Inter
national
J
our
nal
of
Electrical
and
Computer
Engineering
(IJECE)
V
ol.
10,
No.
2,
April
2020,
pp.
1524
1532
ISSN:
2088-8708,
DOI:
10.11591/ijece.v10i2.pp1524-1532
r
1524
P
o
wer
consumption
pr
ediction
in
cloud
data
center
using
machine
lear
ning
Deepika
T,
Prakash
P
Department
of
Computer
Science
and
Engineering
Amrita
School
of
Engineering,
Coimbatore
Amrita
V
ishw
a
V
idyapeetham,
India
Article
Inf
o
Article
history:
Recei
v
ed
Jun
6,
2019
Re
vised
Oct
17,2019
Accepted
Oct
25,
2019
K
eyw
ords:
Cloud
computing
Machine
Learning
Ph
ysical
Machine
Po
wer
consumption
prediction
V
irtual
Machine
ABSTRA
CT
The
flourishing
de
v
elopment
of
the
cloud
computing
paradigm
pro
vides
se
v
eral
ser
-
vices
in
the
industrial
b
usiness
w
orld.
Po
wer
consumption
by
cloud
data
ce
nters
is
one
of
the
crucial
issues
for
service
pro
viders
in
the
domain
of
cloud
computing.
Pur
-
suant
to
the
rapid
technology
enhancements
in
cloud
en
vironme
nts
and
data
centers
augmentations,
po
wer
utilization
in
data
centers
is
e
xpected
to
gro
w
unabated.
A
di-
v
erse
set
of
numerous
connected
de
vices,
eng
aged
with
the
ubiquitous
cloud,
results
in
unprecedented
po
wer
utilization
by
the
data
centers,
accompanied
by
increa
sed
car
-
bon
footprints.
Nearly
a
million
ph
ysical
machines
(PM)
are
running
all
o
v
er
the
data
centers,
along
with
(5
–
6)
million
virtual
machines
(VM).
In
the
ne
xt
fi
v
e
years,
the
po
wer
needs
of
this
domain
are
e
xpected
to
spiral
up
to
5%
of
global
po
wer
produc-
tion.
The
virtual
machine
po
wer
consumption
reduction
impacts
t
he
diminishing
of
the
PM’
s
po
wer
,
ho
we
v
er
further
changing
in
po
wer
consumption
of
data
center
year
by
year
,
to
aid
the
cloud
v
endors
using
prediction
methods.
The
sudden
fluctuation
in
po
wer
utilization
will
cause
po
wer
outage
in
the
cloud
data
centers.
This
paper
aims
to
forecast
the
VM
po
wer
consumption
with
the
help
of
re
gressi
v
e
predicti
v
e
analysis,
one
of
the
Machine
Learning
(ML)
techniques.
The
potenc
y
of
this
approach
to
mak
e
better
predictions
of
future
v
alue,
using
Multi-layer
Perceptron
(MLP)
re
gressor
which
pro
vides
91%
of
accurac
y
during
the
prediction
process.
Copyright
c
2020
Institute
of
Advanced
Engineering
and
Science
.
All
rights
r
eserved.
Corresponding
A
uthor:
Deepika
T
.,
Department
of
Computer
Science
and
Engineering,
Amrita
School
of
Engineering,
Coimbatore,
Amrita
V
ishw
a
V
idyapeetham,
India.
Email:
t
deepika@cb
.students.amrita.edu
1.
INTR
ODUCTION
Cloud
computing
is
a
technological
adv
ancement
that
furnishing
with
e
v
erything
as
a
service
such
as
storage
space
to
the
user
,
netw
orking,
serv
er
as
well
as
applications.
Infrastructure
as
a
Service(IaaS),
Softw
are
as
a
Service(SaaS),
and
Platform
as
a
Service(P
aaS)
are
the
dif
ferent
types
of
service
models,
in
Cloud
computing
that
can
be
deli
v
ered
on
demand.
Cloud
pro
viders
of
fer
a
pool
of
virtualized
computa-
tional
resources
to
customers
in
the
data
center
,
in
a
pay-as-you-go
manner
[1].
The
virtualized
computing
services
pro
vide
IaaS
that
helps
reduce
the
installation
and
maintenance
cost
for
computing
en
vironments.
A
cloud
data
center
is
associated
with
a
group
of
connected
ph
ysical
machines
(PM)
or
host
used
by
the
or
g
anizations
for
netw
ork
processing,
remote
storage
and
access
to
enormous
data.
The
data
centers
are
the
backbone
for
the
cloud
en
vironments.
The
e
xponential
gro
wth
of
cloud
computing,
because
of
emer
ging
tech-
nologies
lik
e
IIO
T
(Industrial
Internet
of
Things)
appli
cations,
big
data
e
v
oluti
o
n,
and
5G
functional
ity
.
In
2020,
J
ournal
homepage:
http://ijece
.iaescor
e
.com/inde
x.php/IJECE
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
r
1525
50
billion
connected
de
vices
will
be
in
the
Internet
of
Things
(IoT)
field,
the
amount
of
internet
traf
fic
as
per
second
is
51974
GB
[2].
Consequently
,
the
cloud
service
pro
viders
lik
e
A
WS,
Google
cloud
and
Azure
are
moti
v
ated
to
e
xtend
data
centers
across
the
globe
to
pro
vide
on-demand
services.
The
virtualization
technique
plays
a
major
role
in
the
data
centers
-
f
acilitate
sharing
resources
am
ong
customers
through
VMs.
Each
virtual
machine
is
isolated
and
used
to
e
x
ecute
customer
applications
with
the
follo
wing
requirements
including
its
storage
capacity
,
main
memory
,
CPU,
I/O
capabilities
and
netw
ork
bandwidth
[3,
4].
Consolidation
of
ph
ysical
machines,
f
ault
t
olerance,
and
load
balancing
are
some
of
the
k
e
y
f
actors
that
impro
v
e
cloud
computing
performance.
The
PM
consolidation
occurs
through
V
irtual
Ma-
chine
(VM)
migration,
when
the
una
v
ailability
of
the
requested
resources
by
virtual
machine
from
the
ph
ysical
machine,
relocation
of
the
virtual
machine
will
tak
e
place.
The
VM
is
relocated
to
another
ph
ysical
machine,
fulfill
the
need
for
VM
[5].
The
proposed
method
forecast
the
po
wer
of
each
VM
preli
minary
to
VM
migra-
tion,
based
on
this
prediction
and
resource
a
v
ailability
of
PM,
then
VM
migrated
to
particular
PM.
The
VM
po
wer
prediction
escalates
the
system
a
v
ailability
,
m
inimizes
the
infrastructure
comple
xity
,
and
reduces
the
operational
cost
for
cloud
pro
viders
which
helps
the
customer
to
pay
less
amount
[6,
7].
There
is
a
need
to
forecast
the
VM’
s
po
wer
in
adv
ance
to
manipulate
the
processes
f
astest
and
pro
vide
more
reliable
services
to
customers
.
The
conserv
ati
on
of
po
wer
can
be
accomplished
t
hrough
po
wer
forecas
t
by
appl
y
i
ng
v
arious
machine
learning
methods.
In
this
w
ork,
the
re
gression-based
ML
strate
gy
is
applied
to
forecast
the
po
wer
consumption
of
the
virtual
machine,
to
enrich
the
cloud
computing
infrastructure
and
to
enhance
service
for
IT
industries.
Moreo
v
er
,
the
po
wer
utilization
of
VM’
s
is
predicted
before
the
VM
allotted
to
ph
ysical
machines.
The
outline
of
the
paper’
s
structure
follo
ws.
Section
2
re
vie
ws
past
literature
w
ork
on
w
orkload
forecast
of
VM,
resource
management
allocation
based
on
v
arious
characteristics
of
VM.
Section
3
deals
with
the
frame
w
ork
for
the
VM
po
wer
prediction
based
on
re
gression-based
methods.
Section
4
illustrates
the
ML
models
and
performance
e
v
aluation
of
the
proposed
approach,
through
empirical
i
n
s
pection,
follo
wed
by
closing
remarks
in
Section
5
as
a
conclusion.
2.
RELA
TED
W
ORK
The
background
research
kno
wledge
in
the
cloud’
s
virtual
machine
such
as
forecasting
of
CPU
utilization,
resource
usage,
and
management
is
the
ef
fecti
v
e
approaches
to
w
ards
the
future
in
adv
ance.
The
po
wer
supply
increases
day
by
day
,
to
run
and
cool
do
wn
the
utilized
de
vices
in
the
cloud
data
center
and
these
phenomena
increase
the
operational
e
xpenses
of
cloud
service
pro
viders.
The
conserv
ation
of
po
wer
by
the
data
center
,
the
v
arious
po
wer
a
w
are
methodologies
were
studied.
Prediction
of
po
wer
consumption
is
used
to
estim
ate
the
non-linear
future
v
alue
for
better
performance
of
a
comple
x
function.
Beloglazo
v
and
Buyya
[8]
ha
v
e
applied
an
Adapti
v
e
Threshold
algorithm,
Local
Re
gression,
and
Rob
ust
Local
Re
gression
to
e
v
aluate
o
v
erloaded
serv
er
,
based
on
CPU
utilization
in
IaaS
infrastructure.
The
threshold
is
adjusted
automati-
cally
based
on
historical
analysis
of
data,
manipulate
with
estimator
lik
e
Mean
absolute
de
viation,
interquartile
range.
Pre
v
ost
et
al.,
[9]
focused
on
netw
ork
load
prediction
using
Autore
gressi
v
e
linear
prediction
and
neural
netw
ork.
The
data
samples
used
in
this
method
w
as
less
for
training
to
learn
the
relati
o
ns
hip
between
attrib
utes.
Chonglin
et
al.,
[10]
presented
a
T
ree
Re
gression(TR)-based
model
to
compute
the
VM
po
wer
utilization,
using
cross-v
alidation,
based
on
black
box
method.
The
VM
and
serv
er
feature
information
are
g
athered
based
on
black
box
method.
The
y
ha
v
e
considered
data
as
linear
v
alues
for
their
predi
ction
model.
Jingqi
et
al.,
[11]
presented
the
Linear
Re
gression
method
to
forecast
the
w
orkload
of
cloud
services.
The
y
also
performed
the
autoscaling
process
reducing
the
operational
cost
of
virtual
resources
through
v
ertical
and
horizontal
scaling.
Jitendra
et
al.,
[12]
proposed
a
self-adapti
v
e
dif
ferenti
al
e
v
olution
algorithm
to
estimate
the
w
orkload
utilized
by
the
cloud
data
center
using
N
ASA
trace
and
Saskatche
w
an
trace.
The
authors
re
vie
wed
fitness
function,
mutation,
and
crosso
v
er
carried
out
in
this
method,
which
w
as
better
than
other
approaches
lik
e
P
article
Sw
arm
Optimization
(PSO),
Genetic
Algorithm
(GA)
and
so
on.
This
method
required
to
minimize
the
Service
Le
v
el
Agreement
(SLA)
violations
for
better
service
processing.
Fla
vien
et
al.,
[13]
e
xplored
the
challenges,
in
a
cloud
en
vironment,
to
diminish
the
po
wer
consump-
tion
of
VMs
and
the
operational
e
xpenses
for
cloud
v
endor
.
The
y
implemented
the
ad-hoc
frame
w
ork
for
VM
consolidation;
b
ut
this
approach
did
not
tak
e
into
acc
o
unt
VM
requirements
lik
e
disk
space,
netw
ork
band-
width
and
time
tak
en
by
VM
to
complete
a
particular
task.
Hao
Xu
et
al.,
[14]
in
v
estig
ated
the
po
wer
of
VM
with
normalized
parameters
that
satisfy
the
correlation
coef
ficient
of
VM’
s
po
wer
using
Radial
Basis
Function
(RBF)
Neural
Netw
ork.
This
method
used
a
small
number
of
samples
for
training
and
testing
data,
which
P
ower
consumption
pr
ediction
in
cloud
data...
(Deepika
T)
Evaluation Warning : The document was created with Spire.PDF for Python.
1526
r
ISSN:
2088-8708
could
not
get
an
accurate
prediction
in
the
neural
netw
ork.
The
estimator
used
for
calculating
prediction
error
w
as
a
v
erage
prediction
and
maximum
prediction
error
.
Minal
P
atel
et
al.,
[15]
proposed
the
Support
V
ector
Re
gression
(SVR)
and
Autore
gressi
v
e
Inte
grated
Mo
ving
A
v
erage
(ARIMA)
method
to
predict
the
dirty
pages
of
VM
during
li
v
e
migration
and
determ
ine
the
migration
time
of
VM
depend
on
time
series
analysis.
The
ARIMA
model
is
applied
to
reduce
the
dirty
pages,
netw
ork
traf
fic,
and
memory
size
based
on
past
statistical
data.
This
approach
has
less
capability
to
ascertain
the
b
uilt-in
features
because
it
formed
with
single
hidden
layer
as
shallo
w
neural
netw
ork
structure.
Cortez
et
al.,
[16]
applied
ML
algorithms
to
estimate
the
resource
management
of
VM,
in
the
cloud
platform
using
the
characteristics
of
Azure
w
orkload
such
as
the
first
party
for
IaaS
and
third
party
for
P
aaS
services.
The
authors
e
xploited
the
F
ast
F
ourier
T
ransform
to
find
the
cate
gory
of
VM
w
orkload
and
plotted
the
graph
for
CPU,
memory
,
CPU
core
usage
per
VM
and
lifetime
of
VM,
using
cumulati
v
e
distrib
ution
function.
This
method
used
the
dynamically
link
ed
library(DLL)
to
accumulate
the
result
after
each
prediction,
while
in
the
ne
xt
predic
tion,
it
checks
whether
the
forecast
w
as
v
aluable
using
the
score
of
the
DLL.
V
erma
et
al.,
[17]
analyzed
the
w
orkload
of
VM
in
order
to
minimize
the
po
wer
consumption
of
VM
using
supervised
learning
algorithms.
The
y
listed
the
v
arious
scheduling
approaches
to
reduce
carbon
dioxide
emissions
from
data
centers.
The
statistical
metrics
such
as
RMSE,
R
squared
and
accurac
y
accomplished
with
an
algorithm
to
calculate
the
prediction
error
.
Chang
et
al.,
[18]
applied
the
recurrent
neural
netw
ork
to
forecast
and
manage
the
resource
allocation
to
a
cloud
serv
er
.
The
y
compared
the
serv
ers
w
orkload
prediction
results
with
T
ime-Delay
Neural
Netw
ork(TDNN)
and
Re
gression
methods.
W
itanto
et
al.,
[19]
proposed
the
adap-
ti
v
e
selector
neural
netw
ork
to
select
the
algorithm
for
reduction
of
the
acti
v
e
VM
and
compared
the
results
with
Linear
re
gression.
This
method
w
as
also
focused
on
Service
le
v
el
agreement(SLA)
between
customer
and
cloud
service
pro
vider
b
ut
still,
SLA
is
not
fulfilled
when
the
customer
requirements
v
ary
.
The
abo
v
e-
mentioned
literature
outline
e
xhibits
the
potential
of
machine
learning
to
predict
v
arious
problems
in
cloud
computing
for
future
e
v
aluation.
The
aforementioned
related
w
orks
are
tab
ulated
belo
w
in
T
able
1.
T
able
1.
Comparison
of
algorithms
for
VM
resource
requirements
prediction
A
uthor(s)
Method
Goal
W
eakness
P
erf
ormance
better
than
John
J.
Pre
v
ost
et
al.,(2011)
Auto
Re
gressi
v
e
Linear
prediction
Netw
ork
load
prediction
Need
for
e
xtension
to
multiple
resources
-
Beloglazo
v
et
al.,
(2012)
Adapti
v
e
Threshold
algorithm
Local
Re
gression,
Rob
ust
Local
Re
gression
Predict
o
v
erloaded
serv
er
based
on
CPU
utilization
Multiple
migration
not
discussed
Heuristic
algorithm
Jingqi
Y
ang
et
al.,(2014)
Linear
Re
gression
W
orkload
prediction
of
service
cloud
Netw
ork
load
is
not
considered
Hidden
Mark
o
v
process
Chang
et
al.,
(2014)
Neural
netw
ork
Resource
allocation
SLA
violation
TDNN
and
Re
gression
method
Chonglin
et
al.,
(2015)
TR-based
method
Compute
VM
po
wer
Examine
only
Linear
v
alues
Linear
Re
gression
Re
gression
tree
Hao
Xu
et
al.,
(2016)
RBF
Neural
Netw
ork
VM
po
wer
prediction
Considered
less
VM
samples
Short
term
prediction
models
Minal
P
atel
et
al.,(2016)
ARIMA
Dirty
pages
prediction
in
VM
Slo
w
e
x
ecution
Support
v
ector
re
gression
V
erma
et
al.,(2017)
Supervised
learning
methods
F
orecast
the
VM’
s
w
orkload
Operational
time
of
VM
and
CPU
usage
not
tak
en
into
account
Gaussian
process,
Ridge
Re
gression
and
so
on
Cortez
et
al.,
(2017)
Gradient
boosting
tree,
Random
F
orest
Resource
management
Resource
e
xhaustion
-
Jitendra
et
al.,
(2018)
Self-adapti
v
e
dif
ferential
e
v
olution
algorithm
W
orkload
prediction
Impro
v
e
the
SLA
for
better
prediction
P
article
Sw
arm
Optimization,
Genetic
algorithm
W
itanto
et
al.,(2018)
Adapti
v
e
selector
Neural
Netw
ork
Resource
management
SLA
v
aries
with
dif
ferent
QOS
requirements
Local
Re
gression
Int
J
Elec
&
Comp
Eng,
V
ol.
10,
No.
2,
April
2020
:
1524
–
1532
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
r
1527
3.
SYSTEM
MODEL
The
aspects
of
the
proposed
method
to
forecast
the
po
wer
utilization
of
VM
in
a
proacti
v
e
manner
.
Figure
1
sho
ws
an
o
v
erall
frame
w
ork
for
the
proposed
system
which
focuses
on
the
prediction
of
VM
po
wer
utilization.
The
proposed
frame
w
ork
is
comprised
of
dif
ferent
components,
which
includes
cloud
information
service
module,
resource
pro
visioner
module,
machine
learning
module,
and
decision-making
module.
A
mul-
tiple
VM
request
from
the
customer
is
re
gistered
in
cloud
information
service
module
to
deplo
y
their
system
and
application.
The
resource
pro
visioner
module
allocates
the
resources
to
the
virtual
machines
based
on
the
decision
of
cloud
manager
,
whene
v
er
needed.
This
module
is
responsible
for
satisfying
the
service
request
for
customers
according
to
the
service
le
v
el
agreement(SLA).
The
ML
module
inspects
the
repository
of
VMs
his-
torical
data
and
then
selects
the
data
for
training
and
testing
phase
.
These
are
retrie
v
ed
by
the
decision-making
module
for
po
wer
prediction.
The
cloud
management
monitors
the
other
modules
and
tak
es
the
decision
in
the
appropriate
situation.
The
cloud
data
center
consists
of
connected
hosts
in
which
each
host
is
allocated
with
multiple
VMs.
The
virtual
machine
monitor
(VMM)
is
a
layer
which
controls
each
virtual
machine
located
in
the
ph
ysical
machines.
The
VMM
recei
v
es
the
result
from
the
cloud
manager
and
allocates
the
VM
to
the
preferable
PM.
The
une
xpected
creation
of
virtual
machine
instance
in
the
ph
ysical
host
or
assignment
of
a
task
to
e
xisting
VM,
ensue
in
changes
of
VM
attrib
utes;
consequent
fluctuations
in
po
wer
consumption
occur
in
the
corresponding
ph
ysical
host.
In
this
scenario,
the
po
wer
anomalies
can
be
re
gulated,
through
prediction,
at
an
y
point
from
the
historical
data,
before
the
change
in
po
wer
consumption.
Figure
1.
Frame
w
ork
of
the
proposed
system
4.
PO
WER
PREDICTION
OF
VM
BY
APPL
YING
MA
CHINE
LEARNING
4.1.
F
or
ecasting
methods
The
machine
learning
models
are
used
to
learn
the
features
of
the
dataset
in
a
fla
wless
manner
,
to
forecast
the
VM
metric
s
lik
e
CPU,
memory
,
and
po
wer
.
The
comple
x
correla
tion
between
the
input
v
ariables
can
be
handled
by
ef
fecti
v
e
learning
algorithms
among
the
massi
v
e
amount
of
traced
data
contains
the
numer
-
ous
VM.
The
dataset
can
be
handled
with
normalization,
feature
selection
and
find
the
relationship
among
features,
through
correlation
method.
The
supervised
machine
learning
algorithms
will
predict
the
tar
get
v
ari-
ables
based
on
input
a
n
d
output
v
ariables
[20,
21].
The
ra
w
dataset
contains
the
tar
get
v
ariable
as
a
continuous
v
alue;
so,
it
comes
under
the
cate
gory
of
Re
gression
predicti
v
e
model.
The
Re
gression
m
o
de
l
is
used
to
predict
the
response
v
ariable
from
analyzing
the
relationship
between
multiple
independent
v
ariables
and
one
depen-
dent
v
ariable
[22].
The
re
gre
ssion
model
can
be
assessed
through
the
root
mean
square
error
using
the
formula
noted
belo
w
R
M
S
E
=
v
u
u
t
1
n
n
X
j
=1
(
y
j
b
y
j
)
2
(1)
P
ower
consumption
pr
ediction
in
cloud
data...
(Deepika
T)
Evaluation Warning : The document was created with Spire.PDF for Python.
1528
r
ISSN:
2088-8708
where
’n’
is
the
number
of
observ
ations
in
the
ra
w
dataset,
’
y
j
’
is
the
forecasted
v
alue
and
’
b
y
j
’
is
the
actual
v
alue
of
t
h
e
observ
ation.
The
v
arious
re
gress
ion
methods
ha
v
e
been
trained
and
tes
ted
on
the
dataset,
to
generate
the
RMSE
v
alue,
using
the
prediction
method
between
the
test
data
and
re
gression
models.
F
or
both
training
and
testing
data
bas
ed
on
the
metrics,
pro
vide
the
output
as
a
score
of
the
model,
prediction
error
,
running
time,
performance
consistenc
y
and
so
on.
The
best
model
is
selected
based
on
the
score
for
prediction
of
attrib
utes
such
as
VM’
s
CPU(MHZ)
usage,
Memory(GBs)
usage,
and
Po
wer
consumption.
The
follo
wing
re
gression
methods
were
used
in
the
prediction
process.
4.2.
Regr
ession
types
The
shrinkage
algorithms
lik
e
Least
Absolute
Shrinkage
and
selection
operator
re
gression,
Ridge
re
gression
are
ef
fecti
v
e
for
multicollinearity
problem.
The
v
ariables
in
the
dataset
are
highly
correlated
with
each
ot
her
that
results
in
poor
prediction,
can
be
o
v
ercome
by
these
algorithms.
The
Elastic
Net
is
the
h
ybrid
of
Lasso
and
Rigid
met
hods.
The
aforementioned
algorithms
are
re
gularisation
techniques
to
a
v
oid
the
o
v
erfitting
of
data
[23].
The
result
of
the
Lasso,
Ridge,
and
Elastic
Net
Re
gression
are
compared
with
other
re
gression
methods.
4.3.
Random
f
or
est
r
egr
essor
Random
F
orest
is
one
of
the
ensemble
ML
algorithms
used
for
re
gression
analysis.
It
uses
the
bag-
ging
technique
and
selects
the
features
for
best
node
splitting
and
to
construct
the
multiple
decision
trees
subsequently
a
v
eraging
the
v
alue
of
all
decision
tree
to
predict
the
accurac
y
[24].
This
approach
will
learn
ho
w
to
predict
the
f
u
t
ure
v
alue
with
the
help
of
cur
rently
observ
ed
v
alue.
The
RMSE
metric
is
used
to
calculate
the
dif
ference
between
the
real
observ
ed
v
alue
and
forecasted
v
alue
by
the
re
gressor
.
4.4.
K
near
est
neighbor
(KNN)
r
egr
ession
The
K
nearest
neighbor
forecasts
the
po
wer
utilized
by
each
VM,
and
based
on
the
feature
simi
larity
,
it
collects
the
a
v
erage
of
the
training
test.
The
distance
metric,
Euclidean
distance
defines
the
distance
between
the
ne
w
v
alue
and
training
v
alue,
using
the
formula
E
ucl
idean
distance
=
v
u
u
t
k
X
i
=1
(
x
i
y
i
)
2
(2)
F
or
tuning
the
h
yperparameter
’k’,
KNN
uses
the
k-fold
cross-v
alidation
to
choose
the
right
v
alue
of
k,
and
sum
up
all
losses
of
each
’k’
v
alue,
to
estimate
the
score
of
the
algorithm.
The
cost
function
of
’k’
drops
in
some
period
of
time,
and
ag
ain
increase
it
further
whilst
find
the
’k’
v
alue
using
the
elbo
w
method.
Figure
2
depicts
the
v
alue
of
RMSE
decreases
while
increasing
the
‘k’
v
alue.
The
optimum
v
alue
of
’k’
is
determined
through
parameter
tuning
to
achie
v
e
a
better
score
of
the
KNN
re
gressor
algorithm.
Figure
2.
Optimum
v
alue
of
k
Int
J
Elec
&
Comp
Eng,
V
ol.
10,
No.
2,
April
2020
:
1524
–
1532
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
r
1529
4.5.
Multi-lay
er
per
ceptr
on
(MLP)
r
egr
essor
The
MLP
is
the
pre
cursor
for
an
artificial
neural
netw
ork.
An
MLP
Re
gressor
uses
back
propag
a
tion
to
train
the
data
based
on
the
perceptron,
which
consist
of
an
input
layer
,
hidden
layer
,
and
the
output
layer
.
The
neurons
composed,
in
each
of
the
layer
and
hidden
layer
with
acti
v
ation
function
to
produce
output
for
a
gi
v
en
input
node
or
neuron;
in
this
model,
it
uses
the
Relu
acti
v
ation
function
within
the
hidden
layer
.
The
performance
of
MLP
Re
gressor
impro
v
ed
high
while
compared
to
other
models.
4.6.
EXPERIMENTS
AND
RESUL
TS
4.6.1.
VM
po
wer
The
utilization
of
the
po
wer
in
a
virtual
machine
can
be
computed
with
the
po
wer
consumption
VM’
s
CPU,
VM’
s
memory
,
VM’
s
IO,
and
so
on
[25].
The
equation
for
VM
po
wer
calculation
is
defined
as
belo
w
,
P
V
M
=
P
C
P
U
U
til
iz
e
+
P
M
emor
y
+
P
I
O
(3)
where
P
V
M
is
the
amount
of
po
wer
consumed
by
VM,
P
C
P
U
U
til
iz
e
,
P
M
emor
y
,
P
I
O
are
the
po
wer
consumption
of
VM
requirements
such
as
CPU,
memory
and
IO
respecti
v
ely
.
4.6.2.
Data
model
The
ra
w
dataset,
col
lected
from
Azure
VM
w
orkload,
cont
ains
VM
requirement
details
lik
e
CPU
utilization
for
minimum
and
maximum
usage,
memory
space,
CPU
core,
and
VM
lifetime
and
so
on,
a
v
ailable
in
Github
[16].
Ov
er
ten
lakhs
of
VM’
s
are
monitored,
and
collected
data
from
each
VM,
for
24
hours
per
day
,
for
four
months
continuously
.
Ev
ery
VM’
s
detail
relates
to
fi
v
e-minute
VM
CPU
utilization
readings
and
other
features.
The
task/data-dri
v
en
model
uses
this
dataset
to
forecast
the
po
wer
consumption
of
VM,
with
the
help
of
error
estimators.
4.6.3.
P
erf
ormance
e
v
aluation
The
input
to
the
dif
ferent
re
gression
algorithms
i
s
ra
w
dataset
di
vided
as
80%
for
training
and
20%
for
the
testing
set.
The
performance
of
each
method
is
v
alidated.
The
e
v
aluation
will
be
ho
w
f
ar
into
the
future
prediction
v
alue.
The
proposed
method
depicts
the
performance
of
machine
le
arning
models
along
with
the
accurac
y
of
testing
and
training
data.
4.6.4.
Result
and
analysis
The
dataset
with
a
collection
of
dif
ferent
services
for
VM
types,
based
on
their
CPU,
and
memory
usage,
the
better
forecasting
me
thod
w
as
chosen,
based
on
the
score
v
alue
of
each
machine
learning
algorithm.
Figure
3,4
depicts
the
amount
of
CPU
and
memory
utilization
per
service.
Figure
5
illustrates
the
plotting
of
learning
curv
e
to
e
xhibit
the
predict
performance
through
v
alidated
error
and
training
error
for
ran-
dom
forest
re
gressor
analysis.
The
prediction
error
and
score
of
the
results
are
used
in
v
estig
ating
the
o
v
erall
performance
of
the
algorithms.
Figure
6,7,8,and
9
represents
the
ef
fect
of
dif
ferent
machine
learning
models
on
the
actual
v
alue,
and
predicted
the
v
alue,
of
the
random
VM’
s
po
wer
.
Figure
10
sho
ws
the
better
performance
of
the
MLP
re
gressor
model
correctly
predicts
the
actual
v
alue
for
each
record
of
VM.
This
profound
approach
observ
es
the
past
VM
data,
and
e
v
aluates
the
upcoming
service,
to
achie
v
e
the
goal
of
the
prediction
process.
Figure
3.
CPU
utilization
Figure
4.
Memory
usage
P
ower
consumption
pr
ediction
in
cloud
data...
(Deepika
T)
Evaluation Warning : The document was created with Spire.PDF for Python.
1530
r
ISSN:
2088-8708
Figure
5.
Random
forest
learning
curv
e
Figure
6.
Lasso,
rigid
and
elastic
net
re
gression
Figure
7.
KNN
re
gressor
Figure
8.
Random
forest
re
gression
Figure
9.
MLP
re
gressor
Figure
10.
Comparison
of
all
algorithms
5.
CONCLUSION
In
this
paper
,
proacti
v
e
methods
can
forecast
the
sudden
fluctuation
in
po
wer
consumption,
due
to
the
changes
in
VM
attrib
utes,
ahead
of
time,
happens
through
histori
cal
performance
data
of
a
lar
ge
number
of
VM’
s.
The
Re
gression
based
dif
ferent
machine
learning
algorithms
were
tested
in
the
historical
dataset
to
predict
the
VM
po
wer
consumption.
The
MLP
re
gressor
model
estimates
the
actual
po
wer
v
al
ue
of
VM
to
o
v
ercome
the
future
uncertainty
in
po
wer
consumption
of
VM
and
the
proposed
frame
w
ork
has
a
potenc
y
to
proceed
the
cloud
manager
as
proacti
v
e
to
forecast
the
future
po
wer
consumption
of
VM
for
ef
ficient
po
wer
management
in
the
cloud
data
center
.
This
scenario
in
cloud
computing
technic
implies
to
pro
vide
a
reliable
en
vironment
to
customers.
The
future
enhancement
will
allo
w
the
data
center
to
understand
the
characteristics
of
VM
in
adv
ance
with
better
prediction
and
VM
migration
model
which
leads
to
po
wer
consumption.
Int
J
Elec
&
Comp
Eng,
V
ol.
10,
No.
2,
April
2020
:
1524
–
1532
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Elec
&
Comp
Eng
ISSN:
2088-8708
r
1531
REFERENCES
[1]
Ismaeel,
R.
Karim,
and
A.
Miri,
”Proacti
v
e
dynamic
virtual-machine
consolidation
for
ener
gy
conserv
a-
tion
in
cloud
data
centres,
”
Journal
of
Cloud
Computing
,
v
ol.
7,
no.
1,
p.
10,
2018.
[2]
K.
Mason,
M.
Dugg
an,
E.
Barrett,
J.
Dugg
an,
and
E.
Ho
wle
y
,
”Predicting
host
cpu
utilization
in
the
cloud
using
e
v
olutionary
neural
netw
orks,
”
Future
Generation
Computer
Systems
,
v
ol.
86,
pp.
162-173,
2018.
[3]
P
.
Prakash,
G.
K
ousalya,
S.
K.
V
asude
v
an,
and
K.
K.
Rang
araju,
”Distrib
uti
v
e
po
wer
migration
and
man-
agement
algorithm
for
cloud
en
vironment,
”
Journal
of
Computer
Science
,
v
ol.
10,
no.
3,
p.
484,
2014.
[4]
F
.
Zhang,
G.
Liu,
X.
Fu,
and
R.
Y
ah
yapour
,
”A
surv
e
y
on
virtual
machine
migration:
Challenges,
tech-
niques,
and
open
issues,
”
IEEE
Communications
Surv
e
ys
and
T
utorials
,
v
ol.
20,
no.
2,
pp.
1206-1243,
2018.
[5]
N.
Janani,
R.
S.
Je
g
an,
and
P
.
Prakash,
”Optimization
of
virtual
machine
placement
in
cloud
en
vironment
using
genetic
algorithm,
”
Research
Journal
of
Applied
Sciences,
Engineering
and
T
echnology
,
v
ol.
10,
no.
3,
pp.
274-287,
2015.
[6]
Z.
Usmani
and
S.
Singh,
”A
surv
e
y
of
virtual
machi
ne
placement
techniques
in
a
cloud
data
center
,
”
Procedia
Computer
Science
,
v
ol.
78,
pp.
491-498,
2016.
[7]
H.
Zhao,
J.
W
ang,
F
.
Liu,
Q.
W
ang,
W
.
Zhang,
and
Q.
Zheng,
”Po
wer
-a
w
are
and
performance-guaranteed
virtual
machine
placement
in
the
cloud,
”
IEEE
T
ransactions
on
P
arallel
and
Distrib
uted
Systems
,
v
ol.
29,
no.
6,
pp.
1385-1400,
2018.
[8]
A.
Beloglazo
v
and
R.
Buyya,
”Optimal
online
deterministic
algorithms
and
adapti
v
e
heuristics
for
ener
gy
and
performance
ef
ficient
dynamic
consolidation
of
vir
tual
machines
in
cloud
data
centers,
”
Concurrenc
y
and
Computation:Practice
and
Experience
,
v
ol.
24,
no.
13,
pp.
1397-1420,
2012.
[9]
J.
J.
Pre
v
ost,
K.
Nagothu,
B.
K
elle
y
,
and
M.
Jamshidi,
”Prediction
of
cloud
data
center
netw
orks
loads
using
stochastic
and
neural
models,
”
in
2011
6th
International
Conference
on
System
of
Systems
Engi-
neering
,
pp.
276-281,
IEEE,
2011.
[10]
C.
Gu,
P
.
Shi,
S.
Shi,
H.
Huang,
and
X.
Jia,
”A
tree
re
gression-based
approach
for
vm
po
wer
metering,
”
IEEE
Access
,
v
ol.
3,
pp.
610-621,
2015.
[11]
J.
Y
ang,
C.
Liu,
Y
.
Shang,
B.
Cheng,
Z.
Mao,
C.
Liu,
L.
Niu,
and
J.
Chen,
”A
cost-a
w
are
auto-scaling
approach
using
the
w
orkload
prediction
i
n
service
clouds,
”
Information
Systems
Frontiers
,
v
ol.
16,
no.
1,
pp.
7-18,
2014.
[12]
J.
K
umar
and
A.
K.
Singh,
”W
orkload
prediction
in
cloud
using
artificial
neural
netw
ork
and
adapti
v
e
dif
ferential
e
v
olution,
”
Future
Generation
Computer
Systems
,
v
ol.
81,
pp.
41-52,
2018.
[13]
F
.
Quesnel,
H.
K.
Mehta,
and
J.-M.
Menaud,
”Estimating
the
po
wer
consumption
of
an
idle
virtual
ma-
chine,
”
in
2013
IEEE
International
Conference
on
Green
Computing
and
Communications
and
IEEE
Internet
of
Things
and
IEEE
Cyber
,
Ph
ysical
and
Social
Computing
,
pp.
268-275,
IEEE,
2013.
[14]
H.
Xu,
X.
Zuo,
C.
Liu,
and
X.
Zhao,
”Predicting
virtual
machine’
s
po
wer
via
a
rbf
neural
n
e
tw
ork,
”
in
International
Conference
on
Sw
arm
Intelligence
,
pp.
370-381,
Springer
,
2016.
[15]
M.
P
atel,
S.
Chaudhary
,
and
S.
Gar
g,
”Machine
learning
based
statistical
prediction
model
for
impro
ving
performance
of
li
v
e
virtual
machine
migration,
”
Journal
of
Engineering
,
v
ol.
2016,
2016.
[16]
E.
Cortez,
A.
Bonde,
A.
Muzio,
M.
Russino
vich,
M.
F
ontoura,
and
R.
Bianchini,
”Resource
central:
Understanding
and
predicting
w
orkloads
for
impro
v
ed
resource
management
in
lar
ge
cloud
platforms,
”
in
Proceedings
of
the
26th
Symposium
on
Operating
Systems
Principles
,
pp.
153-167,
A
CM,
2017.
[17]
N.
V
erma
and
A.
Sharma,
”W
orkload
prediction
model
based
on
supervised
learning
for
ener
gy
ef
ficienc
y
in
cloud,
”
in
2017
2nd
International
Conference
on
Communication
Systems,
Computing
and
IT
Appli-
cations
(CSCIT
A)
,
pp.
66-71,
IEEE,
2017.
[18]
Y
.-C.
Chang,
R.-S.
Chang,
and
F
.-W
.
Chuang,
”A
predicti
v
e
method
for
w
orkload
forecasting
in
the
cloud
en
vironment,
”
in
Adv
anced
T
echnologies,
Embedded
and
Multimedia
for
Human-Centric
Computing
,
pp.
577-585,
Springer
,
2014.
[19]
J.
N.
W
itanto,
H.
Lim,
and
M.
Atiquzzaman,
”Adapti
v
e
selection
of
dynamic
vm
consolidation
algorithm
using
neural
netw
ork
for
cloud
resource
management,
”
Future
Generation
Computer
Systems
,
v
ol.
87,
pp.
35-42,
2018.
[20]
C.
Sammut
and
G.
I.
W
ebb,
Enc
yclopedia
of
machine
learning
and
data
mining
,
Springer
,
2017.
[21]
H.
Brink,
J.
Richards,
and
M.
Fetherolf,
Real-w
orld
machine
learning
,
Manning
Publications
Co.,
2016.
[22]
L.
Breiman,
Classification
and
re
gression
trees.
Routledge
,
2017.
[23]
G.
James,
D.
W
itten,
T
.
Hastie,
and
R.
T
ibshi
rani,
”Linear
model
selection
and
re
gularization,
”
in
An
P
ower
consumption
pr
ediction
in
cloud
data...
(Deepika
T)
Evaluation Warning : The document was created with Spire.PDF for Python.
1532
r
ISSN:
2088-8708
introduction
to
statistical
learning
,
pp.
203-264,
Springer
,
2013.
[24]
J.
Chen,
K.
Li,
Z.
T
ang,
K.
Bilal
,
S.
Y
u,
C.
W
eng,
and
K.
Li,
”A
parallel
random
forest
algorithm
for
big
data
in
a
s
park
cloud
computing
en
vironment,
”
IEEE
T
ransactions
on
P
arallel
and
Distrib
uted
Systems
,
v
ol.
28,
no.
4,
pp.
919-933,2016.
[25]
Z.
Jiang,
C.
Lu,
Y
.
Cai,
Z.
Jiang,
and
C.
Ma,
”Vpo
wer:
Me
tering
po
wer
consumption
of
vm,
”
in
2013
IEEE
4th
International
Conference
on
Softw
are
Engineering
and
Service
Science
,
pp.
483-486,
IEEE,
2013.
BIOGRAPHY
OF
A
UTHORS
Deepika
T
recei
v
ed
the
B.T
ech
and
M.E
de
grees
from
Anna
Uni
v
ersi
ty
,
in
2010
and
2012,
respec-
ti
v
ely
,
where
she
is
currently
pursuing
the
Ph.D.
de
gree
in
Computer
Sci
ence
and
Engineering,
Am-
rita
School
of
Engineering,
Coimbatore.
His
research
interests
include
Cloud
Computing,
Machine
Learning
and
Image
Processing.
Dr
.
Prakash
P
recei
v
ed
the
Ph.D.
de
gree
in
Information
and
Communication
Engineeri
ng
from
Anna
Uni
v
ersity
,
in
2016.
He
is
currently
serving
as
Assistant
Professor
at
department
of
Computer
Science
and
E
ngineering,
Amrita
School
of
Engineering,
Coimbatore.
His
research
interests
include
Cloud
Computing,
Big
data
analytics,
Automata
Theory
and
Analysis
of
Algorithms.
He
is
also
e
xploring
the
inte
gration
and
data
analysis
of
Internet
of
Things
(IoT)
with
cloud
computing.
Int
J
Elec
&
Comp
Eng,
V
ol.
10,
No.
2,
April
2020
:
1524
–
1532
Evaluation Warning : The document was created with Spire.PDF for Python.