A Cross-layer Communcaton Soluton for Multmeda Applcatons n Underwater Acoustc Sensor Networks Daro Pompl Rutgers, The State Unversty of New Jersey Department of Electrcal and Computer Engneerng 94 Brett Rd., Pscataway, NJ 08854 pompl@ece.rutgers.edu Ian F. Akyldz Georga Insttute of Technology School of Electrcal and Computer Engneerng Broadband Wreless Networkng Laboratory 75 5th St., Atlanta, GA 30308 an@ece.gatech.edu Abstract Underwater multmeda acoustc sensor networks wll enable new underwater applcatons such as multmeda coastal and tactcal survellance, undersea exploratons, pcture and vdeo acquston and classfcaton, and dsaster preventon. Because of the dfferent requrements of these applcatons, t s needed to provde effcent dfferentated-servce support to delay-senstve and delaytolerant data traffc as well as to loss-senstve and losstolerant traffc. The objectve of ths paper s twofold: 1) explore the nteractons of dfferent underwater communcaton functonaltes such as modulaton, forward error correcton, medum access control and routng, and 2) develop a dstrbuted cross-layer soluton ntegratng specalzed communcaton functonaltes that cooperate to allow multple devces to effcently and farly share the bandwdthlmted hgh-delay underwater acoustc medum. 1 Introducton Underwater Acoustc Sensor Networks (UW-ASNs) [1 consst of sensors deployed to perform collaboratve montorng tasks over a body of water. UW-ASNs enable applcatons for oceanographc data collecton, polluton montorng, offshore exploraton, and asssted navgaton. Wreless acoustc communcatons are the typcal physcal layer technology n underwater networks due to rado frequency and optcal waves propagaton lmtaton. In addton to the ablty to retreve multmeda data, underwater multmeda sensor networks wll also be able to store, process, correlate, and fuse data orgnated from heterogeneous sources. Thus, underwater multmeda sensor networks wll not only enhance exstng sensor network applcatons but they wll also enable new applcatons such as multmeda coastal and tactcal survellance, undersea exploratons, pcture and vdeo acquston and classfcaton, and dsaster preventon. Many of the above applcatons, however, requre the underwater sensor network paradgm to be re-thought n vew of the need for mechansms to delver multmeda content wth a certan level of Qualty of Servce (QoS). There are several characterstcs of UW-ASNs that make QoS delvery of multmeda content a challengng task such as frequency-dependent transmsson loss, colored nose, multpath, Doppler frequency spread, hgh and varable propagaton delay, sensor battery and resource constrants, varable channel capacty, and cross-layer couplng of functonaltes [1[9. Whle most of research on underwater communcaton protocol desgn so far has followed the tradtonal layered approach, whch was orgnally developed for wred networks, mproved performance n wreless networks can be obtaned wth a cross-layer desgn, especally n a harsh envronment such as the underwater. Gven our research experence n ths area, n ths paper we clam that UW-ASNs requre for a cross-layer communcaton soluton to allow for an effcent use of the scarce resources such as bandwdth and battery energy. However, although we advocate ntegratng hghly specalzed communcaton functonaltes to mprove network performance and to avod duplcaton of functons by means of crosslayer desgn, t s mportant to consder the ease of desgn by followng a modular desgn approach. Ths wll also allow mprovng and upgradng partcular functonaltes wthout the need to re-desgn the entre communcaton system. For these reasons, n ths paper we rely on the abovementoned desgn gudelnes and propose a cross-layer communcaton soluton for UW-ASN multmeda applcatons that s bult upon our prevous work on underwater routng [4 and Medum Access Control (MAC) [5. In partcular, the objectve of ths paper s twofold: 1) explore the nteracton of dfferent underwater communcaton 1-4244-2575-4/08/$20.00 c 2008 IEEE
functonaltes such as modulaton, Forward Error Correcton (FEC), MAC and routng, and 2) develop a dstrbuted cross-layer soluton ntegratng hghly specalzed communcaton functonaltes that cooperate to allow multple devces to effcently and farly share the bandwdth-lmted hgh-delay underwater acoustc medum. To the best of our knowledge, ths work s the frst to coherently propose a cross-layer framework to optmze communcaton wthn the UW-ASN paradgm. The remander of the paper s organzed as follows. In Sect. 2, we descrbe our desgn phlosophy for cross-layerng and we ntroduce our communcaton soluton. In Sect. 3, we analyze the performance results. Fnally, n Sect. 4, we draw the man conclusons and outlne future research drectons. 2 Cross-layer Communcaton Soluton Our objectve s to develop a resource allocaton framework that accurately models every aspect of the layered network archtecture, resultng n theoretcal and practcal mpacts beyond the prevously establshed results. Our prevous experence n modelng functonaltes of the communcaton stack of underwater networks led us to develop a hghly specalzed coherent communcaton soluton that can adapt to dfferent applcaton requrements and seek optmalty n several dfferent stuatons. Our cross-layer soluton reles on a dstrbuted optmzaton problem to jontly control the routng, MAC, and physcal functonaltes n order to acheve effcent communcatons n the underwater envronment. In partcular, the proposed soluton combnes a 3D geographcal routng algorthm (routng functonalty), a novel hybrd dstrbuted CDMA/ALOHA-based scheme to access the bandwdthlmted hgh-delay shared medum (MAC functonalty), and an optmzed soluton for the jont selecton of modulaton, FEC, and transmt power (physcal functonaltes). The proposed soluton s talored for the characterstcs of the underwater acoustc physcal channel, e.g., t takes explctly nto account the very hgh propagaton delay, whch may vary n horzontal and vertcal lnks due to multpath, the dfferent components of the transmsson loss, the mparment of the channel, the scarce and range-dependent bandwdth, the hgh bt error rate, and the lmted battery capacty. These characterstcs lead to very low utlzaton effcences of the underwater acoustc channel and hgh energy consumptons when common MAC and routng protocols are adopted n ths envronment, as analyzed n [4[5. 2.1 Physcal Layer Functonaltes The underwater transmsson loss descrbes how the acoustc ntensty decreases as an acoustc pressure wave propagates outwards from a sound source. The transmsson loss TL(d, f 0 ) [db that a narrow-band acoustc sgnal centered at frequency f 0 [khz experences along a dstance d [m can be descrbed by the Urck propagaton model [10, TL(d, f 0 ) = χ 10 log 10 (d) + α(f 0 ) d. (1) In (1), the frst term accounts for geometrc spreadng 1, whch refers to the spreadng of sound energy caused by the expanson of the wavefronts. It ncreases wth the propagaton dstance and s ndependent of frequency. The second term accounts for medum absorpton, where α(f 0 ) [db/m represents an absorpton coeffcent that descrbes the dependency of the transmsson loss on the frequency. Interestngly, the transmsson loss ncreases not only wth the transmsson dstance, but also wth the sgnal frequency. As a result, gven a maxmum tolerated transmsson loss TL max [db, whch depends on the transmtter output power and the recever senstvty, a maxmum central frequency exsts for each range. In addton, because of the colored structure of the underwater ambent nose power spectrum densty (p.s.d.), N(f) [db re µpa /Hz 2, the useful acoustc bandwdth B [khz 3 dramatcally depends on the transmsson dstance and on the central frequency. Hence, the desgn of the routng and MAC functonaltes of our cross-layer soluton (Sect. 2.3) takes ths characterstc of the underwater channel nto account, whch can be stated as follows: a greater nformaton throughput may be acheved f messages are relayed over multple short hops nstead of beng transmtted drectly over one long hop. Moreover, the unque V structure of the underwater acoustc nose p.s.d. (whch has a mnmum of 20 db re µpa /Hz at about 40 khz), makes non-trval the choce of the optmal bandwdth, Interestngly, when the central frequency s low, e.g., f 0 = 10 khz, a hgher relatve Sgnal-to-Nose-Rato (SN R) s acheved wth a narrow bandwdth (B = 3 as opposed to 9 khz); conversely, when the central frequency s hgh, e.g., f 0 = 100 khz, a hgher relatve SN R s acheved wth a wde bandwdth (B = 90 as opposed to 30 khz). Ths mples that f a hgh central frequency s used, a large bandwdth can be exploted for the communcaton, although a hgh transmt power would be needed to compensate for the hgher transmsson loss. Our communcaton soluton takes nto account ths unque characterstc, whch s caused by the pecular V structure of the nose p.s.d. and by the fact that the dfference between the slopes of N(f) and TL(d, f) decreases as the 1 There are two knds of geometrc spreadng: sphercal (omndrectonal pont source, spreadng coeffcent χ = 2), and cylndrcal (horzontal radaton only, spreadng coeffcent χ = 1). 2 A reference pressure of 1µPa s used to express acoustc source levels n db re µpa. Hence, 0.1, 1, and 10 W correspond to 161, 171, and 181 db re µpa, respectvely. 3 We assume the band to be symmetrcal around the central frequency,.e., the band occupancy of bandwdth B at central frequency f 0 s [f 0 B/2, f 0 + B/2, whch wll be denoted as < f 0, B >.
Optmal Central Frequency (f 0 ) and Bandwdth (B) [khz 90 80 70 60 50 40 30 Optmal Central Frequency and Bandwdth vs. Dstance (@ [db) Optmal f 0 [khz Optmal B [khz @ SNR =5 db th Optmal B [khz @ =10 db Optmal B [khz @ SNR =15 db th Optmal B [khz @ =20 db Optmal B [khz @ =25 db Optmal B [khz @ =30 db Capacty [kbps 2000 1800 1600 1400 1200 1000 800 600 400 Capacty vs. Dstance SNR =5 db th SNR =10 db th =15 db =30 db SNR =25 db th SNR =30 db th p.s.d [db re µpa /Hz 280 260 240 220 200 180 160 140 120 Power Spectral Densty (p.s.d.) of Nf*TL and Sf vs. Frequency (@ =20 db) Nf*TL p.s.d, dst=100 m Sf p.s.d, dst=100 m Nf*TL p.s.d, dst=1000 m Sf p.s.d, dst=1000 m Nf*TL p.s.d, dst=10000 m Sf p.s.d, dst=10000 m 20 200 100 10 Dstance [m 0 Dstance [m 80 0 10 20 30 40 50 60 70 80 Frequency [khz (a) (b) (c) Fgure 1. (a): Optmal central frequency f 0 [khz and chosen bandwdth B [khz vs. dstance d [m, gven a fxed pre-specfed target [5, 30 db; (b): Chosen capacty C [kbps vs. dstance d [m, gven a fxed pre-specfed target [5, 30 db; (c): P.s.d. of N(f) TL(d, f) and S(d, f)[db re µpa /Hz vs. frequency f [khz at d = 10 2, 10 3, 10 4 m when the pre-specfed target s heurstcally set to 20 db. central frequency ncreases (e.g., postve for low frequences and negatve for hgh ones). In [9, the author assesses the bandwdth dependency on the dstance usng an nformaton-theoretc approach that takes nto account the underwater propagaton loss and ambent nose. The author defnes the bandwdth correspondng to optmal sgnal energy allocaton as the one that maxmzes the channel lnk capacty. However, n order to fnd the optmal sgnal power dstrbuton across the chosen band an a pror knowledge of the optmal SN R threshold ( ) at the recever s requred, whch s often a nonrealstc assumpton n practcal systems. In [9, a heurstc pre-specfed value of 20 db s proposed for ths threshold. If we denote by S(d, f) [db re µpa /Hz the p.s.d. of the transmtted sgnal chosen for dstance d,.e., the power dstrbuton across the chosen band < f 0 (d), B(d) >, then the total transmtted power s P(d) = <f 0(d),B(d)> S(d, f)df, and the sgnal to nose rato s, SNR(d, B(d)) = <f 0(d),B(d)> S(d, f) TL(d, f) 1 df <f 0(d),B(d)> N(f)df. (2) By assumng that the nose s Gaussan and that the channel s tme-nvarant for some nterval tme, the channel transfer functon appears frequency-nonselectve n a narrow subband f centered around frequency f n whch the nose can be approxmated as whte (wth p.s.d. N(f )). Under these assumptons, the capacty C [bps s gven by, C(d) = [ f log 2 1 + S(d, f ) TL(d, f ) 1. (3) N(f ) Maxmzng the capacty wth respect to S(d, f), subject to the constrant that the transmtted power be fnte, yelds to the optmal p.s.d S(d, f) = K(d) N(f) TL(d, f), f < f 0 (d), B(d) >, accordng to the water-fllng prncple [6, where K(d) [db re µpa /Hz s a dstance-dependent constant. The SN R correspondng to ths optmal power dstrbuton s thus gven by, SNR(d, B(d)) = K(d) f 0(d),B(d) TL(d, f) 1 df f 0(d),B(d) N(f)df 1. (4) Fgure 1(a) depcts the chosen central frequency f 0 and bandwdth B, whle Fg. 1(b) shows the assocated theoretcal capacty C, when a fxed pre-specfed target rangng n [5, 30 db s chosen. Note that, whle the optmal central frequency f 0 (lower curve n Fg. 1(a)) s ndependent on the target, both the chosen bandwdth B and maxmum theoretcal capacty C computed depend on t, whch makes ther values suboptmal. Therefore, the words optmal and chosen. Fgure 1(c) depcts the p.s.d. of N(f) TL(d, f) and S(d, f), as well as the sgnal band occupancy, when the pre-specfed target s heurstcally set to 20 db. Ths shows that, f the recever SNR s not consdered n the lnk optmzaton at the sender sde, suboptmal decsons are taken. In fact, the lnk (and thus the overall system) performance strongly depends on the selected, as t s emphaszed n Fgs. 1(a-c) and 2(a). 4 For these reasons, our cross-layer soluton jontly controls 4 The dscontnuty of the capacty as well as of the transmtted power n Fgs. 1(b) and 2(a) at 300 and 5000 m, respectvely, s caused by the mnmum frequency f 0 B/2 reachng zero, whch n turns sets the maxmum frequency to f 0 + B/2, gven the constrant on the band symmetry around the central frequency.
physcal transmsson, modulaton, and FEC functonaltes, n such a way as to optmze the overall system performance by mnmzng (or maxmzng) meanngful objectve functons such as the energy per successfully receved bt (or the net bt rate), accordng to the applcaton requrements. In the next sectons, we present the man communcaton functonaltes of our cross-layer soluton. Wthout the need for a target pre-specfed, our algorthm jontly selects, n a dstrbuted manner, the optmal p.s.d. of the transmtted sgnal,.e., K, f 0, and B, and the best combnaton of modulaton technque and FEC scheme as well as MAC and routng, wth the objectve of ether savng energy, thus prolongng the lfetme of the network (Objectve 1), or maxmzng the system performance, thus ncreasng the network end-to-end throughput (Objectve 2). Specfcally, the actual objectve (1 or 2) depends on the specfc applcaton requrements that need to be supported, and s ether decded offlne durng the deployment phase, or onlne through control sgnalng from the surface staton. In order to acheve the selected objectve, our cross-layer communcaton soluton nterfaces wth the modulaton functonalty by choosng the optmal transmtted power and number of bts per symbol, thus tradng power effcency for spectral effcency 5. Moreover, our soluton nterfaces wth the FEC functonalty, tradng overhead,.e., redundant bts, for ncrease of packet protecton,.e., bt error correctng capablty (Sect. 2.2). Last but not least, t jontly decdes the best next hop (routng functonalty) and the modaltes to access the channel and send the data to the chosen next hop (MAC functonalty) (Sect. 2.3). 2.2 Modulaton and FEC Interactons We consdered several classes of modulaton schemes such as PSK, FSK, and QAM (both n ther coherent and non-coherent versons), whose Bt Error Rate (BER) vs. SNR performance s reported n Fg. 2(b). Note that, whle normally BER plots are referred to the receved bt SNR E b /N 0, we defne the p.s.d. of an equvalent whte nose as N 0 = (1/B) <f 0,B> N(f)df, and the receved bt energy as E b = (1/C) <f S(d, f) TL(d, 0,B> f) 1 df. Hence, the equvalent bt SNR s E b /N 0 = (B/C) SNR. As far as the FEC functonalty s concerned, we consdered block codes because of ther energy effcency and lower complexty compared to convolutonal codes [7[8. In fact, the lmted energy consumpton requrements of UW-ASNs and the low complexty n the sensor hardware necesstate energy effcent error control and prevent hgh complexty codes to be mplemented. In [8, the en- 5 By movng to a hgher-order constellaton, t s possble to transmt more bts per symbol usng the same bandwdth (hgher spectral effcency), although at the prce of hgher energy per bt requred for a target Bt Error Rate (BER) (lower power effcency). ergy consumpton profle of convolutonal codes s presented based on µ AMPS archtecture. It s shown that no convolutonal code provdes better energy effcency for BER > 10 5 than uncoded transmsson [8. Smlarly, n [7, convolutonal and BCH (Bose, Ray-Chaudhur, Hocquenghem) 6 codes are compared n terms of energy effcency n a framework to optmze the packet sze n wreless sensor networks. Results of ths work reveal that BCH codes outperform the most energy effcent convolutonal code by almost 15%. Consequently, we do not consder convolutonal codes n our work, although our framework can be extended to support convolutonal codes as well as other codes such as turbo codes or Type I or II Automatc Repeat request (ARQ) schemes. Last but not least, the use of BCH codes leads to smple closed-form expressons to compute the Packet Error Rate (PER) gven the lnk BER, whch can be readly used for the purpose of the optmzaton carred out by the cross-layer soluton proposed n ths paper. A BCH block code s represented by (n, k, t), where n s the block length, k s the payload length, and t s the error correctng capablty n bts. In our experments, we used BCH codes able to correct up to t = 10 bt errors. Fg. 2(c) depcts PER vs. BER for dfferent BCH(n,k,t) codes and for the case of uncoded transmssons (NO FEC), computed as, { n ) BLER(n, k, t) = =t+1 BER (1 BER) n, PER(L P, n, k, t) = 1 [ 1 BLER(n, k, t) L Pk, (5) where BLER represents the BLock Error Rate and L P s the packet length, whch s set to 100 Byte. To qualtatve understand how we capture the cross-layer nteractons between the modulaton and FEC functonaltes to mprove the lnk performance, let us consder the objectve of these functonaltes when they operate n solaton. The FEC functonalty performs the so-called channel codng,.e., ntroduces some controlled bt redundancy, wth the objectve of reducng the PER at the recever gven a certan BER on the lnk. On the other hand, the modulaton functonalty decdes what the best modulaton scheme and ts constellaton should be ether ) to maxmze the lnk raw rate,.e., the rate of transmtted bts (spectrum effcency), or ) to mnmze the lnk BER (power effcency). It s clear that mproved performance can be acheved by jontly selectng the BCH code and modulaton scheme. Hence, our cross-layer desgn s amed at maxmzng the lnk net rate,.e., the rate of successfully receved bts, by jontly decdng: 1) the modulaton scheme and ts constellaton (whch affect the lnk raw rate), 2) the transmt power (whch affects the BER), and 3) the FEC type and strength (whch affect the PER). Whle ths provdes an ntutvely 6 A BCH code s a multlevel, cyclc, error-correctng, varable-length dgtal code used to correct multple random error patterns. ( n
350 Transmt Power vs. Dstance 10 0 Bt Error Rate (BER) vs. Sgnal to Nose Rato (SNR) 10 0 Packet Error Rate (PER) vs. Bt Error Rate (BER) (@ L P =100 Byte) 10 1 10 2 Transmt Power [db re µpa 300 250 200 150 SNR =5 db th =10 db SNR =15 db th =20 db =25 db =30 db BER 10 2 10 3 10 4 10 5 10 6 10 7 10 8 10 9 BPSKco BFSKco BFSKnc DPSKnc 4 FSKco 4 FSKnc 8 FSKco 8 FSKnc 4 QAMco 8 QAMco 16 QAMco 32 QAMco 64 QAMco 128 QAMco 256 QAMco 1024 QAMco PER 10 4 10 6 10 8 10 10 10 12 10 14 NO FEC BCH(63,57,1) BCH(63,52,2) BCH(63,45,3) BCH(63,39,4) BCH(63,33,5) BCH(127,85,6) BCH(127,78,7) BCH(127,71,8) BCH(255,183,9) BCH(255,175,10) 10 16 100 10 10 10 5 10 4 10 3 10 2 10 1 4 6 8 10 12 14 16 18 20 Dstance [m SNR [db BER (a) (b) (c) Fgure 2. (a): Transmt power P [db re µpa vs. dstance d [m, gven a fxed pre-specfed target [5, 30 db; (b): BER vs. SNR for dfferent coherent and non-coherent typcal underwater modulaton technques; (c): Packet error rate (PER) vs. BER for dfferent BCH(n,k,t) codes. explanaton of the cross-layer operaton as far as the physcal layer functonaltes are concerned, n the the next sectons we cast a rgorous mathematcal framework to capture the FEC/modulaton cross-layer nteractons. 2.3 MAC and Routng Interactons The MAC functonalty s based on a novel hybrd medum access scheme that combnes Drect Sequence Code Dvson Multple Access (DS-CDMA) for the data payload and a smple yet effectve ALOHA access for a control header, whch s transmtted back-to-back mmedately before the data packet to help the next hop set ts recever, as explaned n Sect. 2.4. The MAC functonalty ncorporates a closed-loop dstrbuted algorthm that nteracts wth the physcal functonalty, descrbed n Sect. 2.1, to set the optmal transmt power and code length. The objectve s to let sgnals arrve at the recever wth approxmately the same mean power, thus mnmzng the near-far effect 7 [3. DS-CDMA compensates for the effect of multpath, whch may heavly affect underwater acoustc channels (especally n shallow water 8 ), by explotng the tme dversty n the underwater channel. Ths leads to hgh channel reuse and low number of packet retransmssons, whch result n decreased battery consumpton and ncreased network throughput. In such a scheme, however, the major problem encountered s the Multuser Access Interference (MAI), whch s caused by smultaneous transmssons from dfferent users. In fact, the system effcency s lmted by 7 The near-far effect occurs when the sgnal receved by a recever from a sender near the recever s stronger than the sgnal receved from another sender located further. In ths case, the remote sender wll be domnated by the close sender. 8 In oceanc lterature, shallow water refers to depth up to 100 m. Fgure 3. Data and broadcast transmssons. the amount of total nterference and not by the background nose exclusvely [2. Our MAC functonalty, n conjuncton wth other functonaltes such as FEC and modulaton, ams at achevng three objectves,.e., guarantee ) hgh network throughput, ) low channel access delay, and ) low energy consumpton. To do so, t uses locally generated chaotc codes to spread transmtted sgnals on the optmal band,.e., < f 0, B >, whch guarantees a flexble and granular bt rate, secure protecton aganst eavesdroppng, transmtter-recever self-synchronzaton, and good auto- and cross-correlaton propertes. The dstrbuted closed-loop MAC functonalty ams at settng the optmal combnaton of transmt power and code length at the transmtter sde relyng on local per-
odc broadcasts of MAI values from actve nodes, as shown n Fg. 3. Here, sender needs to transmt on the shared medum a data packet to j, and let j receve enough power to correctly decode the sgnal wthout mparng ongong communcatons from h to k and from t to n. Because the system effcency s lmted by the amount of total nterference, t s crucal for to optmze ts transmsson n terms of both transmt power and code length, n order to lmt the near-far problem. These requrements are expressed by the followng set of constrants, NIj (f) TL j (f)df P j mn k K [ ( ˆRk NI k ) TL k. w j Ω(BER j ) (6) In (6), NI j (f) [W/Hz s the nose plus MAI p.s.d. at recever j, whle NI k [W s the nose plus MAI power at nodes k K, wth K beng the set of nodes whose ongong communcatons may be affected by node s transmt power. In addton, w j and w tk k are the bandwdth spreadng factors of the ongong transmssons from to j and from t k to k, respectvely, where t k s the node from whch k s recevng data. The normalzed receved spread sgnal,.e., the sgnal power after despreadng, s ˆR n = R n w tnn Ω(BER tnn), where R k [W s the user sgnal power that recever k s decodng, and Ω() s the MAI threshold, whch depends on the target bt error rate. P j [W represents the power transmtted by to j. Fnally, TL j (f) and TL k are the transmsson losses from to j and from to k K, respectvely,.e., TL j (f) = TL(d j, f) and TL k = TL(d k, f 0k ), as n (1). The left constrant n (6) states that the SINR 1 at recever j needs to be below a certan threshold,.e., the power P j transmtted by needs to be suffcently hgh to allow recever j to successfully decode t, gven ts current nose and MAI p.s.d. NI j (f). The rght set of constrants n (6) states that the SINR 1 at recevers k K must be below a threshold,.e., the power P j transmtted by must not mpar the ongong communcatons toward nodes k K. Consequently, to set ts transmt power P j and spreadng factor w 9 j, node needs to leverage nformaton on the MAI and normalzed recevng spread sgnal of neghborng nodes. Ths nformaton s broadcast perodcally by actve nodes. In partcular, to lmt such broadcasts, a generc node n transmts only sgnfcant values of NI n and ˆR n,.e., out of predefned tolerance ranges. Constrants (6) are ncorporated n the cross-layer lnk optmzaton problem (,j) n (16). In our cross-layer soluton, the level of nterference at potental recevers,.e., ther MAI, s used not only by the MAC functonalty, but also by the routng functonalty to P cross layer 9 We assume the spreadng factor to be proportonal to the chaotc code length,.e., w j = α c j. By usng chaotc codes as opposed to pseudorandom sequences, a much hgher granularty n the the code length can be acheved; code lengths, n fact, do not need to be a power of 2. decde for the best next hop. In fact, whle a routng functonalty mplemented n solaton would fnd the best path from the sender to the destnaton only consderng routnglayer metrcs, our cross-layer routng/mac soluton fnds the best path also consderng the nterference levels at the neghborng nodes (potental next hops). Ths means that a longer path characterzed by a hgher number of hops (or, n general, a path that would be suboptmal accordng to only routng-layer nformaton) may be chosen f the drect path,.e., the one that would guarantee the mnmum number of hops, were composed of nodes characterzed by hgh levels of MAI, whch would requre longer codes and/or hgher transmt power. Also, gven the fact that n underwater acoustc channels the lnk bandwdth ncreases when the range decreases (.e., shorter lnks provde hgher bandwdth, whch n turns leads to hgher data rates, as dscussed n Sect. 2.1), our cross-layer desgn soluton captures ths nteractons between the routng and physcallayer functonaltes by tryng to compose, n a dstrbuted manner, paths usng short lnks to explot ther hgher bandwdth. Ths cross-layer approach leads to a hgher end-toend throughput, as shown n Sect. 3. The proposed routng functonalty reles on a geographcal paradgm, whch s very promsng underwater for ts scalablty feature and lmted requred sgnalng. However, Global Postonng System (GPS) rado recevers, whch may be used n terrestral systems to accurately estmate the geographcal locaton of sensor nodes, do not work properly n the underwater envronment. Stll, underwater devces need to estmate ther current poston, as t s necessary to assocate the sampled data wth the 3D poston of the devce that generates the data to spatally reconstruct the characterstcs of the event. Underwater localzaton can be acheved by leveragng the low speed of sound n water, whch permts accurate tmng of sgnals, and parwse node dstance data can be used to perform 3D localzaton. Accordng to our dstrbuted routng algorthm, a source or relay node wll select j as ts best next hop ff argmn j S P NE(j) j (Objectve 1) = OR (7) argmax j S P NR(j) (Objectve 2), where E (j) [J/bt represents the mnmum energy requred to successfully transmt a payload bt from node to the snk and R (j) [bps represents the maxmum net bt rate that can be acheved from node consderng every outbound lnks n the path towards the snk. In (7), S s the neghbor set of node and P N s the postve advance set, whch s composed of nodes closer to snk N than node,.e., j P N ff d jn < d N. The lnk metrcs E (j) and R (j) n (7) are the objectve functons (8) and (9) of the cross-layer lnk optmzaton problem
P cross layer (,j). These metrcs take nto account the number of packet transmssons ˆN j T assocated wth the optmal lnk (, j ), gven the optmal combnaton of modulaton technque (Mj M) and FEC (Fj F, LF P j ), and transmtted p.s.d. Sj (f) = K j NI j (f) TL j (f), f < f0j, B j >. Moreover, they account for the estmated Hop hop-path length ˆN j from node to the snk when j s selected as next hop. The proposed optmzaton problem s a dstrbuted communcaton soluton for dfferent multmeda traffc classes that optmzes the transmsson consderng every feasble outbound lnk from,.e., (, j), j S P N, by choosng the optmal p.s.d. of the transmtted sgnal as well as band (K, f0, B ), modulaton (M ), FEC (F, L F P ), and code length (c ). The objectve s set dependng on the hgh-level applcaton requrements. We consder two alternate objectves,.e., Objectve 1: mnmze the energy per bt transmtted; and Objectve 2: maxmze the lnk net bt rate, defned as the lnk bt rate R b dscounted by the number of transmssons N T. Whle the frst objectve leads to prolong the lfetme of the network, the second leads to hgh end-to-end throughput. In the followng, we cast the cross-layer lnk optmzaton problem. (,j): Cross-layer Lnk Optmzaton Problem P cross layer Objectve 1 : OR Objectve 2 : Fnd : K j, f 0j, B j, M j, F Subject to : Mnmze E (j) Maxmze R (j) j, LF P j, c j = E b j Πe2e j (8) = R b j Πe2e 1 j (9) (Class-ndependent Constrants/Relatonshps) Rj b = η(m j) B j, Ej b = 2Eelec b + P j c j Rj b L P (10) Π e2e j = L P LH P ˆN T Hop LF j ˆN j (11) P j f SINR j = K 0j,B j TL j (f) 1 df j 1 (12) f 0j,B j NI j (f)df BER j = Φ Mj( ) SINR j (13) PER j = Ψ Fj( L P,L F ) Pj, BER j (14) P mn j ˆN Hop j ( = max d N < d j > N, 1 ) (15) (c j, BER j ) P j mn[pj max, P max (16) where P j = K j B j <f 0j,B j> NI j (f) TL j (f)df (17) P mn j (c j, BER j ) = P max j <f 0j,B j> NI j(f) TL j (f)df α c j Ω(BER j ) = mn k K [ ( ˆRk NI k ) TL k (Class-dependent Constrants/Relatonshps) Class I = { Class III = Class IV = (18) (19) ˆN j T = 1 1 ( ) 1 PER ˆNHop j +N (m) HC j PERmax e2e,(m) Class II = { ˆNT j = (1 PER j ) 1 ˆN j T = 1 1 ( ) 1 PER ˆNHop j +N (m) HC j PER e2e,(m) d j ( q j + δ(γ) σ q j (m) D ˆN Hop j max ) ˆQ j L P R b j ˆN j T = (1 ( PER j) 1 q j + δ(γ) σ q j (m) ) D ˆQ ˆN Hop j L P. R b j d j We envson that underwater multmeda sensor networks wll need to provde support and dfferentated servce to applcatons wth dfferent QoS requrements, rangng from delay-senstve to delay-tolerant, and from losssenstve to loss-tolerant. Hence, we consder the followng four traffc classes: Class I (delay-tolerant, loss-tolerant), Class II (delay-tolerant, loss-senstve), Class III (delaysenstve, loss-tolerant), and Class IV (delay-senstve, losssenstve). Whle for loss-senstve applcatons a packet s locally retransmtted untl t s correctly decoded at the recever, for loss-tolerant applcatons packets are transmtted only once on each lnk, and are protected unequally, dependng on the mportance of the data they are carryng for correct perceptual reconstructon. Notatons used n the class-ndependent constrants/relatonshps: L P = L H P + L F P j + LN P j [bt s the fxed optmal packet sze, soluton of an offlne optmzaton problem presented n [4, where L H P s the header sze of a packet, whle L F P j s the varable FEC redundancy of each packet from to j. Eelec b = Eelec trans = Eelec rec [J/bt s the dstancendependent energy to transt one bt, where Eelec trans s the energy per bt needed by transmtter electroncs (PLLs, VCOs, bas currents) and dgtal processng, and Eelec rec represents the energy per bt utlzed by recever electroncs. E b j [J/bt s the energy to transmt one bt from to j, when the transmtted power and the bt rate are P j [W and R b j [bps, respectvely. The second term P j/r b j s the dstancedependent porton of the energy to transmt a bt. P max [W s the maxmum transmttng power for node. j
BER = Φ M (SINR) represents the bt error rate, gven the SINR and the modulaton scheme M M, whle η(m) s the modulaton spectrum effcency. PER = ψ F (L P, L F P,BER) represents the lnk packet error rate, gven the packet sze L P, the FEC redundancy L F P, and the bt error rate (BER), and t depends on the adopted FEC technque F F. ˆN T j s the number of transmssons of a packet sent by. Hop ˆN j = max ( d N <d j > N,1 ) s the estmated number of hops from node to the surface staton (snk) N when j s selected as next hop, and < d j > N s the projecton of d j onto the lne connectng node wth the snk. Notatons used n the class-dependent constrants/relatonshps: PERmax e2e,(m) s the maxmum allowed end-to-end packet error rate to packet m, whle Nmax Hop s the maxmum expected number of hops. N (m) HC s the hop count, whch reports the number of hops of packet m from the source to the current node. D (m) = D max ( t (m),now t(m) 0 ) [s s the tme-to-lve of packet m arrvng at node, where t (m),now s the arrvng tme of m at, and t (m) 0 s the tme m was generated, whch s tme-stamped n the packet header by ts source, and D max [s s the maxmum end-to-end delay. T j = L P/Rj b + T q j [s accounts for the packet transmsson delay and the propagaton delay assocated wth lnk (, j); we consder a Gaussan dstrbuton for T j,.e., T j N(L P/Rj b + T q j, σq j2 ); ˆQ j [s s the network queueng delay estmated by node when j s selected as next hop, computed accordng to the nformaton carred by ncomng packets and broadcast by neghborng nodes. Note that sender optmally decouples the routng decson, whch s based on (7), from the soluton of P cross layer (,j), whose output s the optmal metrc E(j) (or R (j) ), nput of the routng decson tself. Therefore, sender can optmally decouple the cross-layer algorthm nto two sub-problems: frst, mnmze the lnk metrc E (j) (or maxmze R (j) ) for each of ts feasble next-hop neghbors (Algorthm 1 presents a possble space-search approach) (physcal functonaltes); second, pck as best next hop that node j assocated wth the best lnk metrc (MAC/ Routng functonaltes). Ths means that the generc node does not have to solve a complcated optmzaton problem to fnd ts best route towards a snk. Rather, t only needs to sequentally solve the two aforementoned subproblems wth no loss of optmalty. The frst has a complexty O( S th M F, where S th, M, and F are the number of dfferent thresholds, modulaton technques, and FEC schemes, respectvely, used n combnaton wth Algorthm 2. The second has a complexty O( S P N ),.e., proportonal to the number of the sender s neghborng nodes wth postve advance towards the snk. Moreover, ths operaton does not need to be performed every tme a sensor has to route a packet, but only when the channel or the traffc condtons,.e., the structure of the MAI n the neghborhood, have consstently changed. Algorthm 1 Cross-layer Lnk Optmzaton (gven, j, d j ) 1: E mn = [or R max = 0 {ntalzaton} 2: for th=1 : S th do 3: for mo=1 : M do 4: for fe=1 : F do 5: (SNR, K, f 0, B) Algo 2( = th) 6: PER = ψ fe( L P,L F P(fe),Φ mo (SINR) ) 7: Solve P cross layer, Calculate E (j) (8) [OR R (j) (9) 8: f (E (j) < E mn) [OR R (j) > R max then 9: E mn = E (j) [OR R max = R (j) 10: (fe, mo, K, f 0, B) = (fe, mo, K, f 0, B) 11: end f 12: end for {end FEC cycle} 13: end for {end modulaton cycle} 14: end for {end SNR cycle} Algorthm 2 Lnk Transmsson (gven d and ) 1: f 0 =argmn f [N(f) TL(d, f) {optmal f 0} 2: K (0) = [mn f N(f) TL(d, f 0) {ntalzaton} 3: stop = 0, n = 0 4: whle (stop == 0) do 5: n = ++, Fnd B (n) s.t. K (n 1) N(f) TL(d,f) 6: Calculate SNR (n) from (4) usng K (n 1) and B (n) 7: f (SNR (n) ) then 8: stop = 1 9: else 10: K (n+1) = (1 + ɛ) K (n) {ɛ R + } 11: end f 12: end whle 2.4 Protocol Operaton Once the cross-layer optmzaton problem has been solved at sender, and the optmal communcaton parameters (.e., K, f 0, B, M, F, L F P, and c ) have been found, randomly access the channel transmttng a short header called Extended Header (EH). The EH s sent usng a common chaotc code c EH known by all devces. Sender transmts to ts next hop j the short header EH. The EH contans nformaton about the fnal destnaton,.e., the surface staton, the chosen next hop,.e., j, and the parameters that wll use to generate the chaotc spreadng code of
Energy [J/bt Transmt Energy vs. Dstance 10 20 Cross layer (OBJ.1) BPSKco, NO FEC 10 15 BFSKnc, BCH(63,57,1) 16 QAMco, BCH(63,45,3) 64 QAMco, BCH(63,39,4) 10 10 10 5 10 0 Transmt Power [db re µpa 350 300 250 200 Transmt Power vs. Dstance Cross layer (OBJ.1) BPSKco, NO FEC BFSKnc, BCH(63,57,1) 16 QAMco, BCH(63,45,3) Net Bt Rate [kbps 900 800 700 600 500 400 300 Net Bt Rate vs. Dstance Cross layer (OBJ.1) BPSKco, NO FEC BFSKnc, BCH(63,57,1) 16 QAMco, BCH(63,45,3) 64 QAMco, BCH(63,39,4) 10 5 150 200 10 10 100 10 15 100 0 Dstance [m Dstance [m Dstance [m (a) (b) (c) Fgure 4. (a): Energy per bt E (j) [J/bt (a), transmt power P [db re µpa (b), and net bt rate R (j) [kbps (c) vs. dstance d [m for the proposed cross-layer soluton (Objectves 1 and 2) and for four fxed FEC/modulaton combnatons. length c for the actual data packet that j wll receve from. Immedately after the transmsson of the EH, transmts the data packet on the channel usng the optmal communcaton parameters set by the cross-layer algorthm. Note that the protocol does not have to send control packets before the actual data packet s transmtted. Ths s because the packet - composed of the extra header EH and the actual data packet (payload plus standard header) - uses a hybrd MAC to access the channel,.e., t smultaneously accesses the channel usng ALOHA-lke MAC (the header) and locally adaptng ts power and code length as n standard dstrbuted CDMA MACs (data packet). Ths novel approach s motvated by the huge propagaton delay affectng the underwater envronment (fve orders of magntude larger than n terrestral wreless networks) and by the need to acheve hgh channel utlzaton effcences to compensate for the low-bandwdth shared medum. If no collson occurs durng the recepton of the EH,.e., f s the only node transmttng an EH n the neghborhood of node j, j wll be able to synchronze to the sgnal from, despread the EH usng the common code, and acqure the carred nformaton. Then, f the EH s successfully decoded, recever j wll be able to locally generate the chaotc code that used to send ts data packet, and set ts decoder accordng to the optmal communcaton parameters used by n such a way as to decode the data packet. Once j has correctly receved the packet from, t acknowledges t by sendng an ACK packet to j usng code c A. In case does not receve the ACK before a tmeout T out expres, t wll keep transmttng the packet untl a maxmum transmsson number s reached. If sender does not have updated nformaton about the MAI n j, t ncreases the code length every tme a tmeout expres to mprove the probablty that the packet be decoded. 3 Performance Evaluaton In ths secton we compare the performance acheved by our-cross layer soluton aganst that acheved by ndvdual communcaton functonaltes that do not share nformaton and operate n solaton. Also, we compare results obtaned when the objectve functon of our cross-layer optmzaton problem s ether Objectve 1 or 2. As far as the nteractons between physcal layer functonaltes are concerned, Fg. 4 shows the energy per bt [J/bt (a), transmt power P [db re µpa (b), and net bt rate R (j) [kbps (c) vs. dstance d [m for the proposed cross-layer soluton (when both Objectve 1 and 2 - n the fgures OBJ.1 and OBJ.2 - are consdered) and for those four fxed FEC/ modulaton combnatons that E (j) showed best performance. As can be seen, our soluton outperforms competng fxed schemes when ether objectve s selected, n terms of both energy and throughput. In partcular, n Fgs. 4(a) and 4(b), the curves assocated wth OBJ.1, representng the transmt energy and power, respectvely, for a payload bt to be successfully decoded at the recever, are always above any other curve assocate wth a fxed FEC/ modulaton combnaton such as coherent BPSK/NO FEC, non-coherent BFSK/BCH(63,57,1), coherent 16-QAM/BCH(63,45,3), and coherent 16-QAM/BCH(63,39,4). Moreover, the performance gan of our cross-layer soluton over the best FEC/ modulaton combnaton out of the four consdered ncreases as the dstance ncreases. Note that, out of the many FEC/ modulaton combnatons we tested, for the sake of clarty, n the fgures we reported only those four wth best performance. The same concluson can be drawn lookng at Fg. 4(c), whch reports the net bt rate vs. dstance for
1.8 x 10 4 Average Normalzed Used Energy (LP=100Byte) 1 Normalzed Succesfully Receved Packets (LP=100Byte) 12 Average Packet Delay (LP=100Byte) Normalzed Used Energy [J/bt 1.6 1.4 1.2 1 0.8 0.6 0.4 MAC + Routng Cross layer (OBJ.1) Normalzed Succesfully Receved Packets 0.9 0.8 0.7 0.6 0.5 0.4 0.3 MAC + Routng Cross layer (OBJ.1) Average Packet Delay [s 10 8 6 4 2 Cross layer (OBJ.1) MAC + Routng 0.2 5 10 15 20 25 30 35 40 45 50 55 Number of Sensors 0.2 5 10 15 20 25 30 35 40 45 50 55 Number of Sensors 0 5 10 15 20 25 30 35 40 45 50 55 Number of Sensors (a) (b) (c) Fgure 5. (a): Average normalzed energy vs. no. of sensors (Class II); (b): Normalzed successfully receved packets vs. no of sensors (Class II); (c): Average packet delay vs. no. of sensors (Class III). our soluton as well as for the best four competng fxed schemes. Agan, the curve depctng the performance of our soluton when OBJ.2 s set as objectve functon of the optmzaton problem outperforms any of the other four curves whatever dstance s consdered. As far as the nteractons between MAC and Routng functonaltes are concerned, Fgs. 5(a-c) report the average normalzed used energy, the normalzed successfully receved packets, and the average packet delay, respectvely, vs. number of sensors. We consdered a varable number of sensors (from 10 to 50) randomly deployed n a 3D volume of 500x500x500 m 3. Performance results refer to the three cases of OBJ.1 and OBJ.2 for our cross-layer soluton, and the case where a CDMA-based MAC [5 and geographcally-based routng [4 run ndvdually. In Fgs. 5(a-b) our cross-layer soluton wth OBJ.1 outperforms the MAC+Routng case (Class II), and n Fg. 5(c) ths s agan the case wth OBJ.2 (Class III). These postve results are due to the fact that our soluton jontly optmzes the consdered communcaton functonaltes, thus leveragng synerges that lead to mproved end-to-end system performance. 4 Conclusons and Future Work We explored the nteracton of dfferent underwater communcaton functonaltes, and developed a cross-layer communcaton soluton that allows for effcent utlzaton of the bandwdth-lmted hgh-delay underwater acoustc channel. We showed that end-to-end network performance mproves n terms of both energy and throughput when hghly specalzed communcaton functonaltes are ntegrated n a cross-layer module. As future work, we shall develop an ad-hoc schedulng mechansm to smultaneously handle traffc classes wth dfferent QoS requrements, and we shall ncorporate end-to-end rate control functonaltes to provde far congeston avodance n dynamc condtons. References [1 I. F. Akyldz, D. Pompl, and T. Meloda. Underwater Acoustc Sensor Networks: Research Challenges. Ad Hoc Networks (Elsever), 3(3):257 279, May 2005. [2 C.-S. Chang and K.-C. Chen. Medum Access Protocol Desgn for Delay-guaranteed Multcode CDMA Multmeda Networks. IEEE Transactons on Wreless Communcatons, 2(6):1159 1167, Nov. 2003. [3 A. Muqattash, M. Krunz, and W. E. Ryan. Solvng the Nearfar Problem n CDMA-based Ad Hoc Networks. Ad Hoc Networks (Elsever), 1(4):435 453, Nov. 2003. [4 D. Pompl, T. Meloda, and I. F. Akyldz. Routng Algorthms for Delay-nsenstve and Delay-senstve Applcatons n Underwater Sensor Networks. In Proc. of ACM Conference on Moble Computng and Networkng (Mob- Com), Los Angeles, LA, Sept. 2006. [5 D. Pompl, T. Meloda, and I. F. Akyldz. A Dstrbuted CDMA Medum Access Control for Underwater Acoustc Sensor Networks. In Proc. of Medterranean Ad Hoc Networkng Workshop (Med-Hoc-Net), Corfu, Greece, June 2007. [6 J. Proaks. Dgtal Communcatons. McGraw-Hll, New York, 2000. [7 Y. Sankarasubramanam, I. F. Akyldz, and S. W. McLaughln. Energy Effcency Based Packet Sze Optmzaton n Wreless Sensor Networks. In Proc. of IEEE Internal Workshop on Sensor Network Protocols and Applcatons (SNPA), Seattle, WA, May 2003. [8 E. Shh, S.-H. Cho, N. Ickes, R. Mn, A. Snha, A. Wang, and A. Chandrakasan. Physcal Layer Drven Protocol and Algorthm Desgn for Energy-effcent Wreless Sensor Networks. In Proc. of ACM Internatonal Conference on Moble Computng and Networkng (MobCom), Rome, Italy, July 2001. [9 M. Stojanovc. On The Relatonshp Between Capacty and Dstance n an Underwater Acoustc Communcaton Channel. In Proc. of ACM Internatonal Workshop on UnderWater Networks (WUWNet), Los Angeles, CA, Sept. 2006. [10 R. J. Urck. Prncples of Underwater Sound. McGraw-Hll, 1983.