Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
11
,
No.
3
,
June
2021
,
pp. 2
525~
2534
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v11
i
3
.
pp2525
-
25
34
2525
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om
A smart
method f
or spark
using
ne
ural n
etwork
f
or big data
Md. Ar
ma
nur
Rahm
an
1
,
J.
Ho
sse
n
2
, Az
iz
a
S
ultana
3
, Ab
dull
ah
Al
Mamun
4
, N
or Az
l
ina
Ab
.
A
z
iz
5
1,2,4,5
Facul
t
y
of E
ngine
er
ing
an
d
Te
chno
log
y
,
Multi
m
edi
a
Univer
s
ity
,
Mela
k
a, Ma
l
a
y
si
a
3
Facul
t
y
of
Com
puti
ng
and
Eng
i
nee
ring
,
Dhak
a
I
nte
rna
ti
ona
l
Uni
ver
sit
y
,
Dhaka
,
Bangl
ad
esh
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
ug
24, 202
0
Re
vised
Oct
6
,
2020
Accepte
d
Oct
27
, 202
0
Apac
he
s
par
k
,
f
amous
l
y
known
for
big
data
han
dli
ng
ability
,
is
a
distri
but
ed
open
-
source
fr
a
m
ework
tha
t
u
t
il
izes
th
e
id
ea
of
distri
but
ed
m
emory
to
proc
ess
big
data
.
As
the
per
form
a
nce
of
th
e
spa
rk
is
m
ostl
y
bei
ng
aff
ecte
d
b
y
the
spark
pre
do
m
ina
nt
conf
igur
at
ion
par
amete
rs
,
it
is
cha
llengin
g
to
ac
hie
v
e
the
opt
imal
r
esul
t
from
spark
.
Th
e
cur
r
ent
pr
actice
of
tun
ing
th
e
p
ara
m
et
ers is
ine
ffe
ct
iv
e,
as
it
is
per
form
ed
m
anua
lly
.
Manu
al
tuni
ng
is
challe
nging
fo
r
la
rge
spa
ce
of
par
amete
rs
and
complex
intera
c
ti
ons
with
and
among
the
par
amete
rs
.
This
pape
r
proposes
a
m
ore
eff
ec
ti
v
e,
self
-
tuni
n
g
appr
oac
h
subjec
t
to
a
n
e
ura
l
ne
twork
c
a
ll
ed
Sm
art
m
ethod
for
spark
using
neur
al
net
work
for
big
dat
a
(SS
NN
B)
t
o
avoi
d
the
d
isa
dvant
ag
es
of
m
a
nual
tun
ing
of
the
par
amete
r
s.
The
pape
r
has
sele
cted
five
pr
edominant
par
a
m
et
ers
with
five
d
iffe
r
ent
si
z
es
of
da
ta
to
te
s
t
the
appr
o
ac
h
.
The
proposed
a
pproa
ch
h
as
inc
re
ase
d
th
e
spee
d
of
aro
und
30%
compare
d
with
the
d
efa
u
lt
par
amet
er
conf
iguration.
Ke
yw
or
d
s
:
Ap
ac
he spa
rk
Bi
g
data
Config
ur
at
io
n param
et
ers
Ma
chine
le
a
rn
i
ng
Self
-
c
onfig
ur
at
ion
This
is an
open
acc
ess
arti
cl
e
un
der
the
CC
B
Y
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Md. Arm
anur
Ra
hm
an
Faculty
of E
ngineerin
g
a
nd T
echnolo
gy
Mult
i
m
edia Universit
y
Me
la
ka,
75450, Ma
la
ysi
a
Em
a
il
: ar
m
an.
bd
m
ai
l@gm
a
il.c
om
1.
INTROD
U
CTION
Aroun
d
t
he
world,
the
num
ber
of
onli
ne
us
e
rs
is
i
ncr
easi
ng
at
a
ra
pid
rate
with
t
he
a
dva
ncem
ent
of
so
ci
al
com
m
u
nicat
ion
a
nd
e
-
com
m
erce
bu
s
iness.
Be
sides
,
a
lot
of
us
er
s
are
stori
ng
the
ir
con
te
nt
cons
ta
ntly
for
f
utu
re
us
e.
As
ind
ic
at
ed
by
In
te
r
natio
na
l
data
cor
po
r
at
ion
(IDC)
,
di
gital
sp
ace
is
pro
j
ect
e
d
to
in
crease
m
or
e
than
44
Z.B.
in
volum
e
by
2020
[1
-
3].
In
t
he
era
of
dig
it
al
data,
big
data
is
som
et
hin
g
that
can'
t
be
ov
e
rlo
oked
.
T
her
e
fore
rece
nt
ly
,
the
big
data
era,
di
ff
e
ren
t
industries
an
d
governm
ents
hav
e
giv
e
n
em
ph
a
sis
on
big
data
te
chnolo
gies.
Si
nc
e
the
conven
ti
on
al
com
pu
ti
ng
te
chn
i
qu
e
s
cou
l
d
not
pro
vid
e
the
ex
pected
res
ult
and
ef
fici
ency
to
m
anag
e
bi
g
data.
T
he
diff
e
ren
t
distrib
uted
fr
am
ewo
r
ks
li
k
e
h
ad
oop
[
4],
sp
ar
k
[5
]
,
a
nd
storm
[6
]
hav
e
b
ee
n
i
ntr
oduce
d
to
sa
ti
sfy the prere
quisi
te
o
f
taki
ng care
of the
bi
g
data.
Ap
ac
he
sp
a
r
k
is
one
of
t
he
m
os
t
no
ta
ble
an
d
broa
dly
us
e
d
fr
am
ewo
r
ks
beca
us
e
of
it
s
hi
gh
pe
r
form
ance
and
fle
xib
il
it
y
[7
]
.
Ap
ac
he
s
pa
r
k
ha
s
over
18
0
param
et
ers
with
de
fau
lt
v
a
lues.
The
a
ppr
opriat
e
values
of
the
par
am
et
er
can
be
sel
ect
ed
by
the
use
r
m
anu
al
ly
wh
il
e
pr
ocessin
g
dif
fere
nt
siz
es
an
d
t
ypes
of
data.
Th
e
pe
rfor
m
ance
bec
om
es
un
sat
isfac
tory
due
to
t
he
ina
pprop
riat
e
sel
ect
ion
of
pa
ram
et
er
values.
Ther
e
f
or
e,
a
ddit
ion
al
tu
n
in
g
of
the
pa
ram
eter
is
require
d
f
or
each
pa
rtic
ul
ar
ap
plica
ti
on
[8
]
.
The
use
rs
require
appr
opriat
e
know
le
dg
e
f
or
m
anu
al
tu
ning
of
the
par
am
et
er
s
in
the
sp
ar
k
f
ram
ewo
rk,
ho
wev
e
r,
m
anu
al
tun
in
g
is ver
y t
e
dious
du
e
to
t
he
c
omplex i
nteracti
on
betwee
n
the
m
.
A
s
pe
r
the
c
ur
ren
t
pr
act
ic
e,
par
am
et
er
tun
i
ng
i
n
bi
g
da
ta
is
perform
ed
in
2
ways.
Fir
stl
y,
m
anu
al
tun
in
g
of
the
par
am
et
er
by
tria
l
an
d
er
ror.
This
process
i
s
ve
ry
com
plica
te
d
as
it
re
qu
ires
a
lo
ng
ti
m
e
an
d
dep
t
h
kn
ow
le
dge
due
to
a
la
r
ge
num
ber
of
par
am
et
ers
and
it
s
internal
co
r
relat
ion
with
e
ach
ot
her.
To
a
ddress
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
11
, No
.
3
,
J
une
2021
:
25
25
-
2534
2526
the
m
anu
al
tun
in
g
pro
blem
,
[9
]
aut
hor
pr
opose
d
a
c
os
t
-
ba
sed
m
od
el
for
the
ha
doop
s
yst
e
m
.
Ho
we
ve
r,
the
m
od
el
need
s
t
o
be
perseve
re
d
by
us
ers
bas
ed
on
diff
e
re
nt
po
li
ci
es.
S
ec
ondly,
sel
f
-
tu
ni
ng
pa
ram
et
er
wh
e
n
it
requires
.
This
pap
e
r
proposes
an
ap
proac
h
ba
sed
on
a
ne
ural
networ
k
to
m
ini
m
iz
e
the
dr
a
w
back
of
m
anu
al
tun
in
g.
T
he
re
search
dev
el
oped
a
sel
f
-
tu
ning
ap
proac
h
tha
t
can
per
f
orm
sel
f
-
tu
ning
of
the
par
am
et
er
range
base
d
on
the
neural
netw
ork
m
od
el
.
T
his
app
r
oac
h
has
three
key
ad
van
ta
ges
com
par
e
d
to
the
existi
ng
appr
oach
es
.
Firstl
y,
al
l
ta
sk
s
are
processe
d
by
the
neural
ne
twork
m
od
el
.
Seco
nd
ly
,
al
l
t
ypes
of
dataset
s
that
consi
st
of
str
uc
ture
d
data,
se
m
i
-
structur
e
d
data,
an
d
un
st
r
uctu
red
data
ca
n
be
pro
cesse
d.
Thir
dly,
any
vo
l
um
e
of the
dataset
c
an be
processe
d.
The
trai
ni
ng
da
ta
has
bee
n
c
ollec
te
d
f
or
th
e
sel
ect
ed
five
par
am
et
ers
by
changin
g
th
e
par
am
et
er
range
an
d
var
i
ou
s
in
put
of
da
ta
set
s.
The
trai
nin
g
process
is
on
ly
fo
r
one
tim
e
to
le
arn
the
m
achi
ne
le
a
rn
i
ng
m
od
el
,
w
hich
then
can
pre
dict
the
nu
m
erical
valu
es
f
or
the
s
el
ect
ed
par
am
et
ers.
T
he
m
et
ho
d
ha
s
be
e
n
i
m
ple
m
ented
on
a
te
s
tbed
th
at
us
es
Dell
Po
we
rE
dg
e
R
720
se
rv
e
r,
hos
ti
ng
sp
a
rk
f
ra
m
ewo
r
k,
an
d
r
un
s
as
sp
ar
k
no
des.
T
he
te
st
resu
lt
s p
r
ovide
that
ou
r
pro
posed
m
eth
od
ca
n
pe
rform
eff
ect
ive
sel
f
-
tu
ning b
ased o
n
t
he
neural
netw
ork
m
od
el
so
that
it
m
eet
s
m
axi
m
u
m
reso
urce
us
age
capa
bil
it
y
and
saves
processi
ng
ti
m
e.
The
key comm
itm
e
nts
of
t
he
m
et
ho
d are
as
fo
ll
ows:
It
has
im
ple
mented
a
n
arti
fici
al
neu
ral
net
work
i
n
the
a
ppr
oach
that
processes
s
park
jo
bs
us
i
ng
i
ts
app
li
cat
io
n
ser
vice
base
d
on
the
ne
ur
al
network
m
od
el
.
Hen
ce
,
use
rs
do
not
re
quire
i
n
-
de
pth
knowle
dg
e
of the i
nternal
syst
e
m
f
un
ct
io
n.
Th
us
,
they c
an
sa
ve
ti
m
e b
y avo
i
ding m
anu
al
t
un
i
ng.
The
sel
f
-
tu
ning
facil
it
y
of
the
ap
proac
h
integ
rates
pa
ra
m
et
er
ran
ge
a
ll
ocati
on
.
It
he
lps
to
m
eet
t
as
k
dead
li
ne
s a
nd im
pr
ov
es t
he o
ver
al
l
perform
ance
of sp
a
r
k
.
In
our
eval
uat
ion
us
in
g
sp
a
r
k
w
orkloa
ds
with
five
dif
fe
ren
t
in
pu
t
dataset
s,
the
ap
proach
ac
hieve
d
an
aver
a
ge per
f
orm
ance sp
ee
dup o
f
a
bout
30
%
perform
ance.
The
rem
ai
ns
of
the
pa
per
ar
e
organ
iz
e
d
as
fo
ll
ows.
Sect
ion
2,
prese
nting
the
backg
r
ound
of
th
e
stud
y.
Sect
i
on
3,
the
relat
ed
work,
is
disc
usse
d.
Sect
i
on
4
pr
ese
nts
the
de
ta
il
s
of
the
arti
fici
al
neu
ral
ne
tw
ork
.
S
e
ct
ion
5
pr
es
ents
the
arch
it
e
ct
ur
e
of
SS
NNB
.
T
he
m
e
tho
dolo
gy
is
pr
ese
nt
ed
in
s
ect
ion
6.
Sect
io
n
7
pr
esents
resu
lt
s a
nd a
na
ly
sis
-
finall
y, Con
cl
us
io
ns
a
nd
futur
e
wo
rk pr
esented
in
sect
i
on 8.
2.
BACKG
ROU
ND OF
THE
STUDY
2.1.
Spa
r
k
In
t
he
area
of
big
data
,
“
Ap
a
che
S
park
”
is
t
he
m
os
t
acce
pted
ope
n
-
s
ource
platfo
rm
that
su
pp
or
ts
t
he
idea
of
resil
ie
nt
distrib
uted
dataset
s
(RD
D
s).
T
he
RD
Ds
al
low
rap
i
d
tr
eat
ing
of
the
m
assive
siz
e
of
data
le
ver
a
ging
distrib
uted
m
e
m
or
y.
Data
op
e
rati
on
in
m
e
m
or
y
is
appr
opriat
e
for
re
petit
ive
a
pp
li
cat
io
ns
s
uc
h
as
gr
a
ph
al
gorith
m
s
and
reit
era
ti
ve
m
achine
le
arn
i
ng.
RD
D
is
co
ns
ide
red
as
the
m
ai
n
featur
e
of
s
pa
rk
.
I
t
char
act
e
rizes
a
read
-
only
colle
ct
ion
of
entit
ie
s
al
locat
ed
am
on
g
seve
ral
m
achines.
A
n
RDD
e
xp
li
ci
tl
y
stores
in
the
cac
he
m
e
m
or
y
by
th
e
us
e
r
ov
e
r
se
ver
al
m
achine
s
an
d
ca
n
be
reu
se
d
a
s
the
par
al
le
l
ope
rati
on
i
n
m
ul
ti
ple
Ma
p
Re
du
ce
.
R
DD
ha
s
the
fa
ult
tol
eran
ce
abili
ty
ov
e
r
a
no
ti
on
of
e
xtracti
on.
Wh
ene
ve
r
a
pa
rtit
ion
of
RDD
is
lost,
it
can
re
bu
il
d
it
since
it
has
su
f
fici
ent
inf
o
r
m
at
ion
reg
a
rd
i
ng
it
s
ori
gin.
Th
ough
RD
Ds
do
no
t
hav
e
s
har
e
d
m
e
m
or
y
con
st
ruct
ion
,
on
the
one
ha
nd,
they
can
re
pr
ese
nt
r
el
ia
bili
ty
and
scal
abili
ty
and
,
on
th
e
oth
e
r
ha
nd,
a
s
weet
-
s
po
t
am
on
g
e
xpressi
vity
.
RDDs
a
re
well
-
su
it
ed
f
or
a
div
er
sit
y
of
app
li
cat
io
ns
.
F
igure
1
pr
ese
nts
the
s
park
-
cl
us
te
r
fra
m
ewo
r
k
[10].
A
s
pa
r
k
c
omprises
a
dri
ve
r
node
t
hat
is
equ
i
valent
to
a
m
ast
er
node
a
nd
se
ve
ral
w
orke
r
node
s
that
are
co
r
respo
nd
e
nt
to
sla
ve
nodes.
T
he
dr
i
ver
node
m
anag
es
al
l
worker
nodes
t
hro
ugh
the
w
orke
r
node
proce
s
s.
T
he
w
orke
r
node
s
com
m
un
ic
at
e
with
the
dri
ver
node
t
hro
ugh
t
he
worker
node
proces
s
a
nd
m
a
nag
e
local
e
xe
cuto
rs.
Each
a
pp
li
cat
io
n
c
onsist
s
of
m
ulti
ple
exec
uto
r
s
a
nd
one
dr
i
ver
.
All
the
j
obs
in
an
a
ppli
cat
ion
com
e
fr
om
the
sa
m
e
execu
to
rs.
T
he
sp
a
rk
c
on
te
xt
is
creati
ng
by
the
m
ai
n
j
obs
of
th
e
ap
plica
ti
on
t
hat
are
r
un
by
t
he
dr
i
ver
proce
ss.
Eac
h
of
the
worker
nodes
a
ccom
plishes
one
or
m
or
e
execu
t
or
bac
kend
proc
ess
du
rin
g
la
unchi
ng,
a
nd
a
sing
le
e
xec
utor
back
e
nd
do
e
s
m
anag
ing
ex
ecu
t
or
instance. An
e
xecu
t
or
m
an
ag
es a thr
ead
gro
up
that r
uns ea
ch
of the tasks a
s a sing
le
thr
e
ad.
N
e
ve
rthele
ss,
the
tim
e
of
exec
ution
of
a
s
pecifi
c
ta
sk
in
the
pl
at
fo
rm
of
A
pa
che
dep
e
nds
on
var
io
us
facto
rs
s
uch
as
in
pu
t
data
vo
l
um
e,
data
ty
pe,
CPU
spe
ed,
m
e
m
or
y
siz
e,
nu
m
ber
of
no
des,
c
onfig
ur
at
io
n
pa
r
a
m
et
ers,
desig
n
an
d
i
m
ple
m
entat
io
n
of
the
syst
em
and
so
on.
Ba
sed
on
these
factor
s
,
the
ti
m
e
of
execu
ti
on
tim
e
of
a
sp
eci
fic
jo
b
in
apache
sp
a
r
k
m
a
y
diff
er
c
on
s
pic
uous
ly
[
11
]
.
T
her
e
is
m
or
e
than
180
config
ur
at
io
n
par
am
et
er
in
apach
e
sp
ar
k
t
hat
us
er
can
tu
ne
acc
ordi
ng
to
t
he
ne
ed
of
a
s
pecifi
c
ap
plica
ti
on
t
o
e
nhance
the
perform
ance.
I
t
is
the
m
od
est
and
m
os
t
op
e
rati
ve
a
ppr
oach
to
e
nhance
the
enac
t
m
ent.
Users
t
un
e
these
pa
ra
m
et
ers
ph
ysi
ca
ll
y
by
exp
e
rim
ent
[1
2].
At
prese
nt,
the
par
am
et
ers
are
m
anu
al
ly
tun
e
d
by
e
xperim
entation
that
is
no
t
e
ff
e
ct
ive.
I
t
needs
com
plica
te
d
interact
io
ns
with
the
pa
ram
et
ers
and
ta
kes
a
la
rger
pa
ram
et
er
sp
ace.
A
gain
,
these
par
am
et
ers
m
u
st be
re
-
t
un
e
d f
or v
a
rio
us
a
ppli
cat
ion
s a
nd clusters
.
Ar
ti
fici
al
ne
ural
netw
orks
(
AN
N)
is
a
m
at
hem
a
ti
cal
pr
oc
essing
m
et
ho
d
that
ca
n
be
us
e
d
f
or
bot
h
cl
assifi
cat
ion
and
re
gressi
on
[13,
14]
.
The
neur
on
s
m
ake
it
a
po
we
rful
le
arn
i
ng
m
od
el
fo
r
this
reas
on
f
or
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
A smart
meth
od fo
r s
pa
r
k
us
ing
Ne
ural N
et
work fo
r
big
data
(
M
d.
Arm
anur R
ahm
an
)
2527
regressio
n
a
naly
sis.
It
is
the
best
cho
ic
e,
incl
ud
i
ng
m
ult
iple
inputs
an
d
out
pu
t
data
[15,
16]
.
A
ne
ural
ne
twork
can
predict
nu
m
erical
values
correct
ly
,
and
it
can
pr
ev
ent
over
fitt
ing
easi
ly
.
ANN
is
m
uch
su
it
able
in
se
ver
al
areas, i
nclu
ding
natu
ral la
ngua
ge
a
nd im
age p
r
ocessi
ng, pre
dicti
on
a
s
well
as em
otion
r
ec
ogniti
on [1
7
-
19]
.
Figure
1. A
com
m
on
la
yout
of a
pach
e
sp
a
r
k
3.
RELATE
D
W
ORK
In
recent
ye
ars
,
one
of
the
ke
enest
resea
rc
h
is
in
the
op
t
i
m
iz
ation
of
th
e
perform
ance
of
big
data
syst
e
m
.
Howe
ver,
al
m
os
t
all
the
existi
ng
researc
hes
ha
ve
been
done
on
the
Ha
doop
platfo
rm
or
t
h
e
fr
am
ewo
r
k
of
Ma
pRed
uce
c
om
pu
ti
ng
.
Star
fish
[
9]
util
iz
es
sim
ulati
on
and
a
c
os
t
-
base
d
m
od
el
to
se
ek
t
he
require
d
jo
b
c
onfig
ur
at
io
n
for
the
work
l
oad
of
Ma
pRed
uc
e.
AROM
A
[
20]
us
es
a
n
op
ti
m
iz
at
ion
fr
am
ewor
k
and
tw
o
-
phase
ML
to
automa
te
reso
ur
ce
distribu
ti
on
a
nd
j
ob
co
nf
i
gurati
ons
co
ns
i
der
i
ng
he
te
rog
eneous
cl
ouds
.
T
he
a
uthors
of
[
21
]
,
ind
ic
at
ed
that
hado
op
s
cheduler
in
t
he
heter
ogene
ou
s
e
nviro
nme
nt,
the
perform
ance
r
edu
ct
io
n
an
d
pro
po
se
d
a
nother
sche
dule
r
nam
ed
longest
ap
pro
xim
a
te
tim
e
to
en
d
.
I
n
[
22
]
a
diff
e
re
nt
w
ork
con
ce
ntrate
d
on
e
xam
ining
the
dif
fer
e
nt
re
so
urce
c
onsu
m
ption
e
ff
ect
s
f
or
var
ia
nt
set
f
or
t
he
Re
du
ce
slots
a
nd
Ma
p.
T
hese
pro
blem
s
have
bee
n
a
ddress
ed
in
[
23]
,
th
r
ough
a
f
ram
ewo
r
k
cal
le
d
“P
r
of
il
in
g
and
Per
form
ance
-
base
d
Syst
e
m
”
(P
PA
BS
),
wh
ic
h
can
at
om
ic
al
l
y
tun
e
th
e
config
ur
at
i
on
of
ha
doop
set
ti
ng
by
deducti
ng
th
e
requirem
ents
of
a
pp
li
cat
ion
pe
rfor
m
ance.
M
od
i
fyi
ng
the
popula
r
KMea
ns+
+
cl
us
te
ri
ng
al
ong
with
the
sim
ulate
d
A
nneal
in
g
al
gorithm
are
the
m
ai
n
cont
ribu
ti
ons
of
[
24
]
,
w
hich
we
re
nee
de
d
to
a
dju
st
to
the
Ma
pRed
uc
e
par
a
dig
m
.
Re
fer
e
nce
[
23
]
reco
m
m
end
s
easi
ng
this
iss
ue
by
an
en
gine
that
su
ggest
s
the
config
ur
at
io
ns
for
a
ne
w
a
nal
yt
ic
al
j
ob
ti
m
e
ly
and
intel
li
ge
ntly
.
This
e
ngine
is
em
bedded
i
n
an
ada
pt
ed
k
-
near
est
neig
hbor
(KN
N)
al
go
rithm
to
discover
the
ap
pro
pr
i
at
e
config
ur
at
ion
base
d
on
th
e
past
job
e
xpe
rienc
e
that
is
execu
te
d
well
.
H
ow
e
ve
r,
the
resea
rch
of
optim
iz
ing
apach
e
s
park
pe
rfor
m
ance
is
sti
ll
in
the
beg
i
nn
i
ng
sta
ge.
T
he
aut
hors
of
[
24]
,
present
a
sim
ula
ti
on
dri
ve
n
f
oreca
st
m
od
el
to
antic
ipat
e
the
perform
ance
of
a
jo
b
with
hi
gh
c
orre
ct
ness
f
or
Apa
che
S
park.
T
he
ir
pro
posed
m
od
el
can p
re
dict
m
e
m
or
y
us
age
and
e
xec
utio
n
tim
e
of
s
park
syst
em
s
in
the
case
of
de
fa
ult
para
m
et
ers.
[25]
Show
e
d
that
the
sup
port
vec
tor
re
gr
es
sio
n
(S
VR
)
m
od
el
is
com
pu
ta
ti
o
nally
ef
fici
ent
with
hi
gh
accu
racy.
Acc
ordin
g
t
o
th
ei
r
fin
dings,
it
ca
n
be
c
oncl
ude
d
th
at
us
in
g
the
aut
o
-
tun
in
g
m
et
ho
d
can
offer
c
ompara
ble
or
bette
r
perf
or
m
ance
com
par
ed
to
sta
rf
ish
with
a
few
e
r
nu
m
ber
of p
a
ra
m
et
ers.
4.
AR
TIF
ICIAL
N
EU
R
AL
NETWOR
K (A
N
N)
The
sci
kit
-
le
ar
n
is
an
esse
ntial
too
l
since
it
al
lows
on
ly
a
few
li
ne
s
of
cod
i
ng
an
d
prevalent
data
gro
undwo
rk.
I
n
ord
er
to
proc
eed
with
t
he
e
valuati
on,
th
e
Ker
as
wr
a
pper
s
nee
d
to
be
pr
ov
i
ded
with
a
def
i
ne
d
functi
on
to
cr
eat
e
AN
N
.
I
n
fact,
the
f
un
ct
i
on
is
f
or
m
ulate
d
to
create
a
base
m
od
el
t
hat
is
the
su
bj
ect
of
evaluati
on.
T
he
base
m
od
el
is
connecte
d
wi
th
three
ne
uro
ns
throu
gh
a
hidden
la
ye
r,
as
il
lustrate
d
in
Fig
ur
e
2.
The
hi
dd
e
n
a
nd
outp
ut
la
ye
r
is
act
ivate
d
with
Re
LU
a
nd
s
oft
m
ax
act
i
vation
functi
ons
.
F
ur
t
her
m
or
e,
a
n
eff
ic
ie
nt
opti
m
iz
er
"
A
dam
"
c
an
be
use
d
to
update
netw
or
k
weig
hts
it
era
ti
vely
based
on
trai
ning
data
.
The
obj
ect
i
n
the
K
eras
wr
a
pper,
known
as
KerasR
egr
es
sor,
is
us
e
d
as
a
regr
ession
est
im
a
t
or
in
t
he
sci
kit
-
le
ar
n.
The
f
unct
ion
of
A
NN
is
t
hen
creat
ed
im
m
e
diate
ly
to
pass
par
am
et
ers
including
the
ba
tc
h
siz
e
an
d
e
po
c
hs
nu
m
ber
al
on
g
with
the
f
unct
ion
of
t
he
m
od
el
,
bo
t
h
of
w
hi
ch
are
set
to
def
a
ult.
F
ur
the
rm
or
e,
a
proce
ss
of
arb
it
ra
ry
num
ber
creato
r
with
a
const
ant
ar
bi
trary
seed
has
been
i
niti
al
iz
e
d
to
c
om
par
e
the
co
ns
ist
ency
of
t
he
m
od
el
s.
In
this
researc
h,
the
pr
oce
ss
of
ar
bitrary
num
ber
creator
s
is
rep
eat
ed
f
or
th
e
evaluati
on
of
each
m
od
el
.
A
neur
on
ta
kes
in
pu
t
s,
does
so
m
e
m
at
h
with
them
,
and
pro
duc
es
an
outp
ut.
A
sim
ple
neu
r
on
lo
oks
li
ke
w
hat is s
how
n
i
n
Fi
gure
3.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
11
, No
.
3
,
J
une
2021
:
25
25
-
2534
2528
Figure
2. A
n
e
ur
al
netw
ork w
it
h
hidde
n
la
ye
r
Figure
3. Lay
out o
f
a
sim
ple
neur
on
Thr
ee
thin
gs ar
e h
a
pp
e
ning
he
re. Fi
rst, eac
h
i
nput is m
ulti
pli
ed by a
weig
ht:
1
→
1
∗
1
,
2
→
2
∗
2
,
→
∗
(1)
Nex
t,
all
the
w
ei
gh
te
d i
np
uts
are a
dd
e
d
t
og
et
her with
a
bias
b
:
(
1
∗
1
)
+
(
2
∗
2
)
+
(
∗
)
+
(2)
Finall
y, the s
um
is passe
d
th
r
ough a
n
act
ivat
ion
f
un
c
ti
on:
=
(
1
∗
1
+
2
∗
2
+
∗
)
+
(3)
4.1.
Act
i
vation f
un
ctions ReL
U a
nd
s
oftm
ax
Re
ct
ifie
d
li
nea
r
un
it
(ReLU
),
is
a
rece
ntly
popula
r
act
ivati
on
f
unct
ion
in
neural
netw
ork
s
[
26
-
28
]
.
It
is
well
-
def
i
ned
as
(
)
=
(
0
,
)
.
O
n
e
of
the
ad
va
ntages
of
the
f
unct
ion
i
s,
it
is
al
so
no
n
-
li
nea
r
a
nd
ca
n
run
bac
kw
a
rd
for
er
r
or
m
inim
iz
at
ion
.
A
dd
i
ti
on
al
ly
,
the
functi
on
act
ivate
s
m
ulti
ple
neuron
la
ye
rs
.
Fi
gure
4
sh
ows
the
rec
ti
fied
li
nea
r u
nit
(ReLU
)
act
ivat
ion
f
un
ct
io
n.
So
ftm
ax
is
a
t
ype
of
l
og
ist
ic
functi
on
i
n
m
at
hem
atics.
The
s
of
tm
ax
func
ti
on
accom
m
od
at
es
outp
uts
of
ea
ch
unit
in
betwee
n
0
t
o
1,
disp
la
ye
d
in
a
K
-
dim
ensio
nal
vect
or
of
r
andom
real
nu
m
ber
s
[29
-
31]
.
The
functi
on
is
us
e
d
as
a
n
act
ivat
ion
functi
on
due
to
it
s
cat
eg
or
ic
al
pr
ob
a
bi
li
ty
distribu
ti
on
char
act
e
risti
c.
Th
e
functi
on
is
us
e
d
f
or
a
ny
num
ber
of
cl
asses
and
able
t
o
est
i
m
at
e
the
probabil
it
y
that
any
o
f
the
te
ste
d
c
la
sses
are tr
ue.
The
s
of
tm
ax
functi
on
pro
vid
e
d by
(
)
=
∑
=
1
(4)
b
1
X2
X1
W
2
a
1
1
a
3
1
a
2
1
b
2
a
1
2
a
2
2
W
1
Ŷ
Ou
tp
u
t
Hid
d
en
Inp
u
t
1
X
1
X
2
X
m
∑
f
()
W
0
W
W
2
W
1
Inp
u
ts
W
eig
h
ts
W
eig
h
ted
Su
m
Activ
atio
n
Fu
n
ctio
n
Error
W
eig
h
t Upd
ate
Ou
tp
u
t
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
A smart
meth
od fo
r s
pa
r
k
us
ing
Ne
ural N
et
work fo
r
big
data
(
M
d.
Arm
anur R
ahm
an
)
2529
Figu
re
4. Re
ct
ifie
d
li
nea
r u
nit
5.
SSNNB F
R
A
MEWO
RK
The
s
pa
rk
co
nfi
gurati
on
pa
ra
m
et
ers
are
tu
ne
d
by
the
predi
ct
ed
val
ues
fro
m
the
sel
f
-
t
un
i
ng
ap
proac
h
SSNNB,
wh
ic
h
arc
hitec
ture
is
sh
ow
n
in
Fig
ur
e
5.
SS
N
NB
consi
ders
two
i
nput
val
ues,
w
hich
are
datase
t
siz
e
and ex
ec
utio
n
t
i
m
e.
Fr
om
F
ig
ur
e
5, t
he
re a
re
sev
e
ral b
l
ocks
su
c
h
as
:
Trainin
g data i
s obtai
ne
d
f
r
om
a d
at
abase
The dat
a
has b
een
receive
d,
a
nd the m
od
el
is
g
e
ner
at
e
d by the “
M
od
el
T
ra
ining
”
b
l
oc
k
Gen
e
rated
m
od
el
h
as
bee
n
sto
red in a
f
ixe
d
l
ocati
on b
y t
he
'
Stor
e
Mo
del on
Disk
'
b
l
ock
“Pre
dicte
d
Pa
r
a
m
et
er V
al
ue”,
this
blo
c
k pro
vid
es
the
pr
e
di
ct
ed
opti
m
u
m
par
am
et
er v
al
ue
Finall
y, the
pr
e
dicte
d op
ti
m
um
v
al
ues
are
re
cei
ved
a
nd
upda
te
d
in t
he
“
Spark
Syst
em
” b
l
ock
Figure
5.
SS
N
NB ar
c
hitec
tur
e
6.
METHO
DOL
OGY
6
.1
.
Par
amet
er
sel
ection
The
sel
ect
ed
five
pa
ram
et
ers
are
sho
wn
i
n
Table
1.
The
c
olu
m
n
'
Def
ault
value'
disp
la
ys
the
de
fau
l
t
par
am
et
er
valu
es,
an
d
the
col
um
n
'
Ra
ng
e
va
lue'
disp
la
ys
the
ra
nge
of
th
e
sel
ect
ed
pa
ra
m
et
ers
in
the
sp
ar
k
m
et
ho
d
[
32
-
34]
. S
el
f
-
tu
ning is r
eq
uire
d
whe
n
processi
ng vario
us
sizes an
d
dif
fer
e
nt ty
pes
of
d
at
a to mi
ni
m
iz
e
processi
ng
ti
m
e
an
d
ac
hieve
m
axi
m
u
m
per
f
or
m
ance
f
ro
m
sp
a
r
k
[35].
T
hi
s
pap
e
r
sel
ect
ed
fi
ve
pr
e
do
m
inant
par
am
et
ers
of
the
sp
a
r
k,
base
d
on
t
he
re
vie
w
of
th
e
aut
hors
[
36]
.
Th
e
nota
ble
rea
son
i
s:
firstly
,
the
s
el
ect
ed
five
par
am
et
ers
are
co
ve
red,
includi
ng
CP
U,
m
e
m
or
y
and
dis
k
of
t
he
resou
rce
in
a
cl
us
te
r.
Se
co
ndly
,
in
sche
du
le
an
d
s
huff
li
ng
m
od
ul
es,
it
has
a
gr
eat
i
m
pact.
Third
ly
,
this
par
am
et
er
al
so
has
a
sign
ific
a
nt
i
m
pact
on
the m
ach
ine and clu
ste
r
le
ve
l
[37].
Table
1.
Def
a
ul
t par
am
et
er v
a
lue
of s
pa
rk
wi
th r
a
nge
Sp
ark Para
m
ete
rs
Sp
ark Para
m
ete
r
Ran
g
e Value
Def
au
lt Value
d
river.c
o
res
d
river c
o
res fo
r
a
driv
er
p
rocess
1
-
8
1
d
river.
m
e
m
o
r
y
d
river
m
e
m
o
r
y
f
o
r
a driv
er
p
rocess
1g
-
4
g
1
g
ex
ecut
o
r.
co
res
co
res ar
e
f
o
r
ex
ecut
o
r
p
rocess
10
-
40
1
ex
ecut
o
r.
m
e
m
o
r
y
ex
ecu
to
r
o
f
m
e
m
o
r
y
f
o
r
per
execu
to
r
p
rocess
2g
-
8
g
1
g
redu
cer.
m
ax
SizeIn
Flig
h
t
Max size of
the
m
a
p
ou
tp
u
ts
2
4
m
-
96
m
48
m
Predicted
Para
m
eter
Valu
e
Tr
ain
in
g
Data
Mod
el
Tr
ain
in
g
Test and
sav
e M
o
d
el
Sto
re
Mod
el on
Disk
Databas
e
Sp
ark
Sy
ste
m
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
11
, No
.
3
,
J
une
2021
:
25
25
-
2534
2530
6.2.
Data
c
oll
ectio
n
Trainin
g
data
has
been
colle
ct
ed
by
t
he
s
pa
rk
job,
wh
ic
h
is
com
plete
d
by
cha
ngin
g
t
he
par
am
et
er
and
val
ues
a
nd
va
rio
us
datase
t
siz
es
an
d
ty
pe
s.
Fi
nally
,
the
su
m
of
3,0
00
sam
ple
data
ha
ve
been
c
ollec
te
d
for
trai
ning
a
nd
te
sti
ng
t
he
ne
ur
al
netw
ork
m
od
el
.
F
or
the
hi
gh
accu
racy
of
the
m
od
el
,
t
h
e
norm
al
iz
a
tio
n
has
been d
one.
6.3.
Tr
aining
and
te
stin
g
Fo
r
trai
ning,
th
e
neu
r
al
network
m
od
el
has
r
andom
ly
sle
et
e
d
80%
an
d
the
rem
ai
nin
g
20%
data
have
been
us
e
d
f
or
t
est
ing
.
T
o
get
the
best
acc
ur
a
cy
fr
om
the
m
od
el
,
t
he
trai
ni
ng
cy
cl
e
has
be
en
re
peated
s
ever
a
l
tim
es.
In
trai
nin
g,
the
e
po
c
h
siz
e
has
increased
up
to
25
0,
an
d
the
m
o
del
accuracy
le
vel
was
97.
1%
and
96.7%
f
or
te
sti
ng.
It
ha
s
obse
rv
e
d
that
the
a
ccur
acy
has
be
en
incr
eased
duri
ng
trai
ning
and
te
sti
ng
wit
h
the
nu
m
ber
of
e
pochs
is
inc
rease
d.
It
is
obser
ve
d
from
Figur
e
6
that,
after
250
ep
oc
hs
,
t
he
re
is
no
si
gn
i
ficant
i
m
pr
ovem
ent i
n bo
t
h
m
od
el
a
ccur
acy
a
nd m
od
el
l
os
s.
Figure
6. Mo
de
l acc
ur
acy
a
nd lo
ss in
traini
ng
6.4.
Te
st
be
d
The
SS
NN
B
appr
oach
has
us
e
d
the
Dell
Po
we
rE
dg
e
R720
ser
ver
a
s
a
te
stbed.
The
ser
ver
is
equ
i
pp
e
d
with
I
ntel®
Xe
on
®
CP
U
E
5
-
26
50
ve
rsion
2.0
@
2.6
0
GH
z
16
-
c
or
e
proce
sso
r
a
nd
32
G
B
PC3
m
e
m
or
y.
The
op
e
rati
ng
syst
e
m
was
Ubu
ntu,
a
nd
the
ve
rsi
on
wa
s
17.10
a
nd
ha
do
op
ve
r
sion
2.8
.1
wit
h
sp
a
r
k
ver
si
on
2.2.0.
The
sel
f
-
tu
ning
ta
sk
ca
n
be
r
un
us
i
ng
a
n
in
dep
e
ndent
or
a
diff
e
re
nt
VM.
As
li
ste
d
in
T
able
2,
the
sp
ar
k
jo
b
is
run
with
five
diff
ere
nt
datas
et
s
ran
gi
ng
f
r
om
5
GB,
10
G
B,
15
GB,
20
GB
and
50
GB
,
wh
ic
h
is
colle
ct
ed
f
rom
the
Pu
m
a
B
ench
m
ark
s
uit.
In
orde
r
to
fa
ci
li
ta
te
a
fair
c
om
par
ison
wit
h
the
def
a
ult
s
yst
e
m
,
the
five
pa
ram
et
ers
are
sel
ect
ed.
D
at
aset
s
ra
ng
i
ng
f
r
om
1
GB
to 5
GB h
a
ve
bee
n
us
e
d
duri
ng
trai
ning,
and
t
he
rest of t
he data
set
s up
to
50
G
B ha
ve been
u
s
ed durin
g
t
he
e
valuati
on
proc
ess.
Table
2.
C
onsidere
d
da
ta
set
s
Sp
ark
Size of
datas
et
So
u
rce
o
f
datas
et
W
o
rd co
u
n
t
5
GB
Pu
m
a Ben
ch
m
a
rk
1
0
GB
1
5
GB
2
0
GB
5
0
GB
6.5.
Art
ific
ial
neur
al ne
twork
m
od
el
de
velo
pm
ent
In
A
NN
m
od
el
de
velo
pm
ent,
the
ML
li
brari
es
are
re
qu
ire
d,
w
hich
are
im
ported
f
ro
m
K
eras.
O
ne
of
th
e
well
-
know
n
li
br
a
ries o
f
K
eras
a
nd
b
e
hind
it
Tens
orFlo
w,
is
s
upporte
d
.
Ke
ras
f
ram
e
work
is
m
uch
easi
er
to
us
e
instea
d
of
directl
y
us
in
g
Tens
orflo
w.
I
n
so
m
e
resp
ect
s,
the
va
riables
X,
Y,
a
nd
Z
a
re
us
e
d
to
l
oa
d
an
d
store
the
trai
n
and
te
st
data.
Thu
s
,
X
an
d
Y
com
pr
ise
t
wo
trai
ni
ng
da
ta
;
execu
ti
on
tim
e
and
dataset
siz
e
ob
ta
ine
d
by
m
anu
al
pa
ram
eter
tu
ning.
Sim
il
arly
,
the
va
riable
Z
holds
t
he
siz
e
a
nd
ti
m
e
of
e
xec
utio
n
of
t
he
te
st
data.
The
te
st
dataset
,
as
well
as
the
train
,
are
fill
ed
int
o
the
syst
e
m
.
The
nece
ss
ary
hidden
la
ye
r
is
bu
il
t
from
the
base
m
od
el
.
Fu
rthe
r
m
or
e,
fu
ncti
ons
for
act
ivati
on
are
al
so
a
dded.
I
n
the
ba
se
m
od
el
,
the
dr
opout
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
A smart
meth
od fo
r s
pa
r
k
us
ing
Ne
ural N
et
work fo
r
big
data
(
M
d.
Arm
anur R
ahm
an
)
2531
functi
on
(
0.02)
is
add
e
d
to
pr
e
ven
t
ov
e
r
fitt
ing
.
It
pass
ed
the
opti
on
al
le
arn
in
g
rat
e
of
0.000
1
f
or
the
com
pilat
ion
of
the
m
od
el
,
an
d
the
desi
gn
at
ed
le
arn
i
ng
rat
e
is
0.
01.
A
fter
that,
the
opti
m
iz
er
Ad
am
a
nd
the
m
ean
sq
ua
re
d
error
(l
os
s
f
un
ct
ion
)
a
re
com
piled
with
the
base
m
od
el
.
X
and
Y
data
ar
e
then
fitt
ed
w
it
h
a
scal
e
f
un
ct
io
n. To
predict
the
accur
acy
of
Z
d
at
a,
t
he
base m
o
del
com
bin
es
batc
h
siz
e a
nd
ep
oc
h.
Th
e v
al
idit
y
and
loss
of
an
al
ysi
s
are
pr
int
ed.
The
act
iva
ti
on
functi
on
or
the
num
ber
of
ep
oc
h
or
th
e
opti
m
iz
er
m
us
t
be
change
d
if
the
accuracy
is
lower
tha
n
the
e
xp
ect
e
d
res
ult.
The
a
ccur
acy
of
96.
9%
f
or
te
sti
ng
an
d
97.
8%
for
trai
ning
data
cou
l
d
be
acco
m
pl
ished
by
util
iz
ing
25
0
epo
c
h
a
nd
ap
pro
pr
ia
te
ly
ch
ang
i
ng
the
oth
ers
-
th
e
accuracy
of
i
nc
rem
ents
in
trai
nin
g
an
d
te
sti
ng
segm
ents
w
hen
t
he
qu
a
ntit
y
of
e
po
c
hs
is
increase
d.
Fig
ur
e
6
sh
ows
that
bey
ond
25
0
ep
ochs,
accuracy
or
l
os
s
is
no
t
sub
sta
ntial
ly
i
m
pr
ov
e
d.
T
he
m
o
del
will
be
saved
f
or
ever
y
pa
ram
eter
.
It
has
five
m
od
el
s
buil
t
by
m
od
ify
in
g
th
e
Y
with
five
disti
nct
par
am
et
ers,
w
hich
is
il
lustrate
d
in
Fi
gure
7.
Figure
7. A
NN m
od
el
s to pre
dict t
he op
ti
m
i
zed
par
am
et
er (
P
for
t
he para
m
et
er)
7.
RESU
LT
S
AND A
N
ALYSIS
7.1.
SSNNB
mo
del effici
enc
y
Figure
8
re
pres
ents
the
c
om
puta
ti
on
al
tim
e
of
s
park
work
ind
e
pe
nd
e
ntly
f
or
both
def
a
ult
desig
n
a
nd
SSNNB.
For
va
rio
us
siz
es
of
input
dataset
s.
It
has
bee
n
see
n
that
the
tim
e
necessa
ry
in
e
xecu
ti
ng
s
pa
rk
j
ob
i
s
essenti
al
ly
lowe
r
with
SS
N
N
B
rather
tha
n
t
he
def
a
ult
pa
ra
m
et
er
bounda
r
y
set
ti
ng
s
f
ree
of
i
nfor
m
at
ion
siz
e
in
the sc
op
e
of
5 GB to
50 GB.
Figure
8. Com
par
is
on w
it
h S
SNNB ap
proac
h
a
nd d
e
fa
ult config
ur
at
io
n
7.2.
Ab
il
ity
of S
el
f
-
tuni
ng a
nd
e
xecu
tion time
speedup
To
assess
t
he
a
bili
ty
of
the
SS
NN
B
fr
am
ework,
a
sp
a
rk
job
has
bee
n
e
valu
at
ed
f
or
fi
ve
di
sti
nct
siz
es
of
i
nput
data
e
xt
en
ding
f
ro
m
5
GB
,
10
GB,
15
GB,
20
GB to
50
GB
i
nd
e
pende
ntly
with
bo
t
h
the
S
SNNB
an
d
the
de
fa
ult
de
sig
n.
T
he
pr
e
di
cat
ed
ideal
pa
ram
et
ers
valu
e
has
bee
n
int
rod
uced
in
Figure
9.
Re
fe
rri
ng
t
o
Figure
8,
wit
h
the
de
fau
lt
co
nfi
gurati
on,
for
dataset
siz
es
of
5,
10,
15,
20,
and
50
GB,
spa
rk
ta
kes
8.3
3,
14.8,
19.83,
25.
45,
and
52.
11
m
i
nu
te
s
se
pa
ratel
y.
No
t
withstan
ding,
the
SS
N
NB
fr
am
ewor
k
ta
kes
5.9
8,
10.35,
13.55,
17.
29,
a
nd
35.
21
m
inut
es
sepa
ratel
y.
In
Ta
ble
s
3
a
nd
4,
it
can
be
s
een
from
the
r
esult
that
t
he
S
SNNB
appr
oach
ac
hi
eved
a
n
a
ver
a
ge
30%
faster
com
par
ed
to
the
de
fau
lt
co
nf
i
gurati
on
wit
h
in
dep
e
ndent
dataset
siz
e.
8
.33
1
4
.8
1
9
.83
2
5
.45
5
2
.11
5
.98
1
0
.35
1
3
.55
1
7
.29
3
5
.21
0
20
40
60
5
10
15
20
50
Minu
te
Data
Size
W
ith
Defau
lt Co
n
f
ig
u
r
atio
n
W
ith
SSN
NB Sy
stem
Relo
ad
M
o
d
el
&
Co
rr
esp
o
n
d
Argu
m
en
t
d
ata
“
m
e
m
o
ry
_
m
o
d
el.h5
”
“
co
res_
m
o
d
el.h5
”
“
m
e
m
o
ry
_
m
o
d
el.h5
”
“
m
ax
Size
InFlig
h
t_
m
o
d
el
.h5
”
“
co
res_
m
o
d
el.h5
”
P
-
1
Value 4
P
-
2
Value 4g
P
-
3
Value 30
P
-
4
Value
6g
P
-
5
Value 80
m
Dataset
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
11
, No
.
3
,
J
une
2021
:
25
25
-
2534
2532
Figure
9. Dete
ct
ed
opti
m
u
m
values
for
sp
a
r
k param
et
ers
Table
3.
Proce
ssing t
i
m
e redu
ce f
or
dif
fer
e
nt
d
at
aset
Proces
s with
D
ef
au
lt Co
n
f
i
g
u
ration
Proces
s with
SS
N
NB syste
m
Ti
m
e
Saved
Data I
n
p
u
t
Execu
tio
n
T
i
m
e
(
Min)
Execu
tio
n
T
i
m
e
(
Min)
In Min
5
GB
8
.33
5
.98
2
.35
1
0
GB
1
4
.8
1
0
.35
4
.45
1
5
GB
1
9
.83
1
3
.55
6
.28
2
0
GB
2
5
.45
1
7
.29
8
.16
5
0
GB
5
2
.11
3
5
.21
1
6
.9
Table
4.
Pr
e
dic
te
d
opti
m
u
m
p
aram
et
er v
al
ue
u
si
ng
SS
NN
B
appr
oach
Co
n
f
i
g
u
rable Par
a
m
e
te
rs
Def
au
lt
Para
m
eter
Valu
e
W
ith
SSNNB
5
GB
W
ith
SSN
NB
1
0
GB
W
ith
SSNNB
1
5
GB
W
ith
SSNNB
2
0
GB
W
ith
SSNNB
5
0
GB
Nu
m
b
e
r
o
f
cores o
f
driv
er
p
rocess
1
3
4
6
6
8
Driver proces
s
m
e
m
o
ry size in
Gig
a
Bytes
1
g
2
g
4
g
4
g
4
g
4
g
Nu
m
b
e
r
o
f
cores o
f
execu
to
r
p
rocess
1
20
20
30
30
40
Execu
to
r
p
rocess
m
e
m
o
ry si
ze in
Gi
g
a By
tes
1
g
3
g
4
g
4
g
5
g
6
g
Maxi
m
u
m
nu
m
b
er
of
the
m
ap
to each
r
ed
u
cer
task
48
m
48
m
60
m
60
m
65
m
80
m
8.
CONCL
US
I
O
N
This
resea
rc
h
intr
oduces
a
novel
way
to
dea
l
with
the
sel
f
-
tun
in
g
a
ppr
oac
h
f
or
s
pa
rk
predo
m
inant
par
am
et
ers
to
sp
eed
u
p
t
he
ex
ecuti
on
w
hile
handlin
g
b
i
g
da
ta
,
includi
ng
t
he
dif
fer
e
nt
siz
es
of
the
datas
et
and
var
ie
ty
of
data.
Moreove
r,
est
i
m
ation
of
opti
m
u
m
par
am
e
ter
value
for
five
sel
ect
ed
par
a
m
et
ers
is
enab
le
d
by
the
ap
proac
h.
The
a
ppro
ac
h
r
ecei
ved
the
optim
u
m
value
from
th
e
neu
ral
netw
ork
m
od
el
and
update
d
it
in
the
sp
ar
k
syst
em
bef
ore
proces
sin
g.
Dell
Power
e
dg
e
R
70
ser
ve
r,
inclu
ding
fiv
e
diff
e
ren
t
dat
aset
s,
has
bee
n
us
ed
in
the
pr
ocedu
re.
T
he
pe
rfo
r
m
ance
of
SS
N
NB
is
com
pared
with
t
he
de
fau
lt
co
nf
i
gura
ti
on
,
a
nd
t
he
r
esult
s
hows
the
perf
or
m
ance
i
m
pr
ov
em
ent
is
30%
on
a
n
ave
ra
ge.
It
ha
s
al
so
been
obser
ve
d
that
the
perfor
m
ance
was
im
pr
ovin
g
w
hile
inc
re
asi
ng
t
he
dataset
siz
e.
Fu
t
ure
researc
h
will
fo
c
us
on
ho
w
to
sel
ect
a
m
or
e
appr
opriat
e
nu
m
ber
of
par
am
et
ers
an
d
us
e
be
tt
er
serv
e
r
s
to
ob
ta
in
bette
r
outc
om
es.
Me
ta
heurist
ic
s
al
gorithm
s
are to
b
e
consi
der
e
d for t
his opti
m
iz
at
ion
.
ACKN
OWLE
DGE
MENTS
This
resea
rch
is
fund
e
d
by
t
he
Mi
nistry
of
Higher
E
duca
ti
on
,
Ma
la
ysi
a,
unde
r
the
F
undam
ental
Re
search
G
rant
Schem
e
FRGS/1/2
019/ICT
02/M
MU/0
2/15.
The
a
utho
rs
a
lso
w
ould
li
ke
to
ack
nowle
dge
the
anonym
ou
s r
e
vi
ewer
s
f
or
t
heir
v
al
ua
ble c
omm
ents and insi
gh
ts
.
REFERE
NCE
S
[1]
Archa
na
,
R.
A.,
Ravi
ndra
S.
Heg
adi
,
and
T
.
N.
Manjunath,
"A
Stud
y
on
Big
Data
Privacy
Prot
ecti
on
Models
using
Data
Masking
Metho
ds,"
Int
ernati
onal
Journa
l
of
Elec
tri
cal
a
nd
Computer
E
ngine
ering
(
IJE
CE)
,
vol.
8,
no.
5,
pp.
3976
-
3983
,
2018,
doi
:
10
.
11
591/i
jece
.
v8i5
.
p
p3976
-
3983
[2]
Anagnostopoulos,
Ioa
nnis,
Sher
al
i
Zeada
l
l
y
,
an
d
Erne
sto
Expos
it
o,
"H
andl
ing
b
ig
data:
Rese
arch
challe
ng
es
a
n
d
future
d
irecti
ons
,
"
The
Journal
o
f
Superc
omputing
,
vol
.
72
,
no
.
4
,
p
p.
1494
-
1516
,
2
016.
[3]
Salkut
i,
Surend
e
r
Redd
y
,
"A
surve
y
of
b
ig
data
and
m
ac
hine
l
e
arn
ing,
"
Inte
rnat
ional
Journal
of
El
ectric
al
and
Computer
Engi
n
ee
ring (
IJE
C
E)
,
vol.
10
,
no
.
1
,
pp
.
575
-
5
80
,
2020
,
doi:
10.
11591
/i
j
ec
e
.
v10i1
.
pp575
-
580.
“
Para
m
ete
r
1
Valu
e
4”
“
Para
m
ete
r
2
Valu
e
4g”
“
Para
m
ete
r
3
Valu
e
30”
“
Para
m
ete
r
4
Valu
e
6g”
“
Para
m
ete
r
5
Valu
e
8
0
m
”
“
sp
ark.d
river.c
o
res”
= 4
“
sp
ark.d
river.
m
e
m
o
ry
”
= 4g
“
sp
ark.execu
to
r.
co
res”
= 30
“
sp
ark.execu
to
r.
m
e
m
o
r
y
”
= 6g
“
sp
ark.re
d
u
cer.m
a
x
SizeInFlig
h
t”
= 80
m
Sp
ark Sy
ste
m
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
A smart
meth
od fo
r s
pa
r
k
us
ing
Ne
ural N
et
work fo
r
big
data
(
M
d.
Arm
anur R
ahm
an
)
2533
[4]
Janka
tt
i
,
S.
,
R
ag
hav
endr
a
,
B
.
K.
,
Ragha
v
endr
a
,
S
.
,
and
Me
ena
ksh
i,
M.
,
“
Perform
a
nce
eva
lu
at
ion
o
f
Map
-
red
uc
e
ja
r
pig
hive
and
spark
with
m
ac
hi
ne
le
arn
ing
using
big
dat
a
,
”
In
te
rnational
Journal
of
El
e
ct
rica
l
and
Computer
Engi
ne
ering
(
IJ
ECE
)
,
vol
.
10
,
n
o.
4
,
pp
.
3811
-
3
818
, 2
020
,
doi
:
1
0.
11591/i
j
ece.
v1
0i4.
pp3811
-
381
8.
[5]
Hasan,
R.
A.,
A
l
ha
y
a
li
,
R
.
A.
I.
,
Za
ki,
N.
D.
,
an
d
Ali,
A.
H.,
“
An
ada
pti
v
e
cl
ust
eri
ng
and
c
la
ss
if
ic
a
ti
on
al
gor
it
h
m
for
Twit
t
er
data
strea
m
ing
in
Apa
che
Spark
,
”
T
ELKOMNIKA
Telecomm
unic
ation,
Computing,
El
e
ct
ronics
and
Control
,
vo
l.
17
,
no.
6,
pp.
3
086
-
3099,
2019
.
[6]
Nivash,
J.
P.
,
et
al
.
,
“
Anal
y
sis
on
enha
n
ci
ng
stor
m
to
eff
i
ci
en
tly
proc
ess
big
d
at
a
in
re
al
ti
m
e,”
F
if
th
In
te
rnation
a
l
Confe
renc
e
on
C
omputin
g,
Com
municat
ions a
nd
Net
work
ing
Tec
hnologi
es
(
ICCCNT
)
,
2014.
pp.
1
-
5.
[7]
Raswitha
Bandi,
and
G.
Anitha,
"
Mac
hine
L
earning
with
P
y
Spark
-
Review,"
Indone
sian
Journal
of
El
e
ct
ric
a
l
Engi
ne
ering
and
Computer
Sc
ie
n
ce
(
IJEECS)
,
vol
.
12
,
pp
.
102
-
10
6,
2018
.
[8]
Kal
y
ani
K.
,
and
Pathri
k
ar,
“
Rev
ie
w
on
apa
ch
e
s
par
k
t
ec
hnolog
y
,
”
Int
ernati
onal
Re
search
Journal
of
Engi
n
ee
rin
g
and
Technol
og
y (
IRJ
ET)
,
vol.
04
,
pp.
1386
-
1388,
Oct.
2017
.
[9]
Herodot
ou,
H
.
,
Li
m
,
H.
,
Luo
,
G
.
,
Boriso
v,
N.,
D
ong,
L
.
,
Cetin,
F.
B.
,
and
B
abu,
S.,
“
Starfi
sh:
A
Self
-
tuni
ng
S
y
s
t
em
for
Big
Da
ta Analy
tics,”
Cidr
,
v
ol.
11
,
no
.
2011
,
pp.
261
-
272
,
20
11.
[10]
Riz
a
,
L
.
S.
,
Pra
t
ama,
F.
D.,
Pian
ta
ri
,
E
.
,
and
Fashi,
M.
,
“
Genom
ic
rep
ea
ts
detec
t
i
on
using
Bo
y
er
-
Moore
al
gor
it
h
m
on
Apac
he
Spark
Strea
m
ing,
”
TEL
KOMNIKA
T
el
e
communic
atio
n,
Computing,
El
e
ct
ronics
and
Control
,
vo
l.
18
,
no.
2
,
pp
.
783
-
7
91,
2020
.
[11]
Sall
oum
,
S.,
Daut
ov,
R.
,
Chen
,
X.,
Peng,
P.
X.,
and
Huang,
J.
Z.
,
“
Big
dat
a
ana
l
y
t
i
cs
on
Apac
he
Spark,”
Inte
rnational
Jo
urnal
of
Data
Sc
ie
nc
e
and
Analyt
ic
s,
vo
l.
1
,
no.
3
-
4,
pp
.
145
-
164
,
2016.
[12]
Jonnala
gadd
a,
V
.
S.,
Srikant
h
,
P.
,
Thumat
i,
K.
,
a
nd
Nall
ama
la
,
S.
H,
“
A
rev
ie
w
st
ud
y
of
apac
he
spark
in
b
ig
da
ta
proc
essing,
”
Int
e
rnational
Journal
of
Computer
S
c
ie
n
c
e
Tr
ends
an
d
Technol
og
y
,
v
ol.
4
,
no
.
3
,
pp
.
9
3
-
98,
2016
.
[13]
Naser
and
M
.
Abd
Ulkar
ee
m
,
“
Prediction
prices
of
basra
h
li
ght
o
il
using
art
if
ic
i
al
neur
al
net
works
,
”
In
te
rnat
ional
Journal
of
Elec
t
rical
and
Computer
Eng
ine
ering
(
IJE
CE)
,
vol. 10
,
no
.
3
,
pp
.
2682
-
2689
,
2020.
[14]
Ma
hm
ood,
Maha
,
Be
la
l
Al
-
Kha
te
eb
,
and
W
isa
m
Makki
Alwash,
“
A
rev
ie
w
o
n
neur
al
n
et
w
orks
appr
oac
h
on
cl
assif
y
ing
canc
ers,
”
Int
.
J
.
Arti
f
.
Intell
.
,
vo
l. 9, n
o.
2
,
pp
.
317
-
32
6,
2020
.
[15]
Kw
ia
tkowski,
B
.
,
B
art
m
an,
and
Maz
ur,
“
The
qu
al
ity
of
da
ta
an
d
the
a
cc
ur
acy
of
ene
rg
y
g
ene
r
at
ion
for
ecast
b
y
art
if
ic
i
al
n
eur
a
l
net
works
,
”
In
te
r
nati
onal
Journal
of
E
le
c
tric
al
an
d
Computer
Eng
ine
ering
(
IJE
C
E
)
,
vol.
10
,
no.
4,
pp.
3957
-
3966
,
2020.
[16]
Anh,
Q.
H.,
Ta
n
,
P.
T.
,
and
An,
N.
T.
,
“
A
hy
br
i
d
Artifi
cial
neur
al
net
work
-
gen
e
t
ic
al
gor
it
hm
for
loa
d
shedding,
”
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Compu
t
er
Engi
n
ee
ring
(
IJE
CE)
,
vol
.
10
,
n
o.
3
,
pp
.
2250
-
2
258.
,
2020.
[17]
Goa
y
,
C.
H.,
et
a
l.
,
“
Progress
in
neur
al
n
et
work
ba
sed
te
chni
qu
es
for
sign
al
integrity
an
aly
sis
–
a
sur
ve
y
,
”
Bul
l
et
in
o
f
El
e
ct
rica
l
Eng
in
ee
ring a
nd
Infor
matic
s
(
BE
EI)
,
v
ol.
8
,
no
.
1
,
pp
.
2
76
-
282,
2019
.
[18]
Souri,
A.,
El
Maa
zouzi,
Z
.
,
Al
Achha
b
,
M.
,
and
El
Mohaji
r,
B.
E
.
,
“
Neura
l
net
wo
r
k
dea
li
ng
with
Arabi
c
la
ngu
age,
”
Inte
rnational
Jo
urnal
of
In
formatic
s and
Comm
unic
ati
on
Techno
l
ogy
(
IJ
-
ICT)
,
vol.
9
,
no
.
2
,
pp
.
73
-
78,
2020
.
[19]
Abougara
ir,
A.
J.,
“
Neura
l
N
e
t
works
Ide
nti
fica
ti
on
and
Con
tro
l
of
Mobile
Ro
bot
Us
ing
Adapti
v
e
Neuro
Fuz
z
y
Infe
ren
c
e
S
y
ste
m
,
”
Proc
ee
dings
of the
6th
In
te
rn
ati
onal
Con
fe
ren
ce
on
Eng
ine
eri
ng
&
MIS
,
2020,
pp.
1
-
9.
[20]
La
m
a,
P.,
and
Z
hou,
X.,
“
Aroma
:Automate
d
r
esourc
e
a
ll
oc
at
ion
and
conf
igurati
on
of
m
apr
educ
e
envi
ronm
ent
i
n
the
cl
oud
,”
Proc
ee
dings o
f
th
e
9
t
h
international c
onfe
renc
e
on
Au
tonomic
computing
,
2012
,
pp
.
63
-
72
.
[21]
Za
har
ia,
M.,
Konw
inski,
A
.
,
Jos
eph,
A.
D.,
Ka
t
z,
R.
H.
,
and
Stoic
a
,
I
.,
“
Im
pro
ving
MapReduce
per
form
anc
e
i
n
het
ero
g
ene
ous
e
nvironments,
”
Os
di
,
vol
.
8
,
no
.
4
,
pp.
29
-
42,
2008
.
[22]
W
u,
D.,
and
G
okhal
e
,
A.
,
“
A
self
-
tuni
ng
s
y
st
e
m
base
d
on
ap
pli
c
at
ion
profili
ng
and
p
erf
orm
anc
e
an
aly
sis
fo
r
opti
m
iz
ing
h
ado
op
m
apr
educ
e
c
l
uster
conf
igura
t
i
on,
”
20
th
Annua
l
Int
ernati
onal
Confe
renc
e
on
Hi
gh
Pe
rform
anc
e
Computing
,
201
3,
pp
.
89
-
98
.
[23]
Zha
ng,
R.
,
L
i,
M.,
and
Hild
ebr
and,
D
.
,
“
Findin
g
the
bi
g
d
at
a
s
wee
t
spot
:
Tow
ard
s
aut
om
atic
all
y
re
comm
endi
ng
conf
igurations
f
or
hadoop
c
lust
ers
on
docker
c
onta
in
ers
,”
I
EEE
Inte
rnat
ional
Confe
renc
e
on
Cloud
Engi
n
ee
ri
ng
,
2015,
pp
.
365
-
3
68
.
[24]
W
ang,
K.,
and
Khan,
M.
M.
H
,
“
Perform
anc
e
pre
diction
for
a
pac
he
spark
pl
a
tform
,”
IEEE
1
7th
Inte
rnationa
l
Confe
renc
e
on
High
Pe
rform
an
ce
Comput
ing
a
nd
Comm
unic
ations
,
2015,
pp.
1
66
-
173
.
[25]
Yigit
basi
,
N.,
W
il
lke
,
T
.
L.,
Li
ao
,
G.,
and
Epe
m
a,
D.,
“
T
owards
m
ac
hine
le
arn
ing
-
base
d
aut
o
-
tuni
ng
of
m
apr
educ
e
,
”
20
13
IEE
E
21st
In
te
rnational
S
ym
posium
on
Mode
ll
ing
,
Ana
ly
sis
and
Simulat
ion
of
Computer
and
Tele
communic
a
t
ion
Syst
ems
,
pp.
11
-
20
,
2013
.
[26]
Sahana
,
H
.
P.,
Sanja
na
,
M.
S.
,
Muddasir,
and
N.
M.,
“
Apac
he
Spark
Met
hods
and
Techni
qu
e
s
in
Big
Data
-
A
Revi
ew,”
In
ve
nt
i
ve
Comm
unic
a
tion and
Computa
t
ional
Te
chnol
og
ie
s
,
pp
.
721
-
726
,
2020.
[27]
Naz
m
ul
Haque
,
and
Md
.
Hasn
at
Ri
az.,
"A
uton
om
ous
Vehic
le
Control
S
y
st
em
as
a
Mobil
e
R
obot
b
y
Artif
ic
i
a
l
Neura
l
N
et
work
,"
Inte
rnationa
l J
ournal
of Robot
i
cs
and
Au
tomation (
IJR
A)
,
vol
.
6
,
no
.
3
,
pp
.
200
-
206,
2017
.
[28]
Shatha
A.
Ba
ke
r,
Hesham
H.
Moham
m
ed,
Hana
n
and
A.
Ald
aba
gh
,
“
Im
proving
Face
Rec
ogn
it
ion
b
y
Artifici
al
Neura
l
N
et
wor
k
Us
ing
Princ
ipa
l
Com
ponent
Anal
y
sis
,
”
T
ELKOMNIKA
Tele
communic
a
t
ion,
Comput
ing
,
El
e
ct
ronics
and
Control,
vo
l. 18, no.
6,
pp.
3357
-
3364,
2020
.
[29]
Shaik
h,
E.,
Moh
iuddi
n,
I.,
Aluf
ais
an,
Y.
,
and
Nah
vi,
I
.
,
“
Apac
he
Spark:
A
B
ig
Da
t
a
Proce
ss
ing
En
gine
,
”
2019
2nd
IEE
E
Middle
Ea
st and
North
Af
ri
ca
COMMunic
a
t
ions Conf
ere
n
ce
(
MENA
COMM)
,
2019,
pp.
1
-
6.
[30]
Al
-
Azz
awi
,
D.
S.,
“
Applicat
ion
and
evalu
at
ion
o
f
th
e
neur
al
n
et
work
in
ge
arb
ox,
”
TEL
KOMNIKA
Tele
communic
a
t
ion,
Comput
ing,
El
e
ct
ronics
and
Control,
vo
l.
18
,
no.
1,
pp.
19
-
29
,
2020
.
[31]
Abd
Rahman,
N.
H.,
and
Le
e
,
M.
H.,
“
Artificial
neur
al
n
et
w
ork
fore
c
asti
ng
per
form
anc
e
wit
h
m
issing
val
ue
imputat
ions,
”
I
A
ES
Int
ernati
onal
Journal
o
f Artifi
ci
al
Intelli
g
ence
(
IJ
-
AI)
,
vol.
9
,
n
o.
1
,
pp
.
33
-
39
,
2020
.
[32]
Bhat
tacha
r
y
a
,
A.
,
and
Bhat
nag
ar,
S.,
“
Big
dat
a
an
d
apa
che
spark
:
A
rev
ie
w,”
Int
ernati
onal
Journal
of
Eng
ine
erin
g
Re
search
&
Sci
enc
e
,
vol
.
2
,
no
.
5
,
pp
.
206
-
210
,
2
016.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
11
, No
.
3
,
J
une
2021
:
25
25
-
2534
2534
[33]
Pati
l,
N
.
S.
,
et
al
.
,
“
A
surve
y
on
gra
ph
da
ta
base
m
ana
gement
t
echni
ques
for
hug
e
unstruct
ur
ed
da
t
a,
”
Int
ernati
onal
Journal
of
Elec
t
rical
and
Computer
Eng
ine
ering
(
IJE
CE)
,
vol. 8,
no.
2
,
pp
.
1140
-
1149
,
2018
.
[34]
Nair
,
L.
R.
,
She
tty
,
S.
D.,
and
Shetty
,
S.
D.
,
“
St
rea
m
ing
bi
g
data
ana
l
y
sis
for
re
al
-
ti
m
e
sent
iment
base
d
ta
rge
t
ed
adve
rt
ising,
”
Int
ernati
onal
Jour
nal
of
El
e
ct
ri
cal
and
Computer
Engi
ne
ering
(
IJE
CE)
,
vol
.
7,
no
.
1,
pp.
402
-
407
,
2017.
[35]
Vijay
ar
ekha,
K.
,
“
Acti
va
t
ion
Fun
ct
ions,
NP
TE
L
-
El
e
ct
ron,”
Com
mun.
Eng
.
-
Patte
rn R
ec
og
n
it,
pp
.
1
-
6
,
2015
.
[36]
Md.
Arm
anur
Rahman,
J.
Hos
sen
and
Venka
t
a
seshaia
h
C
.
,
“
SM
BS
P:
A
Self
-
Tuni
ng
Approa
c
h
using
Mac
h
in
e
Le
arn
ing
to
Im
prove
Perform
ance
o
f
Spark
in
Bi
g
Data
Pro
ce
ss
i
ng,
”
7
th
Int
ernati
onal
Con
fe
ren
ce
on
Comput
er
and
Comm
unic
a
ti
on
Engi
ne
ering
,
2018
,
pp
.
274
-
279
.
[37]
Md.
Arm
anur
Rahman1,
Abid
Hos
sen,
J.
Hos
s
en,
Venka
ta
s
esh
ai
ah
C
.
,
“
Towa
r
ds
Mac
hine
L
earning
base
d
Self
-
tuni
ng
of
Hadoo
p
-
Spark
S
y
stem,
”
Indone
sian
Jo
urnal
of
Elec
trical
Engi
n
ee
ring
and
Computer
Sci
en
ce
(
IJEECS)
,
vol.
15
,
no
.
2
,
pp
.
1076
-
1
085
,
20
19.
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Md.
Armanu
r
Rah
man
rec
ei
ved
B.
Sc.
d
egr
ee
in
computer
science
and
eng
inee
ring
from
As
ia
n
Univer
sit
y
of
Banglade
sh
(AU
B)
in
2010,
Mast
ers
(
MEngSc.
)
degr
e
e
in
Big
dat
a
and
Ma
chi
ne
L
ea
rning
f
rom
the
Multi
me
dia
Univer
si
t
y
(
MM
U),
Malay
si
a
in
2019.
Now
he
i
s
per
suing
Ph.D
.
in
Faci
a
l
Expr
ession
Rec
ognition
using
Mac
hine
Le
arn
ing
at
Mul
ti
m
edi
a
Unive
rsit
y
(MM
U).
His
rese
arc
h
int
er
est
i
ncl
ude
pe
rform
ance
opti
m
iz
ation
of
big
da
ta s
y
st
em, data mining, m
ac
hine l
e
arn
ing
a
nd
image
pro
ce
s
sing.
Jakir
Hossen
is
gra
duated
in
M
ec
han
ic
a
l
Eng
in
ee
ring
from
th
e
Dhaka
Univer
sit
y
of
Engi
ne
eri
ng
an
d
Te
chnol
og
y
(1997),
Master
s
in
Comm
unic
at
ion
and
Net
work
Engi
ne
eri
ng
fro
m
Univer
siti
Putra
Malay
sia
(200
3)
and
PhD
in
S
m
art
Te
chnol
og
y
and
Robot
ic
Engi
ne
e
ring
from
Unive
rsiti
Putra
Malay
sia
(2012).
He
is
cur
ren
t
l
y
a
Se
nior
Le
c
ture
r
a
t
the
Facul
t
y
of
En
gine
er
ing
and
Te
chno
log
y
,
M
ult
imedia
Univ
e
rsit
y
,
Malay
s
ia.
His
r
e
sea
rch
int
er
ests
are
in
th
e
a
rea
o
f
Artificial
Int
el
l
ige
nc
e
(Fuzz
y
L
ogic
,
Neura
l
Ne
twork
),
Infe
r
enc
e
S
y
st
ems
,
Patt
ern
Cl
a
ss
ifi
ca
ti
on
,
Mobi
le
Robot
Navig
a
ti
on
and
Int
el
l
ige
nt
C
ontrol
.
A
z
i
z
a
Su
ltana
r
ec
e
ive
d
th
e
B.
Sc.
degr
ee
in
co
m
pute
r
scie
nc
e
and
engi
ne
eri
ng
from
Dhaka
Inte
rna
tional
Univer
sit
y
(DIU
)
in
2016.
She
is
cur
ren
tly
per
sui
ng
Ma
sters
degr
ee
in
Com
pute
r
Scie
n
ce
an
d
Engi
nee
ring
a
t
the
sam
e
unive
rsit
y
.
Her
rese
arc
h
int
er
est
inc
lud
e
per
form
anc
e
opt
imiza
ti
on
of
big
dat
a
s
y
stem,
da
ta
m
ini
ng,
m
ac
h
ine
le
arn
ing
and
ima
ge
proc
essing.
Ab
d
ullah
Al
Mam
un
has
rec
e
ive
d
B
.
Sc.
degr
ee
i
n
E
lec
tri
c
al
and
Elec
troni
c
Engi
ne
eri
ng
fro
m
Pabna
Univer
sit
y
of
Scie
n
ce
and
Te
chno
log
y
in
2018.
Now
he
is
pursuing
M.E
ng
.
Sc.
a
t
Mult
imed
ia
Univ
ersity
(
MM
U)
in
the
F
ac
ul
t
y
of
Engi
n
ee
ring
and
Te
chno
log
y
since
2019.
His
rese
arc
h
int
er
est
inc
lude
s
co
m
pute
r
vis
ion;
image
proc
essing,
sign
al
pro
ce
ss
ing, deep
l
ea
rning
and
m
ac
hine
learni
n
g.
Nor
A
z
li
na
Ab
A
z
i
z
she
is
cur
ren
tly
a
Senior
L
ec
tur
er
in
th
e
Fa
cul
t
y
of
Engi
ne
e
ring
and
Technol
og
y
at
Mult
imedia
Univer
sit
y
,
Me
l
aka
.
She
is
inter
este
d
in
the
f
ie
l
d
of
soft
computing
and
it
s
appl
i
cati
on
in
eng
ineeri
ng
proble
m
s.
More
spec
ifi
c
al
l
y
,
her
foc
us i
s in
th
e ar
ea
of
sw
arm i
nt
e
ll
ige
n
ce a
nd
nature
inspir
ed
op
tim
iz
at
ion
al
gor
ithm
.
Evaluation Warning : The document was created with Spire.PDF for Python.