Journal of VLSI Sgnal Proessng 2007 * 2007 Sprnger Sene + Busness Meda, LLC. Manufatured n The Unted States. DOI: 10.1007/s11265-006-0024-7 A Novel Dynam and Salable Cahng Algorthm of Proxy Server for Multmeda Objets HYUNG RAI OH AND HWANGJUN SONG CSE, POSTECH, San 31, Hyoja-dong, Nam-gu, Pohang, Gyeongbuk 790-784, Korea Reeved: 22 August 2006; Aepted: 29 November 2006 Abstrat. Ths paper presents a novel dynam and salable ahng algorthm of proxy server wth a fnte storage sze for multmeda objets. Among the multmeda suh as text, mage, audo and vdeo, vdeo s a domnant omponent n terms of the performane of proxy server due to ts traff haratersts. For the fast ahng proess, ahng sequenes for vdeos are obtaned to derease both the buffer sze and the requred bandwdth and saved nto metafles n advane. Then, we present a novel ahng and replang algorthms for multmeda objets based on the metafles. Fnally, expermental results are provded to show the superor performane of the proposed algorthm. Keywords: proxy server, ahng, multmeda, network 1. Introduton Proxy server has been wdely employed to redue the response tme delay and mprove the network utlzaton [1, 2]. Proxy server s losely loated to lents. When a lent wants some data, proxy server nterepts the request message and provdes the data f possble. Otherwse t ontats the remote orgnal server, and then saves the reeved data n storage deves and forwards them to the lent. As a result, the response tme s sgnfantly redued n terms of lent, the load of server an be redued, and the network utlzaton an be mproved sne the number of duplated data dereases. Reently, the role of proxy server has been extended to varous Ths researh was supported by the MIC(Mnstry of Informaton and Communaton), Korea, under the ITRC(Center for Moble Embedded Software Tehnology) support program supervsed by the IITA(Insttute of Informaton Tehnology Assessment) (IITA- 2005-C1090-0501-0018) and Grant No. R08-2004-000-10084-0 from the Bas Researh Program of the Korea Sene and Engneerng Foundaton. multmeda serves over network whle Web ahng s a domnant applaton untl now. Vdeo has several dfferent haratersts ompared wth other meda suh as text, mage, and audo, et. Frst, vdeo requres a onsderable amount of data ompared wth text, mage, and audo. Thus, t s almost mpossble to store the whole data of vdeo objet at proxy server whle the entre data are saved n the ase of text and mage. Seond, the gener attrbute of vdeo s hghly VBR (varable bt rate). Aordngly the requred lent_s buffer sze s very large. Thrd, t s tme-onstraned. Atually, vdeo s a domnant omponent n the performane of proxy server. Hene, the effetve vdeo ompresson algorthm s ndspensable to redue the amount of data. So far, dgtal vdeo odng tehnques have advaned rapdly. Internatonal standards suh as MPEG-1, 2 [3] and 4 [4], H.261 [5], H.263/+/++ [6], and H.264 [7] have been establshed or under development to aommodate dfferent needs by ISO/IEC and ITU-T, respetvely. The ompressed vdeo data s generally varable-
Oh and Song bt-rate due to the gener haratersts of entropy oder, frame type and sene hange/nonstant moton hange of the underlyng vdeo. Generally speakng, the varablty of ompressed vdeo traffs onssts of two omponents suh as short-term varablty (or hgh frequeny varablty) and longterm varablty (or low frequeny varablty). Bufferng s qute effetve n redung losses aused by varablty n the hgh frequeny doman whle t s not effetve for handlng varablty n the low frequeny doman [8]. On the other hand, a onstantbt-rate vdeo traff may be generated by ontrollng the quantzaton parameters and t s muh easer to handle over the network, but the qualty of the deoded vdeo may be serously degraded. Furthermore, vdeo requres the strngent network QoS (qualty-of-serve) due to ts gener tme onstraned haraterst for the smooth dsplay. These fats make the problem more hallengng. Untl now, a large amount of researh efforts have been devoted to the networks supportng guaranteed QoS and vdeo odng that s adaptve and robust to network ondton. It s very mportant to determne whh data must be stored at proxy server when only a lmted storage spae s avalable. Cahng algorthm handles ths problem. Several approahes have been proposed so far. LFU (least frequent used) [9] and LRU (least reently used) [10] have been extensvely used untl now. LFU algorthm removes the least frequently used objet and then ahes a new objet when proxy server reeved a new request message. On the other hand, LRU algorthm removes the least reently used objet. Atually, these whole ahng algorthms are unsutable for vdeo objet beause t generally needs a onsderable amount of data as mentoned before, thus a few effetve vdeo ahng algorthms have been proposed n the lterature. The prefx ahng, proposed by Sen et al. [11], s a speal form of seletve vdeo ahng, whh nvolves ahng only a group of onseutve frames at the begnnng of the vdeo sequene to smooth and redue the bt-rate of VBR vdeo. Wang et al. proposed a vdeo stagng algorthm n [12], whh pre-fethes to the proxy a porton of bts from the vdeo frames whose sze s larger than a pre-determned ut-off rate, to redue the bandwdth n the server-proxy hannel. Dan and Staram proposed a GIC (Generalzed Interval Cahng) algorthm n [13], whh ahes short vdeo objets as well as ntervals or fratons of large vdeo objet. The algorthm orders all ntervals n terms of nreasng nterval sze. It alloates ahe to as many of the ntervals as possble. Hene, the GIC algorthm stll maxmzes the number of streams served from ahe wth a larger number vdeo segments and vdeo objets. Yu et al. proposed QoS-adaptve proxy ahng [14], whh adapts farly well to network bandwdth varatons and aheves good qualty under dfferent network ondtons by meda-haraterstweghted replang algorthm, network-ondtonand-meda-qualty-adaptve resoure-management mehansm, pre-fethng sheme and request and send-bak shedulng algorthms. Shen et al. proposed ahng strateges n the transodng-enabled proxy systems [15] that are useful a ahng algorthm over heterogeneous network. B. Wang et al. proposed the proxy-asssted transmsson sheme [16] that redues the ntal lateny and maxmze the utlzaton of server network bandwdth over multast for ontnuous meda delvery. And, Mao and Ortega proposed a salable seletve ahng [17] that s partal ahng strateges spefally desgned for streamng vdeo. They proposed SCQ (Seletve Cahng for QoS networks) and SCB (Seletve Cahng for Best-effort networks). These algorthms are frame-wse ahng shemes n whh SCQ algorthm redues the requred maxmum buffer sze and SCB algorthm prevents underflow at the lent. In ths paper, we propose a dynam and salable ahng and replang algorthms based on the metafle for multmeda objets. In the followng, multmeda are lassfed nto non-real-tme meda suh as text and mage and real-tme meda suh as vdeo and audo, and the whole ahng s performed for non-real-tme meda and the partal ahng s done for real-tme meda. It s assumed that vdeo s transmtted smultaneously wth the orrespondng audo for the synhronzaton. Thus they are handled as an objet (for the smplty, we all the ntegrated vdeo/audo objet as vdeo objet). It s observed that the performane of vdeo ahng algorthm generally depends on the relatve frame postons n the tme axs as well as the frame sze under the general vdeo traff ondton. Based on ths observaton, we proposed an effetve ahng algorthm mnmzng the requred lent_s buffer sze and the hannel bandwdth between the remote orgnal server and lent, whh s aheved by seletng ahed frames mnmzng the normalzed buffer sze
A Novel Dynam and Salable Cahng Algorthm for Multmeda Objets n [18]. Now, the ahng nformaton s stored nto metafles. Ths paper s organzed as follows. Symbol desrptons are desrbed n Seton 2, detals of proposed ahng and replang algorthms are presented n Seton 3, expermental results are provded n Seton 4, and fnally onludng remarks are gven n Seton 5. 2. Symbol Desrptons We assume that the number of non-real-tme meda objets s N N and the number of vdeo objets s N. R N s the sze of the th non-real-tme meda objet and N ndates whether the th non-realtme meda objet s ahed or not. And the th vdeo onssts of L frames where Z[1, N ], the sze of the jth frame s R ;j bytes n the th vdeo, and ts temporal samplng rate s T s seond (typal values are 1/25 or 1/30 s). Hene the total sze of the jth vdeo s alulated by R ¼ XL j¼1 R ;j : ð1þ And let! denote the postons of ahed frames n the th vdeo (t s alled the ahed frame vetor n the followng),.e.,! ¼ ;1 ; ;2 ; ; ;L ; ð2þ where ;j ndates whether the jth frame of the th vdeo s ahed or not, whh s defned by ;j ¼ 0 f the jth frame n the th vdeo s not ahed: 1 f the jth frame n the th vdeo s ahed: ð3þ It s assumed that the hannel bandwdth s onstant-bt-rate BW for the th vdeo (the extenson to varable-bt-rate hannel stll needs a onsderable amount of researh efforts sne t s qute dffult to estmate how muh the network resoures are requred and guarantee the qualtyof-serve for the burst multmeda traff under the varable-bt-rate hannel), and the ntal lateny, the tme delay between the dsplay tme at lent and the tme that the transmsson begns at the sender, s d, that s, R ;j s onsumed for the dsplay the th vdeo at tme t=j. Now the followng ondton must be satsfed when the avalable storage sze of proxy server s M. XN N ¼1 N R N þ XN ¼1 : h!! R T M ; ð4þ where! R ¼ R;1 ; R;2 ; ; R;L and the dot operator s the vetor nner produt. By the way, some of M must be reserved for the nstant dsplay before the data from the remote orgnal server arrve, that s, prefx ahng s requred for at least round trp tme between remote orgnal server and lent (t s assumed that M prefx s reserved for the prefx ahng n the followng). Generally speakng, BW must be larger than the average rate of the th vdeo defned by BW avg ¼ R L T s for the smooth dsplay of the th vdeo at the lent sne the ompressed vdeo traff s burst as mentoned earler. Now, the followng equaton must be satsfed to avod the lent_s buffer underflow to ause the dsplay tme delay. BW Q BW avg : ð5þ For the smplty, t s assumed that the lent starts to dsplay the frst frame of the th vdeo at t=0 and thus the meda server begns transmttng the frst paket at t=jd sne the ntal lateny s d. The umulatve frame rate at tme t n the th vdeo, whh means the vdeo data onsumpton urve at the lent at tme t, s defned by S ðþ¼ t Xt j¼0 R ;j : ð6þ Now, BW! and S! ðþndate t the requred onstant-bt-rate hannel bandwdth and the umulatve frame rate at tme t n the th vdeo when the ahed frame vetor s!, respetvely. Thus, S! s defned by ðþ t S! ðþ¼ t Xt j¼0 1 : ;j R ;j : ð7þ
Oh and Song And, the slope funton of the vdeo s defned by F! ðþ¼ t S! ðþ t t : ð8þ Sne the lent retreves the ahed frames from the proxy server at the tme when those frames need to be dsplayed and t s assumed that the tme delay s very short between the proxy server and the lent, the ondton for the smooth dsplay at lent an be desrbed as follows. X t BW j¼ d ðþq j S! ðþ; t for all t 2 1; L ; ð9þ where BW ðþs j the hannel bandwdth at jth frame_s tme n the th vdeo. Atually, S! ðþs t thought of as the vdeo data onsumpton urve at the lent when the ahed frame vetor s! whle P t j¼ dbw ðþs j the data supply urve from the server-to-lent hannel. Under the CBR (onstant bt rate) hannel whose bandwdth s BW, P t j¼ dbw ðþs j smply alulated by BW ðt þ d Þ. To avod lent_s buffer underflow for dsplay the th vdeo, the followng ondton must be satsfed. BW! ¼ max 1tL F! ðþ t ; t 2 1; L ð10þ sne BW! Q F ðþ; t t 2 ½1; L Š must be satsfed. And, the tme that the peak bandwdth s requred n! the th vdeo s defned by t peak ¼ arg max F 1tL! ðþ t ð11þ The tme that the maxmum buffer sze s needed n the th vdeo s defned by t max ¼ arg max 1tL ( ) ðþ S j! ðþ t : ð12þ X t BW j¼ d It s assumed that request s arrved aordng to a Posson dstrbuton wth an expeted nter-arrval tme of 1/l, where l s request rate. And, vdeo seleton follows Zpf-lke dstrbuton. Now, the probablty of hoosng the th vdeo s alulated by p ¼ f PN k¼1 f k ; for ¼ 1; :::; N ; ð13þ where f ¼ 1 1. Here, q s a parameter to spefy the skew fator (q s typally set to 0.271, whh losely mathes the populartes generally observed by vdeo store rentals [19]). The request rate for the th vdeo s alulated by l ¼ p l: In the real system, l s estmated by ð14þ bl ¼ AC T ; ð15þ where AC s the number of requests n the th vdeo durng the tme nterval T. The same method s used for the other meda. 3. Proposed Cahng Algorthm In ths senaro, server-to-proxy hannel provdes guaranteed onstant bandwdth serve and the hgher bandwdth serve s provded at the hgher ost. On the other hand, proxy-to-lent hannel s fast and relable at the even lower and fxed pre. Thus, the tme delay and the ost between proxy server and lent are neglgble ompared wth those between remote server and proxy server, and the ahed data at the proxy server s almost nstantly avalable for the dsplay at the lent. Thus, the requred lent_s buffer sze n the proxy-to-lent hannel s neglgble ompared wth that n the server-to-lent hannel. The logal system arhteture under the onsderaton s shown n Fg. 1. Now, we address how to determne N ; ¼ 1; 2; ; N N for non-real-tme meda objets and ahng vetors! ; ¼ 1; 2; ; N for real-tme meda objets mnmzng the requred lent_s buffer sze and the hannel bandwdth. Now, we an formulate the multmeda ahng problem as follows.
A Novel Dynam and Salable Cahng Algorthm for Multmeda Objets Fgure 1. Logal system arhteture under the onsderaton. Problem Formulaton Determne N ; ¼ 1; 2; ; N N and! ; ¼ 1; 2; ; N to mnmze the requred lent_s buffer sze and the peak bandwdth : subjet to PN N N R N þ PN h!! T R M ; ¼1 ¼1 d N d max N and d d max ; where d N s the response tme of non-real-tme meda, d max_n s the tolerable maxmum response tme, d s the ntal lateny of real-tme meda, and d max_ s the tolerable maxmum ntal lateny. Hene, the requred bandwdth for non-real-tme meda s alulated by BW N ¼ N R dmax N. Frst of all, total reduton ost s defned as follows to onsder the requred lent_s buffer sze and the peak bandwdth onurrently.! l BW ð0þ BW! þ XNN l N j BWj j N w XN ¼1 þ ð1 w Þ XN ¼1 j¼1! l B ð0þ B ;! ð16þ where w s the weghtng fator between the redued bandwdth and the redued buffer sze. Now, the above ahng problem s equvalent to fndng j N ; j ¼ 1; 2; ; N N and! ; ¼ 1; 2; ; N that maxmzes the above total reduton ost, but t looks lke the knapsak problem. To solve the knapsak problem, Greedy algorthm and Dynam Programmng an be employed. However, the former may provde only a loal mnmum wth a low omputatonal omplexty and the latter gve the optmal soluton wth a huge omputatonal omplexty. To redue the omplexty, we dvde the gven storage spae nto two parts for real-tme meda and non-real-tme meda,.e., frst, we set the ntal storage spae assgnment, then the proposed whole ahng s performed for non-real-tme meda and the proposed partal ahng s done for real-tme meda wth the gven storage assgnment. And the assgnment s teratvely adjusted based on bseton method to obtan the optmal soluton. 3.1. Metafle-based Cahng Algorthm for Multmeda Objets In our prevous work [18], we proposed a salable proxy ahng algorthm for only a vdeo traff to mnmze the requred buffer sze and the requred bandwdth at the lent. The proposed ahng algorthm s summarzed as follows (see [18] for more detals). Rule 1: Cahe the frame maxmzng B norm! aft for 1 j t peak. Rule 2: When several frames have the same normalzed buffersze, ahe the frame mnmzng BW! aft F! ðþ. t aft Where B norm! aft s the normalzed buffer sze defned by B norm! aft. ¼ B! bef B! aft R ;j ; where B! bef and B! aft are the maxmum buffer sze before and after the jth frame ahng n the th
Oh and Song vdeo, respetvely (the only dfferene between! aft and! bef s that ;j ¼ 1 and ;j ¼ 0, respetvely). That s, B! bef ¼ max BW 1tt! bef S! ðþ t ; peak bef B! aft ¼! aft S! ðþ t : max 1tt peak BW Now, the ahng sequenes of vdeo objets are found and saved nto metafles as shown n Table 1. Based on the above metafles, we propose a dynam ahng algorthm that provdes a soluton lose to the globally optmal soluton wth a relatvely low omputatonal omplexty. At frst, the benefal value for real-tme meda s defned by + ;O j ¼ +. ;O ðþ j R ;O ðþ j ðþ ¼ a l BW O ðj 1Þ þfð1 Þ l B O ðj 1 aft ; ð17þ BW O ðþ j Þ B O ðþ j ; where a s the weghtng fator between the redued bandwdth and the redued buffer sze and O (j) s the frame number of the jth ahed frame n the th vdeo. On the other hand, the benefal value for non-real-tme meda s defned as follows sne the redued buffer sze s zero n the ase of whole ahng. N ¼ l BW N R : ð18þ N Table 1. Cahng order An example of metafle. Frame number O j () Requred bandwdth Requred buffer sze Frame rate 1 0(no ahng) 10,000 7,500,000 0 2 1 990 7,490,000 30,000 3 2 980 7,480,000 3,000 4 3 970 7,470,000 3,000 5 4 960 7,460,000 7,000 6 876 955 7,450,000 3,000 7 745 950 7,440,000 30,000 8 1,023 945 7,430,000 27,000 9 1,932 940 7,420,000 34,000 10 534 935 7,410,000 10,000 Now, the pseudo ode of the proposed dynam ahng algorthm s summarzed n the followng. Dynam Cahng Algorthm The proxy server must ahe the frst some frames for ntal lateny n every vdeos. ahed sze of real-tme meda = M prefx ; ahed sze of non-real-tme meda = 0; I = m = the number of the prefx-ahed frames n the th vdeo, Z[1, N ]; whle{ /* Storage assgnment adjustment */ f (normalzed reduton ost s mnmzed) break; adjust M v by bseton method; /* Whole ahng for non-real-tme meda */ whle { fnd k ¼ arg max N ; 1N N f (ahed sze of non-real-tme meda + R k N > M jm v ) break; ahe the kth non-real-tme meda objet; ahed sze += R k N ; /* Partal ahng for real-tme meda */ whle { fnd k ¼ arg max 1N Wndow Sze max ; f (ahed sze of real-tme meda + R k;mk > M v) break; ahe the (m k )th frame of the kth vdeo objet; ahed sze += R k;mk ; mk ¼ m k þ 1; f (m þ W þ 1 > L ) W ¼ L m 1; max s denoted by m þwþ1 P j¼m þ1 m + ðþ. þwþ1 ;O j P R ;O ðþ j, j¼m þ1 where M v s the storage spae for real-tme meda.
A Novel Dynam and Salable Cahng Algorthm for Multmeda Objets Table 2. Ttles of test H.263 trae fles [20]. Vdeo ID Trae fles Vdeo ID Trae fles 1 Aladdn 14 VIVA 2 ARD News 15 N3 Talk 3 ARD Talk 16 Parkng 4 Mr. Bean 17 Robn Hood 5 Boulevard Bo 18 Slene Of The Lambs 6 The Frm 19 Smpsons 7 De Hard III 20 Alpn Sk 8 From Dusk Tll 21 Soer Dawn 9 Star Trek Frst 22 South Park Contat 10 Offe Cam 23 Star Wars IV 11 Formula 1 24 Sus und Strolh 12 Futurama 25 Starshp Troopers 13 Jurass Park I (http://msl.posteh.a.kr/msl/dynam_ahng/table2.htm). 3.2. Metafle-based Replang Algorthm for Multmeda Objets In ths seton, we propose the replang algorthm when a user aess pattern s hanged or a new request s arrved. It s assumed that the proxy server exeutes the replang algorthm perodally every T r. The proposed replang algorthm an sgnfantly redue the omputatonal omplexty wthout the performane degradaton ompared wth the reahng f T r s well determned onsderng the user aess pattern hangng rate. The pseudo ode of the replang algorthm s presented n the followngs Replang Algorthm /* Non-real-tme meda replang*/ whle { k ¼ arg max h ; 2 1; N NN p bef < p ur ; ahed sze of non-real-tme meda += R k N ; whle (ahedn sze > M jm h v ){ N g; 2 1; N Np bef > p ; ur l ¼ arg mn f ( k N >l N ){ remove R l N ; else { ahed sze of non-real-tme meda j= R k N ; ext of non-real-tme meda replang; ahe R k N ; /* Real-tme meda replang*/ whle { k ¼ arg max max ; 2 1; N p bef < p ur ; ahed sze of real-tme meda += R k;mk þ1 whle (ahed sze > M v ){ l ¼ arg mn mn ; j 2 1; N p bef > p ur ; f ( k max >l mn ){ ; remove R l;ml ; ml ; else { ahed sze of real-tme meda -= R k;mk þ1 ; ext of real-tme meda replang; ahe R k;mk ; mk ++; Wndow Sze f (m þ W þ 1 > L ) W ¼ L m 1; f (m W < I ) W ¼ m I ; where, mn s denoted by P m P m ðþ j j¼m W R;O ðþ j j¼m W +;O, p bef s the probablty of hoosng the th vdeo at the frst prevous replang, and p ur s the probablty at the urrent. 4. Experment Results Durng the experment, we used 500 texts, 500 mages and 25 vdeos (vdeo ttles are gven n Table 2). The.
Oh and Song Fgure 2. Performane omparson aordng to the wndow sze. average szes of texts and mages are 54,774 and 261,602 bytes, respetvely. Vdeo sequenes n Table 2 are enoded by H.263 [6]. The enodng frame rate s 25 frames per a seond, and the enodng struture s IBBPBBPBBPBB (.e., 1 GOP onssts of 12 frames). As a result, the average number of frames s 31,465 and the output traffs are varable-bt-rate. Sne the vdeo s a domnant omponent, the performane s shown n Seton 4.1 when the proposed algorthm s exeuted for only multple vdeo objets. And then the performane s presented n Seton 4.2 when the proposed algorthm s extended to multmeda objets. Fgure 4. Performane omparson of replang algorthms for multple vdeo objets. 4.1. Performane of the Proposed Cahng/ Replang Algorthms for Multple Vdeos Durng the experment, the tme nterval for prefx ahng s set to about 334 mllseonds and thus the orrespondng prefx ahng sze (M prefx ) s set to the amount of the frst ten frames n every vdeo. The proxy server_s spae (M ) s set to 60 MB for the dynam ahng, the average number of requests s 100 for a unt tme and the weghtng fator (a) s set to 0.8. The expermental results are presented n Fgs. 2, 3 and 4. Fgure 2 shows the performane omparson aordng to wndow sze. It s observed Fgure 3. Performane omparson the seletve ahng wth the proposed ahng algorthm when the wdow sze s set to 10. Fgure 5. The normalzed reduton ost aordng to the dfferent storage assgnment for non-real tme meda.
A Novel Dynam and Salable Cahng Algorthm for Multmeda Objets that Greedy algorthm shows the worst performane. On the other hand, the proposed algorthm shows a better performane as the wndow sze nreases. Compared wth Greedy algorthm, the total reduton ost s mproved by about 300% when the ahng rato s 100%, where the ahng rato s denoted by the ahed sze/the storage sze. The total reduton ost (Eq. 17) urves are almost same when wndow szes are 5, 10, and 20 as shown n Fg. 2. Fgure 3 shows performane omparson wth SCQ algorthm aordng to wndow sze. The proposed algorthm shows a better performane than SCQ algorthm sne t does not onsder the general traff as shown n the prevous work [18], Atually, the proposed algorthm shows performane mprovement by about 80% when the ahng rato s 100% ompared wth SCQ algorthm. Fgure 4 shows the performanes omparson among re-ahng algorthm, Greedy algorthm, and the proposed replang algorthm when the request rate s hanged. Now, the normalzed ost s defned by the total reduton ost/the total ahed sze. The re-ahng means that the proxy server exeutes the proposed ahng algorthm wth the new request rate after removng all ahed data. As shown n Fg. 4, both the re-ahng and the proposed replang algorthm an provde almost same effeny, but the performane of Greedy algorthm serously deterorates. By the way, t s emprally observed that the omputatonal omplexty of the proposed replang algorthm s about 10% of that of re-ahng although the omputatonal omplexty of the proposed replang algorthm generally depends on the request rates of vdeo objets and vdeo haratersts, et. 4.2. Performane of the Proposed Cahng/ Replang Algorthm for Multmeda Objets Fgure 6. Performane omparson of replang algorthms for multmeda objets. Durng the experment, the proxy server_s spae(m )s set to 100 MB for the dynam ahng, and the average numbers of requests per a unt tme for texts, mages and vdeos are 500, 500 and 20, respetvely (audo objets are nluded n vdeo objets as mentoned earler). The weghtng fator (a) of Eq.(17) s set to 0.8. Fgure 5 shows the performane aordng to the storage sze hange for non-real-tme meda. As shown n the fgure, the normalzed reduton ost s an almost onave urve and the optmal storage assgnment for non-real-tme meda s easly found by usng b-seton method. As a result, the normalzed reduton ost s maxmzed when the storage sze for non-real-tme meda s about 35,000,000 B. Fgure 6 shows the performane omparson among re-ahng, the proposed replang, and Greedy algorthms. The phenomena almost same as Seton 4.1 sobserved,.e.boththe re-ahng and the proposed replang algorthm show almost same effeny, whh s muh better than that of Greedy algorthm, and the proposed replang algorthm stll works well wth lower omplexty than re-ahng. 5. Conluson In ths paper, we have proposed an effetve dynam and salable ahng algorthm of the proxy server for multmeda objets nludng text, mage, audo and vdeo. After alloatng the gven storage spae to non-real-tme meda and real-tme meda, the proposed ahng and replang algorthms have been performed wth the gven storage sze and metafles nludng effetve ahng sequenes to redue the buffer sze and the requred bandwdth. The storage spae re-alloaton based on bseton method has been teratvely exeuted to fnd the optmal soluton. The proposed replang algorthm an keep the performane almost same as re-ahng algorthm wth a low omputatonal omplexty when user aess pattern s hanged. The proposed algorthms have shown a superor performane durng the whole experment.
Oh and Song Referenes 1. L. Rzzo and L. Vsano, BReplang Poles for a Proxy Cahe,^ IEEE/ACM Tran. Netw., vol. 8, no. 4, 2000, pp. 158 170, (Aprl). 2. J. Shm, P. Sheuermann, and R. Vngralek, BProxy Cahe Algorthms: Desgn, Implementaton, and Performane,^ IEEE Trans. Knowl. Data Eng., vol. 11, no. 4, 1999, pp. 549 562, (Jan.). 3. ISO/IEC 13818 (MPEG-2): Gener odng of movng ptures and assoated audo nformaton, 1994, (Nov.). 4. ISO/IEC JTC1/SC29/WG11, BOvervew of the MPEG-4 standard,^ ISO/IEC JTC1/SC29/WG11 N4030, 2001. 5. ITU-T Reommendaton H.261: Vdeo Code Audovsual Serve at p*64 kbps, 1993. 6. ITU-T Reommendaton H.263 verson 2, BVdeo Codng for Low Btrate Communaton,^ 1998. 7. Jont Vdeo Team of ISO/IEC MPEG and ITU-T VCEG, JVT- G050, BDraft ITU-T Reommendaton and Fnal Draft Internatonal Standard of Jont Vdeo Spefaton (ITU-T Re. H.264/ISO/IEC 14 496-10 AVC),^ 2003. 8. Z. Zhang, J. Kurose, J. Saleh, and D. Towsley, BTraff Smoothng, Statstal Multplexng and Call Admsson Control for Stored Vdeo,^ IEEE J. Sel. Areas Commun., vol. 15, no. 6, 1997, pp. 1148 1166. 9. J. Robnson and M. Devarakonda, BData Cahe Management Usng Frequeny-Based Replaement,^ n Proeedng of the ACM SIGMETRICS Conferene on the Measurement and Modelng of Computer Systems, 1990, pp. 134 142. 10. S. Wllams, M. Abrams, C. Standrdge, G. Abdulla, and E. Fox, BRemoval Poles n Network Cahes for World-Wde Web Douments,^ n Proeedng of ACM SIGCOMM, 1996, pp. 293 305. 11. S. Sen, J. Rexford, and D. Towsley, BProxy Prefx Cahng for Multmeda Streams,^ n Pro. IEEE Infoom 99, New York, USA, 1999. 12. Z. Zhang, Y. Wang, and D. H. C. Du, BVdeo Stagng: A Proxy-server-based Approah to End-to-End Vdeo Delvery Over Wde-area Networks,^ IEEE/ACM Tran. Netw., vol. 8, no. 4, 2000, pp. 429 442. 13. A. Dan and D. Staram, BA Generalzed Interval Cahng Algorthm for Mxed Interatve and Long Vdeo Workloads,^ SPIE Pro., vol. 2667, 1996, pp. 344 351. 14. F. Yu, Q. Zhang, and W. Zhu, BQoS-Adaptve Proxy Cahng for Multmeda Streamng Over the nternet,^ IEEE Trans. Cruts Syst. Vdeo Tehnol., vol. 13, no. 3, 2003, pp. 257 269. 15. B. Shen, S. Lee, and S. Basu, BCahng Strateges n Transodng-enabled Proxy System for Streamng Meda Dstrbuton Network,^ IEEE Trans. Multmeda, vol. 6, no. 2, 2004, pp. 375 386. 16. B. Wang, S. Sen, et al., BOptmal Proxy Cahe Alloaton for Effent Streamng Meda Dstrbuton,^ IEEE Trans. Multmeda, vol. 6, no. 2, 2004, pp. 366 374. 17. Z. Mao and A. Ortega, BSalable Proxy Cahng of Vdeo Under Storage Constrants,^ IEEE J. Sel. Areas Commun., vol. 20, no. 7, 2002, pp. 1315 1327. 18. H. Oh and H. Song, BSalable Proxy Cahng Algorthm Mnmzng Clent_s Buffer Sze and Channel Bandwdth,^ J. Vs. Commun. Image Represent., vol. 17, no. 1, 2006, pp. 57 71. 19. C. C. Aggarwal, J. L. Wolf, and P. S. Yu, BOn Optmal Bathng Poles for Vdeo-On-Demand Storage Server,^ n Pro. of IEEE Internatonal Conferene on Multmeda Computng and Systems, 1996. 20. Vdeo Traes Researh Group, http://trae.eas.asu.edu/. Hyung Ra Oh reeved the B.S. and M.S. degree n Eletron and Eletral Engneerng from Hongk Unversty n 2003 and 2005. Currently, he s a Ph.D. student at the Department of Computer Sene & Engneer at POSTECH (Pohang Unversty of Sene and Tehnology). Hs nterests are n the areas of multmeda ahng, overlay multastng, mage/vdeo ompresson and onvergene networks. Hwangjun Song reeved B.S. and M.S. degrees from Dept. Control and Instrumentaton (EE), Seoul Natonal Unversty, Korea n 1990 and 1992, respetvely. From 1992 to 1994, he served (full-tme) n Korea army. He reeved hs Ph.D. degree from the Department of Eletral Engneerng-Systems at Unversty of Southern Calforna n May 1999. From 2000 to 2005, he was an Assstant Professor/Ve Dean of Admsson Affars at Hongk Unversty, Seoul, Korea. Sne February, 2005, he has been wth POSTECH (Pohang Unversty of Sene and Tehnology), Korea. Hs researh nterests nlude multmeda ommunaton (H.32x, H.26x), wred/ wreless network protools neessary to mplement a funtonal real-tme multmeda applaton, multmeda applatons, mage/vdeo ompresson, dgtal sgnal proessng, ontrol system and fuzzy-neural system.