Computer Science Technical Report Archive

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

(USC DC Other)

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

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Online Reorganization of Data in Scalable Con tin uous Media
Serv ers
Shahram Ghandeharizadeh and Dongho Kim
Departmen t of Computer Science
Univ ersit y of Southern California
Los Angeles California April Abstract
The n um ber of sim ultaneous displa ys supp orted b y a con tin uous media serv er eg Mitra
dep ends on the n um b er of disk driv es as wellasthe amoun t of memory in the system T o supp ort
ahighern um b er of displa ys one ma y increase the n um b er of disks in the system and reorganize
the placemen t of data to incorp orate their bandwidth This study presen ts alternativ e online
reorganization tec hniques that mo dify the placemen t of data without disrupting service W e
quan tify the memory and disk storage requiremen ts of these tec hniques using analytical mo dels
In tro duction
A scalable con tin uous media serv er eg Streaming RAID TPBG Mitra GZS
supp orts the
displayof audio and video clips The n um ber of sim ultaneous displa ys supp orted b y a system is a
function of the a v ailable disk bandwidth This is b ecause con tin uous media ob jects in particular
video are large in size and almost alw a ys disk residen t F or example a t w o hour MPEG enco ded
video clip requiring Megabit p er second bandwidth Mbps for its displa y is Gigab yte in size
This researc h w as supp orted in part b y the NIMH gran t PMHA the National Science F ounda
tion under gran ts IRI IRI NYI a w ard and CD A and a HewlettP ac k ard unrestricted
cashequipmen t gift

Giv en a hardw are platform with D disks a scalable serv er w ould strip e eg staggered strip
ing BGMJ a clip across the disks This w ould enable the system to distribute the w ork imp osed
bya displayev enly across the D disks T o displayan ob ject the system w ould sc hedule the band
width of individual disks as a function of time in order to retriev e the appropriate p ortion of an
ob ject at the righ t time for displa y A service pro vider ma y accurately estimate its desired n um ber of
sim ultaneous display toenable av endor to congure a hardw are platform with an adequate n um ber
of disks Ho w ev er the requiremen ts of the service pro vider migh t increase as a function of time This
w ould require an increase in the n um b er of disks that constitute the underlying hardw are platform
Once additional disks are in tro duced the la y out of data should be reorganized based on the new
conguration This w ould enable the system to harness the bandwidth of additional disks in supp ort
of a higher n um b er of sim ultaneous displa ys
The system migh t reorganize data either as an online or an oline pro cess An oline reorga
nization pro cess consumes all system resources to reorganize the data Th us the system b ecomes
una v ailable for the duration of reorganization With an online reorganization pro cess the system
con tin ues to service user requests while reorganizing data This is ac hiev ed b y assigning a fraction
of system resources to the reorganization pro cess The primary adv an tage of online reorganization
is that it pro vides unin terrupted service to clien ts when p erformed during op eak hours
Unin terrupted service is imp ortan t in sev eral emerging application domains CDRS SSU
F or example a con tin uous media serv er mightbean in tegral comp onen t of a healthcare information
system that enables ph ysicians at dieren t hospitals to view surgery pro cedures on demand This
serv er migh t b e deplo y ed for use in emergency ro oms as a reference source Suc h a serv er mightnot
tolerate the do wntime in tro duced b y an oline reorganization pro cess
Hardw are v endors ha v e recognized the imp ortance of unin terrupted service byin tro ducing mass
storage comp onen ts that feature hotsw ap or liv einsertion capabilit y Qui IEEHw a eg
HP NetServ er Storage System P ac and IBM Hot Sw ap Hard Disk Driv es with IBM Hot Sw ap
Expansion Enclosure IBM With these systems one migh tin tro duce new hotsw appable disk
driv es without sh utting do wn the system This study in tro duces online reorganization tec hniques

X
0.0
X
4.0
...
X
0.1
X
4.1
...
X
1.0
X
5.0
...
X
1.1
X
5.1
...
X
2.0
X
6.0
...
X
2.1
X
6.1
...
C
C
C
0
1 2
X
3.0
X
7.0
...
X
3.1
X
7.1
...
C
3
Figure Simple Striping with F our Clusters
that enable a con tin uous media serv er to mo dify the placemen t of data as an online op eration to
incorp orate the bandwidth of new disks
T arget En vironmen t
Our target arc hitecture for supp orting m ultimedia applications eg MitraGZS
is hierarc hical
consisting of a tertiary storage device eg magnetooptical juk e bo x tap e driv e a group of disk
driv es and some memory GZS
GS MWS BGMJ The reason for targeting hierarc hical
storage managers is the cost of storage and the large size of audio and video ob jects A t the time of
this writing the appro ximate cost p er megab yte of memory is disk storage is and tertiary
storage is less than It is economical to stage the data at the dieren tlev els of hierarc h y in the
follo wing manner a small fraction of referenced ob jects in memory for immediate displa yan um ber
of frequen tly accessed ob jects eg the p opular mo vie titles on the disk driv es and the remaining
ob jects on the tertiary storage device
Audio and video ob jects m ust b e retriev ed and displa y ed at a presp ecied bandwidth Otherwise
their displa y migh t suer from frequen t disruptions and dela ys In this pap er w e fo cus on the displa y
of these ob jects from magnetic disks Giv en a system that consists of D disk driv es eac h with
R
D
eectiv e bandwidth and a database that consists of ob jects les that b elong to a single media
t yp e with bandwidth requiremen t R
C
w e can utilize the aggregate bandwidth of at least d d
R
C
R
D
e
disk driv es to supp ort a con tin uous displa y of an ob ject This can b e ac hiev ed b y a metho d termed
Simple Striping BGMJ First the D disk driv es in the system are partitioned in to C b
D
d
c disk
clusters Next eac h ob ject in the database sa y X is organized as a sequence of n equisized blo cks

P arameter Denition
D T otal n um b er of disks
d Num b er of disks that constitute a cluster
R
C
Bandwidth required to displa y ob jects
R
D
Eectiv e Bandwidth of a disk
R
k
Eectiv e Bandwidth of a cluster
C
or ig
Num b er of clusters with original conguration
C
new
Num b er of clusters with new conguration
X
i
Blo cki of le X
P X
i
Placementof X
i
cluster id
LC M Least Common Multiple of C
or ig
and C
new
T
R
X Time to reorganize le X
TT
R
T otal time to reorganize the le system
N
f
Num b er of les on disks in the system
N
TS
Num b er of time slots allo cated for online reorganization
l
obj
X Displaylengthofobject X
Av g l
obj
Av erage displa y length of ob jects in the system
N
displ ay
Num b er of curren t displa ys during online reorganization
N
or ig
Maxim um n um b er of displa ys with C
or ig
N
new
Maxim um n um b er of displa ys with C
new
N
laz y
N
or ig
N
displ ay
N
eag er
Num b er of streams b eing reorganized with lazy
N
eag er
N
or ig
N
displ ay
N
laz y
Num b er of streams b eing reorganized with eager
N
B r eor g
Num b er of ob ject blo c ks in C
or ig
that should b e reorganized
N
B or ig
T otal n um b er of ob ject blo c ks in C
or ig
S
C
Space required p er cycle blo c ks
S
T
T otal space required S
or ig
Original sc heduling paradigm
S
new
New sc heduling paradigm
P
or ig
Original placemen t of blo c ks
P
new
New placementof bloc ks
T able P arameters and their denition
X
0.0
X
6.0
...
X
0.1
X
6.1
...
X
1.0
X
7.0
...
X
1.1
X
7.1
...
X
2.0
X
8.0
...
X
2.1
X
8.1
...
C C C
0 12
X
3.0
X
9.0
...
X
3.1
X
9.1
...
C
3
X
4.0
X
10.0
...
X
4.1
X
10.1
...
C
4
X
5.0
X
11.0
...
X
5.1
X
11.1
...
C
5
Figure Simple Striping with Six Clusters

Time
New Clusters
Introduced
Reorganization
finished
Regular Service Off-line Reorganization Regular Service
(Display + Materialization) (Display + Materialization) (No display)
No New Services
System Evacuation
(2 hours) (20 mininutes)
System Evacuated
(Service existing requests only)
Resume Regualr Service
with higher number of displays
D = 40, d = 1, R_d = R_k = 40 (Mb/s), R_c = 4 (Mb/s)
C_orig = 40, C_new = 60, Capacity of one disk = 9 (GB)
Block size = 256 (KB), Object display length = 120 (min.)
Object Size = 3600 (MB)
Time Slot length = 0.05 (sec), Time Period length = 0.5 (sec)
Total Capacity = 360 (GB) = Total File System Size on disks
Number of Blocks per file = 14400, Number of files = 100
N_orig = 400
Time to Reorganize the file system = 20 minutes
Interruption of Service (2 hours and 20 minutes)
Figure OLine Reorganization
X
X
X
n Eac h blo c k X
i
represen ts a con tiguous p ortion of X When X is materialized
from the tertiary storage device its blo c ks are assigned to the clusters in a roundrobin manner
starting with an a v ailable cluster In a cluster a blo c k is declustered RE LKB GD in to d
pieces fr agments with eac h fragmen t assigned to a dieren t disk driv e in the cluster T o illustrate
in Figure a system consisting of eigh t disks is organized in to four clusters C D eac h
cluster consisting of t w o disks The blo c ks of X are assigned starting with C
Eac h fragmen t of a
blo c k X
i
is denoted as X
i and X
i When the system displa ys X it starts b y using the cluster
that con tains X
sa y C
i
to displa y the rst p ortion of X Next it emplo ys C
i mod C
to displa y
X
This pro cess rep eats itself un til all blo c ks of X ha v e b een retriev ed and displa y ed
T o ensure the con tin uous displa y of an ob ject the system main tains a time p erio d for eac h
cluster A time p erio d is the amoun t of time required to displaya bloc k It consists of at most b
R
k
R
C
c
time slots A time slot is the time required for a cluster to rep osition its readwrite heads to the
appropriate fragmen t and read the blo c kin to memory The duration of a time slot is dep endenton
the ph ysical c haracteristic of the secondary storage device its seek and latency time and transfer
rate and the size of the fragmen ts The fragmen t size is a parameter that migh t b e decided at the
system conguration time BGMJ The n umberofsim ultaneous displa ys supp orted bya con tin uous media serv er eg Mitra GZS

Time
New Clusters
Introduced
Reorganization
finished
Regular Service On-line Reorganization with Regular Service Regular Service
(Display + Materialization) (Display + Materialization)
w/ 0% of N_orig (200 reorganizations): 20 min
w/ 20% of N_orig (80 displays + 160 reorganizations): 25 min
w/ 50% of N_orig (200 displays + 100 reorganizations): 40 min
w/ 80% of N_orig (320 displays + 40 reorganizations): 100 min
w/ 85% of N_orig (340 displays + 30 reorganizations): 133 min
w/ 90% of N_orig (360 displays + 20 reorganizations): 200 min
* N_orig (Number of displays with original configuration) = 400
Regular Service with
higher number of displays
Figure OnLine Reorganization
dep ends on the n um b er of disk clusters as w ell as the amoun t of memory in the system CP R V TPBG R W BGMJ GK T o increase this n um ber one ma y increase the n um ber of disk
clusters in the system Ho w ev er merely adding more clusters do es not impro v e the n um ber of
sim ultaneous displa ys unless the placemen t of data across the disk clusters is reorganized to stage
data on the newly in tro duced disks F or example if t w o clusters are in tro duced in the system of
Figure then the placemen t of data should b e mo died to yield the organization of Figure This study describ es an OnLine R e or ganization tec hnique With this tec hnique a service
pro vider do es not ha v e to sh ut do wn the system ie no in terruption of service The time to
reorganize a le system dep ends on the amoun t of system resources time slot andor memory al
lo cated to the reorganization pro cess Although using time slots for online reorganization migh t
result in temp orary degradation of the n um ber of sim ultaneous displa ys This lo w er capabilit y has
no impact during the op eak hours when only a few displa ys are activ e
Figure sho ws oline reorganization of data While this pro cess requires min utes it is
preceded with a system ev acuation period during whic h no additional requests are serviced The
duration of system ev acuation dep ends on the curren t status of the activ e displa ys and ho w long the
remainder of their displa y lasts In a video on demand application where the displa y time of mo vie
is t ypically t w o hours long a displa ymightbe initiated prior to the ev acuation p erio d resulting in
a p oten tial t w o hour ev acuation p erio d
Figure sho ws online reorganization using the tec hnique prop osed in this study Assuming that
C
or ig
supp orts N
or ig
sim ultaneous displa ys while C
new
supp orts N
new
displa ys the system supp orts
a fraction of N
or ig
displa ys during the reorganization pro cess Once reorganization completes the
n um b er of sim ultaneous displa y is increased to N
new
The duration of the reorganization dep ends on
ho w m uc h resource is allo cated to online reorganization ie what fraction of N
or ig
is supp orted
F or example online reorganization w ould require appro ximately hours and min utes assuming
only of system resources is allo cated for this purp ose The remaining is used to supp ort
hiccupfree displa ys of ob jects see Figure This time can b e reduced to min utes b y allo cating
of resources to reorganization In this case only of N
or ig
is supp orted during reorganization
With of N
or ig
reorganization b ecomes a min ute pro cess
It is imp ortan t to observ e that prior to online reorganization the placemen t of data has a one
toone corresp ondence with a regular sc hedule for activ ation of clusters to retriev e data in supp ort
of a con tin uous displa y Ho w ev er during the reorganization the placemen t of data is indep enden t
of the sc hedule and migh t not b e onetoone as detailed b y Section Ano v elt y of this study is to
describ e tec hniques that ensure con tin uous displa ys in the presence of suc h scenarios
The rest of this pap er is organized as follo ws In Section w e describ e tec hniques to sc hedule
the migration of blo c ks during reorganization and also describ es alternativ e solutions of v arying
disk space and memory requiremen t to supp ort displa ying ob jects during reorganization pro cess
Section discusses the generalization of online reorganization tec hnique to the case of staggered
striping Our conclusions and future researc h directions are presen ted in Section OnLine Reorganization
Online reorganization can be p erformed in either a lazy or an e ager manner With lazy reorgani
zation the system reorganizes an ob ject when it is displa y ed With eager an ob ject is reorganized
indep endentof a displa y W e start with a complete description of lazy Subsequen tly w e state the
dierences b et w een lazy and eager

Lazy Reorganization
With lazy reorganization the system examines a new request to determine if the referenced ob ject
sa y X has b een reorganized If not the system allo cates t w o time slots on cluster C
i
that con tains
X
One slot is allo cated for reading eachbloc ks to displa y X and the other to write the blo c ks to
up date the placemen t of data
The placemen t of blo c ks of X is a function of the n um ber of clusters in the system the blo c k
n um b er and the cluster that con tains X
The original and new placemen t of a blo c k X
i
are dened
as
P
or ig
X
i
P X
i mod C
or ig
P
new
X
i
P X
i mod C
new
The second allo cated time slot remains idle ev ery time P
or ig
X
i
P
new
X
i
Section describ es
ho w these idle slots can be emplo y ed to reorganized blo c ks Assuming a Least Common Multiple
LC Mof C
or ig
and C
new
in Equation X
i
and X
i LC M
reside on the same cluster with b oth C
or ig
and C
new
due to a roundrobin assignmen t
Lazy reorganization rep eats a regular pattern as a function of time for migrating data blo c ks This
is b ecause the placemen t of data for b oth the original and the new conguration rep eats the same
pattern at regular in terv als LC M The length of the pattern ie the n um ber of time periods termed Cycle L ength is
C y cl e Leng th LC M
The extra time p erio d per cycle is attributed to our assumption that a blo c k cannot be read and
written during a time p erio dYCK GK GZS
The p erformance of reorganization migh t be limited b y the bandwidth of added clusters if the
n um b er of added clusters is lo w er than
C
or ig
This is b ecause the bandwidth required from the added
clusters is half of the original clusters the added clusters are used only to write blo c ks while the
The follo wing paragraphs detail the scenarios where t w o idle slots are una v ailable and its impact on lazy
bandwidth of the original clusters are used to b oth read and write blo c ks
The maxim um amoun t of required memory p er displa y p er cycle is C
or ig
blo c ks assuming the
system can reserv e t w o time slots on a cluster Ho w ev er three sp ecial cases migh t arise a at the
curren t time only one idle slot is a v ailable on the cluster con taining X
b eac h time p erio d consists
of only one time slot and c only one time slot is a v ailable in the en tire system Cases a and b
are iden tical b ecause they force the system to reservet w o time slots that are m clusters apart This
forces a lazy reorganization to require m additional blo c ks In case c the system cannot p erform
lazy reorganization The a v ailable slot is used to displa y the referenced ob ject
The p ossible n um b er of les that can b e reorganized using lazy at a sp ecic time p erio d can b e
computed as a function of N
or ig
and N
displ ay
N
laz y
N
or ig
N
displ ay
Where N
laz y
N
displ ay
N
displ ay
is the total n um b er of activ e displa ys When a new request arriv es
and no idle time slots are a v ailable the system ma y susp end a lazy reorganization ie reduce N
laz y
b y one to accommo date the new request A lazy reorganization is susp ended as follo ws The system
con tin ues with the displa y of the referenced ob ject X and abandons the reorganization of X freeing
up a time slot Reorganization can resume when the load decreases This is ac hiev ed b y lo cating a
displa y that references an ob ject that has either b een partially reorganized or not reorganized at all
An idle slot is assigned to start reorganization of blo c ks referenced b y the activ e displa y T o simplify the discussion assume that the system main tains t w o copies of eac h migrated blo c k
during reorganization one on the source cluster and the other on the destination cluster The disk
storage requiremen t of this approac h and other alternativ es is detailed in Section Lazy reorganization is illustrated in Figure In this Figure the original system conguration
consists of disk clusters C
or ig
and eac h cluster has time slots This system can supp ort
sim ultaneous displa ys N
or ig
The new system conguration consists of six disk clusters
C
new
after adding t w o new clusters to the system Once reorganization completes the new

4
5
6
7
8
9
10
11
12
0
1
2
3
Old Configuration Added Clusters
7
6
8
9
0
1
2
3
4
5
10
11
4
4,5
5,6
6,7
6,7,8
7,8,9
8,9,10
9,10,11
11,12
New Configuration
Memory
Time
Time Slot Cluster
Time Period
13
14
15
16
17
18
19
20
21
22
23
24
25
16
17
22
23
18
20
21
: Original Blocks
: New Blocks
: Reorganization Required
C0 C1 C2 C3
Cycle Length
with lazy
Cycle Length
with eager
: Type 1 (Read)
: Type 2 (Read)
: Type 3 (Write)
19
Figure Reorganization of Data Placemen t
conguration supp ort sim ultaneous displa ys N
new
Figure sho ws the activit y of one lazy
reorganization as a function of time The regular pattern of blo c k retriev als for an ob ject can be
depicted as t yp es and blo c k retriev als T yp e blo c ks are those whose placemen t remains
unc hanged with the in tro duction of new clusters T yp e and corresp ond to those blo c ks that m ust
migrate T yp e corresp onds to the retriev al of these blo c ks from their original clusters while corresp onds to their writing to their new destination During eac h time p erio d t yp e orbloc ks
are read at the righ t time to ensure a con tin uous displa y while t yp e blo c ks are written to result in
a new placemen t of data The LC M of this example is hence the cycle length is time p erio ds
long The left column of Figure depicts the memory requiremen t of the system for the example
stream bysho wing the blo c ks that should b e memory residen t p er time p erio d
Eager Reorganization
If the system utilizes lazy reorganization only the time to reorganize the en tire le system migh t
b ecome unpredictable b ecause the infrequen tly accessed les ie those with a lo w he at migh tnot
b e referenced for a long time Also the absolute minim um time to reorganize the en tire le system
is determined b y the displa y time of the longest ob ject Moreo v er the system migh t not utilize the
bandwidth of idle disk driv es when the system load is lo w
Eager reorganization mo dies the placemen t of data indep endentof a displa y With eager the
system reorganizes a le b y emplo ying t w o time slots per time period in a regular pattern Eager
retriev es few er blo c ks than lazy b ecause it do es not need to read those blo c ks whose assignmentis
iden tical with b oth C
or ig
and C
new
t yp e blo c ks in Figure whose total coun t is
LC M C
or ig
LC M
N
B or ig
This reduces the cycle length b y C
or ig
C y cl e Leng th LC M C
or ig

Figure sho ws the cycle length with eager Note that case c of Section ie only one time slot
is a v ailable in the en tire system can no w b e handled b y retrieving blo c ks during a cycle and ushing
them during the next cycle The time to reorganize a le with eager reorganization is shorter b ecause
only part of the le is retriev ed
Eager can be optimized to reorganize a le in a parallel manner b y reserving m ultiple pairs of
time slots and p erforming sev eral sequen tial cycles during one cycle F or example blo c ks to and
blo c ks to in Figure can b e reorganized concurren tly b y utilizing t w o pairs of time slots The
time to reorganize le X termed T
R
X and the en tire le system termed TT
R
with eager are
T
R
X
LC M C
or ig
LC M
l
obj
X N
TS
TT
R
LC M C
or ig
LC M
Av g l
obj
N
f
N
TS
The n um b er of time slots allo cated for reorganization termed N
TS
is divided b yt w o b ecause a pair

of time slots are required to reorganize one stream
A system should emplo y b oth eager and lazy reorganization Lazy is used to minimize the amoun t
of w asted disk bandwidth b y reorganizing the blo c ks that are staged in memory b y a displa y Eager
a v oids the system from w asting the idle disk bandwidth b y emplo ying it to reorganize les that are
not referenced b y a displa y The p ossible n um b er of les that can b e reorganized in an eager manner
at a sp ecic momen t can b e computed as a function of N
or ig
N
displ ay
and N
laz y
N
eag er
N
or ig
N
displ ay
N
laz y
N
laz y
can also b e computed as a function of eager
N
laz y
N
or ig
N
displ ay
N
eag er
Adding Preloaded Clusters
The p erformance of reorganization can b e enhanced b y adding preloaded new clusters with the blo c ks
that w ould b e assigned to them during reorganization No w the system is required to reorganize the
blo c ks among the original clusters only Hence the reorganization is not limited b y the aggregate
bandwidth of the added clusters b ecause they do not participate in reorganization
The new clusters can b e preloaded in v arietyof w a ys p ossibly from a a tertiary device b a
remote serv er on the net w ork and c during reorganization using the idle slots of lazy and eager
see Sections and The last option is realized as follo ws A t the b eginning of reorganization
the system constructs a list of the blo c ks that should be assigned to the new clusters Whenev er
these blo c ks are rendered memory residen t b y a displa y they can be ushed to the new clusters
Alternativ elyev ery time the system lo cates idle slots based on the regular cycles of lazy and eager
it can read the blo c ks on the list and ush them to the new clusters Ev ery time a blo c k is ushed
it is remo v ed from the list The preloading of new clusters is complete when the list is exhausted
With preloaded clusters the memory requiremen t of online reorganization is reduced one blo c k

4
5
6
7
8
9
10
11
12
0
1
2
3
Old Configuration Added Clusters
7
6
8
9
0
1
2
3
4 5
10 11
4
5
6
6,7
6,7,8
7,8,9
8,9,10
9,11
New Configuration
Memory
Time
Time Period
13
14
15
16
17
18
19
20
21
22
23
16 17
22 23
18
19
20
21
New Cycle
with Eager
New Cycle
with Lazy
(Original Blocks, Reorganization Not Required) - Read
(New Blocks) - Write
: Type 1
: Type 2
: Type 3
(Original Blocks, Reorganization Required) - Read
Figure OnLine Reorganization with Preloading of Added Clusters
7 89
Old Configuration Added Clusters
4 5
10 11
New Configuration
Time
Time Slot Cluster
Time Period
: Read
: Write
7 689
20 21 18 19
19 18 20 21
Optimized
Cycle Length
6
16 17
22 23
Figure Optimization of OnLine Reorganization with Preloading Added Clusters

The cycle length for lazy remains almost the same except for the last blo c k p er cycle that is no longer
required to migrate
New C y cl e Leng th LC M
In the case of eager its cycle length is further reduced
New Cycle Length LC M C
new
Accordingly the time to reorganize with eager is reduced New T
R
X and TT
R
can be obtained
b y substituting C
or ig
with C
new
in both Equations and Figure illustrates an example of
reorganization pro cess with preloaded clusters The n um ber of blo c ks that require migration is
reduced compared with that in Figure The new cycle lengths are also sho wn
The p erformance of eager reorganization can further b e optimized b y grouping the reading and
writing of those blo c ks that are reorganized per cycle With this metho d eager no longer w astes
disk bandwidth and requires no extra memory This is b ecause all blo c ks are written immediately
during the next time p erio d instead of b eing staged in memory once they are read Finally the time
to reorganize will b e signican tly reduced
N
Br eor g
N
Bor ig
C
or ig
C
new
C
or ig
LC M
T ime to R eor g aniz e N
Br eor g
N
TS
The n um ber of time slots used for reorganization N
TS
m ust be a m ultiple of C
or ig
eg or in case of Figure This sc heme is b est illustrated using an example In Figure shaded blo c ks and
are read in one time p erio d using time slots one on ev ery cluster and ushed during the next
time p erio d In Figure only blo c ks t yp e blo c ks N
B r eor g
out of N
Bor ig
should b e
reorganized in a cycle In Figure those blo c ks are read together in one time p erio d and written
during the next time p erio d In this case the resulting optimized cycle length is t w o time periods

long
As an example lets use the parameters in Figure again If w e use all of time slots
to perform reorganization the time to reorganize the le system without optimization will be min utes
It requires only min utes to reorganize the en tire le system with this optimization
with no extra memory requiremen t
Storage and Memory Requiremen t
In a scalable con tin uous media serv er that emplo ys round robin data placemen t on its disk clusters
the system has to follo w only one sc heduling paradigm at a giv en time CP R V TPBG BGMJ GK GKSZ During the reorganization pro cess t w o dieren t data placemen ts P
or ig
and P
new
for the dieren t les will co exist in the system A regular sc heduling paradigm exp ects a
sp ecic placemen t of data to supp ort a hiccupfree displa y ie S
or ig
requires P
or ig
and S
new
requires
P
new
If the migrated blo c ks are deleted new tec hniques should be devised to supp ort regular
displa ys during reorganization
There can b e v e alternativew a ys of displa ying ob jects during reorganization dep ending on whic h
blo c ks are main tained on the original clusters These are as follo ws The rst main tains original
blo c ks un til the system completes the reorganization of the en tire le system It emplo ys the original
sc heduling with the original placemen t of data to displa y ob jects during reorganization S
or ig
and
P
or ig
This paradigm w as assumed in the discussions of Sections and The second
deletes all the migrated blo c ks incremen tally as so on as they are ushed to their destination cluster
This approac h main tains a single cop y of the blo c ks in the en tire system This forces the system to
emplo y the new placemen t of data to displa y those requests that reference reorganized ob jects during
the reorganization period These t w o strategies result in dieren t tradeos bet w een the amoun t of
disk storage and memory required during reorganization
Optimized cycle length could b e dieren t with dieren t congurations
Used expression after substitute C or ig with C new Used expression
Main taining the Original Blo c ks
This approac h deletes original blo c ks once the reorganization completes T o displayan ob ject the
system emplo ys the original placemen t of blo c ks P
or ig
using the sc heduling paradigm based on C
or ig
S
or ig
un til reorganization is completely nished see Figure This p olicy is simple to implemen t
but the system m ust main tain m ultiple copies of migrated blo c ks on the original clusters This
migh t force the system to reference the tertiary storage device more frequen tly during reorganization
b ecause few er ob jects are rendered disk residen t
The disk space o v erhead to main tain original blo c ks is a function of C
or ig
and C
new
It is equal
to the total n um ber of blo c ks that constitute the ob jects subtracted b y the n um ber of blo c ks that
are written on the added clusters C
or ig
blo c ks should b e further subtracted from the ab o v e b ecause
the rst C
or ig
blo c ks in a cycle are not required to b e rewritten on the original clusters The n um ber
of blo c ks p er cycle is
S
C
C
or ig
C
new
LC M C
or ig
The p ercen tage of space required from the original clusters is
S
T
S
C
LC M
C
or ig
C
new
LC M

T able sho ws the p ercen tage of space required from C
or ig
for a v arietyof c hoice of v alues for C
or ig
and C
new
When the n um ber of added clusters C
new
C
or ig
is relativ ely small as compared with
C
or ig
the p ercen tage of space requirementfrom C
or ig
increases signican tly If C
new
is a m ultiple of
LC M no extra space will b e required
Deleting P art of Original Blo c ks
With this paradigm the system main tains blo c ks that are migrated within the original clusters only The blo c ks that ha v e b een migrated to the new clusters are deleted dynamically once they are ushed
None of the blo c ks are migrated within the original clusters

C
or ig
C
new
Space Required from C
or ig

T able Example Space Requiremen ts for Displayb y Main taining Original Blo c ks
0
1
2
3
Old Configuration
New configuration
: Original blocks (Retrieved)
: New blocks (Not utilized)
7
6
8
9
4
5
10
11
0
1
2
3
0
1
2
3
4
5
6
4
5
6
7
8
9
10
11
8
9
7
10
11
Read Display
0
1
2
3
4
5
6
7
8
9
10
11
6
7
8
9
Memory
4
5
10
11
Figure Displa y with Main taining Original Blo c ks
Time
New clusters
Introduced
Reorganization
finished
Regular Service On-line Reorganization Regular Service
Follow Original Scheduling Follow New Scheduling
(Display + Materialization) (Display + Materialization) Display with Original Scheduling Paradigm
Old blocks deleted
Regular Service with Increased Number of Displays
Figure Status of the System for Main taining Original Blo c ks

C
or ig
C
new
Space Required from C
or ig

T able Example Space Requiremen ts for Displayb y Deleting P art of Original Blo c ks
In this case the system emplo ys S
or ig
to displa y ob jects Ho w ev er it retriev es b oth the original and
the new blo c ks dynamically P
dy namic
As demonstrated in Figure the system tries to retriev e
original blo c ks but it retriev es new blo c ks from the new clusters whenev er the required blo c ks are
not found on the original clusters
This approac h reduces the disk space requiremen t of the paradigm describ ed in Section Ho w ev er it w astes bandwidth of original clusters b ecause it reserv es time slots on the original
clusters ev en when it retriev es blo c ks from the new clusters Moreo v er N
displ ay
is no w limited b y
the bandwidth of the new clusters b ecause the new clusters participate in the retriev al of data If
all the activ e displa ys are retrieving already reorganized ob ject blo c ks the system cannot supp ort
more than N
new
N
or ig
displa ys during reorganization On the other hand the system can supp ort
N
or ig
displa ys if all the activ e displa ys are retrieving original blo c ks only The amoun t of required space is equal to that of main taining all of original blo c ks subtracted b y
the n um b er of blo c ks on added clusters
S
C
C
or ig
C
new
LC M C
or ig
C
new
C
or ig
C
new
LC M
S
T
S
C
LC M
C
or ig
C
or ig
C
new

Example space requiremen ts for dieren t congurations are sho wn in T able Negativ e n um b ers
denote an increase in the space storage capacit y of the original conguration When C
new
C
or ig
C
or ig
no additional disk space will b e required on the original clusters

0
1
2
3
Old Configuration
New configuration
: Original Blocks
: New Blocks
7
6
8
9
4
5
10
11
R D
0
1
2
3
4
5
6
7
8
9
10
11
6
7
8
9
Memory
C0 C1 C2 C3
: Retrieved Blocks
0
1
2
3
4
5
6
7
8
9
10
11
0
1
2
3
4
5
6
7
8
9
10
11
Figure Displa y with Deleting P art of Original Blo c ks
Deleting Original Blo c ks Dynamically
With this paradigm the migrated blo c ks are deleted from their source once they are ushed to their
destination cluster F or example in Figure the system deletes blo c k X
from C
once it is ushed
to cluster C
The impact of this on the system is as follo ws the system m ust b e able to displa y b oth those
ob jects that are reorganized and those ob jects whose assignmen t has not b een mo died This can
be ac hiev ed in t w o p ossible w a ys either emplo y the roundrobin sc heduling of retriev al based on C
or ig
termed S
or ig
or the one based on C
new
termed S
new
The tec hnique prop osed here is a h ybrid
of the t w o the system emplo ys S
or ig
during the early stages of reorganization and switc hes to S
new
during one time p erio d in the middle of reorganization This is because the displayof reorganized
ob jects with S
or ig
requires prefetc hing of blo c ks that increases the amoun t of memory required bya
displa y During the later parts of reorganization when man y les ha v e b een reorganized the system
minimizes the o v erall memory requiremen ts of the activ e displa ys b y switc hing to S
new
see Figure This raises the follo wing issues
Ho w do es the system displa y a reorganized ob ject P
new
X using S
or ig

0
1
2
3
Old Configuration
New configuration
: Original Block
: New Block
7
6
8
9
4
5
10
11
1
2
3
4
1
2
3
4
5
6
7 5
6
7
8
9
10
11
9
RR D
0
1
2
3
4,6
5,6,7
6,7,8
7,8,9
8,9
9
10
11
Memory
C0 C1 C2 C3
00
6
Figure Displa y with Deleting Original Blo c ks Dynamically S
or ig
P
new
Time
Beginning of Reorganization
End of Reorganization
Switch
from
Old Scheduling with New Placement
to
New Scheduling with Old Placement
Total file system
Reorganized Space
Unreorganized Space
Uses memory
Uses memory Does not use memory
Does not use memory
(Original Blocks)
(New Blocks)
Figure Memory requiremen ts of S
or ig
P
new
and S
new
P
or ig

Ho w do es the system displa y an ob ject that is not y et reorganized P
or ig
X using S
new
Ho w do es the system switc h from S
or ig
to S
new
without disrupting the curren t displa ys
Displa y of P
new
X using S
or ig
A request referencing a reorganized ob ject X reserv es a time slot p er time p erio d on the original
cluster starting with the cluster that con tains X
Acon tin uous displa y is guaran teed b y prefetc hing
blo c ks The system retriev es the necessary blo c ks from the new clusters on demand without reserving
time slots on these clusters In the follo wing w e describ e howtoa v oid the p ossibilit y of collisions in
the new clusters Once S
or ig
retriev es a blo ckfrom C
or ig
recall that clusters are n um b ered from
to C
or ig
the system do es not read blo c ks from C
or ig
for the next b
C new C
or ig
C
or ig
c C
or ig
time
periods Ho w ev er the reserv ed time slot remains reserv ed ie it visits clusters in a roundrobin
manner If b
C new C
or ig
C
or ig
c C
or ig
is zero then the system con tin ues to retriev e blo c ks from C
or ig
using
S
or ig
During certain time p erio ds the system retriev es blo c ks from b oth the original clusters and new
clusters The o v erall eect is that the system prefetc hes the blo c ks from C
or ig
in to memory for
future use Ev ery time the n um ber of prefetc hed blo c ks equals d
C
or ig
C new C
or ig
e the system susp ends
the retriev al of data on behalf of this displa y from C
or ig
for C
or ig
time periods when the displa y
attempts to retriev e a blo c k from C
The retriev al of data is skipp ed in this manner due to the
constrained placemen t of data and the use of S
or ig
T o illustrate this paradigm consider the displa y of Figure The displa y is using S
or ig
and
once it visits C
to retriev e X
it skips no time p erio ds b ecause b
C new C
or ig
C
or ig
c is zero Hence in the
next t w o time p erio ds it prefetc hes blo c ks six and sev en in to memory F rom no w the displa yis t w o
blo c ks C
or ig
b ehind the retriev al Once the system attempts to retriev e a blo c k from C
its blo c k
is retriev al is susp ended b y C
or ig
time p erio ds Hence the n um b er of prefetc hed blo c ks p er displa y
is d
C
or ig
C new C
or ig
e blo c ks
Displa y of P
or ig
X using S
new

T o supp ort a con tin uous displa y the system m ust retriev e P
or ig
X only from the original clusters
while it follo ws S
new
A request referencing an unreorganized ob ject X reserv es m ultiple time slots
p er time p erio d on the original clusters starting with the cluster that con tains X
to prefetchbloc ks
By utilizing m ultiple time slots the system can p erform lazy reorganization if the le has not b een
already reorganized
The required n um ber of time slots that m ust be reserv ed on C
or ig
can be computed as follo ws
The n um ber of time periods a v ailable on C
or ig
during a cycle for one display LC M time periods is equal to
C
or ig
C new
LC M During this p erio d the n um b er of blo c ks to b e read from C
or ig
equals to
LC M and the n um b er of blo c ks to b e written on C
or ig
equals to
C
or ig
C new
LC M C
or ig
Hence the
required n um b er of time slots on C
or ig
is
N
TSreq
LC M C
or ig
C new
LC M C
or ig
C
or ig
C new
LC M
C
new
C
or ig
C
new
LC M
During C
or ig
time p erio ds reserv e time slots up to N
TSreq
p er cluster and prefetc h blo c ks so that
at the end of C
or ig
time p erio ds C
new
C
or ig
N
B pr ef etch
C
new
Where N
B pr ef etch
is the n um ber of prefetc hed blo c ks in memory at the end of C
or ig
time p erio ds
Flush the prefetc hed blo c ks to the new lo cation during the time p erio d that they are displa y ed
If X
resides other than C
on C
i
there will b e a startup latency C
new
i time periods This case ma y be handled as follo ws During the rst C
or ig
i time p erio ds b efore the displa y
starts the system reads only one blo ckduring eac h time p erio d to minimize memory requiremen t
If the system ushes new blo c ks during TP
displ ay
X
j
i
the original lo cation of X
w ould be
preserv ed New X
will reside on C
ifw e ush new blo c ks when they are displa y ed
Figure sho ws the displa y and migration sc hedule for S
new
P
or ig
According to Equation the
required n um b er of time slots p er cluster equals t w o By the algorithm describ ed ab o v e the system
prefetc hes blo c ks and during C
or ig
time p erio ds so that the n um ber of prefetc hed blo c ks
is larger than or equal to the n um b er of added clusters but smaller than the n um ber of C
new
After
TP
display
X j is the time p erio d that X j is displa y ed

0
1
2
3
Old Configuration
New configuration
: Old Blocks
: New Blocks
1
2
3
6
7
1
2
3
4
5
6
7
8
9
10
11
R R D
0,4
1,4,5
2,4,5,6
3,4,5,6,7
4,5,6,7
5,6,7
6,7,8
7,8,9
8,9,10
9,10,11
10,11
11
Memory
4
5
10
11
6
7
6
7
8
9
8
9
8
9
10
11
W
7
8
9
0
6
0
4
5
10
11
4
5
10
11
4
5
Figure Displa y and Reorganize with Deleting Original Blo c ks Dynamically S
new
P
or ig
prefetc hed blo c ks to are ushed to the new lo cation when they are displa y ed
Switc h from S
or ig
to S
new
The switc hing from S
or ig
to S
new
is p ossible only when enough time slots are a v ailable in the
system b ecause S
new
P
or ig
requires m ultiple time slots per stream It also requires LC M time
periods to prepare switc hing if the cycle of eac h stream do es not sync hronize with the switc hing
p oin t Figure illustrates the switc hing pro cess Once the system decides to switchthe sc heduling
p olicy at time t all of the curren tly activ e reorganization streams can pause at the end of eac h cycle
that is nearest to t at a and resume after t from b On the other hand all of the curren tly
activ e displa ys that retriev e original blo c ks should start to prefetc h the blo c ks that are required
after switc hing from time t LC M during A to supp ort hiccupfree displa y Righ t after switc hing
at t the displa y streams will use the buered blo c ks that are required to follo w the new sc hedule
un til the end of the cycle that is o v erlapp ed with t during B T radeos
Compared with the tec hniques in Sections and this sc heme do es not require anydisk
space on original clusters In fact there will b e left o v er space b ecause some of the blo c ks migrate
to the added clusters
This approac h has sev eral undesirable asp ects It requires the complex sc heduling algorithms

Time
New clusters
Introduced
Reorganization
finished
Switch from
Original Scheduling
to
New Scheduling
Regular Service On-line Reorganization
with Regular Service
Regular Service
Follow Original Scheduling Follow New Scheduling
(Display + Materialization) (Display + Materialization)
Regular Service with
Increased Number of Displays
Figure Status of the system for Deleting All of Original Blo c ks
During A: Pre-fetch blocks for B
During B: Consume pre-fetched blocks
t
t - LCM
t + LCM
a
b
a
b
A
B
A
B
S
S
orig
new
Time
(Time Periods)
Figure Switc hing from S
or ig
to
new

describ ed ab o v e additional temp orary memory and additional time slots to supp ort displa y with
S
new
P
or ig
Also it is necessary to ha v e the switc hing mec hanism b et w een t w o dieren tsc heduling
paradigms during the reorganization to reduce temp orary memory requiremen t In addition the
n um ber of displa ys supp orted b y the system migh t be limited b y the bandwidth of added clusters
b ecause some blo c ks should b e retriev ed from added clusters in the rst case S
or ig
P
new

Hybrid Solutions
W e can reduce temp orary memory andor disk space requiremen tb y using h ybrid metho ds of main
taining and deleting original blo c ks
Hybrid
This is a h ybrid of tec hniques discussed Sections and W e can main tain all of the
original blo c ks for some frequen tly accessed les and delete the original blo c ks dynamically for others
This metho d requires less memory than that of deleting all of original blo c ks dynamically and requires
less disk space than that of main taining all of original blo c ks Ho w ev er it still requires complex
sc heduling and N
displ ay
mightbelimited b y the bandwidth of added clusters
Hybrid
This is a h ybrid of tec hniques discussed Sections and The system can main tain all
of the original blo c ks for some frequen tly accessed les and delete part of the original blo c ks for
others This metho d requires no additional memory to displa y and requires less disk space than that
of main taining all of original blo c ks Ho w ev er N
displ ay
migh t b e limited b y the bandwidth of added
clusters
Staggered Striping
The online reorganization tec hnique discussed in this study is presen ted based on a sp ecial case of
staggered striping BGMJ Ho w ev er the tec hnique can be generalized with staggered striping

without an y fundamen tal c hange With staggered striping either eac h logical cluster size of an ob
ject or the n umberof diskdriv es can aect the cycle length but only the n um b er of disks aects the
n um b er of sim ultaneous displa ys supp orted bythe system As the n um b er of disk clusters is deter
mined bythe n um b er of disk driv es in simple striping the cycle length for reorganization of staggered
striping is also determined b y the n um ber of ph ysical disk driv es ie Least Common Multiple of
the n um b ers of disks with the original and new congurations The sc hedule of reorganization can
also b e generalized b y just emplo ying the sc heduling paradigm of staggered striping
In staggered striping w e migh t not be able to read and write a whole blo c k if k d
This
means that a fraction of a blo c k d k
d
migh t be staged in memory whereas a blo c k is the unit of
memory requiremen t in case of simple striping
Conclusion and F uture Researc h Directions
This study presen ts alternativ e online reorganization tec hniques that incorp orate the bandwidth
of additional disks without disrupting service Eager online reorganization can be parallelized b y
taking adv an tage of the mo dularityin the reorganization pro cess W e quan tied b oth the memory
and disk storage requiremen ts of these tec hniques using analytical mo dels
This study can b e extended in sev eral w a ys First w e assumed an en vironmen t where additional
disks are iden tical to those that already exist in the system a homogeneous en vironmen t This
assumption migh t be violated in a real w orld sa vings where the new disks are the next generation
to those that are curren tly in use a heterogeneous en vironmen t In this case w e need to revisit
b oth the displa y and the reorganization tec hniques Second at times a service pro vider mightw an t
to reduce the cost of a system due to a shrinking demand In this case a n um b er of disks mightbe
ph ysically remo v ed Reorganization of data in suc h cases without disrupting service b ecomes c hal
lenging and requires further in v estigation Finallyw ein tend to implemen t the prop osed tec hniques
using Mitra GZS
k is the v alue of stride in staggered striping

References
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 International Confer enceon
Management of Data CDRS M Carey D DeWitt J Ric hardson and E Shekita Storage managemen t for ob jects in
EXODUS In W Kim and F Lo c ho vsky editors Obje ctOriente d Conc epts Datab ases
and Applic ations pages Addison W esley
CP P M Chen and D P aterson A new approac h to IO p erformance ev aluation self
scaling IO benc hmarks predicted IO p erformance In Pr o c e e dings of the A CM
SIGMETRICS Intl Conf on Me asur ement and Mo deling of Computer Systems Ma y
GD S Ghandeharizadeh and D DeWitt A m ultiuser p erformance analysis of alternativ e
declustering strategies In Pr o c e e dings of International Confer enc e on Datab ase Engi
ne ering GK S Ghandeharizadeh and S H Kim Striping in Multidisk Video Serv ers In SPIE
International Symp osium on Photonics T e chnolo gies and Systems for V oic e Vide o and
Data Communic ations Octob er GKSZ S Ghandeharizadeh S H Kim C Shahabi and R Zimmermann PlacementofCon tin
uous Media in MultiZone Disks In S Ch ung editor Multime dia Information Stor age
and ManagementKlu w er GS S Ghandeharizadeh and C Shahabi Managemen t of Ph ysical Replicas in P arallel Mul
timedia Information Systems 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 GZS
S Ghandeharizadeh R Zimmermann W Shi R Rejaie D Ierardi and T Li Mitra
A Scalable Con tin uous Media Serv er Submitte d to VLDB Hw a Kai Hw ang Advanc e d Computer Ar chite ctur e Par al lelism Sc alability Pr o gr ammability pages McGra wHill Inc IBM IBM D ASD Hot Sw ap Expansion Enclosure and Hot Sw ap Hard Disk Driv es August
Pro duct Catalog of Options IBM
IEE IEEE IEEE Standard for Futurebus Ph ysical La y er and Prole Sp ecication IEEE
Std LKB M Livn y S Khoshaan and H Boral MultiDisk Managemen t Algorithms In Pr o
c e e dings of the A CM SIGMETRICS Intl Conf on Me asur ement and Mo deling of
Computer SystemsMa y
MWS D Maier J W alp ole and R Staehli Storage System Arc hitectures for Con tin uous Media
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 P ac Hewlett P ac k ard HP In tro duces The HP NetServ er Storage System Septem ber HP Press Release
Qui Ric hard A Quinnell Liv e Insertion Do it without Killing Your System EDN magazine F ebruary RE D Ries and R Epstein Ev aluation of distribution criteria for distributed database
systems UCBERL T ec hnical Rep ort M UC Berk eleyMa y
R V P Rangan and H Vin E!cien t Storage Tec hniques for Digital Con tin uous Media IEEE
T r ansactions on Know le dge and Data Engine ering August
R W A L N Reddy and J C Wyllie IO Issues in a Multimedia System IEEE Computer
Magazine Marc h
SSU A Silb ersc hatz M Stonebrak er and J Ullmen Database systems Ac hiev emen ts and
opp ortunities Communic ations of the A CM Octob er
TPBG FA T obagi J P ang R Baird and M Gang Streaming RAIDA Disk Arra y Manage
men t System for Video Files In First A CM Confer enc e on Multime dia August
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

Linked assets

Computer Science Technical Report Archive

Conceptually similar

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

PDF

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

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. 578 (1994)

PDF

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

PDF

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

Description

Shahram Ghandeharizadeh and Dongho Kim. "On-line reorganization of data in scalable continuous media servers." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 634 (1996).

Asset Metadata

Creator Ghandeharizadeh, Shahram (author), Kim, Dongho (author)

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

Alternative Title On-line reorganization of data in scalable continuous media servers (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 UC16270812

Identifier 96-634 On-line Reorganization of Data in Scalable Continuous Media Servers (filename)

Legacy Identifier usc-cstr-96-634

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

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

Coverage Temporal 1991/2017

Repository Email csdept@usc.edu

Repository Name USC Viterbi School of Engineering Department of Computer Science

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

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