Computer Science Technical Report Archive

USC Computer Science Technical Reports, no. 592 (1994)

(USC DC Other)

USC Computer Science Technical Reports, no. 592 (1994)

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Con tin uous Displa y of Video Ob jects Using MultiZone Disks
Shahram Ghandeharizadeh Seon Ho Kim and Cyrus Shahabi
Departmen t of Computer Science
Univ ersit y of Southern California
Los Angeles California April Abstract
Video serv ers are exp ected to pla y an imp ortan t role in a n um b er of application domains eg
library information systems en tertainmen t industryscien tic application s etc A c hallengin g task when
implemen ting these systems is to ensure a con tin uous displa y of audio and video ob jects If sp ecial
precautions are not tak en the displayof an object ma y suer from frequen t disruptions and dela ys
termed hiccups A second c hallengin g task is to congure a system to meet the p erformance requiremen ts
of an application its desired n um ber of sim ultaneous displa ys and the w aiting tolerance of a displa y F or applications with large data sets eg video serv ers magnetic disks ha v e established themselv es
as mass storage device of c hoice A tec hnique to increase the storage capacit y of disks is zoning A
sideeect of zoning is that it in tro duces a disk driv e with v ariable transfer rates This pap er describ es
tec hniques to supp ort a con tin uous displa y of video ob jects using a m ultizone disk driv e As compared
to previous approac hes the prop osed tec hniques harness the a v erage transfer rate of the magnetic disk
instead of its minim um transfer rate In addition w e describ e a conguration planner that logically
manipulates the zoning information of a disk driv e to supp ort the p erformance criteria of an application This researchw as supp orted in part b y the National Science F oundation under gran ts IRI IRI IRI
NYI a w ard and CD A and a HewlettP ac k ard unrestricted cashequipme n t gift
In tro duction
A tec hnological trend in the area of magnetic disks is an increase in their storage capacit y a increase
ay ear Zoning is one approac h to accomplish this A side eect of zoning is that it in tro duces a disk driv e
with v ariable transfer rates the dieren t regions of a disk driv e termed zones pro vide dieren t transfer
rates F or example a Seagate STW disk consists of zones with bandwidths v arying from to megab ytes p er second MBs see T able b A n um b er of studies ha vein v estigated tec hniques to supp ort a
hiccupfree displa y of con tin uous media video and audio using magnetic disk driv es AH R VR R V VR CL TPBG R Wa YCK Gem These studies assume a xed transfer rate for a disk
driv e If a system designer elects to use one of these tec hniques the system is forced to use the minim um
transfer rate of the zones for the en tire disk in order to guaran tee a con tin uous displa y of video ob jects W e
term these studies collectivelyasminim um transfer rate zone designs MinZtfr Section describ es a
cyclebased tec hnique TPBG BGM that falls in to MinZtfr category
This study in tro duces t w o alternativ e tec hniques V ARB and FIXB that harness the a v erage transfer
rate of zones while ensuring a con tin uous displa y Assuming the bandwidth required to displa yan object
is Mbs our prop osed tec hniques enable a Seagate W disk to supp ort sim ultaneous displa ys
as compared to with MinZtfr While v ersus ma y app ear insignican t at rst glance this is a
impro v ement

that has a signican t impact on the p erformance of scalable m ultidisk serv ers
F or example with a disk system the prop osed tec hniques harness enough of the disk bandwidth to
supp ort sim ultaneous displa ys as compared to with MinZtfr See GK for factors that
impact the scalabilityc haracteristics of parallel serv ers Ho w ev er b oth FIXB and V ARB suer from t w o
limitations when restricted to emplo y the ph ysical zone conguration pro vided bythe v endor First they
result in a higher latency ie the time elapsed from when the request arriv es un til the onset of its displa y Second they w aste disk space W e eliminate these limitati ons using a conguration planner that consumes
the p erformance requiremen ts of an application and manipulates the zoning information to supp ort the
p erformance criteria of an application The exp erimen tal results indicate that the output of the planner
almost alw a ys pro vides signican tsa vings as compared to MinZtfr MinZtfr cannot outp erform our
planner b ecause it is sim ulated as one of the p ossible congurations
There are three factors that are manipulated b y the conguration planner the p ercen tage of w asted
disk space the n umberofsim ultaneous displa ys ie throughput supp orted b y the disk and the maxim um
latency incurred b y a displa yT rading latency time for a higher throughput is dep enden t on the requiremen ts
of the target application One ma y increase the throughput of the Seagate from to displa ys using the
ph ysical zone c haracteristics with V ARB This causes the maxim um latency to increase from seconds to
min utes see the sev en th column of T able A small lo cal vide oondemand serv er ma y exp ect to ha v e

sim ultaneous displa ys as its maxim um load with eac h displa y lasting t w o hours With MinZtfr the disk
driv e supp orts a maxim um of sim ultaneous displa ys eac h observing a seconds latency During p eak
system loads sa yactiv e requests sev eral requests mayw ait in a queue un til one of the activ e requests
completes its displa y These requests observ e a latency time signican tly longer than a second p oten tially
in the range of hours dep ending on the status of the activ e displa ys and the queue of p ending requests In
this scenario it migh t b e reasonable to force eac h request to observea w orst case latency of min utes in
order to supp ort sim ultaneous displa ys
Alternativ elywitha news on demand application whose t ypical video clips ha v e a displa y time of appro x
imately four min utes a min ute latency time migh t not b e a reasonable tradeo for a higher n um ber of
sim ultaneous displa ys In this case the planner migh tc ho ose an alternativ e zone conguration that supp orts
alo w er latency time b y reducing system throughput
T o illustrate the impact of w asted disk space assume a hierarc hical m ultim edia storage manager consisting
of memory magnetic disk and a tertiary storage device As compared with the magnetic disk the tertiary
storage device pro vides a lo w storage cost desirable Ho w ev er a request observ es a higher latency and a
lo w er transfer rate undesirable when retrieving data from this device The data resides p ermanen tly on
the tertiary storage device The w orking set of an application is rendered disk residen t to minim i ze the
n um b er of references to the tertiary storage device The w orking set Den of an application consists of a
collection of ob jects that are rep eatedly referenced during a xed in terv al of time F or example in existing
video stores a few titles are exp ected to b e accessed frequen tly and a store main tains sev eral sometimes
man y copies of these titles to satisfy the exp ected demand These titles constitute the w orking set of a
videoondemand serv er If the amoun tof w asted disk space is suchthatthe w orking set of an application
can b ecome disk residen t then the impact of w asted space is marginal Otherwise the amoun tof w asted
space ma y reduce system p erformance due to frequen t references to the tertiary storage device that has
alo w er p erformance when compared to the magnetic disk driv e W equan tify this factor and consider its
impact in our exp erimen tal studies
Giv en a disk the prop osed planner logically manipulates zones to satisfy the p erformance ob jectiv es
of an application ie exp ected n um b er of activ e customers as w ell as their w aiting tolerance based on
the size of its w orking set The rest of this pap er is organized as follo ws Section pro vides an o v erview
of a magnetic disk driv e and the concept of zoning Section describ es t w o alternativetec hniques FIXB
and V ARB that ensure a hiccupfree displa y of video ob jects using a m ultizone disk driv e In addition it
quan ties the tradeos asso ciated with these t w o tec hniques and compares them with a cyclebased approac h
Section describ es the conguration planner and its heuristic based searc hmec hanism Section ev aluates
the prop osed tec hniques using MinZtfr as the comparison p oin t The results demonstrate the sup eriorit y

Spindle
Head
Arm
Platter
Cylinder
Arm
assembly
Figure A disk driv e
of our prop osed tec hniques Our conclusions and future researc h directions are con tained in Section Ov erview of a Magnetic Disk Driv e
A magnetic disk driveisamec hanical device op erated b y its con trolling electronics The mec hanical parts
of the device consist of a stac k of platters that rotate in unison on a cen tral spindle see R Wb for details
Presen tly a single disk con tains one t w o or as man y as sixteen platters see Figure Eac h platter surface
has an asso ciated disk head resp onsible for reading and writing data Eac h platter is set up to store data in
a series of trac ks A single stac k of trac ks at a common distance from the spindle is termed a cylinder T o
access the data stored in a trac k the disk head m ust b e p ositioned o v er it The op eration to rep osition the
head from the curren t trac k to the desired trac k is termed se ek Next the disk m ust w ait for the desired
data to rotate under the head This time is termed r otational latency T o meet the demands for a higher storage capacit ydiskdriv eman ufacturers ha vein tro duced disks with
zones A zone is a con tiguous collection of disk cylinders whose trac ks ha v e the same storage capacit yT rac ks
are longer to w ards the outer p ortions of a disk platter as compared to the inner p ortions hence more data
ma y b e recorded in the outer trac ks While zoning increases the storage capacit y of the disk it pro duces a
disk that do es not ha v e a single transfer rate The m ultiple transfer rates are due to the v ariable storage
capacit y of the trac ks and a xed n umberofrev olutions p er second for the platters T able presen ts the
zone c haracteristics of t w o commercial disk driv es Both are SCSI fast and wide disks with a gigab yte
storage capacit y A disk p erforms useful w ork when transfering data and w asteful w ork when p erforming seek op eration
The seek time is a function of the distance tra v eled b y the disk arm BG GHW R Wb Sev eral
studies ha vein tro duced analytical mo dels to estimate seek time as a function of this distance Other studies
ha vein v estigated tec hniques to minim i ze the seek time GKS BMC YCK One approac h is based
on Constr aine d Allo cation It con trols the placemen t of blo c ks of dieren t ob jects on the disk surface to

Zone Size MB T ransfer Rate MBs

Zone Size MB T ransfer Rate MBs
a HewlettP ac k ard C disk b Seagate STW disk
T able Tw o commercial disks and their zoning information
minim ize the distance tra v eled b y the disk arm Our prop osed tec hniques emplo y this approac h to lo calize
blo c k references to a single zone Hence when m ultiplexing the disk bandwidth among retriev al of N blo c ks
the incurred seek time is b ounded b y the maxim um seek time within a zone Let S eek k denote the time
required for the disk arm to tra v el k cylinders to rep osition itself from cylinder i to cylinder i k or i k Hence S eek and S eek cy l denote the time required to rep osition the disk arm b et w een t w o adjacen t
cylinders and the rst and the last cylinder of a disk with cy l cylinders resp ectiv ely Consequen tlyfor a
disk driv e with m zones the a v erage time required to mo v e among the cylinders within a zone is S eek cy l
m

Note that S eek cy l
m
Seek cy l m
This mo del b ecomes more complex b ecause our planner Section logically com bines t woph ysically adjacen t zones and treat them as a single zone The w orst seek time of
this zone is no w dieren t b ecause m is reduced W e eliminate this factor from discussion b y using a constan t
w orst seek time for eac h zone W e made this simplifying assumption in order to fo cus on v ariable transfer
rate of a magnetic disk driv e that deserv es sp ecial atten tion instead of seek time that has b een in v estigated
b y the previous studies The extensions of this study to incorp orate the v ariable w orst seek time for the
reads lo calized to a zone is straigh tforw ard
Con tin uous Displa y
This section presen ts t w o alternativ e designs to supp ort a con tin uous displa y of video ob jects W e assume

T erm Denition
T W S eek W orst seek time of a zone
including the maxim um rotational latency time
T cseek Seek time required to mak e a complete span
cap T otal capacit y of a disk driv e
m Num b er of zones in a disk driv e
Z i Zone i of a disk driv e where i m
R Z i T ransfer rate of Z i in MBs
siz e Z i Capacityof Z i
B Z i Blo c k size of Z i with V ARB
B Fixed blo c k size with FIXB
G i Group i constructed from merging t w o or more zones
R G i T ransfer rate of G i in MBs
siz e G i Capacityof G i
T scan Time to p erform one sw eep of m zones
T MUX Z i Time to read N blo c ks from zone Z i
siz e WS Size of the application w orking set
R C Displa y bandwidth requiremen t Consumption rate in MBs
Mem Required amoun t of memory
N Maxim um n um ber of sim ultaneous displa ys throughput
Maxim um latency time
C T otal system cost
T able List of terms used rep eatedly in this pap er and their resp ectiv e denitions
the follo wing
a disk with m zones Z
Z
Z
m The transfer rate of a zone i is denoted as R Z
i
Z
denotes
the outermost zone with the highest transfer rate R Z
R Z
R Z
m a single media t yp e with a xed displa y bandwidth R
C
the displa y bandwidth is less than or equal to the a v erage transfer rate of the disk R
C
P
m i R Z i m
The t w o prop osed tec hniques partition eac h ob ject X in to f blo c ks X
X
X
X
f The rst termed
FIXB renders the blo c ks equisized With the second tec hnique termed V ARB the size of a blo c k dep ends
on the transfer rate of its assigned zone The adv an tage of FIXB is its ease of implem en tation There are t w o
reasons for this claim First a xed blo c k size simplies the implem en tation of a le system that serv es as
the in terface b et w een memory and magnetic disk driv e Second the design and implemen tation ofamemory
manager with xed frames the size of a frame corresp onds to the size of a blo c k is simpler than one with
v ariable sized frames This is particularly true in the presence of m ultiple displa ys that giv e rise to race con
ditions when comp eting for the a v ailable memory
While V ARB migh t b e more complicated to implemen t
it requires a lo w er amoun t of memory and incurs a lo w er latency as compared to FIXB when the bandwidth
Weha v e participate d in the design co ding and debugging of a m ultiuser buer p o ol manager with a xed blo c ksizefor
a relational storage manager W e explored the p ossibilit y of extending this buer p o ol manager to supp ort v ariable blo c k size
and concluded that it required signican t mo dication s

T
MUX
T
MUX
T
MUX
0 1 N-1 0
1
N-1 0 1 N-1
... ... ... ...
...
T
Scan
Time
T
cseek
Zone ID
Disk
Activity
Z Z
Z
T
T
disk
(Z )
m-1
1 0
W_Seek
1
Figure T
S can
and its relationship to T
MUX
of the disk driv e is exhausted FIXB and V ARB share manycommon c haracteristics In Section w ede scrib e FIXB Subsequen tly Section describ es V ARB and its dierences as compared to FIXB Section compares these t wotec hniques with a cyclebased TPBG BGM approac h Section quan ties the
tradeos b et w een V ARB and FIXB
Fixed Blo c k Size FIXB
With this tec hnique the blo c ks of an ob ject X are rendered equisized Let B denote the size of a blo c k The
system assigns the blo c ks of X to the zones in a roundrobin manner starting with an arbitrary zone FIXB
congures the system to supp ort a xed n um b er of sim ultaneous displa ys N This is ac hiev ed b y requiring
the system to scan the disk in one direction sa y starting with the outermost zone mo ving in w ard visiting
one zone at a time and m ultiplexing the bandwidth of that zone among N blo c k reads Once the disk arm
reads N blo c ks from the innermost zone it is rep ositioned to the outermost zone to start another sw eep of
the zones The time to p erform one suc hasw eep is denoted as T
Scan
The system is congured to pro duce
and displa y an iden tical amoun t of data p er T
S can
p erio d The time required to read N blo c ks from zone
i denoted T
MUX
Z
i
is dep enden t on the transfer rate of zone i This is b ecause the time to read a blo c k
T
disk
Z
i
during one T
MUX
Z
i
is a function of the transfer rate of a zone
Figure sho ws T
S can
and its relationship with T
MUX
Z
i
for m zones During eac h T
MUX
p erio d
N activ e displa ys migh t b e referencing dieren t ob jects This w ould force the disk to incur a seek when
switc hing from the reading of one blo c k to another termed T
W Seek
T
W Seek
also includes the maxim um
rotational latency time A t the end of a T
S can
p erio d the system observ es a long seek time T
cseek
attributed
to the disk rep ositioning its arm to the outermost zone The disk pro duces m blo c ks of X during one T
Scan

MEM
TIME
(Sec)
0
. . .
. . . . . .
. . .
T
T
MUX
Scan
(Z )
i 01
m-2 m-1
T
MUX (Z ) T
MUX
(Z ) T
MUX
(Z ) T
MUX
(Z )
T
Max
Required
Memory
R
(R
C
(Z ) i R
C
-
cseek
)*T i
(Z )
*(T
i
(Z )-T
T
disk
(Z )
i
disk
MUX i
(Z )
disk
)
Figure Memory required on b ehalf of a displa y
p erio d m B b ytes The n um ber of b ytes required to guaran tee a hiccupfree displa yof X during T
Scan
should either b e lo w er than or equal to the n um ber of b ytes pro duced b y the disk This constrain t is formally
stated as
R
C
T
cseek
m X
i T
MUX
Z
i
m B The amoun t of memory required to supp ort a displa y is minimi zed when the left hand side of Equation equals its righ t hand side
During a T
MUX
N blo c ks are retriev ed from a single zone Z
Activ e
In the next T
MUX
p erio d the system
references the next zone Z
Activ e mod m
When a displa y references ob ject X the system computes the
zone con taining X
sa y Z
i
The transfer of data on b ehalf of X do es not start un til the activ e zone reac hes
Z
i
One blo c kof X is transfered in to memory p er T
MUX
Th us the retriev al of X requires f suc h p erio ds
The displa yof X ma y exceed
P
f j T
MUX
Z
i j mod m
p erio ds as describ ed b elo w The displa yof object
X requires some memory due to the v ariable transfer rate This is b est illustrated using an example Assume
that the blo c ks of X are assigned to the zones starting with the outermost zone Z
If Z
Activ e
is Z
then
this request emplo ys one of the idle T
disk
Z
slots to read X
Moreo v er its displa y can start immediately
b ecause the outermost zone has the highest transfer rate The blo c k size and N are c hosen suchthatthe
data accum ulates in memory when accessing outer zones and decreases when reading data from inner zones
on b ehalf of a displa y see Figure In essence the system uses buers to comp ensate for the lo w transfer
rates of inner zones using the high transfer rates of outer zones harnessing the a v erage transfer rate of the

disk Note that the amoun t of required memory reduces to zero at the end of one T
scan
in preparation for
another sw eep of the zones
The displayof an object ma y not start up on the retriev al of its blo c k from the disk driv e This is b ecause
the assignmen t of the rst blo c k of an ob ject ma y start with an arbitrary zone while the transfer and displa y
of data is sync hronized relativ e to the outermost zone Z
In particular if the assignmen tof X
starts with
a zone other than the outermost zone sa y Z
i
i then its displa ym ust b e dela y ed to a v oid hiccups The
duration of this dela y dep ends on the time elapsed from retriev al of X
to the time that blo c k X
m i
is retriev ed from zone Z
termed T
accessZ and the amoun t of data retriev ed during T
accessZ If the
displa y time of data corresp onding to item T
display m i islo w er than T
accessZ then the displa ym ust b e
dela y ed b y T
accessZ T
display m i T o illustrate assume that X
is assigned to the innermost zone Z
m ie i m and the displa y time of eac h of its blo c k is seconds ie T
display seconds If seconds elapse from the time X
is read un til X
is read from Z
then the displayof X m ust b e dela y ed b y
seconds relativ e to its retriev al from Z
m Otherwise its displa yma y suer from a second hiccup
if initiated up on retriev al The dela yto a v oid hiccup is shorter than the duration of a T
scan
Indeed the
maxim um latency observ ed b y a request is T
scan
when the n um b er of activ e displa ys is less than
N T
S can
T
cseek
m X
i T
MUX
Z
i
This is b ecause at most N displa ys migh t b e activ e when a new request arriv es referencing ob ject XIn
the w orst case scenario these requests migh t b e retrieving data from the zone that con tains X
sa y Z
i
and
the new request arriv es to o late to emplo y the a v ailable idle slot Note that the displa yma y not emplo y the
idle slot in the next T
MUX
b ecause Z
i is no w activ e and it con tains X
instead of X
Th us the displa y
of X m ust w ait one T
scan
p erio d un til Z
i
b ecomes activ e again
One can solv e for the blo c k size b y observing from Figure that T
MUX
Z
i
can b e dened as
T
MUX
Z
i
N B
R Z
i
T
W Seek
Substituting this in to Equation the blo c k size is dened as
B R
C
T
cseek
m N T
W Seek
m R
C
P
m i
N
R Z i Observ e that FIXB w astes disk space when the storage capacit y of the zones is dieren t This is b ecause
once the storage capacit y of the smallest zone is exhausted then no additional ob jects can b e stored as they
w ould violate a roundrobin assignmen t
Section quan ties the p ercen tage of disk bandwidth w asted b y
FIXB for the commercial disks of T able
When the n um b er of activ e displa ys exceeds N then this discussion m ust b e extended with appropriate queuing mo dels
Unless the n um ber of blo c ks for an ob ject is less than mW e ignored this case from consideratio n b ecause video ob jects
are t ypically v ery large

V ariable Blo c k Size V ARB
V ARB is similar to FIXB except that it renders the duration of T
MUX
Z
i
iden tical for all zones This is
ac hiev ed byin tro ducing v ariable blo c k size where the size of a blo c k B Z
i
is a function of the transfer rate of
a zone This causes the transfer time of eac hbloc k T
disk
to b e iden tical for all zones ie T
disk
B Z i R Z i B Z j R Z j for i j m Similar to FIXB the blo c ks of an ob ject are assigned to the zones in a roundrobin
manner and the concept of T
Scan
is preserv ed This means that the blo c ks of an ob ject X are no longer
equisized The size of a blo c k X dep ends on the zone it is assigned to Ho w ev er the c hange in blo cksize
requires a sligh t mo dication to the constrain t that ensures a con tin uous displa y
R
C
T
cseek
m T
MUX
m X
i B Z
i
The duration of T
MUX
is no w indep enden t of the transfer rate of a zone and is dened as
T
MUX
N T
disk
T
W Seek
Substituting this in to Equation the size of a blo c k for a zone Z
i
is dened as
B Z
i
R Z
i
R
C
T
W S eek
N m R
C
T
cseek
P
m i R Z
i
R
C
N m
Similar to FIXB V ARB emplo ys memory to comp ensate for the lo w bandwidth of inner zones using the
high bandwidth of the outer zones This is ac hiev ed b y reading more data from the outer zones Moreo v er
the displayof an object X is sync hronized relativ e to the retriev al of its blo c k from the outermost zone and
ma y not start imm ediately up on retriev al of X
V ARB w astes disk space when
siz e Z i R Z i siz e Z j R Z j for i j and i j m The amoun tof w asted space dep ends on the zone that accommo dates the few est blo c ks
This is b ecause the blo c ks of an ob ject are assigned to the zones in a roundrobin manner and once the
capacit y of this zone is exhausted the storage capacit y of other zones cannot b e used b y other video ob jects
A Comparison with Cyclebased Approac h
A general approac h to supp ort con tin uous displa ys of N sim ultaneous video ob jects is a cy cl ic approac h
This approac h strip es eac h ob ject in to a n um b er of blo c ks Subsequen tly the disk bandwidth is m ultiplexed
among retriev al of N blo c ks corresp onding to N dieren t ob jects within a cycle There are t wov ariations
of this approac h cycleb ase d TPBG BGM and pipel ining GR BGMJ With cyclebased the
blo c ks retriev ed in one cycle are displa y ed in the next cycle With pip elining the displa y of eac h blo c k
starts as so on as its retriev al is initiated Hiccups are eliminated b y prefetc hing and c ho osing appropriate
blo c k sizes as describ ed in Section and The ma jor adv an tage of cyclebased is that the order of
blo c k retriev al can b e sh ued within a cycle YCK to force the mo v emen t of the disk arm to sim ulate

FIXB V ARB
N Blo c k Size Minim um Blo c k Maxim um Blo c k Av erage Blo c k
MBytes Size MBytes Size MBytes Size MBytes
T able Seagate STW disk
the scan algorithm T eo minim izing the disk seek time A limitation of this approac h is its high memory
requiremen t attributed to its double buering c haracteristic t wice as m uc h as that of pip elining Both
FIXB and V ARB emplo y pip elining for the follo wing t w o reasons First the transfer rate of a m ultizone
disk driv e is not xed resulting in a v ariable duration of a cycle This complicates the memory managem en t
within a cycle Second our tec hniques con trol the placemen ts of the blo c ks across the disk surface suc h
that anyt w o blo c ks retriev ed within a cycle reside on a single zone Hence the incurred seek time b et w een
retrieving t wobloc ks is m uc h less than the maxim um seek time of the disk see Section This renders the
b enets of a sh uing blo c ks in a cycle marginal
FIXB vs V ARB
While V ARB determines the blo c k size based on transfer rate of the individual zones FIXB determines
the blo c k size as a function of the a v erage transfer rate of the zones In essence V ARB p erforms lo cal
optimization while FIXB p erforms global optimization This enables V ARB to c ho ose smaller blo c k sizes
when compared to FIXB for a xed n um ber of users T able presen ts the required blo c k size as a function of
the n um b er of sim ultaneous displa ys N for the Seagate disk driv e
The bandwidth requiremen tofobjects
that constitute the database is Mbs R
C
Mbs The a v erage transfer rate of the disk can supp ort
a maxim um of sim ultaneous displa ys
A small blo c k size minimi zes the duration of T
S can
whic h in turn reduces the amoun t of required
memory This is reected b y the results presen ted in T able F or eachof V ARB and FIXB this table
presen ts the required memory latency
p ercen tage of w asted disk space and bandwidth as a function of N F or users the memory requiremen t of a system with FIXB is sev en times higher than that with V ARB
Moreo v er the maxim um incurred latency is more than min utes with FIXB as compared to min utes
The n um b ers for the HP disk driv e are con tained in App endix A Latency is equiv alen t to the duration of T
S can

FIXB V ARB
N Memory Latency w asted Avg w asted Memory Latency w asted Avg w asted
MBytes Sec disk space disk bandwidth MBytes Sec disk space disk bandwidth
T able Seagate STW disk
with V ARB
The p ercen tage of w asted disk space with eachof FIXBand V ARB is dep endenton the ph ysical c harac
teristics of the zones While V ARB w astes a lo w er p ercen tage of the Seagate disk space it w astes a higher
p ercen tage of the HP disk space see T able in App endix A The a v erage p ercen tage of w asted disk
bandwidth is lo w er with FIXB b ecause it minimi zes this v alue from a global p ersp ectiv e With V ARB a
n um b er of outer zones w aste a higher p ercen tage of their bandwidth in fa v or of a smaller blo c k size these
zones increase the p ercen tage of w asted disk bandwidth
Conguration Planner
Alternativ e applications ha v e dieren t p erformance ob jectiv es It ma y not alw a ys b e desirable to incur a
high latency to supp ort a large n um b er of displa ys F or examples an application migh t desire few er displa ys
in fa v or of a lo w er latency time Hence it migh t b e b enecial to com bine some zones reduce m to reduce
the latency time This w ould reduce the a v erage transfer rate of the disk that migh t minim ize the n um ber
of sim ultaneous displa ys supp orted b y the system Another application with a v ery small w orking set migh t
require a high throughput and a lo w latency time If the w orking set of this application ts in the outermost
zone of the disk then the system should eliminate all the other zones harnessing the highest transfer rate
for this application By eliminating m zones the incurred latency is also minimi zed These examples
motiv ate the need for a conguration planner that logically manipulates the zoning c haracteristics of a disk
to satisfy the p erformance ob jectiv es of a target application
The inputs to the planner include the c haracteristics of an application its p erformance ob jectiv es and
the ph ysical attributes of the target disk driv e The c haracteristics of an application include R
C
and the
size of its w orking set The p erformance ob jectiv es of the application include its desired latency desir ed

REDUCE
Heuristics
(e.g., MaxT)
Q = < N, l, B, C >
Q1
Q2
.
.
.
.
.
.
Exhaustive
Iteration
ODZA Reduced
DZA
DZA1
DZA2
Core
BYPASS
Heuristics
(e.g., MinW)
Q = < N, l, B, C >
Q1
Q2
.
.
.
.
.
.
DZA1
DZA2
Core ODZA
a. Class of REDUCE heuristics
b. Class of BYPASS heuristics
Figure Tw o alternativ e classes of heuristics
and throughput N
desir ed
The ph ysical attributes of a disk include T
W Seek
T
cseek
and its Disk Zone
A rr angement DZA The outputs of the planner are a new DZA and an appropriate blo c k size that satises
the p erformance ob jectiv es while minim izi ng cost A new DZA can b e constructed b y merging one or more
adjacen t zones in to one termed mer ge eliminating one or more zones from consideration termed eliminate splitting a zone in to sev eral zones termed split and a com bination of the rst three tec hniques The
planner consists of three comp onen ts The rst comp onen t constructs all p ossible DZAs using a com bination
of merge and eliminate A second comp onen t termed core computes the p ossible throughput latency memory requiremen t and costs of using a DZA with either FIXB or V ARB It emplo ys the analytical mo dels
of Section for this computation One can emplo y these t wocomponen ts to nd the c heap est conguration
that supp orts the p erformance ob jectiv es of an application ie desir ed
and N
desir ed
Ho w ev er as describ ed
later in Section the n um b er of p ossible DZAs b ecomes large when a disk consists of a large n um ber of
zones requiring y ears of computation to complete this exhaustiv e searc h As a solution the third comp onen t
of the planner in tro duces heuristics Tw o classes of heuristics are describ ed REDUCE these heuristics
construct a new DZA termed R e duc e d DZA that consists of few er zones than the original DZA that is used
as input to the rst comp onen t of the planner see Figure a BYP ASS these heuristics eliminate the
rst comp onen t of the planner all together b y pro ducing a select n um b er of DZAs based on the kno wledge
of target application see Figure b In the follo wing w e describ e eac h of these comp onen ts starting with
the core
The Core
The input parameters of core are similar to those of the planner see Figure Core emplo ys the analytical
mo dels of Section to output p ossible v alues of N B and C for either FIXB or V ARB It iterates
o v er p ossible v alues of N to generate these quadruples The lo w er and upp er b ound for N are N
desir ed

and b
R Z R C
c resp ectiv ely C denotes the cost of the system It is computed as a function of the amoun tof
memory Mem required b y the target application
C Mem ram disk where ram is the cost of memory p er megab yte and disk is the cost of the single disk driv e required b y
the application
The amountof w asted disk space W is used for t w o purp oses First it is used to eliminate those
DZAs that a v oid the w orking set of an application from b ecoming disk residen t siz e WS cap W where siz e WS and cap denote the size of the w orking set and the disk storage capacit y resp ectiv ely Second the w aste is used to compute the extra latency time observ ed b y requests referencing ob jects that
do not constitute the w orking set of the target application The larger the w aste the higher the probabilit y
that a reference to a nonw orking set ob ject is directed to the tertiary storage device resulting in a higher
a v erage latency time T o compute the extra latencyw e use the a v erage size of an ob ject Size
Av g req
to
estimate the n um b er of ob jects that constitute the database DB b
siz e DB Size Av g req
c and the w orking
set WS b
siz e WS Size Av g req
c Th us the n um b er of ob jects that do not constitute the w orking set is
DB WS A DZA ma y enable a larger n um b er of ob jects than WS to b ecome disk residen t
termed DZA DZA m ust b e greater than or equal to WS Otherwise it w ould ha v e b een eliminated
Let REM denote DZ A WS Let p
b e the probabilit y that a request references ob jects inside
the w orking set Assuming that references are randomly distributed across the ob jects that constitute the
p ortion of the database excluding its w orking set the probabilit y that a reference to one suchobjectis
directed to tertiary is p
ter
RE M DB WS Th us the probabilit y that a request retriev es data from the
tertiary is p
p
ter
It will observ e a minim um extra latency time of
Size Av g req
D T er tiar y
where D
T er tiar y
is
the transfer rate of the tertiary storage device Multiplying these comp onen ts with one another the extra
latency time observ ed due to the w asted space is computed as
p
RE M DB WS Size
Av g req
D
Tertiary
Cores output is a n um b er of quadruples Q N B C where C is the cost computed using Eq and B either denotes blo c k size with FIXB or the largest blo c k size corresp onding to the outermost zone
with V ARB
Note that the latency time incorp orates b oth latency times computed in Eqs and One can in v ok e Core using alternativ e input DZAs to generate a list of quadruples Subsequen tlya
searc h pro cedure can tra v erse the list to lo cate a quadruple Q N B C that is b oth the c heap est in
the list ie has the minim um C v alue and satises the follo wing conditions N
desir ed
N and desir ed
The blo c k size for an y zone Z
i
can then b e computed as B Z
i

B
R Z
R Z
i

Core
Disk-Zone-Arrangement:
m, (R(Zi), size(Zi))
0 <= i < m
< N, l, B, C >
Size of the application WS
N
l
desired
desired
R
C
T ,
cseek T W_Seek
Figure Core the analytical mo dels of the planner
Zone Size MB T ransfer Rate MBs
T able Prepro cessed zone information of Seagate STW disk
Exhaustiv e Iteration
The input to Core is one p ossible DZA The v endorbased zone conguration is the original DZA termed
ODZA If i ph ysically adjacen t zones of the ODZA pro vide an iden tical transfer rates then these i zones
are reduced in to one The size of this zone is the total size of the i zones F or example in T able b
R Z
R Z
MBs These t w o zones are com bined to form a single zone whose size is equiv alen t
to siz e Z
siz e Z
and a transfer rate of MBs This pro cess is also rep eated for Z

and Z
Therefore the ODZA of the Seagate disk driv eof T able is reduced to that of T able with zones
Other DZAs can b e constructed b y using the follo wing t w o op erations el iminate and merge As
suggested b y its name the el iminate op eration simply eliminates zones from a disk driv e Figure a
illustrates an ODZA of a zone disk driv e Figure b sho ws one p ossible DZA b y eliminating zone Zone
elimination w astes disk space ho w ev er it migh t increase the a v erage transfer rate of a disk driveeg if
the innermost zone is eliminated If the w orking set of an application requires the en tire storage capacit y
of a disk eliminating zones migh t not b e appropriate b ecause it w ould increase the frequency of access to
the tertiary storage device Another dra wbac k of eliminating a zone not as signican tas the w asted space

a. 4-zone ODZA
b. eliminate operation
(on zone 2 of a)
0
2
3
1
0
3
1
c. merge operation
(on zones 0 and 1of a)
G0 = {0,1}
G1 = {2}
G2 = {3}
Figure Eliminate and merge op erations
m P Time required for exhaustiv e iteration
Assuming eac h iteration requires ms
seconds
y ears
T able Complexit y of Exhaustiv e
factor is the increase in T
W Seek
when retrieving t wobloc ks from adjacen t zones see Section The mer g e op eration com bines t w o adjacen t zones in to one Figure a illustrates an ODZA of a zone
disk driv e Figure b sho ws one p ossible DZA b y merging zones and The new DZA can b e considered as
a disk drivewiththree zones m
First group G
consists of Z
and Z
G
f g the second group
corresp onds to Z
and the third to Z
T o guaran tee a con tin uous displa y the transfer rate of the group is
determined b y the minim um transfer rate of its constituting zones F or example for this DZA the transfer
rate and size of G
is computed as follo wing R G
R Z
siz e G
siz e Z
siz e Z
Merging
reduces the total n um b er of zones ie m
m in a DZA resulting in a lo w er latency time see Eq Ho w ev er it reduces the a v erage transfer rate of the disk that migh t minim i ze the n um ber of sim ultaneous
displa ys It also increases T
W S eek
Using merge and el iminatea n um b er of DZAs can b e constructed Subsequen tlyeac h DZA can b e
fed to Core to generate a list of quadruples When m is small it migh tbe w orth while to generate all
p ossible DZAs b ecause the planner is in v ok ed once at system conguration time Figure illustrates ho w
all p ossible DZAs can b e generated for a zone ODZA First the el iminate op eration is in v ok ed on the
ODZA constructing all subsets excluding the empt y set of the original set of zones The n um b er of p ossible
nelemen t DZAs is
m
n
A
F or eac h DZA generated b y el iminatew e then consider all p ossible mer g e op erations T o illustrate
consider Figure F rom the DZA consisting of zones Z
Z
and Z
m
w e can generate DZAs
consisting of either one t w o or three groups Supp ose w e are in terested in t w ogroup DZAs There are t w o
This is a c ombinatorial function the binomial c o ecient whic h can b e written as
m
n
m n m n
m = 4
z = 1 s = 0
m z - 1
m
z
z - 1
s
Max number of merge-points
Number of required groups - 1
Max. number of zones
Elimination
2
= 4 possible ways to
do 3-zone elimination
( 1-zone disk drive )
0 1
4
1
3
0
2
0
3
1
2
1
3
2
3
= 6 possible ways to
do 2-zone elimination
( 2-zone disk drive )
4
2
0
1
0
1
2
1
2
3
0
2
3
= 4 possible ways to
do 1-zone elimination
( 3-zone disk drive )
4
3
= 1 possible way to
do 0-zone elimination
( 4-zone disk drive )
4
4
0
1
3
= 1 possible way to
generate 1-group DZA
from a 3-zone DZA
3 - 1
0
3 - 1
1
3 - 1
2
Merging
Number of decided zones
Example
Number of all possible
DZAs:
P =
ODZA
0
2
3
1
0
2
3
1
= 2 possible ways to
generate 2-group DZAs
from a 3-zone DZA
= 1 possible way to
generate 3-group DZA
from a 3-zone DZA
G0 =
{0, 1, 3}
G0 = {0}
G1={1,3}
G0={0,1}
G1 = {3}
G0 = {0}
G1 = {1}
G2 = {3}
Figure Exhaustiv e iteration generating all p ossible DZAs

m
p ossible w a ys to merge zones Hence t w o p ossible DZAs can b e deriv ed In general the n um ber
of g group DZAs that can b e deriv ed using a DZA consisting of m
zones is computed as
m
g A
Therefore the total n um b er of p ossible DZAs is computed as
P m
X
z z X
s m
z
A
z s
A

Note that the order of in v oking op erations el iminate follo w ed b y mer g e is imp ortan t If the order is
rev ersed then duplicate DZAs are pro duced In this case P
is computed as
P
m X
s s X
z m s
A
s
z
A

where P
P The states coun ted b y P
P are duplicates Ob viously considering duplicate DZAs w astes
time
Assume that feeding eac h DZA to Core and generating a quadruple requires milliseconds T ra v ersing
the en tire searc h space to lo cate an optimal conguration with the zone HP disk driv e see T able a
requires seconds Ho w ev er this b ecomes a y ear computation with the zone Seagate disk driv e
see T able b ecause of its large n um b er of zones that results in an explosion of p ossible states This
motiv ates the need for heuristics to reduce the n um b er of p ossible DZAs
Heuristics
The planner emplo ys t w o alternativeapproac hes to minim i ze the n um b er of p ossible DZAs One approac his
to use heuristics to reduce the n um b er of zones m in order to prune the n um b er of p ossible DZAs termed
class of REDUCE heuristics The reduced ODZA is used as input to the exhaustiv e iteration of Section see Figure a An alternativ e approachisto b ypass the exhaustiv e iteration all together termed class of
BYP ASS heuristics This approac h selectiv ely c ho oses a small n um b er of DZAs that are used as input to
the core see Figure b
A large n um b er of heuristics can b e dev elop ed for eac h class A heuristic migh t b e designed with dieren t
ob jectiv es minim ize latency maxim ize throughput minim ize w asted space etc In this pap er w e describ e
t w o The rst is a mem b er of REDUCE that attempts to maxim ize the throughput of the system termed
M axT The second heuristic is a mem berofBYP ASS that attempts to minim i ze the amountof w asted
space termed MinW W e describ e eac h heuristic in turn

SIZE(Zi)
MB
R(Zi)
MBps i
0
1
2
324
112
76
3.40
3.17
3.04
3
4
5
6
7
77
71
145
109
89
2.92
2.78
2.54
2.27
2.02
slice on
zone 2
SIZE(G0) = 512
R(G0) = 3.04
SIZE(G1) = 491
R(G1) = 2.02
R(s=2) = (512*3.04 + 491*2.02) / 1003 = 2.54
SIZE(Zi)
MB
R(Zi)
MBps i
0
1
2
3
4
324
112
76
77
71
3.40
3.17
3.04
2.92
2.78
5
6
7
145
109
89
2.54
2.27
2.02
slice on
zone 4
SIZE(G0) = 660
R(G0) = 2.78
SIZE(G1) = 343
R(G1) = 2.02
R(s=4) = (660*2.78+ 343*2.02) / 1003 = 2.52
Figure MaxT algorithm for m
desir ed

MaxT Heuristic
The inputs to MaxT are the original DZA of the disk with m zones and the desired n um ber of zones
for the output DZA m
desir ed
where m
desir ed
m The output of MaxT is a reduced DZA that consists
of m
desir ed
zones F or this transformation MaxT emplo ys the mer g e op eration Its goal is to pro duce a
m
desir ed
zone DZA with a maxim um exp ected transfer rate hoping that this w ould increase throughput T o
compute the exp ected transfer rate let us dene the probabilit y of data residing on Z
i
as
P
i
siz e Z
i
cap

Hence the exp ected transfer rate of a mzone ODZA is dened as
m X
i P
i
R Z
i

Recall that merging reduces the exp ected transfer rate of the disk driv e b ecause the transfer rate of a
group is determined b y the minim um transfer rate of its constituting zones Fixing the v alue of m
desir ed
there are sev eral w a ys to merge zones of a DZA MaxT is in terested in the one that maximi zes the exp ected
transfer rate T o observ e consider Figure where m
desir ed
Tw o alternativew a ys of merging are
demonstrated dividing the zones in to t w o groups b y slicing the ODZA at zone s and s The
exp ected transfer rate for constructing groups b y slicing at zone k is computed as

R s k
P
m desir ed i siz e G
i
R G
i
cap

F rom Figure and b y applying Eq w e obtain

R s MBs and

R s MBs MaxT
v aries k from to m tond a k
whic h maximi zes

R s k
No w assume m
desir ed
In this case MaxT w orks as depicted in Figure It constructs one group b y
One migh t argue that this probabilit y should b e based on the n um b er of blo c ks accommo da ted b y b oth a zone and the
disk This is infeasible b ecause at this stage the size of a blo c k is unkno wn the size of a blo c k is determined b y the Core

Best of
red_m = 2
Best of
red_m = 2
. . .
Best of
red_m = 2
k = 2
k = 1
k = m - 2
0
k
m-1
0
1
k
m-1
0
k-1
k
m - 1
Figure MaxT algorithm for m
desir ed

float MaxTstar t w c locals lic es int startw
float c
int localsli ce sM AX M int ij
float ci float MaxMu float Mu
if w localsli ce s start
returnZo ne T m R for istart imw i ci ZoneTi siz e Mu ZoneTi R ci MaxTi w c ci localslic es c ci c if Mu MaxMu MaxMu Mu
locals lic es w i for j jw j globals li ces w j localsli ce s j return Max Mu Figure A recursiv e function for MaxT
merging the rst k zones Next it determines the b est w a y to generate t w o groups using the remaining m k
zones T o accomplish this it rein v ok es MaxT on a DZA consisting of zones k to m where m
desir ed
This can b e rep eated recursiv ely based on anyv alue for m
desir ed
In general MaxT algorithm sim ulates
the divideandconquer algorithm Figure illustrates a recursiv e function for MaxT in C language Its
rst reference should b e MaxT m
desir ed
cap N U LL array MinW Heuristic
The input to MinW is an ODZA Its output is a n um b er of DZAs MinW constructs the output DZAs with
the ob jectiv e to reduce the amountof w asted space with either FIXB or V ARB T o simplify the discussion
w e rst describ e this heuristic for FIXB
MinW is based on the observ ation that the zone with the minim um size determines the total useful

a. 4-zone ODZA b. split operation (on zone 0)
c. split operation (on zone 2 of b)
0
2
3
1
2
3
1
0.0
0.1
2.0
2.1
3
1
0.0
0.1
Figure Split op eration
capacit y of a disk driv e By forcing the zones to b e equisized MinW can minimi ze the w asted space The
spl it op eration see Figure partitions a zone sa y Z
i
in to t w o or more equisized subzones The transfer
rate of eac h subzone is iden tical to the transfer rate of Z
i
This increases the n um b er of zones in a DZA
Ho w ev er b y constructing equisized zones the p ercen tage of w asted space with FIXB is reduced
T o construct equisized zones one can force the size of eac h subzone to b e the GCD Greatest Common
Divisor of the p ossible zone sizes Subsequen tly the GCD is considered as the size of eac h subzone termed
subsiz e and eac hzone i is splitted in to
siz e Z i subsiz e
subzones This approac h raises t w o issues First the zone
sizes are real n um b ers in Megab ytes while GCD is meaningless for real n um b ers Second if at least t woof
the zone sizes are prime n um b ers then their GCD will b e ie subsiz e resulting in a large n um ber
of zones There are t w o solutions for the rst problem consider lo w er units suc h as Kilob ytes for the
zone sizes in order to con v ert them to in tegers andor truncate the sizes and suer from some internal
fr agmentationT osolv e the second problem w ev ary subsiz e from the GCD to the size of the smallest zone
When subsiz e GCD some space is w asted p er zone b ecause
siz e Z i subsiz e
migh t not b e an in teger n um ber W e
term this w aste external fr agmentation MinW generates one DZA p er v alue of subsiz e Subsequen tly for
eac h DZA it computes the w aste resulting from b oth in ternal and external fragmen tation and the n um ber
of zones In practice w ev ary subsiz e from the size of the smallest zone do wnto the GCD As subsiz e
approac hes the GCD the n um b er of zones increases Ho w ev er the amoun tof w aste migh t either increase or
decrease MinW main tains only those DZAs with few zones that minimi ze w aste
The complexit y of MinW for FIXB is O cap This is b ecause for eac hv alue of subsiz e MinW tra v erses
all the m zones to apply the spl it op eration Moreo v er the n um b er of p ossible subsiz esis
cap
m
This is
assuming the worst case of GCDand the size ofthe smallest zone b eing
cap
m
Due to this complexit y
O cap
m
m O cap assuming Kilob ytes as disk storage capacitymigh t result in a long computation time
Hence c ho osing a unit for storage capacit y is a tradeo b et w een minimi zing the in ternal fragmen tation as
compared to reducing b oth the computation time of the algorithm and the n um b er of resulting zones
This explains wh y split is not appropriat e for use in the REDUCE class of heuristics
T o force the subzones to b e equisized subsiz e should not exceed the size of the smallest zone

Disk Capacit y cap Gigab yte
Max Rotational Latency Time milliseconds
Max Seek Time milliseconds
T W Seek milliseconds
T cseek milliseconds
T able Seagate and HP disk parameters
The algorithm of MinW for V ARB is similar to that of FIXB The dierence is that the zones should b e
forced to ha v e a xed
siz e Z i R Z i in order to minimi ze the w aste Therefore the GCD is computed for the size
o v er transfer rate of zones F urthermore a rate f actor is v aried from the minim um v alue of
siz e Z i R Z i where
im do wn to the GCD Subsequen tly the size of the Z
i
s subzones subsiz e Z
i
is computed as
subsiz e Z
i
r ate factor R Z
i
The complexit y of this algorithm is O cap
where is the exp ected
transfer rate of the disk see Eq In Section w e sho w that MinW enables V ARB to w aste no space of the HP disk b y increasing its
n um b er of zones to and xing
siz e Z i R Z i at seconds Ho w ev er the in ternal and external fragmen tation
caused b y the spl it op eration w astes of the disk space One migh t b e tempted to apply the exhaustiv e
iteration on the obtained zone DZA This is not recommended b ecause the exhaustiv e iteration migh t
destro y the xed v alue of
siz e Z i R Z i This also explains wh y MinW is considered as a BYP ASS instead of a
REDUCE heuristic
P erformance Ev aluation
W e conducted some exp erimen ts to demonstrate the exibilit y and sup eriorit y of our approachascom pared to the con v en tional approac h MinZtfr whic h assumes the transfer rate of the innermost zone as
the transfer rate of the disk ev aluate our heuristics and compare V ARB with FIXB in the presence
of heuristics F or this purp ose w e considered the system latency time throughput and cost as p erformance
criteria The system cost is determined assuming a memory cost of p er MB and for a onegigab yte
disk driv e In these exp erimen ts w e used b oth Seagate and HP disk driv es see T able Except for the
zoning information the remaining parameters of these t w o disks are almost iden tical T able summarized
the v alues of these parameters
The target application is a newsondemand Its database consists of MPEG compressed news
clips eac h with a Mbs bandwidth requiremen t The duration of eac h news clip is min utes Th us eac h
clip is MBytes in size resulting in a GBytes database W ev aried the size of the w orking set from news clips MB up to MB The probabilit y of reference to the w orking set ob jects is W e
assumed a tertiary storage device with a transfer rate of MBs

Num ber of objects Maximum numb er of users
in w orking set MinZtfr MaxT MaxT MaxT MaxT Exhaustiv e MaxT
WS m
desir ed
m
desir ed
m
desir ed
m
desir ed
m
desir ed

T able Maxim um throughput with xed latency time sec in HP C disk
W e compared our approac h with a con v en tional approac h that assumes the disk consists of a single zone
with the transfer rate of the innermost zone and the storage capacityof the en tire disk MinZtfr MaxT
with m
desir ed
In addition to considering MaxT with dieren t m
desir ed
v alues w e also considered the
full exhaustiv e iteration without heuristics termed E xhaustiv e

Exhaustiv eis iden tical to MaxT in v ok ed
with m
desir ed
using the HP disk driv e T able presen ts the maxim um n um b er of users byeac h tec hnique
columns to as a function of the n um b er of clips that constitute the w orking set rst column W eset
desir ed
at the one observ ed with MinZtfr seconds The throughput with MinZtfr is indep enden tof
the size of the w orking set b ecause it w astes an insignican t fraction of the disk space ie utilize almost
of the disk space When siz e WS MB MaxT and Exhaustiv e construct a DZA that results
in a higher throughput as compared to MinZtfr When siz e WS MB Exhaustiv ec ho oses a DZA
that logically consists of t w o zones with an a v erage transfer rate that can supp ort displa ys outp erforming
MinZtfr Note that a larger v alue for m
desir ed
do es not alw a ys translate in to a higher throughput F or
example when the w orking set consists of ten clips th
rowofT able MaxT with m
desir ed
results in
alo w er throughput when compared with m
desir ed
This is b ecause MaxT is only a heuristic and migh t
fail to compute an optimal throughput This is also explains wh y Exhaustiv e outp erforms MaxT Similar
observ ation w as made with the Seagate disk driv e see App endix B
W e considered the minim um cost conguration that supp orts the latency and throughput iden tical to that
pro vided b y MinZtfr Figure a sho ws the p ercen tage reduction in cost obtained with b oth Exhaustiveand
MaxT m
desir ed
relativ e to MinZtfr This p ercen tage is computed as
cost Min Z tf r cost E xhaustive cost Min Z tf r and ma y not exceed Exhaustiv epro vides a more signican tsa ving when compared to MaxT b ecause
it considers all p ossible DZAs The c heap est conguration do es not necessarily minimi ze the latency time
Figure b sho ws the p ercen tage reduction in latency with b oth Exhaustiv e and MaxT relativ e to MinZ
W e can consider the en tire searc h space for the HP disk b ecause it consists of eigh t zones

#(WS)
Saving
in cost (%)
5 10 15 20
0
5
10
15
20
25
30
Exhaustive
MaxT
#(WS)
Reduction
in latency (%)
5 10 15 20
0
20
40
60
80
100
MaxT
Exhaustive
a Cost b Latency
Figure Latency time and system cost with HP disk N m
desir ed
for MaxT Latency (sec)
N
5 10 15 20
0
10
20
30
40
50
60
70
MaxT
MinW
Figure MinW vs MaxT siz e WS MB
tfr While Exhaustiv e minimi zes cost it also results in a higher a v erage latency for some w orking set size
Relativ e to MinZtfr b oth Exhaustiv e and MaxT pro vide a signican t reduction in latency for small w orking
set size
MinW is insensitivetothe w orking set size This is b ecause it a v oids the mer g e and el iminate op erations
all together It cannot outp erform Exhaustiv e b ecause in the w orst case Exhaustivesim ulates MinZtfr whic h
results in no w aste Ho w ev er for large w orking sets where the amountof w aste should b e minim i zed MinW
outp erforms MaxT This is demonstrated in Figure where the size of the w orking set is xed at MBytes and N is v aried MinW can supp ort a higher n um b er of displa ys b ecause it emplo ys the a v erage
transfer rate of the disk b y splitting zones MaxT constructs a DZA with one zone to accommo date the
large w orking set assuming the minim um transfer rate of the disk
T able demonstrates the n um b er of zones constructed b y MinW and its impact on b oth the p ercen tage of
w asted space and latency The rep orted n um b ers are based on the maxim um n um b er of displa ys supp orted

Num b er of zones m disk storage w aste latency sec
T able MinW heuristic N 2.5 5 7.5 10 12.5 15
0
10
20
30
40
50
60
70
No. of users
Saving
in cost (%)
Figure Cost comparison b et w een VARB andFIXB siz e WS MB sec b y MinW N Observ e the tradeo b et w een latency and p ercen tage of w asted space A higher n um ber
of zones minimi zes the w asted space at the exp ense of a higher latency time
Finallyw e compared FIXB and V ARB in the presence of heuristics Figure summarizes the obtained
results The xaxis of this gure is the desired throughput and the yaxis is the p ercen tage of sa ving in
the system cost for V ARB as compared to FIXB cost F IXB cost V ARB cost FIXB The size of the w orking set and
the latency time w ere xed at MB and seconds resp ectiv elyW eemplo y ed MaxT for the HP disk
driv e with m
desir ed
Since FIXB requires larger memory to satisfy a higher throughput it costs more
when the n um b er of activ e users is higher than These results are in accordance with the observ ations of
Section Conclusions and F uture Researc h Directions
This pap er describ es t wotec hniques that ensure a hiccupfree displa yof con tin uous media data t yp es using
am ultizone disk driv e As compared to the previous studies the prop osed tec hniques harness the a v erage
transfer rate of a disk driv e instead of assuming its minim um transfer rate W e describ ed a conguration
planner that consumes the p erformance criteria of an application and logically congures the activ ezones

of a disk driv e to supp ort the desired criteria of the application Due to the large size of this optimization
space heuristics are in tro duced to sp eedup the planner W eev aluated b oth the prop osed tec hniques and
the planner The results demonstrate the sup eriorit y of the prop osed tec hniques when the w orking set of an
application do es not require the en tire storage capacit y of the disk driv e
W e are extending the prop osed designs in t wow a ys First w e are exploring the p ossibilit y of supp orting
a database consisting of a mix of media t yp es eac h with a dieren t bandwidth requiremen t and p erformance
criteria Giv en suc h a mix the blo c k size and zone conguration should b e c hosen to strik e a compromise for
the alternativ e media t yp es Second w e are in v estigating the role of m ultizone disks in a m ultidisk serv er
W e are dev eloping analytical mo dels to quan tify the throughput and latency time of a system as a function
of the blo c k size and the n um b er of disk driv es W e are also extending the conguration planner to compute
the c heap est system that supp orts a desired p erformance criteria
References
AH D Anderson and G Homsy A cotin uous media IO serv er and its sync hronizati on IEEE Computer Octob er BG Dina Bitton and J Gra y Disk shado wing In Pr o c e e dings of the Internation al Confer enceonV ery L ar ge
Datab ases Septem b er BGM S Berson L Golub c hik and R Mun tz Fault Toleran t Design in Multimedia Serv ers In Pr o c e e dings of
the A CM SIGMOD Internationa l Confer enc e on Management of Data BGMJ S Berson S Ghandeharizadeh R Mun tz and X Ju Staggered Striping in Multimedia Information
Systems In Pr o c e e dings of the A CM SIGMOD Internation al Confer enc e on Management of Data
BMC PBoc hec k H Meado ws and S Chang Disk Partitioning Tec hnique for Reducing Multimedia Access
Dela yIn ISMM Distribute d Systems and Multime dia Applic ation s August 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 Den P J Denning The Working Set Mo del for Program Beha vior Communic ations of the A CM Gem D J Gemmell Multimedia Net w ork File Serv ers Multic hannel Dela y Sensitiv e Data Retriev al In First
A CM Confer enc e on Multime dia pages August GHW J Gra y B Host and M W alk er P arit y striping of disc arra ys Lo wcost reliable storage with acceptable
throughput In Pr o c e e dings of the Internationa l Confer enceon V ery L ar ge Datab ases August GK S Ghandeharizad eh and S H Kim A Unied Framew ork for Placemen t of Data in Parallel Con tin uous
Media Serv ers T ec hnical Rep ort USCCS USC GKS S Ghandeharizadeh S H Kim and C Shahabi On Conguring a Single Disk Con tin uous Media Serv er
In Toapp e ar in Pr o c e e dings of the A CM SIGMETRICS GR S Ghandeharizadeh and L Ramos Con tin uous retriev al of m ultimedia data using parallelism IEEE
T r ansactions on Know le dge and Data Engine ering August R V P Rangan and H Vin Ecien t Storage Tec hniques for Digital Con tin uous Media IEEE T r ansactions
on Know le dge and Data Engine ering August R VR P Rangan H Vin and S Ramanathan Designing an OnDemand Multimeida Service IEEE Commu
nic ations Magazine July
R Wa A L N Reddy and J C Wyllie IO Issues in a Multimedia System IEEE Computer Magazine Marc h R Wb C Ruemmler and J Wilk es An In tro duction to Disk Driv e Mo deling IEEE Computer Marc h T eo TJ T eory Prop erties of Disk Ssc hedulin g Policies in Multiprogrammed Computer Systems In Pr o c
AFIPS F al l Joint Computer Conf pages 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 VR H Vin and P Rangan Designing a Multiuser HDTV Storage Serv er IEEE T r ansactions on Sele cte d
A r e as in Communic ations Jan uary YCK P S Y u MS Chen and DD Kandlur Group ed sw eeping sc heduling for DASDbased m ultimedia storage
managemen t Multime dia Systems Jan uary A FIXB vs V ARB for HP C disk
This section con tains a comparison of FIXB and V ARB for the HP disk driv e The observ ations from these tables
are almost iden tical to those of Section and not rep eated One dierence is the lo w er p ercen tage of w asted disk
space with FIXB as compared to V ARB This is attributed to the ph ysical c haracteristics of the zones on the HP
disk driv e
FIXB V ARB
N Blo c k Size Minim um Blo c k Maxim um Blo c k Av erage Blo c k
MBytes Size MBytes Size MBytes Size MBytes
T able HewlettP ac k ard C disk
FIXB V ARB
N Memory Latency w asted Avg w asted Memory Latency w asted Avg w asted
MBytes Sec disk space disk bandwidth MBytes Sec disk space disk bandwidth
T able HewlettP ac k ard C disk
B Maxim um throughput for Seagate STW disk
This app endix con tains the exp erimen tal results of Section using the Seagate disk driv e The observ ations from
T able is iden tical to those of Section
Num b er of ob jects Maximum numb er of users
in w orking set MinZtfr MaxT MaxT MaxT MaxT MaxT
WS m desir ed m desir ed m desir ed m desir ed m desir ed

T able Maxim um throughput with xed latency time sec in the Seagate STW disk

Abstract (if available)

Linked assets

Computer Science Technical Report Archive

Conceptually similar

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

USC Computer Science Technical Reports, no. 625 (1996)

PDF

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

PDF

USC Computer Science Technical Reports, no. 948 (2014)

PDF

USC Computer Science Technical Reports, no. 618 (1995)

PDF

USC Computer Science Technical Reports, no. 584 (1994)

PDF

USC Computer Science Technical Reports, no. 964 (2016)

PDF

USC Computer Science Technical Reports, no. 600 (1995)

PDF

USC Computer Science Technical Reports, no. 748 (2001)

PDF

USC Computer Science Technical Reports, no. 578 (1994)

PDF

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

PDF

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

Description

Shahram Ghandeharizadeh, Seon Ho Kim, and Cyrus Shahabi. "Continuous display of video objects using multi-zone disks." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 592 (1994).

Asset Metadata

Creator Ghandeharizaden, Shahram (author), Kim, Seon Ho (author), Shahabi, Cyrus (author)

Core Title USC Computer Science Technical Reports, no. 592 (1994)

Alternative Title Continuous display of video objects using multi-zone disks (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 28 pages (extent), technical reports (aat)

Language English

Unique identifier UC16269925

Identifier 94-592 Continuous Display of Video Objects Using Multi-Zone Disks (filename)

Legacy Identifier usc-cstr-94-592

Format 28 pages (extent),technical reports (aat)

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

Internet Media Type application/pdf

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

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

Repository Name USC Viterbi School of Engineering Department of Computer Science

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

Repository Email csdept@usc.edu

Inherited Values

Title Computer Science Technical Report Archive

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)