MPEG4 Traffi Modelig sig The Trasform Expad Sample Methodology Ashraf Matrawy Ioais Lambadaris Chagheg Huag Broadbad Networks Laboratory Departmet of Systems ad Computer Egieerig Carleto iversity {amatrawy, ioais, huag}@se.arleto.a Phoe: 1-613-520-2600 ext. 2913 Fax: 1-613-520-5727 Abstrat The trasfer of digital video will be a ruial ompoet of the desig of future home etworkig appliatios. This trasfer was made feasible by the advaemet of digital video eodig tehiques that redued the badwidth required for this trasfer to a pratial level. MPEG4 is a eodig tehique that is suitable for home etworkig appliatios with its low bit rate. It also has the advatage that allows viewers to iterat with eoded objets. I this paper, we preset our work that eables the study of MPEG4 properties ad performae o the Iteret usig simulatio. We propose a traffi geerator that is able to geerate traffi that has almost the same first ad seod order statistis as a origial trae of MPEG4 frames that is geerated usig a MPEG4 eoder. We model ad geerate this traffi based o the Trasform Expad Sample (TES) methodology usig TEStool. We preset the model ad show the performae of the geerator i terms of good mathig of the harateristis of the modeled trae. 1. Itrodutio The Iteret s impat o daily life is ireasig dramatially. Relyig o the Iteret as a soure of iformatio ad a meas of ommuiatio is umathed i history. This is the reaso why some researhers ad ommuiatio experts believe that sooer or later, all the devies that are parts or our daily life, like home appliaes, will be oeted to the Iteret. Coetig those devies ad appliaes to the Iteret will require the trasfer of video or at least will ilude video trasfer as a ehaemet. This is oe of the motivatios behid establishig the ISO Movig-Piture-Experts-Group (MPEG) family of stadards for digital video eodig. This family of stadards iludes MPEG1, MPEG2, ad MPEG4. MPEG7 is uder developmet. A. Matrawy, I. Lambadaris, ad C. Huag, Carleto iversity, Caada 1
MPEG4 is the oe that is most suitable for the Iteret [1]. It is targeted for low bit rates. It allows real images to o-exist with omputer-geerated outerparts ad also allows their separatio ad their reeivig differet treatmet due to iteratio with the user. The mai feature of importae to the etwork is MPEG4 s apability of real-time adaptive eodig. This ehaes etwork utilizatio ad eables MPEG4 seders to be more resposive to hages i etwork oditios. It geerates video i three differet frame types (I, P, ad B) that serve to eode differet portios of the video sigal i differet levels of quality. I this work, we preset our traffi geerator that a be used for studyig MPEG4 behavior ad performae through simulatio. This is relevat to givig a more realisti sese to simulatios of multimedia trasfer o the Iteret. The traffi we geerate losely mathes the statistial harateristis (i terms of margial distributio ad auto-orrelatio futio) of a origial real trae of video frames that were geerated usig a MPEG4 eoder. We developed a model usig the Trasform Expad Sample (TES) [2], whih is a versatile methodology for modelig ay set of give observatios i a time series. We modeled I frames, P frames, ad B frames usig three differet TES models. We the used iterleavig to geerate the origial sequee of frames for MPEG4. We implemeted this geerator i software ad itegrated it ito the etwork simulator s-2 [3]. The rest of this paper is orgaized as follows: Setio 2 gives a overview of the MPEG4 eodig tehique for digital video ad how it fits ito a etworked appliaes eviromet. I Setio 3, we preset the Trasform Expad Sample (TES) models ad their appliatios. Setio 4 gives the details of how we modeled MPEG4 usig TES ad demostrates how losely the geerated traffi mathes the real traffi. Setio 5 oludes the paper ad presets our future researh i this area. 2. MPEG4 video eodig I [4], where requiremets for etworked appliaes were preseted, the speifiatio of Home Audio ad Video Iteroperability was a mai requiremet. Also, support for media streamig was oe of the ommuiatio mode requiremets. MPEG4 has a lot of features that may fulfil these requiremets [1]: A. Matrawy, I. Lambadaris, ad C. Huag, Carleto iversity, Caada 2
It is targeted to provide high video quality at relatively low bit rates. This makes it suitable for home devies that does ot have high speed aess to the etwork It allows iterativity of the users at the deoder side Salable otet, whih meas eodig oe ad playig bak at differet speeds Real-time adaptive eodig that results i hagig the sedig rate i reatio to hagig etwork oditios Salable eodig that differetiates betwee importat objets ad less importat objets. This results i streams that are suitable for trasmissio over etworks that support priorities MPEG4 eoders geerate three types of frames: I frames, P frames ad B frames. They a be desribed as follows: I frames (itra-oded frames) otai the iformatio that results from eodig a still image, i.e., with o referee to ay other image. They are poits of referee ad radom aess i the video stream. They a be deoded without the eed for ay other frames. The ompressio rate for these frames is the lowest of all frame types. P frames (preditively oded frames) require iformatio from previous I frames ad/or P frames for eodig ad deodig. By exploitig temporal redudaies, P frames ahieve higher ompressio rates tha that for I frames. B frames (bidiretioally preditively oded frames) require iformatio from the previous ad followig I frames ad/or P frames for eodig ad deodig. They are predited from both previous ad followig frames. They have the highest ompressio ratio amog all frame types. Referee frames must be trasmitted first. Therefore the trasmissio order ad display order may differ, as desribed i Fig. 1. At the begiig, there is always a I frame. The first I frame ad the first P frame serve as a referee for the first two B frames. At the first time, the first I frame is the referee for the first P frame. The seod I frame serve as a referee for the B frames that are displayed before it, that s why it has to be set first. Trasmissio order: I P B B P B B P B B I B B Display order: I B B P B B P B B P B B I Fig. 1 Frame trasmissio ad display orders A. Matrawy, I. Lambadaris, ad C. Huag, Carleto iversity, Caada 3
3. Trasform Expad Sample (TES) models TES is a versatile methodology for geeratig data that losely math (i terms of its margial distributio ad auto-orrelatio futio) ay set of give observatio of a time series. The derivatio of TES models is performed i two phases [2]. First, a orrelated sequee, with uiform margial i [0,1], (also alled bakgroud TES proess) is formed as follows: or 0 = 1 V = = 0 > 0 eve 1 odd Where 0 is a radom variable uiformly distributed i [0,1]. {V } is a sequee of iid radom variables idepedet of 0, alled the iovatio sequee (with probability desity f v ). The operator < > deotes the modulo-1 operatio whih for every real x is defied by: x = x max{ iteger ( ) : x} Both ad a geerate lag-1 autoorrelatios i the rage [0,1) ad [-1,0) respetively. Regardless if the probability law of the iovatio desity f v, it a be show [5] that the TES bakgroud sequee forms a statioary Markovia proess with a uiform i [0,1) margial distributio. A smoothig operatio alled the stithig trasform is used to get rid of the iheret disotiuity of the < > operator [2]. So i short, i the first phase of the TES model developmet, the objetive is to geerate a time series of orrelated radom variables with uiform margials [0,1). The amout of orrelatio depeds o the struture of the desity futio f v of the iovatio sequee {V }, whih i geeral may be arbitrary. I the seod phase, sytheti sample data, alled the foregroud sequee, whih resembles the real sample, may be derived from the bakgroud sequee usig the iversio tehique i [5]. This iversio tehique allows the trasformatio of ay uiform radom variable to oe with arbitrary distributio. More speifially, it is assumed that a histogram H(.) of the empirial video sequee has bee built osistig of J ells, where the j th ell is positioed o the iterval [l j,r j ) ad is haraterized by a probability p j. Suh a histogram a be iverted as follows: A. Matrawy, I. Lambadaris, ad C. Huag, Carleto iversity, Caada 4
H 1 ( x) = J I j= 1 [, j 1 ] j ( x) l j ( x j 1 w ) p j Where I A is the idiator futio of the set A, w j = r j - l j is the width of ell j, ad j =, 1 j J, ( 0, 1 ). j = = The radom variate 1 0 J { ( x)} i= 1 p i H = 1 has a margial distributio H(.) ad autoorrelatio that depeds upo the iovatio desity f v. The modelig proess is arried out i two steps: the first step essetially aptures the autoorrelatio of the iput time series usig iovatio f v ad applyig a smoothig tehique usig the stithig trasform [5]. This would result i the geeratio of the bakgroud sequee. The seod step is to ivert the bakgroud sequee usig the empirial histogram, whih always guaratees the mathig of the margials. The task of fidig a suitable f v that approximates the empirial autoorrelatio is arried out via a heuristi searh usig TEStool [5]. The mathig of the histogram is guarateed, while the mathig of the autoorrelatio is arried out i a iterative maer with visual feedbak for the best f v. 4. Modelig MPEG4 usig TES I this setio, we preset our model for MPEG4 traffi geeratio. We used a sequee of MPEG4 frames that was geerated by eodig a ews broadast from the BBC at a rate of 30 frames/se. Here are the steps we followed to geerate the traffi: We used a trae of frame sizes (i bits) of a MPEG4 eoded video. Fig. 2 shows the harateristis of the trae as geerated while developig the model usig the TEStool. Widow w1 of the figure shows the sample path of the frame sizes, i w2 the autoorrelatio futio is show, while w3 shows the histogram of the distributio of the frame sizes. We separated this trae ito three traes, oe for eah frame type. This failitates the geeratio of a trae that mathes the origial trae. Tryig to do that with the origial trae without separatio proved to be very diffiult. This is due to the periodi ature of the autoorrelatio futio of the origial trae. Havig doe that we model eah of the traes usig a separate TES model usig the methodology desribed i Setio 3. A. Matrawy, I. Lambadaris, ad C. Huag, Carleto iversity, Caada 5
Fig. 3, Fig. 4, ad Fig. 5 show the models of the I frames, P frames, ad B frames, respetively. I eah figure, w4 shows the iovatio desity f v that was desried i Setio 3. I all the figures, the geerated data is show i blue (or darker lie) while the origial is i gree. From the figures, we otie that the histogram always mathes the origial data while there is a little differee i both the autoorrelatio futio ad the sample path. To geerate the MPEG4 trae, we iterleave geeratig samples from the three developed Figure 2: The origial trae models to math the sequee that is show i Fig. 1. The fial trae that is geerated by iterleavig a be ompared to the origial trae as show i Fig. 6. We implemeted this geerator i software. First, we use TEStool to geerate the histograms ad the iovatio desity for the three frame types. The we provide this as a iput to the software we wrote i C that uses eah of the three models to geerate a frame size of eah type i the order metioed i Fig. 1. We itegrated this ode ito the s-2 simulator to geerate traffi for simulatio of IP etworks. A. Matrawy, I. Lambadaris, ad C. Huag, Carleto iversity, Caada 6
Figure 3: The I frames model Figure 4: The P frames model A. Matrawy, I. Lambadaris, ad C. Huag, Carleto iversity, Caada 7
Figure 5: The B frames model Figure 6: Compariso betwee the origial ad the geerated traes A. Matrawy, I. Lambadaris, ad C. Huag, Carleto iversity, Caada 8
5. Colusio ad future work We have preseted our work o geeratig video traffi that mathes the statistial properties (first ad seod order) of a origial trae of MPEG4 frames. We used the Trasform Expad Sample (TES) methodology to model the MPEG4 trae. We have suessfully modeled the I, P, ad B frame sequee ad we showed how the geerated traffi mathes the origial trae. This work is relevat to the simulatio MPEG4 performae i etworked appliaes appliatios as well as ay other video appliatio that rus o IP etworks. For future work, we will be developig models that may model more properties of the video trae suh as geeratig differet priority pakets ad apturig self-similarity of the video traffi. Referees 1. R. Koee, MPEG-4 Overview, IEEE SPECTRM, Vol. 36, No. 2, February 1999. 2. B. Melamed, A Overview if TES Proesses ad Modelig Methodology, I Performae Evaluatio of Computer ad Commuiatio Systems, 1993. 3. http://www.isi.edu/sam/s/ 4. S. Tsag, S. Moyer, D. Marples, H. Shulzrie, ad A. Royhowdhury, Requiremets for etworked Appliaes: Wide-Area Aess, Cotrol, ad Iterworkig, Iteret Draft, drafttsag-appliaes-reqs-01.txt. 5. B. Melamed et al., TES-Based Modelig for Performae Evaluatio of Itegrated Networks, I Pro. of INFOCOM, 1992. A. Matrawy, I. Lambadaris, ad C. Huag, Carleto iversity, Caada 9