Quality-Driven Joint Rate and Power Adaptation for Scalable Video Transmissions over MIMO Systems

Transcription

1 Quality-Diven Joint Rate and Powe Adaptation fo Scalable Video Tansmissions ove MIMO Systems Xiang Chen, Student Membe, IEEE, Jenq-Neng Hwang, Fellow, IEEE, James A. Ritcey, Fellow, IEEE, Chung-Nan Lee, Membe, IEEE, and Fu-Ming Yeh Abstact We popose a joint ate and powe adaptation scheme to maximize the decoding quality fo scalable video coding SVC)-based video tansmissions ove multi-input multioutput MIMO) systems. The ate adaptation in ou poposed scheme includes selecting the best modulation and coding schemes MCSs), set of spatial channels, numbe of SVC layes souce coding ates) and thei coesponding application laye fowad eo coection APP-) coding ates. The powe adaptation involves the pope allocation of the powe to each antenna in the MIMO system. SVC-based video tansmissions equie unequal eo potection UEP) fo diffeent SVC layes due to the inte-laye dependency. In most of the pevious wok, the bit steam of each paticula SVC laye is allocated to one spatial channel and the UEP is achieved by tansmitting the moe impotant SVC layes though the spatial channels with highe channel gains. Howeve, in ou poposed scheme, the bit steam of each paticula SVC laye is distibuted to multiple spatial channels so that additional divesity gain can be exploited by applying APP-. The UEP can also be achieved by allocating diffeent APP- coding ate on each video laye. Moeove, tansmit powe allocation is also effectively and jointly detemined to impove the system pefomance. The effectiveness and favoable pefomance of ou poposed scheme ae shown by simulations with H. SVC taces of high definition HD) video clips ove MIMO systems. Index Tems Rate adaptation, powe allocation, scalable video coding SVC), multi-input multi-output MIMO). I. INTRODUCTION As pedicted in [], 7% of all consume mobile taffic will be mobile video taffic in 9, up fom % in. Futhemoe, mobile data taffic will exponentially incease between and 9, epesenting a 7 pecent of compound annual gowth ate CAGR), which is about thee times faste than fixed IP taffic. The development of moden Copyight c) IEEE. Pesonal use of this mateial is pemitted. Howeve, pemission to use this mateial fo any othe puposes must be obtained fom the IEEE by sending an to pubs-pemissions@ieee.og This study is conducted unde the -EC-7-A--S- poject fom the Ministy of Economic Affais MOEA) of Taiwan and Advanced Wieless Boadband System and Inte-netwoking Application Technology Development Poject of the Institute fo Infomation Industy which is subsidized by the Ministy of Economy Affais of Taiwan. Xiang Chen is with Tupl Inc., Bellevue, WA 98, USA xchen8@uw.edu). Jenq-Neng Hwang is with the Depatment of Electical Engineeing, Univesity of Washington, Seattle, WA, 989 USA hwang@uw.edu). James A. Ritcey is with the Depatment of Electical Engineeing, Univesity of Washington, Seattle, WA, 989 USA itcey@uw.edu). Chung-Nan Lee is with the Depatment of Compute Science and Engineeing, National Sun Yat-Sen Univesity, Kaohsiung, Taiwan cnlee@cse.nsysu.edu.tw). Fu-Ming Yeh is with the Boadband Wieless Depatment, Gemtek Technology Co., Ltd., Taiwan fed yeh@gemteks.com). video delivey technologies have been boosted by the apidly inceasing demand of wieless multimedia applications [] []. Nevetheless, video steaming applications ae bandwidth consuming and delay sensitive. Moeove, high packet loss ate, lage delay and jitte, caused by the bandwidth-limited and eo-pone natue of wieless channels [], esult in temendous quality degadation of eal-time video steaming sevices []. Many technologies have been developed to compensate fo the effects caused by vaying wieless channel quality. In the application APP) laye, scalable video coding SVC) [7], [8] is employed so that videos can be encoded with seveal spatial, tempoal and quality scalabilities layes), including one base laye and seveal enhancement layes, whee valid bit steams can still be fomed even when pats of the encoded bit steams highe enhancement layes) ae emoved. With moe decoded enhancement layes, the eceived video quality is impoved. But the video layes ae dependent such that the base laye and lowe enhancement layes ae equied to be successfully decoded in ode to decode highe enhancement layes. Thus, SVC can be used to accommodate vaying teminal capabilities o netwok conditions as well as satisfying the vaious needs o pefeences of end uses. Anothe widely used APP laye technology fo video steaming is application laye fowad eo coection APP-). With the APP-, data packets ae tansmitted along with additional welldesigned edundant packets. By eceiving subsets of data and edundant packets, the eceive can econstuct all of oiginal data packets even when some packets ae lost [9]. Theefoe, APP- can povide cetain coection capability without etansmissions. This is suitable fo delay-sensitive eal-time video steaming applications [9], []. In the physical PHY) laye, multi-input multi-output MIMO) technology is one of the effective solutions to suppot high quality video steaming sevices []. By taking advantage of spatial divesity and/o spatial multiplexing, MIMO can significantly impove tansmission eliability and/o spectal efficiency accodingly [], []. When beamfoming and powe allocation techniques ae applied, MIMO systems can also povide unequal eo potection UEP) on bit steams tansmitted though diffeent spatial channels []. Since SVC video layes have diffeent decoding pioities due to thei inte-laye dependency, MIMO systems ae suitable fo tansmitting SVC-based videos with UEP on each video laye []. To efficiently utilize the vaying wieless channel, it is cucial to adaptively configue the system paametes with coss-laye design so that the chaacteistics of technologies on

2 diffeent open system inteconnection OSI) layes ae jointly optimized. Vaious quality of sevice QoS) measues, such as packet loss ate, delay and jitte etc., can be used as the citeia of netwok optimization fo the geneal pupose designs of wieless tansmission systems [] [8]. Fo video steaming sevices, the ultimate goal is to impove the decoding quality at the use end. In the ecent video tansmission eseach, tansmission and video coding paametes ae adjusted based on well-designed objective video quality measues, which have high coelation with uses subjective peceptual satisfaction. Many studies have been conducted in ate and powe allocation fo video steaming applications. In [9], an oppotunistic layeed multicast system is poposed fo scalable video multicast, whee APP- ates and modulation and coding schemes MCS) ae jointly optimized fo bette video eceiving qualities. A quality-diven esouce allocation scheme fo ealtime video tansmissions is poposed in [], whee the souce coding ate and tansmission data ate ae jointly consideed. The authos in [] poposed a scheme fo wieless video chat applications, whee the system tansmission paametes ae adjusted based on a powe-ate-distotion model. Moeove, many studies have been conducted in tansmitting scalable videos ove MIMO systems. In most of the pevious wok, diffeent UEP schemes ae developed based on the dependency stuctue of SVC video layes []. In [], authos poposed an adaptive channel selection ACS) scheme whee the UEP is achieved by tansmitting highe pioity bit steams i.e., base laye and lowe enhancement layes) though spatial channels with highe channel qualities. The advantage of this scheme is that it is simple, effective, and only equies patial channel infomation. UEP can also be achieved by powe allocation [], [] []. With full channel knowledge, the tansmit powe is adjusted based on empiical design [] o optimized fo SVC video decoding quality [], []. With a well-designed tansmit powe allocation, the SVC video decoding quality at the use end is futhe impoved compaed with ACS []. In [], an unequal powe allocation scheme is developed fo SVC video tansmissions ove massive-mimo system, whee multiple eceives ae consideed. In [], an optimal powe allocation scheme is poposed fo tansmitting multiple SVC-based video steams to multiple end uses though MU-MIMO. A esouce allocation scheme is poposed in [7], whee powe, modulation and bit-level coding ae adjusted accoding to video distotions. In the above wok, each paticula SVC video laye is tansmitted though a single spatial channel and the numbe of video layes is esticted by the numbe of available spatial channels. This limits the flexibility of the system. Futhemoe, the applied in the above can only exploit the divesity gain in time domain since the bit steam of each SVC video laye is not inteleaved to multiple spatial channels. In ou pevious wok, we poposed a novel MIMO tansmission scheme fo SVC-based video steaming in such a way that each bit steam of a paticula SVC video laye can be spead ove multiple spatial channels [8]. By applying, divesity in both the time and spatial domains can be exploited. The UEP of each video laye is achieved by diffeent coding ates, which is optimized based on video decoding quality. Moeove, the numbe of tansmitted video layes and the spatial channels ae jointly selected to bette utilize the wieless esouces. We demonstate that when the spatial divesity gain can be exploited, the system pefomance with equal tansmit powe allocation can even outpefom the ACS scheme with nea-optimal powe allocation poposed in [], []. In this pape, we futhe extend ou pevious wok in [8] in the following aspects: Fist, the adaptive modulation and coding AMC) scheme is consideed so that the system is moe flexible in diffeent channel conditions. Second, since the bit steam of each video laye is spead ove multiple spatial channels, the powe allocation schemes poposed in [], [] ae no longe valid. In this wok, we popose a novel powe allocation scheme, which is specifically suitable in ou poposed tansmission scheme. Thid, we apply APP- to exploit the time-space divesity instead of using PHY- as in [8]. By using APP-, the poposed system can be moe compatible with cuent wieless communication standads. Fouth, high-definition HD) videos, which have diffeent quality pefomance equiements than those of lowe esolution videos, ae used in the simulations. This pape is oganized as follows. System oveviews, including SVC-based video coding, APP- and MIMO systems, ae intoduced in Section II. In Section III, poblem fomulations, including SVC video laye selection, powe allocation, adaptive APP- ate contol, AMC and spatial channel selection, ae descibed. Poblem solving stategies and detailed algoithms ae poposed in Section IV. Simulation esults and concluding emaks ae given in Section V and VI espectively. Notations: Uppe lowe) boldface lettes ae used fo matices column vectos). E [.] denotes the expectation. diag h) is a diagonal matix with the elements of h sitting on the diagonal. N is N unit vecto. I N denotes the N N identity matix..) T denotes the tanspose..) H means the Hemitian.. epesents the ceil.. is the floo. The culed inequality symbol e.g., ) epesents the component-wise inequality. A. SVC-Based Videos II. SYSTEM OVERVIEW As illustated in Fig., SVC can encode a video sequence into up to L max layes, including one base laye and L max enhancement layes with cetain dependencies [7], [8], []. The base laye is the most impotant laye because it is mandatoy to decode the whole video sequence. All the enhancement layes ely on the base laye. And a less impotant highe-layeed) enhancement laye elies on moe impotant lowe-layeed) enhancement layes. The moe successfully decoded enhancement layes, the bette the econstucted video qualities. Due to the inte-laye dependency chaacteistic of SVC-based video, in wieless video tansmission, the base laye and moe impotant enhancement layes equie stonge potection compaed to the less impotant layes. Moeove, SVC can suppot all the tempoal fame ates), spatial pictue esolutions) and quality image fidelity) scalabilities. In this

3 Oiginal Video Laye Base Laye) Low Quality Moe Impotant Laye SVC Laye L max High Quality Less Impotant whee P PLR) is the packet loss ate PLR). One of the wellknown APP- schemes is the Reed-Solomon RS) code, which opeates on non-binay symbols and has ideal coection capability. Fo the most commonly used RS code with 8 bits pe symbol, at most souce packets can be encoded, i.e., K N, which limits the paametes selection in pactice. Moeove, the RS code has high decoding complexity due to non-binay opeations [], which is not attactive fo HD video steaming applications. The Rapto code [] is a moe attactive solution fo HD video steaming sevices due to the flexible choices of paametes and linea decoding cost. Unlike RS code, the coection capability of a Rapto code is t = N + ɛ)k, whee ɛ is eception ovehead efficiency, which makes the coection capability sub-optimal. Howeve, the eception ovehead efficiency of standadized Rapto code is close to ideal []. In this pape, we only conside the APP- scheme with ideal coection capability. Fig.. Concept of SVC-based video. pape, without loss of geneality, the poposed scheme is designed only based on quality scalability. B. Utility and Video Decoding Quality Measue The poposed scheme maximizes the oveall video decoding quality at the eceive, which can be measued in utility values [9]. Much eseach has been conducted in objective video quality assessment VQA) methods to quantify the video decoding quality. Among these VQA schemes, the Multi-Scale-Stuctual SIMilaity index MS-SSIM) developed in [9], oiginally designed as an image quality assessment method, has been epoted as a simple but effective VQA scheme, which has high coelation with subjectively peceived video quality [] []. In VQA applications, the MS-SSIM is applied fame-by-fame to the luminance component of the video and the oveall MS-SSIM index can be calculated by aveaging the fame-level quality scoe []. Theefoe, we adopt the aveage MS-SSIM as the utility value of each encoded SVC video laye. C. APP- APP- can povide eliable end-to-end steaming applications with packet-level potection []. A souce block consisting of K packets is encoded into an block with N N K) packets so that N K edundancy packets ae constucted and the encoding ate is K/N. Fo an ideal APP- scheme, the decode can econstuct the oiginal K souce packets fom any K out of N eceived packets [] with coection capability t = N K. If the packet losses ae independent and identical distibuted i.i.d.), the APP- block coection ate BCR) of an N, K) APP- code can be expessed as [9], [], []: t ) N P BCR) t) = P PLR)) i P PLR)) N i, ) i i= D. MIMO System Model The input-output elation of an M M t MIMO system can be expessed by the following linea equation: y = Hx + n, ) whee y is an M eceive complex symbol vecto. H is an M M t channel matix in which all elements ae i.i.d. zeo mean ciculaly symmetic complex Gaussian ZMCSCG) andom vaiables with zeo mean and unit vaiance, i.e., CN {, }. x is an M t tansmit complex symbol vecto with E [ xx H] = diag p) = diag p, p,..., p Mt ). Hee, p is the M t eal non-negative tansmission powe vecto subject to the nomalized powe constain: p T Mt =. n is an M i.i.d. complex additive white Gaussian noise AWGN) noise vecto with covaiance matix N I M. Theefoe, the system signal-to-noise atio SNR) is ρ = /N. With a singula value decomposition SVD), the known MIMO channel matix H can be decomposed as: H = UDV H, ) whee U and V ae unitay matices. The diagonal matix D is: [ D = diag λ, λ,..., ] λ R,,...,, ) whee R min M, M t ) is the ank of channel matix H. λ λ... λ R ae the eigenvalues of H H H. If accuate and full channel knowledge is available at both tansmitte and eceive sides, a pecode V and a decode U H can be appended on the tansmitte and eceive side espectively so that the MIMO input-output elation can be educed as: ỹ = U H HV + U H n = Dx + ñ. ) Thus, the MIMO system can be decomposed into R independent single-input single-output SISO) channels. The SNR of the th SISO channel is: SNR = ρλ p.

4 SVC Decode & Video Fame Reconstuction APP- Decode PHY- Decode Demodulation Decode & Detection Channel Estimation Wieless Channel Video Laye Selection Packetization APP- Encode Channel Selection PHY- Encode Modulation Powe Allocation Pecode Base Laye st Enhance Utilities L max -) th Enhance Bit-ates L max L L Contol channel Coss laye info. Packet sizes Joint Rate and Powe Adaptation Scheduling info. CSIs Base Laye st Enhance L max -) th Enhance Fig.. Poposed system stuctue fo tansmitting SVC-based videos to the intended eceive. Type m) TABLE I COEFFICIENTS FOR DIFFERENT MCSS [] Modulation Code Rate a m b m Spectal Efficiency c m) BPSK / b/hz QPSK /.. b/hz QPSK /.97.. b/hz QAM 9/.8.. b/hz QAM /.9.8 b/hz QAM / b/hz E. Modulation and Coding Schemes MCSs) In eal wold wieless sevices, diffeent MCSs ae used fo vaying wieless channel conditions. In this pape, we adopt six widely used MCS schemes and thei bit eo ate BER) expessions can be appoximated by []: P BER) SNR; m) a m exp b m SNR), ) whee a m and b m ae coefficients coesponding to the MCS m. The coefficients of all the six MCSs ae listed in Table I. F. Poposed System Stuctue The system stuctue of ou poposed coss-laye joint ate and powe allocation scheme is shown in Fig.. The souce video clip is encoded into L max SVC layes, including one base laye and L max enhancement layes. The souce data ate of the l th SVC laye is R S) l. A video laye selection module chooses a suitable numbe of video layes, i.e., L L max to be tansmitted. Then, the bit steams of selected video layes ae packetized with fixed packet length S, which is chosen to be S = 8 bits in this pape. The souce packets ae then fed into the APP- encode evey T s second to fom an block. T s is assumed to be ms in this wok. To fit the souce data ate of video laye l, the numbe of souce packets pe block is appoximately K l = R S) l T s /S. A channel selection module detemines a set of spatial channels S R to be used fo tansmission, whee R = {,,..., R} denotes the set of indices of available spatial channels and S R. Moeove, as shown in Fig., the channel selection module inteleaves the packets of each video laye to the selected multiple spatial channels, instead of allocating the packets of moe impotant video layes to the channels with highe channel gains as in the pevious wok illustated in Fig. ). Thee ae two advantages of the poposed channel selection method. Fist, the numbe of video layes is not esticted by the numbe of spatial channels because of inteleaving. Second, by inteleaving the packets of each video laye to seveal spatial channels, the divesity gain of both time and space domains can both be exploited with APP-. In the poposed system, packet allocation is based on the thoughput of each spatial channel, which is detemined by the coesponding MCS used fo tansmission. Fo instance, if the thoughput of a spatial channel is R T), S, and in total N l) encoded packets of a video laye l needs to be tansmitted, the numbe of packets of the video laye l

5 Laye Laye Laye Laye Laye Laye Laye Laye Laye Laye λ Fig.. Concept of packet steam allocation in the pevious wok. The moe impotant packet steam is tansmitted though the spatial channels with highe gains. Fig.. Concept of packet steam inteleaving pefomed in the poposed channel selection module. The packet steam of each video laye is inteleaved to multiple spatial channels. tansmitted though the spatial channel is appoximately N,l λ N l) R T) λ λ k S RT) k λ λ λ λ. 7) The exact numbe of packets can be calculated by the algoithm descibed in Section IV. The modulated symbols ae tansmitted though the wieless channel afte powe allocation and pecoding. All the tansmission paametes, including numbe of SVC layes souce coding ates), APP- ate of each video laye, MCS of each spatial channel and its coesponding tansmit powe, ae jointly optimized by the poposed joint ate and powe adaptation scheme based on coss-laye infomation at the tansmitte side and channel state infomation CSI) fom the eceive side. At the eceive, the channel estimation module estimates the channel and feeds back the CSI. In this wok, we assume pefect and full channel knowledge at both tansmitte and eceive sides. Afte MIMO decoding, detection, demodulation, PHY- decoding and APP- decoding, the eceived bit steams ae saved in eceiving buffes and fed into the SVC decode. The bit steam of the video laye l is dopped if a tansmission eo is detected o its dependent layes i.e., fom the base laye to the l th) laye) ae not coectly decoded. We define ou optimization poblem as: Q : max L,m,t,p L l= subject to p, R p =, = u l P V) l m, t, p) t, L R t l +K l = T s R T) m ) /S, l= = whee L is the numbe of SVC video layes selected fo tansmissions. t is an L vecto and its l th element t l [t] l denotes the APP- coection capability of the SVC laye l. m [m, m,..., m R ] T, m =,,...,, denote the MCSs of the th spatial channel with coesponding spectal efficiency c m in Table I. Hee m = indicates that the th channel is not selected fo tansmission i.e., / S). p is an R vecto and its th element p [p] is the tansmit powe allocated on the th channel. u l is the utility value associated with the l th SVC laye. R T) m ) is the thoughput of spatial channel, which is a function of the coesponding chosen MCS. P V) l m, t, p) is the pobability of the successfully decoding SVC laye l, and can be epesented as: P V) l m, t, p) = l j= 8) P BCR) j m, t j, p)) b, 9) whee b denotes the numbe of APP- blocks in each goup of pictues GoP) time peiod. P BCR) j m, t j, p) is the APP- block coection ate of the j th SVC laye. If we define X,j as a andom vaiable of having eoneous packets in one APP- block of the SVC laye j tansmitted though spatial channel, it follows the Binomial distibution and its pobability mass function PMF) is: ) N,j ) x ) N,j x f X,j x) = P PLR) P PLR), ) x whee the PLR of the th spatial channel is given by: ) S P PLR) = P BER) p, m ) = a m exp b m ρλ p )) S. ) Thus, the APP- block coection ate in 9) can be epesented as: P BCR) j m, t j, p) = t j k= f Yj k), ) whee f Yj y) is the PMF of andom vaiable Y j and Y j = S X,j. ) III. PROBLEM FORMULATION The video decoding quality at the eceive side can be optimized by maximizing the aveage utility of the system. Theefoe, f Yj y) can be deived by calculating the convolution sums of f X,j x). In 8), the fist constaint means that the tansmit powe allocated on each spatial channel is non-negative. The second constaint means that the sum of powe is limited by a unit powe. The thid constaint means

6 Q, : up to laye is tansmitted ove, solve Eq. ) Q, : up to laye is tansmitted ove, solve Eq. ), L, m, t, p, * *,, L, m, t, p, * *, Inputs Estimate spatial channel weightings Sub-poblem Q L,C Powe allocation APP- ate adaptation Packet mapping C L, m, t, p L, C * * L, C L, C Q Q, : up to laye is tansmitted ove, solve Eq. ) Q Lmax,C : up to laye L max is tansmitted ove C, solve Eq. ), L, m, t, p * *,, L Lmax, m C, t, p L, max * *, C, C Lmax Lmax Choose the one with lagest utility value L, m, t, p * * * * Fig.. Poposed modules fo each sub-poblem. Figue shows the poposed modules fo each sub-poblem. A weighting vecto is estimated fist. Then, the estimated weightings and necessay paametes ae fed into the powe allocation module and the packet mapping module. The powe allocation module calculates the optimal powe and is used by the APP- ate adaptation module, which also includes packet mapping inteleaving) to each spatial channel. Fig.. Poblem solving stategies: decomposing the oiginal poblem into seveal sub-poblems. that the eo coection capability of each APP- block is non-negative. The last constaint sets the total numbe of tansmitted packets in each sampling peiod of an APP- encoding block i.e., T s seconds) equal to the total numbe of packets that the netwok can tansmit in this time peiod. A. Sub-Poblems IV. PROPOSED ALGORITHMS To educe the complexity of the oiginal optimization poblem in 8), we define the maximum SVC layes to be L, with the MCS vecto being m, and decompose the oiginal poblem into seveal sub-poblems: Q L,C : max t,p P V) L t, p; m C)) subject to p, R p =, = t, L R Ts R T) t l +K l = S l= = m C) ), ) whee m C) is the C th possible combination of the MCS vecto { m. Each sub-poblem } povides one candidate solution set L, m C), t L,C, p L,C. Note that fo each sub-poblem, the paametes L and m C) ae pe-detemined. The solution of the oiginal poblem can be obtained by seaching the candidate solution set with highest utility, i.e., L {L, m, t, p }=ag max u l P V) l t L,C, p L,C;m C)), ) C l= whee C denotes the set containing all L max C candidate solutions. The poblem solving stategies including the elationships between the oiginal poblem, the sub-poblems, the candidate solution set and the final solution ae illustated in Fig.. B. Spatial Channel Weightings and Packet Mapping To exploit the divesity gain of diffeent spatial channels, the APP- encoded packets of each SVC laye ae mapped to multiple spatial channels. As indicated in 7), the numbe of packets of the l th SVC laye ove the th channel i.e., N,l ) is popotional to the nomalized thoughput of the th channel. Let N be an R L packet mapping matix with elements N,l. Let w be an R vecto with each element w = R T) k S RT) k ) epesenting the ate distibution taget weighting) of the th spatial channel. The thoughput of the th channel is R T) = Bc m. 7) Table II shows the packet mapping algoithm to calculate N. The poposed packet mapping algoithm iteatively allocates one packet to the spatial channel whose weighting w temp), calculated in line of Table II, is the smallest compaing to its coesponding taget weighting w. And the packet numbe distibution fo each spatial channel appoaches to the taget weighting w fo each iteation. Inputs: Output: TABLE II ALGORITHM FOR PACKET MAPPING Weighting vecto: w Packet numbe: N l) fo each l =,,..., L Packet mapping matix: N. Initialization: N,l := fo all and l; N, := ;. fo l = L. while N l) >. w temp) N,l fo each S; := l. := ag max. N,l := N,l + ; 7. N l) := N l) ; 8. end while 9. end fo. etun N; N,l /,l ) w w temp) fo each S;

7 7 C. Powe Allocation As indicated in ), the goal of powe allocation is to maximize the successful decoding pobability P V) L t, p; m C)), which is equivalent to maximizing the APP- block coection ate P BCR) j t j, p; m) in 9). When N,j is lage and P PLR) is small, the PMF of X,j can be appoximated by the Poisson distibution with paamete γ X,j = N,j P PLR). Theefoe, Y j in ) also follows Poisson distibution with paamete [7] γ Yj p)= S N,j P PLR) p ) N j) w P PLR) p ). 8) S If the MCS vecto m and the APP- coection capability t ae given, the APP- block coection ate P BCR) j p; m, t j ) is maximized when γ Yj p) is minimized. Theefoe the powe allocation poblem can be witten as Q PA) L,C : p L,C = ag min J p; m C)) p subject to p, 9) R p =, = whee J p; m C)) = S C) w m C) ) P PLR) p ) ) is the weighted sum of packet loss ate of each spatial channel. When P BER) is small, P PLR) in ) can be appoximated by P PLR) p ; m ) S a m exp b m ρλ p ), ) which is convex function with espect to p. Theefoe, the objective function in ) can also be appoximated as a convex function since the non-negative sum of convex functions is still convex [8]. The powe allocation poblem in 9) can thus be efficiently solved by ) when the appoximation is valid see Appendix A fo deivations). [ p L,C ] whee = µ + log Sw a m b m ρλ ), b m ρλ S, / S logsw a m b m ρλ ) b m ρλ µ = S S /b m ρλ ) ). ) Howeve, when the appoximation is not valid, i.e., P BER) is not small enough, moe powe than necessay is allocated on the channels with low channel gains, esulting in wasted powe. Let J appx p) denotes the objective function in 9) with the appoximation in ), the powe allocation solution is tustable if J p L,C; m C)) J appx p L,C; m C)) is less than a theshold T h, which is empiically set as.. Othewise, the channel with smallest gain is emoved fom the selected channel set S and its coesponding powe is set as. Then, p L,C is ecalculated by ) until the solution is tustable o only one channel is left in S, i.e., S =. The algoithm fo powe allocation is descibed in Table III. TABLE III ALGORITHM FOR POWER ALLOCATION Inputs: Channel MCS vecto: m C) Channel gain: ρλ fo each =,,..., R Output: Powe allocation vecto: p L,C. if S [ = ; ]. p L,C := ; [ ]. p L,C := fo =,,..., R;. else. Solve p L,C by );. if J p L,C ; mc)) J appx p L,C ; mc)) T h 7. min) := ag min λ ; { 8. S := S min)} ; [ ] 9. p L,C := ; min). Go back to step ;. end if. end if. etun p L,C ; D. APP- Rate Adaptation Since the tansmitted packets of each SVC laye ae spead ove multiple spatial channels, the channel conditions expeienced by each SVC laye ae simila. Theefoe, the UEP cannot be achieved by channel selection as in []. Instead, unde the oveall allowed packet numbe constaint, detemined by the system thoughput, the UEP can be achieved by assigning diffeent APP- ates on each SVC laye. This esults in the following APP- ate adaptation poblem: Q ) L,C : t L,C = ag max t subject to t, L R t l + K l = l= = P V) L T s R T) t; m C), p L,C m C) ) ) /S. ) Note that the APP- ate adaptation is conducted afte the optimal powe allocation vecto p L,C is obtained. Theefoe, all the paametes L, m C), and p L,C ae detemined and the only vaiable left is t. In this pape, a simple steepest ascent algoithm, shown in Table IV, is poposed to find the optimal solution of the APP- ate of each SVC laye. Note that the objective function of ), P V) L t) is non-deceasing with espect to the coection capabilities t. Theefoe, the steepest ascent algoithm can each the global optimal solution. The idea of the poposed APP- ate adaptation algoithm iteatively adds one APP- coection capability to the SVC laye with the highest gain in tems of video decoding coection pobability. The iteation stops when no additional edundancy packets can be added due to the thoughput limitation. E. Oveall Rate and Powe Allocation Adaptation Algoithm The oveall poposed joint ate and powe allocation adaptation algoithm, which combines all the above mentioned algoithms, is descibed in Table V.

8 8 TABLE IV ALGORITHM FOR APP- RATE ADAPTATION Inputs: Numbe of APP- blocks: b Numbe of souce packets: K l fo each l =,,..., L Channel MCS vecto: m C) System bandwidth: B APP- block sampling time: T s APP- packet size: S Channel SNR gains: ρλ fo each =,,..., R Optimal powe allocations: p fo each =,,..., R Output: APP- coection capability vecto: t L,C. Initialization: t := ; L R ). while + t l + K l T sr T) m C) /S l= =. fo j = L. Do packet mapping algoithm and obtain N;. t temp) := t;. t temp j := t j + ; 7. Calculate P BCR) j by ); ) 8. G j t j ) := P V) L t temp ; m C), p L,C ) P V) L t; m C), p L,C ; 9. end fo. j := ag max G j t j );. t j := t j + ;. end while. etun t L,C := t; TABLE V OVERALL ALGORITHM FOR RATE AND POWER ADAPTATION Inputs: Numbe of APP- blocks: b Maximum SVC laye: L max Souce packet numbe: K l fo each l =,,..., L max Possible MCSs table Utility vecto: u System bandwidth: B APP- block sampling time: T s APP- packet size: S Channel gain: ρλ fo each =,,..., R Output: Optimal solution set: {L, m, t, p }. Initialization: utility max) := ;. fo each MCS combination C. Calculate weighting vecto w by );. Do powe allocation and obtain p L,C ;. fo L = L max. Do APP- ate adaptation and obatin t L,C L 7. utility := u l P V) l l= 8. if utility > utility max) 9. utility max) := utility;. {L, m, t, p } :=. end if. end fo. end fo. etun {L, m, t, p }; m C), t L,C, p L,C ) ; { L, m C), t L,C, p L,C } ; TABLE VI VIDEO CODING AND TRANSMISSION PARAMETERS Cactus Kimono GoP Size Fames) 8 8 Fame Patten IBBBBBBB IBBBBBBB Total SVC Layes QP of SVC Laye Base Laye) QP of SVC Laye QP of SVC Laye QP of SVC Laye Total Encoded Fames 97 Fame Rate fps) System Bandwidth MHz) Coheence Time ms) TABLE VII CUMULATIVE BIT RATE OF DIFFERENT VIDEO LAYERS Video Laye Laye Laye Laye Cactus.87 Mbps.89 Mbps.79 Mbps 9.9 Mbps Kimono.7 Mbps.889 Mbps.8 Mbps.9 Mbps V. SIMULATION RESULTS The effectiveness and favoable pefomance of ou poposed algoithm ae evaluated by intensive simulations. Two high definition HD) video clips Cactus and Kimono ae encoded by the Joint Scalable Video Model JSVM) vesion 9.9 [9]. The detailed video coding and tansmission paametes ae listed in Table VI. Both videos ae encoded with medium-gain scalability MGS). Both motion compensation and estimation ae constained at the cuent laye []. The cumulative souce coding bit ates of the SVC layes ae listed in Table VII. The encoded netwok abstaction laye units NALUs) ae packetized into packets with bytes. A MIMO system is used fo tansmissions. The channel matix changes andomly evey channel coheence time. And the CSIs with full channel knowledge ae fed back evey channel coheence time. The sampling peiod T s of APP- block is ms. At the eceive side, contol messages such as video coding paametes and MCS infomation ae assumed to be coectly eceived. Pefect eo detection is assumed and eoneous packets ae dopped. The undecodable NALU, which is caused by eithe packet loss of its own SVC laye packets o unsatisfied SVC laye dependencies, is discaded befoe SVC decoding. In total diffeent schemes ae simulated. Thei coesponding techniques and abbeviations ae listed in Table VIII. Note that the scheme, which uses APP- to exploit the divesity of diffeent spatial channels, is simila to ou pevious wok in [8]. The scheme in Table VIII efes to a simila system poposed in [], whee the bit steam of the l th SVC laye is tansmitted though the spatial channel with the l th highest channel gain. In this case, to meet the bit ate equiement of each SVC laye, the MCS of tansmitting the SVC laye base laye),, and ae set as: QPSK

9 9 Fame index Fig. 7. SVC laye indices of eceived fames. Cactus, SNR: db. /, QPSK /, -QAM 9/ and -QAM / espectively. Also, the same MCSs ae applied fo all the schemes without AMC fo fai compaisons. The scheme is simila to the wok in [], whee the powe allocation algoithm is applied on the system. Figue 7 shows the eceived SVC laye index of each decoded video fame of Cactus when the system SNR is db. It is obvious that the poposed packet mapping and ate adaptation scheme i.e., scheme) has moe video fames with highe enhancement layes decoded. This is because the poposed scheme can bette exploit the multi-channel divesity gain. If the AMC scheme is applied i.e., scheme), the pefomance of the poposed scheme can be futhe impoved since the best MCSs ae chosen accoding to diffeent channel qualities. The best system pefomance is achieved when both AMC and powe allocation ae applied with ate adaptation i.e., scheme). The cumulative distibution functions CDF) of the powe allocated on each spatial channel ae plotted in Fig. 8. The scheme allocates powe based on the fact Abbeviations [8] TABLE VIII PROPOSED AND CONTROL GROUP SCHEMES Techniques Poposed algoithm in Table V Poposed algoithm in Table V with equal powe allocation on each spatial channel. Poposed algoithm in Table V with fixed MCS on each spatial channel. Poposed algoithm in Table V with fixed MCS and equal powe allocation on each spatial channel. [] Simila system as in [] [] Simila system as in [] CDF CDF CDF Channel Channel Channel Channel Powe Channel Channel Channel Channel Powe Channel Channel Channel Channel Powe Fig. 8. CDF of powe on each spatial channel. Cactus, SNR: db. MS SSIM Fame Index Fig. 9. Pe-fame MS-SSIM indices of video. Cactus, SNR: db. that each SVC video laye is tansmitted though a single spatial channel. While both and schemes allocates powe based on the algoithm in Table III, which minimizes the weighted sum of packet loss ate. Since the scheme adoptes diffeent MCSs in each coheence time, its powe allocation esult is diffeent to the scheme. The pe-fame MS-SSIM and PSNR of decoded video Cactus at system SNR db ae plotted in Fig. 9 and Fig.

10 Channel Channel P 8 Aveage Spectal Efficiency Aveage Spectal Efficiency Fame Index Fig.. Pe-fame PSNR indices of video. Cactus, SNR: db. Aveage Spectal Efficiency Channel Aveage Spectal Efficiency Channel Aveage powe.8.. Channel Aveage powe.8.. Channel Fig.. Aveage spectal efficiency on each spatial channel. Cactus. Video laye Video laye.. Aveage powe Channel Aveage powe Channel Aveage coding ate Aveage coding ate Video laye Video laye Fig.. Aveage powe allocated on each spatial channel. Cactus. espectively. The effectiveness of ou poposed algoithms ae clealy demonstated by compaing with and methods. The aveage powe and the aveage spectal efficiency on each spatial channel unde diffeent channel conditions ae illustated in Fig. and Fig. espectively. Note that the sepctal efficiency depends on both the MCSs and the spatial channel selections. If a specific spatial channel is not selected, its spectal efficiency is. Also, the spectal efficiency can not eflect the successful decoding ate at the APP laye. Fo instance, conside the scheme, the spectal efficiencies of the d and th channel ae highe than the othe schemes. Howeve, due to lack of potection, the tansmitted packets of these two channels cannot be successfully decoded. Figue shows the aveage APP- coding ates of each SVC video laye. The decoded videos of Cactus of, and schemes at system SNR db ae available fo eviews at []. One can obseve that the video decoding quality is the best with method. Aveage coding ate Aveage coding ate Fig.. Aveage APP- coding ate of each video laye. Cactus. Figue shows the aveage MS-SSIM index and the aveage PSNR at diffeent aveage channel SNRs. Appaently, without AMC and powe allocation, ou poposed scheme i.e., scheme) still outpefoms and schemes. Howeve, with fixed MCSs, the system thoughput is limited and the system pefomance cannot be futhe impoved even when channel quality becomes bette. If the AMC scheme is applied i.e., scheme), decoded video quality becomes bette when channel quality impoves. Powe allocation can povide additional decoding

11 MS SSIM P CDF. Channel Channel Channel Channel Powe CDF. Fig.. Aveage MS-SSIM index and PSNR of econstucted video Cactus. Channel Channel Channel Channel CDF Powe. Channel Channel Channel Channel Powe Fame index Fig.. CDF of powe on each spatial channel. Kimono, SNR: db. MS SSIM Fig.. SVC laye indices of eceived fames. Kimono, SNR: db..9 quality gain as shown in the scheme. Simila esults can be demonstated by the video clip Kimono. Figue shows the SVC quality laye indices of the econstucted video when system SNR is db. The CDFs of the powe allocated on each spatial channel is shown in Fig.. The coesponding MS-SSIM and PSNR plots ae shown in Fig. 7 and Fig. 8 espectively. Clealy, the poposed methods outpefom the othes. The aveage powe and spectal efficiency of each spatial channel ae plotted in Fig. 9 and Fig. espectively. The aveage APP- coding ates of each SVC laye ae plotted in Fig.. The visual compaison of the decoded fame of, and methods at system SNR db ae available at []. Since can decoded moe enhancement layes than the othes, its video quality is the best. The aveage.9 Fame Index Fig. 7. Pe-fame MS-SSIM indices of video. Kimono, SNR: db. MS-SSIM indices and aveage PSNR at diffeent system SNR ae plotted in Fig.. Still, the poposed ate adaptation and powe allocation scheme i.e., ) has the best pefomance among the othe schemes. VI. CONCLUSION A quality-diven joint ate and powe adaptation scheme fo scalable video tansmissions ove MIMO systems is poposed. Unlike pevious wok, which achieve UEP by tansmitting highe pioity bit steams though spatial channel with highe

12 P Fame Index Fig. 8. Pe-fame PSNR indices of video. Kimono, SNR: db. Aveage Spectal Efficiency Aveage Spectal Efficiency Channel Channel Aveage Spectal Efficiency Aveage Spectal Efficiency Channel Channel.8 Channel.8 Channel Fig.. Aveage spectal efficiency on each spatial channel. Kimono. Aveage powe... Channel Aveage powe... Channel Aveage coding ate.8.. Video laye Aveage coding ate.8.. Video laye Aveage powe.. Aveage powe..... Video laye. Video laye Fig. 9. Aveage powe allocated on each spatial channel. Kimono. Aveage coding ate.8... Aveage coding ate.8... gains, we popose to inteleave the bit steam of each SVC laye to multiple spatial channels and use APP- to exploit both time and multi-channel divesity gain. In the poposed scheme, numbe of video layes souce coding ate), APP- encoding ate fo each video laye, tansmit powe, the best spatial channel and MCS choices ae jointly optimized. The oiginal optimization poblem is decomposed into seveal sub-poblems, which can be futhe decomposed into packet mapping, powe allocation and APP- ate adaptation poblems. The encoded packets ae inteleaved based on the available thoughputs of the selected spatial channels. Optimal powe allocation is obtained by solving a convex optimization poblem with a closed-fom fomula. UEPs on diffeent video layes can be achieved by assigning diffeent APP- coding ates. Each sub-poblem geneates a candidate solution and the final optimal solution is obtained by seaching the candidate solution with best estimated decoding quality highest utility). Intensive simulations with two HD videos unde diffeent channel conditions demonstate the effectiveness and favoable pefomance of ou poposed system. Fig.. Aveage APP- coding ate of each video laye. Kimono. APPENDIX A DERIVATION OF ) AND ) Conside the optimization poblem in 9) with objective function J appx p), the coesponding Lagangian is: L p, ξ, ν) =S w a m e b M ρλp ) ξ p +ν. S S S ) whee ξ and ν ae Lagange multiplies associated with the inequality constaints and equality constaint espectively. Fo each S, the Kaush-Kuhn-Tucke KKT) conditions can be expessed as:

13 MS SSIM P Aveage MS-SSIM index and PSNR of econstucted video Ki- Fig.. mono ) Pimal feasible: p ; T p =. ) Dual feasible: ξ. ) Complementay slackness: ξ p =. ) Gadient of Lagangian vanishes: L p, ξ, ν) p p,ξ,ν = Sw a m b m ρλ e bm ρλp ξ +ν =. ) Since the optimization poblem in 9) with objective function J appx p) is convex, any point satisfying the above KKT conditions is pimal and dual optimal with zeo duality gap [8]. The above KKT conditions imply: ν Sw a m b m ρλ e bm ρλp, 7) and ) ν Sw a m b m ρλ e bm ρλp p =. 8) If p =, the appoximation J appx p) is not valid and the th channel should be emoved fom S. Thus, we only conside p > in 8). Theefoe, Let µ = log /ν ), ν = Sw a m b m ρλ e bm ρλp. 9) p = µ + log Sw a m b m ρλ ) b m ρλ, ) which is equivalent to ). Since T p =, µ + b m ρλ S S log Sw a m b m ρλ ) b m ρλ =. ) Theefoe, ) is obtained by efoming ). Note that the notation p used in this Appendix denotes the optimal solution of ), which has diffeent meaning to the notaion p used in ) and the est of this pape. REFERENCES [] Cisco Visual Netwoking Index: Foecast and Methodology, - 9,. [] J.-N. Hwang, Multimedia Netwoking: Fom Theoy to Pactice. Cambidge Univesity Pess, 9. [] X. Chen, J.-N. Hwang, C.-N. Lee, and C.-W. Hwang, An efficient CQI feedback esouce allocation scheme fo wieless video multicast sevices, in Poc. of IEEE Global Telecommunications Conf., Atlanta, GA, Decembe 9-, pp. 8. [] X. Chen, J.-N. Hwang, K.-H. Lee, and R. L. de Queioz, Qualityof-content QoC)-diven ate allocation fo video analysis in mobile suveillance netwoks, Xiamen, China, Octobe 9-. [] Y. Yang, Contibutions to smat meteing potocol design and data analytics, Ph.D. dissetation, Univesity of Washington. [] W. Hamidouche, C. Peine, Y. Pousset, and C. Olivie, A solution to efficient powe allocation fo H./SVC video tansmission ove a ealistic MIMO channel using pecode designs, Jounal of Visual Communication and Image Repesentation, vol., no., pp. 7,. [7] T. Wiegand, L. Noblet, and F. Rovati, Scalable video coding fo IPTV sevices, IEEE Tans. on Boadcasting, vol., no., pp. 7 8, 9. [8] H. Schwaz, D. Mape, and T. Wiegand, Oveview of the scalable video coding extension of the H./AVC standad, IEEE Tans. on Cicuits and Systems fo Video Technology, vol. 7, no. 9, pp., 7. [9] D. Juca, P. Fossad, and A. Jovanovic, Fowad eo coection fo multipath media steaming, IEEE Tans. on Cicuits and Systems fo Video Technology, vol. 9, no. 9, pp., 9. [] M. Luby, T. Stockhamme, and M. Watson, Application laye in IPTV sevices, IEEE Communications Magazine, vol., no., pp. 9, 8. [] A. J. Paulaj, D. A. Goe, R. U. Naba, and H. Bölcskei, An oveview of MIMO communicationsa key to gigabit wieless, Poceedings of the IEEE, vol. 9, no., pp. 98 8,. [] L. Zheng and D. N. C. Tse, Divesity and multiplexing: a fundamental tadeoff in multiple-antenna channels, IEEE Tans. on Infomation Theoy, vol. 9, no., pp. 7 9,. [] X. Chen, J.-N. Hwang, P.-H. Wu, H.-J. Su, and C.-N. Lee, Adaptive mode and modulation coding switching scheme in MIMO multicasting system, in Poc. of IEEE Intl. Symp. on Cicuits and Systems, Beijing, China, May 9-. [] X. Chen, J.-N. Hwang, C.-N. Lee, and S.-I. Chen, A nea optimal QoEdiven powe allocation scheme fo scalable video tansmissions ove MIMO systems, IEEE Jounal of Selected Topics in Signal Pocessing, vol. 9, no., pp. 7 88,. [] D. Song and C. Chen, Scalable H. /AVC video tansmission ove MIMO wieless systems with adaptive channel selection based on patial channel infomation, IEEE Tans. on Cicuits and Systems fo Video Technology, vol. 7, no. 9, pp. 8, 7. [] Y. Yang and S. Roy, PMU deployment fo optimal state estimation pefomance, in Poc. of IEEE Global Telecommunications Conf., Anaheim, CA, Decembe -7. [7], Gouping based MAC potocols fo EV chaging data tansmission in smat meteing netwok, IEEE Jounal on Selected Aeas in Communications, vol. 9, no. 7, pp. 8,. [8], PCF scheme fo peiodic data tansmission in smat meteing netwok with cognitive adio, in Poc. of IEEE Global Telecommunications Conf., San Diego, CA, Decembe -. [9] C.-W. Huang, S.-M. Huang, P.-H. Wu, S.-J. Lin, and J.-N. Hwang, OLM: Oppotunistic layeed multicasting fo scalable IPTV ove mobile WiMAX, IEEE Tans. on Mobile Computing, vol., no., pp.,. [] P.-H. Wu, C.-W. Huang, J.-N. Hwang, J. young Pyun, and J. Zhang, Video-quality-diven esouce allocation fo eal-time suveillance video uplinking ove OFDMA-based wieless netwoks, IEEE Tans. on Vehicula Tech., vol., no. 7, pp.,. [] S.-P. Chuah, Y.-P. Tan, and Z. Chen, Rate and powe allocation fo joint coding and tansmission in wieless video chat applications, IEEE Tans. on Multimedia, vol. 7, no., pp ,. [] Y. Huo, C. Hellge, T. Wiegand, and L. Hanzo, A tutoial and eview on inte-laye coded layeed video steaming, IEEE Communications Suveys & Tutoials,. [] Q. Liu, S. Liu, and C. Chen, A novel pioitized spatial multiplexing fo MIMO wieless system with application to H. SVC video, in Poc. of IEEE Intl. Conf. on Multimedia and Expo, Suntec City, July 9-.

14 [] X. Chen, J.-N. Hwang, C.-Y. Wang, and C.-N. Lee, A nea optimal QoE-diven powe allocation scheme fo SVC-based video tansmissions ove MIMO systems, in Poc. of IEEE Intl. Conf. on Communications, Sydney, NSW, June -. [] S.-J. Kim, C.-B. Chae, and J.-S. Lee, On unequal powe allocation fo video communications using scalable video coding in massive MIMO systems, in IEEE Intenational Symposium on Multimedia ISM), Taichung, Decembe -Decembe. [] J. Pak, X. Chen, and J.-N. Hwang, Optimum powe allocation and ate adaptation fo scalable video steaming ove multi-use MIMO netwoks, in Poc. of IEEE Global Telecommunications Conf., San Diego, CA, Decembe -. [7] W. Hamidouche, C. Olivie, Y. Pousset, and C. Peine, Optimal esouce allocation fo medium gain scalable video tansmission ove MIMO channels, Jounal of Visual Communication and Image Repesentation, vol., no., pp. 7 87,. [8] X. Chen, H. Du, J.-N. Hwang, J. A. Ritcey, and C.-N. Lee, A QoEdiven ate adaptation scheme fo scalable video tansmissions ove MIMO systems, in Poc. of IEEE Intl. Conf. on Communications, London, UK,. [9] Z. Wang, E. P. Simoncelli, and A. C. Bovik, Multi-scale stuctual similaity fo image quality assessment, in Poc. of IEEE Asiloma Conf. Signals, Systems and Computes, Novembe 9-. [] S. Chikkeu, V. Sundaam, M. Reisslein, and L. J. Kaam, Objective video quality assessment methods: A classification, eview, and pefomance compaison, IEEE Tans. on Boadcasting, vol. 7, no., pp. 8,. [] A. K. Moothy, K. Seshadinathan, R. Soundaaajan, and A. C. Bovik, Wieless video quality assessment: A study of subjective scoes and objective algoithms, IEEE Tans. on Cicuits and Systems fo Video Technology, vol., no., pp ,. [] M. Vanješ, S. Rimac-Dlje, and K. Ggić, Review of objective video quality metics and pefomance compaison using diffeent databases, Signal Pocessing: Image Communication, vol. 8, no., pp. 9,. [] X. Chen, J.-N. Hwang, C.-J. Wu, S.-R. Yang, and C.-N. Lee, A QoEbased APP laye scheduling scheme fo scalable video tansmissions ove Multi-RAT systems, in Poc. of IEEE Intl. Conf. on Communications, London, UK,. [] X. Chen, J.-N. Hwang, D. Meng, K.-H. Lee, R. L. de Queioz, and F.-M. Yeh, Quality-of-content QoC)-based joint souce and channel coding fo human detections in a mobile suveillance cloud, IEEE Tans. on Cicuits and Systems fo Video Technology. [] A. Shokollahi, Rapto codes, IEEE Tans. on Infomation Theoy, vol., no., pp. 7,. [] H. Luo, D. Wu, S. Ci, H. Shaif, and H. Tang, TFRC-based ate contol fo eal-time video steaming ove wieless multi-hop mesh netwoks, in Poc. of IEEE Intl. Conf. on Communications, Desden, June [7] H. Stak and J. Woods, Pobability, Statistics, and Random Pocesses fo Enginees. Pentice Hall,. [8] S. Boyd and L. Vandenbeghe, Convex Optimization. Cambidge Univesity Pess,. [9] The JSVM joint scalable video model) softwae [Online]. Available at [] The compaisons of decoded videos [Online]. Available at [] Y. Yang, X. Wang, and X. Cai, On the numbe of elays fo othogonalize-and-fowad elaying, in Poc. of IEEE Intl. Conf. on Wieless Communications and Signal Pocessing, Nanjing, China, Novembe 9-, pp.. [] Y. Yang and S. Roy, PMU deployment fo thee-phase optimal state estimation pefoman, in Poc. of IEEE Intl. Conf. on Smat Gid Communications, Vancouve, BC, Octobe -. Xiang Chen eceived the B.S. degee in electonic engineeing fom City Univesity of Hong Kong, Hong Kong, in 9 and the M.S. degee in electical and compute engineeing fom Univesity of Floida, Gainesville, in, and the Ph.D. degee in electical engineeing fom Univesity of Washington in. He is cuently with Tupl Inc., a statup big data company fo telecom opeatos. His eseach inteests include multimedia netwoking, wieless communication and MIMO techniques. Jenq-Neng Hwang F ) eceived the B.S. and M.S. degees, both in electical engineeing fom the National Taiwan Univesity, Taipei, Taiwan, in 98 and 98 sepaately. He then eceived his Ph.D. degee fom the Univesity of Southen Califonia. In the summe of 989, D. Hwang joined the Depatment of Electical Engineeing of the Univesity of Washington in Seattle, whee he has been pomoted to Full Pofesso since 999. He is cuently the Associate Chai fo Reseach in the EE Depatment. He has witten moe than jounal, confeence papes and book chaptes in the aeas of multimedia signal pocessing, and multimedia system integation and netwoking, including an authoed textbook on Multimedia Netwoking: fom Theoy to Pactice, published by Cambidge Univesity Pess. D. Hwang has close woking elationship with the industy on multimedia signal pocessing and multimedia netwoking. D. Hwang eceived the 99 IEEE Signal Pocessing Society s Best Jounal Pape Awad. He is a founding membe of Multimedia Signal Pocessing Technical Committee of IEEE Signal Pocessing Society and was the Society s epesentative to IEEE Neual Netwok Council fom 99 to. He is cuently a membe of Multimedia Technical Committee MMTC) of IEEE Communication Society and also a membe of Multimedia Signal Pocessing Technical Committee MMSP TC) of IEEE Signal Pocessing Society. He seved as associate editos fo IEEE T-SP, T-NN and T-CSVT, T-IP and Signal Pocessing Magazine SPM). He is cuently on the editoial boad of ETRI, IJDMB and JSPS jounals. He was the Pogam Co-Chai of ICASSP 998 and ISCAS 9. James A. Ritcey S 7-M 79-SM -F ) eceived the BSE degee fom Duke Univesity, the MSEE degee fom Syacuse Univesity, and the Ph.D. degee in electical engineeing communication theoy and systems) fom the Univesity of Califonia, San Diego. Since 98, he has been with the Depatment of Electical Engineeing at the Univesity of Washington, whee he now holds the ank of Pofesso. His eseach inteests include modulation and coding fo wieless communications, statistical signal pocessing fo undewate acoustics, and netwok secuity. Pofesso Ritcey seved as the Geneal Chai of the 99 Intenational Confeence on Communications in Seattle. He also seved as Technical Pogam Chai of the 99 and Geneal Chai of the 99 Asiloma Confeence on Signals, Systems, and Computes and is cuently a membe of the Steeing Committee. Chung-Nan Lee eceived the B.S. and the M.S. degees, both in electical engineeing, fom National Cheng Kung Univesity, Tainan, Taiwan, in 98 and 98, espectively, and the Ph.D. degee in electical engineeing fom the Univesity of Washington, Seattle, in 99. Since 99, he has been with National Sun Yat-Sen Univesity, Kaohsiung, Taiwan, whee he was an Associate Pofesso in the Depatment of Compute Science and Engineeing fom 99 to 999; he was the Chaiman of the Depatment of Compute Science and Engineeing fom August 999 to July, ; cuently, he is a Pofesso and Diecto of Cloud Computing Reseach Cente and Wieless Boadband Communication Potocol Coss-Campus Reseach Cente. His cuent eseach inteests include multimedia ove wieless netwoks, cloud computing, and evolutionay computing. Fu-Ming Yeh eceived the Ph.D. in 997 in electical engineeing fom National Taiwan Univesity. He is a CTO of Boadband Wieless Depatment at Gemtek Technology Co., Ltd. He was a deputy head at the Electonic System Reseach Division of Chung-Shan Reseach Institute of Science and Technology fom 997 to. His eseach inteests include LTE Small Cell system development, DSP system design, hadwae veification, VLSI testing, and fault-toleant computing.