Proposed an Adaptive Bitrate Algorithm based on Measuring Bandwidth and Video Buffer Occupancy for Providing Smoothly Video Streaming

Transcription

1 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, ol. 9, No. 2, 218 Proposed a Adaptve Btrate Algorthm based o Measurg Badwdth ad deo Buffer Occupacy for Provdg Smoothly deo Streamg Saba Qasm Jabbar School of Electrocs ad Iformato Egeerg, Huazhog Uversty of Scece ad Techology, Wuha, Cha Dheyaa Jasm Kadhm Electrcal Egeerg Departmet, College of Egeerg, Uversty of Baghdad, Baghdad, Iraq Yu L School of Electrocs ad Iformato Egeerg, Huazhog Uversty of Scece ad Techology, Wuha, Cha Abstract Dyamc adaptve streamg va HTTP (DASH) has bee popular dssemated over the Iteret especally uder the crcumstaces of the tme varyg etwork, whch t s curretly the most challegg for provdg smoothly vdeo streamg va hgh qualty. I DASH system, after completg the dowload of oe segmet, the player estmates the avalable etwork badwdth by calculatg the dowloadg throughput ad the adaptg the vdeo btrate level based o ts estmatos. However, the estmated badwdth the applcato layer s ot accurate due to off-tervals appearace durg the dowloadg process. To avod the ufaress of badwdth estmato by the clets, ths work proposes a logarthmc approach for receved etwork badwdth, whch cludes creasg or decreasg ths badwdth logarthmcally to coverge the far share badwdth (estmated badwdth). After obtag the measured badwdth, a adaptve btrate algorthm s proposed by cosderg ths measured badwdth addto to vdeo buffer occupacy. The vdeo buffer model s assocated wth three thresholds (.e. oe for tal startup ad two for operatg thresholds). Whe the vdeo buffer s level stays betwee the two operatg thresholds, the vdeo btrate wll keep uchaged. Otherwse, whe the buffer occupacy s too hgh or too low, a approprate vdeo btrate s chose to avod buffer overflow/uderflow. Smulato results show that the s able to coverge the measured badwdth to the far share badwdth very quckly. Also the s compared wth covetoal scheme, we foud that our outperforms achevg the best performace terms of effcecy, stablty ad faress. Keywords DASH; vdeo streamg; vdeo buffer; vdeo adaptve btrate algorthm, QoE I. INTRODUCTION I last few years, HTTP adaptve vdeo streamg has bee wdely used Iteret vdeo techologes such as Mcrosoft Smooth Streamg, Netflx, Apples HLS, Adobes HDS ad Akama HD [1], [2], sce the use of HTTP over TCP s easy to cofgure ad smplfy the traversal of frewalls. Moreover, the stadard of HTTP based adaptve vdeo streamg s the Dyamc Adaptve Streamg over HTTP (DASH) [3]. I DASH systems, each vdeo s ecoded to multple represetatos of dfferet btrates ad each represetato s dvded to multple segmets (2-1 secods of vdeo tme). At the clet sde, a clet requests ad receves vdeo segmets from DASH servers that ow the segmets cotuously. Oe of the most mportat features DASH system s adaptg the vdeo btrate to a varyg etwork badwdth dyamcally. I such way, DASH clets ca ejoy wth vdeo at maxmum qualty possble sce they receve vdeo segmets from dfferet versos each of whch beg ecoded wth specfc btrates a way throttlg the vsual qualty to match the avalable etwork badwdth. It s mportat challegg to satsfy the user experece durg a whole vdeo sesso uder tme-vary etwork codtos. Wthout a powerful adaptve btrate algorthm, DASH s clet may face frequet terruptos ad whch t s degraded hs vdeo qualty. For example, f the selected vdeo btrate s hgher tha the avalable badwdth, t wll cause etwork cogesto. O the other had, f the vdeo btrate s lower tha the avalable badwdth, the vsual vdeo qualty would ot reach the maxmum allowed by the avalable badwdth. Besdes, smooth vdeo btrates ad stable btrates swtchg are preferred through vdeo playback [4]. Due to the O-Off phase whch t s a atural pheomeo DASH system (ON meas dowloadg a segmet ad OFF meas stayg dle), the competg clets may face dffcultly estmatg ther badwdths whch usually leads to several performace problems such as effcecy, stablty, ad ufaress [5]. There are may research problems about vdeo streamg for multple DASH clets over TCP etwork stll ope ad challegg. As a example, the tradeoff betwee the stablty ad effcecy whch t s very mportat ssue to have smooth vdeo btrates. Uder tme-varyg badwdth codto, requestg a low vdeo btrate wll produce more opportutes for rate selecto, ad therefore well esure hgh stablty (smoothess) wth cotuous vdeo playback but t also causes a low vdeo qualty wth low badwdth usage (effcecy). Aother tradeoff challegg s betwee sestvty ad stablty, sce the chael badwdth s heretly dfferet tme, the hgh sestvty of the btrate cotrol techque usually makes the vdeo btrate detcal wth the badwdth ad hece leadg to hgh stablty. Such these challeges become more troublesome whe multple clets compete o the bottleeck lk because each clet wll try to optmze the qualty of the vdeo wthout lookg at others. Besdes, the faress problem arses for multple clets DASH 68 P a g e

2 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, ol. 9, No. 2, 218 techology sce they are deployed va HTTP/TCP etwork. The exstg btrate adaptve algorthms for DASH system are amg to ether acheve the effcecy of hgh badwdth usage for vdeo adaptato btrate to match avalable badwdth, or mata cotuous vdeo playback by homogeety the vdeo btrate to avod buffer overflow/uderflow such as badwdthbased approaches dscussed [6]-[8] ad buffer-based approaches [9][1].Recetly, the problems wth vdeo streamg for multple DASH clets competg over a commo bottleeck have bee studed may researches such as [11]-[13]. I [11] ad [12], the authors studed problems of btrate adaptato ad determed the causes of may uwated teractos that arse as a result of modfyg the vdeo btrate over HTTP. Furthermore, a seres of techques have bee developed that have tred to systematcally gude the tradeoffs betwee stablty, faress ad effcecy. However, these techques dd ot look well showg the factors that affect the qualty of experece, such as the vdeo btrate oscllato ad vdeo playback terruptos. Whle [13], a Markov Decso Process (MDP) s used to deal wth the stochastc decso problem, whch reduces both the umber of playback terruptos ad the umber of qualty level swtches whle creases the qualty of experece. Besdes rate adaptato, there are some other research works addressg the faress problem from several aspects such as the work [14] ams at achevg relatve faress at the packet level through the mplemetato of the weghted far lst. At [15], a ew protocol s proposed the cotext of complex mult-server adaptato, amed at mprovg the user experece by provdg the best faress, effcecy ad stablty. The problem of faress has bee addressed by the applcato of a server o traffc shapg as [16]. I our work, a btrate adaptve approach s desged o the bass of stablty ad faress of DASH system uder the scearo that may clets compete for etwork resources. I ths work, to avod the ufar badwdth that was estmated by the clet due to off tervals durg the dowloadg process, a cremet method s desged based o estmated badwdth for covergg the measured badwdth to the estmated badwdth through creasg scheme. Exclusvely, whe the measured badwdth s smaller tha the estmated badwdth (.e. estmated badwdth s equal to the sze of a segmet dvded by the tme t takes to dowload ths segmet), a logarthmc law based cremet scheme s desged for covergg the measured badwdth toward share badwdth effcet way amog competg clets. I cotrast, whe measured badwdth s hgher tha estmated badwdth, a coservatve reducg scheme s desged for reducg the measured badwdth to avod cogesto. O other sde for stable ad smooth vdeo btrate, we propose a adaptve btrate algorthm that takes the measured badwdth based o cremet scheme ad the buffer occupacy level to accout for keepg a cotuous vdeo playback ad selectg the best vdeo btrate to dowload the ext segmet. I our approach vdeo buffer s assocated wth three predefed thresholds for keepg the buffer stable state. Usg ths model, clets wll dowload vdeo segmets cotuously wthout the eed for OFF terval. Especally, whe the share badwdth s hgher tha the selected vdeo btrate (wth the codto the vdeo btrate ot exceed the measured badwdth), the buffer occupacy wll crease ad coverge toward the maxmum level. The adaptve algorthm wll swtch the btrate a way prevetg buffer overflow, ot dle ad vce versa. Through smulato results, our approaches outperform the exstg algorthms measurg the far share badwdth, achevg faress, buffer stablty ad reducg the umber of vdeo btrates swtchg. II. SYSTEM DESIGN Geerally HTTP clets operate the meda cotet that trasferred from streamg server va avalable etwork badwdth. The Server-Clet model for streamg system s show Fg. 1 below, where the etwork badwdth represets the data producg rate ad vdeo cotet s btrate represets the data cosumg rate to the vdeo buffer of the clet. Fg. 1. Server-clet model for streamg system. Assumg at the vdeo playg tme, the curret durato beg playbacked s fully loaded, the ext durato s playback tme s from t s to te as show Fg. 2, ad T deotes the etwork badwdth. Whle R+1 defes the vdeo btrate for the ext durato +1 to guaratee there s o playback terrupto durg the future terval [t s, t e ]. At the begg of each dowload stage, a adaptato algorthm wll select the sutable btrate of the ext segmet to 1 L R 1, R,..., R,..., R,1 L ad dowload 1 2 L R R... R, whch t wll specfy how much tme s gvg for the curret segmet dowload utl the ext dowload request,.e. t est (+1), whch t s the estmated requred tme for dowloadg the ext segmet). Hece the clet tates a HTTP GET request to the server for the segmet of sequece umber +1 wth R+1 the the dowloadg s startg. Let t m =[t e - t s ] be the measured tme that s requred to complete the dowload. Assumg that o ppelg of dowloadg s volved, the ext dowload step starts after the followg tme that ca be expressed as follows: act est m t ( 1) max(t ( 1), t (+1)) (1) where t act (+1) s the actual tme to dowload (+1) th segmet. If the dowload durato t(+1) s shorter tha the t est (+1) the a clet must wat tme [t est (+1)- t m (+1)] whch s the off-terval before startg the ext dowloadg step (case A), otherwse the clet starts the ext dowload step mmedately after the curret dowload s completed (case B) P a g e

3 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, ol. 9, No. 2, 218 Fg. 2. HTTP segmet dowloadg process. III. SHARED BANDWIDTH APPROACH To expla the effect of badwdth cosumpto o adaptg the sutable vdeo btrates, three dfferet scearos are cosdered: deal partcpated, more partcpated ad less partcpated whch they represet respectvely the amout of badwdth asked by clet to be equal to, more tha ad less tha the shared badwdth. We assume a temporal overlap of the ON-OFF tervals amog two competg clets. Suppose the avalable badwdth s T ad two actve coectos share ths badwdth farly,.e. each has T/2.So T 1 ad T 2 are the badwdth receved by clet1 ad clet2 respectvely. I deal partcpated, the total amout of traffc s requested by the two clets perfectly fllg the lk,.e. each clet explots the avalable badwdth durg ts ON perod T 1 = T 2 =T/2. If the shared badwdth s overestmated, ew segmets wth hgh vdeo btrates maybe asked by the clets who lead to cogesto. I ths case, the curret estmated badwdth s less tha the prevous estmato, ad hece clets wll tur to ask segmets wth lower vdeo btrates. Ths fluctuato vdeo qualty level leads to vdeo buffer stablty as show Fg. 3(a). Fg. 3(b) shows the case of more partcpated where the ON durato of oe clet falls the ON durato of the other clet. Ths happeed whe the clets are stll askg vdeo segmets for dfferet vdeo bt-rates. I ths stuato, the motored throughput by the clets s T 1 >T/2 ad T 2 =T/2,.e., frst clet overestmates the far share of badwdth. Whe oly oe clet overestmates the far share of badwdth, the two clets wll coverge to a stable but ufar equlbrum that the clet, who overestmates the far share of badwdth, wll request a hgher vdeo btrate, causg ufaress. Whle Fg. 3(c) dscusses the problem of mperfect usage of the shared badwdth whe the ON durato for the two clets are alged, the avalable badwdth at the server sde s dscrete ad lmted ths case clets estmate the far share of badwdth correctly, t stll may cause mperfect utlzato problem. For example f the avalable shared badwdth s ot T/2, clets wll ask segmets wth vdeo btrates smaller tha T/2 to avod buffer uderflow. The OFF perods should adopt to solve the buffer overflow case ad mperfect utlzato problem happes. Accordgly, the badwdth oscllato wll mpact o vdeo btrates stablty whe the badwdth s more partcpated case (.e., the cogesto occurs), the the badwdths are estmated by the clets approxmately equal, ad faress ca be obtaed. Whe cogesto occurs, the vdeo buffer at the clet sde wll be draed, causg playback stop sce clets may ask segmets wth btrates more tha the shared badwdth, whch may degrade the vdeo qualty of experece. Fg. 3. Two clets are competg shared badwdth durg dowload a segmet. Whe the badwdth s less partcpated case, the requested vdeo btrate s smaller tha the avalable badwdth, ad OFF perods s eeded to stop the trasmsso so as to avod buffer overflow. I ths case, whe the clet buffer reaches the maxmum level Bmax, ths wll lead to badwdth overestmato ad the vdeo btrates wll oscllate accordgly. But the avalable shared badwdth caot cotue wth the same vdeo btrate, the vdeo buffer falls below Bmax, so the clet returs to the buffer growg mode ad the off-tervals dsappear, so the lk aga s gog to be more partcpated case ad the measured throughput starts to coverge to the farshare badwdth. Fally, due to the quatzato mpact, the requested vdeo btrate falls below the far-share badwdth, ad the clet buffer starts growg aga, completg oe oscllato cycle. I. IDEO BITRATE ADAPTIE APPROACH I ths secto, we wll descrbe a four stages model for DASH btrate adaptato algorthm, whch cludes both the covetoal algorthms (e.g., [17], [18]) ad our proposed btrate adaptato algorthm ths work, whch wll serve as a bechmark. Essetally, a btrate adaptato algorthm proceeds the followg four stages: a) Estmatg Stage: The algorthm begs the frst step adaptato process by estmatg the etwork throughput whch be used selectg the vdeo btrate. The curret shared badwdth equates to the prevous TCP throughput that motors durg the ON terval,.e.: ˆT (2) T 1 b) Flterg Stage: The a flterg process s doe by usg a smoothed fucto fs() wth the am of removg outlers. The algorthm takes the measured hstory as a put.e.: T f ({T ˆ : j }) (3) s j c) Adaptg Stage: The cotuous measured throughput s adapted to get the sutable vdeo btrate. I ths step a dscrete vdeo btrate s selected based o sde hstory 7 P a g e

4 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, ol. 9, No. 2, 218 formato such as vdeo buffer occupacy, prevous selected vdeo btrate Rj,, etc. The the curret vdeo btrate s gve by: f ( T ;,B(j) : B(j) B, j ) (4) R1 a 1 Rj max d) Next Segmet Dowloadg Stage: A schedulg tme for dowload the ext segmet s determed by the adaptve algorthm as a mechacal fucto based o vdeo buffer sze. If the curret buffer occupacy B() s less tha the maxmum level B max, the the estmated tme for ext dowload s set to ad the dowload tme for ext segmet begs after the curret dowload s fshed. The the estmated tme for ext dow s gve by: But the tme s τ (..e. : t est, B() B, B() B where τ s the vdeo segmet perod to stop the vdeo buffer from growg. I geeral, all of today s commercal DASH clets mplemet the frst stage ad last stage of the rate adaptato algorthm a smlar way, though they may dffer ther mplemetato of the secod stage ad thrd stage of the algorthm. I ths work we make some modfcatos to these stages sce the off tervals may become a source of ambguty for a clet to correctly perceve ts far share of the etwork badwdth, thus prevetg the clet from makg accurate rate adaptato decsos. To esure faress, we focus o prevetg the happestace of OFF pheomeo, that s, the badwdth s ether the deal partcpated or more partcpated. I the case of more partcpated, faress s guarateed, whle the deal of partcpated, there s a ulmted umber of badwdth-sharg modes. It s worth to ote that although t s easy to avod OFF pheomeo by keepg o dowloadg wthout break wth them, t s mpossble eve to dstgush f the badwdth s more partcpated or the deal partcpated. Moreover, because the avalable vdeo btrate s lmted ad dscrete, stead of usg the estmated badwdth, we suggest a effectve badwdth cremet scheme to gude the adaptve vdeo btrate. The stablty s take to accout through proposg double threshold of buffer level, ad the the frequecy of fllg the buffer occupacy ad the mpact of badwdth varatos o the adaptve vdeo btrate ca be reduced. As per durg the rate adaptato process, the smoothess s also take to accout ad t wll be dscussed deeply ext secto. I DASH system, the vdeo btrate ca be chaged oly whe a segmet s completely dowloaded. The, the vdeo btrate for the ext segmet to be dowloaded s calculated accordgly, we frst estmate the etwork badwdth ad flter t. The, wth ths estmated badwdth, we update the cremeted badwdth whch wll be used to gude the rate adapto. Accordg to the buffer occupacy, the vdeo btrate s adapted whe the buffer occupacy s betwee the two predefed thresholds to avod buffer overflow ad uderflow. max max (5). SMOOTHING NETWORK BANDWIDTH I ths secto, we wll descrbe the prcples of our approach to smooth the etwork badwdth estmato ad how flterg t, also we wll descrbe our Tral cremet approach based o badwdth estmato ad the we wll expla detals our proposed stable ad smooth adaptve btrate algorthm. A. Network Badwdth Estmato ad Flterg Assume the ecodg of a vdeo segmet s doe to dfferet versos of L, wth playback of dfferet vdeo 1 L,...,,..., set. All vdeo versos btrates R R R are dvded to segmets of equal legth, each of whch cosumes the same playg tme τ. For each clet, the streamg process s dvded to sequetal segmet dowloadg steps =1, 2, 3 the, the requred badwdth to dowload (+1) th segmet, s estmated as [6]: ˆT (6) R1 1 1 tm where τ s segmet durato. The dowloaded segmets are stored the vdeo buffer. Furthermore, to remove the mpact of ose ad terferece durg the recevg badwdth estmato, a ose-flter step s adopted [19]. The, we have: Y T ˆ (1 )Y (7) 1 where Y +1 s the avalable modfed badwdth after flterg step. s a smooth coeffcet factor that reflects the chagg etwork codtos; t s measured as: x z (8) where x ad z represet the absolute ad smoothed values of the throughput devato (e ) respectvely ad they are gve by: x e (1 ) x 1 z e (1 ) z 1 where ρ s a costat wth 1. At ay stat e s determed as: e Tˆ Y (1) where e s dfferece betwee the measured value ad adjusted value. If s a small value, the prevous adjusted badwdth may play a more mportat role predctg the badwdth. As gets larger, the adjusted badwdth s closer to the curret measured value. Thus, f the varato of badwdth s large, we should decrease, ad f the varato of badwdth s small, we ca crease to more accurately reflectg the chage of the etwork codtos. (9) 71 P a g e

5 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, ol. 9, No. 2, 218 B. Tral Icremet Approach Due to the o-off pheomeo the DASH system, the badwdth estmated by the clet s dscretoary ad caot be used drectly for the adjustmet rate. Addtoally, oly whe badwdth s uder more partcpated case, the cogesto happes, so badwdth estmated by clet equals to a far-share badwdth sce all clets see almost the same badwdth as avalable. O the other had, whe cogesto occurs, the requred vdeo btrate caot be supported by badwdth, ad vdeo freezg may occur. Thus, how to get a free badwdth quota wthout cogesto s a crtcal ssue to mprove the performace of the rate adaptato for DASH system. I ths work, we propose a logarthmc crease scheme based o receved etwork badwdth, whch cludes creasg the badwdth logarthmc ad decrease t f the cremet badwdth more tha the far share badwdth for makg the measured badwdth quckly coverge to the far share badwdth. The am behd ths scheme s that the estmated badwdth by a clet s always the upper boud of the far share badwdth due to the off tervals. Thus, durg the cremet phase, the measured badwdth wll cotuously crease utl t exceeds the estmated badwdth expressed (7). The, whe t s hgher tha the estmated badwdth, the clet wll swtch to the decrease phase for avodg cogesto. Let Y be the measured badwdth whch would be used 1 for adaptg the sutable vdeo btrate to (+1) th segmet ad t s talzed to zero that Yt. Wheever the dowloaded process of segmet s completed, the estmated badwdth s update accordg to expresso (7). The, the measured badwdth s updated as follows: 1) Icremet Phase: Y Y Y Y max,, f Y Y 2 1 (11) where s a costat to avod slow covergece, the measured badwdth wll quckly approach to the estmated badwdth the logarthmc way. However, whe the gap betwee Y ad Y s small,.e. Y Y 2, the measured badwdth wll be addtvely creased by rather tha by logarthmc approach. Otherwse whe the gap betwee the measured badwdth ad the far share badwdth s large, a aggressve way should be employed to crease the measured badwdth quckly. Practcally, the clet far share badwdth practce caot get, so stead ts upper boud badwdth (.e., the estmated badwdth) s used. 2) Decremet Phase: Y 1 Y Y Y, f Y Y (12) where β s a postve costat satsfyg β > 1. Whe the measured badwdth exceeds the estmated badwdth, a coservatve way should be employed to cotrol Y to be ot hgher tha Y.e. whe the measured badwdth s hgher tha the far share badwdth, cogestos may happe, so leadg to playback freezg. I ths case, a coservatve way s desged to decrease the measured badwdth guarateeg that o cogesto happes. C. Proposed Adaptve Btrate Algorthm At the clet aspect, the vdeo player uses a adaptve algorthm to determe the sutable btrate to be selected for dowload the ext segmet. Every segmet that s dowloaded placed a buffer of maxmum sze B max. The buffer s assocated wth three thresholds (B, B low, B hgh ) where B < B low < B hgh ad represets the set of avalable btrates as cleared Fg. 4. After gettg the measured badwdth the prevous secto, to obta a smooth ad stable vdeo btrate wth hgh badwdth utlzato a btrate adaptato scheme s desged based o the buffer occupacy wth the measured badwdth. I ths work, the buffer occupacy s deoted by the buffered vdeo durato cosderg that the clet buffer may cota segmets from dfferet versos,.e., dfferet vdeo btrates ad there s o loger a drect mappg betwee the buffered vdeo sze ad the buffered vdeo durato. Fg. 4. Clet vdeo buffer. The smooth btrate adaptato scheme s desged wth the am to provde a smooth vdeo qualty ad avod buffer overflow/uderflow. The mpact of short term badwdth oscllatos o vdeo btrates ca be elmated by usg thresholds for buffer occupacy. At the begg of the vdeo sesso (or after a terrupto due to buffer empty state), the umber of segmets the buffer s below B ad the vdeo playback begs durg whch the lowest btrate s selected. Ths step s ecessary because reducg the startup tme s mportat to prevet the user from gvg away the vdeo sesso. Whe the share badwdth lk s more partcpated case,.e., the cogesto occurs, the badwdths estmated by the clets are approxmately equal, ad the faress ca be obtaed. Whe cogesto occurs, the vdeo buffer at the clet sde wll be uderflow, causg playback stop sce clets may ask segmets wth btrates more tha the share badwdth, whch degrades vdeo qualty of experece. So the rate adaptato scheme decreases the vdeo btrate to mmum level. Oce the vdeo buffer exceeds B ad lower tha B low, to avod buffer drag ad provde a cotuous playback, the vdeo btrate should be selected o hgher tha the measured badwdth. Smlarly, whe the buffer occupacy s hgher tha B hgh, to esure o buffer overflow occurs whle achevg hgh badwdth utlzato, a hgh vdeo btrate ca be selected. Whe the vdeo buffer occupacy falls betwee the two thresholds B low ad B hgh, the rsk of buffer overflow/uderflow s low, ad the vdeo btrate should be mataed uchaged so that smooth qualty s provded. Cosder a vdeo cotet s coded to L represetatos wth dfferet vdeo btrates. The 72 P a g e

6 Network Badwdth (Kbps) Network Badwdth(Kbps) (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, ol. 9, No. 2, 218 set of vdeo btrates s deoted by 1 L,...,,...,,1 L. B s the curret buffer R R R occupacy after dowload th segmet ad measured badwdth Y for ext segmet. The vdeo btrate for ( +1) th segmet s 1 selected as: m{, R R R Y 1 } f B B 1 max {, R B B, R R R Y 1 } f B B ˆT low 1 L 1 R 1 m, R B B, R R low R Y 1 f B B ˆT hgh 1 L 1 R f B low B B hgh (13) At the startg of the vdeo sesso, the umber of segmets the buffer s below B ad the rate adaptato scheme decreases the vdeo btrate to mmum level. Oce the buffer level goes beyod B, the amout of tme requred to load the ext segmet s less tha the B - B, the maxmal vdeo btrate whch s o hgher tha the measured badwdth s selected so as to guaratee a cotuous vdeo playback. Whe the buffer occupacy s hgher tha the B hgh, the mmal vdeo btrate whch s o lower tha the measured badwdth s selected so as to mprove vdeo qualty; otherwse whe the buffer occupacy betwee B low ad B hgh the curret bt rate s preserved. I. SIMULATION RESULTS The s evaluated usg s-3 etwork smulator ad the topology mplemeted ths work s show Fg. 5 below. The topology cossts of a HTTP server, two or more HTTP clets ad a par of etwork elemets (.e. x>=2). To acheve adaptve streamg, the HTTP server offers the clet seve levels of represetatos to adapt the vdeo rates these are ={ 356, 5, 8, 12, 15, 24 ad 35Kbt/s}. The legth of vdeo segmet ad vdeo buffer s 2s ad 35 sec, respectvely. The values of other parameters B low, B hgh, B, β, ρ, are 15s, 3s, 5s, 1.25,.5, 32kbps respectvely. the estmated badwdth (share badwdth) quckly through the accurate logarthmc law based cremet scheme. We ca ote from Fg. 6(a) ad 6(b), whe the secod clet eters the system, t eeds oly about 1s to follow the far share badwdth well, the frst clet detects the chage the avalable etwork badwdth such that ther measured badwdth coverge to the far share badwdth quckly. Whle the other scheme speds a log tme to coverge the measured badwdth to the estmated badwdth lear creasg scheme based o estmated badwdth. The fgures also demostrate that whe two clets compete wth each other, the compared scheme could t track the far-share badwdth well because of the slow start ad frequet fluctuatos, whe the badwdth chages Badwdth bottleeck 1 Frst clet measured badwdth Frst clet estmated badwdth 5 Secod clet measured badwdth Secod clet estmated badwdth Tme(sec) a: Proposed scheme 2 15 Fg. 5. Network topology cofgurato the smulato. To evaluate the performace of the, we compare t wth a covetoal scheme [6] as Fg. 6 ad 7 term of measured badwdth, vdeo btrate ad buffer occupacy, respectvely. The avalable etwork badwdth s equal to 4Mbps wth two clets compete o t. The frst clet eters at the tme s whle the secod clet eters the etwork at 2s ad leaves at 4s. Fg. 6(a) ad 6(b) show the ablty of cremet scheme covergg the measured badwdth to 1 Badwdth bottleeck Frst clet measured badwdth 5 Frst clet estmated badwdth Secod clet measured badwdth Secod clet estmated badwdth Tme (sec) b: Covetoal scheme Fg. 6. Network badwdth computed by two clets. I Fg. 7(a) ad 7(b) the vdeo btrate uder the proposed scheme s smooth ad stable tha the secod scheme. The covetoal scheme uses the lear creasg method measurg the share badwdth whch s used for adaptg the vdeo btrate hece ths leads to vdeo btrate oscllato for matchg the measured badwdth. I addto the vdeo buffer 73 P a g e

7 deo btrate(kbps) deo btrate(kbps) (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, ol. 9, No. 2, 218 occupacy s always hgh to esure cotuous vdeo playback. Sce the vdeo btrate s separate, the selecto of sutable btrate s usually lower tha the measured badwdth whch leads to low badwdth utlzato effcecy. Whle the vdeo btrate of our scheme s smoother due to may reasos; at the begg the adaptato of btrate s't depedg o the measured badwdth oly, t assocates wth the buffer occupacy. The vdeo buffer s related wth three operatg thresholds for avodg buffer overflow/uderflow ad compactg the short term badwdth oscllato that effects o vdeo btrates. Besde the btrate adaptve algorthm takes may factors to accout that mpact o the QoE for vdeo streamg, volvg maxmze the vdeo btrates ad mmze the vdeo start tme, the umber of bufferg evets, ad the umber of btrate swtchg evets. So from fgure 8a the btrate curve s smoother tha the other scheme. Fally, o buffer overflow/uderflow occurs,.e., vdeo playback wthout break s guarateed Frst clet vdeo btrate Frst clet buffer occupacy 5 Secod clet vdeo btrate Secod clet buffer occupacy Tme(sec) a: Proposed scheme 1 Frst clet vdeo btrate 5 Frst clet buffer occupacy Secod clet vdeo btrate Secod clet buffer occupacy Tme(sec) b: Covetoal scheme Fg. 7. Buffer occupacy ad vdeo btrate of two clets uder badwdth bottleeck. I Fg. 8, we compare our scheme wth other scheme term of effcecy, stablty, ad faress based o the followg metrcs [11]: a) Istablty metrc: Clets are lkely to be sestve to frequet ad mportat vdeo btrate swtches as dcated by some studes. The stablty metrc s defed as: M stablty z1 a R,a R,a 1 z a1 R,a a a (15) The stablty metrc equals to the weghted sum of all btrate swtch steps motored wth the last 1 segmets dvded by the weghted sum of btrates the last z=1 segmets. The weght fucto ω(a) = -a. b) Ufaress metrc: At tme t, the ufaress metrc s defed as: M 1 jafar (16) ufaress t where JaFart s the dex of Ja faress at tme t ad s calculated based o the btrates R over all clets. c) Ieffcecy metrc: Ths metrc s calculated as: M effcecy j T R, j (17) where R,j s the vdeo btrate of th segmet for j th clet ad T s the avalable badwdth. A value close to zero meas that the clets average are usg as hgh a average vdeo btrate as possble to mprove vdeo qualty. I Fg. 8, the umber of clets s chaged from two to twelve wth the avalable etwork badwdth fxed at 8Mbps order to evaluate the performace of our scheme wth other scheme. From the fgure the s outperformg due to the cremet scheme based o badwdth measuremet also vdeo buffer model s assocated wth three operatg thresholds for prevetg buffer uderflow /overflow ad the robust of the adaptve algorthm based o estmatg the dowload tme of ext segmet based o segmet sze. I Fg. 9, our scheme always acheves the lowest effcecy, stablty, ad ufaress,.e., compared wth covetoal scheme. The ca provde hgher ad smoother vdeo btrate, ad besdes, t ca also better guaratee the faress betwee the clets. As the avalable badwdth creases, the performace does ot learly crease or decrease. Ths s because the vdeo btrate s dscrete so that for a gve avalable badwdth, f the farshare badwdth s approxmately equal to the vdeo btrate, geerally better performace ca be acheved, ad vce versa P a g e

8 Ufaress Ufaress Istablty Istablty Ieffcecy Ieffcecy (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, ol. 9, No. 2, covetoal scheme.4.35 covetoal scheme Number of Clets a: Ieffcecy avalable etwork badwdth(mbps) a: Ieffcecy.2.18 covetoal scheme covetoal scheme Number of Clets b: Istablty Number of Clets c: Ufaress covetoal scheme Fg. 8. Performace evaluato whe the avalable etwork badwdth s fxed at 8 Mbps, ad the umber of competg clets vares from 2 to Avalable etwork badwdth(mbps) b: Istablty Avalable etwork badwdth(mbps) c: Ufaress covetoal scheme Fg. 9. Performace evaluato for the two schemes whe the avalable etwork badwdth s chagg from 1Mbps to 8Mbps wth two competg clets P a g e

9 Ufaress Istablty Ieffcecy (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, ol. 9, No. 2, 218 From the fgure our scheme has better performg all stuatos as (a) ad (b), the other scheme has effcet badwdth utlzato, stablty vdeo btrate because the decso of vdeo btrate s based o the measured far share badwdth, whch s always the same (equal to 1 Mbps) ths expermet. I term of faress, the other scheme fluctuates wth creasg the o. of clets whle the performs better sce faress s easy to be mpacted wth creasg the umber of competg clets (Fg. 1) Number of Clets a: Ieffcecy b: Istablty c: Ufaress covetoal scheme covetoal scheme Number of Clets covetoal scheme Number of Clets Fg. 1. The performace assessmet of two schemes uder chagg both the umber of competg clets ad avalable etwork badwdth. II. CONCLUSION I ths work, a btrate adaptve approach s desged o the bass of stablty ad faress DASH system uder the scearo that may clets compete for etwork resources. The s based o estmated badwdth to obta the far-share badwdth (.e. the measured badwdth) by tral cremet mechasm. The collectg the measured badwdth wth the vdeo buffer occupacy, we propose a adaptve vdeo btrates algorthm based o tme eeded for dowloadg the ext segmet to acheve smooth, stable vdeo qualty ad avod buffer uderflow/overflow. From smulato results, our scheme s able to quck coverge the measured badwdth to the far-share badwdth eve wth creasg the umber of competg clets, hece achevg faress badwdth utlty. Besdes, the vdeo btrate s allowed to be hgher tha the measured badwdth, thereby achevg a hgher badwdth utlzato effcecy ad hgher average vdeo bt-rate compared wth a covetoal scheme. REFERENCES [1] Csco,.N.I., 216. Forecast ad Methodology, Whte Paper, Csco. [2] Patos, R. ad May, W., 21. HTTP Lve Streamg draft-patos-httplve-streamg-5. Apple Ic., IETF draft, 23. [3] Sodagar, I., 211. The mpeg-dash stadard for multmeda streamg over the teret. IEEE Multmeda, 18(4), pp [4] Km, J., Care, G. ad Molsch, A.F., 216. Qualty-aware streamg ad schedulg for devce-to-devce vdeo delvery. IEEE/ACM Trasactos o Networkg, 24(4), pp [5] Akhshab, S., Aatakrsha, L., Bege, A.C. ad Dovrols, C., 212, Jue. What happes whe HTTP adaptve streamg players compete for badwdth?. I Proceedgs of the 22d teratoal workshop o Network ad Operatg System Support for Dgtal Audo ad deo (pp. 9-14). ACM. [6] Lu, C., Bouazz, I. ad Gabbouj, M., 211, February. Rate adaptato for adaptve HTTP streamg. I Proceedgs of the secod aual ACM coferece o Multmeda systems(pp ). ACM. [7] Zhou, B., Wag, J., Zou, Z. ad We, J., 212, Jauary. Badwdth estmato ad rate adaptato HTTP streamg. I Computg, Networkg ad Commucatos (ICNC), 212 Iteratoal Coferece o (pp ). IEEE. [8] L, Q., Lu, Y., She, Y., She, H., Sag, L. ad Yag, D., 214, December. Badwdth estmato of rate adapto algorthm DASH. I Globecom Workshops (GC Wkshps), 214 (pp ). IEEE. [9] Huag, T.Y., Johar, R., McKeow, N., Truell, M. ad Watso, M., 215. A buffer-based approach to rate adaptato: Evdece from a large vdeo streamg servce. ACM SIGCOMM Computer Commucato Revew, 44(4), pp [1] Km, S., Yu, D. ad Chug, K., 216, Jauary. deo qualty adaptato scheme for mprovg QoE HTTP adaptve streamg. I Iformato Networkg (ICOIN), 216 Iteratoal Coferece o (pp ). IEEE. [11] Jag, J., Sekar,. ad Zhag, H., 212, December. Improvg faress, effcecy, ad stablty http-based adaptve vdeo streamg wth festve. I Proceedgs of the 8th teratoal coferece o Emergg etworkg expermets ad techologes (pp ). ACM. [12] Y, X., Bartulovć, M., Sekar,. ad Sopol, B., 217, May. O the effcecy ad faress of multplayer HTTP-based adaptve vdeo streamg. I Amerca Cotrol Coferece (ACC), 217 (pp ). IEEE. [13] Zhou, C., L, C.W. ad Guo, Z., 216. mdash: A markov decsobased rate adaptato approach for dyamc HTTP streamg. IEEE Trasactos o Multmeda, 18(4), pp [14] Ma, H., Hao, J. ad Zmmerma, R., 214, July. Access pot cetrc schedulg for dash streamg multrate wreless etwork. I 76 P a g e

10 (IJACSA) Iteratoal Joural of Advaced Computer Scece ad Applcatos, ol. 9, No. 2, 218 Multmeda ad Expo (ICME), 214 IEEE Iteratoal Coferece o (pp. 1-6). IEEE. [15] Zhag, S., L, B. ad L, B., 215, Jue. Presto: Towards far ad effcet HTTP adaptve streamg from multple servers. I Commucatos (ICC), 215 IEEE Iteratoal Coferece o (pp ). IEEE. [16] Dub, R., Dvr, A., Hadar, O., Shalala, R. ad Ahark, O., 215, Jauary. deo complexty hybrd traffc shapg for HTTP Adaptve Streamg. I Cosumer Commucatos ad Networkg Coferece (CCNC), th Aual IEEE (pp ). IEEE. [17] Ta, G. ad Lu, Y., 212, December. Towards agle ad smooth vdeo adaptato dyamc HTTP streamg. I Proceedgs of the 8th teratoal coferece o Emergg etworkg expermets ad techologes (pp ). ACM. [18] Mller, K., Quaccho, E., Gear, G. ad Wolsz, A., 212, May. Adaptato algorthm for adaptve streamg over HTTP. I Packet deo Workshop (P), th Iteratoal (pp ). IEEE. [19] Dolsky, K. ad Celkovsky, S., 212, Jue. Kalma flter uder olear system trasformatos. I Amerca Cotrol Coferece (ACC), 212 (pp ). IEEE P a g e