Time-Shifted Streaming in a Tree-Based Peer-to-Peer System Jeonghun Noh ASSIA inc., Redwood City, CA, USA

Transcription

1 22 JOURNAL OF COMMUNICATIONS, VOL. 7, NO. 3, MARCH 212 -Shifed Sreaming in a Tree-Based Peer-o-Peer Sysem Jeonghun Noh ASSIA inc., Redwood Ciy, CA, USA jnoh@assia-inc.com Bernd Girod Informaion Sysems Laboraory, Deparmen of Elecrical Engineering Sanford Universiy, CA, USA bgirod@sanford.edu Absrac We propose a peer-o-peer (P2P) sreaming sysem ha mulicass live video as well as provides he live video as Video-on-Demand (VoD) during he live session. The proposed sysem allows users o individually pause and resume a live video sream, and o view video ha was sreamed before he paricular user joined he session. Since live video conen naurally resuls in many users waching i online a he same ime, he P2P concep is leveraged o reduce server load. We exend our Sanford Peer-o-Peer Mulicas (SPPM) proocol, originally designed for live video mulicas, o suppor playback conrol. To achieve his in a P2P fashion, peers sore received video packes in heir local buffer and forward hem a a laer ime when requesed by oher peers, which is called ime-shifed sreaming. To undersand ineracion beween live and ime-shifed raffic, we analyze video availabiliy of he proposed sysem, which refers o how many peers possess video conens of a paricular posiion. Moivaed by he analysis, we propose fas prefeching and a paren selecion scheme suiable for fas prefeching in order o furher reduce server load by disseminaing video quickly. Fas prefeching is possible because he relaxed ransfer deadline of ime-shifed sreams and peers exra uplink bandwidhs are exploied. Wih simulaions and mahemaical analysis, i is shown ha fas prefeching no only alleviaes he requiremen of server bandwidh by expediing video disseminaion, bu also miigaes video disrupion due o peer churn. Index Terms P2P sreaming, video-on-demand, imeshifed sreaming, playback conrol I. INTRODUCTION Two primary forms of video sreaming are live sreaming and video-on-demand (VoD). Live sreaming ofen implies mulicas communicaions as many users concurrenly wach video. Video-on-demand, on he oher hand, has been achieved by unicas communicaions because he emporal correlaion among video segmens users wach ends o be low. In his paper, we propose a P2P sreaming sysem ha allows viewers o individually pause a live video sream, rewind (even o conens prior o heir ime of joining), and jump forward (o conens beyond he Jeonghun Noh was wih Sanford Universiy, Sanford, CA 9435 USA. Manuscrip received February 15, 211; revised July 15, 211; acceped Sepember 3, 211. curren poin of playback), which allows for he VoD of a live sream. As viewers wach he same conen a a differen ime in a raher shor ime scale, he proposed P2P sysem explois high emporal correlaion among users viewing poins in ime. The poenial benefis from high emporal correlaion among user arrival imes for P2P VoD sysems are well sudied in [1]. In he proposed sysem, peers sore received video in heir local buffer. If a peer rewinds or jumps forward o conen no presen in is local buffer, he required video segmen is rerieved from oher peers. The video, buffered a some peer(s), is hen sreamed o he requesing peer asynchronously wih respec o he live sream emanaing from he server. To keep rack of he video conens a peers, he peer daabase a he source peer is augmened wih addiional informaion of peers. The proocol is changed o be able o handle addiional fields, such as a video posiion or a ime samp. This is called ime-shifed sreaming. As menioned earlier, he concep of ime-shifing in a live session is closely relaed o asynchronous video sreaming in VoD sysems. VoD sysems allow users o selec and wach video conen on demand. Wih playback conrol, also known as rick mode, users can also pause a video sream, rewind o an arbirary posiion, and fas forward. Unlike live video, video files requesed ondemand are pre-encoded and sored a video servers. If he available uplink bandwidh a he sender exceeds he video birae, he user can receive video faser han he play-back speed [2]. Tradiionally VoD services were provided by servers. For insance, Chaining [3] and Paching [4] employ media servers and IP mulicas, where IP mulicas is used o reduce server workload. The auhors of [5] sudy rade-off beween he sysem profi and user experiences in a near-vod sysem. In [6], paching and baching of server-based VoD service are augmened by aking ino consideraion heerogeneiy in receivers. As in server-based live sreaming, servers deployed for VoD services may fail in he presence of a large populaion of users. Thus, many P2P-based VoD sysems have been proposed o achieve scalable services. In order o achieve low workload on servers, he resources in he P2P nework are leveraged for efficienly aking doi:1.434/jcm

2 JOURNAL OF COMMUNICATIONS, VOL. 7, NO. 3, MARCH advanage of uplink bandwidh and sorage conribued by peers [7] [1]. For VCR-like operaion (emporal random access), eiher disribuion overlay is combined wih lookup overlay [7], [8], or a separae look-up overlay is buil o avoid a cenral look-up service [11] [13]. In [9], [14], [15], viewers are able o wach a pre-recorded video from he beginning of he conen. Dynamic Skip Lis [8] and osream [7] provide users random emporal access o video conen. The mesh-based P2P VoD sysems are sudied in [16] [18]. The auhors of [15] exploi he exra uplink bandwidhs and peer sorages by downloading video faser han playback speed. In [19], a user-log based prefeching algorihm is proposed. PPB, Peer-o- Peer Baching, combines peer-o-peer and IP mulicas by using IP mulicas for baching while a peer-o-peer nework is used o provide he iniial porion of video o new peers [2]. In [21], he auhors propose a ree-based sysem ha consrucs dynamic mulicas rees, called J- ree. The J-ree srucure explicily disinguishes he live mulicas ree and playback rees ha are implicily buil in our sysem. DRPSS, presened in [22], also suppors live and ime-shifed sreaming in a single sysem based on he Disribued Hash Table overlay. P2TSS [23] suppors ime-shifed sreaming in a live broadcas. In P2TSS, peers have wo separae buffers; a buffer for local video playback and a shared buffer for serving ime-shifed sreams o oher peers. The conens of he shared buffer do no change once he buffer becomes full of he received video. Unlike P2TSS, our sysem requires a single buffer o sore he incoming video sream regardless of is shif in ime (e.g., live or ime-shifed sream). The remainder of he paper is srucured as follows. Secion II inroduces he Sanford Peer-o-Peer Mulicas (SPPM) proocol [24]. In Secion III, SPPM is exended o suppor ime-shifed sreaming. Secion IV presens he analysis of video availabiliy, which represens how many peers possess video conens of a paricular posiion. In Secion V, we describe fas prefeching o improve video availabiliy. Fas prefeching allows peers o disseminae video faser. To faciliae video disseminaion wih fas prefeching, a new paren selecion algorihm, called maximum hroughpu, is proposed and is effecs are demonsraed wih exensive simulaions in Secion VI. packes. Peers ha receive video packes are he children of he peers ha relay he packes. When a paren peer can relay packes o more han one child peer, hen he pah naurally evolves from a chain o a ree srucure, wih he server being he roo and peers being inermediae nodes and boom nodes. Noe ha he desired disribuion srucure becomes a spanning ree because no packe needs o visi he same peer more han once. Fig. 1 illusraes an example overlay in which wo complemenary rees are consruced among a small group of peers. Typically, we use 4 o 8 rees, and peer groups migh be much larger [27]. As peers join he sysem, he rees are incremenally consruced in a disribued manner. When a new peer conacs he video source, he video source replies wih session informaion, such as he number of mulicas rees and he video bi rae. I also sends a lis of candidae parens randomly chosen from he able of paricipaing peers i mainains. The new peer hen probes each candidae paren o know abou heir curren saus. Afer receiving probe replies, he bes candidae paren is seleced and conaced for each ree by minimizing he heigh of he disribuion ree. Once he seleced candidae paren acceps he aachmen reques, a daa connecion is esablished beween he paren and he new peer. Afer daa ransmission sars, each child peer periodically sends hello messages o heir parens. When a peer leaves he sysem ungracefully, is parens deec i by observing consecuive missing hello messages, and sop forwarding video o he child. The deparing peer s children noice ha neiher video packes nor a response o hello messages arrive. Each abandoned child hen iniiaes procedure o connec o a differen paren node in he ree. Video sream P 4 S P 2 P 1 P 3 P 6 P 5 II. STANFORD PEER-TO-PEER MULTICAST Sanford Peer-o-Peer Mulicas (SPPM) is a P2P sysem ha achieves low-laency and robusness in live video mulicas. Like in [25], [26], SPPM adops he push approach for low-laency daa disseminaion. The push approach requires a persisen roue among peers. This roue is buil by connecing a pair of peers using unicas connecions, such as TCP or UDP. Since unicas connecions are esablished on op of he IP layer, he collecion of he unicas connecions is ofen called an overlay. Each daa pah, a chain of peers in he overlay, sars from he server. On he daa pah, all inermediae peers ac as boh parens and children; peers ha relay video packes are he parens of he peers ha receive he P 8 P 7 Tree 1 Tree 2 Figure 1. A peer-o-peer overlay consising of wo complemenary P. Bacciche Sanford Peer-o-Peer Mulicas 16 disribuion rees. The encoded video is packeized and spli ino disjoin subsreams, one for each disribuion ree. To provide he bes video qualiy despie congesion and peer churn, SPPM employs sender-driven packe scheduling [28] in conjuncion wih receiver-driven reransmission requess [29]. Acing as a relay, each peer schedules he nex packe o forward by comparing he imporance of packes in is oupu queue. As a receiver,

3 24 JOURNAL OF COMMUNICATIONS, VOL. 7, NO. 3, MARCH 212 each peer evaluaes he imporance of missing packes by compuing he expeced video qualiy improvemen if he corresponding frame is received before is playou deadline. Reransmission requess of seleced missing packes are hen sen o alernaive parens. This NACKbased reransmission is advanageous as packe error rae is ofen no saionary over ime nor uniform among peers. Since packe losses ypically occur in rare burss, such feedback error conrol is generally superior o forward error conrol. Feedback implosion, prohibiing feedback error conrol in nework-layer mulicas, is no an issue for a P2P overlay because reransmissions are handled locally beween a paren and a child peer. III. TIME-SHIFTED STREAMING In his secion, SPPM is exended o allow users o conrol heir playback in he ime domain [3]. Wih playback conrol, also known as rick mode in videoon-demand sysems, peers can individually pause a live video sream, rewind o an arbirary posiion (even o conens prior o heir ime of joining), and fas forward unil he laes posiion of he live video. In he exended SPPM, peers cache he received video in heir local buffer. For pause/resume, he locally cached video can be played back a a laer ime. When a peer rewinds or jumps forward o conen no presen in is local buffer, he requesed video segmen, buffered a some peer(s), is sreamed o he peer asynchronously unlike he live sream being mulicas o peers. This is called ime-shifed sreaming. As we will see laer, SPPM suppors boh he live sream and ime-shifed sreams while keeping he sysem s scalabiliy. While exending he SPPM sysem for ime-shifed sreaming, he fundamenal archiecure of he sysem is kep inac so ha live video mulicas is seamlessly suppored. In his secion, we describe only he sysem exensions added o he exising archiecure. A. Preliminaries As in live sream mulicas, he video server (source peer) sars o sream a live video saring a ime. Le V (x) denoe a video frame associaed wih posiion x in ime, x. The live video emanaing from he server a ime is hus represened by V (). A ime-shifed sream is represened by V ( d) wih delay d, < d, a ime. Suppose ha a new peer or an exising peer requess video V (x) a ime. If he peer requess he live video V (), he peer is called LS peer. If he peer requess a ime-shifed sream V ( d), < d, hen he peer is called TS peer. For LS peers, he live video is immediaely relayed from peer o peer. For TS peers, he live sream mus be delayed by ime d somewhere in he sysem, eiher a he server or a he peers. Delayed sreaming, or ime-shifed sreaming, is achieved by soring pas video packes in peers buffers while hey wach he video 1 and 1 A peer shares is sorage only when i paricipaes in a session. For a ypical wo hour-long video sream encoded a 6 kbps, abou 54 MB of sorage is required. This is a moderae space requiremen considering oday s ypical personal compuers. ransmiing hem when requesed. The ime-video plo in Fig. 2 illusraes he rajecories of he live sream, Peer 1 (LS peer), and Peer 2 (TS peer). The rajecories indicae he video conens sored a he peers over ime. The verical axis represens he ime samp of video frames (a ime samp is he ime when a video frame is encoded a he server). In Fig. 2, Peer 1 requess he live video V ( 1 ) a ime 1. Is video rajecory overlaps he live sream rajecory. A ime 2, Peer 2 requess he video saring from V (x 2 ). Since he requesed video is sored a Peer 1, Peer 2 can obain he video delayed by 2 x 2 from Peer 1. Video posiion x 2 Live sream 1 2 Peer 1 (LS) Peer 2 (TS) Figure 2. Video rajecories of live video, Peer 1, and Peer 2. Peer 1, an LS peer, requess he live video. Peer 2, a TS peer, requess he video delayed by 2 x 2. B. Peer Daabase In he sysem, peers collecively form a disribued video sorage. To faciliae he use of he videos sored by peers, locaing video among peers is essenial. To locae he video of a paricular posiion, he video server mainains a daabase comprising each peer s idenifier, he ime he video is requesed, he ime samp of he iniial video frame, and he ime samp of he laes video sored in he peer s buffer. The laes ime samp is esimaed by he server based on each peer s iniial video posiion and reques ime. Esimaion by he server reduces he amoun of conrol raffic beween he server and peers by runcaing he number of buffer sae updaes. To preven he esimaion from deviaing oo far from he acual buffer saes, peers may repor heir buffer sae inermienly (see Secion V-A). C. Video Reques and Connecion Se-up When a new peer joins or an exising peer seeks a video of a paricular posiion, i sends a query requesing a cerain video posiion o he server. The server refers o he peer daabase i mainains. I reurns a lis of randomly chosen paren candidaes whose video buffer may conain he requesed (possibly ime-shifed) video. The peer hen probes each paren candidae o learn abou he candidaes curren saus, including ime samps of he sored video and available bandwidh. Afer he peer receives probe replies, i selecs and requess for each

4 JOURNAL OF COMMUNICATIONS, VOL. 7, NO. 3, MARCH mulicas ree he bes paren candidae ha will accep i as a child. We discuss paren selecion crieria in deail in Secion V. Once he paren candidaes accep he video reques, daa connecions are esablished beween he parens and he child. D. Asynchronous Video Transmission Connecions esablished for ime-shifed sreaming are differen from connecions ha consiue he live video mulicas rees. In he laer case, packes are immediaely relayed o he nex hop in he mulicas ree, whereas in he former, packes are asynchronously relayed over he ime-shifed sreaming connecion. Muliple complemenary rees are he basis of daa disseminaion in SPPM. As LS peers, TS peers receive packes from muliple parens. Parens asynchronously relay video packes o heir TS peers by ensuring ha a packe ravels along he ree designaed o iself. A paren peer dedicaes a sreaming window for each TS child peer. Fig. 3 shows a sreaming window of a TS child ha requess video from posiion x 1 a ime 1. The lower end of he window, denoed by W L, is he oldes video packe he paren can ransmi. Since packes ha are pas heir display ime (called he play-ou deadline) are no decoded a he child, he value of W L is se o he minimum of he iniial video posiion in ime and of he ime samp of he video frame displayed a he child. The upper end of he window, denoed by W U, corresponds o he ime samp of he laes video packe he paren can send. The value of W U is seleced in such a way ha he child s downlink is no congesed by he paren, and he child does no suffer from buffer underflow. The paren regularly searches for packes whose ime samps are beween W L and W U in is video buffer. In addiion, only packes ha belong o he paren s ree ID are seleced o consruc he corresponding subsream. To avoid duplicae ransmission, he hisory of ransmied packes is mainained for a finie period of ime. Seleced packes are marked wih heir play-ou deadline and pu ino he nework oupu queue. The oupu queue sors video packes according o heir play-ou deadline. Therefore, urgen packes, such as reransmied packes or live video packes, are sen earlier han packes wih a relaxed deadline, such as packes in ime-shifed sreams. A a child peer, packes ha belong o differen subsreams may arrive ou-of-order. Once received, packes are ordered according o heir posiions in he video sream before being passed o he video decoder. E. Recovery from Disconnec Afer video ransmission sars, each child peer periodically sends hello messages o is parens. When a peer leaves he sysem ungracefully, is parens deec his by observing consecuive missing hello messages and sop forwarding video o he child. The deparing peer s children noice ha neiher video packes nor responses o Video posiion x B W U () W L () Figure 3. A sreaming window for a TS child ha requess video from posiion x 1 a ime 1. W L is is lower end and W U is is upper end. B is he pre-roll delay of play-ou a he child. Packes whose ime samps fi in he sreaming window are ransmied unless hey have been sen already. hello messages arrive. Each abandoned child hen iniiaes a recovery procedure o connec o a differen paren node in he ree. During he recovery procedure, LS peers reques he laes porion of he video in order o keep endo-end ransmission delay as low as possible. TS peers, on he oher hand, reques he nex posiion of he video hey received if he posiion was ahead of heir playback posiion. Oherwise, TS children reques he playback posiion of he video. IV. ANALYSIS OF VIDEO AVAILABILITY This secion presens he analysis of he availabiliy of ime-shifed video segmens among peers, which is closely associaed wih sysem scalabiliy. High video availabiliy lessens he need for making a connecion o he server, hereby reducing server load. The analysis also sheds ligh on he ineracion beween live and ime-shifed sreams. Suppose ha peers join he sysem according o a Poisson process N() wih rae λ [31]. Their lifeime follows an exponenial disribuion 2 wih parameer µ (1/µ is he average peer lifeime). We assume ha he probabiliy ha he nex peer is an LS peer follows a Bernoulli process wih parameer α. When a TS peer requess a video, is video access paern is assumed o be uniform. This assumpion allows us o be free from any paricular video conen and i provides he wors-case sudy because overlaps beween peers buffered conens are minimized, which resuls in minimizing he video availabiliy. Based on his assumpion, he posiion x is seleced, a ime, uniformly beween and. We furher assume ha a peer always receives he video i requess, by connecing eiher o anoher peer or o he video server. Le M(, x; λ, µ, α) denoe he number of peers possessing he video a posiion x a ime. To analyze he 2 A peer lifeime is known o follow a long-ail disribuion, such as he Zipf disribuion [27], [32], [33]. However, he exponenial disribuion is sill a popular choice for boh analysis and simulaion sudy because i reasonably approximaes he acual disribuion. More imporanly, his assumpion allows he sysem analysis o be racable, and he sysem behavior can be prediced by observing only a few primary sysem parameers.

5 26 JOURNAL OF COMMUNICATIONS, VOL. 7, NO. 3, MARCH 212 Video posiion Live sream Video posiion Live sream x (, x) x (, x) x -x Figure 4. Iniial video posiions and reques imes for LS peers (Lef) and TS peers (Righ) ha possess video a posiion x a ime. Lef: An LS peer whose arrival ime is earlier han x has video a posiion x a ime. Righ: A TS peer whose reques ime and iniial video posiion are wihin he area surrounded by (, ), ( x, ), (, x), and (x, x) on he ime-video plo has video a posiion x a ime. availabiliy of video conens among peers, we compue he expeced value of M(, x; λ, µ, α), given by E{M(, x; λ, µ, α)} = k Pr{M(, x; λ, µ, α) = k} = k=1 k Pr{M(, x; ) = k N() = i} Pr{N() = i}, k=1 i=k where N() represens he number of peers ha have arrived a he sysem beween and ime. Given N() = i, since N() is a Poisson process, peers arrival imes S 1,..., S i, considered unordered random variables, are disribued independenly and uniformly in he inerval [, ] [34]. This allows us o sudy each peer s behavior independenly regardless of peers join sequence. Consider a peer ha joins he sysem beween and. Le p denoe he probabiliy ha he peer is sill presen and possesses video a posiion x a ime. Then, E{M(, x; λ, µ, α)} in (1) can be rewrien as (1) and (x, x) on he ime-video plo in Fig. 4. One can easily check his by drawing a video rajecory (in parallel wih he live sream rajecory) from any poin inside he area unil ime and examine wheher he rajecory passes or ouches he horizonal line drawn from (, x) o (, x). Noe ha he video is no available above he live sream rajecory. Le g(s;, x) denoe he probabiliy ha a TS peer joining a ime s conains he video x a ime. The value of g(s;, x) is 1 when s ( x), ( x)/s when ( x) < s x, and ( s)/s when x < s <, respecively. Based on he exponenial disribuion of peer lifeime, he probabiliy ha he peer is sill presen a ime is e µ( s). Then, he probabiliy ha a TS peer ha conains video a x a ime, l(, x; µ), is l(, x; µ) = e µ( s) g(s;, x) 1 ds { x x e µs e µs ds + ( x) x s ds } + e µs s s ds. (3) = e µ x E{M(, x; λ, µ, α)} ( ) i = k p k (1 p) i k λ (λ)i e k i! k=1 i=k ( ) k 1 p = e λ ((1 p)λ) i (k 1)! 1 p (i k)! k=1 i=k ( ) k 1 p = e λ e (1 p)λ ((1 p)λ) k (k 1)! 1 p k=1 = e pλ k=1 (pλ) k 1 (k 1)! pλ = pλ. (2) We compue p by considering he cases of TS peer arrivals and LS peer arrivals separaely. Suppose ha a peer arrives a ime s, s. If i is a TS peer, hen i conains video a posiion x a ime only when is arrival (or reques) ime and is requesed video posiion are inside he area surrounded by (, ), ( x, ), (, x), When an LS peer arrives a ime s, i possesses video a posiion x a ime only when s x. The se of he pairs of he live video posiion and arrival ime is he line segmen from (, ) o (x, x) shown on he lef side of Fig. 4. Then, by incorporaing he likelihood of a peer s deparure before ime, he probabiliy ha an LS peer ha conains video a x a ime is expressed as x e µ( s) 1 ds = e µ (e µx 1). (4) µ Finally, p can be wrien as p = Pr {L c } Pr {x is possessed a ime L c } + Pr {L} Pr {x is possessed a ime L} = (1 α)l(, x; µ) + αe µ (e µx 1), (5) µ where L is he even ha an LS peer joins he sysem. By

6 JOURNAL OF COMMUNICATIONS, VOL. 7, NO. 3, MARCH insering (5) ino (2), we rewrie (2) as E{M(, x; λ, µ, α)} = (1 α)λl(, x; µ) + α λ µ e µ (e µx 1), (6) where (1 α)λl(, x; µ) corresponds o he expeced number of TS peers ha possess video a posiion x a ime and α λ µ e µ (e µx 1) corresponds o he expeced number of LS peers ha possess x a ime. In (6), he video availabiliy is prediced o increase linearly as he peer s join rae λ increases. This resul demonsraes ha he sysem is self-scalable. The availabiliy of he early porion of he video coninues o decrease as, which indicaes ha ime-shifed sreams have o be provided by he video server a a laer ime. In Fig. 5, he video availabiliy is prediced by evaluaing (6) wih λ = joins per second, 1/µ = 12 seconds, and α =.5. In he figure, he video near he live sream is available a many peers, he majoriy of which are LS peers. As ime progresses, more video frames become available and he overall video availabiliy decreases. Video posiion [seconds] [seconds] Figure 5. Video availabiliy on he ime-video plo. The value a (, x), prediced by he model, represens he expeced number of peers available for video a posiion x a ime (λ = joins per second, 1/µ = 12 seconds, and α =.5). Expeced number of peers TS peers only LS peers only Mixure Video posiion [in seconds] Figure 6. Model: video availabiliy a 3 seconds. (λ = joins per second, 1/µ = 12 seconds, and α =.5) Figs. 6 and 7 illusrae he expeced video availabiliy a 3 and 9 seconds, respecively. In hese figures, he Expeced number of peers TS peers only LS peers only Mixure Video posiion [in seconds] Figure 7. Model: video availabiliy a 9 seconds. (λ = joins per second, 1/µ = 12 seconds, and α =.5) conribuion from LS peers and TS peers is separaely shown. The expeced number of TS peers having video a posiion x a ime peaks when x = /2, whereas he expeced number of LS peers a ime peaks a he live video posiion of x = and is value is lim x α λ µ e µ (e µx 1) αλ/µ. Alhough he videos sored a eiher LS or TS peer have high overlaps a ime 3s, he overlap dramaically decreases a ime 9s. This resul implies ha a a laer ime of he session more TS peers have o connec o he server, which was already observed in Fig. 5. V. FAST PREFETCHING In he previous secion, we have observed ha he video availabiliy decreases over ime. Server load is highly affeced by video availabiliy because higher video availabiliy allows more peers o obain video from oher peers, hus reducing he number of connecions made o he server. In his secion we propose fas prefeching, which allows peers o download video more rapidly. Fas prefeching explois he relaxed ransfer deadline of ime-shifed sreams when compared o he live sream. As an inpu device such as a video camera generaes a video signal, he live video is immediaely encoded and ransmied o peers. On he oher hand, ime-shifed sreams are he pas porion of he video, already encoded and sored somewhere in he sysem. Fas prefeching also explois he exra uplink bandwidh of peers. A. Improving Video Availabiliy Suppose a TS child receives a ime-shifed sream from is paren. When is paren has sufficien uplink bandwidh, he sream can be pushed o he child faser han playback speed. Since fas prefeching faciliaes video disseminaion among peers, video availabiliy can be improved accordingly. Fas prefeching also improves he video playback qualiy of TS peers. Wih fas prefeching, peer buffers grow faser han playback speed, which reduces he urgency for acquiring laer packes. When a peer is disconneced due o failure of is paren, i reconnecs o he overlay. Since he peer has more video frames o play back unil i resumes o receive laer video,

7 28 JOURNAL OF COMMUNICATIONS, VOL. 7, NO. 3, MARCH 212 i experiences saisically less video disrupion due o buffer underflow. Since fas prefeching is based on he relaxed ransfer deadline of ime-shifed sreams, i does no benefi LS peers. When LS peers are disconneced from he overlay, hey ofen experience buffer underflow because he playou deadline of a video packe is kep small in live sreaming. To compensae for a igh play-ou deadline, LS peers connec o peers ha can minimize he number of logical hops o he source [35]. Using his crierion, LS peers aemp o reduce disrupions in receiving he video when heir paren fails or leaves unexpecedly. Video posiion x 2 x P() : Paren C 1 (): Child 1 (fas prefeching) C 1 (): Child 1 (no fas prefeching) C 2 (): Child 2 Figure 8. An example of fas prefeching for TS peers. The paren peer can serve wo children concurrenly. A ime, he excess download a Child 1 due o fas prefeching is C 1 () C 1 (). The effecs of fas prefeching are illusraed in Fig. 8. In his illusraion, he uplink of he paren is assumed o be wice he video birae. P () is he video rajecory of he paren. Afer Child 1 connecs o is paren, i prefeches he video faser han he playback speed unil Child 2 connecs o he same paren. The rajecories of he downloaded video for Child 1 and 2 are depiced as C 1 () and C 2 (), respecively. If fas prefeching is no employed, Child 1 receives video a he playback speed, which is marked by C 1 (). Since fas prefeching enables peers o download video faser han a he playback speed, he peer daabase s esimaion on he peers buffer sae may deviae from he acual buffers. To correc mismaches incurred a he daabase, peers inermienly repor heir buffer sae o he video server. B. Selecing Parens ha Maximize Prefeching LS peers choose heir parens according o he number of hops o he server, which is cerainly no opimal for TS peers because he video disrupion is less likely due o fas prefeching. To maximize prefeching, we propose a new paren selecion algorihm for TS peers, which esimaes video download for each paren candidae. When Paren Candidae i is probed by a TS peer ha requess he video from posiion x, he paren candidae repors (1) he download rae q i from is curren paren, (2) he number of bis cached a Paren Candidae i and requesed by he TS peer (e.g., a par of he requesed subsream and beyond he posiion x), denoed by l i, and (3) he maximum upload rae i can offer o he TS peer, denoed by r i. In Fig. 9, q i, r i, and l i repored by Paren Candidae i are depiced over ime. The Y-axis shows he video posiion in bis. Noe ha he raes and he buffer amoun were adjused o reflec he number of rees. As each mulicas ree delivers a 1/k fracion of he video bisream, q i, r i, and l i were muliplied by k for he illusraion (k is he number of rees). Afer probing he candidaes, he TS peer compues he amoun of prefeched video ha i expecs o receive ime D afer connecing o each paren candidae: if he peer s download caches up wih is paren s download before ime D has elapsed, he expeced download is D q i + l i. Oherwise, he expeced download is D r i (see Fig. 9). The value of D is chosen o be sufficienly small, as q i and r i may change over ime. The bes candidae k is chosen according o he rule k = arg max min (D q i + l i, D r i ). (7) i The video server is seleced as a paren only when here is no peer possessing he requesed video or when no peer ha possesses he requesed video has available bandwidh. I should also be noed ha he server does no paricipae in fas prefeching in order o minimize is uplink bandwidh usage. Video posiion x + D Video posiion x + D Figure 9. Expeced download a a child peer afer ime D when a child peer joining a ime connecs o paren candidae i. Lef: child buffer lags behind paren buffer a ime + D. Righ: child buffer caches up wih paren buffer. (q i : download rae of Paren Candidae i from is paren, l i : buffered video in bis a Paren Candidae i, r i : upload rae from Paren Candidae i o he child peer.) C. Uplink Bandwidh Pariioning To suppor fas prefeching, peers pariion heir available uplink bandwidh o children as follows: (Sep 1) Allocae he ree birae r(= R/k) o every LS child. R is he video birae and k is he number of rees. (Sep 2) Allocae r o every TS child who compleed fas prefeching (he child s laes video ime samp is he same as he paren s laes video ime samp). (Sep 3) Compue r = U/n, r r. U is he remaining uplink bandwidh afer Sep 2, and n is he number of TS children who have no been assigned bandwidh in he previous seps. If some TS children have a smaller downlink capaciy han r, he bandwidh allocaed o hem is curailed in order no o overwhelm hem. The remaining bandwidh is evenly allocaed o he remaining TS children.

8 JOURNAL OF COMMUNICATIONS, VOL. 7, NO. 3, MARCH The algorihm is illusraed in Fig. 1. Noe ha he server performs no bandwidh pariioning because i suppors no fas prefeching. coninues o decrease over ime, as expeced from he analysis. 8 5 r U UT r r UT r r r r STEP 2 STEP 3 r STEP 1 r r Figure 1. Uplink bandwidh pariioning. In Sep 1, LS children are assigned he ree birae. In Sep 2, TS children ha have compleed fas prefeching are assigned he ree birae. In Sep 3, TS children ha can perform fas prefeching are assigned he remaining uplink bandwidh evenly. Every peer excep he server regularly performs bandwidh pariioning. Video posiion [seconds] [seconds] Figure 11. Video availabiliy averaged over 1 simulaion runs. (λ = joins per second, 1/µ = 12 seconds, and α =.5 are used.) VI. EXPERIMENTAL RESULTS In he experimens in he ns-2 simulaor [36], 15 peers join and leave repeaedly wih an average lifeime of 12 seconds. We used he Moher and Daugher sequence, encoded a R = 42 kbps, using H.264/AVC wih an inraframe period of 15 frames. Two B frames were encoded beween anchor frames. The video server s uplink bandwidh was se o 2R. 2R was allocaed o LS peers, and he res of he server s uplink bandwidh was allocaed o TS peers. The peer uplink bandwidh was se o 2R and he download bandwidh was se o 4R. On joining a ime, half of he peers wached he live sream. The res of hem wached he sream shifed by a random delay ranging from o. Firs, we examine he sysem wihou fas prefeching. Fig. 11 is he plo of he experimenal resuls for video availabiliy among peers. The value associaed wih he pair (, x) represens he average number of peers ha possess video a posiion x a ime. The resul closely maches he video availabiliy prediced by he model in Secion IV and shown in Fig. 5. Fig. 12 shows he video availabiliy for he model and he experimens in erms of he number of TS peers ha possess video a posiion x ( x 9) a ime 9s (no LS peers were presen). I is noable ha video availabiliy peaks a 45s, he half poin of he video available a 9s, under he assumpion of random video access. As wih Fig. 5 and 11, he video availabiliy prediced by he model closely maches he averaged video availabiliy of he experimens. As he session proceeds, only a small neighborhood of he video near he live sream is cached by he majoriy of he peers. The exen of he neighborhood is deermined by several facors, one of which is he average peer lifeime. The expeced number of peers ha cache a pas porion Number of peers Model Experimenal daa Video posiion [seconds] Figure 12. Video availabiliy a 9 seconds. Averaged over 1 simulaion runs. Only TS peers (α = ) joined he sysem a he rae of λ = joins per second. 1/µ = 12 seconds. Nex, we sudy he sysem behavior when fas prefeching is employed. In Fig. 13, he average number of direc children of he server is ploed for hree differen cases: No fas prefeching is performed. Fas prefeching wih paren selecion minimum hops. For boh TS and LS peers, parens ha minimize he number of logical hops o he server are seleced. Fas prefeching wih paren selecion maximum hroughpu. The paren selecion algorihm described in Secion V-B is used (D is se o 1 seconds). The graphs show ha wih fas prefeching, fewer connecions direcly o he server are esablished because peers find heir requesed video among oher peers more ofen han wihou fas prefeching. The proposed paren selecion algorihm furher reduces he server load, compared o he case where paren selecion minimum hops was used. As he session progresses, more TS peers connec o he server. This causes peers o experience less peer churn, leading o higher video qualiy a he cos of server

9 21 JOURNAL OF COMMUNICATIONS, VOL. 7, NO. 3, MARCH 212 bandwidh. Since only few TS peers need o connec o he video server a he early sage of a session, we evaluae video qualiy for he firs 9 seconds. When here are missing or lae packes ha do no arrive in ime, he las fully decoded frame is displayed unil he nex I frame arrives. Table I summarizes average video qualiy and server load. The average PSNR is compued a he end of he simulaion by aking he average of he PSNR value of he video frames displayed a all peers. For missing frames, copy error concealmen (display of he previously decoded frame) is used. The able indicaes ha fas prefeching achieves higher or similar video qualiy wih reduced server load. We observe ha fas prefeching wih paren selecion maximum hroughpu ouperforms no fas prefeching by 2 db for TS peers wih abou 4% sever load reducion. In Fig. 13, he server load remains low a he early sage of he session because TS peers can easily find peers o connec o. Toward he end of he session, in conras, more TS peers need o receive video from he video server as fewer peers are found o cache video. Finally, we invesigae he sensiiviy of fas prefeching agains D. Recall ha D denoes a look-ahead ime, he ime a TS peer looks ahead o compue he expeced download. In he experimen, we had 75 TS peers wih uplink speed of 2R, wih 6 seconds of average lifeime and 6 seconds of average off-ime. Simulaions ran for 9 seconds and he average download speed a which a peer received video was averaged over 2 simulaions. Fig. 14 shows he average download rae, which is normalized by he video birae. We observe ha he average download peaks near and decreases as D increases, indicaing ha large D may overesimae expeced downloads because seleced parens may depar he sysem. However, he download speed is no very sensiive o he value of D, showing only a 3% drop from he peak value for 1 D 2. The download rae becomes nearly fla as D approaches he average peer lifeime. I is worh noing ha when D, he maximum-hroughpu algorihm in Eq. 7 can be viewed as a maximum-download-rae algorihm; when D and l i is moderae, min (D q i + l i, D r i ) approximaes o D r i, hereby arg max i min (D q i + l i, D r i ) arg max i r i. Server load [number of direc children] w/o prefeching minimum hops maximum hroughpu [seconds] Figure 13. Comparison of server load over ime (averaged over 1 simulaion runs). Normalized average download rae D [seconds] Figure 14. Normalized average download rae wih respec o D, lookahead ime. TABLE I. EFFECTS OF FAST PREFETCHING (COMPUTED OVER THE FIRST 9 SECONDS). VIDEO QUALITY IS REPRESENTED BY THE AVERAGE PSNR OF PEERS. SERVER LOAD IS THE AVERAGE NUMBER OF DIRECT CHILDREN OF THE SERVER. Hops AND hroughpu ARE PARENT SELECTION CRITERIA. Paren Average Average Server selecion PSNR PSNR load crierion LS peers TS peers [num. of [db] [db] children] Wihou prefeching hops Wih hops prefeching hroughpu VII. CONCLUSIONS In his paper, a ree-based P2P live sreaming sysem like SPPM is shown o suppor ime-shifed sreaming wihou any fundamenal changes in he sysem archiecure. To undersand he disribuion of video segmens among peers, we modeled he average number of peers ha possess he required video segmens cached in heir buffers, defined as video availabiliy. The model and he experimens revealed ha few peers reques ime-shifed sreams from he server a he early sage of he session because hey can connec o peers who have been caching he live sream, and ha he server load increases as he session progresses since a smaller number of peers are likely o have he requesed ime-shifed sreams. The video availabiliy improves more significanly when fas prefeching is combined wih he maximum hroughpu paren selecion. This shows ha he overlay may look very differen depending on wha objecives are pursued in he overlay consrucion.

10 JOURNAL OF COMMUNICATIONS, VOL. 7, NO. 3, MARCH In his work, we considered he coninuous bisream consising of pas video packes sored in peers local sorage. One can devise more sophisicaed caching algorihms, which may allow muliple chunks of video ha are no necessarily coninuous in bisream. Server bandwidh allocaion is also an imporan research problem. We have found ha ime-shifed sreaming ofen requires more server bandwidh han live sreaming because peers ha reques pas porion of video need o connec o he server if he requesed video chunk is no available from oher peers. Thus, he server needs o divide is finie uplink bandwidh beween live sreaming and ime-shifed sreaming dynamically in order o maximize he overall video qualiy of peers. REFERENCES [1] M. Cha, H. Kwak, P. Rodriguez, Y.-Y. Ahn, and S. Moon, I ube, you ube, everybody ubes: analyzing he world s larges user generaed conen video sysem, in Proceedings of he 7h ACM SIGCOMM conference on Inerne measuremen, ser. IMC 7. New York, NY, USA: ACM, 27, pp [2] Y. Shen, Z. Shen, S. Panwar, K. Ross, and Y. Wang, On he design of prefeching sraegies in a peer-driven video on-demand sysem, IEEE Inern. Conf. Mulimedia and Expo, pp , Jul. 26. [3] S. Sheu, K. Hua, and W. Tavanapong, Chaining: a generalized baching echnique for video-on-demand sysems, in IEEE Proc. Inern. Conf. Mulimedia Compuing and Sysems, Oawa, Canada, Jun. 1997, pp [4] K. A. Hua, Y. Cai, and S. Sheu, Paching: A mulicas echnique for rue video-on-demand services, in Proc. ACM Inern. Conf. Mulimedia, Universiy of Brisol, Unied Kingdom, 1998, pp [5] F. A. T. S.-H. G. Chan, Tradeoff beween sysem profi and user delay/loss in providing near video-on-demand service, IEEE Trans. Circuis and Sysems for Video Technology, vol. 11, no. 8, pp , Aug. 21. [6] B. Qudah and N. Sarhan, Towards scalable delivery of video sreams o heerogeneous receivers, in Proc. ACM Inern. Conf. Mulimedia, Oc. 26, pp [7] Y. Cui, B. Li, and K. Nahrsed, osream: Asynchronous sreaming mulicas in applicaion-layer overlay neworks, IEEE Journ. on Seleced Areas on Comm., vol. 22, no. 1, pp , 24. [8] D. Wang and J. Liu, Peer-o-peer asynchronous video sreaming using Skip Lis, Proc. of IEEE Inernaional Conference on Mulimedia and Expo, 26. [9] Y. Guo, K. Suh, J. Kurose, and D. Towsley, P2Cas: peero-peer paching scheme for VoD service, in Proc. Inern. Conf. World Wide Web. ACM Press New York, NY, USA, May 23, pp [1] T. Do, K. Hua, and M. Tanaoui, Robus video-on-demand sreaming in peer-o-peer environmens, Compuer Communicaions, vol. 31, no. 3, pp , Feb. 28. [11] H. Chi, Q. Zhang, and S. Shen, Efficien search and scheduling in P2P-based media-on-demand sreaming service, IEEE Journ. on Seleced Areas on Comm., vol. 25, no. 1, pp , Jan. 27. [12] W. Yiu, X. Jin, and S. Chan, VMesh: disribued segmen sorage for peer-o-peer ineracive video sreaming, IEEE Journ. on Seleced Areas on Comm., vol. 25, no. 9, pp , 27. [13] Y. Guo, K. Suh, J. Kurose, and D. Towsley, Direcsream: A direcory-based peer-o-peer video sreaming service, Compuer Communicaions, Jan. 28. [14] T. Do, K. A. Hua, and M. Tanaoui, P2VoD: providing faul oleran video-on-demand sreaming in peer-o-peer environmen, vol. 3, June 24, pp [15] C. Huang, J. Li, and K. W. Ross, Peer-assised vod: Making inerne video disribuion cheap, Feb. 27. [16] X. Xu, Y. Wang, S. Panwar, and K. W. Ross, A peer-opeer video-on-demand sysem using muliple descripion coding and server diversiy, Inern. Conf. Image Processing, vol. 3, pp , Oc. 24. [17] Z. Liu, Y. Shen, S. Panwar, K. W. Ross, and Y. Wang, Efficien subsream encoding and ransmission for P2P video on demand, in IEEE Proc. Inern. Packe Video Workshop, Nov. 27, pp [18] S. Annapureddy, S. Guha, C. Gkansidis, and D. Gunawardena, Exploring VoD in P2P swarming sysems, in IEEE Proc. Inern. Conf. Compuer Comm., Anchorage, AK, Jan. 27, pp [19] C. Zheng, G. Shen, and S. Li, Disribued prefeching scheme for random seek suppor in peer-o-peer sreaming applicaions, in Proc. ACM Inern. Conf. on Mulimedia, P2PMMS Workshop, Singapore, Nov. 25, pp [2] K. M. Ho, W. F. Poon, and K. T. Lo, Video-on-demand sysems wih cooperaive cliens in mulicas environmen, IEEE Trans. Circuis and Sysems for Video Technology, vol. 19, no. 3, pp , Mar. 29. [21] Weihui, J. Xu, and C. Qing, Dynamic mulicas ree o suppor ime-shifing in real-ime sreaming, in Inernaional Conference on Inelligen Compuing and Cogniive Informaics, Los Alamios, CA, USA, 21, pp [22] X. Hou and R. Ding, Research of providing live and imeshifing funcion for srucured P2P sreaming sysem, in Proceedings of he 21 Inernaional Conference on Machine Vision and Human-machine Inerface, Washingon, DC, USA, 21, pp [23] S. Deshpande and J. Noh, P2TSS: -shifed and live sreaming of video in peer-o-peer sysems, in Proc. of IEEE Inernaional Conference on Mulimedia and Expo, 28, pp [24] J. Noh, P. Bacciche, F. Harung, A. Mavlankar, and B. Girod, Sanford Peer-o-Peer Mulicas (SPPM) overview and recen exensions, Proc. of Inernaional Picure Coding Symposium (PCS), Chicago, Illinois, USA, invied paper,, May 29. [25] V. N. Padmanabhan, H. J. Wang, P. A. Chou, and K. Sripanidkulchai, Disribuing sreaming media conen using cooperaive neworking, in Proc. of ACM NOSSDAV, Miami Beach, FL, Miami, FL, May 22, pp [26] Y. Chu, S. Gao, and H. Zhang, A case for end sysem mulicas, in Proc. ACM Sigmerics, Sana Clara, CA, 2, pp [27] J. Noh, P. Bacciche, and B. Girod, Experiences wih a large-scale deploymen of Sanford Peer-o-Peer Mulicas. Seale, WA: IEEE Proc. Inern. Packe Video Workshop, May 29, pp [28] E. Seon, J. Noh, and B. Girod, Congesion-disorion opimized peer-o-peer video sreaming, Inernaional Conference on Image Processing (ICIP), Alana, Georgia, pp , Oc. 26. [29], Low-laency video sreaming over peer-o-peer neworks, IEEE Inern. Conf. Mulimedia and Expo, pp , Jul. 26. [3] J. Noh, A. Mavlankar, P. Bacciche, and B. Girod, shifed sreaming in a peer-o-peer video mulicas sysem, in Proc. of IEEE Global Communicaions Conference (GLOBECOM 29), Honolulu, Hawaii, USA, Nov./Dec 29. [31] B. Chang, L. Dai, Y. Cui, and Y. Xue, On Feasibiliy of P2P on-demand sreaming via empirical VoD user behavior analysis, in Proc. Inern. Conf. Disribued Compuing Sysems Workshops, Beijing, China, 28, pp

11 212 JOURNAL OF COMMUNICATIONS, VOL. 7, NO. 3, MARCH 212 [32] Y. Chu, A. Ganjam, T. Ng, S. Rao, K. Sripanidkulchai, J. Zhan, and H. Zhang, Early experience wih an Inerne broadcas sysem based on overlay mulicas, in USENIX Annual Technical Conference, June 24, pp [33] B. Li, S. Xie, G. Y. Keung, J. Liu, I. Soica, H. Zhang, and X. Zhang, An empirical sudy of he Coolsreaming+ sysem, IEEE Journ. on Seleced Areas on Comm., vol. 25, no. 9, pp , Dec. 27. [34] S. Ross, Sochasic processes, 2nd ed., [35] P. Bacciche, J. Noh, E. Seon, and B. Girod, Conenaware P2P video sreaming wih low laency, in IEEE Proc. Inern. Conf. Mulimedia and Expo, Beijing, China, 27, pp [36] The Nework Simulaor - NS-2, hp:// nsnam/ns, seen on June 1, 21. Jeonghun Noh received an M.S. degree in Elecrical Engineering from Seoul Naional Universiy, Korea, in 1999, and a Ph.D. in Elecrical Engineering from Sanford Universiy, CA, in 21, respecively. He is currenly a senior sysems engineer a ASSIA inc., Redwood Ciy, CA, USA. His research ineress include lowlaency daa disribuion, peer-o-peer sreaming, mobile video sreaming, and broadband access neworks. He has auhored or co-auhored over 17 conference papers and one journal paper, and holds wo US paens graned and seven US paens under review. He worked as a summer research inern a Sharp Laboraories of America, Camas, WA, in 26 and 27. He worked as a consulan a Dyyno Inc., Palo Alo, CA, in 28. He collaboraed wih Deusche Telekom Laboraories in Los Alos, CA, and in Berlin, Germany, in He is a recipien of he ACM Mobimedia Bes Suden Paper Award in 29 and he IEEE CCNC Bes Presenaion Award in 28. Dr. Noh is a member of IEEE, IEEE ComSoc, and IEEE MMTC. Bernd Girod is Professor of Elecrical Engineering and (by couresy) Compuer Science in he Informaion Sysems Laboraory of Sanford Universiy, California, since Previously, he was a Professor in he Elecrical Engineering Deparmen of he Universiy of Erlangen-Nuremberg. His curren research ineress are in he areas of video compression, neworked media sysems, and image-based rerieval. He has published over 45 conference and journal papers, as well as 5 books, receiving he EURASIP Signal Processing Bes Paper Award in 22, he IEEE Mulimedia Communicaion Bes Paper Award in 27, he EURASIP Image Communicaion Bes Paper Award in 28, as well as he he EURASIP Technical Achievemen Award in 24. As an enrepreneur, Professor Girod has been involved wih several sarup venures, among hem Polycom, Vivo Sofware, 8x8, and RealNeworks. He received an Engineering Docorae from Universiy of Hannover, Germany, and an M.S. Degree from Georgia Insiue of Technology. Prof. Girod is a Fellow of he IEEE, a EURASIP Fellow, and a member of he German Naional Academy of Sciences (Leopoldina).