Abstract With thc cmcrgcncc nf 21 variety nf mobile rlntn scwiccs with varinhlc covcrlige, bandwidth, and htindolf stratcgics, aiid thc need for ini)bilc

tcrnminals 10 tomi ainrmg Ihcsc nctwmks, haridoff in hybrid data nctworks has allractcd trcincndous altcntioii. This article presents an rwcrview of isstics rclntcd ti) Iiandofr' with particular cmplinsi!: on hybrid mobilc data nelworks. Issucs arc lugicnlly clividcd inti] archikc~ural ntid IlanrlaiC dc~ision titnc algorithms. Tlic Iinndoff ai-chitccturcs in high-spccd lrical ctivcragc It $1 X 802.11 wirclcss IANs, a id low-spccd wide arc3 coverngc

CDP13 mid GPRS mobilc data nctrvorks arc dcrcribcd nnrl cnniparcd. A s i i n ~ y of traditional algorithnia and an exnmplc of nii adv:inwd dgarithm usiiig ncural ticmnrks f o r HO rlccisioti timc in ~iotnogciiet~us nclworks arc prcscntcrl. The HO architectural issucs rclntcd to hybrid

nc.rwotkn arc dirciisscd through >in cxnniplc of n hybrid rictwoi-k that eiuploys GPRS and IEEE 802.1'1, Fivc architecturcs for ihc example hybrid network, hascd 011 cruulation of GPllS cniitics within tlic WLAN, mobile IP, a virtual iicccss poiiit, aiid a mobilily gntcway (proxy), are dcscribcd atid comptlrcd. Tile Itinhility grilcway aiid inobilc IP npproachcs are sclcctcd for nioi-e dclnilcd discussion. Thc rlifbrcncca in applying a complcx

nilgorilhin for 110 rlccision timc i n a hoinogcricnw and II hybrid nctwork mc shown lhrouglr an exlimple.

Handoff in Hybrid Mubile Datu Networks KAVEH PAHLAVAN, PRASHANT KRISHNAMURTHY, AHMAD HATAMI ,

WORCESTER POLYTECHNIC INSTITUTE M l K A YLVANTTILA, JUHA-PEKKA MAKELA, ROMAN PICHNA, JAR1 VALLSTROM,

UNIVERSITY OF OULU

andoff (HO) [ 11 is extrenicly important in any mobilc nctwork becausc of-thc'default ccllu-- Iar architccture cmplnycd to maximize spcctrum utilization. Whcii ii mobile tcrrninal moves away froiii n base station, the signal lcvcl degrades ancl there i s a nccrl to switch cnminuni- cations to another haw station. Handoff is the mechanism by rvliicli an ongoing conncction hctwccn o mot>ilc tcrmiiial or host (h4H) ;ind ;i correspondcnt tcrminal o r host (CH) is transfcrred from onc point of acccss to the fixed nctwork to another. I n ccllular voicc tclcphony arid mtmbilc data net- works, such points nf attachinent arc rcfcrrctl to 11s hose SIU-

tians (BSs) and in wirclcss LANs (WLANs), they are callcd uccess points ( A h ) . I n either casc, such a point of attachnicnr serves a covcragc iirta callcd a cell. Handoff, in tlic case nf cellular tclcpliuiiy, involvcs tlic transfer of a voicc call from onc BS to another. I t ] thc casc of WLANs, it involves trans- fcrring the conncction from one AP to another. Tti Iiyhrid nct- works il will invoivc thc trarisfcr of ii coiineclion from oiic US to onoher, from an AP to aiinthcr, bctwween a BS and an AP, or vice versa.

For a voicc user, HO rcsults in an audihlc click inlerrupt- ing the conversation for each HO 111; and hccausc oC €IO, data users may losc packets and uunccessary ct~ngcstion con- trol incasurcs may come into play [2]. Dcgradation of the sig- rial Icvcl, howcver, is a randoin process, and simple decision mcchiiihms such as thosc based on signal strength rncasurc- mciils result in thc plrag-pong eflecr. Thc ping-pong cffcct refcrs to scvcrwl 1-10s that occur bwck and forth bctwccn two BSs. This ti1kes.n scvcrc toll on both thc uscr's quality perccp- tion i u ~ d the nctwork load. One way of climinatirig thc ping- pong ellect is to pcrsist with a BS for a s long as possiblc. However, if HO is dcl;iycrl, weak signal reception persists

unnecessarily, rcsirlting in lawcr voicc qualily, incrcnsing the probability of call drops and/or dcgradation of quality of ser- vice (QoS). Cunscqueiitly, tmmorc complex algorithms arc need- ed to dccide on tlie opt imal time for HO. Handoff also itivolvcs a sequence of cvciits in thc hackbone network, iticluiiiiig rerouting ttic conncction and rcrcgistcring with thc iicw RP, which are additional loads on nctwurk traffic. Hand- off has n n impact on traffic matching aiitl traffic dcnsity for individual RSs (sincc tlic load on tlic air interface i s t rans- fcrrcrl from otic US to another). In thc case of random iicccss techriiqucs cinployed tn ~ C C C S S the air interfacc, or in code- divisinn rriultiplc access (CDMA), moving from uiic cc l l to ancmtlicr impacts QoS in both cells since throughput and iritcr- fcrcnce depend on tlic number of tcrmiiials competing for thc nvailable bandwidth. In hybrid data nctworks, R decision on 110 has an impact on the throughput of the system.

Whilc signilicant work has been dotic on HO mechanisms in circuit-switched mobilc nctworks [ l , 31, thcrc is not much litcratiirc availahlc on pnckct-switchcd mobile networks. In this study we are mainly interested in nrwrcal-time applica- tions in wirelcss nctworks. Perfori~iancc incmires snch as call blocking and call dropping are applicablc only to real-time traffic and may not bc suitable for tlic buruty trarfic that cxists in client-servcr applications. Whcii ii voice call i s in progress, allowed latency is very limited, rcsource allocation has to be guarantcctl, and, while occasionfilly some packcts may bc drnppcd ;ind moderate error ratcs nre permissiblo, rctrfmsmis- sions are tint possiblc, and connectivity has to be maintained continuously. On tlie other hand, bnrsty data traffic by dcfini- Lion necds only intermittclit cuniiectivity, and can tolerate greater latcncies and employ rctransmissiori of lost packets. In such nctwurks l I0 is warrantcd only when the tcrniiiial moves out of coverage of thc currcnt point of attnclimcnt or thc traf- fic load is so high that ail 110 may rcsult in grcater through- put and utilization,

Wireless data sciviccs w e becoming increasingly popular, but are not ubiquitous. Consequcntly, the natural trcnd has

hccn toward 11 t i l iz i rig R m al I-cnvc ragc high-banclwitltli data nctwurks such :IS IEEE 802.1 I whcnevcr thcy iirc avail- ahlc atid switching tci an r)vcilay scr- vice such its thc Gct~cral Packct R d i o Scrvicc ( G P R S ) iictuwrk with Iriw I,niiclwidlh w h e n lhc coverage nC B wirclcss loc;il aIca nctwcirk (WLAN) is not availahlc (Fig. I ) . Wc rcfcr to such a pruccdurc as iritcrteclr mnruirig or i-IO if1 h y h d rrerworks. Whcn wc consider I LOS bctwccii hybrid packet- switched networks, a n I 1 0 rrom ii Wl A N Al’ to a C1’KS I1S, for cxmi- ple, aliiltrld tic done 0 1 1 1 ~ w i t h vci-y

. . . . . . .. . .. ... .-

iow priority, wliilc an HO f rom i CiPRS RS to R W[,AN AP should hc c l n m wticncvcr it is Iiossitilc in light of thc ordcrs of inagiiitiidc of diffcr- ciicc iii available Ixwlwidtli hctwccn the two systems.

We prcxcrit an ovcrvicw of the issues relatcd to HO. Tssucr arc logically classified intn architcctural issucs and HO dcci- sinri t imc nlgnrittiins. A tlricf survey of traditinnal HO algo- rithms lxisccl on rcccivcd signal stt‘cngth is provirlcd. Thc HO arcliitcctiircs in low-s~”xl wiilc-nicii-covcragc Ccllular Digital 1’ackc.t I h t ; ~ (CDPII) and OPKS mobilc da t i i iictwurks i i 1 d high-spccd luc;il-covcr;igc 1EEL 802.1 1 WLANs :ire ilcscrilrrcd and compiircrl. An cxamplc of adv;inccrl ntgorittiin using iicural nctworks for H O dccision titnc in hoinogcncous iict- works is prcsciitcd. Thc HO architcctural issucs rcl;itcd t o Iiybrid iictwrrrks arc cliscusscrl through an cxainplc of ii hyl>rid nctwork that employs GPKS ;ind [EEL XU2.I 1. Five architcc- t u w for thc cxamplc hyhrid nctwork, Iiascd on cmulatioii o f GI’RS ctititics within the WLAN, inribilc IF, ii virtu;il IICCCSS

point, :ind ;I inohility gatcwiy (proxy), arc clcscribccl mnd coiii- piirccl. Ttic incibility gatcrvay aiid mobilc IP npproachcs arc sclcctctl for more t letailed discussioti. Ttic dirfcrcriccs in applying il complcx algoritlini for HO dccisioii tinie i r i x IiornogctIcnus arid n Iiylirid network arc shown through an

I Figure ~*Y~7r . i f1nr

cx;lniplc.

Issues in Hnridoff Thcrc iirc ii varicty c i f issiics rclntctl to HO. As shown in Fig. 2, thcsc issiics ;ire diviclcd into two catcgorics: ;irchitcctural issucs ;uid 110 dccision limc algorithms. Architccliird issuch iirc thosc rcliitctl to the mcthodology, wiitrol, i i i i i l softwarcl 1ixrtIw;irc clcmcnts iiivolvcrl in rcrriuting thc conncctiriii. Isstics rcliiIcd 10 the tlccision timc d g o r i h n s il1-c Ihc t y p s of nlgnrithins, mclrics iiscrl hy t h c ;ilgorithms, ; ~ n d pcrformxncc cvaluation rncthodolngics.

Arch i tect III‘CI I Iss U es I IantIoff pruccdurcs iiivolvc ii sct ut’ protocols to notify all thc rclatcd erililics or n piir1iciil;ir coiincclion t h a t ail 1IO has l xcn cxccutcd and that the conticction tms tn tic rcdcfincd. In data networks, thc MH is usiially registered with a particular point o f attaciirncnt. I n voice networks, an idle MH would liiivc sc1cctc.d ii piirticular US t h t is scrving thc ccll in which it is locatcd. l’his is fo r thc p p o w o f routiiig iiicoming data packcts c~i- voicc c;ills ;ipprupri;mAy. Whcii the MIL iiiovcs ;111d executes sa FIO rrnm mic gninl o f ilLtilCllmclIl lii a i i r ~ t l i c i ~ , tlic nld scrvitig point of attachment Iias 10 be ioformctl about tlic clxingc. This is usually callcrl di,wocio!ioti. Thc MH will alsn havc tci rc:issoci;itc itsclf with thc ticw point of access to tlic fixed iiclwork. Ollicr nework cntitics invo1vc.d in routing h t a packcts tu thc MH 01- switching vtiicc c~ills tiwc to bc awarc

of the HO in ordcr to scaiiilcssly continue the ongoing cnn- iiectiori or cal l . Ucliciitling on whcthcr a new connection is created hcrorc brcnking lhc old one or not, HOs arc classificd into hard aiitl scnmlcss [IOs. L11 CDMA, thc cxistcncc of two sitnultatieous cotincclions during L10 rcsults i n soft HO [4]. The dccision riicchtlriisni 01’ hrtnrloff cuntrol inay hc Iric;itcd in a nctwrirk ciitity (as i n cellular voice) or in thc MII itscll (as in mohilc data arid WLANs). These cmes lire c;illcd ~rciwurlc- con trolled h nnrlojf (NCHO) a t i d nrnbile- cot1 truilcii hciridofl (MCHC)), respectively. hi GL’KS, inform:ition scnt by the MI1 c i i i i he employed by the iiclwork ciitity in making ttic hiindoff rlccisinn. This is callcd mobile-nssktcd kwidojf (MRI-IO). In any ciisc, thc cntity that tlccidcs on Lho 110 11sos somc nict- rics, algorithnis, ancl performnticc nicxsurcs in iniiking thc decision. Thcsc arc dincrissed below.

Decisiori Time AIgorifJinis Scvcral nlgorithiiis iirc being crnployctl or investigated to make the corrcct rlccisioi~ to hai ic i o f f [ I , 31. Ttaditiotial algorithms ernploy tlircsholrls to comparc the valiics of ~riet- rica t rom rliffcrcrit points of atlachment iiiid thcn dccitlc on whcii t o inakc t hc HO. A variety nC metr ics have I)cen

HO. Prirriarily, the rcceivetl sigiinl strciigtli (RSS) nic;ist~rc- nicnts from tlic sctving point of altaiclimcnt ;iud iicighboring points of attactirnciit arc uscd i i i tnosl O C lhesc networks. Rltcrnativcly o r i n corijunctioii, tlic p a t h loss, cnrrier-to- intcrfcrcncc ratio (CIR), sigiial-~o-iriterfcrcilcc ratio (SIK), bit crror rntc (BER), hlnck error rate (BLEK), symbol crror ratc (SCK), power budgcts, and cell ranking havc h e n cinpluycrl ;is nictrics in certain rnohilc vnice and data net- works. I n orrlcr to woitl thc ping-pong cffect, adclitional p,arainctcrs arc employcd hy ~ I \ c i ~ l g ~ r i t h i ~ ~ ~ siich :is hystci-e- SIS mirgin, dwell titncrs, and averaging windows. Adtlilioritll paranictcrs (wlicn nvailahlc) may IK cmploycrl t o niakc tiinrc inlclligciit dccisions. Suma uf thcsc parnmcters also includc the distance betwceii ihe M11 and tlic puiiit o f :ittachmcnt, llie vclocity oC thc MII, and traffic c1i;irxtcristics in thc sew- ing cell. The pcrl‘ormaiice O C 110 algorithms is rlctcrniincd by tlicir cffcct on ccrtaiti pcrformaticc measures. Most of the pcrfrirmancc Incitsurec that have been considered, such as call blocking probability, HO hlocking prot)atiility, dclay bctwccii I IO rctlucst and cxcciitioii, and call dropping prub- ability, arc rclatcd to voicc cr)niicctioi~s. Handoff ratc (num- her of HOa per m i l of tinic) is r e l i l l c t l tu thc ping-pong cffcct, and algorithms arc irsually tlesignctl to minimize the iiiimbcr c i f ~iiincccssary HOs. While iniiiiinizitig the HO ratc is iniportant i n moliilc data nctworks, other isslles includc

CtilployctI i1i inobilc voicc iiliil diitii networks to dccidc 011 XI

.-

Handoff I I I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . , .. . . . . . . . . . . . . : i,.jo.e~i:~id~:time~lgorithm:~jj j ;: :

. . . . . . . . . 1 . . . . . . . . . . . . . . . . . . . . . . . .

I

I I

7. Network- controlled

2 . Mobile- assisted

3. Mobile- control led

Hard handoff seamless

handoff Soft handoff

/

throughput ~naximization and maintaining 0.0s guarantccs rluring and a f t e r HO. However , thcsc issues havc not receivccl sufficient nttcntiun in the litcrnturc.

l'ratlitional HO algorithtns arc all based on thc received signal strength (RSS) OT reccivcd power P. Sonic of the tradi- tional algorithms [I] arc as follows:

RSS: Thc US whosc signal is being rcccivcd with thc l q e s t strcngth is selectcd (choose 3s B,,, if I:,,,,, > f o l d ) .

HSS p f w ThreFhokk An HO is made il the RSS of a new RS exceeds that of the currcnt one atid thc signal strength of thc current BS is bclow a tlircshold T (chnosc U,,,," if P3,c,c

is greater ihan that of thc old BS hy a hystcrcsis margin If (ciinosc RSS, Hysrer.e.k, and Thushold: An HO is niadc if thc KSS or ii new RS excccds that of the currwt RSliy a Iiystcrcsis mar- gin H and tlic signal strength of thc current BS is below a threshold 2' (choose R,,, if P,,, > pol, + 11 and Pnld i r). dlgwiir/im plus Dwdl Tirmr: Sometirncs ii dwell titncr is usctl with tlic above algorithins. A timer is startcd lhe instant thc condition in the algorithm is truc. If thc condition cuntin- tics to be truc until tlie timer cxpircs, an HO is pcrformed. Kecently, otlicr tcchniqires arc merging such :IS hypothc-

sis testing [ 5 ] , dynamic programming [ 6 ] , and pnttcrii rccogni- tion tcchniqucs based o n ncural networks or fuzzy logic systems [7] (for an excellent survey of various algorithms, sec [I, 71). Thcse complicatcil algorithms arc necessitated by thc complcxity of the HO problem, espccially in hylirid data or voice networks. Thc molrile tcrininal has to monitor the air for wireless data scrvices that may bc availahlc f u r altacli- iiicnt. As NI example, consider an MtI that could coiiiiect to cither an 802.11 WLAN AP cmiccted tn a LAN ur a FFRS BSS coiinected to il backbone GPRS nctwork. Thcrc milst bc i i mechanism or algorithm wittiin tlic MI1 that will cn;iblc it to choose the licst availablc servicc iinrl switch to this scrvice as snnn iis it is availablc. For wimple, tlic MI1 milst he ahlc to

if P,, > l'ol,~ + H),

RS5. path loss, Cell ranking CIR, SIR BER, BLER, SER Power budget

J-

. . . . . . . . . .

Hysteresis margin. dwell timers. averaging windows, distance, velocity,

Call blocking Handoff blocking Handoff rate

switch from ilia GI'RS scrvice Lo rhe WLAN AP ;Is s m r i as i t detects the availability of ii connection to i i i i AP. Most o f thc emcrgcnt nlgotitlitns arc iu thcir tinsmiit sttigcs, arid h:ivc been atialyzcd or simulated only foi voicc networks :ind only in extretncly siinplc scenarios.

Hn n doff Arch i fect uws on d Algorithms irr

Jfomogeneous Mobile Daia Networks IIntidoR in v:irioiis technologics is iliffercrit. To illuslrate tbc similarities ailti rliffercnccs, wc consider GPRS, CDPn and TEEE 802.2 1 as exainplca bclow. Arcliitccturcs arc opcn m3 staiidardized Init vcry oftcn Ihe HO dccisiun time algorithms arc proprietary. Wc consider an cx;iniplc or a ncural nclwork (NN) algorithm in tlic last suhscctioii which indicates thc clilfer- cnce in performance hetwccii tratliliond and advanced algo- r i thms with lhe HO ratc and clclay as pcrforinancc nicasurcs.

I fnrrduf in Gcrternl k k u i Rnclio Servirc GPRS [ & I n ] is i i i i cnhancemctit of the Clohal Systcin for Mobile Coniinunica~ions (GSiM). It iiser exactly thc s a n e physical radio clxinnels as GSM, and only lieu' logical CI'RS radio channcls arc defincd. Allocation nf thcsc chanticls is flexihlc; from one to cight radio intcrfacc timc slotr; can bc allocatctl per time-divisiun mulliplc acccss (TUMA) frame. Time s lo ts are sharcd by active I I S C ~ S , ; i n c l tlie uplink ;inti downlink arc allocntcd scparately. Pliysic;il channels arc takcn from the coinmoii pool of availahlc channels in tlic ccli. AIIO- cation to circuit-switched serviccs and GPRS Is donc dynami- cdly according tu it c~puci ty oti durnurirl principle. This ineans that the capacity allocation for ( iPKS is hnscd on thc acliinl iicctl Cor packct transfers. CifRS docs no1 reqiiirc pcrin;inent- ly d o c a t e d physiciil cliantiels. Logical network ~i(idcs called GIRT sirppurt iiodcs (GSNs) tire used for packcl routing in

r Corporate

tlic hncklxlnc. ’L’ l iepfawoy GSN (GGSN) iicls as thc interrace to puhlic ( la la tietwcirks stich as thc Iiiternct aurl contains thc routing infotniation to he uscd to iunncl packets to thc M[I through R servirig GSN (SGSN). ‘I‘he SCiSN is responsiblc Cor loctition rriaiiagcmcnt atd dclivcry nf packcts.

‘the GGSN and SGSN can he coiisidcred thc mobile IF’ equivalents o f thc homc iigcnt (HA) aut1 looreign agont (FA). L’ackets originaling from ;in MH arc routed by an SQSN to its destitiation as in any p;ickel-switctwtl network. Packets intcnd- ed for ai1 MI1 reach rhc GGSN associaled with its /rorr?e rict- work. Tlic GGSN dctcrmine5 wliich SGSN is scrviug thc M t l , cticapsulntes thc packet, and forwards (tunucls) it to tlic SGSN. Information related to thc MH is stored i t ) a GI’RS register (GR), which is par t or thc Iioriie locotioti wgister (1I1,R) nf GSM.

A GPKS MH can be in OW of three statcs: idle (uiircacli- ablc), tcntly (wiicrc it is registcrcd with an SGSN), and stand- tiy (iiiiicllve fnr ii long time). I n order to CoImmtiiiicatc, ijn MI I docs a GPRS attach aiirl cnlers a rcatly starc. Thc MH is rcspotisihlc for cell rcsclcction indepcnclcntly, and this is donc in tlie samc way as iti GSM. ‘Ilia MH ineasiires thc RSS nf thc current broiirlcrrs! control cki.rnnel (UCCH), coinp;ires it to the RSS of thc BCCH of adjacent cclls, arid dccidcs to which cell lo attach. ‘Ihcre is, howwcr, an option nvailablc to operr- tors to inakc the BSS iisk for rcports from t h c CvlH (as in GSM), and [hen thc HO is donc as in GSM (MAIIO). Plain GPRS-spccific itiforinitlion can be sciit in ii piiukel R C C l l (PHCCH), hut tlic KSS iq always measurcd from the I3CCII.

Thcre are also othcr principlcs which inay bc coiisidcrcd in HO rlccision (patti loss, cell tanking, ctc.).

Thc location is upiliilcd with ;I routing upclatc prnccdnrc, ils shown in Fig. 3. A rnuiing arcn ( R A ) cor tcspor i r ls to ;I gruup nf cells. Whcn an M H clmges i t s RA, it scrids an RA update reqiimt containing thc cell idcntity ant1 tlic idcnlity o t l h e prcviotis HA to tbc IICW SGSN { 1) . Nolc that iin itltrit- SGSN routing area 11pdiitc is also possiblc wlicti the samc SGSN scrvcs lhe ncw RA. ‘ lhe new SGSN asks ihc ulil SGSN to provida thc routing conicxt ( ( iCiSN xltlress and turincliug inform;itioii) of the MH (2). The ncw SGSN tlicn updatcs thc GGSN uf the hornc iictwork with the iicw SGSN iirldiws niitl new tuuneling informntinti (3). l h c iicw SGSN i i lso updalcs thc 11I.R. Tlic I11 .K caticcls thc MH infurmatinn contcxt in thc old SGSN aud loads thc subscriber data to thc IICW SGSN. Tlic ncw SGSN acknowlctlges the Mll. ‘lhe previous SGSN i s rcqucstetl lo transmit uridelivcrcd diita to tlic iicw SGSN.

Handoffjii M i r l n r D.$d Pocket D a h The CDPU nctwork [ I 1-13] operatcs as ii coiiticctionlrss ncl - work that is ;I wireless cxtciision tn cxisting wired ct~iinectioii- ICSS rietwrirks. 11 shares thc cxisting infrxlructurc ;is well :is spectriim uC the Ailvanccrl Mohilc Phoric Scrvice (AMPS) ilnalog cellular tcleplionc ncrwork in tlic United Stales. Con- sequcntly, it employs difkrent physic;il chariticls ot freqticncy batids for uplink and downlink transmissions. Haiirloff occiirs when a n MtT m o w s from otic cc l l to nnothcr, [Ire CDI’D channcl quality dctcrioratcs, thc ciirrciit CDI’T, ch;iniicl is

MSF: Mobile serving functiori MHF' Mobile homc ftrnction MD-IS, Mobile data interindiate system MNLP: Mobile Network Location Protocol MNRP: Mobile Network Registration Protocol

MSF

Old serwiiin MD is .-- ..p,

./--

.._---- ~- -. - - ~ -... -...

RSSl falls: start scanrjirlg ~

MSF: Mobile scming function MI IF : Mobile home furictiun MD-IS: Mobile d d l a intermediate syrtem MNRP: Mobile Nctwnik Rcgistration Protocol Scan RSS1 falls l ist and below tunc thicsholtl to thr: hcrt i

Beacon i s on the same physical channel as data

Decision on handoff is made at the . -- I ;nMe;rrninaI

IAPP protocot to infotm old AP about

.- I-- The multiple access is CSMA, and the channel is manitorcd all the time before , packet transmission

Bcacon (Beet I) is on a separate physical channel from data traffic

Decision on handoff is made at the basc station iantrollcr or rnnbile host

.--.--I..-...-_.

. - - .- . -.

updates GGSN in GPRS

The multiple access is TDMA-bassd. and channel monitoring is done durirlq times when the MH does not transmit or receive

Beacon (reference) is usually on a separate physical channel from data traffic

Derision on handoff i s made at the mobile twininill

FlustVtedirect packet message from MHF to MSF of the old RS

The multiple accc55 is DSMNCD. and the chatincl is rnoqitnred all tho time bcfore ~

packet tianmission I

_I 4

.~ -. -..-

- - - .

. .. .. . . . . . . .. .. . . . . . . .. . .

Input r-

Mobile movcs from DS1 l a BS2 along a straight

line

1

2

3

1G

- F Output

works h a w N U J 0 iuul h t a iiclworks prcfcr MCHO. I n hoth cilccs, channcl mniiitnritig is always pcrfi)riiicd :it h e terminal. Whcii the 110 control is with lhc network, l l i c tnohilc has to tl';insniit l l ~ c iiicasiirccl iiiformation ti) thc ticcision cntity.

HO dccisicin timc algorithms arc iisiidly proprichry nncl scc- iinri{r-rlcpciitlcrii, Thcy ~ c , howcvcr, iiitlcpcntictrt nf the HO arclritcclurc. 111 CDPU a n d Cil'RS, the algoriltims arc similiir to the AMPS a n d CiSM nlgoritlims simply bccnusc it is ticsir- ahlc for ;I data MH til make nri HO at around the same 10c;i- lioii at which ii vuicc terminal would. This keeps lhc lrarfic loails balnnccrl in each cell, As mcntioiied earlicr, thcrc arc scvcral cmcrging ;~lgnritlims that cniploy complicatctl tech- niqucs to iirrivc nt the corrcct time Cor HO. L'iiltern recogni- tion 130 algoriltiinx [ I ] train a syslcni ming av;iilablc trictrics (c.g,, RSS) ;iiitl llic locations wlierc HOs shuuld IK riiarlc so tha t thc systcm acquires knowledge nP thc KSS patterns ;it such Iocatiiiiis. Neural nctwtrrks (NNs) can hc iiscd for such pattcrii classificLilioii 118, 191. 'lhc basic it lc;i of nri NN is to design a systcm thwl tnkcs a fcw inputs that appcai' to be ran- tlom but Iiwc somc partcrn associatcil W i t h thcm and, regard- less of rlic ii;itiirc of thc inputs, a d j u s t s rhc paramctcrs nf the systcni in rrrdcr to gct soinc clcsirctl outputs thruiigli n learri- ing prciccss. tly adjusting thcsc syslem parainctcrs the proccss of 1c;irniiig i s complctcd, mid llic systcrn ciiii bc considered iis

:I black hnx [lint is c;ip;iblc As an cxamplc, we cnnsidcr a scciiarin of four idciitical RSs

in a microccllulnr ~~iviroiiriiciit and an M H which is ninvii!g frniti ttic ncighborliood of BSI toward RS2 ;ilong a direct path, a s sliriwii in Fig. 7. It j umcd that all thc USs can prtiviilc thc samc scrvkc t o the M11. The nt)jcctivc is to dcvclop ; i n

N N systcm which takcs i i numbcr of powcr samples frnrn all the RSs and, iisiiig R pattcrn recogtiition tcclinique, selects tlic BS which is mos t suit;thlc, while minimizing thc HO dclay xntl ping-pung eflcct. Tlic distancc nf cacli bluck is aswmcd tu bc R = 250 m. For tlic input of ltic h" wc iisc [he RSS from cacli US. '1'hc output is it coiitrnl sigtial that is ?.ern as long i i s

the MH is closcr trr HS I iltld 011c whcncvcr thc MI-[ is closer to US*. Thc RSS lcvcl dcpcniln 011 tlic chaiinel hctwccn thc M H and c a d i US. t-or (tic purposc uC this simulatiuii wc 11nvc uscd a iiiicrocclliilar piitti loss nioclcl witti lognornial fading. h r l l i c r rlctails about this modcl can be fnuiirl it1 [7, 201. Soinc prcpro- ccnxing i)f tlic inlnil vcctnrs also improvcs ttic pcrforniancc of l l ic ncural Iictwoi.ks. Figiitc X slio\vs i i sarriplc of tlic t iSS frnrn lour RSs BS ii Cunctiori nf locatioii oC llic M H :ilong Ihc straight linc cnii~iccting US] arid RSz. Thc MIT tiiovcs frrim ii rlistnlicc of 12.24 111 f rom 13S1 to 487.70 m Croni BSI. 'I'his dixlancc is split into roiiglily 100 cqu;illy spiiccd points, ;und liSS sanrplcs :ire ohhiiictl cvcry 4.76 111.

A chi-ce-layer hack~propagiition NN has been dcsigncd ;iiitl is shown in Pig. 7. l ' l i c inputs t o this 5yslcIn arc sanipIc< oC RSS f r o m c;icIi oC tlic four USs. Thcsc samples arc takci i using a sliding wiiiilnw o f 4 sainplcslwinrlow. 'l'lic NN is traincrl wi l l i lhe previously inciilioiicci RSS and dcsircd,out- piits. Once thc ncural nctwork "hlnck hox" is dcsigncrl, wc t'cctl iis inpiits B freshly gciicixtcrl input vector of RSS fruni t h c four RSs as ttic inobilc mnvcs frum l l S l lo BSl. Wc iisc thc came inpills lo tcst two trarliliounl algilrithnis lhat com- paw ntily tlic KSS aiid usc tlic RSS will1 a liystcrcsis margin. As i)ur pcrCornisticc tiicasurcs, wc use the numhcr of HOs and tlic tlci;iy i n executing thc 110. Figure 9 dcinniistrntcs sonic oC t h c rcsults of our simulations. In Fig. 9a tiic irienl corilrol signd ttiat switclics thc mohilc Trtm BSI to US? ilt thc inidpoint lictwccii thc two I,;isc statiolis is sliowii. If tlic sini-

prnducing t hc desired output.

~ 20

10

0

- 1 0

"20 E % -30

-40

-50

60

Microcell model for channel LOS and NON-LOS

o I O 29 30 40 50 60 70 80 90 loo Points ( R = 250 in)

ple RSS critcrion is used (the m t h i l c is swilcticrl to thc US piwitling tlic largest LISS), sevcral tiniicccss;iry 110s occiir. lu Fig. 9b, R samplc ciisc is sliowri wlicrc scvcii 110s L R ~ C placc, six of llicni unncccssary. I[ n liystcrcsis margin or 10 d B is cmploycd tn rcducc this ping-pong cffcct, only oiic HO t;ikcs place. Hnwcvcr, thc rlcliiy in 170 cxccution is large. hi Fig. 9c, lhe el'fcct of iiiltling ;I Iystercnis is shown. 'L'hc [IO takcs place at around 767 m from BSI that is ; h u t 17 m away from the corrcct lociitiun wlici'c thc HO should lii ive l i cc i i cxcculed. Pigiiic 911 sliows tlic pcrforinance O F an N N with sonic prcpro- ccssed input. Tlicrc is exactly ni ic H O cxccuretl at approxi- mate ly 255 111 f rom US'l. Ttic pcnalty f o r lhis is iti the iricrcascd a1gi)rithin coinplcxity and ti-;iiniiig o f thc NN lliat has to hc donc bcrorcllaricl.

Hnn doff A rch itectu re in Hybrid Ne f works

(Intertech or Verticnl Roaming) Thc niolivatinii fnr liytirirl neiworks arises f r o m tlic fact that no rmc tcclinoligy or scwicc can providc iihiqiiilous covcragc, and it will be tieccssary for ;L mobile tcrniiwl to miploy vari- oiis points o f attacliincnt til m;iiiilnin conticctivity to Lhc nct- wurk a l till iiiiics. Ttw ii:itiirnI t w i r l Iias bccii tuwird utilizing Iuciil-cnvcragc h ig l i -h~~n t~wj i l t l i d a t a iictwtrrks sucli as IEl.III' 802.1'1 whcncvcr ;ivuiIablc nrrd to switch to kiii ovcrlny scrvicc such LIS R GPRS nctwtrrk with IOW i x " i d t h wlicii tlic cover- age of a WLAN is iiot uvailahle. For cxiiniplc, i l i c Bay Arca ksci1I.ch \Virclcss hcccss Network (UAKWAN) 121 1 is implc- incnting romiiing bclwccti WavcLAN wiixlcss LANs iintl Mct- riconi IIiiCkCt t lntn scrviccs. 'l'hcy riiffci,cntiatc bc lwccn Iiorizorifd ;md iw6iuol tlOs [22]. I louiztmtnl HOs rcl'ccr til HO bclwecn RSs using thc satiic kind of network iutcrfacc. Vcrti- cal LIOc refer to HOs bctwccri BSs cmpIoying tlilfcrcnt wirc- lens tcchnologics. An ntpwwd vo.fical i f 0 ( i c c w froin ii 13s of a scrvict? with ii smallcr-size c c l l t o il RS of a scivicc with widcr covcragc. A ciowtrwnrd vwiicul /ift/tdq/T takes placc i n thc I c~"c rsc direction. Whilc t l ic iipwarrl vertical [IO o c c m whcri t l i ~ MH moves r ~ i l or covcr;igc r)f ii sci-vicc, thc tlowt~- w:irrl vertical HO has to h k c plxcc whcn covcrage of a sei-vicc willi sniallcr covcr;igc I)ccrrrncs av:iil;iblc when the IISCI' still has cmiicction to the scrvicc with w i t h co\~ctngc.

In tlic fulluwii~g scctions WC discuss cxninplc arcl\itccturcs for a hybrid ric~work coiisisiiiig of CiPRS i i i i i l IEEE 802.1 1 that wcrc consiilercd i i i 123, 241. Five architcctiires for the

2

1.5

1

E 0.5 I

m bl

I-

% 0 -a

T $ -0.5

-1

-1.5

................. ,. ............ ,. .... , , , , , , , , , , . . . . . . . . . . . . . . . . . , . . . , , , , , , , . , . ........................ .............. *..._ ., ...... ,

-. Hahdofi usini onli RSSi j j j j ................................................................ , , , , . , , , , , I I , . . I I I I

Meal handoff control signal

, , , , . , , , , , , , , . , , , , ................................................................. . . . . . . . .

5 4

3

2 - m .- s 1

t o 71 c -1

3: -2

-3

4

.. o i o 20 30 40 5~ 60 70 80 90 loo

(a)

Paints (R = 250 m)

Handoff control signal using absolute power with 10 dB hystersis

2

. . . . . . . . . . . . . , . , , , , . , .

-- . - - .L .... .: ..... :. .... 1. .... .:. .... ,:. ... .: ..... 1. .... .:. .... - I I i l I I i I i

. . . . . . . . . . . . . . . . . . . . . . . . . . .

1.5

Delay 4 - m & 0.5 .- n

g o U E -0.5 -

-1.5

-7

A

-1 . . . . . . . . . , . ' . . , . . , ,

-. ... .; .... .;.. . ..:.. .... :. ..... L.. .. ,:. .... :. .... .:... . . .;. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I t l I I

. . . . . . . . . , , , , . , , . . . , , , . , , . . ' -1.5 ...... ' ...... ' ...... 4 .......

1 . 1 i I i 1 1 I I

. . . . . , , . , . , , . , , , . , . . , . . , , , , . , , . , . . -7 c 0 10 20 30 40 50 60 70 80 90 100

Points (R = 250 rn) k)

Handoff control signal using NN 2

1.5

1 - g 0.5 ._ m 0 5

5 4 . 5 7l

I -1

0 10 20 30 40 50 60 70 80 90 100 Points (R = 250 m)

(d 1

examplc hybrid nctwork, hascd 011 cinulation of (iPRS cnti- ties within the WI.AN, tnohile TP, a virtual AP, and a mohility gatcway (prnxy) arc dcscrihcd aiid compared. Thc mobility gatcway and ~iiobilc 1P ;ipproachcs arc sclcctcd f o r morc d c t d c d discussion.

Figure 10 stlows fivc rliffcrcnt archirccturcs 123, 241 f w iinplc- nicntiug 1-10 bctwccn QPKS i i d ILEL 802.11 nctworks. l'hc objjcctivc 11crc is to rcducc, ;IS far a s pussiblc, niajur chaugcs io existing nctworks and technologies, cspccially at the lower laycrs such as MAC and physical laycrs. This will ctisure that cxisting networks will crmtiiiuc tn function as hcf'nnre withnut rcquiring currciit iiscrs to chmgc to thc ncw approach. Thc iinplcinciitatiori invulvcs incorporating ncw cntitics or proto- cols thal opcratc at the nctwork or highcr laycrs to cniiblc intcrtech roaming that will be transparent to the mobile user to the extcnt possible.

The first two ;irchitccturcs invrhvc coii~iccting thc WLAN to thc GPRS nctwork through GPRS cntitics such as the SGSN and CiCiSN. I n these cascs, the WLAN will appear tn bc a GPKS ccll or RA, rcspcctivdy. GPRS will bc thc rrinster iiriwork wnd thc WLAN will hc thc . s h e nefwork. This tiicaiis th;it mobility will bc handlcd by GPRS, consirlcring thc

WLAN onc of its cclls or RAs. This a a y require dual-inodc PCMCIA cards to acccss two different physical laycrs. Tn addition, all traffic will first rcach the GPRS SGSN nr GGSN bcforc rcaching its f ina l dcstiriiition cvcn if thc final dcstina- tiori is in tlic WLANILAN itsclf. 'llik Will putciitiiilly causc bottlenecks in thc GPKS nclwork. ' l hc virtual AP (3) rcvcrscs thc rolcs pltiycd by thc GPKS anti W I A N in thc first two arcliitcctut'cs. Hcrc, the WLAN is a master nctwork and thc GPRS is the slave network. Molility is managed according to thc IEEE X(12.1 I and TAPP spccifications hy thc.WLAN. Thc fourth approach introduccs ti mohili@ grateway (MC;) bctwccn thc GPKS xnd WLAN nctworks. Thc MCi is a proxy implc- mentctl on citlier Ihc GPHS or thc WLAN sidcs, i ind will han- dle the mobility ancl rnutitig issues. The l a s l nrchitccturc cmpIoys mobiIc IP tn handlc the issuc of mobility managc- mcnt. Hcrc, GPRS and WLAN arc pccr nctworks, Ccrtain changcs will bc nccdcd t o support intcrtcdi roaming on hoth thc tcrniinal and network sidcs. Wc consillcr cinly thc last twu approaches since the first lhree are ineflicient xi id rcndct one of thc [WO nctworks as n slavc nctwork. Wc do not ;idrlrcss thc issuc of location managcincnt. It is assumcd that whcn an MH is attachcd to one particular network, the lncatiori manage- ment functionality of that tnetwnrk is used. Either the proxy tracks tlic nctwork to whicli thc MH is ;ittaclicd, or mobilc IP Iocatioii inatiagcinctit fcaturcs iirc uscd to dctcrminc to which nchvork ttic MH is conricctcd at a particular timc.

42 11Ct;l: I'crsonal Corntnunicatioris April 2000

~. . -

I HUMANIST 8/8

I t i Fig. 11, Ihc gerieral architeclure for a proxy- or mohility- ga t eway-bad approach to iiitcrlcch roaming is shown. An intcrinccliatc server is pliiccd in thc Iietwnrk s o that any traffic to a n d from Ihc MH is forccd to pass throuEh it. L'otiseqiicnt- ly, it is possible for this cntity lo p c r h r m ccrtain imctioiiali- tics t h t may he reqiiircd on bclialI nf thc MI1 in ti manncr that is lransparcnt to thc rcst of tlic nctwurk. Wheri ilic M H is attnchcd t o iiii AP (and thus ctiniiccted to ii WLAN), lhc communication path betwcen thc Mtl and R cnrrcspondcii~ host (CH) o n the lnlernet will l ~ c (1,,)-(3). The CH-MI1 com- niunicntion path wil l hc (4)-(2L). When thc MI1 is on tlic GPRS network, these paths will be (lr;)-(3) aiicl (4)-(2c;), rcspcctivcly. 11 should I>c ohscrvctl that scgincnts (3) and (4) i n any of thc pilths do )lot chnrige regiirdlcr;~ cif wherc the M H is localed. Otdy links (I) atid (2) will be c~nlinually c11;inging dcpcnding on thc movcmcnt nf thc MH. Clcarly, thcn, only llic cnminiitiicatiou links betwccn tlic MI 1 and the proxy serv- er (PS) arc subjccl tn changca, w1i;ikver tlicy inay bc. Allow- ing thc proxy-ninl>ilc cotincction lo cliangc while maintaining the proxy-CH cntincctioii uncliatiged suppurrs mohility. Thc chwngesriecdcil arc w i t h respcci t i l thc comiiiunication protocols Iictwccn the MH and the PS. Uolh the M H and PS arc prcsuniwbly undcr common ciwncrship 1251, and licncc it w i l l be quitc casy to tune ttrc rcquirccl char- actcristics t o the specific ticctls u l llie u s u s of that syxtcm.

Thcrc ;ire several advaiitiiges I n cniploying a proxy architecture for interlecli roaniing. l'here is tlic possi- bility nf furthcr minimi/.iiig ttic cnc;ip- sul~i t ion and routing incfficicncies associatctl w i h mobilc IP. Ilowcvcr, lhe real rcduction i n ovcrhead inay nut bc vcry significant due tn tlic nccd for additicmil cuiilrol pmtncols. If thc pmxy is undcr Ihe control of tlic sanie organization that owns the MHs, it is pnssihlc tu configure ttic proxy to sup- port thc pcculiar riecds id its poptiln- tion cif Mlls. An optimizcd protcicol tnay hc r u n bctwceri thc rnohilc and prtixy tlcpcnding o n ttic l i n k being cmploycd. TIw proxy ciiI1 miln;lgl: 1 1 1 ~

limited resotirces of certain con- I ~ C ct io ii s iii o r c e f c i c i e 11 t 1 y. I o r example, wtieii the mtihilc is con- ncctcd to GPRS, tlic prrixy can drop structured da ta such :IS e- mail headers, framcs in iiii MPEG stream, ancl s o on sclectivcly o r drup niistructuretl data iisiiig suiiie licirristics, EIS in the casc of quoted lext in e-mails 1261. Thc proxy cnri compress data or delay d a t a by spooling, resegment packcts, and respond to ICMP Incssiiges oil hchalf of a n MH to i m p r o w the pcrformmce of the wirclcss link. Proxies a re alrcsrdy in place in many organizations i i s fircwalls or Web caching servers. Thcsc may be reirsed for inability managc- inciit and intcrtecli ro;rmin(r. Prox- -

its can bc usctl for logging thc charactcristics o l ii cnnticction, dctails of which may tie irscfully cniployctl in various applicii- tions incl~iditig acconthig and frraurl management.

'I'hc main tlisadvniitagcs of thc proxy architccturc arc as follows. The architccturc is nut stantlarclized, requiring propri- ctary prolncnls for intcrtcch roaming. The perfnrmancc or a pioxy i? poor sincc significant lnlency is added to thc client- s e r ve r corn m U H i cat i 011 p i i t h . Pot en t i all y. t 11 e c n d - t 0- en d seniaiitics of tlic transport protocol may also hc viointctl. 1I ;I siiiglc proxy is employed and it fails, it nxiy res~rll in tlic fail- urc o f tlic cntirc network, and tlicrc is a nccd to have sonic fault tulcrancc. In addition to thc sigiiificaiil isare of dcvclop- iiig prolocols Cor tnohility inanagcment will1 the proxy archi- tccturc, thcre are some murc open issucs. Thc placcmcnt and iiunibcr u l proxies to hc cmploycrl may t lq~cnd on tlic sitita- tion. It is pret-ei-ablc to havc thc proxy connected to tlic last links nf cadi scrvicc the M H inay iisc so that it can gatticr i!irnrttiation abuut tlic qualily nf each last link. flowcver, tlic nwncrship of such w proxy will hc coiitciitiuus. 'lhe numhcr of prrixics thal have to he placcd for optimum perfotmancc is iilso subject tn rictwork criiirlitions, and an easy ~r iswcr is not possible.

-

(1)-(3): Client to server communication (4)-(2): Server to client communication

Corresoondent host

This approach cniployx mubilc LP [27] lo rcs(riiclure coiiiiec- tinns wlicii i i i i MI1 roaiiis Irom one data network to anrithcr. Outsidc of its homc iictwork Ihc mobile hosl is idetitificd hy n car+c-o~uddrc,~s associ:itcd with ik poinl of amdime t i t and a cnllocatcd jbreign ngcrit (IA) (hid ninnagcs dccapsulation arid dclivery of packcts. 'Ilic iiicibile liosl regislers its care-of address with ;I lrorire qe:orr (Ilh). 'l'hc IIA resides in the h a m

network ol the MI1 anti i s rcsponsiblc Cor intcrccpliug &ita- grsuns addrcssed to t l i c MH's Iiomc address as w e l l as cricap- sulariiig and tuniicliiig thcm to thc associatcd care-of arldrcss. Uiitagriims to an M H arc ;ilw;iys roiitcd tl irough thc HA. Dotagrattis froin thc MH arc rclaycd along mi optimal path hy thc Iritcrtict routing systcm, although it is pnssitilc to crnploy rcvcrsc tunneling ttirongh tIic HA. Figurc 12 sliows tlic gcncr- al architccturc of tlic i i irrhilc-IP-hascd approach. Aa sliiwn, therc it is assuincil that tlic WLAN is thc liorric nctwrjrk (with

+ \ Working in W A N

HO algorithm

r Prepare 8

i

.L

Working in GPRS I-i Monitor

Nhr or FLS

44 1EGD PcIsr~inl Coininuriicntions 9 April 2000

Mobile On GPRS I Application 1

. ............... .............................. ~ .~ .....................................

1 SNDCP

802.11 PHY

.. ,

SGSN Base station

BSSGP

On the LAN vlpl... Access point . . . . . . . . . . . . . . . . . . . . . . . . . . . ...

E l 802 L l C

Home or fore rg n

GGSN agent ...

. . . . . . . . . . . . . . . . . . . . . . . . . . . 802 LLC

.- I- .. = Figure 13. The ~""710~01 sicrck orweinfed with mobile-ll'-hased arcckitt'ctiitc.

the homc agciit rcsiding on the homc LAN), atid tha CiPlZS network is thc foreign or visitcd iictwork. Figurc 13 shows [he user planc protocol stack associatcd with lhis implcmciitation. Clearly, both the GPRS ['I] and WLAN networks are peer nctwtrrks. 'I'he protocol stack, however, docs iiot show the functionality of thc HA ant1 FAs, which cxist ai the IT' laycr in cach neLwnrk.

The gcncrd ;ircliilecturc dcscribcs thc functions uf ele- tiicnts o f tlic network whcn wc have an H n from WLAN to GPRS or the other way a round . Firsl, we dcscribc the step-hy- step procedure for tlic MLI handing off fruiii the WLAN coli- neclion to tlic GPKS connection. Thc following stcps occw while tlic mobile moves away from Ihe coveragc rif t l~c WLAN wilhin the GPRS cowage. The signal rcccived from thc AP in the WLAN is initially strong. Thc signal hom thc AP hccclmes weaker as tlic. MI1 inoves away. Tlic HO algorithm in thc MI1 decides to dissociate from tlic WLAN and associntc with GPRS. Tlic PA in Lhc MH is activated, and thc MH uses [he ncw visiting 1P addrcss. Thc 1IA in the WLAN is inlnrmccl itbout the new IP addrcss. In lhe reverse situation, when the MI1 is cntinectcd to GPKS and rcalizcs t1i;it ;I WI,AN i s avail- ahlc, thc following steps will uccw. Thc sign;il from any WLAN is initially not clctcctcd. ' lhc MH tlicn dctccts a signal from thc AP of a WLAN. ?'hc I10 algorithm dacides 011 tnak- iiig nn I10 Irom GPRS to the WT,AN. Tlic 1JA in the M H is deactivated, and tlic home 11' addrcss is uscd. 'I'he HA i n thc WLAN i s instructcd by the M H to n u lotiger d n a pruxy address rzsolutinn mi its bchalC. Hcrc we are assutninr movcmciit onlv hetwccn the home -.... .............

expcricncc scvcrc ticgradation of scrvicc and will have to HD w r y fast to maintain the higher-laycr conncction. For tlicsc reasons thc triggcring algorittim is dcctmiposcd into two parts, undcrhy (WLAN) and overlay (GPRS). To enable reuse of thc cotlc, both parts use thc samc NN; only the frequency of its invocation niid thc action on its output are diffcrciit. The larger block i t ) Fig. 13 shows the block diagram of the algo- rithm. Simulating the results for a complcx systcm of several BSs and APs is complicatcd, and a sirnplc system ns describctl bcluw provides preliminaiy insights into what might hc appro- priiitc performnncc mcilsiircs for initiating HCI.

Figurc 14 siwws a simple "moving away sccnario" wherc thc MI-I moves away from a WLAN AP. One can infer in this scenario (bascd on 1281) that an efficient algorithm will try to use thc scrviccs of the AP as long as possiblc and do the HO ta the BS as thc last alternative. This is uiilikc thc microcellu- lar sccnario in thc earlier section, whcrc thc best possible timc t o HO wiis when the MH mas midw;iy bctween Iwo identical BSs. In this case the AP has riiuch higher priority than tlic BS. l 'he reason for t h i s diffcrence in HO strategy is that it is almosl. always prcfcrablc nut to make mi HO whcii the wide- area scrvicc provides n data ratc two orders or magnitiitle smnllcr than the local-ami scrvice, since tratisrnittiiig at 2 Mb/s for 1 s is preferablo to transmitting at 19.2 kb/s fur 100 s. One possible impicincntation of such a sclicmc will be to employ timc Iiystcrcsiu. Here, the MH will take thc samples of [he RSS froni thc AP and compare it with a predefincd

-- .................... - ... ~ .............

WTAN &I GPKS. Tlic HO problem is tmnfnld: from WI,AN

to GPRS xid from GPRS tu WLAN. l'he dif- fcccnce between thcsc is thal A user opcrating in GPKS does not h w c to worry ahout losing the connection. Thcrcfore, the uscr attached l o the overlay nchvork just occasionally checks for thc availability of the underlay nctwork. Thc WLAN-to-GI'RS HO triggering algo- rithm is morc crucial for rcli;iblc operation of thc system, since an MH moving away from tlic uiidcrlay nelwork covcragc may suddenly W Figurc 14.A moving away scetiario.

IFHE I'crsrinai Ctnnniunicalions April 2000 45

-90 I 42.35 44.85 54.7 62.6 65.1

42.35 44.85 53.7 62.6 65.1

42.35 44.85 53.7 Time (s)/distance (m)

62.6 65.1

~~

thrcsholtl (c.g., x = -80 dRm); if R predcfincd nuinbcr rif COLI- scciitive sainples are beIow the rhrcshr~ld, thc MH initiatcs tlic 1-10; otherwise, it will pcrsist with thc AP. This is also cnlled il dwell timer (timc for which tlic MI-I persists with a point of attachtnetit cvcn if thc signal strcngth is low) ['l]. The region wlierc Ihe RSS first falls helow 1: and is last atiovc is cdlcd the iransitian region TI?. Aii'HO should t h u s be made only oncc: a1 the ctlge or T,.

The first plot of Fig. 15 shows thc RSS prtifilc withinlj, nnd R threshold of -80 dBm. A titnc liystcrcsis dgorithm docs the HO at the last crossing of thc thrcshrild x (in this cnsc a t t = 62 R). R u t since this timc is random, :in NN is again prekr- ahle. Figurc 16 sliows a suggcstcd architecture for sucli a net- work. This iictwork consists ol three layers. The iriput tn t h c systcin ctiiisisis of simplcs of the RSS Crom the AI' (in a slid- ing window nl rivc samplcs). Thc output of tlic systcin is ii

binary signal, zcro mciliiing thc Ml-I should continue comiiiti- nicating with tlic AP and 0111: iinplying that the MI-I should ni;~kc thc 1-10 :uid comiiiunicale wit11 the BS. The accond plot in Fig. 15 shows ;ill 110 coiitrol signal using orily RSS, and the third plol, an HO control signal that is thc output cif tlic NN architectiirc OF Fig. 16. Clearly tile NN architccturc is prcfcr-

5 0 x 5

I

50 x 1 c

r-- -I +I

able since it eliminates the ping-pong effect. There is a need for further work in this arca for tiinre complicatcd scenarios a n d performancc tneasurcs.

conczusions This articlc prcsciits nti ovcrvicw of thc issucs rclatcd to handoff with particulqr cinpliasis u11 hybrid mobilc data nctworks. Haiidnff issucs can bc c1;issificd iiito two indcpciidcnt pirrts: archi- tcctural issues and I10 dccision timc algorithms. l ' h e former are open and standardized, but ihe latter are usually proprietary. IInndolf architec- tures in mnhile data networks arid high-speed WLANs liavc scvcral siinihritics. Traditional HO algorithms have poor pcrfonnance and arc Iwiiig replaced hy emerging advanced HO algorithms. An examplc neural nctwork algorithm not only rcduccs thc iiumhcr of nnncccssary HOs, hut atso minimizes the H n delay in a micrncellular scenario. No one techtinlogy or service can pro- vide ubiquitous coveragc, and it will be ncccssary for a mobile terminal to ctnplny various points of attachmcnt to ~naintnin crmncctivity to tlic nct- work ;it ;dl timcs. resulting iu thc nccd for I10 in

a hybrid network of mobiie data and GLAN. ?'he I 1 0 archi- tectural issues related to hybrid networks are discussed through an example of a hybrid network that employs GPRS and lEEE 8112.11. Moldc IP is thc most suitable architccturc for HO in a hyhrid network. Again, ncural-nctwurk-bascd algorithiiis perEorm better than traditional algorithms for hartdoff time in hybrid nctworks.

Ai*knowldgmcn fs l'he Rillhors would like Lo thank TEKES, Nokia, Sonera Lid., and Finnish Airforce root supporting this project. We would like to tliatik Dr. Jaakko Talvitic, Dr. Ali Zaliedi and Yaii Xu for participation i t1 this prjcct.

References [ I ] G. P. Pollini. "Trends in Handover Design," / H E Commun. Mag,, Mar. 1996. [ Z ] R. Cacarcs and L. Iftodc. "improving the Performance of Reliable Trans-

port Protocols in Mobile Computing Environments," l P € € JSAC. June 1995, pp, 850-57.

[3] N. D. Tfipathi, J. H. Read, and H. F. VanLandingham, "Handoff in cellu- lar Systems," !€FE Pen. Commun , Dec. 1998.

[4] 5. Tekinay and 8. Jabbari, "Handover Policies and Channel Assignment Strategies in Mobile Cellular Networks," /E€€ Commun. Mag., vol. 29, no, 11. NOV. 1991.

50 x 1 t

1 x 1

46 IGBB Persriiial Cwnnunicatirms April 2OUO

IS] G. Liodakis and P. Straviouiakis, "A Novel Approach in Handovcr Initiation for Microcallular Systems," Proc. WC '94, Stockholm, Sweden, 1994.

[61 R. Rczaiifar, A. M. Makowski, and S . Kumar, "Optimal Control of Hand- dffs in Wireless Networks," Ptoc. /FEE VTC '95, 1995, pp 887-91.

[71 N. Tripathi, "Generic Adaptive Handoff Algorithms using Fuzzy Logic anti Neural Networks," P l i D. Thesis. VA Polytechnic Inst. a n d State Univ., Aug, 1997.

181 J. Vallstrom, "GPRS," WiLU tech, rep., Oct. 1997. 191 C. Bettslcttcr, H:J. Voyel, and 1. Elmspacher, "GSM Phase 2+ General

Packet Radio Servlcc GPKS: Architecturc, Protocols. and Air Interface." IEEE Commun. Surveys. 3rd qtr. 2999.

[IO) J. Cai and D. J. Goodman, "Gcncral Packet Radio Scrvice in GSM," / € € E Commun. Mag.. Oct. 1997.

[ l l l "CDPD System Spccification." Rel. 1.1. COP0 Forum Inc. Jan. 19. 1995, [I21 M.S. Taylor, W. Waung, and M. Banan, lntefnetwork Mobility: The

COPD Approach, Prenticc Hail, 1997. [13] K. C. Budka. H. Jiang, and 5. E. Sommers. "cellular Digital Packet Data

Networks," Bcf/ Labs Tech. I., Summer 1997. [14! The I E E E 802.1 1 Standard: h t t p llgrouper.ieee.org/groups/802/1 11

indcx.htmi t151 K. Pahlavan, A. Zahcdi, and P. Krishnamtirthy, "Wideband local Access:

WLAN and WATM,"!E€€ Commun. Mag., Nav. 1997. 1161 B. P, Crow cta l . , I E E E 802.11 Wireless Local Arca Networks." f€€€

Commun. Mag., Sept. 1997. 11 71 H. Moelard. "Inter Acccrs Point Protocol: Wirclcss Networking Distribu-

t ion System Communication," Stds. track internet draft, Mar. 1998. [181 5. Haykin, "Ncural Networks Expand 5P's Horizons," I € € € Signal Proc.,

Mor. 1996. [191 0. Cox, "Pattern Recognition Handoff Algorithms for Cellular C o m m u ~

njcation Systems," project description submitted t o Stanford Ctr. f o r Telecommun., Sept. 1996.

[2D1 L E . Bcrg, R Bownds. and F. Cotso, "Path Loss and Fading Models far Micro- cc1iIs at 900 MHz.." 4Znd /€E€ WC, Denver, CO, May 1992. pp, 666 ,71

[211 E.A. Drewer et al., "A Network Architecture for tleterogcneous Mobile Computing." IEEE Pers. Commun.. Oct. 1998.

t22l M. Stemm a n d R . Katr, "Vertical Handnffs in Wirciejs Overlay Nct- works." ACM MONET, Summer 1 g98, pp, 335-50.

12.31 P. Krishnamurthy et a l . , "Handoff i n 3G Non-Homogcneous Mobi le Data Networks." Euro. Microwave Week, Netherlands. Oct. 1998

I241 P. Krishnamurthy et al., "Scenarios far inter-Tcch Mobility," WiLU tech. rep.. Jan. 1998.

[251 D. A. Maitz and P. Bhagwat. "MSOCKS: An Architccture for Transport Laycr Mobility," /FEE INFOCOM, 1998.

[L61 6 , Zenel a n d 0. Ouchamp. "General Purpose Proxies: SoIv@d a n d Unsolved Problems," Pmc. l/OT-OS VI, May 1997.

[271 C. E. Pcrkins. Mobile IP: Design Principles and Practices. Addison We+ ley Communications Sories, 1997,

[28] A. Hatami et a i , "Analytical Framework far Handofl in Non-tlomagc. neotis Mobile Data Networks." PIMRC 'W, Osaka, Japan, Sept, 1999.

Uiq my h ics KAVEN PAHMUAN [F] (kaueh@cce.wpi.edu) is a profossor of ECE and director of tlre Center for Wirdess information Network Studies. Worccstcr Polytech- nic Institute, Massachusetts. His arca of research i s broadband wireless local nctworks. His previous research background i s o n niodulation. coding, and adaptive signal proccssing far digital communications, He has contributed to marc than 200 technical papers and presentations in various countrics. Ho is the principal author of Wireless Inforrmtion Networks (Wiley, 1995. He has been a conwltant t o many companies in the Unitcd States, Finland, ancl

Japan. Before joining WPi, he was dircctor of advanced development at Infi- nite Inc., Andover, Massachusetts,. working on dara communications. He started his career a s an assistant profcssor at Northeastern University, Boston. Massachusetts. tie is Editor.in-Ctricf of the international Jowrnal on Wirclcss lnformatiun Networks. He was program chair and organizer of the IEEE Wireless LAAN Workshop, Worcester, in 1991 and 1996 and thc organiz- er and the technical program chairman o f the lEEE International Symposium on Personal, Indoor, and Mobile Radio Communications, Boston, MA, 1992 and 1998. For his contributions t o wireless networks hc was the Westin Hadden Professor o f Electrical anti compu te r Engineering a t WPI dur ing 1993-1996, and became a fcl iow,of Nokia in 1999.

PRA5HANl KKISHNAMURTHY [MI lprashmt@tclc.pitt.edu) is an assistant profcs- sor in the Department of information Science and Telecommunications a t tlre University o f Pittsburgh. Earlier. he was a research assistant at the Cen- ter for Wirciess informat ion Network Stutlics (CWiNS) at Worcester Poiy- tcchnic Institute working o n r x i o propagation modeling fo r geolocation applications and mobii ity management in mobilc data networks, where ho also go t his Ph.D. in 1999. His interests are In thc arcas of radio propaga- tion, wifeless data networks, and wireless nctwork security.

AHMAO HATAMI ( a h a t ~ m i ~ l u c e n t . c o m ) is a sof tware engincor i n Lucent Technologies intcrnctworking systems division. H e jo inct l Lucent in 1999 and has workcd on layer 2 and 3 protocols for muitiservices core switches. His intcmsts are in the areas of communication and data networking. Prior t o j o i n i n g Lucent he w o r k c d w i t h 3Com developing var iahio hit rate sources. Prior to that he was a research assistant a t the Center for Wireless Information Network Studies working o n Mobile iP and Handover in mobile data networks. Hc received his M,Sc. in communication and camputcr net- working froin Worcester Polytechnic institute, and his 0 . 5 ~ . in electrical cirgineering f rom the University of Tehran.

Miw YLIANIU (mika.ylianttiIa~oulu.ii) received his M.Sc. degrcc in electri- cal engineering f rom Oulu University, Finland. i n 1998. Hc is n o w working a5 a project managcr in t h e Centre for Wireicss Communications at Oulu University. and 15 working o n his P h D in the arca of mobil ity managcmcnt and SystQm architecture issues in fourth-generation wireless networks. His professional interests includc 1P protocol evolution, wireicss optimizations, location-based serviccs, and real-time architectures.

JUHA-PEKKA MAKELA (juha-pekka.makcla@ee.oulu fi) received his MSc. degree in electrical englnccring f rom Ouiu University. Finland. in 1997 tie is cur- rently a research scientist in the Centre for Wircicss Communications at Oulu University. working for his Lich. Tech. degree. His interests are intelligent handoff techniques, ad hoc nclworks, and COMA network isstics,

ROMAN P~CHNA (roman.pichna~nokia.rom) received an ing. degree from Elo- vak Technical Univcrsity, Bratislava, Slovakia. in 1987 and a Ph.D. dcgree from the University o f Victoria, Canada in 1996. He worked at the Centre for Wircicss Communications, Oulu Univetsity, Finland until 1998. He is cur- rently w i th Nokia Networks, Radio Access Systems Rosearch in tlelsinki. His professional intcrcsts are in the areas of cellular radio network simulations. medium access control protocols, notwork architecture design, and net- work protocols.

J A R I VALLSTKOM (iari.vallstrom~nokia,com) received an M S c . degree in elec- trical engineering f rom Oulu University. Finland. in 1995. He worked at thc Centre for wireless Communications, Oulu University. unti l 1998. Hc is cur- rently w i t h Nokia Mobile Phones. His professional interests arc in the areas of GSM evolutioti. standardization, and notwork architectures.