Guaranteeing Quality of Service in Interactive Video-on-Demand Servers

396 IEEE Transactions on Consumer Electronics, Vol. 45, No. 2, MAY 1999

GUARANTEEING QUALITY

OF

SERVICE IN INTERACTIVE

VIDEO-ON-DEMAND SERVERS

Chih-Yuan Cheng and Yen-Jen Oyang*

Department of Computer Science and Information Engineering National Taiwan University, Taipei, Taiwan, R.O.C.

and Meng-Huang Lee

Department of Information Management Shih Chien University, Taipei, Taiwan, R.O.C.

Abstract

One of the main challenges in disk system design for interactive VOD server.s is how to achieve CL good quality- ofservice (QoS) guarantee. This paper proposes a queue- ing model for analyzing the I/O bandwidth required in or- der to achieve a certain level of QoS f o r a disk system de- sign scheme that provides VCR em.ulation. The disk system design scheme employs a practice that requesfs no extra bandwidth to support interactive operutions such as ,fast forward search and fast backward search. The proposed queueing model provides a complete analysis too1,for p i t i -

ning down the resource requireinerit issue and is verijied through simulation for its validity.

1 Introduction

Guaranteeing quality of service (QoS) is one of the main challenges in the design of video-on-demand (VOD) systems that support interactive features such as fast for- ward search and fast backward search. In recent years, there have been a number of studies focusing on this is- sue [ l , 2, 3, 4, 51. The study done by Jayanta K. Dcy- Sircar and et al. assumed a Fast-mode stream plays lrames at a rate higher than a normal-mode stream and proposed a queueing model for analyzing thc system capacity re- quired in order to achieve a certain level of QoS[I]. The major disadvantage with the operation scheme that Jayanta

K. Dey-Sircar and et al. assumed is that a stream will claim higher bandwidth when it switches from the normal play- back mode to a fast search mode. As a result, a certain amounts of I/O bandwidth and network bandwidth must hc

* Corresponding author: Yen-Jen Oyaiig

reservcd for meeting thc additional handwidth required by streams switching to the list search modes. This leads to less efficient utilization 01 system resources.

To avoid this problem, scvcral schemes that support in-

tcractive VOD operations based on sampling MPEG[6, 71

GOPs(Group of Pictures) were proposed [2, 3 , 41. These

schemes require no extra handwidth to support interactive operations hut suffer the deficiency of unfavorable visual impact, especially in the fast backward search mode.

Shenoy and Vin then proposed a scheme that also re- quires no cxtra handwidth to support interactive operations and provides a good VCR emulation[5]. The problem with the scheme proposcd by Shenoy and Vin is that if the sys- tem needs to support hoth fast forward search and fast backward search or to support more than one rates of fast scarch modes, then multiple video files of about the same size must he created for each video program in the sys- tem. Sincc each vidco file created i n accordance with the scheme is of size no less than the original MPEG file of the program, the disk capacity required are increased by several folds.

This paper proposes a queueing model to analyze the resource requirement for a disk system design scheme that alleviates the deficiencies mentioned above. The disk sys- tem scheme employs a practice that provides VCR emula- tion without requiring extra disk bandwidth to support in- teractive opcrations [ X , 91. The practicc employed in this papcr to support the fast forward and backward searches plays separate MPEG tiles derived from sampling video frames at a certain rates and compressing the sampled vidco frames in accordance to the MPEG standard [6, 71. Figure 1 illustrates the process of creating separate MPEG files to be played in fast search modes. The major advan- tagcs of this practicc are two folds. First, it facilitates run- time resourcc allocation. Second, from the aspect of disk

Cheng, Oyang and Lee: Guaranteeing Quality of Service in lntcractive Video-On-Demand Servers 391

T h e original MPEG

file

for

the

normal

playback mode

I

B

B

P

B

B

P

B

B

P

B

B

P

B

B

Decoding

I

1

I I

'0 '1 F2 3' '4 '5 F 6 ' 7 F S '9 10' 11' F12 F13 F14 Sampling

I

/

/

/

'0 F 3 F 6 '9 F12 F15 F I S F21 '24 ' 2 7 F30 F33 F36 F39 F42 Encoding

I

1

1

I

I

B

B

P

B

B

P

B

B

P

B

B

P

B

B

A s e p a r a t e

file to

be played

in

the fast mode

Figurc I : The creation of special MPEG lilcs to he played in fast scarch modes.

capacity usage, it only introduces a small amount of over- head for supporting fast search modes. When generating a separate MPEG file to be playcd i n a fast search mode, one can select a hit rate no higher than the hit rate of thc orig- inal MPEG Ale. As a result, a video stream entering a fast search mode does not need to claim more system resources from the VOD server and the network. This alleviates a lot of run-time resource management problems for supporting interactive VOD operations. Since tlic quality of individual frames is less sensitive in last search modes, this practice

2 Disk system design for guaranteeing QoS

The discussion in this section first focuses on the pro- disk system design schcmc that supports interactive

VOD featurcs requiring higher data bandwidth,

Then, i n 2.2, a queueing Inode] is proposecl for determining the required I/O bandwidth of the disk system.

2.1 ~

i

system architecture

~

k

provides a good tradeoff hctwcen system rcsource manage-

ment and visual impact. As far as disk capacity is cow cerned, it is obvious that an additional MPEG tile crcatcd to support the N-time fast forward or backward scarch oc- cupies only 1/N the amount ofdisk space required to store the original video program.

With the practice descrihed above employcd, the next problem to address is how to figure out thc disk system bandwidth required in order to achieve a certain levcl of QoS. This paper presents a queueing model to tackle this problem and uses simulation to vcrify the validity of the queueing model. Thc simulation results reveal that the queueing model provides good estimates of the IIO band- width required

The remaining part 01 this paper is organizcd a s follows. Section 2 elaborates thc disk system design scheme and the proposed queueing model to determine the systcin re- sources required for achieving a certain levcl of QoS. Sec- tion 3 verifies the validity of the queucing model through simulation. Finally, the discussion of this papcr is con- cluded i n Section 4.

Figurc 2 depicts the general disk system architecture. The entire disk systcin consists of two disk arrays. One disk array stores MPEG files for normal playback. An-

other disk array storcs specially generated MPEG Ales to be played in fast search modes. The disk array that stores nor- mal playhack files provides the disk bandwidth for servic- ing strcains in the normal playback mode, while the disk ar- ray that stores fast mode files provides the disk bandwidth for servicing streams in fast search modes. Though there arc two disk arrays, a stream entering a fast search mode does not relinquish disk bandwidth that it claims from the disk array storing the normal mode files. The retained disk bandwidth can be used by the same stream when it returns from the fast search mode.

With thc disk system configuration described above and with the hit rate 01' a Fast mode file being carefully se- lected to be no higher than the hit rate of the corresponding normal playback file, a stream switching from the normal playhack mode to a fast scarch mode requires no extrarun- time system resources from the VOD server and the net- work. This alleviates a lot 0 1 run-time resource macage-

398 IEEE Transactions on Consumer Electronics, Vol. 45, No. 2, MAY 1999

Disk array storing

/

normal mode files

ment problems in interactive VOD systems design. In par- ticular, the VOD server does not need to pump more data per unit of time and the network does not need to allocate more data bandwidth to servicc a stream that switches from the normal playback mode to a fast search mode.

2.2 The

queueing model

In order to guarantee quality ol.servicc (QoS), a qucue- ing model must be developed to determine the amounts 0 1 system resources required. Figurc 3 shows a 2-dimcnsional queueing model that emulates the behavior of thc system. However, the 2-dimensional queueing model is too. com- plicated to be analyzed mathematically. Thercfore, an ap-

proximate approach is adopted. Figure 4 shows thc two Markov processes used to model the behavior of thc sys- tem. The first Markov process, which has only one state as shown i n Figure 4(a), models arrival of new streams and termination of existing streams in the VOD system. Thc second Markov process, which has two states as shown in

Figure 4(b), models streams switching back and forth be- tween the normal playback mode and either one of the two fast search modes, the fast forward search mode and the fast backward search mode. In the first Markov process, the arrival behavior of new stremw is a Poisson process with parameter XO and the duration of a stream staying in

the system is exponentially distributed with mean

k.

The first Markov process is used to figure out the amount of disk bandwidth that the VOD syslem must provide in order to reduce the average waiting time of an incoming requcst for creating a new stream to the level given in thc QoS specification. In the second Markov proccss, the interval during which a stream stays in the normal playback mode before switching to a fast search mode is exponentially distributed with mean and the interval during which a stream stays in a fast search mode is exponcntially dis-

Figure 2: The disk system architecturc.

/

Disk array storing fast mode files

\

tributed with mean

k.

The second Markov process is used

to figure out the amount of bandwidth that the disk array storing fast mode MPEG files must provide.

According to the queueing theory developed for the Markov process showri in Figure 4(a) [ I O ] , i i the average waiting time is the primary QoS criterion, then the num- ber of streams that the disk array storing normal playback files is able to servicc at one time, denoted by m, must sat- isfy the following inequality in order to reduce the average waiting time of a new stream to a certain level:

where

and Q I L is the allowed averagc waiting time given in the QoS specification. The remaining problcm after applying the incquality ( 1 ) abovc is that the video streams in the sys- tem may require different amounts of bandwidth to play. Thcrcfore, just calculating the number of strcams that thc system needs to support does not yield the amount of band- width requircd. This paper USCS the product ofthe weighted

average bandwidth and the number of streams that the disk array must be able to service at one time, which is denoted by m, to determine the total amount of bandwidth required. That is,

total amount of disk bandwidth requircd = m x p i b i ,

where 1 is the number of video programs stored in the VOD system, pi the probability that program i is selected by a

Cheng, Oyang and Lee: Guaranteeing Quality of Service in Interactive Video-On-Demand Servers

switching to

a fast mode switching back to

the normal mode

incoming new streams terminating streams

c

Figure 3: A 2-dimensional qucueing model that crnulates the system behavior.

( a ) The Markov proccss that inodcls arrival of new streams and lerinination of existing

.

streams i n the VOD system.

hi

Normal

U

P I

(h) The Markov process that models strcams

switching back and forth between the norinal playback mode and the fast search modes.

400 IEEE Transactions on Consumer Electronics, Vol. 45, No. 2, MAY 1999

newly-created stream, and bi is the bandwidth requircd to play program

i.

Nevertheless, using the weighlcd average bandwidth introduces a source of discrepancy betwecn the required amount of bandwidth derived from the qucucing model and the real world requirement and must be vcrificd through simulation.

With the amount of bandwidth that the disk array storing normal mode files must provide determined, the nexl issuc is to determine the amount of bandwidth that the disk array storing fast mode files must provide in order to reduce the average waiting time observed by a strcam switching from the normal playback mode to a fast mode to less than the value given in the QoS specification, denoted by

Ql,

in the following discussion. According to the queucing theory de- veloped for the Markov process shown in Figure 4(h)[ IO],

if there are s streams present in the VOD system and the disk array storing fast mode files can service

t

streams at one time, then the expected waiting time observed by a stream switching from the normal mode to a fast mode is equal to

where qs3 is the probability that j streams are eithcr in thc fast search mode or are in the queue waiting to switch to a fast search mode under thc condition that therc are totally

s streams present i n the system and

where

X I

P I = - P1

Because the number of streams in the VOD system does change from time to time, this papcr uses the following in- equality to figurc out the number of streams that the disk array storing fast search files must he able to service at onc time, denoted by

n,

in order to reduce the thc avcragc wait- ing time observed by a stream switching from the normal mode to a fast search mode to less than &f

system based on the Markov process model in Figure 4(a). According to thc queueing theory,

and

Once the number of streams that the disk array storing fast- mode tiles must provide is determined, denoted by 11, n is inultiplicd by the weighted average bandwidth to figure out the amount of bandwidth required.

Obviously, applying the mechanism described above in

determining the numbcr of streams that the disk array stor- ing the fast mode filcs must be able to support at one time is not in fully accordancc with the condition on which thc Markov process shown in Figure 4(b) is based. The Markov process shown in Figure 4(b) assumes the number

of strcams in the system remains a constant. Therefore, the validity of this mcchanism iri subject to verification. The verification of the qucucing model presented in this section

is presented i n next section.

2.3

Verification of the queueing model

In this paper, the validity of the queueing model is veri- fied through simulation. Figures 5-9 depict the simulation results based on different sets of parameters and compare thein with the rcsults derived from applying the queueing inodel shown in Figure 4. In all cascs, it is assumed that thc lcngth of a video program is 90 minutes, which is about the length of a typical movie. Parts (a) of the figures show the parameters used in simulation runs. Figures 5 shows a case in which all the programs require the same amount of bandwidth for playing back. Figures 6 shows a case in which the bandwidth required to play back the programs is uniformly distributed. Figures 7-9 show three cases in which the bandwidth rcquircd to play back the programs is

of the form delined as follows:

Probability[x

= k ] = P , L ( L L , u ) [ X

5

IC

+

0.51 P 7 L ( , L , o ) [ X

5

IC - 0.51 1 -

Pn(,,,)[X

5

k - 0.51 if k = 2 if 3

I

IC

5

7 if k = 8

e L ( L L , c ) [ ~

5

k:

+

0.51-

IO

otherwise, s=n+l

2

p , j=n ( g q s , j - " + l 7 w i

I

Q f ,

where rn is the maximum number of streams that thc disk array storing the normal mode files is able to service at one time and p , is the prohahilily that R streanis arc in the

where P,L(,L,c) is the probability function of the normal dis- tribution with mean p and variance v z . In Figures 7, 8,

Cheng, Oyang and Lee: Guaranteeing Quality of Service in Interactive Video-On-Demand Servers 40 1

and 9, the means of the normal distributions arc all equal to 5 and the variances are 1, 2, and 3, respectively.

Parts (c) of the figures show how thc average waiting time observed by

a

client changcs with rcspect to the hand- width of the disk array. Here, the handwidth of the disk array is measured i n units of the weighted averagc of the playback bandwidth of the vidco programs in the system, The simulation results presented in the figures arc derived from averaging 100 independent simulation runs with the same set of parameters but different random number seeds. Parts (d) of the figures show the nuniber of disk bandwidth units that the disk array storing thc normal-modc files needs to provide in order to reduce the average waiting time lor

a new stream to less than 5 seconds. and parts (e) show the number of disk bandwidth units that the disk array storing the fast-mode files needs to provide i n order to reduce thc average waiting time of a stream switching to a fast mode to less than 1 second. As the figures reveal, in all S cases, the values determined hy applying the proposed queueing model consistently fall within S % range from the values derived from simulation runs.

3

Conclusions

This paper discusses guaranteeing quality of service in

interactive video-on-demand servers. This paper proposes a queueing model for analyzing the I/O bandwidth requircd in order to achieve a certain level of QoS for a disk sys- tem design scheme that provides VCR emulation without requiring extra disk bandwidth to support interactive opcr- ations. The major advantages of thc proposed disk system design scheme include:

1. providing VCR emulation with similar visualization effects;

2. requesting no extra handwidth to support a stream switching from the normal playhack mode to a fast search mode;

3. introducing little overhead i n respect to disk space uti- lization.

The proposed queueing model provides a coniplctc analy- sis tool for pinning down the resourcc requircmcnt issuc. Since the proposed queueing model cmploys sevcral ap- proximations, its validity must be verified. This paper uses simulations to check the accuracy of the queueing model. The simulation results show that the queueing model is

quite accurate in determining the disk bandwidth requircd for achieving a certain level of QoS.

References

111

Jayanta K. Dcy-Sircar, Jamcs D. Salehi, James

E

Kurose, and Don Towsley. “Providing

VCR

capa- bilities in large scale vidco server,” In Proceesings of the Second ACM International Confereme on Multi- meditr, pp. 25-32, October 1994.

[2] Ming-Syan Chen, Dilip D. Kandlur, and Philip S. Yu. “Support for fully interactive playout in a disk-array- based vidco server.” ACM Multimedia, Vol. 3, No. 3, pp. 126-135, April 1995.

[ 3 ] Chih-Yuan Cheng, Chun-Hung Wen, Meng-Huang

Lee, and Yen-Jcn Oyang. “Effective utilization of disk bandwidth for supporting interactive video-on- dcmand,” IEEE Transaction on Consumer Electron- ics, Vol. 42, No. I, pp. 71-79, February 1996.

[41 Ming-Syan Chen and Dilip D. Kandlur. “Stream con- version to support interactivc video playout,” IEEE Multinzediu Magazine, Vol. 3, No. 2, pp. 51-58, Sum- mer 1996.

151 Prashant J. Shenoy and Marrick M. Vin. “Efficient support for scan operations in video servers,” In Pro-

ceedings of the Third .4CM Conference on Multime- dia, November 1995.

[6] “International standard iso/iec 1 1172-1,”. The Inter- national Organization for Standardization and the In- ternational Electronical Commision, August 1993.

171

“International standard iso/iec dis 13818-1,”. The In- ternational Organization for Standardization and the Intcrnational Electronical Commision. October 1995. [8] Chih-Yuan Cheng, Meng-Huang Lee, and Yen-Jen

Oyang. “Disk system design for guaranteeing quality

01‘ service in interactive vidco-on-demand systems,” In Proceedings of IS&T/SPIE Symposium on Elec- tronic Imaging: Science and Technology, January

1998.

L9J George Apostolopoulos, Marwan Krunz, and Satish Tripathi. “Supporting interactive scanning operations

in vod systems,” In Proceedings of IS&T/SPIE Sym-

posium on Electronic Imaging: Science and Technol- C J ~ Y , Conference on Multimedia Computing and Net- working, January 1998.

[ 101 Leonard Kleinrock. Queueing Systems Vol.1. A Wilcy-Intcrscience Publication, 1975.

402

in

disk bandwidth units

lEEE Transactions on Consumer Electronics, Vol. 45, No. 2, MAY 1999

1515 1515

(a) Simulation parameters.

. . . . . .

By queueing model Required bandwidth

disk handwidth units in

1 I O

1

1495 15001505 15101515 1520 1525 1530 1535 1540 The bandwidth of the normal-mode disk array in disk bandwldth units

By simulation

1 I O

..

2 3 4 5 6 7 8

Bandwidth in number of units

(b) Distribution of video programs bandwidth

Queueing mode -+-. -

,

Simulation + 1 Queueing mode -+-. - n n i I I I I I " . - I 95 100 105 110 115 120 125 130

The bandwidth of the fast-mode disk array in disk bandwidth units

(c) Average waiting time V.S. disk array bandwidth

I

By queueing model

I

By simulation

Keouired bandwidth

I

I

Cheng, Oyang and Lee: Guaranteeing Quality of Service in Interactive Video-On-Dcmand Scrvers

A1 11'1 A 2 / I 2

1/20 sec-'

I

1/5400 sec-'

I

1/300

see-'

1

1/20 sec:-'

m 0 0.1 0 a

c

By queueing niodcl Rcquircd , bandwidth

disk bandwidth units

I I1 I S 3 5 Simulation + Queueing mode

4

By simulation 151s .- _+. P ._ _m Required bandwidth disk bandwitlLh units in

115

1495 1500 1505 1510 1515 15201525 1530 1535 1540

The bandwidth of the normal-mode disk array 01 disk bandwlth units

simulation 110 - 403 . . . 95 100 105 110 115 120 125 130

The bandwidth of the fast-mode disk array of disk bandwidth unils

(c) Average waiting timc V.S. disk array bandwidth

( d ) Thc bandwidth rcquircd by thc iiormal-Inode disk array

Tor reducing thc averagc waiting tiinc of a newly-created stream to less than 5 seconds

(e) The bandwidth required by the last-mode disk array lor reducing

thc avcragc waiting time of a strcam switching to a last search modc to lcss than 1 second.

404 IEEE Transactions on Consumer Electronics, Vol. 45, No. 2, MAY 1999

By queueing model Rcquired bandwidth

disk bandwidth units

in 1520

A1

I

P1

I

Aa

I

PL2

1/20 s e c 1

I

1/5400 sec-'

I

1/300 sec-'

I

1/20 sec-'

By slmulatlon

1515

(a) Simulation parameters

, I l l /

Simulation +

20 Queueing mode - e - -

B y B y

Required bandwidth disk bandwidth units in

1 I O

-

0 VI

-

.-

E

m c m m ._ m E

k

simulation I10 1495 1500 1505 1510 1515 1520 1525 1530 1535 1540 The bandwidth of the normal-mode disk array In disk bandwidth units

0.4 0.35 0.3

2

0.25 E

3

0.15 0.1 0.05 0

2

0.2 2 3 4 5 6 7 8

Bandwidth in number of units.

Distribution od video programs bandwidth

0.01

95 100 105 110 115 120 125 130

The bandwidth of the fast-mode disk array in disk bandwidth units

(c) Average waiting timc V.S. disk array bandwidth

Cheng, Oyang and Lee: Guaranteeing Quality of Scnrice in Interactive Video-On-Demand Servers

A1 I A y P2

1/20 sec-'

I

1/5400 sec-'

I

1/300

.sew-'

I

1/20 sec-'

~ 405

2

0.2 2 0.15 a By queueing model Required bandwidth

disk bandwidth units in I525

2 3 4 5 6 7 8

Bandwidth in number of units.

( a ) Simulation pariunctcrs (b) Distrihution of video programs bandwidth

By simulation 1515 , . . . Simulation +

4

Queueing mode Required bandwidth i n 115 1495 1500 1505 1510 1515 1520 1525153015351540 The bandwidth of the normal-mode disk array in disk bandwidth units

I I O

n " 4 I

Y." I

95 100 105 110 115 120 125 130

The bandwidth of the last-mode disk array in disk bandwidth units

(c) Average waiting time V.S. disk array bandwidth

(d) Thc bandwidth rcquircd by the normal-mode disk array

for rcducing thc average waiting time of a newly-crcatcd stream to lcss than 5 seconds

...

I

disk bandwidth units

1

(c) The bandwidth required hy the fast-mode disk array for reducing

the average waiting timc o l a strcam switching to a fast search mode to less than 1 second.

406

By qucucing model Required bandwidth

disk bandwidth units

in IS30

IEEE Transactions on Consumer Electronics, Vol. 45, No. 2, MAY 1999

By simulation 1515 (a) Simulation parameters

Required,bandwidth 1 I1 I

IS

I-

I I O Simulation -+- Queueing mode -+--

4

1 1495 1500 1505 1510 1515 1520 1525 1530 1535 1540 The bandwidth of the normal-mode disk array in disk bandwidth units

0.3 Z .- 0.25 0 0.15 a 0.1 0.05 n > a 2 0.2 2 3 4 5 6 7 8

Bandwidth in number of units.

(b) Distribution of video programs bandwidth

. . ... .. .. .~ ...

95 100 105 110 115 120 125 130

The bandwidth of the fast-mode disk array in disk bandwidth units

(c) Average waiting time V.S. disk array bandwidth

(cl) The bandwidth required by the normal-mode disk array

for reducing the average wailing time of a newly-created stream to less than 5 seconds

(c) The bandwidth required by the fast-mode disk array for reducing

the average waiting timc o f a stream swilching to a fast search mode Lo lcss than 1 second

Cheng, Oyang and LCC: Guaranteeing Quality of Service in Interactive Video-On-Demand Servers

media storage systems,

Chili-Yuan Cheng rcccivcd the

B.S. degree i n Computer Sci- ence and Information Enginecr- ing lrom National Taiwan U n - vcrsity in 1991, the M.S. degree in 1993, and P1i.D. dcgrcc in

1998. He is currently a project manager i n Bridgcwell Inc. His rcscarch intcrests includc niulti- and operating systems.

Mcng-Huang Lcc rcccivcd the

B.S. degree in Elec~rical En-

gineering from National Clicng Kung University i n 1987, the M.S. tlcgrec in 1989, and the P1i.D. dcgrec in Coinputer Sci- ence and Inforination Enginccr- ing from National Taiwan Uni- versity i n 1996. He is currently

an Associate Professor i n the Department of Inforination Management, Shih Chicn University. His research interests includc multimedia storage systems and opci-aring systems.

Yen-Jcn Oyang rcccivcd thc B.S. degree in Inforination Enginccr- ing froni National Taiwan Uni- versity in 1982, the M.S. dc-

g m in Coinputer Scicncc froni the Caliiornia Institute (11 Tech- nology in 1984, and the P1i.D. degree in Elcctrical Engineer- ing from Stanlortl University in

1988. Hc is currently a professor i n the Dcprtmcnt of Computer Science and Inforination Engineering, National Taiwan University. From I989 to 1996, he was an associatc professor in the same dcpartment. His rcscarch interests in- cludc design of video scrvcr systems and digital libraries.