732 IEEE Transactons on Consumer Electroncs, Vol. 50, No. 2, MAY 2004 Hybrd Transcodng for QoS Adaptve Vdeo-on-Demand Servces Ilhoon Shn and Kern Koh Abstract Transcodng s a core technque that s used n provdng qualty-of-servce (QoS) adaptve vdeo-on-demand servce. Many studes have examned how best to perform transcodng and reduce computaton overhead. However, the queston of when to transcode has not been adequately studed n prevous research. Ths paper addresses ths queston and presents a smple and ntellgent approach that can be used to reduce both dsk bandwdth and space requrements. Our approach determnes the optmum tme to apply transcodng by consderng the potental benefts that can be realzed. For nstance, n order to save dsk bandwdth for frequently accessed content, t pre-creates and stores multple QoS versons. On the other hand, n order to save dsk space for rarely accessed content, t stores only a sngle QoS verson and performs transcodng on the fly. The key s to fnd the optmal threshold between pre-created multple QoS versons and on-demand transcodng. We compute the optmal threshold by usng a mathematcal model. A smulaton-based experment to evaluate the effectveness of our new approach hghlghts three advantages. Frst, our method effectvely reduces both dsk bandwdth and space requrements. Second, our technque s more effcent for skewed access patterns. Thrd, the threshold computed by our mathematcal model results n mproved performance regardless of envronmental parameters. 1. Index Terms vdeo-on-demand, qualty of servce, transcodng I. INTRODUCTION Usng computer technology n the consumer electroncs ndustry s becomng more wdespread wth the penetraton of hgh technology nto all aspects of socety. Analogue TV has been evolvng nto dgtal TV, and PDAs (Personal Dgtal Assstants) and cellular phones work lke small computers. These computerzed products enable us to enoy useful servces that could be provded n the past only by the presence of computers. One example s vdeo-on-demand (VOD) servce. Nowadays, we can use VOD servces wth cellular phones, PDAs, or dgtal TV as well as personal computers. Therefore, upcomng VOD servers should support a varety of clents: cellular phones, PDAs, dgtal TV, laptops, desktop computers, etc, as seen n Fg. 1. 1 Ths research was supported by Bran Korea 21 and by the Korea Sanhak Foundaton. Fg. 1. Future vdeo-on-demand servces. The most mportant challenge n the upcomng VOD servces s that these clents have heterogeneous capabltes n processng power, storage capacty, battery lfe, and connecton bandwdth. For example, cellular phones and PDAs that are connected to wreless networks have lmted capabltes compared to desktop computers that are connected to wred networks. Therefore, VOD servers cannot provde the same hgh qualty vdeo fles to these heterogeneous clents. Rather, multple qualty-of-servce (QoS) versons n accordance wth the clent's capabltes should be provded. Transcodng s a core technque that s used n provdng multple QoS versons accordng to the clent's capabltes. Transcodng transforms exstng vdeo fles to dfferent QoS versons by changng resoluton, bt rates, or fle format. The smplest form of transcodng s to decode and re-encode an exstng fle. However, snce ths naïve method uses a consderable amount of computaton, many researchers have presented ntellgent methods that can be used to reduce computatonal requrements [1]-[8]. Ther prmary focus has been on how to perform transcodng and on how to reduce the computatonal requrements. In ths paper, we focus on when to perform transcodng, a subect that has not been adequately studed n prevous research. The problem of when to transcode s one of optmzng dsk bandwdth and space requrements; the deal s to reduce both, but exstng solutons have succeeded n reducng only one or the other. In ths paper, we present a smple and ntellgent method to acheve the approxmate deal. I. Shn and K. Koh are wth the Computer Scence and Engneerng Dvson, Seoul Natonal Unversty, Seoul, South Korea, 151-744 (e-mal: {ee kernkoh}@ oslab.snu.ac.kr). Contrbuted Paper Orgnal manuscrpt receved March 2, 2004 Revsed manuscrpt receved Aprl 9, 2004 0098 3063/04/$20.00 2004 IEEE
I. Shn and K. Koh: Hybrd Transcodng for QoS Adaptve Vdeo-on-Demand Servces 733 Exstng approaches can be categorzed as ether dynamc or transcodng. Dynamc transcodng s performed on the fly, as requred, whereas transcodng s performed n advance n antcpaton of future requests. In dynamc transcodng, a server stores only the hghest QoS verson, as transcodng a low QoS verson to a hgh QoS verson s almost mpossble 2. For requests that need a low QoS verson, the server reads the stored verson and transcodes on the fly accordng to the clent s capabltes. Dynamc transcodng requres a relatvely small amount of dsk space, as the server stores only a sngle QoS verson. However, dsk bandwdth requrements are hgh, because the server must read the hghest qualty verson even when a low qualty verson s all that s requred. In transcodng, the server pre-creates and stores multple QoS versons. When requested, the server reads and sends the approprate verson wthout addtonal transcodng. Although dsk bandwdth s utlzed more effcently, ths method wastes dsk space, as t must store multple versons. As a compromse between these two extremes, we present a hybrd scheme that saves both dsk bandwdth and space. For frequently accessed content, the hybrd scheme uses transcodng to save dsk bandwdth, whereas for nfrequently accessed content t uses dynamc transcodng to save dsk space. The goal of our hybrd scheme s to fnd the optmal threshold between transcodng and dynamc transcodng. In ths paper, we present a method that fnds the optmal threshold by usng a mathematcal model. The remander of ths paper s organzed as follows. Related work s descrbed n Secton Ⅱ. Secton Ⅲ descrbes our hybrd scheme and presents a method to fnd the optmal threshold. Secton Ⅳ nvestgates the effectveness of the hybrd scheme through a smulaton-based experment. Fnally, Secton Ⅴ states our conclusons. II. RELATED WORK Transcodng s a technque that transforms exstng multmeda content to dfferent QoS versons. It s classfed nto bt-rate reducton transcodng [1]-[3], spatal-temporal resoluton reducton transcodng [4]-[6], and dfferent format transcodng [7],[8] n accordance wth the technque used to perform the transcodng. The prmary focus of the research work noted above has been the queston of how fast the transcodng can be performed whle mantanng vdeo qualty at an acceptable level. In contrast, our focus s on reducng dsk bandwdth and space requrements. The results of the above referenced works can be easly ntegrated wth our work to produce a hghly effcent transcodng technque. Scalable codng [9],[10] s another technology that provdes QoS adaptaton. Scalable codng transforms low QoS content to hgh QoS content by usng a multple-layered approach, whereas transcodng transforms hgh QoS content to low QoS content. The problem wth scalable codng s that t needs to re-encode the exstng multmeda content to multple-layers. A more detaled explanaton of scalable codng s beyond the scope of ths paper. Kartan, Roger, and Josp compared dynamc transcodng and transcodng for mage servers [11]. However, they dd not present any ntellgent scheme to consder the advantages and dsadvantages of the two transcodng methods. They dd, however, propose the nvestgaton of such a scheme as possble future work. Lum and Lau presented a method that fnds the optmal tradeoff pont between dynamc and transcodng [12]. However, they dd not consder skewed access patterns to multmeda content and dsk bandwdth cost. III. THE HYBRID SCHEME The am of our hybrd scheme s to save both dsk bandwdth and dsk space by usng dynamc and transcodng selectvely. The followng examples descrbe typcal cases for whch our hybrd scheme can be used to select the approprate transcodng method. Cellular phone clents wth low capabltes access some content frequently. On each request, the server has to supply the approprate QoS verson for the clent s capabltes. In ths case, transcodng,.e., precreatng the correspondng QoS verson, s more benefcal than transcodng on the fly. Storng a low QoS verson requres only a small amount of dsk space, whereas the dsk bandwdth requrements for readng the hghest QoS verson would be much hgher due to the frequent accesses. Thus, our hybrd scheme selects transcodng. Some content s rarely accessed by clents wth hgh capabltes, n whch case pre-creatng hgh QoS versons s not benefcal, as such versons are generally large. In addton, the dsk bandwdth requrements for dynamc transcodng are not excessve as access s nfrequent. Thus, our hybrd scheme selects dynamc transcodng. The most sgnfcant test of our hybrd scheme s whether t can make ntellgent decsons nvolvng cases wth ambguous content that fall between the two extremes noted above. In the followng subsecton, we present a mathematcal soluton that enables our hybrd scheme to make ntellgent decsons for cases wth ambguous content. A. How to decde on dynamc transcodng vs. transcodng Our hybrd scheme determnes when to perform transcodng by evaluatng the benefts of both and dynamc transcodng. If the benefts realzed by usng transcodng are greater than the benefts of dynamc transcodng, then t selects transcodng. Otherwse, t selects dynamc transcodng. The benefts of the transcodng methods are calculated as follows. 2 The excepton s scalable codng, whch s explaned n related work.
734 We defne the multmeda content set, M, and the possble QoS versons set, τ, as (1) and (2), respectvely. τ 1,, and τ N2 denote the playback rate of each QoS verson. Note that M and τ are sorted n descendng order by popularty and by playback rate, respectvely. The arrval rate of QoS class τ s defned as λ. Μ = M, M,..., M } (1) { 1 2 N1 τ = τ, τ,..., τ } (2) { 1 2 N 2 Let us frst compute the per unt tme beneft of pre-creatng the τ verson of the M content. The request arrval rate of the τ verson of the M content, λ, s computed usng (3) from the defnton of λ. Pr() s the probablty that a clent chooses the M content. If the ncomng requests are all processed by the server, the number of concurrent sessons of the τ verson of the M content, N, s calculated usng (4) from (3), where T play () s the playback length of the M content. From (4), the per unt tme bandwdth saved by usng transcodng, SB, s computed usng (5). From (5), the per unt tme beneft of pre-creatng the τ verson of the M content, B_, s computed usng (6), where C bandwdth s the byte-prce of dsk bandwdth. λ = λ Pr( ) (3) N SB = λ T ( ) (4) play ( 1 = N τ τ ) (5) B _ C (6) = SB bandwdth The per unt tme beneft of transcodng the τ verson of the M content on the fly s calculated as follows. The amount of dsk space that s saved by transcodng the τ verson of the M content dynamcally, SS, s computed usng (7). From (7), the per unt tme beneft of dynamc transcodng, B_dynamc, s computed usng (8), where C space s the byte prce of dsk space. SS = τ T ( ) (7) play B _ dynamc C (8) = SS space The condton that our hybrd scheme selects transcodng for the τ verson of the M content s therefore descrbed as n (9). Inequalty (9) s solved as n (10) by usng (6) and (8). In concluson, for the content whose access probablty satsfes (10), our hybrd scheme selects transcodng,.e., pre-creates and stores the τ verson. In contrast, for the content whose access probablty does not satsfy (10), our hybrd scheme selects dynamc transcodng, IEEE Transactons on Consumer Electroncs, Vol. 50, No. 2, MAY 2004.e., creates the τ verson on the fly. B B _ dynamc _ (9) Pr( ) τ Cspace λ ( τ 1 τ ) C bandwdth (10) IV. EXPERIMENTAL RESULT In ths secton, we nvestgate the effectveness of our hybrd scheme by smulatng a QoS adaptve vdeo-on-demand (VOD) servce. The model server conssts of four nodes, each wth ts own IBM Deskstar 120 GXP hard dsk [13]. The prmary attrbutes of the dsk are shown n Table Ⅰ. We assumed that the dsk requests were scheduled usng a cyclebased dsk schedulng. The server contaned 200 move fles, each 100 mnutes n length. The requests arrved accordng to a Posson dstrbuton and the popularty of the moves exhbted a Zpf dstrbuton wth a parameter of 0.271. The model smulated 8 hours servce. TABLE I IBM DESKSTAR 120 GXP ATTRIBUTES Capacty 80 GB Sector sze 512 bytes Aggregate sustaned data rate 35 MB/s Aggregate rotatonal delay 4.17 ms Aggregate seek tme 8.5 ms Full seek tme 15 ms A. The number of streams servced vs. space requrements We evaluated the performance of the transcodng methods by consderng the number of streams servced and dsk space requrements. Intutvely, transcodng servces the greatest number of streams, by effcently utlzng dsk bandwdth, whle dynamc transcodng requres less dsk space. The focus of our experment was to determne whether our hybrd scheme could servce a smlar number of streams as that realzed by usng only transcodng whle at the same tme reducng dsk space requrements. We assumed that there were three QoS classes: 1.5 Mbps, 512 Kbps, and 128 Kbps playback rates. Dynamc transcodng stored only the QoS class wth the 1.5 Mbps playback rate and generated other QoS classes by transcodng on the fly. Statc transcodng precreated and stored all QoS classes. Our hybrd scheme precreated and stored content only when the beneft of transcodng was greater than the beneft of dynamc transcodng, as determned by (10). The mean nter-arrval tme of each QoS class was 20 seconds. Table Ⅱ and Fg. 2 show the results. Table Ⅱ shows the space requred by each transcodng method and Fg. 2 llustrates the number of streams servced as a functon of cycle length. The results show that our hybrd scheme servced almost the same number of streams as transcodng (Fg.
I. Shn and K. Koh: Hybrd Transcodng for QoS Adaptve Vdeo-on-Demand Servces 735 2), whle at the same tme reducng dsk space requrements by about 47 GB (Table Ⅱ ). Our hybrd scheme was able to enoy the advantages of transcodng whle at the same tme reducng ts dsadvantages. Dynamc transcodng delvered the best performance from the aspect of dsk space requrements; t consumed about 49 GB less dsk space than our hybrd scheme, a gan of about 18%. However, dynamc transcodng servced about 16-43% fewer streams than our hybrd scheme. Due to ts sgnfcant dsk bandwdth requrements, dynamc transcodng dd not prove to be a good canddate method for transcodng. TABLE Ⅱ SPACE REQUIREMENTS dynamc hybrd Space (GB) 219.73 311.28 268.43 was able to servce the same number of streams (Fg. 3). Ths means that our hybrd scheme was able to make ntellgent decsons by also consderng the skew n the popularty of multmeda content. # of streams 3500 3000 2500 2000 dynamc 5000 4500 4000 1500 0.1 0.2 0.3 0.4 zpf parameter Fg. 3. The number of streams servced as a functon of zpf parameter. 3500 # of streams 3000 2500 2000 1500 350 320 dynamc 1000 500 0 dynamc 0 5 10 15 20 cycle length (sec) space (GB) 290 260 230 Fg. 2. The number of streams servced as a functon of cycle length. B. The mpact of skew n the popularty of multmeda content The amount of skew n the popularty of multmeda content vares n accordance wth the knd of VOD servce. For nstance, a lecture streamng servce s lkely to be less skewed than a move streamng servce. In ths subsecton, we study the mpact of skew n the popularty of multmeda content on the performance of our hybrd scheme. We vared the zpf parameter from 0.1 to 0.4 and measured the space requrements and the number of streams servced. Fg. 3 llustrates the number of streams servced, and Fg. 4 llustrates the space requrements. The results show that the benefts of our hybrd scheme were greater when the request pattern was skewed. It requred less dsk space by servng more content wth dynamc transcodng when the access pattern was skewed (Fg. 4). At the same tme, regardless of the zpf parameter, t 200 0.1 0.2 0.3 0.4 zpf parameter Fg. 4. Space requrements as a functon of zpf parameter. V. CONCLUSION Ths paper addressed the queston of when to perform transcodng and presented the hybrd scheme, whch determnes the optmal tme to apply transcodng by consderng the potental benefts. Our hybrd scheme computes the benefts of and dynamc transcodng and selects the more benefcal method. A smulaton-based experment showed the followng. Frst, our hybrd scheme effectvely reduced both dsk bandwdth and space requrements. It was able to servce almost the same number of streams as that realzed by usng only transcodng, whle
736 consumng less dsk space by 45-90 GB, a gan of 14-28%. Second, our hybrd scheme was benefcal n cases wth skewed access patterns; t was able to save dsk space whle servcng the same number of streams. Thrd, the threshold computed by the mathematcal model delvered consstently good performance, regardless of the access skew and the request arrval rate. Fourth, dynamc transcodng was not a good choce for transcodng owng to ts sgnfcantly hgher dsk bandwdth requrements. REFERENCES [1] H. Sun, W. Kwok, and J. Zdepsk Archtecutures for MPEG compressed btstream scalng, IEEE Trans. Crcuts Syst. Vdeo Technol., vol.6, pp.191-199, Apr. 1996. [2] P. Assuncno and M. Ghanbar Post-processng of MPEG-2 coded vdeo for transmsson at lower bt-rates, n Proc. IEEE Int. Conf. Acoustcs, Speech and Sgnal Processng, Atlanta, GA, 1996, pp. 1998-2001. [3] P. Assuncao and M. Ghanbar A frequency-doman vdeo transcoder for dynamc bt-rate reducton of MPEG-2 btstreams, IEEE Trans. Crcuts Syst. Vdeo Technol., vol.8, pp.953-977, Dec. 1998. [4] P. Yn, A. Vetro, B. Lu and H. Sun, Drft compensaton for reduced spatal resoluton transcodng, IEEE Trans. Crcuts Syst. Vdeo Technol., vol.12, pp.1009-1020, Nov. 2002. [5] J. N. Hwang, T. D. Wu, and C. W. Ln, Dynamc frame-skppng n vdeo transcodng, n Proc. IEEE Workshop Vdeo Sgnal Processng, Redondo Beach, CA, Dec. 1998, pp.616-621. [6] K. T. Fung, Y. L. Chan, and W. C. Su, New archtecture for dynamc frame-skppng transcoder, IEEE Trans. Image Processng, vol.11, pp.886-900, Aug. 2002. [7] N. Memon and R. Rodla, Transcodng GIF mages to JPEG-LS, IEEE Trans. Consumer Electon., vol.43, pp.423-429, Aug. 1997. [8] N. Feamster and S. Wee, An MPEG-2 to H.263 transcoder, n Proc. SPIE Conf. Voce, Vdeo Data Communcatons, Boston, MA, Sept. 1999. IEEE Transactons on Consumer Electroncs, Vol. 50, No. 2, MAY 2004 [9] W. L Overvew of fne granularty scalablty n MPEG-4 vdeo standard, IEEE Trans. Crcuts Syst. Vdeo Technol., vol.11, pp.301-317, Mar. 2001 [10] E. Barrau, MPEG vdeo transcodng to a fne-granular scalable format, n Proc. IEEE Int. Conf. Image Processng, Rochester, NY, vol.1, Sept. 2002, pp.717-720. [11] R. Kartan, M. Roger, and Z. Josp, Transcodng versus multple mage versons n mage servers wth QoS Capabltes, n Proc. Norsk Informatkkonferanse, Tromsø, Norway, Nov. 2001, pp. 141-151. [12] W. Y. Lum and F. C. M. Lau, "On balancng between transcodng overhead and spatal consumpton n content adaptaton", n Proc. ACM Mobcom, 2002. pp. 239-250. [13] IBM Deskstar 120GXP hard dsk drves, http://ssddom01.hgst.com/tech/techlb.nsf/techdocs/ E0B26749E1A7728C87256B290055ECA5/$fle/D120GXP_ds.PDF Ilhoon Shn receved the B.S. and the M.S. degrees n computer scence and engneerng from Seoul Natonal Unversty, Korea. He s currently a Ph.D. canddate n the School of Computer Scence and Engneerng, Seoul Natonal Unversty. Hs research nterests nclude vdeoon-demand systems and operatng systems. Kern Koh receved the B.S. degree n appled physcs from Seoul Natonal Unversty, Korea, and the Ph.D. degree n computer scence from the Unversty of Vrgna. He s currently a professor of Computer Scence and Engneerng at Seoul Natonal Unversty. Hs research nterests nclude operatng systems, web applcatons, and vdeo-on-demand systems.