HYBRID ERASURE ERROR PROTOCOLS FOR WIRELESS VIDEO

Transcription

1 HYBRID ERASURE ERROR PROTOCOLS FOR WIRELESS VIDEO Shirish S. Karade* ad Hayder Radha Departmet of Electrical & Computer Egieerig / 220 Egieerig Buildig Michiga State Uiversity, East Lasig, MI USA {karades, radha}@egr.msu.edu Ph: , Fax: Abstract Recet cross-layer protocols ad related schemes that are desiged for wireless video lead to both errors ad erasures at the compressed video applicatio layer. Hece, we geerically refer to such schemes as hybrid error-erasure protocols (HEEPs). I this paper we aalyze the utility of HEEPs for efficiet trasmissio of video over wireless chaels. The ature of the error/erasure impairmets observed at the applicatio layer ad their impact o wireless video is a fuctio of the particular protocol stack. Thus, ad i order to maitai the geeric ature of the deductios i this paper, we base our aalysis o two (rather abstract) commuicatio schemes for wireless video, hybrid error-erasure cross-layer desig (CLD) ad hybrid error-erasure cross-layer desig with side-iformatio (CLDS). We make a comparative aalysis of the chael capacities of these schemes over sigle ad multi-hop wireless chaels to idetify the coditios uder which the HEEPs ca provide improved performace over covetioal (CON) protocols. I additio, we employ Reed Solomo (RS) ad Low Desity Parity Check (LDPC) code based FEC schemes to illustrate that the improvemet i capacity ca easily eable a FEC scheme employed i cojuctio with a HEEP to provide improved throughput. Fially we compare the performace of CON, CLD ad CLDS i terms of video quality usig the H.264 video stadard. The simulatio results show a sigificat advatage for the HEEPs. Idex Terms- chael capacity, chael codig, video codig, wireless etworks Part of this work is preseted as: S. S. Karade ad H. Radha, Does Relay of Corrupted Packet Lead To Capacity Improvemet?, at IEEE Wireless Commuicatios & Networkig Coferece Shirish Karade ad Hayder Radha, "The Utility of Hybrid Error Erasure LDPC (HEEL) Codes for Wireless Multimedia," IEEE Iteratioal Coferece o Commuicatios (ICC), May This work has bee supported by the Natioal Sciece Foudatio uder NSF Theoretical Foudatio Award CCF

2 I. INTRODUCTION The challeges faced by high bitrate trasmissios (such as video) over wireless etworks have ecessitated the desig of more flexible protocols. Lack of iformatio trasmissio across layer boudaries is cosidered to be a key shortcomig of the covetioal protocol stacks for the wireless media. I order to provide support for high bitrate trasmissio, better power maagemet, improved QOS, more efficiet routig, packet schedulig ad improvig other etwork fuctioalities over wireless etworks, there is a icreased willigess to allow iter-layer commuicatio. Network strategies, which allow cross-layer iformatio trasmissios, ca be broadly classified as cross-layer protocols. Examples of these protocols ad related wireless schemes ca be foud i []-[3]. A cross-layer desig ca provide vital iformatio to higher protocol layers, which ca i tur facilitate a improvemet i performace. A complete redesig of the etire protocol stack would most aturally facilitate a vastly improved performace, but o accout of the large-scale deploymet of the already existig stadards, such approaches are deemed impractical. Hece, the curret attempts at desigig cross-layer protocols have to comply with already existig architectures to a large extet. However, clever maipulatio of some of the already available fuctioalities ca lead to a vast improvemet. A example of such a reassigmet of a existig fuctioality is the UDP-lite protocol ad its extesios [6]-[3]. I this scheme the lik-layer checksums are tured off (i.e., the MAC layer passes corrupted packets to the trasport layer [2],[3]), while the trasport layer checks are ru oly over the packet header istead of the etire packet. This leads to a reductio i the umber of packet drops o accout of residue errors 2 ad thus a reductio i umber of erasures at the applicatio layer. Badwidth hugry video applicatios ca beefit immesely from such modificatios. Larzo et al. [6]-[8] were the first to explicitly advocate the use of partial checksums for multimedia trasmissio whe they preseted the UDP-lite protocol. Their experimetal work was primarily based o stream- I this documet all protocols that relay corrupted packets to the higher layers shall be collectively referred to as hybrid erasure-error protocols (HEEPs). Furthermore, it should be highlighted that i this paper the discussio o cross-layer protocol desigs is withi the paradigm of HEEPs oly. Hece, all the work ot withi the domai of HEEPs is cosidered outside the scope of this paper. Thus from here o i all discussios ad aalysis the term crosslayer should be iterpreted strictly i the cotext of HEEPs. 2 It should be oted that errors are ecoutered at the lik layer oly whe the physical layer error cotrol mechaism fails. Thus errors see at the lik-layer ca be termed as residue errors.

3 ig over the Iteret. The UDP-lite approach has subsequetly bee exteded to wireless architectures [9]- [3]. Sigh et. al. exteded this work for cellular video [9]. Zheg et. al. [0]-[] preseted the Complete UDP Protocol (CUDP) which used the frame error rate from the Radio Lik Physical Layer for improved performace. They also provided some aalytical expressios that estimated the performace of UDP, UDPlite ad CUDP for the theoretical chael they cosidered. Khayam et. al. exteded this work to 802.b WLANS [2]-[3]. I most of the previous work it has bee observed that, to haress the beefits of the trasport layer modificatios due to UDP-lite, a lik-level modificatio is also required. I particular, the lik-layer checksums are tured off. Thus some of the previous work has also itroduced modified lik-layer protocols such as PPP-lite [9] ad MAC-lite [2]-[3]. These lik-layer protocols are idetical to the origial protocols, with the oly differece beig the disablig of lik-level checksum. It should be oted that the modificatio suggested by HEEPs lead to the existece of errors at the applicatio layer. If the umber of errors i the corrupted packets relayed to the applicatio layer is ot large, the the total umber of chael failures is reduced. (Here, a chael failure could be a error or a erasure.) This reductio i the total umber of failures ca lead to a icrease i chael capacity. Previous studies [6]-[3] have assumed that the probability of bit-error i the corrupted packets is always low eough to dramatically reduce the total umber of bit/symbol impairmets ad thus reder the performace of a cross-layer approach to be better tha a covetioal oe. However, it is well kow that the erasure correctig capability of ay chael code is much better tha its ability to correct symmetric errors [4] 3. Thus if the corrupted packets which are relayed to the applicatio layer have a high corruptio level (i.e. high percetage of errors) the allowig the existece of corrupted packets at the applicatio layer ca i fact lead to a reductio i throughput ad chael capacity. Thus there is a iheret tradeoff i relayig corrupted packets to the applicatio layer. Therefore the utility of a HEEP approach is strogly coupled with the corruptio level i the relayed packets. I may realistic scearios (ad especially the chael coditios cosidered by [6]-[3]), the corruptio level is low; however the choice betwee a cross-layer scheme ad the covetioal oe is ot always straightforward. Oe such example is the case of multi-hop wireless sceario, which has ot bee co- 3 dmi For example, it is kow that a liear code with miimum distace dmi ca correct upto 2 the ability to correct upto d erasures [4], [5]. mi errors, as agaist

4 sidered thoroughly by previous efforts i this area. The level of corruptio i a packet icreases as it iteratively udergoes impairmets whe relayed over multiple wireless hops. If the umber of hops are large eough, it is quite possible that the covetioal schemes actually perform better tha HEEP. Furthermore, the aalysis provided i [6]-[3] is architecture/stadard specific ad thus the possibility of high corruptio levels i corrupted packets relayed to the applicatio layer caot be ruled out i a geeric sese. Thus the above-metioed tradeoff eeds to be aalyzed ad studied i greater detail. Thus the primary goal of the preseted work is to aalyze the above highlighted tradeoff ad idetify the chael coditios uder which the existece of corrupted packets at the applicatio layer ca be preferred over droppig them etirely. I order to keep the discussio geeric ad ot depedet o a particular implemetatio (or stadard), i this paper we cosider three rather abstract commuicatio schemes (a) trasmissio over erasure chaels, which represets the covetioal protocols like UDP (CON) (b) trasmissio i presece of erasures ad errors usig a cross-layer desig (CLD), which is represetative of schemes like UDP-lite (c) side-iformatio ehaced trasmissio i presece of erasures ad errors usig a crosslayer desig (CLDS). Detailed descriptio of these schemes is provided i sectio II. Noe of the previous studies has attempted to directly quatify the gai i chael capacity o accout of employig a HEEP ad have istead limited the aalysis oly to idirect measuremets. Thus i sectio III we evaluate the chael capacity (i a iformatio theoretic sese) of each of the above metioed abstract schemes. A comparative aalysis of the chael capacities, i order to idetify the coditios uder which the cross-layer protocols ca provide improved performace, is preseted. I sectio III we also show that the chael characterizatio described i sectio II ca be easily exteded to characterize trasmissio over multiple hops. Thus we evaluate the chael capacities of the above-metioed schemes whe the trasmissio is over multiple hops too. We prove that the capacity of the CLDS scheme is always better tha either CLD or CON. We also idetify the chael coditios uder which CLD scheme ca perform better tha CON. Discussio pertaiig to some subsidiary issues related to the depedece of the deduced applicatio layer chael capacities o the lik-layer chael model have bee deferred to the Appedix. Despite icrease i capacity, the utility of a cross-layer approach ca be udermied, if a applicatio layer Forward Error Correctio (FEC) is uable to take advatage of it. As the HEEPs permit the occurrece of

5 errors ad erasures at the applicatio layer, ay FEC scheme used i cojuctio with these protocols must be capable of hybrid erasure ad error recovery. Previous work o HEEPs does ot address chael codig related issues i sufficiet detail. Algorithms capable of hybrid erasure-error decodig of Reed Solomo (RS) codes have existed for a log time (see [4], [5] ad citatios withi), similarly the sum-product decodig algorithm used for decodig Low Desity Parity Check (LDPC) codes [6]-[22] ca also be easily exteded to facilitate hybrid error-erasure decodig, however their utility at the applicatio layer especially i a cross-layer sceario is uder explored. Thus i sectio IV we employ these hybrid decodig algorithms to i cojuctio with HEEPs ad evaluate the throughput improvemets. Packet throughput improvemet does ot ecessarily traslate ito improvemet i video quality. Thus it is importat to explicitly establish the impact of cross-layer protocols o video quality. Thus i sectio V we compare the performace of the three schemes, CON, CLD, ad CLDS, i terms of the video quality provided to the ed receivers. We use the H.264 [35] stadard for all the simulatios i sectio V. Fially, Sectio VI summarizes the key coclusios of the preseted work. II. CHANNEL CHARACTERIZATION I this sectio we primarily characterize the three abstract protocols uder cosideratio. The three commuicatio schemes cosidered i this paper ca be explaied by cosiderig a geeric Logic Trasmissio Uit (LTU) as show i Figure. The geeral packet structure ca be segregated ito two parts ) the header iformatio 4 ad 2) the data payload. I additio to traditioal header iformatio (e.g., ode addresses etc), the LTU header cotais two sets of checksums CK-HDR ad CK-DATA. The CK-HDR checksum is applied to ad depedet o the header iformatio oly; while the CK-DATA check sum is applied to ad depedet o the data payload oly. Hece, uder this geeric LTU model: The covetioal (o-cross layer) protocols drop a packet if either of the checksums CK-HDR, or CK- DATA is ot satisfied. 4 I may practical implemetatios the header ad CK-HDR might be further partitioed ito multiple headers ad checksums. I additio a direct correlatio of a specific stadard/architecture/implemetatio with the abovecosidered abstract LTU may ot always be possible. Reallocatio (or eve additio) of some header fields might be required. However, we itetioally do ot address such implemetatio details to maitai the geeric ature of argumets preseted i this sectio.

6 CLD, which represets protocols like UDP-lite [8], turs the CK-DATA checksum off ad drops the packet oly if CK-HDR is ot satisfied. Therefore, a CLD chael exhibits both erasures (due to CK-HDR violatios) ad possible errors i some of the delivered packets. It is importat to ote that (without further iformatio or additioal parity bits) the CLD chael receiver does ot kow which delivered packets are error-free ad which packets are corrupted. It oly distiguishes betwee erasures ad delivered packets. CLDS is a alterative to the above schemes. Similar to CLD, a CLDS chael drops a packet oly if CK- HDR is ot satisfied. However, i CLDS the CK-DATA is ot tured off but either is the decisio to drop a packet depedet o this checksum. Moreover CK-DATA ad iformatio about the success or failure of this check-sum is made available to the applicatio layer as side iformatio. Therefore, ad ulike a CLD receiver, the CLDS receiver ca distiguish corrupted packets from error-free packets. Thus the chaels uder cosideratio ca be characterized as give below: ( i ) δ is the probability that at least a sigle bit is i error i the header ad/or the data payload. Thus δ is the probability of a packet beig dropped i a covetioal (o-cross layer) protocol because at least oe of the checksums, CK-HDR ad/or CK-DATA, was ot satisfied. ( ii ) λ is the probability that the packet header cotais at least a sigle bit i error. Thus λ is the probability of packet beig dropped i the cross-layer schemes because the check CK-HDR was ot satisfied. (Note that this evet could occur regardless if there is a error withi the packet data or ot.) Thus δ λ represets the probability of a corrupted packet beig delivered to a CLD/CLDS chael receiver. ( iii ) ε is the coditioal probability of a bit i the data payload beig i error give that the checksum CK-HDR is satisfied ad checksum CK-DATA has failed. Give a corrupted packet at a CLD/CLDS receiver, ε represets the probability of havig a radom bit selected from that packet to be i error. It is worth otig that ε specifically represets the probability of error i corrupted packets relayed to the applicatio layer. Thus the overall probability of bit error i packets that are received at the applicatio layer (i.e. all packets that do t get dropped due to corruptio i the header) is give by ( δ λ) ε p =. I ( λ)

7 other words p is the coditioal probability of a bit i the data payload beig i error give that the checksum CK-HDR is satisfied. ( iv ) For a cross-layer protocol with side-iformatio (CLDS) let Z be a discrete radom variable that takes o three possible outcomes: S = { 0,,? } z. Where, (i) Z=? if the header cotais at least sigle bit error ad CLDS drops the packet. Thus pz ( =?) = λ (ii) Z=0 if a packet cotais o errors i the header but cotais at least a sigle bit error i the data payload. Thus pz ( = 0) = ( δ λ) ad (iii) Z= if either the header or the data payload cotai eve a sigle erroeous bit. Thus pz ( = ) = ( δ ). 5 Thus the CON, CLD ad CLDS schemes ca be represeted by Figure 2 (a), (b), (c) respectively. III. CAPACITY EVALUATION Popular iformatio chaels [32] ad some extesios of these chaels (as outlied below) ca provide ivaluable represetatio ad isight regardig the performace of the covetioal ad cross-layer protocols. Our focus here is o providig a uifyig framework (albeit with simplifyig assumptios 6 ) for a comparative aalysis amog the differet protocols i terms of chael capacity. We begi by briefly outliig the chael capacities of the covetioal ad cross-layer protocols usig the geeric error ad erasure parameters described i the previous sectio: The covetioal protocol (CON), which is the simplest amog the three (CON, CLD, ad CLDS), ca be represeted by a Biary Erasure Chael (BEC). It is well kow [32] that the chael capacity of a BEC is give by δ ; hece, the capacity of a covetioal protocol is give by CCON = δ ( ) The cross-layer chael (CLD) ca be represeted as a cascade of a BEC chael with probability of erasure equal to λ followed by a Biary Symmetric Chael (BSC) with probability of bit error equal to p. Such a cascade ca hece be termed as Biary Symmetric/Erasure Chael (BSEC). It ca be easily show 5 I this paper we assume that probability of false alarm ad the probability of missed detectio of the checksums is egligibly small. 6 I strict iformatio theoretic terms, the capacity should actually be derived by treatig a etire packet as a symbol ad the describig the chael formed by CON, CLD, ad CLDS i terms of this symbol. However by derivig the capacities i terms of bit-equivalet chaels we sigificatly simply the capacity deductios, while still mimickig the overall capacity scalig treds ad thus facilitatig a appropriate compariso betwee CON, CLD ad CLDS.

8 that the chael capacity of such a cascade is give by the product of the chael capacities of the idividual chaels: C = ( λ) ( h ( p)) ( 2 ) CLD b where h ( b p ) is the etropy of a biary radom variable with parameter p. The chael capacity of the cross-layer chael i presece of side iformatio Z is as follows: whe Z= all the bits are trasmitted reliably; ad i this case, which occurs with probab ility ( δ ), the (coditioal) capacity is. Whe Z=? all the bits get erased ad the coditioal capacity is 0 while whe Z=0 the chael reduces to a BSC with a cross-over probability ε ad the coditioal capacity is ( ( ε )). Thus the chael capacity of CLDS is give by: h b C = ( δ ) + ( δ λ) ( h ( ε)) ( 3 ) CLDS b A. Trasmissio over multiple-hops A multi-hop wireless chael ca be represeted as a cascade of chaels. For example, a cascade of BEC chaels ca represet the covetioal scheme, where a distict BEC chael i the cascade represets each lik. Similar cascades of BSEC chaels ca be made for the cross-layer schemes too. However, i case of CLDS it should be oted that the side iformatio Z shall deped o the cumulative errors ecoutered over all the hops. Thus i geeral we ca express the ed-to-ed parameters for chael capacities over multiple hops as give by equatios (4), (5), (6). C = ( δ ) CON ( hop) i ( 4 ) i= C = ( λ ) h * p 7 ( 5 ) CLD( hop) i b i i= i= C CLDS ( hop) = ( δi ) + B A ( 6 ) i= * 2 = ( 2) + 2 ( ) pi = i= p p2 p3 pi p 7 Note: Operator * is s.t. p p p p p p ad * ( * )* ) )* ) )* )

9 ( ) λi i where B= ( λi) ( δi) & A = = h b * p i i= i= i= ( λi) ( δi) i= i= Equatios (4), (5), (6) ca be covert them ito equatios (), (2), (3) respectively by employig the followig substitutios: Let δ = ( δi ), λ = ( λi ) i=, λ p = * pi i= ad ε = p δ λ i= Thus for the remaider of this paper, uless explicitly stated, we fid it sufficiet to maitai the discussio i terms of δ, λ, p, ε. A.. Hop-Hop retrasmissios It should be oted that the above deductio of chael capacities are ot applicable to schemes that employ hop-hop retrasmissios. Hop-Hop retrasmissios represet a possible way of embeddig error cotrol mechaisms withi the etwork. Though curretly this is the most commoly deployed method, other methods to embed error cotrol mechaisms withi the etwork do exist [30], [3]. Thus a etwork architecture, which allows embedded error cotrol, eeds separate attetio ad should be distiguished from ed-to-ed error cotrol. Thus to provide a fair compariso betwee the cross-layer approach ad the covetioal approach we should permit the existece of etwork embedded error cotrol i a cross-layer architecture too. However detailed aalysis of such schemes is outside the scope of this paper. Meawhile, it is worth otig the followig: a) The capacity of covetioal chael employig hop-hop retrasmissios is give by C CON ( retras) [, ] ( δ ) = argmax ( 7 ) i i b) Hop-Hop retrasmissios ca be employed i a cross-layer sceario too. Oe possible method to do so is to request retrasmissios whe the CK-HDR fails. The capacity of a cross-layer scheme employig hophop retrasmissios as described above is

10 C = argmax( λ ) h * p [, ] ( 8 ) CLD( retras) i b i i i= C = ( δ ) + argmax( λ ) ( δ ). i= i [, ] i= CLDS ( retras) i i i h b argmax( λi ) i [, ] argmax( λi) ( δi) i [, ] i= * i= pi ( 9 ) c) Wireless LAN s usually cosist of a sigle wireless hop. Thus for most WLANs it should be expected that the reliability/throughput offered by the cosidered cross-layer schemes i cojuctio with FEC should be eve better tha a retrasmissio based scheme o the covetioal stack. A.2. Hybrid Architecture Throughout this paper we assume that the etire etwork path from the source to the destiatio is capable of supportig a cross-layer approach. However, as show i Figure 3 that eed ot be the case always. If we assume that the total path cosists of hops; ad hops m to s lie iside the o cross-layer sectio the the expressio for capacity deductio would get modified as described below: C s = ( δ ) ( λ ) h * p ( 9 ) CLD( hop) i i b i i= s+ i= i= s+ C = + s ( -δ ) ( - δ ) (- λ ) ( -δ ) CLDS ( - hop) i i i i i= i= i= s+ i= s+ (- λ ) i i s - h = + b * p i i= s+ (- λi) ( - δi) i= s+ i= s+ (0) Thus it should be oted that the capacity improvemet o accout of a cross-layer approach ca be severely dimiished over multiple hops. This ca be avoided by updatig all the CK-DATA s at the last ode (i this case ode m) before the o cross-layer sectio begis. I additio, a header field that keeps a cout of the

11 checksum updates should be icluded. At the ed receiver the umber of updates field i additio to the updated CK-DATA ca be used to determie the corruptio status of the payload. B. Comparative Aalysis Figure 4 shows a compariso of the applicatio layer capacities offered by the three protocols cosidered here, uder various chael coditios. It ca be clearly see that the cross-layer protocols ca provide dramatic improvemets i capacity uder a variety of chael coditios. Furthermore it ca be observed, (i) as show i Figure 4 (a), the relative performace of the cross-layer schemes improves whe the packet drops due to data payload corruptio icrease (i.e. δ λ icreases), while (ii) as show i Figure 4 (b) ad (c), the relative performace deteriorates as the packet drops due header corruptios icreases (i.e. λ icreases) or whe the corruptio level i the u-erased but corrupted packets icreases (i.e. ε icreases). Two additioal importat observatios that ca be derived from Figure 4 (c) are: (i) I the absece of CK-DATA sideiformatio, whe the packets are highly corrupted, the capacity of covetioal protocol ca actually be better tha that of the cross-layer. Thus cross-layer schemes that advocate a simple disablig of the payload checksum, should esure that the bit-error rate i the corrupted packets is withi acceptable limits, else the performace ca ifact worse due to deploymet of partial check sums. (ii) If the CK-DATA checksum is ot tured off, ad this iformatio is provided to higher layers, the the resiliece of HEEPs to high corruptio levels, sigificatly improves. Ifact it ca be observed that the capacity of CLDS is always better tha that of CON ad CLD. This improvemet i capacity is substatially magified as the corruptio level icreases. The above two observatios ca be formalized by cosiderig the followig two propositios. Propositio states that the performace of CLDS is always better tha CON ad CLD uder all chael coditios. While Propositio 2 will allow us to idetify the chael coditios uder which CLD ca perform better tha CON. Propositio : a) C CLDS C with equality occurrig iff ε = 0.5 or δ = λ. (b) C C with CON CLDS CLD equality occurrig iff δ = λ.

12 Proof: a) Sice ( δ λ) ( h ( ε)) 0. Equality occurs iff ( δ λ) ( h ( ε)) = 0, i.e. iff δ = λ or h ( ε ) = i.e. ε = 0.5. b b) This result ca be proved by usig the followig fact: b b If f ' ( x ) is strictly mootoically decreasig over [ ab, ] the α, β [ ab, ], f( α ) f( β ) α, β 0, α < β α > β. Notice that ' h ( ) b x is strictly mootoically decreasig for x [0,]. As λ δ it ca be show that p ε, which implies that hb( p) hb( ε ). Thus, substitutig for p we p ε ca coclude that ( δ ) + ( δ λ) ( h ( ε)) ( λ) ( h ( p)). Note that equality ca occur oly whe p = ε, which is possible iff δ =. b b Propositio 2: C > C h ( p) > ( p / ε ), CON CLD b Proof: Ca be directly derived from expressios (2) ad (3). For a give value of δ, λ ad for ε 0.5 the equatio h ( p) = ( p/ ε ) i.e. b h b ( δ λ) ε ( δ λ) = ( λ) ( λ) has sigle uique solutio. Thus combied with the above observatio, Propositio 2 tells us that for a give δ, λ, there exists a threshold ε mi ( δλ, ) such that if the level of corruptio i the corrupted (but ot dropped) packets is greater tha this threshold the the covetioal schemes (CON) shall perform better tha the cross-layer scheme (CLD). Thus s δ, λε, ( δλ, ) = mi defies a surface i the three dimesioal parameter Ω= (i.e. all possible combiatios of δ, λε, such that space represeted by { δ, λ, ε δ λ, ε 0.5} δ λ, ε 0.5). The surface s = δ, λε, mi ( δλ, ) divides the parameter space ito two distict regios, oe that cotais all possible values of δ, λε, s.t. C CON > C CLD ad aother that cotais all possible values of δ, λε, s.t. C CON < CCLD. Similarly, for a give λ, we ca idetify a curve δ, εmi ( δ) such that it divides the two dimesioal parameter space λ { δ, ε δ λ, ε 0.5} Ω = ito two regios. Figure 5 shows such curves for various values of λ. I Figure 5 the regio above the curve correspods to all possible values

13 [, ] δ ε that lead to C C, while the regio below the curve correspods to C < C. All the CON > CLD CON CLD poits o the surface mi ( δλ) δ, λε,, correspod to the chael coditios whe C = C, similarly CON CLD for a give λ, all the poits o δ, ε ( δ) mi correspod to the chael coditio whe CCON = CCLD. The threshold defied by ε ( δλ ) ca be used to improve the efficiecy of a commuicatio scheme, i mi, scearios where side-iformatio from the physical layer or some steady state chael statistics (or some other method) may make it possible to acquire a estimate of the corruptio level i a packet. This estimate i cojuctio with ε ( ) mi, δλ ca allow us to idetify the highly corrupted packets, which might be preferred to be dropped. Fially, it should be oted that sice a appropriate choice of parameters ca represet a multi-hop sceario, the propositios proved i this sectio are valid for multi-hop sceario also. Thus it ca be observed i Figure 4 (d) that eve i the multi-hop case the performace of CLDS is the best. However, ulike CLDS the performace of CLD is ot always better tha CON. If the corruptio level icreases, the performace of CLD ca i fact drop below CON, however as show i Figure 4 (d), there ca exist may operatig coditios whe the capacity of CLD is better tha CON eve over multiple hops. Extesive modelig of lik-layer chaels is a topic which has bee ivestigated by recet studies. These studies have ofte focused o models for a specific architecture/stadard. (e.g. [3], [26], [27] primarily cocetrate oly o the 802.b). This motivated our focus o a rather abstract set of chael parameters, amely, δ, λ, ε. However, i geeral, these parameters are ot idepedet of each other ad are i fact a fuctio of the uderlyig lik-layer chael. Assumig a specific lik-layer chael ca ideed lead to more precise coclusios ad reder some of the regios (combiatios of δ, λ, p ) uachievable. I the Appedix we cosider two simple example chaels to exhibit as to how a specific lik-layer chael model ca facilitate more precise coclusios. I additio i order to show the utility of the preseted discussio for actual lik-level chaels, the Appedix also presets results of some cross-layer video simulatios o actual 802.b error traces ad aalyzes the observed performace vis-à-vis the lik-level parameters δ, λ, ε.

14 IV. PERFORMANCE EVALUATION OF FEC WITH HEEPS So far, we have exhibited the capacity improvemet that ca be achieved o accout of relayig corrupted packets to the applicatio layer. I order to exploit this icreased capacity offered by the cross-layer approach, efficiecy of FEC schemes suitable for the HEEPs eeds to be evaluated. As may curret deploymets of FEC schemes for low-delay applicatios employ RS codes (e.g. [6]-[7]) it is importat to establish the gais of usig a cross-layer protocol with a RS based FEC scheme. O the other had, the past few years have see tremedous research iterest i LDPC codes [9]-[25]. These codes are show to be capacity achievig over a variety of chaels. Thus the utility of LDPC codes i the desig of efficiet FEC schemes for the cross-layer chaels also eeds to be evaluated. I this sectio we shall make comparative aalysis of all combiatios of protocols ad chael codig schemes. For all the simulatios i this paper, uless stated otherwise, we use a packet block-legth of 30 packets, a packet size of 500 bytes ad codig rate of 0.66 (i.e. 20 message packets i each block). The packet based FEC is obtaied o the basis of a simple iterleavig scheme, where the etire block of packets ca be broke dow ito c code words. Each codeword cosists of u symbols derived from each packet, where each symbol ca be represeted by v bits. Thus for the RS based FEC, we choose c = 00, u = 5, v = 8 (i.e. RS code is based o 8 GF(2 ) ). Hece the legth of each RS code is 50 symbols. Similarly for LDPC based FEC, we choose c = 4, u = 000 ad v = (i.e. LDPC code is a biary code). Hece the legth of each LDPC code is symbols. Though our choice of code legths is quite arbitrary, care has bee take to esure that the performace treds are represetative ad the time-complexity of LDPC based FEC is lesser tha the RS based FEC 8. The LDPC code we use i this sectio is a regular LDPC code with 3 checks per variable bit. The parity check matrix was costructed usig the Progressive Edge Growth Techique (PEG) [23], [24]. 8 More specifically, we have esured that if b bits correspod to (i) c RS umber of RS codewords of block-legth RS symbols ad (ii) c LDPC umber of LDPC codewords with block-legth LDPC symbols, the they satisfy the coditio 2 cldpc LDPC < crs RS. For the codes cosidered i this sectio, it ca be verified that ( ) ( ) ( ) ( ) 2 cldpc LDPC , ie.. = c RS RS

15 Similar to [3], the decodig algorithm used for simultaeous erasure ad error decodig of RS code is the Berlekamp-Massey Algorithm (see e.g. [5] for backgroud), while the decodig algorithm used for LDPC is based o the Log-Likelihood Ratio (LLR) domai [22] 9 implemetatio of the sum-product decoder. LDPC decodig i presece of erasures is achieved by settig the iitial LLR value of the erased bit to zero. For the LDPC decodig, we used a stoppig criteria of checkig for a valid codeword after each iteratio ad the maximum umber of iteratios has bee limited to 25. Side-Iformatio for Improved decodig efficiecy: Performace over CLDS ca be improved by takig advatage of the side-iformatio provided by CK- DATA for FEC decodig. This ca be doe as described below (a) FEC block decodig failure: O occurrece of a block decodig failure the chael decoder is uable to recover all the dropped/corrupted packets. However as the FEC block is systematic, message packets that ca be idetified as ucorrupted ca still be forwarded to the evetual applicatio. I a cross layer desig like CLD as we do ot have iformatio about CK-DATA, if a decodig failure is ecoutered, the etire packet block is dropped. However i a CLDS scheme the CK-DATA iformatio is used to forward the correctly received message packets to the applicatio layer. (b) Apriori estimates for improved FEC decodig: CK-DATA side-iformatio ca be utilized to acquire improved apriori estimate of chael impairmets ad thus improve FEC decodig: (i) I case of RS based FEC, we use this iformatio rather simplistically, if the umber of packets for which CK-DATA fails is less tha the total umber of redudat packets, the eve if all the corrupted packets are treated as erasures the etire FEC packet block is recovered by pure erasure decodig. Thus the CLDS scheme ca switch to the pure erasure decodig wheever possible to avoid decodig failure due to a high corruptio level i the packets. Further improvemets i performace ca be obtaied by employig softdecodig of RS codes (e.g. see [8]). However RS based soft-decodig algorithms usually have higher time complexity ad may ot always be suitable for low-latecy video applicatios. (ii) The sum-product LDPC decodig algorithm ca also be modified to take advatage of the side iformatio provided by CK-DATA. If the CK-DATA of a particular packet is satisfied tha we set the apriori 9 The software implemetatio used for the simulatios i this sectio is a modified versio of that provided at[25].

16 probability of the bits i this packet beig i error to 0, this is achieved by settig the magitude of the LLR to. Figure 6 shows the performace of various FEC ad protocol combiatios averaged over a trasmissio of 0,000 blocks of 30 packets. All the results i Figure 6 are expressed i terms of message throughput: ( message packets recieved after chael decodig ) ( message packets tramitted ) τ =. It ca be observed that for a give FEC scheme, the performace of CLDS is better tha both CON ad CLD. I Figure 6 (a), it should be observed, that though the codig rate is well below capacity, the CON scheme is uable to provide 00% packet recovery, as agaist that it ca be observed i Figure 6 (a), (b) ad (d) that whe the corruptio level is low, eve the most straight forward of cross-layer combiatios (i.e. RS- CLD), is sufficiet to provide substatial improvemet i throughput over covetioal schemes ad almost a 00% reliability. However, cotrary to some of the previous works (e.g. [2], [3]), Figure 6 (c) shows that, whe the corruptio level icreases, the performace of (as well as ) ca actually drop well below the performace of the covetioal scheme (CON). It ca be observed that CK-DATA side-iformatio is essetial to guaratee that the performace of a HEEP is ever worse tha CON. A simplistic usage of the side-iformatio as doe by is sufficiet to provide improved performace (as ca be observed for ε > 0.0 i Figure 6 (c)) ad i additio to guaratee that a HEEP does ot ever perform worse tha a covetioal scheme. (as ca be observed for ε > 0.09 i Figure 6 (c)) Fially it should be oted that amog all the cosidered combiatios, provides the best performace. The performace improvemet o accout of usig, as compared to other combiatios, is especially high whe the corruptio level i the packets is high. Uder poor chael coditios, the decodig algorithm fails to coverge for the CLD scheme, while still successfully decodig i the case of the CLDS scheme. Similarly, eve uder chael coditios whe both ad are able to provide almost 00% reliability, the umber of iteratios used by the CLDS scheme are lesser. This ca play a critical role, whe the ed-receiver is a low-power device. Detailed aalysis of the complexity versus throughput tradeoff is deemed outside the scope of this paper due to brevity cocers.

17 V. VIDEO SIMULATIONS Discussios i previous sectios have cocetrated o exhibitig the capacity ad packet throughput improvemets that ca be achieved by usig cross-layer protocols. At this stage, it is ecessary to clearly establish whether these improvemets ideed traslate ito improvemet i the quality of video available at the ed receivers. Thus i this sectio we use the emergig H.264 [35] video stadard to preset some results based o video simulatios. The results preseted i this sectio are a subset of the examples we cosidered ad thus it should be stated that the choice of test sequece, quatizatio, frame frequecy ad frame size do ot compromise o the geeraless of the coclusios derived o the basis of the video aalysis preseted here. Thus here we preset results based oly o the Stefa, Carphoe, Paris, test sequeces. A CIF (352x288) frame size ad a frequecy of 30 frames/sec have bee chose for all sequeces. We used a costat quatizatio size of QP = 28 for all sequeces. We used a GOP size of 5 frames ad a GOP structure of IPPP. It is also worth otig that all the ecoded streams are made up of video packets of size 500 bytes each. The referece software versio of H.264 used for our simulatios was TML 9.0. The ecodig of the sequece was RDoptimized for 30% losses to icrease the error-resiliece. The FEC ecodig parameters used i this sectio are idetical to those used i the previous sectio. The video quality performace was studied uder a variety of chael codig parameters ad chael coditios. However i this sectio we preset results based oly o a sample of the various scearios we cosidered. Thus we specifically focus o chael coditios, which were similar to the oes used to geerate Figure 6 (c) ad (d). Results preseted i Figure 7 evaluate the impact of corruptio level o the video quality ad correspod to chael coditios similar to those used i Figure 6 (c), while results preseted i Figure 9 correspod to the chael coditios similar to those used i Figure 6 (d) ad evaluate the degradatio i video quality as the data is relayed over multiple hops. Each data poit i Figure 7 ad Figure 9 is obtaied o the basis of averagig the performace over three repetitios of trasmittig the first 300 frames of a test sequece. I each experimet ad for each protocol-fec combiatio we esured that the first I frame is trasmitted i a lossless maer.

18 Uder severe chael coditios, may of the picture frames are completely u-recovered. This leads to sigificat amout of motio discotiuity ad also block-distortio due to lack of a good referece frame. Thus i Figure 7 (d), (e), (f) ad Figure 9 (d), (e), (f) the motio cotiuity has bee measured i terms of the proportio of successfully recovered picture frames γ ; while the block-distortio i Figure 7 (a), (b), (c) ad Figure 9 (a), (b), (c) has bee calculated i terms of average PSNR, which has bee obtaied by assigig 0dB to the missig frames. Naturally, the presece of error cocealmet ad improved errorresiliece techiques ca improve the video quality performace of each of the cosidered scheme. However, detailed aalysis of all the curretly available error cocealmet/resiliece tools i H.264 is outside the scope of this paper. Istead we focus o presetig the results as per the above described experimetal setup ad hope that the treds i video quality degradatio may be useful i desig of future error-resiliece ad cocealmet features specifically caterig to commuicatio over HEEPs. I Figure 7, the chael coditios ( δ = 0.33, λ = 0.0) reder the operatio of the CON scheme to be extremely close to capacity. Thus eve whe the corruptio levels are low, the CON scheme provides a very poor video quality. As agaist that, at low corruptio levels, all the FEC-HEEP combiatios ca provide a performace improvemet i excess of 4 db for Stefa, 9dB for Carphoe ad 7dB for Paris. The motio discotiuity for the CON scheme is also very high as more tha 20% of the picture frames remai completely u-recovered. As the decodig algorithm employed for RS is a hard-decodig algorithm, the impact of errors o RS based FEC is sigificat. With a slight icrease i the corruptio level, the performace of both ad drops drastically. It may be oted that the performace of ca ifact drop below eve at modestly high corruptio levels. Thus we re-iterate that whe CK-DATA checksum is completely tured off ad the applicatio layer FEC is a RS based hard-decodig algorithm, its essetial to verify that the chael coditios do ot reder the corruptio levels to be to high, else the performace of a HEEP may ifact be worse tha a covetioal scheme. As should have bee aticipated, the performace of ever drops below 0. The soft-decodig algorithm employed i cojuctio with LDPC makes it more resiliet to errors. However, after a certai corruptio threshold, the perform- 0 I practice, whe the errors are bursty ad clustered together, we expect the impact of errors o RS based FEC to be less sigificat.

19 ace of eve drops below that of CON. As agaist this ca cotiue to provide performace improvemet i video quality i excess of 5dB over all other schemes, eve whe the corruptio level is as high as 5%. I Figure 8 we provide a detailed compariso betwee the video quality provided by ad for chael coditios correspodig toδ = 0.33, λ = 0.0, ε = 0.07, which represets a data poit i Figure 7. From the temporal sapshots i Figure 8 (a), (b), (c) it ca be observed that at high corruptio levels the video quality for ca frequetly drop below uacceptable levels. It should be oted that the video quality is able to recover, despite frequet deterioratio, because source codig is RDoptimized for 30% losses. If the source codig is less robust/error resiliet, the performace of compared to would be eve worse. It should also be oted that i Figure 8 (a), (b), (c) the dowward spikes that reach all the way upto 0 db represet the loss of a etire picture frame. As has bee elaborated before, loss of etire picture frames ca lead to severe motio discotiuity. Thus the primary artifact, especially i a video sequece such as Stefa, was motio discotiuity. However, eve whe a picture frame is ot dropped etirely, severe block-distortio ca be observed. Figure 8 (d), (e), (f) show subjective examples for comparig the block-distortio. For the particular images show i Figure 8, the differece i video quality betwee ad is greater tha 8dB for Stefa, 4dB for Carphoe ad 8dB for Paris. I Figure 9, the chael coditios o each hop are represeted byδ = 0.06, λ = 0.0, ε = 0.0. Thus the chael coditios o each hop are ot very severe ad it ca be see that, o the iitial 2-3 hops all the schemes icludig CON ca provide reasoable video quality. However, as the hops icrease, the performace of the CON scheme drops drastically. It was observed that by hop 6, the performace of CON as compared to all the FEC-HEEP combiatios is extremely poor. Thus, eve if the chael coditios o every sigle hop are ot very bad, the cumulative effect of a multi-hop commuicatio o a covetioal scheme ca be quite drastic. Hop 6 owards all the FEC-HEEP combiatios provide atleast 6dB improvemet i video quality over the CON scheme. Though the corruptio level i each packet is at ε = 0.0, the average bit error probability amog all the u-erased packets is as low as p = A packet that is corrupted o

20 oe hop, does ot ecessarily get corrupted o aother. Thus eve as the umber of hops icrease, the average corruptio level, eve by the 0 th hop does ot icrease drastically. Thus it ca be observed that all the FEC-HEEP combiatios cotiue to provide reasoable video quality eve till the 0 th hop. Never the less, by the 0 th hop, the soft decodig algorithm employed i cojuctio with LDPC allows it to provide -2dB performace improvemet over all other FEC-HEEP combiatios. The performace of ad is almost idetical o all the hops, however if we keep icreasig the umber of hops over which the iformatio is relayed, evetually shall start performig better (ad at times sigificatly better) tha. VI. CONCLUSION I this paper we aalyzed the tradeoff ivolved i allowig the relay of corrupted packets to the applicatio layer. Chael coditios uder which HEEPs ca provide improvemet were idetified. The capacity improvemets were quatified ad it was established that a dramatic icrease i capacity ca be achieved i may realistic scearios by usig HEEPs. Similarly, performace of RS ad LDPC based FEC i cojuctio with HEEPs was show to be sigificatly better tha RS based FEC over covetioal protocols (CON) uder several chael coditios. H.264 based video simulatios were used to clearly establish that improvemet i capacity ad packet throughput does actually traslate ito sigificat improvemet i video quality. It was observed that video quality improvemet provided by HEEPs ca be dramatic ad i excess of 6-0dB uder several chael coditios. It was observed that the performace of HEEPs ca be affected by packet drops due to header corruptio ad icrease i the bit error rate of the corrupted packets. It was show that at high corruptio levels the sideiformatio provided by CK-DATA ca improve the performace of HEEPs sigificatly. Combiatio of LDPC ad CLDS was observed to provide the best performace. At high corruptio levels absece of CK- DATA based side iformatio reder a CLD like HEEP to perform worse tha CON. Thus, o the basis of our observatios, it ca be recommeded that: HEEPs should be further explored ad deployed as they have the potetial to be extremely useful i dramatically icreasig the throughput of video applicatios

21 Future desigs of HEEPs should: o o icrease header robustess ot discard ay chael state iformatio from lower layers, which ca be utilized as sideiformatio at higher layers o use error cotrol algorithms capable of efficietly utilizig the side-iformatio from lower layers I absece of CK-DATA like side-iformatio the performace of a HEEP ca be worse tha that of covetioal protocol ad thus a HEEP should be employed with utmost care. ACKNOWLEDGEMENTS The author s would like to ackowledge the cotributio of Kira Misra for providig us with software related to LDPC codes ad also for some helpful techical discussios. Lik Layer Biary Symmetric Chael: APPENDIX If we assume the lik-layer chael to be a BSC chael with cross-over probability q ad, let h ad d ( ) represet the umber of header ad payload bits respectively, the we ca say ( ) h + δ = q d, λ = ( q) h ad ε = q. Thus, the chael capacities are give by C = q + CON ( ( ) h d ) d ( b (( ( ) ) )) ( ) ( ) h C = q h q q ( 2 ) CLD d ( (( ( ) ) b ( ))) ( ) h C = q q h q ( 3 ) CLDS Similar to propositio 2, we ca deduce a threshold, which divides the parameter space ito 2 halves. Note that this is equivalet to deducig a threshold d, sice for a give q ad h as the payload size icreases the susceptibility of the CON scheme relative to CLD icreases. The followig propositio proves the existece of such a threshold. Propositio 3: q (0,0.5), dmi 0 s.t., d dmi C CLD C CON equality occurs iff d = dmi. ( - - d ) ( - ( - d b ) ) Note: ( ) ( ) q h q q CCLD CCON q mi mi

22 The above propositio ca be proved by showig that the equatio ( ) d at maximum have a sigle o-egative solutio d sol (( ( ) ) ) d = b q which is less tha ( q) lim h - - q q h ( ) d b d ( ) ( ) ( - - ) ( - ( - d q hb q) ) = q ca. Moreover it should be observed that lim - - = q (0,0.5). Thus if a solutio to the d above equatio does ot exist it ca be cocluded that d d sol >, ( ) ( ) ( - - d ) > ( - ( - d b ) ) q h q q. C CLD C CON while if a solutio does exist tha The value of d eve for q = 0.4 (value of d is strictly mootoic with q ) is less tha 20 bytes. mi Thus if the lik-layer chael resembles a BSC the CLD scheme is better tha CON almost always. mi Lik Layer 2- state Chael: The deductio made above ca be geeralized by cosiderig a chael over which all the packet trasmissios are ot ecessarily susceptible to errors. Such a behavior is possible o time-varyig chaels ad over wireless etworks with dyamic topologies. Thus let s cosider a chael over which oly a fractio η umber of the total trasmitted packets are susceptible to errors. Whe packet trasmissios are ideed susceptible to errors, the behavior of the lik-layer chael ca be assumed to be equivalet to a BSC with ( crossover probability of θ. For such a chael it ca be show that ε = θ, δ η ( θ ) h+ d ) = ad λ = η ( θ) h. Thus it ca be show that the chael capacities of the three schemes are C C = η + θ CON ( ( ) h + d ) ( 4 ) d ( ( θ) ) h ( θ) θ h = ( η+ ( θ) ) h ( 5 ) η+ ( θ) CLD b h C CLDS (( ( ) ( h + ) η θ d ) +( θ ) h ( ( θ ) d ) ( hb ( ))) = + θ ( 6 ) It s agai quite evidet that a cross-layer approach ca provide improvemets i capacity. It should be oted that the expressios for the BSC example could be obtaied from the above equatios by settig η =. For chaels with memory the capacity is determied by the available state iformatio. If o state iformatio is available the the capacity is the least. Thus equatio (8), (9) ad (0) ca be cosidered to be lower bouds

23 o a lik-layer chaels that ca be described by a Gilbert-Elliot model [32] or some Fiite State Markov Model [34]. The actual 802.b Chael: I this sub-sectio we evaluate the performace of, ad i a 3-hop sceario, whe the lik-level chael at each hop is described by the Mbps traces provided at [28]. We used a header size of 48 bytes for our simulatios. The FEC parameters used i this subs-sectio are idetical to those i sectio IV, while the video ecodig parameters are idetical to those i Sectio V, with the key differece beig the use of QP=6 istead of 28. I additio we limit our experimetal setup to a sigle ru of 300 frames of the test sequece stefa. Noe of the other experimetal parameters were chaged. Table shows the degradatio i video quality for each of the RS based scheme. It ca be observed that the performace of the cross-layer schemes is sigificatly better tha the covetioal oe. As the hops icrease, the relative performace improves dramatically. If we kew the parameters δ, λ, ε for each hop the the relative performace of CON, CLD ad CLDS over 802.b could have bee predicted. The average values of δ, λ, ε for each hop are listed i Table. I sectio III it was show that i a multi-hop sceario, the cascaded chaels ca be characterized by a sigle δ, λ, ε ad thus the performace over multiple hops ca be predicted o the basis of the cumulative values of δ, λ, ε. Thus such cumulative parameter values have also bee evaluated ad are listed i Table. It ca be see that the cumulative value of ε is always lesser tha -2%. Thus o the basis of Figure 4 ad Figure 6 it could have bee predicted that (i) the improvemet provided by over the scheme will be miimal. This is exactly what the actual PSNR values also depict. Furthermore as the value of δ is always greater tha almost 30% we should have bee able to predict that the performace of RS- CON is goig to be very poor. Fially, the combiatio of the fact that ε is small ad the differece δ λ is quite large should have eabled us to predict that the relative performace of the cross-layer schemes will be dramatically better tha the covetioal scheme. Thus the average performace of CON, CLD ad CLDS over sufficietly complex chaels ca ofte be predicted oly by evaluatig δ, λ, ε.

24 REFERENCES [] H. Balkrisha, Ph.D Dissertatio: Challeges to Reliable Data Trasport over Heterogeeous Wireless Networks, 998. [2] S. Shakkottai, T. S. Rappaport, P. C. Karlsso, "Cross-Layer Desig for Wireless Networks," IEEE Commuicatios Magazie, Volume 4, No. 0, October 2003, pp [3] M. Schaar, S. Shakar, Cross-layer optimized multimedia trasmissio over wireless etworks IEEE Iteratioal Coferece o Image Processig (ICIP) [4] M. Schaar, S. Krishamachari, S Choi, X. Xu, Adaptive Cross-Layer Protectio Strategies for Robust Scalable Video Trasmissio Over 802. WLANs, IEEE Joural o Selected Areas i Commuicatio, Volume 2, No.0, December [5] E. Setto, T. Yoo, X. Zhu, A. Goldsmith, ad B. Girod, "Cross-Layer Desig of Ad Hoc Networks for Real-Time Video Streamig," IEEE Wireless Commuicatios Magazie, vol. 2, o. 4, pp , Aug [6] L. Larzo, M. Degermark, ad S. Pik, UDP Lite for Real Time Multimedia Applicatios, IEEE Iteratioal Coferece of Commuicatios (ICC), Vacouver, Jue 999. [7] L. Larzo, M. Degermark, ad S. Pik, Efficiet Use of Wireless Badwidth for Multimedia Applicatios, IEEE MoMUC, November 999. [8] L. Larzo, M. Degermark, S. Pik, The UDP lite protocol, IETF Iteret Draft, February 200. [9] A. Sigh, A. Korad, A. D. Joseph, Performace Evaluatio of UDP Lite for Cellular Video, NOSSDAV, 200. [0] H. Zheg ad J. Boyce, A Improved UDP Protocol for Video Trasmissio Over Iteret-to-Wireless Networks, IEEE Tras. o Multimedia, vol. 3, o. 3, pp , September 200. [] H. Zheg, Optimizig Wireless Multimedia Trasmissios through Cross Layer Desig, IEEE ICME, July [2] S. A. Khayam, S. S. Karade, M. Krappel ad H. Radha, "Cross-Layer Protocol Desig for Real-Time Multimedia Applicatios over 802.b Networks," IEEE Iteratioal Coferece o Multimedia ad Expo (ICME), July [3] S. A. Khayam, S. S. Karade, H. Radha ad D. Loguiov, Performace Aalysis ad Modelig of Errors ad Losses over 802.b LANS for High-Bitrate Real-Time Multimedia, Sigal Processig: Image Commuicatio, vol. 8, Aug [4] S. A. Vastoe ad P. C. va Oorschot, A Itroductio to Error Correctig Codes with Applicatios, Kluwer Academic Publishers. [5] R. H. Morelos-Zaragoza, The Art of Error Correctig Codig, Joh-Wiley & Sos Ltd. [6] L. Rizzo, Effective Erasure Codes for Reliable Computer Commuicatio Protocols, Computer Commuicatio Review, April 997. [7] J. C. Bolot, S. Fosse-Parisis, D. Towsley, Adaptive FEC-based error cotrol for Iteret telephoy, Proceedigs of IEEE INFOCOM '99, vol. 3, pp , 999. [8] Iterative algebraic soft decisio list-decodig of Reed-Solomo codes, M. El-Khamy ad R. J. McEliece, to appear i IEEE Joural o Selected Areas i Commuicatios. [9] IEEE Trasactios o iformatio theory, Special Issue o codes o graphs ad iterative algorithms, vol. 47, Issue 2, Feb 200. [20] D. J. C. Mackay, Iformatio Theory, Iferece, ad Learig Algorithms, Cambridge Uiversity Press, [2] R. Urbake, Iterative Codig Systems, August 200. ( [22] W.E. Rya, A itroductio to LDPC codes, i CRC Hadbook for Codig ad Sigal Processig for Recodig Systems (B. Vasic, ed.), CRC Press, [23] Xiao-Yu Hu, Evagelos Eleftheriou, Dieter M. Arold, "Regular ad Irregular Progressive Edge-Growth Taer Graphs", IEEE Trasactios o Iformatio Theory, vol. 5. o., pp , [24] [25] [26] S. S. Karade, S. A. Khayam, M. Krappel, ad H. Radha, "Aalysis ad Modelig of Errors at the 802.b Lik- Layer," IEEE Iteratioal Coferece o Multimedia ad Expo (ICME), July 2003.

25 [27] S. A. Khayam ad H. Radha, Markov-based Modelig of Wireless Local Area Networks," ACM MSWiM, Sept [28] [29] S. S. Karade, U. Parrikar, K. Misra ad H. Radha, O Modelig of 802.b Residue Errors, CISS, [30] M. Wu, S. S. Karade, ad H. Radha, "Network-Embedded Chael Codig for Optimum Throughput of Multicast Packet Video," Iteratioal Packet Video Workshop (PV), December [3] P. A. Chou, Y. Wu, ad K. Jai, "Practical etwork codig," Allerto Coferece o Commuicatio, Cotrol, ad Computig, Moticello, IL, October Ivited paper. [32] T. M. Cover ad J. A Thomas, Elemets of Iformatio Theory, Wiley Series i Telecommuicatios. [33] M. Mushki ad I. Bar-David, Capacity ad Codig for the Gilbert-Elliot Chaels, IEEE Trasactios o Iformatio Theory, vol. 35, Nov [34] A. Goldsmith ad P. Varaiya, Capacity, Mutual Iformatio, ad Codig for Fiite-State Markov Chaels, IEEE Trasactio O Iformatio Theory, vol 42., No. 3, pp , 996. [35] ISO/IEC JTC /SC29/WG ad ITU-T SG6 Q.6, Draft ITU-T Recommedatio ad Fial Draft Iteratioal Stadard of Joit Video Specificatio (ITU-T Rec. H.264 ISO/IEC AVC), Doc. JVT-G050, March CK-HDR Header Iformatio (APP) DATA PAYLOAD CK- DATA Figure A sigle Logic Trasmissio Uit (LTU). -δ (-λ) (-p) δ δ? (-λ) p (-λ) p λ λ? 0 -δ 0 0 (-λ) (-p) 0 ( a ) ( b ) -δ ( Z=) δ - λ * (Z * = 0) λ λ? ( Z=?) -ε ε ε? 0 0 * (Z 0 * = 0) δ - λ -ε 0 -δ ( c ) 0 ( Z=) Figure 2 (a) Biary Erasure Chael (BEC) represetig the UDP chael (b) Hybrid Biary Symmetric/Erasure Chael (BSEC) (c) BSEC with Side iformatio Z

26 Source (ode) Destiatio (ode +) (ode m) No Cross- Layer (ode s+) Figure 3 A Hybrid Network. I the above Figure the liks outside the star-cloud are cross-layer eabled but the liks iside are ot. applicatio layer capacity CON CLD λ = 0.0 ε = 0.0 CLDS λ = 0.0 ε = 0.0 CLD λ = 0.05 ε = 0.05 CLDS λ = 0.05 ε = 0.05 applicatio layer capacity CON CLD ε = 0.0 CLDS ε = 0.0 CLD ε = 0.05 CLDS ε = δ ( a ) λ ( b ) δ = applicatio layer capacity CON CLD λ=0.0 CLDS λ=0.0 CLD λ=0.05 CLDS λ=0.05 applicatio layer capacity CON CLD ε = 0.0 CLDS ε = 0.0 CLD ε = 0.05 CLDS ε = ε ( c ) δ = umber of hops ( d ) δ = 0.06, λ = 0.0 Figure 4 Compariso of chael capacity for CON, CLD ad CLDS

27 Figure 5 ε mi as a fuctio of δ τ τ δ ( a ) λ = 0.05, ε = λ ( b ) δ = 0.33, ε = τ ε δ = 0.33, λ = 0.0 ( c ) τ hops ( d ) δ = 0.06, λ = 0.0, ε=0.0 Figure 6 Throughput performace of RS/LDPC based FEC schemes over CON/CLD/CLDS. I (a), (b), (c) the simulatios are over a sigle hop ad oly oe parameter is varied while i (d) all three parameters are maitaied costat ad performace over multiple hops is recorded.

28 average psr γ ε (a) Stefa ε (d) Stefa average psr γ ε Carphoe (b) ε Carphoe (e) average psr γ ε ε (c) (f) Paris Paris Figure 7 Impact of corruptio level ε o the video quality performace. (a), (b), (c) show the blockdistortio i terms of average PSNR. (d), (e), (f) show the motio cotiuity γ i terms of the proportio of picture frames recovered. The simulatios are over a sigle hop ad the chael parameters are maitaied costat at δ = 0.33, λ=

29 (a) Stefa: Temporal Sapshot 7.78 db 26.3 db (d) Stefa: Subjective Compariso based o Frame 08 (b) Carphoe: Temporal Sapshot 6.95 db 3.87 db (e) Carphoe: Subjective Compariso based o Frame 64 (c) Paris: Temporal Sapshot 9.2 db db (f) Paris: Subjective Compariso based o Frame 48 Figure 8 Video quality compariso betwee ad o the basis of Temporal Sapshots (a) (c) ad picture frame samples (d)-(f) for δ = 0.33, λ=0.0 ad ε = 0.07.