TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.6, Jun
e
201
4, pp. 4361 ~ 4
3
6
7
DOI: 10.115
9
1
/telkomni
ka.
v
12i6.453
1
4361
Re
cei
v
ed O
c
t
ober 1, 20
13;
Revi
se
d Ja
n
uary 11, 201
4
;
Accepte
d
Febru
a
ry 3, 20
14
Load Balancin
g Algorithm of GPU Based on Genetic
Algorithm
Zhang Xiang
y
ang*, Feng
Chaomin, Zh
ao Shugui. Wen Ling. Li Chang
c
hun
PetroC
hi
na R
e
searc
h
Institute of Petrol
e
u
m Exp
l
or
ation &D
evel
opme
n
t-No
rth
w
e
s
t,
Lanz
ho
u, Gansu Provinc
e
73
0
020, Ch
in
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: zxy_
petro@
1
63.com
A
b
st
r
a
ct
As the dev
elo
p
m
e
n
t of GPU/CPU par
all
e
l c
o
mputi
ng
i
n
re
cent years, l
o
a
d
bal
anc
e of GPU serve
r
has bee
n more
an
d mor
e
i
m
p
o
rtant,
so
w
e
p
r
omote a
g
e
n
e
tic al
gor
ith
m
-ba
s
ed
loa
d
bal
an
cing
al
gorith
m
fo
r
GPU RT
M. T
h
e alg
o
rith
m tak
e
s the server
status
and j
ob
assig
n
m
ent i
n
to accou
n
t, and
desig
n a co
di
n
g
m
e
c
h
anis
m
and
genetic
m
a
nipul
ation,
as well as
fitness function. Th
e experim
ents s
how
that,the algor
ithm
can re
ach
a
b
e
tter effect of
efficiency
an
d l
oad-
bal
anc
ing.
It can h
i
d
den
data tra
n
s
m
issi
on i
n
th
e p
a
ral
l
e
l
computi
ng, an
d duri
ng server
dow
nt
ime, it can prev
ent the
idle
of other co
mp
utin
g resour
ces.
Ke
y
w
ords
:
reverse-ti
me
migrati
on, gen
eti
c
algorit
hm,
lo
ad ba
lanc
in
g, GPU/CPU het
erog
ene
ous p
a
r
alle
l
computi
ng, GPU Server
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
With the i
n
creasi
ng difficu
lty of oil expl
oration,
reve
rse time
mig
r
ation, full wa
vefor
m
inversi
on a
n
d
other n
e
w p
r
ocessin
g
technolo
g
ies
as
well a
s
GPU
and othe
r eq
uipment
s hav
e
been a
pplie
d
in sei
s
mic
data processing. Relati
ve
to the CPU serve
r
, GP
U se
rver
gre
a
tly
improve
d
the
parall
e
lism
a
nd computati
onal effici
e
n
cy of reverse time migration,
and ha
s a
go
od
effect of reve
rse
time mi
gration. Howev
e
r, du
e
to th
e la
ck
of loa
d
bala
n
ci
ng
strategy
or j
u
st
throug
h
simpl
e
metho
d
to
achi
eve loa
d
balan
cing
of t
he reverse
-
time mig
r
ation
algorith
m
ba
sed
on GPU serv
er, there i
s
n
o
t suitable lo
ad balan
ci
n
g
algorith
m
for the algorith
m
s ba
sed on G
P
U
serve
r
. In thi
s
pa
pe
r, a
Geneti
c
Algo
rithm ba
se
d
load b
a
lan
c
i
ng alg
o
rithm
of reverse
-
ti
me
migratio
n ha
s been
de
sign
ed to ad
apt to the cha
r
ac
t
e
risti
c
s of GP
U serve
r
s, a
n
d
ma
ke different
batch
es a
nd
different mod
e
ls of GPU
servers dy
nam
ically assig
n
tasks d
epe
ndi
ng on the loa
d
of
servers, so that all serv
ers
can be fully utilized.
Geneti
c
algo
rithm (GA G
enetic Algo
rithm) is a cl
a
ss of self
-org
anizi
ng and
adaptive
artificial intelli
gen
ce tech
ni
que by simul
a
ting evol
utio
n and mecha
n
ism
s
of natural biolo
g
ical
to
solve the
pro
b
lem [1-5]. By the en
codin
g
, fitness fun
c
tion a
nd g
e
netic m
anipul
ation alg
o
rith
ms,
GA ca
n be
u
s
ed to
obtain
environ
ment
al inform
atio
n
to se
arch a
n
d
adju
s
t the
search
dire
ctio
n.
Thro
ugh thi
s
self-o
rg
ani
zin
g
, adaptive chara
c
te
risti
c
s,
GA can a
u
tomatically di
scover th
e la
ws of
the enviro
n
m
ent, so that
it is very suitable
for
real-time stat
us cha
ngin
g
environ
ment
of
appli
c
ation a
nd serve
r
s, a
c
cordi
ng
to
the current st
ate of the
se
rver de
cid
e
t
a
sk allo
catio
n
. In
summ
ary, this pa
per
pre
s
ents a
Gen
e
tic Algo
rithm b
a
se
d dynami
c
loa
d
bala
n
cing algo
rithm
of
reverse
-
time
migratio
n tha
t
takes the e
a
ch
se
rv
er status an
d
se
rver
type i
n
to co
nsi
deration
together, it ca
n flexibly adjust task for ea
ch se
rv
er. On
the one han
d, based o
n
different types of
serve
r
s, it allocate diffe
ren
t
amounts of
task,
on th
e other h
and, a
d
just
s assig
n
m
ents b
a
sed
on
serve
r
lo
ad
st
ate, so th
at, it can
ma
ke ful
l
use
of serve
r
s
and
achiev
e the effe
ct of
load
-bal
an
ce
,
t
hereby
in
cr
e
a
sin
g
t
he loa
d
cap
a
cit
y
of
serv
e
r
s [
6
-11
]
.
2. Parallel Algorithm for
Rev
e
rse-tim
e
Migration
Reverse
time
migration i
s
the seismi
c
d
a
ta
mig
r
ation
method
that
usin
g two-wa
y wave
equatio
n, the basi
c
form
ula
is as follo
ws:
22
22
2
2
11
1
1
11
1
1
()
()
(
)
()
P
PP
P
P
P
x
xy
y
x
x
y
y
vt
v
t
(1)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4361 – 4
367
4362
Whe
r
e,
,,
,
PP
x
y
z
t
is the pre
s
sure
field of medium
，
(,
,
)
x
yz
is the medium
den
sity
，
(,
,
)
vv
x
y
z
is
the veloc
i
ty field
，
,,
,
s
xy
z
t
is the
source te
rm.
With the hi
g
h
order
differen
c
e o
r
comp
act diffe
ren
c
e sch
e
m
e
is used to solve the Equa
tion (1), it can
be use
d
for the
nume
r
ical si
mulation of wave pro
pagati
on. Dabl
ain(1986
) already
discusse
d in
detail the hig
her
orde
r finite differen
c
e soluti
on of the thre
e-dim
e
n
s
iona
l two-way wa
ve equation.
Here just list the
basi
c
calculat
ion formul
a of the forwa
r
d a
nd inverse ex
trapolatio
n [12-15].
The 3d
hig
h
ord
e
r fini
te differen
c
e
wave eq
u
a
tion, with
truncation e
rro
r is
2
,
,
,
t
z
y
x
O
M
M
M
, is
as
follows
:
2
1
,
,
,
,
,
,
0
2
2
1
,
,
,
,
,
,
0
2
2
1
,
,
,
,
,
,
0
2
1
,
,
,
,
1
,
,
2
1
2
1
2
1
2
M
m
n
m
k
j
i
n
m
k
j
i
m
n
k
j
i
M
m
n
k
m
j
i
n
k
m
j
i
m
n
k
j
i
M
m
n
k
j
m
i
n
k
j
m
i
m
n
k
j
i
n
k
j
i
n
k
j
i
n
k
j
i
u
u
u
z
t
v
u
u
u
y
t
v
u
u
u
x
t
v
u
u
u
(2)
The im
pleme
n
tation of
rev
e
rse-time
mi
gration
with
n
o
load
-b
alan
ce is
sh
own in
Figu
re
1, the so
urce
data ha
s b
e
en tota
lly dist
ributed
at the
begin
n
ing,
with no
co
nsi
deratio
n of th
e
serve
r
type a
nd se
rver lo
a
d
state,
re
sul
t
ing in the se
rvers with faster pro
c
e
s
sin
g
spe
ed waiting
the serve
r
s with slo
w
e
r
pro
c
e
ssi
ng speed for
a l
ong time. And once a server failu
re,
job
pro
c
e
ssi
ng ti
me will
be lo
nger agai
n; al
l the othe
r se
rvers mu
st be
waiting fo
r th
e fault server to
finis
h
its
task
.
Figure 1. Parallel Algorith
m
of Reverse
-
time Migratio
n with on Lo
a
d
Balance
Becau
s
e of the pro
b
lem
s
of reverse-ti
me
migratio
n
with no load
-bala
n
ce, through the
optimizatio
n
of software,
desi
gn the
p
o
lling al
go
rithm ba
se
d lo
ad b
a
lan
c
ing
algo
rithm f
o
r
reverse
-
time
migratio
n, a
s
sho
w
n
in Fi
g
u
re
2. Thi
s
al
gorithm
can
a
ssi
gn ta
sk d
u
r
ing th
e
sei
s
mic
data processi
ng, and avoi
d
t
hat all the other
serve
r
s wait for t
he faul
t serve
r
, but becau
se of th
e
algorith
m
without
con
s
id
ering the
serve
r
lo
ad
st
ate
a
nd the
type
o
f
se
rver, j
ob
compl
e
tion ti
me
differen
c
e is
bigge
r, and st
ill cann
ot make full use of the se
rvers.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Load Bala
nci
ng Algorithm
of GPU Base
d on Gen
e
tic
Algorithm
(Zh
ang Xiang
ya
n
g
)
4363
Figure 2. The
Polling Algori
t
hm Based L
oad Bala
n
c
in
g Algorithm for Reve
rse-ti
me Migratio
n
The above al
gorithm
still cannot solve the pro
b
lem o
f
load balan
ci
ng very well, so this
article d
e
si
gn
ed a gen
etic
algorith
m
ba
sed load bal
a
n
cin
g
algo
rith
m of reverse-time migratio
n
.
The algo
rith
m con
s
ide
r
s the different p
r
ocessin
g
ca
pacity of each GPU se
rv
er type, for different
GPU serve
r
s
Sets in differe
nt servi
c
e ma
tching d
egree
. The algorith
m
flow cha
r
t is as follo
ws:
Figure 3. The
Genetic Algo
rithm Base
d Load Bala
nci
ng Algorithm
of Reverse
-
time Migratio
n
3. Algorith
m
Desig
n
for th
e G
e
n
e
tic
Algor
ithm Ba
sed
Load
Balan
c
ing Algo
rithm of
Rev
e
rse-tim
e
Migration
In the pro
c
e
s
s of solvin
g p
r
acti
cal p
r
obl
ems,
ge
netic algorith
m
s ca
nnot
deal
dire
ctly with
the data i
n
the p
r
obl
em
spa
c
e, it
can
handl
e
o
n
ly data exp
r
e
s
sed i
n
the f
o
rm of
gen
o
m
e
chromo
som
e
, and the
r
efo
r
e to u
s
e
g
enetic
algo
rithm to solve
the proble
m
, first of al
l is
conve
r
ting th
e sol
u
tion of
probl
em into
the org
ani
zat
i
on form
of chrom
o
some,
namely codin
g
[16-19].
3.1. Encoding Mechanis
m
Each ta
sk in t
he form of a "
n
(l
) p" d
e
scri
p
tion, re
sp
ect
i
vely task
nu
mber,
and th
e si
ze
of
the task, ta
sk matching d
egre
e
. The n
u
mbe
r
of s
h
o
t
data is use
d
to descri
b
e
the size of the
task,
acco
rdi
ng to the typ
e
of serve
r
setups ta
sk m
a
tchin
g
de
gre
e
, due to
proce
s
sing
ca
p
a
city
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4361 – 4
367
4364
different type
of task of
GPU
se
rver
matchin
g
d
e
g
ree
varie
s
.
In the alg
o
rit
h
ms, e
a
ch g
ene
rep
r
e
s
ent
s a
task to b
e
distributed;
a g
r
oup to
be
a
s
sign
ed ta
sks
that ma
ke
up
a
ch
romo
so
me,
each
ch
romo
some
rep
r
e
s
e
n
ts a
scheme
.
So codin
g
p
r
oble
m
i
s
th
e
prima
r
y p
r
obl
em o
c
currin
g
in
the
u
s
e of
ge
netic algo
rith
m.
The algo
ri
thm
u
s
es
the
de
cimal
en
coding,
su
ch
a
s
a
cl
uste
r of
N
serve
r
s,
in
clu
d
ing GPU
S2
090 se
rvers, GPU
S1
070
serve
r
s
a
nd
GPU k1
0 servers. Du
e
to the
different pro
c
essing capa
city, di
fferent serve
r
types have differ
ent service matchin
g
deg
ree,
servi
c
e
matching
deg
ree
i
s
d
e
scribed
b
y
usin
g p,
If
a
task n
a
me
d
M, ea
ch ta
sk
informatio
n
< n,
l, p
> info
rmat
ion
contai
ns three
info
rmati
on, n fo
r ta
sk
numbe
r, l fo
r
task si
ze,
and
p o
n
behalf
o
f
the task mat
c
hin
g
de
gre
e
.
Obviously,
n, l and
p el
ement a
r
e d
e
c
imal
numb
e
r, a ch
rom
o
so
me
encode
d as
shown:
Figure 4. Chromosome En
codi
ng
3.2.
Fitness Function
Server'
s
statu
s
is th
e impo
rtant factors in
fluenci
ng the
load bal
an
cin
g
. First, a
s
su
me the
total time for
each
s
e
rver
completes
task
s as
s
um
T
, its maximum as
ma
x
T
, the minimu
m v
a
lue a
s
mi
n
T
,average a
s
T
, the minimum differen
c
e
of
s
,The sm
aller of
s
, means that the
task
allocation mo
re bala
n
ced; Secon
d
ly, the load e
rro
r
rate refle
c
ts t
he overall di
stributio
n of the
load of GPU serve
r
s, the
higher
utilization of GP
U serve
r
’s a
n
d
smalle
r load
erro
r rate, t
he
better p
e
rfo
r
mance of th
e
serve
r
, a
se
rv
er fitne
s
s fun
c
tion fo
r
f
,set the
current
se
rver’
s
loa
d
fo
r
i
CL
, new load for
i
NL
, server’
s
utilization rate of GPU for
i
GP
, with mean
GP
, there
are:
p
l
n
l
n
2
1
0
k
q
0
j
p
i
N
N
i
i
NL
CL
T
）
（
(
1
)
N
T
T
N
i
i
0
(
2
)
min
max
T
T
s
(3)
max
T
T
GP
i
i
(
4
)
N
GP
GP
N
i
i
0
(
5
)
The fitness fu
nction of alg
o
r
ithm as follo
w:
i
GP
T
s
f
)
GP
-
1
(
)
1
(
2
i
(6)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Load Bala
nci
ng Algorithm
of GPU Base
d on Gen
e
tic
Algorithm
(Zh
ang Xiang
ya
n
g
)
4365
3.3.
Genetic Opera
t
ions
(1) Sele
ct op
eration. In order to ma
ke the
individual
s with large
r
fitness deg
ree
can be
dire
ctly retain
ed to the ne
xt generation
of gr
oup, se
lect " determ
i
ne the type of samplin
g to
cho
o
se" meth
od, sp
ecifi
c
steps to
see lit
eratu
r
e [3],se
tting thre
shol
d, kee
p
the
server
whi
c
h
i
GP
is bigg
er than
the threshold
direct
ly to the next genera
t
ion of group
s.
(2) T
he cro
s
sover o
p
e
r
ati
on. For con
v
eni
ent of crossover o
p
e
r
ation and m
u
tation
operation, so
the above chrom
o
some
s recom
b
ine,
combi
ned int
o
one dimen
s
ion
a
l encod
ed
string. As
sho
w
n in figure 5
is one-dime
n
s
ion
a
l co
ding
gene
s u
s
ed i
n
the algorith
m
.
Figure 5. One
-
dime
nsi
onal
Codi
ng Ge
ne
s of Ch
romo
somes
Select two p
a
rent individ
u
a
ls from a p
opulatio
n, make cro
s
sover operatio
n, and meet
that interse
c
tion ha
s th
e same g
ene l
o
cation.
Ran
d
o
mly exch
an
ge ge
ne
s whi
c
h a
r
e
on the
left
of the interse
c
tion a
nd a
r
e
not the
sam
e
, after
that t
o
get two ne
w individ
ual
chromo
som
e
s.
Whe
n
a
cro
s
sover op
eration i
s
compl
e
ted, the
ch
an
ge of
gen
e lo
cation
will
lea
d
to the
chan
ge
of task allo
cat
i
on, so the
i
p
value
s
sh
ould
cha
nge
corre
s
po
ndin
g
ly.
(3) M
u
tation.
Due to the
st
ate of the se
rv
er i
s
a re
al
-time chang
e
,
so value
s
o
f
i
p
will
also
ch
ang
e. Whe
n
a
se
rver failure,
i
p
to 0; If the server
re
place
m
ent value
s
of
i
p
is al
so
cha
ngin
g
. Accordingly, after the
com
p
l
e
tion of a m
u
tation, valu
es of
i
p
also
need to
be
modified.
4.Experimen
t
and An
aly
s
is
4.1. Test Cas
e
Article ta
ke
s
an a
c
tual
rev
e
rse-time
mig
r
at
ion ta
sk a
s
the exam
ple;
equip
m
ent in
clud
es
24
serve
r
s of
S2090
GP
U, 24
se
rvers o
f
GPU K1
0,
and
12
se
rve
r
s of
S10
70 GPU,
op
eration
para
m
eters a
s
follo
ws: Co
ntrast
dia
g
ra
m of the
ta
sks total
time a
s
follo
ws, through
the
act
ual
test re
sult, we ca
n
see th
at the effect
of the
alg
o
rit
h
m obvio
usly
is b
e
tter tha
n
the oth
e
r t
w
o
kind
s of algo
rithms, esp
e
ci
ally for the prese
n
ce
of server failure, th
e load-bala
n
ce perfo
rman
ce
of algorithm
effect is bette
r, and with th
e increa
se
of
Work L
oad, the effect of the algo
rithm
will
be more obvi
ous.
Table 1. Para
meter Li
st of Reverse
-
time
Migration Ta
sk
Area
（
Km
2
）
300
Data amount
(
G
b
)
487
Sampling interval
（
ms
）
2
FLOD
120
Trace length
（
ms
）
6000
Bin size
（
m
）
25x25
Shot number
29328
Continuation dep
th
5000
Time continuation
0.4ms
Main frequenc
y o
f
wa
v
e
l
e
t
F
=
2
0
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4361 – 4
367
4366
Figure 6
.
Co
mparison Chart o
f
Algorithm‘s
Ef
fect in
T
r
o
uble-Free Co
ndition
Figure 7
.
Co
mparison chart o
f
Algorithm‘S
Ef
fec
t
with
O
ne Ser
v
er Fa
ilure
for 12
H
ours
Figure 8. Co
mpari
s
o
n
Ch
art of Algorith
m
‘S
Effec
t
with Two Servers
Failure for 12
Hours
Figure 9. Co
mpari
s
o
n
Ch
art of Algorith
m
‘S
Effec
t
with One Server Failure for 24 Hours
4.2. Perform
a
nce Compa
r
ison
Relative to th
e othe
r two
a
l
gorithm
s, ge
netic al
go
rith
m ba
sed l
o
a
d
bala
n
ci
ng a
l
gorithm
of reverse
-
time migratio
n can have bette
r run
n
i
ng effe
ct, the algorit
hm ca
n gen
e
r
ate re
asona
ble
distrib
u
tion schem
e by se
rver loa
d
sta
t
e and t
he type of serve
r
; The pollin
g algorithm j
u
st
assign
ed
a fi
xed nu
mbe
r
of task, a
nd
can't
adj
ust
according
to
the serve
r
st
atus,
so
that
it’s
effect of load balan
cing i
s
not as goo
d a
s
gen
etic
alg
o
rithm; witho
u
t of load balanci
ng algo
rit
h
m,
sei
s
mic d
a
ta
is di
stribute
d
in
one
-tim
e, and
th
e
al
gorithm
com
p
letely do
es
not con
s
ide
r
the
serve
r
state.
Whe
n
serve
r
failure, th
e g
enetic al
go
rit
h
m ba
se
d al
gorithm
can
have the
effect of
load-bala
n
ce, and
comp
ared
with the
p
o
lling
algo
rith
m, it ha
s b
e
tter
effect of
lo
ad-b
a
lan
c
e,
a
nd
as fo
r the
alg
o
rithm
withou
t load b
a
lan
c
i
ng, on
ce
a se
rver failu
re, al
l the othe
r
se
rvers mu
st
wait
until the serv
er is no
rmal.
5. Conclusio
n
The alg
o
rith
m fully con
s
i
dere
d
the typ
e
of
se
rver,
serve
r
lo
ad
state, and the
averag
e
pro
c
e
ssi
ng
time of
ea
ch
serve
r
, thi
s
a
l
gorithm
can
distrib
u
te ta
sks a
c
cording
to the
stat
u
s
of
different serv
ers, m
a
ke full use of
the
se
rvers, and av
oid the influe
nc
e
of se
rver
failure, so it h
a
s
a good a
ppli
c
ation effect.
Referen
ces
[1]
Z
hang W
e
n
x
iu
, Lian
g Yi. T
h
e mathem
atic
al bas
is of
g
e
netic al
gor
ithm
. Xi '
an:
xi '
a
n jia
o ton
g
univ
e
rsit
y
pres
s.
2000.
[2]
Dai W
e
n
h
u
a
. Rese
arch o
n
text cl
assific
a
ti
on a
nd s
e
rver
ing
base
d
o
n
gen
etic al
gor
ithm. Beiji
ng
:
scienc
e press. 200
8.
[3]
Yang Pi
ng, Z
h
eng Ji
nh
ua. Co
mpariso
n
of
ge
netic sel
e
ctio
n
oper
ator an
d stud
y.
Co
mp
uter
engi
ne
erin
g
and a
p
p
licati
o
n
. 2007; 43(
15): 59-6
2
.
[4] Z
hu
Li
，
Sh
en Weiming
，
Pa
n
Shaom
ing,
et al.
A Dyn
a
m
i
c
Loa
d Ba
lanc
ing M
e
tho
d
for
Spatia
l D
a
t
a
Netw
ork Service.
T
he 5th Internatio
nal C
o
nferenc
e on
W
i
reless Com
m
unic
a
tions. N
e
t
w
o
r
ki
ng an
d
Mobil
e
Com
put
ing. Bei
jin
g. 20
09.
66
68
70
72
74
76
Tasks
total
time(hours)
withou
t
load
balanc
i
n
g
Pollin
g
algori
t
h
m
Geneti
c
algori
t
h
m
0
20
40
60
80
100
Tasks
total
time(hours)
withou
t
load
balanc
i
n
g
Pollin
g
algori
t
h
m
Geneti
c
algori
t
h
m
0
20
40
60
80
100
Tasks
total
time(hours)
without
load
balancing
Polling
algorithm
Genetic
algorithm
0
20
40
60
80
100
120
Tasks
total
time(hours)
without
load
balancing
Polling
algorithm
Genetic
algorithm
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Load Bala
nci
ng Algorithm
of GPU Base
d on Gen
e
tic
Algorithm
(Zh
ang Xiang
ya
n
g
)
4367
[5]
Pico
CAG, Wain
w
r
ight RL
．
Dy
namic
Scheduling of Computer
T
a
sks Using Genetic A
l
gor
ithms.
IEEE
W
o
rld Co
ngres
s Computati
o
n
a
l Intell
ige
n
ce.
199
4; (2): 829-
833.
[6]
Z
o
ma
ya AY, Y
ee-H
w
e
i
T
.
Observatio
ns o
n
Using
Gen
e
tic
Algor
ithms for
D
y
nam
ic L
o
a
d
-Bal
anci
n
g
.
Parall
el a
nd Di
stributed Syste
m
s.
20
01; 12(
9
)
：
899-91
1.
[7]
GB Z
heng. Achievi
ng H
i
gh
Performance
on Ex
trem
el
y Lar
ge Par
a
l
l
el Mach
in
es: Performan
c
e
Predicti
on a
nd
Loa
d Bal
anci
n
g
．
Urban
a
：
UIUC. 200
5
[8] JM
Bahi
，
et al
. D
y
n
a
mic lo
a
d
bal
anc
ing
a
nd efficie
n
t lo
ad estimators
for as
y
n
c
h
ro
nous iter
ativ
e
algorithms.
IEEE T Parall Distrib.
200
5; 16(4)
：
289-2
9
9
．
[9]
YIN W
en, YIN Xi
ng-
ya
o, Z
H
ANG F
an-cha
ng. A st
ud
y
o
n
seism
i
c attri
bute o
p
timiz
a
ti
on b
a
se
d o
n
para
lle
l ge
netic
algor
ithm.
Jou
r
nal of Jil
i
n Un
i
v
ersity (Earth Scienc
e Editio
n).
2005; 3
5
(5)
:
672-67
6.
[10] Yu
Lei
，
Li
n Z
ong-K
a
i
，
Guo Yu-Ch
a
i
，
Li
n Shuo-
Xun. Lo
ad ba
lanc
in
g and fau
l
t-toler
ant services i
n
multi-server s
ystem.
Journal
of System Si
mulati
on
. 20
01; 13(3): 32
5-3
2
8
(in Chi
nese)
[11]
Li Shu
a
n
g
-Qin
g, Gu Ping, C
hen
g Da
i-Jie.
Anal
ys
is an
d researc
h
on l
o
a
d
bal
anc
ing str
a
teg
y
in W
e
b
server s
y
stem.
Journ
a
l of Co
mputer Eng
i
n
eeri
ng an
d App
lica
t
ion
. 200
2; (19)
: 40-43 (in ch
in
ese)
[12] Liu
M
i
n
，
W
u
Che
ng, Yi
n
W
enju
n
. Solvi
ng i
d
e
n
tical
p
a
rall
el m
a
chi
n
e pro
ducti
on l
i
ne sc
he
dul
ing
prob
lem
w
i
th
s
peci
a
l
proc
edu
re co
nstraint
b
y
ge
netic
al
gor
ithm.
Acta Automatica Sinica
. 200
1; 2
7
(3)
:
381-
386 (i
n Ch
ines
e)
[13]
Liu Mi
n, Hao J
i
ngh
ua, W
u
Ch
eng. A g
enetic
algor
ithm for para
lle
l mach
in
e sche
dul
ing
p
r
obl
ems
w
i
t
h
proce
dure c
o
n
s
traint and its a
pplic
atio
ns.
Chi
nese Jo
urn
a
l o
f
Electronics.
2
006; 15(
3): 463
-466.
[14]
Ramamrith
am
KJ, Stankovic
A, Shia
h PF
. E
ffici
ent sch
ed
ul
ing
al
gorit
hms
for rea
l
-time m
u
ltipr
o
cesso
r
s
y
stems.
IEEE Trans on Parallel an
d Distributed System
s
. 1
990; 1(2): 1
84-
194.
[15]
Manimar
an G, Murth
y
CS
R. An efficie
n
t d
y
namic
sch
ed
uli
ng al
gor
ithm for multipr
o
ces
s
or real-ti
m
e
s
y
stems.
IEEE Trans on Parallel
and Distributed System
s.
1998; 9(3): 3
12-
319.
[16]
Xu
eso
ng Ya
n, W
e
ighted K-
Near
est Neig
h
bor Cl
assificati
on Alg
o
rithm
Based o
n
Gen
e
tic Algor
ithm
.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
ng.
2013; 1
1
(10):
617
3-61
78.
[17]
Yanb
ei
Li,
Lei
Yan,
Hu
a Qi
an. A Ga
it R
e
c
ogn
ition
S
y
st
em us
ing
GA-base
d
C-SV
C
an
d Pl
anta
r
Pressure.
T
E
L
K
OMNIKA Indones
ian J
ourn
a
l of Electrica
l
Engi
neer
in
g.
2013; 11(
10): 61
35-6
142.
[18]
YU Z
h
ij
un. R
B
F
Neura
l
N
e
t
w
orks O
p
timi
zation A
l
g
o
rith
m and A
p
p
lic
ation
on T
a
x F
o
recastin
g.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2013; 1
1
(7): 3
491-
349
7.
[19]
Jings
ong Ya
ng
. A personifica
tion he
uristic
Genetic Alg
o
rit
h
m for Digit
al Microflui
d
ics-b
a
sed
Bi
ochi
p
s
Placem
ent.
T
E
LKOMNIKA Indon
esia
n Jour
nal
of Electric
al
Engin
eeri
ng.
2
013; 11(
6): 318
7-31
93.
Evaluation Warning : The document was created with Spire.PDF for Python.