TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 11, Novembe
r
2014, pp. 78
6
9
~ 787
5
DOI: 10.115
9
1
/telkomni
ka.
v
12i11.66
40
7869
Re
cei
v
ed
Jul
y
10, 201
4; Revi
sed Septe
m
ber
4, 2014
; Accepte
d
Septem
ber 25,
2014
An Adaptive Genetic Algorithm for Mesh-Based NoC
Application Mapping
Yizhuo Wan
g
, Zhibiao Zhang*
Schoo
l of Com
pute Scie
nce a
nd T
e
chnol
og
y, Beijin
g Institut
e of
T
e
chnol
og
y,
Beiji
ng, 10
08
1, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: carldotz@
16
3.com
A
b
st
r
a
ct
Appl
icatio
n ma
ppi
ng is
o
ne o
f
the
key prob
l
e
ms
of N
e
tw
ork-on-Ch
ip (
N
o
C
) des
ign. It
ma
ps th
e
cores of app
li
cation to the
process
i
ng e
l
e
m
e
n
ts
of the NoC top
o
lo
gy. T
h
is paper p
r
esents a nov
e
l
appr
oach for N
o
C ap
plic
atio
n ma
pp
ing, w
h
ic
h uses ad
apt
iv
e gen
etic al
gor
ithm (AGA) in the map
p
in
g. T
h
e
prop
osed
a
ppr
oach
a
daptiv
el
y vari
es the
p
r
oba
bil
i
ties
of
crossover
an
d
mutatio
n
ope
rators i
n
g
e
n
e
t
ic
alg
o
rith
m, ai
mi
ng to reduc
e the over
all co
mmu
n
ic
ation
cos
t
of NoC. Experi
m
ent
al resu
l
t
s show
that
th
e
prop
osed ap
pr
oach decre
ase
s
the
co
mmu
n
i
c
ation c
o
st by
3% to 7
%
o
n
a
v
erag
e, co
mp
a
r
ed to the
exist
i
ng
appr
oach
usin
g Standar
d Ge
netic Alg
o
rith
m (SGA).
Ke
y
w
ords
: net
w
o
rk-on-chi
p
, app
licati
on
ma
ppi
ng, ad
apt
ive
genetic a
l
g
o
rit
h
m, g
enetic a
l
g
o
rith
m
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Network-on-Chip (NoC)
is
the de
sign of mod
u
lar a
nd
scalable
com
m
unication
architectu
re
s whe
r
e va
rio
u
s p
r
o
c
e
ssi
n
g
eleme
n
ts
are
con
n
e
c
te
d to a route
r-b
ased n
e
twork
usin
g app
rop
r
iate network
interface [1-3
]. Since appli
c
ation Ma
ppi
ng ha
s dire
ct
effect on del
ay,
power con
s
u
m
ption and ot
her pe
rforma
nce of NoC, i
t
is a very sig
n
ificant a
s
pe
ct of NoC d
e
s
ign
[4]. In its gen
eral form, the
proble
m
of NoC
ap
plication mappin
g
is
an NP-hard p
r
oble
m
[4].
Geneti
c
Alg
o
r
ithm
(GA),
a sto
c
h
a
sti
c
se
ar
ch
alg
o
rithm
ba
se
d on
n
a
tural
gen
etic
operation
s
provides
a sol
u
tion to
the problem of
No
C map
p
ing.
Sometimes t
he pe
rform
a
n
c
e of
Standard G
e
netic Alg
o
rith
m (SGA
) b
a
sed
No
C ma
p
p
i
ng
ca
nnot
meet the
ne
e
d
of p
r
a
c
tical
work
becau
se the
paramete
r
s
of gen
etic o
peratio
n a
r
e
fixed values.
The
r
efore,
several a
dap
tive
geneti
c
algo
rithms ba
se
d
on SGA are pro
p
o
s
ed.
The Adapti
v
e Genetic
Algorithm
s (AGA)
prop
osed by
M. Srinivas
a
nd L. M. Patn
ak [5] is
co
nsi
dere
d
to be t
he mo
st re
prese
n
tative, and it
reali
z
e
s
th
e
t
w
in goal
s of maintainin
g
t
he
dive
rsity
o
f
popul
ation and su
staini
n
g
the
capa
cit
y
o
f
conve
r
ge
nce with the use of adaptive pr
obabilitie
s of cro
s
sove
r an
d mutation.
In this pap
er, we use AGA in the NoC
ap
plicatio
n mappi
ng. First, we extend the
rep
r
e
s
entatio
n for chromo
some
s propo
sed by W. Zh
ou, et al [6]
to a general situation in wh
ich
the num
ber
of appli
c
ation
co
re
s is l
e
ss than
or
eq
ual
to
the nu
mber of
processing eleme
n
t
s
(PEs)
of the
No
C topolo
g
y
. Second, o
u
r ma
ppi
ng a
ppro
a
ch
redu
ce
s
the com
m
unication cost
with an
imp
r
o
v
ed ada
ptive
geneti
c
alg
o
ri
thm, F-AGA,
whi
c
h i
s
p
r
op
ose
d
by X.
Cheng
[7], and
it
amends the
expre
ssi
ons f
o
r
crosso
ver probability
pc
and mutation probability
pm
to s
o
lve the
probl
em that
pc
and
pm
are
ze
ro fo
r
the solution
with the
max
fitness. Experime
n
tal results
sho
w
that o
u
r approa
ch d
e
c
re
ases th
e
communi
ca
tio
n
co
st by 3%
to
7% on ave
r
age,
com
pared
to the appro
a
c
h u
s
ing Stan
dard G
eneti
c
Algorithm (S
GA).
The re
st of this pap
er i
s
organi
zed a
s
fo
llows
. Sectio
n 2 introd
uce
s
the rel
a
ted
work in
No
C m
appin
g
an
d GA
b
r
iefly. Sectio
n 3
pres
ents the d
e
finitio
n
s
used i
n
t
h
is
pap
er an
d
descri
b
e
s
o
u
r AGA ba
se
d
approa
ch. Se
ction
4
sho
w
s the
expe
ri
mental
re
sult
s, an
d Se
ctio
n 5
summ
ari
z
e
s
our mai
n
co
ntribution.
2. Related Work
The
No
C a
p
p
lication
map
p
i
ng te
chni
que
s
can
be
cl
assified
as dyn
a
mic map
p
in
g [8-10]
and stati
c
ma
pping, de
pen
ding on the ti
me at whic
h the co
re
s of application are
assi
gne
d to the
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 78
69 – 787
5
7870
PEs of the
NoC to
pology.
Static map
p
in
g is
gen
erally
re
comm
end
ed, a
s
excess
comm
unica
tion
co
st in dyn
a
m
ic ma
ppin
g
signifi
cantly
affects
ov
era
ll perfo
rma
n
ce of sy
stem
[11]. The
sta
t
ic
mappin
g
te
ch
nique
s m
a
inl
y
includ
e Inte
ger
Line
ar P
r
ogra
mming
(I
LP) [12], Bra
n
ch
-an
d
-Bo
u
n
d
(BB) [13], Partic
le Swarm Op
timization (PSO
) [1
4], Ant Colo
ny optimizati
on (A
CO
) [1
5],
Simulated An
nealin
g (SA) [16], Genetic
Algorithm (G
A), etc. The rest of this se
ction will focu
s o
n
the related
work of GA for
No
C mappi
ng
.
Lei et
al. pro
pose a
two
-
st
ep G
eneti
c
A
l
gor
ithm
(GA
)
for
No
C m
a
pping,
whi
c
h
redu
ce
s
the overall ex
ecutio
n time [
17]. In the fi
rst step,
the
tasks a
r
e
assign
ed o
n
to different Intelle
ctu
a
l
Prope
rties (I
Ps), assumi
n
g
that
the edge delays a
r
e
const
ant and
equal to the averag
e valu
e o
f
the ed
ge
del
ays. In the
seco
nd
step, t
he IPs are
m
appe
d to th
e
PEs of
No
C,
taking
the
act
ual
edge d
e
lay base
d
on the
netwo
rk traffic model, an
d
the total system delay is minimized [1
1].
Zhou et al. propo
se a ge
n
e
tic algo
rithm
base
d
on a
delay model f
o
r NoC ma
pp
ing [6]. Zhou
et
al. also propose a representation
for the chromosom
e
s to ensure t
hat the offspri
ng will meet the
con
s
trai
nt, a one-to
-o
ne constraint bet
wee
n
IP
core
s an
d PEs. A pareto
ba
se
d multi-obj
ect
i
ve
evolutiona
ry comp
uting te
chni
que is p
r
opo
sed by
Ascia et al. [18], which o
p
timize
s perfo
rma
n
ce
and po
we
r consumption
of NoC.
Je
n
a
et al. propose MGAP, a genetic algorith
m
ba
sed
optimizatio
n t
e
ch
niqu
e, wh
ich mi
nimizes the po
we
r co
nsum
ption an
d
maximizes the
thro
ugh
p
u
t
by redu
cing t
he numb
e
r of
switche
s
in the com
m
uni
cation path bet
wee
n
co
re
s [19].
The majo
r drawb
a
ck of the genetic al
g
o
rithm
s
is the
slow rate of conve
r
ge
nce. It often
requi
re
s the
GA to evolve a large
num
ber of ge
ne
rations to
con
v
erge to a
solution. The
best
solutio
n
at the end is take
n to be the solution of the mappin
g
pro
b
lem. To accelerate the ra
te o
f
conve
r
ge
nce, the mutation
rate can b
e
increa
sed.
H
o
wev
e
r,
it
most
ly
co
nv
erg
e
s t
o
lo
cal
b
e
st
solutions,
rat
her than finding the global best
[11].
AGA has better capability of locating t
h
e
global
optimu
m
than SGA.
Ho
weve
r, we have
not
f
ound
any lite
r
ature that h
a
s
use
d
AG
A in
No
C mappi
ng
.
3. Technique
In this
se
ctio
n, we
start
wi
th the definiti
ons
used in
this p
ape
r. Th
en we de
scri
be the
adaptive ge
n
e
tic algo
rith
m for mesh-based No
C
mappin
g
. Th
e rep
r
e
s
entat
ion of popul
ation
(chrom
osomes) and the
expressi
ons
of probability are very im
portant aspects of AGA. We
introduc
e
them s
eparately.
Defini
tion 1
:
Application
Cha
r
a
c
teri
stic Graph A
C
G
(V, E) i
s
a
directed
graph,
whe
r
e
each vertex c
i
∈
V represe
n
ts an IP core and ea
ch e
dge e
i,j
∈
E repre
s
e
n
ts the
commu
nicati
on
band
width b
e
t
ween a
sou
r
ce core (c
i
)
and a de
stin
ation co
re (c
j
). Each e
i,j
is tagged with
v
ij
whi
c
h re
presents the com
m
unication volume from
c
i
to c
j
.
Defini
tion 2:
No
C Archite
c
ture
Graph
NAG (R, C)
i
s
a di
re
cted
grap
h, wh
ere
each
ver
t
ex r
i
∈
R,
rep
r
e
s
ent
s a
PE. Each di
rected
edg
e
c
i,
j
∈
C
rep
r
e
s
e
n
ts a
physi
ca
l unidi
re
ction
a
l
cha
nnel
which con
n
e
c
ts a
n
output port
of r
i
to an input port of r
j
.
Defini
tion 3
:
Virtual IP
C
o
re
(VIC
) i
s
an IP
co
re
t
hat do
es not
exist in
ACG. The
comm
uni
cati
on volu
me from a
VIC to
an IP
co
re i
s
alway
s
ze
ro,
as i
s
the
com
m
unication f
r
om
an IP c
o
re to
a VIC.
Defini
tion 4:
Extended Application Ch
ara
c
teri
stic G
r
aph EA
CG (C, E) is a di
rected
grap
h, wh
ere
each vertex
c
i
∈
V re
prese
n
ts a
n
IP core o
r
a
VIC and e
a
ch edg
e e
i,j
∈
E
rep
r
e
s
ent
s the commu
nica
tion band
widt
h betwee
n
a sou
r
ce co
re (c
i
) an
d a dest
i
nation co
re (c
j
).
Eac
h
e
i,j
is ta
gged
with v
ij
whi
c
h re
presents the com
m
unication volume from
c
i
to c
j
.
3.1. Population Repr
ese
n
t
ation
No
C map
p
in
g has
a on
e-to-one
con
s
traint b
e
twe
en IP cores and PEs i
n
No
C
appli
c
ation m
appin
g
, which mean
s that different
IP
cores
can
not
be mapped
to the same PE
and diffe
rent
PEs ca
nnot
map the
sam
e
IP co
re. Th
us, some
offspring
will viol
ate the
con
s
traint
if they are generate
d
by parent
s ran
d
o
m
ly. In or
der to overcome
this proble
m
, Zhou et al.
[6]
has p
r
o
p
o
s
ed
a conve
n
ient
populatio
n re
pre
s
entat
io
n
scheme.
Usi
n
g this sch
e
m
e
, the offsprin
g
will sati
sfy the con
s
traint a
s
long
as th
e pare
n
ts
d
o
. Howeve
r, the situation in whi
c
h the n
u
mbe
r
of IP cores i
s
less than the
numbe
r of PEs ha
s not bee
n con
s
id
ere
d
in this schem
e.
An extended
populatio
n repre
s
e
n
tation
sch
eme is
sugge
sted h
e
re. This sch
e
m
e is
based on the
sch
eme in [6], and is abl
e to repre
s
e
n
t the situation in whi
c
h the numb
e
r o
f
IP
cores i
s
le
ss than the nu
mber of PEs. The
mapp
e
d
2D-me
s
h
No
C is rep
r
e
s
ente
d
by o
ne-
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Adaptive
Geneti
c
Algorithm
for Mesh
-Bas
ed NoC
Applicatio
n M
appin
g
(Yizhu
o Wan
g
)
7871
dimen
s
ion
a
l array usin
g a
left to right scan p
e
rfo
r
me
d in a top-d
o
w
n fashion. T
he ch
rom
o
so
me
is an intege
r array, and the value at the i
th
position denote
s
the mappin
g
of the
i
th
I
P
core or
V
I
C.
For the value
of the i
th
position an integ
e
r betwee
n
1 a
nd i is allo
we
d.
Next, we illustrate the decoding
procedure of
the
chromosome
structure with the help
of an exa
m
pl
e. Figu
re 1
(
a
)
sho
w
s the
ACG of
an
a
pplication with
7
IP cores,
and Figu
re 1(b
)
sho
w
s the NAG of a 3x3 2D-m
esh No
C.
(a) AC
G
(b)
NAG
Figure 1. ACG with 7 IP Core
s and 3x3
2D-me
s
h NA
G
The firs
t
s
t
ep is
to cons
truc
t EACG from
ACG. In
this exam
ple
the n
u
mbe
r
of PEs
minus the n
u
m
ber of IP core
s is 2. Th
us, two
VIC, ‘h’ and ‘I’, should be ad
de
d to the EACG.
Figure 2 sh
o
w
s the EACG
.
Figure 2. EACG
Then,
fo
r a chrom
o
some
structu
r
e (1, 2,
2,
4, 5,
4,
7,
1, 3), th
e first
integer i
s
a
p
p
lied to
map IP
co
re ‘
a
’; the
se
con
d
integ
e
r i
s
a
pplied
to
ma
p
IP co
re ‘
b
’; a
nd
so
on.
Cle
a
rly, the la
st t
w
o
integers a
r
e
applie
d to m
ap the VIC ‘
h
’ and ‘i’. Th
e
value of th
e first po
sitio
n
is ‘1’, an
d
all
solutio
n
s
will
have a val
ue
‘1’ into the fi
rst po
siti
on. T
he second
int
eger is
appli
e
d to map
‘b’,
and
there
are two
situatio
ns: v
a
lue ‘
1
’ me
an
s ‘b’
is bef
ore ‘a’ and value ‘2’ means ‘b’
is
after ‘a’.
T
h
e
se
con
d
int
e
g
e
r in t
h
e
chr
o
mos
o
me
st
r
u
ct
ur
e i
s
‘2
’.
So [a, b] is obtaine
d. Similarly, the t
h
ird
integer i
s
ap
plied to map
‘c’, and there are thr
ee situations: val
ue ‘1’ mean
s ‘c’ is before
‘a’,
value ‘2’ mea
n
s ‘c’ i
s
bet
ween ‘a’ an
d ‘b
’ and value ‘3
’ mean
s ‘c’ is after ‘b’. The
third intege
r
in
the sample
chrom
o
some
structu
r
e i
s
‘2’.
So [a, c,
b] is o
b
tained.
Continuin
g
in t
h
is fa
shio
n, the
followin
g
pe
rmutation of th
e nin
e
alp
hab
ets [h, a, i,
c,
b, f, d, e, g] i
s
o
b
taine
d
. Core
s a
r
e
pla
c
ed
in a
3x3
2D-mesh
No
C a
c
cordi
n
g
to th
is p
e
rm
utatio
n
as
s
h
o
w
n in
F
i
gu
r
e
3 (
a
)
.
F
i
n
a
lly, VICs
‘h
’
and ‘i’ shoul
d
be re
moved
from the
core
placement. T
hen the final
core pla
c
em
e
n
t is obtain
e
d
in
Figure 3 (b
).
(a)
Core pla
c
ement
(b) T
he final core pla
c
e
m
en
t
Figure 3. Core Placem
ent of a Sample Chromo
som
e
Structure
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 78
69 – 787
5
7872
Note that thi
s
extend
ed p
opulatio
n re
p
r
es
entation
schem
e can e
a
sily de
al wit
h
the
above con
s
traint con
d
ition
even if the numbe
r
of IP core
s is le
ss th
an the numb
e
r
of PEs.
3.2. Expressi
ons of Probabilit
y
In the AGA
proposed by M.
Srinivas [5] t
he probabiliti
e
s
of crossover and
mutati
on are
adju
s
ted by the expre
s
sio
n
(1) a
nd (2
).
(1)
(2)
Whe
r
e
P
c
i
s
the proba
bil
i
ty of cro
s
so
ver;
P
m
is the probability of mutation;
is the
maximal fitne
ss; i
s
the av
erag
e fitness;
is
the bi
gge
r fitness val
u
e
betwe
en two
pare
n
ts; i
s
the
fitness valu
e of the mutated individual;
k
1
,
k
2
,
k
3
and
k
4
are co
nsta
n
t
s betwe
en 0
and 1.
The AGA p
r
o
posed in [5] i
s
mo
re p
r
o
p
itious to th
e p
r
obability adj
u
s
tment of the
later
perio
d evol
ution, be
ca
use
the p
r
o
babil
i
ty of
crosso
ver an
d m
u
tation a
r
e
ze
ro for the
be
st
solutio
n
. Here we introdu
ce
an im
pro
v
ed ad
aptive
gen
etic
alg
o
rithm F
-
AG
A [7], and t
h
e
probability expr
essi
ons can be
described as
(3)
(4)
Whe
r
e
P
c
i
s
the probability of crossover;
P
c_
m
a
x
is the maximal cros
sover probability;
P
c_
m
i
n
is the minimal crossover probability;
P
m
is
the probabilit
y of mutatio
n
;
P
m_
ma
x
is the
maximal mutation pro
babil
i
ty;
P
m_
mi
n
is the minimal m
u
tation probabilit
y; is the
maximal fitness;
is the averag
e fitness; i
s
the bigg
er fitn
ess va
lue b
e
twee
n two p
a
rents; is the f
i
tness value
of
the mutated i
ndividual. Th
e paramete
r
s are
set a
s
P
c_max
= 0.9,
P
c_
m
i
n
= 0.6,
P
m_
m
a
x
= 0.2,
P
m_
m
i
n
= 0.01.
3.3. The Oth
e
r Asp
ects o
f
the
Algorithm
a) Initial
populati
o
n
The initial
p
opulatio
n is
gene
rated
ra
ndomly.
Note that the in
teger val
ue
at the i
th
positio
n of ea
ch individ
ual in the
initial populatio
n mu
st betwe
en 1
and i.
b) Fitness
fun
c
tion
The goal of th
is pap
er is to
redu
ce the
co
mmuni
cation
co
st (CommCost) me
asure
d
by
(5)
Where
is the Manhattan
di
stance bet
ween
sour
ce
core
‘i’ and destination core
‘j’,
and th
e di
st
ance b
e
twe
e
n
two
a
d
jacent PEs
is
one
hop.
is the
requi
rement
of
comm
uni
cati
on bandwidt
h
from source
core ‘i’ to destination core ‘
j’.
c
)
R
e
pr
o
d
u
c
tion
Tourname
n
t sele
ction i
s
used in thi
s
pa
per.
Tou
r
na
m
ent sele
ction
rand
omly sa
mples
k
individual
s from pare
n
t po
pulation, and
then sele
cts the best one
into the mating pool. In this
pape
r k is
set
to 2.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Adaptive
Geneti
c
Algorithm
for Mesh
-Bas
ed NoC
Applicatio
n M
appin
g
(Yizhu
o Wan
g
)
7873
d) Cro
s
sove
r
Our
expe
rim
ents
sh
ow th
at the unifo
rm cro
s
sover i
s
b
e
tter tha
n
singl
e-p
o
int
crossove
r
and m
u
lti-poi
nt crossove
r.
So we
u
s
e
un
iform
cro
s
sov
e
r in
this pap
er. After
cal
c
ulating
crossover
prob
ability, a uniform
cro
ssover is do
ne
with the ne
w cro
s
sove
r pro
bability.
e) Mutation
After cal
c
ulati
ng mutation
probability a mutation
is done by selecting a position ‘i
’ at the
rep
r
e
s
entatio
n of
ch
romo
some
with t
he n
e
w cro
s
sover p
r
ob
a
b
ility, and th
e integ
e
r value
assign
ed at i
th
position is a
rando
m integ
e
r between 1
and i.
f) Elite-pre
s
e
r
v
a
tion
strate
gy
In orde
r to speed u
p
the conve
r
ge
nce
rate
of AGA and en
sure that the offspring i
s
alway
s
better than the pa
rent, elite-preservation
st
rat
egy is u
s
ed in
this pap
er. El
ite-preservati
on
strategy e
n
su
res th
e be
st individual al
ways su
rvives
to be an indi
vidual of the next gene
rati
on.
Then, the b
e
st individu
al
will repla
c
e
the
worst individual of
next generation in the AGA
prop
osed in this pa
per.
g) Terminal
criterion
The termin
ation crite
r
io
n is set to be 500
generation
s
.
4. Experimental Re
sults
To evalu
a
te
the pe
rform
a
nce
of the
propo
sed
AGA
ba
sed
NoC
appli
c
ation
m
appin
g
approa
ch,
we
impleme
n
ted
the AGA a
n
d
the SG
A i
n
C++. T
he
crossover prob
ability and th
e
mutation probability of SGA are fixed as
P
c
= 0.9 and
P
m
= 0.05.
We g
ene
rate
6 ACG
s
wit
h
TGFF [20]
rand
omly, an
d gene
rate
6
NAGs fo
r th
e ACG
s
.
The num
ber
of core
s an
d the numb
e
r of
PEs of
the 6
test ca
se
s are sho
w
n in T
able 1.
Table 1. The
6 Mappin
g
Case
s
Cases
Number of
IP cor
e
s
Number of PEs
1 60
8*8
2 46
7*7
3 20
5*4
4 19
8*8
5 44
7*7
6 19
4*5
We ma
de 1
0
0
simul
a
tion runs fo
r ea
ch
ca
se in T
able
1. Con
s
ide
r
i
ng the rando
mization
effect of S
G
A and
AGA,
the alg
o
rithm
s
we
re
ex
ecu
t
ed for 50
0
g
eneration
s
in
ea
ch
sim
u
lat
i
on
run.
Figure 4
sh
o
w
s the
comm
unication
co
sts at ea
ch
ge
neratio
n of th
e gen
etic
alg
o
rithm
s
for the test ca
se 1. Each p
o
int in the fig
u
re
re
present
s the averag
e
val
ue of 100 simulatio
n
ru
ns.
From th
e figu
re, we can
se
e that the AG
A outper
fo
rm
s the S
G
A, a
nd the SGA t
end
s to g
e
t st
uck
at a local opti
m
um at point A while the AGA doe
s until point B.
Figure 4. Vari
ations of the
Co
mm
uni
cat
i
on Co
st
s
(t
es
t
case 1
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 78
69 – 787
5
7874
Table 2 sho
w
s the final re
sult of the mappin
g
ca
se
s in Table 1. As we can see from
Table 2, after 500 gen
erati
ons, the AGA
base
d
app
ro
ach d
e
crea
se
s the co
mmu
nicatio
n
co
st by
3% to 7% on averag
e, com
pare
d
to the SGA base
d
a
ppro
a
ch.
Table 2. Co
m
m
unication Costs of the Fi
nal Mappi
ng
s
Cases
SG
A
AG
A
AG
A/
S
G
A
1
3645.63
3416.16
0.93705615
2
4025.68
3824.88
0.95012023
3
994.962
978.439
0.98339334
4
1812.01
1769.11
0.97632463
5
2828.26
2726.67
0.96408039
6
1671.73
1649.82
0.98689382
5. Conclusio
n
In this pape
r
an ada
ptive genetic al
gorit
hm
(AGA) b
a
sed a
pproa
ch
for No
C appl
ication
mappin
g
is p
r
opo
se
d. We
present a populatio
n
rep
r
esentation
schem
e whi
c
h
can ea
sily deal
with the
co
nstraint conditio
n
in
No
C a
p
p
lication
map
p
i
ng even
if th
e num
ber of I
P
co
re
s is le
ss
than the
num
ber
of PEs. T
he p
r
op
osed
approa
ch
ada
pt
ively varies
the proba
bilities
of cro
s
sover
and mutation
ope
rato
rs. Our
expe
rim
ents sh
ow
that the A
G
A
enh
an
ce
s t
he
cap
ability of
prem
ature co
nverge
nce
prevent
ion an
d prod
uces lo
wer co
mmuni
cation co
st tha
n
SGA.
The p
r
op
ose
d
app
roa
c
h
only con
s
id
ers the
comm
unication cost. In future work, th
e
approa
ch will
be extende
d by con
s
ide
r
in
g the other
a
s
pect
s
of No
C su
ch a
s
the n
e
twork del
ay.
Ackn
o
w
l
e
dg
ement
The autho
rs thank the anonymo
us reviewe
r
s for their valuab
le comme
nts on the
manu
script.
This
wo
rk wa
s p
a
rtially
su
pporte
d by th
e National
Natural S
c
ien
c
e Fou
ndatio
n
of
Chin
a unde
r
grant NS
FC- 6130
0011.
Referen
ces
[1]
L Ben
i
n
i
, G De
Miche
li. Net
w
orks on
Ch
ips:
a n
e
w
So
C p
a
rad
i
gm.
IEEE
Co
mp
uter.
20
02; 3
5
(1): 70
–
78.
[2]
WJ Dally
,
B To
w
l
es.
R
oute
p
a
ckets, not w
i
res: on-ch
ip i
n
tercon
nectio
n
n
e
tw
orks
. Proceedi
ngs of t
h
e
38th Des
i
g
n
Automatio
n
Co
nferenc
e.
Las Ve
gas, USA. 200
1: 684–
68
9.
[3]
S Kumar, A Ja
ntsch, JP Soi
n
i
nen, M Forse
ll,
M Millb
erg, J
Oberg, K T
i
en
s
y
rj
a, A Hem
a
ni.
A netw
o
rk
on c
h
ip
arc
h
ite
c
ture a
n
d
des
i
gn
metho
dol
og
y
. Procee
di
ngs
of ISVLSI. Pi
t
t
sburgh, USA. 200
2:
1
17–
124.
[4]
R Pop, S Kumar. A surve
y
of techniq
ues for
m
appin
g
an
d sched
uli
ng a
p
p
licatio
ns to net
w
o
rk on chi
p
s
y
stems. Scho
ol of Engi
ne
eri
ng, Jönkö
p
i
ng
Univers
i
t
y
. R
e
p
o
rt Number: 04
:4. 2004.
[5]
M Srinivas, LM
Patnak. Adapt
ive Prob
abi
litie
s of
Crossover
and Mutatio
n
i
n
Genetic Alg
o
rithm.
IEEE
T
r
ans. on Systems, Man
and
Cyber
netics.
1
994; 24(
4): 656
- 666.
[6]
W
Z
hou, Y
Z
hang, Z
Ma
o.
An
ap
plic
a
t
ion s
pecific
NoC
map
p
in
g
for o
p
ti
mi
z
e
d
de
lay.
IEEE
Internatio
na
l C
onfere
n
ce
on
Desig
n
an
d T
e
st of In
tegr
ated
S
y
stems i
n
N
anosc
a
le.
La
Marsa, T
unisia
.
200
6: 184
–1
88
.
[7]
Che
ng
Xi
an-
yi
, Yang C
h
a
n
g
-
y
u
.
R
o
b
o
t
pa
th
pla
nni
ng b
a
sed on ada
p
t
ive
isol
atio
n nich
e
ge
neti
c
alg
o
rith
m
. Pro
c
eed
ings
of
2nd
Intern
atio
nal
S
y
mpos
iu
m on
Intell
ig
ent Informati
o
n
T
e
chnol
o
g
y
Appl
icatio
n. Sh
ang
hai, C
h
in
a. 200
8: 151-
154.
[8]
G Chen, F
Li, M Kandemir.
Co
mp
iler
-
directe
d
a
p
p
l
icatio
n
ma
ppi
ng for
No
C
bas
ed c
h
i
p
mu
ltiproc
e
ssor
s
. Proceedi
ngs
of LCT
ES. San Die
go, Cal
i
fo
rnia, USA. 200
7: 155–
15
7.
[9]
CL Cho
u
, UY Ogras, R Marculesc
u
. Energ
y
- and
perform
a
n
ce-a
w
a
re i
n
cr
ementa
l
mapp
i
ng for NoCs
w
i
t
h
multip
le v
o
ltag
e leve
ls.
IEEE Transactions on Computer-Aided
design of Integrat
ed Circuits and
System
s
. 20
08
; 27(10): 18
66–
187
9.
[10]
E Carval
ho, F
Moraes.
Con
gestio
n
-aw
a
re
task ma
ppi
ng
in heter
og
ene
ous MPSoCs
. In
te
rn
a
t
i
onal
S
y
mp
osi
u
m on
SoC.
T
a
mpere
,
F
i
nland. 20
08
: 1–4.
[11]
PK Sah
u
, S C
h
attopad
h
y
a
y
. A
surve
y
on
ap
pl
icat
io
n ma
ppi
n
g
strateg
i
es for
net
w
o
rk-on-c
h
ip
desi
gn.
J.
Syst. Archit..
2013; 59(
1): 60–
76.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Adaptive
Geneti
c
Algorithm
for Mesh
-Bas
ed NoC
Applicatio
n M
appin
g
(Yizhu
o Wan
g
)
7875
[12]
C Ostler, KS C
hatha.
An ILP form
ulation for
system
-lev
el
application m
a
pping on network process
o
r
architectur
e
.
Procee
din
g
s of Desig
n
, Autom
a
tion a
nd T
e
st
in Euro
pe (DA
T
E). Nice, F
r
a
n
ce. 200
7: 1
–
6.
[13]
J Hu, R Marculesc
u
.
Energ
y
-aw
a
re ma
pp
i
ng for tile-b
as
ed No
C archit
ectures un
der
performanc
e
constrai
nts.
Asia a
nd So
uth
Pacific Des
i
g
n
Automatio
n
C
onfere
n
ce (AS
P
-DAC). Kitak
y
us
hu, Ja
pa
n.
200
3: 233
–2
39
.
[14]
I Kenne
d
y
,
RC
Eberh
a
rt.
Particle sw
ar
m opti
m
i
z
at
io
n
. Proceedings
of IEEE In
ternational Conferenc
e
on Ne
ural N
e
tw
o
r
ks. Ne
w
Je
rse
y
, USA. 19
9
5
: 1942
–1
94
8.
[15]
A Col
o
rn
i, M
Dorig
o
, V M
a
niezz
o
. Distri
b
uted
opt
imiz
ati
on
b
y
a
n
t col
oni
es. actes
d
e
la
pr
emièr
e
confére
n
ce e
u
r
opé
en
ne sur la
vie artificie
lle.
Paris, F
r
ance. 199
1: 134
–1
42
.
[16]
HM Harm
ana
n
i
, R F
a
ra
h.
A
meth
od
for effi
cient
ma
pp
ing
and r
e
li
ab
le r
o
uting f
o
r No
C
architectur
e
s
w
i
th
mi
ni
mu
m ban
dw
idth an
d
are
a
.
IEEE Int
e
rnational Wor
kshop on
Circuits and s
y
stem
s and T
A
ISA
Confer
ence (N
EW
CAS-T
A
ISA). Los Alamito
s
, CA. 2008: 29–3
2.
[17]
T
Lei, S Kumar.
A tw
o-step gen
etic a
l
g
o
r
i
thm for
ma
pp
ing task
gra
p
h
s to a
netw
o
rk on
chi
p
architectur
e
. P
r
ocee
din
g
s of
the Eur
o
micro
S
y
m
posi
u
m
on D
i
git
a
l S
y
s
t
em Desi
gn (
D
SD). Bel
e
k-
Antal
y
a, T
u
rkey. 20
03: 18
0–1
87.
[18]
G Ascia, V C
a
tani
a, M Pal
e
si.
Multi-o
b
j
e
ctive map
p
in
g
for mes
h
-b
as
ed N
o
C arc
h
it
ectures
. ACM
Internatio
na
l Confer
ence
o
n
Hard
w
a
re/S
oft
w
ar
e
Co
de
sign a
nd S
y
s
t
em S
y
nth
e
sis
.
Stockholm,
S
w
e
d
en. 20
04:
182–
18
7.
[19]
RK Jena, GK Sharma.
A
mul
t
i-obj
ective ev
o
l
utio
nary a
l
gor
ithm
bas
ed o
p
ti
mi
z
a
t
i
o
n
mod
e
l
for Netw
ork-
on-Chip sy
nthesis.
IEEE Internati
ona
l C
o
n
f
erence
on Inf
o
rma
tio
n
T
e
chnol
og
y
(IT
N
G). Las Ve
gas
,
Neva
da, USA. 200
7: 977
–9
82
.
[20]
RP Dick, DL
Rho
des, W Wolf.
T
G
F
F
: task grap
hs for free.
Proce
edi
n
g
s of the 6th
Internatio
na
l
W
o
rkshop o
n
Hard
w
a
re/Soft
w
a
r
e C
o
-Des
ig
n. Ne
w
York, U
SA. 1998: 97-
1
01.
Evaluation Warning : The document was created with Spire.PDF for Python.