TELKOM
NIKA
, Vol.11, No
.11, Novemb
er 201
3, pp. 6563
~6
569
e-ISSN: 2087
-278X
6563
Re
cei
v
ed Ma
rch 2
8
, 2013;
Re
vised June
20, 2013; Accepte
d
Jul
y
2
0
, 2013
A Resource Scheduling Strategy in Cloud Computing
based o
n
Multi-agent Genetic Algorithm
Wuxu
e Jian
g
1,2
, Jing Zh
ang*
1
, Junhu
ai Li
1
, Hui Hu
3
1
School of Co
mputer Scie
nc
e and En
gi
neer
ing,
Xi’
an Un
iv
ersit
y
of T
e
chn
o
lo
g
y
,
Xi
’an
71
004
8, Shaa
n
x
i,
Chin
a
2
Departme
n
t of Computer En
g
i
ne
erin
g, Don
g
gua
n
Pol
y
tec
h
nic, Don
g
g
uan
523
80
8, Guan
gdo
ng, Ch
ina
3
Departme
n
t of Science a
nd R
e
searc
h
, Huizh
ou Un
iveres
it
y
,
Huizh
ou 5
160
07, Guan
gdo
n
g
, Chin
a
*Co
rre
sp
ondi
ng autho
r, e-mail: zhan
gjin
g@xaut.ed
u
.cn
A
b
st
r
a
ct
Reso
urce sch
e
duli
ng strate
gi
es in cl
oud c
o
mp
ut
in
g are
us
ed eit
her to i
m
prove syste
m
oper
ating
efficiency,
or
to i
m
pr
ove
us
er satisfacti
on
. T
h
is
p
a
p
e
r
prese
n
ts a
n
i
n
tegrate
d
sc
h
edu
lin
g strate
g
y
consi
deri
ng
bo
th resourc
e
s cr
edi
bil
i
ty an
d us
er satisfac
tio
n
. It takes user s
a
tisfaction
as o
b
jectiv
e functi
o
n
and res
ourc
e
s credi
bil
i
ty as a part of the user
satisfac
tion, a
nd rea
l
i
z
e
s
o
p
ti
ma
l sche
dul
in
g
by using g
e
n
e
tic
alg
o
rith
m. W
e
integr
ate this sc
hed
uli
ng strategy into Ag
ent
subs
eq
ue
ntly and pr
opos
e clou
d co
mputi
n
g
system
architecture based on Mult
i-
agent. The num
eric
al res
u
lts show
that this sc
heduling strat
e
gy
improves
not o
n
ly the syste
m
oper
ating
effi
ci
ency, but als
o
the user satisfa
c
tion.
Ke
y
w
ords
: cloud co
mputi
n
g
,
resources cr
edi
bil
i
ty, custom
er s
a
tisfactio
n
, resources s
c
hed
uli
ng,
mul
t
i
-
age
n
t
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Clou
d compu
t
ing syste
m
take
s a
d
vant
age of
the
tens
of millio
ns of i
d
le
co
mputing
resou
r
ces on
the Inte
rnet
so
that it
can b
e
mo
re
powerful th
a
n
the
ce
ntral
i
zed
computi
n
g
sy
st
em.
Ho
w
e
v
e
r,
it
s
r
e
s
our
ce
s
c
h
e
d
u
ling f
a
ce
s
c
hallen
ges
d
u
e
t
o
it
s ma
s
s
iv
e
re
sou
r
c
e
s,
hetero
gen
eo
us natu
r
e, an
d netwo
rk
co
mmuni
cation
delay.
Sched
uling
o
f
reso
urce
s i
s
to assig
n
the
jobs
su
b
m
itted by users to
app
ro
priate
resou
r
ces re
aso
nably to meet the ne
eds of u
s
ers and maximi
ze sy
stem o
peratin
g ben
efits.
Schola
r
s h
a
ve put
forwa
r
d a
serie
s
of sche
dulin
g
strategi
es for the
sche
duli
ng p
r
o
b
lem
s
of
Clou
d Comp
uting an
d di
stributed
re
so
urces. [1
,
2] prop
ose a
di
stribute
d
resource
sche
d
u
ling
strategy fo
re
casting m
odel
based on
ant
colo
ny al
go
ri
thm to improv
e the dynami
c
an
d re
al-tim
e
perfo
rman
ce
of the distri
buted comp
u
t
ing and r
eal
-time pe
rformance an
d effectivene
ss of
r
e
sour
ce sc
heduling. [3
,
4
]
prop
ose a b
a
lan
c
ed
sche
duling
algo
rithm. The
algo
rithm calculat
es
job orde
r val
ues
ba
sed
o
n
job d
epe
n
den
cie
s
, and
adju
s
ts the
jobs
acco
rdin
g to that ord
e
r
values. As
a
result, key jobs
will
be fini
shed as soon as possible
,
the wait time
between jobs is
redu
ce
d, an
d
the job
co
m
puting
wo
rkfl
ow exe
c
uti
o
n
time is fin
a
ll
y redu
ce
d. It prove
s
that
th
e
sho
r
t job co
mputing exe
c
ution time o
f
this
algorith
m
in job ma
nagem
ent sy
stem ha
s
strong
s
u
pe
r
i
or
ity. [5
,
6] provide
a di
stribute
d
syste
m
scheduli
ng mo
del for th
e p
r
oble
m
of
weak
dynamic
re
g
u
latory
capa
city of fixed pro
c
e
ssi
ng n
ode
s di
stribu
ted syste
m
s.
The
simulat
i
on
experim
ents
sho
w
that the algorit
hm ha
s good dyn
a
m
ic co
ntrol a
b
ilit
y. It can redu
ce processor
load and im
p
r
ove the dela
y
of task pro
c
e
ssi
ng a
s
neede
d and m
a
ke u
s
e of system re
sou
r
ce
s
more rationally. [7
,
8] propo
se a
task
sched
uling
load
bala
n
cing alg
o
rithm
ba
sed
on f
a
ir
indexe
s
. It d
edu
ce
s the
task a
ssi
gn
ment me
th
o
d
un
der mul
t
i-node
condi
tions, a
nd t
hen
improve
s
the
load b
a
lan
c
in
g algo
rithm b
a
se
d on fai
r
i
ndexe
s
with t
h
is meth
od. [
9
,
10] ind
u
ce
s a
gene
ral mod
e
l of load bal
anci
ng sche
d
u
ling on the
basi
s
of in-d
epth study of
load balan
ci
ng
sched
uling p
r
oble
m
in di
stribute
d
sy
stems a
nd
co
ndu
cts a det
ailed an
alysi
s
of the vari
ous
factors that affect load bal
a
n
cin
g
.
For the curre
n
t proble
m
s
of cloud com
puting
and di
stribute
d
re
source sch
edu
ling, this
pape
r p
r
e
s
en
ts an inte
grat
ed cl
oud
co
mputing
re
so
urces sche
du
ling strategy
con
s
id
erin
g b
o
th
resou
r
ces credibility
and use
r
satisfa
c
tion
ba
se
on
the Multi-ag
ent Gen
e
tic
Algorithm. T
he
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 656
3 – 6569
6564
strategy deal
s with
both user
sa
ti
sfacti
on and resource
s
credibilit
y in the objective function. It
integrate
s
re
source
credibil
i
ty into the fu
nction
of
u
s
e
r
sat
i
sf
act
i
o
n
a
nd t
a
ke
s u
s
e
r
sat
i
sf
act
i
o
n
a
s
optimizatio
n obje
c
tive. Then it will calculate solu
tio
n
s
with a gen
etic algo
rithm
.
By taking the
above m
e
a
s
ure
s
, the
strategy ca
n m
a
ximize th
e
i
n
tere
sts of
u
s
ers and co
mputing effici
ency.
After integrat
ing this
sche
duling
strate
gy into
Agen
ts we p
r
o
p
o
s
e a M
u
lti-a
gent sche
dul
ing
frame
w
ork, whi
c
h can complete reso
urce sche
dul
ing in clou
d
computin
g with Multi-a
g
ent's
excelle
nt autonomy, learni
ng ability and
soci
ality.
2. Resource
Credibilit
y
Re
sou
r
ce cre
d
ibility is an index that refl
ects
the
pro
c
essing
cap
a
ci
ty and reliabili
ty of the
resou
r
ces. It reflect
s
the p
r
ocessin
g
ca
pacity
and
re
liability of the res
ource
during a pe
riod
of
time (fro
m no
w to th
e future). With
the
model
of
re
source
credibil
i
ty, the resou
r
ce
s
whi
c
h
h
a
ve
highe
r cre
d
ibi
lity are mo
re
likely to be
called, while t
he resource
s whi
c
h h
a
ve l
o
we
r
cre
d
ibili
ty
are le
ss li
kel
y
to be called or even ha
ve no chan
ce
to be called.
So the load con
d
ition
s
of the
resources will
be more bal
anced. The credibility of
the resource i
s
dynamic
and will vary with its
perfo
rman
ce,
utilizatio
n,
online
rate
and
su
cce
s
s rate
of
co
mpleted j
o
b
s
. By introdu
cing
resource credibility into the
scheduling strategy,
the scheduling becomes
dynamic and will
cha
nge
wit
h
r
e
so
ur
ce
s.
Thi
s
i
s
v
e
ry
imp
o
rt
ant
f
o
r
dist
ribut
ed
sy
st
e
m
s
wh
ere
r
e
sou
r
c
e
s
cha
n
ge
freque
ntly.
Re
sou
r
ces
credibility is def
ined a
s
follows:
()
()
*
(
)
*
()
*
(
)
rt
p
t
i
t
h
t
v
t
(1)
in which
t
is
time,
()
rt
is t
h
e credibility of
the resources
at the m
o
ment
t
, which
con
s
i
s
ts of four p
a
rts:
()
p
t
,
()
it
,
()
ht
,
()
vt
.
()
p
t
is a pe
rfo
r
man
c
e fun
c
t
i
on of re
sou
r
ce
s. It
reflect
s
the
proce
s
sing
ca
p
a
city of the
re
sou
r
ces.
For
example, if th
e re
so
urce i
s
a CP
U, then
i
t
will be the CPU Clo
ck Sp
eed.
()
it
is the availability rate of resources. Same as
()
pt
, it
reflec
t
s
the processi
ng capability of the
resources. If the resource i
s
CP
U, then it
wil
l
be (1 -
CPU
utilization).
()
ht
is the online
rate of resour
ces, and it refl
ects the reli
ability of the resources. A
s
distrib
u
ted system
s
empl
o
y
res
o
urc
e
s
on the Internet, it is
diffi
cult to pre
d
ict
whe
t
her a
re
sou
r
ce
is on
-line
or o
ff-line at one
moment. But we
can o
b
tai
n
()
ht
by cal
c
ulat
ing the on
-lin
e rate of the
resou
r
ces. T
he highe
r th
e on-lin
e rat
e
is, the higher reliability the reso
urce has, the hi
gher
credibility it has.
()
vt
is the su
ccess rate of
the reso
urce
s
to compl
e
te
the job. Afte
r acceptin
g
jobs,
t
h
e r
e
s
our
ce
s
can
n
o
t
alw
a
y
s
co
mplet
e
t
h
e
m
su
c
c
e
ssf
ully
.
Like
wi
se,
t
he
su
cc
es
s
r
a
t
e
reflect
s
the reliability of the
resource.
3. User Satis
f
ac
tion
Most sch
edul
ing algo
rithm
studie
s
only
focu
s
on op
timizing ove
r
all system o
peratin
g
indicators, bu
t they ignore
the sati
sfacti
on from t
he
p
e
rspe
ctive of an individu
al use
r
. In fact, for
different
users, p
r
eferen
ce
s of
its
ope
rat
i
ng time
pe
rfo
r
man
c
e
an
d
eco
nomi
c
co
sts are diffe
re
nt.
Take
ani
mati
on rend
erin
g
system
as
an
example,
user A
ca
re
s a
b
out re
nde
ring
co
sts while
u
s
er
B is more co
nce
r
ne
d ab
o
u
t the ren
deri
ng time. Even for the sam
e
use
r
, the re
quire
ment
s of job
operating tim
e
and
e
c
on
o
m
ic
co
sts va
ry over time.
Whe
n
u
s
e
r
A
is fa
cing
an
urge
nt case, the
cu
stome
r
C want
s to see
the rend
erin
g
s
as
soon a
s
possi
ble an
d
does not care about mon
e
y.
At this p
o
int, Use
r
A
wo
ul
d p
r
efer fa
st
rend
eri
ng
performan
ce;
When
user A i
s
fa
cing
a
no
n-
emergen
cy case, and
the cu
stome
r
D
h
a
s
limite
d
fun
d
s
and
ne
eds A to
control t
he b
udg
et, a
nd
he/sh
e ha
s
n
o
stri
ct time
requi
rem
ent,
we
can
tell that A will
pa
y more
atten
t
ion to re
du
ci
ng
co
sts; if the custome
r
E asks u
s
e
r
A to balan
ce
the e
c
on
omic
co
sts and
spe
ed, obviou
s
ly A
will
cho
o
se a bal
anced solutio
n
.
From thi
s
poi
nt of view, in orde
r to meet
the need
s of
different u
s
ers, we n
eed to
desig
n
satisfa
c
tion f
unctio
n
s a
ccordin
g to their different
req
u
irem
ents. A
nd the obje
c
ti
ve function of
th
e
sched
uling
st
rategy
sho
u
ld
maximize th
e u
s
er's
overall satisfa
c
tio
n
. Wh
en th
e j
obs of u
s
e
r
i
ar
e
assign
ed to reso
urce
j
, the
us
er
s
a
tis
f
action
ij
S
is cal
c
ul
ated as follo
ws:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Reso
urce
Sched
uling S
t
rategy in
Clo
ud
Com
putin
g based on M
u
lti-age
nt… (Wu
x
ue
Jian
g)
6565
j
b
ij
it
ie
bj
r
c
sw
w
rc
(2)
1
it
ie
ww
(3)
it
w
and
ie
w
are the
perfo
rma
n
ce
prefe
r
e
n
ce a
nd
co
st p
r
efe
r
en
ce
re
sp
ect
i
vely and th
ei
r
rang
e i
s
[0
,
1
]
.
j
r
is
the credi
bility of resource
j
. The bi
gge
r
j
r
is, the
bigge
r user
sati
sfaction
is.
b
r
is th
e m
ean
re
so
urce
credi
bilit
y
in
the system
history.
j
c
is the
quotatio
n of
j
th
resou
r
ce for the use
r
i
, and the lower
j
c
is, the bigge
r use
r
sati
sfa
c
tion is.
b
c
is the mean
quotation in t
he syste
m
history.
Dep
endin
g
o
n
the u
s
e
r
types, you
can
set different
it
w
and
ie
w
. For
u
s
ers
wh
o care
more
abo
ut perform
an
ce,
you ca
n set
it
w
=0.8,
ie
w
=0.2 ; Fo
r users
who
care m
o
re
abo
ut co
sts,
you can
set; For u
s
e
r
s
wh
o want a bal
a
n
cin
g
system,
you can set
it
w
=
0
.5,
ie
w
=0.
2
.
Cre
d
ibility is an
ind
e
x that reflect
s
t
he
p
r
o
c
e
ssi
ng cap
a
city and relia
bility
of
th
e
r
e
sour
ces
.
4. Multi-agen
t Sy
stem
Multi-ag
ent techn
o
logy is
a
n
importa
nt b
r
an
ch of a
r
tificial intellig
en
ce te
chn
o
log
y
. Agent
is an age
ncy.
It is a software entity that
can
co
mpl
e
te
its function
s unde
r ce
rtain
circum
stan
ces
contin
uou
sly
and
sp
onta
neou
sly. It also
can
com
m
unicate
with the
releva
nt Agents a
nd
pro
c
e
s
ses. A
gent is auto
n
o
mou
s
. It can com
p
lete
mo
st of its fu
nction
s an
d control its inte
rnal
state without
intervention
o
f
peopl
e
o
r
other
Agent
s. Another
imp
o
r
t
ant attrib
ute
of Age
n
t i
s
i
t
s
social ability. It can take the initiative to intera
ct with
other Agents
to ac
hi
eve their goals. Agent
also h
a
s the
ability to learn. It can c
han
ge with the e
n
vironm
ent a
daptively.
Multi-ag
ent system is a system com
p
ose
d
of
multiple Agents. It can solve
compl
e
x
probl
em
s whi
c
h
can
not b
e
solve
d
by
single Ag
ent. Agents
com
p
ete, coo
r
di
nat
e and
coope
rate
with ea
ch oth
e
r to solve p
r
oblem
s toget
her.
Traditio
nal sche
duling
m
e
thod
s, su
ch as b
r
a
n
ch and b
oun
d method,
dynamic
prog
ram
m
ing
and heuri
s
tic graph
sea
r
ch algorithm, a
r
e mathem
atically perfe
ct, but they can not
adapt to
com
p
lex pro
d
u
c
tion environm
e
n
ts be
ca
use t
hey ha
s ma
d
e
a lot of
sim
p
lification
of the
real
wo
rld. At
the
sam
e
ti
me, la
ck of di
versit
y ha
s limited thei
r
scope
of ap
plication. Sched
uling
strategy ba
sed
on multi-agent
te
chn
o
logy can o
v
erco
me the
sho
r
tcoming
s
of the a
b
o
ve
algorith
m
s. T
he strategy relies o
n
the
swarm in
telli
gen
ce effect
of Agents an
d has ve
ry g
ood
simulatio
n
of
compl
e
x issu
es
and
proce
ssi
ng
po
wer.
At the same t
i
me it ha
s
po
werful
ad
apti
v
ity
and ca
n cha
nge
th
eir stat
us and
lea
r
n relevant
kn
o
w
led
ge
adapt
ively to achie
v
e the exp
e
cted
effect of sch
e
duling. As a result, it has very signifi
cant
robu
stne
ss.
The Multi
-
ag
ent sche
duli
ng st
rategy
pre
s
ente
d
in
this p
ape
r
contai
ns th
e
followin
g
Agents.
User A
gent:
Submit jo
b
s
to
Jo
b
Re
spo
n
se
Age
n
t and
a
c
ts as the
co
n
s
ume
r
of
resources. It
will occupy
resources when the job appl
ication i
s
successful
and
release resources
after the job is co
mpleted.
Job
Re
spon
se Agent: It is
the middle la
yer bet
we
en the system a
n
d
use
r
s. It is use
d
to
accept u
s
er-submitted job
appli
c
ation a
nd retu
rn re
sults to the user after the jo
b is co
mplete
d.
Job
De
com
p
osition a
nd
Distributio
n Ag
ent: A
cce
pt jobs from Job
Re
spon
se A
gent, and
decompo
se
s
jobs into the
smalle
st particles
a
nd di
stribute th
e decompo
se
d
sub
-
job
s
to Jo
b
Sched
uling A
gent.
Re
sou
r
ce Search Age
n
t
:
Search on
line
re
so
urces in the
n
e
twork a
n
d
regi
ster
r
e
sour
ces
.
Job S
c
he
duli
ng Agent: Job Sch
edulin
g Agent
will
accept sub-jobs that ha
s bee
n
decompo
se
d
into the sm
allest pa
rticle
s and
se
le
ct sched
uling al
gorithm, an
d assign the
sub-
jobs
to the Res
o
urc
e
Agent.
Re
sou
r
ce A
gent: Sen
d
online
me
ssage
s to
Re
source
Search Agent
s
an
d sele
ct
Re
sou
r
ce Se
arch Ag
ent t
o
dete
r
min
e
the regi
st
rati
on. Acce
pt j
obs from
the
Jo
b S
c
he
du
ling
Agent and
complete the
m
. Detect its own
statu
s
and
send it
s own
re
sou
r
ce
s credi
bility to
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 656
3 – 6569
6566
Re
sou
r
ce Search Age
n
t. Submit job transfe
r sig
nal to Job
Monitori
ng Agent wh
en th
e
sched
uling p
ours or jo
b fails to apply for the transfe
r of jobs.
Job
Monito
rin
g
and
Coordi
nation Ag
ent: Monito
r the
status of the
Re
sou
r
ce Ag
ent and
re-allo
cate jo
bs after recei
v
ing the job transfe
r
appli
c
ation sig
nal sent by the Re
sou
r
ce Agent
.
The overall framework is
shown in Figu
re 1:
Figure 1. Multi-age
nt Sche
duling Strate
gy Frame
w
o
r
k
5. Job-Reso
urce Sche
du
ling Strategy
Becau
s
e
of
space con
s
trai
nts, we de
scribe
job
re
so
u
r
ce
sch
eduli
n
g strategy h
e
r
e o
n
ly.
Job
re
sou
r
ce
sched
uling
st
rategy
will be
encap
sulate
d in the
Job
Sched
uling A
gent. For th
e
Job
Agent, it has resource
s
1
,
2
,
..
..
..
,
m
and credi
bility matrix
12
[
,
,
.
..
...
,
]
T
m
Rr
r
r
of those
re
sou
r
ce
s
and co
st
m
a
trix
12
[
,
,
...
...,
]
T
n
Cc
c
c
, and
acce
pted jo
bs
1
,
2
,
..
..
.
.
,
n
a
nd jo
bs'
satisf
action
matrix
12
12
...
...
tt
n
t
ee
n
e
ww
w
S
ww
w
. Its mission
is to assi
gn
the accepte
d
jobs
1
,
2
,
..
...
.,
n
to th
e res
o
urces
1
,
2
,
..
..
..,
m
res
p
ec
tively.
Since th
e
scheduli
ng p
r
o
b
lem i
s
NP
probl
em, it i
s
often
very
difficult for tradition
al
algorith
m
s t
o
obtain glo
bal optimal
solutio
n
s. G
enetic Alg
o
ri
thm stand
s
out from m
o
st
sched
uling
al
gorithm
s fo
r i
t
s excell
ent g
l
obal o
p
ti
mization capa
bili
ty. The main
module
s
of the
algorith
m
a
r
e
:
ch
romo
so
m
e
en
co
ding,
p
opulatio
n
initi
a
lizatio
n, ge
n
e
tic m
anipul
a
t
ion an
d fitne
s
s
asse
ssm
ent. The flow of th
e algorith
m
is sho
w
n in Fig
u
re 2.
Figure 2. Genetic Algo
rith
m Flow Chart
base
d
on Re
sou
r
ce Sch
e
d
u
ling Strategy
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Reso
urce
Sched
uling S
t
rategy in
Clo
ud
Com
putin
g based on M
u
lti-age
nt… (Wu
x
ue
Jian
g)
6567
5.1. Chromo
some Encod
i
ng
The
chromo
some in thi
s
pape
r is
a n
-
dime
nsi
onal
vector
12
n
Dd
d
d
(all
element
s are
integers).
(1
~
)
i
di
n
is an intege
r, it
represents t
hat the
i
th job is assig
ned
to
the re
so
urce
i
d
.
i
d
ran
ges in
1~
m
. We
ca
n
see
that the ve
ctor
D
compl
e
tely rep
r
e
s
ent
the
allocation of reso
urce
s for the cu
rrent job
s
1
,
2
,
..
..
.
.
,
n
.
5.2. Population Initializati
o
n
After setting
the si
ze
of th
e po
pulatio
n:
popN
um
,
the algo
rit
h
m will
initiali
ze ra
ndomly
and ge
nerate
po
pNum
ch
romo
so
mes. And ele
m
ents in e
a
ch ch
romo
som
e
are al
so g
e
nerate
d
rand
omly, the numbe
r of elements
ran
g
e
s
in
1~
m
.
5.3. Fitness
Function
After the initi
a
lizatio
n is co
mplete, we n
eed to
asse
ss the fitne
s
s
of ch
rom
o
so
me in th
e
popul
ation. Rational fitness functio
n
is t
he key
of the
genetic al
go
rithm be
cau
s
e the fitness will
guide the
pop
ulation evoluti
onary di
re
ctio
n. Improp
er fi
tness fun
c
tio
n
may lead to
large
deviati
on
betwe
en po
p
u
lation conve
r
ged
solutio
n
s
and the o
p
timal solutio
n
s.
As de
scribe
d
in se
ction 2,
when th
e th job is a
ssi
g
ned to the re
sou
r
ce
i
d
, the
use
r
sat
i
sf
a
c
t
i
on
,
i
id
s
is cal
c
ul
ated a
s
follows:
,
i
i
i
d
b
i
d
it
ie
bd
r
c
sw
w
rc
(4)
The
same
re
sou
r
ce may
be a
ssi
gne
d
to many job
s
, as a
re
sult, multiple job
s
may
actually share
the credi
bility
of
sam
e
resource.
So, we need
to ad
just
user satisfaction function
according
to
the num
ber o
f
jobs
assig
n
ed to
e
a
ch reso
urce. If we set the n
u
m
ber of job
s
i
d
h
assign
ed to
reso
urce
i
d
to the same val
ue, then
we
can
obtain th
e modified
u
s
er satisfa
c
tion
function:
,
1
i
i
ii
d
b
i
d
it
ie
db
d
r
c
sw
w
hr
c
(
5)
The fitness e
v
aluation fun
c
tion is a
s
foll
ows:
,
1
i
n
id
i
f
itn
e
s
s
s
(
6)
5.4. Genetic
Manipulatio
n
The main op
eration
s
of the genetic al
g
o
rithm
are: selectio
n, hybridizatio
n and
mutation.
Selection i
s
to sele
ct indi
vidual ch
rom
o
som
e
from
the chromo
some pop
ulati
on to rep
r
od
uce
according
to
ce
rtain
rul
e
s a
nd
prefe
r
ences.
Th
e
gene
ral rule
is
to ch
oo
se
those
in
dividual
chromo
som
e
s who h
a
ve
better fitne
s
s. We ad
opt t
he roulette-b
ase
d
sele
ctio
n me
chani
sm
, in
whi
c
h the
value, indivi
dual
ch
rom
o
som
e
fitne
ss/total fitne
ss,
act
s
a
s
the individ
ual
chromosom
e
's probability of being selected.
A
f
t
e
r sele
ct
in
g t
w
o or mo
r
e
chr
o
mo
som
e
s,
Hy
br
i
d
ization is to sele
ct a rand
om bit-point
in
ea
ch ch
ro
moso
me and divide
the ch
romosome
s
in
to two
se
ction
s
at th
at point
. Subse
que
ntly,
excha
nge th
e front o
r
rear
se
ction
s
of two ch
romosome
s t
o
pro
d
u
c
e t
w
o n
e
w offspring
chr
o
mo
som
e
s.
The mutation
operation ta
ke
s mutation
rate a
s
the probability and
will reg
ene
rat
e
som
e
bit-co
de in o
ne of the ch
romosome
s randomly. Pr
o
per mutatio
n
rate can en
h
ance the glo
bal
optimization
ability of the
algorithm,
however, if the
mutation
rate is
too big, it
will
cause that the
algorith
m
is d
i
fficult to converge.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 656
3 – 6569
6568
6. Numerical
Example
Simulation e
x
perime
n
ts a
r
e carried o
u
t
in the Clou
dSim enviro
n
m
ent ba
sed
on the
JACK ag
ent langu
age. Cl
o
udSim is a si
mulation
software of clo
u
d
comp
uting which a
nno
un
ced
by Grid
bu
s p
r
oje
c
t in Ap
ril
2009.
Clou
d
S
im softw
a
r
e
frame
w
o
r
k
consi
s
ts
of SimJava, G
r
id
Sim,
Clou
dSim, UserCo
de. Clo
udSim is an
extensib
l
e
si
mulation tool
kit that enabl
es mod
e
ling
and
simulatio
n
of clou
d com
put
ing system
s a
nd appli
c
atio
n provisi
onin
g
environ
men
t
s.
In this
pap
e
r
, the
sched
uling
algo
rith
m is implem
ented i
n
Clo
udSim laye
r of the
simulatio
n
pla
tform, and th
e JACK ag
en
t simulation p
r
og
ram i
s
imp
l
emented in
UserCod
e
layer.
The su
bmitte
d job types a
nd qua
ntities
are sho
w
n in
Table 1.
Table 1. Jo
b Type and Qu
antity
Job t
y
pe
Quantit
y
Basic
running
time / s
Giving
priorit
y
to
perform
ance
Giving
priorit
y
to
costs
Balanced T
y
pe
eas
y 3000
2
15%
70%
15%
medium 2000
5
33%
33%
34%
difficult
1000
10
70%
15%
15%
Seen fro
m
Table
1, job
types a
r
e
divided into
three: e
a
sy,
medium
an
d difficult
according
to
the ba
si
c run
n
ing time
of
jobs (the
ba
sic time i
s
th
e
job
run
n
ing
time at 1G
HZ
CPU). Th
e
compo
s
ition
of users i
s
also
different
in
e
a
ch
type of j
o
bs. F
o
r
exam
ple, in th
e e
a
s
y
job type, 70%
of the job
s
v
a
lue
co
sts,
while 15%
thi
n
k mo
re
of performan
ce
and
15% thin
k m
o
re
of balanced
type. This is becau
se th
e runni
ng ti
me of easy job is sh
ort, the use
r
is m
o
re
intere
sted in
their e
c
on
o
m
ic requi
rem
ents. In
the
medium jo
b
type,
perform
ance, co
sts
and
balan
ce
d type accou
n
t for approxim
ate 1/3 each. Fin
a
lly, for the difficult job type, beca
u
se of its
longe
r run
n
in
g time, users are more co
nce
r
ne
d
abo
ut the performance. So 70% of the jo
bs
value perfo
rm
ance, while 1
5
% think more of cost
s an
d 15% think
more of bal
an
ced type.
The above jo
bs will n
o
t be
relea
s
ed to the Mu
lti-a
g
e
n
t system at the same tim
e
s. The
system
will re
lease them to the Multi-
age
nt rando
mly during a p
e
ri
od
of 1 hour.
Set the n
u
m
ber
of
Re
so
urce Ag
ent i
n
the
Mu
lti-a
gent
system
to 10.
The
co
mpo
s
ition
of
Re
sou
r
ce Ag
ents is in T
a
b
l
e 2.
Table 2. Age
n
ts of The Re
sou
r
ces
Number of
agent
s
Freque
nc
y
/ Hz
Running time
agent 1~3
500M
basic running time * 2
agent 4~7
1G
basic running time
agent 8~10
2G
basic running time / 2
Table 3. Average Respon
se Time and
User Sati
sfacti
on
T
y
pe of
algorithm
Average
response
time/s
Overall user
satisfaction
Multi-agent
5.42
1.68
Minimum load
scheduling
5.74 1.59
Random
scheduling
5.85 1.53
Among th
e th
ree type
s of j
ob
sho
w
n
in
Table
1, m
e
dium type
jo
bs
are
m
o
st
obje
c
tive
and re
presen
tative as they do not con
s
i
der u
s
e
r
s’
p
r
eferen
ce
s. T
o
comp
are this algo
rithm
with
rand
om
sche
duling
algo
rit
h
m an
d mi
ni
mum lo
ad
scheduli
ng
algo
rithm, me
diu
m
type jo
bs
are
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Reso
urce
Sched
uling S
t
rategy in
Clo
ud
Com
putin
g based on M
u
lti-age
nt… (Wu
x
ue
Jian
g)
6569
taken
as in
st
ance. The
calcul
ated ave
r
age
re
spo
n
se time and u
s
er
sati
sfacti
on is
sho
w
n
in
Table 3.
From
the a
b
o
ve re
sult
s,
the Multi-age
nt's
avera
g
e
re
spo
n
se ti
me an
d u
s
e
r
overall
satisfa
c
tion
are sh
orte
r than
that of
Ra
ndom
sche
duling
al
gorithm
an
d
Minimum
l
oad
sched
uling al
gorithm, e
s
p
e
cially on the
users’
satisf
action, Multi-agent ge
neti
c
algo
rithm than
Ran
dom
sch
edulin
g alg
o
ri
thm and
Mini
mal loa
d
sch
edulin
g alg
o
ri
thm ha
s g
r
e
a
t
ly improved,
and
the in
crement
between
Mul
t
i-agent
gen
e
t
ic alg
o
rithm
and Ra
ndom
sched
uling
al
gorithm
is
2.5
0
times of
that betwe
en Mini
mum
lo
ad scheduli
ng
algo
rithm a
nd
Ra
ndom
sch
edu
ling al
gorith
m
.
Obviou
sly, the use
r
overall
satisfa
c
tion
of the Mult
i-a
gent gen
etic
algorith
m
is superi
o
r to tha
t
o
f
the other two
algorith
m
s.
7. Conclusio
n
An integ
r
ate
d
a
s
sessme
nt model
co
nsid
erin
g b
o
t
h re
sou
r
ce
credibility
and
user
satisfa
c
tion i
s
establi
s
h
ed i
n
this
pap
er,
and a
re
so
urce sched
ulin
g
strategy
ba
sed on gen
etic
algorith
m
is desi
gne
d on the basi
s
of this m
odel. T
hen a clo
ud
comp
uting sy
stem re
so
urce
sched
uling a
r
chite
c
ture i
s
prop
os
ed ba
sed on Multi-a
gent frame
w
o
r
k.
The nu
meri
cal re
sults
sh
ow that thi
s
syst
em
enha
nce
s
u
s
e
r
sa
tisfaction
and
redu
ce
s
the ave
r
age
j
ob a
c
ce
ss time. The
op
erating effici
en
cy of th
e
clo
ud
com
puting
syste
m
i
s
fin
a
lly
improve
d
.
Ackn
o
w
l
e
dg
ement
This wo
rk
wa
s sup
porte
d by
the
Nation
al
Natural
Science Fo
und
ation of
Chin
a (G
ra
nt
No. 611
720
18), an
d by
the Nation
al Natu
ral S
c
ien
c
e F
oun
dation of China (Gra
nt
No.
6117
2124
), and by the Natural Sci
ence Fou
n
d
a
tion of Gu
angd
ong Province
(G
rant
No.
S20120
100
0
9865
) and G
uang
dong Province Sci
e
n
c
e an
d techn
o
logy plan p
r
oject (G
ra
nt No.
2012B0
101
0
0040
).
Referen
ces
[1]
Z
hang Y
i
n
g
, Gao Mi
n, Z
hou
Chi. Stud
y
on
Clou
d
R
e
sourc
e
Sche
dul
in
g
Probl
em Base
d on
Lo
gistic
s
Park Distributi
o
n S
y
stem.
Jour
nal of Co
mp
ut
ation
a
l Intell
ig
e
n
ce an
d Electri
c
Systems
. 20
12; 1(1):12
2
-
125.
[2]
De
w
a
k
i
P, Val
a
rmathi ML. J
ob Sch
edu
li
ng
using Ge
netic
Algorithm
w
i
t
h
Qos Satisfa
c
tion in Gri
d
.
Europ
e
a
n
Jour
nal of Scie
ntific
Researc
h
. 201
2; 74(2): 27
2-2
85.
[3]
Son D
u
y Da
o,
Kazem Ab
har
y, Romeo M
a
ria
n
. Optimisatio
n
of Reso
urce S
c
hed
uli
ng
in V
C
IM S
y
stem
s
Using
Gen
e
tic
Algorit
hm.
Inte
rnatio
nal
Jo
urn
a
l
of Adv
ance
d
R
e
searc
h
i
n
Artificial I
n
tell
ig
ence
. 20
12
;
1(8): 49-5
6
.
[4]
Jian
hua Gu, Ji
nhu
a Hu, T
i
an
hai Z
h
ao, Guof
ei Su
n. A Ne
w
Reso
urce Sch
edu
lin
g Strateg
y
Bas
ed
o
n
Genetic Alg
o
rit
h
m in Cl
oud C
o
mputi
ng Envir
onme
n
t.
Journ
a
l of Co
mp
uter
s.
2012; 7(1): 4
2
-52.
[5]
Pa
yal
Sin
g
h
a
l
,
Ravin
der
Si
ngh, P
i
nk
y
R
o
sema
rr
y. A
n
Improve
d
C
onstrai
nt Bas
ed R
e
so
urce
Sched
uli
ng A
p
proac
h Usi
ng
Job Grou
pi
ng Strateg
y
Grid Comp
uting.
Int
e
rnati
ona
l Jo
u
r
nal
of Grid
Co
mp
uting & A
pplic
atio
n
. 201
2; 3(4): 33-42.
[6]
Bu
yya
R, Ranj
an R. Speci
a
l Se
ction: F
eder
ated Res
ource
Manag
em
ent i
n
Grid and Cl
o
ud Com
puti
n
g
S
y
stems.
F
u
tur
e
Generati
on C
o
mputer Syste
m
s
. 20
10; 26(
8
)
: 1189-1
1
9
1
.
[7]
M Mezmaz, N Mela
b, Y Kessaci, et al. A pa
ralle
l
bi-
obj
ecti
ve h
y
bri
d
meta
heur
istic for en
erg
y
-a
w
a
r
e
sched
uli
ng for
clou
d computi
n
g s
y
stems.
Jou
r
nal of Para
lle
l Distribut
e Co
mputin
g
. 201
2; 71(1): 149
7-
150
8.
[8]
Sotoma
yor B,
Montero R
S
, Ll
orente IM, et al
. Virt
ual infrastr
ucture ma
nag
e
m
ent in pr
ivate
and h
y
b
r
i
d
clou
ds.
IEEE Internet Computing
. 200
9; 12(
5): 14-22.
[9]
Veen
a Gos
w
a
m
i, Sudh
ans
u
Shekh
a
r Patra
,
GB
Mund. Performanc
e An
al
ys
is of
Clo
u
d
Com
putin
g
Centers f
o
r Bu
lk Servic
es.
Internati
ona
l Jo
ur
nal
of Cl
ou
d A
pplic
atio
n a
nd
Co
mp
uting
.
20
12; 2(
4): 13
-
26.
[10]
RF
Sun, Z
W
Z
hao. R
e
sou
r
ce Sche
du
lin
g Strateg
y
B
a
sed
on
Clo
u
d
Com
putin
g.
Aeron
autic
a
l
Co
mp
uting T
e
c
hni
que.
2
010;
40(3): 10
3–
105
.
Evaluation Warning : The document was created with Spire.PDF for Python.