Inparticular,we con-sidertheexploitationoftwoadvantages: thereduced interfer-enceduetothenarrowerbeamwidths,andtheextendedrange dueto higher signal-to-noise ratio (by virtue of the higher gainandreducedmultipath). Ourfocusisonexaminingthe relative performanceimprovementoveromni-directional an-tennas, and specically on establishing a lower bound. As willbecomeapparent later,this goaldrivesmanyofour as-sumptionsand experimental methodology. Whileour work isinthecontextofaproactiveadhocroutingprotocol,most ofourresultsindependentoftheroutingprotocolperse.
2.
RELATED WORK
Whiletherehas beenavastamount ofresearchonphysical layer issues related to beamforming antennas, only a small amount of work considers the medium access and network layerimplications ofbeamforming antennas for adhoc net-working. Zander [2] hasproposed theuseof directional an-tennasinaslottedALOHAmultihoppacketradionetwork. Morerecently[3]presentseveralmediumaccesscontrol pro-tocols forxeddirectionalantennas. Theseprotocolsutilize physical location information to implement collision avoid-ance. Subsequently,[4] suggest a protocol that usessignal strengthinformationinsteadofpositioninformation. Beam-forming antennas forsingle-hop packetradionetworks have beenstudiedin[5,6]andotherpapers. Theuseofdirectional oodstolimitthescopeofrouterequestsinan\on-demand" adhocroutingprotocolissuggestedin[7],andexploredin[8].
Ourworkis uniqueinseveralways. First,mostofthework citedabovehaverestrictedthemselvestoaspecicaspectof utilizingdirectionalantennas,suchasmediumaccess,whereas wemodeltheeectsofanumberofmechanismsapartfrom medium access such as directional neighbor discovery, link power control, etc. and consider their interactions and ef-fectiveness within a full-edged ad hoc networking system. Second,noneof theabove worksevaluatetheimpactofthe longer-range links possible withdirectional antennas in im-proving the connectivity and reducing latency. Third, the onlymultihopspecicworksaboveconsidermultiplexed di-rectionalantennaswhereasweconsiderthesupportformore sophisticated beamforming such as steered beams. Finally, ourstudyisbasedonafarmorecomprehensiveantennaand propagationmodelwithmultipleantennapatterns,modeling ofsidelobes,andtheuseofantennatransmitandreceivegains inthebit-error-rate calculation at eachnode. In summary, this is the rstbroad-based, realistic analysis of how much beamformingwillhelp,andwhatitrequiresoftheMACand
3.
BEAMFORMING ANTENNAS: CONCEPTS,
TERMINOLOGY, RELEVANCE
Inthis section, we discuss some concepts related to beam-formingantennas. This isnot intendedto coverall aspects ofthis technology,nor do we coverit preciselyor formally. Rather,the ideais togivethebasics inaninformaland in-tuitivefashiontoequipthereaderunfamiliarwiththistopic withjustenoughknowledgetounderstandtheremainderof thispaper. Readersfamiliarwithbeamformingantennasmay skipthissection. Readerswishingtoexplorethiseldin de-tailarereferredto[9]andthecitationstherein.
3.1
Antenna Concepts
Radioantennascouple energyfromonemediumtoanother. Anomnidirectionalantenna(sometimesknownasanisotropic antenna)radiatesorreceivesenergyequallywellinall direc-tions
1
. Adirectionalantennahascertainpreferred transmis-sionandreceptiondirections,thatis,transmits/receivesmore energyinonedirectioncomparedtotheother.
Thegainofanantennaisanimportantconcept,andisused toquantifythedirectionality ofanantenna. Thegainofan antennainaparticulardirection
~
d=(;)isgiven[9]by
G( ~
d)=
U( ~ d) Uave
(1)
where U( ~
d)is the powerdensity inthe direction ~
d,Uave is theaveragepowerdensityoveralldirections,andisthe ef-ciencyoftheantennawhichaccountsforlosses. Informally, gainmeasuresthe relativepowerinonedirectioncompared toanomnidirectionalantenna.Thus,thehigherthegain,the moredirectionalis theantenna. Thepeakgain isthe maxi-mumgain taken overall directions. Whena singlevalue is givenforthegainofanantenna,itusuallyreferstothepeak gain. Gainisoften measuredinunitless decibels(dBi),that is,G
dBi
=10log 10
(G abs
). Anomni-directionalantennahas againof0dBi.
Anantennapattern isthespecicationofthegain valuesin eachdirectioninspace,sometimesdepictedasprojectionson theazimuthalandelevationplanes. It typicallyhasa main lobeofpeakgainand(smallergain)sidelobes. Asiscommon practice,weusethewordbeamasasynonymfor\lobe", es-peciallywhendiscussingantennaswithmultiple/controllable beams/lobes.
Arelated conceptisthe antennabeamwidth. Typically,this means the \3 dB beam width", which refers to the angle subtendedbythe twodirections oneitherside ofthe direc-tion of peak gain that are 3 dB down in gain. Gain and beamwidthare related. Typically, the moredirectional the antenna,higherthegain andsmallerthe beamwidth. How-ever, two antennas with thesame gain could have dierent beamwidths{forinstance,theantennawiththesmallermain lobewidthmayhavemoreorlargersidelobes.
1
intention-3.2
“Smart” Beamforming Antennas
Thesimplestwayofimprovingthe\intelligence"ofantennas isto havemultiple elements. Theslightphysical separation betweenelements,ordiversitycanbeusedcounteract multi-patheects. Therearetwowell-knownmethods. Inswitched diversitythesystemcontinuallyswitchesbetweenelementsso astoalwaysusetheelementwiththebestsignal. Whilethis reduces the negative eects of fading and multipath, there isnoincreaseingain. Indiversitycombining,the phase er-rorofmultipathsignalsiscorrectedandthepowercombined tobothreducemultipathandfading,aswellasincreasethe gain.
Thenextstepinsophisticationinvolvesincorporatingmore controlinthewaythesignalsfrommultipleelements(the an-tennaarray)areusedtoprovideincreasedgain,morebeams andbeamagility. Again, therearetwomainclassesof tech-niques,asdescribedbelow.
Inswitched beam systems, multiple xedbeams are formed by shifting the phase of each element's signal by a prede-terminedamount (thisis done by abeamforming network). Thetransceivercanthenchoosebetweenoneormorebeams fortransmittingorreceiving. Whileprovidingincreased spa-tialreuse,switchedbeamsystemscannottrackmovingnodes whichthereforeexperienceperiodsoflowergainastheymove betweenbeams.
Inasteered beamsystem,the mainlobe canbe pointed vir-tually in any direction, and often automatically using the receivedsignalfromthetargetusingsophisticated \direction-of-arrival" techniques. One may distinguish between two kindsofsteeredbeamsystems{dynamicphasedarrayswhich maximizesthe gain toward the target, and adaptive arrays whichadditionallyminimizethegain(producenulls)toward interferingsources.
Inthis paper,weconsiderswitchedbeamandsteeredbeam antennas,jointlyreferredtoasabeamformingantenna.
3.3
Relevance for Ad Hoc Networks
Whenconsidering the use of beamforming antennas for ad hoc networks, a question is: Aren't beamforming antennas too expensive and/or too big for ad hoc networks?. In this section,wearguethattheredoexistantennatechniqueswith suitablepriceandform-factorcombinations.
Applicationsforadhocnetworkingmaybeclassiedbroadly intothreecategories:military,commercialoutdoor,and com-mericalindoor,eachwithitsowndistinctiveprole,andable toaccommodatedierentantennatechnologies.
Militarynetworks,whicharebyfarthemostprevalent appli-cationofmobileadhocnetworks,containasignicant num-beroflargenodes(suchastanks,airplanes). Thesizeofthese platformsmakestheformfactorofmostantennasquite irrel-evant. Further,eachplatformbyitself is soexpensive that the costof eventhe most sophisticated antennais dwarfed bycomparison. Thus,beamforming antennasare extremely relevanttomilitarynetworks. Anaddedbonusisthatuseof directionaltransmissions have better immunity to jammers
Fixedadhocnetworksforcommercialoutdoorinsfrastructure extendthereachofbasestationsusingwirelessrepeaters or-ganizedintoameshnetwork. Heresteeredbeamapproaches maybetooexpensive. However,switchedbeamsusing inex-pensivebeamformingnetworkssuchastheButlermatrix[10] are easily manufactured using inexpensive hybrid couplers etc.[9]makingswitchedbeamformingquiterelevant.
Thebiggestdeterrenttousingbeamformingantennasfor net-workingsmallnodessuchasPDAsandlaptopswithinan in-door environmentis the size. At 2.4 GHz, and the typical half-wavelengthelement spacing, an eight element cylindri-calarraywouldhavearadiusofabout8cm,makingitquite unwieldy. However, asthe operatingfrequencycontinues to increase (already the IEEE 802.11a is working on wireless LANsinthe5GHzband),theantennasizeswill shrink. At the5.8 GHz ISM band,the 8-elementcylindrical arraywill havearadiusofonly3.3 cm,and atthe24GHz ISMband, amere 0.8 cm. Thus, the futurelooks bright for applying beamformingtechnologyeventosuchapplications.
Thus,whileatrstglanceitmayseemthatadhocnetworks andbeamformingantennasarenotcompatible,amorecareful examinationopensupanumberofpossibilities.
4.
PROTOCOL MODELS
In this section, we describe the details of each key proto-col used in the simulation system. The simulation system was developed using the OPNET modeling and simulation tool, version6.0. Comparedtomorecommonlyusedadhoc networksimulation tools suchas ns-2 and GloMoSim, OP-NET oers better support for directional communications. Specically,OPNETallowsthecreationofarbitraryantenna gain patterns in 3-D, allows the pointing of the main lobe towardanarbitrary positionin3-Dandautomatically com-putestheenergyreceivedateverynodeinthesystemtaking intoaccountthedierentgainsindierentdirections. Italso supports a detailed modeling of path loss, SNR, SIR, and bit-error-ratecomputations. These featuresallow foravery realisticmodelingofdirectionalradiocommunications.
Oursimulationsystemcontainssomeabstractmodels.An ab-stractmodelofaprotocolisonewhere someoftheprotocol dynamicsare replacedwith\short-cut"equivalents. For in-stance,insteadofmodelingthegeneration,transmissionand receptionofacontrolmessageoverthechannel,themessage couldbesent througha \back-door",thatis, asingleevent that delivers the necessary contents to the relevant mod-ule. Anotherexample is whenmultiple shortmessages are replacedwithasinglelongmessage (ora\packettrain").
Abstractionisawayoftradingodelityforshorterrunning timesand quickerdevelopment. It is appropriate whenthe purposeistodeterminetherough-order-of-magnitude perfor-mancegures,orwhencomparingmechanismswherethe ab-stractiontendstoaecteachmechanismmoreorlessequally, so that the relative performance is not aected as much
2 . Boththesecasesapplytoourwork,aswellastheubiquitous 2
needforshortrunningtimes. Therefore,someofourmodels areabstract.
We now discuss the protocol models. We begin with the beam forming and pointing model, which, although not a \protocol",formsakeycomponentofanyadhocnetworking studywithbeamformingantennas.
4.1
Antenna Patterns and Beam Control
Eachnodeinthesimulationsystemhasanantennathatcan be associated with one of a number of predened antenna patterns. Beamsteeringismodeledbyorientingthepattern sothatthecenterofthemainlobepointstowardthetarget node. Werst describeour antenna patternsand then the pointingtechniques.
4.1.1
Antenna Patterns
Sincemodelinga real antenna withprecise values for main andsidelobesisdiÆcult anddeviatesfromthe focus ofthis paper, we use an approximate antenna pattern as follows. Given a gain value gm, the antenna pattern for this gain consistsof amainlobe ofbeamwidth
m
, anda sidelobe of gaingsofbeamwidth(2-m). Thatis,the mainlobeisa coneofuniform gain,and thesidelobesare aggregated toa single\bulb"atthebaseofthecone. Figure1illustratesone ofthepatternsusedinoursimulations.
Figure1: Antennapattern20dBi
Asdiscussedinsection3,adirectionalantennamerely redis-tributestheenergy. Thus,thechoiceof
m andg
s
mustbe chosentoprovidearealisticmodelofanantennabeam,while at thesame time notbreaking any laws of nature. Speci-cally, the question is: in order to model anantenna as in gure1withagiven(mainlobe)gainofgm,whatshouldthe itsbeamwidthmand sidelobe gaings be,sothatthe total energyisthesame. Thisquestionwillariseinanyworkthat wishes to use a quick, yetreasonable model for directional antennas. Wederivebelowaformulaforcomputingm and g
s
,givenagain value g m
,and hopethatit mightbeof use tootherresearchers.
Firstwederivetheapproximatemaximumbeamwidthmax possible for g
m
. Consider a sphere of some radius r (no loss of generality). The surface area A on the sphere for abeamwidthofmcanbeapproximatedasacircleofradius rtan(
max
=2). LetS bethesurfacearea ofthesphere, and
m m s
10dBi 60deg -7.4dBi
14dBi 40deg -7.6dBi
20dBi 20deg -6.5dBi
26dBi 10deg -4.0dBi
Table 1: Antennapatternparametersusedinourstudy
g m
= P=A P=S
=
4r 2 (r
2 tan
2 (
max =2))
(2)
Solvingfor max
,
max=2tan 1
r 4 gm
(3)
Wethenpicka m
< max
. Thedierenceleavesenergythat canbeusedforsidelobes, andisdependentuponhowmuch sidelobe gain we want to model. In this study, we simply choose
m
tobethe largestmultipleof 10that is lessthan max,forpurelypragmaticreasons(OPNETantennapattern granularity).
Given m
andg m
,wenowderiveg s
tocompletethe model, asfollows. Bydenition,
gmUavA+gsUav(S A)=P (4)
whereU av
istheaveragepowerdensity,andisgivenbyP=S. Letting represent S=A, substituting, and simplifying, we have
g s
=
g
m
1
(5)
where=
4 tan
2 m
2
(usingexpressionsforS andA,similarto equation2). Asanaside,isanoftenusedquantitycalled thedirectivityoftheantenna.
Usingthisformula,wehavegeneratedandusedfourantenna patternswithparametersillustratedinthetable1.
4.1.2
Beam Steering and Switching
Beamsteeringismodeledasfollows. Assumethatapattern P correspondingto gain ghasbeenassociatedwith anode S. Supposethat S wantsto sendapacketusingasteerable beamtoanothernodeR .
tennacontainsKswitchedbeams. Weassumethatthebeams are identical (have a pattern P
g
), and are equally spaced. Withoutlossofgenerality,thedirectionofeachbeamisxed upon startup,withdirection of antennai isgivenby D
i = (i 1)360=Kdegreesrelativetodueeast. WhenKnot di-vide360,thedirectionsareapproximatelygivenbytheabove equation.
WhenanodeS wantsto transmitapacketto anodeR , it determinesthedirectionDthatRisrelativetoitself. Itthen determinesD
j
such that jD D
j
jis the minimum overall Di,i=1throughK. Thatis,itndsthebeamofminimum angularseparationwiththeintendeddirection. AfterS has obtainedaccesstothechannel(seesection4.2below),itsets itspatterntoP andthemainbeamispointedinthedirection Dj. Thepacketis thensent for transmission. Immediately after the packet is transmitted, S sets its pattern back to anomni-directionalpattern. Thus,weonlyusetransmit di-rectionality,all receivesare omni-directional. We note that switchedbeammodelingdoesnotnotrequireconstructionof aK-lobedpatternorKbeams.
Inboth steered and switched beam models, it is necessary to nd the relative direction of a neighboring node. This couldbedoneeitherusingpositioninformationor direction-or-arrivaltechniques within the antenna controller. We as-sume that the latter is used, and hence do not model the positioninformationdissemination
3 .
4.2
Channel Access
OurchannelaccessprotocolmodelisbasedontheCSMA/CA (CarrierSenseMultipleAccesswithCollisionAvoidance)
ap-proach. Traditional CSMA/CA works as follows. When
the carrier is free, asending node sends a Request-to-Send (RTS) packetto the intended receiver. The receiving node replies withaClear-to-Send(CTS). Thisexchangeacquires the channel \oor" for the communication by having any nodethat hearsthisexchangedesistfromtransmitting,and henceavoidscollision. Severalprotocolspublishedinthe lit-erature are representativesof this approach, including [11], and the the standardized wireless LAN mechanism IEEE 802.11[12].
ThegoalofCSMA/CAistohaveanodetransmitapacketif and only if there will be no collisions. In particular, not only should it be blocked if there will be a collision, but it should also not be blocked if there will be no collision. CSMA/CA meets this goal admirably in ad hoc networks when omni-directional antennas are used. However, with directionaltransmissions, this is not true whether oneuses omni-directional or directional RTS/CTS. One example of eachisillustratedingure2.
Ingure2(a),AwantstosendapackettoB,andCwantsto sendapackettoD.TheRTSfromAissentomni-directionally or directionallyto B, butis heard by C and Cis inhibited from sending even though it can do so without interfering withthetransmissionfromAtoB.
3
Ifonewishestousepositioninformation,thenthepredicted performanceimprovementshouldbedowngraded slightlyto reecttheoverheadofconveyingthisinformation,since
tradi-Figure2: Exampleswhen traditionalRTS/CTSis insuf-cientwhen beamformingisemployed.
InFigure2(b),supposeAissendingapackettoBafter hav-inginitiatedanRTS-CTSexchange. NeithertheRTSnorthe CTSisheardbyC,whichproceedstoinitiateatransmission toD.TheRTS-CTSexchangebetweenCandDisdirectional andnotheardbyA,which,aftercompletingitstransmission toB,nowinitiatesatransmissiontoE. TheRTSfromA to Einterfereswiththedatabeing receivedbyDfromC.
Thus,withdirectionalcommunications,thereceiptofanRTS orCTSdoesnot implythatyoumustbeblocked(refer g-ure2(a))andthenon-receiptofanRTSorCTSdoesnot im-plythatyoucantransmit(refergure2(b)). Moregenerally, thereisaspectrumoftradeosbetweenhavingbetterspatial reusewith parallel transmissions (but more collisions), and havinglessernumberofcollisions (butlessspatialreuse). A goodsolutiontotheproblemshouldbeparametrizedtoallow foraselectabletradeo inthisspectrum.
Thispaperdoesnotattemptto solvethis general problem. Rather,weconsiderthetwoendsofthespectrum,andhave modeledtwoprotocolscalledaggressiveCAandconservative CA,describedbrieybelow. Bothareoverandaboveabasic CSMAprotocol,andoperatewhenthepackethaspassedthe carriersensing. Inbothcases,the RTSand CTSare trans-mittedomni-directionally and are assumedto havea range at least that of the directional antenna
4
Both are abstract models,thatis,theRTSandCTSarenotactuallysentover theair{thishelpedusreducetherunningtimebyanorder ofmagnitude.
Aggressive Collision Avoidance Model. This models a protocolinwhichanodeisneverblockedupon receiv-ing an RTS or CTS. The handshake is used only for ensuringthat the receiveris notalready busysending or receiving. Inparticular,noattentionis paidto the receivestatus ofothernodes. Thus, thiscould poten-tially causecollision at nodesotherthanthe intended receiver.
ConservativeCollisionAvoidanceModel.Thismodelsa 4
protocolinwhichanodeisalwaysblockedupon receiv-ing an RTSor CTS. Thisis similar tothe traditional collisionavoidanceapproach. Inparticular,a transmis-sion takesplaceonlyif none ofthe nodes initsrange are busy. When used with directional antennas, this modesometimespassesuponcollison-freetransmission opportunities.
Weemphasizethatwedonotsuggesttheseasasolutionfor the beamforming MAC problem. Rather,by analyzing the performance for these two extremes, we can place a lower boundontheperformance ofafuturesolutiontothis prob-lem, presumablyonethat combinesthe bestcharacteristics ofeach. Workingtowardalowerboundonthe performance improvementwith beamformingantennas isconsistent with thegoalsofthispaper(seethelast paragraphinsection1). Aswillbeseeninsection5,evenwithsuchatrivialprotocol, onecangethugeperformancegains.
4.3
Link Power Control
The system can be congured to do link powercontrol on aper-packetbasis. Theidea hereis tocontrol the transmit powersothat itis justabout suÆcientto activatethe link. Thisreducesinterference,aswellasbatteryconsumption.
A straightforward way of doing this within the RTS/CTS frameworkisasfollows. TheRTSissentatapredetermined power(forinstance,the maximumpower). Thereceiver de-terminesthedierenceÆbetweenthereceivedpowerforthe RTSandits receive threshold. Thevalue ofÆ issent along withtheCTS.WhentransmittingtheDATA,thesenderuses apowerthatisÆlessthanthepowerusedfortheRTS. Op-tionally,itmayaddasmall\margin"tothevaluetoaccount for fading and mobility. This mechanism is widely imple-mentedinmilitaryradios. Avariantofthisideaisdescribed in[13].
Inourmodel,thereare noRTS/CTSpacketsactuallysent, andtherefore,weuseanabstractpowercontrolmodelwhich monitorsthereceivedpoweratthetargetnodeusinga \back-door"mechanism,andadjustsitstransmitpoweraccordingly.
4.4
Neighbor Discovery
NeighbordiscoveryisdoneusingaHelloprotocol. Eachnode periodically (with a small random jitter)transmits a Hello packet. ReceivedHellos aretrackedovera history(sliding) windowofsizeN. Aneighborisdeemedupifthenumberof Hellosreceivedfromthat neighboroverthe currentwindow isgreaterorequalto aconguredvalueKup. Ifthenumber ofHellosreceivedislowerthanaconguredvalueK
dow n the neighborisdeemedunreachableandthelinkiserased.
Whileneighbordiscoveryisfairlystraightforwardwith omni-directionalantennas,anumberofinteresting research prob-lems arise when beamforming is considered. For instance, consider two nodes A and B who are out of each other's rangewhenomni-directionaltransmissionsare usedbutcan communicatewhentransmitand/or receivebeamforming is used. The problem is that A and B mustboth determine independently { without communicating with each other {
ateachother .
Inthispaper,weconsidertwocases{omni-directional neigh-bordiscovery,anddirectionalneighbordiscovery. Intherst, allHellos aresentomni-directionally(but datapacketsmay besentdirectionally). Thus,theroutingtopologyisexactly thesameasonewouldgetwhenbeamformingdoesnotexist. Inthesecondcase,whichweonlyusewithswitchedbeams
6 , Hellos are sent directionally. In particular, a Hello is sent oneachofthe beams. SincetheseHellostravelfurther,the potential topologyis richerthaninthe rstcase. Thisis a detailedmodel, and incorporates all control messages with highdelity.
4.5
Routing
Weusethewell-knownlink-stateroutingprotocol.Although weareawareofitsscalabilitylimitations,itisnotofconcern tousinthisstudyasweconsideronly40nodenetworkswith nomobility. Further, the overhead inducedin most of our experiments(theones basedonomni-directional neighbors) isthesamewhetherornotbeamformingisusedfordata,and doesnotaecttherelativeperformance.
Briey,link-stateupdatesaretriggeredbyanodewhenoneof itslinksgoesup ordown, asdeemedbyneighbor discovery. Updatesare oodedthroughout the network, updating the forwarding table. Route generationis done usingDijkstra's shortest-pathalgorithm. For adhoc networks, scalable ver-sionsofthis\proactive"or\table-driven"approachabound, suchas [14]. Thisisadetailedprotocolmodeland incorpo-ratesall controlmessageswithhighdelity.
5.
EXPERIMENTAL RESULTS
Werstdescribe oursimulationenvironment,andthen dis-cusstheresults. Due tothe abstract natureof someof the models, we caution the reader that while the relative per-formance improvements are predicted with some degree of accuracy,theabsoluteperformancenumbersarenot.
5.1
Simulation Environment
OPNEThasadetailedpropagationmodelwherethebiterror rateonapacketiscomputedbasedonthesignal to interfer-enceandsignal tonoiseratios. Theradiomodelisbasedon adirectsequencespreadspectrumradio,withQPSK modu-lationandadatarateof1.6 Mbps. Thepropagationmodel uses the following equation, from [15], to compute the re-ceivedpower.
Pr(d)= Pt
2 d
2 ref
GtGr 4
2 d
4
(6)
where Pt and Pr are the transmitand received powers re-spectively,isthe wavelength,Gt andGr are thetransmit antennaand receiveantennagainsrespectively,disthe dis-tance between the nodes, and d
ref
is a reference distance 5
Alternatively,one transmitomni-directionally butusing a higherprocessing gain, therebytrading dataratefor longer range. We do not consider this or other methods in this paper.
6
given by dref =2D= 2
, where D is the maximumantenna dimension.
Allsimulationsarewitha40nodeadhocnetwork. Thenodes areplacedrandomlyina2-dimensionalsquareareaofvarying size depending uponthe density parameter. We study the dependanceofthesystemonthefollowingparameterranges.
Densities 4,8, 16, 32, 48, 80, and 112nodes/sq mile. These correspondto average nodedegrees(numberof neighborsinthenetworktopology,withomni-directional neighbordiscovery)ofapproximately3,5,9,16,22,30, and38respectively.
Gains 10, 14, 20, and 26 dBi. These correspond to beamwidthsof 60, 40, 20, and 10degreesrespectively (refertable1). Thebaselinecaseisanomni-directional antenna,representedbyabeamwithgain0intheplots.
Fortheswitchedbeamformingcase,antennaswith4,8, and12beamsareconsidered.
Weconsider two performance metrics: throughput, the per-centage of packets sent by any sourcethat was successfully receivedat theintendeddestination;and delay, theaverage timeelapsed, for all successful packets, betweenthe packet being sentby asourceandit being received. Eachresultis anaverageover3runswithrandomseeds.
For all of the results presented inthe network, 20 streams areused,betweenrandomlychosensource-destinationpairs. Eachstreamconsistsofpacketsofsize1800bitsandthe inter-arrivaltimeisuniformlydistributedaroundameanrateof90 kbpsperstream. Nodeshavebuersof30packets,and pack-etsthat nd the queuefull are dropped. This representsa highbutnotexcessiveloadonthenetwork. Theuseof buer-ingimpliesthatcapacityimprovementsareoftenreectedas reduceddelayratherthanincreasedthroughput.
In all cases, receiving is done omni-directionally. Further, inall steered beam cases, control messages are sent omni-directionally. Animplicationofthisisthatthenetwork topol-ogyuseddoesnotchangeasthegainisincreased. Onesetof resultsreportedherehascontrolpacketsbeingsent direction-allywithswitchedbeamantennasinwhichcasethetopology getsricherwithincreasinggain.
Onlystationary networks are simulated. The main reason for this is that the advantages we wish to evaluate { spa-tialreuse, and longerrange { are largely orthogonal to the mobility ofthe nodes. Inotherwords, sincethe same rout-ing algorithm is used for both omni-directional and direc-tionalexperiments,and theMAClayerismobility impervi-ous,thenumbersfortherelativeperformancebetween direc-tionaland omni-directionalcommunicationswill notchange signicantlyforamobilescenario. Giventhis,itmakessense tousetheavailablesimulationresources(CPUtime)tomodel awiderrangeofparametersettingsandalargervarietyof ca-pabilitiessuchaspowercontrolandswitchedbeams,which, asweshallsee, aecttherelativeperformance more
signi-5.2
Simulation Results
Webegin by consideringtheperformance ofanadhoc net-workwith CSMAchannelaccess, nopowercontrol,steered beams, and omni-directional neighbor discovery. We then progressively changeeachoftheseparameterstodiscernthe eectthateachonehasontheperformance.
5.2.1
CSMA, No Power Control, Steered Beams,
Omni-directional Neighbors
Asillustratedingure3thereisnoimprovementin through-putwhenbeamformingantennas are employed. Thisis be-cause CSMA is even more unsuitable with beamforming { the\hiddenterminal"problemisexacerbatedduetoreduced gainsonall butthemainlobe. Thus,anodeoftenendsup transmittingtoaneighborthatisbusytransmittingor receiv-ing. Thisseemstonegateanygainsduetoslightlyincreased spatialreuse.
Asillustratedingure4thedelaywithbeamforming anten-nasis afair bit lower, especially at high densities. This is because,duetodirectionaltransmissions,nodesbackoless andare able access the channelsooner. Althoughmany of thesepacketscollide,nodesnonethelessgettheirpacketsout quicker. Thedelaydierenceisafactorofabout17incaseof density112,betweenomni-directionalandthe20dBibeam.
0
20
40
60
80
100
10
20
30
40
50
60
70
80
90
100 110 120
Tput%
Density
Tput% vs Density for various Gain; 40 nodes; NumAnt=1
’Gain=0’
’Gain=10’
’Gain=14’
’Gain=20’
’Gain=26’
Figure3: Throughput:Steeredbeams,CSMA,nopower control
Thus, even with one of the simplest MAC protocol and a typicalad hoc networking algorithm, there are some gains inperformance. However, theyare probably notenough to justifydeployingbeamformingantennas,andhencemotivate theenhancementsconsideredbelow.
5.2.2
Adding Aggressive Collision Avoidance
Weconsidertheperformancewithaggressivecollision avoid-ance(refersection4.2),nopowercontrol,steeredbeams,and omni-directionalneighbors. Theresultsareingure5,and6.
0
50
100
150
200
250
300
10
20
30
40
50
60
70
80
90
100 110 120
Delay-ms
Density
Delay-ms vs Density for various Gain; 40 nodes; NumAnt=1
’Gain=0’
’Gain=10’
’Gain=14’
’Gain=20’
’Gain=26’
Figure4: Delay: Steered beams,CSMA,no power con-trol
when the receiver is not busy, thereby drastically increas-ing the number of succesful transmissions. Throughput is increasedby15%(20dBiat112nodes/sqmile).
Thedelayofbothdirectionalandomni-directionalantennas increases,whencomparedwithusingCSMA,duetothe ad-dtionalwaitingfor collisionavoidance. Beamforming anten-nashavethesameorlessdelaythanomni-directional anten-nas, withthe dierence increasingat higherdensities when omni-directional antennasstart suering from agreater de-creaseinspatialreuse.
0
20
40
60
80
100
10
20
30
40
50
60
70
80
90
100 110 120
Tput%
Density
Tput% vs Density for various Gain; 40 nodes; NumAnt=1
’Gain=0’
’Gain=10’
’Gain=14’
’Gain=20’
’Gain=26’
Figure 5: Throughput: Steered beams, AggressiveCA, Nopowercontrol
WefoundthattheConservatveCollisionAvoidancedescribed insection4.2hadconsistentlylessperformancethanthe Ag-gressive Collision Avoidance, even for the omni-directional case. Tounderstandthis,recall thatweusedanitequeue size of30packetsat each node. Packets arrivingat anode ndthequeuefullwithagreaterprobabilitycomparedto Ag-gressiveCollisionAvoidance,andsomorepacketsaredropped than would have collided. We noticed that increasing the buersize doesincreasethethroughputsomewhat,but
pro-0
50
100
150
200
250
300
10
20
30
40
50
60
70
80
90
100 110 120
Delay-ms
Density
Delay-ms vs Density for various Gain; 40 nodes; NumAnt=1
’Gain=0’
’Gain=10’
’Gain=14’
’Gain=20’
’Gain=26’
Figure 6: Delay: Steered beams, Aggressive CA, No
powercontrol
ConservativeCollisionAvoidancefurther.
5.2.3
Adding Link Power Control
Weconsidertheperformancewithaggressivecollision avoid-ance, link power control (refer section 4.3), steered beams, andomni-directional neighbors. Theresultsare ingure7, and8.
Addingpowercontroltothesteered,aggressiveCAcase im-provestheperformance forbothdirectionalaswellas omni-directionalbeams. However,theimprovementisgreaterwhen beamforming is used. The throughput (see gure 7) with beamformingantennasissignicantlyhigherthanwith omni-directionalantennas. Interestingly,thegreatestdierence oc-cursat middledensities { e.g., at density of 48, usinga 26 dBiantennagives28%betterthroughputthantheomni di-rectionalantenna. Thisis aresultofcounteracting forces{ theshorternumberofhops,whosebenecialeectsincrease withincreasingdensity,versusthedetrimentaleectof side-lobes, whichalsoincreaseswithincreasingdensity,resulting inapeakatthemiddledensities. Thedierencebetweenthe variousgainsismuchless-to within6%inmostcases,and about10%inthedensity16case.
The delay (see gure 8) for steered beams with aggressive CAandpowercontrolisdramaticallylowerthanwith omni-directional antennas (also using aggressive CA and power control). When the density is 112 nodes/sq mile, there is a reductionby a factor of about 28 inthe delay when the 26dBiantennaisused. Thedierenceislessatlower densi-ties(aboutafactorof2-5atdensity16). Thedelayofboth omni-directionalaswellasbeamformingantennasishigherat lowerdensities. Thisisduemainlyto theincreasednumber ofhopsatlowerdensities.
0
20
40
60
80
100
10
20
30
40
50
60
70
80
90
100 110 120
Tput%
Density
Tput% vs Density for various Gain; 40 nodes; NumAnt=1
’Gain=0’
’Gain=10’
’Gain=14’
’Gain=20’
’Gain=26’
Figure 7: Throughput: Steered beams, AggressiveCA, PowerControl
0
50
100
150
200
250
300
10
20
30
40
50
60
70
80
90
100 110 120
Delay-ms
Density
Delay-ms vs Density for various Gain; 40 nodes; NumAnt=1
’Gain=0’
’Gain=10’
’Gain=14’
’Gain=20’
’Gain=26’
Figure8: Delay: Steered beams,AggressiveCA,Power Control
inboth omni-directional and directional communications is thesame,causingroughlythesameamountofinterference.
5.2.4
Using Switched instead of Steered Beams
Weconsidertheperformancewithaggressivecollision avoid-ance, power control, switched beams, and omni-directional neighbors. Theresultsareingure9,10,and11.
As seen in gure 9 and gure 10, the switched beamcase behaves very similar to the steered case, both in terms of throughputand delay. Animportant contributor to this is thenumberofbeams used{ 16inthis case {which allows good coverage of the azimuthal plane. However, as seen in gure11, performance at highergain is reducedwhenonly 8beamsareused. Thisisbecause thebeamwidthofthe20 dBiantennaisonly20degrees. Thelargeamountof\gap"in thecoverage(only160degreesoutof360,whichislessthan 50%) implies that some of the neighbors discovered omni-directionallyarenotreachablenow,andnegates anyspatial reusegains. Infact,wenoticedinexperimentsnotpresented
resultsinlessthroughputthantheomni-directionalcase.
0
20
40
60
80
100
10
20
30
40
50
60
70
80
90
100 110 120
Tput%
Density
Tput% vs Density for various Gain; 40 nodes; NumAnt=16
’Gain=0’
’Gain=10’
’Gain=14’
’Gain=20’
Figure9: Throughput: Switchedbeams,AggressiveCA, PowerControl
0
50
100
150
200
250
300
10
20
30
40
50
60
70
80
90
100 110 120
Delay-ms
Density
Delay-ms vs Density for various Gain; 40 nodes; NumAnt=16
’Gain=0’
’Gain=10’
’Gain=14’
’Gain=20’
Figure 10: Delay: Switched beams, Aggressive CA,
PowerControl
Our results here indicate that one may consider switched beamsasalessexpensivealternativetofullyadaptivebeams, atleastintermsof capacity gainsfromspatialreuse. How-ever,aminimumnumberofbeams,dependingonthebeamwidth, iscrucialforgoodperformance.
5.2.5
Using Directional Neighbor Discovery
Weconsidertheperformancewithaggressivecollision avoid-ance,powercontrol, switchedbeams,and directional neigh-bors. Recallthatwithdirectionalneighbordiscovery,Hellos are sent outdirectionally oneach beamand travel further, enablingatopology withlonger-range links. Weconsider a lowerdensityrangefortheseexperiments,asthatisthemore interestingrangeinthiscase. Thereare 12beamspernode. Theresultsareingure13andgure14.
tioned (left), whereasuse of directional neighbor discovery with 10 dBi beams (middle) and 20dBi beams
(right) provides good connectivity and commensurate performance. We note that each link depicted is a
directional link, andhence,unlike theomni-directional case,ahighaveragedegreeisnot bad.
0
20
40
60
80
100
0
5
10
15
20
Tput%
Gain
Tput% vs Gain for various NumAnt; 40 nodes; Density=80
’NumAnt=8’
’NumAnt=16’
Figure 11: Throughput: Switched beams, Aggressive
CA,PowerControl
consider that the beamwidth of the 20 dBi antenna is 20 degrees, and therefore the 12 switched beams donot cover theazimuthalplanecompletely.
Fordensity4nodes/sqmile,thenetworkispartitionedwith omni-directionalantennasbutconnected withbeamforming antennas,asillustratedingure12.
Theperformancedropsatmiddledensitiesbeforerisingagain. This reects a playing-out of the interference versus range forces.
We also note that for higher densities, the throughput is about the same or worse with beamforming. This is due to thefact that directionalneighbor discoverytends to use thelongerlinksbyvirtueoftheshortestpathrouting,which inturn causesmoreinterference(recallthe sidelobes). This bearsoutandextendstodirectionalcommunicationsthe con-clusionin[1]thatall thingsbeingequal,oneshouldusethe smallestpower(shortestlinks)thatprovidesaconnected net-work. Thismotivatesthe useof noveltopology control and routingalgorithms that use shorter links even when longer
0
20
40
60
80
100
0
5
10
15
20
25
30
35
40
45
50
Tput%
Density
Tput% vs Density for various gains; 40 nodes; NumAnt=12
’Gain=0’
’Gain=10’
’Gain=14’
’Gain=20’
’Gain=26’
Figure 13: Throughput: Switched beams, Aggressive
CA,PowerControl,DirectionalNeighbors
6.
CONCLUDING REMARKS
Basedonoursimulationstudyoftheperformanceofadhoc networks with beamformingantennas, we arrive at the fol-lowingconclusions.
1. Beamformingantennashavetremendouspotentialwithin adhoc networks. Fora typical40node stationaryad hocnetwork,andmoderate-to-highload,weobservean improvementofupto118%inthroughputandaupto factorof28reductioninend-to-enddelay.
2. Evenwithsimplechannelaccesstechniques,the perfor-manceimprovementisdrastic. Inthetradeospectrum between parallel transmissions and collisions, leaning towardmorecollisions butmoreparallel transmissions seemstopayo. Thismaybetheappropriateoperating pointforreal-timetraÆc.
3. Linkpowercontrolisessentialinexploitingthebenets ofbeamformingantennastotheirfullest.
0
50
100
150
200
0
5
10
15
20
25
30
35
40
45
50
Delay-ms
Density
Delay-ms vs Density for various gains; 40 nodes; NumAnt=12
’Gain=0’
’Gain=10’
’Gain=14’
’Gain=20’
’Gain=26’
Figure 14: Delay: Switched beams, Aggressive CA,
PowerControl,DirectionalNeighbors
5. Themarginal utility of bothsteered and switched an-tennasdecreases withincreasinggain. Consideringthe costandlargerform-factorofhighergainantennas,this isgoodnews.
6. Directional neighbor discovery works wonders at low densities where it makesall the dierence between a connectedandapartitionednetwork.
Our preliminary results point to a dense mesh of low cost transceivers equipped with switched beam antennas as the bestcandidate for awireless extension of gigabit networks. Needlesstosay,muchworkremainstobedonetorealizethe truepotentialofbeamformingantennas. Ourworkpointsto anumberof exciting areasat theMACand networklayers thatrequireresearch,including:
1. Channelaccessprotocolsforsteeredandswitchedbeams thatallowforatradeobetweenspatialreuseand col-lisions(forrealandnon-real-timetraÆc).
2. Techniquesforexploitingthelargerrangeofdirectional communications for better connectivity and lower la-tency. This in turn would require neighbor discovery usingsteered beamssuchaseÆcient \scanning" using transmit and/or receive beams, or use of higher pro-cessing gain,etc.
3. New analytical and simulation modelsthat donot as-sume that the the range is equal inall directions (no more\unitdisks").
4. Newtechniquesforcharacterizingdirectionallinksand theiruseinsupportingquality-of-service.
5. For slowly steerable antennas, topology control using beamforming in a mannersimilar to topology control usingpower[16].
Fromapracticalviewpoint,thereisaneedfordening
stan-for networking protocols to be able to use smart antennas. Finally,often-usedsimulationtoolssuchasNS-2needto be modiedtoincorporateantennapatternsandbeamsteering.
7.
ACKNOWLEDGEMENTS
WethankCesarSantivanezforhisxestotheOPNETpipeline stageswhichresultedinmoreaccurate simulation. We also thankDr. NitinVaidyaforsuggestingtheexampleingure 3(b),andDr. MarthaSteenstrupforhervaluablecomments.
8.
REFERENCES
[1] P.GuptaandP.R.Kumar,\TheCapacityofWireless Networks",IEEETransactionsonInformationTheory,vol. IT-46,no.2,pp.388-404,March2000.
[2] J.Zander,\SlottedALOHAmultihoppacketradionetworks withdirectionalantennas,"ElectronicLetters,vol.26,no. 25,1990.
[3] Y.B.KoandN.H.Vaidya,\MediumAccessControl ProtocolsUsingDirectionalAntennasinAdHocNetworks," ProceedingsofIEEEInfocom,March2000.
[4] A.Nasipuri,S.Ye,andR.E.Hiromoto,\AMACProtocol forMobileAdHocNetworksUsingDirectionalAntennas," ProceedingsoftheIEEEWirelessCommunicationsand NetworkingConference(WCNC),2000.
[5] N.Pronios,\Performanceconsiderationsforslotted spread-spectrumrandomaccessnetworkswithdirectional
antennas,"inProc.ofIEEEGLOBECOM,November1989. [6] J.WardandR.Compton,\Highthroughputslotted
ALOHApacketradionetworkswithadaptivearrays,"IEEE TransactionsonCommunications,vol.41,pp.460-470,Mar 1993.
[7] Y-B.KoandN.H.Vaidya,\Location-AidedRouting(LAR) inMobileAdHocNetworks",ACM/BaltzerWireless Networks(WINET),Vol.6,No.4,pp.307-322.
[8] A.Nasipuri,J.Mandava,H.Manchala,andR.E.Hiromoto, \On-DemandRoutingUsingDirectionalAntennasinMobile AdHocNetworks,"inProceedingsoftheIEEE
InternationalConferenceonComputerCommunication and Networks(ICCCN2000),October,2000,LasVegas. [9] J.C.LibertiandT.S.Rappaport,\SmartAntennasfor
WirelessCommunications,"Prentice-HallPTR,1999. [10] J.BulterandR.Lowe,\BeamformingMatrixSimplies
DesignofElectronicallyScannedAntennas,"Electronic Design,Apr1961.
[11] ,V.Bhargavan,A.Demeers,S.Shenker,andL.Zhang, \MACAW-AmediaaccessprotocolforwirelessLANs,"in Proc.ofACMSIGCOMM'94,September1994.
[12] \WirelessLANmediumaccesscontrol(MAC)andphysical layer(PHY)specications,"1997.DraftStandardIEEE 802.11,P802.11/D1:TheeditorsofIEEE802.11 [13] J.Monks,V.BharghavanandW.Hwu,\Transmission
PowerControlforMultipleAccessWirelessPacket
Networks."IEEEConferenceonLocalComputerNetworks (LCN),Tampa,FL,November2000
[14] P.Jacquet,P.Muhlethaler,andA.Quayyum,\Optimized linkstateroutingprotocol",IETFMANETWorkingGroup Internet-Draft,WorkinProgress
[15] T.S.Rappaport,WirelessCommunications,Principlesand Practice,Prentice-Hall,1996.
