TELKOM
NIKA
, Vol.14, No
.4, Dece
mbe
r
2016, pp. 15
65~157
4
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v14i4.3964
1565
Re
cei
v
ed Ma
y 10, 201
6; Revi
sed O
c
tob
e
r 22, 201
6; Acce
pted No
vem
ber 8, 20
16
DCR: Double Component Ranking for Building Reliable
Cloud Applications
Lixing Xue, Zhan Zha
ng*
, Dechen
g Zuo
Schoo
l of Com
puter Scie
nce
and T
e
chno
log
y
, Harb
in Instit
ute of T
e
chnol
og
y,
Harbi
n
15
00
01
, Heilo
ngj
ian
g
Provinc
e
, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: zz@ftcl.hit.edu.cn
A
b
st
r
a
ct
Since cl
oud a
p
p
licati
ons ar
e u
s
ually l
a
rg
e-sc
ale, it
is too expens
ive to en
h
anc
e the re
lia
bi
lity of all
compo
nents for buil
d
i
ng hi
g
h
ly reli
ab
le cl
oud a
ppl
icat
i
o
ns. T
herefore,
w
e
need to ide
n
tify signific
ant
compo
nents w
h
ich
have
gre
a
t
imp
a
ct on th
e system r
e
li
a
b
ility. FTClou
d
, an existi
ng
ap
proac
h, ranks t
h
e
compo
nents o
n
ly cons
id
erin
g the i
m
p
a
ct o
f
comp
on
ent
i
n
terna
l
fail
ures
and i
gnor
in
g error pro
p
a
gati
on.
How
e
ver, error
propagation is
also
an i
m
portant factor on the system
rel
i
ability. To attack the problem
,
w
e
prop
ose
an
i
m
prove
d
co
mpo
nent r
anki
n
g
framew
ork,
na
med
DCR, to
i
d
entify si
gnific
a
nt co
mp
on
ents
in
clou
d ap
pl
icati
ons. DC
R e
m
p
l
oys tw
o indiv
i
d
ual
alg
o
rith
ms t
o
rank th
e co
mpon
ents tw
ice
and
deter
mi
ne
s a
set of the
most
sign
ifica
n
t co
mp
on
ents b
a
s
ed
on th
e tw
o
r
anki
ng r
e
sults.
In ad
ditio
n
, DC
R do
es
not re
q
u
ire
infor
m
ati
on of compo
nent inv
o
catio
n
frequ
e
n
cies.
Extensiv
e exper
iments
are pr
ovi
ded
to evalu
a
te D
C
R
and co
mpare it
w
i
th
FTCloud.
The experi
m
e
n
tal result
s sh
o
w
that DCR outperforms FTCl
oud i
n
al
most all
cases.
Ke
y
w
ords
: Co
mp
on
ent Ra
nki
ng, Clo
ud Ap
pli
c
ation, Syste
m
Relia
bi
lity, Error Propa
gati
o
n
,
Internal F
a
ilur
e
Copy
right
©
2016 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Clou
d co
mp
uting is a
n
Internet
-ba
s
e
d
com
puting
para
d
igm,
whi
c
h p
r
ovid
es
sha
r
e
d
pro
c
e
ssi
ng
reso
urce
s a
n
d
data to
co
mputers a
n
d
other
device
s
on
dem
and
[1, 2]. In re
cent
years,
cl
oud
co
mputing
i
s
b
e
coming
more
an
d m
o
re
pop
ula
r
and
many
e
n
terp
rises an
d
individual
s p
r
efer to
build
their
syste
m
s
in t
he clou
d environ
ment. The softwa
r
e
system
s
i
n
t
he
clou
d a
r
e
na
med a
s
cl
ou
d ap
plication
s
whi
c
h
usu
a
lly co
nsi
s
t
of variou
s cl
oud
com
pon
ents
comm
uni
cati
ng
with e
a
ch othe
r. Th
e
clo
ud
appli
c
ation
s
are
usu
a
lly large
-
scale
an
d v
e
ry
compl
e
x [3],
whi
c
h may p
o
se a th
reat
to the sy
ste
m
reliability and hin
der transfe
rri
ng cri
t
ical
system
s to t
he cl
oud.
Nowa
days, e
n
d
-u
se
rs
hate
appli
c
ation
s
with lo
w-rel
i
ability and t
h
e
deman
d for
high reliabilit
y is co
ntinual
ly incre
a
si
ng.
Building hi
g
h
ly reliabl
e cl
oud a
pplications
has b
e
come
a chall
engin
g
and req
u
ired
rese
arch p
r
o
b
lem.
The m
a
jor a
ppro
a
ch fo
r i
m
provin
g the
clo
ud
appli
c
ation reliabilit
y is to
enh
a
n
ce
the
reliability of
each individual component. This
may be accom
p
lished either by empl
oying
function
ally equivalent but
more
reliabl
e comp
one
nts to redu
ce
com
pone
nt failure
s or by ad
din
g
fault-toleran
c
e st
rategi
es to tole
rate
co
mpone
nt
failu
res.
Unfortun
ately, both of
them
will i
n
cu
r
extra cost. A
s
cloud
ap
pli
c
ation
s
u
s
u
a
l
l
y invo
lve a
large
num
be
r of
comp
on
ents, it is to
o
expen
sive to provide alt
e
rnative
com
pone
nts o
r
add fault-tol
e
ran
c
e
strat
egie
s
for all
the
comp
one
nts.
Based
on the
80-2
0
rul
e
[4], FTCloud
-an
existing ap
proach [5] attempts to impro
v
e
the reliability
of cloud a
pplication
s
by rankin
g the comp
one
nts to identify a small set
of
signifi
cant
co
mpone
nts a
n
d
enh
an
cing
their reli
abilit
y. Howeve
r, FTClo
ud o
n
ly con
s
id
ers th
e
impact
of the
co
mpo
nent i
n
ternal
failu
res
on th
e
sy
stem
and
do
es
not ta
ke
i
n
to a
c
count
error
prop
agatio
n whi
c
h is al
so
a seri
ou
s thre
at to the global reliability [6].
To attack th
e pro
b
lem,
we p
r
op
ose
a com
pon
ent
ran
k
ing f
r
a
m
ewo
r
k for i
dentifying
signifi
cant
co
mpone
nts an
d hel
ping
de
sign
ers to
bu
ild hig
h
ly reliable
clo
ud
a
pplication
s
. T
h
i
s
frame
w
ork in
clud
es two compon
ent ra
nkin
g algo
rith
ms
, tak
i
ng into acc
o
unt the direc
t
impac
t
of
comp
one
nt i
n
ternal
failu
res
on
the
sy
stem
and
th
e ha
rm
of e
r
ror p
r
op
agati
on, respe
c
tively.
Based
o
n
the
two
alg
o
rith
ms, two
ran
k
i
ng
re
sults a
r
e obtai
ned
a
nd the
n
a
sm
all set of th
e
most
signifi
cant co
mpone
nts wh
ich have
g
r
e
a
t
impa
ct
on
the sy
stem
reliability a
r
e
determined.
By
enha
nci
ng th
e reli
ability of
these
signifi
cant
co
m
pon
ents, the
sy
stem reli
ability ca
n b
e
g
r
ea
tly
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1565 – 157
4
1566
improve
d
. Du
e to the fa
ct t
hat the
com
p
onent
s
a
r
e
ra
nke
d
twi
c
e,
we nam
e the f
r
amework
DCR-
doubl
e com
p
onent ra
nki
n
g
.
The main
con
t
ribution of this pap
er is:
1. Thi
s
p
ape
r identifie
s the
impo
rtan
ce
o
f
error propa
g
a
tion in
lo
cati
ng
signifi
cant
co
mpo
nents
of
clou
d ap
plications whi
c
h
is n
o
t con
s
id
e
r
ed
by
FTClo
ud, a
n
d
propo
se
s
an imp
r
ove
d
component ranki
ng framework, named DCR.
DCR only employs component invocation
relation
shi
p
s to in
depe
n
dently ran
k
the
com
p
o
nents twi
c
e
and
sele
cts the
critical
components
whi
c
h have great
impact on the system
reliability from the two ranking results.
2. Extensive experiment
s are provid
e
d
to ev
aluate the impact of significa
nt compo
nen
ts
identified by
DCR
on the
reliability of cl
oud
appl
i
c
ati
ons an
d d
r
a
w
p
e
rfo
r
man
c
e
com
p
a
r
iso
n
betwe
en DCR
a
nd FTCl
oud.
T
he re
sults
sho
w
t
hat DCR i
s
effective an
d
outpe
rforms
FTClo
ud in al
most
all ca
se
s.
The re
st of this pap
er i
s
organi
zed a
s
fo
llows
. Sectio
n 2 introd
uce
s
the two d
e
scriptio
ns
of
significant comp
one
nts, the
sy
stem a
r
chite
c
tu
re of
DCR an
d rel
a
ted wo
rk. Section 3
detai
ls
the double
compon
ent ra
nkin
g frame
w
ork. Se
cti
on
4 sho
w
s the experim
ent
s t
o
evaluate DCR.
Section 5 d
r
a
w
s the
con
c
lu
sion a
nd future work finally.
2. Preliminaries
2.1. Significant Compo
n
e
n
ts
A failure of a comp
onent i
n
softwa
r
e
system
s ca
n b
e
attributed t
o
two re
ason
s [6], as
sho
w
n in Fig
u
re 1. One is that an error caused
by faults in the compon
ent (such a
s
bug
s) is
delivere
d
at the output inte
rface,
i.e. co
mpone
nt internal failure.
Th
e other is that
the compo
n
e
n
t
receives
an i
n
co
rrect inp
u
t and ge
nerates an
erro
n
e
ous o
u
tput, n
a
mely error
p
r
opa
gation le
ads
to a comp
on
ent failure. A system failure occurs
only
if an erro
r e
v
entually rea
c
he
s the sy
stem
interface, no
matter h
o
w the e
r
ror i
s
produ
ced
an
d
prop
agate
d
. In a
wo
rd,
co
mpone
nt inte
rna
l
failure
s and e
rro
r propa
gati
on are two m
a
jo
r threats to
the system reliability.
Figure 1. Two threats to reli
ability
It is apparen
t that not only the direct
impac
t of co
mpone
nt internal failu
re
s on the
system
sh
oul
d be
redu
ce
d
,
but also th
e
harm
of
erro
r propa
gation
sho
u
ld al
so
be minimi
ze
d
,
if
we
wa
nt to
b
u
ild hi
ghly rel
i
able
clo
ud
a
pplication
s
. T
herefo
r
e, th
e
signifi
cant
co
mpone
nts in t
h
is
pape
r are abl
e to be descri
bed from two perspe
c
tives.
1·
The sig
n
ifica
n
t compo
nent
s are th
e one
s wh
ose failures have g
r
eat
impact on th
e system.
2· The
signif
i
cant comp
o
nents a
r
e al
so the o
nes
which may severely affect a lot of othe
r
comp
one
nts
and furthe
r h
a
rm the glo
b
a
l
reliability
by prop
agatin
g e
rro
rs o
u
t whe
n
they fail.
2.2. Sy
stem
Architec
ture
The sy
ste
m
arc
h
itectu
re
of DCR i
s
s
how
ed in Fig
u
re 2, whi
c
h
include
s thr
ee part
s
:
stru
cture g
r
a
ph buil
d
ing,
comp
one
nt ranki
ng a
nd
signifi
cant
co
mpone
nt det
ermin
a
tion. T
h
e
pro
c
ed
ures of
DCR are a
s
f
o
llows:
1. The
syste
m
de
signe
r
provide
s
the
stru
cture info
rmation
of
a
clou
d ap
plica
t
ion to DCR.
A
stru
cture gra
ph is ge
nerated ba
sed o
n
t
he com
pon
en
t invocation relation
ship
s.
2. Two
seri
es of si
gnifican
c
e val
u
e
s
of t
he
cl
ou
d com
pone
nts are
cal
c
ulate
d
by
employin
g
t
w
o
different
com
pone
nt ra
nkin
g algo
rithm
s
whi
c
h a
r
e
pro
posed in te
rms of the
two
descriptio
n
s
of significant comp
one
nts i
n
the last su
b
s
e
c
tion,
re
sp
ectively. Accordin
g to the two se
rie
s
of
signifi
can
c
e v
a
lue
s
, the co
mpone
nts a
r
e
ranked twi
c
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
DCR: Do
uble
Com
ponent
Ran
k
in
g for Building
Re
liab
l
e Clou
d Appl
ication
s
(Li
x
in
g Xue)
1567
3. Based on t
he two ra
nki
n
g results, the most
sig
n
ifica
n
t compo
nent
s whi
c
h have
strong imp
a
ct
on the glob
al
reliability are determi
ned
and retu
rne
d
to the syst
em de
signe
r
for buildin
g a
reliabl
e clo
u
d
application.
Figure 2. System architect
u
re of DCR
2.3. Relate
d Work
In traditio
nal
softwa
r
e
reli
a
b
ility engin
e
e
r
ing, th
ere
a
r
e fou
r
comm
on m
e
thod
s t
o
buil
d
reliabl
e
software
sy
stems,
nam
ely faul
t preve
n
ti
on,
fault remov
a
l, fault tole
rance a
n
d
fa
ult
forecastin
g [7
]. Howeve
r, fault preve
n
tion and fault re
moval are n
o
t able to be a
pplied
whe
n
we
build
clou
d a
pplication
s
. This i
s
be
ca
use buildi
ng
clo
ud ap
plicatio
ns u
s
u
a
lly uses exi
s
ting
cl
oud
comp
one
nts
and we ca
n
not partici
pat
e in the de
velopment of
them. But
we can sele
ct
comp
one
nts
with high reli
ability acco
rd
ing to
desig
n
requi
reme
nts. Another m
e
thod we ca
n
employ is sof
t
ware fault tol
e
ran
c
e. Software faul
t-tol
e
ran
c
e te
chni
que
s, such a
s
re
covery bl
ock
[8] and
N-Versi
on P
r
og
ramming
(N-Modula
r
Re
d
unda
ncy) [9], are
widely
use
d
in
vari
ou
s
system
s. In the cl
oud
envi
r
onm
ent, a g
r
eat numb
e
r o
f
functionally equivalent bu
t
indepe
nde
ntly
desi
gne
d co
mpone
nts ca
n be used for
desi
gning fa
u
l
t-tolera
nce mech
ani
sms.
As
clou
d
com
puting i
s
be
coming
pop
ula
r
, a
num
ber o
f
wo
rks
have
been
carried
out on
it. Service co
mpone
nt sele
ction and
co
mpositio
n
is
one of the ho
tspots. Many
approa
che
s
h
a
ve
been p
r
op
osed, su
ch as QoS-a
w
are web servi
c
e comp
ositio
n [10], web se
rvice reputati
o
n
model [1
1], OWL
-
S
servi
c
e
profile
ba
sed
web
se
rv
i
c
e
sel
e
ction
[12] an
d web
servi
c
e
sel
e
ction
based o
n
co
ncu
r
rent re
q
uest
s
[13]. Compon
ent
ra
nkin
g is
a prereq
uisite fo
r applying th
e
s
e
resea
r
ch find
ings
and
so
me stu
d
ies h
a
ve bee
n carried o
u
t. Ho
wever, th
e a
ppro
a
che
s
d
o
not
take i
n
to a
ccount e
r
ror propag
ation,
which
is al
so
a majo
r th
re
at to the
reli
ability of cl
o
u
d
appli
c
ation
s
.
In additio
n
, t
hey requi
re t
he
stru
ctur
e
informatio
n a
s
well
a
s
the
inform
ation
of
comp
one
nt in
vocation
fre
q
uen
cie
s
. Ou
r
approa
ch
whi
c
h
attacks th
e wea
k
ne
ss
requires only
the
stru
cture info
rmation
and
take
s into a
c
cou
n
t t
he ha
rm of e
rro
r
prop
agatio
n i
n
the sy
ste
m
,
obtainin
g
wo
nderful
re
sult
s.
3. Double Co
mponent
Ra
nking
As shown in
Figure 2,
DCR incl
udes three
part
s
, which
will be
detailed in thi
s
section,
respe
c
tively. Structu
r
e
gra
ph b
u
ild
ing
is introd
uced
within Sectio
n
3.1. The
n
the
two
co
mpon
en
t
ran
k
ing
alg
o
ri
thms
are
p
r
o
posed
acco
rd
ing to
the
t
w
o
de
scriptio
ns
of sig
n
ifica
n
t
comp
one
nts i
n
Section 3.2
and 3.3 resp
ectively. In Section 3.4,
d
e
termin
ation
of significant
comp
one
nts is
discu
s
sed.
3.1. Structu
r
e Graph
Building
The st
ruct
ure
of a cloud
ap
plicatio
n, that
is, the com
p
onent invo
cat
i
on rel
a
tionsh
i
ps, ca
n
be mo
deled
as a
directe
d
gra
ph
,
GC
E
, whe
r
e a
nod
e
i
C
in the no
de
set
C
den
otes
a
comp
one
nt a
nd a
di
re
cted
edge
ij
e
from
i
C
to
j
C
in the edge
set
E
repre
s
e
n
ts that
i
C
invo
k
e
s
j
C
(denote
d
as
ij
CC
).
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1565 – 157
4
1568
A
stru
cture g
r
aph co
ntaini
ng
n
nodes
ca
n be de
scri
b
ed by an
nn
ad
jace
ncy mat
r
ix
ij
nn
Aa
. Eac
h
entry
ij
a
in the matrix is
defined by:
1
0
ij
ij
C
f
a
s
C
i
el
e
(
1
)
In the matrix
,
1
ii
a
represe
n
ts a self-i
nv
ocation of co
m
pone
nt
i
C
. The
numbe
r of
edge
s
starting
from
node
i
C
is call
e
d
out-deg
ree
of
i
C
, denote
d
as
de
g
(
)
i
C
. It is
able
to be obtained
by:
1
de
g
(
)
n
ik
k
i
Ca
(
2
)
Similarly, the numbe
r of edge
s en
din
g
at node
i
C
is call
ed in-d
egre
e
of
i
C
, denoted a
s
de
g
(
)
i
C
. It can be cal
c
ulate
d
by:
1
de
g
(
)
n
ki
k
i
Ca
(
3
)
A com
pon
ent
i
C
inv
o
ke
s a
t
o
t
a
l of
de
g
(
)
i
C
com
p
o
nents an
d
i
C
is invoke
d
by a total
of
de
g
)
(
i
C
compo
nent
s.
3.2. Failure-Bas
ed Comp
onent
Ran
k
ing
In a clou
d appli
c
ation,
some
com
p
o
nents
a
r
e freque
ntly invoke
d by ma
ny other
components. It is obvious that t
heir failures
will direct
ly affect
the
system reliability much more
than other
compon
ents [
14]. These comp
one
nts
fo
llow the first descriptio
n
of significant
comp
one
nts
discu
s
sed in t
he la
st se
ctio
n. Intuitiv
ely, these si
gnifica
nt co
mp
one
nts in a
stru
ctu
r
e
grap
h a
r
e
th
e on
es which
have
many i
ngoin
g
lin
ks f
r
om
other im
portant
comp
onent
s. On
the
basi
s
of the Page
Ran
k
alg
o
rithm [15], we prop
ose
an algorith
m
to calcul
ate the first se
rie
s
of the
signifi
can
c
e v
a
lue
s
of the cloud compo
n
ents, name
d
as failure-ba
s
ed sig
n
ifica
n
ce values.
For a clou
d appli
c
ation
which
contain
s
n
comp
one
nts, the
failure-based
sig
n
ifican
ce
value
()
i
VF
C
of a co
mpone
nt
i
C
is defined a
s
:
(
1
()
(
1
)
de
g
(
)
)
ji
j
i
CC
j
C
C
C
VF
VF
n
(
4
)
Whe
r
e
1
n
is the basi
c
sig
n
i
f
ican
ce value
of
i
C
itself while
(
deg
(
)
)
ji
j
CC
j
C
VF
C
is the significa
nce
value de
rived
from othe
r
comp
one
nts
that invoke
i
C
. The parameter
(
01
)
in (4)
is
utilized to adj
ust the
proportion of
the two values,
which i
s
usuall
y
set to be 0.85. By (4),
a
comp
one
nt
i
C
has a larg
e failure
-ba
s
ed sig
n
ifican
ce value if the sum of
failure-ba
s
e
d
signifi
can
c
e v
a
lue
s
of the comp
one
nts t
hat invoke
i
C
is large, indi
cat
i
ng that
i
C
is in
voked by
many other
si
gnifica
nt com
pone
nts.
Equation (4)
can b
e
writte
n in matrix form:
11
()
(
)
1
/
(1
)
()
(
)
1
/
nn
CC
CC
VF
V
F
n
W
VF
V
F
n
(
5
)
Whe
r
e the m
a
trix
()
ij
n
n
Ww
is defined by:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
DCR: Do
uble
Com
ponent
Ran
k
in
g for Building
Re
liab
l
e Clou
d Appl
ication
s
(Li
x
in
g Xue)
1569
1
deg
(
)
0
ij
j
i
j
C
if
w
el
s
e
C
C
(
6
)
The pro
c
e
d
u
r
es of cal
c
ula
t
ing the failure-ba
s
e
d
sig
n
ifican
ce value
s
are
simple.
First,
rand
omly a
s
sign initial valu
es b
e
twe
en 0
and 1
to the
failure
-ba
s
ed
signifi
can
c
e v
a
lue
s
()
i
VF
C
(
1,
2
in
). Then, solv
e (5) by repe
at
ing the co
mputation un
til all significa
nce valu
es b
e
co
m
e
stable.
Usi
ng the
a
bove ap
pro
a
c
h, the fail
ure
-
ba
se
d si
gnifica
nce value
s
of th
e clo
u
d
comp
one
nts
can b
e
obtai
ned. Accordi
ng to these
values, the
comp
one
nts are ran
k
ed.
A
comp
one
nt with a la
rge
r
value is
co
nsid
ere
d
to
be mo
re
sig
n
ificant. The
failure
s of
the
signifi
cant co
mpone
nts sel
e
cted fro
m
this ran
k
in
g re
sult will have
great
impa
ct on the syste
m
reliability.
3.3.
Propaga
tion-Bas
e
d Compon
ent Rankin
g
In a cloud ap
plicatio
n, there must be so
me
comp
one
nts that frequ
ently invoke a lot of
other compo
nents.
Thei
r failures ma
y affect a
l
o
t of
sub
s
e
quent
co
mp
onent
s by
e
rro
r
prop
agatio
n and
furth
e
r h
a
rm
the syste
m
relia
bilit
y. So these
co
mpone
nts a
r
e co
nsi
dered
to be
importa
nt an
d they accord
with the se
cond de
scri
pt
ion of sig
n
ificant com
pon
e
n
ts in Sectio
n
2.1.
Intuitively,
these
signifi
ca
nt comp
onen
ts in
a stru
ct
ure g
r
ap
h are the one
s
whi
c
h have
many
outgoin
g
lin
ks to othe
r im
portant
com
p
onent
s. I
lluminated by the
Tru
s
tRa
n
k
al
gorithm [1
6], we
prop
ose anot
her al
gorithm
to calculate
the se
con
d
serie
s
of sig
n
i
f
ican
ce value
s
of the clo
u
d
comp
one
nts, named a
s
p
r
o
pagatio
n-b
a
sed sig
n
ifica
n
ce values.
Assu
ming th
at a clo
ud a
pplication co
ntains
n
com
pone
nts, the
prop
agatio
n
-
ba
sed
signifi
can
c
e v
a
lue
()
i
VP
C
of a compone
nt
i
C
is defined a
s
:
(
1
()
(
1
)
deg
(
)
)
ij
j
i
CC
j
C
C
C
VP
VP
n
(
7
)
Whe
r
e
1
n
is the basi
c
sig
n
i
f
ican
ce value
of
i
C
itself while
(
deg
(
)
)
ij
j
CC
j
C
VP
C
is the significa
nce
value
de
rived
from other comp
one
nts whi
c
h are
in
voked
by
i
C
. Similarly, the
para
m
eter
(
01
) in (7
) i
s
e
m
ployed to
a
d
just the
pro
portion
of the
two value
s
,
whi
c
h i
s
u
s
ua
lly set as
0.85. By (7
),
a
com
pon
en
t
i
C
ha
s
a la
rg
e p
r
opa
gatio
n-ba
se
d
signi
fican
c
e val
u
e
if the
sum
of
prop
agatio
n-based
sig
n
ifican
ce
value
s
of
the
comp
onent
s
whi
c
h
are invo
ke
d
by
i
C
is large,
sho
w
in
g that
i
C
invoke
s a large qua
ntity of
other sig
n
ificant comp
one
nts.
The equival
e
nt matrix equ
ation of (7) i
s
:
11
()
()
1
/
(1
)
()
(
)
1
/
nn
VP
V
P
n
W
VP
V
P
n
CC
CC
(
8
)
Whe
r
e the m
a
trix
()
ij
n
n
Ww
is defined by:
1
deg
(
)
0
i
j
j
i
j
if
w
el
C
s
C
C
e
(
9
)
The procedu
res of cal
c
ul
ating the pro
p
a
gati
on-ba
sed
significan
c
e
values a
r
e id
entical
with tho
s
e of
cal
c
ulating t
he failure-b
a
s
ed
signifi
ca
nce valu
es.
First, ra
ndo
m
l
y assig
n
init
ial
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1565 – 157
4
1570
values betwe
en 0
and
1
to the p
r
op
a
gation-ba
sed
sig
n
ifican
ce
value
s
()
i
VP
C
(
1,
2
in
).
Then, solve (8) by rep
eatin
g the comp
utation
until all signifi
can
c
e v
a
lue
s
be
com
e
stable.
With the
a
b
o
ve ap
proach, the
pro
p
a
gation-
ba
sed
sig
n
ifica
n
ce
value
s
of
the
clou
d
comp
one
nts
can be o
b
tai
ned. On the
basi
s
of the values, the comp
one
nts are ran
k
ed
. A
comp
one
nt is co
nsi
dered t
o
be
mo
re
si
gnifica
nt if
it
has a l
a
rger
value. Th
e fa
ilure
s of
the
s
e
signifi
cant co
mpone
nts sel
e
cted
from
thi
s
ran
k
ing
result will
seve
re
ly affect othe
r co
mpon
ents
in
the cloud application a
nd further affect the system rel
i
ability.
3.4. Significant Com
p
o
n
e
n
t De
term
ina
t
ion
Based
on th
e
two
seri
es
of signifi
can
c
e
values, the
compon
ents in
the cl
oud
ap
plicatio
n
can be
ran
k
ed
re
sp
ectiv
e
ly.
The
failu
re-ba
s
ed sig
n
ifican
ce
val
ues enabl
e us
to
ide
n
tify
the
signifi
cant co
mpone
nts which have g
r
eat dire
ct
impact on the
system reli
ability while the
prop
agatio
n-based si
gnificance value
s
help u
s
loca
te the signifi
ca
nt compo
nent
s whi
c
h
severely
affect other components and furt
her harm the
system reliability.
Whi
c
h
ranking result i
s
m
o
re
importa
nt? We believed th
at there is no
accurate an
swer.
To re
du
ce b
o
th of the d
i
rect a
nd in
d
i
rect th
reat
s and bette
r
improve the
system
reliability, Top-
2
k
(
2
kn
and
k
is e
v
en) comp
on
ents a
r
e
re
sp
ectively sel
e
cted from the t
w
o
ran
k
ing
re
sul
t
s and
he
nce
a total of
k
comp
one
nts are determin
ed
a
s
the m
o
st significan
t
comp
one
nts.
In this way, the de
sign
er o
f
the
cloud
ap
plicatio
n ca
n i
m
prove the
system relia
bili
ty
efficiently by enha
nci
ng th
e reliability of these
comp
o
nents.
4. Experiments and Ev
aluation
In this
se
ctio
n, extensive
experi
m
ent
s are
pr
o
v
id
ed
to
va
lid
a
t
e DC
R
,
e
v
a
l
ua
te
th
e
impact of different pa
ram
e
ter setting
s o
n
DCR an
d compa
r
e DCR with FTCl
oud
.
4.1. Experimental Setup
In this se
ction
we co
mpa
r
e
the followin
g
approa
che
s
:
1.
DCR:
The comp
one
nts are ran
k
ed b
y
DCR
and
t
he Top
-
K pe
rcent
comp
on
ents a
r
e
sele
cted
as the si
gnificant
components for
enhancing the reliability.
2. Ran
d
CR:
K
pe
rce
n
t comp
o
nents
a
r
e ra
n
domly sele
ct
ed
a
s
the sig
n
ificant co
mp
onent
s
fo
r
enha
nci
ng th
e reliability.
3.
FTClo
ud: T
h
e comp
one
nts a
r
e
ran
k
ed
by
FTCloud
and the Top
-
K perce
nt co
mpone
nts are
selected as the significant compone
nts for enhancing
the reliability.
The system reliability
is consi
dered
to be
the p
r
o
b
a
b
ility of gene
rating
co
rre
ct
output
with corre
c
t input [17]. Fo
r a fair
co
mp
arison, we
a
s
sume th
at the internal fa
ilure p
r
o
babili
ty
(
int
f
) of the selected components
can be
reduced to 80% afte
r enhanci
ng the reliability no
matter whi
c
h
approa
ch is
e
m
ployed. In a
ddi
tion, in DCR and F
T
Clo
ud, the para
m
eter
is used
to bala
n
ce th
e si
gnifican
c
e value
s
deri
v
ed from
oth
e
r
com
pon
en
ts an
d the
ba
sic value
s
of
the
comp
one
nts t
hemselves. I
n
previo
us
st
udie
s
[
18, 19
], it has bee
n
proved th
at 0.85 is
a goo
d
choi
ce. Thu
s
,
in our expe
ri
ments,
is also set to be 0.85.
A scale
-fre
e grap
h is a graph wh
ose d
egre
e
distri
b
u
tion follows a powe
r
law,
at least
asymptoticall
y
.
Previous studies have
d
e
mon
s
trate
d
that not only
the Internet [
20] but al
so t
h
e
internal struct
ure
s
of comm
on
softwa
r
e such as
Linux
Kernel, M
o
zill
a, Xfree86
an
d MySQL [21
,
22] ap
pea
r t
o
be
scal
e-free. Th
erefo
r
e
,
the net
wo
rk analy
s
is software Paj
e
k [23] is utilized
to
gene
rate
scale-free di
re
cted g
r
a
p
h
s
as
stru
ct
u
r
e graph
s of
clou
d ap
pli
c
ation
s
in t
h
e
experim
ents.
Thre
e scale
-
free di
re
cted
grap
hs
with d
i
fferent settin
g
s of nod
e n
u
mbe
r
s
(i.e. 500, 100
0
and 200
0) a
r
e gen
erate
d
by Pajek in our ex
pe
ri
ments. The
n
the compo
nent invocati
on
freque
nci
e
s
of each grap
h are
ran
d
o
m
ly gene
rate
d
to simul
a
te
the statisti
cal data du
rin
g
a
perio
d of run
n
ing onli
ne. These comp
onent invo
cat
i
on freq
uen
ci
es a
r
e u
s
ed
in FTClo
ud a
n
d
cal
c
ulating the system reli
ability.
4.2. Validation and Perfor
mance Com
p
arison
In order to va
lidate
DCR
a
nd comp
are DCR with FT
Clou
d,
the
ap
proa
ch
es a
r
e
appli
ed
to the three graphs respective
ly and the experi
m
ental result
s of application reliability are
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
DCR: Do
uble
Com
ponent
Ran
k
in
g for Building
Re
liab
l
e Clou
d Appl
ication
s
(Li
x
in
g Xue)
1571
repo
rted in T
able 1.
in
t
f
represe
n
ts the in
ternal failure prob
ability of the cloud compon
ents,
inclu
d
ing two
value setting
s (i.e. 0.05 and 0.1).
Top-K
(K=2%, 6%,
10% and 20
%) indicate
s that
K
pe
rce
n
t mo
st si
gnificant
co
mpon
ents in
DCR
an
d FT
Clou
d,
and
K
pe
rcen
t ran
domly
sele
cted
com
pone
nts in
Rand
CR
are
se
lecte
d
for en
han
cing th
e
reliability.
ep
w
h
ic
h
is
no
t l
i
s
ted
in the table repre
s
e
n
ts th
e error p
r
op
a
gation
proba
bility of the cloud compo
n
ents. It is usually
clo
s
e to 1 [24
]
and thus we
set it to be 0.99.
Compared with
RandCR
, DCR obt
ains better
reliability performance in all the
experim
ental
setting
s. Thi
s
ob
se
rvation
sho
w
s th
at enha
nci
ng th
e relia
bility of the significant
comp
one
nts
i
dentified by DCR ca
n obt
ain
b
e
tter s
ystem reli
ability than
en
han
ci
ng the
reliabil
i
ty
of randomly selecte
d
com
p
onent
s. In other word
s,
DCR is a
b
le to effectively id
entify significant
components
whi
c
h have gr
eat impact on the system
reliability.
Table 1. Perf
ormance Compari
s
on of Application Reliability
Node
Numbers
Methods
Component intf=
0.05
Component intf=
0.1
Top2%
Top6%
Top10
%
Top20
%
Top2%
Top6%
Top10
%
Top20
%
500
RandCR
0.7381
0.7403
0.7410
0.7418
0.5416
0.5481
0.5499
0.5512
FTCloud
0.7453
0.7597
0.7641
0.7714
0.5572
0.5750
0.5818
0.5930
DCR
0.7478
0.7681
0.7725
0.7818
0.5568
0.5883
0.5952
0.6096
1000
RandCR
0.7247
0.7298
0.7312
0.7391
0.5201
0.5263
0.5234
0.5365
FTCloud
0.7322
0.7460
0.7510
0.7590
0.5398
0.5552
0.5626
0.5746
DCR
0.7354
0.7522
0.7581
0.7680
0.5396
0.5650
0.5739
0.5889
2000
RandCR
0.6974
0.7015
0.7098
0.7156
0.4881
0.4901
0.4928
0.5006
FTCloud
0.7147
0.7251
0.7307
0.7383
0.5098
0.5248
0.5329
0.5440
DCR
0.7206
0.7322
0.7390
0.7494
0.5185
0.5355
0.5455
0.5610
Compared wi
th
FTCloud, DCR
pr
ovides better reliability perfo
rm
ance in all
the settings
except for th
e case that Top-K eq
ual
s to 2% while
in
t
f
is
s
e
t as
0.1. In thi
s
c
a
s
e
, the
perfo
rman
ce
of DCR may be slightly wo
rse tha
n
that of FTClou
d
when the nod
e
numbe
r is 5
00
or 1
000
and
the differen
c
e
is n
o
t more th
an 0.00
04. In
ord
e
r to
furth
e
r
study the
p
e
rform
a
n
c
e
o
f
DCR and d
r
a
w
perfo
rma
n
c
e compa
r
iso
n
, more inve
stigation
s
int
o
the impact
of internal fail
ure
prob
ability, error p
r
op
agati
on pro
bability
and Top
-
K will be followed
.
4.3. Impact of Relev
a
nt Parameters
To study the impact of the
compo
nent i
n
ternal failu
re prob
ability (
in
t
f
) on the system
reliability, we
comp
are the
approa
che
s
i
n
int
f
settings of
0.01 to 1
with
a ste
p
value
of 0.01. In
this expe
rime
nt, the node
numbe
r is 1
0
00 and
ep
is se
t as 0.99. Th
e experim
ent
al re
sults of
clou
d appli
c
at
ion relia
bility in Figure 3 sh
ow that:
1·
DCR
consi
s
t
ently
provides
better
reliability perform
ance than
F
T
Cloud in al
l
cases
when
Top-K
=
20%, Top-K
=
10% or
Top
-
K=6%,
and
in almo
st all ca
se
s when Top
-
K=2
%
.
2· O
n
ly if To
p-K=2% an
d
in
t
f
is
not le
ss than 0.
09, the reliability
perfo
rman
ce
of FTCl
oud
approa
che
s
o
r
slightly exce
eds that of
DCR. It exceed
s a maximum
of 0.0004.
To study the
impact
of the com
pon
ent error propagation
probability
(
ep
) o
n
the
sy
stem
reliability, we
comp
are the approa
che
s
i
n
ep
settings
of 0.9 to 0.99 wi
th a step valu
e of 0.01. In
this exp
e
rime
nt, the nod
e
numbe
r i
s
al
so 100
0 a
nd
in
t
f
is
set a
s
0.1.
The exp
e
rim
ental resul
t
s
of cloud a
ppli
c
ation reliabili
ty in Figure 4 sho
w
that:
1·
DCR
consi
s
t
ently
provides
better
reliability perform
ance than
F
T
Cloud in al
l
cases
when
Top-K
=
20%, Top-K
=
10% or
Top
-
K=6%,
and
in almo
st all ca
se
s when Top
-
K=2
%
.
2·
Only if To
p-K
=
2%
and
0.
9
ep
, the reliability p
r
ovided
by FT
Clou
d i
s
0.0
0
1
mo
re th
an
that
provide
d
by DCR.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1565 – 157
4
1572
Figure 3. Impact of com
p
o
nent intern
al failure p
r
o
babi
lity
Figure 4. Impact of com
p
o
nent error p
r
o
pagatio
n pro
b
ability
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
DCR: Do
uble
Com
ponent
Ran
k
in
g for Building
Re
liab
l
e Clou
d Appl
ication
s
(Li
x
in
g Xue)
1573
Figure 5. Impact of Top-K
To
study the
impact
of T
o
p-K o
n
th
e
system reliabil
i
ty, we
comp
are
the
app
ro
ach
e
s in
different Top
-
K settings. In this experim
e
n
t, the node numbe
r is stil
l 1000 and
ep
is set as 0.99,
too. The expe
rimental resul
t
s of cloud a
p
p
licatio
n relia
bility in Figure 5 sho
w
that
:
1. DCR
consi
s
t
ently
provides
better
reliability perform
ance than
F
T
Cloud in al
l
cases
when
0.05
in
tf
and almo
st all cases
whe
n
0.1
in
t
f
.
2. Only if
0.1
in
t
f
and Top
-
K is set as 2%, the
reliability pro
v
ide by FTCl
oud is
0.000
2 more
than that prov
ide by DCR.
To sum u
p
, DCR
outpe
rforms FT
Clo
ud i
n
almo
st
all
cases. Only if Top-K
=
2% as well
as
ep
is
small or
in
t
f
is larg
e, the pe
rforma
nce of
FTClo
ud m
a
y app
roa
c
h
or
slightly excee
d
that of
DCR. Thi
s
o
b
se
rvation i
s
due to the
signifi
cant co
mpone
nts d
e
t
erminatio
n o
f
DCR and t
h
e
inequ
ality bet
wee
n
imp
a
ct
of
comp
one
nt internal
fai
l
ure
s
on th
e
reliability a
n
d
impa
ct of
erro
r
prop
agatio
n in clou
d appli
c
ation
s
. DCR treats impa
ct of compone
nt internal fai
l
ure
s
and e
r
ro
r
prop
agatio
n equally,
a
nd sele
cts
Top
-
2
k
comp
one
nts f
r
om th
e two
ranki
ng
re
sult
s
re
spe
c
tively.
In these extreme cases, t
he impa
ct of the first Top
-
2
k
comp
one
nts of the prop
a
gation-ba
sed
ran
k
ing resul
t
on the reliability may be a little weaker than the i
m
pact of the
second To
p-
2
k
comp
one
nts
of the failure-based ra
nki
n
g re
sult,
cau
s
ing the perfo
rmance of DCR to be slig
htly
worse th
an th
at of FTClo
ud
in this case. The ob
se
rvati
on can only b
e
found
whe
n
node n
u
mb
e
r
is not
big. When the
scal
e of the
clou
d appl
i
c
atio
n
rea
c
h
e
s
20
00 no
de
s, DCR
outpe
rform
s
FTClo
ud
with
out any exce
ption. Anywa
y
, the neglig
i
b
le pe
rform
a
nce
differen
c
e in a few
extreme
ca
se
s doe
s n
o
t cover the e
ffect
iveness a
nd advanta
g
e
s
of DCR.
5. Conclusio
n
This p
ape
r
prop
oses
a
comp
one
nt ran
k
ing f
r
a
m
ewo
r
k for identifying signifi
cant
comp
one
nts
whi
c
h have g
r
eat impa
ct on the cloud a
pplication reli
ability to help design
e
rs bu
ild
reliabl
e clo
u
d
application
s
. This fram
ework ta
ke
s
into
account the i
m
pact of com
pone
nt intern
al
failure
s a
s
well as the h
a
rm of erro
r p
r
opag
ati
on to
ran
k
the
co
m
pone
nts twi
c
e only em
plo
y
ing
the sy
stem
structu
r
e
information. Th
e
signifi
cant
co
mpone
nts
are dete
r
mine
d
ba
sed
on th
e two
ran
k
ing
re
sult
s. The reli
abi
lity of cloud appli
c
at
ion
s
can b
e
greatl
y
improved b
y
enhan
cing
the
reliability of these signifi
cant comp
one
nts.
Comp
ared with FTCl
oud, the pro
posed frame
w
o
r
k
con
s
id
ers mo
re but
req
u
ires le
ss. Plen
ty of
experim
ents a
r
e
con
ducte
d to draw p
e
rfo
r
ma
nce
comp
ari
s
o
n
and the re
sults sh
ow that our fram
ework is effective
and outperfo
rms FT
Clo
u
d
in
almos
t
all cases
.
The future work i
n
cl
ude
s:
a) imp
r
oving
the
determin
a
tion of si
gnificant
comp
on
ents, b
)
more
expe
ri
mental a
naly
s
is of a
c
tual
clo
ud
appl
i
c
ations,
and
c) con
s
ide
r
ing
more fa
ctors to
identify signifi
cant compo
n
ents.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1565 – 157
4
1574
Ackn
o
w
l
e
dg
ements
The
wo
rk
de
scribe
d in
th
is p
ape
r
wa
s
sup
porte
d
by the Natio
nal Natural
Scien
c
e
Found
ation o
f
China
(No. 6117
3020
)
and
Chine
s
e
Nation
al Progra
m
s fo
r
High T
e
chnol
ogy
Re
sea
r
ch an
d Develo
pme
n
t (No. 201
3
AA01A215
).
Referen
ces
[1]
Armbrust M, Fox A, Griffith
R, Joseph AD,
Katz RH, Konw
inski
A, Le
e
G, Patterson
D, Rabk
in A,
Stoica I, Z
ahari
a
M. A Vie
w
of
Clou
d
Com
puti
ng.
Co
mmu
n
ic
ations of the A
C
M
. 2010; 53(
4): 50-58.
[2]
Prasad MP, Naik PR, Bapu
ji
V. Cloud Co
m
putin
g: Rese
arch Iss
ues an
d Implicati
ons.
International
Journ
a
l of Clo
u
d
Co
mp
uting
a
nd
Servic
es Scienc
e (IJ-CLOSER)
. 2013; 2(
2): 134-1
40.
[3]
Z
heng Z
,
Z
h
a
n
g
Y, L
y
u M
R
.
Clou
d
R
ank: A
QoS-Drive
n
C
o
mpo
nent
Ran
k
ing F
r
a
m
ew
ork for Clo
u
d
Co
mp
uting
. 2
0
10 29th IEEE
S
y
mp
osi
u
m on
Relia
bl
e Distri
buted S
y
stems
.
Ne
w
D
e
l
h
i. 20
10: 184-
19
3.
[4]
Rooney
P. Microsoft's CEO: 8
0
-20 Ru
le A
ppl
ies to Bugs, No
t Just F
eatures.
Chann
elW
e
b
.
2002.
[5]
Z
heng
Z
,
Z
hou
T
C
, L
y
u
MR,
King
I.
F
T
Clou
d
: A C
o
mp
one
nt Ra
nkin
g F
r
a
m
ew
ork f
o
r F
a
ult-T
o
ler
a
n
t
Clou
d
Ap
plic
at
ions
. 2
0
1
0
IEEE 21st Intern
ation
a
l S
y
mp
o
s
ium o
n
Soft
w
a
re R
e
li
abi
lit
y
Engi
neer
in
g.
San Jose. 2
0
1
0
: 398-4
07.
[6]
Avizie
nis A, Laprie JC, Ran
d
e
ll B, Lan
d
w
e
h
r
C. Basic Concepts and T
x
o
nom
y
of Dep
e
nda
ble a
n
d
Secure Computing.
IEEE Transactions on Dependa
ble and
Secure Com
p
uting
. 200
4; 1(1)
: 11-33.
[7]
Ly
u
MR
.
Han
d
book of Soft
w
a
re Reli
ab
ilit
y E
ngi
neer
in
g. Ne
w
York: McGra
w
-
H
il
l. 199
6.
[8]
Ran
del
l B, Jie
X. T
he Evol
uti
on of th
e R
e
co
ver
y
Block
Co
ncept. In: L
y
u
MR.
Editor
. Soft
w
a
re Fault
T
o
lerance. Ne
w
York: W
ile
y.
199
5: 1-21.
[9]
Avizie
nis A. T
he Met
hod
ol
o
g
y
of N-V
e
rsi
on Pro
g
rammi
ng. In: L
y
u
M
R
.
Editor
. Sof
t
w
ar
e F
a
u
l
t
T
o
lerance. Ne
w
York: W
ile
y.
199
5: 23-4
6
.
[10]
Z
eng
L, Ben
a
talla
h B, N
g
u
A
HH, Kal
a
g
nan
a
m
J, Cha
ng
H.
QoS-a
w
ar
e Mi
ddl
e
w
ar
e for W
eb S
e
rvices
Comp
ositio
n.
IEEE Transactions on So
ftwar
e Engineering
. 200
4; 30(5): 31
1-27.
[11]
Maximil
i
en EM
, Singh MP. C
once
p
tu
a
l
Mod
e
l of W
eb S
e
r
v
ice Re
putati
o
n.
SIGMOD Record
. 20
02
;
31(4): 36-
41.
[12]
Naji-
Hasa
n AH
, Shu G.
OW
L
S
-CSM: A Service Profile Ba
sed Simil
a
rit
y
F
r
ame
w
ork for
W
eb Service
Discover
y
.
T
E
L
K
OMNIKA T
e
leco
mmu
n
icati
o
n C
o
mputi
n
g
E
l
ectron
ics a
n
d
Contro
l
.
20
14; 12(4): 107
9-
108
7.
[13]
Lu G, Hai Y,
Sun Y. A Relia
bl
e W
eb
Servic
es Se
le
ction Meth
od
for Concurr
ent
Requ
ests.
T
E
LKOMNIKA T
e
leco
mmunic
a
tion C
o
mputi
n
g Electron
ics a
nd Co
ntrol
. 20
14; 12(4): 1
053
-106
3.
[14]
Che
ung R
C
. A User-Oriented
S
o
ft
w
a
re
Re
lia
bil
i
t
y
M
ode
l.
IEEE T
r
ansactions
on Software
Engi
neer
in
g
. 1980; SE-6(2): 1
18-1
25.
[15]
Page
L, Brin S,
Mot
w
a
n
i
R, W
i
nogr
ad T
.
T
he Page
Rank
Cita
t
ion R
anki
ng: B
r
ingi
ng Ord
e
r t
o
the W
e
b
.
Standford Info
Lab. Re
port nu
mber: 199
9-6
6
. 1999.
[16]
G
y
o
n
g
y
i Z
,
Ga
rcia-Mol
i
na
H,
Peders
en J.
C
o
mbati
ng W
e
b
Spa
m
w
i
th T
r
u
s
trank
. Proce
e
d
in
gs of the
T
h
irtieth Internation
a
l Co
nfere
n
ce on Ver
y
La
r
ge Data
bas
es (VLDB). T
o
ronto. 2004; 3
0
: 57
6-58
7.
[17]
Cortell
e
ssa V,
Grassi V. A
Mode
lin
g Ap
pr
oac
h t
o
An
al
yze the Imp
a
ct of Error Pro
pag
atio
n o
n
Reli
ab
ilit
y of C
o
mpo
nent-B
as
ed S
y
stems. In: Schmid
t H
W
, Crnkovic I, Hei
nema
n
GT, Stafford JA.
Editors
. Comp
one
nt-Base
d
Soft
w
a
re En
gi
neer
ing. Ber
lin
, Heide
l
b
e
rg: Sprin
ger Ber
l
i
n
Hei
del
ber
g
.
200
7: 140-
156.
[18]
Brin S, Page
L. Reprin
t of:
T
he anatomy of a Larg
e
-
scale H
y
p
e
rte
x
tu
al W
eb Se
arch Eng
i
ne.
Co
mp
uter Net
w
orks
. 2012; 56(18): 38
25-
38
33.
[19]
Inoue K, Yoko
mori R, Yama
moto T
,
Matsushita
M, Kusu
moto S. Ranki
ng Sig
n
ific
anc
e of Soft
w
a
r
e
Comp
one
nts Based o
n
Use
Relati
ons.
IEEE Transactions on So
ftware Engineering
. 200
5;
31(3)
:
213-
25.
[20]
F
a
loutsos M,
F
a
loutsos P,
F
a
loutsos
C. On Po
w
e
r-
La
w
R
e
l
a
tio
n
shi
p
s of the Inter
net T
opolo
g
y
.
SIGCOMM Comp
uter Co
mmunic
a
tion R
e
vi
ew
. 1999; 29(4
)
: 251-26
2.
[21]
D H
y
l
a
n
d
-W
oo
d, D Carringto
n
, Y Kaplan.
Scale-F
r
e
e
Natu
re of Java Softw
are Package,
Class an
d
Method
Co
lla
b
o
ratio
n
Grap
hs
. T
he 5th Inter
natio
nal
S
y
mp
osium
on Em
pi
rical S
o
ft
w
a
re
Engi
neer
in
g.
Rio d
e
Jan
e
iro.
2005: 4
39-4
4
6
.
[22]
M
y
ers
CR. Softw
a
r
e S
y
stems
as Com
p
le
x
N
e
t
w
o
r
ks:
Structure, Functio
n
a
nd
Evo
l
vab
ilit
y of
Soft
w
a
r
e
Coll
ab
oratio
n Graphs.
Physic
a
l Review E
. 2003; 68(
4): 46-
116.
[2
3
]
D
e
N
ooy
W, Mrva
r
A, Ba
ta
g
e
l
j
V. Ex
pl
o
r
a
t
o
r
y
So
cia
l
Ne
tw
ork Ana
l
y
s
is
w
i
th
Pa
je
k. R
e
vi
sed
and
Exp
a
n
d
e
d
Sec
ond Ed
itio
n. Ne
w
Y
o
rk: Cam
b
rid
ge Un
iversi
t
y
Press. 201
1.
[24]
Pham T
T
, Xe
fago D.
Rel
i
a
b
ility Pr
edicti
o
n for C
o
mpo
n
ent-Bas
e
d
Sy
stems: Incor
p
oratin
g Error
Propa
gati
on A
nalysis
an
d Dif
ferent Executi
o
n Mod
e
ls
. 20
1
2
12th Inter
nati
ona
l Co
nfere
n
c
e
on Qu
alit
y
Soft
w
a
re (QSIC). Xi’
an. 20
12
: 106-11
5.
Evaluation Warning : The document was created with Spire.PDF for Python.