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.
11
,
No.
1
,
Febr
uar
y
2021
, pp.
686
~
698
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v11
i
1
.
pp
686
-
698
686
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om
Optimi
zation
of
energy
consu
mp
tion in cl
oud com
pu
ting
datac
enters
Ah
med
Osm
an
,
Assim S
agahyr
oon
,
R
aafat
A
buru
kb
a,
F
ad
i
Alou
l
Depa
rtment
o
f
C
om
pute
r
Scie
n
ce a
nd
Engi
n
ee
rin
g,
Am
eri
c
an
Uni
ver
sit
y
of
Sharj
a
h,
Uni
Emirat Ar
ab
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
l
e hist
or
y:
Re
cei
ved
Ja
n 1
3,
2020
Re
vised
Ju
l
29
,
2020
Accepte
d
Aug 8, 2
020
Cloud
computing
has
emerge
d
as
a
pra
ct
i
ca
l
p
ara
digm
for
providi
ng
IT
resourc
es,
infr
astruc
tur
e
and
servic
es.
Th
is
has
le
d
to
the
esta
b
l
ishm
ent
of
dat
a
ce
nt
ers
th
at
have
subs
ta
nt
ia
l
ene
rg
y
de
m
ands
for
thei
r
oper
a
ti
on
.
Thi
s
work
inv
e
stiga
te
s
the
optim
iz
at
ion
of
en
e
rg
y
consum
pti
on
in
c
loud
dat
a
ce
nt
er
using
ene
rg
y
eff
ic
i
ent
al
lo
cation
of
t
a
sks
to
resourc
es
.
The
work
see
ks
to
dev
el
o
p
form
al
opti
m
iz
a
ti
on
m
odel
s
tha
t
m
ini
m
ize
t
he
ener
g
y
c
onsum
pti
on
of
computational
resourc
es
and
e
val
ua
te
s
th
e
use
of
ex
isti
ng
opti
m
iz
ation
sol
ver
s
in
te
sting
the
se
m
odel
s.
I
nte
ger
l
inear
progra
m
m
ing
(IL
P)
te
chn
ique
s
are
used
to
m
odel
th
e
sche
du
li
n
g
proble
m
.
Th
e
obje
c
ti
ve
is
to
m
ini
m
iz
e
th
e
tot
a
l
power
c
onsum
e
d
by
th
e
active
and
id
le
cor
es
of
the
serve
rs’
CP
Us
while
m
ee
ting
a
set
of
con
strai
nts.
Nex
t,
we
use
the
se
m
odel
s
to
c
arr
y
out
a
d
et
a
il
ed
per
form
anc
e
co
m
par
ison
bet
we
en
a
sel
ec
t
ed
set
of
Gen
eri
c
I
LP
and
0
-
1
Bo
ole
an
sati
sfi
abi
l
i
t
y
base
d
solvers
in
solving
the
ILP
form
u
la
ti
ons
.
Sim
ula
t
ion
result
s
ind
ic
a
te
tha
t
in
s
om
e
ca
ses
the
develope
d
m
odel
s
have
save
d
up
to
38%
in
ene
rg
y
consu
m
pti
on
when
compare
d
to
co
m
m
on
te
chni
que
s
such
as
round
robin
.
Further
m
ore
,
result
s
al
so
show
ed
tha
t
generic
I
LP
solvers
had
sup
eri
or
per
form
anc
e
when
compare
d
to
SA
T
-
base
d
I
LP
sol
ver
s
espe
c
ia
l
l
y
as
the
num
ber
o
f
ta
sks
and
resourc
es
grow
i
n
size.
Ke
yw
or
d
s
:
Boo
le
a
n
sat
isfi
abili
ty
(
SA
T
)
Cl
oud
c
om
pu
ti
ng
Energy
op
ti
m
i
zat
ion
In
te
ger
li
ne
ar
pro
gr
am
m
ing
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
B
Y
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Assim
Sag
ahyr
oon,
Dep
a
rtm
ent o
f C
om
pu
te
r
Scie
nce a
nd E
ng
i
ne
erin
g,
Am
erican Univer
s
it
y o
f Sha
r
j
ah
, Uni
ver
sit
y C
it
y,
Sh
a
rj
a
h,
U
ni E
m
irat
Ar
ab
.
Em
a
il
: asagahyroo
n@
a
us
.e
du
1.
INTROD
U
CTION
In
rece
nt
ye
ar
s,
cl
oud
com
pu
ti
ng
has
em
e
rg
e
d
as
a
pr
ac
ti
cal
par
adigm
fo
r
ho
sti
ng
a
nd
deliveri
ng
serv
ic
es
ove
r
the
inter
net.
Th
is
in
turn
has
l
ed
to
the
creati
on
of
huge
da
ta
centers
wh
i
ch
are
powe
r
hungry
e
ntit
ie
s
with
com
plex
requir
e
m
ents
and
ope
rati
on
al
nee
ds
[1
]
.
A
crit
ic
al
desig
n
pa
ram
eter
for
data
cen
te
rs
i
s
their
power co
ns
um
ption cha
racteri
sti
cs an
d ene
rg
y
dem
and
s.
Cl
oud
c
om
pu
ti
ng
ser
vices
a
r
e
prov
i
de
d
t
hro
ugh
datace
nters.
Datace
nt
ers
c
on
ta
in
la
rg
e
s
cal
e,
netw
orke
d
co
m
pu
te
infr
ast
r
uctu
re.
Th
e
m
ajor
physi
cal
resour
ces
in
cl
oud
com
pu
ti
ng
can
be
cl
assif
ie
d
into
three
cat
e
gorie
s:
com
pu
te
-
rel
at
ed
res
ources
,
stora
ge
-
relat
e
d,
a
nd
netw
ork
res
ources.
F
or
the
cl
ou
d
pro
vid
e
rs
to
be
a
ble
to
m
eet
the
con
s
um
ers’
re
qu
est
s
with
high
qual
it
y
of
se
rv
i
ce
(QoS
),
t
he
y
need
t
o
have
la
rg
e
nu
m
ber
s
of
th
ese
resou
rces
with
hi
gh
ca
pa
bili
ti
es.
Hen
c
e,
datace
nte
rs
ho
sti
ng
these
resour
ce
s
en
d
up
consum
ing
sig
nificant
am
ou
nt
s
of
ene
rg
y.
I
n
2010
f
or
e
xa
m
ple,
the
el
ectr
ic
it
y
us
age
of
d
at
acente
rs
ar
ou
nd
the
w
or
ld
re
presented
(1.1
%
-
1.5%
)
of
the
total
wo
rl
dw
i
de
el
ect
rici
ty
us
a
ge
w
hile
in
the
U
nited
Stat
es,
the
datace
nters
us
age
re
pr
e
se
nted
up
to
2.2
%
of
the
total
energy
consum
ed
in
the
US
[
2]
.
This
in
turn
stresses
the n
ee
d f
or
e
ne
rg
y e
ff
ic
ie
nt
da
ta
centers t
o
m
ini
m
iz
e the en
e
rg
y c
os
ts a
nd s
us
ta
in t
he
e
nv
i
ronm
ent.
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
Op
ti
miz
atio
n o
f ener
gy
c
on
s
umptio
n
i
n
cl
ou
d
c
omp
uting
datace
nters
(
Ah
med Os
m
an
)
687
So
m
e o
f
t
he
te
chn
i
qu
e
s that c
an
c
on
t
rib
ute to e
nergy ef
fici
ency in cl
oud c
om
pu
ti
ng
i
nclu
de:
Use
of
sm
art
t
asks
al
locat
ion
strat
egies
includi
ng
sta
ti
c
as
well
as
dynam
ic
app
ro
ac
he
s
to
red
uc
e
en
erg
y
consum
ption
[
3,
4].
Use
of
virt
ualiz
at
ion
te
ch
niques
to
ena
ble
the
creati
on
of
m
ul
ti
ple
virtu
al
reso
urces
on
a
sing
le
physi
cal
resou
rce.
T
his
al
lows
t
he
re
du
ct
io
n
of
the
num
ber
of
physi
cal
resour
ce
s
nee
de
d
to
pe
rfor
m
a
ta
sk
a
nd
hen
ce
can
c
on
t
rib
ute in red
uci
ng the e
ne
rg
y c
on
s
um
ption
[5
]
.
Wor
klo
a
d
Co
nsoli
datio
n
to
m
ini
m
iz
e
the
nu
m
ber
of
res
ources
us
e
d
a
nd
m
axi
m
iz
es
the
util
iz
at
io
n
of
the r
es
ources
[6].
Othe
r
te
ch
niques
inclu
de
dynam
ic
VM
s
placem
ent
and
VM
m
ig
rati
on,
heat
m
anag
em
ent
a
nd
tem
per
at
ur
e
-
a
war
e
all
ocati
on, and
l
oad b
al
a
ncin
g
a
nd task
sche
du
li
ng
[6
]
.
Sche
du
li
ng
is
t
he
process
of
m
aking
decisi
ons
to
al
locat
e
r
eso
ur
ces
t
o
ta
s
ks
ov
e
r
ti
m
e
with
the
go
al
of
o
ptim
izing
on
e o
r
m
or
e
ob
j
ect
ives
[
7].
I
n
cl
oud
com
pu
ti
ng,
sche
duli
ng
addresses
t
he
pro
blem
of
al
locat
ing
netw
ork
res
our
ces
,
c
om
pu
te
r
eso
ur
ces
,
or
stora
ge
nee
ds
to
ta
sk
s
over
a
pe
rio
d
of
ti
m
e,
and
it
enc
om
passes
the foll
owin
g
[
8]:
Re
so
urces:
ph
y
sic
al
/virtual de
vices
with the
abili
ty
to
exec
ute or
process
tasks;
Tasks: i
ns
tr
uctions t
o
be
ex
ec
uted by the
r
es
ources.
Con
st
raints:
co
nd
it
io
ns
m
us
t
be
co
ns
ide
re
d
wh
e
n
sc
heduling
th
e
ta
sk
s
to the
res
ources.
They
m
a
y
be
t
ask
-
base
d,
res
ourc
e
-
base
d,
or
a
com
bin
at
ion
of
these
.
T
he
y
cou
l
d
al
s
o
be
hard
c
on
strai
nts,
m
eaning
const
raints t
hat m
us
t be fu
ll
-
fi
ll
ed
or s
of
t c
on
strai
nts that ca
n be
relaxe
d.
Objecti
ves: t
he
ev
al
uati
on crite
ria that
needs
to b
e
m
easur
ed
in or
der to a
ss
ess the
syst
e
m
p
er
form
ance.
So
m
e
of
the
relat
ed
schedul
ing
ob
j
ect
ives
can
be
m
ini
m
iz
ing
the
com
ple
ti
on
tim
e,
m
axi
m
iz
ing
the
pr
o
fit,
m
ini
m
izing
t
he
e
ne
rg
y
c
onsu
m
ption
,
m
axi
m
iz
e
the
res
ource
util
iz
at
ion
,
or
it
can
hav
e
m
ulti
ple
obj
ect
ives
. Our researc
h
ai
m
s
to f
orm
ulate
the sch
ed
ulin
g
pr
ob
le
m
(
with th
e o
bject
ive
being
t
he
m
ini
m
izati
on
of
power
c
ons
um
ption
)
as
a
0
-
1
IL
P
m
od
el
.
The
ge
ne
rated
ILP
m
od
el
will
be
transfo
rm
ed
into
a
B
oo
le
a
n
sat
isfia
bili
ty
pr
oble
m
(S
AT)
that
can
be
s
olv
e
d
us
in
g
S
AT
s
olv
e
rs
wi
th
the
res
ults
bein
g
m
app
ed
bac
k
to
the
or
i
gin
al
scheduli
ng
pro
bl
e
m
.
W
e
will
va
li
date
the
ILP
m
od
el
s
an
d
exam
ine
the
c
apab
il
it
ie
s
of
c
urrent
sta
te
of
the
a
rt
gen
e
ric
IL
P
a
nd
SAT
so
l
vers
when
us
e
d
to
s
olv
e
f
orm
ul
at
ion
s
of
t
he
s
cheduli
ng
pro
blem
.
Pr
el
im
inary res
ults relat
ed
to
t
his wo
rk w
e
re
repor
te
d
in
[
9].
Op
ti
m
izing
t
he
pow
er
co
nsum
ption
in
a
cl
oud
com
pu
ti
ng
e
nv
ir
on
m
ent
can
be
achieve
d
by
consi
der
i
ng
it
as
a
sche
duli
ng
pro
blem
wh
e
re
t
he
ta
s
ks
s
ubm
i
tt
ed
to
the
datace
nt
er
are
sche
du
le
d
t
o
the
di
ff
e
ren
t
r
eso
ur
ces
in
t
he
datace
nter
i
n
su
c
h
a
m
ann
e
r
to
opti
m
iz
e
the
powe
r
c
ons
um
ed
by
c
ompu
ti
ng
resou
rces.
H
oweve
r,
t
he
sc
he
du
li
ng
pro
ble
m
can
ha
ve
m
ulti
ple
optim
izati
on
s
obj
ect
iv
es
su
c
h
as
incr
easi
ng
the
util
iz
at
ion
of
res
ources,
reducin
g
the
tim
e
ta
ken
to
pe
rfo
rm
the
ta
sk
s,
m
axim
iz
ing
the
pr
of
it
,
or
m
ini
m
iz
ing
the
en
e
rg
y c
onsum
pt
ion
.
In
rece
nt
ye
ars
,
the
sc
hedulin
g
as
pects
of
a
datace
nter
ha
ve
be
e
n
ta
ckled
with
di
ff
e
ren
t
op
ti
m
iz
ation
obj
ect
ives in
m
ind
. For
e
xam
ple, in
[
10
]
, th
e
ta
sk
s o
r proces
ses are
assigne
d
in a r
ound r
obin f
as
hion b
et
ween
the
avail
able
process
or
s
.
T
he
ta
sk
s
are
di
stribu
te
d
eq
ual
ly
between
al
l
the
process
ors;
prob
le
m
is,
so
m
e
p
r
ocess
or
s
will
end
up
bein
g
hea
vily
loaded
com
par
ed
to
oth
e
rs.
T
he
auth
or
s
propos
ed
overc
om
ing
this
lim
it
at
ion
by
usi
ng
a
weig
hte
d
rou
nd
r
obin
appr
oach.
T
he
Mi
n
-
Mi
n
sta
ti
c
loa
d
balanci
ng
al
gorithm
is
us
e
d
in
[1
1
]
.
He
re,
par
am
et
ers
relat
ed
to
the
jo
b
or
ta
s
k
are
ass
um
ed
to
be
kn
own
in
a
dv
a
nc
e.
The
al
go
rith
m
sta
rts
by
ide
ntifyi
ng
the
m
ini
m
u
m
com
pleti
on
tim
e
nee
ded
by
a
ta
sk
a
nd
assig
ning
ta
s
ks
by
the
cl
ou
d
m
anag
er
on
ascen
ding
com
pleti
on
ti
m
e
n
eeds.
A
dra
wback
is
the
fact
that
ta
sk
s
re
quirin
g
lo
ng
ex
ecuti
on
ti
m
es
hav
e
t
o
wait
and
this
m
ay
le
ad
to
sta
rv
at
io
n.
Liu
et
.
al
[1
2
]
pro
po
se
d
an
op
ti
m
iz
ation
m
od
el
fo
r
m
i
nim
iz
ing
the
ene
r
gy
co
nsum
ption
of
cl
oud
c
om
pu
ti
ng
datace
nters
us
in
g
ta
s
k
sc
he
du
li
ng.
T
he
a
uthors
f
or
m
ulate
d
an
integer
pro
gr
a
m
m
ing
m
od
el
of
th
e
ta
sk
sc
he
du
li
ng
pro
blem
with
the
goa
l
of
m
ini
m
iz
ing
the
nu
m
ber
of
act
ive
serv
e
rs
an
d
th
us
m
ini
m
iz
ing
the
overall
energy
con
s
um
ption
of
the
datace
nter
with
out
breakin
g
the
res
ponse
tim
e
con
strai
nt
s.
The
f
or
m
ulati
on
process
consi
der
e
d
tw
o
cases:
an
init
ia
l
case
w
here
no
ta
sk
is
receive
d
befor
e
a
nd
the
new
ta
s
ks
are
assigne
d
im
mediat
el
y
and
a
seco
nd
case
w
her
e
th
e
datac
enter
ha
s
bac
kl
ogged
ta
sk
s
that
res
ul
t
in
qu
eui
ng
delay
s.
In
a
dd
it
ion
,
the
aut
hors
pro
po
se
d
a
scheduli
ng
s
chem
e
na
m
ed
m
os
t
-
eff
ic
ie
nt
-
s
er
ver
first
th
at
assigns
t
he
m
os
t
eff
ic
ie
nt
ser
ve
r
first.
T
he
m
od
el
a
nd
the
e
nv
i
ronm
ent
are
the
n
si
m
ulate
d
with
heter
ogene
ou
s
ta
sk
s.
T
heir
resu
lt
s
s
howe
d
that
the
pro
po
se
d
sc
hem
e
m
ini
m
iz
es
the
ser
ve
r
energy co
ns
um
ption
70
ti
m
es
bette
r
than
a ra
ndom
-
based
sc
hedulin
g
sche
m
e o
n
aver
a
ge.
In
[1
3
]
a sch
e
du
li
ng
al
gorithm
fo
r
on
li
ne
sche
duli
ng
of
ser
vers
of
cl
oud
c
ompu
ti
ng
pro
vid
e
rs
wa
s
pr
opose
d.
T
he
al
gori
thm
is
cal
le
d
pr
e
dict
op
ti
m
iz
e
disp
a
tc
h
(P
O
D
)
bas
ed
on
the
thre
e
m
ai
n
ste
ps
t
hat
the
al
gorithm
fo
ll
ow
s.
It
was
assum
ed
that
the
se
rv
e
rs
us
e
d
hav
e
m
ulti
ple
proce
s
so
rs
wit
h
different
s
pee
ds
and
di
ff
e
ren
t
pow
e
r
consum
ption
,
and
eac
h
se
rv
e
r
ca
n
be
i
n
busy,
idle,
or
s
w
it
ched
off
m
ode.
T
he
obj
ect
i
ve
of
the
sc
he
du
li
ng
al
gorithm
is
t
o
m
ini
m
iz
e
the
ene
rg
y
c
ons
um
ption
of
a
cl
oud
c
om
pu
ti
n
g
en
vir
on
m
ent.
T
he
al
go
rithm
is
i
m
ple
m
ented
in
th
ree
ste
ps
:
first
pre
dict
ion
of
the
w
orkload
assig
ne
d
t
o
the
datace
nter.
T
he
w
orkload
is
acqu
i
red
us
in
g
est
i
m
a
te
s
based
on
the
cl
oud
pr
ovid
er
’s
inf
or
m
at
ion
or
usi
ng
a
ne
w
kind
of
m
ult
i
-
arm
bandit
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.
11
, No
.
1,
Febr
uar
y
2021
:
6
86
-
698
688
work
l
oa
d
pre
di
ct
ion
.
T
he
n,
t
he
al
gorit
hm
m
od
ifie
s
and
op
ti
m
iz
es
the
set
of
avail
abl
e
ser
ver
s
to
be
st
fit
the
pre
dicte
d
work
l
oa
d.
T
he
final
ste
p
is
t
o
sc
he
du
le
t
he
job
s
to
t
he
a
va
il
able
process
or
s
us
i
ng
a
m
od
i
fied
ver
si
on
of
Ro
und
R
obin
sc
he
du
li
ng
m
et
ho
d.
The
a
uthor
s
th
en
sim
ulate
d
their
pro
po
se
d
so
luti
on
a
nd
te
ste
d
it
against
oth
e
r
onli
ne
a
nd
offli
ne
sc
he
du
li
ng
al
gorithm
s
and
presente
d
thei
r
res
ults
with
a
rec
omm
end
at
ion
on
wh
ic
h
ty
pe
of
scheduli
ng
to
us
e
f
or
dif
fe
ren
t
ty
pes
of
work
l
oa
d.
A
nothe
r
te
ch
niqu
e
that
is
us
ed
with
sche
du
li
ng
t
o
m
ini
m
iz
e
the
energy
c
onsum
pt
ion
i
n
cl
ou
d
c
om
pu
ti
ng
e
nv
i
ronm
ent
is
dynam
ic
vo
lt
age
a
nd
fr
e
qu
e
ncy
scal
i
ng
(
DVFS
).
In
[1
4
]
a
DVFS
ba
sed
gr
ee
n
e
ne
rg
y
-
e
ff
ic
ie
nt
sc
hedulin
g
al
go
rithm
was
pro
posed.
It
work
s
on
inc
reas
ing
the
util
iz
at
ion
facto
r
of
the
dif
fer
e
nt
serv
e
rs
in
the
datace
nter
in
a
n
effor
t
to
re
du
ce
it
s
energy
co
nsu
m
pt
ion
.
T
he
e
nv
i
ronm
ent
con
sist
ed
of
hete
roge
neous
se
rvers
with
eac
h
s
erv
e
r
ha
ving
di
ff
ere
nt
process
or,
m
em
or
y
,
and
sto
rag
e
spe
ci
ficat
ion
s
.
S
e
rv
e
r
’s
has
f
reque
nc
y
is
va
ried
or
re
du
ce
d
t
o
m
ini
m
ize
the
ene
rg
y
m
eet
ing
the
re
qu
i
rem
ents
of
the
wor
klo
a
d.
Ne
xt,
a
pr
io
rit
y
-
ba
sed
sc
hedulin
g
al
gorithm
is
us
e
d
to
assign
j
obs
to
create
d
VMs
.
E
xperim
ental
resu
lt
s
ind
ic
at
ed
that
the
pro
po
s
ed
al
gorith
m
red
uce
the
e
nergy
consum
ption
by
a
r
an
ge
th
at
var
ie
s
from
5
to
25%
w
hile
m
ai
ntaini
ng
th
e
re
qu
ire
d
pe
rfor
m
ance
a
nd
execu
ti
on ti
m
e
.
Var
i
ou
s
com
bin
at
ion
of
DV
F
S
an
d
node
vary
-
on
va
ry
-
off
(
VOV
O)
m
et
hods
wer
e
e
valua
te
d
in
[1
5
]
.
The
ob
j
ect
ive
of
the
resear
c
h
was
t
o
m
inim
iz
e
the
po
we
r
co
nsum
ption
in
cl
us
te
rs
of
serv
e
rs
du
rin
g
tim
es
wh
e
re
the
w
orkl
oad
is
re
duced.
T
he
a
uthors
ex
plored
f
ive
dif
fe
ren
t
po
li
ci
es
wit
h
diff
e
re
nt
de
grees
of
com
plexity
at th
e im
ple
m
enta
ti
on
level
. T
he five
poli
ci
es ar
e:
Inde
pende
nt
volt
age
scal
in
g
(IVS)
:
us
es
process
or
s
t
ha
t
ha
ve
volt
age
scal
in
g
pro
per
ty
an
d
c
an
ind
e
pende
ntly
ch
an
ge
t
heir v
oltage a
nd fre
quency acc
ordi
ng to
their
work
load.
Coordi
nated
volt
age
scal
in
g
(CVS):
ex
pl
oits
volt
age
s
cal
ed
proces
sors
but
co
ordi
na
te
their
volt
age
scal
ing
al
gorithm
s.
Var
y
-
on
var
y
-
off
(V
OVO
): turnin
g
t
he who
le
ser
ve
r o
n or
off depe
ndin
g on the
siz
e of t
he wor
klo
a
d.
Com
bin
at
ion
of IVS
and
V
O
VO.
Com
bin
at
ion
of CV
S a
nd VO
VO.
These
dif
fer
e
nt
poli
ci
es
were
stu
died
an
d
com
par
ed
with
a
c
om
bin
at
ion
of
VOV
O
an
d
C
V
S
pro
du
ci
ng the
m
os
t saving
s
in
e
n
er
gy c
on
s
um
pt
ion
.
In
[
16
]
a
uthor
s
de
fine
d
a
n
e
nergy
co
nsum
ption
rati
o
(E
CR
)
to
e
valuat
e
the
ef
fici
enc
y
of
dif
fer
e
nt
fr
e
qu
e
ncies
under
wh
ic
h
to
e
xecu
te
a
ta
sk
.
They
co
nverte
d
the
e
nergy
-
ef
fici
ent
ta
sk
sc
he
du
li
ng
pro
blem
into
m
ini
m
iz
ing
th
e
total
ECR
.
Nex
t
t
hey
transfor
m
ed
the
pro
blem
to
the
var
ia
ble
siz
e
bin
pack
i
ng,
an
d
s
howe
d
that
the
m
ini
m
iz
at
ion
of
EC
R
is
NP
-
ha
rd
and
propose
d
pro
po
se
a
ta
sk
al
locat
ion
and
scheduli
ng
m
et
hods
base
d
on
the
fe
at
ur
e
of
this
prob
le
m
.
They
also
pro
posed
a
process
or
-
le
vel
m
igrat
ion
al
go
rithm
to
resch
edul
e
rem
ai
nin
g
ta
sks
a
m
on
g
proce
sso
r
s
on
a
n
in
di
vid
ual
se
rv
e
r
and
dynam
ic
ally
balance
the
work
l
oa
ds
an
d
lower
the
total
ECR
on
a
giv
e
n
s
erv
e
r
.
In
[17],
researc
hers
pro
pose
a
co
nsoli
dation
al
gor
it
h
m
wh
ic
h
fa
vors
the
m
os
t
eff
ec
ti
ve
m
i
gr
at
ion
am
on
g
Virt
ua
l
Ma
chines,
con
ta
ine
rs
a
nd
app
li
cat
io
ns
;
and
i
nvest
igate
how
m
igrati
on
deci
sion
s
sho
uld b
e m
ade to
sav
e
en
e
rg
y
with
out any ne
gative
im
pact o
n
the
s
erv
ic
e
perf
or
m
ance.
Linear
pro
gr
a
m
m
ing
is
a
m
at
hem
atical
m
et
hod
for
m
axi
m
izing
or
m
i
nim
i
zi
ng
a
li
ne
ar
obj
ect
ive
functi
on
unde
r
certai
n
co
nd
it
i
on
s
or
co
ns
trai
nts
co
ns
ist
ing
of
diff
e
re
nt
va
riables.
I
nte
ger
li
near
pro
gr
a
m
m
ing
(I
L
P)
is
a
sp
ec
ia
l
ty
pe
of
li
near
pro
gr
am
m
ing
w
her
e
s
om
e
or
al
l
the
va
ria
bles
in
the
obje
ct
ive
functi
on
and
const
raints
a
re
lim
i
te
d
to
int
eger
val
ues.
I
LP
pr
ob
le
m
s
can
be
cat
eg
ori
zed
i
nto
t
wo
cat
egories:
0
-
1
I
LP
pro
blem
s
wh
er
e
the
var
ia
bles
in
the
pro
blem
are
bin
a
ry
vari
ables
an
d
ha
ve
value
s
of
0
or
1,
a
nd
Gen
e
ri
c
ILP
wh
e
re t
he vari
ables are
not li
m
it
ed
to b
i
nar
y
v
a
riables
[
1
8
].
In
t
his
w
ork,
t
he
ge
ne
rated
I
LP
m
od
el
will
be
tra
ns
f
or
m
ed
into
a
bo
olean
sat
isfia
bili
ty
prob
le
m
(S
A
T)
eq
uatio
ns
.
T
hese
eq
ua
ti
on
s
will
be
so
lve
d
us
in
g
S
AT
so
l
ver
s
an
d
the
resu
lt
wi
ll
be
m
app
ed
back
to
the
or
i
gin
al
sc
hedulin
g
pro
bl
e
m
.
Boo
le
an
s
at
isfia
b
il
it
y
pr
ob
le
m
ref
ers
t
o
fin
ding
a
sat
isfyi
ng
assi
gnm
ent
fo
r
a
pro
blem
or
pro
ving
that
none
exist.
A
sat
i
sfying
a
ssig
nme
nt
f
or
a
ce
rtai
n
Bo
olean
for
m
ula
is
an
assignm
ent
of
t
he
bin
a
ry
var
ia
bles
in
t
he
form
ula
to
0
or
1
to
m
ake
the
ove
rall
for
m
ula
evaluate
to
1.
T
he
assi
gnm
ent
m
us
t
sat
isfy
all
the
co
ns
trai
nts
of
the
pro
ble
m
s
i
m
ultaneousl
y
[1
9
]
.
Bot
h
the
S
AT
pro
ble
m
and
t
he
0
-
1
I
LP
pro
blem
are
rep
rese
nted
usi
ng
bi
nar
y
vari
ables
but
in
SA
T,
the
Bo
olean
f
or
m
ulas
us
e
d
t
o
represen
t
the
pr
ob
le
m
and
it
s
co
ns
trai
nt
s
m
us
t
al
l
be
in
a
pro
duct
-
of
-
s
um
s
fo
rm
or
a
c
onju
nctive
norm
al
fo
rm
(
CNF)
and
t
he
ope
rator
s
us
e
d
in
t
he
se
form
s
are
log
ic
al
operato
rs
(AND,
OR
,
a
nd
N
OT)
w
hile
in
0
-
1
ILP
prob
le
m
the
c
onstrai
nts
are
sim
ple
in
equ
al
it
ie
s
of
t
he
f
or
m
≤
wh
e
r
e
∈
ℤ
,
∈
ℤ
×
ℤ
,
a
nd
∈
{
0
,
1
}
.
These
c
onstrai
nts
are
ref
e
rr
e
d
to
as
Pse
udo
Boo
le
a
n
(
PB)
const
raints
a
nd
they
a
re
co
nsi
der
e
d
as
a
ge
ner
al
form
o
f
the
CN
F constrai
nts [
20
].
The
SAT
pro
blem
is
con
sidere
d
to
be
NP
com
plete
pr
o
blem
.
The
so
lvin
g
tim
e
increases
e
x
p
o
n
e
n
t
i
a
l
l
y
w
i
t
h
t
h
e
i
n
c
r
e
a
s
e
i
n
t
h
e
n
u
m
b
e
r
o
f
v
a
r
i
a
b
l
e
s
.
T
h
e
S
A
T
s
o
l
v
e
r
s
a
r
e
a
l
g
o
r
i
t
h
m
s
t
h
a
t
d
e
t
e
r
m
i
n
e
i
f
a
S
A
T
p
r
o
b
l
e
m
i
s
s
a
t
i
s
f
i
a
b
l
e
o
r
n
o
t
a
n
d
f
i
n
d
a
s
a
t
i
s
f
y
i
n
g
a
s
s
i
g
n
m
e
n
t
i
n
c
a
s
e
t
h
e
p
r
o
b
l
e
m
w
a
s
s
a
t
i
s
f
i
a
b
l
e
[
1
9
-
21
].
In
t
he
pa
st
fe
w
ye
ars
t
he
us
e
of
SAT
m
od
el
s
to
re
pr
ese
nt
diff
e
ren
t
pro
blem
s
in
var
io
us
do
m
ai
ns
in
creased
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
Op
ti
miz
atio
n o
f ener
gy
c
on
s
umptio
n
i
n
cl
ou
d
c
omp
uting
datace
nters
(
Ah
med Os
m
an
)
689
rap
i
dly
du
e
t
o
the
huge
im
prov
em
ent
in
the
SA
T
s
olv
i
ng
al
gorithm
s
and
the
avail
abili
ty
of
bette
r
an
d
m
or
e
powe
rful
com
pu
ti
ng
ca
pa
bili
ti
es.
On
e
of
th
e
gr
eat
im
pr
ov
e
m
ents
to
SAT
-
ba
sed
s
ol
vers
is
the
exte
nsi
on
t
o
handle
PB
co
nst
raints
beca
use
they
are
easi
er
to
w
rite
and
unde
rstan
d
an
d
can
re
place
an
expo
nen
ti
al
num
ber
of
C
NF
c
onstr
ai
nts.
Be
f
or
e
this
exte
ns
io
n
S
AT
m
od
el
s w
e
re
li
m
it
ed
to
de
ci
sion
p
r
oble
m
s
b
ut
with
the h
el
p
of
PB
con
st
raints
SA
T
can
now
m
od
el
op
ti
m
i
zat
ion
pro
ble
m
s
and
so
l
ve
them
us
ing
the
diff
e
ren
t
S
AT
-
base
d
PB so
l
ver
s
[
20
].
Pseudo
Bo
olea
n
S
AT
S
olv
e
r
s:
w
hich
a
re
ba
sic
al
ly
SA
T
so
lve
rs
t
hat
ca
n
handle
ps
e
udo
Bo
olean
const
raints
al
lowi
ng
t
he
S
A
T
so
lve
rs
to
be
us
ed
i
n
so
l
vi
ng
op
ti
m
iz
a
tio
n
pro
blem
s.
In
[
22
]
,
f
or
e
xam
ple,
two
powe
r
est
i
m
ation
prob
le
m
s
wer
e
form
ulate
d
as
SA
T
pro
blem
s
and
evaluate
d
us
ing
ge
ner
ic
IL
P
a
nd
SA
T
-
bas
ed
s
olv
ers
.
It
was
s
how
n
that
the
S
AT
base
d
s
olve
r
perf
or
m
ed
be
tt
er
than
ra
ndom
based
appr
oach
e
s
in term
s o
f fin
ding the
opti
m
al
v
al
ues
of t
he
prop
os
ed
pr
oblem
.
In
t
his
w
ork
,
PB
-
S
AT
s
olv
e
rs
are
us
e
d
t
o
so
lve
the
pro
po
s
ed
optim
izati
on
m
od
el
.
The
te
sti
ng
process
in
volv
es
us
in
g
dif
fere
nt
so
lve
rs
an
d
com
par
in
g
their
res
ults.
T
he
so
l
ver
s
a
re
sel
ect
ed
accor
ding
t
o
the foll
owin
g
c
rite
ria:
A prove
n pe
rfo
rm
ance o
f
t
he
s
olv
e
r
f
r
om
the r
ep
or
te
d
li
te
ratur
e
.
The
a
bili
ty
o
f
t
he
s
olv
e
r
to
h
a
nd
le
la
rg
e
prob
lem
size.
Ba
sed on t
he
a
bove
c
rite
ria,
we
ide
ntifie
d t
hr
ee
PB
-
SA
T
s
olv
e
rs,
nam
el
y:
NaPS
[2
3
]
:
N
ago
ya
Pse
udo
Boo
le
a
n
So
l
ve
r
,
is
a
so
lv
er
f
or
PB
c
onstrai
nts
that
are
li
near
a
nd
c
onta
i
n
bin
a
ry
var
ia
ble
s
on
ly
.
It
was
the
best
pe
rform
ing
so
lve
r
in
the
Pseudo
B
oo
le
a
n
com
petit
ion
2016
in
bot
h
cat
egories
of
op
ti
m
iz
ation
with
sm
all
int
eger
s
a
nd
la
r
ge
integers
an
d
was
a
m
on
g
t
he
best
so
l
vers
in
the o
t
her cat
eg
or
ie
s.
T
he
in
form
ation
a
nd r
e
su
lt
s of t
he
c
om
pet
it
ion
is av
ai
la
ble in [2
4
].
Mi
nisat+
[2
5
]
:
is
a
m
od
ifie
d
ver
si
on
of
t
he
Mi
nisat
so
lv
er
to
be
a
ble
to
deal
with
pse
udo
B
oo
le
a
n
const
raints,
a
nd
des
pit
e
bein
g
a
si
m
ple
so
lver
it
per
f
orm
ed
well
in
t
he
ps
e
udo
Bo
olean
com
petit
ion
thr
ough the
yea
rs.
Sat4j
[
2
6
]
:
is
a
Boo
le
a
n
reasonin
g
j
ava
li
br
a
ry
that
is
us
ed
to
so
l
ve
B
oo
le
an
sat
isfact
io
n
and
opti
m
iz
at
i
on
pro
blem
s.
The
sat
4j
so
l
ver
is
diff
e
ren
t
f
rom
the
oth
e
r
two
s
olv
e
rs
be
cause
it
do
es
no
t
pri
ori
ti
ze
bein
g
the
fastest
i
n
s
olv
in
g
a
pr
ob
le
m
.
Ho
we
ver,
t
he
s
olv
e
r
is
fu
l
l
featur
e
d,
r
ob
us
t,
a
nd
use
r
fri
end
ly
.
T
he
s
ol
ver
desig
n fo
ll
ows
the Ja
va desig
n g
uid
el
ines
and
code c
onve
ntion
s
.
Fu
rt
her
m
or
e,
Com
m
ercial
iz
e
d
an
d
ope
n
so
urce
opti
m
i
zat
ion
to
ols
with
va
rio
us
e
m
bed
de
d
al
gorithm
s
exist
tod
ay
.
Th
ose
too
ls
deal
wi
th
ge
ne
rati
ng
a
so
luti
on
t
o
a
def
i
ned
li
near,
non
-
li
near
,
or
i
ntege
r
m
od
el
s.
So
m
e o
f
the
popula
r
t
oo
ls
inclu
de:
CPLEX
[2
7
]
:
is
a
hig
h
-
pe
rfor
m
ance
so
l
ver
for
l
inear
program
ing
,
m
ixed
intege
r
program
ing
,
and
qu
a
drat
ic
pro
gram
ing
.
It
is
a
well
-
kn
own
a
nd
wildly
us
ed
c
omm
ercial
so
lver
de
velo
ped
ov
e
r
20
ye
a
rs
a
go
and
is
c
urre
ntly
owned
by
I
BM
.
The
s
olve
r
us
e
s
the
Si
m
plex
exact
al
gorithm
,
and
t
he
nam
e
CPLEX
i
s
act
uall
y
ta
ken
from
C
-
Si
m
pl
ex
w
hich
is
Si
m
plex
al
go
rith
m
i
m
ple
m
ente
d
in
C
la
ngua
ge.
T
he
so
l
ver
is
i
m
pr
oved
a
nd
updated
pe
riod
ic
al
ly
to
in
crease
it
s
perf
or
m
ance
an
d
include
dif
fer
e
nt
al
gorithm
s
for
diff
e
re
nt
ty
pes
of
pr
ob
le
m
s.
so
m
e
of
the
va
r
iou
s
m
et
hods
us
e
d
in
C
PLE
X
ar
e:
pri
m
a
l
si
m
plex
al
gorithm
,
du
al
sim
plex
al
gorithm
,
netw
ork
sim
plex
al
gorithm
,
and
t
he
barrier
m
eth
od.
T
his
s
olve
r
will
be
us
e
d
in
te
sti
ng
the
pr
opose
d m
od
el
.
LING
O
[2
8
]
:
is
an
ef
fici
ent
com
pr
e
hensi
ve
to
ol
us
e
d
in
bu
il
di
ng
an
d
s
olv
i
ng
dif
fer
e
nt
ty
p
es
of
op
ti
m
iz
ation
m
od
el
s.
It
was
dev
el
op
e
d
by
LIND
O
Syst
em
s
as
an
op
ti
m
iz
at
ion
m
od
el
ing
softwa
re
f
or
li
near
,
no
nline
ar,
an
d
intege
r
pro
gr
am
m
in
g.
LI
N
GO
prov
i
des
it
s
us
e
rs
with
a
la
ngua
ge
to
ex
pr
ess
the
opti
m
iz
at
i
on
m
od
el
s,
an
env
i
ronm
ent
to
buil
d
a
nd
e
dit
their
prob
l
e
m
,
and
a
set
of
s
olve
rs.
Th
ese
so
lve
rs
inclu
de
bu
t
no
t
lim
it
e
d
to:
Pr
im
a
l
and
dual
si
m
plex
so
lvers
,
Ba
rr
ie
r
so
lve
r,
I
ntege
r
so
lve
r,
Ge
neral
nonlinea
r
s
olv
e
r,
a
nd
Qu
a
dr
at
i
c so
l
ver
.
Th
e
re
st
of
t
he
pap
e
r
is
or
ga
nized
as
fo
ll
ows
:
in
Sect
io
n
2
we
disc
us
s
the
resea
rch
m
et
hodo
l
og
y
fo
ll
owe
d
in
this
work
;
Sect
io
n
3
inclu
des
the
res
ults
ob
ta
ined
an
d
a
thoro
ugh
disc
us
si
on
of
it
,
the
pa
per
is
con
cl
ud
e
d
i
n
S
ect
ion
4.
2.
RESEA
R
CH MET
HO
D
2.1.
En
vironm
ent
s
et
ting a
nd
m
od
el
fo
r
mulati
on
T
his
sect
io
n
presents
t
h
e
cha
racteri
sti
cs
of
the
c
loud
c
om
pu
ti
ng
en
vir
onm
e
nt,
a
naly
ses
the
sche
d
uling
pro
blem
in
the
cl
oud
da
ta
center,
a
nd
f
or
m
ulate
s
the
sc
he
du
li
ng
pro
blem
as
an
intege
r
li
nea
r
pro
gr
am
(I
LP
) m
od
el
.
A
cl
oud
c
om
pu
ti
ng
e
nviro
nm
ent
m
ay
consi
st
of
a
sin
gle
or
m
ulti
pl
e
datace
ntres
that
m
ay
belon
g
to
a
sing
le
or
m
ulti
ple
pro
vid
e
rs.
T
his
w
ork
deals
with
a
sing
le
cl
ou
d
da
ta
centre
that
belo
ngs
to
a
s
ing
le
pro
vid
er
.
T
he
cl
oud
datace
nt
re
in
fr
ast
ru
ct
ure
co
ntains
t
hr
e
e
res
ource
ty
pe
s:
com
pu
te
,
s
tora
ge,
a
nd
ne
twork
resou
rces.
O
ur
aim
in
the
wor
k
pr
ese
nted
he
re
is
to
f
oc
us
on
the
c
om
pu
ti
ng
c
om
po
ne
nt
of
the
res
ources
us
e
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.
11
, No
.
1,
Febr
uar
y
2021
:
6
86
-
698
690
since
it
is
sh
own
t
o
c
on
s
ume
about
26
%
of
the
total
e
nergy
of
a
datace
ntre
[
17]
.
T
he
m
ai
n
com
pu
ti
ng
unit
s
are
the
server
s.
Th
e
prop
os
e
d m
od
el
is form
ulate
d
bas
e
d o
n t
he
f
ollo
wing a
ssu
m
ption
s:
Each
CP
U
can
run
two
or
four
ta
s
ks
at
the
sam
e
tim
e
dep
end
i
ng
on
num
ber
of
c
or
es
it
con
ta
in
s
(
du
al
or
qu
a
d
c
ore); eac
h
c
or
e
w
il
l r
un
on
ly
on
e
task
a
t 100%
u
ti
li
zat
ion
.
A
ser
ver
is
tur
ned
ON
if
at
le
ast
a
sing
le
core
in
on
e
of
it
s
CPUs
is
exec
ut
ing
a
ta
sk
an
d
is
switc
hed
O
FF
oth
e
rw
ise
.
A
co
re
is
ei
ther
O
N
(s
witc
he
d
on
an
d
e
xe
cuting
a
ta
s
k),
idle
(sw
it
che
d
on
but
not
r
unning
a
ny
ta
sk
)
,
or com
plete
ly
OF
F
(
s
witc
he
d o
ff).
Each c
or
e
is as
su
m
ed
to
hav
e
two r
unning m
od
e
s: Hi
gh and
Lo
w.
In
High
m
od
e,
the
co
re
r
uns
with
the
m
axim
u
m
avail
able
fr
e
qu
e
ncy
a
nd
volt
age
,
w
hi
le
in
Lo
w
m
od
e,
the core
runs
at the m
ini
m
u
m
avail
able f
re
quency an
d v
oltage.
Cores
inside
th
e sam
e serv
er c
an ru
n
i
n diff
e
r
ent m
od
es.
Tasks
are
al
lo
c
at
ed
in
batc
hes
tha
t ar
rive
at d
ist
inct t
i
m
es (
t
0
, t
1
, …
, t
n
).
All
the
ta
sk
s
in
a
sp
eci
fic
ba
tc
h
will
be
scheduled
im
m
ediat
el
y
and
will
sta
rt
ru
nn
ing
sim
ultaneou
sl
y
(no
ta
s
k wil
l b
e assig
ned to
sta
rt lat
er)
.
A
ta
s
k
ca
n on
l
y be e
xecu
te
d by a c
or
e
that
m
eet
s it
s f
requ
ency an
d
m
e
m
or
y
requirem
ents.
Fo
r
a
ny task t
he
re is at le
ast
one c
or
e
that m
e
et
s the r
e
quire
m
ents o
f
the ta
sk
(no
ta
s
k wil
l be
re
j
ect
ed
).
The
e
xecu
ti
on
of
a
ta
s
k
on
a
certai
n
co
re
is
assum
ed
to
co
ns
um
e
a
sp
eci
f
ic
a
m
ou
nt
of
e
nergy
that
will
be
cal
le
d
the e
nergy co
s
t i
n
t
his
m
od
el
.
Ne pre
-
em
ption
of tasks
is all
ow
e
d.
2.2.
Op
timi
z
at
ion
mo
del f
or
mulati
on
This
sect
ion
di
scusses
the
intege
r
li
near
pro
gr
am
m
ing
(I
L
P)
m
od
el
that
m
ini
m
iz
es
the
energy
consum
ption
of
the
ser
ve
rs
in
a
cl
oud
datace
ntre.
T
he
inte
nded
obj
ect
ive
i
s
to
fin
d
the
s
pe
ci
fic
assignm
ent
of
ta
sk
s
to
c
or
es
that
gi
ves
t
he
m
ini
m
u
m
ener
gy
val
ue.
The
t
otal
powe
r
c
on
su
m
ption
is
t
he
powe
r
c
onsum
ed
by
the
co
res
w
he
n
a
ta
s
k
is
al
locat
ed
to
it
plu
s
t
he
powe
r
co
nsum
ed
by
the
idle
c
ores.
Ta
ble
1
present
s
the
var
ia
bl
es
us
e
d
in
dev
el
op
i
ng
the
m
od
el
and
thei
r
m
eaning
.
Fi
gure
1
il
lustrate
s
a
si
m
ple
datace
ntr
e
config
ur
at
io
n
with
two
ser
ve
rs
S
1
and
S
2
.
S
1
has
two
dual
-
c
or
e
CPU
s
and
S
2
has
one
quad
-
co
re
CPU
as
sh
ow
n.
A
ser
ve
r
has
t
wo
sta
te
s
ei
ther
O
N
or
OF
F
.
Dep
e
nd
ing
on
t
h
e
se
r
ve
r
sta
te
,
ea
ch
c
or
e
is
ass
um
ed
to
be
ei
ther
act
ive,
i
dle,
or
switc
he
d
off,
a
s
repres
ented
by
t
he
green
,
blu
e,
an
d
re
d
c
oloure
d
-
s
qu
a
res
res
pecti
vely
.
In
the
sce
na
rio
dep
ic
te
d
by
Figure
1,
al
l
cor
e
s
wer
e
init
ia
ll
y
of
f
,
the
n
two
t
asks
are
assi
gned
to
C
1
,1,1
and
C
1,2,1
with a c
os
t
of
X
1,1,1,1
and X
1,2
,1,2
resp
ect
ively
.
T
a
bl
e
1.
M
od
e
l
not
a
t
i
on
s
V
a
r
i
a
b
le
s
D
e
s
cr
i
pt
i
o
n
B
i
n
a
ry
v
a
r
ia
b
l
e
e
q
u
a
l
s
t
o
0
i
f
s
er
v
er
i
i
s
o
ff
,
a
n
d
1
i
f
i
t
i
s
o
n
.
,
B
i
n
a
ry
v
a
r
ia
b
l
e
e
q
u
a
l
s
t
o
0
i
f
C
PU
j
in
s
e
r
v
er
i
i
s
n
o
t
a
c
ti
v
e
a
n
d
1
o
t
h
e
rw
i
se
.
,
B
i
n
a
ry
v
a
r
ia
b
l
e,
e
q
u
a
l
t
o
1
i
f
C
PU
j
i
n
s
e
r
v
e
r
i
i
s
i
d
le
a
n
d
0
o
t
h
er
w
i
se
.
,
,
B
i
n
a
ry
v
a
r
ia
b
l
e,
e
q
u
a
l
t
o
1
i
f
c
o
r
e
k
i
n
C
P
U
j
i
n
s
er
v
e
r
i
i
s
a
c
ti
v
e
,
a
n
d
0
o
t
h
e
rw
i
s
e.
,
,
B
i
n
a
ry
v
a
r
ia
b
l
e,
e
q
u
a
l
t
o
1
i
f
c
o
r
e
k
i
n
C
P
U
j
i
n
s
er
v
e
r
i
i
s
i
d
l
e,
a
n
d
0
o
t
h
er
w
is
e
.
,
,
,
B
i
n
a
ry
v
a
r
ia
b
l
e,
e
q
u
a
l
t
o
1
i
f
ta
s
k
n
is
a
s
s
i
g
n
e
d
t
o
c
o
re
k
in
C
P
U
j
i
n
s
er
v
er
i,
a
n
d
0
o
t
h
e
rw
i
s
e.
,
,
,
B
i
n
a
ry
v
a
r
ia
b
l
e,
e
q
u
a
l
t
o
1
i
f
ta
s
k
n
is
e
x
e
c
u
t
e
d
i
n
c
o
r
e
k
in
C
P
U
j
i
n
s
er
v
er
i
i
n
L
o
w
m
o
d
e
a
n
d
0
o
t
h
e
rw
i
s
e.
,
,
,
B
i
n
a
ry
v
a
r
ia
b
l
e,
e
q
u
a
l
t
o
1
i
f
ta
s
k
n
is
e
x
e
c
u
t
e
d
i
n
c
o
r
e
k
in
C
P
U
j
i
n
s
er
v
er
i
i
n
H
i
g
h
m
o
d
e
a
n
d
0
o
t
h
e
rw
i
s
e.
,
,
,
T
h
e
e
n
e
r
g
y
c
o
s
t
v
al
u
e
o
f
a
s
s
i
g
n
i
n
g
t
a
sk
n
t
o
c
o
r
e
k
i
n
CP
U
j
i
n
se
r
v
er
i
i
n
Lo
w
m
o
d
e
.
,
,
,
T
h
e
e
n
e
r
g
y
c
o
s
t
v
al
u
e
o
f
a
s
s
i
g
n
i
n
g
t
a
sk
n
t
o
c
o
r
e
k
i
n
CPU
j
i
n
se
r
v
er
i
i
n
H
ig
h
m
o
d
e
.
,
,
T
h
e
e
n
e
r
g
y
c
o
s
t
v
al
u
e
o
f
c
o
r
e
k
i
n
C
PU
j
i
n
se
r
v
er
i
w
h
en
i
t
is
i
d
le
.
T
h
e
m
i
n
i
m
u
m
a
m
o
u
n
t
o
f
R
AM
r
e
q
u
e
st
e
d
b
y
ta
s
k
n
.
T
h
e
m
i
n
i
m
u
m
a
m
o
u
n
t
o
f
F
r
e
q
u
e
n
c
y
r
e
q
u
e
s
t
e
d
b
y
t
a
s
k
n
.
,
,
T
h
e
a
v
a
i
la
b
l
e
a
m
o
u
n
t
o
f
R
AM
i
n
c
o
re
k
i
n
CP
U
j
i
n
s
erv
e
r
i
.
,
,
T
h
e
a
v
a
i
la
b
l
e
L
o
w f
r
e
q
u
e
n
c
y
i
n
c
o
re
k
i
n
C
PU
j
i
n
s
er
v
e
r
i
.
,
,
T
h
e
a
v
a
i
la
b
l
e
H
i
g
h
fr
e
q
u
e
n
cy
i
n
c
o
r
e
k
i
n
CP
U
j
i
n
s
er
v
er
i
.
We
de
s
c
r
i
be
t
he
f
or
m
ul
at
i
on
s
t
e
ps
t
a
ke
n
t
o
de
ve
l
op
t
he
c
om
pre
he
ns
i
ve
m
od
e
l
.
I
ni
t
i
a
l
ly
,
t
he
f
or
m
ul
a
t
i
on
c
om
m
e
nc
e
s
w
i
t
h
a
n
ob
je
c
t
i
ve
f
un
c
t
i
on
w
h
os
e
p
r
i
m
ar
y
go
a
l
i
s
t
o
m
i
ni
m
iz
e
t
he
e
ne
r
gy
c
on
s
um
pt
i
on
w
he
n
a
s
s
i
g
ni
n
g
t
a
s
ks
t
o
c
or
e
s
i
n
a
da
t
a
c
en
t
e
r
.
F
r
om
F
ig
ur
e
1
w
e
o
b
s
e
r
ve
t
ha
t
t
he
e
ne
r
gy
c
on
s
um
pt
i
on
c
om
e
s
f
r
om
t
he
e
ne
r
gy
c
os
t
of
a
c
t
i
ve
c
or
e
s
.
T
hi
s
e
ne
r
gy
c
os
t
c
on
s
i
s
t
s
of
t
he
c
os
t
of
c
or
e
s
r
un
ni
ng
i
n
L
ow
m
od
e
(
,
,
,
)
a
n
d
t
he
c
os
t
of
c
or
e
s
r
u
nn
i
n
g
i
n
hi
gh
m
od
e
(
,
,
,
)
.
I
n
a
d
di
t
i
on
t
o
a
c
t
i
ve
c
or
e
s
t
he
r
e
i
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
Op
ti
miz
atio
n o
f ener
gy
c
on
s
umptio
n
i
n
cl
ou
d
c
omp
uting
datace
nters
(
Ah
med Os
m
an
)
691
a
l
s
o
t
he
e
ne
r
gy
c
os
t
o
f
i
dl
e
c
o
r
e
s
(
,
,
,
)
.
F
ur
t
he
r
m
or
e
,
t
he
of
f
c
or
e
s
a
r
e
a
s
s
um
e
d
t
o
ha
ve
z
e
r
o
e
ne
r
gy
ne
e
ds
.
T
hi
s
i
s
r
e
pr
e
s
e
nt
e
d
a
s
f
ol
l
ow
s
i
n
(
1
)
:
(
∑
∑
∑
∑
,
,
,
,
,
,
=
1
+
=
1
=
1
=
1
∑
∑
∑
∑
,
,
,
,
,
,
=
1
+
=
1
=
1
=
1
∑
∑
∑
,
,
,
,
=
1
=
1
=
1
)
(
1
)
T
he
ob
je
c
t
i
ve
f
un
c
t
i
on
de
s
c
r
i
be
d
i
n
(
1)
ha
s
t
hr
e
e
pa
r
t
s
,
t
he
f
i
r
s
t
t
w
o
p
a
r
t
s
r
e
pr
e
s
e
nt
t
he
e
ne
r
gy
c
on
s
um
e
d
by
t
he
a
c
t
i
ve
c
or
e
s
.
T
he
a
c
t
i
ve
c
or
e
s
e
ne
r
gy
i
s
c
a
l
c
ul
at
e
d
by
s
um
m
i
ng
t
he
e
ne
r
gy
c
os
t
of
a
l
l
t
he
c
or
e
s
t
ha
t
a
r
e
r
un
ni
ng
a
t
a
s
k
a
c
r
os
s
t
he
w
ho
l
e
da
t
a
c
e
nt
e
r
.
T
he
t
hi
r
d
c
om
po
ne
nt
of
t
he
ob
je
c
t
i
ve
f
u
nc
t
i
on
s
um
s
t
he
e
ne
r
g
y
c
on
s
um
e
d
by
t
he
r
e
m
a
i
ni
ng
i
dl
e
c
or
e
s
i
n
t
h
e
da
t
a
c
e
nt
e
r
.
T
hi
s
ob
je
c
t
i
ve
f
un
c
t
i
on
i
s
s
ub
j
e
c
t
t
o
t
he
f
ol
l
ow
i
ng
s
e
t
of
c
on
s
t
r
a
i
nt
s
:
I
f
a
c
o
r
e
k
i
s
a
c
t
i
ve
,
t
he
n
t
he
C
P
U
j
t
ha
t
c
on
t
a
i
ns
t
he
c
or
e
k
i
s
c
on
s
i
de
r
e
d
a
c
t
i
ve
a
s
w
e
l
l
.
,
=
1
∑
,
,
≥
1
=
1
∀
∀
I
f
a
C
P
U
j
i
s
a
c
t
i
ve
,
t
he
n
a
t
l
e
as
t
on
e
o
f
i
t
s
c
or
e
s
i
s
a
c
ti
ve
.
∑
,
,
≥
1
,
=
1
=
1
∀
,
∀
I
f
a
C
P
U
i
s
a
c
t
i
ve
,
a
l
l
t
he
c
or
e
s
i
ns
i
de
t
he
C
P
U
m
us
t
be
e
it
he
r
a
c
t
i
ve
or
i
dl
e
bu
t
n
ot
s
w
i
t
c
he
d
of
f
.
∏
(
,
,
+
,
,
)
=
1
=
1
,
=
1
∀
,
∀
A
c
or
e
c
a
nn
ot
be
a
c
t
i
v
e
a
nd
i
dl
e
a
t
t
he
s
am
e
t
im
e
.
,
,
+
,
,
≤
1
∀
,
∀
,
∀
I
f
a
C
P
U
i
s
i
dl
e
,
t
he
n
a
l
l
i
t
s
co
r
e
s
m
us
t
be
i
dl
e
.
∑
,
,
=
,
=
1
=
1
∀
,
∀
I
f
a
s
e
r
ve
r
i
s
of
f
a
l
l
t
he
c
or
e
s
i
ns
i
de
t
he
s
e
r
ve
r
a
r
e
o
f
f
.
∑
∑
,
,
+
,
,
=
0
=
1
=
1
=
0
∀
I
f
o
ne
o
r
m
or
e
of
t
he
C
P
U
s
i
n
s
e
r
ve
r
i
a
r
e
a
c
t
i
ve
,
t
he
n
s
e
r
ve
r
i
m
us
t
be
a
c
ti
ve
.
=
1
∑
,
≥
1
=
1
∀
I
f
a
s
e
r
ve
r
i
s
a
c
t
i
ve
,
t
he
n
a
t
l
e
as
t
on
e
o
f
i
t
s
C
P
U
s
i
s
a
c
t
i
ve
.
∑
,
≥
1
=
1
=
1
∀
I
f
a
s
e
r
ve
r
i
s
a
c
t
i
ve
,
t
he
n
a
l
l
t
he
C
P
U
s
i
n
a
n
a
c
t
i
ve
s
e
r
ve
r
a
r
e
e
i
t
he
r
a
c
t
i
ve
or
i
dl
e
bu
t
n
ot
s
w
i
t
c
he
d
of
f
.
∏
(
,
+
,
)
=
1
=
1
=
1
∀
I
f
a
s
e
r
ve
r
i
i
s
of
f
,
a
l
l
t
he
C
PUs
i
ns
i
de
t
he
s
e
r
ve
r
a
r
e
of
f
.
∑
,
+
,
=
0
=
1
=
0
∀
A
C
P
U
c
a
n
no
t
be
a
c
t
i
ve
a
nd
i
dl
e
a
t
t
he
s
am
e
t
im
e
.
,
+
,
≤
1
∀
,
∀
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.
11
, No
.
1,
Febr
uar
y
2021
:
6
86
-
698
692
A
t
a
s
k
n
m
us
t
be
a
s
s
i
gn
e
d
t
o
on
e
c
o
r
e
o
nl
y
a
t
a
ny
gi
ve
n
t
im
e
.
∑
∑
∑
,
,
,
=
1
=
1
=
1
=
1
∀
A
c
or
e
c
a
n
e
xe
c
ut
e
a
s
i
ng
l
e
t
as
k
o
nl
y
a
t
a
ny
gi
ve
n
t
im
e
or
n
ot
e
xe
c
ut
e
a
ny
t
a
s
k.
∑
,
,
,
≜
{
1
,
,
=
1
0
,
,
=
0
=
1
∀
,
∀
,
∀
T
he
n
um
be
r
of
a
c
t
i
ve
c
or
e
s
i
s
e
qu
a
l
t
o
t
he
t
ot
a
l
nu
m
be
r
of
t
a
s
ks
t
o
be
s
c
he
d
ul
e
d.
∑
∑
∑
,
,
=
=
1
=
1
=
1
A
c
or
e
c
a
nn
ot
be
r
u
nn
i
n
g
i
n
H
i
gh
a
nd
L
ow
m
od
e
a
t
t
he
s
a
m
e
ti
m
e
.
,
,
,
+
,
,
,
≤
1
∀
,
∀
,
∀
,
∀
I
f
a
t
a
s
k
i
s
e
xe
c
ut
e
d
o
n
a
s
pe
c
i
f
i
c
c
or
e
,
t
he
n
i
t
m
us
t
be
e
xe
c
u
t
e
d
a
t
e
i
t
he
r
H
ig
h
or
L
ow
m
od
e
.
,
,
,
+
,
,
,
=
1
,
,
,
=
1
∀
,
∀
,
∀
,
∀
i
f
a
t
a
s
k
i
s
no
t
e
xe
c
ut
e
d
o
n
a
n
y
c
or
e
t
he
n
i
t
c
a
nn
ot
be
e
xe
c
u
t
e
d
a
t
ne
i
t
he
r
H
i
gh
n
or
L
ow
m
od
e
.
,
,
,
+
,
,
,
=
0
,
,
,
=
0
∀
,
∀
,
∀
,
∀
T
he
r
e
q
ui
r
e
d
f
r
e
qu
e
nc
y
by
t
a
s
k
n
do
e
s
no
t
e
x
c
e
e
d
t
he
a
va
i
l
ab
l
e
hi
gh
or
l
o
w
f
r
e
q
ue
nc
y
of
t
he
c
or
e
.
,
,
,
≤
,
,
∀
,
∀
,
∀
,
∀
,
,
,
≤
,
,
∀
,
∀
,
∀
,
∀
T
he
a
m
ou
nt
of
R
A
M
r
e
qu
i
r
e
d
by
t
a
s
k
n
i
s
l
e
s
s
t
ha
n
o
r
e
q
ua
l
t
o
t
he
am
ou
nt
of
R
A
M
a
va
i
l
a
bl
e
i
n
c
or
e
k
.
,
,
,
≤
,
,
∀
,
∀
,
∀
,
∀
Figure
1. A
sim
ple d
at
acente
r
c
onf
ig
ur
at
io
n
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
Op
ti
miz
atio
n o
f ener
gy
c
on
s
umptio
n
i
n
cl
ou
d
c
omp
uting
datace
nters
(
Ah
med Os
m
an
)
693
2.3.
D
ynamic
sche
duli
n
g
F
ol
l
ow
i
ng
t
he
s
c
he
du
l
i
ng
of
t
he
f
i
r
s
t
ba
t
c
h
a
t
ti
m
e
t
0
,
i
n
th
e
s
e
c
on
d
ph
a
s
e
of
t
hi
s
w
or
k,
w
e
a
s
s
um
e
t
he
a
r
r
i
va
l
of
a
ne
w
ba
t
c
h
t
o
be
s
c
he
d
ul
e
d
a
t
t
im
e
t
1
.
H
e
r
e
,
t
he
r
em
a
in
i
ng
f
r
e
e
c
or
e
s
a
r
e
us
e
d.
F
o
l
l
ow
i
ng
t
he
s
c
he
du
l
i
ng
of
t
he
f
i
r
s
t
b
a
t
c
h
a
nd
w
hi
l
e
pr
e
pa
r
i
n
g
t
o
s
c
he
du
l
e
t
he
s
e
c
on
d
ba
t
c
h,
t
he
c
on
s
t
r
a
i
nt
s
be
l
ow
a
r
e
a
dd
e
d:
A
l
l
t
he
c
or
e
s
t
h
a
t
a
r
e
s
t
il
l
r
un
n
i
ng
a
t
a
s
k
f
r
om
ba
t
c
h
t
0
a
r
e
m
a
r
ke
d
a
s
bu
s
y
,
∑
∑
∑
,
,
=
1
=
=
1
=
1
w
he
r
e
i
s
t
he
n
um
be
r
of
t
a
s
ks
f
r
om
t
he
f
i
r
s
t
ba
t
c
h
t
ha
t
a
r
e
s
t
il
l
r
un
ni
n
g
a
t
t
1
.
I
f
a
c
o
r
e
,
,
i
s
bu
s
y
t
he
n
no
t
a
s
k
f
r
om
t
he
s
e
c
on
d
ba
t
c
h
i
s
a
s
s
i
gn
e
d
t
o
i
t
,
∑
,
,
,
=
1
=
0
,
,
=
1
w
he
r
e
i
s
t
he
n
um
be
r
of
t
he
n
e
w
t
a
s
ks
i
n
t
he
s
e
c
on
d
ba
t
c
h
a
t
t
1
.
T
he
t
ot
a
l
nu
m
be
r
of
a
c
t
i
ve
c
or
e
s
i
s
no
w
e
q
ua
l
t
o
t
he
t
ot
a
l
n
u
m
be
r
of
t
a
s
ks
t
o
be
s
c
he
d
ul
e
d
pl
us
t
he
n
um
be
r
of
t
a
s
ks
t
ha
t
a
r
e
s
t
i
ll
r
un
ni
n
g
f
r
om
t
he
f
i
r
s
t
ba
t
c
h.
∑
∑
∑
,
,
=
1
=
=
1
=
1
+
3.
RESU
LT
S
AND DI
SCUS
S
ION
T
he
ge
ne
r
a
t
e
d
op
t
im
i
z
a
ti
on
m
od
e
l
of
(
1)
i
s
va
l
i
da
t
e
d
a
nd
t
e
s
t
e
d
us
i
ng
S
A
T
-
ba
s
e
d
0
-
1
P
B
s
ol
ve
r
s
na
m
e
d
N
a
P
S
,
M
i
ni
s
a
t+
,
a
nd
S
a
t
4j
,
a
nd
t
he
ge
ne
r
i
c
I
L
P
s
o
l
ve
r
C
P
L
E
X
.
T
he
s
ol
ve
r
s
c
o
m
p
ut
e
t
he
m
i
ni
m
u
m
e
ne
r
gy
c
on
s
u
m
pt
i
on
of
t
he
da
t
a
c
e
nt
e
r
a
nd
t
he
c
or
r
e
s
po
n
di
ng
a
s
s
i
gn
m
en
t
of
t
a
s
ks
t
o
t
he
c
or
e
s
t
ha
t
w
ou
l
d
a
c
hi
e
ve
t
hi
s
po
w
e
r
.
T
o
t
e
s
t
t
he
m
od
e
l
,
s
e
ve
r
a
l
da
t
a
c
e
nt
e
r
i
ns
t
a
nc
e
s
w
e
r
e
c
r
e
a
t
e
d
w
i
t
h
di
f
f
e
r
e
nt
n
um
be
r
o
f
s
e
r
ve
r
s
,
di
f
f
e
r
e
nt
nu
m
be
r
a
nd
t
y
pe
s
of
C
P
U
s
(
du
a
l
-
c
o
r
e
or
q
ua
d
-
c
or
e
)
i
n
e
a
c
h
s
e
r
ve
r
,
a
nd
di
f
f
e
r
e
nt
nu
m
be
r
of
t
a
s
ks
t
o
be
s
c
h
e
du
l
e
d.
T
he
t
e
s
t
s
a
r
e
c
a
r
r
i
e
d
o
ut
a
s
de
s
c
r
i
be
d
i
n
t
he
f
ol
l
ow
i
n
g
s
e
c
t
i
on
s
.
The
com
pr
eh
ensive
m
od
el
in
(1
)
is
te
ste
d
us
in
g
th
e
three
PB
-
S
AT
so
l
ver
s
a
nd
CPL
E
X
.
The
t
est
ing
i
nst
ances
us
e
d
c
on
sist
of
a
nu
m
ber
of
ser
ve
rs
with
eac
h
s
erv
e
r
co
ntaini
ng
dual
or
quad
co
re
CPUs.
The
i
ns
t
ances
al
s
o
incl
ud
e
t
he
nu
m
ber
of
ta
sks
to
be
assigne
d
t
o
th
e
cor
e
s
of
the
datace
nter
.
T
he
ta
sk
s
to
be
sche
dule
d
hav
e
seve
ral
char
act
e
risti
cs
in
cl
ud
i
ng
:
ta
s
k’s
siz
e
(in
num
ber
of
inst
ru
ct
i
on
s
),
ta
sk’s
re
qu
i
red
fr
e
qu
e
ncy,
an
d
ta
sk
’
s
re
quire
d
R
AM.
Eac
h
cor
e
in
t
he
dat
acente
r
r
un
s
in
one
of
tw
o
m
od
e
s:
Hi
gh,
or
Lo
w
.
As
m
entioned
earli
er,
in
High
m
od
e
the
cor
e
runs
at
the
m
axi
m
u
m
fr
equ
ency
an
d
volt
age
al
lowe
d,
wh
il
e
i
n
low
m
od
e
,
the
cor
e
r
uns
at
t
he
m
ini
m
u
m
fr
equ
e
ncy
a
nd
volt
age
possi
ble.
The
te
sti
ng
i
nst
ances
a
re
generate
d
rand
om
l
y
fr
om
a
po
ol
of
serv
e
rs.
The
ta
sk
s
requirem
ents
are
creat
ed
dep
e
ndin
g
on
the
ta
s
ks’
siz
e.
Table
2
sho
w
s
the
ta
sk
siz
e
ran
ges
a
nd
the
co
rr
es
pond
ing
ta
s
ks
’
r
e
q
ui
r
em
e
nt
s
.
T
ab
l
e
3
i
nc
l
ud
e
s
pa
r
t
i
a
l
s
e
r
ve
r
s
’
s
pe
c
i
f
i
c
a
t
i
on
s
.
T
a
bl
e
4 l
i
s
t
s
t
he
i
ns
t
a
nc
e
s
c
r
e
a
t
e
d f
o
r
t
hi
s
t
e
s
t
.
E
a
c
h
s
e
t
up
i
s
t
e
s
t
e
d
t
w
i
c
e
us
i
ng
di
f
f
e
r
e
nt
nu
m
be
r
of
t
a
s
k
s
.
T
he
i
ns
t
a
nc
e
s
a
r
e
f
i
r
s
t
s
ol
v
e
d
us
i
ng
a
l
l
t
he
s
ol
ve
r
s
t
o
f
i
n
d
t
he
m
i
ni
m
um
e
ne
r
gy
c
on
s
um
pt
i
on
a
nd
c
om
pa
r
e
th
e
s
ol
ve
r
s
’
pe
r
f
or
m
a
nc
e
.
T
he
n,
t
he
s
am
e
i
ns
t
a
nc
e
s
a
r
e
s
ol
ve
d
us
i
ng
r
ou
nd
r
ob
i
n
a
n
d
m
ax
im
u
m
po
s
s
i
bl
e
va
l
ue
m
e
t
ho
ds
t
o
c
o
m
pa
r
e
t
he
i
r
pe
r
f
or
m
a
nc
e
a
ga
i
ns
t
t
he
pr
o
po
s
e
d
m
od
e
l
.
T
a
bl
e
2.
T
a
s
ks
’
s
i
z
e
r
a
n
ge
s
a
n
d
r
e
q
ui
r
e
m
e
nt
s
us
e
d
i
n
T
e
s
t
A
Task
Size
Ran
g
e
(#in
stru
ctio
n
s)
Req
u
ired Frequ
en
cy
(M
Hz
)
Req
u
ired RAM
(GB)
500
–
2
0
0
0
800
1
2500
–
4
0
0
0
1000
2
4500
–
6
0
0
0
2400
2
6500
–
8
0
0
0
2600
3
8500
–
1
0
0
0
0
2800
4
T
a
bl
e
3.
T
e
s
t
A
s
e
r
ve
r
s
’
i
nf
o
r
m
a
ti
on
Se
rver
No
.
Server
Ty
p
e
Hig
h
Fr
eq
.
(
MHz
)
Hig
h
Volt.
(Volt
)
Low Freq.
(
MHz
)
Low Vo
lt.
(Volt)
RAM
(
GB)
Server1
Du
al
2400
1
.45
800
0
.9
2
Server2
Du
al
3000
1
.35
1000
1
.0
4
Server3
Qu
ad
2400
1
.35
1000
0
.9
2
Server4
Qu
ad
2600
1
.45
800
1
.0
4
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.
11
, No
.
1,
Febr
uar
y
2021
:
6
86
-
698
694
T
a
bl
e
4.
T
he
t
e
s
t
i
ng
i
ns
t
a
nc
e
s
f
or
T
e
s
t
A
Ins
tan
ce No.
No
.
o
f
servers
Total n
o
.
o
f
CPUs
Total n
o
.
o
f
cores
No
.
o
f
Task
s
Total Task
s Siz
e
(#in
stru
ctio
n
s)
1
2
4
12
2
7
,50
0
2
2
4
12
6
2
5
,50
0
3
3
6
20
5
2
3
,00
0
4
3
6
20
10
4
1
,00
0
5
4
8
24
8
2
9
,50
0
6
4
8
24
12
4
7
,00
0
7
8
16
52
18
6
8
,50
0
8
8
16
52
26
9
8
,50
0
I
n
t
he
f
i
r
s
t
pa
r
t
of
t
he
t
e
s
t
,
t
he
P
B
-
S
A
T
s
ol
ve
r
s
a
n
d
C
P
L
E
X
s
ol
ve
d
t
h
e
t
e
s
t
i
ng
i
ns
t
an
c
e
s
t
o
f
i
n
d
t
he
m
i
ni
m
um
e
ne
r
gy
c
on
s
u
m
pt
i
on
a
nd
t
he
r
e
s
ul
t
s
a
r
e
pr
e
s
e
nt
e
d
i
n
T
a
bl
e
5.
A
l
l
s
ol
ve
r
s
m
a
na
ge
d
t
o
f
i
nd
t
he
m
i
ni
m
um
e
ne
r
gy
c
on
s
u
m
pt
i
on
i
n
t
he
f
i
r
s
t
s
i
x
i
ns
t
a
nc
e
s
.
T
he
C
P
U
t
im
e
s
f
or
t
he
P
B
-
S
A
T
s
ol
v
e
r
s
a
r
e
r
e
l
a
t
i
ve
ly
c
l
os
e
t
o
e
a
c
h
ot
he
r
,
w
i
t
h
N
a
P
S
t
a
ki
ng
t
he
l
e
a
d
i
n
m
os
t
of
t
he
i
ns
t
a
nc
e
s
.
H
o
w
e
ve
r
,
i
f
w
e
i
nc
l
ud
e
C
P
L
E
X
i
n
t
he
c
om
pa
r
i
s
on
w
e
c
a
n
s
e
e
t
ha
t
e
x
c
e
pt
i
n
i
ns
t
a
nc
e
1
i
t
w
a
s
a
l
w
a
y
s
f
a
s
t
e
r
t
ha
n
t
he
ot
he
r
s
ol
ve
r
s
w
i
t
h
a
n
a
pp
r
e
c
i
a
bl
e
di
f
f
e
r
e
nc
e
a
s
c
a
n
be
s
e
e
n
i
n
F
i
gu
r
e
2
.
T
a
bl
e
5.
T
he
m
i
ni
m
um
e
ne
r
gy
c
on
s
um
pt
i
on
a
nd
C
P
U
t
im
e
af
t
e
r
a
dd
i
n
g
H
i
g
h
a
n
d
L
o
w
e
xe
c
ut
i
on
m
od
e
s
Ins
tan
ce No.
Min Energ
y
Co
n
su
m
p
tio
n
CPU Ti
m
e
(secon
d
s)
NaPS
Minis
at+
Sat4
j
CPLEX
1
151
1
.05
0
.05
0
.18
0
.33
2
536
2
.63
1
.94
7
.79
0
.35
3
488
9
.86
1
5
.31
1
3
.71
0
.33
4
690
4
4
.0
1
8
4
.74
1
5
1
.68
0
.53
5
516
6
5
.09
1
4
0
.63
1
5
0
.36
0
.51
6
730
2
1
5
.9
7
0
5
.13
6
8
.21
0
.68
7
1101
> 10
0
0
(11
0
8
)
> 10
0
0
(12
6
1
)
> 10
0
0
(11
4
3
)
3
.02
8
1567
> 10
0
0
(20
8
0
)
> 10
0
0
(18
3
3
)
> 10
0
0
(17
5
6
)
4
.96
F
i
gu
r
e
2.
T
he
p
e
r
f
or
m
a
nc
e
of
t
he
P
B
-
S
A
T
s
ol
v
e
r
s
a
n
d
C
P
L
E
X
i
n
T
e
s
t
A
I
n
t
he
l
a
s
t
t
w
o
i
ns
t
a
nc
e
s
(
7
a
nd
8)
t
he
P
B
-
S
A
T
s
ol
ve
r
s
a
r
e
s
t
op
pe
d
a
f
t
e
r
e
xc
e
e
di
ng
t
he
10
00
s
e
c
o
n
d
m
a
r
k.
T
he
va
l
u
e
s
be
t
w
e
e
n
t
he
br
a
c
ke
t
s
i
n
T
a
bl
e
5
a
r
e
t
he
b
e
s
t
s
ol
ut
i
on
f
o
un
d
by
t
he
s
ol
ve
r
s
u
p
t
o
t
ha
t
po
i
nt
i
n
t
im
e
.
O
nl
y
C
PL
E
X
ha
ve
s
uc
c
e
s
s
f
ul
l
y
f
ou
n
d
t
he
m
i
ni
m
um
e
ne
r
gy
c
o
ns
um
pt
i
on
i
n
t
h
os
e
i
ns
t
a
nc
e
s
a
nd
w
i
t
hi
n
l
e
s
s
t
ha
n
5
s
e
c
on
ds
.
F
r
om
F
ig
ur
e
3
w
e
no
t
i
c
e
t
ha
t
t
he
m
in
im
um
e
ne
r
gy
c
on
s
um
pt
i
on
i
s
di
r
e
c
t
ly
r
e
l
at
e
d
t
o
t
he
t
ot
a
l
nu
m
be
r
of
i
ns
t
r
uc
t
i
on
s
of
t
he
t
a
s
ks
t
o
be
s
c
he
d
ul
e
d.
T
he
t
im
e
t
a
ke
n
by
N
a
P
S
i
s
pr
op
or
t
i
o
na
l
t
o
t
he
t
ot
a
l
num
be
r
of
i
ns
t
r
uc
t
i
on
s
a
s
w
e
l
l
.
T
a
bl
e
6
de
pi
c
t
s
t
he
r
e
s
ul
t
s
of
t
he
s
e
c
on
d
pa
r
t
of
t
he
t
e
s
t
w
he
r
e
t
he
pr
o
po
s
e
d
m
od
e
l
i
s
c
om
pa
r
e
d
w
i
t
h
m
a
xi
m
um
po
s
s
i
bl
e
va
l
ue
(
M
P
V
)
a
n
d
r
ou
n
d
r
o
bi
n
(
R
R
)
.
We
ob
s
e
r
ve
t
h
a
t
ov
e
r
a
l
l
t
he
pr
o
po
s
e
d
m
od
e
l
ha
s
c
on
s
i
s
t
e
nt
l
y
r
e
t
ur
ne
d
t
he
m
in
im
um
e
ne
r
gy
c
on
s
um
pt
i
on
w
he
n
c
om
pa
r
e
d
w
i
t
h
t
he
ot
he
r
t
w
o
m
et
ho
ds
.
T
he
a
ve
r
a
ge
a
m
ou
nt
of
e
ne
r
gy
s
a
ve
d
i
s
2
7
.
3%
of
t
he
m
ax
im
um
e
ne
r
gy
c
on
s
um
pt
i
on
w
i
t
h
a
m
a
xi
m
um
of
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
Op
ti
miz
atio
n o
f ener
gy
c
on
s
umptio
n
i
n
cl
ou
d
c
omp
uting
datace
nters
(
Ah
med Os
m
an
)
695
35
.
6%
e
ne
r
gy
s
a
ve
d
i
n
i
ns
t
a
nc
e
7.
We
a
l
s
o
no
t
i
c
e
t
ha
t
r
ou
nd
r
ob
i
n
m
e
t
ho
d
f
a
i
l
e
d
t
o
s
c
he
du
l
e
t
w
o
o
f
t
he
i
ns
t
a
nc
e
s
(
4
a
nd
6)
.
T
hi
s
c
a
n
be
a
t
t
r
i
bu
t
e
d
t
o
t
he
f
a
c
t
t
ha
t
r
ou
n
d
r
o
bi
n
m
et
ho
d
s
c
he
du
l
e
s
t
he
t
a
s
ks
t
o
t
he
f
i
r
s
t
a
va
il
ab
l
e
c
or
e
t
ha
t
c
an
e
xe
c
ut
e
t
he
s
e
t
a
s
ks
e
ve
n
i
f
th
e
t
a
s
k
i
s
sm
al
l
a
nd
c
a
n
b
e
e
x
e
c
ut
e
d
on
s
om
e
ot
he
r
l
ow
e
r
c
a
pa
bi
l
i
ty
c
or
e
.
T
hi
s
l
ea
ds
t
o
l
a
r
ge
r
t
a
s
ks
no
t
f
i
n
di
ng
a
ny
c
or
e
t
ha
t
c
a
n
be
a
s
s
i
gn
e
d
t
o
t
he
m
be
c
au
s
e
a
l
l
t
he
c
a
pa
bl
e
c
o
r
e
s
a
r
e
e
xe
c
ut
i
ng
s
m
a
ll
e
r
t
a
sks
a
n
d
t
ha
t
e
x
a
c
t
ly
w
ha
t
ha
pp
e
ne
d
i
n
bo
t
h
i
ns
t
a
nc
e
s
4
a
nd
6.
T
he
o
nl
y
w
a
y
r
ou
nd
r
o
bi
n
c
a
n
s
c
he
d
ul
e
t
he
t
a
s
ks
o
f
i
ns
t
a
nc
e
s
4
a
n
d
6
i
s
b
y
w
a
i
t
i
ng
un
t
i
l
on
e
of
t
he
b
us
y
c
or
e
s
f
i
ni
s
he
s
i
t
s
t
a
sk.
H
e
nc
e
,
i
n
a
l
l
c
a
s
e
s
t
he
pr
o
po
s
e
d
m
od
e
l
pr
ov
e
d
i
t
s
s
up
e
r
i
o
r
i
t
y
.
F
ig
ur
e
4
i
s
a
g
r
a
ph
i
c
a
l
r
e
pr
e
s
e
nt
a
t
i
on
of
t
he
r
e
s
ul
t
s
of
T
a
bl
e
6.
F
i
gu
r
e
3.
I
l
l
us
t
r
a
t
i
on
of
hi
gh
a
nd
l
o
w
m
od
e
l
r
e
s
ul
t
s
T
a
bl
e
6.
C
om
pa
r
i
s
on
o
f
t
he
p
r
op
os
e
d
m
od
e
l
w
i
t
h
r
o
un
d
r
ob
i
n
a
nd
m
a
xi
m
um
po
s
s
i
bl
e
va
l
ue
m
et
ho
ds
Ins
tan
ce No.
Total Task
s Siz
e
Energy Co
n
su
m
p
ti
o
n
MPV
Ro
u
n
d
Ro
b
in
Prop
o
sed
M
o
d
el
1
7
,50
0
233
189
151
2
2
5
,50
0
584
536
536
3
2
3
,00
0
599
491
488
4
4
1
,00
0
939
UNSAT
690
5
2
9
,50
0
742
560
516
6
4
7
,00
0
1081
UNSAT
730
7
6
8
,50
0
1710
1383
1101
8
9
8
,50
0
2277
1830
1567
F
i
gu
r
e
4.
E
ne
r
gy
c
on
s
um
pt
i
on
o
f
t
he
p
r
o
po
s
e
d
m
od
e
l
,
r
ou
n
d
r
ob
i
n,
a
nd
m
a
xi
m
um
po
s
s
i
bl
e
va
l
ue
m
e
t
ho
ds
Evaluation Warning : The document was created with Spire.PDF for Python.