TELKOM
NIKA
, Vol.14, No
.4, Dece
mbe
r
2016, pp. 14
54~146
1
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v14i4.3539
1454
Re
cei
v
ed Ma
rch 5, 2
016;
Re
vised July
29, 2016; Accepted Augu
st
17, 2016
Resear
ch on Batch Scheduling in Cloud Computing
Jintao Jiao
*
1
, Wensen Yu
2
, Lei Guo
3
1
School of Mathematics a
nd
Comp
uter Scie
nce,
W
u
y
i
U
n
iv
ersit
y
, 3
543
00,
W
u
y
i
sh
an, Ch
i
na
2
T
he Ke
y
L
a
b
o
r
ator
y
of Cogn
i
t
ive Comp
uting
and Inte
ll
ig
ent Information Pr
ocessi
ng of F
u
j
i
an Ed
ucati
on
Institutions, 35
430
0, W
u
y
i
s
h
a
n
, Chin
a
3
Colla
bor
ative Innov
atio
n Cent
er
of Chin
ese
Oolon
g
T
ea Industr
y
,
3
543
00
, W
u
yisha
n
, Ch
ina
*Corres
p
o
ndi
n
g
author, e
_
ma
il: jiao
j
i
n
tao@
1
63.com
A
b
st
r
a
ct
In the existi
ng
clou
d co
mp
uti
ng e
n
viro
n
m
e
n
t, bat
ch sched
ulin
g strateg
i
e
s
ma
inly foc
u
s
on the
ma
na
ge
me
nt o
f
resources
al
l
o
catio
n
. T
h
is p
aper
provi
des t
he task sc
he
d
u
lin
g a
l
g
o
rith
m base
d
o
n
serv
ice
qua
lity w
h
ich
fully cons
id
ers
priority a
nd
sched
uli
ng d
e
adli
ne. T
he i
m
pr
ove
d
al
gor
ithm c
o
mbi
nes
the
adva
n
tag
e
s of
Min-
min
al
gorit
hm w
i
th hi
gh
er
throu
ghp
ut a
n
d
li
ne
ar pr
ogra
m
mi
ng w
i
th
glo
bal
opti
m
i
z
a
t
io
n,
consi
ders n
o
t only al
l the task
s but also th
e h
i
gh pr
iori
ty task
s. T
he experi
m
ent re
sult shows that compared
with the Min-min and DBCT the
com
p
leted tasks of the im
pr
oved
algorithm
increase about 10.6%
and
22.0%, on the other hand the
complet
ed high priority tasks also incr
eas
es approx
imatel
y 20% and 40%.
Ke
y
w
ords
: clo
ud co
mp
utin
g, sched
uli
ng stra
tegy, batch sch
edu
lin
g
Copy
right
©
2016 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Clou
d
comp
u
t
ing is a
ne
w
busi
n
e
s
s mo
del, an
d it i
s
a p
r
odu
ct
of i
n
formatio
n te
chn
o
logy
[1]. In cloud
comp
uting
market, use
r
scal
e
is
ve
ry large, and
the QoS(Q
u
ality of Service)
requ
est
s
of each ta
sk a
r
e
not identical. These
facto
r
s make task
sched
ule in cloud computi
ng
very compli
cated. So it i
s
a technical diffi
cult
y to
sche
dule th
e m
a
ssive ta
sks in
clou
d
comp
uting
in orde
r to ge
t a better sch
edulin
g re
sult
s with the con
s
train of bu
dg
et and dea
dli
ne.
The main b
a
tch sch
edulin
g algrithm
s are Min-min [2]
and DBCT [3] (unde
r the
deadli
n
e
and b
udg
et constrain time
optimizat
io
n
algorith
m
). T
he ide
a
of Mi
n-min
algo
rith
m is th
at the
best
resou
r
ces
wil
l
be allo
cate
d to the min
i
mum task, thus the
syst
em throu
ghp
ut rate can
be
improve
d
a
n
d
the mi
nimu
m time of th
e job
can b
e
obtain
ed.
While the
idea
of DB
CT i
s
to
sched
ule th
e
task on
the
reso
urce th
at
can
comple
te the tas
k
earlies
t, als
o
reques
t
the cos
t
no
more th
an u
s
er'
s
bu
dget a
nd the fini
sh t
i
me no m
o
re
than de
adlin
e
.
With the po
pularity of
clo
ud
comp
uting,
some u
s
e
r
s wi
th less
budg
e
t
may also
want to u
s
e
clo
ud
comp
uting
;
also
there a
r
e
some
tasks
with sufficien
t budget a
n
d
intensive
de
mand
of dea
dline. Fo
r th
ese
ca
se
s, the
numbe
r of
co
mpleted ta
sks of Mi
n-mi
n
and
DBCT al
gorithm
s
will
be le
ss. The
r
e are al
so
a l
o
t of
peopl
e put fo
rwa
r
d
othe
r a
l
gorithm
s, but
som
e
did
no
t con
s
ide
r
th
e dea
dline [4
], some di
d n
o
t
con
s
id
er th
e
use
r
’s bu
dget
[5]; they did
not bal
an
ce t
he
relation
shi
p
bet
wee
n
th
e de
adlin
e a
n
d
the budg
et.
From the pe
rspe
ctive of QoS, this
paper propo
se
s DBCMNCT
(dea
dline an
d budget
con
s
trai
ned
maximize
d n
u
mbe
r
of
co
mpleted
tasks
sched
uling
algorith
m
) al
gorithm
in
order to
meet the b
u
dget an
d de
adline
co
nstrains.
Co
mp
a
r
ed
with Min
-
min a
nd
DBCT alg
o
rith
m,
DBCM
NCT
a
l
gorithm ca
n accompli
sh h
i
gher
pri
o
ri
ty tasks
better,
at the same
time the ne
w
algorith
m
can
maximize th
e numb
e
r of
compl
e
ted ta
sks. So the u
s
er’
s
d
e
ma
nd
s can b
e
me
et;
on the other
hand, the cl
o
ud com
puting
can a
ttra
c
t more con
s
um
ers and g
e
t more benefits.
2. Cloud Co
mputing Sch
e
duling Mar
ket Mod
e
l
Clou
d comp
u
t
ing sche
duli
ng ma
rket m
odel[6-7]co
u
l
d
man
age
an
d evaluate
re
sou
r
ces
allocation m
o
re effe
ctively. The
mod
e
l
bring
s
fo
ur b
enefits: (1) u
s
er’
s
fai
r
use
of computin
g
resou
r
ces [8], (2
) adj
ustin
g
the b
a
lan
c
e
of su
pply an
d
dema
nd
of reso
urce
s in
cloud
com
puti
ng.
Whe
n
dema
n
d
excee
d
s
su
pply, raisin
g the pri
c
e of
re
sou
r
ces
coul
d help to red
u
ce the n
u
m
ber
of users an
d tasks, on the
other
ha
nd, whe
n
su
pply excee
d
s d
e
m
and, lower pri
c
e could hel
p
to
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Re
sea
r
ch on
Batch Sch
e
d
u
ling in Clo
u
d
Com
puting (Jintao
Jiao
)
1455
attract more use
r
s [9], (3) providing qu
ality of se
rvice to users, such a
s
task deadli
ne, co
st to
compl
e
t
e
t
h
e
t
a
sk,
se
cu
rit
y
of
resou
r
ce
s [10], (4
) p
r
oviding the
effective re
so
urce m
ana
gem
ent
and allo
catio
n
mech
ani
sm
[11].
The mod
e
l contain
s
clie
nt, broker, re
sour
ce
s, re
so
urces
su
ppo
rter and info
rmation
servi
c
e, the a
r
chite
c
tu
re sh
ows in Figu
re
1.
Figure. 1 Clo
ud Com
putin
g Sched
uling
Model
Usually the t
a
sks that ne
e
ded to
be
run
can
be
divid
ed into
se
rial
appli
c
ation
s
,
parallel
appli
c
ation
s
, para
m
eter
scannin
g
appli
c
ations, colla
b
o
rative ap
plications a
nd et
c. The sy
ste
m
allows users to set the re
sou
r
ce req
u
irements
a
nd
prefe
r
en
ce
s for pa
ramete
rs, and the cli
ent
pay for the using re
sou
r
ce
s.
Broker i
s
the
interme
d
iate i
n
terface bet
ween
c
lient an
d re
sou
r
ce, functio
n
of wh
ich i
s
to
discover re
so
urces, sele
ct
resou
r
ces,
receiv
e ta
sks,
return sch
e
duling
re
sults and ex
chan
ge
informatio
n. And
b
r
o
k
er sup
port
s
diff
erent
sche
du
ling poli
c
ie
s
whi
c
h
can
find resource
and
sched
ule tasks a
c
co
rdin
g to use
r
’s d
e
m
and
s. Bro
k
er is m
a
inly compo
s
ed
of job co
ntrol a
g
ent,
sched
ule advi
s
or, explo
r
e
r
, trade ma
nag
er and d
eploy
ment agent.
There is a variety of compu
t
ing resou
r
ce
s in
clo
ud co
mputing, re
so
urce ha
s the attribute
of comp
uting
performan
ce and comp
uting pri
c
e,
perfo
rman
ce
is mea
s
ure
d
by MIPS(Million
Instru
ction
s
Per Second
), price is me
asu
r
ed
by
unit time c
o
s
t
G$. Generally the better the
resource perf
o
rmance, the higher the pri
c
e will be.
Information service mainly
re
cords avai
lable
resources. Broker
will query information
servi
c
e when
search
es for appro
p
riate
resou
r
ces,
th
en intera
ct with reso
urce
s providers after
getting information of re
sou
r
ces th
at meet
use
r
’s demand. If resou
r
ces p
r
oviders have
ne
w
resou
r
ce to lease, they must regi
ste
r
the re
so
u
r
ce
s information i
n
informatio
n
service so t
hat
bro
k
e
r
co
uld find it.
In
th
e
pr
oc
es
s o
f
tr
an
sac
t
io
n
be
tw
een
c
lien
t
a
nd r
e
s
o
ur
ce
s
p
r
o
v
id
ers
,
r
e
s
o
ur
ce
s
provide
r
s reg
i
ster the reso
urces info
rm
ation in
the informatio
n service. After
client submit
s a
task to broker, broker will get av
ailable resources information from
information serv
ice, and then
sched
ule
s
th
e task to the
approp
riate reso
urce
s
a
ccordin
g to the
sched
uling
al
gorithm. Broker
also e
s
timate
s the com
p
let
i
on time and the co
st
before the task i
s
execute
d
. If the time exce
ed
s
the deadline
or the
co
st is
higher than
user’
s
budget,
broker
will
ref
u
se to
accept
the task. If the
task is
executed successfully, br
oker
will return results to the
client and
obtain the
profit,
otherwise ret
u
rn the e
rro
r i
n
formatio
n.
3. Descrip
tio
n
of Schedul
ing Tasks
In the clo
u
d
comp
uting
environ
ment,
reso
u
r
ces
provide
r
can
benefit by
leasi
ng
comp
uting re
sou
r
ces, me
anwhile clie
n
t
can co
m
p
l
e
te the task of quality requireme
nts by
purcha
s
in
g the right to use
reso
urce
s. We a
s
su
me that the task
budg
et and d
eadlin
e is limited,
client
re
quire
s total
cost
within bu
dget
a
nd h
ope
a
s
much
a
s
po
ssible
task co
uld b
e
com
p
l
e
ted
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1454 – 146
1
1456
unde
r the premise of ta
sk completio
n
time befor
e de
adline. If clie
nt has suffici
ent budg
et and
long en
ough
deadli
ne, the tasks would b
e
compl
e
ted
within bu
dget
and dea
dline
.
Due to the
b
udget i
s
less or the ta
sk
is
urgent, m
a
ybe some o
f
the tasks
could be
compl
e
ted. So, this pape
r mainly discu
s
se
s ho
w
to maximize the
numbe
r of tasks
compl
e
ted.
Also, this pa
per
co
nsi
ders the ta
sk pri
o
rity
unde
r
sp
ecial
ci
rcum
stance
s
, b
e
ca
use
cli
ent ho
pes
that urge
nt and impo
rtant
tasks
s
houl
d
be co
mplete
d
in time, whil
e the re
st of the tasks ca
n
be
delayed.
With the abo
ve assumptio
n
s, clie
nt sep
a
rate
s
the tas
k
s
into two
k
i
nds
by priority: the
highe
r p
r
io
rity tasks
and
the lo
we
r p
r
io
rity tasks.
T
h
is p
ape
r u
s
e
s
HS to
rep
r
ese
n
t the
hig
her
prio
rity tasks
and
LS to
re
p
r
esent th
e lo
wer p
r
io
ri
ty tasks. Fi
gure
2
sho
w
s the
de
adline
of the
HS
and LS ta
sks. We can figu
re out th
at t
he HS ta
sks
must b
e
com
p
leted b
e
fore
T=
HS
T
, and th
e
LS tasks mu
st be com
p
let
ed before T=
LS
T
, so the tasks are formali
z
ed as (HS,
HS
T
) a
nd (LS,
LS
T
).
I
f
t
a
sks t
y
pe is (0,
HS
T
) an
d (LS,
LS
T
), it means that there
are
no
HS tasks, as many tasks
as po
ssible
should b
e
com
p
leted withi
n
budg
et and b
e
fore
LS
T
.
Figure. 2 Dea
d
line of HS a
nd LS tasks
4. Task Sche
duling Algor
ithm Base
d on Qualit
y
of Ser
v
ice
In clo
u
d
com
puting
marke
t, resource
s
sup
porte
r
ca
n’t co
mplete
all tasks
wh
e
n
cli
ent
requi
re
s l
e
ss
budg
et or sh
orter time th
a
n
ne
ede
d. So
this
pap
er
propo
se
s the
task
sched
uli
ng
algorith
m
ba
sed on qu
ality of se
rvi
c
e to solve the pro
b
lem.
4.1. Deadline
and Budg
et
Cons
train
e
d Maxi
mum Number of
Complet
e
d Tasks(D
B
C
M
NCT)
Suppo
se tha
t
there are
m machi
n
e
s
m
r
r
r
,
,
,
2
1
, n t
a
sk
s
n
e
e
e
,
,
,
2
1
and th
e
followin
g
vari
able
s
in the cl
oud computin
g environ
men
t
.
Table. 1 Defi
nition of Varia
b
les
variable definition
n
number
of tasks
m
number of comp
uting resources
i
L
length of task i
B task
budget
j
Re
computing perfor
m
ance of resour
ce j
j
P
price of resource
j
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Re
sea
r
ch on
Batch Sch
e
d
u
ling in Clo
u
d
Com
puting (Jintao
Jiao
)
1457
Acco
rdi
ng to the client’
s
re
quire
ment
s,
HS tasks sho
u
ld be co
mpl
e
ted before
HS
T
, LS
tasks shoul
d be com
p
leted
before
LS
T
. In order to en
su
re
that the high priority tasks HS can b
e
compl
e
ted a
c
cording to
cli
ents’ d
e
man
d
good
re
sou
r
ce
s shoul
d b
e
allo
cated to
sho
r
t tasks, so
that more tasks can be compl
e
ted in
a period
of
time. Based on the two kinds of tasks,
DBCM
NCT gi
ves hig
h
e
r
p
r
i
o
rity to the al
locatio
n
of
re
sou
r
ces to th
e HS ta
sks throu
gh Mi
n-min
algorith
m
. Becau
s
e e
a
ch
machi
ne p
e
rf
orma
nce is d
i
fferent and t
he allo
cation
of the task
size
and num
be
r are not the
same,
so th
e com
p
letion
time and cost of each machi
ne will
be
different. If
j
t
and
j
b
are used
to represent
the com
p
leti
on time and
co
st whe
n
m
a
chi
ne j is
used to com
p
lete one
of the HS tasks and b repr
esent the
serv
ice cost of al
l the HS tasks,
m
j
j
b
b
1
.
Ho
wever, if th
e bu
dget i
s
li
mited o
r
the
HS ta
sks d
e
a
d
line i
s
urg
e
n
t, part of th
e
HS ta
sks
may not be
completed, th
en the u
n
co
mpleted
HS
tasks
sh
ould
be ad
ded to
the LS tasks
set.
Whe
n
exe
c
ut
ing the
LS t
a
sks,
a lin
ea
r pl
anni
n
g
m
odel
LPM i
s
used to
get
the m
appi
n
g
relation
shi
p
betwe
en LS
tasks a
nd re
sou
r
ces.
T
h
e
definition of
task
structu
r
e contain
s
an
attribute of so
rt, when
sort
=1 the task typ
e
is LS, when
sort
=2 the ta
sk type is
HS.
There are
so
me assum
p
tions
in the
scheduli
ng process:
Tasks
in
exe
c
ution can
not be sei
z
ed. O
n
ce
t
he se
rvi
c
e provider
b
egi
n to
exe
c
u
t
e a ta
sk,
it cannot interrupt the curre
n
t task to pe
rf
orm othe
rs u
n
til the task is complete
d.
There is no p
r
iority for the same type ta
sks,
the clien
t
only conce
r
ns abo
ut the numbe
r
of complete
d tasks befo
r
e
deadli
ne.
So the
algo
ri
thm of d
eadli
ne a
nd
bud
g
e
t co
nst
r
ain
e
d
maximu
m
numbe
r
of co
mpleted
tasks (DBCM
NCT
) is
sho
w
n in algorith
m
1.
Algorithm 1: DBCM
NCT
Step 1:
S
t
he set
of
all t
a
sks;
Step 2:
HS
S
t
he set
of
t
a
sks t
h
at
sort
=1;
LS
S
the s
e
t of tas
k
s
that s
o
rt
=
2
;
Step 3: initialize
b
B
b
t
T
T
j
j
LS
HS
,
,
,
,
,
;
Step 4:
)
(
HS
S
if
goto step 6;
Step 5: estimate the num
ber of com
p
let
ed
HS tasks by Min
-
m
i
n algorithm
, return b,
j
t
;
Step 6:
}
\
{
HS
i
i
i
LS
LS
S
e
d
uncomplete
e
e
S
S
;
Step 7: call L
P
M(
b
B
t
T
S
i
LS
LS
,
,
,
,
) m
odule, estim
a
te the
num
ber of
com
p
leted LS tasks
and retu
rn th
e m
apping rel
a
tionship bet
wee
n
tasks a
nd re
sou
r
ce
s;
Step 8: acco
rding to the
m
apping rel
a
tions
hi
p, sch
edule ta
sks
t
o
the co
rrespondi
ng
resou
r
ces to
perfo
rm
ance;
Step 9: accep
t
the sche
dul
e result returned from
the resou
r
ces p
r
o
v
ide
r
;
Step 10: end;
4.2. LPM Module
The matrix
nm
A
is defined in formula 1,
nm
n
n
m
m
nm
a
a
a
a
a
a
a
a
a
A
,
,
,
,
,
,
,
,
,
2
1
2
22
21
1
12
11
(1)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1454 – 146
1
1458
;
0
;
1
otherwise
j
resource
to
scheduled
is
i
task
if
a
ij
Each ta
sk i
s
execute
d
at most on on
e reso
urce, so t
here i
s
formul
a 2.
m
j
LS
i
ij
S
e
n
i
a
1
,
,
,
2
,
1
,
1
(2)
Before th
e ex
ecutio
n of LS
tasks,
each
reso
urce
ha
s
alrea
d
y worked for
j
t
and g
a
ined
servi
c
e
co
st b due to the
executi
on
of HS tasks. B
a
se
d on d
e
a
d
line
LS
T
and the available
budg
et left B-b, there are formul
a 3 and
formula 4.
n
i
LS
i
j
LS
j
i
ij
S
e
m
j
t
T
L
a
1
,
,
,
2
,
1
,
Re
/
*
(3)
LS
i
n
i
m
j
j
j
i
ij
S
e
b
B
P
L
a
,
Re
/
*
*
11
(4)
Whe
n
exe
c
ut
ing the LS ta
sks, the
sche
duling
o
b
je
ctive is to maxi
mize the
nu
mber
of
compl
e
ted ta
sks, so the
r
e
is formul
a 5.
LS
i
n
i
m
j
ij
S
e
a
,
max
11
(5)
Acco
rdi
ng to
formula 2 to f
o
rmul
a 5, a li
near plan
nin
g
model can be
esta
blishe
d,
LPM
algorith
m
ca
n
be use
d
to solve the LPM module
in al
g
o
rithm 1, as
shown in algo
rithm 2.
n
i
m
j
ij
a
11
max
.
.
t
s
m
j
ij
n
i
a
1
,
,
1
,
1
n
i
j
LS
j
i
ij
m
j
t
T
L
a
1
,
,
1
,
Re
/
*
n
i
m
j
j
j
i
ij
b
B
P
L
a
11
Re
/
*
*
m
j
n
i
or
a
ij
,
2
,
1
,
,
2
,
1
,
1
0
LS
i
S
e
Algorithm 2: LPM
Step 1: M
the set of all m
a
chine
s
;
Step 2: initialize
b
B
t
T
L
P
R
A
j
LS
i
j
e
nm
i
,
,
,
,
,
,
,
;
Step 3: establ
ish an
d cal
c
ul
ate the linear
planni
ng m
odel, return
nm
A
;
Step 4: for(i=1;i<=n;i++)
for(j=
1
;j<=
m;j+
+){
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Re
sea
r
ch on
Batch Sch
e
d
u
ling in Clo
u
d
Com
puting (Jintao
Jiao
)
1459
i
f
(
1
ij
a
&&
LS
j
S
e
) sc
hedule task
i to mac
h
ine j;
else
contin
ue
;
}
Step 5: return
the expe
cted
com
p
letion tim
e
and m
apping re
sult;
Step 6: end;
5. Experiment and Resul
t
Analy
s
is
Two
expe
rim
ents
are
de
si
gned
in
ord
e
r to test th
e p
r
opo
sed
DB
CMNCT al
gorit
hm. The
simulatio
n
pla
tform is
Clou
dsim [12
-
1
3
]. Clou
dsim i
s
a
n
event drive
n
clo
ud comp
uting sim
u
lati
on
toolkit ba
se
d
on Java; its
main go
al is
to re
sea
r
ch i
n
the effectiv
e re
sou
r
ce al
locatio
n
meth
od
based on the
computin
g e
c
on
omy mod
e
l. In the desi
gned
cloud
computing m
a
rket, the pri
c
e is
proportinal to the
computing ability.
5.1. Experiment
w
i
th
out
HS Task
s
The first exp
e
rime
nt is d
e
sig
ned to test the total numbe
r of completed ta
sks u
nde
r
different bud
get and dea
dline wh
en there a
r
e no
HS tasks. And the size
of each task i
s
distrib
u
ted ra
ndomly between.
between
600 MI(Milli
on Instructio
n
)
and 1
260
0 MI. There a
r
e 7
para
m
eter g
r
oup
of b
udg
et and
d
eadli
ne; the
pa
ra
meter i
s
in
creased
step
b
y
step,
su
ch
a
s
(450
000,5
0
0
0
) 70
000
0,7000
) (8500
00,90
00
) (10
000
0
0
,1000
0)
(115
000
0,13
000)
(130
000
0,16
000) (14
500
0
0
,1900
0).
Figure. 3 Nu
mber of Com
p
leted Ta
sks
Corre
s
p
ondin
g
to Each Group of Budge
t and Dea
d
lin
Figure 3 sho
w
s the result of the first
exper
im
ents.
With the increa
se of bud
get an
d
deadli
ne the numbe
r of tasks
compl
e
ted
by Min-min
algorithm is in
crea
sed, but n
o
t absolute; the
compl
e
ted ta
sks of grou
p 4 is less than grou
p 3,
also the com
p
l
e
ted tasks of
group 6 is l
e
ss
than group 5
;
the budget
and de
adlin
e
is increa
se
d
while the n
u
mber of
com
p
leted tasks
is
decrea
s
e
d
. T
he main
re
ason is
wh
en
some of t
he ta
sks
can
not b
e
com
p
leted,
the bud
get a
nd
deadli
ne
will
have an
imp
a
ct on
the
schedul
e of ta
sks t
hat the n
u
mbe
r
of
co
mpleted ta
sks is
decrea
s
e
d
. In comp
ari
s
on,
the numbe
r o
f
comple
te
d tasks of DB
CMNCT and
DBCT algo
rith
m is
increa
sed
wit
h
the i
n
crem
ent of b
udg
et and
de
ad
lin
e. For ea
ch
grou
p of
bu
d
get an
d d
ead
line,
the perfo
rma
n
ce of DB
CM
NCT i
s
better than that
of
DBCT a
nd M
i
n-min, there are ab
out 22
%
and 10.0
6
% perfo
rman
ce i
m
provem
ent.
5.2. Experiment
w
i
th
HS Tasks
The p
r
opo
rtio
n of HS tasks in the
se
co
nd
expe
rime
nt chan
ge
s from 10% to 8
0
%; and
the numbe
r o
f
all tasks i
s
r
andomly di
str
i
buted from 5
00 to 1000.
Figure 4
sho
w
s th
at the HS tasks
comp
letion
rate
of DBCM
NCT al
gorithm i
s
mu
ch m
o
re
highe
r than
the othe
r two
algorith
m
s, th
is is mainly
b
e
ca
use the p
r
iority is
give
n to HS ta
sks i
n
DBCM
NCT al
gorithm while
not in the other two
al
gorithms. This m
ean
s more u
r
gent tasks ca
n
be com
p
leted
with DBCM
NCT algo
rithm
than the othe
r two.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1454 – 146
1
1460
Figure 5 sh
o
w
s that
with the increa
se
o
f
the HS
tas
k
s
,
the total tas
k
s
completion rate of
Min-min
a
n
d
DB
CT
algo
rithm i
s
relat
i
vely stable.
And th
e to
tal tasks
co
mpletion
rate
of
DBCM
NCT al
gorithm i
s
hig
her th
an the
other t
w
o al
g
o
rithm
s
if the
prop
ortio
n
of
HS tasks i
s
le
ss
than 50%, b
u
t when the
prop
ortion
of HS tas
ks is more th
an 50% the
performan
ce of
DBCM
NCT
a
l
gorithm
is worse th
an th
e othe
r t
w
o.
This is mainl
y
becau
se th
e HS
tasks
wil
l
occupy th
e g
ood
re
sou
r
ce
s, so the
co
mpletion if
affected
with th
e in
cre
a
se of
the HS
tasks. In
the actu
al
cl
oud
com
puti
ng sy
stem t
he nu
mbe
r
o
f
HS tasks should
be limi
t
ed so
that t
he
perfo
rman
ce
coul
d be bett
e
r.
Figure. 4 HS Tasks Com
p
l
e
tion Rate
with
D
i
ffer
ent Pr
opor
tion of H
S
Tasks
Figure. 5 Total Tasks
Completion Rate with
D
i
ffer
ent Pr
opor
tion of H
S
Tasks
5.3. LPM Algorithm Perfo
r
mance Anal
y
s
is
The b
r
an
ch a
nd bou
nd al
g
o
rithm [14] is
use
d
to solve
the linear
pla
nning m
odel i
n
LPM
Algorithm, th
e time
and
sp
ace
cost
s
of the
whol
e al
g
o
rithm
are al
so mai
n
ly con
s
ume
d
i
n
sol
v
ing
the linea
r pl
annin
g
mod
e
l
. Table 2 a
nd Tabl
e
3
sho
w
that th
e time and
spa
c
e
co
sts
are
increa
sed
wh
en the numb
e
r
of tasks is in
cre
a
sed.
Table 2. Time
and Space Cost with Diffe
rent Numb
er o
f
Tasks
Number of
Tasks
50
100
500
1000
Time
Cost(ms)
10 10 20
40
Space
Cost(K)
101 192 922
1825
Table 3. Time
and Space Cost with Diffe
rent Numb
er o
f
Reso
urce
s
Number of
Reso
urces
50
100
500
1000
Time Cost(ms)
10
20
20
40
Space Cost(K)
866
1704
5100
8421
The time co
st of the algorithm is in the
millise
c
on
d
level, which
is far less than the
executio
n tim
e
of the
ta
sk, so
the time
co
st
of
the
sched
uling
al
gorithm
can
be ig
nored.
Th
e
spa
c
e
co
st is increa
sed lin
early, so the
num
be
r of tasks o
r
re
so
urces shoul
d be
limited.
Although the
linear pl
anni
n
g
model is a
NP pr
o
b
lem,
it has alre
ady
been u
s
ed i
n
other
sched
uling a
l
gorithm.
No
wad
a
ys the
r
e are
a lot
of resea
r
ch
results
can
make th
e time
compl
e
xity of serial alg
o
rit
h
m cont
rol in
polynom
ial level. And with the improv
ement of parallel
techn
o
logy, the se
rial line
a
r plan
ning a
l
gorithm
can
be easily pa
rallelize
d
, whi
c
h also gre
a
tly
redu
ce
s the time and spa
c
e co
sts.
6. Conclusio
n
This
pape
r di
scusse
s h
o
w
to maximize t
he num
be
r of
com
p
leted ta
sks
with con
s
train of
budg
et and d
eadlin
e; also
when the
r
e
are
some u
r
gent or imp
o
r
tant tasks, the algo
rithm
can
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Re
sea
r
ch on
Batch Sch
e
d
u
ling in Clo
u
d
Com
puting (Jintao
Jiao
)
1461
ensure th
ese
tasks
compl
e
ted a
c
cordi
n
g to user
’
s
d
e
mand.
HS tasks
will be
a
llocate
d on g
ood
resou
r
ces; th
en hi
ghe
r
HS task
compl
e
tion rate a
nd t
o
tal task
com
p
letion
rate
can b
e
a
c
hiev
e
d
by using the
DBCM
NCT al
gorithm b
a
se
d on LPM.
In this p
ape
r, we
pay mo
re attention to
t
he u
s
er.
He
ncefo
r
th by
combinin
g the
cha
r
ate
r
of the sy
stem itself, we will
continue to impr
ove the algorithm in or
der t
o
make it m
o
re
comp
re
hen
si
ve, more wid
e
ly used a
nd
more
simple t
o
use.
Ackn
o
w
l
e
dg
ements
This work was supp
orted
by Fund of Fuji
an Edu
c
a
t
ion Institutions(JB1
410
1),
F
und of
Wuyi Unive
r
sity(XD2
014
0
6
) and
re
se
arch proje
c
t of universit
y teaching reform in Fuj
i
an
Province (JA
S
15134
2).
Referen
ces
[1]
Shuo
W
a
n
g
,
Shih
ong
L
i
u.
Ke
y T
e
chno
lo
g
y
of
Agric
u
lt
ural
Prod
uctio
n
a
n
d
Market
Informatio
n
Matchin
g
i
n
Bi
g D
a
ta Era.
T
E
LKOMNIKA Indo
nesi
a
n
Jou
r
nal
of El
ectric
al E
ngi
ne
erin
g
. 20
14; 1
2
(3)
:
221
2-22
18.
[2]
Patel G, Meht
a R, Bhoi U.
Enha
nce
d
L
o
ad Ba
la
nced
Min-
min
Alg
o
ri
thm for St
atic
Meta T
a
sk
Sched
uli
ng i
n
Clou
d
Co
mputi
n
g
. Proced
ia C
o
mputer Sci
e
n
c
e. 2015; 5
7
: 545-5
53.
[3]
Chint
apalli, V
e
nkataram
i Reddy
. A
deadline and
budget
constrained c
o
st
and tim
e
optimiz
ation
algorithm for c
l
oud computing.
Co
mmun
icati
ons i
n
C
o
mp
uter an
d Infor
m
a
t
ion Sci
enc
e
. 2
011;
193(
4):
455-
462.
[4]
Suja
n S, Devi
RK.
A batchmode dy
na
mic s
c
hed
uli
ng sch
e
m
e for cl
oud c
o
mputi
n
g
. Pro
c
eed
ings
of
201
5 Glob
al C
onfere
n
ce o
n
Commun
i
cati
o
n
T
e
chnolo
g
ies
.
Nagerco
il. 20
15: 297-
30
2.
[5]
F
an Yu
an
yu
an
, Lia
n
g
Qingz
h
ong,
Ch
en Y
u
n
s
ong, Y
a
n
Xu
e
s
ong,
Hu
Ch
en
g
y
u, Ya
o
Hon
g
,
Liu
Ch
ao,
Z
eng D
e
ze.
E
x
ec
utin
g time
and c
o
st-a
w
a
r
e
task sche
d
u
l
ing
in h
y
b
r
i
d
c
l
ou
d usi
ng
a
modifi
ed DE
algorithm.
Co
mmu
n
ic
ations i
n
Co
mp
uter and
Information Sci
ence
. 20
16; 57
5: 74-83.
[6]
Rabi
Pras
ad P
adh
y, M
anas
Ranj
an
Patra.
Architec
ture
&
Desig
n
of Affordab
le
an
d Hi
g
h
l
y
Ava
i
l
abl
e
Enterpris
e
Cloud Serv
ice.
Internati
ona
l Jo
ur
nal
of Clo
ud
Co
mp
uting
an
d Servic
es Sci
ence
. 20
13
;
2(2): 85-1
05.
[7]
Hon
g
she
ng Su
, Ying Qi. A chaos clo
ud p
a
rti
c
le s
w
arm a
l
go
rithm base
d
av
aila
bl
e transfer
capab
ilit
y.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2014; 1
2
(1): 3
8
-47.
[8]
Mazal
o
v V, L
u
k
y
an
enk
o A,
Lu
ukkai
n
e
n
S. Equi
libr
i
um
in c
l
o
ud c
o
m
putin
g mark
et.
Perfor
ma
nce
Evalu
a
tion
. 2
0
15; 92: 40-
50.
[9]
Hamsa
nan
dh
in
i S, Moh
a
n
a
R
S
.
Maxi
mi
z
i
n
g
the rev
enu
e w
i
t
h clie
nt cl
assifi
cation
in
Cl
ou
d
Co
mp
utin
g
mar
k
et
. 20
15
Internati
o
n
a
l
C
onfere
n
ce
on
Comp
uter
C
o
mmunicati
on
a
nd Inform
atics. Co
imbator
e.
201
5: 1-7.
[10]
Manso
u
ri
Dou
El Kefe
l, Ben
y
ettou Mo
hamm
ed. Pr
ic
e o
p
ti
misatio
n
of c
l
o
ud c
o
mputi
ng
service
in
a
mono
pol
istic competitiv
e mar
k
et.
Multiage
nt and Grid Syste
m
s
. 20
16; 11(
4
)
: 259-27
1.
[11]
Xu
Bol
e
i, Qin
T
ao,
Qiu Guopi
ng, L
i
u T
i
e-Yan.
Co
mpeti
t
ive pric
ing f
o
r clou
d co
mp
uting
in
an
evol
ution
a
ry market
. Procee
di
ngs of the Inte
rnatio
nal J
o
int Confer
ence on
Autonom
ous Agents
a
nd
Multia
gent S
y
s
t
ems. Ist
anbul. 201
5; 3: 1755-
175
6.
[12]
Rodri
go
N. Cl
oudS
im: A No
vel F
r
ame
w
o
r
k
fo
r the Mod
e
l
i
ng
and S
i
mul
a
tion
of Clo
ud
Computi
n
g
Infrastructures and Serv
ices. T
he Univer
sit
y
of Melbo
u
rn
e Australi
a. 200
9.
[13]
RN Ca
lh
eiros,
R Ra
nja
n
, A B
e
lo
glaz
ov, CAF
De R
o
se, R B
u
yya
. Cl
ou
dSi
m
: a toolkit for
mode
lin
g a
n
d
simulati
on of
clou
d computi
ng env
ironm
e
n
ts and eva
l
u
a
tion of reso
u
r
ce provisi
o
n
i
n
g
alg
o
rithms.
Softw
are Practice and Exp
e
ri
e
n
ce
. 201
1; 41(
1): 23-50.
[14]
Jin F
ang, Yu
an Z
h
i-h
ua, Li
u Qing-Qua
n
, Xin
g
Z
han
g. Quantized fe
edb
ack contro
l of net
w
o
rk
empo
w
e
rme
nt ammuniti
on
w
i
t
h
data-r
a
te lim
itations.
TELK
OMNIKA Telecommunic
a
tio
n
Co
mp
utin
g
Electron
ics an
d Contro
l
. 201
4; 12(3): 52
5-5
32.
Evaluation Warning : The document was created with Spire.PDF for Python.