TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.4, April 201
4, pp. 2850 ~ 2
8
5
9
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i4.4731
2850
Re
cei
v
ed Se
ptem
ber 5, 2013; Re
vi
sed
No
vem
ber 3,
2013; Accept
ed No
vem
b
e
r
21, 2013
Performance Evaluation of Dynamic Load Balancing
Algorithms
Tianshu You
,
Wenhui Li, Zhiy
i
Fang*,
Hongbin
Wa
ng, Guanna
n
Qu
Coll
eg
e of Co
mputer Scie
nc
e and T
e
chno
l
o
g
y
, Jili
n Univ
e
r
sit
y
, Cha
ngc
h
un, Chi
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: fangz
y@j
l
u.e
du.cn
A
b
st
r
a
ct
Efficient task sched
uli
ng
mec
han
is
m can b
e
tter
meet the
u
s
ers
’
Q
o
S req
u
ire
m
e
n
t, and
achi
ev
e
the loa
d
ba
la
nc
ing i
n
physic
a
l
hosts, so the cl
uster syst
em w
i
th hig
h
scal
abi
lity and re
lia
bil
i
t
y can effective
l
y
improve the inform
at
ion utili
z
ation of
the system
, thereby enhanc
e the ov
erall perfor
mance of the Web
server cluster
system. In ord
e
r to buil
d
a n
e
tw
ork se
rvice system w
i
th be
tter scalabi
lity and re
lia
bil
i
ty, thi
s
paper describes the r
ound-robin sc
heduling algorithm
of
L
VS cluster syst
em
,
least-connection scheduling
alg
o
rith
m, w
e
ighted l
east-co
nnecti
on sch
e
duli
ng a
l
g
o
rith
m an
d a pri
o
r propos
ed ne
w
w
e
ighted v
a
lu
e
assig
ned
sche
duli
ng
al
gorith
m
, dy
na
mic
ad
aptive fe
ed
bac
k loa
d
b
a
la
nci
ng strate
gy. Meanw
hi
le, it ta
k
e
s
simulati
on ex
p
e
ri
ment for the roun
d-rob
i
n
schedu
lin
g al
gorith
m
of LV
S, least-conn
e
c
tion sche
dul
i
n
g
alg
o
rith
m, w
e
ighted
least-co
nnecti
on sch
e
duli
ng
alg
o
rith
m a
nd the
n
e
w
w
e
ighted
valu
e assi
gne
d
sched
uli
ng a
l
g
o
rith
m, dyna
mi
c ada
ptive fee
dback l
o
a
d
ba
l
anci
ng strateg
y
and take c
o
mp
arativ
e an
al
ysis
for the exper
i
m
ental d
a
ta, thro
ugh th
e an
alys
is and
a
ssess
me
nt, effectivel
y point o
u
t the
advant
ages
a
nd
disa
dvant
ages
of the
existin
g
lo
ad
ba
lanc
i
ng strate
gy. It is co
nduc
ive
to better i
m
p
r
ove the
existi
n
g
equ
ali
z
at
io
n al
gorith
m
p
e
rfor
ma
nce d
e
ficie
n
c
ies
an
d pro
p
o
s
e opti
m
u
m
l
o
a
d
bal
anci
ng str
a
tegy.
Ke
y
w
ords
: QoS, load ba
la
nci
ng, LVS cluster
system,
perfor
m
a
n
ce a
nalys
i
s
, performance
evalu
a
tio
n
Copy
right
©
2014 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
With the ra
pi
d developm
e
n
t of network tec
hnol
ogy, the appli
c
atio
n of the network
ha
s
been
wi
dely
popul
ar. T
h
is tren
d ma
ke
s high
-spee
d I
n
ternet
service be
com
e
ve
ry po
pula
r
in
the
worl
d of pu
bli
c
net
work
services. Fo
r ex
ample: Ama
z
on, Da
ngda
n
g
, Jing
dong
Mall and T
a
o
bao
e-comme
rce
system
s, onli
ne trai
n ticket
s b
o
o
k
i
ng
system, online
p
anic buying
system an
d ot
he
r
netwo
rk se
rv
ice
s
.
These netwo
rk services
sometim
e
s e
n
co
unt
er in
st
ant and rapi
d increa
sed
visits,
c
a
us
ing
th
e en
tir
e
s
y
s
t
e
m
s
l
ow
do
wn
or e
v
e
n
w
e
b
s
it
e cra
s
h. T
o
p
r
ovide
a la
rge
numb
e
r of
cli
ent
requ
est
s
a
n
d
better p
e
rfo
r
mance, op
erators u
s
e
th
e We
b
serve
r
cl
uste
r to
solve the
serv
er
perfo
rman
ce
issue
s
. Expand the serve
r
’s ba
nd
width
through the
load balan
ci
ng techn
o
log
y
,
increa
se the
throug
hput of
the syste
m
, stren
g
t
hen th
e network da
ta pro
c
e
ssi
ng
cap
ability, an
d
increa
se net
work flexibility and availabi
lity.
Load
balan
ci
ng technolo
g
y
make
se
rver loa
d
a
c
hi
eve a rel
a
tively balan
ce,
thereby
reducing the response ti
me of
client request
s
ta
sk, increasing the ut
ilization of sy
stem
resou
r
ces, so
that the performa
n
ce of the
syste
m
ca
n be improve
d
.
A good scheduli
ng strategy
can effe
ctively solve the probl
em of n
e
twork
co
n
g
e
s
tion, the ne
are
s
t provid
e
d
whe
n
se
rvi
c
e
need
ed to a
c
hieve l
o
cati
on ind
epen
d
ence; prov
id
es the
users with better acce
ss
qua
lity;
improve the
speed
of response f
r
om t
he
server; im
prove the
utilizat
ion efficiency of
the
server
and othe
r re
source
s; avoid
network’
s
ke
y
parts havin
g a singl
e poi
nt failure [1].
Efficient task sche
dulin
g
mech
ani
sm
s
can
bette
r
sa
tisfy the u
s
e
r
'
s
Q
o
S
requi
rements,
reali
z
ing the l
oad bala
n
ce betwe
en ea
ch physical ho
st, improving
the overall pe
rforma
nce of the
Web
serv
e
r
clu
s
t
e
r sy
st
e
m
.
Task
sc
h
edulin
g is an
important p
a
rt of t
he cluster
system
, it
execute
s
thro
ugh map
p
ing
the task o
r
tasks
subm
itt
ed by use
r
s t
o
approp
riate
resource
s, a
nd
its efficiency
will directly af
fect
the performance
of the entire
Web
server cluster system
[2-4].
By
usin
g load b
a
l
anci
ng techn
o
logy, it effectively im
proves the resource ut
ilization of
each se
rver,
improve the
service avail
a
b
ility of
the system [5].
The next ge
neratio
n of Internet
servi
c
e
s
might ne
ed more hig
h
-scal
ability and hig
h
-
availability system
s on the
high-sp
eed In
ternet a
s
pe
ct. In orde
r to p
r
ovide a mo
re
gene
ral a
nd
a
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Perform
a
n
c
e
Evaluatio
n of Dynam
ic Loa
d Balanci
ng
Algorithm
s (T
ianshu You
)
2851
variety of ne
xt-gene
ration
Internet
se
rv
ice
s
, we n
e
e
d
to d
e
velop
a hig
h
ly scal
able
and
reli
able
c
l
us
ter s
y
s
t
em
[6].
In orde
r to a
nalyze
and e
v
aluate the d
y
namic
lo
ad
balan
cing
alg
o
rithm p
e
rfo
r
mance of
Web
serve
r
clu
s
te
r
syst
em, we p
e
rf
orm
seve
ral
expe
riment
s and
me
asu
r
e va
riou
s l
oad
balan
cing
al
gorithm p
e
rf
orma
nce of the clu
s
te
r
sy
stem from lo
ad bala
n
ci
ng
indicato
rs. We
comp
ared th
e laboratory
prop
osed we
ighted lea
s
t con
n
e
c
tion sche
duling
alg
o
rithm with t
h
e
gene
ral LVS
(Linux Vi
rtual
Server)
clu
s
ters u
s
i
ng th
e
cycle, th
e le
ast conn
ectio
n
and
wei
ght
ed
least conn
ecti
on sche
dulin
g algorith
m
s.
The pa
per i
s
organi
ze
d a
s
follows: Se
ction two d
e
s
cribe
s
the d
o
mesti
c
and
foreign
resea
r
chers’
work o
n
the l
oad bal
an
cin
g
algo
rith
m p
e
rform
a
n
c
e a
nalysi
s
; Secti
on thre
e of this
pape
r
will be
use
d
fo
r pe
rforma
nce a
n
a
lysis
of
loa
d
bala
n
ci
ng
strategy; Se
ction four takes
perfo
rman
ce
analysi
s
for the algo
rithm
s
decri
bed in
section three; Section five is the su
mma
ry.
2. Related Work
In the
appli
c
ation e
n
viron
m
ent of
We
b
se
rver
clust
e
r
system,
fo
r the
cli
ent's
requ
est
task, it is
an
importa
nt factor affe
cting
t
he overall
system’
s
pe
rforman
c
e
on
how th
e lo
ad
balan
ce
r opti
m
um sche
du
les the load.
Among the
solutio
n
s of high-pe
rf
orm
ance distri
bu
ted
load, some
ta
ke hardware scheme,
othe
rs
take software sche
me. No
m
a
tter wh
ich one
to
ta
ke,
we have to consi
der the fo
llowing i
s
sue
s
[1]:
(1) After takin
g
the load ba
lanci
ng sche
me, t
he spee
d of server to
receive a
nd transmit
datagram an
d the overall
detectio
n
ca
p
ability of
load balan
cing a
r
e
the primary
consi
deration
s
.
(2) T
he lo
ad
balan
cing
scheme
sho
u
ld
be able to
m
eet the growi
ng dem
and
o
f
network
traffic, bal
an
ce th
e l
oad
of differe
nt o
peratin
g
syst
ems an
d h
a
rdwa
re
platforms, a
s
well
as
different load
flow.
(3)
Loa
d bal
anci
ng e
quip
m
ent sh
ould
have go
od
redu
nda
ncy
solutio
n
whe
n
failure
happ
en
s, ensure the availa
bility, av
oid th
e system
suffer a hug
e loss.
(4) A flexible
, intuitive and safe man
a
gement
me
a
s
ure is ea
sy
to install, configure
,
maintain an
d monitor, imp
r
ove work efficiency an
d avoid errors.
At present,
a large n
u
m
ber
of do
mestic
an
d
foreign
do
cu
ments
de
scribe the
perfo
rman
ce
evaluation
an
d an
alysi
s
of
load
bala
n
ci
n
g
st
rategy. In
these
docum
ents, they
take
perfo
rman
ce
analysi
s
an
d evaluation of
the load
bal
a
n
cin
g
algo
rith
ms and exi
s
ting algo
rithm
s
,
measure the
advanta
g
e
s
and
di
sadv
antage
s
of
the al
gorith
m
on
perfo
rm
ance, hel
p p
eople
desi
gn better
algorith
m
to meet the actu
al appli
c
ation
requi
rem
ents.
Surde
anu
M
and
othe
r p
e
ople [7]
prop
ose
d
a
kind
of dist
ribute
d
Q/A sy
stem,
usi
n
g
interqu
e
stio
n parall
e
lism a
nd dynami
c
load bala
n
ci
n
g
to improve the system’
s
throug
hput, a
nd
redu
ce the p
e
r
so
nal p
r
oble
m
s’ re
sp
on
se
time.
Mode
rn
para
llel ad
aptive
grid
co
mputi
ng
simulatio
n
ha
s va
riou
s
sizes, Iq
ba
l S and
Carey GF [8]
studie
d
and
compa
r
ed the
cha
r
a
c
teri
stics which pe
rfo
r
med
by the d
y
namic
cha
n
g
e
of the
pro
c
e
s
sors’
numb
e
r
wh
en th
e
four loa
d
-b
alan
cing
alg
o
rithm
s
a
r
e
having
parall
e
l
comp
utation.
In the docum
ents of Koyama K, Shimizu K, Ashihara
H [9],
the authors u
s
e
d
si
mulation
method to
comp
are an
d evaluate
variou
s a
d
a
p
tive load
balan
cing
st
rategy in
a
c
tual
environ
ment,
the ada
ptive load
bal
a
n
c
ing
strategy
has a b
e
tter loa
d
bal
an
ce, an
d bett
e
r to
improve
the
overall
pe
rf
orma
nce of
the cl
uste
r
system, and
can
be
re
ali
z
ed
in
pra
c
ti
cal
appli
c
ation
s
.
Shan Z, Li
n
C, Mari
ne
scu
DC’
s
docum
ents [10] i
n
trodu
ced th
e l
oad b
a
lan
c
in
g strategy
about the web se
rver cl
uster a
nd HTTP requ
est
content an
d prio
rity proce
s
s sche
d
u
ling
mech
ani
sm. This meth
od
can e
n
sure th
e quality of load bala
n
cin
g
and net
work
servi
c
e (QoS).
Yang
J, Ji
n
D, Li Y ‘
s
do
cume
nts [1
1] ta
ke
mod
e
li
ng a
nd
simul
a
tion for cl
uster-b
ased
real
we
b serv
er, buil
d
syst
em pe
rforma
nce
analy
s
is
model, fo
r da
ta se
nt to server, ea
ch
part of
the system
si
mulate an
d calcul
ate the p
a
cket del
ay. Introdu
ce th
e netwo
rk
add
ress tran
slatio
n,
ip
tunnel and
dire
ct
ro
uting three kin
d
s of
load bal
an
ci
ng. Del
a
y as
part of the
in
put mod
e
l of t
h
e
runni
ng
syste
m
and
the
sy
stem m
odel t
o
mea
s
u
r
e
th
e tran
smi
s
sio
n
data
pa
cke
t
in acco
rd
an
ce
with the syste
m
and the an
alog sy
stem perfo
rm
an
ce.
Throu
gh the
perfo
rman
ce
asse
ssm
ent and
the analy
s
is
of possibl
e p
e
rform
a
n
c
e
b
o
ttlenecks to
adju
s
t the sol
u
tion to the
p
r
oble
m
, in o
r
der
to find the maximum pro
c
e
ssi
ng capa
cit
y
of the system.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2850 – 2
859
2852
Teo YM, Ayani R’
s
do
cuments [12]
use
16
PC and
a d
r
ive tra
c
k sim
u
lator two
experim
ent platforms to take cl
uster-based web serv
er perform
a
nce and scalability experiment
analysi
s
. Th
rough th
e p
e
rforman
c
e
an
alysis
and
e
v
aluation it
can be
seen that
rou
nd-ro
bin
sched
uling
al
gorithm i
s
n
o
t as
good
a
s
the othe
r two
algorith
m
s, th
e lea
s
t-con
n
e
c
tion al
gorith
m
is e
a
sy to
im
plement, a
nd
is
suitabl
e for high
loa
d
situation. Ho
we
ver,
wi
th
the
decrea
s
e
of l
oad,
the lea
s
t-con
nectio
n
sch
e
duling
algo
rithm’s
waiting
time is m
o
re than 2
-
6 time
s of the lea
s
t-l
oad
sched
uling al
gorithm. The
performan
ce
of least-
loa
d
sch
eduli
ng
algorith
m
is the be
st, but it
need
s the time informatio
n of each
requ
e
s
t se
rvice.
Dist
ri
but
ed
s
e
rv
er
cl
ust
e
r
sy
st
e
m
i
s
a
co
st
-effective solution
to
provide
scal
able
and
reliabl
e Internet se
rvice. In order
to a
c
hieve hig
h
-q
u
a
lity servi
c
e,
it is ne
ce
ssary to adju
s
t the
system
config
uration p
a
ra
meters and u
s
e differe
nt algorithm
s to improve
syste
m
perfo
rman
ce.
3. Load Bala
ncing Stra
te
gies
After readi
ng
and stu
d
ying
a large
num
ber of p
e
rfo
r
mance an
alysis
articl
es, t
h
is a
r
ticle
elabo
rated th
e rou
nd-ro
bi
n sched
uling
algorithm, t
he lea
s
t-con
nectio
n
sche
duling al
gorit
hm,
weig
hted le
a
s
t-conn
ectio
n
sched
uling
algorith
m
an
d a ne
w la
borato
r
y pro
posed
weig
h
t
ed
distrib
u
tion sche
duling al
g
o
rithm an
d dynamic ad
ap
tive feedback load balan
ci
ng sche
dulin
g
algorith
m
of LVS cluste
r system.
3.1. Scheduling Algorith
m
of LVS Cluster Sy
stem
(1)
Rou
nd-ro
bin sche
dulin
g algorit
h
m
(Rou
nd-Ro
bin
Scheduli
ng, rr).
Rou
nd-ro
bin sched
uling al
gorithm
is
also known a
s
1
:
1 sche
duling
algo
rithm, when the
load b
a
lan
c
e
r
re
ceives
a
service
re
que
st from the
cli
e
nt, it will sen
d
the re
que
st to the ba
ck-e
nd
real
serve
r
t
o
pro
c
e
s
s in
accordan
ce
with 1:1
p
r
op
ortion. In the
reali
z
ation o
f
algorithm,
we
con
s
id
er the
state of the re
al
back-end
server to en
su
re that
the task execute pro
perly.
A
ssu
me t
hat
a clu
s
t
e
r
sy
st
em ha
s n
se
rver nod
es S
0
, S
1
, S
2
, …,
Sn
-1
, k mean
s th
e ID of
the last task assign
ed to the se
rvice n
o
de,
the algorit
hm is de
scrib
ed as follo
ws:
Input: the last sele
ct task a
ssi
gne
d se
rver nod
e and t
he total serve
r
data;
Output: sele
ct the server to
be assign
ed
task.
Rou
nd_
Robi
n
_
Algorithm
(n
ode k, Serve
r
s n).
1. Settings an
d initialize variable
s
j=k
;
do
:
2. j=
()
;
j+1
m
od n
3. if
(
S
j
is alive
)
4. k=j
;
5. return S
k
;
6. else nothi
n
g
to handle
;
7. while
(!
)
j
=
k
//if agree return to execute do, not re
turn NULL
8. return
NU
L
L
;
Cod
e
De
s
c
rip
t
ion
:
Rou
nd_
Robi
n
_
Algori
h
tm(n
ode k, Serve
r
s n)
{
int j=
k
;
do {
j=
(j+
1
)mod
n;
if(S
j
is alive)
{
k
=
j;
return S
k
;
}
els
e
{
}
}while(j!
=
k
)
return NULL;
}
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Perform
a
n
c
e
Evaluatio
n of Dynam
ic Loa
d Balanci
ng
Algorithm
s (T
ianshu You
)
2853
Rou
nd-ro
bin
algorith
m
is a
simple
sche
d
u
ling
alg
o
rith
m, load bal
an
cer
sel
e
ct a
server to
pro
c
e
ss
cu
st
omers’ requ
e
s
ts in tu
rn, so that
only when the b
a
ck-end
se
rv
er configuration has
the sa
me
co
ndition a
nd e
a
ch
user
req
uest al
most
h
a
s
t
he sa
me sy
st
em re
sou
r
ce
s co
st
,
it
c
an
provide the b
e
st load b
a
la
ncin
g efficien
cy.
(2) L
e
a
s
t-Con
nectio
n
Sche
duling Algo
rithm (Le
a
st-Co
nne
ction Sch
edulin
g, lc).
Lea
st-Conn
e
c
tion Sch
edu
ling Algorith
m
assess
th
e real
-time lo
ad statu
s
of back-end
serve
r
n
ode
s throug
h
the conne
ction nu
mber re
co
rde
d
on the l
oad
balan
ce
r, del
ivering the
ne
w
requ
est to the
serve
r
nod
e whi
c
h ha
s the
least numb
e
r of conne
ctio
ns to proce
ss.
Assu
me th
at
a cl
uste
r
syst
em ha
s
n
se
rver no
de
s S
0
, S
1
, S
2
, ...
...
, S
n-1
, C (S
i
) m
ean
s
the
numbe
r of co
nne
ction
s
of the i se
rver
node
now, S
m
means th
e
serve
r
a
s
sig
ned for th
e n
e
w
requ
est, and t
he lea
s
t-conn
ection
sched
uling
algo
rith
m can be d
e
scrib
ed a
s
follows.
Input: the last sele
ct task a
ssi
gne
d se
rver nod
e and t
he total serve
r
data;
Output: sele
ct the server to
be assign
ed
task.
Lea
st_Conn
e
c
tion_Alg
orith
m
(nod
e m, Servers n)
1. for (trav
e
r
s
e serv
e
r
cl
ust
e
r);
2. Determi
ne
wheth
e
r the
m serve
r
wo
rks p
r
op
erly; if not, return to
step 1 and m
o
ve to
the next serv
er, if so, conti
nue to step 3;
3. for (traverse the se
rver cl
uster from th
e m+1
serve
r
);
4. find the server with the least conn
ecti
ons;
5. find the correspon
ding
server, retu
rn
s S
m
;
6. return
s NULL if not fi
nd the co
rrespon
ding serve
r
.
Cod
e
De
s
c
rip
t
ion
:
Lea
st_Conn
e
c
tion_Alg
orith
m
(no
de m, Servers n)
{
for(m=
0
;m<
n
;m+
+
)
{
if(S
m
is alive)
{
for(i=
m
+
1
;i<n;i++
)
{
if(C(S
i
)< C(
S
m
))
m=
i;
}
return S
m
;
}
}
return NULL;
}
Lea
st-Conn
e
c
tion algo
rith
m is a simpl
e
dynam
ic lo
ad balan
cin
g
algorithm, it can ta
ke
con
n
e
c
tion
a
s
u
n
it an
d
dynamically tra
n
sfer the
loa
d
requ
est
wit
h
different le
ngths to th
e
back-
end serve
r
for pro
c
e
s
sing,
when all the
serve
r
s’
p
r
o
c
e
ssi
ng pe
rfo
r
man
c
e a
r
e
similar, it can
be
obtaine
d goo
d load bal
an
cing efficien
cy.
(3)
Wei
ghted
Least
-
Conn
ection S
c
he
duli
ng Alg
o
ri
thm (Weight
ed Lea
st-Co
nne
ction
Sched
uling, wlc)
The
weig
hted
least
-
conn
ection sche
duli
ng al
g
o
rithm
is the i
m
provement fo
r th
e lea
s
t-
con
n
e
c
tion al
gorithm, n
o
t only con
s
ide
r
t
he re
al-time
n
u
mbe
r
of co
n
nectio
n
s fo
r e
a
ch
se
rver, b
u
t
also
consi
d
e
r
the
pro
c
e
s
s
cap
a
city o
f
each se
rve
r
, take
divisi
on o
peration
for
con
n
e
c
tion
numbe
r of
ea
ch
se
rver
and
their p
r
o
c
e
ssing pe
rfor
ma
nce
which m
ean
s the l
oad
of ea
ch
se
rver,
and thu
s
to
determi
ne th
e real
-time
status of ea
ch
serve
r
, take
the lea
s
t rat
i
o of se
rvers to
handl
e the co
nne
ction re
qu
est se
nt by clients.
Assu
me th
at
a cl
uste
r
syst
em h
a
s n
server n
ode
s S
0
, S
1
, S
2
, .....
.,
S
n-1
, W (S
i
) m
ean
s the
weig
ht value
of the i
serve
r
no
de,
C (S
i
) mean
s th
e n
u
mbe
r
of
con
nectio
n
s
of th
e i serve
r
n
o
de
now, Sm me
ans the
serve
r
assign
ed for the new re
qu
est.
S
m
server me
ets the followi
ng co
ndition
s:
(C (S
m
) / CS
UM) / W (S
m
)
= min {(
C (S
i
)
/ CSUM) / W
(S
i
)} (i =
0,1,
...,
n-1)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2850 – 2
859
2854
W
h
er
e W (
S
i
) i
s
n
o
t zero,
be
cau
s
e
in t
h
is
rou
nd l
o
o
k
up
CS
UM i
s
a
con
s
tant,
so th
e
judgme
n
t con
d
itions
can b
e
simplified a
s
:
C (S
m
) / W (S
m
) = min {C
(
S
i
) / W (S
i
)} (i
=
0,1, ..., n-1)
Whe
r
e W (S
i
) is n
o
t zero, the jud
g
me
nt
con
d
ition
can
be
cha
nge
d i
nequ
ality ope
ration to
comp
are.
C(S
m
) / W
(
S
m
) > C
(
S
i
) / W(
S
i
)
As divisi
on
compa
r
ison o
peratio
n
con
s
ume
mo
re CPU re
so
ur
c
e
s
than m
u
ltiplication
comp
ari
s
o
n
o
peratio
n, and
the float point divisi
on is
not allowed i
n
the Linux kernel,
conn
ection
numbe
r an
d
weighte
d
di
vision comp
arison o
per
a
t
ion can
be
conve
r
ted to
the con
n
e
c
tion
numbe
r
and
weig
hted m
u
l
t
iplication
co
mpari
s
o
n
o
p
e
ration
to co
mpare,
serve
r
weight
s are
all
greate
r
than
zero, so the j
udgme
n
t con
d
ition ca
n be
further o
p
timized a
s
follo
w:
C(S
m
)
×
W(S
i
) >
C
(
S
i
)×
W
(
S
m
)
At the same time ensure that the server has
a weight of zero, the server will
not be
sched
uled,
a
nd the
weig
hted le
ast
-
co
nne
ction
sch
edulin
g al
gorithm can
be
de
scrib
ed
as
follows
:
Input: the last sele
ct task a
ssi
gne
d se
rver nod
e and t
he total serve
r
data;
Output: sele
ct the server to
be assign
ed
task.
Weig
hted_
Le
ast_
Con
n
e
c
tion_Algo
r
ithm
(nod
e m, Servers n
)
.
1. for (travers
e server c
l
us
ter);
2. determin
e
wheth
e
r the
weig
ht value of m
serve
r
is above 0,in a
nother
way whether it
works p
r
op
erl
y
; if not, return to st
ep 1
and move to
the next se
rver, if so,
cont
inue to
step 3;
3. for (traverse the se
rver cl
uster from th
e m+1
serve
r
);
4. find the server with the least ratio b
e
twee
n co
nne
ction numbe
r a
nd wei
ght value;
5. find the correspon
ding
server, return
s S
m
;
6. return
s NULL if not fi
nd the co
rrespon
ding serve
r
;
Cod
e
De
s
c
rip
t
ion:
Weig
hted_
Le
ast_
Con
n
e
c
tion_Algo
r
ithm
(nod
e m, Servers n
)
{
for(m=
0
;m<
n
;m+
+
)
{
if(W(S
m
)>
0)
{
for(i=
m
+
1
;i<n;i++
)
{
if(C(S
m
)*
W(
S
i
)>
C(S
i
)*W(S
m
))
m=
i;
}
return S
m
;
}
}
return NULL;
}
Weig
hted le
a
s
t-conn
ectio
n
algo
rithm i
s
t
he mo
st effici
ent algo
rithm
on loa
d
b
a
lan
c
ing
of
LVS clu
s
ter system, it not
only con
s
ide
r
the n
u
mb
e
r
of real
-time
conne
ction
of
the serve
r
, b
u
t
also th
e differen
c
e
between the
ba
ck-e
nd
se
rv
er pro
c
e
s
sing
perfo
rman
ce,
truly reali
z
e
a
dynamic lo
ad
balan
cing.
3.2. A Ne
w
Weigh
t
ed Va
lue
Assign
e
d
Schedulin
g Algorithm
In the new
weighted valu
e
assigne
d sch
edulin
g algo
ri
thm [13], we use the
CP
U idle rate
and
me
mory idle
rate of
b
a
ck-e
nd sub
-
serve
r
whi
c
h perio
dically
a
c
qui
red by
lo
ad
bal
an
cing to
cal
c
ulate the
weig
ht of the
back-end
sub
-
se
rver. Th
e weig
hted
valu
e function ex
pre
ssi
on can be
expre
s
sed a
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Perform
a
n
c
e
Evaluatio
n of Dynam
ic Loa
d Balanci
ng
Algorithm
s (T
ianshu You
)
2855
F(S
i
)=
0.6×
(
1
-C
(
S
i
))+
0.4×
M(
S
i
)
(1)
C(S
i
), M
(
S
i
),
W(S
i
)
den
ote
the CPU uti
lization, mem
o
ry idle rate
and the
se
rver no
de
weig
ht of the each ba
ck-e
nd se
rver, F(S
i
) denote
s
the new
weig
ht of back-en
d serve
r
nod
e.
After the ne
w weig
ht cal
c
ul
ation we dete
r
mine
wh
ethe
r the n
e
w
wei
ght sh
ould
be
written to th
e
IPVS schedul
ing according to the following formul
a:
()
()
()
;
()
()
(
)
()
.
ii
i
ii
i
i
WS
F
S
B
W
S
WS
F
S
B
W
S
F
S
(2)
B is a pre-giv
en boun
da
ry value, the value of
B indicates the algo
rithm effective assess
value after the improvem
ent of weig
h
t
ed least
-
con
nectio
n
sche
duling al
gorit
hm. From th
e
formula
abov
e it is
not diffi
cult to
se
e th
at the
sm
alle
r the value
of
B, the high
er
freque
ncy
of the
weight value to write to IPVS. If the B
value is
0, then every load inf
o
rmat
ion result co
llected can
satisfy the bo
unda
ry value con
d
ition. Th
e load re
su
lts collecte
d
will
also be written to the IPVS
sched
uling. If
B =
1, then
the lo
ad info
rmation
re
sult
colle
cted
can
not satisfy th
e bo
unda
ry v
a
lue
condition, wit
h
IPVS without any change, st
ill maintain the original scheduling m
ode.
3.3. D
y
namic Adap
tiv
e
F
eedba
ck Loa
d Balancing
Strateg
y
A new wei
ght
ed distri
butio
n sched
uling
algorith
m
is p
e
riodi
cally ma
king a
requ
est by the
load bal
an
ce
r to colle
ct the load info
rmation of
ba
ck-en
d
sub-server n
ode,
this app
roa
c
h
increa
se
s the
commu
nication overhea
d
of the w
hol
e clu
s
ter
system, also in
crea
se
s the lo
ad
balan
ce
r’s bu
rden.
F
eed
ba
ck dynami
c
adaptive
lo
ad
balan
cin
g
st
rategy [14] p
e
riodi
cally a
n
d
adaptively col
l
ect their
own
load
by the
back-end
se
rver nod
e, an
d se
nd the lo
ad informatio
n to
the loa
d
b
a
la
nce
r
. Thi
s
ap
proa
ch
redu
ces th
e
comm
unication
ove
r
hea
d of th
e
wh
ole
clu
s
te
r
system, an
d also
red
u
ce
s
the burden of
the load bal
a
n
ce
r. In ord
e
r to avoid the instant ove
r
lo
ad
of the l
oad
b
a
lan
c
er an
d
a si
ngle
serv
er
nod
e,
al
so
intro
d
u
c
e th
e serve
r
nod
e loa
d
red
u
n
dan
cy
value this pa
rameter to furt
her optimi
z
e t
he perfo
rma
n
ce of the clu
s
ter system.
About the
we
ight cal
c
ul
ation of the b
a
c
k-en
d sub
-
server,
we u
s
e se
rver
perf
o
rma
n
ce
and dyna
mic
load value th
ese t
w
o imp
o
r
tant pa
ramet
e
rs
to ma
ke more accu
rat
e
asse
ssmen
t
of
the cu
rre
nt lo
ad ca
pa
city of the se
rve
r
n
ode
s, we a
ssume that the
weig
ht value i
s
W, the
weig
h
t
W ab
out the
serve
r
pe
rf
orma
nce and
dynamic l
o
ad value
s
th
ese t
w
o pa
rameters
can
be
cal
c
ulate
d
as
follows:
W(S
i
)=
L(
S
i
)/P
(
S
i
)
(3)
P(S
i
) is the
performance
indicator of t
he se
rver, decided by serv
er’s CPU utilization,
available di
sk sp
ace and
memory spa
c
e. L(S
i
) is the dynamic lo
ad value of the se
rver no
de,
deci
ded by CPU utilization, disk I/O
read and
wr
ite spee
d, memory utilization, net
work
band
width util
ization a
nd re
que
st respon
se time.
4. Performa
nce An
aly
s
is
and Ev
aluation
In the web
se
rver
clu
s
ter
system, it mai
n
ly us
e
effect
ive mea
s
ure i
ndicators of
server to
analyze an
d
evaluate
whether it i
s
good
or
b
ad the
clu
s
ter
system’
s
load b
a
lan
c
ing
perfo
rman
ce.
These effective measu
r
e
indicators
[1
] mainly incl
ude
CPU util
ization, me
m
o
ry
usa
ge, ban
d
w
idth utilizati
on, disk IO throu
ghp
ut an
d netwo
rk IO
through
put, the numbe
r
o
f
compl
e
ted service p
e
r u
n
it time, the numbe
r
of con
n
e
c
ted cl
ients pe
r uni
t time, and the
respon
se tim
e
to complete
a task reque
st.
This arti
cle
fo
cu
se
s o
n
a n
e
w
wei
ghted
value a
s
sign
ed
sched
ulin
g alg
o
rithm,
dynamic
adaptive
fee
dba
ck strat
e
gy
and roun
d-robin sche
du
ling
algo
rithm of LVS cluster sy
stem
, the
least-co
nne
ct
ion sched
ulin
g algorithm a
nd weig
hted
least-co
nne
ct
ion sched
ulin
g algorithm f
o
r
perfo
rman
ce analysi
s
an
d evaluation.
Compa
r
e
an
d
analyze the a
d
vantage
s an
d disa
dvanta
ges
of existing a
l
gorithm
s in
orde
r to better im
prove and optimi
z
e
the algorith
m
. In Choi E’s
pape
r
,
in
ord
e
r to
analy
z
e
the dynami
cs of serve
r
p
e
rforman
c
e
du
e
to the
wo
rklo
ad, the
autho
rs
perfo
rm seve
ral expe
rime
nts. With
so
me kn
owl
edg
e of ho
w to
sele
ct the p
r
oper
pe
rform
ance
cou
n
ters
and
their prope
r
threshold
val
ue, they
com
pare
the
pe
rforma
nce
results of t
he AL
BM
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2850 – 2
859
2856
clu
s
ter with
the g
ene
ral
L
VS clu
s
ter u
s
ing
RR,
LC a
nd
WL
C
sche
duling
algo
rit
h
ms. T
he A
L
BM
clu
s
ter a
c
hie
v
es the better pe
rform
a
n
c
e than the
LVS by balancin
g the lo
ads am
ong t
he
serve
r
s. It also has ve
rified
and an
alyze
d
that
among
roun
d-robi
n sched
uling al
gorithm of LV
S
clu
s
ter sy
ste
m
, least-con
n
e
ction
sched
uling al
go
rith
m and wei
ght
ed lea
s
t-conn
ection
sched
uling
algorith
m
, we
ighted lea
s
t-conne
ct
ion
scheduli
ng alg
o
r
ithm ha
s bet
ter pe
rform
a
n
c
e in the th
re
e
algorith
m
s [6]
.
Here we ta
ke simulatio
n
experim
ent for the ne
w weighted valu
e
assi
gne
d scheduli
n
g
algorith
m
,dynamic
ada
ptive feedba
ck
load b
a
la
n
c
in
g strate
gy an
d wei
ghted l
east-co
nne
cti
on
sched
uling al
gorithm
on t
he re
que
st resp
on
se
tim
e
, and ta
ke
comp
arative analysi
s
for t
he
experim
ental
data. Labo
ratory
expe
ri
ments we
re done unde
r
l
o
cal enviro
n
m
ent,
use
r
a
c
cess
client
uses WAS (Micro
soft We
b A
ppli
c
a
t
ion Stre
ss T
ool) to
si
mula
te the u
s
e
r
p
r
essure. Divid
e
d
to 7 expe
rim
ental g
r
ou
ps,
set th
e first
grou
p a
c
ce
ss numb
e
r
of u
s
ers to
10
0, the
se
cond
g
r
oup
acce
ss count
is 20
0, an
d
set
30
0,400
,500,600,7
0
0
re
sp
ectively
in
ord
e
r. E
a
ch
g
r
ou
p ta
ke
simulatio
n
te
sts
with wei
g
hted lea
s
t-co
nne
ction sch
edulin
g algo
ri
thm , a new
weig
hted val
u
e
assign
ed sch
edulin
g algo
ri
thm and dyna
mic ada
pt
ive feedba
ck loa
d
balan
cing
strategy [13].
0
100
200
300
400
500
60
0
700
0
1000
2000
3000
4000
5000
6000
TT
FB
Av
g(
ms
)
The
numb
er o
f co
nnect
ion
A n
ovel
weig
ht v
alue
Dyn
amic
adap
tive
feed
back
wlc
Figure 1. Re
spon
se Time
Comp
ari
s
o
n
of the Two Algorithm
s
0
1
00
20
0
3
0
0
40
0
5
0
0
60
0
7
0
0
0
200
400
600
800
1000
1200
1400
1600
1800
2000
T
TFB A
vg(m
s)
t
he
num
ber
o
f c
onn
ect
ion
wl
c
dy
na
mic
ad
apt
iv
e f
eed
bac
k
Figure 2. Re
spon
se Time
Comp
ari
s
o
n
of the Two Algorithm
s
We can cle
a
rly see tha
t
from Figure 1,
when the numb
e
r
of user requ
est task
con
n
e
c
tion
s is less than
230, the average
wait
re
spo
n
se time of the impro
v
ed algorith
m
is
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Perform
a
n
c
e
Evaluatio
n of Dynam
ic Loa
d Balanci
ng
Algorithm
s (T
ianshu You
)
2857
slightly longe
r than which o
f
the wlc algo
rithm.
The m
a
in rea
s
o
n
for this phe
nom
enon i
s
that the
improved algorithm
has
a new
weight
c
a
lc
ulat
ion and IPVS write proces
s
,
that c
o
nsumes a
certai
n
amou
nt of time.
Whe
n
the
n
u
m
ber of ta
sk co
nne
ction
i
s
g
r
e
a
ter tha
n
23
0,
with t
h
e
increa
se n
u
m
ber of the
re
q
u
ire ta
sk
co
n
nectio
n
,
the a
v
erage
wait
resp
on
se time
of the improved
algorith
m
sig
n
ificantly less than th
at of wlc
al
go
rithm. No m
a
tter the
wlc
al
gorithm, o
r
the
improve
d
alg
o
rithm, with the increa
se
numbe
r
of task
requ
est
conne
ction, when it rea
c
he
s a
certai
n number, the
whol
e clus
ter sy
stem will reach the satu
ration. Because the resource
utilization
of each server
cluster
system is al
ready saturated.
When
this occurs,
we should
con
s
id
er th
e
clu
s
ter sy
ste
m
to b
e
exp
a
nded,
addi
ng
a ce
rtain am
ount
of ba
ck-end real
serv
ers
to
in
c
r
ea
se
ava
ila
b
l
e
se
r
v
er
r
e
s
o
ur
ce
s
.
From th
e ave
r
age
wait re
spon
se
req
u
e
s
t time
comp
arison
of u
s
e
r
reque
st in
Figure 2,
we
can
see t
hat when th
e
numbe
r of
user
requ
est
s
i
s
less tha
n
2
4
0
, the ave
r
ag
e wait re
sp
on
se
time of the i
m
prove
d
alg
o
rithm i
s
slig
htly l
onge
r th
an
whi
c
h of
the wl
c al
go
rithm. The m
a
in
rea
s
on fo
r th
is ph
enom
en
on is th
at the improv
ed a
l
gorithm
ha
s
weig
ht value
cal
c
ulatio
n
and
IPVS write process, this proces
s consumes a
cert
ain amount of
time. When t
he user request
numbe
r is g
r
eater tha
n
24
0, with the in
cre
a
se num
b
e
r of the u
s
e
r
req
u
e
s
ts, d
y
namic a
dapt
ive
feedba
ck lo
ad bal
an
cin
g
strategy resp
ond
s fa
ster than th
e wei
ghted
least
-
conne
ction
sched
uling
al
gorithm, dyn
a
mic
adaptiv
e feedb
ack l
oad b
a
lan
c
in
g strategy is more
effecti
v
ely
enha
nce the
throug
hput
of the syste
m
, and bette
r
bala
n
ce th
e system lo
a
d
, can effe
ctively
improve the p
e
rform
a
n
c
e o
f
the whole cl
uster
system.
In the following we mai
n
ly take com
pari
s
on a
nal
ysis for a n
e
w weighted
value
assignm
ent sched
uling
al
gorithm, dyn
a
mic ada
ptiv
e feed
ba
ck
strategy
and
wei
ghted
le
ast-
con
n
e
c
tion scheduli
ng algo
rithm from th
e respon
se ti
me and sy
ste
m
throug
hput
aspe
cts.
Figure 3. Not
e
Ho
w the Ca
pti
on is Cente
r
ed in the Col
u
mn
As can b
e
se
en from Figu
re 3, when the
numbe
r of user’
requ
est conne
ction is
not big,
a
ne
w weig
hted
all
o
catio
n
sched
uling al
gorithm
an
d
dynamic ad
a
p
tive feedb
ack lo
ad
balan
cing
strategy hav
e longe
r re
spon
se time for user
s’ re
q
uest than th
e weig
hted l
east-co
nne
cti
o
n
sched
uling al
gorithm, the
main re
ason
is the form
e
r
two algo
rithm
s
incre
a
se th
e burden of the
load sch
edul
er when
cal
c
ulate the wei
ghts val
ue,
write and
rea
d
the weig
ht value
s
, and lo
ad
balan
cing
assign tasks,
cau
s
ing a
ddition
al comm
uni
cation overhea
d, cost
some
resou
r
ces. Wi
th
the increa
sin
g
numbe
r of acce
ss
requ
e
s
ts conn
ectio
n
s of the cu
rrent use
r
s, th
e new
weig
hted
assign
ed
sch
edulin
g algo
rithm and dy
namic
ada
ptive feedba
ck load bal
an
ci
ng strategy
are
sup
e
rio
r
to
th
e weighte
d
le
ast-con
n
e
c
tio
n
sch
eduli
ng
algorith
m
o
n
perfo
rman
ce.
In a
ddition, i
t
can
also be
seen from the
diagr
am that
the dynami
c
adaptive fe
e
d
back lo
ad b
a
l
anci
ng
strate
gy
has lea
s
t
re
spon
se tim
e
, i
t
mean
s th
at dynami
c
ad
aptive feed
ba
ck st
rategy
i
s
su
pe
rior to t
he
new
weig
hted
distributio
n sche
duling al
g
o
rithm on p
e
rforman
c
e.
0
100
200
300
400
500
60
0
700
0
200
400
600
800
1000
1200
1400
1600
Byte
s Recv Rate
(Kb/s)
th
e
nu
mb
er
o
f
co
nn
ec
tio
n
a
n
ov
el
w
ei
gh
t
va
lu
e
d
yn
am
ic
a
da
pt
iv
e
fe
ed
ba
ck
w
lc
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2850 – 2
859
2858
0
100
200
300
400
500
600
7
0
0
0
200
400
600
800
1
000
1
200
1
400
1
600
Byt
es
Recv
Rat
e(Kb
/s)
th
e n
umb
er
of
con
ne
cti
on
a
nov
el
wei
gh
t v
alu
e
dy
nam
ic
ada
pt
ive
fe
edb
ack
wl
c
Figure 4. Not
e
Ho
w the Ca
pti
on is Cente
r
ed in the Col
u
mn
For dyna
mic
page
s, we
can se
e from
t
he above di
agra
m
, whe
n
the numbe
r of use
r
acce
ss
req
u
e
st conn
ecti
ons i
s
le
ss t
han 2
00,
the
throug
hput
of a ne
w we
ighted a
s
sig
ned
sched
uling al
gorithm and dynamic
ada
ptive feedba
ck loa
d
bal
an
cing poli
c
y is
slightly le
ss t
han
whi
c
h of the
weig
hted le
ast-con
n
e
c
tio
n
sche
du
lin
g
algorith
m
, mainly due
to the first two
algorith
m
s in
cre
a
se the b
u
rde
n
of the
load sch
edul
er in the p
r
o
c
e
ss of
weig
ht calculation
,
the
weig
ht write,
weight
rea
d
and the t
a
sk a
ssi
gnm
ent of load
balan
ce
r, ca
usin
g additio
nal
comm
uni
cati
on overhea
d
,
spent
som
e
re
sou
r
ce
s.Whe
n
u
s
er acce
ss req
u
e
st
conn
ecti
on
rea
c
he
s abo
ut 400, th
e t
h
rou
ghp
ut of
the sy
st
em
reache
s the
maximum. O
v
erall the
dy
namic
adaptive feed
back lo
ad bal
anci
ng
strate
gy has th
e m
a
ximum thro
u
ghput. It mea
n
s the
dynam
ic
adaptive fee
dba
ck
strate
gy is su
peri
o
r to
the n
e
w weighte
d
sch
eduli
n
g
algorithm
on
perfo
rman
ce.
In summ
ary, throu
gh the
compa
r
ative a
nalysi
s
of th
e three
algo
ri
thms o
n
pe
rforma
nce,
dynamic
ada
ptive feedba
ck lo
ad b
a
la
ncin
g strateg
y
has bette
r perfo
rman
ce than the
new
weig
hted a
s
sign
ed sch
e
duling al
gorit
hm and rou
nd-robi
n sch
edulin
g algo
rithm of the LVS
clu
s
ter
syste
m
, the lea
s
t-co
nne
ction
sched
u
ling a
l
gorithm
and
weig
hted l
east-co
nne
cti
o
n
sched
uling al
gorithm.
5. Conclusio
n
This p
ape
r an
alyzed the p
e
r
forma
n
ce an
alysis of the d
o
mesti
c
and f
o
reig
n do
cum
ent on
the clu
s
ter
sy
stem loa
d
bal
anci
ng alg
o
rit
h
m, its main
purp
o
se is to
resea
r
ch and
desi
gn a b
e
tter
perfo
rman
ce
of the load balanci
ng a
l
gorithm,
whi
c
h ma
ke
s the cluste
r sy
stem with hi
gh
scalability an
d reliability, and can effectively
impro
v
e the utilization and pe
rforma
nce of the
system
information. Th
ro
ugh th
e d
e
scriptio
n,
an
al
ysis
and
ev
aluation
of a
ne
w
weig
hted
distrib
u
tion
sche
duling
al
gorithm, dyn
a
mic
adapt
iv
e feedb
ack l
oad b
a
lan
c
in
g strategy a
nd
roun
d-robi
n
sche
duling
alg
o
rithm
of LV
S clu
s
ter
syst
em,
lea
s
t-con
nectio
n
sch
e
duling algo
rit
h
m
and
weig
hted
least
-
conn
ection sche
duli
ng alg
o
rith
m
on pe
rforman
c
e. Fin
a
lly th
e con
c
lusi
on
is
that dynami
c
ada
ptive fee
dba
ck loa
d
b
a
lan
c
in
g
st
rat
egy ha
s bett
e
r
perfo
rma
n
c
e. T
h
ro
ugh
t
h
e
analysi
s
a
nd
evaluation
of the pe
rf
orma
nce
of the loa
d
bala
n
ci
ng
st
rategy of the
clu
s
ter
syste
m
,
whi
c
h
can
effectively point
out the a
d
van
t
ages
and
di
sadvantag
es
o
f
the existing
load b
a
lan
c
in
g
strategy, is
condu
cive to
better imp
r
ov
e the def
ici
e
ncie
s of the
existing eq
ua
lization al
gori
t
h
m
and propo
se
optimum pe
rforma
nc
e load
balan
cing
strategy.
Ackn
o
w
l
e
dg
ements
This research is
su
pport
ed by
China Jilin Province
Natural
Science
Funds under
prog
ram of
parall
e
lization
and dynam
ic sche
du
lin
g model of CPU/GP
U coope
rative hi
gh
perfo
rman
ce comp
uting
cl
uster (20
121
5189
).
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Perform
a
n
c
e
Evaluatio
n of Dynam
ic Loa
d Balanci
ng
Algorithm
s (T
ianshu You
)
2859
Referen
ces
[1]
Xu
Lan. L
oad
b
a
la
ncin
g techn
o
lo
g
y
r
e
searc
h
.
Agriculture&T
e
chn
o
lo
gy.
200
9; 29: 155-
157.
[2]
Sun H, C
h
u
n
g
uan
g T
.
Optimi
zing M
u
lti-a
g
e
n
t
MicroGrid Re
source Sc
he
du
ling
b
y
C
o
-Evol
u
tion
ar
y
w
i
t
h
Preferenc
e.
T
E
LKOMNIKA Indon
esia
n Jour
nal
of Electric
al
Engin
eeri
n
g
. 2
013; 11(
12): 72
22-7
229.
[3]
Jian
g W
,
Z
han
g J, Li J, et al.
A Resourc
e
S
c
hed
uli
ng Strat
e
g
y
i
n
Cl
oud
Comp
uting B
a
sed o
n
Multi-
age
nt Genetic
Algorit
hm.
T
E
LKOMNIKA Indones
ian J
our
n
a
l of
El
ectrical
Engi
ne
erin
g
. 201
3; 11(
11
)
:
656
3-65
69.
[4]
Quan B, L
i
H,
Le Z
.
D
y
nam
ic
Routi
ng a
nd
R
e
sourc
e
Assig
n
ment Al
gorit
h
m
In sloted
opt
ical N
e
t
w
orks.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2013; 1
1
(4): 1
813-
182
1.
[5]
ZHU Jia
y
u,
XI
AO, Dan, WANG Fei. Multi-di
m
ensi
ona
l Qo
S constrai
ne
d
sched
uli
ng m
e
chan
ism bas
ed
on lo
ad b
a
la
nci
ng for Clo
ud co
mputin
g.
Co
mp
uter Engi
ne
erin
g and Ap
pl
icati
ons
. 201
2.
[6]
Choi E. Perfor
mance test an
d ana
l
y
sis for
an
ada
ptive l
oad b
a
l
anci
n
g
mechan
ism o
n
distrib
u
te
d
server cluster s
y
stems.
F
u
ture
Generatio
n Co
mp
uter Syste
m
s
. 2004; 20: 23
7-24
7.
[7]
Surdeanu M,
Moldovan DI,
Ha
ra
bag
iu SM.
Performa
nce
ana
l
y
sis
of a
d
i
stribute
d
q
ues
tion/ans
w
e
ri
n
g
sy
s
t
e
m
.
IEEE Transactions on Parallel and
Distribut
ed Sys
t
em
s
. 20
02; 13
: 579-59
6.
[8]
Iqbal
S, Car
e
y
GF
. Performan
c
e a
nal
ys
is of
d
y
nam
ic l
oad
bal
anci
n
g
al
go
rithms
w
i
th
vari
abl
e n
u
mb
er
of processors.
Journ
a
l of par
a
llel a
nd d
i
strib
u
t
ed co
mputi
n
g
.
2005; 6
5
: 934-
948.
[9]
Koy
a
ma K, Shimizu K, As
hihara H,
et al.
Per
f
orma
nce
ev
al
uatio
n
of ad
apt
ive l
o
a
d
bal
anc
ing
po
lici
e
s
in distributed s
ystem
s.
Proc
e
edi
ngs of IEE
E
Sing
ap
ore I
n
ter
nati
o
n
a
l C
onfere
n
ce
on.
IEEE. 1993; 2:
606-
611.
[10]
Shan Z
,
Li
n C,
Marinesc
u
D
C
. Mode
lin
g a
nd p
e
rf
ormanc
e an
al
ysis
of QoS-a
w
ar
e lo
a
d
ba
lanc
in
g o
f
w
e
b-server clusters.
Comp
ute
r
Netw
orks
. 2002; 40: 23
5-25
6.
[11]
Yang J, Ji
n D,
Li Y. Mod
e
li
ng
and s
i
mul
a
tio
n
of performa
nc
e an
al
ysis for
a cluster-
base
d
w
e
b server
.
Simulati
on Mo
dell
i
n
g
Practice
and T
heory
. 2
006; 14: 1
88-2
00.
[12]
T
eo YM, Ay
a
n
i
R. Compar
iso
n
of loa
d
b
a
la
n
c
ing strate
gies
on clust
e
r-bas
ed
w
e
b serv
ers.
Sim
u
lation
.
200
1; 77: 185-
195.
[13]
W
ang Ho
ng
bi
n, F
ang Z
h
i
y
i,
Qu Guanna
n, Z
hang
Xi Z
h
ang, Yu
nchu
n. A novel
w
e
ig
ht distributi
o
n
sched
uli
ng a
l
g
o
rithm.
Journ
a
l
of Informatio
n and C
o
mput
ati
ona
l Scienc
e
. 201
1; 8: 1235-
124
4.
[14]
W
ang H
o
n
gbi
n, F
ang Z
h
i
y
i,
Cui S
h
u
ang.
D
y
namic
ad
ap
tive feed
back
of loa
d
b
a
l
anc
ing strate
g
y
.
Journ
a
l of Infor
m
ati
on a
nd C
o
mp
utatio
nal Sc
ienc
e
. 201
1; 8: 1901-
19
08.
[15]
W
ang Ho
ng
bi
n. Rese
arch o
n
ad
aptive
lo
a
d
bal
anc
ing s
c
hed
uli
ng strat
e
g
y
in W
e
b s
e
rver cluste
r
s
y
stem. Jili
n U
n
iversit
y
. 20
13.
Evaluation Warning : The document was created with Spire.PDF for Python.