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. 630 (1996)
(USC DC Other)
USC Computer Science Technical Reports, no. 630 (1996)
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
Co ordinated Displa y of Structured Presen tations Using a
MultiDisk Arc hitecture
Martha L EscobarMolano and Shahram Ghandeharizadeh
Departmen t of Computer Science
Univ ersit y of Southern California
Los Angeles California Jan uary Abstract
With the structured approac h to represen ting video clips a presen tation consists of a collection of
bac kground ob jects and actors D represen tations constrained using spatial and temp oral constructs
along with rendering features eg shading audiences view p oin t While the spatial constrain ts dene
the p osition of displa y ed ob jects on the screen the temp oral constrain ts describ e when the ob jects are
rendered As compared with an alternativ e approac h termed streambased that conceptualizes a video
clip as a sequence of frames the structured approac h pro vides for b oth reusabilit y of ob jects in other
presen tations and eectiv e query pro cessing tec hniques for retriev al of relev antdata The displayof a
structured presen tation is termed c o or dinate d when the displa y of its ob jects resp ects its presp ecied
temp oral and spatial constrain ts Otherwise the displa ymigh t suer from failures that translate in to
meaningless scenarios F or example a c hase scene b et w een a dinosaur and a jeep lled with screaming
c haracters b ecomes meaningless if the system fails to render the dinosaur when displa ying the scene
Assuming a m ultidisk hardw are platform congured with a xed amoun t of memory this study
describ es a taxonom y of resource sc heduling tec hniques that guaran tee a co ordinated displa y of structured
presen tations A subset of the prop osed tec hniques manipulates the placemen t of data across the a v ailable
disks W e emplo y a sim ulation study to quan tify the tradeo asso ciated with the alternativ etec hniques
The results are compared with a theoretical minim um demonstrating that one of the prop osed tec hniques
pro vides a p erformance almost iden tical to this minim um
This researchw as supp orted in part b y NSF gran ts IRI IRI NYI a w ard and CD A and a
unrestricted cashequipmen t gift from HewlettP ac k ard
Figure System Arc hitecture
In tro duction
One ma y representvideo using t w o alternativeapproac hes streambased and structured Gha With the
streambased approac h a video clip consists of a sequence of frames that are displa y ed at a presp ecied
rate eg frames p er second to fo ol the h uman p erception to observ e motion With the structured ap
proachEM a presen tation is represen ted as a collection of ob jects eg D represen tations of a dinosaur
with spatial and temp oral constrain ts eg p ositions of the dinosaur and the time of their app earances along
with their rendering features eg lightin tensit y view p oin t Presen tly the structured approac h is used
to pro duce animated sequences F or example T o y Story Bat and Reb o ot Ber are animations
generated using the structured approac h
The structured approac h pro vides for b oth reusabilit y of information and dev elopmen t of eectiv e query
pro cessing tec hniques EMG They enable a user to extract a c haracter eg a dinosaur a motion path
eg the tra jectory and timing of the dinosaurs motion from one presen tation and reuse it in another In
addition one can devise algorithms to supp ort pro cessing queries that reason ab out the temp oral and spatial
information of a structured presen tation T o illustrate consider the animation The Lion King Assuming it
w as represen ted using the structured approac h to retriev e the scene where Sim ba nds his father Mufasa
dead a user can p ose the follo wing query select scenes that con tain b oth Sim ba and Mufasa suc h that
Mufasa is static while Sim ba is mo ving The system lo cates the relev antdata b y analyzing the temp oral and
spatial constrain ts imp osed on the D represen tations of Mufasa and Sim ba
With structured presen tation the database con tains D represen tations of dieren t ob jects spatial and
temp oral constrain ts and rendering features The system renders a presen tation in real time to a computer
screen b y retrieving and displa ying ob jects at the appropriate time and lo cation on the screen satisfying
the temp oral and spatial constructs describ ed b y the author This database migh t b e shared byav ariet y
of users who author presen tations eg pro ducers animation artists etc Once a structured presen tation
is dened its author ma y w an t to displa y it in order to rene the presen tation This pro cess is almost
iden tical to a programmer authoring a program and desiring to execute it in order to ev aluate and rene
its functionalit y A c hallenging task for the system is to ensure a co ordinated displa y where the system
satises the authors presp ecied temp oral and spatial constrain ts imp osed on the presen tation of ob jects If
these constrain ts are not satised then the presen tation migh t suer from errors T o illustrate consider the
sequence of p ostures p
p
n
that pro vide the illusion that a dinosaur is w alking These p ostures mightbe
a collection of p ersisten t delta up dates on the dinosaurs D represen tation c hanging its facial expressions
morphing and mo ving its b o dy along a curv e as a function of time TTa TTb Lev If the system
fails to render the original D ob ject the deltas will yield a partially missing c haracter with a c hanging face
and mo ving b o dy parts
One approachto resolv e the p ossibilit y of errors is to render an en tire structured presen tation memory
residen t prior to initiating its displa y This approac h suers from the follo wing limitations First it w astes
memory b ecause it requires ob jects that are displa y ed at the end of the presen tation to b e rendered memory
residen t for the en tire duration of the displa y Second it fails to displa y those presen tations whose total ob ject
sizes exceed the amoun t of a v ailable memory some systems use sw ap memory and emplo y the op erating
system to sw ap pages in and out of memory Third it incurs a high startup latency
b ecause the displa y
is dela y ed un til all referenced ob jects are rendered memory residen t This study describ es a taxonomyof
sc heduling tec hniques that o v erlap the displa y of ob jects with their retriev al from disk in to memory These
tec hniques consume a sc hedule that dictates the displayof ob jects and construct a resource sc hedule suc h
that the incurred latency and the amoun t of memory required for a displa y are minimized and the
resource sc hedule satises the temp oral constrain ts imp osed on the presen tation Wequan tify the tradeo
asso ciated with these tec hniques b y measuring their memory disk bandwidth and disk space requiremen ts
Our target hardw are platform Figure is a m ultidisk arc hitecture consisting of D disks Eac h disk
can p erform indep enden t readswrites The unit of transfer from eac h disk is xsized and termed a page
or blo c k The D disk driv es ma y retriev e D pages in to D dieren t memory frames at the same time An
ob ject x is rst partitioned in to k d
siz e of x
siz e of a pag e
e disk pages Subsequen tly these pages are assigned to
the D disks This assignmentmigh t partially decluster ob ject x when k D An alternativew ould ha v e
b een to decluster ob ject x across all the D disks This w ould ha vebeen w asteful of b oth memory and disk
The time elapsed from the arriv al of a request to the onset of its displa y
bandwidth sp ecially when ob ject x is smaller than a disk page This is b ecause ev ery time x is referenced
the system w ould b e forced to read D disk pages with
D
of these pages con taining the relev an t data
and allo cate D memory frames instead of one
A limitation of data placemen t that facilitates partial declustering of ob jects is as follo ws The reference
pattern of asc hedule to pages migh t be suc h that a subset of disks migh t b ecome temp orary b ottlenec ks
during the displa y These temp orary b ottlenec ks w ould result in system errors when an ob ject cannot b e
rendered memory residen t in supp ort of the displa y These b ottlenec ks can b e detected in adv ance b ecause
when the displa y is activ ated both the sc hedule of page references and the placemen t of data are pre
sp ecied The system ma y resolv e these b ottlenec ks in sev eral w a ys to eliminate these errors This study
describ es the taxonom y of p ossible tec hniques Wequan tify the tradeo asso ciated with these tec hniques
using a sim ulation study Our results indicate that those tec hniques that mo dify the placementofdatacan
signican tly reduce the amoun t of memory required b y a displa y as compared to those that ha v e no impact
on placemen t of data
The rest of this pap er is organized as follo ws Section pro vides further details of a co ordinated displa y
and a taxonomyof sc heduling tec hniques that enable a m ultidisk hardw are platform to supp ort this func
tionalit y These tec hniques striv e to utilize all a v ailable disk bandwidth to minimize the latency and the
memory requiremen t As mec hanisms to reduce latency and memory requiremen t they either cac he disk
pages referenced more than once during a displa y replicatemigrate disk pages that migh t cause a b ottlenec k
during the displa y or b oth Section quan ties tradeo asso ciated with these tec hniques using a sim ulation
study Brief conclusions are pro vided in Section Related W ork
A n um ber of studies ha v e in v estigated tec hniques to ensure a con tin uous displa y of streambased presen
tations AH R VR TPBG GR R W BGMJ GLM MKT DS These studies concep
tualize a presen tation as a le that is read sequen tially at a presp ecied rate Alternativ ely a structured
presen tation consists of a collection of ob jects that are referenced and displa y ed based on a presp ecied
sc hedule A single ob ject migh t b e referenced at dieren t times byasc hedule Moreo v er it migh t b e shared
bysev eral presen tations These conceptual dierences render all these studies irrelev an t
In EMGIng w e describ ed a resource sc heduler that realizes a co ordinated displa y for a structured
presen tation assuming a single disk system congured with a xed amoun t of memory This pap er ex
tends EMGIng b y fo cusing on a m ultidisk hardw are platform The placemen t of ob jects across the disks
impacts both the amoun t of memory required to supp ort a displa y and its incurred latency W e prop ose
no v el sc heduling tec hniques that manipulate the placemen t of data for the target application and hardw are
platform The obtained results demonstrate that the sc heduling tec hnique of EMGIng is not appropriate
for a m ultidisk arc hitecture b ecause it is inferior as compared with the other prop osed tec hniques
An um b er of studies ha v e analyzed tec hniques to sc hedule tasks with realtime deadlines HL CRS RS These studies sc hedule the CPU with the ob jectiveto minimize the total cost to the system the
n um b er of tasks that miss their realtime deadline and their asso ciated cost This study is no v el b ecause
w e fo cus on IO sc heduling data placemen t and cac hing that are k ey pro cessing comp onen ts of a data
in tensiv e application There are no deadlines on when a displa y should start Instead once a displa y
starts the computed resource sc hedule should supp ort the constrain ts dened on when ob jects should be
rendered memory residen t in preparation for their displa y In addition a presen tation should emplo y the
disks con taining relev an t data with CPU sc heduling literature a task ma y access an y CPU as there is no
dep endency on data
Co ordinated Displa y
When a user requests the displa y of a structured presen tation the system has adv anced kno wledge of the
iden tit y of ob jects that should b e memory residen t at sp ecic times to supp ort a co ordinated displa y This
sc hedule is termed a display sche dule Without loss of generalit y and in order to simplify the discussion w e
partition time in to xsized units termed time intervals The duration of eac h time in terv al is t A collection
of ob jects represen ted as a set of disk pages are displa y ed during eachtime in terv al The b eginning of a
time in terv al i is termed time instant i Figure W e representa displaysc hedule as a sequence of time
instan ts Eac h time instan t i is tagged with a collection of disk pages termed P
i
displa y ed during time
in terv al i W e assume that the system can supp ort a co ordinated displa y if the set P
i
of pages displa y ed at
in terv al i is memory residentat i i m Our prop osed sc heduling tec hniques o v erlap the displa y and the retriev al of disk pages and manipulate
the placemen t of data on disks to reduce b oth the observ ed startup latency and the required amoun t of
memory Ideally the collection of pages that constitute P
i
should be retriev ed in to memory during time
in terv al i This w ould minimize the amountof required memory Ho w ev er this ideal situation migh t
be infeasible at times b ecause the pages that constitute P
i
migh t be unev enly disp ersed across the disks
exhausting the bandwidth of one or more disks while other disks are idle suc h that they fail to retriev e the
set P
i
during a time in terv al Note that in this scenario the total bandwidth of the disks is sucien t the
primary limitation is the placementof datain com bination with the displaysc hedule that results in formation
of b ottlenec k disks The system ma y pursue t w o alternativ e solutions to a v oid b ottlenec ks retriev esome
pages of P
i
during earlier time in terv als i i etc these pages are termed pr efetche d pages or
T erm Denition
m Num ber of time in terv als in the video
min pg es Minim um n um b er of pages in a in terv al
max pg es Maxim um n um b er of pages in a in terv al
sty pes Num b er of shot durations
n pag es Num b er of pages in the database
t Duration of a time in terv al
Time In terv al i P erio d i i Instan t i The b eginning of time in terv al i
S i P ages in memory at instan t i
P i P ages con taining ob jects displa y ed during in terv al i
F i P ages retriev ed from disk on to memory during time in terv al i
F
d
i
P ages retriev ed from disk d on to memory during time in terv al i
K i P ages sw app ed out during time in terv al i
R
a
j
Read page a from disk j
U
d
i
P ages written to disk d during time in terv al i
W
a
j
W rite page a to disk j
D Disk driv es in the system
B Maxim um n um b er of pages read b y a driv e during a time in terv al
C Num b er of memory frames
T able List of terms used in this pap er and their denitions
manipulate the placemen t of data prior to time in terv al i so that P
i
is more ev enly distributed across
disks This study describ es a resource sc heduler that consumes a displaysc hedule and pro duces as output
asc hedule of page retriev als from disks to memory and page writes to c hange the data placemen t
Similar to a displaysc hedule a resource sc hedule is represen ted as a sequence of time instan ts The duration
of a time in terv al for the resource sc hedule is iden tical to that of the displaysc hedule Asso ciated with time
instan t i is a collection of pages retriev ed from eachofthe D disks during time in terv al i denoted F
d
i
a collection of pages written to eachofthe D disks during time in terv al i to c hange the placemen t of data
denoted U
d
i
a collection of pages sw app ed out
of memory to accommo date these retriev als denoted as
K
i
The set of pages retriev ed and sw app ed out during a time in terv al are disjoin t Otherwise the system
w ould b e p erforming w asteful w ork
The n um b er of pages that can b e retriev ed from a disk during a time in terv al is b ounded denoted as B b ecause eac h disk has a xed bandwidth the duration of a time in terv al is xed and eac h page is xsized
When computing the resource sc hedule the sc heduler ensures that the n um ber of pages retriev ed from a
disk during time instan t i do es not exceed B Otherwise the displaymigh t suer from errors The state
of memory ie pages o ccup ying memory frames at eac h instan t i denoted as S
i
is dened based on the
curren t set of pages that o ccup y memory those sw app ed out of memory those ushed to disks and those
retriev ed from dieren t disks Giv en a system with D disks the state of memory at eac h instan t is dened
Notice that sw ap outs do not require disk bandwidth
Time
01 i i+1 m
Request Display
Arrives Starts
Display
Ends
Time
Interval i
S0 Si
Sm
Time
Intervals
-p
S-p
Time
Instant i
Figure Time in terv al and time instan t
R
1
a
Display starts Display ends
Display
-6 -5 -4 -3 -2 -1 0 1 2 3 4 5
Time
a
c
d
b
1
e
1
f
1
b
2
b
3
e
2
f
2
Pages
in
Memory
Request Arrives
R
1
c
R
1
d
R
1
b1
R
1
e1
R
1
f1
R
1
b2
R
1
b3
R
1
e2
R
1
f2
objects in
interval 0
Figure A resource sc hedule using memory
as
S
i
S
i K
i U
i U
D i F
i F
D i The n um b er of pages in S
i
should b e lo w er than or equal to the C frames that constitute the memory T o illustrate these concepts assume a displa ysc hedule that consists of v etimein terv als P
fa b
cd g P
fa b
cd g P
fa b
cd g P
fa e
f
g and P
fa e
f
g Assume that the system consists of
four disks D eac h with sucien t bandwidth to retriev e one disk page during a time in terv al B
Assuming that all the referenced pages reside on disk one Figure sho ws a resource sc hedule that facilitates
a co ordinated displa y R
a
denotes that disk page a is read from disk one In this gure a negativ e time
instan t corresp onds to page retriev als p erformed prior to start of the displa y A page migh t either be
-6 -5 -4 -3 -2 -1 012345
Display Starts Display Ends
R
1
e1
c
d
b
1
e
1
f
1
b
2
b
3
Display
objects in
Time
Pages
in
Memory
R
1
f2
R
1
a
R
1
c
R
1
d
R
1
b1
R
1
b2
R
1
b3
R
1
f1
R
1
e2
interval 0
W
2
e1
W
3
f2
f
2
a
R
2
e1
R
3
f2
e
1
f
2
e
2
Request Arrives
Figure A resource sc hedule that manipulates placemen t of data
retriev ed during the time in terv al prior to its displa y eg b
or prefetc hed at an earlier time in terv al eg
a Prefetc hing increases the memory requiremen ts of the system F or example sev en frames of memory
are allo cated at instan t one a c d b
e
f
b
while the displa ysc hedule dictates that only v e should b e
allo cated a b
b
cd The t w o prefetc hed pages e
f
increase the memory requiremen ts of the system
This example illustrates that an un balanced sc hedule of references to disks migh t result in formation of
b ottlenec k disks that requires the system to prefetc h pages while other disks remain idle In our example
while the bandwidth of four disks could accommo date the retriev al of four pages the system w as forced to
prefetc h pages b ecause they all resided on disk one The sc heduler ma y manipulate the placementofdata
b y utilizing the idle disk bandwidth to construct replicas of ob jects to minimize the n um b er of prefetc hed
pages Figure sho ws one suchsc hedule With this sc hedule the system reads page e
from disk one during
time in terv al and ushes it to disk t w o denoted as W
e during time in terv al This allo ws the system
to free the memory frame o ccupied b y e
at time instan t and utilize disk n um ber t w o to retriev e e
during
time in terv al t w o to satisfy the displaysc hedule With this sc hedule only memory frames are required
at instan t one for pages a b
b
c and d A disadv an tage of this approachis that it w astes b oth disk
bandwidth and disk space b y constructing an additional copyof e
and f
Once the displa y is complete
the system ma y either allo w b oth copies of e
and f
to con tin ue to exist main tain one copyofeac h
e
and f
b y deleting the other copies or main tain one copyof e
f
and b oth copies of f
e
In
the rst case a subsequen t displa y of the sc hedule w ould incur a four time in terv al startup latency b ecause
Coordinated
Resource
Memory-Based
Replication-Based
Migration-Based
Transient
Persistent
Transient
Persistent
Single Copy
Multiple Copies
Schedule
Figure T axonom y of p ossible approac hes to resource sc heduling
there is not need to prefetc h e
and f
Figure sho ws a taxonom y of the p ossible tec hniques They are categorized in to memorybased
replicationbased and migrationbased Memorybased emplo ys prefetc hing to supp ort a co ordinated dis
pla y It has no impact on the placemen t of data across the a v ailable disks Example of Figure is an
illustration of this approac h Replicationbased detects b ottlenec ks and iden ties those pages that cause the
b ottlenec ks It resolv es b ottlenec ks b y constructing copies of pages during earlier time in terv als Example
of Figure is an illustration of this tec hnique With a transien t replication the system deletes the newly
constructed replicas at the end of a displa y to restore the placemen t of data prior to the displa y In the
example of Figure transien t replication w ould delete copyof e
from disk and cop yof f
from disk at
the end of the displa y With p ersisten t replication the displa y of a presen tation migh t pro duce a new organization of data This
tec hnique is categorized in to singlecop yand m ultiple copies With singlecop y p ersisten t replication disk
space is a scarce resource and the system main tains only one copyof the replicated ob jects at the end of
displa y In the example of Figure this tec hnique migh t delete copies of e
and f
from disk n um ber one One approac h to decide whichcopytomain tain is as follo ws The system monitors the n um b er of references
to dieren t copies during the displa y and deletes the least frequen tly referenced copies With m ultiple copies
p ersisten t replication the system allo ws the additional replicas of a page to con tin ue to reside on disk The
exp ectation with b oth singlecop yand m ultiple copies p ersisten t replication is that the new placemen t with
p ossibly additional copies minimizes b oth the incurred startup latency and the amoun t of required memory In the example of Figure with new copies of e
and f
on disk t w o and three resp ectiv ely the subsequen t
displayofthe sc hedule incurs a four time in terv al startup latency and requires memory frames
With migration when the system migrates e
to disk t woduringin terv al it deletes the original cop y
of e
from disk one This is useful when disk space is scarce and cannot accommo date additional replicas
during the displa y With p ersisten t migration the new assignmen tof e
p ersists at the end of displa y With
transien t migration the system remem b ers the original placemen t of data across the disks and restore it at
the end of the displa y This w ould require additional disk op erations at the end of a displa y p oten tially
dela ying the displa y of other presen tations The fo cus of this study is on resource sc heduling tec hniques
that utilize memory and replication to minimize latency and memory requiremen ts Resource sc heduling
tec hniques with migration are not included in this study F or a v ailabilitypurposes the system uses extra disk space con taining redundan t parit y information to
reco v er the original information when a disk fails PGK When writing a replica of a page the system
ma y a v oid up dating the parit y information b y allo cating a region of all disks to con tain replicated data
MK This region do es not ha v e to b e protected b ecause in case of disk failure the system ma y access the
original cop y protected using the parit y information The singlecop y p ersisten t replication m ust up date
the redundan t information asso ciated to the disk page whose lo cation c hanged at the end of the displa y W e
assume that this up date is p erformed once the displa y completes An alternativeis to sc hedule this up date
an y time after the page is replicated measure the cost of this up date The adv an tage of this approachis that
it migh t reduce the disk bandwidth cost if the up date is done while the page is still in memory otherwise
the system has to read the page In this studyw edonot in v estigate tec hniques that ensure a co ordinated
displa y in the presence of disk failure This topic constitutes our future researc h direction Section Summary
T o summarize in a formal manner a displa y sc hedule consists of m time instan ts eac h mark ed with P
i
pages that are displa y ed during time in terv al i The sc heduler consumes the displaysc hedule and constructs
a resource sc hedule The resource sc hedule consists of p m time in terv als m of these o v erlap with the
displa y p of them either prefetc h pages in to memory or mo dify the placemen t of data across the disks
in preparation for the displa y The p time in terv als are n um b ered from to p in order to sho w that their
retriev als are p erformed prior to initiating the displa y In essence p denotes the incurred latency Asso ciated
with eac h time in terv al i are
a collection of pages retriev ed from eachof the D disks during time in terv al i denoted as F
i
F
D i
a collection of pages written to eachofthe D disks during time in terv al idenotedas U
i
U
D i
a collection of pages sw app ed out of memory to accommo date these retriev als denoted as K
i
The resource sc hedule is v alid and supp orts a co ordinated displa y if it satises the follo wing three conditions
i Once the displa y starts the set of pages in memory at eac h instan t i is a subset of those required b y
the displaysc hedule F or eac h i m P
i
S
i
and P
i
S
i ii The n um b er of pages retriev edwritten to a disk during a time in terv al i do es not exceed B F or eac h
i p m and eac h d D jF
d
i
j jU
d
i
j B iii The n um ber of memory residen t pages at eac h time instan t i is lo w er than the n um ber of a v ailable
memory frames F or eac h i p m jS
i
j C The sc heduler mayemplo y memory and page replication to pro duce a resource sc hedule When the sc heduler
emplo ys only memory its resulting resource sc hedule has no impact on the placemen t of data across the disks
Ho w ev er with both replication and migration the resulting resource sc hedule ma y impact the placemen t
of data The system migh t b e designed suc h that this impact is either transien t or p ersisten t relativeto a
displa y The computation of a resource sc hedule that supp orts the co ordinated displa y of a structured presen tation
and minimizes latency is an NPhard problem EMG This study describ es our prop osed heuristics to
compute resource sc hedules that manipulate the placemen t of data and striv e for minim um startup latency
and memory requiremen ts In Section w e describ e a tec hnique that emplo ys only mainmemory to
supp ort a co ordinated displa y Subsequen tly Section describ es a heuristic for deciding what pages to
replicate the disk driv es that should con tain these replicas and when should replication b e p erformed
MemoryBased Resource Sc hedule
This approac h emplo ys memory to prefetc h pages in order to resolv e b ottlenec ks It has no impact on the
placemen t of data across the disks Giv en a displaysc hedule and the curren t placemen t of data across the D
disks this tec hnique extracts the displaysc hedule for eac h disk based on the pages that reside on that disk
It in v ok es the greedy sc heduler of EMGIng using the displa ysc hedule of eac h disk to compute a retriev al
sc hedule
for that disk The union of these D retriev al sc hedules yield a nal retriev al sc hedule for the
displa y The disk with the longest startup latency p determines the o v erall system latency F or the giv en
placemen t of data this retriev al sc hedule minimizes the amoun t of memory required b ecause the greedy
sc heduler minimizes the memory requirementateac h instan t i for a single disk EMGIng By minimizing
the n um b er of pages that constitute S
the greedy sc heduler minimizes the incurred latency This tec hnique
fails to displa y an ob ject when the n um b er of pages that constitute S
i
exceeds C In this case the system
m ust emplo y the replication tec hnique of Section In the follo wing w epro vide an o v erview of the greedy
algorithm W e refer the in terested reader to EMGIng for the optimalit y pro of
A resource sc hedule that consists of retriev als and sw ap outs ie for eac h i U
i
U
D i
Greedy Sc heduler hP
P
m iB S
m
P
m for i m to do
if i then RS
i
P
i
else RS
i
P
i
P
i Let N
i
b e the sequence of pages in S
i RS
i
sorted b y last i
if jN
i
j B then F
i
N
i
else F
i
fthe rst B pages in N
i
g
S
i
RS
i
N
i
F
i
end for
Figure The greedy algorithm
The greedy algorithm Figure consumes a displa ysc hedule fP
P
m g and the n um b er of pages
the target disk can retriev e during a time in terv al B as input to pro duce a sequence
fF
F
m g
of page retriev als for that disk and a sequence fS
S
m
g of memory states It tra v erses the displa y
sc hedule bac kw ards to determine the disk pages to retriev e during eac h time in terv al It assumes that only
pages referenced during the nal time in terv al are in memory at the end of the display S
m
P
m T o
minimize prefetc hing it striv es to retriev e during in terv al i all pages in memory at instan t i S
i that are not referenced during in terv als i and i P
i
P
i If all pages in S
i
P
i
P
i cannot
b e retriev ed during in terv al i then it pushes the retriev al of some pages to an earlier time in terv al P ages
in S
i
P
i
P
i with the lo w est most recen t instantat whic h they are referenced are retriev ed during
in terv al i the others retriev als are pushed to earlier time in terv als F or eachtime in terv al i and page aw e
dene last a i to b e the most recen t instan t j j iatwhic hpage a w as referenced More formally let
last a and for eac h i dene
last a i i if a P
i last a i otherwise
ReplicationBased Resource Sc hedule
This tec hnique solv es the b ottlenec ks created byun balanced data placemen t using replication It rst detects
the b ottlenec ks and then sc hedule replications to resolv e them It consists of three stages During the rst
stage it emplo ys the greedy metho d It tra v erses the displaysc hedule in rev erse order starting with time
instan t m constructing a retriev al sc hedule for eac h disk When it starts it constructs D retriev al
sc hedules one for eac h disk for instan t m Disk d
i
is a b ottlenec k if the displaysc hedule requires more
The retriev als b efore the displa y are not included b ecause they are determined b y the state of memory at the b eginning of
the display S
than B pages from this disk during time in terv al m jN
m j B pages in N
m see the co de of greedy
for the denition of N
i
should b e replicated The disks that can serv e as target candidates to hold these
pages are those with a v ailable disk bandwidth during in terv al m In the presence of candidate disks the
follo wing record is app ended to a list L of replications fpageid time in terv al m b ottlenec k disk d
i
a
list of target candidate disksg If no target candidate disks exist then the system is forced to prefetc hthese
jN
m j B This pro cess is rep eated for eachof the remaining in terv als A t the end of this pro cess the
sc heduler has a retriev al sc hedule for eac h disk free from replication requests a list L of replications
describing the pages to b e replicated their curren t lo cation and p ossible target disks and the time in terv al
prior to whic h the replication m ust b e done in order to resolv e the b ottlenec k Stage one is no w complete
The second stage selects a target from the candidate disks for eac h replication in L W e start with a
simple description of this stage Subsequen tly w e describ e optimizations that impro v e its p erformance
During stage t w o this heuristic computes a demand and supply v alue for eac h disk driv e While the demand
for a disk sp ecies ho w frequen tly that disk migh t b e used as the target for con taining the replica of a page
its supply sp ecies ho w frequen tly that disk mightbe a v ailable as the target for replication Assuming that
r tar g ets denotes the collections of disks that can serveas target for replicating a page and jr tar g ets j
denotes the n um b er of disks in that set the demand for disk d
i
is
demand d
i
replications r suchthat d i r tar g ets jr tar g ets j
The recipro cal w eigh ts the demand of a candidate disk based on the n um b er of alternativ es that the repli
cation has F or example a replication with one target candidate d
i
mak es the demand for d
i
higher than
if it w ould ha v e four target candidates Assuming that jF
d i
j
j denotes the n um b er of pages retriev ed from d
i
during in terv al j the supply of d
i
is estimated as
suppl y d
i
p i m B j F
d i
j
j A disk that is almost completely utilized during eac h time in terv al will ha vea lowsupplyv alue A disk that
participates as a p ossible candidate for man y page replications mightha v e a high demand v alue dep ending
on the n um b er of disks that comp ete with it to serv e as a p oten tial candidate Next replication computes
the demand to supply ratio of eac h disk
A disk with a high demand to supply ratio has the maxim um n um b er of constrain ts The bandwidth of
these disks should b e emplo y ed with care b ecause they could p oten tially b e used to resolvesev eral b ottlenec ks
The heuristic to accomplish this is as follo ws Replication pic ks the disk with the lo w est demand to supply
ratio least constrained disk termed d
free
It iden ties pages from list L that con tain d
free
as a p ossible
candidate to resolv e b ottlenec ks F rom this set of pages it c ho oses the page with the few est target candidate
disks Still sev eral pages ma y qualify b ecause their n um ber of target candidate disks is iden tical and this
is the smallest v alue F rom these it c ho oses to replicate the page on d
free
whose other candidate disks are
highly constrained ie ha v e the highest demand to supply ratio The supply v alue for d
free
is adjusted
and this pro cess is rep eated un til eac h disk is left with one candidate target disk
In the nal stage this heuristic extends the retriev al sc hedule with appropriate read and write op erations
to construct replicas A t this time eac h page to b e replicated has a source and target disk termed d
sour ce
and d
tar g et
resp ectiv ely Moreo v er this page m ust b e replicated prior to the time in terv al that the b ottlenec k
w as encoun tered termed i
bottleneck
This information is main tained for eac h requested replication in list
L see the rst stage W e prefer to p erform replication during the displa y b et w een time in terv al and
the one corresp onding to the b ottlenec k in order to minimize the incurred latency Moreo v er the duration
of replication should b e minimized in order to minimize the amoun t of memory required from the system
the n um ber of in terv als elapsed from when a replicated page is read from d
sour ce
to the time that this page
is written to d
tar g et
m ust b e minimized Based on these t w o ob jectiv es the system emplo ys the follo wing
heuristic to replicate a page It lo cates t w o adjacen t p ositiv e time in terv als j j prior to the o ccurrence
of the b ottlenec k j i
bottleneck
suc h that d
sour ce
and d
tar g et
ha v e a v ailable disk bandwidth during
in terv als j and j resp ectiv ely If this step is successful it extends the j and j in terv als of the retriev al
sc hedule to replicate the target page Otherwise it p erforms the replication as close as p ossible to i
bottleneck
in order to a v oid using a negativ e time in terv als The read and write of this page are p erformed as close as
p ossible to minimize the memory requiremen t
This v ersion of the replication pro cess is further optimized as follo ws First a page migh t app ear m ultiple
times in list L b ecause it participates in sev eral b ottlenec ks detected during stage one The curren t heuristic
replicates this page as man y times as necessary In fact it ma y replicate a page on the same disk driveat t w o
dieren tpoin ts in time In order to a v oid this p ossibilit yw emak e a second pass on the retriev al sc hedule
and eliminate suchw asteful op erations The heuristic can minimize the n um b er of replicas constructed for
a page b y computing the in tersection set of the candidate disks that can serv e as p oten tial targets for that
page Assuming that this in tersection set is not empt yb y replicating the page once on one of these disks
no additional replicas are required If this in tersection set is empt y then the heuristic w ould b e required to
compute m ultiple in tersections of the candidate lists that are not empt y A Hybrid of Memory and ReplicationBased Resource Sc hedule
With a h ybrid approac h the system uses either memory or replications to resolv e a b ottlenec k Hybrid
emplo ys replication as so on as the displayreac hes a presp ecied threshold on the amoun t of memory This
tec hnique is v ery similar to the heuristic of Section The only dierence is that during stage one it
emplo ys prefetc hing if the amoun t of utilized memory is less than or equal to the threshold Otherwise it
emplo ys replications When constructing the resource sc hedule h ybrid migh t toggle b et w een replication and
memorybased tec hnique m ultiple times dep ending on when the utilized memory reac hes a threshold
Ev aluation
W e quan tied the trade o asso ciated with memorybased replicationbased and h ybrid sc heduling tec h
niques using a sim ulation study The p erformance of eac h tec hnique dep ends on b oth the placementofdata
across the disks and the sc hedule of ob ject references As a y ardstic kto ev aluate the alternativ e tec hniques
w e compute a theoretical minim um on the amoun t of memory and latency required to supp ort a co ordinated
displa y This theoretical minim um conceptualizes D disks as a single disk that can supp ort the retriev al of
B D pages during eac h time in terv al Th us it renders the placemen t of data across the disks irrelev an t
It emplo ys the greedy algorithm of Section to compute b oth the minim um amoun t of required memory
and incurred latency This tec hnique is theoretical b ecause it migh t b e imp ossible to emplo y the aggregate
bandwidth of D disks with arbitrary sc hedules and data la y outs
This section starts with a description of our sim ulation mo del and exp erimen tal design Next w e analyze
the results obtained
Sim ulation mo del
The sim ulation mo del consists of four mo dules Figure a display sche dule gener ator that pro duces a
disk page reference pattern for a presen tation a gr e e dy sche duler that computes the theoretical minim um
a data plac ement gener ator that assigns the pages that constitute the database to the disk driv es and
a multidisk r esour c e sche duler that generates a resource sc hedule that satises the temp oral constrain ts
imp osed b y the displa y sc hedule This comp onen t implemen ts the memorybased replicationbased and
h ybrid tec hniques along with their p ersisten t and transien t mo des of op eration
The displaysc hedule generator assumes that a presen tation consists of a sequence of scenes Eachscene
consists of a collection of bac kground ob jects and a collection of motions A motion is represen ted as a
sequence of deltas to an ob ject the rst p osture where a delta is implemen ted either as a computation or
a page retriev al that describ es the required c hanges Once applied to an ob ject they pro vide the illusion of
amo v emen t The duration of eac h scene dur
scene
is selected randomly from a collection of p ossible v alues
These v alues w ere generated b y analyzing the duration of sev eral scenes for t w o dieren t animations Lion
King and Reb o ot The n um b er of pages that constitute the rst in terv al of a scene ie the pages that
Display Schedule
Generator
(P
0
, ... , P
m-1
)
Number of intervals
Number of pages
per interval
Shot Durations
Greedy Scheduler
Retrieval
Data Placement
Generator
({0, ..., n
pages
-1}
2
{0, 1, ...,D-1}
)
Number of
(n
pages
)
Multi-disk
Placement
ResourceScheduler
Display Schedule
(m)
Number of
(D)
(p
max *
D)
Agreggate bandwidth
Minimum
Requirements
Latency
Minimum
Memory
Schedule
Pages
Drives
Number of
Replicas
Placement
Retrieval
Requirements
Latency
Memory
Schedule
Figure Sim ulation Mo del
con tain bac kground ob jects and the rst p ostures of the motions are generated randomly using a v alue in
the range of
The n um b er of motions n
motions
is a fraction of the n um b er of pages referenced during
the rst in terv al This fraction is a randomly c hosen real v alue bet w een and to represen t scenes that
range from static bac kground to those that c hange rapidly Deltas can translate in to either disk retriev als
or computations The displaysc hedule generator selects a p ercen tage p of the total n um b er of deltas as
retriev als assuming a delta per motion per time in terv al is required It assumes that this p ercen tage is
distributed uniformly The n um ber of deltas that translate in to disk retriev als are computed as follo ws
p n
motions
dur
scene
Finally it assigns page iden tiers to all pages in the rst in terv al and all pages
corresp onding to deltas
The data placemen t mo dule assigns pages to D disks with the ob jectiv e to distribute them ev enly P erformance results
F or the purp oses of this ev aluation w e generated a single presen tation using our displaysc hedule gener
ator This presen tation is min utes long and references pages The total n um b er of pages in the
database is The displaysc hedule consists of time in terv als eac h
seconds long Figure This range w as based on the n um ber of c haracters app earing sim ultaneously in these t w o animations and assuming that
there is one ob ject p er page F or the ev aluation purp oses a large ob ject in the bac kground is equiv alentto sev eral small ob jects
in the bac kground Similarly a motion of a large ob ject is equiv alen t to sev eral small ob jects motions
One time in terv al is the duration of the displa y of frames in the streambased approac h assuming a frames p er second
displayrate
0 1000 2000 3000 4000 5000
Intervals
0
10
20
30
40
50
60
Number of Pages
Figure Num b er of pages referenced during eachtime in terv al b y Presen tation Resource Requiremen ts
T ec hnique Lat Memory Bandwidth Ov erhead
Max Avg Max Avg Band Mem Disk Sp
theoretical minim um NA NA NA
memorybased NA NA NA
rst displa y m ultiple copies second displa y replication third displa y based rst displa y single cop y second displa y fourth displa y eigh th displa y rst displa y
m ultiple copies second displa y
h ybrid rst displa y single cop y second displa y third displa y T able P erformance of alternativ e heuristics for Presen tations using disks
sho ws the n um b er of page references p er time in terv al for this presen tation With this presen tation most of
page references o ccurred during the rst few scenes of the mo vie W e analyzed an alternativ e presen tation
termed Presen tation that w as the rev erse of the one sho wn in Figure With Presen tation most of
page references o ccurred during its last few scenes
Our target system w as congured with a Kilob yte disk page size Its buer p o ol w as Megab ytes in
size partitioned in to disk page frames C The n um b er of disks in the system v aries from to D Eac h disk supp orts a m bps transfer rate millisecond seek times and millisecond
latency Th us a disk can read nine random pages during a time in terv al B In the follo wing w e
compare the alternativ e tec hniques memorybased replicationbased and h ybrid based on their startup
latency memory requiremen t disk bandwidth and disk space requiremen ts W e start with an analysis of a
system congured with disks Subsequen tlyw e describ e the observ ed trends as a function of the n um ber
of disks Finallyw e establish the tradeo asso ciated with alternativ e tec hniques
Resource Requiremen ts
T ec hnique Lat Memory Bandwidth Ov erhead
Max Avg Max Avg Band Mem Disk Sp
theoretical minim um NA NA NA
memorybased NA NA NA
rst displa y replication m ultiple copies second displa y based rst displa y single cop y second displa y T able P erformance of alternativ e heuristics for Presen tation using disks
0 1000 2000 3000 4000 5000
Intervals
0
10
20
30
40
50
60
Number of Pages
0 1000 2000 3000 4000 5000
Intervals
0
500
1000
1500
2000
2500
3000
Number of Pages
a Bandwidth requiremen t b Memory requiremen t
Figure Co ordinated displa y of Presen tation using memorybased with a disk system
Sev enDisk System
T ables and showthe results obtained for the t w o alternativ e presen tations using a sev endisk system
These tables quan tify the latency the amoun t of memory and disk bandwidth required to supp ort a co
ordinated displa y with alternativ e tec hniques In addition it rep orts on the o v erhead due to replication
p ercen tage of disk bandwidth disk space and memory used b y replications The theoretical minim um
and the memorybased sc hedules do not dep end on the systems memory C The theoretical minim um
determines the minim um amoun t of memory required to supp ort a co ordinated displa y with or without ma
nipulation of the data placemen t The memorybased sc hedule determines the minim um amoun t of memory
required to supp ort a co ordinated displa y without c hanging the data placemen t On the other hand the
replicationbased and h ybrid sc hedules dep end on the memory capacit y of the system C A system with
higher memory capacit y has more c hances to sc hedule the replications during the displa y than a system with
lo w er memory capacit y Replications main tain the replicated page memory residentbet w een the read and
write op erations All measuremen ts for replicationbased and h ybrid tec hniques quan tify the p erformance of
the tec hniques in either p ersistentmode The measuremen ts for transien t mo de are as the p ersisten t mo des
except for last column p ercen tage of extra disk space that is alw a ys for the transien tmode The memorybased tec hnique mightha v e excessiv e memory requiremen ts to supp ort a co ordinated dis
pla y Un balanced placemen t of pages referenced byapresen tation during time in terv al i increases the memory
requiremen t during previous in terv als j j i Similarlyun balanced placemen t during in terv al i in creases the memory requiremen t during previous in terv als j j i Therefore the increase of memory
requiremen t migh t b e cum ulativ e if consecutivein terv als result in formation of b ottlenec ks Figure a sho ws
the total n um b er of disk pages retriev ed from the disks b y the memorybased resource sc hedule Figure b
sho ws the n um ber of pages rendered memory residen t with this tec hnique Presen tations and require
memory pages th us a system with few er memory frames cannot supp ort a co ordinated displayofeither
presen tation
The startup latency observ ed b y both replicationbased and h ybrid tec hniques dep ends on the time
in terv als when the b ottlenec ks o ccur Presen tations whose reference pattern results in the formation of
b ottlenec ks at the b eginning of the displaymigh t incur a higher latency than those whose sc hedule results
in a b ottlenec k at the end of their displa y Presen tation has a lo w disk bandwidth requiremen t at the
b eginning of its displa y This enables the system to replicate pages during the displa y of this presen tation
resulting in a lo w er startup latency as compared with for presen tation with memorybased
If a system conguration can supp ort the co ordinated displa y of a structured presen tation then there is a
replicationbased resource sc hedule that supp orts the presen tation The replicationbased sc hedule tec hnique
resolv es all b ottlenec ks with replications and sc hedule the replications so that the memory requiremen t
do es not exceed the memory a v ailable to the system C Therefore if the minim um memory requiremen t
of a co ordinated displa y theoretical minim um
is lo w er than the systems memory C then there is a
replicationbased sc hedule for the presen tation The memory a v ailable to the system is greater than
the theoretical minim um for b oth Presen tations and enabling replicationbased to supp ort b oth
The replicationbased tec hnique requires and memory frames to displa y Presen tation and
resp ectiv ely Figures sho ws the n um ber of disk pages retriev ed and the n um ber of memory frames required per
time in terv al for the theoretical minim um If the sev en disks w ere replaced b y one disk whose bandwidth
w as equiv alen t to the aggregate bandwidth of sev en disks rendering the placemen t of data as irrelev an t
a co ordinated displa y of Presen tation could ha v e b een accommo dated with a maxim um of memory
frames and a startup latency equiv alen t to one time in terv al
In a system that allo ws m ultiple copies of a page the replicationbased tec hnique in p ersisten t mo de
appro ximates the theoretical minim um as the n um ber of presen tations increases In the absence of disk
space limitations the p erformance of this tec hnique will be iden tical to the theoretical minim um after a
nite n um ber of presen tations eg three consecutiv e displa ys of Presen tation This is b ecause the new
Assuming that disk space is not a limitation for replication
0 1000 2000 3000 4000 5000
Intervals
0
10
20
30
40
50
60
Number of Pages
0 1000 2000 3000 4000 5000
Intervals
0
20
40
60
80
100
Number of Pages
a Bandwidth requiremen t b Memory requiremen t
Figure Theoretical minim um for Presen tation in a system with disks
replicas eliminate some b ottlenec ks in subsequen t displa ys reducing b oth the memory and disk bandwidth
requiremen t of a displa y along with its incurred startup latency Figure sho ws the bandwidth and memory
requiremen t of replicationbased sc hedules for eac h of three consecutivedispla ys of Presen tation Note that
the maxim um amoun t of memory required for the rst displa y frames decreased b y a factor of six for
the second displa y of presen tation T able The last prole Figure e and f is almost iden tical to
that of the theoretical minim um compare with Figure enabling the sc heduler to harness the bandwidth
of D disks The resulting replicated data requires an additional of the disk space
With the single cop y replicationbased tec hnique in p ersisten t mo de the disk bandwidth o v erhead also
decreases as a function of subsequen t displa ys This tec hnique main tains the cop y of a page that w as most
frequen tly referenced b y the displa y and deletes the rest This c hange in data placemen t decreases the
n umberofbottlenec ks in subsequen t displa ys reducing the n um b er of page replications from for the
rst displa y to for the second displa y and for the eigh t displa y this is reected as a decrease in
the p ercen tage o v erhead in disk bandwidth requiremen t see T able Note that with this tec hnique b oth the
memory requiremen ts and the memory o v erhead due to replications in subsequen t displa ys migh t increase
T able The data placementc hanges induced bythistec hnique migh t cause a b ottlenec k at dieren t time
forcing the replications to b e sc heduled during a p erio d of scarce resources Moreo v er the heuristic tries to
minimize latency then it migh tsc hedule the replications so that the length of time a page remains memory
residen t prior to b e written is made longer This increases the memory requiremen ts of the sc hedule F or
example in T able the memory requiremen t increased from in the rst displa y replicationbased to
in the second displa y replicationbased singlecop y second displa y
The h ybrid tec hnique requires less disk bandwidth columns BandwidthAvg and Overhead Band
in T able and disk space than the replicationbased tec hnique Instead of solving all b ottlenec ks with
replications that increase disk bandwidth and space requiremen ts it prefetc hes pages un til a threshold
of the total memory for our exp erimen ts is reac hed The p ercen tage of extra disk space attributed
0 1000 2000 3000 4000 5000
Intervals
0
10
20
30
40
50
60
Number of Pages
0 1000 2000 3000 4000 5000
Intervals
0
100
200
300
400
500
600
700
Number of Pages
a Bandwidth requiremen t b Memory requiremen t
P ersisten t replication rst displa y
0 1000 2000 3000 4000 5000
Intervals
0
10
20
30
40
50
60
Number of Pages
0 1000 2000 3000 4000 5000
Intervals
0
20
40
60
80
100
120
Number of Pages
c Bandwidth requiremen t d Memory requiremen t
P ersisten t replication second displa y
0 1000 2000 3000 4000 5000
Intervals
0
10
20
30
40
50
60
Number of Pages
0 1000 2000 3000 4000 5000
Intervals
0
20
40
60
80
100
120
Number of Pages
e Bandwidth requiremen t f Memory requiremen t
P ersisten t replication third displa y
Figure Three consecutiv e displa ys of Presen tation in a system with disks
Resource Requiremen ts
T ec hnique Lat Memory Bandwidth Ov erhead
Max Avg Max Avg Band Mem Disk Sp
theoretical minim um NA NA NA
memorybased NA NA NA
rst displa y m ultiple copies second displa y replication third displa y based rst displa y single cop y second displa y third displa y T able Measuremen ts obtained for Presen tation using disks
Resource Requiremen ts
Driv es Lat Memory Bandwidth Ov erhead
Max Avg Max Avg Band Mem Disk Sp
T able Measuremen ts obtained b y memorybased and h ybrid tec hniques for Presen tation in a system
with and driv es Both tec hniques yielded iden tical measuremen ts
to replication is for the rst displa yof hybrid while for the rst displa y of replicationbased
Hybrid reac hes a stable state so oner than the replicationbased tec hnique The second displa y using the
h ybrid tec hnique do es not require an y replication Ov erhead in T able while the third displa y using the
replicationbased tec hnique still p erform replications
T rends
With few er than sev en disks none of the prop osed tec hniques can supp ort a co ordinated displa y of the
presen tation in either forw ard or rev erse mo de b ecause there is insucien t disk bandwidth and memory With more than disks the a v ailable disk bandwidth is abundan t rendering the dierence bet w een the
alternativetec hniques negligible T ables and An increase in the n um b er of disks has the follo wing eects decreases the memory requiremen ts of
the memorybased tec hnique columns Memory in T ables and decreases the disk bandwidth
requiremen ts of the replicationbased tec hnique columns Bandwidth and Ov erhead Band in T ables
and and migh t decrease the memory requiremen ts of the replicationbased tec hnique columns
Memory in T ables and The reason is that an increase in the n um b er of disks reduces the c hances
of a b ottlenec k The database is distributed ev enly across the disk driv es therefore as the n um b er of disks
increases the c hances of parallelism increases Then the n um b er of prefetc hings in memorybased sc hedules
decreases Moreo v er the n um b er of replications in replicationbased sc hedule decreases Ho w ev er for the
Resource Requiremen ts
T ec hnique Lat Memory Bandwidth Ov erhead
Max Avg Max Avg Band Mem Disk Sp
theoretical minim um NA NA NA
memorybased NA NA NA
replication m ultiple copies rst displa y based single cop y rst displa y T able Measuremen ts obtained for Presen tation using disks
Resource Requiremen ts
T ec hnique Lat Memory Bandwidth Ov erhead
Max Avg Max Avg Band Mem Disk Sp
theoretical minim um NA NA NA
memorybased NA NA NA
rst displa y replication m ultiple copies second displa y based rst displa y single cop y second displa y third displa y T able Measuremen ts obtained for Presen tation in a system with driv es
replicationbased sc hedules the memory requiremen t dep ends not only on the n um ber of prefetc hes but
also on the sc heduling of the replications The heuristic presen ted in this study striv es to reduce the startup
latency of a displa y This migh t increase the memory requiremen ts of the displa y A h ybrid of replication and memorybased beha v es the same as the memorybased resource sc hedule
if the memory requiremen t is belo w the threshold In our exp erimen ts the memory requiremen t of the
memorybased tec hnique for a system with at least disk driv es is lo w er than the threshold memory
frames Then this h ybrid tec hnique b eha v es lik e memorybased for a system with driv es Figures and
Summary
The replicationbased tec hnique in either transien t or p ersisten t mo de allo ws the system to render a co
ordinated displa y that could ha v e b een imp ossible to displa y with a memorybased sc hedule In a system
that allo ws m ultiple copies of a page the p ersisten t replicationbased tec hnique appro ximates the theoretical
minim um as the n um ber of displa ys increases The p ersisten t replicationbased sc hedule m ultiple copies
trades disk bandwidth and space for memory and latency In a system with disk space limitations the
p ersisten t replicationbased tec hnique single cop y is an alternativ e that migh t reduce the latency in subse
quen t displa ys The h ybrid tec hnique uses memory and replications to solv e the b ottlenec ks in a con trolled
fashion so that resources are utilized eectiv ely Eac h replication requires t w o IOs and a memory frame
bet w een the read and write op erations The h ybrid tec hnique detects b ottlenec ks and resolv es them b y
either prefetc hing or replicating data dep ending on whether the memory requiremen ts is excessiv e Memory
requiremen ts b elo w the threshold signals that solving a b ottlenec k with replications migh t imp ose unnec
essary demands for resources memory disk space and disk bandwidth therefore it resolv es b ottlenec ks
using memory On the other hand memory requiremen ts ab o v e the threshold signals that the b ottlenec ks
are demanding excessiv e memory and migh t increase the latency Therefore it w ould b e b etter to solv e the
b ottlenec ks with replications
Conclusions and F uture Researc h
In this studyw e presen ted a taxonomyof sc heduling tec hniques to supp ort a co ordinated displa y of a struc
tured presen tation These tec hniques are categorized in to memorybased replicationbased and migration
based The last t w o categories manipulate the placemen t of data across the a v ailable disks W e prop osed
three tec hniques that compute a memorybased a replicationbased and a h ybrid of memory and replication
based resource sc hedules The memorybased tec hnique emplo ys prefetc hing to resolv e b ottlenec ks The
replicationbased emplo ys idle disk bandwidth to resolv e b ottlenec ks The h ybrid tec hnique uses either
memory or replication dep ending on the a v ailabilit y of system resources when the b ottlenec k o ccurs
Weev aluated the prop osed tec hniques using a sim ulation study W e compared these tec hniques based on
their startup latency memory requiremen t disk bandwidth and disk space requiremen ts W e also compared
them with a theoretical minim um that ignores the placemen t of data and its impact on a co ordinated displa y a singledisk with the aggregate bandwidth of D disks The replicationbased tec hnique pro v ed to supp ort
co ordinated displa ys that the memorybased tec hnique is unable to supp ort b ecause they require excessiv e
memory In a system that emplo ys m ultiple copies the replicationbased tec hnique in p ersisten t mo de
appro ximates the theoretical minim um on memory requiremen t and latency as a function of the n um ber
of times a presen tation is displa y ed In a system with scarce disk space the single cop y replicationbased
tec hnique in p ersisten t mo de reduces the disk bandwidth o v erhead caused b y replications in subsequen t
displa ys The h ybrid tec hnique reduces the disk bandwidth and disk space requiremen t of the replications
b y resolving some of the b ottlenec ks with prefetc hing and the rest with replications Ev aluation of these
tec hniques based on traces from actual animations is curren tly in progress
In the future w e w ould lik e to address the follo wing issues in tegration of fault tolerance in to the
sc heduling tec hniques dev elopmen t and ev aluation of migrationbased tec hniques resource sc heduling
tec hniques for m ultiuser en vironmen ts where sev eral users displa y dieren t presen tations
References
AH D Anderson and G Homsy A con tin uous media IO serv er and its sync hronization IEEE
Computer pages ! Octob er Bat J Bates Stev e jobs to get executiv e pro ducer credit on disney animated lm L os A ngeles Times
Se ction D Octob er Ber S Bernstein Tec hnoArtists To oning Up L os A ngeles Times Se ction FNo 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 A CMSIGMOD pages ! CRS C Chen K Ramamritham and J Stank o vic Resource Reclaiming in Multipro cessor RealTime
Systems IEEE T r ansactions on Par al lel and Distribute d Systems ! April DS A Dan and D Sitaram An online video placemen t p olicy based on bandwidth to space ration
bsr In Pr o c e e dings of A CMSIGMOD pages ! EM M L EscobarMolano Managemen t of Resources to Supp ort Con tin uous Displa y of Structured
Video Ob jects T ec hnical Rep ort USC EMG M L EscobarMolano and S Ghandeharizadeh A framew ork for conceptualizing structured
video In First International Workshop on Multime dia Information Systems Arlington
Virginia
EMG M L EscobarMolano and S Ghandeharizadeh On the complexit y of resource sc heduling for
co ordinated displa y of structured presen tations T ec hnical rep ort USC EMGIng M L EscobarMolano S Ghandeharizadeh and D Ierardi An Optimal Resource Sc heduler for
Con tin uous Displa y of Structured Video Ob jects IEEE T r ansactions on Know le dge and Data
Engine ering Corr esp ondenceon R e c ent Developments F orthcoming
Gha S Ghandeharizadeh Streambased Versus Structured Video Ob jects Issues Solutions and
Challenges In S Ja jo dia and VS Subrahmanian editors Multime dia Datab ase Systems Issues
and R ese ar ch Dir e ctions Springer V erlag GLM L Golub c hik JCS Lui and R Mun tz Reducing io demand in videoon demand storage
serv ers In Pr o c e e dings of A CM SigmetricsMa y
GR S Ghandeharizadeh and L Ramos Con tin uous Retriev al of Multimedia Data Using Parallelism
IEEE T r ansactions on Know le dge and Data Engine ering ! August HL K Hong and J Leung Online Sc heduling of RealTime Tasks In R e alTime Systems Symp osium Decem b er Lev E L Levitan editor Handb o ok of A nimation T e chniques V an Nostrand Reinhold Co MK K Mogi and M Kitsurega w a Hot blo c k clustering for disk arra ys with dynamic striping In
Pr o c e e dings of V ery L ar ge Datab ases MKT K Ramamritham M Kamath and D T o wsleyCon tin uous media sharing in m ultimedia database
systems In Pr o c e e dings of the F ourth International Confer enceon Datab ase Systems for A dvanc e d
Applic ations D ASF AA
PGK D P atterson G Gibson and R Katz A case for Redundan tArra ys of Inexp ensiv e Disks RAID
In Pr o c e e dings of A CMSIGMOD Ma y RS K Ramamritham and J Stank o vic Sc heduling algorithms and op erating systems supp ort for
realtime systems Pr o c e e dings of the IEEE ! Jan uary R VR P Rangan H Vin and S Ramanathan Designing an OnDemand Multimedia Service IEEE
Communic ations Magazine ! July
R W A L N Reddy and J C Wyllie IO Issues in a Multimedia System IEEE Computer Magazine ! Marc h 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 pages ! August TTa N M Thalmann and D Thalmann editors Computer Animation The ory and Pr actic e Springer
V erlag TTb N M Thalmann and D Thalmann editors Synthetic A ctors in ComputerGener ate d D Films SpringerV erlag
App endix
Abstract (if available)
Linked assets
Computer Science Technical Report Archive
Conceptually similar
PDF
USC Computer Science Technical Reports, no. 629 (1996)
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. 616 (1995)
PDF
USC Computer Science Technical Reports, no. 633 (1996)
PDF
USC Computer Science Technical Reports, no. 625 (1996)
PDF
USC Computer Science Technical Reports, no. 623 (1995)
PDF
USC Computer Science Technical Reports, no. 592 (1994)
PDF
USC Computer Science Technical Reports, no. 612 (1995)
PDF
USC Computer Science Technical Reports, no. 653 (1997)
PDF
USC Computer Science Technical Reports, no. 587 (1994)
PDF
USC Computer Science Technical Reports, no. 864 (2005)
PDF
USC Computer Science Technical Reports, no. 627 (1996)
PDF
USC Computer Science Technical Reports, no. 634 (1996)
PDF
USC Computer Science Technical Reports, no. 863 (2005)
PDF
USC Computer Science Technical Reports, no. 628 (1996)
PDF
USC Computer Science Technical Reports, no. 590 (1994)
PDF
USC Computer Science Technical Reports, no. 589 (1994)
PDF
USC Computer Science Technical Reports, no. 791 (2003)
PDF
USC Computer Science Technical Reports, no. 622 (1995)
Description
Martha L. Escobar-Molano and Shahram Ghandeharizadeh. "Coordinated display of structured presentations using a multi-disk architecture." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 630 (1996).
Asset Metadata
Creator
Escobar-Molano, Martha L.
(author),
Ghandeharizadeh, Shahram
(author)
Core Title
USC Computer Science Technical Reports, no. 630 (1996)
Alternative Title
Coordinated display of structured presentations using a multi-disk architecture (
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
27 pages
(extent),
technical reports
(aat)
Language
English
Unique identifier
UC16270146
Identifier
96-630 Coordinated Display of Structured Presentations Using a Multi-Disk Architecture (filename)
Legacy Identifier
usc-cstr-96-630
Format
27 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
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/