Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
10
,
No.
5
,
Octo
be
r
2020
,
pp.
4881
~
4891
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
5
.
pp
4881
-
48
91
4881
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
An
e
ne
rgy
o
ptim
izat
i
on with
i
mp
roved Q
OS
a
pp
roach f
or
a
dapti
ve
c
l
oud
r
esour
ces
Danthul
uri S
udh
a
1
, San
j
ay C
hit
nis
2
1
CMR
Instit
ute
of
Technol
og
y
,
Visvesvara
y
a
Technol
ogi
ca
l
Uni
ver
sit
y
(
VTU
)
,
I
ndia
2
Da
y
ana
ndasa
g
a
r
Univer
sit
y
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
13
, 201
9
Re
vised
Ma
r
6
,
2020
Accepte
d
Ma
r
1
8
, 202
0
In
recent
t
imes,
the
utilizati
on
of
cl
oud
computing
VM
s
is
ext
rem
e
l
y
enha
nc
ed
in
ou
r
da
y
-
to
-
d
a
y
l
if
e
due
to
the
a
m
ple
uti
l
izati
on
of
digital
appl
i
ca
t
ions,
n
et
work
applian
ce
s,
por
ta
bl
e
gadge
ts,
and
informati
on
devi
c
es
et
c
.
In
thi
s
cl
oud
computing
VM
s
num
ero
us
diffe
ren
t
sche
m
es
ca
n
be
imple
m
ent
ed
li
k
e
m
ult
imedia
-
signal
-
proc
essing
-
m
ethods
.
Thus,
eff
icient
per
for
m
anc
e
of
the
se
cl
oud
-
computin
g
VM
s
bec
om
es
an
ob
li
ga
to
r
y
constra
i
n
t,
pr
e
ci
sel
y
for
th
ese
m
ult
imedi
a
-
signal
-
pro
ce
ss
i
ng
-
m
et
hods.
How
eve
r,
la
rg
e
amount
of
ene
rgy
consum
pti
on
a
nd
red
uct
ion
in
e
ffic
i
ency
of
the
se
cl
oud
-
co
m
puti
ng
VMs
a
re
the
ke
y
issues
fac
ed
b
y
dif
fer
ent
c
loud
computing
orga
niz
a
ti
ons.
Th
erefore,
her
e,
we
have
int
rodu
ce
d
a
d
y
n
amic
volt
ag
e
and
fr
eque
nc
y
s
ca
l
in
g
(DV
FS
)
bas
ed
ada
p
ti
v
e
c
l
oud
resourc
e
re
-
conf
igur
abi
l
ity
(
)
te
chn
ique
fo
r
cl
oud
computi
ng
device
s,
whi
ch
eff
icientl
y
r
educes
ene
rg
y
cons
um
pti
on,
as
we
l
l
as
p
erf
orm
oper
ations
in
ver
y
le
ss
ti
m
e.
W
e
have
demons
tra
te
d
an
eff
i
cient
resourc
e
all
oca
t
ion
and
uti
lization
t
ec
hn
ique
to
op
ti
m
ize
b
y
r
educ
ing
di
ffe
ren
t
costs
of
the
m
odel
.
W
e
have
al
so
demons
tra
te
d
ef
fic
i
ent
en
erg
y
opti
m
iz
ation
tec
hnique
s
b
y
red
uci
ng
ta
sk
lo
ads.
Our
expe
r
i
m
ent
al
out
comes
show
s
t
he
superior
ity
of
ou
r
proposed
m
odel
in
t
erms
of
av
er
age
run
ti
m
e
,
po
wer
consum
pti
o
n
and
ave
rag
e
power
r
equi
red
tha
n
an
y
othe
r
sta
te
-
of
-
ar
t
t
ec
hniqu
es.
Ke
yw
or
d
s
:
ACRR
,
Cl
oud
c
om
pu
ti
ng,
DVFS,
Energy c
onsum
pt
ion
,
Re
so
urce al
loc
at
ion
Copyright
©
202
0
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Dan
t
hu
l
ur
i
Sud
ha,
CM
R In
sti
tute
of Tech
nolo
gy,
Visv
es
va
raya
Tech
no
l
og
ic
al
Un
i
ver
sit
y
(
VT
U
)
,
Be
ng
al
uru,
Ka
rn
at
a
ka 560
091,
I
nd
ia
.
Em
a
il
:
dan
thu
l
ur
is
udha
15@
gm
ai
l.co
m
1.
INTROD
U
CTION
Du
e
t
o
eve
r
-
e
nhanci
ng
dem
a
nd
a
nd
po
pu
la
r
it
y
of
cl
oud
c
om
pu
ti
ng
ap
plica
ti
on
s,
var
i
ous
com
pan
ie
s
has
m
ov
ed
thei
r
f
oc
us
to
cl
ou
d
c
om
pu
ti
ng
t
o
decr
ea
se
c
os
ts
an
d
f
or
the
bet
te
r
util
iz
at
ion
of
res
ources
he
nc
e
i
t
ref
e
rr
e
d
as
a
nex
t
ge
ner
at
io
n
c
om
pu
ti
ng
app
li
cat
io
n.
C
loud
c
om
pu
ti
ng
a
pp
li
cat
io
n
te
rm
ed
as
a
novel
com
pu
ta
ti
on
al
m
od
el
,
wh
ic
h
pro
vid
es
on
-
de
m
and
res
ource
s
and
re
qu
ire
d
inform
at
ion
,
netw
ork,
st
or
a
ge
a
nd
data
to
the
s
ub
s
cribe
rs.
Th
e
cl
oud
-
c
om
pu
ti
ng
m
od
el
c
om
bin
es
har
dware
de
vice
locat
ions
an
d
var
i
ous
so
ft
war
e
res
ou
rces
over
the
c
loud
netw
ork
to
decr
ea
se
the
m
anag
em
ent
c
os
ts.
Cl
oud
Co
m
pu
ti
ng
is
a
hig
hly
e
m
erg
ed
te
ch
no
l
og
y
w
hich
offe
rs
high
a
m
ou
nt
of
st
or
a
ge
ca
pacit
y,
instant
sca
la
bili
ty
and
work
on
the
pri
nciple
of
pa
y
-
per
-
us
e
wh
ic
h
is
on
l
y
fo
r
the
ti
m
e
per
i
od
subs
cribers
are
util
iz
ing
it
[1
]
.
Cl
oud
com
pu
ti
ng
a
pp
li
cat
ion
s
are
di
stribu
te
d
int
o
t
hr
ee
sect
ions
s
uch
as
Infr
ast
r
uctu
re
-
as
-
a
-
se
r
vice
(
IaaS
),
Plat
fo
rm
as
-
a
-
se
r
vice
(Pa
aS)
a
nd
S
of
t
war
e
a
s
-
a
-
ser
vi
ce
(S
aaS
)
.
Virt
ualiz
at
ion
is
t
he
m
os
t
essenti
al
te
chn
iq
ue
for
cl
oud
com
pu
ti
ng
a
pp
li
cat
ion
s,
w
hic
h
use
d
t
o
decre
ase
res
ource
util
iz
at
ion
.
Mo
reover
,
virtu
al
iz
at
ion
helps
t
o
act
ive
the
nu
m
erous
virtu
al
m
achines
(V
Ms)
on
on
e
m
achine
by
al
locat
ing
ever
y
res
ource
,
wh
ic
h
belo
ng
s
to
the b
a
sic
h
a
rdwar
e
[2].
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
5
,
Oct
ob
e
r 2
020
:
48
81
-
48
91
4882
At
the
sam
e
tim
e,
en
or
m
ou
s
grow
t
h
of
cl
ou
d
c
om
pu
ti
ng
a
pp
li
cat
io
ns
in
r
ecent
ye
ars
ha
s
le
d
t
o
ris
e
of
nu
m
erous
da
ta
centers
w
hi
ch
can
le
d
to
m
assive
a
m
ount
of
e
nergy
con
s
um
ption
.
T
he
vast
util
iz
at
ion
of
el
ect
rici
ty
le
d
to
va
ri
ous
e
nv
i
ronm
ental
issues
du
e
t
o
release
of
huge
ca
r
bon
as
well
as
it
enh
a
nce
s
the
m
ai
ntenance
cost
of
in
form
ation
proces
sing
center
s
[
3]
.
The
key
reas
on
be
hind
m
assive
c
on
s
um
pti
on
of
powe
r
is
the
insuffici
e
nt
util
iz
at
ion
of
res
ources
[4
]
.
F
or
an
instance
,
in
a
recent
re
por
t
it
is
con
cl
ude
d
that
sever
al
researc
her
s
ha
ve
m
onit
or
ed
m
or
e
th
an
50
00
ser
vers
for
a
l
ong
6
m
on
ths’
ti
m
e
-
per
i
od
a
nd
obs
erve
d
that
m
os
t
of
the
se
rv
e
rs
op
e
rates
only
at
10%
to
50%
of
t
heir
whole
capaci
ty
.
It
i
s
al
so
ob
s
er
ve
d
that
gen
e
rall
y
m
or
e
nu
m
ber
of
re
s
ources
are
assi
gn
e
d
to
VMs
t
han
us
ually
re
qu
i
red
at
that
tim
e
[4
]
.
Moreo
ve
r,
cl
oud
com
pu
ti
ng
a
pp
li
cat
io
ns
exp
e
riences
e
xtrem
el
y
var
ia
ble
ta
sk
-
l
oad
s
du
e
t
o
w
hich
s
erv
e
r
can
not
con
ti
nu
e
for
a
l
ong
-
tim
e
pe
rio
d
a
nd
there
is
a
nee
d
of
res
our
ce
re
-
m
anag
em
ent
course
a
fter
a
s
pecific
ti
m
e
-
per
io
d
t
o
enh
a
nce t
he pr
esent re
source
sh
ari
ng [3,
5
]
.
Du
e
to
m
od
er
nizat
ion
i
n
el
e
ct
ronic
te
ch
no
l
og
ie
s
,
the
pe
rfor
m
ance
cost
of
va
rio
us
devi
ces
is
highl
y
e
m
erg
ed
as
we
ll
as
the
ener
gy
con
s
um
ption
in
these
de
vi
ces
is
hig
hly
inc
reased
[6
]
.
Mo
reover
,
seve
ral
issues
can
be
occ
urr
ed
su
c
h
as
ecolo
gical
,
eco
no
m
ic
,
and
te
chn
ic
al
issues
du
e
to
high
a
m
ou
nt
of
e
nergy
consum
ption
.
Fu
rt
her
m
or
e,
e
xtensi
ve
powe
r
c
on
s
um
ption
can
gi
ve
rise
to
c
ooli
ng
co
sts
an
d
if
num
ero
us
com
pu
ti
ng
f
aci
li
ti
es r
e
m
ai
n
ac
ti
ve
fo
r
a long
per
i
od, th
en
ch
ances of h
eat
in
g
pro
blem
o
ccurrence ca
n
inc
rease
du
e
t
o
w
hich
s
yst
e
m
reli
abilit
y
and
avail
abil
it
y
can
increase.
A
cl
oud
c
om
pu
ti
ng
ap
plica
ti
on
nee
ds
s
uppo
rt
of
com
piler
tim
e
and
e
xecu
ti
on
tim
e
to
ru
n
a
ppli
cat
ion
s
prec
ise
ly
.
Ho
we
ver,
du
e
to
high
-
e
nergy
co
ns
um
ption,
the ex
ec
utio
n
t
i
m
e o
f
ta
sk
-
loa
ds
ca
n be i
ncr
e
ased
from
the as
usual
ti
m
e.
In
cl
oud
c
om
pu
ti
ng
ap
plica
ti
on
s
,
m
os
t
essenti
al
thing
is
the
al
locat
io
n
of
res
ources
e
ff
ic
ie
ntly
t
o
increase
the
pe
rf
or
m
ance
of
cl
oud
in
f
or
m
ation
ce
nter
s.
T
her
e
fore,
to
e
nhance
the
pe
rfor
m
ance
an
d
c
ount
e
r
these
issues
in
recent
tim
e
di
ff
e
ren
t
te
ch
niques
are
intr
od
uced
by
var
i
ous
resea
rch
e
rs
su
c
h
as
Hierar
chical
reli
abili
ty
-
dr
iv
en
sc
he
du
li
ng
(H
R
DS
)
te
ch
ni
qu
e
[
7],
C
on
st
rai
ne
d
earli
est
finish
tim
e
(CEFT)
al
gorith
m
[8
]
,
Con
te
ntion
-
aw
are
e
nergy
-
e
ff
i
ci
ent
du
plica
tio
n
(
FastC
EED
)
te
ch
nique
[
9]
,
Dynam
ic
volt
age
an
d
fr
e
quenc
y
scal
ing
(
DVF
S)
[
10
]
an
d
volt
age
an
d
fr
e
qu
e
ncy
isl
and
(V
F
I)
te
ch
niq
ue
[
11]
.
In
above
al
l
te
chn
iq
ues
,
Dynam
ic
vo
lt
age
a
nd
fr
e
qu
e
nc
y
scal
ing
(
D
VF
S
)
is
one
of
the
m
os
t
wide
ly
e
m
erg
ed
te
chn
i
qu
e
f
or
e
ffi
ci
ent
sche
du
li
ng
of
ta
sk
-
loa
ds.
DVFS
is
a
hi
gh
ly
est
ablish
ed
ene
r
gy
co
ns
um
ption
optim
iz
ation
sche
m
e
for
e
m
bed
de
d
cl
oud
syst
em
s
and
th
e
e
nergy
optim
iz
at
ion
can
be
ac
hi
eve
d
by
scal
in
g
do
wn
the
volt
age
of
a
ny
chip
dyna
m
icall
y.
DV
F
S
t
echn
i
qu
e
help
s
to
ac
hieve
high
QoS
In
t
ern
et
s
er
vices
by
reducin
g
ene
rg
y
consum
ption
i
n
cl
ou
d dev
ic
e
s.
Howe
ver,
thes
e
existi
ng
te
c
hniq
ues
works
upon
the
cl
assic
prototype
w
hich
is
not
ge
nuine
pr
oto
ty
pe
and
the
c
omm
un
ic
at
io
n
cost
of
inter
proces
so
rs
bec
om
es
ver
y
hi
gh
wh
il
e
us
in
g
these
te
chn
i
qu
e
s.
I
n
[12],
an
ene
rg
y
a
nd
run
-
ti
m
e
op
tim
iz
at
ion
te
chn
i
que
is
intr
oduce
d
to
decr
ease
e
nergy
co
nsum
ption
by
sche
duli
ng
of
va
rio
us
ta
s
k
-
l
oad
s
on
nu
m
ero
us
em
bed
de
d
syst
e
m
s.
I
n
[
13]
,
var
i
ou
s
e
nergy
ef
fici
ent
sta
te
-
of
-
a
rt
-
te
chn
iq
ues
ar
e
disc
us
se
d
and
basic
req
ui
rem
ents
of
e
m
bed
de
d
syst
e
m
s
are
p
resen
te
d
to
increa
s
e
the
pe
rfo
rm
ance
of
em
bed
de
d
com
pu
ti
ng.
In
[14],
a
ta
sk
-
sc
hedulin
g
al
go
rithm
is
pr
esent
ed
to
achie
v
e
t
rad
e
-
off
bet
ween
energy
an
d
pe
rfor
m
ance.
T
his
sche
duli
ng
al
go
rithm
helps
to
increa
s
e
the
flexibili
ty
of
heter
og
e
ne
ou
s
com
pu
ti
ng
de
vi
ces
(H
CD
).
T
his
te
chn
i
qu
e
a
tt
enu
at
es
the
volt
age
of
var
i
ous
ta
sk
s
acc
ord
ing
to
com
pu
ti
ng
m
od
el
ta
sk
l
oad.
Hen
ce
,
the
po
we
r
c
onsu
m
ption
is
re
du
ce
d
and
a
no
vel
in
te
racti
on
a
ware
DAG
prototype
is
al
so
prese
nted
f
or
heter
ogene
ou
s
com
pu
ti
ng
de
vices
(
HCD
)
to
re
du
ce
e
nerg
y
con
s
um
ption
[15].
Howe
ver, thes
e tec
hn
i
ques
do
no
t
ha
ve feat
ur
es
li
ke
or D
y
nam
ic
N
et
wo
r
k
S
hutd
own
(
)
[
16]
.
Th
us
,
t
o
en
ha
nc
e
perf
or
m
ance
an
d
re
duce
e
nergy
co
nsum
ption
of
heterogen
e
ous
c
om
pu
ti
ng
de
vices
an
e
ff
ect
ive
ta
sk
sc
he
du
li
ng
te
chn
i
qu
e
is
nee
ded
an
d
rely
upon
D
VF
S
te
chn
i
qu
e
w
hich
offer
s
va
rio
us
energ
y
opt
im
iz
ation
fa
ci
li
ti
es
and
hel
ps
to
achie
ve
be
tt
er
res
ource
util
iz
at
ion
an
d
trade
-
off
betw
een
perform
ance
an
d
energy.
T
heref
or
e
,
we
ha
ve
pr
ese
nted
a
D
ynam
ic
vo
lt
ag
e
an
d
f
reque
nc
y
scal
ing
(DVFS)
based
a
dap
ti
ve
cl
oud
res
ource
re
-
c
onfig
ur
a
bi
li
ty
(
)
te
ch
nique
for
cl
oud
c
ompu
ti
ng
dev
ic
es
,
wh
ic
h
eff
ic
ie
nt
ly
red
uces
energy
c
onsu
m
ption,
a
s
well
as
perform
op
e
rati
on
s
in
ver
y
le
ss
ti
m
e.
Our
pr
opos
e
d
a
ppro
ac
h
hel
ps
t
o
achieve
high
trade
-
off
bet
ween
e
nergy
and
perf
or
m
a
nce
in
hete
rogen
e
ous
com
pu
ti
ng
dev
ic
e
s.
This
sche
du
li
ng
te
c
hn
i
qu
e
hel
ps
t
o
ac
hieve
high
by
ope
rati
ng
on
l
ower
tra
ns
m
issi
on
rate
s
an
d
c
on
ce
di
ng
m
ini
m
u
m
del
ay
.
This
te
ch
nique
sup
ports
eff
ic
ie
ntly
t
o
achie
ve
obje
ct
ive
of
en
ha
ncin
g
the
co
st
of
interact
ion
an
d
com
pu
ta
ti
on
al
ene
rg
y.
This
t
echn
i
qu
e
ad
apt
ively
sche
du
le
s
ta
sk
s,
al
locat
es
it
s
res
ource
s,
a
nd
pro
vid
es
scal
able
im
plantat
i
on.
Our
pro
posed
te
ch
niqu
e
pe
rfor
m
s
m
uch
bette
r
c
om
par
e
to
ot
he
r
sta
te
-
of
-
art
-
te
c
niques.
This
pa
pe
r
is
orga
nized
i
n
f
ollow
in
g
sect
io
ns,
w
hich
a
re
as
fo
ll
ows.
In
se
ct
ion
2,
we
pr
e
sent
relat
e
d
work
in
the
fi
el
d
of
ene
r
gy
consum
ption
i
n
cl
oud
en
vir
onm
ent.
In
sect
ion
3,
we
des
cribe
d
ou
r
pro
po
s
ed
m
et
ho
do
l
og
y.
I
n
sect
ion
4,
exp
e
rim
ental
resu
lt
s
a
nd
perform
ance
eval
uation
s
how
n,
an
d
sect
io
n
5
con
cl
ud
e
s
our pape
r.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
An
e
ne
r
gy op
ti
miza
ti
on wi
th i
mp
r
ove
d QOS
approac
h f
or
adapti
ve cl
ou
d
r
eso
ur
ces
(
D
anthu
l
ur
i
Sudha
)
4883
2.
RELATE
D
W
ORK
In
rece
nt
ye
ars,
the
dem
and
of
cl
ou
d
com
pu
ti
ng
ap
plica
ti
on
s
has
ta
ken
i
m
m
e
ns
e
gro
wt
h.
Ther
e
f
or
e,
to
c
on
t
ro
l
high
de
m
and
f
ro
m
the
cl
ie
nts,
there
i
s
a
nee
d
of
e
xc
essive
res
ource
s
of
dif
fer
e
nt
ty
pes
.
All
these
res
ou
rces
co
nsum
es
hi
gh
am
ou
nt o
f
el
ect
rici
ty
and
hen
ce p
ow
e
r
c
on
s
um
ption
is m
or
e.
Accord
i
ng
t
o
a
20
13
re
sea
rch
in
U
nited
Stat
es
of
A
m
erica,
the
inf
or
m
at
ion
cl
oud
proce
ssin
g
ce
nters
c
onsu
m
es
910
ℎ
el
ect
rici
ty
,
w
hich
is
al
m
os
t
sam
e
as
t
he
s
umm
at
ion
of
34
t
her
m
al
en
erg
y
pla
nts
w
ho
s
e
el
ect
rici
ty
gen
erati
on
ca
pacit
y
is
500
per
ye
ar.
This
c
on
s
um
e
d
ene
r
gy
was
su
f
fici
ent
f
or
entire
N
e
w
York
Ci
ty
f
or
tw
o
ye
a
rs
a
nd
ti
ll
2020,
th
is
ene
rg
y
co
nsum
ption
will
r
ise
to
1400
ℎ
w
hich
is
enorm
ou
s
el
ect
rici
ty
con
su
m
ption
a
nd
al
m
os
t
sam
e
as
t
he
su
m
m
ation
of
50
therm
a
l
ener
gy
plant
s
[17].
Hen
ce
,
t
he
e
ne
rg
y
c
onsu
m
ption
i
n
cl
oud
da
ta
centers
an
d
c
om
pu
ti
ng
proces
sors
has
t
aken
drast
ic
gro
wth.
Ther
e
f
or
e,
c
on
trolli
ng
of
pow
er
co
ns
um
ptio
n
in
data
cente
rs
an
d
em
bed
de
d
proces
sors
is
a
vital
and
cr
it
ic
a
l
requirem
ent,
wh
ic
h
nee
d
to
be
fo
c
us
ed
s
oon.
Th
us
,
an
e
xtensiv
e
re
searc
h
work
desc
ribe
d
in
this
sect
ion
on
energy
balance
d
sc
he
du
li
n
g
al
gorithm
s an
d
t
heir
c
onnecti
on
with
f
or
va
riou
s
em
bed
de
d dev
ic
es
.
In
[
18
]
,
le
vel
of
powe
r
c
ons
um
ption
in
i
nfor
m
at
ion
proc
essing
cente
rs
of
c
hin
a
an
d
pe
rfor
m
ance
is
m
easur
ed
.
T
he
y
con
cl
ude
tha
t
the
power
co
ns
um
ption
in
i
nfor
m
at
ion
pro
cessi
ng
cente
rs
of
CH
I
NA
is
ver
y
high
an
d
var
io
us
te
ch
niques
are
intr
oduce
d
to
reduce
pow
er
co
nsum
ption
an
d
e
nh
a
nce
perform
ance.
I
n
[
19
]
,
an
e
nergy
co
nsum
ption
m
od
el
is
pr
ese
nte
d
ba
sed
on
th
e
ser
ver
m
axim
u
m
po
wer
a
nd
de
gr
ee
of
CPU
util
iz
at
ion
to
pr
e
dict
the
tot
al
power
in
t
he
pr
ese
nt
ser
ve
r.
I
n
[
20]
,
m
ob
i
le
cl
oud
c
om
pu
ti
ng
prot
ot
ype
is
introd
uced
to
reduce
the
en
erg
y
co
nsum
pt
ion
at
the
time
of
wireless
com
m
un
ic
at
ion
base
d
on
dynam
ic
energy
-
a
war
e
cl
oudlets.
T
he
y
pro
vid
e
sim
ulati
on
s
res
ults
base
d
on
t
he
pract
ic
al
ex
per
im
ents.
H
oweve
r,
execu
ti
on
ti
m
e
is
ver
y
hi
gh
us
i
ng
this
te
chn
i
qu
e
,
wh
ic
h
m
ay
deg
ra
de
it
s
pe
rfor
m
ance.
I
n
[21],
an
ef
fici
ent
r
eso
ur
ce
al
loca
ti
on
m
od
el
is
introd
uced
i
n
cl
oud
e
nvir
on
m
ent
and
a
rev
ie
w
on
e
xisti
ng
sche
du
li
ng
a
nd
energy
c
on
s
um
pt
ion
strat
e
gies
is
pr
e
sente
d.
To
offe
r
bett
er
res
ources
in
cl
oud
e
nv
i
ron
m
ent
and
im
pr
ov
e
r
el
at
ion
sh
i
p
wit
h
us
ers
,
sc
heduling
of
re
sou
rces
is
a
n
ext
r
e
m
el
y
essenti
a
l
top
ic
,
wh
ic
h
can
perform
ance o
f
cloud c
om
pu
ti
ng V
Ms
.
In
[22],
a
no
ve
l
energy
awar
e
based
on
VM
sche
du
li
ng
te
c
hn
i
qu
e
is
i
ntrodu
ce
d.
He
re,
both
netw
o
rk
com
po
ne
nts
a
nd
resou
rces
both
are
co
ns
i
dered
to
prov
i
de
an
e
ff
ic
ie
nt
sc
hedulin
g
te
c
hniqu
e.
VM
plac
e
m
ent
and
VM
m
igrati
on
are
t
he
tw
o
esse
ntial
sch
edu
li
ng
ste
ps
t
o
achie
ve
obj
e
ct
ive.
This
te
c
hn
i
qu
e
hel
ps
to
re
duce
energy
co
nsu
m
pt
ion
as
w
el
l
as
traff
ic
ov
e
r
netw
ork.
In
[
23
]
,
e
nergy
aw
are
res
ource
-
sc
hedulin
g
te
ch
ni
qu
e
is
pr
ese
nted
bas
ed
on
D
VF
S
netw
orke
d
inf
or
m
at
ion
pr
ocessin
g
ce
nt
ers
f
or
cl
oud
com
pu
ti
ng
VMs
.
Her
e
,
two
ty
pe
s
of
e
nergies
a
re
m
a
inly
op
ti
m
iz
ed
su
ch
as
com
pu
ti
ng
e
ne
rg
y
an
d
c
ommun
ic
at
io
n
ene
r
gy
to
reduce
ov
e
rall
energy
co
nsum
ption
w
hile
f
ollow
in
g
S
LA
c
onstrai
nts
.
T
his
te
chn
iq
ue
is
di
ff
ic
ult
to
im
plem
ent
in
real
-
tim
e.
In
[24],
an
ef
fic
ie
nt
cost
m
ini
m
iz
at
ion
and
r
eso
ur
ce
util
iz
at
ion
ap
proac
h
is
introdu
ce
d
f
or
cl
oud
com
pu
ti
ng
devi
ces
us
in
g
sta
ble
par
al
le
l
ap
plica
ti
on
s.
T
hi
s
appr
oach
preci
sel
y
decr
eases
cost
by
ch
oo
s
i
ng
dev
ic
es
,
w
hic
h
f
ollo
w
the
m
et
ho
ds
of
le
ast
resou
rce
ut
il
iz
ation
.
Howev
e
r,
t
he
dif
ficult
y
is
to
m
ai
ntain
trade
-
off bet
we
en
pe
rfor
m
ance an
d
e
nergy con
s
um
ption
. In
[
25]
, a
pr
eci
se
sch
e
duli
ng alg
or
it
hm
is intro
duc
e
d
wh
ic
h
rely
upon
DV
F
S
-
e
na
bled
net
work
processi
ng
ce
nters.
T
his
al
gorithm
helps
to
achieve
eff
ic
ie
nt
sche
du
li
ng
for c
loud c
om
pu
ti
ng V
Ms
. Ho
we
ver, this tec
hniqu
e
intr
oduces
op
ti
m
iz
ation
prob
le
m
as w
el
l.
In
a
bove
w
ork
s,
dif
fer
e
n
t
re
s
earche
rs
hav
e
util
iz
ed
dif
fer
e
nt
ene
rg
y
c
onsu
m
ption
an
d
sche
du
li
ng
te
chn
iq
ues
.
Howev
e
r,
only
fe
w
sc
he
du
li
ng
t
echn
i
qu
e
s
a
re
well
-
kn
own
t
o
be
offer
e
d
i
n
r
eal
-
tim
e
app
li
cat
ions
du
e
to
va
rio
us
pro
blem
s
occurre
d
i
n
te
ch
niques
[18,
20
,
24
,
25]
,
li
ke
la
ck
of
bala
ncin
g
be
tween
pe
rform
ance
and
po
wer
c
onsu
m
ption,
op
tim
iz
at
ion
co
m
plexit
y,
la
rge
run
-
tim
e.
Ther
e
fore,
t
o
c
on
t
ro
l
the
se
issues
,
we
ha
ve
int
rod
uced
a
novel
dy
nam
ic
vo
lt
ag
e
and
fr
e
quenc
y
scal
ing
(
DVFS)
base
d
ada
ptive
cl
oud
res
ourc
e
re
-
c
onfig
ur
a
bil
it
y
(
)
te
ch
niqu
e
f
or
cl
oud
com
pu
ti
ng
de
vices,
w
hich
eff
ic
ie
ntly
re
duces
e
ne
rg
y
consum
ption
,
a
s
well
as
per
f
orm
op
erati
on
s
i
n
ve
ry
le
ss
tim
e.
The
refor
e
,
this
te
chn
i
qu
e
i
s
ver
y
m
uch
ef
fici
ent
to establi
sh a t
r
ade
-
off bet
wee
n per
form
ance an
d ene
r
gy con
su
m
pt
ion
.
3.
PROP
OSE
D E
NERGY B
A
LAN
CED
S
C
HEDU
LI
NG
ARCHITE
CT
UR
E
This secti
on
de
fines pr
opose
d
arch
it
ect
ure an
d
it
s v
a
rio
us
m
odules.
This se
ct
ion
also
d
e
sc
ribes
about
the
opti
m
iz
at
ion
of
co
m
pu
ta
ti
on
al
and
re
-
co
nf
i
gurat
ion
c
os
t
in
inf
orm
ation
proces
sing
ce
nt
ers.
Figure
1
dem
on
strat
es
the
pro
pose
d
arch
it
ect
ure.
Her
e
,
w
e
in
tro
du
ce
a
no
vel
ada
ptive
cl
oud
re
sourc
e
re
-
c
onfig
ur
a
bil
it
y
(
)
te
chn
i
qu
e
for
cl
oud
com
pu
ti
ng
VMs
.
The
pro
pose
d
te
ch
nique
w
orks
on
the
pr
inci
ple
of
pa
rall
el
co
m
pu
ti
n
g
wh
ic
h
ca
n
ha
nd
le
nu
m
erous
cl
oud
c
om
pu
ti
ng
VMs
a
nd
they
can
be
con
t
ro
ll
ed
by
a
central
res
ou
rce
ha
ndle
r.
E
ver
y
cl
ou
d
co
m
pu
ti
ng
VM
f
inishes
the
pre
sent
al
locat
ed
ta
sk
as
a
sel
f
-
gove
rn
i
ng
process
or
by
sel
f
-
c
on
tr
olli
ng
it
s
m
e
m
or
y
and
res
ources
.
Me
ssage
pass
i
ng
m
et
ho
d
is
use
d
for
In
tra
-
cl
us
te
r
in
te
racti
on
.
Wh
e
nev
e
r
a
ne
w
ta
sk
is
assigne
d,
a
central
resour
ce
handler
si
m
ul
ta
neo
usl
y
sta
rts
i
m
ple
m
enting
r
eso
ur
ce
distrib
ution an
d
a
dm
i
ssion g
over
ning.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
5
,
Oct
ob
e
r 2
020
:
48
81
-
48
91
4884
Figure
1
.
A
rch
i
te
ct
ur
e
diag
ra
m
o
f
our p
rop
ose
d
m
od
el
Ther
e
a
re
t
hr
e
e
vital
c
om
ponen
ts
in
our
pro
posed
te
ch
ni
qu
e
w
hich
he
lps
t
o
ac
hieve
bette
r
resou
rce
util
izati
on
from
an
inf
rastr
uctu
re
per
c
eptio
n
s
uch
as
in
f
or
m
at
ion
st
or
a
ge,
switc
he
d
lo
c
al
area
netw
ork
(LAN
)
an
d
virt
ua
l
m
achine
ha
ndle
r
(
VMC)
as
dem
on
strat
ed
in
F
ig
ur
e
1.
Wh
enev
e
r,
a
ne
w
ta
sk
i
s
assigne
d
t
he
ar
rival
ti
m
e
of
th
at
ta
sk
is
de
fin
ed
by
and
siz
e
of
that
ta
s
k
is
denoted
by
in
bits.
T
he
tota
l
processi
ng
ti
m
e
of
assig
ne
d
ta
sk
is
le
ss
tha
n
or
e
qual
t
o
the
est
i
m
at
ed
e
m
plo
ye
d
tim
e
(
)
wh
ic
h
is
very
essenti
al
f
or
a
ny
te
chn
i
qu
e
to
be
a
dopt
in
rea
l
tim
e
scenarios.
Our
pro
pose
d
te
ch
nique
w
orks
on
so
m
e
essenti
al
par
a
m
et
ers
w
hich
is
nece
ssary
to
w
ork
in
rea
l
tim
e
scena
rios
su
c
h
as
process
in
g
ta
sk
siz
e
,
the
m
axi
m
u
m
al
lowed
delay
in
sec
an
d
the
ta
sk
gr
a
nu
la
rity
w
hich
s
hows
the
m
axi
m
u
m
num
ber
of
ta
s
ks
(
≫
1
)
can
be gr
oupe
d i
nto
t
he
assi
gned wor
k.
Assum
e
that
the
m
axi
m
u
m
VMs,
w
hich
c
an
be
util
iz
ed
in
the
assi
gn
e
d
ta
s
ks
us
i
ng
our
pro
pose
d
m
et
ho
ds,
can
be
ex
presse
d
by
≫
1
and
pr
ese
nte
d
in
Fi
gure
1.
Our
sc
hedulin
g
te
ch
nique
w
orks
on
the
pr
i
nciple
that
ever
y
VM
can
be
dem
on
strat
ed
as
a
virtu
al
ser
ve
r,
wh
ic
h
can
pro
cess
(Bit
s
pe
r
Seco
nd).
The
op
e
r
at
in
g
rate
can
be
pa
rall
el
ly
scal
ed
at
the
tim
e
of
exec
ution
depen
ding
upon
the
ta
s
k
s
iz
e
in
bits.
As
sum
e
that
al
l
the
ta
sk
f
ollo
ws
t
he
i
nter
val
[
0
,
↑
]
w
her
e
↑
belo
ngs
to
t
he
m
axi
m
um
per
m
issi
ble
op
e
rati
ng r
at
e
.
More
ov
e
r,
the
ta
sk
siz
e
does
not
a
ff
ect
the
est
i
m
at
ed
tim
e
to
com
plete
th
at
assig
ned
ta
s
k
by
VM,
wh
ic
h
is
fi
xed
pri
or
only
to
be
a
dopt
our
m
od
el
in
rea
l
tim
e
scenari
os
a
nd
de
no
te
d
by
in
sec
onds.
Fu
rt
her
m
or
e,
a
VM
can
ha
nd
le
bac
kgr
ou
nd
ta
sk
-
loa
ds
of
a
present
a
ssign
e
d
ta
sk
whose
siz
e
is
an
d
the
bac
kgr
ound
ta
sk
-
loa
d
si
ze
is
.
T
his
backg
rou
nd
ta
sk
-
l
oad
com
es
unde
r
OS
(
operati
ng
syst
e
m
)
pro
gr
am
s.
It
is
assum
ed
t
hat
the
bac
kg
rou
nd
ta
s
k
-
lo
ad
is
store
d
by
basic
m
e
m
or
y
of
VM.
Thus,
the
bac
kgr
ound
ta
s
k
-
loa
d
only
re
q
uire
d
c
om
pu
ti
ng
c
os
t
an
d
do
e
s
no
t
pe
rsu
a
de
i
nteracti
on
cost.
Th
us,
the u
ti
li
zat
ion
par
am
et
er
can
be
e
xpresse
d
a
s,
≜
.
(
↑
)
−
1
∈
[
0
,
1
]
,
(1)
w
he
re,
(1)
represents
t
hat
th
e
dynam
ic
el
e
m
ents
of
t
he
com
pu
ti
ng
e
ne
rg
y
ar
e
the
m
os
t
esse
ntial
pa
rt
to
decr
ease
t
he
c
om
pu
ta
ti
on
al
c
os
t.
A
ssu
m
e
that
the
total
ener
gy
co
nsum
pti
on
by
VM
is
to
finis
h
a
sin
gl
e
ta
sk
of
ti
m
e
interval
(
)
is
denoted
by
in
jo
ule
at
the
op
e
rati
ng
rat
e
.
Th
us
,
t
he
dim
ension
le
ss
ra
ti
o
can
be
expresse
d
as
,
(
)
≜
(
)
.
(
↑
)
−
1
≡
(
)
.
(
↑
)
−
1
,
(2)
Vi
r
t
ua
l
m
ac
hi
ne
c
o
n
t
r
o
l
l
e
r
an
d
t
as
k
s
c
he
du
l
e
r
S
w
i
tch
C
P
U
,
Se
r
v
e
r
s
an
d
st
o
r
ag
e
d
e
v
i
c
e
s
B
as
i
c
l
ay
e
r
(
Vi
r
t
ua
l
i
z
at
i
o
n
)
…
….
.
…
…
T
o
t
he
i
nt
e
r
ne
t
F
r
o
m
t
he
i
nt
e
r
ne
t
O
ut
pu
t
t
as
k
l
o
ad
Inp
ut
t
as
k
l
o
ad
G
at
e
way
R
o
u
te
r
Vm
(
1)
Vm
(
2)
Vm
(
m
)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
An
e
ne
r
gy op
ti
miza
ti
on wi
th i
mp
r
ove
d QOS
approac
h f
or
adapti
ve cl
ou
d
r
eso
ur
ces
(
D
anthu
l
ur
i
Sudha
)
4885
w
he
re,
(
2)
re
presents
the
Tot
al
Energy
Co
nsum
ption
by
the
c
on
ce
r
n
V
M.
For
a
n
in
sta
nce,
t
he
DVF
S
bas
e
d
CPU a
naly
ti
cal
f
or
m
can
be
desc
ribe
d by the
foll
ow
i
ng equ
at
ion
,
(
)
=
2
,
[
0
,
1
]
,
(3)
Her
e
, we ca
n
a
lso
us
e
to c
ompu
te
relat
ive e
nergy c
os
t by c
on
ce
r
ned V
M
f
or the c
om
pletio
n o
f
ta
s
k.
3.1.
Model
li
ng
fo
r
ta
s
k
-
lo
ad red
uctio
n u
sin
g p
roposed
tec
h
nique
In
t
his
sect
ion,
m
od
el
li
ng
f
or
ta
sk
-
loa
d
re
du
ct
io
n
is
discuss
e
d.
Assu
m
e
that
≜
↓
{
,
}
is
the
num
ber
of
ta
sk
s
w
hich
a
re
not
ov
e
rlap
ped
an
d
ca
n
be
perf
or
m
ed
in
par
al
le
l
to
e
xecu
te
va
rio
us
ta
sk
s
.
Assum
e
that
is
the
ta
sk
siz
e
wh
ic
h
a
re
assi
gn
e
d
t
o
the
co
m
pu
ti
ng
(
)
.
T
he
process
t
im
e
of
di
ff
e
rent
ta
sk
s
does
no
t
rely
up
on
t
he
ta
s
k
le
ng
t
h
.
T
her
e
f
or
e,
t
he
processi
ng
rate
ca
n
be
def
i
ned
as
in
bits
per seco
nds,
(
)
=
.
(
)
−
1
,
(4)
This
(4)
s
ho
ws
t
hat
the
m
axi
m
u
m
l
eng
t
h
per
m
it
t
ed
for
a
ta
sk
is
↑
=
.
↑
(
)
.
And
≜
(
)
.
(
↑
)
−
1
≡
.
(
↑
)
−
1
.
T
he
t
otal
siz
e
of
a
jo
b
ca
n
be
ref
e
rred
as
in
bits
a
nd
ta
s
k
siz
e
of
the
ta
s
k,
w
hic
h
is
assig
ne
d
to
the
(
)
by
ta
sk
sche
dule
r
as
sho
wn
in
F
i
gure
1,
ca
n
be
re
ferr
ed
as
≫
0
,
=
1
,
…
.
,
.
.
.
T
o
reduce
ta
sk
l
oad
s
,
we
dis
tribu
te
t
otal
jo
b
siz
e
into
pa
rall
el
ta
sk
s
w
hose
siz
e
bounda
ry li
m
it can
be de
fine
d as
∑
=
.
3.2.
Opt
im
iz
at
io
n
of
reco
nf
ig
urati
on cos
t usi
ng pro
po
se
d
te
c
hnique
This
sect
io
n
pro
vid
es
detai
le
d
m
od
el
li
ng
f
or
the
opti
m
izati
on
of
rec
on
fig
ur
at
io
n
c
os
t
.
T
he
VM
m
od
ule
co
ntr
ol
le
r
is
us
e
d
to
perform
two
ke
y
op
e
rati
ons
s
uch
as
balanci
ng
the
ta
s
k
loa
ds
a
nd
co
ntr
olli
ng
of
virtu
al
m
achin
es
.
F
or
the
c
on
trolli
ng
of
virt
u
al
iz
at
ion
la
ye
r
as
dem
on
stra
te
d
in
F
ig
ur
e
1,
the
virt
ual
m
achine
con
t
ro
ll
e
r
(
)
is
re
qu
i
red
w
hich
he
lps
to
ac
hieve
final
m
app
ing
of
VM
res
our
ces
on
num
ero
us
c
om
pu
ti
ng
VMs. T
he VM
’s
c
har
act
erist
i
cs p
a
ram
et
ers
can
be descri
be
d by (
9) as,
{
,
↑
(
)
,
(
)
,
↑
(
)
,
(
)
,
(
)
,
=
1
,
…
…
.
.
}
(
5)
w
he
re,
al
l
thes
e
par
am
et
ers
can
be
sta
te
d
usi
ng
Virtuali
zat
ion
Lay
er
a
nd
then
they
a
re
transm
it
te
d
to
virtu
al
m
achine
c
ontr
oller
(
)
as
dem
on
strat
ed
i
n
F
ig
ur
e
1.
Th
e
op
erati
ng
r
at
e
can
be
scal
ed
up
or
scal
ed
dow
n
us
in
g
a
n
eff
ic
ie
nt
fr
e
que
ncy
-
scal
in
g
schem
e,
wh
ic
h
is
c
on
t
ro
ll
ed
by
.
The
powe
r
co
nsum
ptio
n
wh
il
e
switc
hin
g
from
op
er
at
ing
f
reque
nc
y
1
to
fr
e
qu
e
nc
y
2
can
be
(
1
:
2
)
in
jou
le
.
Th
is
powe
r
consum
ption
m
ai
nly
rely
upon
the
te
c
hn
i
que
us
ed
an
d
on
the
CP
U’
s
pre
sent
in
t
he
wor
ks
ta
ti
on.
T
his
f
un
ct
io
n
(
1
:
2
)
consi
sts
of
s
om
e
pr
operti
e
s
su
c
h
as
t
he
functi
on
(
1
:
2
)
rely
upon
the
e
nt
ire
fr
e
qu
e
ncy
ga
p
|
1
−
2
|
,
it
beco
m
es
zero
at
1
=
2
and
re
m
ai
n
non
-
decre
asi
ng
i
n
the
entire
f
reque
nc
y
gap
|
1
−
2
|
,
it
is
com
bin
ed
c
onve
xly at
1
an
d
2
. Our m
od
el
ha
ve
so
m
e char
act
e
risti
cs w
hich
c
an be s
how
n us
ing
(
10),
(
1
:
2
)
=
(
1
−
2
)
2
,
(6)
w
he
re,
re
pr
ese
nts
rec
onfig
ura
ti
on
c
o
st
f
or
t
he
unit
switc
hing
of
fr
e
quency
and
the
values
of
is
boun
de
d
on
ly
t
o
so
m
e
hund
red
s
o
f
/
(
)
2
. I
n our
m
od
el
, f
or ever
y
jo
b
t
he
s
iz
e
rem
ai
ns
sam
e
ove
r
the
res
pecti
ve
op
e
rati
ng
tim
e
and
a
ny
kind
of
fl
uctuati
ons
not
occ
urre
d
in
the
ta
s
k
-
l
oads
duri
ng
ta
sk
-
execu
ti
on.
Va
r
iou
s
ta
s
ks
can
be
pa
rall
el
execu
te
d
at
run
-
tim
e,
du
e
to
the
induce
d
ti
m
e
ov
er
hea
d
us
in
g
fr
e
qu
e
ncy
-
scal
ing
te
c
hn
i
qu
e
is
ve
ry
le
ss
in
f
ew
for
DVFS
-
enab
le
d
a
rch
it
e
ct
ur
es.
T
he
a
bove
-
m
entioned
pr
e
dicti
on
that
the
util
iz
at
ion
par
am
et
er
can
be
c
onti
nuous
value
d
a
nd
it
r
equ
i
res
c
on
ti
nuous
c
om
pu
ta
ti
on
al
rates,
wh
ic
h
is
denoted
b
y
. T
he
can
off
e
r
a
n i
ns
ta
nce
of CP
U’
s
which
off
e
r
a
finite
set as
,
ℍ
≜
{
̂
(
0
)
≡
0
,
̂
(
1
)
,
…
,
̂
(
−
1
)
≡
↑
}
,
(7)
w
he
re,
the
se f
i
nite set
co
ns
ist
s o
f discret
e co
m
pu
ta
ti
on
al
r
at
e
. Th
e
op
ti
m
ali
ty
loss
fr
om
b
ot
h
co
ntinuo
us
a
nd
discrete
DVFS
en
a
bled
te
c
hn
iqu
e
s,
ca
n be el
i
m
inate
d
by
(
8) as,
≜
{
̂
(
0
)
≡
0
,
̂
(
1
)
,
…
…
,
̂
(
−
1
)
≡
1
}
,
(8)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
5
,
Oct
ob
e
r 2
020
:
48
81
-
48
91
4886
w
he
re,
the
dis
crete
val
ue
set
of
w
hich
re
pr
ese
nt
t
he
fr
e
qu
e
ncy
set
ℍ
as
sh
ow
n
i
n
(7).
A
vi
rtu
al
powe
r
consu
mp
ti
on
c
ur
ve
ca
n
be
denoted
a
s
̃
(
)
and
f
or
m
ed
us
i
ng
piecew
ise
li
nea
r
inter
polati
on
a
nd
the
per
m
it
te
d
op
e
r
at
ing
po
i
nts a
r
e,
{
(
̂
(
)
,
(
̂
(
)
)
)
,
=
0
,
…
…
.
.
,
(
)
−
1
}
(9)
w
he
re,
the c
orr
esp
onding
ver
t
ex po
i
nts ca
n b
e presente
d
as
,
(
̂
(
)
,
(
̂
(
)
)
)
(
̂
(
+
1
)
,
(
̂
(
+
1
)
)
)
(10)
This
a
bove
m
entione
d
m
ain
ta
ins
t
he
c
on
ti
nu
it
y
an
d
ca
n
be
us
e
d
f
or
the
prov
isi
on
i
ng
of
resou
rces.
T
he
us
e
of
su
gge
sts
that
with
the
help
of
,
the
ave
ra
ge
en
erg
y
c
os
t
of
DVFS
e
nab
le
d
te
ch
niques
rem
ai
ns
under
the
est
im
at
ed
i
nter
val
of
ti
m
e
durati
on
.
Her
e
,
e
ver
y
co
nfi
gurati
on
rely
up
on
CP
U
ty
pe
,
siz
e
of
m
e
m
or
y
and
c
os
t
pe
r
tim
e.
The
cos
t
rely
upon
t
he
ty
pe
of
c
onfi
gurati
on.
The
i
nter
nal
cost
of
assum
ed
to
be
zer
o
in
all
info
rm
ation
clo
ud
centers.
3.3.
Model
li
ng
of
e
ff
ic
ie
nt re
sour
ce alloc
at
i
on
In
t
his
sect
io
n,
m
od
el
li
ng
f
or
eff
ect
ive r
es
ource
al
locat
io
ns
is
pr
ese
nted
. Here,
offers
tw
o
ty
pes
of
se
rv
ic
es
s
uc
h
as
balanci
ng
of
the
load
and
s
har
i
ng
of
com
pu
ta
ti
on
al
resour
ces
.
Pr
e
ci
sel
y,
these
serv
ic
e
us
e
d
t
o
fine
-
tu
ne
t
he
rate
of
com
pu
ta
ti
on
(
,
=
1
,
…
…
…
.
,
)
a
nd
siz
e
of
ta
s
k
(
,
=
1
,
…
…
…
.
,
)
for
t
he
DVFS
enab
le
d
cl
oud
com
pu
ti
ng
as
dem
on
str
at
ed
in
F
i
gure
1.
T
he
m
ai
n
obj
ect
ive
is
to
re
du
c
e
the total
c
om
pu
ta
ti
on
al
e
nerg
y i
n
jo
ule,
wh
i
ch
is
def
i
ned in
(
11)
,
≜
∑
(
)
=
1
,
(11)
wh
e
re
,
th
e
tot
al
com
pu
ta
ti
onal
energy
co
nsum
ption
rely
on
t
he
r
un
-
ti
m
e/
j
ob
in
seco
nd
s
,
ope
rati
ng
tim
e
/
ta
sk
(
)
needed
by
dem
on
strat
ed
in
F
i
gure
1.
Pre
ci
sel
y,
in
F
ig
ur
e
1
al
l
the
li
nk
s
are
operate
d
by
the s
witc
hing
unit
ad
a
ptively
. T
he
e
ntire c
ompu
ta
ti
onal
ov
e
r
head f
or the
ℎ
li
nk can
b
e
exp
r
essed
a
s,
2
(
)
+
,
(12)
w
he
re,
the
c
onditi
on
on
tot
al
ru
n
-
ti
m
e
per
job
to
sat
isf
y
the
so
luti
on
of
opti
m
iz
at
i
on
pro
blem
c
an
be
expresse
d
as
,
ma
x
1
≤
≤
{
2
(
)
}
+
≤
,
(13)
Assum
e that t
he
total
com
pu
t
at
ion
al
e
nergy
op
ti
m
iz
ation
is
su
es ca
n be e
xpress
ed
as
fo
ll
ow
i
ng,
min
{
ℂ
,
,
}
∑
(
.
(
↑
)
−
1
)
(
)
=
1
↑
+
(
−
)
2
,
(14)
w
he
re it sta
te
s
that,
(
+
(
)
)
≤
,
=
1
,
…
…
…
,
(15)
∑
=
,
=
1
(1
6
)
0
≤
≤
↑
,
=
1
,
…
…
…
,
(1
7
)
≥
0
,
=
1
,
…
…
.
,
,
(1
8
)
w
he
re,
in
(14
)
sta
rting
first
te
r
m
rep
rese
nts
com
pu
ta
ti
on
al
e
nergy
and
t
he
f
ollowi
ng
te
rm
rep
r
esents
re
-
c
onfig
ur
at
i
on
e
nergy,
w
hic
h
ca
n
be
ex
pre
ssed
by
(
)
j
oi
ntly
by
t
h
e
c
om
pu
ti
ng
VMs
.
M
oreov
e
r,
in
(
14),
re
pr
es
ents
the
prese
nt
rate
of
c
om
pu
ta
ti
on
and
is
the
require
d
rate
of
c
om
pu
ta
ti
on
.
H
ere,
rem
ai
ns
const
ant
w
hile
processi
ng
of
t
ask
an
d
s
hows
pr
ese
nt
sta
te
of
(
)
w
her
eas
can
be
var
ia
ble
and
c
ha
ng
e
s
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
An
e
ne
r
gy op
ti
miza
ti
on wi
th i
mp
r
ove
d QOS
approac
h f
or
adapti
ve cl
ou
d
r
eso
ur
ces
(
D
anthu
l
ur
i
Sudha
)
4887
ta
sk
to
ta
s
k.
H
ere,
(15
)
s
how
s
the
c
onditi
on
for
w
hich
as
sign
e
d
ta
s
k
m
us
t
be
e
xec
utes
in
seco
nds
w
herea
s
(16) sho
ws
t
he c
onditi
on in
which t
he
assi
gned job m
us
t be
div
ide
d
i
nto
pa
rall
el
tasks.
3.4.
So
luti
on
f
or o
pt
im
iz
at
ion
p
r
ob
le
m
using
This
sect
ion
pro
vid
e
va
rio
us
so
l
utio
n
f
or
handlin
g
optim
iz
at
ion
prob
le
m
us
ing
our
ef
fici
ent
sche
du
li
ng
a
rc
hitec
ture
.
Fir
stl
y,
le
t
the
ti
m
e
delay
occ
urr
ed
w
hile
switc
hing
of
f
re
qu
e
ncies
rem
ai
ns
const
ant
wh
ic
h
is
in
du
ce
d
by
DVF
S
-
e
na
bled
-
te
ch
niques.
H
ow
e
ve
r,
t
he
ti
m
e
delay
occ
urre
d
wh
il
e
s
witc
hing
of
f
reque
ncies
bet
ween
var
i
ou
s
can
be
op
ti
m
iz
ed
by
a
no
n
-
negat
iv
e
f
un
ct
i
on
(
,
)
(
)
an
d
the
so
luti
on
for
optim
iz
at
ion
pro
blem
,
wh
ic
h
is
sta
te
d
in
(
15)
an
d
(
18)
,
can
be
der
i
ved.
This
non
-
ne
ga
ti
ve
functi
on
(
,
)
(
)
helps t
o
retai
n t
w
o
c
har
act
erist
ic
s s
uch as,
a.
The
non
-
ne
gati
ve fu
nctio
n
(
.
,
.
)
rem
ai
ns
non
-
dec
reasin
g
th
r
oughout t
he
inter
va
l
|
−
|
.
b.
The pr
oduct
of
an
d
non
-
neg
a
ti
ve
f
un
ct
io
n
(
.
,
.
)
re
m
ai
n
conve
x
i
n
.
To
c
ontr
ol
op
tim
iz
at
ion
pr
oble
m
,
the
first
te
rm
in
(
14)
w
hich
re
pr
e
s
ent
c
om
pu
ta
tio
nal
ene
rg
y
∑
(
.
(
↑
)
−
1
)
.
(
)
=
1
.
↑
ca
n
be
interc
hange
d
by
the
e
ne
r
gy
functi
on
(
1
,
…
…
…
.
,
)
.
Sim
i
la
rly
,
the
re
-
co
nf
i
gur
at
ion
e
nergy
(
−
)
2
can
be
interc
ha
ng
e
d
by
the
e
ne
rg
y
functi
on
(
1
,
…
…
…
.
)
wh
ic
h
are
co
njo
i
ntly
conve
x
in
t
he
interval
{
,
=
1
,
…
…
…
.
.
}
2
.
Sim
i
la
rly
,
in
(
15)
t
he
pa
ram
et
er
can
be
intercha
nged
by
(
)
without
changin
g
a
ny
co
nd
it
io
n
or
li
near
e
qu
at
i
on.
T
he
op
ti
m
iz
at
ion
pro
ble
m
pr
ese
nted
from
(14)
to
(
18)
is
no
t
a
co
nvex
t
ype.
I
n
fact,
t
hi
s
prob
le
m
is
a
loo
sel
y
co
up
le
d
opti
m
iz
at
ion
ty
pe,
wh
e
re
the
pa
ram
et
er
,
=
1
,
…
…
…
re
presents
t
he
co
m
pu
ta
ti
on
al
pro
blem
s.
The
s
olu
ti
on
f
or
op
ti
m
iz
ation
of com
pu
ta
ti
on
al
and
re
-
c
onfig
urat
ion
al
pro
ble
m
is pr
ese
nted
as,
min
{
,
}
∑
(
(
.
(
↑
)
−
1
)
(
)
↑
+
(
−
)
2
)
=
1
(19)
Assum
e
that,
{
∗
,
∗
,
=
1
,
…
…
…
}
re
pr
ese
nts
the
s
olu
ti
on
set
for
c
om
pu
ta
ti
on
al
and
re
-
c
onfig
ur
at
i
onal
op
ti
m
iz
a
tio
n
prob
le
m
,
w
hich
is
s
how
n
in
(1
5
)
to
(
1
8
)
.
At
la
st,
we
present
our
pro
po
s
ed
m
od
el
in
an
ef
f
ic
ie
nt algorit
hm
f
or
m
, w
hich
is as
fo
ll
ows:
Algorithm :
algorithm
1. Fix
(
,
,
)
2. Fix
(
,
(
)
,
1
↑
,
)
3.
Fix
(
,
,
,
(
)
,
,
)
Parameter for channel processing of
VMs
4. Collect
5. Verify the attainable constraints of (14
)
6.
1
≅
(
+
(
)
)
≤
,
=
1
,
…
…
…
,
7.
2
≅
∑
=
,
=
1
8. if
≈
(
1
&
2
)
ℎ
,
9.
(
)
10. else
11. Special optimization complexity :
12.
min
(
,
,
(
)
)
13. subjected to:
14. conditions in (15) and (16)
15. end if
16. return
,
,
(
)
basic SLA constraints
Paramete
r processing of VMs
4.
PERFO
R
MANC
E E
V
ALU
ATIO
N
Now
days,
th
e
req
ue
st
of
cl
oud
com
pu
ti
ng
dev
ic
es
ha
s
hig
hly
em
er
ged
in
real
-
ti
m
e
du
e
to
the
extensi
ve
ut
il
iz
ation
of
inf
or
m
at
ive
dev
ic
es,
dig
it
al
instr
um
ents,
network
a
pp
li
anc
es
a
nd
portable
ga
dg
et
s
et
c.
Mult
i
m
edia
-
sign
al
-
proc
e
ssing
m
et
ho
d
is
well
-
known
te
c
hn
i
qu
e
,
wh
ic
h
can
be
util
iz
ed
in
these
cl
oud
-
c
om
pu
ti
ng
de
vices.
T
he
refor
e
,
the
pe
rfor
m
ance
of
t
hese
com
pu
ti
ng
de
vices
m
us
t
be
superi
or
due
to
the
extensi
ve
dem
and
of
t
he
se
com
pu
ti
ng
dev
ic
es
in
day
-
t
o
-
day
li
fe.
H
ow
e
ve
r,
hi
gh
-
e
nergy
co
ns
um
ption
i
n
these
com
pu
ti
ng
dev
ic
es
ca
n
distu
rb
their
per
f
or
m
ance.
Thu
s
,
this
sect
ion
disc
us
se
s
about
the
ba
la
ncing
betwee
n
pe
rform
ance
and
power
c
on
s
um
pti
on.
To
ac
hieve
these
obj
ect
iv
es,
we
ha
ve
in
tro
du
ce
d
a
Dy
nam
ic
vol
ta
ge
a
nd
Fr
e
qu
e
ncy
Sc
al
ing
(DVF
S)
base
d
A
da
ptive
Cl
oud
Re
so
urce
Re
-
Co
nf
i
gurab
il
it
y
(
)
te
chn
iq
ue
for
heter
og
e
ne
ou
s
com
pu
ti
ng
de
vices,
wh
ic
h
e
ff
ic
ie
ntly
re
duces
ene
rg
y
c
onsu
m
ption,
as
well
as
pro
vid
e
s
uperi
or
pe
rfor
m
ance.
The
r
un
-
ti
m
e
can
be
e
va
luate
d
c
onsid
erin
g
va
rio
us
j
obs
a
s
30,
50,
10
0,
and
1000.
G
ra
ph
ic
al
re
pr
ese
nt
at
ion
of
our
outc
om
es
is
a
lso
pr
e
sente
d
co
ns
ide
rin
g
exec
ution
ti
m
e,
nu
m
ber
of
ta
sk
s
an
d
ene
r
gy
consum
ption
.
T
he
r
un
-
ti
m
e
and
total
po
wer
c
on
s
um
ed
can
be
ev
al
ua
te
d
us
in
g
differe
nt
par
a
m
et
ers
in
T
able
1,
wh
ic
h
is
dem
on
strat
ed
i
n
t
he
f
ollo
wing
se
ct
ion.
Our
pro
posed
m
od
el
is
te
ste
d
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
5
,
Oct
ob
e
r 2
020
:
48
81
-
48
91
4888
on
ℎ
sci
entifi
c
da
ta
set
.
W
e
have
co
ns
i
der
e
d
di
ff
ere
nt
siz
es
of
sci
e
ntific
workflo
w
e
xperi
m
ents
as
30,
50,
100
a
nd
10
00.
O
ur
pro
posed
m
od
el
e
m
plo
ye
d
on
64
-
bit
window
s
10
OS
with
16
GB
RAM
wh
ic
h
co
ns
ist
s
on
I
NTEL
(R
)
co
re
(
TM)
i
5
-
4460
proces
so
r
.
It
c
onsist
s
of
3.2
0
G
Hz
CPU.
T
his
pro
j
ect
is
si
m
ulate
d
us
i
ng
Ne
on.3 edit
or an
d
c
ode is
wr
it
te
n
i
n
J
AVA.
4.1.
Co
m
pa
r
ati
ve
stu
d
y
In
this
m
od
er
n
era,
com
pu
ti
ng
de
vices
has
ru
le
d
m
ark
et
i
n
dif
fer
e
nt
fiel
ds
li
ke
m
edica
l,
healt
hcare
so
luti
ons,
t
rad
i
ng,
softwa
re
c
om
pan
ie
s
et
c.
Th
us
,
f
uture
e
xperti
se
is
cl
early
in
favor
o
f
t
he
se
cl
oud
-
c
ompu
ti
ng
dev
ic
es
due
t
o
their
e
xte
ns
iv
e
re
qu
i
rem
ents
.
H
oweve
r,
th
e
eff
ic
ie
ncy
of
these
c
om
pu
ti
ng
de
vices
m
ay
be
reduce
d
du
e
to
hi
gh
-
ene
r
gy
c
on
s
um
ption
a
nd
la
c
k
of
ef
fici
ent
re
source
util
iz
at
ion
te
chn
i
qu
e
s.
C
onseq
ue
ntly
,
these
iss
ues
c
an
be
s
ort
ed
out
usi
ng
e
ff
ic
i
ent
ta
sk
sc
heduling
te
ch
niqu
es.
T
her
e
f
or
e,
to
al
locat
e
res
ource
pro
per
ly
an
d
s
chedule
al
l
the
ta
sk
s
e
ff
ic
ie
nt
ly
to
overc
om
e
powe
r
c
onsum
p
t
ion
pro
blem
,
we
hav
e
pr
esente
d
a
no
vel
Dy
nam
ic
volt
age
a
nd Fr
e
qu
e
ncy
Sca
li
ng
(
D
VFS)
ba
sed
A
dap
ti
ve
Cl
oud
Re
s
ourc
e
Re
-
Co
nfi
gur
abili
ty
(
)
te
ch
niq
ue.
A
preci
se
ta
s
k
s
cheduli
ng
te
c
hniq
ue
ca
n
en
ha
nce
t
hro
ughp
ut
of
the
syst
e
m
,
increase
connecti
ons
w
it
h
subscri
ber
s
,
offe
r
bette
r
r
eso
ur
ce
util
iz
at
ion
a
nd
ca
n
aff
ord
to
ha
nd
le
m
ulti
ple
ta
s
ks
at
a
tim
e
et
c.
T
he
res
ults
are
dem
on
strat
e
d
in
c
ontrast
t
o
ot
her
sta
te
-
of
-
art
te
c
hn
i
qu
e
s
in
te
rm
s
of
energy
consum
ption
,
run
-
ti
m
e,
pow
er
s
um
and
a
ver
a
ge
powe
r
as
s
how
n
i
n
T
able
1
us
i
ng
sci
entifi
c
m
od
el
Cyb
er
Shak
e
for
var
i
ou
s
j
obs
as
30,
50,
100
a
nd
1000
.
E
nergy
c
on
s
um
ption
us
in
g
our
pro
pose
d
te
chn
iq
ue
f
or
Cyb
er
S
hake
30
is
1303.
74
Watt
s,
Cyb
er
Shake
50
is
13
30.92
Watt
s,
Cyb
er
Shak
e
100
is
1436.
83
W
at
ts
an
d
Cyb
er
Shak
e
1000
is
32
28.60
Watt
s
de
m
on
strat
ed
in
T
able
1,
w
hich
is
ver
y
le
ss
i
n
con
t
rast
to
othe
r
sta
te
-
of
-
art
te
chn
i
qu
e
s.
Tab
le
2
dem
on
stra
te
s
A
ver
a
ge
E
xecu
ti
on
t
i
m
e
evaluati
on
us
i
ng
our
te
chn
i
que
a
nd
com
par
iso
n
with
oth
e
r
sta
te
-
of
-
a
rt
-
t
echn
i
qu
e
s
pr
e
sented
f
or
sc
ie
ntific
m
od
el
Cyb
er
Shak
e
f
or
va
rio
us
job
s
as
30,
50
,100
an
d
10
00.
The
A
ve
ra
ge
E
xec
ution
Ti
m
e
us
in
g
our
te
chn
iq
ue f
or
s
ci
entifi
c m
od
el
Cy
be
rShak
e
100
is
31.16
2
sec
and
Cyb
er
Shake
1000
is 4.9
74 s
ec.
Table
1
.
Var
i
ous
par
am
et
ers
com
par
ison f
or prop
os
ed
A
C
RR
techn
iq
ue
vs
DVFS
us
in
g
sci
e
ntific
m
od
el
ℎ
P
a
ram
e
ters
DV
F
S
ℎ
3
0
ℎ
50
ℎ
100
ℎ
1000
ℎ
30
ℎ
50
ℎ
100
ℎ
1000
VM
=
2
0
VM
=
3
0
VM
=
6
0
VM
=
2
0
VM
=
2
0
VM
=
3
0
VM
=
6
0
VM
=
2
0
T
o
tal
E
x
e
c
u
ti
o
n
T
ime
(s
)
6
3
5
9
.4
1
1
4
4
4
8
.90
3
0
1
2
4
.41
7
4
5
4
3
.57
2
9
3
8
.2
2
2
9
5
3
.8
6
3
1
1
6
.2
2
4
7
9
4
.3
9
P
o
we
r S
u
m
(W
)
1
2
1
7
5
9
2
2
.6
4
2
9
0
6
8
5
5
2
.8
9
6
1
1
7
7
3
3
8
.4
0
1
4
9
9
6
8
1
2
2
.08
4
6
4
6
3
9
0
.51
4
6
9
6
9
6
3
.35
4
9
5
5
1
3
3
.77
7
6
2
3
6
1
0
.31
Avera
g
e
P
o
we
r (
W)
1
9
.1
4
6
3
2
0
2
0
.1
1
8
1
8
3
2
0
.3
0
8
2
2
9
2
0
.1
1
8
1
8
4
1
5
.8
1
3
6
2
3
1
5
.9
0
1
1
0
3
1
5
.9
0
1
1
0
3
15.
9
0
1
1
0
5
P
o
we
r
C
o
n
su
m
p
ti
o
n
(
ℎ
)
3
4
9
5
.4
2
8
5
1
8
.3
9
1
8
9
6
6
.33
2
3
6
3
0
3
.2
8
1
3
0
3
.7
4
1
3
3
0
.9
2
1
4
3
6
.8
3
3
2
2
8
.6
0
Table
2
.
A
ver
a
ge
e
xecu
ti
on
ti
m
e com
par
ison o
f propose
d
te
chn
i
qu
e
w
it
h
oth
e
r
Stat
e
-
of
-
art
-
te
ch
niques
us
in
g
sci
e
ntific
m
od
el
ℎ
DAGs
Nu
m
b
e
r
o
f
n
o
d
es
Av
erage E
x
ecut
io
n
ti
m
e
(s)
EMO
-
b
ased
algo
ri
th
m
[
2
6
]
DVFS
ℎ
100
100
3
1
.53
3
0
1
.244
3
1
.16
2
ℎ
1000
1000
2
2
.71
7
4
.54
4
.97
4
4.2.
Graphi
cal
rep
resent
at
i
on
This
sect
i
on
dem
on
strat
es
th
e
gr
a
phic
al
re
pr
ese
ntati
on
of
our
e
valuated
ou
tc
om
es.
Her
e,
F
i
gure
2
sh
ows
run
ti
m
e
com
par
ison
of
our
pr
opos
e
d
t
echn
i
qu
e w
it
h
DVFS
te
ch
nique
us
i
ng
sci
ent
ific
work
l
oa
d
ℎ
for
dif
fer
e
nt
jo
bs
as
30,
50,
100
a
nd
1000.
Her
e
,
F
ig
ur
e
3
sh
ows
P
ower
Su
m
Com
par
ison
of
our
pro
po
se
d
te
chni
qu
e
with
D
VFS
te
chn
i
qu
e
usi
ng
sci
e
ntific
work
l
oa
d
ℎ
f
or
diff
e
re
nt
job
s
a
s
30,
50,
100
a
nd
1000.
Here,
F
ig
ur
e
4
s
hows
Av
e
ra
ge
Po
w
er
Re
qui
red
C
om
par
ison
of
our
pro
po
s
ed
te
chn
iq
ue
with
D
VF
S
te
ch
ni
qu
e
us
in
g
sci
e
ntific
w
orkl
oa
d
ℎ
f
or
diff
e
re
nt
job
s
as
30,
50,
100
an
d
10
00.
Si
m
il
arly
,
F
igure
5
s
how
s
P
ower
Co
nsum
pt
ion
Com
par
iso
n
of
our
pro
po
sed
te
ch
nique
with
D
VFS
te
chn
i
qu
e
us
i
ng
sci
entif
ic
w
ork
load
ℎ
for
diff
e
ren
t
j
obs
as
30,
50,
100
a
nd
1000.
Fu
rt
her
m
or
e,
F
igure
6
sho
ws
Av
e
ra
ge
Ru
n
Ti
m
e
Co
m
par
ison
of
our
pro
po
s
ed
te
ch
nique
wit
h
D
VFS
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
An
e
ne
r
gy op
ti
miza
ti
on wi
th i
mp
r
ove
d QOS
approac
h f
or
adapti
ve cl
ou
d
r
eso
ur
ces
(
D
anthu
l
ur
i
Sudha
)
4889
te
chn
iq
ue
us
i
ng
sci
e
ntific
w
orkloa
d
ℎ
f
or
di
ff
ere
nt
j
obs
a
s
3
0,
50,
10
0
and
10
00.
T
hi
s
res
ult
dem
on
strat
es
the
s
up
e
rio
rity
of
our
pro
po
se
d
te
chn
iq
ue
i
n
te
rm
s
of
po
wer
c
onsu
m
ption,
exec
utio
n
tim
e, av
erag
e
pow
e
r
a
nd ti
m
e.
Figure
2
.
Run
-
t
i
m
e co
m
par
iso
n usin
g o
ur
AC
RR
techn
iq
ue
wi
th
DVFS
Figure
3
.
P
ow
e
r
s
um
co
m
par
ison u
sin
g our
ACRR
tech
nique
with
DVFS
Figure
4
.
A
verage
p
owe
r
c
omparis
on usi
ng
our
A
CR
R t
ech
nique
with
DVFS
0
20000
40000
60000
80000
V
M=2
0
V
M=3
0
V
M=5
0
V
M=3
0
CYB
ERSHA
KE 3
0
CYB
ERSHA
KE 5
0
CYB
ERSHA
KE 1
0
0
CYB
ERSHA
KE 1
0
0
0
Run
-
tim
e
comparis
on
of ou
r
ACR
R
tech
niqu
e
wi
th
DV
FS
DV
FS (
s)
ACR
R (
s)
0
20000000
40000000
60000000
80000000
100000000
120000000
140000000
160000000
V
M=2
0
V
M=3
0
V
M=5
0
V
M=3
0
CYB
ERSHA
KE 3
0
CYB
ERSHA
KE 5
0
CYB
ERSHA
KE 1
0
0
CYB
ERSHA
KE 1
0
0
0
Pow
er Sum
comparsi
on of
our
ACR
R
tech
niqu
e
wi
th DVF
S
DV
FS (
W)
ACR
R (
W)
0
5
10
15
20
25
V
M=2
0
V
M=3
0
V
M=5
0
V
M=3
0
CYB
ERSHA
KE 3
0
CYB
ERSHA
KE 5
0
CYB
ERSHA
KE 1
0
0
CYB
ERSHA
KE 1
0
0
0
Ave
rage Po
wer
comparsi
on of our ACRR t
echnique
with DVF
S
DV
FS (
W)
ACR
R (
W)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
5
,
Oct
ob
e
r 2
020
:
48
81
-
48
91
4890
Figure
5
.
P
ow
e
r
c
on
s
um
pt
ion
com
par
ison u
sing o
ur A
CR
R t
echn
i
qu
e
w
it
h DVFS
Figure
6
.
A
verage
e
xecu
ti
on
t
i
m
e co
m
par
iso
n usin
g o
ur
AC
RR
techn
iq
ue
with
DVFS
5.
CONCL
US
I
O
N
The
sig
nifican
ce
of
c
on
tr
olli
ng
high
-
e
ne
rgy
con
s
um
ption
and
ta
sk
l
oad
al
locat
ion
f
or
ever
y
cl
ou
d
com
pu
ti
ng
V
Ms
is
ver
y
essenti
al
.
The
refor
e
,
to
at
ta
in
balanci
ng
betwee
n
po
wer
c
on
s
um
ption
a
nd
perform
ance
fo
r
c
om
pu
ti
ng
process
or
s
i
n
c
loud
en
vir
on
m
ent,
we
hav
e
i
ntr
oduce
d
a
no
vel
dynam
ic
vo
lt
age
and
fr
e
qu
e
ncy
scal
ing
(
DVFS
)
base
d
a
da
ptiv
e
cl
o
ud
resou
r
ce
re
-
co
nfi
gura
bili
ty
(
)
te
ch
niqu
e
f
or
cl
ou
d
com
pu
ti
ng
V
Ms.
Eff
ic
ie
nt
m
od
el
li
ng
f
or
al
l
the
thre
e
ty
pes
of
c
os
t
su
c
h
as
c
om
pu
ta
ti
on
co
st
and
re
-
c
onfig
ur
at
i
on
c
os
t
is
pr
ese
nted.
T
he
perf
or
m
ance
of
th
e
m
od
el
is
en
han
ce
d
by
reducin
g
al
l
th
re
e
costs.
Fu
rt
her
m
or
e,
m
od
el
li
ng
fo
r
eff
ic
ie
nt
resour
ce
al
locat
io
n
an
d
util
iz
at
i
on
is
presente
d
an
d
ta
sk
lo
ads
ar
e
reduce
d
w
hich
is
a
gr
eat
cha
ll
eng
e
f
or
oth
e
r
sta
te
-
of
-
a
rt
te
chn
i
qu
e
s.
Num
ero
us
res
our
ces
can
be
ef
fici
ently
al
locat
ed
ada
pt
ively
at
a
tim
e
us
i
ng
this
te
c
hn
i
qu
e
.
T
he
ex
per
im
ental
resu
lt
s
are
sho
wn
in
te
rm
s
of
run
tim
e
ta
ken
,
re
duct
io
n
in
ene
rg
y
c
onsu
m
ption
an
d
avera
ge
powe
r
re
quire
d
f
or
c
loud
c
om
pu
ti
ng
VMs.
T
he
A
ver
a
ge
Run
Tim
e
us
ing
our
propos
ed
m
od
el
fo
r
sci
entifi
c
m
od
el
ℎ
100
is
31.16
2
sec
an
d
ℎ
1000
is
4.
97
4
sec.
Si
m
il
arly
,
Po
wer
C
on
s
um
ption
f
or
ℎ
30
is
1303.74
Watt
s,
ℎ
50
is
1330.
92
Watt
s,
ℎ
100
is
1436.
83
Watt
s
an
d
ℎ
100
0
is
3228.
60
Watt
s
w
hich
is
ver
y
lo
w
com
par
e
to
ot
her
sta
te
-
of
-
a
rt
-
te
ch
niques.
Our
e
xperim
e
ntal
res
ults
ve
rifies
the
supe
rior
it
y
of
our
m
od
el
in
te
rm
s
of
pe
rfor
m
ance
an
d
po
wer
co
nsum
ption
in
c
ontrast
to
oth
e
r
sta
te
-
of
-
art
-
te
c
hn
i
qu
e
s.T
her
e
f
or
e
,
a
trade
-
of
f
betwee
n
perf
or
m
ance
and
energy
co
nsu
m
pt
ion
is
m
a
i
ntained
us
in
g
our
pr
opose
d
al
go
rith
m
.
In
f
uture,
eff
ic
ie
nt
m
od
e
ll
ing
to
op
ti
m
i
ze
com
m
un
ic
at
ion
c
os
t
will
be prese
nted
.
0
50000
100000
150000
200000
250000
V
M=2
0
V
M=3
0
V
M=5
0
V
M=3
0
CYB
ERSHA
KE 3
0
CYB
ERSHA
KE 5
0
CYB
ERSHA
KE 1
0
0
CYB
ERSHA
KE 1
0
0
0
Pow
er Consumption
comparsi
on of
our
ACR
R
tech
niqu
e
wi
th DVF
S
DV
FS (
Wh)
ACR
R (
Wh)
0
100
200
300
400
EM
O
DVFS
PS1
Av
er
age
Ex
ec
ution
Time compar
sio
n
of
our
ACRR
techn
ique
with
DVFS
CYB
ERSHA
KE 1
0
0
CYB
ERSHA
KE 1
0
0
0
Evaluation Warning : The document was created with Spire.PDF for Python.