TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 13, No. 3, March 2
015,
pp. 568 ~ 57
3
DOI: 10.115
9
1
/telkomni
ka.
v
13i3.723
7
568
Re
cei
v
ed O
c
t
ober 2
8
, 201
4; Revi
se
d Decem
b
e
r
25, 2014; Accept
ed Ja
nua
ry 1
0
, 2015
Performance Analysis of Load Balancing Techniques in
Cloud Computing Environment
V Rav
i
Teja
Kana
kala*, V.Krishna
Re
dd
y
Rese
arch Sch
o
lar, KL Un
iver
sit
y
, Vad
des
w
a
ram, India
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: raviteja.ka
n
a
k
ala@
gma
il.co
m
A
b
st
r
a
ct
Clou
d
co
mp
uti
ng is a foreru
n
ner a
m
o
ng the
tec
hnol
ogi
es emergi
ng tod
a
y
in the IT
w
o
rld. It has
alre
ady b
een
some years
after the emerg
ence of cl
oud
technol
ogy b
u
t still the w
o
rld di
dn
’
t
g
e
t t
h
e
compl
e
te fruits
of clo
ud c
o
mp
uting. Out of th
e
many
revo
lut
i
on
ary resu
lts
expecte
d as
o
u
tcomes fro
m
t
h
e
clou
d co
mputi
ng tec
hno
lo
gy
there w
e
re v
e
ry few
ac
hi
ev
ed a
nd t
he re
st of t
he ex
pe
cted are
still
u
nde
r
researc
h
. T
w
o
of the main
ob
stacles in th
e u
s
age of
cl
ou
d computi
ng ar
e Clou
d
Sec
u
rity and P
e
rforman
c
e
stability. Loa
d Bala
ncin
g
is
o
ne of
the
el
e
m
ents that s
how
s i
m
pact
on
th
e p
e
rfor
manc
e
stabi
lity of c
l
o
u
d
computi
ng. In this pa
per w
e
discuss a
b
o
u
t loa
d
ba
l
anc
in
g
and d
i
fferent
alg
o
rith
ms that
are pro
pose
d
for
distrib
u
ting th
e
loa
d
a
m
o
ng th
e no
des
and
al
so the p
a
ra
meters that are t
a
ken i
n
to acc
o
u
n
t for calcu
l
ati
n
g
the best al
gorit
hm to b
a
la
nce
the loa
d
.
Ke
y
w
ords
:
lo
a
d
bal
anci
ng, cl
oud co
mputi
n
g
,
cloud secur
i
ty, performance
stability
Copy
right
©
2015 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Cloud computing provides
the availabilit
y of IT resources
which are
at
different parts of
the world
to
use
r
s who
wants to
a
c
cess tho
s
e
reso
u
r
ce
s from
the
i
r work
pla
c
e
in the fo
rm
a
s
a
servi
c
e throu
gh an optimi
z
ed and reliabl
e servi
c
e
p
r
o
v
ider maintai
n
ing conveni
ence and u
b
i
quity
[1]. The cloud
comp
uting te
chn
o
logy inte
nds to p
r
ovid
e “computin
g
as a utility” [2
] in near future.
Clou
d com
p
uting provid
e
s
its se
rvice
s
on
lin
e on-deman
d and
pay-as-you
-
go ba
sis. Cl
oud
comp
uting
ha
s
cha
nge
d th
e way the
IT
comp
anie
s
way of tra
d
ing
and
de
signi
n
g
thei
r p
r
od
u
c
ts.
Clou
d comp
u
t
ing ha
s mini
mized
the
pa
yment of pr
e
m
ium
cost
s f
o
r p
r
od
uct
s
with its pay
as
you
go basi
s
which increa
sed t
he traffic in the IT se
rvice
s
maki
ng Loa
d Balanci
ng a central poin
t
of
resea
r
ch. To balan
ce the l
oad amo
ng
multiple re
so
urces in a
clo
ud there a
r
e
several algo
ri
thms
prop
osed
but
till now
no
al
gorithm
wa
s
able to
balan
ce the
loa
d
in
a cl
oud
with
out pe
rform
a
nce
degradin
g
. Many research
ers a
r
e worki
ng on the
issue of load ba
lanci
ng and t
here a
r
e ma
ny
algorith
m
s be
ing p
r
op
osed
day by d
a
y
as
outcome
of their
re
sea
r
ch. In
this p
aper we
surv
ey
some
of th
e
optimistic alg
o
rithm
s
whi
c
h ha
d
sho
w
n
so
me im
pro
v
ement in
lo
ad b
a
lan
c
in
g
and
increa
sed the
level of performan
ce.
2. Defini
tion
Load
bal
an
cing i
s
a
pe
rformance im
pro
v
ing me
tho
d
applie
d in
the
area
of Networki
ng t
o
distrib
u
te the
work load a
c
ro
ss multipl
e
resour
ce
s
that are invol
v
ed in the computation o
f
a
netwo
rki
ng task. He
re L
oad can be
memory,
proce
s
sor
cap
a
city, network load et
c. Loa
d
balan
cing o
p
timize
s the use of re
so
urce
s, redu
ce
s re
spo
n
se time, avoids ove
r
lo
ad on any on
e
system by d
i
stributin
g th
e load to m
u
ltip
le co
mp
onent
s [3]. Clou
d co
mp
uting ca
n se
rve
compli
cate
d t
a
sk that requi
res hu
ge
com
putational
re
source
s by u
s
i
ng di
stribut
ed
re
sou
r
ces in
a
decentrali
ze
d
mann
er. In
a
network
with
high
comput
ational
req
u
irements we
can b
a
lan
c
e
a
n
d
distrib
u
te the
load evenly
across all t
he no
de
s co
ntentedly by
usin
g app
ro
p
r
iate sch
eduli
ng
algorith
m
s. T
here
a
r
e sev
e
ral algo
rith
ms pro
p
o
s
ed
ea
rlier in th
e
area
of op
erating
system
s to
sched
ule
re
source
s to th
e proces
se
s.
But in
clou
d
com
puting
to bala
n
ce th
e load
the
r
e
are
some i
s
sue
s
like se
cu
rity, service sp
eed, re
lia
bility etc. So to
develop an
optimized l
o
a
d
balan
cing
alg
o
rithm o
ne m
u
st try to cre
a
te an envi
r
o
n
ment in
whi
c
h total loa
d
of a system
can
be re
assign
e
d
to multiple comp
one
nts t
hat wo
rk
coll
ectively in that system so that there will
be
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Perform
a
n
c
e
Analysis of L
oad Balan
c
in
g Tech
niqu
es in Cloud
… (V Ravi Tej
a
Kana
kala
)
569
no overlo
ade
d and und
er l
oade
d node
s
in a system whi
c
h re
du
ce
s the overall resp
on
se time
o
f
the syste
m
i
m
provin
g the
spe
ed,
se
cu
rity
and
relia
bility. While
developin
g
a
load
balan
ci
ng
algorith
m
we
mu
st con
s
i
der some i
m
portant
thing
s
li
ke
estim
a
ting the l
oad
app
rop
r
iatel
y
,
monitori
ng th
e perfo
rma
n
ce and
stabilit
y of the sy
ste
m
while pe
rfo
r
ming a
task, and sele
ction
of
prop
er n
ode
s accordi
ng to the nature of
work.
3.
Challeng
es in Cloud Co
mputing Loa
d Balancing
Finding
a sol
u
tion for p
r
o
b
l
ems in l
oad
balan
cing i
s
never a
n
ea
sy process the
r
e will
be
many challenges to be fa
ced while developing a
sol
u
tion. He
re i
n
this
section we
will di
scuss
some
of the
comm
on
chal
lenge
s that m
i
ght be fa
ce
d
while
develo
p
ing a
solutio
n
for
a p
r
oble
m
of load bala
n
c
ing in
clou
d comp
uting.
Distri
bution
of
Clou
d No
des:
T
h
e
r
e
are
many al
gorithm
s
bei
ng p
r
op
osed
for lo
ad
balan
cing
in
clou
d
comp
uting. Amon
g t
hem
som
e
al
gorithm
s mi
g
h
t pro
d
u
c
e
efficient results with
small
net
works o
r
a
n
e
twork
with
clo
s
ely located
n
ode
s. Su
ch
a
l
gorithm
s
are
not
suita
b
le
for
large n
e
two
r
ks b
e
cau
s
e
those al
gorit
hms cann
ot
prod
uce the same effici
e
n
t result
s when
applie
d to larger net
wo
rks.
There a
r
e m
any rea
s
on
s that affect the efficiency in
large
r
net
works
like
spee
d of
the network, distan
ce b
e
twee
n the
cli
e
nts an
d se
rve
r
nod
es
and
also th
e dista
n
ce
betwe
en
all t
he n
ode
s i
n
t
he n
e
two
r
k [
2
]. So while
developin
g
a
load
bal
an
ci
ng al
gorith
m
one
sho
u
ld try for better re
sult
s in spatially distri
b
u
ted n
o
des b
a
lan
c
in
g the load eff
e
ctively red
u
c
ing
netwo
rk d
e
la
ys.
Migratio
n Ti
m
e
:
In cloud
co
mputing
the
servi
c
e
-
o
n
-de
m
an
d m
e
thod
will
be
followed
whi
c
h means when there i
s
a demand for a
reso
urce the service
will be
provided to the required
client. So wh
ile se
rving th
e client o
n
hi
s dem
and
s sometime
s we
need to mig
r
ate re
so
urce
s
from long di
stances du
e to unavailability in near lo
ca
ti
ons. In su
ch
ca
se
s the time of migration
o
f
the resources from far locations
will be more which will affect
the performance of the system.
While
develo
p
ing a
n
alg
o
rithm one
sh
o
u
ld note
that
re
sou
r
ce mi
gration
time i
s
an
impo
rta
n
t
factor that greatly affects
the perfo
rma
n
c
e of the syst
em.
Perform
a
n
c
e
of the S
y
ste
m
:
It doesn’t
mean
that if
the
compl
e
xity of an al
go
rithm i
s
high then the
perfo
rman
ce
of the system
will be hi
gh. Any time load balan
cing al
gorithm mu
st be
simple
to im
plement
and
ea
sy to o
p
e
rate. If the
co
mplexity of algo
rithm
is hi
gh th
en
the
implementation
cost will
al
so
be m
o
re and even
after implem
enting the
system performance will
be de
cre
a
sed
due to more
delays in the
function
ality of the algorith
m
.
Failure of controller:
Def
i
nitely centra
lized lo
ad b
a
lan
c
ing alg
o
rithm
s
(Having one
controlle
r)
ca
n provid
e efficient results
while b
a
lan
c
i
ng the loa
d
than the di
stri
buted alg
o
rith
ms.
But in ce
ntral
i
zed l
oad
bal
anci
ng al
gorit
hms
wh
en th
e co
ntrolle
r f
a
ils the
wh
ol
e syste
m
will
be
halted, in
su
ch ca
se
s the
r
e
will be a
hug
e loss fo
r
bot
h clie
nt and
service
provid
er. So, the lo
ad
balan
cing
alg
o
rithm
s
mu
st be de
sign
ed i
n
a decent
ralized a
nd di
stri
buted fashion
so that whe
n
a node acting as
a
cont
roller fa
il
s
the system will not halt [5]. In
such cases the control will be
given to other node
s and th
ey will ac
t as
controlle
rs of
the system.
Energ
y
Ma
na
gem
ent:
A load bala
n
ci
ng
algorith
m
sh
o
u
ld be d
e
sig
n
ed in a way such that
the operation
a
l co
st and the ene
rgy
co
nsum
ption of the algorithm
must be low.
Increa
se in t
he
energy con
s
umption is o
ne of the main probl
em
s t
hat cloud
co
mputing is fa
cing today. Even
though
by usi
ng en
ergy effi
cient h
a
rd
wa
re archit
e
c
ture
s which sl
ows do
wn th
e p
r
oce
s
sor
sp
ee
d
and tu
rn
off machi
n
e
s
tha
t
are
not u
n
d
e
r u
s
e
the
en
ergy m
anag
e
m
ent is be
co
ming difficult. So,
to achieve b
e
tter re
sult
s in
energy ma
na
gement
a
lo
a
d
bal
an
cing
a
l
gorithm
shou
ld be
de
sign
ed
by following
Energy Awa
r
e Job S
c
hed
u
ling method
ol
ogy [6].
Sec
u
rity
:
Security is one
of the probl
e
m
s that
clo
u
d com
puting
has in its t
op mo
st
prio
rity. The
clou
d is al
wa
ys vulne
r
abl
e
in on
e o
r
th
e othe
r
way t
o
security attacks like DDOS
attacks et
c. While
bala
n
cing the lo
ad
there
are
many ope
rati
ons th
at take pla
c
e li
ke
VM
migratio
n etc at that time
there i
s
a
h
i
gh p
r
oba
bility of se
curity
attacks. So
an efficie
n
t l
oad
balan
cing al
g
o
rithm mu
st be stro
ng en
ough to re
du
ce the se
cu
ri
ty attacks bu
t should not
be
vulnera
b
le.
4.
Classific
a
tio
n
of Load
Ba
lancing
Algo
rithm
s
In Gene
ral l
oad b
a
lan
c
in
g algo
rithm
s
can
be
cate
gori
z
ed
into
two different
se
ction
s
calle
d static l
oad bal
an
cin
g
algorith
m
s
and dynami
c
load bal
anci
n
g algorith
m
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 3, March 2
015 : 568 – 5
7
3
570
Static Load B
a
lan
c
ing:
In a
static loa
d
b
a
lan
c
ing al
go
rithm whil
e a
ssi
gnin
g
tasks to the
node
s it will not che
ck the
state and fun
c
tionality
of the node in p
r
evious ta
sks [4]. The process
of assigni
ng t
he tasks will
be pu
rely ba
sed
on
the
system’s p
r
io
r kno
w
le
dge o
n
the prope
rti
e
s
and the cap
abilities of the nod
e like pro
c
e
ssi
ng
powe
r
, storage capa
city and memo
ry
availability. Even though t
he above
li
sted prope
rties of a node are
con
s
id
ere
d
before assign
ing
a task they cannot ad
apt to the
dynami
c
ch
ang
es in
the attributes and the allotted load on t
he
node d
u
rin
g
runtime [5].
Rou
nd Ro
bin
Algorithm
(RR):
Thi
s
is a static load b
a
l
anci
ng algo
ri
thm beca
u
se
before
assigni
ng a task to a node
it will not take into ac
coun
t the previous state
and fun
c
tionality of that
node. T
o
allo
cate the
job
s
the first n
ode
will be
sel
e
ct
ed rand
omly and the
n
the
remai
n
ing
no
de
s
are all
o
cated
jobs in
a ro
un
d robi
n man
n
e
r. Thi
s
way
of sched
uling
the load
will
cre
a
te p
r
oble
m
s
becau
se
whil
e allo
cating t
he job
s
o
ne
node m
a
y b
e
heavily loa
ded a
nd on
e
may be ligh
t
ly
loade
d irresp
ective of their capa
city. To solve th
is in
e
quality in loa
d
distrib
u
tion
“wei
ghted
rou
n
d
robin algorithm was proposed” in thi
s
al
gorithm
every
node
will be
assi
gned wei
ghts respecti
ve
of their capa
city then according to that m
eas
ure th
e lo
ad will b
e
a
s
signed to th
e n
ode
s [5]. Even
though the lo
ad is dist
ribut
ed equally it
is not possibl
e
to predict the
execution time of a proce
ss.
So, this algori
t
hm is not sui
t
able for effici
ent load bal
a
n
cin
g
.
Central Loa
d
Balancin
g Deci
sion Mo
de
l (CLB
DM):
T
h
is alg
o
rithm
is a develo
p
m
ent for
roun
d
robin a
l
gorithm co
nfigure
d
with
weight
an
d se
ssi
on-switchi
ng in
whi
c
h
we
can’t find
the
con
n
e
c
tion time between
the pro
c
e
s
s and the no
de
. In central lo
ad bala
n
ci
ng
deci
s
ion m
o
del
algorith
m
the
con
n
e
c
tion time between
the client
a
nd the no
de
in the clo
ud i
s
calculated
by
impleme
n
ting
a softwa
r
e
module
calle
d Ce
ntral
Lo
ad Bala
nci
n
g
De
cisi
on M
o
dule. Thi
s
m
odule
will interact
with all
parts of
the sy
stem and
collects informatio
n regarding the load balancers
and serve
r
s
etc. After col
l
ecting the d
a
ta a se
n
s
o
r
will be impl
emented, thi
s
monito
rs t
h
e
perfo
rman
ce
of a node and mea
s
ures the required
time for a task to be co
mpleted by the
respe
c
tive no
de. In this
way CLB
D
M al
gorithm
prod
uce
s
b
e
tter
result
s than
the weighte
d
and
se
ssi
on switch config
ure
d
algorith
m
.
Ant Colon
y
O
p
tim
i
zation Algorithm
(ACO):
Ant Colony Optimization algorithm
s
i
mulates
the ant fo
ragi
ng be
havio
r. In this
algo
rith
m the
b
ehavi
o
r of
ants i
s
u
s
ed
for
gathe
ring i
n
form
ation
from diffe
rent
nod
es in
the
syste
m
.
Wh
en the
exe
c
u
t
ion be
gin
s
th
e ant
an
d its
pheromo
ne
will
get initiated f
r
om the
hea
d
node
and m
o
ves to the
n
e
xt node. If the ant find
s
any of the no
des
unde
r loade
d
it will move
forwa
r
d to an
other no
de a
nd if that node is overlo
a
ded it will co
m
e
back to the p
r
eviou
s
nod
e
[8], in this way the in
form
ation abo
ut different nod
e
s
is g
a
thered
by
the system. Due to the m
o
ving of ant
s forward
and backward there
will be
some delay in the
traffic ma
ny rese
arche
r
s p
r
opo
se
d
solut
i
ons like
exiti
ng the
ant i
n
stead
of m
o
ving b
a
ckward
in
the path.
Map Re
du
ce
Algorithm
(MR):
Map
-
Redu
ce is a
prog
ram
m
ing
model de
signed for
handli
ng la
rg
e volume
s of
data by divid
i
ng the h
uge
tasks into
sm
all and in
dep
ende
nt one
s.
In
this al
gorith
m
there
a
r
e t
w
o tasks
calle
d Map
an
d
Red
u
ce, Map
functio
n
i
s
u
s
ed
to ma
p t
he
tasks an
d pa
rtitions the ta
sks into in
de
pend
ent one
s and th
en compa
r
e
s
ea
ch task
with e
v
ery
other ta
sk
an
d then the
Re
duce functio
n
grou
ps th
e
si
milar ta
sks a
nd re
du
ce
s th
e re
sults
of the
tas
k
s
.
Her
e
in this
algor
i
thm the
pr
oblem is
the Map func
tion
r
e
ads multiple tasks
simultaneousl
y which
will becom
e
an ov
erhead on th
e Reduce function to mini
mize the results.
Load Bala
nce Min-Min Al
gorithm
(LB Min-Min
)
:
In the traditiona
l Min-Min al
gorith
m
among
all the
available
re
source
s a
re
so
urce with
mini
mum exe
c
uti
on time
will b
e
opted
out a
n
d
then a ta
sk
with minim
u
m load i
s
a
s
sign
ed to th
e
re
spe
c
tive reso
urce
hen
ce the n
a
me
of the
algorith
m
is
min-min. But
while imple
m
enting th
is
algorith
m
we
are facin
g
p
r
oble
m
s li
ke load
imbalan
ce an
d unco
n
cern
of user prio
rity. To
solve these pro
b
l
e
ms the
r
e are improvem
e
n
ts
prop
osed
for the Mi
n-Mi
n
algo
rithm
s
called
User-Priority gui
ded
load
bala
n
ci
n
g
alg
o
rithm
b
y
Hua
n
kai in [9
]. In the User-Prio
r
ity guid
ed loa
d
bal
a
n
cin
g
alg
o
rith
m the imp
r
ov
ements are t
a
sk
rea
ssi
gnm
ent
whil
e sele
ct
i
ng smalle
st
siz
ed
t
a
s
k
a
c
c
o
rdi
n
g
t
o
t
he re
sou
r
c
e
make
spa
n
a
n
d
while dividin
g
the tasks accordin
g to the use
r
pri
o
rity demand.
Load Bala
nce Max-Min
Algorithm
(L
B Min-Ma
x):
In the Max-Min load
balan
cing
algorithm large tasks
will
be having
hi
ghest pri
o
rity. Here in thi
s
algorithm the process begins
with cal
c
ul
ating the inform
ation abo
ut the execut
io
n time of all the reso
urce
s an
d
then amon
g
all
the resource
s a node wit
h
large
s
t execution time
will be sele
cte
d
. After selecting a node with
highe
st execution time a task with
com
p
letion time
suitable to the
resource i
s
assign
ed. O. M.
Elzeki
propo
sed i
n
[10] t
hat in this
a
l
gorit
hm
wh
e
n
re
sou
r
ce
s that are
ca
pable
of fast
er
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Perform
a
n
c
e
Analysis of L
oad Balan
c
in
g Tech
niqu
es in Cloud
… (V Ravi Tej
a
Kana
kala
)
571
execution are assi
gned sm
all task
s and resources
with slow ex
ecution are assigned large tasks
will improve t
he perform
a
nce of the
algorithm be
cause duri
ng the
executio
n of
a large task
on a
slow perform
i
ng resource
there
will be many
small tasks
executed on the
fa
st performi
ng
r
e
sour
ce.
Dynam
ic Loa
d Balanci
ng:
A dynamic load bala
n
ci
n
g
algorithm
che
c
ks the p
r
eviou
s
state a
nd b
e
havior
of a
n
ode
while
a
c
hieving a
task
. It will ta
ke
the differe
nt runtime p
r
op
e
r
ties
of the node
s while p
r
o
c
e
s
sing th
e tasks into a
c
cou
n
t and
will a
ssi
gn the ta
sks
ba
sed
on
the
prop
ertie
s
col
l
ected o
n
the
node in the
runt
ime [4]. T
he advanta
g
e
of dynamic load bal
an
cing
algorith
m
is whe
n
a node
in the system fails it
will
not stop the whol
e syste
m
it will only
affect
the perfo
rma
n
ce of the
system. A dynamic
loa
d
balan
cing
algorith
m
s requires
con
s
tant
monitori
ng of
the functiona
lity of a node and is very
d
i
fficult in implementation.
A dynamic L
oad
balan
cing
al
gorithm
can
be i
m
plem
e
n
ted in
two
form
s di
stri
buted
and
n
on-di
stri
buted
. In
distrib
u
ted type of algorith
m
all
the nodes in the syst
em will intera
ct with each o
t
her and the t
a
sk
is distri
buted
among the
node
s but in
non-di
stri
but
ed type of algorithm all t
he nod
es wo
rk
indep
ende
ntly for achievin
g the task.
Throttled
Lo
a
d
Balan
c
e
r
Algorithm
:
This i
s
a
sim
p
le al
gorith
m
mainly
use
d
for l
oad
balan
cing
in
virtual ma
chi
nes. T
he
op
eration
of thi
s
alg
o
rithm
begin
s
with l
oad b
a
lan
c
e
r
; it
colle
cts the i
ndexing valu
es of all the virtual
machi
nes in the
system and
sto
r
es in in
dexi
n
g
table. When
a re
que
st i
s
received
by the loa
d
b
a
la
nce
r
fo
r reso
urce all
o
catio
n
it pa
rses the
indexing
tabl
e an
d
alloca
tes
re
sou
r
ce
acco
rdi
ng t
o
the
re
qui
re
ments of the
clie
nt [5]. A
fter
allocation th
e load bal
an
cer
update
s
the indexin
g
table. After the complet
i
on of task t
h
e
resou
r
ce d
e
-allocation
sta
r
ts. T
h
is alg
o
r
ithm h
e
lp
s i
n
a
c
hievin
g
better
perfo
rmance
and
high
utilization of resources.
Hon
e
ybee F
o
ragi
ng Algo
rithm
:
This algorith
m
is inspi
r
ed from the beh
avior of
honeyb
e
e
s
. Hon
e
ybee
s fi
rst g
o
o
u
t an
d search
fo
r
honey
sou
r
ce
s a
nd the
n
re
ap the
hon
ey
from
the s
o
urc
e
s
.
After that they c
o
me
bac
k
to the ho
ney
c
o
mb and calc
ulates
the food left, if there i
s
a sufficient a
m
ount of food they w
ill stay in the comb otherwise
they will go
out in search
of
more
ho
ney. Inspi
r
e
d
fro
m
this M.Ra
ndle
s
in
[11]
propo
sed
a
decentrali
ze
d
self
organi
zation
algorith
m
. In this alg
o
rith
m all the resource
s
a
r
e g
r
oup
ed a
s
vi
rtual resources. Ea
ch virt
ual
resou
r
ce mai
n
tains a p
r
o
c
ess q
ueue
an
d a
c
cepts
re
q
uest
s
fro
m
th
e qu
eue
and
pro
c
e
s
ses th
ose
request
s
. After processing each
requ
est every resource
will cal
c
ul
ate its profit, if it is high the
resou
r
ce
stays el
se it
mov
e
s to
the fo
rage. Th
at
is
why thi
s
alg
o
r
ithm i
s
na
m
ed a
s
Hon
e
ybee
foragin
g
alg
o
r
ithm. Maintai
n
ing a
se
parate que
ue
for each no
de a
nd computati
on of p
r
ofit after
processi
ng requests will be
come overhead for the sy
stem and
al
so
there
i
s
no improvem
ent in
the throug
hp
ut of the system.
Biased Ra
nd
om
Sam
p
ling
Algorithm
:
In this algo
rith
m the metho
d
of ran
dom
sampli
ng
the system d
o
main is follo
wed for a
c
hi
eving
self organi
zation to
balance the
load among
the
node
s i
n
the
syste
m
. Thi
s
al
go
rithm
works
with
th
e hel
p of
a v
i
rtual
dire
cte
d
g
r
aph
whi
c
h is
con
s
tru
c
ted
based o
n
th
e co
nne
ctivity of node
s i
n
the
syste
m
. In the graph e
a
ch n
ode
rep
r
e
s
ent
s a
vertex. When
the load bal
ancer receiv
es a requ
est
it compa
r
e
s
the wal
k
len
g
th
(traversal b
e
twee
n node
s)
of the reque
st and the
threshol
d value of the node in the system a
nd
if the walk len
g
th is eq
ual
o
r
g
r
eate
r
th
an
the
th
re
shol
d
value th
en
th
e reque
st
will
be
processe
d
at that node o
t
herwi
se the
reque
st will be
fo
rwa
r
de
d to other no
de
s in the grap
h [7].
Expon
ential
Sm
ooth Fore
ca
st ba
se
d
o
n
Weighte
d
Lea
st Conn
e
c
tion:
T
h
is al
gorithm
ESWLC (sho
rtly) is an i
m
provem
ent fo
r WLC p
r
opo
sed
by REN in
[13].
In WLC algo
rithm
the
load
will b
e
d
i
stribute
d
o
n
the re
so
urce
s b
a
sed o
n
t
he n
u
mbe
r
of
co
nne
ction
s
that re
sou
r
ce
is
having. When
the loa
d
bal
a
n
ce
r
re
ceives a reque
st fro
m
a
client
it g
a
thers info
rm
ation ab
out th
e
numbe
r of co
nne
ction
s
for
all node
s an
d
from that
a node with le
ast number of
conne
ction
s
wi
ll
be assigned t
he task. S
o
,
here the
drawback
for WLC algorithm is
it
will not check the node’s
cap
abilities li
ke Pro
c
e
s
sin
g
powe
r
, Di
sk spa
c
e
availa
bility, and memory. In ESWL
C the sele
ction
of a node will
be don
e after con
s
ide
r
ing
all the above
mentione
d ca
pabilitie
s [12].
Index nam
e
Serve
r
Alg
o
rithm
:
This algo
rithm
is develo
ped
aiming
to
re
duce the
dupli
c
ation
a
nd
red
unda
n
c
y of
data. It i
n
tegrate
s
the
process
of d
e
-du
p
lic
ation and acce
ss
p
o
int
sele
ction o
p
timization. To
optimize th
e acce
ss p
o
int
many ope
rati
ons li
ke do
wn
loadin
g
the h
a
sh
cod
e
blo
c
k o
f
data and
a
llocatio
n
of sufficient ba
n
d
width to
do
wnlo
ad fro
m
a re
sou
r
ce
b
y
establi
s
hi
ng conne
ction
s
wi
th the resource are to be d
one.
Join Idl
e
Q
u
eue Alg
o
rith
m
(JIQ):
Thi
s
algo
rithm i
s
suitabl
e for
large
scal
e d
i
stribute
d
system
s and
dynamically scala
b
le we
b service
s
. Th
is
algorith
m
is a
n
improvem
e
n
t propo
se
d fo
r
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 3, March 2
015 : 568 – 5
7
3
572
a basi
c
lo
ad
balan
cing
alg
o
rithm that works
with
dist
ributed
dispat
che
r
s. In the
basi
c
alg
o
rith
m
the ideal
processors
ha
s to
inform a
bout
their id
l
ene
ss to the
dispa
t
cher with
out
the kn
owl
edg
e
of job requ
ests whi
c
h
rem
o
ves the l
oad
balan
cin
g
tas
k
f
r
om the critic
al path [14]. To s
o
lve t
h
is
probl
em in JI
Q algorith
m
a
n
improvem
e
n
t was p
r
op
o
s
ed to first load bala
n
ce the pro
c
e
s
sors
on
the disp
atch
e
r
s an
d then lo
ad bala
n
ce the job que
ue l
ength at ea
ch
processo
r.
Table 1. Re
prese
n
tation of the perfo
rma
n
ce of some
popul
ar Lo
ad
Balanci
ng Alg
o
rithm
s
in
Clou
d Com
p
u
t
ing
CLBDM
ACO
Hone
ybee
Foraging
JIQ
ESWLC
MR
LB Min-
Min
LB Min-
Max
Throug
hput
High
Lo
w
Lo
w
Moderate
High
High
Moderate
High
Speed
Moderate
Lo
w
Moderate
Lo
w
High
Moderate
Lo
w
High
Complexit
y
Moderate
Moderate
High
Lo
w
Lo
w
High
High
Moderate
Fault Tolera
nce
Lo
w
High
Moderate
Lo
w
High
Moderate
Lo
w
High
Net
w
ork
overhead
High High
High
Moderate
Moderate
High
High
Moderate
Response time
Moderate
High
High
High Lo
w High
Lo
w
Lo
w
Resource
Utilization
High Moderate
High
Lo
w
High
High
Moderate
High
Performance
Moderate
Lo
w
Lo
w
Lo
w
High
Moderate
Lo
w
High
Migration time
High
Lo
w
Lo
w
High Lo
w High
Moderate
Moderate
5. Conclu
sion
In this pape
r we discu
s
se
d the significance of load balan
cing in
clou
d com
put
ing and
also
we di
scussed
vario
u
s
chall
enge
s that o
c
cur
while
bala
n
ci
ng lo
ad i
n
a
clo
ud
com
p
uting
netwo
rk.
In th
is p
ape
r the
classificatio
n
o
f
load
bala
n
ci
ng al
gorith
m
s wa
s
discu
ssed in
core
a
n
d
the existing algorith
m
s fo
r balan
cin
g
the load
in cl
oud computi
ng are b
r
iefe
d along with
the
improvem
ent
s p
r
op
osed t
o
improve th
eir p
e
rfo
r
ma
nce. Pe
rform
ance of the
algorith
m
s
which
were bei
ng
impleme
n
ted
for balan
ci
ng the loa
d
in clou
d computing
wa
s not up to
the
requi
rem
ents of cloud. Th
ere a
r
e different are
a
s
inv
o
lved in a
c
hi
eving the loa
d
balan
cin
g
of a
clou
d. The
p
r
oblem th
at we are fa
cing
with the
existi
ng lo
ad
balan
cing
alg
o
rith
ms i
s
th
ey are not
able to
pe
rfo
r
m
well i
n
all
the
requi
re
d
are
a
s of
lo
a
d
bal
an
cing.
For
example,
co
nsi
der Lo
ad
Balance Max
-
Min Algo
rith
m, it is goo
d in the thro
ughp
ut but its complexity is hig
h
which
degrade
s th
e
pe
rform
a
n
c
e
of the
cl
oud.
So, in
ou
r f
u
ture
work
we will
imple
m
ent chan
ge
s
to
algorith
m
s like Exponentia
l Smooth Foreca
st bas
ed
on Weig
hted
Least Co
nn
ection alg
o
rit
h
m
and Lo
ad Bal
ance Max-Mi
n Algorithm
a
nd imp
r
ove
their p
e
rfo
r
ma
nce to m
eet the re
quireme
nts
of load bala
n
c
ing in
clou
d comp
uting an
d to incre
a
se the perfo
rma
n
ce of cl
oud.
Referen
ces
[1]
NIST
T
he NIST defin
ition
of cloud c
o
m
puting. Available: http://csr
c.nist.gov/groups/SNS/cloud
computi
ng/ clo
ud-d
e
f-v15.d
o
c
.
2009.
[2]
Klaith
em Al
N
uaimi,
Na
der
Moham
ed, Ma
riam Al
N
uaim
i
, Jame
ela
Al-
J
aroo
di. A
Sur
v
e
y
of
Lo
a
d
Bala
ncin
g i
n
C
l
ou
d C
o
mputi
n
g: Ch
all
eng
es
and
Alg
o
rithm
s
.
IEEE Second Sy
m
p
osium
on Networ
k
Clou
d
Co
mputi
ng an
d App
lica
t
ions
. 201
2.
[3] w
i
kip
e
d
i
a.org,
http
://en.
w
i
k
i
p
e
d
ia.or
g
/
w
iki/L
o
ad_
bal
anc
ing
_
(
computi
ng).
[4]
V Krishna Reddy
,
Srikant
h Reddy
. A Sur
v
ey
of
Various
T
a
sk Scheduling Algorithms in Cloud
Comp
uting.
i-
mana
ger
’
s
Jo
urn
a
l on C
o
mput
e
r
Science (JCO
M).
2013; 1(1)
.
[5]
Div
ya Ch
au
dh
ar
y
,
Ra
jen
der
Sing
h Chh
ill
ar. A Ne
w
Lo
ad
Bala
ncin
g T
e
chni
que for Virtual Mach
in
e
Clou
d
Com
puti
ng Envir
onme
n
t.
International Journ
a
l
of
Co
mp
uter App
lica
t
ions.
201
3; 69
(23): 097
5 –
888
7.
[6]
Vasic N. Makin
g
cluster ap
plic
ations e
ner
g
y
-
a
w
a
r
e
. Proc of automate
d
ctrl
for datacent
ers
and clo
uds
.
200
9.
[7]
Shanti s
w
a
r
oo
p mohar
ana, R
a
ja
dee
pa
n d Ramesh, Dig
am
ber po
w
a
r. Anal
ysis Of Load
Balanc
ers I
n
Clou
d
Comp
uti
ng.
Intern
atio
n
a
l J
ourn
a
l
of C
o
mputer
Scie
n
c
e a
nd
Eng
i
n
e
e
rin
g
(IJCSE).
ISSN
22
78-
996
0. 201
3; 2(2).
[8]
Z
hang Z
,
X Z
h
ang.
A l
o
a
d
ba
l
anci
ng
mec
h
a
n
is
m b
a
sed
on
ant col
ony a
n
d
co
mp
lex n
e
tw
ork theory i
n
ope
n cl
oud
co
mp
utin
g fed
e
r
a
tion.
In
proc.
2nd I
n
ternati
o
nal
Conf
erenc
e on
Ind
u
strial
Mechatro
nic
s
and Autom
a
tio
n
(ICIMA), IEEE. 2010; 2: 240
-243.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Perform
a
n
c
e
Analysis of L
oad Balan
c
in
g Tech
niqu
es in Cloud
… (V Ravi Tej
a
Kana
kala
)
573
[9]
Hua
n
kai
Ch
en
, Proff F
r
ank W
ang, Dr
N
ahe
lia
n, Gbol
a
Akanmu.
Us
er-Priority Gu
i
ded M
i
n-Mi
n
Sched
uli
ng a
l
g
o
rith
m for Loa
d Bala
ncin
g in
Clou
d
Co
mp
u
t
ing.
proc. Nati
ona
l confer
enc
e on Para
lle
l
Comp
uting T
e
chno
log
i
es (PA
RCOMPT
E
CH), IEEE. 2013; 1-8.
[10]
V Krishn
a Re
dd
y, B T
h
irumal
a R
ao, Dr
LSS Re
dd
y,
P Sai Kir
an.
Rese
arch Iss
ues i
n
Cl
ou
d
Comp
uting.
Gl
oba
l Journ
a
l of
Computer Sci
ence a
nd T
e
ch
nol
ogy.
20
11; 11(1
1
).
[11]
OM Elzeki
, M
Z
Resh
ad, MA
Elso
ud. Im
pr
o
v
ed M
a
x-Min
A
l
gorit
hm i
n
C
l
o
ud
Comp
utin
g.
International
Journ
a
l of Co
mputer App
lic
ations
. 201
2; 50(
12): 097
5 – 88
87.
[12]
M Ran
d
les,
D
Lamb, A
T
a
leb-Be
ndi
ab.
A
Co
mp
arativ
e
Study i
n
to D
i
stribute
d
L
o
a
d
Bal
anc
in
g
Algorit
h
m
s for
Clou
d
C
o
mp
uti
ng.
Proc
eed
in
gs of 2
4
th IEE
E
Internati
ona
l
Confer
enc
e o
n
Adva
nce
d
Information N
e
t
w
ork
i
n
g
and A
pplic
atio
ns W
o
rkshops, Perth,
Australia. 20
1
0
; 551-5
56.
[13]
T
Y
,
WT
. Lee, YS Lin, YS
Lin, HL Cha
n
, JS Huan
g.
Dyn
a
mic loa
d
ba
la
ncin
g mecha
n
i
s
m bas
ed o
n
clou
d storag
e
. proc. Computing, Communic
a
tions and Applications C
onf
erenc
e (ComComAp), IEEE.
201
2: 102-
106.
[14]
Ren
X, R L
i
n,
H Z
ou.
A dy
na
mic l
o
a
d
ba
lanci
ng strate
g
y
for cloud c
o
mp
utin
g pl
atform b
a
se
d o
n
expo
ne
ntial s
m
o
o
thi
ng for
e
c
a
st.
proc. Internatio
nal
Co
nfe
r
ence
on.
Cl
ou
d Com
putin
g a
nd Intel
lig
en
t
S
y
stems (CCI
S), IEEE. 2011
; 220-22
4.
[15]
Yi Lua, Qia
o
m
i
n
Xie
a
, Gabri
e
l Kli
o
tb, Al
an
Gellerb, Jam
e
s R Larus
b, Albert Green
ber
gc.
Join-Idl
e-
Queue: A Nov
e
l Lo
ad Bal
a
n
c
ing Al
gorith
m
fo
r Dynamical
ly Scala
b
le W
eb Servic
es.
Proc. T
he
29t
h
Internatio
na
l
S
y
mp
osi
u
m o
n
Com
puter
Perform
anc
e,
Mode
lin
g, Me
asurem
ents a
nd Ev
alu
a
tio
n
,
Elsevi
er. 2011.
Evaluation Warning : The document was created with Spire.PDF for Python.