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.
8
, No
.
6
,
Decem
ber
201
8
, p
p.
5359
~
5370
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp
5359
-
53
70
5359
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Predi
ction B
ased Effici
en
t Resourc
e Pr
ovisionin
g and
Its
Impact o
n QoS Par
am
eters
in th
e Cloud E
nvironm
ent
Lata
J
G
adha
vi
,
Madhuri
D
Bhavs
ar
In
sti
tute
of Tec
hnology,
Nirm
a
U
niv
e
rsity
, Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n
2
3
, 2
01
8
Re
vised
J
un
2
9
, 201
8
Accepte
d
J
ul
24
, 2
01
8
The
purpose
of
thi
s
pape
r
is
to
provision
the
on
demand
resourc
es
to
the
end
users
as
per
t
hei
r
n
ee
d
usin
g
pre
di
ct
ion
m
et
hod
in
c
loud
computing
envi
ronm
ent
.
Th
e
provisioni
ng
of
virt
ual
i
ze
d
res
ourc
es
to
cl
oud
consum
ers
ac
cor
d
ing
to
th
eir
nee
d
is
a
cru
cial
step
in
the
de
plo
y
m
ent
of
app
li
c
at
ions
on
the
cl
oud
.
How
eve
r,
th
e
d
y
n
amical
m
ana
gement
of
resourc
es
for
var
ia
bl
e
workloads
remai
ns
a
ch
allengi
ng
proble
m
for
cl
o
ud
provide
rs.
Th
is
problem
ca
n
b
e
solved
b
y
using
a
pre
di
ct
ion
b
ase
d
adaptive
resourc
e
p
rovisioni
ng
m
ec
hani
sm
,
which
c
an
est
imate
th
e
upcomi
ng
resourc
e
d
emands
of
appl
i
ca
t
ions.
Th
e
pre
sent
rese
ar
ch
int
roduc
es
a
pre
dic
t
ion
base
d
resourc
e
provisioni
ng
m
odel
for
th
e
alloc
at
ion
of
r
esourc
es
in
adva
n
ce
.
T
he
proposed
appr
oac
h
fa
cilita
te
s
the
relea
se
o
f
unused
resourc
es
in
the
pool
with
qual
i
t
y
of
service
(QoS
),
which
is
de
fined
base
d
on
pre
di
ct
ion
m
ode
l
to
p
erf
orm
th
e
al
lo
ca
t
ion
of
re
source
s
in
adva
nce.
In
th
is
work,
the
m
odel
is
used
to
det
ermine
the
fu
ture
workload
p
red
iction
for
use
r
req
uests
on
w
eb
serve
rs,
and
it
s
impact
t
oward
ac
h
ie
ving
eff
i
ci
en
t
r
esourc
e
prov
isioni
ng
in
te
rm
s
o
f
resourc
e
exp
loi
t
at
ion
and
QoS
.
The
m
ai
n
contributi
on
of
th
is
pape
r
is
t
o
deve
lop
the
pre
d
ic
ti
on
m
odel
for
eff
icient
and
d
y
n
amic
resourc
e
pr
ovisioni
ng
to
m
ee
t
th
e
r
equirem
ent
s of end
u
sers.
Ke
yw
or
d:
D
ynam
ic
r
esource
prov
isi
on
i
ng
E
ff
ic
ie
nt
res
ource
prov
isi
on
i
ng
R
esources
de
plo
ym
ent
R
esources e
xploit
at
ion
Wor
klo
a
d pr
e
di
ct
ion
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Lat
a J G
a
dhavi
,
In
sti
tute
of Tec
hnology,
Nirm
a U
niv
e
rs
it
y, Ah
m
edab
a
d,
G
uj
a
rat,
Indi
a.
Em
a
il
: 12
extp
hde9
2@nirm
aun
i.ac.i
n
1.
INTROD
U
CTION
In
the
cl
oud
c
om
pu
ti
ng
env
i
r
on
m
ent,
cl
oud
us
ers
of
te
n
have
tim
e
-
changin
g
requirem
ents
fo
r
vi
rtual
resou
rces.
T
o
pro
vision
res
ources
for
dy
na
m
ic
and
uncert
ai
n
cha
ng
e
s
in
the
wor
klo
a
d
and
t
o
react
to
these
changes
ac
co
r
dingly
,
the
cl
oud
pro
vid
e
r
s
houl
d
m
anag
e
r
eso
ur
ces
base
d
on
the
re
qu
ire
m
ent
of
t
he
use
rs.
In
the
pr
ese
nt
stu
dy,
a
wor
klo
a
d
pr
e
dicti
on
m
od
el
for
dynam
ic
and
ef
fici
ent
resour
ce
prov
i
sion
i
ng
is
pro
pos
e
d
to m
anag
e the
nu
m
ber
of user
r
e
qu
est
s
and t
he req
uire
d res
ources.
On
e
pro
blem
with
s
uch
a
re
s
ource p
r
ovisi
onin
g
sc
hem
e
is
the
oc
cu
rr
e
nce
of
thra
sh
i
ng,
i
n
wh
ic
h,
d
ue
to
fr
e
qu
e
nt
vari
at
ion
of
the
w
orkloa
d
(
nu
m
ber
of
jo
b
reque
sts),
m
achines
can
be
ad
de
d
a
nd
release
d
to
m
eet
each
re
quirem
ent
w
hile
sat
isfyi
ng
th
e
Q
oS
m
et
rics.
So
lvi
ng
t
his
pro
ble
m
req
uire
s
an
abili
ty
to
pr
edi
ct
the
incom
ing
w
orkloa
d
on
the
s
yst
e
m
and
to
al
locat
e
resource
s
a
pr
i
or
i
by
us
ing
pr
e
dic
ti
on
m
et
ho
ds
f
or
th
e
require
d
re
sour
ces.
T
he
m
a
i
n
con
t
rib
ution
s
in
this pa
per
a
r
e;
(a)
T
he
desi
gn
of
a p
re
dicti
on
m
echan
ism
and
of
the
flo
w
of
t
he
pr
e
dicti
on
m
od
el
for
dif
fer
e
nt
per
io
ds
;
an
d
(
b)
T
he
us
e o
f
predict
io
n
m
e
tho
ds
t
o
determ
i
ne
the
work
l
oa
d base
d on a
histor
ic
a
l database
.
The
rest
of
t
he
pa
pe
r
is
or
gani
zed
as
fo
ll
ow
s
.
Sect
io
n
II
pre
sents
a
su
m
m
a
ry
of
the
relat
e
d
work
in
this
dom
ai
n.
S
ect
ion
III
pro
vi
des
the
dom
ain
a
naly
sis
f
or
the
w
orkl
oad
patte
rn.
Sect
io
n
IV
intr
oduce
s
the
pr
e
dicti
on
m
e
chan
ism
.
Sect
i
on
V
descr
i
be
s
the
case
stud
y
us
ed
to
va
li
d
at
e
the
pr
opose
d
ap
proac
h
an
d
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.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5359
-
5370
5360
nu
m
ber
of
jo
b
request
requir
e
m
ent
fo
r
resour
ces
.
I
n
sect
ion
VI,
we
pre
sent
the
eval
ua
ti
on
res
ults
f
or
the
resou
rce
prov
is
ion
in
g
a
nd all
oc
at
ion
. Fi
nally
,
secti
on VII
pr
esents the
conc
lud
in
g rem
ark
s
and
fu
t
ur
e
sco
pe.
2.
RELATE
D
W
ORK
Cl
oud
com
pu
ti
ng
c
har
act
e
rizes
the
delivery
of
c
om
pu
ta
ti
on
as
a
serv
ic
e,
in
w
hich
re
sour
ces,
su
c
h
a
s
the
CPU
,
s
of
t
war
e
,
ha
r
dw
a
r
e,
inf
orm
ation
,
and
de
vices,
a
re
gra
nted
to
use
rs
as
se
r
vices
thr
ough
the
I
nter
net.
The
c
har
act
e
risti
cs
of
cl
oud
com
pu
ti
ng
,
s
uc
h
as
aut
o
-
scal
ing
,
the
pro
vision
i
ng
of
se
r
vices
base
d
on
de
m
and
,
and
t
he
util
it
y
m
echan
ism
,
has
been
la
rg
el
y
adopted
by
dif
fer
e
nt
analy
sts
[1
]
.
Cl
oud
c
om
pu
ti
ng
pro
vi
des
a
scal
able
and
flexible
platf
orm
fo
r
pro
vidi
ng
hi
gh
-
perform
ance
co
m
pu
ti
ng
require
m
ent
s
[2
]
,
[
3
]
,
[
4
].
Howe
ver,
the
pool
of
res
our
ce,
fle
xib
le
se
r
vice
pr
ov
isi
oni
ng,
a
nd
el
ast
ic
it
y
are
not
the
on
ly
po
te
ntial
s
of
the
cl
oud.
The
cl
oud
com
pu
ti
ng
syst
e
m
pr
ovisi
ons
a
hom
og
eneous
an
d
heter
og
e
ne
ous
operati
ng
syst
e
m
env
i
ronm
ent by
u
sin
g virtuali
zat
ion
.
In
the
IT
in
du
stry,
m
ulti
ple
cl
oud
pro
vid
e
r
s
deliver
Q
oS
to
end
use
rs
a
ccordin
g
to
th
ei
r
dem
and
.
Howe
ver,
a
n
i
ssu
e
t
hat
em
erg
es
from
the
init
ia
l
to
the
de
plo
ym
ent
sta
ge
is
t
he
reali
t
y
that
the
patte
rn
of
acce
ss
to
the
app
li
cat
io
n
by
va
rio
us
en
d
us
e
r
s
var
ie
s
f
reque
n
tl
y.
As
a
resu
l
t,
there
is
an
unpre
dicta
ble
num
ber
of
us
er
s
an
d
var
ia
ble
w
orkload
duri
ng
s
om
e
per
io
ds
.
I
n
a
sta
ti
c
and
ineff
ic
ie
nt
re
so
urce
prov
isi
on
i
ng
m
echan
ism
,
durin
g
pe
rio
ds
of
low
dem
and
f
or
res
ources,
there
will
be
a
n
ov
e
rloa
d
of
ob
ta
ina
ble
res
ource
s,
wh
e
reas
duri
ng
per
i
od
s
of
hi
gh
dem
and,
th
e
avail
able
res
ources
will
be
inade
qu
at
e
.
T
hi
s
issue
le
a
ds
t
o
poor
QoS
a
nd
high
costs.
A
cu
rr
e
nt
researc
h
chal
le
ng
e
is
to
de
ve
lop
an
ada
ptive
m
et
ho
d
of
r
eso
ur
ce
pro
visi
on
i
ng
in term
s o
f
c
ost
an
d pe
rfo
rm
a
nce.
Cl
oud
com
pu
ti
ng
can
a
ddre
ss
the
above
chall
enge
by
pro
vision
i
ng
r
eso
ur
ces
dy
na
m
ic
al
l
y
and
eff
ect
ively
to
end
-
us
er
a
pp
li
c
at
ion
s
base
d
on
the
predict
io
n
of
wor
klo
a
d
and
resou
rces
accor
ding
to
th
e
us
e
r
request
a
rr
ival
r
at
e
an
d
t
he
se
r
vice
res
ponse
r
at
e.
This
m
e
an
s
that
a
dd
it
io
na
l
resou
rces
ca
n
be
pro
vid
e
d
duri
ng
per
i
od
s
of
high
dem
and
a
nd
r
el
eased
durin
g
per
i
od
s
of
l
ow
dem
and
,
wit
hout
loss
of
Q
oS
to
en
d
us
e
rs
[5
]
Th
e
chall
enge
with
su
ch
a
dynam
ic
resour
ce
pro
visio
ning
strat
egy
is
the
dete
rm
inati
on
of
th
e
prop
e
r
qua
ntit
y
of
resou
rces
to
be
set
up
an
d
pro
vision
e
d
in
a
par
ti
cular
per
i
od
to
m
ee
t
the
exp
ect
ed
Qo
S
for
a
va
riabl
e
work
l
oa
d.
A
n
ideal
so
l
utio
n
would
re
quire
the
ca
pa
bili
ty
to
predict
th
e
inc
om
ing
w
orkl
oad
an
d
re
quire
d
resou
rces in
advance
.
The
ex
pected
outc
om
e is the d
et
erm
inati
on
of
the n
um
ber
of
virtu
al
m
achines that sh
ould
be
create
d,
config
ur
e
d,
a
nd
pro
vid
e
d
to hand
le
the v
aria
ble
w
orkloa
d.
Th
is
chall
en
ge
has
bee
n
a
ddre
ssed
th
rou
g
h
va
rio
us
ways,
s
uc
h
rea
ct
ive
[
6],
pr
oa
ct
ive
[
6],
an
d
pr
e
dicti
ve
[
7
]
ap
pr
oach
es
.
E
f
fecti
ve
re
sourc
e
pro
visio
ning
is
not
an
easy
a
nd
un
com
plica
te
d
ta
sk
.
To
m
eet
the
ab
ove
r
eq
uir
e
m
ents,
the
f
ollow
in
g
crit
eria
sho
uld
be
co
nsi
der
e
d
in
de
velo
ping
t
he
al
gorithm
:
(a)
the
c
om
pu
ta
ti
on
of
t
he
use
r
re
quest
rate,
(b)
the
m
ini
m
i
zat
ion
of
t
he
r
equ
e
st
rej
ect
io
n rate
, (
c) th
e a
ve
rag
e
respo
ns
e ti
m
e
and the
num
ber
of
requests ti
m
ed
out, (
d) th
e p
erce
nta
ge
of
tim
ed
ou
t
re
qu
e
sts
(P
TOR)
for
a
nu
m
ber
of
us
e
r
s,
and
(e
)
the
accurate
com
pu
ta
ti
on
in
ad
va
nce
of
the
re
so
urc
e
s
require
d
for
a
va
riable
w
ork
l
oad.
T
he
c
urren
t
resea
rc
h
pr
ese
nts
a
n
eff
ic
ie
nt
an
d
effe
ct
ive
res
ource
pro
vision
i
ng
al
gorithm
that
us
es
a
pr
e
dicti
on
te
ch
nique
to
prov
isi
on
a
nd
rem
ov
e
resou
rces
dynam
ic
ally.
A
strat
egy
to
im
pr
ov
e
the
re
source
util
iz
at
ion
is
pr
opose
d
a
nd
c
om
par
ed
wi
th
the
conve
ntion
al
a
ppr
oach.
3.
DOMAI
N A
N
ALYSIS
The
cl
oud
c
ompu
ti
ng
f
orm
at
i
s
m
ai
nly
descr
ibed
by
us
in
g
t
hr
ee
se
rv
ic
e
m
od
el
s:
(a
)
I
nfra
structu
re
as
a
Serv
ic
e
(
Iaa
S)
,
wh
ic
h
ex
pl
ai
ns
ab
ou
t
the
resou
rces
offe
r
ing
procedu
res
by
cl
oud
pro
vi
der
s;
(
b)
Plat
f
or
m
as
a
Se
rv
ic
e
(P
aa
S)
,
wh
ic
h
desc
ribes
how
cl
oud
pro
vid
e
rs
pro
vid
e
a
n
entir
e
cl
oud
en
vir
onm
ent
i
m
ple
mente
d
and
dep
l
oyed
in
a
certai
n
pro
gr
am
m
ing
la
ng
ua
ge
f
or
a
spe
ci
fic
ty
pe
of
app
li
cat
io
ns
;
and
(c
)
Softwa
r
e
as
a
Ser
vice
(S
aaS)
,
wh
ic
h
ref
e
rs
to
app
li
cat
ions
that
can
be
offer
e
d
to
custo
m
ers
accor
ding
to
their
nee
d.
The
dep
l
oyed
cl
ou
d
m
od
el
s
are
m
ai
nly
cl
assified
into
f
our
f
or
m
s:
(a)
publ
ic
cl
ou
d,
w
hi
ch
is
avail
able
fo
r
ever
y
on
e;
(b)
pr
i
vate
cl
oud,
wh
ic
h
is
hoste
d
s
olely
fo
r
on
e
industry;
(c
)
com
m
un
it
y
cl
o
ud,
w
hich
is
a
cl
oud
env
i
ronm
ent
m
ade
acce
ssib
le
on
ly
to
a
c
ertai
n
gr
oup
of
in
dustry
or
i
nd
i
viduals
in
colla
borati
on;
and
(
d)
hybri
d
cl
ou
d,
wh
ic
h
re
fer
s
to
var
i
ou
s
cl
ouds
that
are
interc
onnected
with
the
hoste
d
a
ppli
cat
ion
s
de
plo
y
ed
i
n
the
cl
oud
e
nvir
on
m
ent
[8]
.
In
the
present
w
ork
,
the
te
rm
“work
l
oa
d”
re
fe
rs
to
th
e
num
ber
of
ar
rive
d
re
qu
e
sts
to access t
he re
so
urce
of the
cloud syste
m
.
The
a
ppli
cat
ion
or
jo
bs
a
re
ne
ed
t
o
be
s
witc
hed
a
uto
m
at
i
cal
ly
wh
ic
h
ar
e
acce
sse
d
by
us
ers
.
In
th
e
cl
oud
com
pu
ti
ng
sce
na
rio,
t
he
w
orkloa
d
cat
egory,
cl
ou
d
ser
vice
m
od
el
s,
an
d
depl
oy
m
ent
m
od
el
s
are
interco
nnect
ed
with
each
ot
he
r,
as
sho
wn
i
n
Fig
ure
.
1.
T
he
ap
plica
ti
on
work
l
oa
d
patte
rn
descr
i
bes
di
ver
se
us
er
beh
a
viors
,
wh
ic
h
res
ult
in
the
util
iz
a
ti
on
of
IT
res
ources
in
va
ri
able
form
s.
T
he
w
or
klo
a
d
c
an
be
determ
ined
bas
ed
on
the
num
ber
of
us
e
r
re
quest
s,
the
l
oad
cal
culat
ion
on
the
ser
ver
s
,
the
netw
ork
tra
ff
i
c,
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Predict
ion B
ase
d
Eff
ic
ie
nt Re
so
urc
e Pr
ovisi
on
i
ng an
d Its I
mpact
on Qo
S
…
(
La
ta J
Ga
dhavi
)
5361
the
data
st
or
a
ge
[9
]
.
In
t
he
cl
oud
c
om
pu
ti
ng
env
i
ronm
ent,
the
res
ource
pr
ov
isi
on
i
ng
de
pe
nds
on
the
inc
om
ing
work
l
oa
d
[
10]
.
For
the
p
red
i
ct
ion
of
t
he
w
orkloa
d
a
nd
re
so
urces
i
n
this
stud
y,
the
w
orkl
oad
patte
r
ns
are
cat
egorized
ba
sed
on
t
he
inc
om
ing
w
orklo
ad,
w
hich
incl
ud
e
s
the
nu
m
ber
of
use
r
re
quest
s,
the
am
ou
nt
of
work
over
a
s
pecified
pe
rio
d,
the
us
er
re
quest
arr
ival
rate,
an
d
the
ti
m
e
betwee
n
tw
o
consecuti
ve
re
qu
e
sts
(inter
-
ar
rival
ti
m
e).
Accord
i
ng
to
the
ab
ov
e
char
act
erist
ic
s
,
the
wor
klo
a
d
is
cat
ego
rized
into
the
five
ty
pes:
sta
ti
c, p
eri
od
ic
,
unpre
dicta
ble,
conti
nuously
c
ha
ngin
g, an
d o
nce in a l
ifet
im
e [
11
]
.
3.1.
St
ati
c W
orkl
oad
The
w
ork
loa
d
for
resou
rces w
it
h
c
orres
pondin
g
or
e
qu
al
usage of
a hoste
d
a
ppli
cat
ion
is
cal
le
d
sta
ti
c
work
l
oa
d.
In
this
w
orkl
oad
ty
pe,
the
re
is
no
re
qu
i
rem
ent
to
inc
rease
or
dec
rease
t
he
processi
ng
powe
r
,
netw
ork
ba
ndwidth,
data
st
orage
,
or
m
e
m
o
ry
beca
us
e
the
w
ork
loa
d
is
c
on
sta
nt.
A
sta
ti
c
wor
klo
a
d
does
not
require
the
fun
ct
ion
al
it
y
of
th
e
cl
oud,
s
uch
as
el
ast
ic
it
y.
Pr
ivate
Websi
te
s
an
d
wikis
ar
e
exam
ples
of
sit
es
with a
sta
ti
c wor
klo
a
d.
3.2.
Peri
od
ic
W
or
kload
Perio
dic
jo
bs
are
ve
ry
com
m
on
in
our
da
il
y
ro
utine.
Ye
arly
incom
e
tax
pay,
m
on
thl
y
util
ity
bill
s,
traff
ic
duri
ng
ru
s
h
h
o
ur
s
et
c
are
the
exam
ples
of
per
io
di
c
ta
sk
s
or
jobs.
It
is
ob
ser
ve
d
that
in
the
sa
m
e
interval,
m
any
people u
sin
g
t
he
se
ta
sk
s
. D
uri
ng
s
uch
pe
rio
di
c
ta
sk
s,
it
is
di
ff
ic
ult
t
o
prov
i
de
e
nough
re
sourc
es
for
the
pea
k
lo
ad
an
d
t
o
ha
nd
le
the
unus
e
d
r
eso
ur
ces
for
th
e
non
-
pea
k
lo
a
d.
T
his
prob
le
m
le
ads
to
ove
r
-
or
unde
r
-
pro
visio
ning
of
res
ourc
es to th
e
hoste
d
a
pp
li
cat
ion.
Perio
dic task
s
occur at th
e sa
m
e interval
of
t
he day
,
m
on
th,
or
ye
ar;
howev
e
r,
th
ey
con
sist
of
a
hig
he
r
num
ber
of
re
qu
e
sts
ov
e
r
a
peak
per
io
d.
A
pe
rio
dic
work
l
oa
d
oft
en
requires
el
ast
ic
it
y
and
scal
ab
il
ity
to
ha
nd
le
the
num
ber
of
requests
durin
g
the
pea
k
inte
rv
al
s
.
The
c
om
pu
tim
g work
l
oa
d patt
ern
is
de
picte
d o
n
Fi
gure
1.
Figure
1. The
Cl
oud
Com
pu
t
ing
Wor
klo
a
d Pat
te
rn
3.3.
Un
pre
dict
ab
le
Wo
r
klo
ad
Si
m
il
arly
to
a
per
i
od
ic
w
orkl
oad,
an
un
pr
e
di
ct
able
wo
r
kloa
d
co
ns
ist
s
of
increasin
g
an
d
decr
easi
ng
incom
ing
wor
kl
oad
s
f
ro
m
the
us
ers
.
This
w
orkl
oad
is
desc
ribed
as
un
pr
e
di
ct
able
becau
se the
var
ia
ti
ons
occur
rand
om
l
y.
As
a
res
ult,
cl
oud
pro
vid
e
rs
have
to
deal
with
th
e
un
st
ru
ct
ured
a
nd
un
plan
ned
pro
visio
ni
ng
of
resou
rces
t
o
m
eet
the
cha
ngin
g
requirem
ents.
Acc
ur
at
e
predict
io
n
is
the
m
ai
n
chall
eng
e
to
obta
in
ing
t
he
scal
ing
requir
e
m
ents
for
re
so
urces
for
a
n
un
pr
e
dicte
d
wor
klo
a
d.
T
o
ac
hieve
acc
ur
at
e
resu
lt
s,
const
an
t
ob
s
er
vation o
f t
he
w
orkl
oad is re
qu
i
red f
or e
xam
ple, u
np
re
dicta
ble tra
ff
ic
,
for
ecast
in
g
et
c
.
3.4.
Continu
ousl
y Ch
anging
W
or
kload
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.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5359
-
5370
5362
The
resou
rce
ut
il
iz
ation
m
ay
be
m
axi
m
iz
ed
or
m
ini
m
iz
ed
ov
e
r
ti
m
e,
and
the
w
orkl
oad
m
ay
chan
ge
con
ti
nu
ously
.
Fo
r
m
any
appl
ic
at
ion
s,
the
re
m
ay
be
lon
g
-
te
rm
chan
ges
in
the
w
orkl
oad.
A
c
onti
nuously
changin
g
w
ork
load
ca
n
be
co
ns
ide
red
as
a
c
on
sta
nt
escal
at
ion
or
decli
nation
of
the
res
ource
util
iz
at
ion.
T
he
el
ast
ic
ity
of
the
cl
oud
e
na
bles
the
pro
vi
sion
i
ng
or
de
com
m
issi
on
ing
o
f
res
ources
fo
r
a
c
on
ti
nuously
changin
g w
ork
load.
3.5.
Once
-
in
-
a
-
li
fe
time W
orkloa
d
The
peak
w
ork
load
util
iz
at
ion
m
ay
occu
r
on
l
y
on
ce
i
n
a
li
fe
tim
e
wh
en
the
resou
rces
are
di
stribu
te
d
equ
al
ly
.
In
su
c
h
a
s
pecial
ca
se,
the
res
ourc
e
pro
visio
ning
an
d
deco
m
m
i
ssion
i
n
g
can
oft
en
be
a
m
anu
al
or
dynam
ic
task at a know
n po
i
nt
in
tim
e.
4.
PREDI
CTIO
N MEC
HANI
SM
To
pr
e
dict
the
incom
ing
wor
klo
a
d
a
nd
to
identify
t
he
work
l
oad
ty
pe,
t
he
us
e
r
requests
ha
ve
to
be
analy
zed.
A
n
analy
sis
of
the
histo
rical
data
base
is
al
s
o
ne
eded
to
m
ee
t
t
he
resou
rce
re
qu
i
rem
ents
a
pr
iori
.
Ba
sed
on
t
he
use
r
re
quest
s
an
d
their
nee
d
a
na
ly
sis
are
cons
idere
d
f
or
host
ed
ap
plica
ti
on.
Fig
ure
2
s
how
s
the
proce
dure
of
predict
io
n
at
ti
m
e
t
-
1
a
nd
inte
rval
t.
The
previ
ou
s
pe
rio
d
for
interval
(t
-
1)
a
nd
the
li
ve
pe
ri
o
d
f
or
interval
t
are
obser
ve
d
for
a
nu
m
ber
of
act
ive
us
e
rs.
T
he
r
eso
ur
ce
re
quir
e
m
ents
fo
r
the
ob
se
rv
e
d
pe
riod
are
m
on
it
or
ed,
c
ol
le
ct
ed,
an
d
ent
ered
int
o
the
hi
storical
databa
se
after
the
pro
cedure.
Ba
se
d
on
t
he
data
bas
e,
the
work
l
oa
d
for
tim
e
intervals
t
-
1
an
d
t
are
cal
culat
ed,
an
d
pr
e
dicti
on
m
eth
ods
are
a
pp
li
ed
to
determ
ine
the
work
l
oa
d
f
or t
he next
pe
rio
d (t+1)
.
Figure
2.
T
he Fl
ow of t
he
P
r
edict
ion
M
odel
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
Predict
ion B
ase
d
Eff
ic
ie
nt Re
so
urc
e Pr
ovisi
on
i
ng an
d Its I
mpact
on Qo
S
…
(
La
ta J
Ga
dhavi
)
5363
As
show
n
in
the
ab
ov
e
Fi
g
ure
2,
pre
dicti
on
of
work
l
oa
d
for
the
res
ourc
e
dem
and
s
is
descr
i
bed
in
diff
e
rent
interval.
F
or
ti
m
e
interval
t
−
1
,
th
e
m
ean
CPU,
RAM
w
orkl
oa
d,
a
nd n
um
ber
o
f
act
ive u
ser
s ar
e
cal
culat
ed
as
fo
ll
ows:
α
−
1
(
c
pu
)
=
CPU
Wor
kload
Me
an
at
th
e
in
t
er
va
l
t
−
1
β
−
1
(
R
A
M
)
=
RAM
W
orkl
oad
Me
a
n
at
th
e
in
t
er
val
t
−
1
γ
−
1
(
a
c
tive
use
r
)
=
Me
a
n
num
be
r
of
act
iv
e
user
s
at
t
ime
t
−
1
λ
−
1
′
=
Fo
r
wh
ole sce
na
rio
at
ti
m
e t
-
1
(1)
Fo
r
ti
m
e interval
t
, th
e m
ean CPU
,
RAM
w
orkloa
d,
a
nd
num
ber
o
f
act
i
ve users
are
calc
ul
at
ed
as foll
ow
s:
α
(
c
pu
)
=
Me
a
n
of
C
PU
W
orkl
oad
at
ti
me
t
β
(
R
AM
)
=
Me
a
n
of
RAM
Wor
k
loa
d
at
t
ime
t
γ
(
a
c
t
ive
u
ser
)
=
Me
a
n
of
numb
er
of
ac
t
iv
e
user
s
at
ti
me
t
λ
′
=
Fo
r wh
ole s
cenario
a
t time
t
(2)
To
c
onside
r
th
e wor
klo
a
d
at
t
he
init
ia
l l
evel,
a pseu
do
-
c
od
e
is d
esc
rib
e
d
i
n Alg
ori
thm
1
.
Algori
th
m
1
:
I
nitial Le
vel
Inp
ut: N
um
ber
of r
e
quest
s
Re
qu
irem
ent: C
al
culat
ion
of
the inter
ar
riva
l t
i
m
e fo
r
t
he p
erio
ds
-
1
a
nd
per
i
od
1.
Be
gi
n.
2.
M
on
it
or
acti
ve user
r
e
quest
s
3.
Wh
il
e
(tim
e inter
val is
-
1).
4.
Ca
lc
ula
te
the inter
ar
rival tim
e o
f
act
i
ve user
s.
5.
For
(
each inte
r
a
rr
iv
al
tim
e in p
erio
d
-
1).
6.
M
on
it
or
w
orkloa
d
7.
Ca
lc
ulate
m
ean v
al
ue
λ
−
1
′
for week
1.
8.
For
(
each inte
r
a
rr
iv
al
tim
e o
f
pe
rio
d)
.
9.
M
on
it
or t
he wor
klo
a
d.
10.
Ca
lc
ulate
the m
ea
n value
λ
′
f
or w
e
ek2.
11.
En
d for
12.
En
d for
13.
E
nd.
Now
f
or
each
week
of
t
he
m
on
t
h,
cal
culat
e
the
m
ean
value
of
the
w
orkl
oad
a
nd
predic
t
fo
r
ne
xt
wee
k
base
d
on curre
nt a
nd
the h
ist
or
ic
al
da
ta
base.
T
hus,
the pre
dicte
d v
al
ue
f
or ti
m
e
+ n
is:
λ
+
n
=
λ
−
1
′
+
λ
′
+
λ
+
1
′
+
⋯
+
λ
+
(
n
−
1
)
′
(3)
λ
+
n
=
∑
λ
′
+
(
n
−
1
)
−
1
(4)
(3)
a
nd
(
4)
ex
pr
ess
the
su
m
m
at
ion
of
the
pr
e
vious
w
ork
loa
d
t
o
pr
e
dict
the
re
sour
ce
dem
and
s
f
or
t
he
nex
t
per
i
od. T
he ps
eudo
-
co
de fo
r
t
he wor
klo
a
d pre
dicti
on
is
pres
ented
i
n Alg
or
i
thm
2
.
Algori
th
m
2
: Workl
oad Pre
diction
Inp
ut: N
um
ber
of r
e
quest
s
Re
qu
irem
ent: C
al
culat
ion
of
the m
ean CPU
,
RAM, a
nd
num
ber
o
f
act
i
ve users
at ti
m
e in
te
rv
al
-
1
a
nd
per
i
od
1.
Be
gin
.
2.
M
on
it
or
t
he nu
m
ber
of act
ive
us
er
r
e
quest
s at tim
e
-
1
a
nd p
e
rio
d
.
3.
Ca
lc
ulate
λ
−
1
′
an
d
λ
′
//
by
us
in
g (
1) an
d (2).
4.
F
or eac
h w
eek,
cal
c
ulate
the m
ean workl
oad.
5.
P
re
dict t
he
workloa
d for
the n
e
xt
week
+ n
,
∑
λ
′
+
(
n
−
1
)
−
1
.
6.
E
nd fo
r.
7.
E
nd.
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.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5359
-
5370
5364
To
al
locat
e
res
ources t
o
the
jo
b,
i
n
this se
ct
io
n nu
m
ber
of
jo
b
ide
ntific
at
ion are c
onsidere
d for
var
i
ous r
e
gi
on
s
as sho
wn in t
he
Fig
ure
. 3.
Figure
3.
The
Flow o
f
t
he
Re
so
urce
Allocat
i
on
Determ
inati
on
of the
num
ber
o
f
re
qu
e
sts:
Jo
b p
r
of
il
e:
n
i
t
j
j
J
,
.
.
.
.
.
.
,
, whe
re i=1
...n, an
d
t
J
is
num
ber
of
j
obs
f
or
t
he c
urren
t t
im
e
L
L
oc
at
i
on
A
I
P
a
dd
D
D
a
t
e
T
T
i
m
e
j
,
,
,
1
Let
,.......
,
,
3
2
1
T
T
T
T
tim
e series
.
.
.
.
,
2
1
d
d
D
Date o
f req
uest
arr
i
ved
n
n
i
i
c
r
c
r
L
,
,
.
.
.
.
.
.
.
.
,
, whe
re
L
is t
he
locat
ion o
f
t
he a
rr
ive
d re
qu
est
c
r
,
Re
gion a
nd coun
t
ry, r
e
sp
ect
i
vely
Total
num
ber
of jo
b
re
quest
s
from
o
ne regi
on=
R
R
Nu
m
ber
of r
e
quest
s
per re
gion
(
5)
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
Predict
ion B
ase
d
Eff
ic
ie
nt Re
so
urc
e Pr
ovisi
on
i
ng an
d Its I
mpact
on Qo
S
…
(
La
ta J
Ga
dhavi
)
5365
Figure
4.
N
umber
of Re
qu
est
s b
y R
e
gion
Figure
4
s
hows
the num
ber
of
j
ob
requests
by
r
egi
on
(R)
. D
i
ff
e
ren
t
pr
e
dicti
on
m
od
el
s a
re
app
li
ed
for
the
res
ource
r
equ
i
rem
ents
i
n
va
rio
us
pe
ri
od
s
,
inclu
ding
the
first
an
d
second
m
ov
ing
a
ver
a
ge
(FSM
A),
weig
hted
m
ov
i
ng
ave
rag
e
(
W
MA),
a
nd
e
xpon
e
ntial
m
ov
ing
a
ver
a
ge
(
EMA)
m
od
el
s
[11
]
.
The
histor
ic
al
data
and
pr
e
dicti
on
m
od
el
ou
tp
ut
are
us
e
d
as
in
pu
t
data,
w
hich
a
re
trai
ne
d
to
i
m
pr
ov
e
the
acc
ur
acy
of
the
re
so
urce
pro
vision
i
ng
st
rategy.
T
he
pre
dicti
on
m
od
el
us
es
t
he
ou
t
pu
t
of
the
pr
edict
e
d
value
t
o
provi
sion
th
e
res
ou
rces.
At
the
sam
e
t
i
m
e,
pr
ovisi
one
d
res
ources
res
ults
will
be
sto
red
in
t
he
histo
rical
database
reposit
ory
for
fu
t
ur
e
pr
e
dicti
on.
Th
e
pr
e
dicti
on
is
done
by
ap
pl
yi
ng
va
rio
us
m
ov
ing
ave
rage
m
et
ho
ds
.
In
par
ti
cula
r,
the
m
ov
i
ng
aver
a
ge
m
et
ho
ds
FSM
A
an
d
WMA
are
recom
m
end
ed
to
c
on
si
der
non
-
fl
uctuati
ons
in
the
short
-
te
rm
dem
and
,
as
sh
ow
n
in
ou
r
pr
e
vious
wor
k
[
12
]
.
A
n
on
t
ology
-
based
dy
nam
ic
reso
ur
ce
prov
isi
on
i
ng
f
or
pu
blic
cl
oud
wa
s
i
m
ple
m
ented
in the
pre
vious
work [
13]
.
First
and Sec
ond M
ovin
g
Ave
ra
ge
(FSM
A)
meth
od:
In
t
he
fir
st
m
o
ving
a
ver
a
ge
m
et
ho
d,
the
i
th
us
er
re
qu
e
st
f
or
res
ource
at
tim
e
interval
t
is
m
easur
ed
,
and N i
s c
on
si
der
e
d
as
a m
oving
a
ve
rag
e
p
e
rio
d
f
or ti
m
e in
te
rv
al
t [
9], as
descr
i
bed in
(
6):
+
1
=
(
(
)
+
−
1
(
)
+
⋯
+
−
(
)
)
(6)
Her
e
,
y
is
co
nsi
der
e
d
as
the
ori
gin
al
val
ue
f
or
each
per
io
d,
an
d
t
is
re
ga
rded
as
the
cu
rr
e
nt
per
io
d.
I
n
(
6),
F
t+1
denote
the
fir
st
m
ov
in
g
ave
ra
ge
val
ue
f
or
t
he
i
th
us
er
at
ti
m
e
t+
1.
The
r
e
so
urce
require
m
ent
of
the
i
th
us
er
at
tim
e t+
δ
can be predict
ed
as:
+
(
i
)
=
(
i
)
+
(
)
(7)
wh
e
re
δ
is
the
pr
e
dicte
d
ti
m
e
seq
uen
ce
num
ber,
an
d
+
de
no
t
es
the
pr
e
dicte
d
value
at
ti
m
e
+
,
as
s
ho
w
n
in (7). T
he
n,
t
o p
red
ic
t t
he
+
va
lue, the
sec
ond m
ov
ing
a
ve
ra
ge va
lue c
an
be m
easur
ed by
app
ly
in
g (8):
+
1
=
+
1
(
)
+
(
)
+
⋯
+
−
(
−
1
)
(
)
(8)
In (7)
,
(
)
and
(
)
are calc
ulate
d b
y usin
g (9) a
nd
(10):
(
)
=
2
+
1
(
)
−
+
1
(
)
(
9
)
(
)
=
2
+
1
(
)
−
+
1
(
)
−
1
(
10
)
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.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5359
-
5370
5366
B
ased
on the a
bove
e
quat
io
n,
the total
nu
m
ber
of
res
ource
r
equ
i
rem
ents f
or
n user
s
is m
e
asur
e
d
a
s:
+
=
(
11
)
The
total
num
ber
of
res
ource
s
requested
by
al
l
the
n
us
e
rs
at
tim
e
+
can
be
cal
culat
ed,
a
nd
t
he
predict
e
d
value
at
tim
e
per
i
od
s
and
+
for
f
or
al
l
us
e
r
s
can
be
deter
m
ined
by
us
i
ng
eq
uatio
n
.
(
11)
.
Furthe
r,
th
e
resou
rces
are
a
ll
ocated
to
t
he
us
ers
acco
r
ding
to
t
heir
act
ua
l
need
a
nd
ba
sed
on
t
he
pr
e
dicte
d
value,
a
nd
t
he
resu
lt
s a
re c
ompare
d.
Wei
gh
te
d
M
ovi
ng Aver
ag
e
Me
thod (
WMA)
:
Wh
e
reas
the
first
an
d
seco
nd
m
ov
in
g
ave
ra
ge
m
od
el
al
locat
es
the
sam
e
weig
ht
to
each
el
e
m
ent
of
the
m
ov
ing
a
ve
rag
e
databa
se
,
the
wei
ghte
d
m
ov
in
g
a
ver
a
ge
m
od
el
al
lows
a
ny
weig
ht
to
be
placed
on
eac
h
el
e
m
ent,
and
t
h
e
s
um
of
al
l
weig
hts
eq
ua
ls
1
[
11
]
.
He
r
e,
the
act
ive
use
r
re
quest
s
a
nd
t
he
util
iz
ation
of
resou
rces
for p
erio
d
F
t
are
d
e
scribe
d
in
(
12):
F
=
W
A
−
+
W
+
1
A
−
(
+
1
)
+
…
.
W
A
−
(
12
)
Wh
e
re
F
t
=
the
for
ecast
for
t
he
co
m
ing
p
e
rio
d,
n=the t
otal n
um
ber
o
f pe
rio
ds i
n
t
he
f
oreca
s
t,
A
i =
the act
ual
occurre
nce
f
or
the p
e
rio
d
t
-
i,
Her
e
,
∑
W
=
1
=
1
=
∑
(
(
ℎ
)
−
(
)
)
∑
ℎ
(
1
3
)
Exp
on
e
ntial M
ovin
g
Aver
age
Met
hod(EMA)
:
The
e
xpone
ntial
m
ov
in
g
a
ve
rag
e
(EM
A)
m
et
hod
is
m
ai
nly
log
ic
al
an
d
i
s
the
easi
est
a
ppr
oach
f
or
pr
e
dicti
ng
fluc
tuati
on
s
i
n
the
sh
ort
-
te
rm
dem
and
.
T
his
m
e
thod
is
eff
ect
i
ve
for
both
s
hor
t
-
te
rm
and
tim
e
-
series
pr
e
dict
ion.
As
each
ne
w
pie
ce
of
data
is
a
dd
e
d,
t
he
old
e
st
ob
se
rv
at
io
n
is
dr
op
pe
d,
an
d
a
ne
w
f
or
ec
ast
is
cal
culat
ed
[
14
]
.
The
pr
e
dicte
d
values
a
re
cal
culat
ed
by
us
i
ng
t
he
sm
oo
thi
ng
c
onsta
nt
α
.
The
EM
A
m
eth
od
is
expresse
d
as:
F
=
F
−
1
+
α
(
A
−
1
−
F
−
1
)
F
=
α
A
−
1
)
+
(
1
−
α
)
F
−
1
(
14
)
whe
re
F
is
t
he
exp
onen
t
ia
ll
y
smoot
he
d
fore
cast
for
prev
ious
p
er
iod
t
F
−
1
is
t
he
e
xponentia
ll
y sm
oo
thed for
ecast
m
ade f
or
the
prio
r
per
i
od,
A
−
1
=
The
a
ctua
l
dema
nd
in
t
he
p
ri
or
p
er
iod
,
α
=
The
desir
ed
re
spon
se
ra
t
e
or
s
mo
othi
ng
c
ons
t
an
t
, 0
<
α
<
1
This
m
e
tho
d
gi
ves
to
a
hig
he
r
weig
ht
to
the
la
te
r
m
easur
e
d
value
a
nd
a
lowe
r
weig
ht
to
the
earli
er
m
easur
ed
val
ue
.
The
EMA
m
et
ho
d
is
a
ble
to
res
pond
qu
ic
kly
to
fluct
ua
ti
on
s
in
the
s
hort
-
te
rm
dem
and.
If
request
f
ro
m
act
ive
us
er
is
great
er
than
t
he
def
i
ned
per
i
od
then
it
will
be
con
si
der
e
d
in
the
long
-
te
rm
per
i
od
and
if
it
is
le
ss
than
the
n
it
w
il
l
be
con
si
dered
as
a
s
hort
te
rm
per
iod
.
Lo
ng
-
te
rm
and
short
-
te
rm
per
iod
s
ar
e
def
i
ned as L
t
a
nd S
t
,
r
es
pecti
ve
ly
.
5.
NUMBE
R
OF
J
OB
R
EQ
UE
ST REQ
UI
RE
MENT
FO
R RESO
URCES
Her
e
,
the
nu
m
ber
s
of
re
qu
es
ts
fr
om
var
io
us
reg
i
on
s
a
re
cal
culat
ed,
as
sh
ow
n
in
(1)
and
(
2)
.
T
he
nu
m
ber
of
a
rr
i
ved
us
e
r
requ
e
sts
in
the
de
plo
ye
d
cl
oud
for
ed
ucati
on
is
e
valuated
,
as
show
n
i
n
ou
r
pre
vious
work,
a
nd
the
avail
abili
ty
of
resou
rces
in
th
e
sp
eci
fie
d
regi
on
s
is
dete
rm
ined.
T
he
a
vaila
ble
res
ources
ar
e
gen
e
rated
an
d
m
on
it
or
ed
by
us
in
g
t
he
c
omm
ercial
Am
az
on
EC2
cl
ou
d
pla
tfo
rm
[1
5,
16,
17
]
.
The
c
apab
il
it
y
of
t
he
avail
a
ble
VMs
is
cal
cu
la
te
d;
then,
t
he
nu
m
ber
s
of
jo
b
re
quest
s
are
al
locat
ed
to
th
e
resou
rces.
O
nce
th
e
nu
m
ber
s
of
jo
bs
are
al
locat
e
d
to
the
VMs,
the
current
loa
d
is
cal
culat
ed.
If
the
VM
bec
om
es
ov
er
-
or
unde
r
-
loa
de
d,
res
ourc
es are a
uto
m
at
i
cal
ly
ad
de
d or
release
d
a
s
nec
essary. T
he w
hole
pr
ocedur
e
i
s d
esc
ribe
d bel
ow.
Let
V
M=
{VM
1
,V
M
2
,V
M
3
,
…
,V
M
N
}
be
a set
of N v
i
rtual m
achines
and
Task (
nu
m
ber
of jo
b
re
quest
s
)
= {t
as
k1, tas
k2, tas
k3, …,
K
} of K t
ask
to be
regular a
n
d processe
d
i
n V
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
Predict
ion B
ase
d
Eff
ic
ie
nt Re
so
urc
e Pr
ovisi
on
i
ng an
d Its I
mpact
on Qo
S
…
(
La
ta J
Ga
dhavi
)
5367
Her
e
,
the
fitne
ss
of
the
num
ber
of
r
eq
uests
(i)
m
a
y
be
determ
ined
by
the
capaci
ty
of
V
M
j,
wh
ic
h
is
cal
culat
ed by u
sing
(
15):
(15)
j
ob
ty
pe
le
ngth
sh
ows
the
ho
w
m
any
ta
sk
s
are
subm
i
tt
ed
on
VM
j
with
the
cal
culat
ion
of
ca
pacit
y
of
sp
eci
fic
VM to ha
ndle
su
bm
it
te
d
j
obs.
Th
e
capacit
y o
f
the
V
M i
s
m
e
asur
e
d by a
pp
l
yi
ng
(16)
:
(16)
wh
e
re
Nu
m
_pr
oc
is
the
num
ber
of
pr
ocess
or
s
in
th
e
VM
,
N
um
_MIP
S
i
nd
ic
at
es
t
he
num
ber
of
MIP
S,
a
nd
B
W
_VM
re
pre
sents
band
w
idt
h of t
he VM.
(17)
In
(
17),
t
he
input
file
le
ng
th
is
al
so
con
si
der
e
d
to
determ
ine
the
le
ng
th
of
t
he
job
befor
e
e
xecu
ti
on.
Tas
ks
are
assigne
d
t
o
the
v
irt
ual m
achine b
y
us
in
g
t
he m
os
t eff
ect
ive
fitness
value
fr
om
the
(17).
Loa
d
Ca
lc
ulati
on
:
Wh
e
n
ta
sks
are
subm
itte
d
to
the
be
neath
loa
ded
VM,
the
present
work
of
al
l
offe
re
d
V
M
will
be
m
easure
d
by
victi
m
izati
on
of
knowl
edg
e
wh
ic
h
is
receive
d
from
the
database
[
18]
.
T
o
cal
cul
at
e
the
dev
ia
ti
ons i
n
l
oad v
a
riance
on VM
s,
t
he foll
ow
i
ng stan
dard
d
e
viati
on
(SD
)
(
18)
is u
s
ed:
(18)
wh
e
re
X
j
is t
he
pro
ces
sin
g
ti
m
e of the
V
M,
a
s d
esc
ribe
d
i
n
(
19):
(19)
The
n,
t
he
m
ea
n processi
ng ti
m
es f
or all
V
M
s ar
e cal
c
ulate
d by ap
plyi
ng
(
20):
(20)
If
the
S
D
of
th
e
loade
d
VM
is
s
m
al
le
r
than
or
e
qu
al
to
the
m
ean,
then
th
e
syst
e
m
is
in
a
balance
d
sta
te
.
On
the
ot
her
ha
nd,
if
the
SD
is
high
er
than
the
m
e
an,
the
n
the
syst
e
m
is
in
i
m
b
al
ance,
an
d
th
e
auto
-
scal
ing
m
echan
ism
fo
r
resou
rce
pro
visio
ning
will
be
ap
plied.
He
re,
on
e
VM
is
con
si
de
red
to
be
capa
ble
of
serv
i
ng
50
re
quest
s
at
a
ti
m
e
.
T
hus,
w
hen
t
he
num
ber
of
requests
>=
50,
the
n
scal
e
up
ne
w
in
sta
nc
e
an
d
m
igrate
reque
sts
on
new
i
nst
ance.
It
is
r
efr
es
hing
ar
riv
al
of
re
qu
e
st
for
eve
ry
1
m
inu
te
.
Ba
se
d
on
t
he
pr
e
dicte
d
val
ue
,
the
res
ource
s
are
scal
ed
up
or
dow
n
acc
ordin
g
t
o
the
pr
e
dicte
d
dem
and
.
The
pse
ud
o
-
c
od
e
f
or
scal
ing
up a
nd
dow
n
is s
how
n i
n Alg
or
it
hm
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.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5359
-
5370
5368
Algori
th
m
3:
Scaling
up
an
d dow
n
__
____
______
______
______
______
______
Be
gin
.
M
on
it
or
λ
−
1
′
and
λ
′
for
the
pre
vious a
nd t
he
curre
nt ti
m
e,
r
especti
vely
.
Pr
e
dict w
ork
load
λ
+
n
at
ti
m
e
+
n
.
Let
R= r
i
ᴜ r
i+
1
ᴜ
r
i+
2
ᴜ
r
i
+3
.....
ᴜ r
m
befor
e
scali
ng, wh
ere R is
the set
of
virtu
al
resour
ces
.
L
et
λ
+
n
=
∑
λ
′
+
(
n
−
1
)
−
1
=
x
.
Let
λ
+
(
n
−
1
)
′
=
y
.
If
>
y
,
t
he
n
//
the
wor
klo
a
d
i
ncr
ease
s.
A
dd r i
nto t
he
R
(scale u
p
t
he vi
rtual r
e
source
);
el
se if
<
, th
e
n
//
the
work
l
oa
d decr
eases.
Re
m
ov
e extra r
from
R (s
cal
e do
wn the
virt
ual re
source
);
el
se.
Re
fr
es
h
t
he
re
sou
rces.
En
d.
6.
RESU
LT
S
O
F THE Q
oS
METRI
CS
F
OR RES
OUR
CE A
LL
OCA
TION
Fig
ure
6
sho
w
s
the
eff
ic
ie
nt
resou
rce
pro
vi
sion
i
ng
with
t
he
us
e
of
the
pr
e
dicti
on
m
od
el
achieve
s
m
axi
m
u
m
CP
U
util
iz
at
ion
com
par
ed
wit
h
the
existi
ng
(conv
e
ntio
nal)
cl
oud.
Her
e
,
the
aver
a
ge
CPU
util
iz
at
ion
is
ob
ta
ined
for
50
requests
on
the
conven
ti
on
al
cl
oud
an
d
10
0
requests
on
the
eff
ic
ie
nt
cl
oud.
The
resu
lt
s
in
dicat
e
that
the
effi
ci
ent
cl
ou
d
use
s
a
m
ini
m
u
m
nu
m
ber
of
instances
but
prov
i
des
m
a
xim
u
m
util
iz
at
ion
.
Fig
ure
7
s
hows
that
an
e
ff
ic
ie
nt
respo
ns
e
ti
m
e
is
obta
ined
f
or
50
re
quest
s
on
the
c
onve
ntion
al
cl
oud
an
d
10
0
request
s
on
the
eff
ic
ie
nt
cl
oud.
Fig
ure
8
i
nd
i
cat
es
that
eff
ic
ie
nt
resour
ce
pro
visio
ning
wit
h
the
us
e
of
the
pr
e
dicti
on
m
od
el
achieves
the
m
axi
m
u
m
through
pu
t
com
par
ed
with
the
e
xisti
ng
(c
onve
ntion
al
)
cl
oud.
Her
e
,
th
e
aver
a
ge
t
hro
ughput
is
m
easur
e
d
for
50
re
quest
s
on
t
he
c
onve
ntio
nal
cl
oud
an
d
100
requests
on
the
ef
fici
en
t
cl
oud.
T
he
r
esults
s
how
t
ha
t
the
e
ff
ic
ie
nt
cl
oud
use
s
a
m
ini
m
u
m
nu
m
ber
of
i
ns
ta
nc
es
but
pro
vid
es t
he
m
axim
u
m
thr
oughput.
Figure
6
.
A
verage CP
U uti
li
zat
ion
for 50 re
qu
e
sts o
n
t
he
c
onve
ntion
al
and
100 re
qu
est
s
on the e
ff
ic
ie
nt cloud
Figure
7
.
Re
spon
s
e ti
m
e fo
r 5
0 req
uests
on t
he
c
onve
ntio
na
l cl
oud
a
nd 10
0 req
uests
on t
he
e
ff
ic
ie
nt cl
oud
Evaluation Warning : The document was created with Spire.PDF for Python.