Computer Science Technical Report Archive

USC Computer Science Technical Reports, no. 666 (1998)

(USC DC Other)

USC Computer Science Technical Reports, no. 666 (1998)

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content An Ev aluation of Alternativ e Disk Sc heduling
T ec hniques in Supp ort of V ariable Bit Rate
Con tin uous Media
Jab er A AlMarri and Shahram Ghandeharizadeh
Computer Science Departmen t
Univ ersit y of Southern California
Los Angeles California falmarrishahr amgp ol luxusce du
Abstract An um b er of recen t studies ha vein v estigated sc heduling tec h
niques in supp ort of v ariable bit rate VBR video When compared with
constan t bit rate CBR video VBR has a lo w er storage and bandwidth
requiremen t while pro viding the same qualit y of images Ho w ev er a VBR
video clip migh t exhibit a signican tv ariance in the bit rate required to
supp ort its con tin uous displa y The previous studies ha v e prop osed tec h
niques to supp ort the displayofa VBR clipfrom t w o dieren t p ersp ec
tiv es disk storage subsystem and the net w ork In this study w e prop ose
a taxonom y of VBR disk sc heduling tec hniques that includes those pro
p osed for the net w ork The results demonstrate that a new class of disk
sc heduling tec hniques termed Atomic VR
VIT AL is sup erior Al
gorithms used to represen t this class w ere adopted from the net w orking
literature
In tro duction
Due to sev eral adv ances in computer pro cessing storage p erformance and high
sp eed comm unications a n um ber of data in tensiv e applications ha v e b ecome
viable Examples include digital libraries distance learning videoondemand
shopping and en tertainmen t services etc Con tin uous media digital audio and
video pla y a ma jor role in these applications The principle c haracteristics of
con tin uous media is their sustained bit rate requiremen t F or example digital
comp onen t video based on the CCIR standard requires Megabits per
second Mbps for its con tin uous displa y If a system deliv ers a clip at a rate
lo w er than its presp ecied rate without sp ecial precautions eg prefetc hing
the user migh t observ e frequen t disruptions and dela ys with video and random
noises with audio These artifacts are collectiv ely termed hic cups The bandwidth
requiremen t of a clip along with its size can be reduced using compression
tec hniques due to redundancy in data With con tin uous media compression
tec hniques can b e classied as either CBR or VBR With b oth approac hes the
data m ust still b e deliv ered at a presp ecied rate T ypically CBR sc hemes allo w
some b ounded v ariation of this rate based on some amoun t of memory at the
displa y With VBR this v ariation is not b ounded The VBR sc hemes ha vethe
adv an tage that for the same a v erage bandwidth as CBR they can main tain a
more constan t qualit y in the deliv ered images b y utilizing more megabits per
second when needed eg when there is more action in a scene The fo cus of this
study is on VBR enco ded con tin uous media and tec hniques that ensure their
hiccupfree displa y T o supp ort a hiccupfree displa y of VBR enco ded con tin uous media one m ust
analyze b oth the net w ork and the disk subsystem A n um ber of studies ha v e
analyzed eac h comp onen t in isolation F or the net w ork the previous studies ha vestriv en to reduce the rate v ariabilit y of the VBR clips
F or the disk storage subsystem the previous studies ha v e fo cused on
data placemen t admission con trol and sc heduling tec hniques In this study w e
pro vide a framew ork that aliates disk subsystem with net w ork through the
in v estigation of the role of bandwidth smo othing algorithms to sc hedule the disk
bandwidth for con tin uous displa y of VBR video clips The con tribution of this
study is a taxonom y that includes all p ossible tec hniques In this taxonom y U niv er sal CR
and Atomic VR
FI T tec hniques ha v e b een applied to the
disk subsystem Ho w ev er the applications of U niv er sal CR
FI T U niv er sal CR
VIT AL Atomic CR
and Atomic VR
VIT AL tec hniques to the disk
subsystem are no v el W eev aluate these tec hniques based on three criteria
Throughput Num ber of sim ultaneous displa ys supp orted b y a tec hnique
assuming a xed system conguration
Startup latency The amoun tofdela y incurred from when a request refer
ences a video clip to the onset of its displa y Cost p er stream The amoun t of resources required byatec hnique computed
as cost to supp ort a xed n um b er of sim ultaneous displa ys
The results demonstrate that Atomic VR
VIT AL is sup erior to others
The rest of this pap er is organized as follo ws After a brief description of
the related studies in Section Section presen ts a taxonom y of disk retriev al
tec hniques eigh t sp ecic algorithms in supp ort of VBR video and a discussion
ab out data placemen t admission con trol and m ultizone disks This section
describ es ho w the dieren t algorithms map in to a single taxonomyand whic h
algorithm is used to represen t a tec hnique in the taxonom y In Section a
case study is presen ted where eac h algorithm is applied to mo vie clips obtained
from the Univ ersal Studios In Section w e describ e analytical mo dels used to
compute the cost p er stream and the a v erage startup latency for the dieren t
approac hes These analytical mo dels are applied to the case study presen ted in
Section Brief conclusions and future researc h directions are oered in Section Related W ork
Tw o studies ha v e prop osed alternativ e approac hes for placemen t of data
with VBR ob jects constan ttime length and constan t data length In the rst
approac h the blo c ks of the video data are v ariable in length with constan t

realtime displa y The blo c ks in the other approac h are xed in length with
v ariable realtime displa yW e refer to the rst approac h in this study as Atomic VR
FI T The constan t data length approac h is referred to as U niv er sal CR
in this study A statistical admissions con trol based on a n um ber of service
classes corresp onding to v arious probabilities of loss is describ ed in Three
tec hniques are pro vided for computing loss probabilities histogram con v olution
Cen tral Limit Theorem and Cramers rule Scalabilit yis ac hiev ed b y allo wing
appropriate frames of MPEG to b e dropp ed without fully susp ending service to
an activ e displa yA n um b er of deterministic and statistical admissions con trol
p olicies for eac h of the data placemen t approac hes is outlined in Based on one
disk exp erimen ts the t woapproac hes are compared using a statistical admission
con trol p olicy Astoc hastic mo del that guaran tees an upp er b ound of the n um b er of hiccup
observ ed b y the streams is describ ed in All the data fragmen ts stored on
the disks ha v e the same displa y time W e refer to this approachas Atomic VR
FI T in this study The LaplaceStieltjes transform of the service time
distribution is deriv ed for the mo del based on batc hed disk service under a
m ultiuser load of the concurren tly serv ed con tin uous data streams Cherno
b ounds are applied to the tail of the service time distribution and the resulting
distribution of the glitc h rate p er stream
Sev eral other studies ha vein v estigated the trac generated b y VBR
streams Tw o net w ork service mo dels are prop osed b y these studies determinis
tic guaran teed service and renegotiated CBR service These studies fo cus
on net w ork sc heduling while w e fo cus on disk sc heduling In the net w ork sc hedul
ing the researc hers aim to minimize the p eak bandwidth minimize the n um ber
of bandwidth increases and maximize the n um b er of bandwidth decreases Ho w
ev er in disk sc heduling w estriv e to maximize the n um b er of concurren t users
minimize the clien t buer and minimize the startup latency time
Our w ork incorp orates and extends these sc hemes W e prop ose a taxonom y
that rev eals all p ossible strategies that can b e used to supp ort a deterministic
disk retriev al of VBR ob jects W eev aluate eac h strategy individuallyThen w e
compare and con trast the dieren t strategies with one another
A T axonom y of Disk Retriev al of VBR Ob jects
A taxonomyof disk retriev al tec hniques that can guaran tee a hiccupfree dis
pla y of VBR ob jects is sho wn in Figure There are t w o main tec hniques that
pro duce a retriev al plan to transmit an ob ject from a serv er to a clien t The
rst tec hnique emplo ys a constan t transmission rate sa y Mbps W e term
this tec hnique C onstant R etr iev al R ate CR
The second tec hnique termed
V ar iabl e R etr iev al R ate VR
utilizes more than one transmission rate With
VR
the retriev al plan of an ob ject is divided in to a n um b er of time in terv als
The serv er ma y use a dieren t transmission rate for eac h of these in terv als If the
in terv als length is constan t the tec hnique is termed F ixed tI me inT er v al FI T F or example with a hour presen tation the serv er migh t transmit the ob ject at

VBR Objects
Retrieval Strategies
Universal Atomic
Constant Retrieval
Rate (CR )
Variable Retrieval
Rate (VR )
Fixed Time
Interval (FIT)
Variable Time
Interval (VITAL)
2 2
Constant Retrieval
Rate (CR )
Variable Retrieval
Rate (VR )
2 2
Fixed Time
Interval (FIT)
Variable Time
Interval (VITAL)
Fig A taxonom y of disk retriev al of VBR ob jects
Mbps for the rst hour Mbps for the second hour and Mbps for the third
hour If the in terv als ha v e v ariable length the tec hnique is termed V arI able
T ime inter v AL VIT AL A disk retriev al strategy determines the manner in
whic h these tec hniques are emplo y ed While a U niv er sal strategy enforces a sin
gle retriev al plan on all ob jects in the rep ositoryan Atomic strategy pro duces
a dieren t retriev al plan for eac h of these ob jects
Time
Size
CR
VR
2
2
Object Cumulative Data
Consumption Rate
Amount of Prefetched Data
Client Startup Latency
Retrieval
Ends
Retrieval
Ends
Display
Ends
Fig CR
vs VR
Figure sho ws CR
and VR
In this gure the solid line represen ts the
n umberofb ytes yaxis consumed b y the clien t as a function of time xaxis
T o ensure a hiccupfree displa y the amoun t of data deliv ered b y a tec hnique
m ust suce that consumed b y the clien t This fact is depicted in the gure
b y dra wing the lines dotted lines corresp onding to CR
and VR
ab o vethe

solid line CR
is represen ted as a straigh t line b ecause it retriev es data at
a constan t rate constan t slop e Ho w ev er VR
is represen ted as a piecewise
line due to the v ariation of rates slop es The in tersection of CR
line with
the yaxis sp ecies the amoun t of data that a clien t m ust prefetc h before its
displa y is started The in tersection of CR
line with xaxis determines the time
clien t startup latency required to materialize the prefetc hed amoun t of data
Ob viouslyone ma y dra w dieren t CR
lines with dieren t slop es The slop e of
a line sp ecies the constan t bandwidth required to supp ort a hiccupfree displa y
Dieren t lines w ould resem ble a tradeo b et w een the bandwidth required bythe
sc hedule and the amoun t of data that a clien t m ust be prefetc hed ie clien t
startup latency It is p ossible to devise algorithms with dieren t ob jectiv es eg
minimize the bandwidth requiremen t while the prefetc hed amoun t of data is less
than Megab yte Figure sho ws b oth VIT AL and FI T While FI T divides
a retriev al plan in to in terv als of equal length VIT AL attempts to adjust to the
need of an ob ject
Time
Size
t1 t2 t3 t4
Transmission Rates
Time
Size
t1 t2 t3 t4
Transmission Rates
t5
Figure a FI T t t t t t t t Figure b VIT AL t t t t t t t t t Fig VR
With Atomiceac h ob ject in the rep ository has an individual retriev al plan
Ho w ev er with U niv er sal a single retriev al plan is emplo y ed for all ob jects This
plan is computed as follo ws U niv er sal constructs a univ ersal ob ject whose
consumption rate is computed as dened b y the maxim um consumption rate
required b y an ob ject p er unit of time refer to T able for the denition of the
parameters in this equation
S i i T
max C j i j O
Assuming that the rep ository consists of t w o ob jects Figure illustrates the
cum ulativ e data consumption of the univ ersal ob ject The cum ulativ e bandwidth

T erm Denition
O T otal n um b er of ob jects in the rep ository
T Displa y time length of the longest ob ject
C j i Amoun t of data consumed at time i byobject j
S i Amoun t of data consumed at time i b y univ ersal ob ject
D j i Av ailable transfer bandwidth for logical disk j at time i
K T otal n um b er of logical disks
R D T ransfer bandwidth of a ph ysical disk
R C Av erage w eigh ted displa y bandwidth requiremen t for all ob jects
Consumption rate
R Cj Displa y bandwidth requiremen t for ob ject j
R Cji Displa y bandwidth requiremen t for ob ject j at in terv al i
N Maxim um n um ber of sim ultaneous displa ys supp orted b y a single disk
SM T otal amountof serv er memory needed to supp ort N streams
CM T otal amountof clien t memory needed to supp ort one stream
T p Duration of a time p erio d
B Size of a blo c k
l max Maxim um serv er startup latency
l av g Av erage serv er startup latency
heatj F requency of access to ob ject j
leng th j Displa y time length of ob ject j
leng th ji Displa y time length of ob ject j at in terv al i
cy l T otal n um b er of cylinders of a disk
T
SeekM ax
W orst case seek time
T
Seek
c Seek time to tra v erse c cylinders
P Num ber of ph ysical disks
R
l
D
T ransfer bandwidth of a logical disk
Mbps Megabits p er second
MBps Megab ytes p er second
MB Megab ytes
GB Gigab ytes
T able List of terms used rep eatedly in this researc h and their denitions
requiremen t of the univ ersal retriev al plan is greater than or equal to that of an y
ob ject in the rep ository As a result a serv er ma y pro duce an ob ject at a higher
rate than its displa y bandwidth requiremen t increasing the amoun t of required
buer at a clien t site Since the univ ersal retriev al plan is used b y the serv er
to transmit all ob jects its duration corresp onds to the largest ob ject in the
rep ository Smaller ob jects ma y not need to utilize the en tire retriev al plan th us
terminating earlier The univ ersal ob ject is used b y either CR
VR
VIT AL or VR
FI T and their asso ciated algorithms to pro duce a univ ersal retriev al
plan
CR
VR
VIT ALand VR
FI T ma y pro duce retriev al plans b y utilizing
bandwidth smo othing algorithms The smo othing algorithms striv e to minimize
the burstiness in the transmission of the ob jects b y using the a priori kno wledge

Time
Size
Time
Size
Object 1
Object 2
Universal Object Curve
Object 1
Object 2
Generated
Universal Object
Fig Cum ulativ e data consumption of the univ ersal ob ject
of the ob jects frame sizes In Section w e describ e eigh t dieren t bandwidth
smo othing algorithms that ha v e app eared in the literature F or eac h algorithm
w e describ e ho w it can representa giv en classication sho wn in the taxonomyof
Figure Finally Section describ es whic h algorithm is c hosen to representa
classication of Figure and the reasons b ehind our decision
Eigh t Bandwidth Smo othing Algorithms
Bandwidth smo othing algorithms are created to compute a retriev al plan for the
deliv ery of an ob ject whic h simplies the allo cation of resources in the serv er
and the net w ork They striv e to remo v e the burstiness in VBR ob jects without
causing startup latency By eliminating burstiness they impro v e the utilization
of serv er and net w ork resources A bandwidth smo othing algorithm pro duces a
retriev al plan whic h reduces the v ariance of the transmission rates for a com
pressed ob ject based on the a priori kno wledge of the frame sizes in the com
pressed ob ject while a v oiding b oth undero wand o v ero w of the clientbuer The smo othing is done b y prefetc hing frames in adv ance of eachburstConse quen tly a larger clien t buer results in a less burst y retriev al plan for an ob ject
Bandwidth algorithms generate retriev al plans that
consist of a n um ber of time in terv als based on the frame lengths and the clien t
buer During eac hin terv al a p ortion of an ob ject is transmitted using a con
stan t rate The time duration of an in terv al dep ends on the clien t buer and
the degree of burstiness in the ob ject Dieren t smo othing algorithms pro duce
dieren t retriev al plans for an ob ject These plans ma y dier in the n um ber of
in terv als length of the in terv als and the consumption rate requiremen tin eac h
in terv al This is due to the dieren t metrics that they aim to optimize T able depicts the classication of the algorithms in to three p ossible categories CR
VR
FI Tand VR
VIT ALW eno w describ e eac h of these algorithms in
turn

Smo othing Algorithm CR
VR
FI T VR
VIT AL
ABA
p p p
SWS
p p p
CR TT
p
PCR TT
p p p
R CBS p
CBA p
MCBA p
MVBA p
T able Classication of bandwidth smo othing algorithms
Algorithm Av erage Bandwidth Allo cation ABA
ABA is a simple algorithm that groups an arbitrary n um b er of frames to
gether in to a c h unk and computes its a v erage bandwidth requiremen t The
a v erage bandwidth of a c h unk is used to transmit its frames Basically ac h unk
is a collection of disk blo c ks that are retriev ed at the a v erage bandwidth re
quirementofac h unk T o guaran tee a hiccupfree displa y the rst c h unk m ust
b e buered in its en tiret y at the clien t prior to displa y The buer requiremen t
b ecomes impractical when ABA assumes large c h unk size Smo othing of the
bandwidth across c h unks is not considered byABA ABA algorithm ma y pro duce a plan for CR
b y grouping the en tire frames
of an ob ject together in to one c h unk This implies that a clien tm ust cac he an
ob ject in its en tiret y prior to initiate its displa y incurring a high startup latency ABA algorithm generates a plan for VR
when an ob jects frames are group ed
in to more than one c h unk If eac hc h unk con tains the same n um b er of frames
the generated plan is VR
FI T Otherwise it is VR
VIT AL Algorithm Sliding Windo w Smo othing SWS
SWS is similar to ABA except that it attempts to smo oth the bandwidth
within c h unks using a sliding windo w It groups an arbitrary n umberofc h unks
in to a windo w and smo othes them The smo othing pro cess is based on prefetc h
ing of frames with large sizes F rames m ust b e prefetc hed in order to guaran tee
a hiccupfree displa yThe smo othing pro cess starts from the last c h unk w ork
ing itself bac k to the rst c h unk When the bandwidth of a c h unk is ab o vethe
a v erage bandwidth of the windo w as a whole the dierence b et w een the c h unk
bandwidth and the a v erage bandwidth is transfered to an earlier c h unks whose
bandwidth is b elowthe a v erage bandwidth This results in prefetc hing of those
frames that ha v e large sizes SWS smo othes the bandwidth of an ob ject b etter
than ABA when the windo w size is sucien tly large Ev en though it ma y require
buers at the clien t to accommo date its prefetc hing of frames its total memory

requiremen t at the clien t is far less than ABA Plans for CR
VR
FI Tand
VR
VIT AL can b e pro duced b y SWS in a similar manner to ABA
Algorithm Constan tRate T ransmission and T ransp ort CR TT
CR TT is an algorithm that pro duces a constan t transmission rate for an
ob ject As input this algorithm consumes the clien t buer size and the startup
latency time in terms of n um b er of frames As output it pro duces t wocurv es
The rst represen ts the minim um bit rate requiremen t that guaran tees a hiccup
free displa y The second designates the maxim um bit rate requiremen t whic h
guaran tees no o v ero w at the clien t buer The in tersection of these t wocurv es
iden ties the feasible region of constan t bandwidth that can b e used to represen t
the ob ject This algorithm migh t require a large clien t buer It supp orts CR
Algorithm PiecewiseConstan tRate T ransmission and T ransp ort
PCR TT
PCR TT algorithm divides an ob ject stream in to xed time in terv als The
bandwidth rate of eachin terv al is represen ted b y the slop e of the line connecting
in terv al p oin ts T o guaran tee a hiccupfree displa yPCR TT algorithm v ertically
osets the lines of the in terv als un til they are ab o v e the ob ject consumption
curv e see Figure This algorithm results in a startup latency The piecewise
linear curv e that is generated b y PCR TT is used to determine the minim um
clien t buer whic h migh t b e large
Time
Size
t1 t2 t3 t4
Object Curve
Initial Bandwidth Plan
Hiccup-free Bandwidth Plan
Client Startup
Latency
Amount of Prefetched Data
Fig Creation of PCR TT in terv als
PCR TT algorithm can generate a plan for CR
b y considering an ob ject
stream as one in terv al In addition PCR TT algorithm can pro duce a plan for

VR
b y dividing an ob ject stream in to more than one in terv al If all in terv als
ha v e the same time length the resulting plan is VR
FI T Otherwise it is
VR
VIT AL Algorithm RateConstrained Bandwidth Smo othing R CBS
Giv en a maxim um bandwidth rate constrain t R CBS examines eac h frame in
an ob ject and ensures its bandwidth do es not exceed this maxim um b y prefetc h
ing data R CBS examines the frames starting from the end of the ob ject and
main tains the excess data that should b e prefetc hed If the bandwidth of a frame
is greater than the bandwidth constrain t the excess bandwidth is distributed
among the preceding frames as long as it do es not cause the bandwidth of the
preceding frames to violate the constrain t An example of the eect of R CBS
on an ob ject bandwidth is sho wn in Figure Conceptually R CBS runs a knife
across the frames bandwidth R CBS minimizes the amoun t of prefetc hed data
Ho w ev er it maximizes the v ariabilit y of the bandwidth allo cation and the n um
b er of bandwidth c hanges It also increases the v ariabilit y of the time b et w een
bandwidth c hanges R CBS pro duces a retriev al plan for an ob ject b y smo oth
ing its frames Since frames are considered xed time in terv al of the same time
length the plan pro duced b yR CBS is VR
FI T Frame Number
Bandwidth
Frame Number
Bandwidth
Effect of RCBS
on an Object
Fig R CBS example
Algorithms and MVBA CBA and MCBA
Minim um V ariabilit y Bandwidth Allo cation MVBA Minim um Changes
Bandwidth Allo cation MCBA and Critical Bandwidth Allo cation CBA

algorithms are similar in nature They consumes as input the accum ula
tiv e data consumption rate curv e of an ob ject and the clien t buer size in order
to pro duce a maxim um transmission curv e that denotes the upp er limits b ey ond
whic h the clien t buer o v ero ws They na vigate bet w een these t w o curv es to
generate a piecewiselinear curv e They dier in ho w they c ho ose the starting
p oin t for the next in terv al when bandwidth either increases or decreases see
Figure Increase
Bandwidth
Decrease
Bandwidth
(a) MVBA (b) CBA (c) MCBA
frontiers
Fig MVBA CBA and MCBA dierences
MVBA starts the bandwidth c hanges at the leftmost p oin t along the fron tier
see Figure for b oth bandwidth increases and decreases It pro duces a plan
that has the smallest p ossible p eak bandwidth It gradually results in a large
n um b er of small bandwidth c hanges
CBA searc hes the fron tier to nd a starting poin t that permits the next
tra jectory to extend as far as p ossible in the case of bandwidth increases F or
bandwidth decreases CBA b egins at the leftmost p oin t on the fron tier It results
in segmen ts that ha v e the smallest p ossible peak bandwidth requiremen t and
the minim um n um ber of bandwidth increases Ho w ev er the segmen ts do not
necessarily ha v e the minim um n um b er of bandwidth decreases
MCBA is similar to CBA except that in the case of bandwidth decreases it
searc hes the fron tier to nd a starting p oin tthatallo ws the next tra jectory to
extend as far as p ossible In addition to the prop erties of CBA MCBA results
in a plan with the smallest p ossible n um b er of bandwidth c hanges
Due to the na vigation bet w een the data consumption rate curv e and the
maxim um transmission curv e the retriev al plan pro duced b y MVBA CBA or
MCBA is divided in to v ariable time length in terv als Consequen tly the plan
pro duced b y these algorithms is of t yp e VR
VIT AL Eigh t Algorithms and One T axonom y
CR
VR
VIT ALand VR
FI T ma y pro duce retriev al plans using one of
the bandwidth smo othing algorithms These algorithms striv e to minimize the

burstiness in the transmission of the ob jects b y using the a priori kno wledge of
the ob jects frame sizes Next w e iden tify the algorithm used to represen t CR
VR
VIT ALand VR
FI T The CR
strategy can be implemen ted using either ABA SWS CR TT or
PCR TT Both ABA and SWS are inappropriate b ecause they require a clien t
to cac he an ob ject in its en tiret y prior to initiating the displa y in order to
guaran tee hiccupfree displa y This increases the cost p er stream and results in
a long clien t startup latencyEv en though CR TT pro vides a larger n um ber of
alternativ e plans than PCR TT it is more complex and has a longer execution
time In addition a plan pro duced b yPCR TT matc hes w ell with the b est plan
pro duced byCR TT W ec hose PCR TT to represen t CR
class of algorithms
The VR
FI T strategy can b e implemen ted using either PCR TT or R CBS
W e c hose PCR TT b ecause R CBS requires the designer to input a maxim um
threshold for the data pro duction An optimal maxim um bandwidth rate con
strain t is not easy to iden tify and suchac hoice migh t dier from one ob ject to
another
The VR
VIT AL strategy can b e realized using either ABA SWS PCR TT
MVBA CBA or MCBA W e selected MVBA CBA and MCBA to pro duce
plans for VR
VIT ALW e eliminated ABA SWS and PCR TT b ecause they
require the in terv als to b e iden tied man ually ahead of time prior to their exe
cution whic h migh t b e impractical for certain class of applications Moreo v er
MVBA CBA and MCBA minimize the amoun t of buer required at the clien t
site and require no clien t startup latency when compared with ABA SWS and
PCR TT
Discussion
In addition to c ho osing a sc heduling algorithms a system designer m ust w orry
ab out design decisions suc h as admission con trol placemen t of data and the dif
feren t zones of a disk driv e Sev eral studies ha vein v estigated
issues p ertaining to these topics In this section w e describ e eac h topic in turn
and discuss ho w it ts with the alternativ edisksc heduling tec hniques
Data Placemen t
In this researchwefocus on a m ultidisk arc hitecture W e consider striping with
a roundrobin data placemen t since it maximizes p erformance b y
balancing the load of a displa y across the a v ailable resources
Ob jects are assigned a blo c k size according to the strategies discussed in
Section The blo c k size is computed based on the bandwidth of an ob ject and
the time p erio d of the serv er whic h will b e describ ed in the next section With
ABA and SWS require to prefetc h the rst in terv al to guaran tee a hiccupfree displa y When the in terv al length is long they incur a long startup latency b y requiring the
clienttoprefetc h a signican t amoun t of data They b eha v e similar to PCR TT when
the in terv al length is short Hence w e ignore ABA and SWS

U niv er sal the serv er ma y use one common blo c k size or xed pattern of blo c k
sizes for the retriev al of its ob jects With U niv er sal CR
all ob jects ha vethe
same constantbloc k size On the other hand with U niv er sal VR
the blo c k
size of an ob ject mayv ary The same v ariation of the blo c k sizes is applied to
all ob jects
Atomic is exible in assigning blo c k sizes to the ob jects Eac h ob ject migh t
ha v e its o wn constan t blo cksizeas pro vided b y Atomic CR
With Atomic VR
an ob ject ma yha v e its o wn v ariable blo c k sizes
Curren tly there is not a w a y to determine an optimal blo c k size for an ob ject
All strategies ha v e their o wn adv an tages and disadv an tages Ha ving a constan t
blo c k size migh t be easier to implemen t but it ma y require a larger amoun t
of memory at a clien t and incur a longer startup latency V ariable blo c k size
attempts to adjust to the need of an ob ject ho w ev er it ma y cause disk frag
men tation and bandwidth fragmen tation
Admission Con trol
Sc heduling of a con tin uous media serv er can b e divided in to t w o categories de
terministic and statistical retriev al W e fo cus on the deterministic sc heduling
where the blo c ks are retriev ed in cyclic rounds termed time p erio ds The read
time of a disk for eac h time p erio d is partitioned in to slots The time duration of
eac h slot is guaran teed to handle the retriev al of a single blo c k Since all ob jects
ha v e the same blo cksizewith U niv er sal CR
the time duration of eac hslot
is the same Ho w ev er with the other tec hniques where the blo c k size c hanges
the time duration of eac h slot mayv ary The admission con trol for U niv er sal CR
is simple b ecause it uses a con
stan t bandwidth for all ob jects Consequen tly new requests can be admitted
to the serv er based on either a rstcomerstserv e shortestrequestrst or
longestrequestrst p olicy In addition to these p olicies Atomic CR
can also
utilize a rsttrstserv e p olicy b ecause eac h ob ject has its o wn constan t band
width and a limited amoun t of idle bandwidth mightbe a v ailable
On the other hand sc heduling with VR
requires a prior kno wledge of the
future load of the serv er and the v arying bandwidth needed b y ob jects The
serv er only needs to main tain future load of the system up to the displaytime
length of the longest ob ject Consequen tly assume that at time t a new user
requests ob ject j where j starts at disk d and j s en tire pla ybacktime in terv al
is L The new user needs to b e admitted to the serv er at time tF or this to b e
feasible the follo wing constraintm ust b e satised
C j i i L
D d i mo d K t i mo d T The denition of the parameters of this equation can be found in T able
U niv er sal VR
can use similar p olicies as the ones men tioned ab o vefor CR
after satisfying the constrain t Ho w ev er in addition to p olicies describ ed ab o v e
Atomic VR
ma y use other p olicies that are based on individual ob ject band
width plan F or instance new requests can b e admitted based on the minim um

or the maxim um p eak bandwidth the n um ber of bandwidth increases or the
n um b er of bandwidth decreases
With VR
this deterministic admission con trol is not applicable with the
existence of V CR functions suc h as pause resume fastforw ard viewing and
fastrewind viewing This is b ecause the deterministic admission con trol relies
on the fact that the serv er has full kno wledge of the future rates needed b y all
requests If V CR functions are pro vided the full kno wledge of the future rates
is not p ossible
MultiZone Disks
Zones are in tro duced in disks to meet the demands for higher storage capacit yA
zone is a con tiguous collection of disk cylinders whose trac ks ha v e the same stor
age capacit y While zoning increases the storage capacit y of the disk it results in
a disk with v ariable transfer bandwidth where the data in the outermost region
if pro duced at faster bandwidth Sev eral studies ha vein v estigated the
la y out of data across the disk surface and the sc heduling for its retriev al to har
ness the a v erage transfer bandwidth of the zones These studies are orthogonal
to the U niv er sal and Atomic strategies
As a demonstration w e describ e a tec hnique that is prop osed in and
explain ho w it can b e utilized in our study Assuming that the n um b er of trac ks
in ev ery zone is a m ultiple of some xed n um b er constructs Logical T rac k
L T from the n um b ered ph ysical trac k of the dieren t zones The order of
the trac ks in a L T is byzonen um ber When displa ying an ob ject the system
reads a L T on its b ehalf p er time p erio d This forces the disk to retriev e data
from the constituting ph ysical trac ks in immediate succession b y zone order A
displa y observ es a constan t disk transfer bandwidth for eachL T retriev al With
U niv er sal and Atomic strategies a blo c k is assigned to a L T
A Case Study with Mo vie Clips from Univ ersal Studios
Resources F rame Sizes
Size Time Rate Avg Max Min Std
GB min Mbps b ytes b ytes b ytes b ytes
Mo vie Mo vie Mo vie Mo vie Mo vie T able Mo vies statistics

In order to pro vide an eectiv e comparison of the dieren t strategies w e require
a large collection of video trace data to represen t the div erse m ultimedia band
width requiremen ts W e obtained MPEG trace data for v erealmo vies from
Univ ersal Studios that are stored in Digital V ersatile Discs D VDs The frame
rate of eachmo vie is framessecond Eac hmo vie is segmen ted in to c hapters
The n um ber of c hapters in a mo vie t ypically range b et w een to Eac hc hapter
is sa v ed as a le T able sho ws some statistics ab out these mo vies Using these
v emo vies w e constructed a rep ository of mo vies The rst v emo vies are
the original traces The remaining mo vies consist of c hapters The c hapter
of eachmo vie w as selected randomly from those of the v e original mo vies
Clien t Buer Size MB Clien t Startup Latency sec
Max Min Avg Max Min Avg
CR
PCR TT min
VR
In terv al min
PCR TT min FI T Length min min MB
Buer MB
CBA MB
Size MB
MB
MB
VR
Buer MB
MCBA MB
VIT AL Size MB
MB
MB
Buer MB
MVBA MB
Size MB
MB
T able Smo othing bandwidth algorithms statistics for U niv er sal
Using PCR TT CBA MCBA and MVBA algorithms w e generated retriev al
plans for eac h ob ject in the rep ository These retriev al plans corresp ond to
Atomic and U niv er salIn the case of VR
FI T w e tried dieren t in terv al
lengths in order to ev aluate the approac h appropriatelyW e pro duced in terv al
plans for v e dieren tin terv al lengths min ute min ute min ute min ute
and min ute With VR
VIT ALw e attempted dieren tclien t buer sizes
that are commonly used in the curren t p ersonal computers MB MB
MB MB and MB W e selected the in terv al lengths with VR
FI T in
terms of min utes rather than seconds b ecause wew an ted the n um ber of in terv als
to b e relativ ely similar to the ones generated b y VR
VIT AL The maxim um
n um ber of in terv als pro duced for an ob ject b y VR
VIT AL w as using
MVBA with MB clien t buer size Consequen tly w e generated retriev al

plans for eac h ob ject in the rep ositoryHalf of these retriev al plans b elong to
Atomic and the other half b elong to U niv er sal One out of the retriev al plans
in a strategy represen ts CR
Fiv e retriev al plans corresp ond to VR
FI T The other fteen retriev al plans represen t VR
VIT ALThe total n um ber of
retriev al plans of the rep ository is A retriev al plan migh t consist of one or more in terv als An in terv al is repre
sen ted b y a length and a constan t consumption rate The length of the in terv al is
expressed in terms of frames A frame is displa y ed in second since the frame
rate of eachmo vie is framessecond The constan t consumption
rate is expressed in terms of bits p er frame bits p er second
Clien t Buer Size MB Clien t Startup Latency sec
Max Min Avg Max Min Avg
CR
PCR TT min VR
In terv al min PCR TT min FI T Length min min MB Buer MB CBA MB Size MB MB MB VR
Buer MB MCBA MB VIT AL Size MB MB MB Buer MB MVBA MB Size MB MB T able Smo othing bandwidth algorithms statistics for Atomic
T ables and summarize the retriev al plans for U niv er sal and Atomic resp ectiv ely The rst column designates the approac h The second column in
dicates the name of the smo othing algorithm used The fourth one describ es
either the constantin terv al length for VR
FI T or the buer size that is used
in the smo othing algorithms for VR
VIT AL The clien t buer refers to the
actual clien t buer size that is required in order to a v oid a memory o v ero w F or
instance ev en though w e used a MB buer with MVBA algorithm to smo oth
the univ ersal ob ject curv e in the case of U niv er sal the actual a v erage clien t
size is MB The reason of this dierence is b ecause the serv er transmits
data based on the bandwidth plan for the univ ersal ob ject curv e while the con
sumption rate is based on the individual ob jects Columns to con tain the
clien t startup latency information F or example in the case of U niv er sal CR

Time
Size
t1 t2 t3
Client Startup
Latency
Amount of Prefetched Data
First
Interval
Second
Interval
Third
Interval
Bandwidth Retrieval Plan
Fig PCR TT startup latency
when the system transmits an ob ject to a clien t the clien tm ust w ait sec
onds from the transmission time b efore it initiates the displa y of the ob ject
This allo ws the system to accum ulate enough data in the clien t that guaran tees
hiccupfree displa y Figure sho ws ho w the system computes the clien t startup
latency with PCR TT PCR TT imp oses an amoun t of data that is required to b e
prefetc hed to a v oid hiccups The system transmits that amoun t of data using the
consumption rate slop e of the rst in terv al The startup latency of an ob ject
is computed based on the amoun t of data to b e transfered and the consumption
rate of the rst in terv al
Analytical Mo dels
This section quan ties the buer requiremen t and the a v erage startup latency for
our taxonom y This formalism is crucial to the design of approac hes that w ould
minimize the system cost p er stream and the a v erage startup latencyIn order
to compare the dieren t approac hes w e designed analytical mo dels to compute
the cost per stream These mo dels quan tify the amoun t of memory required
b y eachapproac h as a function of a presp ecied amoun t of disk bandwidth to
supp ort a xed n um b er of sim ultaneous displa ys In addition w eha v e dev elop ed
mo dels to estimate the a v erage startup latency incurred byeac h approachas a
function of a theoretical upp er bound These mo dels are applied on the case
study men tioned in Section The rep orted n um ber of sim ultaneous displa ys is
a theoretical upp er b ound hence the presen ted cost p er stream is a theoretical
minim um

Seek Time Mo del
When a blo c k is requested from a disk driv e the disk arm is mo v ed to the
p osition of the blo c k The mo v emen t of the disk arm is termed seek Sev eral
studies ha v e prop osed analytical mo del to estimate the seek time as
a function of the distance tra v eled b y the disk arm An approachis to mo del
the seek time as a com bination of squarero ot and linear function F or example
the mo del that w e use to describ e the seek c haracteristic of Seagate STW
disk consisting of cylinders is
T
S eek
c
p
c if c c otherwise
T
Seek
c denotes the time in millisecond required for the disk arm to tra v erse
c cylinders W e assume that rotational latency time is included in T
Seek
c The
maxim um seek time denoted T
S eek M ax
is observ ed when c is equal to
cylinders The Seagate STW disk has a transfer bandwidth of MBps
An elev ator algorithm can b e emplo y ed to optimize the seek time In the w orst
case scenario when N requests are required to be retriev ed from a disk they
mightbe ev enly scattered across the disk surface Therefore the maxim um seek
time observ ed byeac h request is cyl designates the total n um b er of cylinders
of a disk
T
S eek
cy l
N
Cost Mo del
Serv er memory clien t buer size and disk storage are the three comp onen ts
that are impacted b y the U niv er sal and Atomic strategies This section deriv es
equations to estimate the cost p er stream for eac h strategy The ob jectiv eisto
minimize the system cost p er stream
T o supp ort con tin uous displa y of an ob ject X that requires a constantcon sumption rate R
C
the system partitions X in to n equisized blo c ks X
X
X
n where n is a function of the blo cksize B and the size of XW eassume
that a blo c k size is the unit of transfer from disk to main memory and then to
the clien t through the net w ork The time p erio d whic h is the displa y time of a
blo c k is dene as
T
p
B
R
C
Up on the arriv al of a request referencing ob ject X the system stages X
in
memory and transmits it to the clien t Then the system initiates its displa y Prior to completion of a time p erio d the system initiates the retriev al of X
in to memory in order to ensure a hiccupfree displa y This pro cess is rep eated
un til all blo c ks of an ob ject ha v e b een retriev ed and displa y ed Assuming that
all ob jects ha v e the same consumption rate eg U niv er sal CR
a disk can

supp ort N sim ultaneous ob ject displa ys
N T
p
T
S eek M ax
B
R D
T o guaran tee a hiccupfree displa y w e assume the maxim um seek time ie
T
S eek M ax
R
D
is the transfer bandwidth of a disk
Seek time is an o v erhead whic h can greatly reduce the n um b er of sim ultane
ous displa ys supp orted b y a disk if it is not con trolled from the ab o v e describ ed
tec hnique the disk p erforms N seeks per time p erio d Consequen tly the per cen tage of time reserv ed for seeking in order to guaran tee hiccupfree displa ys
is
N T
S eek M ax
T
p
By substituting T
p
from the p ercen tage of w asted disk bandwidth is dened
as
w astef ul N T
SeekM ax
R
C
B
The system can supp ort a higher n um ber of sim ultaneous displa ys b y reduc
ing the w asted disk bandwidth p ercen tage This can be accomplished b y ma
nipulating t wov ariables minimize the distance tra v ersed b y a seek andor
maximize the blo c k size B A larger blo c k size leads to a higher memory
requiremen t p er displa yBelo w w e describ e a tec hnique that decreases the im
pact of seeks The role of a larger blo c k size is analyzed in the con text of this
tec hnique
The impact of seeks can be reduced through the use of Group Sw eeping
Sc heme GSS GSS minimizes the n um b er of cylinders tra v ersed
b y eachseek op eration In its simplest form the data that is retriev ed during
one time p erio d is deliv ered for displa y during the next time p erio d During eac h
time p erio d data for eac hactiv e displa y is retriev ed from the disk and in to main
memory Sim ultaneously the data retriev ed during the previous time p erio d is
transmitted o v er the net w ork to clien ts Since no data blo c k is required to be
presen t in memory un til the end of the time p erio d in whic h it is retriev ed the
disk sc heduling is able to optimize the retriev al of the blo c ks for eac h time p erio d
using disk sc heduling p olicies suchas elev ator algorithm
In a more general form GSS divides N activ e displa ys of a time p erio d in to
g groups This partitions a time p erio d in to g subp erio ds eac h corresp onding to
the retriev al of
N
g
blo c ks F or simplicit yw e consider GSS with g is equal to one
in this researc h More discussion on this sub ject can b e found in T o compute
the blo c k size with GSS w e rst compute the duration of time con tributed to
the seek times during a time p erio d This can b e done using Since N blo c ks
are fetc hed during a time p erio d the system incurs N seek times in addition to
N blo c k retriev als during a time p erio d ie T
p
N B
R D
N T
S eek
cy l
N
By
substituting T
p
in and solving for Bw e obtain
B R
C
R
D
R
D
N R
C
N T
Seek
cy l
N

Assuming few er than N activ e displa ys the maxim um p ossible serv er startup
latency o ccurs when the request arriv es when the time p erio d with emptyslot
has just started In this case the dela y is the summation of t wotime periods
one for the missed time p erio d and the second one for the retriev al of the blo c k
l
max
T
p

The a v erage serv er startup latency is dened as
l
av g
l
max
T
p

If the n um b er of activ e displa ys exceeds N then should b e extended with
appropriate queuing mo dels
The system requires N memory blo c ks in the serv er to supp ort N displa ys
This is b ecause eac h displa y needs t w o buers of size B The rst one con tains
data that is b eing transmitted to the clien t and displa y ed it w as fetc hed during
the previous time p erio d The second is used b y the system to store the data
for the curren t time p erio d The system sw aps b et w een these t w o buers un til a
displa y nishes Hence the total amoun tofserv er memory required bya displa y
is
SM B
Clien t buer size denoted CM dep ends on the strategies and their asso ci
ated bandwidth smo othing algorithms Column in tables and in Section con tains the computed clien t buer size for all dieren t strategies Notice that
w e select the maxim um clien t buer size in order to prev en t clien t buer o v er
o w Therefore the cost p er stream for a disk is dened as
cost SM CM N M D N

The M is the cost of memory p er MB whic h is assumed to b e The D is the cost of a disk driv e whic h is assumed to be The obtained results
remain the same with c heap er disk driv e or memory So far w e assumed that all ob jects in the rep ository ha v e the same constan t
consumption rate eg U niv er sal CR
R
C
No w assume that eac h ob ject
sa y j has a dieren t constan t consumption rate eg Atomic CR
R
Cj
An approac h to compute the cost p er stream for the system is to compute the
a v erage consumption rate for all ob jects in the rep ository as follo ws assuming
O ob jects in the rep ository
R
C
P
O
j R
Cj
O

Equation do es not exploit the c haracteristics of an ob ject T o optimally com
pute the consumption rate for all ob jects w e consider t w o factors The rst one

is the probabilityof access of eac h ob ject j denoted heat
j
The second is the
length of eac h ob ject j denoted leng th
j
heat
j
j O
O
X
j
heat
j

W
j
l eng th
j
P
O
i l eng th
i
j O
The a v erage w eigh ted consumption rate for the system is dened as
R
C
O
X
j
heat
j
W
j
R
Cj

In the t w o factors are considered to ha v e the same impact on the cost
This migh t not b e alw a ys true The equation can b e trivially extended to include
w eigh ts for the t w o factors
So far w e describ ed ho w to compute the cost per stream for U niv er sal CR
and Atomic CR
No wwefocus on computing the cost p er stream for
U niv er sal VR
and Atomic VR
Assume that eac h ob ject sa y j has a
n um ber of in terv als sa y L
j
During eac hin terv al sa y i ob ject j has a dieren t
constan t consumption rate R
Cj i
The duration of in terv al i is denoted leng th
ji
The a v erage w eigh ted consumption rate for the system can be computed as
follo ws
W
ji
l eng th
ji
P
O
k l eng th
k
j O and i L
j

R
C
O
X
j
L j
X
i heat j
L j
W
ji
R
Cj i

Equation is a general equation and can b e trivially extended to include w eigh ts
for the t w o factors
Due to the limitation of storage capacit y and sp eed of a ph ysical disk m ultiple
ph ysical disks can be group ed to constitute a logical disk A blo c k size of an
ob ject is distributed ev enly bet w een the ph ysical disks of a logical disk The
transfer bandwidth of a logical disk that consists of P ph ysical disks is
R
l
D
P
X
i R
D

T able and summarize the cost analysis of all approac hes using the case
study describ ed in Section W ev aried the n umberofph ysical disks in a logical
disk from one to t w en t y in order to alloweac h approac h to optimize its cost p er
stream The t yp e of a ph ysical disk is a Seagate STW disk The frequency

Avg Startup Cost p er
P N T
p
Latency sec Stream
CR
PCR TT min VR
In terv al min PCR TT min FI T Length min min MB Buer MB CBA MB Size MB MB MB VR
Buer MB MCBA MB VIT AL Size MB MB MB Buer MB MVBA MB Size MB MB T able Cost p er stream analysis for U niv er sal
of access to the ob jects is mo deled using Zipf distribution with parameter
this distribution matc hes w ell with empirical data on video ren tal
The rst column in T able designates the approac h The second column
indicates the name of the smo othing algorithm used The fourth one describ es
either the constantin terv al length for VR
FI T or the buer size that is used
in the smo othing algorithms for VR
VIT AL Column has the n um ber of
ph ysical disks in a logical disk at whic h the minim um cost is obtained Column
represen ts the n um ber of sim ultaneous displa ys supp orted b y a logical disk
Column con tains the optimal time period The a v erage startup latency for
the system is sho wn in column whic h is the summation of the a v erage serv er
startup latency and the a v erage clien t startup latencyThe a v erage serv er startup
latency is computed using The a v erage clien t startup latency is obtained
from column in T able in Section Column con tains the minim um cost
p er stream The information con tained in T able b elongs to U niv er salT able is similar to T able Ho w ev er it con tains information that b elongs to Atomic The a v erage clien t startup latency for Atomic is obtained from column in
T able in Section
Avg Startup Cost p er
P N T
p
Latency sec Stream
CR
PCR TT min VR
In terv al min PCR TT min FI T Length min min MB Buer MB CBA MB Size MB MB MB VR
Buer MB MCBA MB VIT AL Size MB MB MB Buer MB MVBA MB Size MB MB T able Cost p er stream analysis for Atomic
As sho wn in T able the cost p er stream for U niv er sal CR
is This conguration supp orts sim ultaneous streams per logical disk and in
curs a seconds a v erage startup latency The minim um cost per stream
of U niv er sal VR
FI T is where the in terv al length is min ute
The system is congured to supp ort sim ultaneous streams p er logical disk and
incurs a seconds a v erage startup latency With U niv er sal VR
VIT AL the minim um cost p er stream is using MVBA algorithm and a clien t
size of MB The system is congured to supp ort sim ultaneous streams p er
logical disk and incurs a second a v erage startup latency Due to the large cost per stream U niv er sal is an infeasible strategy The
c heap est approac h with U niv er sal costs appro ximately times that of the most
exp ensiv e approac h with Atomic Figure sho ws a comparison of U niv er sal CR
U niv er sal VR
FI T using min ute in terv al length and U niv er sal VR
VIT AL using MVBA and MB clien t buer In these gures the xaxis
is the n um ber of sim ultaneous displa ys supp orted b y the system In Figure a
the yaxis is the cost p er stream while in Figure b the yaxis is the a v erage
startup latency
1 2 3 4 5
24000
24500
25000
25500
26000
26500
27000
27500
28000
28500
29000
Cost per Stream [$]
Number of Streams
CR
2
VR FIT
2
VR VITAL
2
Avg Startup Latency [sec]
CR
2
VR FIT
2
VR VITAL
2
Number of Streams
1 234 5
0
100
200
300
400
500
600
Figure a Cost p er Stream Figure b Av erage startup latency
Fig Comparison of the tec hniques of U niv er sal
1 2 34 5 6 78 9 10 11
0
500
1000
1500
2000
2500
3000
3500
4000
Cost per Stream [$]
Number of Streams
CR
2
VR FIT
2
VR VITAL
2
Number of Streams
Avg Startup Latency [sec]
1 2 3 4 5 678 9 10 11
0
20
40
60
80
100
120
140
160
180
200
CR
2
VR FIT
2
VR VITAL
2
Figure a Cost p er stream Figure b Av erage startup latency
Fig Comparison of the tec hniques of Atomic

Figure sho ws a comparison of Atomic CR
Atomic VR
FI Tand
Atomic VR
VIT AL In these gures the xaxis is the n um ber of sim ultaneous
displa ys supp orted b y the system In Figure a the yaxis is the cost p er stream
while in Figure b the yaxis is the a v erage startup latency The minim um
cost p er stream of Atomic CR
is when the system is congured to
supp ort sim ultaneous streams per logical disk This conguration incurs a
startup latency of seconds on the a v erage The minim um cost p er stream
of Atomic VR
FI T is when the in terv al length is min ute This
conguration supp orts sim ultaneous streams per logical disk and incurs a
seconds a v erage startup latency The minim um cost p er stream of Atomic VR
VIT AL is using MVBA and a MB cac he at eac h clien t With
this tec hnique the system supp orts sim ultaneous streams p er logical disk and
incurs a second a v erage startup latency When Atomic VR
VIT AL is compared with Atomic CR
it reduces the
cost p er stream b y a factor of and the a v erage startup latency b y a factor of
When compared with Atomic VR
FI T Atomic VR
VIT AL reduces
the cost p er stream b y ! and the a v erage startup latency b y a factor of It is clearly a sup erior approac h
Retriev al Plan Calibration
The retriev al plan of an ob ject ma y consist of one or more in terv als An in terv al
length is expressed in terms of frames as w e describ ed in Section Also the
constan t consumption rate is expressed in terms of bits p er frame The optimal
time p erio d for an approac h is expressed in terms of seconds as sho wn in column
in T ables and Consequen tlythe in terv als length and their consumption
rate of a retriev al plan are required to b e expressed in terms of the optimal time
p erio d The n um ber of frames in an optimal time period can b e computed b y
m ultiplying the optimal time p erio d b y the frame rate of an ob ject Notice that
the n um b er of frames in an optimal time p erio d mightbea real n um ber One of three cases is encoun tered when an in terv al length is adjusted to b e
expressed in terms of the optimal time p erio d The rst case is when the n um ber
of frames in an in terv al is equal to or a m ultiple of the n umberofframesin the
optimal time p erio d In this case the new in terv al length is obtained b y dividing
the old in terv al length b y the n um b er of frames in the time p erio d The result
of the division is an in teger n um b er The consumption rate can b e adjusted for
the new in terv al bym ultiplying it bythe n um b er of frames in the time p erio d The second case is when the n um ber of frames in an in terv al is less than the
n um ber of frames in the optimal time p erio d In this case frames from the
next in terv al are b orro w ed to mak e the n um b er of frames in the currentin terv al
equals to the n um b er of frames in the time p erio d The length of the new in terv al
b ecomes equal to one time p erio d and its consumption rate is the summation of
the consumption rates of the frames that constitute it If the next in terv al do es
not ha v e enough frames to makethe n um ber of frames in the curren t in terv al
equals to the n um ber of frames in the time p erio d then the curren t and the
next in terv al are com bined together and the same scenario is rep eated The last

case is when the n um b er of frames in an in terv al is greater than the n um ber of
frames in the time p erio d and it is not a m ultiple of the n um b er of frames in the
time p erio d This in terv al is divided in to t woin terv als The rst one con tains the
maxim um n um b er of frames that are a m ultiple of the n um b er of frames in the
time p erio d The second in terv al con tains the remainder n um b er of frames whic h
are less than the n um b er of frames in the time p erio d The rst in terv al is treated
similar to the rst case and the second in terv al is handled lik e the second case
After the adjustmen t the new retriev al plan migh t ha v e same greater or less
n um ber of in terv als than the old one dep ending on the frequencies of the ab o v e
cases If the rst case is encoun tered more frequen tly then the new retriev al plan
mightha veappro ximately the same n um ber of in terv al as the old one When the
frequency of the second case is higher than the other t w o cases the n um ber of
in terv als in the new retriev al plan b ecomes less than the one in the old retriev al
plan The n um ber of in terv als in the new retriev al plan b ecomes greater than
the old in terv al when the third case is encoun tered more frequen tly 0 43 45 100 frames
R = 25 Kbpf
C1
R = 30 Kbpf
C2
R = 15 Kbpf
C4
Interval 1 Interval 3 Interval 4
(T = 0.43 second = 10.75 frames)
P
(a) Retrieval Plan before Calibration
(b) Retrieval Plan after Calibration
50
Interval 2
R = 10 Kbpf
C3
0 4 5 10 Time Periods
R = 268.75 Kbpt
C1
R = 166.22 Kbpt
C2
R = 48.75 Kbpt
C4
Interval 1 Interval 3 Interval 4
9
Interval 2
R = 161.25 Kbpt
C3
Kbpf : Kilobits per frame
Kbpt : Kilobits per time period
Fig An example of ob ject retriev al plan calibration
T o demonstrate these cases w e describ e an example whichis sho wn in Fig
ure Figure a sho ws an ob ject retriev al plan whic h consists of four in
terv als The length of the ob ject is assumed to b e frames seconds The
frame rate of an ob ject is assumed to be framessecond The consumption

rates of the in terv als are Kbpf Kilobits p er frame Kbpf Kbpf and
Kbpf resp ectiv ely The length of the in terv als are frames frames frames and frames resp ectiv ely Assume the optimal time p erio d is com
puted to b e second frames The rst in terv al length is
am ultiple of the n um ber of frames in the time p erio d Hence the rst case is
applied whic h results in an in terv al of length time p erio ds and a consumption
rate of Kbpt Kilobits per time p erio d see Figure b
The second in terv al has frames whic h is less than the n um ber of frames in
the time p erio d Therefore the second case is utilized Since the third in terv al
has frames and the summation of the second and the third in terv als frames is
still less than the n um b er of frames in the time p erio d these t woin terv als are
com bined to pro duce an in terv al with frames and a consumption rate of Kbpf Then the second case is applied to the com bined in terv als
Since the fourth in terv al has enough frames frames are b orro w ed from it
and added to the com bined in terv als No w the new in terv al has frames
and a consumption rate of Kbpt The
fourth in terv al has frames after borro wing frames Since the n um
ber of frames in the in terv al is not a m ultiple of the n um ber of frames in the
time p erio d the third case is applied This in terv al is divided in to t woin terv als
The rst one has frames and the second one has frames Using the rst
case the rst in terv al is con v erted to an in terv al of length time periods and
a consumption rate of Kbpt The second in terv al has few er
frames than the time p erio d and it is adjusted to an in terv al of length one time
p erio d and a consumption rate of Kbpt The second case is
not applied since it is the last in terv al Notice that the new in terv al plan has a
length of seconds while the original one has a length of seconds b ecause of
the adjustmen t to the optimal time p erio d
The retriev al plans in the rep ository are expressed in terms of their
resp ectiv e optimal time p erio ds The increase in the clien tbuer dueto the ad justmen t is less than ! of its original size whic his v ery marginal This increase is
due to the clien t startup latency The clien t startup latency is expressed in terms
of n um ber of frames in the original retriev al plans Since these retriev al plans
are expressed in terms of the optimal time periods the clien t startup latency
is required to be expressed in terms of the time p erio ds as w ell T o guaran tee
hiccupfree displa ythe con v ersion is rounded up The amountof rounding up
is the cause of the increase in the clien t buer T o illustrate assume the clien t
startup latency time is seconds and the time period is second The
clien t startup latency time is required to b e expressed in terms of time p erio d
since the data arriv es at the end of eac h time period As a result the clien t
startup latency time b ecomes d

e seconds time p erio ds
The amoun t of rounding up is second The amoun t of data whic h is accu
m ulated in the clien t is increased b y second w orth of data This increase
in clien t buer can be eliminated bynot transmitting more data to the clien t
buer during the adjusted startup latency time In the example ab o v e the
second w orth of data is not sen t to the clien t Retriev al plans whic h do not ha v e
clien t startup latency do not cause increase in the clien t buer
Conclusion and F uture Researc h Directions
In this studyw e classied and ev aluated disk sc heduling tec hniques in supp ort
of hiccupfree displa y of VBR enco ded clips These tec hniques are categorized in
t w o groups U niv er sal and AtomicEv aluation of the dieren t approac hes w as
based on analytical mo dels that compute the cost p er stream and the a v erage
startup latency Our cost analysis demonstrated the sup eriorit yof Atomic o v er
U niv er salF rom the list of alternativ e Atomic approac hes VR
VIT AL is
the most cost eectiv e tec hnique While the cost of b oth memory and disk is
decreasing Mo ores La w our nal observ ations remain unc hanged as long as
the cost reduction for both t yp es of storage is almost iden tical In the future
w e in tend to ev aluate our taxonom y based on sim ulation mo dels and include
designs to supp ort V CR functions
References
Vide o Stor e Magazine Decem b er
S Berson S Ghandeharizadeh R Mun tz and X Ju Staggered stripping in
m ultimedia information systems In Pr o c e e ding of the A CM SIGMOD International
Confer enc e on Management of Data S Berson L Golub c hik and R Mun tz F ault toleran t design of m ultimedia serv ers
In Pr o c e e dings of the A CM SIGMOD International Confer enc e on Management of
Data Yitzhak Birk T rac kpairing a no v el data la y out for v od serv ers with m ultizone
recording disks In Pr o c e e dings of the International Confer enc e on Multime dia
Computing and Systems pages C S Chang Stabilit y queue length and dela y of deterministic and sto c hatic
queueing net w orks IEEE T r ansactions on A utomatic Contr ol Ma y
E Chang and A Zakhor Admissions con trol and data placemen t for vbr video
serv ers In Pr o c e e dings of the IEEE International Confer enc e on Images Pr o c essing
ICIP v olume pages Austin T exas No v em b er E Chang and A Zakhor V ariable bit rate mp eg video storage on parallel disk
arra ys In First International Workshop on Community Networking Inte gr ate d
Multime dia Servic es to the Home pages San F rancisco July W F eng Rateconstrained bandwidth smo othing for the deliv ery of stored video
T o app e ar in ISTSPIE Multime dia Networking and ComputingF ebruary W F eng F Jahanian and S Sec hrest Optimal buering for the deliv ery of
compressed prerecorded video In Pr o c e e dings of the IASTEDISMM International
Confer enc e on NetworksJan uary W F eng and S Sec hrest Critical bandwidth allo cation for deliv ery of compressed
video Computer Communic ations Octob er W F eng and S Sec hrest Smo othing and buering for deliv ery of prerecorded
compressed video In Pr o c e e dings of the ISTSPIE Symp osium on Multime dia
Computing and Networking pages F ebruary
D J Gemmell H M Vin D D Kandlur P V Rangan and L A Ro w e Multi
media Storage Serv ers A T utorial IEEE ComputerMa y S Ghandeharizadeh Streambased v ersus structured video ob jects Issues so
lutions and c hallenges In Multime dia Datab ase Systems Issues and R ese ar ch
Dir e ctions a b o ok chapter Springer V erlag S Ghandeharizadeh S H Kim and C Shahabi On Disk Sc heduling and Data
Placemen t for Video Serv ers USC T ec hnical Rep ort USCCSTR Univ er
sit y of Southern California S Ghandeharizadeh S H Kim C Shahabi and R Zimmermann Placementof
con tin uous media in m ultizone disks In So on M Chung e ditor Multime dia Infor
mation Stor age and Management chapter Klu w er Academic Publisher Boston
S Ghandeharizadeh J Stone and R Zimmermann T ec hniques to quan tify scsi
disk subsystem sp ecications for m ultimedia T ec hnical Rep ort USCCSTR
USC M Grossglauser S Kesha v and D Tse Rcbr A simple and ecien t service for
m ultiple timescale trac In Pr o c e e dings of the A CM SIGCOMM pages August S R Heltzer J M Menon and M F Mitoma Logical data trac ks extending
among a pluralityof zones of ph ysical trac ks of one or more disk devices US
P aten t No April E W Knigh tlyD E W rege J Lieb eherr and H Zhang F undamen tal limits and
tradeos of pro viding deterministic guaran tees to vbr video trac In Pr o c e e dings
of the A CM SIGMETRICS pages Ma y J M McMan us and K W Ross Prerecorded vbr sources in atm net w orks
Piecewiseconstan trate transmission and transp ort Submitte d for public ation Septem b er J M McMan us and K W Ross Video on demand o v er atm Constan trate
transmission and transp ort In Pr o c e e dings of IEEE INF OCOM pages Marc h G Nerjes P Muth and G W eikum Sto c hastic service guaran tees for con tin uous
data on m ultizone disks In Pr o c e e dings of the th Symp osium on Principles of
Datab ase SystemsT ucson Arizona Ma y R Ng and J Y ang Maximizing buer and disk utilization for news ondemand
In Pr o c e e dings of the International Confer enceon V ery L ar ge Datab ases
C Ruemmler and J Wilk es An in tro duction to disk driv e mo deling IEEE Com
puter pages Marc h
J D Salehi Z L Zhang J F Kurose and D T o wsley Supp orting stored video
Reducing rate v ariabilit y and endtoend resource requiremen ts through optimal
smo othing In Pr o c e e dings of A CM SIGMETRICS pages Ma y
K Salem and H GarciaMolina Disk striping In Pr o c e e dings of International
Confer enc e on Datab ase Engine eringF ebruary FA T obagi J P ang R Baird and M Gang Streaming RAIDA Disk Arra y
Managemen t System for Video Files In First A CM Confer enc e on Multime dia August B L W orthington G R Ganger Y N P att and J Wilk es Online extraction of
scsi disk driv e parameters In Pr o c e e dings of the A CM SIGMETRICSMa y PS Y u M S Chen and D D Kandlur Design and analysis of a group ed sw eeping
sc heme for m ultimedia storage managemen t In Pr o c e e dings of the Thir d Interna
tional Workshop on Network and Op er ating System Supp ort for Digital A udio and
Vide oNo v em b er
PS Y u M S Chen and D D Kandlur Group ed sw eeping sc heduling for dasd
based m ultimedia storage managemen t Multime dia Systems Jan uary

Linked assets

Computer Science Technical Report Archive

Conceptually similar

PDF

USC Computer Science Technical Reports, no. 659 (1997)

PDF

USC Computer Science Technical Reports, no. 864 (2005)

PDF

USC Computer Science Technical Reports, no. 610 (1995)

PDF

USC Computer Science Technical Reports, no. 590 (1994)

PDF

USC Computer Science Technical Reports, no. 650 (1997)

PDF

USC Computer Science Technical Reports, no. 634 (1996)

PDF

USC Computer Science Technical Reports, no. 602 (1995)

PDF

USC Computer Science Technical Reports, no. 612 (1995)

PDF

USC Computer Science Technical Reports, no. 863 (2005)

PDF

USC Computer Science Technical Reports, no. 627 (1996)

PDF

USC Computer Science Technical Reports, no. 615 (1995)

PDF

USC Computer Science Technical Reports, no. 865 (2005)

PDF

USC Computer Science Technical Reports, no. 653 (1997)

PDF

USC Computer Science Technical Reports, no. 628 (1996)

PDF

USC Computer Science Technical Reports, no. 630 (1996)

PDF

USC Computer Science Technical Reports, no. 587 (1994)

PDF

USC Computer Science Technical Reports, no. 622 (1995)

PDF

USC Computer Science Technical Reports, no. 766 (2002)

PDF

USC Computer Science Technical Reports, no. 623 (1995)

PDF

USC Computer Science Technical Reports, no. 619 (1995)

Description

Jaber A. Al-Marri and Shahram Ghandeharizadeh. "An evaluation of alternative disk scheduling techniques in support of variable bit rate continuous media." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 666 (1998).

Asset Metadata

Creator Al-Marri, Jaber A. (author), Ghandeharizadeh, Shahram (author)

Core Title USC Computer Science Technical Reports, no. 666 (1998)

Alternative Title An evaluation of alternative disk scheduling techniques in support of variable bit rate continuous media (title)

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)

Tag OAI-PMH Harvest

Format 30 pages (extent), technical reports (aat)

Language English

Unique identifier UC16270556

Identifier 98-666 An Evaluation of Alternative Disk Scheduling Techniques in Support of Variable Bit Rate Continuous Media (filename)

Legacy Identifier usc-cstr-98-666

Format 30 pages (extent),technical reports (aat)

Rights Department of Computer Science (University of Southern California) and the author(s).

Internet Media Type application/pdf

Source 20180426-rozan-cstechreports-shoaf (batch), Computer Science Technical Report Archive (collection), University of Southern California. Department of Computer Science. Technical Reports (series)

Access Conditions The author(s) retain rights to their work according to U.S. copyright law. Electronic access is being provided by the USC Libraries, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Repository Email csdept@usc.edu

Inherited Values

Title Computer Science Technical Report Archive

Description Archive of computer science technical reports published by the USC Department of Computer Science from 1991 - 2017.

Coverage Temporal 1991/2017

Repository Email csdept@usc.edu

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)