TELKOM
NIKA
, Vol.11, No
.11, Novemb
er 201
3, pp. 6547
~6
555
e-ISSN: 2087
-278X
6547
Re
cei
v
ed Ap
ril 15, 2013; Revi
sed
Jun
e
19, 2013; Accepted July 2
0
,
2013
Software Aging Prediction based on Extreme Learning
Machin
e
Xiaoz
h
i Du
1
, Huimin Lu*
2
, Gang Liu
2
1
School of Softw
a
r
e En
gin
eer
i
ng, Xi’a
n Jia
o
t
ong U
n
iver
ist
y
,
Xi’
an 7
1
0
049,
Shaa
n
x
i, Chi
n
a
2
School of Softw
a
r
e En
gin
eer
i
ng, Cha
ngc
hun
Universit
y
of T
e
chn
o
lo
g
y
, Ch
angc
hu
n, 130
0
12, Jili
n, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
:
luhm.cc@
gm
ail.com
A
b
st
r
a
ct
In the researc
h
on softw
are a
g
in
g an
d rej
u
v
enati
on, on
e of
the most i
m
po
rtant questi
ons
is w
hen
to trigger th
e r
e
juv
enati
on
action. An
d it is
u
s
eful to
pr
ed
ict the syste
m
re
source
utili
z
a
t
i
on state
efficie
n
tly
for deter
mi
nin
g
the rej
u
ve
nati
on ti
me. In th
is
pap
er,
w
e
pro
pose s
o
ftw
are agi
ng
pred
ictio
n
mod
e
l
base
d
o
n
extreme learning
m
a
c
h
ine (
E
LM) fo
r a real VOD system
. First, the dat
a on the par
ameters of sy
stem
resourc
e
s and applic
ation s
e
rver ar
e c
o
llected. T
hen, the data is
preprocessed by
normali
z
a
t
i
on and
princi
pa
l co
mp
one
nt a
nalys
is
(PCA). T
h
e
n
,
ELMs ar
e
co
n
s
tructed to
mo
del
the
extract
ed
data
seri
es
of
system
atic
par
a
meters. Finally, we get
the predicted data
of system
r
e
s
ource by
com
p
uting the s
u
m of t
h
e
outputs of thes
e ELMs. Experiments
show
that the prop
os
ed softw
are
ag
ing pr
edicti
on
meth
od b
a
se
d on
w
a
velet transfo
rm a
nd E
L
M is
super
ior to th
e artifici
a
l
n
eur
al n
e
tw
ork (ANN) an
d sup
port
vector mach
in
e
(SVM) in the aspects of predi
ction
prec
isio
n and effici
ency. Based o
n
the mo
de
ls empl
oy
ed her
e, softwa
r
e
rejuv
enati
on p
o
lici
e
s can b
e
trig
g
e
red by
actual
me
asure
m
ents.
Ke
y
w
ords
:
sof
t
w
a
re aging, ex
treme l
ear
nin
g
mac
h
i
ne, pred
i
c
tion
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The soft
ware reliability and availability
are increasi
ngly being demanded in
present
softwa
r
e sy
stems [1]. Whil
e recent stu
d
ies sh
o
w
that when software appli
c
at
ion is executed
contin
uou
sly for long inte
rvals of time, some e
r
ro
r
condition
s in them are accumulated to result
in perfo
rma
n
c
e d
egradati
on or
even a
cra
s
h fail
ure
,
which is
ca
lled soft
ware
aging [2]. T
h
e
phen
omen
on
has
been
ob
serve
d
in ma
ny softwa
r
e
systems,
su
ch
as o
perating
system [3], web
serve
r
[4], SOA se
rver [5
], and so on.
Becau
s
e of
the effect of softwa
r
e a
g
in
g, the syste
m
reliability decrea
s
e
s
. To counteract software agin
g
and its related
transie
nt sof
t
ware failu
re
s, a
preventive a
nd
p
r
oa
ctive techni
que, ca
lled softw
a
r
e
rej
u
venation
,
has be
en
p
r
opo
se
d a
n
d
is
becoming p
o
pular [2]. It
involves sto
p
p
ing t
he ru
n
n
ing software occa
sionall
y
, cleaning i
t
s
internal
state
and
or it
s e
n
vironm
ent a
nd resta
r
ting
it. An extreme but
well-kn
own
example
of
softwa
r
e
rej
u
venation i
s
th
e ha
rd
wa
re
reboot [6]. In
gene
ral, the
cost of
software rejuvenatio
n is
sub
s
tantially l
o
we
r than the
cost of a syst
em
failure foll
owe
d
by a re
active re
cove
ry.
Over the re
cent years, q
uantitative studie
s
of softwar
e agin
g
a
nd rejuve
nati
on have
been ta
ken,
and ma
ny different ap
proa
che
s
h
a
ve b
een devel
ope
d and the
effects
of software
rejuven
a
tion
have bee
n st
udied. Th
ese
studie
s
can
be catego
rize
d into two
ki
nds, time
-ba
s
ed
rejuven
a
tion policy and
m
easure
-
b
a
sed
rej
u
venat
ion
policy. T
he ti
me-b
ased
rej
u
venation
pol
icy
is
cha
r
a
c
teri
zed by th
e fa
ct that the
soft
ware
i
s
p
e
ri
o
d
ically
rejuve
nated
every ti
me a
predefi
ned
time co
nsta
nt ha
s ela
p
sed
.
In these ap
proa
ch
es
,
a
certai
n failu
re
time di
stribu
tion is assu
m
e
d
and
co
ntinuo
us tim
e
Ma
rkov ch
ain
(CT
M
C) [2],
se
mi
-Markov
process [7], Ma
rkov reg
ene
rati
ve
pro
c
e
s
s (M
RGP) [8] , Ma
rkov de
ci
sion
pro
c
e
s
s (M
DP) [9] or sto
c
ha
stic Petri
net (SPN) m
odel
[10] etc is
d
e
velope
d to
comp
ute an
d
optimize
sy
stem availa
bi
lity or relate
d
measures.
Th
e
measure-b
a
sed rejuve
nati
on poli
c
y appl
ies stati
s
ti
cal
analysi
s
to the measured d
a
ta on re
sou
r
ce
availability to
predi
ct the ex
pected
time t
o
resource ex
hau
stion [11], and it pr
ovides a window o
f
time during
whi
c
h a reju
venation acti
on is adv
ise
d
. The basi
c
idea of the measu
r
e
-
ba
sed
rejuven
a
tion
policy is to m
onitor a
nd co
llect data o
n
the attribute
s
and pa
ram
e
ters,
whi
c
h a
r
e
respon
sibl
e for dete
r
minin
g
the health o
f
the running
softwa
r
e sy
st
em. Garg et a
l
. [3] propose
d
a
methodol
ogy
to dete
c
t an
d
estim
a
te the
aging
in th
e
UNIX
system
, they implem
ented
an S
N
MP
based tool to
colle
ct data,
and ad
opted
non-pa
rame
t
r
ic stati
s
tic m
e
thod to dete
c
t and e
s
tima
te
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 654
7 – 6555
6548
aging. An
drzejak
et al. [6] use
d
a
spli
n
e
-ba
s
e
d
de
scription
of the
aging
profile
s and
adopte
d
a
s
t
atis
tical test to verify its c
o
rrec
tness of th
e
SOA
P
se
rver. An
drzeja
k et
al.
[12] al
so
u
s
ed
machi
ne le
arning meth
od
s to model a
n
d
predi
ct t
he
software
agin
g
of a we
b ap
p
lication. G
r
ott
k
e
et al. [4] use
d
non-param
etric st
ati
s
tical method
s to detect and
estimate tre
nds of agin
g
, and
adopte
d
AR
model to
pre
d
ict the
agin
g
of a
We
b
server.
Hoffm
ann et
al. [13
]
gave a
pra
c
tice
guide to reso
urce fore
ca
sti
ng, they ado
pted seve
ral
method
s to
model a
nd p
r
edict the
software
aging
of a
Web serve
r
, th
ey found th
at pro
babili
stic
wra
ppe
r
(PWA) wa
s
a bet
ter metho
d
f
o
r
variable
sele
ct, and
sup
port ve
ctor
machi
ne
(SV
M
)
wa
s a
b
e
tter a
pproa
ch fo
r
re
sou
r
ce
forecastin
g. For the predi
ction of time seri
es,
artifici
al neural network
(A
NN) [14] and su
pp
ort
vector m
a
chi
ne (SVM
) [1
5] are
wid
e
l
y
adopted.
Artificial ne
u
r
al n
e
two
r
ks,
esp
e
ci
ally BP
netwo
rks, are powerful t
ools fo
r fitting nonli
nea
r time serie
s
. Howeve
r, there
are
so
me
disa
dvantag
e
s
in impleme
n
ting of artificial neu
ra
l n
e
tworks. First
l
y it’s hard to determin
e
the
para
m
eters
o
f
neuron
s a
n
d
the
netwo
rk
stru
cture. F
u
rthe
rmo
r
e, t
he trai
ning
p
r
oce
s
s of
neu
ral
netwo
rks i
s
time-con
sumi
ng and th
e converg
e
n
c
e
rate is sl
ow,
becau
se the
netwo
rks oft
e
n
settle in un
de
sira
ble lo
cal
minima of the
error
surfa
c
e
.
When
sup
p
o
rt vecto
r
ma
chin
e is u
s
e
d
for
predi
ction, it also fa
ce
s so
me disa
dvant
age
s and
cha
llenge
s, su
ch
as slo
w
lea
r
ning rate, mu
lti-
para
m
eters to be determi
ned, and so on. In order
t
o
overcome
some
chall
e
n
ges of ANN
and
SVM, extreme learni
ng m
a
chi
ne (ELM
) propo
se
d by
Huang
e et al.[16] has attracted the mo
re
and m
o
re att
ention
re
cent
ly. ELM is
a
dopted
fo
r g
eneralize sin
g
le-hi
dde
n
la
yer
feedfo
r
ward
netwo
rks (S
LFNs),
and
the
hidde
n
layer ne
ed
not
be tu
ned,
whi
c
h
results in
b
e
tter
gene
rali
zatio
n
perfo
rman
ce, faster lea
r
n
i
ng sp
eed, an
d least hum
a
n
intera
ction.
In this
pap
er,
we
an
alyze t
he
softwa
r
e
aging
phe
no
menon
of a
real VO
D
syst
em, and
prop
ose a
software
agin
g
predictio
n
model b
a
sed
on extrem
e
learning m
a
chin
e. The
main
contri
bution
s
of this pa
per
are
1) p
r
op
osing a
softwa
r
e agi
ng
predi
ction
mod
e
l b
a
se
d o
n
ELM
,
2)
applying P
C
A to redu
ce t
he dime
nsi
o
n
of input vari
able
s
of ELM
s
, 3)
usin
g th
e data
colle
ct
ed
from a re
al VOD sy
stem to
evaluate t
he softwa
r
e agi
n
g
predi
ction p
e
rform
a
n
c
e.
2. Soft
w
a
r
e
Aging Predi
c
tion Model
based on EL
M
In
ord
e
r
to
p
r
ovide su
ppo
rt
for
tri
gge
ri
ng softwa
r
e rejuven
a
tion action
s, we need
to
predi
ct the
sy
stem resource utiliz
ation
st
ate preci
s
ely
to reflect
the
software
syst
em state i
n
the
future.
Extreme lea
r
nin
g
machi
ne, whi
c
h i
s
a le
arni
ng alg
o
rithm,
pro
p
o
s
ed
by Hua
ng et al.
[16]
wa
s d
e
velop
ed fo
r
singl
e
-
hidd
en
layer feedfo
r
ward
neu
ral
net
works (SL
F
Ns). ELM p
r
ovid
es
good
gen
eral
ization
pe
rformance at ext
r
emely fa
st l
earni
ng
spe
e
d
by choo
sin
g
hidd
en n
o
d
e
s
rand
omly an
d determinin
g
the out
put
weig
hts of S
F
LNs an
alytica
lly. The
r
ef
ore, we p
r
esent a
softwa
r
e agin
g
predi
ction model based on
extrem
e
l
earni
ng m
a
ch
ine, sh
own in
Figure 1,
wh
ich
illustrates the
k
-step p
r
edi
cti
on pro
c
e
d
u
r
e
.
Figure 1. Software Aging P
r
edi
ction Mo
d
e
l
From Fi
gure
1, we
see th
at the data a
r
e firs
tly prep
rocesse
d
afte
r they are co
llected.
Then th
e time se
rie
s
of target pa
ramete
r
()
yt
is inputted i
n
to wavel
e
t transfo
rm m
o
d
u
le, and th
e
detail co
mpo
nents
()
,
1
,
.
.
.
,
i
D
ti
p
and th
e app
roximat
i
on compo
n
e
n
t
p
A(
)
t
of
()
yt
are g
o
tten,
whe
r
e
p
is the decompo
si
tion level assign
ed by u
s
er,
t
is the sampl
e
inde
x. The othe
r
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Software Agi
ng Predi
ction
based on Ext
r
em
e Learnin
g
Machi
ne (Xi
aozhi Du
)
6549
para
m
eters
(
)
,
1
,
2
,
.
..,
i
x
ti
n
are inp
u
tted into PCA module, a
nd the first
m
princi
pal
s
(
)
,
1
,
2
,
...,
i
zt
i
m
are
sel
e
cte
d
, whe
r
e
mn
.
Then we co
nstru
c
t
1
p
ELMs
to forec
a
s
t
the
decompo
se
d
com
pone
nts of target
param
ete
r
sep
a
rat
e
ly. Each
com
pone
nt (
p
A(
)
t
or
()
,
1
,
.
.
.
,
i
D
ti
p
) and all the first
m
prin
cipal
s
(
)
,
1
,
2
,
.
..,
i
zt
i
m
are the inpu
ts of the resp
ective ELM,
and th
e out
put of the E
L
M is the
k
-step predi
ctio
n value
of the
comp
one
nt (
p
A*
(
)
tk
or
*
(
)
,
1
,
..
.,
i
D
tk
i
p
). Finally, all
the outputs o
f
these
1
p
ELMs are
summ
e
d
to obtain the
k
-
step predi
ctio
n value of the object pa
ram
e
ter
*(
)
yt
k
.
The detail ste
p
s of ou
r app
roa
c
h a
r
e a
s
follows:
1) D
a
ta
Pr
epr
o
cess
Data preprocess inclu
d
e
s
two pha
se
s: (1) Pa
ramete
r redu
ction a
nd sele
ction.
In this
pha
se, the para
m
eters
whi
c
h are consta
nt va
lues du
ring th
e monitori
ng
period a
n
d
the
para
m
eters t
hat have
the
same
me
anin
g
s
are
excl
ud
ed. (2)
No
rm
alizatio
n. In t
h
is
pha
se,
all
the
sele
cted
pa
ra
meters a
r
e
n
o
rmali
z
e
d
to
eliminate
dim
ensi
on influ
e
n
ce. After no
rmali
z
ation, t
h
e
rang
e of all these p
a
ra
mete
rs a
r
e limited
into [-1, 1].
2)
Princi
pal Co
mpone
nt Ana
l
ysis
In the experi
m
ents, we collect 30
pa
ramete
rs of m
e
mory, 15 p
a
r
amete
r
s
of CPU, 17
para
m
eters
o
f
disk a
nd
2
7
pa
ram
e
ters of ap
p
licatio
n serve
r
. So
me pa
ram
e
te
rs
are
con
s
ta
n
t
durin
g the
ob
servatio
n inte
rval, su
ch
a
s
Com
m
it
Limi
t of memory, C2
Transitions/sec of CP
U
and
so o
n
. Some p
a
ram
e
ters have th
e
same
mea
n
in
g, su
ch a
s
A
v
ailable Kbytes a
nd Availa
ble
Mbytes of m
e
mory etc. After all these para
m
eters a
r
e excl
uded,
there
a
r
e still
32 paramete
r
s
left.
The relat
i
onship between software
aging
and
these param
e
ters
can
be
expressed as the
Equation (1):
12
(,
,
,
)
n
yf
x
x
x
(1)
Whe
r
e
y
denotes the availa
ble bytes of memory an
d
12
,,
,
n
x
xx
are the impa
ct factors
of softwa
r
e a
g
ing in the VOD sy
stem, a
nd here
n
equ
als 32.
If we take all
these 3
2
pa
rameters a
s
the i
nputs of th
e ELMs di
re
ctly,
the netwo
rk
scale
is very large,
and its effici
ency i
s
very
low. PC
A [1
7] is an e
ssential metho
d
of multivari
a
te
statistical ana
lysis, whi
c
h
select
s seve
ral
repres
entative prin
ciple
co
mpone
nt
s to
explain mo
st of
the data ch
a
nge
s. Therefore, in
orde
r to improve th
e efficien
cy
and to kee
p
the accu
racy of
our
predi
ction m
o
del, PCA is a
dopted to re
d
u
ce the in
put para
m
eters h
e
re.
Let the
sampl
e
s
of the
s
e
fa
ctors
be
12
(,
,
)
T
n
XX
X
X
, the
n
the
proced
ure
of P
C
A i
s
as
follows
:
(1)
Normali
z
e
the sample
s
X
to remove di
mensi
on influ
ence.
(2) Calculate
the
relative
matrix
P
and
covarian
ce
ma
trix
C
of
the
sample data, and
eigenvalu
e
s
12
,,
n
and eig
enve
c
tors a
r
e obtai
ned.
(3)
Cal
c
ulate
the contrib
u
t
ion rate of eac
h com
p
o
nent re
spe
c
ti
vely. The choice of
prin
cipal
co
mpone
nt is determi
ned
based on varian
ce
cont
ribution rate
and cumul
a
tive
contri
bution
rate. The varia
n
ce
contri
buti
on
rate
s are calcul
ated by the Equation
(2):
1
/(
)
(
1
,
2
,
,
)
n
kk
i
i
kn
(2
)
And the cum
u
lative cont
ri
bution rate for ea
ch
prin
ciple
com
pon
ent is gotten
by the
Equation (3):
11
/(
)
(
1
,
2
,
,
)
mn
mj
i
ji
mn
(3)
The high
er of
the
1
means t
he
stronger of the ability fo
r the first principal
component to
abstract the
informatio
n o
f
12
,,
,
n
x
xx
. If the accumul
a
tion
contributio
n ra
te of the first
m
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 654
7 – 6555
6550
comp
one
nts
is mo
re tha
n
a p
r
edete
r
mined th
re
shold (su
c
h
as 8
5
pe
rce
n
t), the first
m
comp
one
nts
are sele
cted
as the inp
u
ts
of the ELM.
After PCA is finish
ed, form
ula (1
) ca
n be
redu
ced to th
e Equation (4
):
2
(,
,
)
im
yf
z
z
z
(4)
whe
r
e
y
deno
tes the avail
able bytes o
f
memory an
d
12
,,
,
m
zz
z
is the pri
n
cip
a
l
comp
one
nts
of aging impa
ct factor
s of the VOD
syst
em, and he
re
mn
.
3)
Extreme Lea
rning Ma
chin
e
s
Extreme learning ma
chin
e
,
which i
s
a l
earni
ng
alg
o
ri
thm, propo
se
d by Huan
g e
t
al. [16]
wa
s devel
op
ed for
singl
e-hidde
n layer
feed forw
a
r
d
neural net
works
(SLF
Ns). ELM provid
es
good
gen
eral
ization
pe
rformance at ext
r
emely fa
st l
earni
ng
spe
e
d
by choo
sin
g
hidd
en n
o
d
e
s
rand
omly and
determinin
g
the output
we
ights of SFLNs an
alyticall
y
. Here we consi
der a si
n
g
le-
hidde
n layer feed forwa
r
d network
wi
th L hid
den
neuron
s. The
input
12
(
,
,
...
,
,
)
m
X
zz
z
y
is a
vector with m
+
1 elem
ent
s, the output of the
th
i
hidden ne
uron i
s
(,
,
)
ii
Ga
b
X
, where
i
b
is the bias,
and
12
(,
,
.
.
.
,
,
)
ii
i
i
m
i
y
aa
a
a
a
is th
e we
ight vecto
r
,
(
1
,
2
,
...
,
,
)
is
as
m
y
is the
conn
ect
i
on
weight
be
tween
the
th
i
hidde
n neuron and
the
th
s
input
ne
uron.
The
n
t
he o
u
tput
of
the SLF
N
i
s
given by th
e
Equation (5):
1
()
(
)
(
,
,
)
L
ii
i
i
yt
k
f
X
G
a
b
X
(5)
Whe
r
e,
1
(
,
.
.
.
,,)
'
ii
i
n
i
y
is the weight vect
or co
nne
cting
hidden layer
with output la
yer,
ik
is
th
e con
n
e
c
tion wei
ght betwe
en
th
e
th
i
hidd
en
neu
ron a
nd th
e
th
k
o
u
tput ne
uro
n
. For the
ca
se of additi
ve hidden n
e
u
ron
s
,
(,
,
)
ii
Ga
b
X
take
s
the followin
g
form sho
w
n b
y
the Equation (6):
(,
,
)
(
'
)
ii
i
i
Ga
b
X
g
a
X
b
(6)
Whe
r
e
:
gR
R
is the activation function.
A
ssu
me t
hat
N
arbitrary sa
mples
(,
)
mn
ii
XY
R
R
are
gi
ven, the weig
ht vectors
i
a
and
bias
i
b
are ra
n
domly assign
ed. Then the
SLFN wi
th L
hidde
n neu
ro
ns can ap
pro
x
imate the
N
sampl
e
s
with zero erro
r if and only if there exists
i
, so that we get
j
Y
by Equation (7
):
1
(,
,
)
,
1
,
2
,
.
.
.
,
L
ji
i
i
j
i
YG
a
b
X
j
N
(7)
The above
N
equatio
ns ca
n be
rewritten in th
e follo
wing
compa
c
t form
sh
own by
Equation (8):
H
Y
(8)
whe
r
e
11
(
,
,)
(
,
,)
(
,
,)
(
,
,)
ii
L
L
ii
N
L
L
N
NL
Ga
b
X
G
a
b
X
H
Ga
b
X
G
a
b
X
,
1
T
T
L
L
m
,
1
T
T
N
NM
Y
Y
Y
.
Once the hid
den nod
e pa
ramete
rs
(,
)
ii
ab
are gene
rated
rand
omly, they remain fixed.
Then trai
ning
the SLFN is equivalent to findi
ng the minimum no
rm least-sq
ua
res
solutio
n
*
,
whi
c
h is give
n by the Equation (9
):
*
H
Y
(9)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Software Agi
ng Predi
ction
based on Ext
r
em
e Learnin
g
Machi
ne (Xi
aozhi Du
)
6551
whe
r
e
H
is the Moore–Pen
ro
se ge
nera
lized inverse of matrix
H
.
For the trai
ning data set
1
,
2
,..
.,
{(
,
)
}
ii
i
N
XY
, the ELM algo
rithm [16] is describ
ed a
s
follows:
Step 1: Assig
n
the hidde
n node n
u
mb
e
r
L, and the activation function
g(.)
;
Step 2: For
1
,
2
,
...
,
iL
, randomly ge
n
e
rate the inp
u
t
weight vecto
r
i
a
and the bia
s
i
b
Step 3: Calcu
l
ate the hidde
n layer output
matrix
H
and
H
;
Step 4: Acco
rding Equatio
n
(9), ca
l
c
ul
ate
the output weight vector
*
.
For
any in
pu
t sam
p
le
n
x
R
, the outp
u
t valu
e
*
y
is calculate
d
by u
s
in
g t
he Equ
a
tion
(10
)
:
**
1
()
L
ii
i
i
yg
a
x
b
(10)
Then, we get
the
k
-step
pre
d
iction valu
e of the obje
c
t para
m
eter, a
nd the proce
dure i
s
finis
h
ed.
3. Results a
nd Analy
s
is
3.1. Experimental Setup
The exp
e
rim
ental environ
ment is
a re
al
VOD
syst
em, and it
s
stru
cture is shown in
Figure 2. The
VOD system
con
s
ist
s
of a web se
rver,
an appli
c
atio
n serve
r
(o
r video se
rver)
and
a disk array. The web
se
rver a
c
ts a
s
the pr
e
s
entat
ion layer, wh
ich provide
s
media data t
o
client
s. Wh
en
it receives th
e or
der re
que
st from
a cli
e
nt, the we
b
s
e
rv
er
re
dire
ct
s this
re
que
st
to
the appli
c
ati
on serve
r
. T
hen the
cli
e
nt makes
co
nne
ction
with the a
ppli
c
ation serve
r
and
receives the
media
data
from the
ap
plication
serve
r
d
i
rectly if th
e
cl
ient get
s the
permi
ssion.
T
h
e
disk a
rray i
s
use
d
to sto
r
e
the medi
a d
a
ta. In our
e
x
perime
n
ts,
we a
dopt
Apac
he
a
s
the
web
serve
r
, and
Helix server
as the application se
rver.
Figure 2. Structure of the V
O
D System
3.2. Data
Col
l
ections
Durin
g
the e
x
perime
n
ts, we colle
ct 30
par
am
eters
of memory, 1
5
paramete
r
s of CPU,
17 paramete
r
s of disk an
d 27 parame
t
ers of appl
i
c
ation
serve
r
, the sampli
ng interval is 3
minutes, 75
0
0
sampl
e
s of
the system
p
a
ram
e
ters are colle
cted fo
r 375 ho
urs.
The nu
mbe
r
of client a
c
ce
ss
of the VO
D sy
st
em is il
lustrate
d in F
i
gure
3, and t
he time
seri
es of
syst
em available
memory is
sh
own in Fig
u
re
4.
From
Figu
re
3, it is foun
d
that the num
ber
of
clie
nt
ac
ce
ss
sh
ow
s
a ce
rtain
p
e
riodi
city,
however it ha
s a la
rge
ran
domne
ss an
d
it fluctuat
es f
r
equ
ently. From Figu
re 4,
we
see th
at the
available m
e
mory shows
a frequ
ent a
nd large fl
uct
uation, whi
c
h
is ca
used b
y
the stoch
a
s
tic
arrival
of clie
nt acce
ss an
d the
softwa
r
e agin
g
.
Accordin
g to the
time se
rie
s
of the collect
ed
system
availa
ble mem
o
ry a
nd the n
u
mb
er of
c
lient a
c
ce
ss,
we u
s
e
Mann
-Kend
a
ll method [4] t
o
test whethe
r t
here
i
s
software
agin
g
phe
nomen
on
i
n
t
he VO
D
syst
em. Tabl
e 1
shows th
e
re
sults
of trend te
st for the
availa
ble mem
o
ry
and the
num
ber
of client acce
ss.
F
r
om
Table
1, we find
that there is a
down
w
a
r
d trend in the time seri
es
of th
e available m
e
mory, and th
e time serie
s
o
f
the num
ber
of client
access al
so
ha
s a do
wn
wa
rd
trend. T
hat i
s
, the d
o
wn
ward tre
nd
o
f
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 654
7 – 6555
6552
system
availa
ble me
mory i
s
n
o
t cau
s
ed
by the
cli
ent
s. Therefore,
it ca
n b
e
con
c
l
uded
that the
r
e
exists softwa
r
e aging p
hen
omeno
n in the VOD sy
ste
m
.
Figure 3. Nu
mber of Cli
e
n
t
Acce
ss
Figure 4. System Available
Memory
Table 1. Tren
d Test
data Zstatistic
comment
available memory
-41.1374
Do
w
n
w
a
rd tre
nd
detected
number of client
access
-30.7877
Do
w
n
w
a
rd tre
nd
detected
3.3. Soft
w
a
r
e
Aging Prediction
In orde
r to e
v
aluate the p
e
rform
a
n
c
e o
f
softwa
r
e a
g
i
ng predi
ction
,
root mea
n
squ
a
re
error (RMSE) is adopte
d
a
s
indi
cator. RMSE is t
he square ro
ot of the vari
an
ce
of the resid
u
a
ls,
and it can b
e
interpreted a
s
the stan
da
rd deviation of
the unexplai
ned varia
n
ce. The lower th
e
values of RM
SE, the better the predi
ctio
n re
sult. RMS
E
is defined b
y
the Equation (11
)
:\
2
1
ˆ
((
)
(
)
)
N
i
yi
yi
RM
SE
N
(
1
1
)
Whe
r
e,
()
yi
deno
tes the
actu
al
value of the t
i
me serie
s
of
the availabl
e
memory,
ˆ
()
yi
is
the respe
c
tive predi
ction v
a
lue, and
N
is the point
s of data set.
In the experi
m
ents, the nu
mber of pri
n
ci
pal com
pon
e
n
ts
m
is
s
e
t to 2. For the ELM, we
use
a
sig
m
oi
d a
c
tivation f
unctio
n
a
nd
the nu
mbe
r
of hidd
en
ne
uron
s is 47.
The first
450
0
sampl
e
s a
r
e
ado
pted to
train th
ese E
L
M, an
d the
re
sidu
al 3
0
00
sampl
e
s
are
u
s
ed
to
test
wheth
e
r ou
r
method
is effective. Fig
u
re
5
sh
ows th
e
one
-step fo
rward
pre
d
icti
on valu
e
of the
available me
mory.
(a) Predi
ction
data
(b) Pr
edi
ction
erro
r
Figure 5. Pre
d
iction
Re
sult
s of System Available Me
mory
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Software Agi
ng Predi
ction
based on Ext
r
em
e Learnin
g
Machi
ne (Xi
aozhi Du
)
6553
From Fi
gure
5, we see tha
t
the erro
r b
e
twee
n the p
r
e
d
iction d
a
ta a
nd the a
c
tual
data of
the availa
ble
memo
ry re
sou
r
ce i
s
ve
ry low.
The
n
we
con
c
lud
e
that the
p
r
edi
ction
mo
del
prop
osed by
us is
suitabl
e
for softwa
r
e
aging fo
reca
sting. We al
so
find that the predi
ction e
r
ror
tends to be large
r
at the valley of the reso
urce
con
s
umption. The
reason is th
at there are
more
client
s in the system at the
valley of the
reso
ur
ce
co
nsum
ption, which results i
n
more me
m
o
ry
resou
r
ce co
n
s
umptio
n and
more fluctu
ati
on, so the predictio
n pre
c
i
s
e de
crea
se
s.
Table 2 sho
w
s the a
ppro
x
imation perf
o
rma
n
ce
of our software a
g
ing predi
ctio
n model
comp
ared wit
h
sup
port ve
ctor ma
chi
n
e
(SVM),
artificial neu
ral n
e
twork mo
del
(ANN) with
BP
algorith
m
. The simulatio
n
s for our mod
e
l
, ANN are ca
rrie
d
out in MATLAB R200
7a enviro
n
me
nt
runni
ng in a Core2 Duo CPU, 3GHz. The simul
a
ti
on for SVM is carri
ed out
by using the LIBSVM
[18] impleme
n
ted in C
co
d
e
run
n
ing i
n
the sa
me PC.
The nu
mbe
r
of hidden
ne
uron
s of A
N
N is
set 42, an
d the ke
rn
el fun
c
tion u
s
ed i
n
SVM is radi
al
basi
s
fun
c
tio
n
. In our exp
e
rime
nts, all the
inputs
and th
e output
s ha
ve been
normalize
d
into [
-
1, 1]. 20 trial
s
have
been
con
d
u
c
ted for all
the method
s and the aver
a
ge re
sults a
r
e
adopted.
Table 2. Pred
iction Results of Various M
odel
s
Model
Training data
Testing data
RMSE Time
(s) RMSE
Time(s)
Our
m
e
thod
0.0033
0.084
0.0406
0.021
SVM
0.0186
3.797
0.0445
3.734
ANN
0.0214
139.629
0.0698
0.0391
From Ta
ble 2
,
it can be se
en that the predict
io
n pre
c
i
s
ion of ou
r m
e
thod is
sup
e
rior to
that of ANN
and SVM. T
h
e traini
ng tim
e
an
d the te
sting time of
o
u
r m
e
thod
are far l
o
wer th
an
that of SVM, and the t
r
aini
ng time of o
u
r
mod
e
l is
fa
r lowe
r than
that of ANN. For the t
r
aini
ng
data, the
RM
SE of our
m
e
thod i
s
0.0
0
33, and
fo
r t
he testin
g d
a
t
a, the RMS
E
is 0.04
06.
th
e
training time
and the testi
ng time of our me
thod are
0.084 se
con
d
s and 0.0
2
1
seco
nd
s, wh
ich
sho
w
that th
e efficien
cy o
f
our m
e
thod
is a ve
ry hig
h
. Therefore, the metho
d
we p
r
e
s
e
n
ted
is
effective to forecast the
sof
t
ware a
g
ing p
r
ocess.
3.4. Sensitivi
t
y
Anal
y
s
is
We p
e
rfo
r
m
sen
s
itivity analysis fo
r ou
r pr
esented
model that p
r
edi
cts the
a
v
ailable
memory of the VOD sy
stem by addin
g
a prin
ci
pal
compo
nent
at a time, a
nd cal
c
ulate
the
model’
s
ch
an
ge in RMSE. The re
sult is
sho
w
n in Fig
u
re 6.
Figure 6. RM
SE versu
s
the Numb
er of
Princi
pal Co
mpone
nts
From Fi
gure 6, it’s sho
w
n
that with the i
n
cr
ea
se of th
e numb
e
r of
prin
cipal
com
pone
nts,
the RMSE de
cre
a
ses,
whe
n
the numb
e
r of princi
pal
compon
ents a
rrive
s to a ce
rtain value, th
e
minimum
RMSE is gott
en. Then
wit
h
the contin
uou
s in
cre
a
se of the nu
mber
of prin
cipal
comp
one
nts,
the RMSE i
n
crea
se
s. Th
e rea
s
o
n
is
that with the
increa
se of
the numb
e
r of
prin
cipal
com
pone
nt, the
more
info
rma
t
ion of in
put
variable
s
i
s
i
n
clu
ded,
but
except th
e first
several
prin
ci
pal
com
pone
nts, the
re
sid
ual
com
pone
nts h
a
ve trivi
a
l info
rmation
,
and
with
m
o
re
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 654
7 – 6555
6554
prin
cipal
com
pone
nts, the ELMs are be
comin
g
more
complex, wh
ich re
sult
s in the increa
se
of
RMSE. The
r
e
f
ore, we
sho
u
ld
sele
ct the
app
rop
r
iate
numbe
r
of pri
n
cip
a
l
comp
o
nents ba
se
d
on
requi
rem
ent.
4. Conclusio
n
In this pa
pe
r, we have i
n
vestigated th
e software a
g
ing p
hen
om
enon
of a re
al VOD
system, an
d
have pro
p
o
s
ed a
software a
g
ing
p
r
edictio
n mod
e
l base on
extreme lea
r
ning
machi
ne. T
h
e expe
riment
al re
sult
s h
a
v
e sh
owe
n
t
hat the p
r
o
p
o
se
d
softwa
r
e agin
g
p
r
e
d
iction
model is effe
ctive to forecast the agin
g
pr
og
re
ss, and the PCA
is an import
ant and useful
method to
re
duce the
red
unda
ncy of d
a
ta. Com
pare
d
with SVM a
nd ANN p
r
edi
ction m
odel,
our
model is mo
re effective on preci
s
io
n tha
n
ANN an
d SVM. And the time for trainin
g
and testing
of
our mo
del is f
a
r lower tha
n
that of ANN a
nd SVM.
ELM is a
qui
ck
and
efficie
n
t method fo
r re
solvin
g th
e pre
d
ictio
n
probl
em, but
how to
determi
ne th
e nu
mbe
r
of
hidde
n n
euro
n
s i
s
an
op
en
issue,
thou
g
h
several met
hod
s h
a
ve b
e
e
n
pre
s
ente
d
, it is still a
chall
enge fo
r the
ELM. Theref
ore, in the fu
ture we will
study this issu
e.
Thoug
h in thi
s
pap
er
we h
a
ve study the
softw
a
r
e agi
ng predictio
n
probl
em, the cau
s
e
re
sulting
in software aging is
still pending, so
nex
t we will expl
ore this probl
e
m.
Ackn
o
w
l
e
dg
ements
This
work wa
s sup
p
o
r
ted by
the Nation
al
Natural
Science F
ound
ation of
Chi
n
a un
de
r
Grant
No. 6
0933
003, the
Nation
al Na
tural Sci
e
n
c
e
Foun
dation
of Chin
a u
nder Grant
No.
6124
0029,
the
Nation
al
Postd
o
cto
r
al Scie
nc
e Found
ation of
Chi
na u
nder
G
r
ant
No.
2011M
500
61
1, the Indust
r
ial Te
chn
o
lo
gy Re
sea
r
ch
and Develo
pment Spe
c
i
a
l Proje
c
t of Jilin
Province un
der G
r
ant No. 201100
6-9; the Fund
amental Research Fun
d
s for the Central
Universitie
s
; the Natio
nal
Colle
ge Stud
ents' Inn
o
vative Trainin
g
Prog
ram of Ch
ina und
er G
r
ant
No. 201
210
1
9003
7.
Referen
ces
[1]
Pauls
on LD. C
o
mputer S
y
ste
m
, Heal T
h
y
s
el
f.
IEEE Computer. 2002; 3
5
(
8
): 20-22.
[2]
Hua
ng Y, Kintala C, Kol
e
ttis N, F
u
lton D.
Softw
are Rejuve
natio
n: Analysi
s, Module a
nd
Appl
icatio
ns
.
Procee
din
g
s of
the 25th Symp
. On F
ault
T
o
lerant Comp
utin
g. Pasade
na,
USA. 1995; 3
8
1
-39
0
.
[3]
Garg S, Puliafito A, T
e
lek M, T
r
ivedi KS.
A Method
ol
ogy
for Detectio
n a
nd Esti
mati
on
of Softw
ar
e
Agin
g. Procee
din
g
s of the Intl. Symp.
On Soft
w
a
r
e
Re
lia
bil
i
ty En
gin
eeri
ng.
NJ, USA. 1998
; 283-29
2.
[4]
Grottke M, Li
L, Vai
d
ya
nath
an K, T
r
ivedi
KS.
Anal
ys
is
of Soft
w
a
re
A
g
in
g i
n
a W
e
b
Server.
IEEE
Transactio
n
s o
n
Reli
ab
ility
. 20
06; 55(3): 4
11-
420.
[5]
Silva
L, Mad
e
ir
a H, Silv
a JG.
Softw
are Agin
g
and
Rej
u
ve
nat
ion
in a
Soa
p
-
base
d
Serv
er.
Procee
din
g
s
of the Fifth IEEE International S
y
mpos
ium
on Net
w
ork Computing and A
pplic
ations. Massachus
etts,
USA. 2006; 5
6
-
65.
[6]
Alons
o J, Matias R, Vicente
E, et al. A Compar
ativ
e Exp
e
riment
al Stud
y of Soft
w
a
r
e
Rejuv
e
n
a
tio
n
Overhea
d.
Performanc
e Evalu
a
tion.
20
13; 70
: 231-25
0.
[7]
Bao Y, Sun X,
T
r
ivedi KS. A W
o
rkload-b
a
s
ed
Ana
l
ysis of
Soft
w
a
re Ag
in
g and Re
juv
e
n
a
tion.
IEEE
Transactio
n
s o
n
Reli
ab
ility.
20
05; 54(3): 5
41-
548.
[8]
Garg S, Pu
li
afito A, T
e
lek
M, et al. A
n
a
l
ysis
of Pr
eve
n
tive M
a
inte
na
nce
in T
r
ansa
c
tions B
a
se
d
Soft
w
a
re S
y
stems.
IEEE
Trans. on Com
puters
. 1998; 47(
1): 96-10
7.
[9]
Okamura H, D
ohi T
.
D
y
n
a
mi
c Soft
w
a
re R
e
j
u
ven
a
tio
n
Pol
i
c
ies i
n
a T
r
ansaction-
base
d
S
y
stem
und
e
r
Markovia
n Arri
val Process
e
s.
Performanc
e Evalu
a
tion
. 2
013
; 70: 197-2
11.
[10]
W
ang D,
Xie
W
,
T
r
ived KS. Pe
rformabi
lit
y Anal
ysis
of Cl
uster
ed S
y
ste
m
s
w
i
t
h
Re
juv
enati
on u
n
d
e
r
Var
y
in
g W
o
rklo
ad.
Perfor
man
c
e Evalu
a
tio
n
.
200
7; 247-
265.
[11]
Vaid
ya
nat
han
K,
T
r
ivedi KS. A Compre
he
ns
ive Mod
e
l for
Soft
w
a
re R
e
ju
venati
on.
IEEE Transactions
on De
pe
nda
bl
e
and Secur
e
C
o
mputi
ng.
20
0
5
; 124-1
37.
[12]
Andrze
jak A, S
ilva
L.
Usi
ng M
a
chi
ne
Le
arni
n
g
for N
on-i
n
tru
s
ive Mo
de
lin
g
and
Pred
ictio
n
of Softw
are
Aging.
Proceedings of the IEEE Net
w
or
k Operations and M
anagement S
y
mposium. Salvador, Brazil.
200
8; 25-3
2
.
[13]
Hoffmann GA,
T
r
ivedi KS, Malek M. A Best
Practi
ce Gui
d
e to Res
ourc
e
F
o
recastin
g for Com
putin
g
S
y
stems.
IEEE Transactions
on Reliability
. 2007; 56: 6
15-6
28.
[14]
Goh CK, T
eoh
EJ, T
an KC. H
y
brid
Multi
obj
ective
Evo
l
uti
o
nar
y D
e
sig
n
fo
r Artificial
Ne
ural N
e
t
w
orks
.
IEEE Transactions on Neur
al
Networks
. 200
8; 19: 153
1-15
48.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Software Agi
ng Predi
ction
based on Ext
r
em
e Learnin
g
Machi
ne (Xi
aozhi Du
)
6555
[15]
Cao
LJ, T
a
y
F
E
H. Support
Vect
or Mach
i
ne
w
i
t
h
Ad
apt
ive Par
a
meter
s
in F
i
n
anci
a
l
T
i
me Series
Forecasting.
IEEE Transactions on Neural Networks
. 2003; 14(6): 15
06-
15
18.
[16]
Hua
ng GB, Z
h
u QY, Sie
w
CK
. Extreme
Lear
nin
g
Machi
ne: T
heor
y
an
d Ap
plicati
ons.
N
euro
c
om
pu
ti
n
g
.
200
6; 70(1-
3): 489-
501.
[17]
W
ang
X, Krug
er U, Ir
w
i
n GW
, McCullo
ug
h G, McDo
w
e
l N.
Nonl
in
ear PC
A
w
i
t
h
the L
o
c
a
l Ap
proac
h fo
r
Diese
l E
ngi
ne
F
ault D
e
tecti
on
and
Di
ag
n
o
sis.
IEEE Transactions
on Cont
r
o
l System
Tec
h
nology
.
200
8; 16: 122-
129.
[18]
F
e
rrari S, Sten
gel
RF
. Smoot
h F
uncti
on A
p
p
r
oximatio
n Us
i
ng N
eur
al
Net
w
o
r
ks.
IEEE Transactions
on
Neur
al Netw
or
ks
. 2005; 16(
1)
: 24-38.
Evaluation Warning : The document was created with Spire.PDF for Python.