Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
6, N
o
. 1
,
Febr
u
a
r
y
201
6,
pp
. 20
5
~
21
1
I
S
SN
: 208
8-8
7
0
8
,
D
O
I
:
10.115
91
/ij
ece.v6
i
1.9
341
2
05
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
A P
r
ot
ot
ype of
Online Dynami
c Scaling Scheduler for Real-
Time Task based on Virtual Machine
J
u
nho Seo
,
Ky
ong Ho
on Ki
m
Department o
f
I
n
formatics, G
y
e
ongsang Nation
a
l University
, Korea
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Aug 20, 2015
Rev
i
sed
No
v 9, 201
5
Accepted Nov 28, 2015
In this pap
e
r, w
e
provide an
architect
ure of pow
er-awar
e
schedu
ler for r
e
al-
tim
e virtual m
a
chine s
y
st
em using d
y
namic fr
eque
nc
y
s
cal
ing m
echanis
m
This architecture provides that ho
w to
manage real-time resource and how to
control pro
cesso
r frequen
c
y
.
In additi
on, we prop
ose two scheduling schemes
that u
tilize slack
time to
adjust
pro
cessor frequency
without violating
tasks
deadl
i
ne.
Bas
e
d
on the prov
ided
archi
t
ec
ture
, we
im
plem
ent a v
i
rt
uali
zat
io
n
framework with
online power-
a
ware
sch
e
duler using RT-Xen real-time
h
y
perv
isor and
Litmus-RT real-
time OS
. Our implementation manages entire
of real-
tim
e res
ource and
proce
s
s
o
r
frequency
depending on s
y
stem policy
.
F
o
r tras
ferring
real-
tim
e r
e
quir
e
m
e
nts
,
we
im
plem
ent
an in
te
rface
us
ing
h
y
per
c
a
ll m
e
cha
n
is
m
.
To
evalu
a
t
e
provid
e
d s
y
s
t
e
m
, we an
al
yz
e p
e
rform
anc
e
evalu
a
tion
in
var
i
ous aspects.
Keyword:
D
y
n
a
m
i
c f
r
e
quen
c
y scaling
P
o
w
e
r-
a
w
ar
e
Real-tim
e schedule
r
Virtu
a
lizatio
n
Copyright ©
201
6 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
K
yon
g Hoo
n
K
i
m
,
Depa
rtm
e
nt of
In
fo
rm
atics,
Gy
eo
ngs
an
g N
a
t
i
onal
U
n
i
v
ers
i
t
y
,
Jin
j
ud
ae-r
o
501
, Ji
nj
u,
G
y
eong
san
g
n
a
m
d
o
,
Rep
u
b
lic of
Ko
r
e
a.
kh
ki
m
@
gn
u.ac
.k
r
1.
INTRODUCTION
As com
put
i
n
g
t
echn
o
l
o
gy
i
s
gr
ow
n, t
h
e p
o
we
r co
ns
um
pt
i
on has
bec
o
m
e
an im
port
a
nt
i
ssue i
n
em
bedde
d sy
st
em
. As t
h
e co
m
put
i
ng per
f
o
r
m
ance i
s
i
n
creased,
po
wer c
ons
um
pt
i
on i
s
al
so i
n
crease
d
.
Thi
s
pr
o
b
l
e
m
i
nduc
es i
n
crea
si
n
g
m
a
nagem
e
nt
cost
s
u
ch
as el
e
c
t
r
i
c
i
t
y
charge
and
co
ol
i
n
g c
o
st
.
Dy
nam
i
c Vol
t
a
ge F
r
eq
ue
ncy
Scal
i
ng
(
D
V
FS) t
e
c
hni
que
i
s
m
o
st
com
m
onl
y
used t
e
chni
que
f
o
r
red
u
ci
n
g
t
h
e
p
o
we
r c
ons
um
pt
i
on i
n
c
o
m
put
i
ng sy
st
em
s. It
al
l
o
ws t
o
c
o
nt
rol
pr
ocess
o
r
f
r
eq
ue
ncy
an
d
vol
t
a
ge
dy
nam
i
cal
ly
.
Gene
ral
l
y
, wh
en t
h
e pr
ocess
o
r f
r
eq
ue
ncy
o
r
vol
t
a
ge i
s
decresed
, t
h
e p
o
w
er co
ns
um
pt
ion i
s
al
s
o
decrese
d
. The
DVFS-supported process
o
r c
a
n easily ach
i
e
ve a r
e
d
u
ce
d
po
we
r co
ns
u
m
pti
on
whe
n
e
v
er t
h
e
m
a
xim
u
m
freque
ncy
i
s
not
req
u
i
r
e
d
by
sim
p
l
y
scal
i
ng d
o
w
n
t
h
e o
p
e
rat
i
n
g
f
r
eq
ue
ncy
o
f
t
h
e
p
r
ocess
o
r
.
H
o
w
e
v
e
r,
d
ecreasin
g fr
equ
e
ncy o
f
pr
o
c
essor in
du
ces th
at task
s sp
en
d m
o
re ti
m
e
to
co
m
p
lete.
The
virt
ual m
a
chine tec
h
nol
ogy can e
x
ecute
differe
n
t s
o
ft
ware
platform
s in a
single sy
ste
m
, whic
h
has
been c
o
m
m
onl
y
used a
s
a basi
c
pl
at
fo
rm
i
n
vari
o
u
s a
ppl
i
cat
i
o
n
area t
o
s
u
pp
ort
re
so
urce
s
h
ari
n
g
,
p
a
rtitio
n
i
n
g
, an
d
etc. Th
is tech
no
log
y
in
cl
ud
es reall
o
cating
en
tire
reso
urce b
y
v
i
rt
u
a
lized
h
a
rd
ware reso
urce.
In
[1
], th
ey p
r
ov
id
e a strateg
y
to
ad
op
t serv
er v
i
rtu
a
lization fo
r
u
s
ing
syste
m
reso
urce efficien
tly. Th
u
s
it can
b
e
u
s
ed
for a so
lu
tion
o
f
efficien
t sch
e
du
lin
g in
co
m
p
lex
real-ti
m
e syste
m
s.
In re
al
-t
im
e syst
em
s, a t
a
sk has real
-t
im
e requi
rem
e
nt
such as dea
d
l
i
n
e a
nd
peri
od
. The
t
a
sk sh
oul
d
be operate
d
to satisfy real-time requ
irem
en
t.
To
gu
aran
tee work
i
n
g
o
f
task
, th
e sch
e
du
ler m
a
n
a
g
e
s all o
f
tasks
by
sc
hed
u
l
i
n
g
al
go
ri
t
h
m
.
Thu
s
, t
o
st
u
d
y
res
o
urce
sha
r
i
n
g
al
g
o
rith
m
is i
m
p
o
r
tan
t
issu
e in
real-tim
e syste
m
s.
Many recent st
udies
have
bee
n
provide
d
for
real-
tim
e
syste
m
s
based on vi
rtual m
achine. RT-Xe
n
[2]
i
s
a ki
nd of re
al
-t
im
e hy
perv
i
o
sr base
d o
n
Xen [
3
]
.
The
ori
g
i
n
al
Xen
h
y
p
er
vi
so
r sche
dul
i
n
g al
g
o
ri
t
h
m
i
s
Credit-based C
P
U sc
heduling, where eac
h
virtu
a
l m
ach
in
e is allo
cated
CPU ti
m
e
in
p
r
opo
rtion
to
the cred
it.
In
[
2
]
[
4
]
,
t
h
ey
m
odi
fi
ed t
h
e
C
r
edi
t
-
based
s
c
hed
u
l
e
r
t
o
s
u
p
p
o
r
t
real
-t
im
e requi
rem
e
nt
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
6, No
. 1, Feb
r
uar
y
20
1
6
:
20
5 – 21
1
20
6
In t
h
i
s
pa
per,
we p
r
o
v
i
d
e a
pr
ot
ot
y
p
e
of
o
n
l
i
n
e p
o
w
er
-a
ware sc
he
dul
e
r
fo
r real
-t
i
m
e task ba
sed
o
n
hy
pe
rvi
s
or
. Th
e pr
op
ose
d
sy
s
t
em
can reduce
po
wer c
o
nsum
pt
i
on
usi
n
g
D
V
FS m
echani
s
m
whi
l
e
gua
ra
nt
eei
ng
real-tim
e requirem
e
nts
2.
RELATED WORKS
Xen
hy
per
v
i
s
o
r
i
s
ki
n
d
o
f
vi
r
t
ual
m
achi
n
e f
r
am
ework
t
o
s
u
p
p
o
rt
vari
ous
sche
dul
i
ng al
go
ri
t
h
m
suc
h
as cre
d
it, earliest deadline
firs
t, and
ar
i
n
c
653
. I
n
[4
],
th
ey ex
p
a
nd
ed
th
e X
e
n
h
y
p
e
rv
isor
schedu
ler
t
o
su
ppo
rt
real-tim
e requi
rem
e
nt for s
o
f
t
real-t
im
e applications. They im
ple
m
ented th
e lo
ad
b
a
lancer to
m
i
n
i
miz
e
th
e
laten
c
y o
f
real-ti
m
e ap
p
licatio
n
s
b
y
sp
ecifying
th
eir requ
ired
lax
ity v
a
lu
es. Xi, et al. [2
] p
r
op
o
s
ed
RT-Xen
t
o
su
ppo
rt real
-time sch
e
d
u
ling alg
o
r
ith
m
su
ch
as Earliest Dead
lin
e First (EDF) and
Rate Mo
no
ton
i
c (RM) in
Xen
hy
per
v
i
s
o
r
.
The Dy
nam
i
c
Vol
t
a
ge
Fre
q
u
e
ncy
Scal
i
ng
(
D
V
FS) m
echani
s
m
i
s
a ki
nd
of s
o
l
u
t
i
o
n w
h
i
c
h scal
es t
h
e
pr
ocess
o
r
fre
q
u
ency
a
nd
v
o
l
t
a
ge dy
nam
i
cally
for
re
duci
ng
dy
nam
i
c powe
r
co
ns
um
pt
i
on.
B
u
t
t
h
i
s
m
ech
ani
s
m
h
a
s tr
ad
e-
of
f betw
een
po
w
e
r
co
nsu
m
p
tio
n
an
d ex
ecu
tio
n ti
me.
W
h
e
n
t
h
e
fre
que
ncy is
decresed, tas
k
needs
m
o
re t
i
m
e
t
o
execut
e
.
T
h
i
s
pr
o
b
l
m
e i
s
fatal
t
o
real
-t
i
m
e sy
st
em
. Inc
r
e
a
si
ng
exec
ut
i
o
n t
i
m
e i
nduce
s
by
vi
ol
at
i
o
n
o
f
de
adl
i
n
e.
To s
o
l
v
e
re
d
u
c
i
ng
p
o
we
r c
o
n
s
um
pt
i
on
on
re
al
-t
im
e sy
st
em, m
u
ch resea
r
c
h
has
pr
o
pose
d
sche
d
u
l
i
n
g
al
go
ri
t
h
m
usi
ng D
V
FS t
ech
n
i
que f
o
r
peri
o
d
i
c
t
a
sks [
5
]
-
[
7
]. In
[6
], th
ey p
r
ov
id
ed
a alg
o
rith
m
th
at selects
opt
i
m
al
proces
sor
fre
q
u
ency
r
a
t
i
o
. In
[7]
,
t
h
e
y
pro
v
i
d
e
d
a
m
e
t
hod t
h
at
ho
w t
o
cha
n
ge p
r
ocess
o
r
fre
que
ncy
at
running ti
m
e
of each task. For reduci
ng
power c
onsum
ption, they utilize sl
ack ti
m
e
which is rem
a
ining ti
m
e
of real-tim
e resource.
In [
5
]
,
they
de
fine
d the p
o
w
e
r-a
ware sc
hem
e
s fo
r pe
riod
ic real-ti
m
e task
s in
v
a
ri
o
u
s
lev
e
ls, su
ch
as
sy
st
em
l
e
vel
,
com
pone
nt
l
e
ve
l
and t
a
sk l
e
vel
.
In sy
st
em
l
e
vel
,
wh
ol
e sy
st
em
are execut
e
d
on sam
e
freq
u
e
ncy
.
On t
h
e ot
her
h
a
nd
, i
n
c
o
m
ponent
l
e
vel
a
nd
t
a
sk l
e
vel
,
t
h
e
com
pone
nt
s w
h
i
c
h l
i
k
e
vi
rt
u
a
l
m
achi
n
es or
t
a
sk
s
are execute
d on differe
n
t fre
que
ncy de
pending on re
quire
m
e
nts of each com
pone
nt or
task. Each level has
ow
n sch
e
d
u
l
e
r
and p
o
w
er
-a
ware m
odul
e.
The sche
d
u
l
e
r
ope
rat
e
s ow
n
sched
u
l
i
n
g al
go
ri
t
h
m
usi
n
g
gi
ven
reso
u
r
ce fr
om
up
per
-
l
a
y
e
r sched
u
l
e
r an
d
pr
o
v
i
d
es res
o
u
r
ce t
o
un
der
-
l
a
y
e
r com
pone
nt
s. The p
o
w
e
r-a
ware
m
odul
e m
a
nages p
o
w
e
r
-
awa
r
e i
n
f
o
rm
at
i
on a
n
d
co
nt
r
o
l
s
t
h
e
p
r
oces
so
r
fre
q
u
ency
de
pen
d
i
n
g
o
n
sy
st
em
pol
i
c
y
.
3.
SYSTE
M
MO
DEL
The p
r
o
p
o
sed
archi
t
ect
u
r
e fo
r onl
i
n
e
po
wer
-
awa
r
e
sc
he
dul
er
i
s
s
h
ow
n
i
n
Fi
gu
re 1.
I
n
hy
per
v
i
s
o
r
,
w
e
defi
ne a
Di
scr
e
t
e
Fre
que
ncy
C
ont
r
o
l
(D
FC
)
m
odul
e w
h
i
c
h
m
a
nages real
-
t
im
e requi
rem
e
nt
o
f
t
a
s
k
s i
n
g
u
est
OS a
n
d
co
nt
r
o
l
s
pr
ocess
o
r f
r
e
que
ncy
depe
n
d
i
ng
o
n
real
-time requirem
ent whe
n
e
v
er t
h
e real-tim
e require
m
e
nt
i
s
chan
ge
d. T
h
e DFC
m
odul
e
cont
rol
s
C
P
U
fre
que
ncy
t
o
r
e
duce
p
o
we
r c
ons
um
pt
i
on a
n
d m
odi
fi
es real
-t
im
e
tasks re
quirement for
guara
n
tee real-tim
e requi
rem
e
nt
. And we a
d
d
hypercalls
whic
h trans
f
er
real-t
im
e
req
u
irem
ent fr
om
guest VM
s. In
Gue
s
t VM
, we im
plem
ent DFC API t
o
collect real-time requirem
e
nts. The
DFC
AP
I i
s
usi
n
g
a
hy
pe
rcal
l
i
n
t
e
rface
t
o
t
r
a
n
sfe
r
t
h
e re
q
u
i
r
em
ent
s
t
o
DFC
m
odul
e i
n
hy
p
e
rvi
s
or
.
In VM
, the DFC API m
a
nages real-t
im
e requirem
e
nt which is sum of
rea
l
-tim
e require
ment in each
t
a
sk. T
h
e
real
-t
im
e requi
rem
e
nt
de
fi
nes
by
T(
Π
,
Θ
)
whe
r
e
Π
i
s
peri
o
d
of
t
a
sk a
n
d
Θ
is ex
ecu
tion
tim
e
o
f
task.
We co
n
s
i
d
er t
h
at d
e
ad
lin
e is sam
e
to
p
e
riod
.
If a g
i
v
e
n
jo
b is failed
t
o
run
u
n
til a
p
e
rio
d
tim
e, it is
o
u
t
of
d
ead
lin
e. Th
e
DFC API in
cl
u
d
e
s DFC h
y
percall in
terface wh
ich
is inv
o
k
e
d
to
info
rm
th
e ch
ang
e
o
f
task
status whe
n
the task is c
r
eate
d
or term
inated.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
A Pr
ot
ot
ype
of
Onl
i
n
e
Dy
n
a
mi
c Sc
al
i
n
g
Sc
he
dul
er
f
o
r
Real
-
T
i
m
e T
a
sk
b
a
s
e
d
on
Vi
rt
u
a
l
Mac
h
i
n
e
(
K
.H. Kim)
20
7
Fig
u
r
e
1
.
Th
e
ar
ch
itectur
e o
f
pr
opo
sed
system
To re
d
u
ce p
o
w
er c
o
n
s
um
pt
i
o
n
,
we
defi
ne
t
w
o
pol
i
c
i
e
s b
a
sed o
n
sy
st
e
m
l
e
vel
and co
m
ponent
l
e
vel
pri
n
ci
pl
es.
3.
1.
Sys
t
em
Static Fre
quenc
y
(SSF
)
The
SSF
policy picks t
h
e m
i
nim
u
m
speed l
e
vel am
ong
w
h
ich all the
virtual
m
achines ar
e sche
duled.
Th
us, t
h
e sy
st
e
m
freque
ncy
i
s
de
fi
ne
d as
f
o
l
l
o
ws:
∈
Θ
Π
∙
∈
1
whe
r
e
i
i
s
nu
m
b
er of spee
d
l
e
vel
,
j
i
s
num
ber of vi
rt
ua
l
m
achi
n
e,
f
i
s
lev
e
l o
f
frequ
en
cy scaling
,
Θ
is
ex
ecu
tion
tim
e
o
f
task, an
d
Π
i
s
pe
ri
o
d
of t
a
sk.
i
s
m
a
xim
u
m
freque
ncy
o
f
p
r
ocess
o
r,
i
t
i
s
de
pen
d
i
ng
o
n
syste
m
proces
s
o
r.
The
SSF
p
o
l
i
c
y
ru
ns
at
sam
e
f
r
eq
ue
ncy
f
o
r al
l
vi
rtu
a
l m
ach
in
es.
Wh
en th
e tas
k
s
are
created
or
termin
ated
, the sch
e
du
ler
re-calcu
lates
op
ti
m
a
l frequ
en
cy. Th
is
p
o
l
i
c
y h
a
s a little ov
erh
e
ad
less th
an
Com
pone
nt Static Frequency
(CSF) m
e
thod which is
e
x
plained
bel
o
w because it is invoke
d
once
per
one
variation of task.
Howe
ver, t
h
e SSF
policy
has any wa
st
eful power c
o
nsum
ption m
a
rgi
n
. To
kee
p
a sam
e
freq
u
e
n
c
y is
no
t satisfied to
fu
lly redu
ce
power con
s
u
m
p
t
i
o
n fo
r all of the v
i
rt
u
a
l m
ach
i
n
e.
3.
2.
C
o
mp
one
n
t S
t
ati
c
Fre
q
uency
(
C
SF
)
Th
e CSF
p
o
licy allo
ws th
at each
vi
rt
ual
m
achi
n
e t
o
r
u
n
o
n
i
t
s
m
i
nim
u
m
spee
d l
e
vel
.
E
ach
vi
rt
ual
m
achi
n
e f
r
e
q
u
e
ncy
i
s
defi
ne
d
by
:
,
∈
,
Θ
Π
∙
,
∈
1
In the CSF pol
i
cy, each virtual
m
achine runs at differe
nt freque
n
cy. The
CSF policy can cover the
wast
ef
ul
p
o
we
r co
ns
um
pt
i
on
m
a
rgi
n
i
n
S
S
F
beca
use i
t
u
s
es di
f
f
ere
n
t
s
p
eed l
e
vel
.
H
o
we
ve
r, i
t
has
m
o
re
ove
rhead
tha
n
SSF beca
use
it shoul
d
c
h
ange CPU fre
qu
enc
y
whe
n
t
h
e
virt
ual m
achines a
r
e s
w
itched.
Aft
e
r t
h
e c
h
a
ngi
ng
fre
q
u
e
n
cy
, exec
ut
i
on
t
i
m
e
s of
tasks
are cha
n
ged
so that to avoid dea
d
line
vi
ol
at
i
o
n
.
Th
e DFC
m
odul
e c
h
an
ges e
x
ecut
i
on t
i
m
es of VM
s and t
h
e
D
F
C
API c
h
a
n
g
e
s execut
i
o
n t
i
m
e of
task
s.
Gen
e
rally, ex
ecu
tion
time o
f
task
is in
inv
e
rse
pr
o
p
o
rt
i
o
n t
o
pr
oce
ssor
fre
q
u
ency
.
Thu
s
, exec
ut
i
on t
i
m
e
i
s
cha
nge
d
by
r
a
t
i
o
o
f
decreasi
n
g
p
r
ocess
fre
q
u
ency
,
w
h
e
r
e r
a
t
i
o
i
s
de
fi
ne
d
by
⋅
/
whe
r
e
is ch
ang
e
d ex
ecu
tion
ti
m
e
an
d
is a
g
i
v
e
n
ex
ecu
tion
tim
e o
f
task
.
Ch
ang
i
ng
ex
ecu
tio
n tim
e is p
a
rt
o
f
po
licy, thu
s
th
is fu
n
c
tion
is in
vok
ed wh
en
th
e frequ
e
n
c
y i
s
ch
ang
e
d.
We im
p
l
e
m
en
t th
e ad
ap
tiv
e po
wer-aware sch
e
du
le
r
b
a
sed
o
n
rt
p
a
rtitio
n
sch
e
du
ler in
RT-Xen
h
y
p
e
rv
iso
r
. The rt
p
a
rtitio
n is real-tim
e sch
e
du
ler in
RT-Xen
b
a
sed
on
EDF algo
rit
h
m
.
W
e
m
o
d
i
fy th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
6, No
. 1, Feb
r
uar
y
20
1
6
:
20
5 – 21
1
20
8
sche
dul
er t
o
s
u
pp
o
r
t
po
wer
-
a
w
are f
u
nct
i
o
n
.
We al
so use a
Li
tm
us-R
T [8]
real
-t
i
m
e l
i
nux
OS usi
n
g G
S
N-E
D
F
algo
rithm
for p
o
we
r-a
wa
re V
M
.
To s
u
p
p
o
rt
t
h
e
pr
op
ose
d
arc
h
i
t
ect
ure, we a
d
d t
h
e
DFC
m
odul
e,
DFC
hy
p
e
rcal
l
i
n
R
T
-X
en an
d
DFC
API
in
VM
.
T
h
e
DFC m
odu
le
m
a
nages
p
o
w
er
-awa
re i
n
f
o
rm
ation o
f
al
l VM
s in t
h
e
sy
stem
. Whe
n
VM
s
pr
o
v
i
d
e t
h
ei
r
r
eal
-t
im
e requi
r
e
m
e
nt
, t
h
e m
odul
e
deci
des
o
p
t
i
m
a
l
freq
u
en
cy
of VM
s acc
or
di
n
g
as
pol
i
c
y
and
m
odifies execution tim
e of VMs.
Whe
n
the freque
n
cy is
selected, changing fre
qu
en
cy
fu
nct
i
on i
s
i
n
vo
ke
d
bef
o
re
co
ne
xt
s
w
i
t
c
h.
Th
e DFC h
ypercalls co
n
s
ist
o
f
three h
y
p
e
rcalls:
d
f
c_
in
it, d
f
c_op
,
and
df
c_re
new
. T
h
e
d
f
c_
in
it
is
in
itializat
io
n
fu
n
c
tion
t
h
at prep
ares fo
r supp
orting
DFC.
At th
e end
o
f
fu
n
c
tion
,
th
e
h
y
p
e
rcall
o
b
t
ains list o
f
fre
que
ncy
l
e
ve
l
by
ret
u
rn
val
u
e.
It
can al
s
o
use cl
ea
nu
p
f
unct
i
o
n t
h
at
f
r
ees t
h
e
VM specific DFC m
e
m
o
ry.
Thi
s
hy
percal
l
i
s
i
n
v
oke
d
w
h
en t
h
e
VM
i
s
b
oot
i
n
g a
n
d
sh
ut
i
ng
d
o
w
n
. T
h
e
df
c_
op
trans
f
ers
real
-tim
e
requirem
ents whic
h c
ontain
list of utilization
at each
fre
quency le
vels. This re
quir
em
ent lists are
m
a
naged by
DFC
m
odul
e
a
n
d
D
F
C
m
odu
l
e
pi
cks
t
o
c
o
m
p
are t
h
ese l
i
s
t
s
. To
t
r
a
n
sfe
r
dat
a
at
o
n
e t
i
m
e
, we de
vi
d
e
6
4bi
t
arg
u
m
e
nt
t
o
8 bl
oc
ks
o
f
8bi
t
arg
u
m
e
nt
.
Each
bl
ock has ra
nge
d bet
w
ee
n 0
an
d 25
5 unsi
gne
d
i
n
t
e
ger
,
b
u
t
we
j
u
st u
s
e
fro
m
0
to
10
0
b
ecau
s
e it co
n
t
ain
s
u
tilizatio
n
th
at
sh
ows
p
e
rcen
tag
e
. Fi
g
u
re
2
sho
w
s t
h
at arch
itetu
re
of
argum
e
nt. The
df
c_re
new
is to chec
k the syste
m
changes s
u
ch as
freque
ncy and VM execution tim
e.
Whe
n
a
task is created or term
inated, the
freq
u
e
n
c
y
and
VM ex
ecu
tio
n ti
m
e
is ch
ang
e
d
,
t
h
is
hyp
ercall in
form
s th
e
ch
ang
e
s. Th
an th
e DFC API catch
es th
is in
fo
rm
atio
n
an
d
u
p
d
a
tes ex
ecu
t
io
n
ti
m
e
o
f
task
s. Th
is h
y
p
e
rcall is
i
n
v
oke
d pe
ri
o
d
i
cal
l
y
.
Fig
u
r
e
2
.
A
stru
ctur
e o
f
df
c_op
ar
gu
m
e
n
t
In
VM,
we add
DFC API to
man
a
g
e
real
-time req
u
i
rem
e
n
t
s o
f
task
s in
kern
el. Th
is
API is prov
id
ed
real-tim
e
requi
rem
e
nts and ge
nerates list of
utilization at
each freque
ncy whe
n
the
real
-t
i
m
e task is created or
terminated. To check system
chan
ges, we m
odi
fy
sched
u
l
er
in
VM k
e
rn
el to
in
vok
e
df
c_re
new
hy
percall
peri
odi
cal
l
y
. Whe
n
t
h
e sy
st
em
change
s ar
e up
dat
e
d
,
t
h
e
DFC API
upda
tes execution t
i
m
e
of every real-tim
e
task
.
For t
h
e sim
p
lic
ity, we restrict som
e
argum
ent. To
cal
cul
a
t
e
sched
u
l
i
n
g o
p
er
at
i
on ra
pi
dl
y
and t
r
a
n
s
f
er
real
-t
i
m
e requi
rem
e
nt
sim
p
ly
, t
h
e peri
od i
n
h
y
p
er
vi
so
r sche
dul
e
r
i
s
fi
xe
d b
y
100m
s. Seco
nd
, we re
p
r
ese
n
t
t
h
e
real-tim
e
req
u
i
rem
e
n
t
o
f
h
ypercall in
terface as u
tiliza
tio
n
.
In
CSF
p
o
l
icy, we
m
o
d
i
fy to
p
i
ck
frequ
e
n
c
y
qui
c
k
l
y
usi
n
g
heu
r
i
s
t
i
c
ap
pr
o
ach.
4.
EX
P
E
R
I
M
E
NT
S
In
o
r
d
e
r to
simu
late p
e
riod
ic real-tim
e
task
s, we u
s
e th
e
π
calcu
latio
n
prog
ram
th
at calc
u
lates th
e
π
t
o
a
gi
ve
n
num
ber
o
f
di
gi
t
s
af
t
e
r t
h
e
deci
m
a
l p
o
i
n
t
a
n
d
a si
m
p
l
e
su
m
pro
g
r
am
t
h
at
adds
a ra
nd
om
num
ber
i
n
do
u
b
l
e
l
o
o
p
.
A
peri
odi
c t
a
s
k
a
ppl
i
cat
i
o
n i
s
ge
nerat
e
d s
o
t
h
at
i
t
exec
ut
es a
π
calcu
latio
n
pr
og
r
a
m
or
sim
p
le
sum
pro
g
r
am
every
peri
od
.
To
e
v
aluate powe
r
c
o
nsum
ption, we de
fine
two
sce
n
arios for ea
ch pol
i
cy. In each s
cenari
o
,
we
defi
ne two virt
ual m
achines. In first sce
n
a
r
io,
each virtual machine
has
t
w
o real-tim
e
tasks:
The
has t
a
sk
set
(7
,
2
)
and
(5,
1
)
where t
h
e
fi
rst
o
n
e i
s
pe
r
i
od a
nd t
h
e sec
o
n
d
one i
s
e
x
e
c
ut
i
on t
i
m
e. The
executes 3
tim
e
s in a single VM executi
on a
n
d
exec
utes 4 tim
e
s. The
has task set
(5
,
1
)
and
(6
,
1
)
. The
executes
2 times and
e
x
ecutes a single ti
me. Each task exec
utes t
h
e
sim
p
le sum
program
.
In sec
o
nd
scenari
o
, eac
h
virtual
m
achine has t
h
ree
real-tim
e ta
sks
tha
t
exec
utes
a
PI calculation program
:
The
has
task set
(
3
, 0.
2)
,
(
5
, 0.
5)
and
(7
,
1
)
.
The
executes
6 times in a si
ngl
e VM exec
ution, while
executes
4 times and
exec
ut
es 3 tim
es. The
has task set
(
2
, 0.
2)
,
(
4
, 0.
5)
and
(
6
, 0.
8)
. T
h
e
executes
6 times,
exec
utes
4 tim
e
s and
ex
ecu
t
es
3
tim
es. All of task
s ar
e term
inated when the
tasks
a
r
e
do
ne a
n
d
re-c
r
eat
ed be
f
o
re
t
h
e ne
xt
pe
ri
o
d
i
s
st
art
e
d.
To c
o
m
p
are o
u
r
pol
i
c
y
,
we
us
e p
o
we
r m
odel
[
9
]
w
h
i
c
h t
h
e
po
we
r c
ons
um
pt
i
o
n
E i
s
de
fi
n
e
d
by
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
A Pr
ot
ot
ype
of
Onl
i
n
e
Dy
n
a
mi
c Sc
al
i
n
g
Sc
he
dul
er
f
o
r
Real
-
T
i
m
e T
a
sk
b
a
s
e
d
on
Vi
rt
u
a
l
Mac
h
i
n
e
(
K
.H. Kim)
20
9
⋅
/
⋅
⋅
/
whe
r
e
α
is
a c
o
efficient
of a s
y
ste
m
,
t
is ex
ecu
tio
n tim
e, an
d
S
i
s
pr
ocess
o
r f
r
e
que
ncy
l
e
v
e
l
.
I
n
t
h
i
s
pa
pe
r,
we
assum
e
that
α1
for th
e sak
e
of si
m
p
l
i
city.
The figure 3 and fi
gure 4 s
h
ow that result in each s
cenari
o
using powe
r model. In the fi
gures
,
powe
r
consum
ption
value m
eans calculating
val
u
e
by power m
odel.
Fi
gu
re
3.
P
o
we
r c
ons
um
pt
i
on
of
scena
r
i
o
1
In Fi
g
u
re
3
,
i
t
sho
w
s
t
h
e res
u
l
t
s
of p
o
we
r
c
ons
um
pt
i
on us
i
ng p
o
we
r
m
odel
.
T
h
e pr
o
p
o
s
ed pol
i
c
i
e
s
sp
en
t less p
o
wer th
an
no
rm
al
sch
e
du
ling
policy. Esp
eciall
y
, th
e CSF spen
t th
e least po
wer.
It sho
w
s th
at
decreasi
n
g frequency is
m
o
re efficient t
o
re
duce
powe
r c
o
nsu
m
p
tio
n
ev
en
if to
tal ex
ecu
tio
n ti
m
e
is in
creased
.
Fi
gu
re
4.
P
o
we
r c
ons
um
pt
i
on
of
scena
r
i
o
2
As s
h
o
w
n
on
Fi
g
u
re
4,
p
o
w
er c
o
ns
um
pti
on i
s
m
o
re
d
ecreased
by
t
h
at
va
ri
at
i
ons
are
bi
g
g
er
.
Ho
we
ver
,
t
h
e po
we
r co
nsum
pt
i
on
of
pr
o
p
o
s
ed p
o
l
i
c
i
e
s are alm
o
st sa
m
e
. beca
use it cannot re
duce e
n
ergy
fu
rt
he
r un
der
t
h
e fre
que
ncy
p
r
o
v
i
d
e
d
by
sy
s
t
em
.
Al
t
h
o
u
g
h
i
t
can red
u
ce
m
u
ch
m
o
re po
wer
co
ns
um
pt
ion
,
i
t
j
u
st ru
ns
o
n
th
e min
i
m
u
m
f
r
e
qu
en
cy
pr
ov
id
ed
b
y
system
.
In addition, we evaluate the ove
rhead of power-awa
re control at
each part. The control ti
me
means
ope
rt
at
i
on t
i
m
e of sche
dul
i
ng
and m
a
nagi
n
g
i
n
DFC
.
The
DFC
AP
I re
qu
est
t
i
m
e
i
s
t
h
at ho
w l
o
n
g
i
t
t
a
kes t
o
i
n
v
oke
hy
pe
rc
al
l
.
The chan
gi
ng f
r
e
que
ncy
t
i
m
e
i
s
t
h
e requ
i
r
ed t
i
m
e
t
h
at
cont
rol
s
C
P
U
f
r
eq
ue
ncy
.
Ge
n
e
ral
l
y
,
real-tim
e
task
h
a
s m
o
re th
an
a m
i
llisec
o
n
d
ex
ecu
tion ti
m
e
. Tab
l
e 1
shows th
e resu
lt of overh
ead
measurem
ent. We a
n
alyzed t
h
at the
propose
d system
ha
s a
n
y
i
n
fl
ue
nce
o
n
t
h
e
r
u
nni
ng
o
f
real
-t
im
e t
a
sks.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
6, No
. 1, Feb
r
uar
y
20
1
6
:
20
5 – 21
1
21
0
Tabl
e
1. T
h
e
re
sul
t
o
f
ove
r
h
ea
d m
easurem
ent
Average
Maxim
u
m
Control ti
m
e
4
μ
s 7
μ
s
DFC
API r
e
quest
ti
m
e
4
μ
s 7
μ
s
Changing freque
ncy ti
m
e
2
μ
s 4
μ
s
5.
CO
NCL
USI
O
N
In t
h
i
s
pa
per
,
we p
r
op
ose
d
a
n
o
n
l
i
n
e
po
we
r
-
awa
r
e sc
hed
u
l
e
r f
o
r
real
-t
i
m
e hy
pe
rvi
s
or
. T
h
e p
r
op
ose
d
sy
st
em
provi
de
s ho
w t
o
red
u
c
e
pr
ocess
o
r
fre
que
ncy
ba
sed
on
real
-t
i
m
e requi
rem
e
nt
of t
a
sks,
h
o
w t
o
t
r
ansm
i
t
real
-t
i
m
e requi
rem
e
nt
of t
a
s
k
s i
n
t
o
hy
per
v
i
s
or
an
d
h
o
w to
gua
ra
ntee exec
ution tim
e
so
th
at to avo
i
d
d
e
ad
lin
e
vi
ol
at
i
o
n
.
In
SSF
policy, all the virt
ual m
achines are
execute
d
on sa
m
e
fre
que
n
cy.
On the
othe
r
hand,
CSF
pol
i
c
y
,
al
l
t
h
e
vi
rt
ual
m
achi
n
es are e
x
ec
ut
ed
on
di
f
f
ere
n
t
fre
que
ncy
dep
e
ndi
ng
o
n
vi
rt
ual
m
achi
n
e’s
reso
u
r
ce
dem
a
nd. T
o
c
o
m
p
are t
w
o
pol
i
c
i
e
s, we e
x
pect
t
h
at
t
h
e C
S
F
p
o
l
i
c
y
i
s
bet
t
e
r t
o
re
d
u
ce
po
we
r co
ns
um
pt
i
on
t
h
an
SSF policy.
Thr
o
ug
h
out
t
h
e expe
ri
m
e
nt
s, we anal
y
zed t
h
e p
o
w
er c
ons
um
pt
i
on an
d o
v
er
hea
d
. Es
pe
ci
al
l
y
, C
SF
policy spe
n
t less power t
h
an SSF polic
y. According to powe
r m
odel, po
we
r consum
ption is inc
r
ea
sed in
pr
o
p
o
r
t
i
on t
o
e
x
ecut
i
o
n t
i
m
e.
Ho
we
ver
,
p
o
w
e
r co
ns
um
pt
i
on i
s
i
n
crease
d
i
n
p
r
o
p
o
rt
i
o
n t
o
squa
re
of f
r
e
q
uency
.
It
m
eans, reducing freque
ncy is
m
u
ch
m
o
re profitable.
T
hus, the CSF policy is
m
o
re profitable because
CSF
uses
res
o
urce t
h
at unuse
d rem
a
ining exe
c
ution tim
e in SSF.
Based
on
th
e
propo
sed
system
, we will i
m
p
r
ov
e th
e system fu
rth
e
r. For
ex
am
p
l
e, we will ad
d
m
o
re
syste
m
p
o
licie
s su
ch
as task lev
e
l static. I
n
add
itio
n, we p
l
an
to
d
e
v
e
l
o
p
v
a
ri
o
u
s
power-aware sched
u
ling
algorithm
s
such as rate m
onotonic
on propos
ed system
and s
u
pport
m
u
lti-core m
echanism
that operate
s
po
we
r-a
ware
s
c
hed
u
lin
g at ea
ch c
o
re
.
REFERE
NC
ES
[1]
J.C. Song et al, "A
Strateg
y
fo
r Adop
ting Server Virtualization
in the Pub
lic Sector
: NIPA Co
mputer Center",
Journal of Infor
m
ation and Com
m
uni
cation
Con
vergence
Engineering,
vo
l. 10-1,
p.61-65, 2012.
[2]
S
.
Xi et al
, "
RT-Xen: Towards Real-tim
e Hy
pervisor Scheduling in Xen
", in Embedded Software (
E
MSOFT), 2011
Proceedings of
t
h
e Int
e
rna
tiona
l
Conferenc
e
on
. I
EEE
, p
.
39-48
, 2
011.
[3]
Xen H
y
pervisor, http:
//www.
xenproject.
o
rg/
[4]
M
.
Lee e
t
al
,
"
Supporting soft real-time ta
sks in the Xen
hypervisor
", i
n
P
r
oceeding o
f
the 6th ACM
S
I
GP
LAN/S
I
GOP
S
Internation
a
l Conf
eren
ce
o
n
Virtua
l
Execu
t
i
on Env
i
ronm
ent
s
,
p. 97-108, 201
0.
[5]
G.M. Tch
a
mgoue et al, "
Dynamic
Vo
ltage Scaling for
Po
wer-Awar
e Hierarchical
Real-
T
ime Schedu
lin
g
Framework
", in
P
r
oceeding of
1
5
th IEE
E
Int
e
rn
ation
a
l Confer
en
ce on Com
putat
i
onal S
c
i
e
nce
an
d Engine
ering
,
p
.
540-547, 2012
.
[6]
Y. Shin and K.
Choi, "
Power C
onscious Fixed
Priority Sch
e
duling for Hard Real-Time Systems
", in P
r
oceed
ing
of
36th Design Automation Confer
ence, p. 134-139,
1999.
[7]
P. Pillai and K.
G. Shin, "
Real-
T
ime Dynamic Voltage S
c
aling
for
Low Power
Embedded Operating Systems
", in
Proceeding
of 1
8
th ACM S
y
m
p
osium
on Operat
ing S
y
st
em
Prin
cipl
es, p
.
89-102
, 2001
.
[8]
J. Calandr
ino et
al. "
LITMUS RT: A Testbed for
Empiricall
y Co
mparing Real-Ti
m
e Multiproc
e
ssor Schedulers
", in
Proceedings of
t
h
e 27th
IE
EE R
e
al-T
im
e S
y
st
em
s S
y
m
posium
,
p.
111-123, 2006
.
[9]
L
.
Ni
u a
n
d G.
Qu
a
n
,
"
Reducing
both Dynamic a
nd Leakage En
ergy
Consumption for Hard Rea
l
-
T
ime Systems
", in
Proceeding
of In
terna
tiona
l Conf
erenc
e
on
Com
p
ilers, Archi
t
ec
tur
e
,
and S
y
nth
e
sis for Em
bedd
ed
S
y
stem
s, p. 140-
148, 2004
.
BIOGRAP
HI
ES OF
AUTH
ORS
Junho Seo is currently
a Ph.D
. student at
the De
p
a
rtment of Infor
m
atics in G
y
eon
g
sang National
Unive
r
sity
,
Jinju-si,
Sout
h Korea
,
where h
e
a
l
s
o
rece
ived his
B
a
chelor of Science and Master of
science degr
ees
in 2013
and 2
015, respectiv
ely
.
Hi
s research
inter
e
sts in
clude virtualization,
real-
tim
e s
y
s
t
em
s
,
and
power-aw
a
re
com
puting.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
A Pr
ot
ot
ype
of
Onl
i
n
e
Dy
n
a
mi
c Sc
al
i
n
g
Sc
he
dul
er
f
o
r
Real
-
T
i
m
e T
a
sk
b
a
s
e
d
on
Vi
rt
u
a
l
Mac
h
i
n
e
(
K
.H
. Kim)
21
1
K
y
ong Hoon Kim
received his
B.S., M.S., an
d
Ph.D. degree
s in Com
puter
Science an
d
Engineering fro
m POSTECH, Korea,
in 1998, 20
00,
2005, respectively
.
Since 2007, he has been
an assistant prof
essor at th
e Dep
a
rtment of
In
for
m
atics, G
y
eongs
ang Natio
n
a
l University
, Jinju
,
South Korea. Fr
om 2005 to 2007, he was
a post-doc
toral research fellow
at CLO
UDS lab in the
Department of
Computer Scien
ce
and Softwar
e
Engin
eer
ing,
the Univ
ersity
of Melbourne,
Australia. His research interests
include real-time sy
stems, avi
onics, Grid and Cloud computing,
and s
ecur
i
t
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.