Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
Computer Science Technical Report Archive
/
USC Computer Science Technical Reports, no. 627 (1996)
(USC DC Other)
USC Computer Science Technical Reports, no. 627 (1996)
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
On Minimizing Startup Latency in Scalable Con tin uous Media
Serv ers
Shahram Ghandeharizadeh Seon Ho Kim W eifeng Shi Roger Zimmermann
Departmen t of Computer Science
Univ ersit y of Southern California
Los Angeles California Marc h Abstract
In a scalable serv er that supp orts the retriev al and displayof con tin uous media audio and
video clips b oth the n um b er of sim ultaneous displa ys and the exp ected startup latency of a dis
pla y increases as a function of additional disk bandwidth This pap er describ es ob ject replication
and request migration as t w o alternativ e tec hniques to minimize startup latency In addition to
dev eloping analytical mo dels for these t wotec hniques w e rep ort on their implemen tation using
a scalable serv er The results obtained from b oth the analytical mo dels and the exp erimen tal
system demonstrate the eectiv eness of the prop osed tec hniques
In tro duction
During the past decade the information tec hnology has ev olv ed to where it is economically viable
to store and retrievecon tin uous media suc h as audio and video clips Con tin uous media serv ers are
exp ected to supp ort h undreds of sim ultaneous displa ys and pla y a ma jor role in div erse applications
suc h as those en visioned bythe en tertainmen t industry educational applications library and health
care information systems to name a few The no v elt y of these systems is attributed to t w o require
men ts of con tin uous media that are dieren t from traditional textual and recordbased data First
the retriev al and displayof con tin uous media are sub ject to realtime constrain ts that impact b oth
a the storage sc heduling and deliv ery of data and b the manner in whic h m ultiple users ma y
share resources If the realtime constrain ts are not satised then a displa y migh t suer from dis
ruptions and dela ys that result in jitter with video and random noises with audio These disruptions
This researchw as supp orted in part b y the National Science F oundation under gran ts IRI IRI
NYI a w ard and CD A and a HewlettP ac k ard unrestricted cashequipmen t gift
and dela ys are termed hic cups Second ob jects of this media t yp e are t ypically large in size F or
example a t w o hour MPEG enco ded video requiring Megabits p er second Mbps for its displa y
is Gigab yte in size Three min utes of uncompressed CD qualit y audio with a Mbps bandwidth
requiremen t is Megab yte MByte in size Magnetic disks ha v e established themselv es as the
mass storage of c hoice for these media t yp es
In scalable serv ers the n um ber of sim ultaneous displa ys supp orted b y the system increases as a
function of a v ailable disk bandwidth desirable Ho w ev er the exp ected startup latency observ ed b y
the users increases as w ell undesirable This is due to the constrained placemen t of data across the
disks In this study w e dev elop tec hniques to minimize the exp ected startup latency of a scalable
serv er Minimizing the a v erage startup latency migh t b e desirable for sev eral reasons One ob vious
reasons is to meet the w aiting tolerance of clien ts A less ob vious reasons is to implemen t com
plex features As an example consider the fastforw ard or fastrewind V CR functionalit y Sev eral
studies ha v e prop osed to implemen t this functionalit y b y main taining a fastforw ard fastrewind
v ersion of a clip BGMJ ORS HewlettP ac k ard emplo ys this tec hnique in its commercial pro d
uct And An index main tains the relationship b et w een the dieren t p ortions of a clip X X
displ ay
and its corresp onding fastforw ard v ersion X
ff
When the user references a video clip X the sys
tem retriev es X
displ ay
at the bandwidth presp ecied b y its media t yp e to supp ort a hiccup free
displayof X When the user requests a fastforw ard displa y the system indexes to the appropriate
lo cation in X
ff
and initiates the retriev al and displa y of data from this le This switc h is iden ti
cal to terminating the curren t request for X
displ ay
and issuing a new one for X
ff
When the user
switc hes bac k to normal displa y the system once again is forced to terminate the curren t retriev al of
data from X
ff
and issue a new request to the appropriate lo cation in X
displ ay
using the index By
minimizing the exp ected startup latency the system minimizes the a v erage dela y observ ed b y users
prior to b oth the fastforw ard request and resume to normal pla y This study presen ts t w o alternativ e tec hniques to minimize the exp ected startup latency of the
system request migration and data replication W e dev elop analytical mo dels to quan tify the
tradeos asso ciated with these t w o tec hniques Both the prop osed tec hniques and their mo dels are
no v el The dev elop ed analytic mo dels dier from those prop osed for either parallel le systems or
database managemen t systems MY WYS in that they assume a deterministic usage of disk
bandwidth for a displa y and a displa ymigh t b e activ e for a long time p oten tially for t w o hours
for a mo vie In addition w e describ e and ev aluate an implemen tation of the prop osed tec hniques
using a scalable system Mitra GZS
The analytical and exp erimen tal results demonstrate
that these tec hniques reduce the a v erage startup latency In addition they sho w that replication is
sup erior to request migration
T erm Denition
T
W S eek
W orst seek time with maxim um rotational latency
m Maxim um n um b er of sim ultaneous displa ys in a system
R
C
Displa y bandwidth requiremen t Displa y rate
R
D
T ransfer rate of a single disk driv e
B Size of a blo c k
G
i
Group i
T
p
Time to displa y a blo c k
N Maxim um n um b er of sim ultaneous displa ys in a cluster
L Latency
D T otal n umberof diskdriv es in a system
d Num b er of disk driv es in a cluster
C Num b er of clusters in a system
k Num ber of activ e requests busy serv ers in a system
i Num b er of failures that a request migh t observ e b efore a success
Av erage arriv al rate requests p er second
s Av erage service time of a request seconds
T able List of terms used rep eatedly in this pap er and their resp ectiv e denitions
The rest of this pap er is organized as follo ws Section describ es a paradigm for con tin uous
displa y of audio and video ob jects It separates the bandwidth of disks from their storage capacityb y
in tro ducing the concept of groups Using this abstraction Section describ es analytical mo dels based
on queuing theory to estimate the exp ected startup latency observ ed b y a request Subsequen tly Section describ es request migration and replication as t w o dieren t tec hniques that minimize
the exp ected startup latency W e quan tify the tradeos asso ciated with these t w o alternativ es in
Section Brief conclusions are oered in Section Con tin uous Displa y
T o supp ort con tin uous displayof an object X it is partitioned in to n equisized blo c ks X
X
X
n where n is a function of the blo c k size B and the size of X A time p erio d T
p
is dened
as the time required to displa y a blo c k
T
p
B
R
C
With a m ultidisk platform consisting of D disks the w orkload of a displa y should be ev enly dis
tributed across the D disks in order to a v oid formation of b ottlenec ks Striping BGMJ GK is a tec hnique to accomplish this ob jectiv e This tec hnique partitions the D disks in to C clusters
of disks with eac h cluster consisting of d disks C b
D
d
c Next it assigns the blo c ks of ob ject X
to the clusters in a roundrobin manner The rst blo c k of X is assigned to an arbitrarily c hosen
disk cluster Eac h blo c k of an ob ject is declustered GRA Q across the d disks that constitute a
X
d
0
0.0
X
3.0
...
X
d
1
0.1
X
3.1
...
X
d
2
1.0
X
4.0
...
X
d
3
1.1
X
4.1
...
X
d
4
2.0
X
5.0
...
X
d
5
2.1
X
5.1
...
CC C
0
1
2
Figure Simple Striping
cluster F or example in Figure a system consisting of six disks is partitioned in to three clusters
eac h consisting of t w o disk driv es The assignmen t of the blo c ks of X starts with cluster This
blo c k is declustered in to t w o fragmen ts X
and X
When a request references ob ject X the system stages X
from cluster C
i
in memory and
initiates its displa y Prior to completion of a time p erio d it initiates the retriev al of X
from cluster
C
i mod C
in to memory in order to ensure a con tin uous displa y This pro cess is rep eated un til
all blo c ks of an ob ject ha v e b een displa y ed T o supp ort sim ultaneous displa y of sev eral ob jects a
time period is partitioned in to xsized slots with eac h slot corresp onding to the retriev al time of
a blo c k from a cluster The n um ber of slots N in a time p erio d denes the maxim um n um ber of
sim ultaneous displa ys supp orted b y a cluster
With C clusters the systems main tains C time p erio ds eac h consisting of N time slots It is
trivial to compute N Band T
p
GKS W e conceptualize a time p erio d as a group Eac h group
has a unique iden tier T o supp ort a con tin uous displayina m ulticluster system a request maps
on to one group and the individual groups visit the clusters in a roundrobin manner Figure If
group G
accesses cluster C
during a time p erio d G
w ould access C
during the next time p erio d
During a giv en time p erio d the requests occup ying the slots of a group retriev e blo c ks that reside
in the cluster that is b eing visited b y that group
Therefore if there are C clusters C groups in the system and eac h cluster eac h group can
supp ort N sim ultaneous displa ys then the maxim um throughput of the system is m N C
sim ultaneous displa ys The maxim um startup latency is T
p
C b ecause groups are rotating
ie pla ying m usical c hairs with the C clusters using eachfor a T
p
in terv al of time and at most
C failures migh t occur b efore a request can be activ ated when the n um ber of activ e displa ys
is few er than N C Th us b oth the system throughput and the maxim um startup latency scale
linearly Note that system parameters suchasbloc ks size time p erio d throughput etc for a cluster
can b e computed using displa y tec hniques suc h as REBECA GKS GSS YCK These displa y
C CC C CC
G
G G GG G
2 1 0
0
1
2 3 4 5
3 4 5
X XX X X X
X
YY Y YY Y
Y
0 12 3 4 5
6
0 1 2
3 4 5
6
...
...
Figure Rotating groups
tec hniques are lo cal optimizations that are orthogonal to the optimization tec hniques prop osed b y
this study Throughput and Latency
Our approac h in the previous section demonstrates the scalabilit y of a system in terms of both its
throughput and maxim um latency It migh t b e p essimistic to consider the maxim um latency as the
latency that a request exp eriences b efore b eing serviced This section quan ties the c haracteristics
of latency and dev elops a probabilistic approachto determine the exp ected latency of a request
In general a m ultidisk platform cannot b e mo deled as a m serv er queueing system b ecause not
all serv ers are iden tical Up on the arriv al of a request it should b e assigned to the group accessing
the cluster that con tains its rst blo c k and not an arbitrarily c hosen cluster Ho w ev er due to a
random distribution of the rst blo c ks of ob jects across disks and a roundrobin access pattern a
request can be assigned to a time slot of an y group W e can conceptualize our striping system as
a queueing system with m iden tical serv ers where a serv er corresp onds to a time slot and not a
cluster Hence w e can compute the probability p k that there are exactly k busy serv ers in the
system at a giv en p oin t in time b y applying a queueing mo del F or example with a P oisson arriv al
pattern and an exp onen tial service time the probabilit yof k busy serv ers in an m serv er loss system
is Kle p k Probfk busy serv ers in the system g k
k P
m
k k
k
where and are the arriv al rate of requests and the service rate of the serv er a v erage service
time resp ectiv ely When a request for an ob ject X arriv es at time tthe system determines the cluster con taining
the rst blo c kof X sa y C
x
and the group curren tly accessing this cluster sa y G
x
If G
x
has at
least one a v ailable slot the request is assigned to G
x
and its displa y is initiated If the time slots of
G
x
are exhausted o ccupied b y other requests the request cannot b e serv ed b y this group failure
Next the system c hec ks the a v ailabilit y of cluster G
x
mod C Note that in con trast to ho w
the clusters are n um b ered the n um b ering of the groups is descending see Figure If the time slots
of this group are also fully exhausted the system c hec ks the a v ailabilit y of cluster G
x
mod C This pro cedure is rep eated un til a group with an idle slot is found success Hence a request migh t
ha vesev eral failures b efore b eing assigned to a sp ecic group in the system
This results in a longer
latency for the requests b ecause sev eral time periods migh t pass b efore the assigned group reac hes
cluster C
x
If a request exp eriences i failures b efore a success the latency for the request is
L T
p
i i T
p
i Let p
f
i k be the probabilit y that the request has i failures b efore a success when there are k
busy serv ers in the system Foragiv en k the probabilit y that a request exp erience i failures b efore
a success is
p
f
i k
m i N
k i N
m i N
k i N
m
k
where k m and i b
k
N
c The calculation of p
f
i k migh t best be explained using an
example
Assume that a system consists of four clusters with eac h group supp orting serv ers Hence
there are t w elv e serv ers a v ailable If a request arriv es when serv ers are busy then this request
migh t exp erience at most failures b
c b efore succeeding The n um ber of all the p ossible w a ys
to allo cate busy serv ers in to serv ers is
Note that this is iden tical to the n um ber of
w a ys to c ho ose busy serv ers from serv ers
F or a request to exp erience failures relativ e to the group con taining the rst blo c k of the
referenced ob ject the serv ers of consecutiv e groups m ust b e fully exhausted This implies that Assuming a rstcomerstserv e p olicy for activ ating requests
consecutiv e serv ers should be busy Therefore once consecutiv e busy serv ers are allo cated to consecutiv e groups there are
w a ys to allo cate a busy serv er to remaining serv ers Hence
the probabilityof ha ving failures is
the n um ber of w a ys to allo cate busy serv ers
to serv ers with the condition of consecutiv e busy serv ers divided bythe n um b er of all p ossible
w a ys to allo cate busy serv ers in to serv ers
F or a request to exp erience failures consecutiv e groups consecutiv e serv ers m ust b e busy Once consecutiv e serv ers are allo cated to consecutiv e groups there are
w a ys to allo cate
busy serv ers to remaining serv ers Ho w ev er this also includes the previous failure cases
Therefore
is the n um ber of w a ys to ha v e exactly failures Hence the probabilit y
of ha ving failures is
Th us when m N and k failures p
f
failures p
f
failure p
f
failure p
f
Let a random v ariable L dene the latency for a request with i failures The probabilit y that
a request has a latency of L is the summation of the probabilit y of i failures conditioned b y all k
v alues Hence the exp ected latency is
E L
m X
k
p k p
f
k T
p
m X
k b
k
N
c
X
i p k p
f
i k i T
p
Tw o Alternativ e T ec hniques to Minimize Startup Latency
Ev en though the w ork load of a displayis relativ ely w ell distributed across the clusters with a round
robin assignmen t of blo c ks a group migh t exp erience a higher w ork load as compared to other groups
F or example in Figure if the system service a request for ob ject X using group G
then all serv ers
C CC C CC
G G G GG G
2 1 0
0
1
2 3 4 5
3 4 5
X XX X X X
X
YY Y YY Y
Y
0 12 3 4 5
6
0 1 2
3 4 5
6
...
...
Z
0
Z 1 ...
Figure Load balancing
in G
b ecome busy while sev eral other groups ha vet woidle serv ers This im balance migh t result in a
higher startup latency for future requests F or example if another request for Z arriv es then it w ould
incur a t w o time p erio d startup latency b ecause it m ust be assigned to G
This section describ es
request migration and replication as t w o alternativ e tec hniques to minimize startup latency These
t wotec hniques are orthogonal to one another enabling a system to emplo y b oth at the same time
Request Migration
By migrating one or more requests from a group with zero idle slots to a group with man y idle slots
the system can minimize the p ossible latency incurred b y a future request F or example in Figure if the system migrates a request for X from G
to G
then a request for Z is guaran teed to incur a
maxim um latency of one time p erio d Migrating a request from one group to another increases the
memory requiremen ts of a displa y This is b ecause the retriev al of data falls ahead of its displa y Migrating a request from G
to G
increases the memory requiremen t of this displayb y three buers
This is b ecause when a request migrates from G
to G
see Figure G
reads X
and sends it to
the displa y During the same time p erio d G
reads X
intoabuer sa y B
and G
reads X
in to
a buer B
During the next time p erio d G
reads X
in to a buer B
and X
is displa y ed from
memory buer B
G
reads X
b ecause the groups mo v e one cluster to the righ t at the end of eac h
time period to read the next blo c k of activ e displa ys o ccup ying its serv ers During the next time
period G
reads X
in to a memory buer B
while X
is displa y ed from memory buer B
This
roundrobin retriev al of data from clusters b y G
con tin ues un til all blo c ks of X ha v e b een retriev ed
and displa y ed
With this tec hnique if the distance from the original group to the destination group is B then
the system requires B buers A request ma y migrate bac k to its original group once a request
in the original group terminates and relinquishes its slot ie time slot b ecomes idle This reduces
the memory requiremen t of the displa y When k C N with the prob of
P
C N
k p k request migration can b e applied due
to the a v ailabilityof idle slots This means that Probfa group is full g Hence p
f
k If
k C N with the prob of
P
m k C N p k no request migration can b e applied b ecause
no idle slot is a v ailable in some groups and the load is already ev enly distributed Hence the
probabilit y of failures is
p
f
i k
C i
k
i
C i k
i C
k
where k
k C N The exp ected latency with request migration is
E L
C N X
k
p k T
p
m X
k C N
p k p
f
k
T
p
m X
k C N
k
X
i
p k p
f
i k
i T
p
Ob ject Replication
T o reduce the startup latency of the system one ma y replicate ob jects W e term the original copyof
an ob ject X as X
s primary cop y All other copies of X are termed its secondary copies The n um ber
of secondary copies for ob ject X is r Eac h of its copies is denoted as R
Xi
where i r Instance
of X refers to the n um ber of times X is copied It equals r r secondary plus one primary Assuming t w o instances of an ob ject b y starting the assignmentof R
X with a cluster dieren t than
the one con taining the rst blo c k of its primary copy X the maxim um startup latency incurred b y
a displa y referencing X can be reduced b y one half see Figure This also reduces the exp ected
startup latency The assignmen t of the rst blo c k of eac h instance of X should b e separated b y a xed n um ber
of clusters in order to maximize the b enets of replication Assuming that the primary cop y of X
is assigned starting with an arbitrary clusters sa y C
i
con tains X
the assignmen t of secondary
copies of X is as follo ws The assignmen t of the rst blo ckof cop y R
Xj
should starts with cluster
C
i
jC
r
mod C F or example Figure sho ws the placemen t for t w o secondary copies of ob ject Y
R
Y R
Y assuming its primary cop y is assigned starting with cluster C
R
Y is assigned starting
C CC C CC
2 1 0 3 4 5
X XX X X X
X
YY Y YY Y
Y
0 12 3 4 5
6
4 5 0 1 2 3
6
...
...
X XX X X X
X
3 45 0 1 2
6 ...
YY Y Y Y Y
Y
2 3 4
5
0 1
6
...
Y Y Y YY Y
Y
0 1 2
3 4 5
6
...
Figure An example of m ultiple cop y placemen t
with cluster C
while R
Y is assigned starting with cluster C
With t w o instances of an ob ject the exp ected startup latency for a request referencing this ob ject
can b e computed as follo ws T o nd an a v ailable serv er the system sim ultaneously c hec ks t w o groups
corresp onding to the t w o dieren t clusters that con tain the rst blo c ks of these t w o instances A
failure happ ens only if both groups are full reducing the n um ber of failures for a request The
maxim um n um b er of failures b efore a success is reduced to b
k
N
c due to t wosim ultaneous searc hing
of groups pro ceeding in parallel Therefore the probabilit yof i failures in a system with eac h ob ject
ha ving t w o instances is iden tical to that of a system consisting of
C
clusters with N serv ers per
cluster
F or an arbitrary n um b er of copies sa y j for an ob ject in the system the probabilit y of a request
referencing this ob ject to observ e i failures is
p
f
j
i k
m j i N
k j i N
m j i N
k j i N
m
k
where i b
k
j N
c Hence the exp ected startup latency is
E L
m X
k p k p
f
j
k T
p
m X
k b
k
j N
c
X
i p k p
f
j
i k i T
p
Ob ject replication increases the storage requiremen t of an application One imp ortan t observ ation
in real applications is that ob jects mayha v e dieren t access frequencies F or example in VideoOn
Demand system more than half of the activ e requests migh t reference only a handful of recen tly
released mo vies Selectiv e replication for frequen tly referenced ie hot ob jects could signican tly
reduce the latency without a dramatic increase in storage space requiremen t of an application The
optimal n um ber of secondary copies per ob ject is based on its access frequency and the a v ailable
storage capacit y The formal statemen t of the problem is as follo ws Assuming n ob jects in the
system let S be the total amoun t of disk space for these ob jects and their replicas Let R
j
be
the optimal instances for ob ject j S
j
to denote the size of ob ject j and F
j
to represen t the access
frequency of ob ject j The problem is to determine R
j
for eac h ob ject j j n while
satisfying
P
n
j R
j
S
j
S A solution can b e computed using sev eral algorithms as describ ed in IK A simple one kno wn
as Hamilton metho d computes the n um ber of replicas per ob ject j based on its frequency ie
Q
j
F
j
S in Figure a It rounds the remainder of the quota Q
j
b Q
j
c to compute R
j
Figure a Ho w ev er this metho d suers from t w o parado xes namely Alabama and P opulation
parado xes Generally sp eaking with these parado xes the Hamilton metho d ma y reduce the v alue
of R
j
when either S or F
j
increases in v alue The divisor metho ds pro vide a solution free from these
parado xes see Figure b F or further details and pro ofs of this metho d see IK Using a divisor
metho d named W ebster d R
j
R
j
w e classify ob jects based on their instances Therefore
ob jects in a class ha v e the same instances An example of classication is sho wn in T able Let R
i
be the optimal instances of ob ject i and F
i
the access frequency of ob ject i The exp ected startup
latency in this system with n ob jects is
E L
n
X
i
F
i
E L
R
i
where E L
R
i
is the exp ected startup latency for ob ject ha ving R
i
instances This is computed using
Equation A Comparison
W e conducted sev eral exp erimen ts to compare request migration with replication and ev alu
ate the analytical exp ectations with observ ations from an implemen tation of these tec hniques using
Mitra W e observ ed the follo wing from these exp erimen ts A system that emplo ys either request
migration or replication incurs a lo w er a v erage startup latency When compared with one another
Step Let Q j F j S and
R j Max l j bQ j cfor j n
l j is a lo w er b ound of ob ject j Step Find index j
ha ving the greatest
remainder Q j b Q j c among those
satisfying R j bQ j c Let R
j
R
j
Step If
P
n
j R j S output R and stop
Otherwise return to step Step Let R j l j for j n and
minsiz e Min S j maxsiz e Max S j l j isalo w er b ound of ob ject j Step Compute
F
j
d R
j
for all j Find index j
ha ving Max F
j
d R
j
Step Let rem S P
n
j
S j R j
If rem minsiz e Then output R and stop
If rem S
j
Then R
j
R
j
Else nd index j
whichhas Max F
j
d R
j
among those satisfying S j rem and
let R
j
R
j
Return to step a Hamilton metho d b Divisor metho d for v ariable ob ject size
Figure Tw o tec hniques to compute the n um b er of replicas p er ob ject
replication is a sup erior alternativ e to request migration The analytical mo dels predict that replica
tion in com bination with migration w ould result in a lo w er a v erage startup latency when compared
with replication alone With a high system utilization the obtained exp erimen tal results are con tra
dictory to this exp ectation replication alone outp erforms the com bination of replication and request
migration The explanation for this is as follo ws Ev ery time a request migrates it consumes ad
ditional disk bandwidth during the time p erio d that it migrates With a high system utilization
there is a high probabilit y of a new request arriving during this time p erio d The cluster con taining
the rst blo c k of the secondary copies of the ob ject referenced b y this request migh t b e una v ailable
due to the extra bandwidth requiremen t causing this request to incur a higher startup latency The
analytical mo dels are con tradictory b ecause they do not capture the bandwidth o v erhead of request
migration
In the next section w e pro vide an o v erview of the system used for this ev aluation Next w e
detail our exp erimen tal design and the obtained results
An Ov erview of Mitra
Mitra is a scalable storage manager that supp orts the displayof con tin uous media data t yp es It is
a soft w are based system that emplo ys o the shelf hardw are comp onen ts Moreo v er it emplo ys a
hierarc hical organization of storage devices to minimize the cost of pro viding online access to a large
v olume of data It is curren tly op erational on a cluster of HP w orkstations It emplo ys
a HP Magneto Optical Juk eb o x as its tertiary storage device Eac h w orkstation consists of a MHz P ARISC CPU MByte of memory and four Seagate STW magnetic disks
PM
Audio
Player
PM
MPEG-1
Player
PM
MPEG-2
Player
Catalog
DM 1 DM 2
DM 3
Scheduler /
EVER-
EST
Rel.DB
User Interface
NN N
NN
N
ATM Switch
N
N : HP-NOSE DM : Device Manager
DM 4
N
SCSI-2 (160 Mb/s)
Volume 9
Volume 10
Volume 11
DM 9 DM 10
DM 11
NN
N
EVEREST
EVEREST
EVEREST
DM 12
N
Volume 5
Volume 6
Volume 7
EVEREST
EVEREST
EVEREST
EVEREST
Volume 8
Volume 1
Volume 2
Volume 3
EVEREST
EVEREST
EVEREST
EVEREST
Volume 4
DM 5
DM 6
N
N
DM 7
DM 8
N
N
DM 0
N
Volume 12
EVEREST
DM 13
N
DM 14
N
HP Magneto-Optical Disk Library
(2 drives, 32 platters)
HP 9000/735
125 MHz PA-RISC
...
SCSI-2 (80 Mb/s)
fast & wide
fast
i 2 1
... ...
SCSI-2
Figure Hardw are and soft w are organization of Mitra
Mitra consists of three soft w are comp onen ts
Sc heduler this comp onentsc hedules the retriev al of the blo c ks of a referenced ob ject in supp ort
of a hiccupfree displa y at a PM In addition it manages the disk bandwidth and p erforms
admission con trol Curren tly Sc heduler includes an implemen tation of EVEREST le sys
tem GIZ staggered striping BGMJ and tec hniques to manage the tertiary storage
device It also has a simple relational storage manager to insert and retriev e information
from a c atalo g regarding the clips that are managed This comp onen t implemen ts the request
migration and sc hedules requests in telligen tly in the presence of m ultiple replicas of an ob ject
Mass storage Device Manager DM P erforms either disk or tertiary readwrite op erations
Presen tation Manager PM Displa ys either a video or an audio clip It mightin terface with
hardw are comp onen ts to minimize the CPU requirementofadispla y F or example to displa y
an MPEG clip the PM migh t emplo y either a program or a hardw arecard to deco de and
displa y the clip
Mitra uses UDP for comm unication bet w een the pro cess instan tiation of these comp onen ts UDP
is an unreliable transmission proto col Mitra implemen ts a ligh tw eigh t k ernel named HPNOSE
HPNOSE supp orts a windo wbased proto col to facilitate reliable transmission of messages among
pro cesses In addition it implemen ts the threads with shared memory p orts that m ultiplex messages
using a single HPUX so c k et and semaphores for sync hronizing m ultiple threads that share memory An instan tiation of this k ernel is activ e p er Mitra pro cess
F or a giv en conguration the follo wing pro cesses are activ e one Sc heduler pro cess a DM pro cess
p er mass storage readwrite device and one PM pro cess p er activ e clien t F or example in our t w elv e
1 10 20 30 40 50
Clip Id
0
50
100
150
200
250
300
350
400
Length [sec]
a Length in seconds of eac h clip
1 10 20 30 40 50
Clip Id
0
50
100
150
200
250
300
Total # of Votes
1 10 20 30 40 50
Clip Id
0
50
100
150
200
250
300
350
Total # of Votes
b V otes p er clip during Jan uary c Syn thetic exp onen tial distribution
Exp ected displa y time sec Exp ected displa y time sec
Figure Characteristics of the CD audio clips
disk conguration with a magneto optical juk e bo x there are sixteen activ e pro cesses fteen DM
pro cesses and one Sc heduler pro cess see Figure There are t w o activ e DM pro cesses for the
magneto juk eb o x b ecause it consists of t w o readwrite devices and optical platters that migh t
be sw app ed in and out of these t w o devices
Exp erimen tal Design
In these exp erimen ts w e assumed that the en tire database is disk residen t no references are
made to the tertiary storage device Mitra w as congured with one disk per cluster ie C
The target database and access pattern w ere based on a WWW page main tained b y Daniel T obias
h ttpwwwsoftdiskcomcomphits that ranks the top ft ysongs ev ery w eek W e assumed that
eac h audio clip is CD qualit y and requires Mbps for its displa y bit stereo Figure b
and a sho w the frequency of access to the clips and the size of eac h clip in seconds resp ectiv ely The ranking of the clips is determined through v oting b y the In ternet comm unit y via Email
Class Instances of T otal access T otal
per ob ject ob jects frequency F
i
storage MB
T able Classied replication of ob jects with WWW distribution gigab yte storage for secondary
copies
W e also analyzed a sk ew ed distribution of access based on that of Figure c The bandwidth of
eac h cluster can supp ort t w elv e sim ultaneous displa ys N W e assume a P oisson arriv al rate
sec for system utilization and sec for for user requests The
n um b er of requests p er ob ject follo w ed the distribution pattern of either Figure b or Figure c
Up on arriv al of a request if the sc heduler fails to nd an idle serv er in the system then this
request is rejected This mo dels an en vironmen t consisting of sev eral systems where the rejected
request is directed to another system
Exp erimen tal Results
In the rst exp erimen t w e assumed that t w o gigab ytes of disk space w ere a v ailable for secondary
copies of ob jects T ables and presen t the n um b er of replicas constructed p er ob ject b y the divisor
tec hnique of Figure b for WWWT obias and exp onen tial distribution resp ectiv ely T able presen ts b oth the analytical and exp erimen tal results for eac h access pattern with system utilization In these exp erimen ts Mitra rejected of requests the analytical mo dels pre
dicted of requests w ould b e rejected The rst column of this table represen ts the tec hnique
emplo y ed Standard refers to a system that emplo ys neither replication nor request migration The
last ro w represen ts a system that emplo ys b oth request migration and replication The second col
umn represen ts the approac h emplo y ed to compute the n um bers in the remaining columns either
exp erimen tal using Mitra or analytical using the Equations of Section With Mitra for b oth migration third ro w and a com bination of migration with replication sev en th ro w the n um ber
in paren theses denotes the a v erage n um b er of memory buers required p er migration op eration
The obtained results demonstrate that a system congured with either request migration or
Class Instances of T otal access T otal
per ob ject ob jects frequency F
i
storage MB
T able Classied replication of ob jects with exp onen tial distribution gigab yte storage for sec
ondary copies
T ec hnique Ev aluation Av erage Startup Latency
Approac h WWWT obias Exp onen tial
Standard Mitra
Analytical
Migration Mitra Analytical
Replication Mitra
Analytical
Migration Mitra Replication Analytical
T able System utilization the n um b er in paren thesis is the a v erage n um b er of memory buers
required p er request migration
replication or a com bination of these t w o results in a lo w er a v erage startup latency compare
Standard with other ro ws This latency is further reduced byask ew ed distribution of access to the
ob jects compare WWWT obias with exp onen tial A surprising result in this table is as follo ws
While the theoretical results predict that a system congured with b oth migration and replication
w ould result in the minim um startup latency the exp erimen tal results sho w that replication alone
yields the minim um startup latency W e attribute this to exp erimen tal errors eg net w orking
dela ys random collisions in tro duced b y measuring elapsed times This w as v alidated b y measuring
the n um b er of failures observ ed b y eac h request prior to a success This failure p ercen tage w as lo w er
for Mitra congured with b oth replication and migration when compared with replication alone
T able presen ts the results obtained with a high system utilization In these exp erimen ts
Mitra rejected appro ximately of requests in all exp erimen ts the analytical mo dels predicted as w ell The general observ ations are almost iden tical to those of T able Once again Mitra
congured with replication observ es a lo w er a v erage startup latency as compared with a com bination
T ec hnique Ev aluation Av erage Startup Latency
Approac h WWWT obias Exp onen tial
Standard Mitra
Analytical
Migration Mitra Analytical
Replication Mitra
Analytical
Migration Mitra Replication Analytical
T able System Utilization the n um ber in paren thesis is the a v erage n um ber of memory
buers required p er request migration
0.5 1 1.5 2 2.5 3 3.5 4
Space for Secondary Copies [GB]
0
0.5
1
1.5
2
2.5
Latency [sec]
Latency with Migration only
Replication
Replication + Migration
Figure Impact of a v ailable space on a v erage startup latency with replication utilization
exp onen tial distribution
of replication and migration compare ro ws v e and sev en The reasoning for this is dieren t than
that of T able With a high system load when request migration requires additional disk bandwidth
when migrating a request if a new request arriv es during the time p erio d that Mitra is p erforming a
migration the rst blo c k of secondary copies of the referenced ob ject migh t b e rendered una v ailable
due to migration This mightforce the new request to observ e a higher n um ber of failures prior to
a success resulting in a higher startup latency This observ ation w as v alidated b y measuring the
n um b er of failures observ ed b y eac h request prior to a success This failure p ercen tage w as lo w er for
Mitra congured with replication alone
In a nal exp erimen t w e analyzed the amoun t of space allo cated for replication on the a v erage
startup latency of the system This exp erimentw as conducted using analytical mo dels The obtained
results are presen ted in Figure In this gure the xaxis represen ts the amoun t of space allo cated
for replicating ob jects The yaxis represen ts the a v erage startup latency With zero space for
replication replication alone is inferior to request migration b ecause it is iden tical to Standard no
secondary copies can be constructed With additional space the a v erage startup latency with
replication starts to decrease This decrease lev els o as the a v erage startup latency approac hes
seconds ie one half of a time p erio d b ecause this is the theoretical minim um for the a v erage
startup latency Conclusions and F uture Researc h Directions
This study describ es request migration and replication as t w o alternativ e tec hniques to minimize
the a v erage startup latency of the system W e dev elop ed analytical mo dels to quan tify the trade
os asso ciated with these tec hniques In addition w e presen ted exp erimen tal results based on an
implemen tation of these tec hniques using Mitra The obtained results demonstrate that the pro
posed tec hniques minimize the a v erage startup latency With the startup latency dropping belo w
one second a system can supp ort complex V CR op erations suc h as fastforw ard and fastrewind
As part of our future researc h w ein tend to extend the replication tec hnique with a metho dology
that incorp orates the role of a tertiary storage device This metho dology should consider ho w the
amoun t of space allo cated to secondary copies w ould increase the n um b er of references to the tertiary
storage device to retriev e the primary copies of colder ob jects Ideally there should b e no increase as
references to the tertiary storage device w ould incur a higher startup latency In addition w ein tend
to dev elop online algorithms that analyze the frequency of access to ob jects in order to delete
secondary copies of those ob jects that ha v e b ecome cold and construct secondary copies of those
ob jects that ha v e b ecome p opular
Ac kno wledgmen ts
W e wish to thank Jab er AlMarri for implemen ting p ortions of request migrations in Mitra
References
And D Andersen A Prop osed Metho d for Creating V CR Functions Using MPEG Streams
In Pr o c e e dings of the Data Engine ering Confer enc e IEEE
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 pages GIZ S Ghandeharizadeh D Ierardi and R Zimmermann An online algorithm to optimize
le la y out in a dynamic en vironmen t Information Pr o c essing L etters GK S Ghandeharizadeh and SH Kim Striping in Multidisk Video Serv ers In Pr o c e e dings
of SPIE HighDensity Data R e c or ding and R etrieval T e chnolo gies Confer enc e Octob er
GKS S Ghandeharizadeh S H Kim and C Shahabi On Conguring a Single Disk Con tin
uous Media Serv er In Pr o c e e dings of the A CM SIGMETRICSMa y
GRA Q S Ghandeharizadeh L Ramos Z Asad and W Qureshi Ob ject Placemen t in Parallel
Hyp ermedia Systems In Pr o c e e dings of the International Confer enc e on V ery L ar ge
Datab ases pages Septem ber GZS
S Ghandeharizadeh R Zimmermann W Shi R Rejaie D Ierardi and TW Li Mitra
A Scalable Con tin uous Media Serv er T ec hnical Rep ort USCCSTR USC IK T Ibaraki and N Katoh R esour c e Allo c ation Pr oblems A lgorithmic Appr o aches The
MIT Press Kle L Kleinro c k Queueing Systems V olume I The ory page WileyIn terscience MY A Merc hantand P S Y u Analytic Mo deling and Comparisons of Striping Strategies for
Replicated Disk Arra ys IEEE T r ansactions on Computers Marc h
ORS B
Ozden R Rastogi and A Silb ersc hatz The Storage and Retriev al of Con tin uous Me
dia Data In V S Subrahmanian and S Jo jo dia editors Multime dia Datab ase Systems Springer WYS JL W olf P S Y u and H Shac hnai DASD Dancing A Disk Load Balancing Sc heme
for VideoonDemand Systems In Pr o c e e dings of A CM SIGMETRICS YCK P S Y u M S Chen and D D Kandlur Design and Analysis of a Group ed Sw eep
ing Sc heme for Multimedia Storage Managemen t In Pr o c e e dings of the Thir d Interna
tional Workshop on Network and Op er ating System Supp ort for Digital A udio and Vide o No v em ber
Abstract (if available)
Linked assets
Computer Science Technical Report Archive
Conceptually similar
PDF
USC Computer Science Technical Reports, no. 659 (1997)
PDF
USC Computer Science Technical Reports, no. 628 (1996)
PDF
USC Computer Science Technical Reports, no. 615 (1995)
PDF
USC Computer Science Technical Reports, no. 634 (1996)
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. 653 (1997)
PDF
USC Computer Science Technical Reports, no. 592 (1994)
PDF
USC Computer Science Technical Reports, no. 864 (2005)
PDF
USC Computer Science Technical Reports, no. 610 (1995)
PDF
USC Computer Science Technical Reports, no. 766 (2002)
PDF
USC Computer Science Technical Reports, no. 625 (1996)
PDF
USC Computer Science Technical Reports, no. 685 (1998)
PDF
USC Computer Science Technical Reports, no. 584 (1994)
PDF
USC Computer Science Technical Reports, no. 598 (1994)
PDF
USC Computer Science Technical Reports, no. 666 (1998)
PDF
USC Computer Science Technical Reports, no. 587 (1994)
PDF
USC Computer Science Technical Reports, no. 911 (2009)
PDF
USC Computer Science Technical Reports, no. 748 (2001)
Description
Shahram Ghandeharizadeh, Seon Ho Kim, Weifeng Shi, Roger Zimmermann. "On minimizing startup latency in scalable continuous media servers." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 627 (1996).
Asset Metadata
Creator
Ghandeharizadeh, Shahram
(author),
Kim, Seon Ho
(author),
Shi, Weifeng
(author),
Zimmermann, Roger
(author)
Core Title
USC Computer Science Technical Reports, no. 627 (1996)
Alternative Title
On minimizing startup latency 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
19 pages
(extent),
technical reports
(aat)
Language
English
Unique identifier
UC16269694
Identifier
96-627 On Minimizing Startup Latency in Scalable Continuous Media Servers (filename)
Legacy Identifier
usc-cstr-96-627
Format
19 pages (extent),technical reports (aat)
Rights
Department of Computer Science (University of Southern California) and the author(s).
Internet Media Type
application/pdf
Copyright
In copyright - Non-commercial use permitted (https://rightsstatements.org/vocab/InC-NC/1.0/
Source
20180426-rozan-cstechreports-shoaf
(batch),
Computer Science Technical Report Archive
(collection),
University of Southern California. Department of Computer Science. Technical Reports
(series)
Access Conditions
The author(s) retain rights to their work according to U.S. copyright law. Electronic access is being provided by the USC Libraries, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.
Repository Name
USC Viterbi School of Engineering Department of Computer Science
Repository Location
Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089
Repository Email
csdept@usc.edu
Inherited Values
Title
Computer Science Technical Report Archive
Coverage Temporal
1991/2017
Repository Email
csdept@usc.edu
Repository Name
USC Viterbi School of Engineering Department of Computer Science
Repository Location
Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089
Publisher
Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA
(publisher)
Copyright
In copyright - Non-commercial use permitted (https://rightsstatements.org/vocab/InC-NC/1.0/