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.
2
,
A
pr
il
2020
, p
p.
18
05
~
1813
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
2
.
pp1805
-
18
13
1805
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Reinf
orcement
l
earning
based mu
lti core
sch
eduli
ng
(RLB
M
CS)
for re
al tim
e systems
Dh
ru
va
R.
Ri
nku
1
,
M.
Asha
Rani
2
1
Depa
rtment of
El
e
ct
roni
cs
and
Com
m
unic
at
ion Engi
ne
eri
ng,
CV
R
Coll
eg
e
of
En
gine
er
ing, H
y
d
er
aba
d,
Indi
a
2
Depa
rtment of
El
e
ct
roni
cs
and
Com
m
unic
at
ion Engi
ne
eri
ng,
JN
T
Univer
si
t
y
,
H
y
der
aba
d
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ma
r 13
, 201
9
Re
vised
Oct
29
,
2019
Accepte
d
Nov 6, 2
019
Em
bed
ded
sy
stem
s
with
m
ul
ti
cor
e
process
or
s
are
increasin
gly
popula
r
beca
use
of
the
di
ver
si
ty
of
a
ppli
cat
ion
s
that
ca
n
be
r
un
on
it
.
In
this
work,
a
reinfor
cem
ent
le
arn
in
g
ba
sed
sche
duli
ng
m
et
ho
d
is
pro
po
se
d
to
ha
nd
le
the
real
tim
e
ta
sk
s
in
m
ulti
cor
e
syst
e
m
s
with
eff
ect
ive
CP
U
us
age
a
nd
lowe
r
res
ponse
tim
e.
The
pr
iority
of
the
ta
sk
s
is
vari
ed
dynam
ic
al
l
y
to
ensure
fai
rn
ess
with
rei
nfo
rcem
ent
le
arn
in
g
base
d
pr
io
rity
ass
ign
m
ent
and
Mult
i
Core
Mult
iLe
vel
Feed
back
qu
e
ue
(MCM
LFQ
)
to
m
anag
e
the
ta
sk
execu
ti
on
in
m
ulti
cor
e
syst
em
K
eyw
or
d
s
:
Mult
i cor
e
sch
edu
li
ng
Mult
il
evel feedback
que
ue
Re
inforcem
ent lea
rn
i
ng
Sy
m
m
e
tric
m
ulti
pr
ocess
ors
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
:
Dhruva
R. Rin
ku
,
Dep
a
rtm
ent o
f
Ele
ct
ro
nic
s
and C
omm
un
ic
ation
En
gin
ee
rin
g
,
CVR Coll
ege
of Enginee
rin
g
,
Hyde
rab
a
d, I
ndia
.
Em
a
il
:
rink
ud
hru
va.ravi
@g
m
ai
l.com
1.
INTROD
U
CTION
Sche
du
li
ng
is
t
he
pro
cess
of
al
locat
ion
of
t
asks
to
a
set
of
res
ources
under
ce
rtai
n
co
nst
raints
li
ke
dead
li
ne
,
util
iz
at
ion
,
respo
ns
e
tim
e
e
tc
.
In
R
TOS
(Real
Tim
e
Op
erati
ng
Syst
e
m
s),
ta
sk
sche
du
li
ng
is
a
cor
e
com
po
ne
nt
res
pons
i
ble
f
or
a
ssign
i
ng
res
ou
rces
to
the
ta
s
k
s
o
as
e
nsure
tim
e
gu
ara
nte
e.
The
cu
rr
e
nt
ta
sk
sche
du
li
ng
al
gorithm
s
in
RT
OS
ca
n
be
cl
as
sifie
d
i
nto
ti
me
-
dri
ve
n,
pri
ori
ty
-
dr
i
ven
a
nd
hy
br
id
-
dri
ve
n.
P
rior
it
y
dr
i
ven
sc
hedul
ing
assig
n
pri
ori
ti
es
to
ta
sk
s
in
su
ch
a
way
that
hig
h
pr
i
ori
ty
ta
sk
is
selected
for
sche
du
li
ng
ahead
of
lo
w
pri
or
it
y
ta
sk
s
.
T
he
pr
io
rity
can
be
assi
gned
i
n
two
ways
–
st
at
ic
and
dyna
m
ic
.
In
sta
ti
c
pri
or
it
y
schem
e,
the
pri
or
it
y
is
assig
ned
duri
ng
ta
s
k
creati
on
ti
m
e
an
d
it
does
no
t
c
ha
ng
e
du
rin
g
the
li
feti
m
e
of
the
ta
sk
.
I
n
dynam
ic
pr
iority
schem
es,
the
pri
or
it
y
of
ta
s
k
changes
dep
e
ndin
g
on
syst
em
sta
tus
and
r
eso
ur
ce
avail
abili
ty
.
Pri
or
it
y
inv
e
rsion
is
a
prob
le
m
in
pri
ori
ty
dr
iv
en
sc
hedulin
g.
The
pro
blem
a
rises
w
he
n
ta
sks
with
diff
e
re
nt
pri
ori
ti
es
sh
are
com
m
on
data.
Othe
r
iss
ues
li
ke
dead
l
ock
a
nd
m
ul
ti
ple
blo
c
ki
ng
of
ta
s
ks
is
quit
e
com
m
on
in
pr
iority
dr
i
ven
s
cheduli
ng.
M
ul
ti
pr
ocess
or
ar
chite
ct
ur
es
(e.
g.
,
Sym
m
et
ric
Mult
ipr
ocess
or
s
or
SMPs,
Sin
gle
Chip
Heter
oge
neous
Mult
ipr
ocess
or
s
or
S
CHMs)
are
in
creasin
gly
popula
r
to
cost
re
du
ct
io
n
and
div
e
rse
ap
plica
ti
on
s
that
can
be
r
un
on
it
.
Schedulin
g
in
Mult
ipr
ocess
or
syst
em
is
m
or
e
c
halle
ng
i
ng
as
it
mu
st
con
side
r
m
ult
i
par
am
e
te
r
opti
m
iz
at
i
on
li
ke
e
ff
ect
i
ve
resou
rce
ut
il
iz
ation
,
ta
s
k
res
pons
e
tim
e
e
tc
.
The
sc
he
du
li
ng crite
ria f
or
op
t
i
m
iz
ation
in
m
ulti
p
r
o
cess
sys
tem
is g
iven
in
Tab
le
1.
The
cu
rr
e
nt
scheduli
ng
al
gor
it
h
m
s
fail
fo
r
the
cases
of
uncertai
n
ta
sk
a
rr
ivals
a
nd
uncertai
n
ta
s
k
natu
re.
In
t
he
se
sit
uations,
there
is
a
n
in
crease
in
dea
dline
m
iss
rati
o
an
d
poor
r
eso
ur
ce
util
iz
at
ion
.
To
ov
e
rc
om
e
these
chall
e
nges,
the
sc
he
duli
ng
al
gorith
m
m
us
t
be
adap
ti
ve
an
d
m
us
t
be
sel
f
-
le
arn
i
ng
.
The
sc
heduler
m
us
t
l
earn
the
se
dynam
ic
s
fr
om
the
env
iro
nm
ent
and
m
u
st
adap
t
it
sel
f
to
ens
ure
there
is
ver
y
low
dead
li
ne
m
iss
rati
o
and
higher
resou
r
ce
util
iz
at
ion
.
This
w
ork
is
m
ot
ivate
d
by
above
prob
le
m
s
and
t
o
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
.
2
,
A
pr
i
l 202
0
:
1805
-
1813
1806
desig
n
a
sel
f
-
le
arn
i
ng
sc
he
du
l
er.
P
rev
i
ous
w
orks
on
sel
f
-
le
arn
i
n
g
sc
he
du
l
er
are
al
s
o
avai
la
ble
bu
t
no
e
xi
sti
ng
work
has
c
on
sidere
d
m
ulti
cor
e
syst
em
s
and
dynam
ic
ta
sk
ar
rivals
with
ada
ptati
on
ba
sed
on
MLFQ.
In
this
wor
k,
a
rein
forcem
e
nt
le
ar
ning
ba
sed
dynam
ic
pr
iority
sc
hem
e
with
m
ulti
le
vel
fee
dback
queue
i
s
us
e
d
f
or
sc
he
du
li
ng
t
he
ta
s
ks
to
m
ulti
pr
ocess
or
s
.
T
he
pr
i
or
it
y
to
al
locat
e
f
or
t
he
ta
sk
is
m
od
e
le
d
as
con
ti
nu
ous
le
arn
i
ng
act
ivit
y
with
rew
a
r
d
an
d
pe
nalty
score
al
locat
ed
bas
ed
on
the
m
ult
i
-
obj
ect
ive
para
m
et
e
r
op
ti
m
iz
ation
.
The
pro
posed
sche
du
le
r
is
i
m
ple
m
ented
i
nt
o
F
ree
-
RT
OS
operati
ng
syst
e
m
and
te
ste
d
f
or
it
s
eff
ect
ive
ness
with
job
m
ix
of
CP
U
a
nd
IO
intensi
ve
ta
s
ks
.
To
t
he
best
of
our
knowle
dge
,
this
w
ork
is
the
first
to consi
der
m
ulti
o
bj
e
ct
ive r
e
al
iz
at
ion
u
si
ng
MLFQ a
nd sel
f
-
le
ar
ning
on
MLFQ.
Table
1.
Sche
du
li
ng c
rite
ria
CPU Utiliz
atio
n
Keep th
e CPUs
as
b
u
sy
as p
o
ss
ib
le.
Maxi
m
iz
e the CP
U Utilizat
io
n
Thro
u
g
h
p
u
t
N
u
m
b
e
r
o
f
task
s c
o
m
p
leted
pe
r
u
n
it t
i
m
e.
M
ax
i
m
ize
th
e thro
u
g
h
p
u
t
Turn
Ar
o
u
n
d
T
i
m
e
Ti
m
e
to ex
ecut
e a
p
articular
task
f
ro
m
su
b
m
iss
io
n
to c
o
m
p
letio
n
.
Mini
m
i
ze the turn
arou
n
d
ti
m
e
W
aitin
g
T
i
m
e
A
m
o
u
n
t of
ti
m
e pr
o
cess
is waiting
in
th
e r
eady
qu
eu
e.
Mini
m
ize
the wai
ti
n
g
ti
m
e
Res
p
o
n
se Ti
m
e
A
m
o
u
n
t of
ti
m
e it
tak
es f
ro
m
when
a
requ
est was su
b
m
it
ted
un
til
th
e f
irst r
e
sp
o
n
se is p
rod
u
ced
.
Mini
m
ize
the r
esp
o
n
se
ti
m
e.
2.
LIT
ERATUR
E SU
RV
E
Y
The
existi
ng
s
olu
ti
ons
sche
duli
ng
in
RT
O
S
is
analy
zed
belo
w,
Seife
di
ne
Ka
dr
y
et
al
.
[1
]
ap
plied
dynam
ic
qu
ant
um
instea
d
of
sta
ti
c
qu
ant
um
in
Mult
i
L
evel
Feed
bac
k
Qu
e
ue
(ML
F
Q)
base
d
sche
du
li
ng.
The
schem
e
was
able
to
accom
m
od
at
e
l
ow
pr
io
rity
task
s,
wh
ic
h
ta
kes
longe
r
ti
m
e
fo
r
com
pleti
on
an
d
m
axi
m
iz
e
the
resou
rce
util
iz
at
ion
.
Xiao
j
ie
Li
et
al
.
[2
]
pro
posed
im
pr
o
ved
ED
F
al
go
rithm
wh
ic
h
r
edu
c
e
s
the
ta
sk
switc
hing
over
hea
d
and
num
ber
of
ta
sk
s
witc
h.
Con
tr
olled
p
reem
ption
is
pro
po
se
d
b
y
J
ink
y
u
Lee
et
al
.
[3
]
,
wh
ic
h
regulat
es
the
pr
eem
ption
f
or
bette
r
EDF
.
T
he
ap
proa
c
h
ap
plies
heurist
ic
s
to
deci
de
a
bette
r
value
of
preem
ption
tim
e
fo
r
eac
h
ta
sk
in
case
of
Un
i
pr
ocess
or
syst
e
m
s.J.
Lee
et
al
.
[4
]
pr
op
os
e
d
a
sche
duli
ng
poli
cy
to
m
ake
eff
ect
ive
us
e
of
c
on
te
ntion
f
r
ee
slots.
Tas
ks
in
c
onte
ndin
g
slots
a
re
m
igr
at
ed
t
o
con
te
ntion
f
ree
slots
to
red
uc
e
the
nu
m
ber
of
com
peting
ta
sk
s
in
co
ntent
ing
slots.
J.
Le
e
et
a
l.
[5
]
pr
opose
d,
two
im
po
rtant
aspects:
inter
-
job
c
on
c
urre
nc
y
and
job
ur
ge
ncy
are
cons
id
ered
f
or
sc
he
duli
ng.
A
naly
sis
of
LLF
(Least
La
xity
First)
sc
he
du
li
ng
on
m
ulti
process p
la
tf
or
m
s
is
co
ns
i
der
e
d
in [
6] b
y
J.
Le
e
et
al
.
La
xity
of
a
ta
s
k
at
any
tim
e
is
def
i
ned
as
rem
ai
nin
g
tim
e
to
dead
li
ne
m
inu
s
the
am
ou
nt
of
rem
ai
nin
g
e
xe
cution.
Algori
th
m
s
e
m
plo
yi
ng
Lax
it
y fo
r sche
duli
ng are
call
ed
Z
L
-
ba
sed
sch
e
duli
ng alg
or
it
hm
.
S.
K.
Lee
propose
d,
ED
ZL
sc
hedulin
g
[
7]
em
plo
ys
the
EDF
strat
egy
un
ti
l
ta
sk
s
hav
e
ze
r
o
-
la
xity
and
the
ZL
po
li
cy
.
M.
L.
De
rto
uz
os
et
al
.[8]
pr
opos
e
d,
Lea
st
Laxit
y
First
(L
LF)
sc
he
d
ulin
g
is
a
ZL
sc
he
du
li
ng
al
gorithm
,
wh
ic
h
globall
y
assigns
a
hi
gh
e
r
pr
i
or
it
y
to
a
t
ask
with
lo
we
r
la
xity
.
A
.
Ea
swar
a
n
et
al
.
s
how
n
Cl
us
te
r
-
base
d
sche
du
li
ng
assi
gn
s
eac
h
ta
sk
to
process
or
cl
us
te
r.
Tas
ks
in
each
cl
us
te
r
are
globall
y
sched
ul
e
d
a
m
on
g
them
sel
ve
s,
a
nd
cl
ust
ers
in
tu
rn
are
sche
dule
d
on
the
m
ulti
pr
oc
es
so
r
platfo
rm
.
Cl
us
te
ri
ng
pl
aces
a restrict
io
n on the am
ou
nt
of
con
c
urre
ncy.
[
9]
Ma
rko
Be
rto
gna
et
al
.
[1
0]
pro
po
se
d,
A
li
m
it
ed
-
pr
eem
ption
sche
duli
ng
te
chn
iq
ue
with
be
nef
it
s
of
bo
t
h
preem
pti
ve
an
d
non
-
pr
ee
m
ptive
sche
du
li
ng.
T
he
r
untim
e
beh
avi
or
of
pr
eem
ptive
EDF
sche
du
li
ng
is
i
m
pr
oved
by
a
vo
i
ding
pr
eem
ptions
that
are
no
t
nee
de
d
to
sat
isfy
the
sc
he
du
la
bili
ty
of
t
he
syst
em
.
Algorithm
for
aut
om
at
ic
a
l
ly
set
ti
ng
a
num
ber
of
preem
ption
points
within
the
c
ode
of
each
ta
s
k,
to
m
ini
m
iz
e
the
overal
l
pr
eem
ption
c
ost
unde
r
sc
he
du
la
bili
ty
const
raints
is
pro
po
s
ed
by
M.
Be
rto
gn
a
et
al
.
in
[11].
A
ta
sk
deco
m
po
sit
io
n
al
go
rithm
is
pr
op
os
ed
by
Abusayeed
Saif
ul
la
h
et
al
.to
deco
m
po
se
each
pa
rall
el
ta
sk
into
a
set
of
se
qu
e
nti
al
t
asks
with
deadl
ine
distribu
te
d
fo
r
the
se
qu
e
nt
ia
l
ta
sk
s.
EDF
scheduli
ng
is
execu
te
d
f
or
e
ach
of
the se
qu
e
ntial
tasks [1
2].
An
ef
fici
ent
s
cheduli
ng
a
pp
ro
ac
h
is
pro
po
sed
by
Kaly
an
Ba
it
al
et
al
.
[13]
in
t
he
he
te
rogen
e
ous
m
ul
ti
cor
e
syst
e
m
to
m
ap
the
ta
sk
s
int
o
a
ppr
opriat
e
co
res
base
d
on
t
he
ty
pe
of
co
res
(b
i
g
hi
gh
po
w
er
an
d
sm
a
ll
low
po
wer)
as
well
a
s
ty
pe
of
ta
s
ks
(lo
w
powe
r
and
high
po
w
er)
.
A
r
un
-
ti
m
e
disp
at
ch
er
di
sp
at
ch
diff
e
re
nt
jo
bs
i
nto
sm
al
l
cor
e
s.
Tas
k
m
igrati
on
is
use
d
to
m
igrate
the
re
je
ct
ed
jo
bs
fro
m
s
m
al
l
cor
es
into
big
cor
es
thro
ugh dispatc
her
.
K.
Ba
it
al
et al
. [
14]
presente
d
a
sche
du
li
ng al
gorithm
w
he
re random
tasks gener
at
e
d
at
diff
e
re
nt
tim
e
interv
al
s
wit
h
diff
e
ren
t
pe
riod
ic
it
y
an
d
e
xe
cution
tim
e
can
be
acc
ommod
at
e
d
int
o
a
s
yst
e
m
,
wh
ic
h
is
al
rea
dy
r
unning
a
s
et
of
ta
s
ks
,
m
e
et
ing
the
dea
dl
ine
crit
eria
of
t
he
ta
sks.
Using
the
co
nce
pt
of
Pf
ai
r
sche
du
li
ng,
r
a
ndom
n
ew
tasks
h
a
ve been
d
i
vi
ded to
fit i
nto t
he
idle t
im
es o
f
the
d
if
fer
e
nt
cor
es
of t
he
sy
stem
.
Gen
et
ic
Algo
r
it
h
m
based
scheduli
ng
of
ta
sk
s
to
m
ult
i
-
pr
oces
sor
syst
em
is
pr
opos
e
d
by
A.
S
.
Ra
dh
am
ani
et
al
.
in
[
15
]
with
the
obj
e
ct
ive
of
m
inim
iz
ing
the
tot
al
com
pleti
on
tim
e
of
al
l
t
ask
a
nd
m
axi
m
iz
ing
the
util
iz
at
ion
of
pr
oc
esso
r.
Fit
ness
f
un
ct
i
on
i
s
def
ine
d
ba
se
d
on
obj
ect
ive
and
gen
et
ic
al
gorith
m
to u
se to sea
rc
h
f
or
the
pros
pe
ct
ive so
luti
on
s
m
at
ching
the
obj
ect
ives
. J
. R
egehr in
[
16]
sh
ow ho
w
to sc
hedule
two
novel
sch
edu
li
ng
a
bs
tra
ct
ion
s
that
ov
erco
m
e
lim
it
a
t
ion
s
of
e
xisti
ng
w
ork
on
preem
ption
thre
sh
ol
d
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
Rei
nfo
rce
men
t
le
ar
ni
ng base
d m
ulti
co
re sc
he
du
li
ng
(
RL
B
MCS)
for
real t
ime syste
m
s…
(
Dh
r
uva R
.
Rin
ku
)
1807
sche
du
li
ng.
Th
e
i
m
po
rtant
w
eakn
e
ss
of
ML
FQ
sc
he
du
le
r
is
the
ina
bili
ty
to
de
fine
th
e
optim
iz
ed
nu
m
ber
of
the
qu
e
ues
an
d
quant
um
of
each
queue
al
s
o
this
al
gorith
m
has
sta
rv
at
ion
f
or
so
m
e
pr
oces
s.
These
factors
aff
ect
the
respon
s
e tim
e d
irect
ly
. I
n
[
17
]
Pa
r
var
M
oh
am
m
a
d
et
al., a
ne
w al
gorithm
is p
res
ented
t
o
so
l
ve
these
pro
blem
s
and
m
ini
m
iz
e
the
r
esp
on
se
tim
e
si
m
ultaneou
sly
.
The
pro
pose
d
idea
to
so
l
ve
t
he
sta
r
vatio
n
pro
ble
m
wh
il
e
co
ns
i
deri
ng
the
process
es
pr
i
or
it
y
too.
In
this
al
gorithm
Re
cur
ren
t
Neural
Net
wor
k
has
bee
n
util
iz
ed
to
fin
d
bo
t
h
the
nu
m
be
r
of
que
ues
a
nd
the
optim
iz
ed
quant
um
of
eac
h
queue
.
K.
Ba
it
a
l
et
al
.
[
18
]
presents
a sche
du
li
ng
al
gorithm
w
her
e
rand
om
tasks gener
at
e
d
at
d
i
fferent ti
m
e intervals w
it
h di
ff
e
r
ent p
e
rio
dicit
y and
execu
ti
on
ti
m
e
can
be
ac
co
m
m
od
at
ed
int
o
a
syst
em
,
wh
ic
h
is
al
rea
dy
r
unni
ng
a
set
of
ta
sk
s
,
m
eet
ing
the
dead
li
ne
c
rite
ria
of
the
ta
sk
s.
A
hybri
d
lim
i
te
d
-
pree
m
pt
ion
real
-
ti
m
e
sched
ulin
g
al
go
rithm
i
s
der
i
ved
in
[
19
]
by
M.
Be
rtogna
et
al
.
,
that
ai
m
s
to
ha
ve
lo
w
r
unti
m
e
ov
er
hea
d
wh
il
e
sc
heduling
al
l
syst
em
s
that
can
be
s
che
dule
d
by
fu
ll
y
pr
eem
ptive
al
go
rit
hm
s
.
This
hy
br
i
d
al
gorithm
per
m
i
ts
pr
eem
ption
wh
e
re
nece
ssar
y
fo
r
m
ai
ntaining
fe
asi
bili
ty
, b
ut at
tem
pts to
av
oi
d unnecessa
ry
pr
eem
ption
s
duri
ng run
ti
m
e.
The
pr
ob
le
m
of
sche
duli
ng
pe
rio
dic
real
-
ti
m
e
ta
sk
s
on
m
ulti
process
o
rs
unde
r
the
f
ork
j
oi
n
str
uctu
re
us
e
d
in
O
penM
P
has
bee
n
discusse
d
by
K.
La
kshm
anan
et
al
.
[20].
A
n
I
nte
ger
Line
ar
P
rogr
am
m
i
ng
(
ILP)
form
ulati
on
an
d
tw
o
non
-
it
er
at
ive
he
ur
ist
ic
s
f
or
sche
duli
ng
a
ta
s
k
-
base
d
a
pp
li
cat
io
n
on
t
o
a
heter
og
eneous
m
any
-
cor
e
arc
hitec
ture
has
de
scribe
by
C.
Tan
et
al
.[21
]
.
Pr
ic
opi
M
et
a
l.
pro
posed
work
[
22
]
t
o
pro
ve
that
adap
ti
ve
m
ulti
-
cor
e
arc
hitec
tu
res
can
substan
ti
al
l
y
decr
ease
the
m
akesp
a
n
com
par
ed
to
both
sta
ti
c
sy
m
m
et
ric
and
a
sym
m
et
ri
c
m
ulti
-
cor
e
a
r
chite
ct
ur
es
.
A
.
Hatam
i
e
t
al
.
[23]
pro
posed
sche
du
li
ng
al
gorithm
that
able
to
perform
bette
r
than
tra
diti
on
al
Earli
est
Dead
li
ne
First
(EDF
)
and
m
ini
m
iz
e
the
ov
erall
co
m
ple
ti
on
tim
e
wh
e
n
the
syst
em
in
ov
e
rloa
de
d
c
onditi
on.
T
V
AIR
AM
et
al
.[24]
pr
op
os
e
d
w
or
k
w
hic
h
prese
nt
s
a
hybri
d
ve
r
sion
of
DP
S
O(Discret
e
Partic
le
Sw
ar
m
Op
tim
iz
a
ti
on
)
w
hich
is
a
com
bin
at
ion
of
DP
S
O
an
d
Cy
ber
S
war
m
Algo
rithm
(CSA).
The
ef
f
ic
ie
ncy
of
the
al
gorithm
is
tested
us
in
g
the
perform
ance
m
et
rics
su
c
h
as
m
akesp
a
n,
m
ean
flo
w
tim
e,
reli
abilit
y
cost,
fitness
and
res
ource
util
iz
at
ion
.
T
he
pr
opos
e
d
An
t
Col
ony
Op
ti
m
iz
ation
(A
C
O)
al
gorithm
s
pr
ov
i
de
inh
e
re
nt
par
al
le
li
s
m
,
wh
ic
h
is
ve
ry
us
ef
ul
in
m
ul
ti
pr
ocess
o
r
env
ir
on
m
ents
[2
5].
An
op
ti
m
iz
ed
real
tim
e
sched
ulin
g
al
gorit
hm
fo
r
ta
sk
s
al
locat
ion
in
gr
i
d
env
i
ronm
ent
has
pro
po
sed
by
Aghaza
rian
V
et
al
.[
26
]
.
Sc
he
du
li
ng
of
s
ha
red
m
e
m
or
y
with
m
ulti
-
cor
e
perform
ance
i
n
dynam
ic
al
l
ocato
r
us
in
g
a
rm
pr
oc
essor
ha
s
im
pl
e
m
ented
by
Ve
nk
at
a
Siva
P
ra
sad
C
h.
et
al
.
[
27
]
.
Sur
vey
of
diff
e
re
nt
sche
duli
ng
al
gorithm
s h
av
e b
ee
n discus
s
ed by
Im
ran
Q
ur
es
hi
[28].
3.
RLBM
CS
SCHE
DU
LI
NG
Diff
e
rin
g
f
r
om
pr
e
vious
s
olut
ion
s
f
oc
us
in
g
on
ly
on
dea
dline,
the
pro
posed
RLB
MC
S
sche
du
li
ng
,
sche
du
le
s tas
ks t
o
m
ulti
co
res i
n
a
way to
optim
iz
e
m
ult
iple objecti
ves
of
,
-
CPU
Util
iz
at
ion
(C
)
-
Thro
ughput
(T
)
-
Turna
rou
nd ti
m
e (TA)
-
Wait
ing
ti
m
e (W
T)
-
Re
sp
onse
tim
e (RT)
-
Dead
li
ne
m
eet
rati
o
(
DR)
3.1.
Rein
f
orc
ement le
arnin
g
Re
inforcem
ent lea
rn
in
g
is a
ki
nd
of
m
achine
learnin
g
te
c
hniqu
e
wh
e
re a
n agent lear
ns
a
nd f
i
ne
-
t
un
es
it
s
beh
a
vior
i
n
en
vir
on
m
ent
base
d
on
the
r
esults
of
it
s
pa
st
act
ion
s
.
It
is
goal
-
or
ie
nted
al
gorithm
gu
id
ed
by
pr
i
nciple
of
r
ei
nfor
cem
ent
with
re
wards
for
rig
ht
act
io
ns
an
d
pen
al
t
y
fo
r
wrong
act
ion
s.
By
th
is
way,
it
at
ta
ins
it
s
obj
ect
ive.
Re
inf
orcem
ent
le
arn
ing
is
m
od
el
le
d
in
te
rm
s
of
ag
ents,
e
nviro
nm
ents,
sta
te
s,
ac
ti
o
ns
and re
wards
as
shown i
n
Fi
gu
re
1.
Figure
1. Re
inf
or
cem
ent lea
rni
ng
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
.
2
,
A
pr
i
l 202
0
:
1805
-
1813
1808
-
Ag
e
nt: A
ge
nts
ta
kes
act
io
ns
.
-
Acti
on (A):
A
i
s the set
of all
po
s
sible m
ov
es
the a
gen
t
can
m
ake.
-
En
vironm
ent:
The
a
ge
nt m
ov
es in t
his wo
rld
.
-
Stat
e(S)
:
A
sta
t
e is a c
on
c
rete
and
im
m
ediat
e
sit
uation i
n w
hich
t
he
a
ge
nt
fin
ds
it
sel
f.
-
Re
ward
(R):
A
r
e
ward is t
he
f
eedb
ac
k by
whic
h
we
m
easur
e the s
uccess
or f
ai
lu
re
of an
agen
t
’s
act
io
ns
Re
inforcem
ent
le
arn
i
ng
re
pr
e
sents
a
n
a
ge
nt’
s
at
tem
pt
to
a
ppr
ox
im
at
e
the
en
vir
on
m
ent’s
f
un
ct
io
n,
s
uch
that
act
ion
s
can
be
sent
into
the
black
-
box
e
nv
i
ronm
ent
that
m
axi
m
iz
es
the
rew
a
rd
s
it
sp
it
s
ou
t.
It
le
arns
seq
uen
ces
of a
ct
ion
s t
hat w
il
l l
ead a
n
a
ge
nt to
ac
hieve
it
s
goal
or m
axi
m
ize
it
s obj
ect
ive
fun
ct
io
n.
3.2.
RLB
MCS
The
arc
hitec
tu
re
of
the
propose
d
RLB
MC
S
scheduli
ng
a
lgorit
hm
is
giv
en
in
Fig
ur
e
2.
Disp
at
c
her
runs
i
n
on
e
c
ore
an
d
it
is
res
pons
i
ble
f
or
sc
he
du
li
ng
of
ta
s
ks
to
t
he
m
ulti
-
cor
es
.
A
global
m
ul
ti
le
vel
feedback
qu
e
ue
is
us
e
d
to
place
the
ta
sk
s
an
d
m
ulti
-
cor
e
ret
rieves
the
ta
sk
s
from
n
on
-
em
pty
posit
ion
of
the
Q
ueu
e
.
The
num
ber
of
slots
to
be
a
ll
ocated
to
the
ta
sk
on
the
proces
sor
is
de
pende
nt
on
th
e
qu
e
ue
f
ro
m
wh
e
re
the
ta
sk
is
fet
ched.
By
this
way,
the
pr
ee
m
pt
ion
inter
va
l
is
var
ie
d
for
each
ta
sk.
M
os
t
of
the
pr
e
vious
so
luti
ons,
perf
or
m
ed
poorl
y
because
of
to
o
m
any
pr
eem
ption
s
an
d
perf
or
m
ance
was
i
m
pr
ov
e
d
if
dy
nam
i
c
pr
eem
ption
peri
od
ca
n
be
c
hosen
base
d
on
t
ask
natu
re
a
nd
pr
eem
ption
is
con
t
ro
ll
ed
.
F
rom
this
po
i
nt
of
vie
w
,
the
pr
eem
ption
pe
rio
d
is
m
ad
e
disc
rete
in
th
e
pro
pose
d
sys
tem
by
al
locat
in
g
di
ff
e
ren
t
ti
m
e
qu
a
ntu
m
for
eac
h
qu
e
ue
a
nd
proc
essor
all
ocate t
hat tim
e q
uan
tum
to the task
base
d
on the queu
e
from
w
hich
the task is fe
tc
hed.
The
inc
om
ing
ta
sk
s
an
d
the
pr
eem
pted
ta
sk
s
are
sch
ed
ul
ed
by
the
dis
pa
tc
her
to
m
ultilevel
feedback
qu
e
ue
base
d
on
the m
ulti
o
bject
ive optim
iz
at
ion
r
eal
iz
ed
us
in
g
Re
inf
or
cem
ent lea
rn
i
ng. Th
e Rei
nfor
cem
ent lea
rn
i
ng
is
m
app
ed
t
o
RLB
MC
S
as
f
ollows.
Ba
sic
a
ll
y,
reinfor
cem
ent
le
arn
i
ng
is
base
d
on
the
envo
rironm
ent,
agen
t
,
sta
te
,
act
ion
a
nd
re
ward
bas
is.
To
m
ap
the
reinfo
rce
m
ent
le
arn
in
g
a
pp
li
cat
ion
to
our
s
cheduli
ng
the
sam
e
hav
e
b
ee
n
m
ap
ped as s
how
n
i
n
Ta
ble
2.
Figure
2.
RLBM
CS
arc
hi
te
c
ture
Table
2.
Decla
rati
on of
obj
ect
s for
rein
f
or
ce
m
ent lea
rn
in
g
Env
iron
m
en
t
Multico
re
P
rocess
o
r
Ag
en
t
Disp
atch
er
State
Fo
u
r
states
ar
e def
in
ed
1
.
Initial Stat
e(S1)
2
.
Ob
jectiv
e Degr
a
d
atio
n
stag
e(S2)
3
.
Ob
jectiv
e Pr
o
g
r
ess
io
n
Stage(S3
)
4
.
Ob
jectiv
e Stabil
izatio
n
Stage(S4
)
Actio
n
Mapp
in
g
the task
to o
n
e of
the N qu
eu
es
Reward
The v
alu
e of
the
m
u
lti ob
jectiv
e f
u
n
ct
io
n
.
The
m
ulti
obj
ect
ive
fun
ct
ion
is m
odel
l
ed
as
,
O=w1
*C+w
2*T
-
w
3*(
[T
A/TAe]
-
1)
-
w
4*(
[
WT/
W
T
e]
-
1)
-
w5*([RT/
RTe
]
-
1)
+
w
6*DR
TAe is t
he
S
L
A
(
Ser
vice le
ve
l agr
eem
ent) defi
ned turna
round
ti
m
e
WTe is t
he
S
L
A defi
ned w
ai
t
ing
ti
m
e
RTe
is the
SL
A defi
ned res
ponse
ti
m
e
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
Rei
nfo
rce
men
t
le
ar
ni
ng base
d m
ulti
co
re sc
he
du
li
ng
(
RL
B
MCS)
for
real t
ime syste
m
s…
(
Dh
r
uva R
.
Rin
ku
)
1809
The
weig
hts
va
lues
(
w1,
w2…w
6)
a
re
ass
ign
e
d
ba
sed
on
the
im
po
rta
nce
to
t
he
pa
r
a
m
et
ers
as
def
i
ned
by
envi
ronm
ent.
Say
if
DR
is
i
m
po
rtant,
the
w
6
val
ue
will
be
giv
e
n
a
higher
val
ue
com
par
ed
to
oth
e
r
weig
hts.
F
or
e
ach
it
erati
on
,
we
get
the
ne
xt
O
value,
it
cou
ld
be
posi
ti
ve
in
case
of
bette
r
or
neg
at
ive
in
case
of
worse
scen
ario,
acc
ordin
gly
the
sta
te
would
be
cha
ng
e
d.
T
he
sta
te
transiti
on
is
sh
ow
n
in
T
able
3.
The
val
ue
of
M
is
the
deg
re
e
of
co
nf
i
de
nc
e
fo
r
m
igrati
ng
from
S2
to
S3
.
T
he
val
ue
of
P
is
the
de
gr
ee
of
tolerance
from
S3
to
S
2
a
nd
t
ran
sit
io
n
f
ro
m
S4
to
S2.
The
value
of
T
is
t
he
de
gree
of
sta
bili
ty
fo
r
m
ig
rati
ng
from
S3
to
S4
.
Ther
e
a
re
N
a
ct
ion
s
co
rresp
ondi
ng
t
o
al
loca
ti
on
of
ta
sk
to
any
of
t
he
N
queue
.
At
eac
h
sta
te
,
the
act
ion
t
o
be
invok
e
d
is done
with
c
onstra
ints
on
the
ta
sk
natu
re
(its
CP
U/IO
r
at
io
),
th
e
tim
e
it
has
spe
nt
s
o
far
,
dea
dline
et
c.
Ba
sed
on
t
he
co
ns
trai
nts
,
a
ct
ion
to
place
the
ta
sk
to
c
or
respo
nd
i
ng
qu
eue
is
im
ple
m
ented,
wh
ic
h
a
re
giv
e
n
in
Ta
ble 4.
Table
3.
Stat
e transiti
on
of tasks
from
on
e
qu
eue to an
othe
r qu
e
ue
S1
S2
S3
S4
S1
X
Neg
ativ
e
O
v
alu
e
Po
sitiv
e O
v
alu
e
X
S2
X
Neg
ativ
e O
v
alu
e
Po
sitiv
e O
v
alu
e ov
er
last
M
step
s
X
S3
X
Neg
ativ
e O
v
alu
e
o
v
er
last P step
s
Po
sitiv
e O
v
alu
e
Po
sitiv
e O
v
alu
e ov
er
last T
step
s
S4
X
Neg
ativ
e O
v
alu
e
o
v
er
last P step
s
X
Po
sitiv
e O
v
alu
e
Table
4.
Acti
ons t
o be take
n by ta
sk
s
A1
Fo
r
a new
task
whi
ch
jus
t ar
rived
allo
cate to Q0
(hig
h
est p
riority
)
A2
Fo
r
task
pree
m
p
t
e
d
,
allo
cate to on
e l
ev
el belo
w its last
Qu
eu
e allocated
A3
Fo
r
task
pree
m
p
t
e
d
,
allo
cate to on
e l
ev
el abo
v
e its last
Qu
eu
e allocated
A4
Fo
r
task
pree
m
p
t
e
d
,
allo
cate to sa
m
e
level o
f
last Queu
e allocated
The
act
io
ns
ar
e
done
in
eac
h
sta
te
f
or
al
loc
at
ion
of
ta
sk
s
to
the
co
rr
e
spondin
g
que
ue
and
the
best
po
s
sible ac
ti
on
s to be take
n
for
achie
ving t
he
stable sta
te
of
S4
is lea
rn
t
for
d
if
fer
e
nt n
at
ure of jo
bs
a
nd a
rr
ival
in
a
te
st
env
ir
on
m
ent.
The
r
ules
le
arn
t
are
then
use
d
i
n
producti
on
r
un
f
or
ta
s
ks
in
that
correspo
nd
i
ng
env
i
ronm
ent. Th
e
ps
e
udo
c
ode
of the lea
rn
i
ng pr
ocedur
e
is g
i
ven
bel
ow
,
The
proposed
R
LBMCS
sche
dule
r
is
re
alize
d
a
s
bel
ow.
Th
ere
are
5
qu
eu
es
de
fine
d
[Q0,
Q1,
Q2,
Q3,
Q4]
.
All
ta
sks
whi
ch
arr
ive
as
eve
nts
are
f
irst
put
in
Q0.
Tas
ks
in
Q
0
are
ta
ke
n
by
sche
du
le
r
an
d
pr
i
or
it
y
of
the
ta
sk
is
cal
culat
ed
us
i
ng
the
propose
d
RLM
BC
S
pr
i
or
it
y
cal
culat
ion
al
gorithm
.
The
pri
ori
ty
is
fr
om
1
to
4.
Ba
sed
on
pr
i
or
it
y,
the
ta
sk
s
a
re
al
locat
ed
to
que
ues.
Hen
ce
Q
1
gets
the
ta
sks
of
pri
or
it
y
1,
Q
2
ge
ts
pr
i
or
it
y
of
2,
an
d
s
o
on.
Eac
h
of
Q
ueu
e
from
Q1
to
Q4
is
al
lo
cat
ed
to
one
proc
es
sor
each
.
Fo
r
eac
h
pr
oc
esso
r,
the
re
is
a
ta
sk
routine
w
hich
ta
kes
the
ta
sk
s
fr
om
their
correspo
nd
i
ng
ta
s
ks
an
d
us
es
it
for
exec
ution
i
n
Rou
nd
Ro
bin
with
a
fixed
qu
a
ntum
.
The
slots
are
al
locat
e
d
to
process
or
of
ea
ch
queue
as
,
3
m
s
fo
r
Q
1,
2.5
m
s
fo
r
Q2,
2
m
s
fo
r
Q3,
an
d
1.5
m
s
fo
r
Q
4.
T
he
ta
sk
s
in
Q
1
are
giv
e
n
m
or
e
tim
e
slot
fo
r
exec
utio
n
tha
n
ta
sks
in
Q
4.
Af
te
r
the tim
e sli
ce com
pleti
o
n,
t
he
ta
s
ks
a
re
ag
ai
n se
nt to Q0 i
f n
ot co
m
plete
d
.
4.
PERFO
R
MANC
E
RES
UL
T
The
pro
posed
RLB
MC
S
s
cheduli
ng
is
i
m
ple
m
ented
on
F
ree
RT
OS
operati
ng
syst
e
m
an
d
the
perf
or
m
ance
of
the
s
ol
ution
is
e
valuated
a
gainst
MLFQ
sc
he
du
l
in
g
al
go
rithm
pr
opos
e
d
in
[1
]
.
The per
f
or
m
ance is co
m
par
e
d i
n
te
rm
s o
f
,
-
Av
e
ra
ge
tu
rn
a
r
ound ti
m
e
-
Com
pleti
on
ti
m
e
-
Re
sp
onse
tim
e
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
.
2
,
A
pr
i
l 202
0
:
1805
-
1813
1810
In
Fr
ee
RTO
S,
one
th
rea
d
is
m
ade
to
s
pa
wn
ta
sk
s
with
dif
f
eren
t
pr
i
or
it
y
a
nd
exec
utio
n
ti
m
e
at
diff
e
ren
t
rates
per
sec
onds
an
d
these
ta
s
ks
a
re
exec
uted
on
the
co
res
a
vaila
ble
in
the
m
achine.
The
te
st
is
cond
ucted
wit
h
fo
ll
owin
g para
m
et
ers
as
sho
w
n
in
T
a
ble 5 an
d
the
r
es
ults
obta
ined
a
re tab
ul
at
ed
in
T
a
bles
6,
7, an
d 8.
Table
5.
In
it
ia
li
zat
ion
No
of
pro
cess
o
rs
2
to 4
Nu
m
b
e
r
o
f
T
ask
s
100
Qu
eu
e Size
10
SLOT
I
n
terval
10
m
ill
iseco
n
d
s
Table
6.
T
he
perf
or
m
ance r
es
ult f
or
t
he
case
of
2
CP
U
X
RLBMC
ML
FQ
Av
erage turn
arou
n
d
ti
m
e(s
)
90
276
Co
m
p
letio
n
ti
m
e(s
)
801
1231
Res
p
o
n
se ti
m
e
(
m
s
)
365
4
8
9
.3
Table
7.
T
he
perf
or
m
ance r
es
ult f
or
t
he
case
of
3
CP
U
X
RLBMC
ML
FQ
Av
erage turn
arou
n
d
ti
m
e(s
)
4
6
.79
1
4
5
.80
Co
m
p
letio
n
ti
m
e(s
)
541
911
Res
p
o
n
se
ti
m
e
(
m
s
)
245
4
0
7
.8
Table
8.
T
he
perf
or
m
ance r
es
ults f
or the cas
e of
4
CP
U
X
RLBMC
ML
FQ
Av
erage turn
arou
n
d
ti
m
e(s
)
1
9
.2
90
Co
m
p
letio
n
ti
m
e(s
)
411
721
Res
p
o
n
se ti
m
e
(
m
s
)
1
8
1
.2
3
3
7
.2
The
ave
ra
ge
tur
naro
und
ti
m
e
fo
r
dif
fer
e
nt
nu
m
ber
of
proces
sor
s
is
pl
otted
bel
ow
i
n
Fig
ur
e
3.
The
a
ver
a
ge
t
urnaro
und
ti
m
e
in
the
pr
opos
e
d
RLB
MC
S
sche
duli
ng
i
s
far
bette
r
than
t
hat
of
MLFQ.
The
reas
on
f
or
this
perform
a
nce
is
due
to
a
dap
ti
ve
m
ov
e
m
ent
of
the
Ta
sk
s
ac
r
os
s
t
he
Qu
e
ues
in
RL
BM
CS
wh
e
n
com
par
e
d
to
co
ntinuo
us
downg
rad
e
of
pr
i
or
it
y
in
M
LFQ
.
T
he
com
pleti
on
tim
e
of
the
ta
sk
s
fo
r
var
ie
d
num
ber
of pr
oc
esso
rs
is
giv
e
n
in
Fig
ure
4.
Figure
3. Com
par
is
on of t
urn
arou
nd tim
e for
RLB
MC
S w
it
h M
LFQ
Figure
4. com
par
iso
n of t
asks
com
pleti
on
tim
e
fo
r
RLB
MC
S w
it
h M
LFQ
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
Rei
nfo
rce
men
t
le
ar
ni
ng base
d m
ulti
co
re sc
he
du
li
ng
(
RL
B
MCS)
for
real t
ime syste
m
s…
(
Dh
r
uva R
.
Rin
ku
)
1811
Fr
om
the
res
ul
ts,
the
c
om
pletio
n
ti
m
e
in
RLB
MC
S
is
lowe
r
tha
n
that
of
MLFQ
t
her
e
by
increasi
ng
the
pr
ocess
or
ut
il
iz
ation
.
T
he
reason
f
or
re
duct
ion
in
com
pleti
on
ti
m
e
is
du
e
to
preem
pti
on
of
t
he
ta
sks
w
he
n
any
IO
is
pres
ent
an
d
dow
ng
rad
i
ng
the
pr
i
ori
ty
on
I
O
in
c
ase
of
MLF
Q.
The
res
ponse
tim
e
m
easur
em
ent
for
diff
e
re
nt
proce
sso
r
s
is
plo
tt
ed
belo
w.
F
ro
m
the
resu
lt
,
the
respon
se
ti
m
e
in
the
RLB
M
CS
is
far
bette
r
than
MLFQ.
T
his
re
du
ct
io
n
in
res
ponse
tim
e
is
du
e
to
ada
ptive
pri
or
it
y
ad
j
ust
m
ent
an
d
m
ov
em
ent
in
qu
e
ue
le
vels,
so
that
fair
re
sp
onse
ti
m
e
is
ens
ur
e
d.
The
res
pons
e
ti
m
e
is
m
easur
ed
for
diff
e
re
nt
a
rr
ival
of
ta
sks
an
d
the
res
ult
are
s
how
n
ab
ove
in
Figures
5
a
nd
6.
T
he
res
ponse
tim
e
is
low
in
case
of
RL
BM
CS
com
par
ed
t
o
MLFQ.
T
he
c
on
te
xt
switc
h
ov
e
r
head
in
te
rm
s
of
CPU
cy
cl
es
is
low
in
the
RLB
MC
S
com
par
ed
to
MLFQ
Sche
du
le
r
as
gi
ven
in
T
able
9.
T
he
CP
U
Util
iz
at
ion
acr
os
s
al
l
co
re
is
m
easur
ed
an
d
plo
tt
ed
i
n
Fig
ur
e
7.
The
CP
U
Util
iz
at
ion
is
inc
r
eased
in
RLB
MC
S
sche
du
li
ng
com
par
ed
to
MLF
Q
Sc
he
du
le
r.
T
he
a
ver
a
ge
ma
kesp
a
n
f
or
diff
e
re
nt
ta
sk
a
rr
ival
rate
is
s
how
n
in
Fi
gure
8.
T
he
ave
ra
ge
m
ake
sp
an
is
r
edu
ce
d
in
RLB
MC
S
Sche
du
li
ng c
om
par
ed
to ML
FQ
.
Figure
5. com
par
iso
n of Re
spon
s
e ti
m
e fo
r
RLB
MC
S w
it
h M
LFQ
Figure
6. com
par
iso
n of Re
spon
s
e ti
m
e fo
r
diff
e
re
nt
ar
riva
l t
i
m
e fo
r
RLB
MC
S w
it
h
ML
FQ
Table
9.
C
on
te
xt sw
it
c
h ov
e
r
head
RLBMC
ML
FQ
1
0
task
s av
erage
464
496
1
0
task
s stan
d
ard d
ev
iatio
n
32
33
3
0
task
s av
erage
578
612
3
0
task
s stan
d
ard d
ev
iatio
n
102
102
Figure
7. Com
par
is
on of CP
U
util
iz
at
ion
fo
r
RLB
MC
S w
it
h M
LFQ
Figure
8. Com
par
is
on for A
ve
rag
e
Make
spa
n for
RLB
MC
S w
it
h M
LFQ
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
.
2
,
A
pr
i
l 202
0
:
1805
-
1813
1812
Su
m
m
ary of
th
e res
ults is gi
ve
n belo
w
:
-
The res
pons
e ti
m
e is reduced
i
n pro
po
se
d
RL
BM
CS com
par
ed
to
MLF
Q
-
The
c
onte
xt s
w
it
ch
over
hea
d
i
n
te
rm
s o
f
CP
U
cy
cl
es is
lo
w
in pr
opos
e
d
R
LBM
CS
-
Av
e
ra
ge
m
ake sp
a
n
is l
ow i
n pro
po
se
d
RLB
MC
S
-
The
ta
s
k
c
om
pleti
on
ti
m
e and
turnar
ound ti
m
e is reduced
i
n pro
po
se
d
RL
BM
CS
5.
CONCL
US
I
O
N
A
Re
inf
orcem
ent
le
arn
i
ng
ba
sed
sc
heduling
of
ta
s
ks
t
o
m
ult
i
cor
e
env
i
ronm
ent
i
s
pro
po
se
d.
The
so
l
ution
optim
iz
es
m
ul
ti
obj
ect
ives in pla
ce
m
ent o
f
ta
sk
s.
Dy
nam
ic
qu
ant
um
f
or
each
ta
sk
with r
e
duct
ion
of
pr
eem
ption
,
Sp
eed
up
or
slow
dow
n
base
d
on
the
cu
rr
e
nt
load
a
nd
t
he
natur
e
of
ta
s
k
are
al
l
the
sal
ie
nt
featur
e
s
in
t
hi
s
m
e
thod
of
sche
du
li
ng.
T
he
prop
os
e
d
s
cheduler
w
as
i
m
ple
m
ented
on
Fr
ee
RTO
S
an
d
the
pe
rfor
m
ance
was
eval
uat
ed
agai
ns
t
nu
m
ber
of
co
res.
The
pr
opos
e
d
so
luti
on
pe
rfo
rm
s
bette
r
than
MLFQ
sche
du
li
ng
for m
ul
ti
co
res.
ACKN
OWLE
DGE
MENTS
We
ac
knowle
dge
the
dili
ge
nt
effo
rts
of
M
r
YVS
Ra
vi
ka
nth
i
n
guidi
ng
t
ow
a
r
ds
im
ple
m
entat
ion
of
this idea.
Dh
r
uv
a
is tha
nkf
ul to Dr
. Asha
ra
ni for m
entor
in
g
a
nd h
e
r
s
upport a
nd
gu
i
dance f
or
t
his
wor
k.
REFERE
NCE
S
[1]
Seife
din
e
Kadr
y,
Arm
en
Bagdas
ar
y
an
“
New
D
esign
and
Im
plem
ent
at
ion
of
M
LFQ
Schedul
in
g
Algorit
hm
for
Op
era
ti
ng
S
y
s
tem
s
U
sing
Dy
n
amic
Quantum”
Stat
isti
cs,
Opti
mizat
ion
And
I
nformation
Computing
.
Vol.
3
,
pp.
189
-
196
,
June
2015
.
[2]
Xiaoj
i
e
Li
and
Xianbo
He
“
The
improved
E
DF
Schedul
ing
Algorit
hm
f
or
Embedde
d
Re
al
-
ti
m
e
S
y
s
te
m
in
the
Unc
ertain
En
vironment”
ICA
CTE
20
10
.
[3]
Jink
y
u
L
ee
and
Kang
G.
Shin,
“
Pree
m
pt
a
Job
or
Not
in
E
DF
Schedul
ing
of
Uniproc
essor
S
y
stems
,
”
IE
E
E
Tr
ansacti
on
on
c
omputers
,
vol
.
6
3,
no
.
5
,
Ma
y
20
14.
[4]
J.
Le
e
,
A.
Ea
sw
ara
n,
and
I.
Shin,
“
Maximizi
ng
c
onte
nti
on
-
fre
e
e
xec
ut
ions
in
m
ult
iprocess
or
sche
duli
ng,
”
in
Proc
.
IEE
E
Re
a
l
-
Time
Technol. Appl.
Symp
.
(RTAS),
pp.
235
–
244
,
20
11
.
[5]
J.
Lee,
A
.
E
aswara
n,
I.
Shin,
a
nd
I.
Le
e
,
“
On
opti
m
al
m
ult
ipr
oce
ss
or
sche
dul
i
ng
conside
r
ing
conc
urre
n
c
y
and
urge
nc
y
,
” in
RTSS
Work
-
in
-
Pro
gress
Session
,
p
p.
21
–
24
,
2010
.
[6]
J.
Lee,
A.
Ea
s
wara
n,
and
I.
Shin,
“
LL
F
sc
hedul
ab
il
ity
an
a
l
y
sis
on
m
ultip
roc
essor
pla
t
for
m
s,”
in
RTSS
,
pp.
25
–
36
,
2010
.
[7]
S.
K.
Lee,
“
On
-
l
ine
m
ult
ipro
ce
ss
or
sche
duli
ng
algorithms
for
real
-
ti
m
e
t
asks,”
in
IEE
E
Re
gion
10
’s
Nint
h
Annual
Inte
rnational
Co
nfe
renc
e
,
pp.
60
7
–
611
,
1994
.
[8]
M.
L.
Dert
ouzos
and
A.
K.
Mok,
“
Multi
proc
essor
on
-
li
ne
sche
duling
of
har
d
-
rea
l
-
t
ime
ta
sks
,
”
IEEE
Tr
ansacti
ons
on
Soft
ware
Eng
ine
ering
,
vol
.
15
,
p
p
.
1497
–
1506
,
1989
.
[9]
A.
Ea
sw
ara
n
,
I.
Shin,
and
I.
Le
e
,
“
Optimal
virt
ua
l
cl
usterb
ase
d
m
ult
iprocess
or
sche
duli
ng
,
”
R
eal
-
T
ime
Syste
ms
,
vol
.
43,
no
.
1
,
pp
.
25
–
59,
2009
.
[10]
M
ark
o
Bert
ogna
and
Sanjo
y
Ba
rua
h
,
“
Li
m
ited
Pree
m
pti
on
ED
F
Schedul
ing
of
Sporadic
T
ask
S
y
stem
s,”
IE
EE
Tr
ansacti
ons on Indus
trial
Infor
matic
s,
Vol
.
6
,
n
o.
4
,
Novem
ber
2010.
[11]
M.
Bert
ogna,
O.
Xhani,
M.
Marinoni
,
F.
Esposit
o,
and
G.
Butt
a
zz
o,
“
Optimal
select
ion
of
pre
e
m
pti
on
point
s
to
m
ini
m
iz
e
pre
emption
ov
erh
ea
d
,
”
in
Proc.
23rd
E
urom
ic
ro
Conf.
Re
al
-
Ti
me
Syst
.
(
ECR
TS’11
)
,
Porto,
Portug
al,
Ju
l
.
6
–
8,
pp
.
217
–
22
7
,
2011
.
[12]
Abus
a
y
e
ed
Saif
ull
ah
,
Kunal
Agrawal
,
Chen
y
a
n
g
Lu
“
Multi
-
cor
e
Re
al
-
T
ime
Sc
hedul
ing
fo
r
Ge
ner
alize
d
Para
llel
Ta
sk Model
s”
I
EE
E
Re
a
l
-
Time
Syste
ms
,
2011
.
[13]
Kal
y
an
Baital
,
Am
la
n
Chakra
bar
ti
,
“
D
y
nam
ic
Schedul
ing
of
Rea
l
-
T
ime
Tas
ks
in
Hete
roge
neous
Multi
co
re
S
y
stems
”
IE
EE
Embe
dded
S
ystem
s Let
te
rs
,
2018
.
[14]
K.
Bai
t
al
and
A
.
Chakr
aba
rt
i,
"A
n
Eff
icient
D
y
namic
Schedu
ling
of
Ta
sks
for
Multi
-
cor
e
R
eal
Ti
m
e
S
y
stems
",
in
Adv
anc
es
in
Computing
Appl
ic
at
ions,
Chakr
abarti
A.
,
Sh
arm
a
N.
,
Bal
as
V.
(
eds
)
,
Sing
apor
e:
Springe
r
,
pp.
31
-
47
,
2016
.
[15]
A.
S.
Radh
aman
i
and
E.
Babur
aj,
“
Perform
anc
e
Eff
icient
Het
ero
gene
ous
Multi
co
re
Schedu
li
ng
St
rat
eg
y
base
d
on
Gene
tic
Algor
it
h
m
”,
A
RP
N
Journal
of
Engi
n
ee
rin
g
and
App
li
ed
Sc
ie
nc
es
,
vo
l. 8, no
.
1
,
pp
.
26
-
32
,
2
013.
[16]
J.
Rege
hr
,
"S
che
duli
ng
ta
sks
with
m
ixe
d
pre
emption
re
la
t
ions
for
robustness
to
ti
m
ing
fau
l
ts",
Proc.
23rd
IE
E
E
Re
al
-
Time
Syst
.
Symp.
,
pp.
315
-
3
26,
De
c
.
2002
.
[17]
Parva
r
Moham
m
ad,
M.E
.
Parv
ar,
and
Safa
r
i
Saee
d
,
“
A
Starva
ti
on
Free
IMLFQ
Schedul
ing
Algorit
hm
Based
on
Neura
l
N
et
work,
”
Int
.
J
.
Computa
t.
In
te
l
l. Res
.
,
vo
l.
4
,
no
.
1,
27
–
36,
2008
.
[18]
K.
Bai
t
al,
A.
C
hakr
aba
r
ti,
A.
C
hakr
aba
r
ti,
N.
Sharm
a,
V.
Ba
la
s
,
"A
n
eff
icient
d
y
nami
c
sche
du
li
ng
of
t
asks
for
m
ult
ic
ore
re
al
-
t
i
m
e
s
y
stems
" i
n
Advanc
es
in
Co
m
puti
ng
Applica
ti
ons,
Sing
apor
e:S
pringe
r,
pp
.
31
-
47,
2016
.
[19]
M.
Bert
ogna
an
d
S.
Barua
h,
“
Li
m
it
ed
pre
empti
on
EDF
sche
dul
ing
of
sporadic
ta
sk
s
y
stems
,
”
IEE
E
Tr
ans.
Ind.
Informat
.
,
vol. 6, no. 4, pp. 579
–
5
91,
Nov.
2010.
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
Rei
nfo
rce
men
t
le
ar
ni
ng base
d m
ulti
co
re sc
he
du
li
ng
(
RL
B
MCS)
for
real t
ime syste
m
s…
(
Dh
r
uva R
.
Rin
ku
)
1813
[20]
K.
La
kshm
ana
n,
S.
Kato,
and
R.
R.
Raj
kum
ar
,
"S
che
duli
ng
p
ara
l
le
l
re
al
-
ti
m
e
ta
sks
on
m
ult
i
-
cor
e
proc
essors
,
"
in
2010
31st
IE
E
E
R
eal
-
Time
Sys
te
ms
Symposium,
San
Diego, CA
,
pp
.
259
-
268
,
2
010
.
[21]
C.
Ta
n,
T
.
S.
Muthukaruppa
n,
T.
Mitra
,
L
.
Ju,
"A
pproximati
on
-
awa
re
sche
dul
ing
on
het
ero
ge
neou
s
m
ult
i
-
cor
e
arc
hi
te
c
ture
s"
,
P
roc.
AS
P
-
DAC/I
EE
E
,
pp
.
618
-
62
3,
2015
.
[22]
Pricopi
M,
Mitr
a
T.,
“
T
ask
sc
hed
uli
ng
on
ad
apt
iv
e
m
ult
i
-
cor
e
,
”
IE
EE
Tr
ans Comput
63(10):2590
–
2603
,
2014
.
[23]
A.
Hata
m
i,
S.
C
hupra
t,
H.
Md
Sarka
n,
N.
Fird
au
s
Mohd
Az
m
i
"
H
y
brid
Re
al
-
T
i
m
e
Ta
sk
Sc
hedul
ing
Algorit
hm
i
n
Overl
oad
Si
tua
t
i
on
for
Mult
iprocess
or
S
y
stem"
,
J
TEC
,
Vol
9
,
No
3
-
4
,
2017
.
[24]
T
.
Vai
ram,
S
.
Sara
tha
m
bekai
"M
ult
iproc
essor
ta
sk
sche
du
li
n
g
proble
m
using
h
y
br
id
discr
e
te
par
ti
c
le
sw
ar
m
opti
m
iz
ation",Så
dhanå
,
43:
206
,
Dec
ember
2018
.
[25]
A.
Shah
and
K.
Kotec
h
a,
“
ACO
base
d
d
y
n
amic
sche
du
li
ng
a
l
gorit
hm
for
rea
l
-
ti
m
e
m
ult
iprocess
or
sy
st
ems
,
”
Inte
rnational
.
Jo
urnal.
of
Gr
id
an
d
High
P
erformance
Comput
ing
,
vol
.
3
,
no
.
3
,
pp
.
20
–
30
,
2011
.
[26]
Aghaz
arian
V,
Ghor
banni
a
A,
Motla
gh
NG
,
Nae
ini
MK
,
“
RQSG
-
I
:
an
opti
m
iz
e
d
rea
l
t
ime
sche
duli
ng
al
gor
it
hm
for
ta
sks
al
locati
on
in
grid
envi
ron
m
ent
s,”
2011
IEE
E
3rd
Inte
rna
tional
Confer
en
ce
on
Comm
unic
ation
Software
an
d
Networks,
Xi
'
an
,
pp.
205
-
210
,
20
11
.
[27]
Venka
ta
Siva
Pr
asa
d
Ch.
and
S.
Ravi
“
Schedul
in
g
Of
Share
d
Me
m
ory
W
it
h
Mult
i
-
Core
Perform
anc
e
In
D
y
n
amic
Alloc
a
tor
Us
ing A
rm
Proce
ss
or”
AR
PN
Journal
o
f
E
ng
ine
ering
an
d
Applied
S
cienc
es,
Vol
.
11
,
No.
9,
Ma
y
2016.
[28]
Im
ran
Qureshi,
“
CP
U
Schedul
in
g
Algorit
hm
s:
A
Surve
y
”
Int
.
J
.
Adv
anc
ed
Ne
two
rking
and
Appli
cat
ions
,
Vol
.
05
,
Iss
ue
04,
Pa
g
es: 19
68
-
1973,
201
4.
BIOGR
AP
HI
ES OF
A
UTH
ORS
Dh
ruva
R.
Rin
ku
was
born
in
Ahm
eda
bad,
In
dia
in
1973
.
She
recei
v
ed
her
M.T
ec
h
.
degr
ee
in
digi
tal
s
y
stems
&
Com
pute
r
Elec
tron
ic
s
from
JN
TU,
Hy
d
era
b
ad,
India
in
2009.
Presently
she
is
purs
uing
her
r
ese
arc
h
work
on
s
che
dul
ers
for
R
TOS
on
m
ult
ic
o
re
proc
essors
.
She
has
b
ee
n
wor
king
as
As
socia
te
Profess
or
in
CVR
Coll
ege
of
Engi
ne
eri
ng.
H
y
de
aba
d
and
is
responsible
fo
r
m
ic
roproc
essors
and
embedde
d
s
y
st
ems
rel
a
ted
subjects
in
t
he
dep
art
m
ent
of
Ele
ct
ron
ic
s
and
Com
m
unic
at
ion Engi
ne
eri
ng.
Dr.
M.
As
haR
ani
was
born
in
Srikakul
am,
I
ndia
.
She
re
cei
ved
her
M.T
ech.
degr
e
e
in
di
git
al
s
y
stems
&
Com
pute
r
E
lectr
oni
c
s
from
JN
TU,
H
y
der
aba
d
,
Indi
a
in
1997
.
She
r
ec
e
ive
d
h
er
Ph
D
in
El
e
ct
roni
cs
&
Com
m
unic
at
ion
Engi
nee
r
ing
from
JN
TU
Hy
der
abad i
n
2008.
Presently
she
is worki
ng
as
Profess
or
in
Coll
ege
of
Enginee
ring
,
JN
TU,
H
y
der
aba
d
in
th
e
depa
r
tment
of
El
e
ct
roni
cs
and
Com
m
unic
at
ion
Engi
neering
with
over
26
y
e
a
rs
of
te
ac
hing
expe
ri
ence.
Her
rese
arc
h
int
er
e
sts
inc
lu
d
e
Faul
t
T
ole
ran
t
S
y
s
te
m
s,
Design
for
Te
s
ta
bilit
y
,
Re
al
Tim
e
Embedde
d
S
y
stem
Design
a
nd
Multi
-
cor
e
Emb
edde
d
S
y
stem D
esign.
Evaluation Warning : The document was created with Spire.PDF for Python.