Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
6, N
o
. 4
,
A
ugu
st
2016
, pp
. 15
60
~
1
569
I
S
SN
: 208
8-8
7
0
8
,
D
O
I
:
10.115
91
/ij
ece.v6
i
4.1
067
6
1
560
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Building Fault Tolerance within
Clouds at Network Level
DBK
Kame
sh
, J
K
R
S
a
str
y
,
Ch.
De
vi
Anu
s
ha,
P.
P
a
dmi
n
i, G.
Siv
a
An
jan
e
yul
u
Department o
f
Electronics
and
C
o
mputers Engineering, KL Univ
ersity
, Vadd
esew
aram, Ind
i
a
Article Info
A
B
STRAC
T
Article histo
r
y:
Received
Mar 29, 2016
Rev
i
sed
May 18
, 20
16
Accepted
May 30, 2016
Cloud computin
g techno
logies and infrast
ructure facilities ar
e co
ming up in a
big wa
y m
a
king
it cost
effe
ctiv
e
for the users to
im
plem
ent the
i
r IT based
solutions to run
business in most ec
onomical
way
.
Man
y
intr
icate issues
however, h
a
ve cropped-up which must be
addr
essed to be able to use clouds
the purpose for
which they
are
designe
d and im
plemented
.
Among all, fault
toler
a
nce and s
ecuring
the data stor
ed on
th
e clouds takes
most of the
im
portance. Co
ntinuous availab
ilit
y
of
the serv
ices is dep
e
ndent on m
a
n
y
factors. Faults b
ound to happ
en
within
a network, software,
and
platfo
rm or
within th
e infr
astructure which
are
a
ll used
for
establishing
the
cloud.
The
network that co
nnects various servers,
dev
i
ces,
peripher
a
ls etc., have to be
fault to
ler
a
nt to
start-with so tha
t
intend
ed and u
n
-interrup
t
ed ser
v
ices to t
h
e
user can be made available. A novel
network d
e
sign method that leads to
achieve
high availability
of the
n
e
twork
and
th
ere
b
y
the
cloud
i
t
s
e
lf has
b
een
presented
in
this
paper.
Keyword:
B
u
t
t
e
rfl
y
t
o
pol
ogy
C
l
ou
d c
o
m
put
i
n
g
Fau
lt tree
Reliab
ilit
y
Copyright ©
201
6 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
JKR Sast
ry,
Depa
rt
em
ent
of El
ect
r
oni
cs
a
n
d
C
o
m
put
er
E
ngi
neeri
n
g
,
KL Uni
v
er
sity
,
Vad
d
es
waram
,
G
unt
ur
Di
st
ri
c
t
, I
ndi
a
Em
a
il: d
r
sastry@k
l
u
n
i
v
e
rsity.in
1.
INTRODUCTION
Clo
u
d
h
a
s
b
e
en
a m
a
j
o
r p
l
atfo
rm
h
o
s
ting
d
i
fferen
t
k
i
nd o
f
serv
ices fo
r
facilitatin
g
in
fo
rm
atio
n
pr
ocessi
ng
. It
i
s
bei
n
g
use
d
q
u
i
t
e
ext
e
nsi
v
el
y
even
w
h
e
n
c
o
m
p
ared t
o
gri
d
c
o
m
put
i
n
g
,
a
t
y
pe
of c
o
m
put
i
n
g
w
h
er
e
u
n
u
s
ed
p
r
o
cessi
n
g
cycles o
f
all co
mp
u
t
er
s in a netw
or
k ar
e h
a
r
n
essed
t
o
so
lv
e p
r
ob
lem
s
. I
n
cloud
com
puting, the
word
“cloud”
is
used as a m
e
taphor
for
Inte
rnet
and i
n
a
way cloud com
puting can
be te
rm
ed
as at
t
y
pe o
f
i
n
t
e
rnet
based
co
m
put
i
ng. C
l
o
u
d
c
o
m
put
i
ng
p
r
o
v
i
d
es
di
ffe
re
nt
t
y
pes
of
ser
v
i
ces s
u
ch
as s
u
ch
as
serve
r
s, storage and a
pplications
whic
h are
del
i
v
ere
d
t
o
c
o
m
put
er an
d
de
vi
ces with
wh
i
c
h
th
e
users
int
e
racts.
The g
o
al
o
f
cl
ou
d com
put
i
ng i
s
t
o
a
p
pl
y
t
r
adi
t
i
onal
supe
rc
om
put
i
n
g
,
o
r
hi
gh
-
p
erf
o
rm
ance
co
m
p
u
tin
g power,
n
o
rm
all
y
u
s
ed
b
y
m
i
l
itary an
d re
search facilities, to
p
e
rfo
rm
ten
s
o
f
trillio
ns of
com
put
at
i
ons
per sec
o
n
d
, i
n
co
nsum
er-
o
ri
ent
e
d a
ppl
i
c
atio
n
s
su
ch
as fin
a
n
c
ial po
rtfo
lio
s, to
d
e
liv
er
pers
onalized inform
ation, t
o
provide
d
a
ta st
orag
e
or to
po
wer larg
e, imm
e
r
s
iv
e on
lin
e com
p
u
t
er g
a
m
e
s. To
do
t
h
i
s
, cl
ou
d co
m
put
i
ng uses
net
w
or
k o
f
l
a
rge gr
o
up
of s
e
rve
r
s t
y
pi
cal
l
y
run
n
i
n
g l
o
w
-
cost
co
ns
um
er PC
technology with specialized
connections to spread
data-processi
ng c
h
ores ac
ross the
m
. This shared IT
i
n
fra
st
ruct
ure
cont
ai
n
s
l
a
r
g
e
po
ol
s
o
f
sy
st
e
m
s t
h
at
are l
i
n
ked
t
o
get
h
e
r
.
Oft
e
n,
vi
rt
ual
i
zat
i
on t
e
c
hni
q
u
es a
r
e
use
d
to m
a
ximize the power of cloud c
o
m
puting.
The n
e
t
w
or
k p
l
ay
s a prom
i
n
ent
rol
e
i
n
cl
ou
d i
n
f
r
ast
r
uct
u
r
e
s. C
l
ou
d c
o
m
put
i
n
g
uses t
h
e net
w
or
k t
o
gain on-dem
a
n
d
access
t
o
com
puting res
o
urces,
an
d t
h
e
network
becom
e
s the conduit for e
n
orm
ous
co
m
p
u
tin
g
capab
ility. Th
is critical ro
le o
f
the n
e
two
r
k
in
cl
o
u
d
co
m
p
u
tin
g d
e
m
a
n
d
s
th
at n
e
two
r
k
is right an
d
th
e n
e
twork
m
u
st b
e
co
n
f
ig
ured
t
o
achiev
e
th
e d
e
sired
lev
e
l of
perfo
r
m
a
n
ce, secu
rity, av
ailab
ility,
respon
siv
e
n
e
ss, and
m
a
n
a
g
e
ab
ility.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Building F
a
ult
Toler
ance
within Clo
uds
at N
e
twork Level (
D
BK K
a
mes
h
)
1
561
Fau
lts can o
c
cu
r
with
in th
e
n
e
two
r
k
s
and
as a resu
lt th
e n
e
two
r
k
fails. Th
e
failu
re of a
n
e
two
r
k
co
nn
ecting
th
e clo
u
d
related
i
n
frastru
c
ture lead
s to
d
i
srup
tio
n
i
n
th
e serv
ices p
r
ov
id
ed
t
o
th
e
u
s
er wh
i
c
h
m
a
y
u
lti
m
a
tel
y
affect th
e
b
u
s
i
n
esses b
e
i
n
g conducted
b
y
th
e
u
s
ers.
Th
e 100
% av
ailab
ility o
f
t
h
e
n
e
two
r
k
is
o
n
e
of
t
h
e m
o
st
im
port
a
nt
i
s
s
u
es t
h
at
m
u
st
be
ha
n
d
l
e
d
f
o
r
p
r
o
v
i
d
i
n
g
hi
g
h
l
y
rel
i
abl
e
an
d c
o
nt
i
nue
d
ser
v
i
ces
t
o
t
h
e
user
.
Th
e prob
ab
ility o
f
failure o
f
a
n
e
twork d
u
ring
th
e
d
e
l
i
v
e
ry of th
e serv
ices is
n
o
rmally
m
o
re esp
e
cially
whe
n
t
h
e
n
u
m
b
er
of
use
r
s i
n
creases.
It
nea
r
l
y
im
possi
bl
e to
pr
e
v
en
t f
a
ilu
r
e
s
and
th
a
t
h
a
pp
en
a
t
ru
n-ti
me
.
Sin
ce it is in
ev
itab
l
e th
at
fau
lts can o
c
cu
r, it b
eco
m
e
s ne
cessary
to
m
a
ke Ha
rd
ware
,
N
e
two
r
k
an
d
So
ftware
t
h
at
fo
rm
a cl
oud
t
o
be m
a
de faul
t
t
o
l
e
ra
nt
.
Faul
t
t
o
l
e
ra
nce
has
bec
o
m
e
a m
a
jor t
a
s
k
fo
r
com
put
er e
ngi
neer
s
and
s
o
ft
wa
re
devel
ope
rs
bec
a
use t
h
e
occ
u
r
r
ence
o
f
fa
ults
increa
ses
the cost of
usi
n
g resources. Als
o
the
p
r
ob
le
m a
r
is
e
w
h
en
a s
e
rv
er
is
ov
e
r
lo
ad
ed
,
a
s
e
rv
e
r
or
a
n
ode
fai
l
e
d
et
c.
The
Faul
t
s
m
u
st
be
ha
ndl
e
d
f
i
rst
so
th
at th
e
n
o
rmal o
p
e
ration
of cloud
is
not
effecte
d
a
n
d
then a
r
e
rectified to
bring t
h
e system
to norm
a
l
o
p
e
ration
.
Du
rin
g
th
e tim
e
wh
en
th
e fau
lts are h
a
nd
led,
it
is p
o
ssib
l
e
th
e syste
m
b
e
o
p
e
ration
with fewer
th
ro
ugh
pu
ts, perfo
r
m
a
n
ce, resp
on
se tim
e e
t
c., till su
ch
ti
me
th
e fau
lt are fu
lly rectified
an
d
b
r
i
g
th
e syste
m
to
no
rm
al
operat
i
on
y
i
el
di
ng
t
h
e
desi
gne
d c
h
ar
act
eri
s
t
i
c
t
o
f
u
l
l
e
st
ext
e
nt
.
Thus the m
o
st
im
portant thing is to
m
a
ke the netw
ork
fault tolerant in the first place. T
h
e network
con
n
ect
i
v
i
t
y
, t
o
p
o
l
o
gy
,
use o
f
p
r
ot
ocol
s
,
p
r
ot
oc
ol
co
nve
rs
i
ons
, r
out
i
n
g,
con
g
est
i
o
n c
o
n
t
rol
m
u
st
be d
one i
n
su
ch
a way th
at altern
ate p
a
th
s ex
ists for data tran
sm
i
ssi
on bet
w
ee
n t
h
e user a
nd t
h
e cl
ou
d an
d vi
ce vers
a
.
The c
o
m
m
uni
cat
i
on ca
n t
h
e
n
be m
a
de t
o
m
ove i
n
s
h
o
r
t
e
st
pat
h
p
o
ssi
bl
e.
Once
t
h
e
pat
h
t
h
at
has
f
a
i
l
e
d i
s
rect
i
f
i
e
d,
t
h
e c
o
m
m
uni
cat
i
on
can
be m
a
de t
o
m
ove i
n
t
h
e s
c
hem
e
of
ori
g
i
n
al
de
si
g
n
.
I
n
t
h
i
s
pape
r,
a m
e
t
h
o
d
of m
a
ki
ng t
h
e
net
w
or
k fa
ul
t
tol
e
ra
nt
by
im
pl
em
ent
i
ng B
u
t
t
e
rfl
y
t
opol
ogy
has bee
n
pres
ent
e
d an
d i
t
has bee
n
sho
w
n
ho
w t
h
e
net
w
o
r
k
has
b
een m
a
de t
o
be
fa
ul
t
t
o
l
e
ra
nt
.
2.
RELATED WORKS
F. T
hom
son
L
e
i
ght
o
n
,
an
d B
r
uce M
.
B
a
g
g
s
c
[1]
has
desc
ri
be
d
basi
c d
e
t
e
rm
i
n
i
s
t
i
c
al
gori
t
h
m
s
fo
r
routing Thes
e algorithm
s
are
vigorous
agai
nst faults eve
n
in worst case a
nd
a
r
e effective from
practica
l
point
o
f
v
i
ew. Th
ey fo
und
th
at
m
u
l
ti-b
u
tter-fl
y is an
ex
cellen
t
cand
i
d
a
te
forach
ei
v
i
ng
hig
h
b
a
ndwid
th, lo
w
di
am
et
er swi
t
c
hi
n
g
net
w
or
k e
t
c.
W. Sh
i,
P. K. Srim
an
ic
[2
]
h
a
s
ex
pre
ssed
t
h
e use o
f
but
t
e
rfl
y
n
e
t
w
or
k i
n
V
L
SI
poi
nt
of
vi
e
w
a
n
d
according to t
h
em
, interconnected net
w
orks
can ha
ve
only fixe
d num
b
er
of i
n
puts a
n
d out
puts
.
They have
ex
p
l
ain
e
d th
e
d
r
awb
ack
s of
su
ch
a
n
e
twork
esp
ecially
th
e in
ab
ility to
ach
iev
e
th
e desired
lev
e
l
o
f
fau
l
t
t
o
l
e
rance
base
d o
n
t
h
e
num
b
e
r o
f
n
o
d
es co
nt
ai
ned i
n
t
h
e
net
w
or
k. F
o
r l
a
rge
net
w
or
ks
t
h
e desi
gni
ng
of t
h
e
no
des
is
pr
obl
em
atic. To
ov
ercom
e
the
dr
awbac
k
s
s
een
in
b
u
tter-fly
netw
or
ks,
they
ha
ve im
plem
ente
d
hy
pe
rcu
b
e
net
w
o
r
k
t
o
re
duce
t
h
e
faul
t
s
.
Jin-F
u
Li
et
.,
al. [3], expres
sed a
ve
ry large scale i
n
t
e
g
r
at
i
o
n
t
ech
nol
ogy
k
n
o
w
n as
fast
fo
uri
e
r
tran
sform
n
e
twork (FFT) into
a si
n
g
l
e ch
i
p
.
Actu
ally th
i
s
chi
p
is
very
big. T
o
get effectiveness
of t
h
e c
h
ip
t
h
ey
ha
ve rec
o
m
m
e
nded
usi
n
g fa
ul
t
t
o
l
e
ra
nc
e net
w
or
k.
A
n
e
t
w
o
r
k “C
-t
est
a
bl
e FFT
” ha
s
been
desi
g
n
ed
whi
c
h
sh
ow
h
i
gh
er reliab
ility an
d
the n
e
ed
fo
r lower h
a
rd
ware.
Th
e co
m
b
in
atio
n
o
f
si
n
g
l
e cell
fau
lt in
terconn
ected
n
e
two
r
k
d
e
p
e
nd
s
o
n
t
h
e size
o
f
t
h
e testin
g
p
a
ttern. It
h
a
s
b
een
shown
t
h
at a fau
lty row in
m
u
ltip
le su
b
t
racts
an
d add
s
can be rep
a
ired b
y
3-b
it lev
e
l cell.
Richard L
.
Gra
h
am
, et. al. [4], co
nvey that the LA-MPI (L
os Alam
os
message passi
ng int
e
rface
) is a
peer t
o
pee
r
net
w
or
k fa
ul
t
t
o
l
e
rant
sy
st
em
desi
gned s
p
eci
fi
cal
l
y
for
t
h
e Tera scal
e cl
ust
e
rs. Th
ey
have
p
r
esen
ted a sy
ste
m
wh
ich
is h
i
gh
ly to
lerant to
erro
r rel
a
t
e
d t
o
net
w
o
r
k
s
, net
w
o
r
k t
r
a
n
sm
i
ssi
on err
o
rs a
n
d
wi
re
d-
net
w
or
k
erro
rs. L
A
-M
PI, su
p
p
o
r
t
e
d
m
u
lt
i
l
a
y
e
red ne
twork
in
terfaces. LA-M
PI’s main
featu
r
e is th
at it
can
tran
sm
it
messag
e
s in
a rel
i
ab
le
way throu
g
h
m
u
ltip
le network p
a
t
h
s.
C
hua
n
x
i
o
ng
G
u
o
et
. al
.
[
5
]
,
e
x
p
r
esse
d t
h
at
t
h
e
basi
c c
h
al
l
e
nge
i
n
net
w
o
r
k
i
ng i
s
h
o
w
t
o
i
n
t
e
rc
on
nec
t
exp
o
n
ent
i
a
l
l
y
increasi
ng
num
ber o
f
ser
v
er
s and cl
i
e
nt
s ve
ry efficiently. They have
prese
n
t
e
d DC
el
l
,
a no
vel
net
w
or
k st
ruct
ure
t
h
at
has
t
h
e re
qui
red
feat
ures
f
o
r
net
w
o
r
ki
ng
.
DC
el
l
i
s
a r
ecu
rsi
v
el
y
defi
ned
st
r
u
ct
u
r
e, i
n
wh
ich
a
h
i
g
h
-lev
el DCell is
co
nstru
c
ted
from
man
y
lo
w-l
e
v
e
l Dcells. DCell scales d
o
u
b
l
y ex
pon
en
ti
ally as
th
e no
d
e
d
e
gree in
creases.
DCell is fau
lt to
leran
t
, since it d
o
e
s no
t h
a
v
e
si
n
g
l
e
po
in
t of failure an
d
its
di
st
ri
b
u
t
e
d
fau
l
t
-
t
o
l
e
rant
r
o
ut
i
ng
pr
ot
oc
ol
p
e
rf
orm
s
near s
h
o
r
t
e
st
-
p
at
h r
o
ut
i
ng e
v
e
n
i
n
t
h
e p
r
ese
n
ce o
f
severe
lin
k
o
r
nod
e failu
res. DCell
also
p
r
ov
id
es
h
i
gh
er
n
e
t
w
or
k cap
acity th
an th
e trad
ition
a
l tree-b
ased
stru
ct
u
r
e fo
r
vari
ous t
y
pes
of ser
v
i
ces.
Furt
he
rm
ore,
DC
el
l
can
be ex
pan
d
e
d
.
R
e
sul
t
s
fro
m theoretical analysis,
si
m
u
latio
n
s
, sho
w
th
at
Dcell is a
v
e
ry
reliab
l
e in
terco
n
n
ectio
n stru
cture for
d
a
ta cen
t
r
es.
Vi
nce
n
t
Li
u
et
.
al
. [
6
]
,
e
x
pres
sed t
h
at
,
i
n
cl
ou
d c
o
m
put
i
n
g
,
t
h
e
dat
a
cent
r
e net
w
o
r
ki
ng
i
s
ev
ol
vi
ng
into hi
ghly cos
tly, reliable, and hi
gh
perform
a
nce co
m
puting. E
v
en t
h
ough m
u
lti-tree topol
ogies ca
n provide
scal
abl
e
ba
nd
wi
dt
h a
n
d t
r
a
d
i
t
i
onal
r
o
ut
i
n
g al
g
o
ri
t
h
m
s
can provide e
v
entu
al fa
ult tolerance, t
h
e re
cove
ry
spee
d can
be v
e
ry
hi
g
h
wi
t
h
c
o
m
b
i
n
at
i
on an
d desi
gni
ng
of
vari
ous
net
w
or
k t
o
pol
ogi
es,
r
out
i
n
g al
g
o
ri
t
h
m
and
faul
t
det
ect
or
. They
hav
e
pre
s
ent
e
d a n
ovel
net
w
o
r
k t
o
p
o
l
ogy
t
h
at
has al
l
t
h
e desi
rabl
e
charact
eri
s
t
i
c
s of
a
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
60
–
1
569
1
562
faul
t
t
o
l
e
ra
nt
t
o
p
o
l
o
gy
an
d h
a
s very
hi
gh s
p
eed
reco
ve
ry
by
im
pl
em
ent
i
ng a
fai
l
ove
r p
r
ot
ocol
w
h
i
c
h
hel
p
i
n
connecting t
o
t
h
e
network ins
t
antly even
aft
e
r
occu
rre
nce
of
a
fai
l
u
re
an
d al
s
o
hel
p
i
n
l
o
ad
bal
a
nci
n
g
on
t
h
e
net
w
or
k.
Huas
ha Z
h
a
o
,
and J
o
hn C
a
n
n
y
[7]
has re
f
e
rre
d i
n
c
r
em
ent
a
l
m
odel
s
t
h
at
are u
pdat
e
d
m
a
ny
t
i
m
e
s
usi
n
g sm
al
l subset
s o
f
t
r
ai
ni
ng
dat
a
. The
m
odel
present
e
d by
t
h
em
supp
o
r
t
s
bot
h st
ocha
st
i
c
gra
d
i
e
nt
an
d
M
C
M
C
(M
ark
o
c
h
ai
n m
a
nt
e corl
o) m
e
ssagi
ng
an
d
hel
p
s i
n
f
a
st
seq
u
e
n
t
i
a
l
per
f
o
r
m
a
nce b
u
t
i
t
can
not
han
d
l
e
paral
l
e
l
o
r
cl
u
s
t
e
r set
t
i
ngs.
They
ha
ve e
x
press
e
d t
h
at
b
u
t
t
e
rfl
y
m
i
xi
ng ap
p
r
oac
h
es
l
eads t
o
i
n
t
e
rl
eaved
com
m
uni
cat
i
on an
d com
put
at
i
on. T
h
ey
hav
e
eval
uat
e
d b
u
t
t
e
rfl
y
m
i
xi
ng
on st
oc
hast
i
c
g
r
adi
e
nt
al
gori
t
hm
s
t
o
get
l
ogi
st
i
c
re
g
r
essi
o
n
.
It
has
been s
h
ow
n t
h
at
but
t
e
rfl
y
m
i
xed st
e
p
s are
f
a
st
and fai
l
ure
t
o
l
e
rant
.
3.
3x
spee
ds
whi
c
h a
r
e m
o
re t
h
an
a
ful
l
m
i
x
on
an
Am
azon EC
2 cl
u
s
t
e
r
have
bee
n
a
c
hi
eved
.
R
a
vi
J
h
awa
r
et
. al
.
[8]
,
p
r
op
ose
d
a
c
o
m
p
re
he
nsi
v
e
ap
pr
oac
h
fo
r i
m
pl
em
ent
a
t
i
on of
hi
gh
l
e
ve
l
tech
n
i
qu
es
f
o
r
f
a
u
lt to
ler
a
n
ce. I
n
th
e ap
pr
o
a
ch
pr
esen
ted
b
y
th
em th
e u
s
er
s n
eed
n
o
t
know
th
e f
a
u
lt to
ler
a
n
c
e
requ
irem
en
ts of th
eir app
licatio
n, th
ey
wo
u
l
d lik
e to
ju
st
k
n
o
w
how th
e fau
lt to
leran
ce
has b
e
en
im
p
l
emen
ted
.
Mo
h
a
m
e
d
Abu
Sh
ark
h
et.
al. [9
], h
a
s ex
pressed
th
at
clo
u
d
co
m
p
u
tin
g
is a u
tility p
r
o
cessing
para
di
gm
t
h
at
has t
u
rne
d
i
n
t
o
a st
r
o
n
g
bas
e
fo
r wi
de ex
hi
bi
t
o
f
en
d-cl
i
e
nt
appl
i
cat
i
o
ns. P
r
ovi
ders
’
of
ft
en
chan
ge
p
o
rt
f
o
l
i
o
s
of t
h
e em
pl
oy
ees w
h
i
c
h re
qui
re di
ffe
rent
ki
n
d
of
ser
v
i
ces. A
n
e
x
cel
l
e
nt
res
o
u
r
ce al
l
o
t
m
ent
m
o
d
e
l is th
e
k
e
y to
an
y clou
d
co
m
p
u
tin
g syste
m
. An
y
asset
allo
catio
n
m
o
del n
e
ed
s to co
nsid
er co
m
p
u
t
atio
n
a
l
assets as
well as system
assets
to
p
r
ecisely mak
e
th
e p
e
op
le
ad
h
e
r t
o
th
ei
r respo
n
s
i
b
ilities.
S. Gi
ri
esh et
.
a
l
. [1
0]
,
have
p
r
esent
e
d
va
ri
o
u
s
t
y
pes o
f
fai
l
u
r
es t
h
at
ca
n
ha
ppe
n
whi
l
e
e
n
f
o
rci
ng t
h
e
cl
ou
d com
put
i
ng a
nd m
a
i
n
tai
n
i
ng t
h
e sy
s
t
em
. Fail
uers
can hap
p
e
n
at
di
ffere
nt
l
e
vel
s
w
h
i
c
h i
n
cl
ud
e
com
pone
nt
fai
l
u
res
,
net
w
o
r
k
fai
l
u
res a
n
d se
curi
t
y
fai
l
u
re
s
whi
c
h al
l
m
u
st be co
nsi
d
ere
d
and
deal
t
wi
t
h
t
o
make the entire syste
m
of cloud c
o
m
p
u
ting
effectiv
e and
fail safe. To re
d
u
ce t
h
e fau
lts, a n
e
w fau
lt to
llren
ce
mech
n
a
sim
cal
led
co
llaborativ
e fau
lt to
lera
nce m
echanis
m
(CFTM)
has
been
in
t
r
odu
ced. In
th
is m
ech
an
ism
,
t
h
e dat
a
i
s
upl
o
a
ded i
n
t
o
t
h
e cl
ou
d u
nde
r aut
h
ent
i
cat
i
o
n by
the users. The
r
e is less ch
ance of loss at the serve
r
end
by
fra
gm
ent
i
ng a
nd
re
pl
i
cat
i
ng i
n
t
o
t
h
e vi
rt
ual
st
ora
g
es. T
h
e m
a
i
n
key
feat
ures o
f
C
F
T
M
are
au
th
en
ticatio
n
an
d
d
a
ta reco
very.
P. Pa
dm
akum
ari
an
d
A.
U
m
am
akeswari
[1
1]
ha
ve
des
c
ri
be
d t
h
at
t
h
e
cl
ou
d c
o
m
put
i
ng c
o
ul
d
be
m
a
de feasi
b
l
e
by
usi
n
g fa
ul
t
t
o
l
e
ra
nce an
d m
oni
t
o
ri
n
g
ser
v
i
ces. They
ha
ve
descri
b
e
d t
w
o
m
easures w
h
i
c
h are
proactive a
n
d reactive that takes place
with-i
n cloud. For cl
oud
provide
r
a
n
d cloud c
u
st
omers, fa
ult tolerance
i
s
im
port
a
nt
as
t
h
ey
are pre
-
r
e
qui
si
t
e
s f
o
r
pr
ovi
di
n
g
co
nt
i
n
uo
us se
rvi
ces
. In t
h
ei
r fi
n
d
i
n
gs t
h
ey
has
de
scri
b
e
d
that the
reliability of the
cloud by
diffe
re
nt or
dive
rse
fa
ult tolera
nce m
e
thods c
o
uld
be i
n
c
r
eased.
3.
A
SSESSIN
G FA
ILUR
E RA
TE
OF
AN
EX
ISTIN
G
CLOU
D
An
ex
istin
g
cl
o
u
d
h
a
s b
e
en
co
n
s
i
d
ered
fo
r effectin
g
fau
lt t
o
leran
ce and
in
crease the reliab
ility o
f
th
e
sam
e
.
Th
e
m
o
re th
e reliab
ility
o
f
a n
e
two
r
k
co
nn
ecting
a clo
u
d
,
th
e m
o
re co
n
tinu
e
d
services can
b
e
p
r
ov
id
ed
to
th
e end
-
u
s
ers. Th
e
n
e
two
r
kin
g
d
i
agr
a
m
w
h
ich
fo
llow
s
a
tr
ee lik
e to
po
log
y
and
conn
ects a u
n
i
v
e
r
s
ity cloud
i
s
sh
ow
n i
n
t
h
e
Fi
g
u
re
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Building F
a
ult
Toler
ance
within Clo
uds
at N
e
twork Level (
D
BK K
a
mes
h
)
1
563
Fig
u
r
e
1
.
Topolo
g
y
o
f
a u
n
i
v
e
r
s
ity
clo
ud
The users
of t
h
e cloud a
r
e c
o
nnected t
h
rough i
n
te
rnet c
o
nnections that ar
e pr
ov
id
ed
by BSN
L
and
NK
N. T
h
e
ban
d
wi
dt
h f
r
o
m
b
o
t
h
t
h
e
net
w
or
ks i
s
r
out
e
d
t
h
r
o
u
g
h
a com
m
o
n
swi
t
c
h a
n
d C
Y
B
E
R
C
A
M
s
w
i
t
c
h
.
The ba
n
d
wi
dt
h
from
t
h
i
s
poi
nt
i
s
di
st
ri
but
e
d
i
n
4 c
h
an
nel
s
. In
one
of t
h
e chan
nel
s
6
HP B
l
ade ser
v
ers an
d
5TB
di
s
k
st
ora
g
e co
nne
ct
ed t
h
r
o
ug
h an
INF
I
N
I
TE s
w
i
t
c
h has bee
n
co
n
n
ect
ed t
o
fo
rm
int
o
cl
o
u
d
. O
n
e
of t
h
e
bl
ades
ha
s
bee
n
use
d
as
t
h
e
M
i
ddl
ewa
r
e se
rve
r
whi
l
e
othe
r se
rve
r
s
are
us
ed as
wi
ndows
-
Oracle se
rve
r
; UNIX
base
d serve
r
,
wi
n
d
o
w
s
-
S
Q
L
server a
nd
o
t
her ap
pl
i
cat
i
on ser
v
ers
.
The
perf
orm
a
nce of t
h
e ent
i
r
e
cl
ou
d
com
put
i
ng
pl
a
t
form
i
s
based
o
n
t
h
e
pr
o
p
er
fu
nct
i
o
ni
n
g
o
f
t
h
e C
Y
B
E
R
C
AM
swi
t
c
h
.
A fa
ul
t
t
r
ee
ha
s bee
n
co
nstru
c
ted
for th
e n
e
two
r
k
i
n
g
d
i
ag
ram sh
o
w
n
in
Figu
re
1
,
b
a
sed
on
wh
ich
reliab
ility o
f
th
e n
e
two
r
k
h
a
s
been com
pute
d
. T
h
e fa
ult tree
is shown i
n
Fi
gure
2.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
60
–
1
569
1
564
Fig
u
re
2
.
Fau
lt tree
d
i
ag
ram
fo
r th
e
Un
iv
ersity Network
Using
th
e Fau
lt tree d
i
ag
ra
m
reliab
ilit
y
o
f
th
e n
e
t
w
o
r
k
h
a
s
b
een co
m
p
u
t
ed
. Th
e reliab
ility
com
put
at
i
on
re
sul
t
s
are
s
h
o
w
n i
n
t
h
e
Ta
bl
e
1.
T
h
e
pr
ocee
di
n
g
devi
ces
f
o
r
eve
r
y
devi
c
e
ha
ve
bee
n
c
o
nnect
e
d
usi
n
g a
n
a
p
pr
o
p
ri
at
e
gat
e
bas
e
d
on
t
h
e
pat
h
s avai
l
a
bl
e
f
o
r
safe
wo
r
k
i
n
g
o
f
t
h
e
de
si
g
n
at
ed
devi
ce
. Ei
t
h
e
r
OR
gat
e
or
A
ND
gat
e
i
s
used
fo
r est
i
m
a
t
i
ng t
h
e FTA.
In t
h
e case of c
o
n
n
ec
t
i
ons est
a
bl
i
s
h
e
d t
h
r
o
ug
h
OR
gat
e
,
hi
g
h
est
fai
l
u
re
rat
e
o
f
i
n
c
o
m
i
ng
devi
ces
ha
s bee
n
co
nsi
d
e
r
ed t
o
be t
h
e f
a
i
l
u
re rat
e
of t
h
e co
n
n
ect
ed
devi
ce
whe
r
eas
whe
n
an AND gate i
s
use
d
, the c
o
m
b
ined failure
rate has
been
considere
d
as t
h
e failure rate
of t
h
e
con
n
ect
ed
devi
ce. The fai
l
u
re
rat
e
of t
h
e Un
i
v
ersi
t
y
cl
oud
net
w
or
k i
s
est
i
m
a
t
e
d t
o
be t
h
e fai
l
u
re rat
e
o
f
t
h
e
APE
X
no
de
w
h
i
c
h i
s
t
o
p
gat
e
way
.
T
h
e s
u
c
cess rat
e
of
t
h
e ent
i
r
e
net
w
or
k
has
been
est
i
m
at
ed t
o
be
0
.
22
2 a
s
can be
see
n
from
the
Table 1.
4.
IMPLEME
N
TING BUT
TERFLY NETWOR
K
TOPOLOGY FOR THE UNIVERSITY
N
ETW
OR
K
Mu
lti-stag
e n
e
twork
s
are commo
n
l
y u
s
ed
t
o
conn
ect a
set
o
f
i
n
pu
ts to a
set o
f
ou
tpu
t
s; th
e con
c
ep
t
as such is sim
ilar to cloud c
o
m
puting. T
h
e
connectiv
ity is sued through links bet
w
e
e
n the com
puting /
swi
t
c
hi
n
g
sy
st
em
s. These net
w
o
r
k
s
use
2 X
2 swi
t
c
hes
.
Ea
ch swi
t
c
h t
a
ke
s t
w
o i
n
put
s an
d p
r
o
d
u
ces 2 o
u
t
p
ut
s
vi
a di
ffe
rent
c
o
n
n
ect
i
o
ns
(St
r
ai
ght
,
cr
oss,
u
ppe
r
br
oa
dcast
an
d t
h
e l
o
we
r
br
oa
dcast
.
A
b
u
t
t
e
r
fl
y
net
w
o
r
k
i
s
a
m
u
lti-stage networks
. Num
b
er of st
ag
es
used
d
e
p
e
nd
s on
th
e k
i
nd
of co
nn
ectiv
ity
r
e
qu
ir
ed
. A
bu
tter
f
l
y
t
o
p
o
l
o
gy
w
h
i
c
h
uses a
n
8 st
a
g
e
net
w
or
k
ha
s bee
n
c
o
nsi
d
e
r
ed and
t
h
e same is u
s
ed to
fit in
to
th
e Un
i
v
ersity
net
w
or
k.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Building F
a
ult
Toler
ance
within Clo
uds
at N
e
twork Level (
D
BK K
a
mes
h
)
1
565
Tab
l
e
1
.
Reliabilit
y co
m
p
u
t
ati
o
n
a
l
resu
lts fo
r a Un
iv
ersity network
Sl.
no
Device
Success
Rate
Gates used
For
Connection
Preceding Devices
Device
nam
e
D1
Device
nam
e
D2
Device
nam
e
D3
Device
nam
e
D4
Device
nam
e
D5
Co
m
b
ined
Success
Rate
Success
Rate S1
Success
Rate S2
Success
Rate S3
Success
Rate S4
Success
Rate S5
1
Stor
age
0.
8
0.
8
2 I
n
finite
Switch
0.
8
OR
Storage
0.
8
0.
8
3 HP
Blade
0.
9
OR
Infinite
Switch
0.
8
0.
8
4
5 Dell Blade
0.
9
0.
9
5
Multi Processors
Architecture
(2 GP
Us, 4
CPUs)
0.
9
0.
9
6
SW
ITCH Ra
ck (All
Blockds,
L
a
dies Hostel
0.
8
0.
8
7
M
a
in L
A
N Switch
0.
8
0.
8
8 Siem
ens
Contr
o
l
0.
8
OR
Switch
Rack
Main
Lan
Switch
0.
8
0.
8
0.
8
9
HP Core S
w
itch H
P
X
3506 S1
0.
7 OR
5 DE
LL
Blades
0.
7
0.
9
10
HP Core S
w
itch H
P
X
3506 S2
0.
7 OR
HP Blad
e
0.
7
0.
9
11
HP Core S
w
itch H
P
X
3506 S3
0.
7 OR
2 GPUs,
4CPUs
0.
9
0.
9
A switch
b
o
x
in
stag
e-I is co
nn
ected with
the links that
are at a
dista
n
ce
of
2
i
apart
.
The
8 X 8
but
t
e
r
f
l
y
net
w
or
k i
s
achi
e
ve
d t
h
r
o
u
g
h
t
w
o
4 X 4 net
w
o
r
k
s
and f
u
rt
her f
o
u
r
n
u
m
b
er of
2 X 2 net
w
o
r
ks. T
h
e
p
r
ob
ab
ility th
at on
e
o
f
th
e
p
a
t
h
s ex
ists for con
n
ecting
a cloud
co
m
p
u
tin
g
pro
cesso
r can b
e
co
m
p
u
t
ed
as
A
c =
2
k
ρ
l
Φ
(
k
)
(
1
)
whe
r
e k = Num
b
er of stages
,
ρ
= proba
b
ility that a node fails and
Φ
(k) is th
e p
r
ob
ab
ility
th
at th
at a
switch
box i
n
the
stage K ca
n
fail.
Φ
(k
) ca
n
be c
o
m
put
e
d
usi
n
g t
h
e
eq
uat
i
o
n
(
2
)
.
Φ
(k
) =
1
–
( 1-
ρ
l
Φ
(k
-
1
)
2
(2)
Th
e
bu
tterfly network conn
ected
fo
r fittin
g
u
n
i
v
e
rsity
cloud
h
a
s
b
e
en
sh
own in
Figu
re 3. Th
e
b
u
tter
fl
y
net
w
o
r
k
ha
s been est
a
bl
i
s
hed
usi
n
g 8 X
8 net
w
or
k co
nt
ai
ni
n
g
8 st
ag
es. The 8
X 8
net
w
or
k has
b
ecom
e
n
ecessary d
u
e
to
th
e av
ailab
ility o
f
8
ele
m
e
n
tary lev
e
ls
o
f
in
pu
ts an
d
8
d
i
fferen
t
typ
e
s of o
u
t
p
u
t
s requ
ired
to
m
a
ke t
h
e net
w
or
k rel
i
a
bl
e a
n
d avai
l
a
bl
e.
A
d
di
t
i
onal
swi
t
c
h
e
s have
bee
n
a
dde
d t
o
m
a
ke i
t
possi
bl
e t
o
co
nnect
th
e Un
iv
ersity clo
u
d
in
t
o
a butefy n
e
two
r
k
.
Using the equa
tion (1) and
equation
(2) the probability of success th
at at least one path exists from
i
n
p
u
t
p
o
i
n
t
t
o
a
n
out
put
has
be
en c
o
m
put
ed a
s
0.
2
8
.
5.
MODIFIE
D
UNIVE
RSIT
Y NETWORK
TO FIT BUTTERFLY T
O
POLOGY
Co
n
s
i
d
er
i
n
g
t
h
e bu
tter
f
l
y n
e
t
w
or
k show
n in th
e Figur
e 3 dev
e
lop
e
d fo
r
Un
iv
er
sity n
e
two
r
k
,
t
h
e ex
tr
a
switch
e
s requ
ired
h
a
v
e
b
e
en
id
en
tified
and
th
e sam
e
are
consi
d
e
r
ed
fo
r i
n
cl
udi
n
g
i
n
the t
r
ee like struct
ured
of
th
e
un
iv
ersity
clo
u
d
.
Th
e rev
i
sed
KLU n
e
twork
t
h
at
h
a
s b
een fitted
with
b
u
tterfly top
o
l
o
g
y
is shown
i
n
Fi
gu
re 4
.
It
i
s
seen f
r
om
t
h
e fi
g
u
re t
h
at
3
ext
r
a s
w
i
t
c
he
s have
bee
n
a
dde
d a
nd t
h
e i
n
t
e
r
n
et
ba
nd
wi
dt
h i
s
literally broken into t
w
o hal
f
s; each
half
working as a
back
up to t
h
e
othe
r.
Faul
t
t
r
ee has
been c
o
nst
r
uct
e
d f
o
r t
h
e m
odi
fi
ed U
n
i
v
e
r
si
t
y
net
w
o
r
k
di
ag
ram
and t
h
e sa
m
e
i
s
sho
w
n
i
n
Fi
g
u
re
5. T
h
e con
n
ect
i
v
i
t
y
i
s
achi
e
ved t
h
r
o
u
g
h
OR
a
nd
AN
D gat
e
s as
descri
bed t
o
p
r
od
uce a
n
FT
A fo
r t
h
e
o
r
i
g
in
al Un
i
v
ersity n
e
two
r
k
.
Th
e
p
r
ob
ab
ility o
f
t
h
e su
ccess of th
e
rev
i
sed
n
e
two
r
k
is
o
n
c
e ag
ai
n
com
p
u
t
ed
and t
h
e c
o
m
p
u
t
at
i
onal
resul
t
s
are sh
ow
n i
n
T
a
bl
e 2. F
r
om
the Table, it can be seen th
at the success rate
of the
revi
se
d
Uni
v
er
si
t
y
cl
oud
has
been
i
n
c
r
ease
d
fr
om
0.2
2
2
t
o
0.
27
6
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
60
–
1
569
1
566
Fi
gu
re 3.
Bu
tterfly n
e
t
w
ork
for th
e Un
iv
ersity Clo
ud
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Building F
a
ult
Toler
ance
within Clo
uds
at N
e
twork Level (
D
BK K
a
mes
h
)
1
567
Fig
u
re
4
.
Mod
i
fied
Un
iv
ersity Clo
u
d
t
h
at fits in
to
it the Bu
tt
erfly top
o
l
o
g
y
Fig
u
re
5
.
Mod
i
fied
Un
iv
ersity Clo
u
d
i
n
term
s of Bu
tterfly to
po
log
y
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
60
–
1
569
1
568
Tab
l
e
2
.
Fau
lt
Rate calcu
latio
n
s
for t
h
e Modified
Un
i
v
ersit
y
Clo
u
d
n
e
twork th
at fits a B
u
tterfly topo
log
y
Sl.
no Device
Success
Rate
Gates used
For
Connection
Preceding Devices
Device nam
e
D1
Device nam
e
D2
Co
m
b
ined
Success
Rate
Success Rate S1
Success Rate S2
1 SW
IT
CH
R
A
CK(FE
D
,
C
SE
)
0.
9
0.
9
2
5 DE
LL
BLADE
0.
9
0.
9
3 ST
ORAGE
0.
8
0.
8
4 I
N
FINI
TY
SW
ITCH
0.
8
OR
STORAG
E
0.
8
0.
8
5 HP
CL
OUD
0.
9
OR
INFI
NIT
Y
SWI
T
CH
0.
8
0.
8
6 2GPUS,
4CPUS
HPC
0.
9
0.
9
7
HP CORE S
W
ITCH S1
0.7
OR
5 DEL
L
BL
ADE
0.
7
0.
9
8
HP CORE S
W
ITCH S2
0.7
OR
H
P
CL
OU
D
0.
7
0.
8
9
HO CORE S
W
ITCH S3
0.7
OR
2GPUS,4CPUS
HPC
0.
7
0.
9
10
SW
IT
CH
R
A
CK
(FUL
L
B
L
O
CK)
0.
8
0.
8
11
M
A
IN LAN SWI
T
CH
0.
8
0.
8
1
2
SEI
M
ENS
CON
T
ROLL
ER
0
.
8
OR
MAI
N
L
AN
SWI
T
CH
SWI
T
CH
RACK(FULL) 0.8
0.
8 0.
8
13
SEIMENS
CORE
SW
IT
CH 0.
8
AND
SEI
M
EN
TS
CONTRO
LLER
0.
64
0.
8
1
4
HP
WIF
I
CON
T
O
LLER
0
.
7
5
AND
SWI
T
CH
RACK(FED)
0.
675
0.
9
6.
CO
MP
AR
AT
IVE A
N
A
LYSIS
OF
RELIABILIT
Y EV
ALU
A
T
ION O
F
THE CLO
U
D
CO
MP
UTING NETW
ORKS
The c
o
m
put
at
i
on
o
f
s
u
ccess
r
a
t
e
s of
di
f
f
ere
n
t
t
o
p
o
l
o
gi
es us
ed t
o
de
vel
o
p t
h
e U
n
i
v
ersi
t
y
cl
ou
d rel
a
t
e
d
net
w
or
k i
s
s
h
o
w
n i
n
t
h
e Ta
bl
e 3.
It
can
be s
een f
r
om
t
h
e t
a
bl
e t
h
at
B
u
t
t
erfl
y
t
o
p
o
l
ogi
e
s
w
h
en i
n
co
r
p
orat
e
d
i
n
t
o
U
n
i
v
e
r
si
t
y
cl
oud
rel
a
t
e
d
net
w
or
k has i
n
creased t
h
e suc
cess rat
e
m
a
ki
ng a
v
ai
l
a
bl
e m
o
re c
ont
i
n
ui
t
y
of t
h
e
services
as require
d
by
the us
er.
Tabl
e
3. C
o
m
p
ari
s
o
n
of
succe
ss rat
e
s
o
f
C
l
o
u
d
rel
a
t
e
d
Net
w
o
r
k
w
h
e
n
des
i
gne
d
wi
t
h
di
ff
erent
t
o
p
o
l
o
gi
e
s
T
opology
Serial
T
opology
Success
Rate
[1]
T
r
ee topology
– Or
iginal Univer
sity
cloud r
e
lated Networ
k
0.
227
[2]
Butterfly
topology
built-in
with extra
switches
0.280
[3]
T
r
ee T
opology
enhanced with identified r
e
dundancies included into butter
f
l
y
networ
k
0.
277
7.
CO
NCL
USI
O
NS
Networks that
connects va
ri
ous
res
o
urces
that
form
a cl
ou
d pl
ay
s a
m
a
jor r
o
l
e
i
n
p
r
o
v
i
d
i
n
g
cont
i
n
ue
d ser
v
i
ces t
o
t
h
e user
s of t
h
e cl
ou
ds
whi
c
h l
eads t
o
hea
v
y
user s
a
t
i
s
fact
i
on. H
o
weve
r i
f
t
h
e ne
t
w
o
r
k
is eith
er fau
lty o
r
wh
en
d
i
fferen
t
fau
lts o
ccurs wh
ile sy
stem is ru
nn
ing
,
t
h
e serv
ices to
u
s
ers
will b
e
disrup
ted
till
th
e ti
m
e
t
h
e n
e
t
w
ork
is
m
a
d
e
o
p
e
ratio
n
a
l. Mak
i
ng av
ailab
l
e contin
u
o
u
s
serv
ices to
th
e user is
m
o
st
i
m
p
o
r
tan
t
p
r
e-
req
u
i
site of
im
p
l
e
m
en
tin
g
cloud
b
a
sed
ser
v
ices to
th
e user
.
Th
e netwo
r
k
used
to
h
o
st th
e cloud
com
put
i
n
g
bas
e
d se
r
v
i
ces m
u
st
be
rel
i
a
bl
e a
n
d
t
h
e
net
w
o
r
k
m
u
st
be
desi
g
n
ed
t
o
be
fa
ul
t
t
o
l
e
ra
nt
s
o
t
h
a
t
t
h
e
serv
ices to
t
h
e u
s
ers
will b
e
p
r
ov
id
ed
co
n
t
i
n
uou
sly ev
en
i
n
th
e ev
en
t
of
o
ccurren
c
e
of
th
e fau
lts at
n
e
twork
l
e
vel
.
Al
t
e
rn
at
i
v
e pat
h
s o
f
co
m
m
uni
cat
i
on bet
w
ee
n t
h
e u
s
er an
d t
h
e cl
ou
ds ar
e t
o
be
est
a
bl
i
s
hed t
o
m
a
ke
av
ailab
l
e un
-i
nterrup
t
ed
serv
ice to
th
e u
s
ers. Mu
lti stag
e n
e
twork
s
h
e
lp
s
in
i
m
p
r
ov
ing
th
e reliab
ility
man
y
fol
d
. B
u
t
t
e
rfl
y
net
w
or
ki
n
g
t
o
pol
ogy
s
u
pp
ort
s
m
u
l
t
i
s
t
a
ge n
e
t
w
o
r
ks t
h
r
o
ug
h 2
X
2 s
w
i
t
c
hes
whi
c
h p
r
o
v
i
d
e
4
altern
ativ
e p
a
t
h
s of switch
i
n
g
an
d
m
a
n
y
in
termitten
t
switch
e
s m
u
lti
p
lies
m
a
n
y
o
t
h
e
r
p
a
th
s. B
u
tterfl
y
n
e
two
r
k
i
ng
topo
log
i
es in
creases th
e fau
lt to
leran
ce cap
a
b
ili
ty with
lease c
o
st wh
ich
cou
l
d
b
e
th
e co
st of few
switches.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Building F
a
ult
Toler
ance
within Clo
uds
at N
e
twork Level (
D
BK K
a
mes
h
)
1
569
REFERE
NC
ES
[1]
F. T.
Leigh
t
on and B. M. Baggs
, “Fast Algorithm
s
for
Routing Around Faults i
n
Multibuffer
fli
e
s and Random
l
y
-
W
i
red Split
ter
Networks,”
IEEE
TRANSACTIONS ON COMPUTER
, vol. 41
, pp
.
578-587, 1992
.
[2]
W
.
Shi and
P.
K. Srim
ani,
“
H
yper
-butt
e
rfl
y n
e
twork:
a sc
ala
b
le op
tim
all
y
f
a
ult
tol
e
rant
ar
c
h
ite
cture
,
”
IE
EE
TRANSACTION ON Parallel Processing Symposium
, 1998. IPPS/SPDP 1998.
Proceedings of the First Merged
International
and
S
y
mposium on
Para
llel
and Distributed
Processing 1998,
pp. 732
-736, ISSN: 106
3-7133, 1998
.
[3]
J.
F.
Li,
et al.
, “
E
asil
y T
e
sta
b
le and Faul
t-T
o
leran
t
FFT Butterf
l
y
N
e
twork
s
,”
IEEE TRAN
SACTIONS ON
CIRCUITS AND S
Y
STEMS—II:
ANALOG AND
DIGITAL SIGN
AL PROCESSIN
G
, vol. 47
, pp
. 9
19-929, 2000
.
[4]
R. L
.
Graham
,
et al.
, “A Networ
k-Failure-
T
oler
ant Message-Pa
ssing S
y
stem for
Tera
s
c
al
e Clus
te
rs
,”
Internationa
l
Journal of Parallel Programming
, vol/issue: 31(
4), pp
. 285-303
,
2003.
[5]
C. Guo,
et al.
,
“
D
Cell: A Sca
l
abl
e
and Fault-
Toler
a
nt
Netwo
r
k Structure for
Data Centres,
”
SIGCOMM '08
Proceed
ings of the ACM SIGCOMM 2008
conferenc
e
on Data communication
, pp. 75-86, 2008. ISBN: 978-1-
60558-175-0.
[6]
V.
Liu,
et al.
, “
F
10: A Fault-Toleran
t
Engin
eer
ed Network,”
1
0
th USENIX S
y
mposium on Networked Systems
Design and Implementation
, pp.
339-412, 2013
.
[7]
H. Zhao
and J. Cann
y
,
“Butterfl
y M
i
xing:
Acce
ler
a
ting In
cr
em
ental-
Update Algorithms on Clusters,”
SIAM
International Co
nference on
Data Mining
, 2013
.
ISBN: 978-1-61197-262-7.
[8]
R. J
h
awar,
et al.
, “Fault To
lerance Manag
e
ment in
Cloud C
o
mputing: A Sy
stem-L
evel Perspectiv
e,”
IEEE
S
Y
STEMS JOURNAL
, vol/issue:
7(2), pp
. 288-29
7, 2013
.
[9]
M. A. Sharkh,
et al.
, “Resource Allocation
in a Network-Ba
sed Cloud Computing Environment: Design
Challeng
es,”
IEEE Communica
tions Magazine
,
pp. 46-52
, 2013
. 0163-6804/13.
[10]
S. Giriesh,
et al.
, “D
y
n
am
i
c
Dat
a
Fault
To
leran
c
e Mech
anism
to
Enhan
ce R
e
li
ab
ilit
y
and
Avail
a
bilit
y
in C
l
oud,”
Indian Journal o
f
Sc
ience and
Technology,
vo
l/issue:
8(S9), pp. 30
0–305,
2015
.
[11]
P. Padmakumar
i
and A. Umamakeswari, “Methodical Rev
i
ew
on Various Fault Toleran
t
and Monitoring
Mechanisms to improve Reliab
ility
on Cloud
Env
i
ronment,”
Ind
i
a
n
Journal of Science and Techn
o
logy
, vo
l/issue:
8(35), 2015
. DOI: 10.17485
/IJSTt/2015/v8i35
/80
130.
Evaluation Warning : The document was created with Spire.PDF for Python.