Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
Computer Science Technical Report Archive
/
USC Computer Science Technical Reports, no. 618 (1995)
(USC DC Other)
USC Computer Science Technical Reports, no. 618 (1995)
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
Av oiding Retriev al Con ten tion for Comp osite Multimedia Ob jects
Sura jit Chaudh uri
HewlettP ac k ard Labs
P alo Alto CA Shahram Ghandeharizadeh Cyrus Shahabi
Univ ersit y of Southern California
Los Angeles California Abstract
An imp ortan t requirementfor m ultim edia presen tations is the abilit y to comp ose new m ul
timedia ob jects from the existing ones using temp oral relationships When comp ositions of
con tin uous media ob jects are sp ecied dynamically the task of displa ying these ob jects p oses
new c hallenges These c hallenges are addressed in this pap er W esho w that in the case of a sin
gle comp osite ob ject retriev al a prefetc hing tec hnique simple sliding pro vides an approachto
reduce latency and buering requiremen ts W e extend this prefetc hing tec hnique to the problem
of retrieving m ultiple comp osite ob jects sim ultaneously This new tec hnique is termed buer e d
slidingW e consider sev eral v arian ts of the buered sliding algorithm A sim ulationbased study
is used to compare their usage pattern of a v ailable memory and in determining their relativ e
merits in reducing latency and increasing disk bandwidth utilization
In tro duction
An imp ortan t requirementfor m ultimedia information systems is the abilit y to comp ose new m ul
timedia ob jects from the existing m ultimedia ob jects LG T emp oral primitiv es eg b efore
after o v erlaps All pro vide one of the most p o w erful and natural w a ys of authoring comp osition
Suc h comp osition is necessary in the domain of electronic publishing computer m usic news editing
and man y other applications
In this pap er w ein v estigate howa m ultimedia storage system can displa y a comp osite ob ject
W e fo cus on comp osite ob jects that are authored dynamicallyT o illustrate an example en viron
men t consider a TVnews editor preparing to presen t new fo otage on unrest in Bosnia He requires
This pap er app eared in the Pro ceedings of the VLDB Conference Researc h supp orted in part b y the
National Science F oundation under gran ts IRI IRI NYI a w ard and CD A and a Hewlett
P ac k ard unrestricted cashequipmen tgift
bac kground material to pro vide the audience with a con text He considers pla ying a sequence of
clips one after another from dieren t fo otage tak en at dieren t times to author a thirt y second
presen tation He ma y decide to accompan y a fo otage with appropriate m usic in parts ie m usic
o v erlaps video He ma y conclude his presen tation with split windo ws that concurren tly displa y
short clips that lea v e us with the images of div erse scenes in Bosnia During editing of suc hapre sen tation he w ould try sev eral p ossible comp osition p ossibly pic king dieren t sets of clips or m usic
from the rep ository Surely the editor w ould lik e to displa y his comp osition during the authoring
pro cess to ev aluate his c hoice Th us the pro cess of editing a news story consisted of sp ecifying
comp osite ob jects using temp oral relationships and then displa ying those
Note that displa ying atomic ob jects of highbandwidth con tin uous media ob jects suc h as video
requiring no comp osition is a c hallenging task in itself Video clips require a con tin uous bandwidth
for their displa yF or example the bandwidth required b yNTSC
for net w orkqualit y video is
ab out megabits p er second m bps Has Ev en with a compression tec hnique that pro vides a
reasonable qualit y of presen tation the bandwidth requiremen t of the compressed ob ject is t ypically
quoted as to m bps Nat T ri Video ob jects are large in size and almost alw a ys residenton
a secondary storage device T o supp ort a con tin uous displa y an ob ject should b e retriev ed at its
presp ecied rate Otherwise its displa yma y suer from frequen t disruptions and dela ys termed
hic cupsSev eral studies ha v e describ ed tec hniques to supp ort con tin uous displa y of a video clip
P ol SG TPBG CL BGMJ GKS These studies con trol the la y out of data on a
secondary storage device to ensure the con tin uous displa y of an ob ject
When displa ying a comp osite ob ject the system should ensure that the temp oral constrain ts
are satised Since the comp osition is sp ecied dynamically no assumption can b e made ab out
the placemen t of the atomic ob jects F urthermore w ew ould not b e able to mo dify the placemen t
of data eac h time a new presen tation is authored b ecause it ma y force the user to observ e a long
w ait time b efore the displaystarts In fact retriev al of the atomic ob jects that constitute the
comp osite ob ject ma y conict with one another making signican t demands on buering as will
b e indicated later It is this unique asp ect of retriev al con ten tion that distinguishes this study
from the previous w ork in b oth realtime sc heduling and sync hronization in distributed m ultimedia
information systems See Buf for a surv ey Ho w ev er our results need to b e used in conjunction
with the past tec hniques for a complete solution to the problem of displa ying comp osite ob jects
The general problem of displa ying comp osite ob jects is a complex one In this pap er w e fo cus
on an imp ortan t sp ecial case of this problem where the atomic ob jects that comprise the comp osite
ob ject ha v e the same bandwidth requiremen ts eg o v erlapping video newsclips Ho w ev er w e
m ust still takein to accoun t that the p ossibilit y that the bandwidth requiremen ts can b e high F or
The US standard established b y the National T elevision System Committee
the sak e of ease of exp osition w e will presen t the results in the con text of binary comp osite ob jects
ie where the comp osite ob ject consists of only t w o atomic ob jects Ho w ev er w eha v e included
sk etc hes of generalizations necessary to retriev e arbitrary comp osite ob jects
W e b egin with a discussion on the c hallenges of retriev al of comp osite ob jects Section W e showhowb y using a tec hnique termed simple sliding w e need no more than a constan t
extra buering ie indep enden t of the sizes of the individual atomic ob jects and the extentof
o v erlap to supp ort displa y of a single comp osite ob ject Section Our new tec hnique uses the
prop erties of la y outs of atomic ob jects to compute a sc hedule of retriev al W e prop ose algorithms
based on prefetc hing using a generalization of simple sliding that enable the system to displa y
m ultiple comp osite ob jects Section The p erformance c haracteristics of these algorithms using
the metrics of usage patterns for a v ailable memory bandwidth utilization and latency are studied
in Section The rest of this section pro vides a discussion of the framew ork for our prop osed
tec hniques
F ramew ork
As men tioned in the in tro duction w e assume that all ob jects b elong to a single media t yp e with
iden tical bandwidth requiremen t B
D isplay
W e assume that the simple striping SGM P at tec hnique is used to retriev e video and audio ob jects for con tin uous displa y In this tec hnique w e
partition the aggregate bandwidth of the a v ailable disks in to c hannels eac h with the bandwidth
requirementof B
D isplay
Hence denoting the ee ctive disk bandwidth considering the maxim um
seek and latency time see BGMJ with B
Disk
the n umberofc hannels R is computed as
R b
D
M
c
where D is the n um b er of disk driv es and M B
D isplay
B
Disk
Note that R determines the n um ber of
sim ultaneous requests supp orted b y the system W e will denote these R c hannels b y C
C
C
R
When M eac hc hannel pro vides a fraction of a disk bandwidth see Figure a If M then eac h sub ob ject is declustered GRA Q in to M fragmen ts and the aggregate bandwidth of
M disk driv es determine the required bandwidth of a c hannel see Figure b
In order to distribute the load imp osed b y a displayof object X ev enly across the c hannels
simple striping tec hnique strip es X in to sub ob jects and assigns the sub ob jects of X to the c hannels
in a roundrobin manner The size of eac h sub ob ject is xed for all ob jects and determined at system
conguration time F or example in Figure a sub ob jects of X X
X
X
are assigned
to an c hannel system starting with C
Simple striping also minimizes the amoun tof memory
required b y pip elinin g the data from the disk to a displa y When the system displa ys an ob ject X
Figure Logical c hannels in simple striping
it emplo ys a single c hannel during eac h time in terv al Th us to ensure a con tin uous displayof an
ob ject X the displa yof X
i
is o v erlapp ed with the retriev al of X
i BGMJ The duration of
the retriev al of a sub ob ject is xed for all sub ob jects and is termed a time intervalEac h displa y
iterates o v er the c hannels emplo ying eac h in a roundrobin manner T o illustrate in Figure a
assume a time in terv al that corresp onds to C
is a v ailable Emplo ying that in terv al X
can b e
displa y ed In the next in terv al the system emplo ys C
to retriev e and displa y X
This pro cess
rep eats itself un til X is displa y ed en tirelyA t the same time R other in terv als can b e utilized
to service R other requests
A tomic Ob jects
An atomic ob ject X is a video clip that is retriev ed in a sequen tial manner Its sub ob jects are
assigned to the c hannels in a round robin manner starting with an arbitrary c hannel W e use the
function channel X
i
to determine the c hannel that con tains sub ob ject X
i
Let channel X
i
be c
i
In a system that consists of R c hannels if X consists of k sub ob jects then w e can also represen t
the la y out of X as the list c
c
k
where eac h c
i
f R gT o illustrate assume a system
consisting of three c hannels R The database consists of t w o ob jects X and Y Both ob jects
are assigned to the c hannels in a roundrobin manner starting with c hannel Eac hof X and Y
consist of v e and eigh t sub ob jects resp ectiv ely In this case X is represen ted as while Y is represen ted as
Sc hedule
The sche dule for the storage system reects the state of the disks It can b e represen ted b y sp ecifying
the set of busy c hannels for eachtimein terv al W e representasc hedule S as an arra y The index of
the arra y is time v arying from no w to T where T is time un til whic h at least one of the c hannels
is busy Eac h elemen t of the arra y is the set of busy c hannels at that time in terv al F or example
assume that T and consider a sc hedule for the time p erio d through S f gS f g S f gS f g This states that at time c hannels and w ere busy at time c hannels and w ere busy etc
Problem Denition Retriev al of Comp osite Ob jects
In this pap er w e consider the class of comp osite m ultimedia ob jects that consist of t w o atomic
ob jects The temp oral relationships b et w een the displayof t w o atomic ob jects ma y b e sp ecied in
sev eral w a ys F or simplicit yw e assume the follo wing mo del of sp ecication
A c omp osite obje ct is a triplet X Y j indicating that the comp osite ob ject consists of atomic
ob jects X and Y The parameter j is the lag p ar ameter It indicates that the start time of ob ject
Y ie displa yof Y
is sync hronized with the displa y of sub ob ject X
j
F or example to designate
a complex ob ject where the displayof X and Y m ust start at the same time w e will use the
notation X Y Lik ewise the comp osite ob ject sp ecication X Y indicates that the displa y
of Y is initiated with the displa y of the third sub ob ject of X see Figure a This denition of a
comp osite ob ject supp orts the alternativ e temp oral relationships describ ed in All T able lists
these temp oral relationships and their represen tation using our notation of a comp osite ob ject Our
prop osed tec hniques supp ort all temp oral constructs b ecause they solv e for arbitrary j v alues
arbitrary sizes for b oth X and Y and placemen tof X and Y starting with an arbitrary
c hannel
Since t woo v erlapping ob jects need to b e displa y ed at the same time during retriev al of the
o v erlapping part of a comp osite ob ject request for retriev al of suchan object m ust reserv e the righ t
n um ber of c hannels at appropriate times in order to ensure a con tin uous displa y Moreo v er the
iden tit y of reserv ed c hannels is imp ortan t due to placemen t of data Without prop er precautions
the system ma y fail to supp ort con tin uous displa y of comp osite ob ject if the placementofits
participating atomic ob jects collide and reference sub ob jects that are stored on the same c hannel
T o illustrate assume a system consisting of c hannels Consider a single request displa ying the
comp osite ob ject X Y where X is and Y is Note that
the retriev al of X
and Y
conict resulting in hiccups This illustrates the problem of retriev al
X before Y
X equal Y
X meets Y
X overlaps Y
X during Y
X starts Y
X finishes Y
XXX YYY
XXX
YYY
XXXYYY
XXX
YYY
XXX
YYYYYY
XXX
YYYYY
XXX
YYYYY
(X, Y, j)
(X, Y, j)
(X, Y, j)
(X, Y, j)
(Y, X, j)
(X, Y, j)
(Y, X, j)
size(X) < j
size(X) = size(Y) & j = 1
j = size(X) + 1
1 < j <= size(X)
j > 1 & size(X) <= size(Y) - j
j = 1 & size(X) < size(Y)
j = size(Y) - size(X) + 1 & size(X) < size(Y)
Allen Relations
Composite Object Construct
Figure Allen temp oral relationships and their represen tation using our notation of a comp osite
ob ject
of comp osite ob jects There could b e c ontentions within a single comp osite ob ject b ecause eac h
atomic ob ject is laid out indep enden tly Suchcon ten tion arises ev en though there are no other
activ e requests This con ten tion is formalized b elo w
Denition A comp osite ob ject X Y j has an internal c ontention if there is some i suc h that channel X
i j channel Y
i
Note that our notation extends naturally to sp ecication of comp osite ob jects that con tain
more than one atomic ob jects Th us a comp osite ob ject con taining n atomic ob jects can b e
c haracterized by n lag factor eg X
X
n
j
j
n
where j
i
denotes the lag factor of X
i
with resp ect to the b eginning of the displayof X
Note that Denition can b e generalized in a
straigh tforw ard fashion for the case where the comp osite ob ject con tains m ultiple atomic ob jects
One migh t attempt to address the in ternal con ten tion problem similar to retriev al of atomic
ob jects in a m ultiuser en vironmen t This is not appropriate for the follo wing reason The displa y
of eac h atomic ob ject in a m ultiuser en vironmentis indep endent Dieren tsc heduling p olicies
ma y result in increased latency but it is alw a ys p ossible to obtain a sc hedule that requires no
additional buer Ho w ev er b ecause of the sync hronization constraintbet w een the displa ys of
o v erlapping ob jects treating the ob jects in the comp osite ob ject as indep enden t ob jects ma y require
signican t buering Indeed when there is con ten tion within a comp osite ob ject buering migh t
b e una v oidable In our previous example where X and Y collide while sc heduling X and Y
indep enden tly do es not need an y buer if latency is acceptable retriev al of X Y requires at
least one buer
In the rest of the pap er w e will address the question of retriev al of comp osite ob jects W e
will study the solutions using three metrics latency time memory and disk utilization Latency
time is dened as the amoun t of time elapsed from the arriv al time of a request to the onset of
the displa y of its referenced comp osite ob ject Both memory and disk utilization are dened as
the amoun t of time that the memory and disks are busy supp orting displa ys as a function of total
time resp ectiv ely In the follo wing section w e study the problem of displa ying a single comp osite ob ject Section extends the study to en vironmen ts that supp ort sim ultaneous displa y of sev eral comp osite ob jects
Single DisplayEn vironmen t
In this section w e describ e ho w to minimize b oth the buer requiremen ts and latency of a displa y
referencing a single comp osite m ultimedia ob ject
Naiv e Prefetc hing Strategy Let us assume that the sp ecication of the comp osite ob ject
is X Y j where X and Y ha v e kand k sub ob jects resp ectiv ely ie X X
X
k and
Y Y
Y
k The naiv e solution will b e to prefetc h the en tire o v erlapp ed part ie sub ob jects
Y
Y
u
where u min k j k Ha ving fetc hed the o v erlapping part w e can no w b egin fetc hing
the sequence of sub ob jects for X and then conclude byfetc hing the rest of Y Th us the sequence
of retriev al ma y b e depicted as follo ws
Y
Y
u
X
X
k Y
u Y
k Ho w ev er note that this naiv e metho d incurs a latency of u time in terv als and requires u buers
Th us in the example of Figure a w e will incur an o v erhead of buers and a latency of time in terv als Note that if a system is congured with a single c hannel R then naiv e
prefetc hing is the only alternativ e that is a v ailable In suc h a case naiv e prefetc hing reduces
to retrieving the t w o ob jects one after the other The con ten tion prefetc hing strategy is the
ob vious impro v emento v er the naiv e prefetc hing Giv en a comp osite ob ject X Y j it prefetc hes
only those sub ob jects Y
i
for whic h there is a retriev al con ten tion with the corresp onding sub ob jects
of X ie channel Y
i
channel X
i j Ho w ev er naiv e prefetc hing and con ten tion prefetc hing
b eha veinan iden tical manner b ecause the sub ob jects of eac h atomic ob ject are assigned to the
c hannels in a roundrobin manner Ho w ev er con ten tion prefetc hing is a desirable metho d when the
la y out is not roundrobin Assume that w eha vetodispla y the comp osite ob ject X Y where the
la y out of X and Y are and resp ectiv ely In that case the
naiv e prefetc hing w ould prefetc h the rst sub ob jects of Y whic h is the o v erlapping part of the
ob ject Y In con trast the con ten tion prefetc hing strategy will prefetc h only the third and the sixth
sub ob jects of Y since these are only sub ob jects whose retriev al w ould conict with the retriev al
of the corresp onding sub ob jects of XCon ten tion prefetc hing is appropriate when a presen tation
consists of sev eral comp osite ob jects The scenarios where con ten tion prefetc hing is desirable are
describ ed in CGSon
Simple Sliding Algorithm The amoun t of prefetc hing in naiv e and con ten tion prefetc hing
tec hnique gro ws as the o v erlap b et w een the atomic ob jects increase Moreo v er neither of these
tec hniques exploit the roundrobin assignmen t of sub ob jects In con trast b y exploiting the la y out
simple sliding tec hnique minimizes b oth latency and buer requiremen t of a displa y indep endent
of the size of o v erlap as long as R
T o illustrate the idea of simple sliding consider Figure Figure a sho ws a comp osite ob ject
Z whic h comprises of t w o atomic ob jects X and Y where the lag parameter is Assuming
X
and Y
reside on t w o dieren tc hannels the retriev al and displayof Z can b e depicted as in
Figure b T rivially no extra buer is required in this case No w assume that channel Y
channel X
Exploiting the roundrobin la y out of the ob jects and assuming R w ekno w
that b oth channel Y
channel X
and channel Y
channel X
Th us the system has
t woc hoices either slide the retriev al of Y up for one in terv al upslide Y and prefetc h one
sub ob ject of Y see Figure c or slide the retriev al of Y do wn for one in terv al do wnslide Y and prefetc h one sub ob ject of X see Figure d Note that do wnsliding Y in tro duces latency The
rest of this section pro vides a formal description of simple sliding
Wesa y that an ob ject X has a r e gular layout cycle if the sub ob ject X
j
is placed at a c hannel
adjacen t to X
j
channel X
j channel X
j mod R
F rom this denition it follo ws that regular la y outs dier from one another only in the placemen tof
their rst sub ob jects Observ e that simple striping enforces precisely this prop ert yF or example
t w o regular la y outs of X o v er v ec hannels mightbe An alternativ ela y out
for X mightbe Th us regular la y out and simple striping are iden tical
Lemma Consider the r etrieval of the c omp osite obje ct X Y j wher eb oth X and Y have a
r e gular layout Ther eis a r etrieval c ontention i channel X
j
channel Y
X1 X1
X2, Y1 Y1 X2
X3, Y2 Y2 X3, Y1
X4, Y3 Y3 X4, Y2
X5, Y4 Y4 X5, Y3
Y4
c. Case II: upslide Y
disk channels Memory Display Time
d. Case III: downslide Y
X1 X1 --
X2 X2 X1
X3 X3 X2
X4, Y1 X4 X3, Y1
X5, Y2 X5 X4, Y2
Y3 X5, Y3
Y4 Y4
disk channels Memory Display Time
b. Case I (No prefetching)
disk channels Memory Display Time
X1 X1
X2 X2
X3, Y1 X3, Y1
X4, Y2 X4, Y2
X5, Y3 X5, Y3
Y4 Y4
a. Object Z
X1 X2 X3 X4 X5
Y1 Y2 Y3 Y4
Overlap
Composite Object Z: (X, Y, 3)
0
1
2
3
4
5
0
1
2
3
4
6
0
1
2
3
4
5
Figure Simple Sliding
As men tioned earlier naiv e prefetc hing is the only option when there is single c hannel Ho w
ev er the more in teresting case is where R In suc h a case the follo wing iden tit y holds j since the sub ob ject X
j m ust b e on a dieren tc hannel from the sub ob ject Y
channel X
j
channel Y
channel X
j channel Y
In other w ords when R then w ekno w that X Y j has no con ten tion if X Y j has
con ten tion Therefore in suc ha casew ecan pr efetch sub ob ject Y
during retriev al of X
j Suc h
upsliding results in zer o latencyHo w ev er it requires one extrabuer during the displayof Y In case j thenw e use the observ ation similar to ab o v e that Y X do es not ha vean y
con ten tion whenev er X Y do es ha vecon ten tion This motiv ates a tec hnique that prefetc hes
Y
and then b egin fetc hing sub ob jects of X concurren tly with the retriev al of sub ob jects of Y This do wnsliding results in latency of one time in terv al In addition it requires one extr a buer
during the displa yof X
W e refer to this strategy as simple slidingF or the class of binary comp osite ob jects of the
form X Y j it is the optimal algorithm for retriev al when only a single displa y request is
activ e and ob jects ha v e regular la y outs as in simple striping It is notew orth y that b oth the
buer requiremen t and latency of a displayis indep endent of the size of the comp osite ob ject This
algorithm is summarized b elo w W eha v e omitted the details of buering as they ha v e already b een
explained
Algorithm Simple SlidingXYj
if there is a single c hannel then
do naiv e prefetc hing
else
if j then fetch Y X
else fetch X Y j Assuming that the system m ust o v erlap the displayoft w o atomic ob jects that constitute a
comp osite ob ject this algorithm requires at most one extra buer Moreo v er its incurred latency
is b ounded b y one time in terv al The tec hnique of simple sliding is surprisingly robust ev en where
the atomic ob jects do not satisfy the regular la y out assumption F or example the tec hnique can
b e adapted with mo dications for the case where new c hannels are added
Weno w briey sk etc h the generalization of Simple Sliding when the comp osite ob ject consists
of more than t w o atomic ob jects eg X
X
n
j
j
n
F or simplicit y assume that all the ob jects
X
i
i n are m utually o v erlapping The generalization ensures that the comp osite ob ject can
b e displa y ed with the n um b er of extra buers as w ell as latency indep enden t of the dur ation of
o v erlap among atomic ob jects In tuitiv ely since the sliding tec hnique m ust use buers indep enden t
of the extentof o v erlap it m ust assign a c hannel for displa y of eachofthe n ob jects Therefore
R n F or binary comp osite ob jects the condition reduces to R whic hweha v e already
iden tied as a necessary condition for use of simple sliding for binary comp osite ob jects It can b e
sho wn that the buer requiremen ts are at most n n and the latency is at most n The
details of the generalization of the algorithm will b e rep orted in CGS
Multi DisplayEn vironmen t
The problem of retriev al con ten tion b ecomes c hallenging in the presence of m ultiple users displa ying
ob jects sim ultaneously This is due to comp etition for disk bandwidths b ym ultiple indep enden t
requests arriving at random times In particular in the case of a single user en vironmen t w e
established that using an extra buer it is p ossible to b ound the latency for retriev al with simple
sliding Suc h is no longer the case for m ultiuser case While a single extra buer is still sucien t
to remo vein ternal con ten tion there ma y b e signican t latency due to demands on bandwidths
b y activ e users and activ e users colliding and referencing data items that reside on one c han
nel Increased a v ailabilit y of memory ma y b e utilized to prefetc h additional ob jects that lead to
con ten tion Therefore a k ey issue in m ultiuser en vironmen t will b e to study the role of memory
to reduce latency
W e assume that the sc heduler in a m ultiuser system consists of t w o mo dules A priorit y
go v erning and A task assignmen t mo dule The prioritygo v erning mo dule is resp onsible for
determining the priorit y of tasks among the queued jobs Examples of heuristics that go v ern the
prioritygo v ernor ma y include First Come First Serv e Shortest Job First Giv en a request to displa y
a comp osite m ultimedia ob ject the task assignmen t mo dule sc hedules it b y taking in to accoun t
the curren t status of the c hannels The fo cus of our study is on the task assignmen t mo dule
F or simplicit yw e will assume that sc heduling is nonpr e emptive ie once a task is sc heduled its
sc hedule remains unc hanged
Buered Sliding Algorithms
The class of algorithms that w e presen t here try to minimize latency b y using a v ailable buers Let
us assume that due to con ten tion w ecannot sc hedule X Y j j b eginning at time t but can
sc hedule X Y j b eginning at time t In that case w e can use an extra buer to prefetc h Y
during retriev al of X
j By virtue of prefetc hing w e are then able to display X Y j b eginning
at time tW e can generalize the idea to the case where if the displayof X Y j b is p ossible
b eginning at time t then using b buers w e can display X Y j b eginning at time t assuming
j b W e can think of suc h use of buers as sliding upwar ds Let us no w assume that it is also
p ossible to sc hedule X Y j b eginning at time tIn suc h a case displa yof X will b e ahead
of displayof Y by j and b y using an extra buer to prefetc h X it is p ossible to displa y
X Y j b eginning at time t W e can refer to suchasc heduling tec hnique as sliding downwar d In the general case giv en b extra buers if displa yof X Y j b is p ossible b eginning at time t then displayof X Y j can start b eginning at t b The ab o vetec hnique of sliding up w ards and
do wn w ards is a generalization of prefetc hing based on simple sliding algorithm that w as presen ted
in Section Ho w ev er w e are nowallo wing use of m ultiple buers for prefetc hing Informally w e
refer to these as buer e d sliding Observ ethat iffor sc heduling of a displa y sliding up w ards and
do wn w ards require the same amountof memory then w em ust prefer sliding up w ards b ecause it
do es not adv ersely impact latency W e will no w describ e the algorithms that use buered sliding to a v oid con ten tion so that
latency can b e reduced and disk utilization can b e increased Giv en the currentsc hedule S these
algorithms tak e as input a parameter B whic h is the maxim um n um b er of extra buers used for
sliding as w ell as the sp ecication of the comp osite ob ject X Y j to be sc heduled A k ey function
in v ok ed b y buered sliding algorithms is atomic schedul e af ter X S u S
u
It consumes as its
parameters an atomic ob ject X the curren tsc hedule S a time u It outputs a time u
whichis the
earliest p oin t in time when X can b e displa y ed ie u
is the minim um v alue greater than usuc h
that channel X
i
is free during in terv al u
i for eac h i k The other output parameter
Algorithm First Matc h
u
rep eat
atomic schedul e af ter X S u
S
u
atomic schedul e bef or e Y S
u
j S
u
if u
u
j B then return S
atomic schedul e af ter X S
u
j S
u
if u
u
j B then return S
u
u
forev er
a First Match A lgorithm
Algorithm Exhaust B
u
rep eat
atomic schedul e af ter X S u
S
u
atomic schedul e af ter Y S
u
j B S
u
if u
u
j B then return S
u
u
forev er
b Exhaust BA lgorithm
Figure Algorithms for MultiUser En vironmen t
S
indicates the up dated sc hedule The pro cedure atomic schedul e bef or e X S u S
u
is similar
to atomic schedul e af ter except that it computes an u
that precedes u but is closest to it Th us
atomic schedul e af ter and atomic schedul e bef or e corresp ond to sliding up and do wn resp ectiv ely
using at most B extra buers
The k ey functions b oth atomic schedul e af ter and atomic schedul e bef or e should examine
the a v ailabilityof both c hannels and memory p er in terv al Although the corresp onding c han
nels migh t b e free the system ma y not b e able to sc hedule X b ecause of the memory deciency This is b ecause b y sliding ob ject Y up w ard do wn w ard the system is using prefetc hing During
prefetc hing the memory requiremen ts of a displa y has three states a gro wing state a steady state
and a shrinking state Once the time in terv als that the ob jects displa y and retriev al starts are
xed the memory requiremen t p er in terv al can b e computed accurately see SG for a detailed
description
The functions atomic schedul e af ter and atomic schedul e bef or e emplo y the mo d
els in tro duced in SG to compute the memory requiremen t for ev ery in terv al Subsequen tly if the memory requiremen t for an in terv al exceeds the maxim um a v ailable memory they con tin ue
searc hing for another candidate in terv al to initiate the displa y of a candidate ob ject
First Matc h and Exhaust B
Giv en B extra buers there could b e a family of algorithms based on alternativ e approac hes
that tra v erse the searc h space to incorp orate the sc hedule of a displa y with the curren t system
sc hedule corresp onding to activ e displa ys In our w ork w eha vein v estigated t w o algorithms The
In SG w e assumed a graph that describ es temp oral relationshi ps b et w een video clips Using this graph w e
describ ed tec hniques that con trol the placemen t of the data with the ob jectiv e to minimize b oth the latency and
the buer requiremen ts of a displaytra v ersing a path of this graph While this study is dieren t as it assumes the
existence of no graph the computation of its memory requiremen t is iden tical to that of SG
First-match
Exhaust-B
(no prefetching)
(2 subobject prefetching)
Z : (X, Y, 4)
R = 8
X starts from Channel 1
Y starts from Channel 6
B > 3
Busy
Free
Channels
Schedule
1
2
3
4
5
6
7
8
9
10
8 12 34 5 6 7
X1
X2
X3
X4
X5
Y1
Y2
Y3
Y4
Y1
Y2
Y3
X6
Y4
Figure ExhaustB vs First Matc h
algorithm First Match is c onservative in its approac h of using extra buers On the other hand
the algorithm Exhaust B tries to gr e e dily use all B extra buers for sliding Th us the latter tries
to minimize latency at the cost of increased memory whereas the former tries to minimize the use
of extra buers for the curren t displa yGiv en a w orkload it is in teresting to see whic h of the t w o
approac hes lead to lo w er o v erall latency for a giv en amountofmemory In the p erformance section
w e will address these questions
The algorithm First Match see Figure a assigns ob ject X to the earliest lo cation where it
can b e sc heduled as an atomic ob ject sa y u It then tries to see whether Y can b e sc heduled
at u j If so then none of the extra B buers will b e needed Otherwise it lo oks for a matc h
nearest to u j sliding up w ards un til it exhausts all B buers The reason for sliding up w ards is
to a v oid increasing latencyIf w e fail to nd a suitable matc h then attempts are made to sc hedule
Y b y sliding do wn w ards using at most B buers This increases latency If this also fails a new
time in terv al is sough t to displa y X and the ab o v e steps are rep eated
In con trast the algorithm Exhaust B see Figure b uses the extra B buers greedily Using
the buers it slides Y up as high as p ossible u j B for sc heduling On failure it slides do wn
incremen tally un til i j B This results in utilization of c hannels prior to the currentin terv al in
fa v or of freeing up future c hannels F or example Figure demonstrates a p ortion of the schedul e
arra y for an eightc hannel system Comp osite ob ject Z is dened as X Y and the retriev al
of X is sc heduled to start during the rd in terv al First Matc h will sc hedule the retriev al of Y
from the th in terv al utilizing no memory for prefetc hing This is done although B Ho w ev er
ExhaustB slides Y up w ard for in terv als It cannot slide it higher b ecause the th c hannel is busy
for higher in terv als Finally observ ethat for agiv en B in the w orst case the space of p ossibilities
examined b y these t w o algorithms is iden tical
In the full pap er w e describ e algorithms that implemen t First Matc h and Exhaust B ecien tly
Channels
Schedule
1
2
3
4
5
6
7
8
9
10
Z : (X, Y, 4)
R = 8
X starts from Channel 1
Y starts from Channel 6
Algorithm: Exhaust-B
Memory = 8
Busy
Free
FunB (B = 2)
8 12 34 5 6 7
X1
X2
X3
X4
X5
Y1
Y2
Y3
Y4
Y1
Y2
Y3
Y4
X6
Y1
Y2
Y3
Y4
MaxB (B = 8)
MinB (B = 1)
(No prefetching)
Figure Impact of B
The a v ailabilit y information of c hannels is indexed on the duration and the starting time in terv al
from whichitis a v ailable F or ecien t access the p ossible duration v alues are partitioned in to a
set of buc k ets These data structures help sp eedup the execution time of atomic schedul e bef or e
and atomic schedul e af ter F urthermore for large B these functions can use appro ximating
tec hniques for determining a new sc hedule The regular la y out prop ert y of simple striping is also
utilized extensiv ely in the design of these algorithms CGSon
Estimating B
If the comp osite ob ject has no in ternal con ten tion then the minim um r e quir e d v alue of B is zero In
the w orst case suc hav alue of B w ould result in starting the displa y after all the curren tly sc heduled
displa ys ha v e terminated Ho w ev er as explained in the single user case if the comp osite ob ject has
in ternal con ten tion then the minim um r e quir e d v alue of B is one Increasing B poten tially reduces
latency By adjusting B the algorithm can b e dynamic al ly adjusted for a v ailable buers W e use
the follo wing three simple heuristics to compute the v alue of B See Figure MinB In this case the v alue of B is xed at whic h is the minim um buer requiremen t
This limits the exten t of sliding p ermitted
MaxB This is a greedy approac h where the v alue of B is alw a ys equal to the en tire system
memory This results in the maxim um exibilit y for the buered sliding algorithms As sho wn
in Section this exibilit y cannot b e rendered eectiv eif w e use the First Matc h algorithm
F unB The v alue of B is a function of the system memory and the maxim um n um ber of
displa ys supp orted b y a system
B Max MaxM em
R
The ab o v e function is based on the follo wing in tuition During a time in terv al eac h ob jects
share of memory is M axM em
R
This is b ecause during eac h time p erio d at most R atomic
ob jects are activesim ultaneously Moreo v er eac h comp osite ob ject consists of atomic
ob jects where only one of them is using buers for prefetc hing
Generalization
The structure of the algorithms First Matc h and Exhaust B consist of sc heduling X and Y that
comprise the comp osite ob ject X Y j one after the other If the atomic sc heduling fails to
satisfy the lag factor j and the upp er b ound B on memory then the sc hedulng of X is attempted
with a dieren t temp oral windo wSee Figure The same structure of the algorithms can b e
extended to the case where the comp osite ob ject consists of more than t w o atomic ob jects eg
X
X
n
j
j
n
In suc hacasew e can creat a sc hedule b y deciding an or der of atomic ob jects
This order determines the sequence in whic h the atomic sc heduling of ob jects are done The order
also reects the r everse order in whic h the resc heduling of displa y of atomic ob jects is attempted
The total consumption of buering for a partial sc hedule is accoun ted for If the total consumption
exceeds B then previously determined sc heduling ma y need to b e redone as in the case of a binary
comp osite ob ject In our future w ork w e will b e in v estigating p ossible approac hes to determining
the ab o v e order of atomic sc heduling based on the structure of the ob ject Y et another concern
is the time to determine the sc hedule Regarding assignmentof B to a displa y request w ecan
use the alternativ es discussed in the previous section Observ e that minB is no longer but ma y
dep end on RW e will also in v estigate heurisitcs that ma y b e sensitiv e to structure of the comp osite
ob jects
P erformance Ev aluation
W e implemen ted a sim ulation mo del to compare First Match with ExhaustB and in v estigate
the impact of B and system memory on their p erformance Our exp erimen ts fo cused on binary
comp osite ob jects only Therefore it remains an op en question whether our conclusions generalize
to arbitrary comp osite ob jects First w e describ e the sim ulation mo del Next w e rep ort the results
of our exp erimen ts
An alternativ e heuristic is to dene B as a function of the comp osite ob ject structure Assuming a comp osite
ob ject X Y j lets dene B as Max j
W ein v estigated this heuristic in our exp erimen ts and eliminated it from
this discussion b ecause it did not result in go o d p erformance
Sim ulation Mo del
F or the purp oses of this ev aluation w e assumed a platform of c hannels ie R The
eectiv e bandwidth of eachc hannel is m bps supp orts a single displa y of MPEG compressed
clips The system w as congured with a kilob yte blo c k size In other w ord the size of eac h
sub ob ject is Kb ytes and the duration of a time in terv al is seconds
The database consists of atomic ob jects eac hwitha m bps bandwidth requiremen t The
size of the ob jects w as xed at sub ob jects except for the last exp erimen t where the ob jects had
dieren t sizes Consequen tly the displa y time of eac h atomic ob ject is min utes The rationale
is that ev ery min utes the scene is c hanged for example b y using fadein or fadeout requiring
a small amountof o v erlap The ob jects are assigned to the c hannels in a roundrobin manner
starting with a random c hannel
The total n um b er of requests submitted to the system is Eac h request references a
comp osite ob ject consisting of t w o atomic ob jects W e inquired v et yp es of o v erlaps for the t w o
atomic ob jects within a comp osite ob ject
Small o v erlap the t w o atomic ob jects o v erlap for seconds sub ob jects
Mo derate o v erlap the t w o atomic ob jects o v erlap for seconds sub ob jects
Complete o v erlap the t w o atomic ob jects o v erlap for min utes sub ob jects
Zero o v erlap the t w o atomic ob jects do not o v erlap but meet All
V ariable o v erlap the amoun tof o v erlap b et w een t w o atomic ob jects X and Y are c hosen
randomly b et w een and Min siz e X siz e Y sub ob jects
W e emplo y ed an op en sim ulation mo del for our ev aluation requests arriv eev ery think time in ter
v als W e manipulated the parameter l oad to mo del t w o alternativ e loads on the system Hea vy
and ligh t system load The v alue of think is a function of R Obj Siz e O v er l ap the amoun tof
o v erlap and load think Obj Siz e Ov er lap
R
think think l oad
If l oad the requests arriv e so far apart that at least one idle c hannel is a v ailable when a
request arriv es ligh tly loaded system F or load the compartmen t of requests arriv al is tigh ter
smaller think time imp osing a higher load on the system
0 100 200 300 400 500
1700
1800
1900
2000
2100
Size of Memory (in number of subobjects)
Avg. Latency (in Seconds)
MaxB
MinB
FunB
0 100 200 300 400 500
1800
2000
2200
2400
2600
2800
3000
3200
Size of Memory (in number of subobjects)
Avg. Latency (in Seconds)
MinB
MaxB
FunB
0 100 200 300 400 500
4500
5000
5500
6000
Size of Memory (in number of subobjects)
Avg. Latency (in Seconds)
MinB
FunB
MaxB
a Ov erlap b Ov erlap c Ov erlap Figure First Matc h Hea vy loaded system
Exp erim en t Results
F or all the exp erimen ts wev ary the amountofa v ailable buers from one to sub ob jects the
xaxis of all the graphs and w e measured the a v erage latency time in n um ber of in terv als the
yaxis Note that the disk utilization is in v ersely prop ortional to the a v erage latency time That
is as the a v erage latency time decreases the total disk utilization increases This is b ecause for
lo w er a v erage latency the same n um b er of requests w ere serviced in a relativ ely smaller n um ber of
in terv als Hence the total disk utilization w as higher
Figure demonstrates the results of the rst set of exp erimen ts where the system w as hea vily
loaded
load and First Matc h algorithm w as emplo y ed In Figure a all the comp osite
ob jects consisting of atomic ob jects with t w o sub ob ject o v erlap When B MinB the graph
lev els o at M axM em b ecause the extra amoun t of buer cannot b e utilized Ho w ev er b oth
MaxB and F unB con tin ue reducing the latency time b y utilizing the buers MaxB also lev els o at M axM em b ecause the rst matc hin terv al can b e found with smaller v alue of B Hence
the extra amountof B is not b enecial Ho w ev er since the v alue of B for F unB is a function of the
a v ailable buers it will reac h the lev el o p oin t later than MaxB Note that the eectiv e searc h
space of F unB is smaller than that of b oth MinB and MaxB MinB suers from rep eated failures
b ecause B is to o small to pro vide sucien t exibilit y for prefetc hing Therefore man yin v o cations
of atomic schedul e bef or e and atomic schedul e bef or e need to b e made On the other hand a
large B v alue requires a signican tn um b er of p ossible sliding p ositions F unB byha ving a exible
v alue of B alw a ys tries to pro vide eac h request with no more than its o wn share of buers Hence
W e p erformed the same set of exp erimen ts for a ligh tly loaded system and although the a v erage latency time w as
reduced signican tl y the basic observ ations w ere iden tical
0 100 200 300 400 500
500
1000
1500
2000
Size of Memory (in number of subobjects)
Avg. Latency (in Seconds)
MinB
MaxB
FunB
0 100 200 300 400 500
4500
5000
5500
6000
Size of Memory (in number of subobjects)
Avg. Latency (in Seconds)
MinB
FunB
MaxB
0 100 200 300 400 500
500
1000
1500
2000
Size of Memory (in number of subobjects)
Avg. Latency (in Seconds)
MinB
MaxB
FunB
B=0
a Ov erlap b Ov erlap c Ov erlap Figure ExhaustB Hea vy loaded system
it pro vides a smaller searc h space as compared to b oth MinB and MaxB Figures b and c sho w
the same observ ations when the amountof o v erlap is and sub ob jects resp ectiv ely W e rep eated the ab o v e set of exp erimen ts for the ExhaustB algorithm The basic observ ations
for MinB MaxB and F unB remained the same Figure a compares First Matc h with ExhaustB
for F unB and MaxB with o v erlap ob jects An in teresting phenomenon is that with Exhaust
B neither MaxB nor F unB lev els o Instead they con tin ue reducing the latency time as the
system memory gro ws This is b ecause ExhaustB do es not stop when it nds the rst matc hand
con tin ues searc hing for an a v ailable time in terv al earlier than that determined b y First matc h By
doing so it utilizes more of the a v ailable buers Note that this will result in a large searc h space
for ExhaustB T o reduce its searc h space w e used a heuristic Instead of sliding the second ob ject
do wn w ard one in terv al at a time the heuristic rst detects the time in terv al that w as resp onsible
for the una v ailabilit y of buers failedin terv al and then slides the second ob ject all the w a y to the
failedin terv al
ExhaustB in com bination with MaxB is a greedy algorithm that tries to allo cate all the
a v ailable buers to curren tly activ e requests When memory is a scarce resource this migh t force
other requests to starv e since displa y of comp osite ob jects with in ternal con ten tion requires at
least one extra buer In other w ords these requests require at least one buer to b e sc heduled
Figure a supp orts this fact and it sho ws that for small amoun t of system buers less than
sub ob jects F unB outp erformed MaxB Ev en MinB outp erformed MaxB when the amoun t
of system buer is less than sub ob jects An in teresting observ ation is that the b eha vior of
First Matc h and ExhaustB are iden tical for large o v erlaps compare Figure c with b This
is b ecause First Matc h diers from ExhaustB when the second ob ject slides up w ard F or large
0 100 200 300 400 500
500
1000
1500
2000
Size of Memory (in number of subobjects)
Avg. Latency (in Seconds)
MaxB, Exhaust-B
MaxB, First-match
FunB, Exhaust-B
FunB, First-match
0 100 200 300 400 500
1500
2000
2500
3000
Size of Memory (in number of subobjects)
Avg. Latency (in Seconds)
MinB
MaxB
FunB
0 100 200 300 400 500
1500
2000
2500
3000
Size of Memory (in number of subobjects)
Avg. Latency (in Seconds)
MinB
MaxB
FunB
a Ov erlap b First Matc h V ariable o v erlap c ExhaustB V ariable o v erlap
Fixed Ob j Size V ariable Ob j Size V ariable Ob j Size
Figure ExhaustB vs First Matc h Hea vy loaded system
o v erlaps the amountofup w ard sliding is v ery small usually in terv als in our exp erimen ts Hence
b oth algorithms end up nding the same in terv al for sc heduling the second ob ject
In order to sho w that prefetc hing reduces the latency ev en when there is no o v erlap sim ulating
the meet temp oral relationship prop osed b y All w e p erformed an exp erimen t when Ov erlap
see Figure c Since in this case there w as no in ternal con ten tion w ein v estigated B
no prefetc hing as w ell As sho wn in Figure c heuristics based on buered sliding ac hiev es a
signican tly lo w er a v erage latency time as compared to no prefetc hing
Finally to ensure that our tec hniques are indep enden t of the xed structure of comp osite
ob jects w e conducted another exp erimen t In this exp erimen t the atomic ob jects w ere no longer
equisized Instead w e randomly c hose the size of atomic ob jects to b e from to sub ob jects
Similarly the amoun tof o v erlap b et w een t w o ob jects X and Y w as a random n um ber b et w een
and Min siz e X size Y sub ob jects The observ ations remain v alid for b oth ExhaustB and
First Matc h The results are rep orted in Figures b and c
Conclusion
Wein v estigated the problem of con tin uously displa ying comp osite ob jects that are dynamically
sp ecied The abilit y to displaysuc h ob jects is lik ely to b e imp ortan t in manym ultimedia ap
plications The problem is c hallenging b ecause a comp osite ob ject consists of overlapp e d atomic
ob jects Therefore in order to supp ort con tin uous displa y of comp osite ob jects w e need to solv e
the problem arising due to con ten tion during retriev al of o v erlapp ed ob jects
In this pap er w e prop osed tec hniques based on simple sliding and buered sliding that help
supp ort con tin uous displayb y p artial prefetc hing of o v erlapping ob jects instead of the naiv e strategy
of prefetc hing o v erlapp ed ob jects en tirely The k ey idea in b oth these metho ds is to exploit the
strip ed la y out of the m ultimedia ob jects Our strategies apply for single displa y and m ulti displa y
en vironmen ts
Our study can b e extended in sev eral w a ys Wemen tion a few of those here First our
exp erimen tal study in the m ultidispla y scenario has fo cused on comp osite ob jects consisting of
only t w o atomic ob jects Iden tifying new approac hes to sc heduling when the comp osite ob ject has
more complex structure and comparing them exp erimen tally are imp ortan t problems Next w e
need to extend our approac h to the case where not all atomic ob jects in the comp osite ob ject ha v e
the same bandwidth
As a nal remark w e should note that the problem of displa ying a comp osite ob ject is a
complex problem since temp oral sync hronization is sub ject to sev eral other system parameters
that ha v e not b een considered in our retriev al mo del Th us our prop osed tec hnique should b e
view ed as a to ol for retrieving comp osite ob jects and not as a complete solution in itself
Ac kno wledgemen t W e thank Umesh Da y al for useful discussions and commen ts on the draft
References
All James F Allen Main taining Kno wledge ab out Temp oral In terv als Communic ations of the
A CM No v em ber BGMJ S Berson S Ghandeharizadeh R Mun tz and X Ju Staggered Striping in Multimedia Infor
mation Systems In Pr o c e e dings of the A CM SIGMOD International Confer enc e on Management
of Data Buf John F Ko egel Buford editor Multime dia SystemsA CM Press and AddisonW esley CGSon S Chaudh uri S Ghandeharizadeh and C Shahabi Retriev al of Comp osite Multimedia Ob jects
T ec hnical rep ort HP Lab Tec hnical Rep ort In preparation
CL HJ Chen and T Little Ph ysical Storage Organizations for TimeDep enden t Multimedia Data
In Pr o c e e dings of the F oundations of Data Or ganization and A lgorithms F ODO Confer enc e Octob er GKS S Ghandeharizadeh S H Kim and C Shahabi On Conguring a Single Disk Con tin uous
Media Serv er In Pr o c e edingsofthe A CM SIGMETRICSPERF ORMANCEMa y
GRA Q S Ghandeharizadeh L Ramos Z Asad and W Qureshi Ob ject Placemen tinP arallel Hy
p ermedia Systems In Pr o c e e dings of the International Confer enceon V ery L ar ge Datab ases Has B Hask ell In ternational standards activities in image data compression In Pr o c e e dings of
Scientic Data Compr ession Workshop pages NASA conference Pub NASA
Oce of Managemen t Scien tic and tec hnical information division
LG Thomas DC Little and Arif Ghafo or Spatiotemp oral comp osition of distributed m ultim edia
ob jects for v alueadded net w orks IEEE Computer pages No v em ber Nat K Natara jan Video Serv ers TakeRoot IEEE Sp e ctrum April P at D P atterson Massiv e Parallelism and Massiv e Storage Trends and Predictions for to Keynote Sp eak er In Se c ond International Confer enceon Par al lel and Distribute d Information
SystemsJan uary P ol VG P olimenis The Design of a File System that Supp orts Multimedia T ec hnical Rep ort
TR ICSI SG C Shahabi and S Ghandeharizadeh Con tin uous Displa y of Presen tations Sharing Clips T o
app e ar in A CM Multime dia Systems Journal SGM K Salem and H GarciaMolina Disk striping In Pr o c e e dings of International Confer enceon
Datab ase Engine eringF ebruary TPBG 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 T ri C T ristram Bottlenec k Busters Newme dia pages April
Linked assets
Computer Science Technical Report Archive
Conceptually similar
PDF
USC Computer Science Technical Reports, no. 622 (1995)
PDF
USC Computer Science Technical Reports, no. 600 (1995)
PDF
USC Computer Science Technical Reports, no. 578 (1994)
PDF
USC Computer Science Technical Reports, no. 587 (1994)
PDF
USC Computer Science Technical Reports, no. 650 (1997)
PDF
USC Computer Science Technical Reports, no. 766 (2002)
PDF
USC Computer Science Technical Reports, no. 610 (1995)
PDF
USC Computer Science Technical Reports, no. 584 (1994)
PDF
USC Computer Science Technical Reports, no. 592 (1994)
PDF
USC Computer Science Technical Reports, no. 748 (2001)
PDF
USC Computer Science Technical Reports, no. 694 (1999)
PDF
USC Computer Science Technical Reports, no. 602 (1995)
PDF
USC Computer Science Technical Reports, no. 601 (1995)
PDF
USC Computer Science Technical Reports, no. 615 (1995)
PDF
USC Computer Science Technical Reports, no. 590 (1994)
PDF
USC Computer Science Technical Reports, no. 865 (2005)
PDF
USC Computer Science Technical Reports, no. 733 (2000)
PDF
USC Computer Science Technical Reports, no. 943 (2014)
PDF
USC Computer Science Technical Reports, no. 612 (1995)
PDF
USC Computer Science Technical Reports, no. 619 (1995)
Description
Surajit Chaudhuri, Shahram Ghandeharizadeh, and Cyrus Shahabi. "Avoiding retrieval contention for composite multimedia objects." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 618 (1995).
Asset Metadata
Creator
Chaudhuri, Surajit
(author),
Ghandeharizadeh, Shahram
(author),
Shahabi, Cyrus
(author)
Core Title
USC Computer Science Technical Reports, no. 618 (1995)
Alternative Title
Avoiding retrieval contention for composite multimedia objects (
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
21 pages
(extent),
technical reports
(aat)
Language
English
Unique identifier
UC16271195
Identifier
95-618 Avoiding Retrieval Contention for Composite Multimedia Objects (filename)
Legacy Identifier
usc-cstr-95-618
Format
21 pages (extent),technical reports (aat)
Rights
Department of Computer Science (University of Southern California) and the author(s).
Internet Media Type
application/pdf
Copyright
In copyright - Non-commercial use permitted (https://rightsstatements.org/vocab/InC-NC/1.0/
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)
Copyright
In copyright - Non-commercial use permitted (https://rightsstatements.org/vocab/InC-NC/1.0/