OFFLOAD  
MAKING
DECISION  



FOR WEB
   

APPLICATIONS



2023 International Conference on Intelligent Systems for Communication, IoT and Security (ICISCoIS) | 979-8-3503-3583-5/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICISCOIS56541.2023.10100419

0
3$(!
!$!1)4!,



Dr. Sudha Sadasivam

G !0!2(
3,!0



Barath Kumar G !03-
)14!


Tarun Visva R
 $#!



Computer !

Science and Engineering  $#!



Computer !


Engineering
Science and 







$#!



Computer !

Science and Engineering

.++%'%
.&
%#(-.+.'7

PSG College of Technology 
.++%'%
.&
%#(-.+.'7

PSG College of Technology 
.++%'%
.&
%#(-.+.'7

PSG College
of Technology
.),"!2.0%
-$)!

Coimbatore, .),"!2.0%
-$)!
.),"!2.0%
-$)!


India
'11#1%/1'2%#(!#)-

Coimbatore,

India
 8 /1'2%#(!#)-


Coimbatore, India
 8 /1'2%#(!#)-

[email protected] [email protected] [email protected]




0)2()*


Hrithik B 0!4)-$


Aravind M (..1(!!-


Bhooshaan A
Computer
and Engineering
 $#!



Science !
 $#!



Computer !

Science and Engineering  $#!



Computer !

Science and Engineering

.++%'%
.&
%#(-.+.'7

PSG College of Technology 
.++%'%
.&
%#(-.+.'7

PSG College of Technology 
.++%'%
.&
%#(-.+.'7

PSG College of Technology
.),"!2.0%
-$)!

Coimbatore, India .),"!2.0%
-$)!

Coimbatore, India .),"!2.0%
-$)!

Coimbatore, India
l9z32l @psgtech.ac.in

 8/1'2%#(!#)-  8 /1'2%#(!#)-

[email protected]  8 /1'2%#(!#)-

[email protected]

?4A5>A<0=24
0=3
DB4A
4G?4A84=24
performance "=4
>5
C74
<>BC
8<?>AC0=C

and user experience.
One of the most important
  
Mobile
Abstract: 
"
%


devices play a significant


!
%
role, in our day-to-day%
C0B:B
in
tasks 8=
the
C74
process
?A>24BB
of
>5
offloading
>55;>038=6
is
8B
solving
B>;E8=6
the
C74
decision­
3428B8>=
"
enabling
lives, 
us !
to
work
#
with# 
computationally
!  %
intensive
 "
<0:8=6
?A>1;4<
>5
834=C85H8=6
C74
4G02C
4364
34E824
C>
F7827

making problem of identifying the exact edge device to which

ap 
Implementation
plications.   



of these 
!



applications requires a rich C74
compute
2><?DC4
units
D=8CB
must
<DBC
be
14
offloaded
>55;>0343
to
C>
optimize
>?C8<8I4
the
C74

the


!
!



#

%


set of resources such as highprocessingpower, large memory, and

e
?4A5>A<0=24
'74
45542C8E4=4BB
>5
>55;>038=6
8B
34C4A<8=43

performance. The effectiveness of offloading is determined
an fficient
power
#
supply.
!%
However,
#"
a
huge
!
range

o


f mobile
"


devices fails to

meet
these

!
requirements.
Of
""

jloading involves the 
1H
C74
38554A4=24
14CF44=
C74
C8<4
>5
8<?;4<4=C8=6
0
C0B:
8=

by the difference between the time of implementing a task in
 


delegation 
!
of the compute
!units

#

in web applications

#

from low-end C74
27>B4=
A4<>C4
34E824
C>
C70C
8=
C74
;>20;
34E824

the chosen remote device to that in the local device.


"

local devices to


"

high-end edge devices, to
"


improve performance. In



#

"!



this paper, we propose various machine learning-based prediction 
'74
?DA?>B4
>5
C78B
A4B40A27
8B
C>
34E4;>?
0
3428B8>=<0:8=6

The purpose of this research is to develop a decision-making

which
models, #
take
into
 
account
!
parameters
 
such!
as
execution
$! 
BHBC4<
DB8=6
<0278=4
;40A=8=6
C427=8@D4B
08<43
0C
27>>B8=6

system using machine learning techniques, aimed at choosing



time 

and trans 
to

fer time predict
# 

!
whether a compute
! unit



can be C74
>?C8<0;
A4<>C4
34E824
5>A
>55;>038=6
C74
2><?DC4
D=8CB

the optimal remote device for offloading the compute units,
$! 
locally
executed %
or

must
!
be
o

fjloaded to
a
remote
 
device.
"
The 
?A4B4=C
F8C78=
C74
F41
0??;820C8>=B
'74
4G?4A8<4=C
8=E>;E4B

present within the web applications. The experiment involves
  %

""




strategy proposed involves a machine learning algorithm, 
that



is C4BC8=6
the
testing C74
performance
?4A5>A<0=24
of
>5
various
E0A8>DB
machine
<0278=4
models
<>34;B
on
>=
a0


%
$

trained by examining the 
time


$!
ofexecution 




ofa sample application 


! 
"
# B0<?;4
30C0B4C
2DA0C43
1H
AD==8=6
0
B8<?;4
B>AC8=6
0;6>A8C7<

sample dataset, curated by running a simple sorting algorithm
in multiple devices with

different
!  

computational 
Once
capabilities. 


$!
the execution
time




"


of the local device is predicted 
!

!
using a suitable
DB8=6
0E0&2A8?C
5D=2C8>=B
>=
<D;C8?;4
74C4A>64=4>DB
34E824B

using JavaScript functions on multiple heterogeneous devices

the
regressor, 




proposed model can predict
the 
time


$!
of execution


of 70E8=6
D=8@D4
2>=586DA0C8>=B
'74
<>34;B
DB43
A0=64
5A><

having unique configurations. The models used, range from


"

%


each edge device and correspondingly offload the 
!
compute
!units
0??;820C8>=B?428582
A46A4BB>AB
C70C
?A4382C
C74
4G42DC8>=
C8<4

application-specific regressors that predict the execution time
%


by comparing this 
$!
execution
time




and the aforementioned
 

factors. 10B43
on
based >=
the
C74
parameters
?0A0<4C4AB
of
>5
the
C74
application
0??;820C8>=
to
C>
a0
generic
64=4A82

!
the
Thus, 
!  %

computationally "

intensive 



applications "%

can be selectivel y <0278=4
;40A=8=6
<>34;
C70C
022DA0C4;H
?A4382CB
C74
C8<4
>5

machine learning model that accurately predicts the time of
o

fjloaded to
the

remote

"

edge servers, to
"
!
improve productivity." %
4G42DC8>=
>5
0;;
4364
34E824B
>=
0
B8=6;4
AD=

execution of all edge devices on a single run.

%#
Compute
Keywords: ! 
unit
!
o

com
fjloading, edge ! 
machine
puting, 
'74
final
58=0;
solution
B>;DC8>=
proposed
?A>?>B43
to
C>
the
C74
offload
>55;>03
decision-making
3428B8>=<0:8=6

The



!
learning, 

regression, and clustering 

algorithms. ?A>1;4<
8=
C78B
A4B40A27
8=E>;E4B
CA08=8=6
0
<0278=4
;40A=8=6

problem in this research involves training a machine learning
<>34;
over
model >E4A
a0
dataset,
30C0B4C
created
2A40C43
by
1H
recording
A42>A38=6
the
C74
time
C8<4
of
>5


INTRODUCTION
I. !'%"('"!
4G42DC8>=
>=
<D;C8?;4
4364
34E824B
5>A
0
B8=6;4
0??;820C8>=

execution on multiple edge devices for a single application.
'74
drastic
The 3A0BC82
improvements
8<?A>E4<4=CB
in
8=
the
C74
field
584;3
of
>5
computer
2><?DC4A
science
B284=24
'74
implementation
The 8<?;4<4=C0C8>=
of
>5
the
C74
proposed
?A>?>B43
strategy
BCA0C46H
initiates
8=8C80C4B
with
F8C7

1>C7
5A><
70A3F0A4
0=3
B>5CF0A4
?>8=CB
>5
E84F
70E4
144=

both from hardware and software points of view have been C74
?A4382C8>=
>5
C74
C8<4
>5
4G42DC8>=
>=
C74
;>20;
34E824B

the prediction of the time of execution on the local devices
CA8C4
in
trite, 8=
recent
A424=C
years.
H40AB
However,
>F4E4A
the
C74
increase
8=2A40B4
in
8=
demand
34<0=3
for
5>A
DB8=6
0??;820C8>=B?428582
A46A4BB>AB
0=3
C74=
2><?DC8=6
C74

using application-specific regressors and then computing the
B>5CF0A4
applications
software 0??;820C8>=B
seems
B44<B
to
C>
outrun
>DCAD=
the
C74
hardware
70A3F0A4
C8<4
of
time >5
execution
4G42DC8>=
using
DB8=6
a0
generic
64=4A82
model
<>34;
for
5>A
all
0;;
the
C74
remote
A4<>C4

2><?>=4=CB
A4@D8A43
C>
8<?;4<4=C
C74<

components >34A=30H
DB4AB

required to implement them. Modem-day users 34E824B
8=2;D38=6
C74
0E4A064
CA0=B54A
C8<4
'74
0E0&2A8?C

devices including the average transfer time . The JavaScript
DA64
using
urge DB8=6
computationally
2><?DC0C8>=0;;H
intensive
8=C4=B8E4
software
B>5CF0A4
applications
0??;820C8>=B
2><?DC4
D=8C
20=
14
>55;>0343
1H
2><?0A8=6
C74
4G42DC8>=

compute unit can be offloaded by comparing the execution
>=
B8<?;4
?>AC01;4
34E824B
=
BD27
B8CD0C8>=B
C74
2><<>=

on simple portable devices. In such situations, the common C8<4
on
time >=
the
C74
local
;>20;
and
0=3
various
E0A8>DB
remote
A4<>C4
devices
34E824B
and
0=3
their
C748A

8BBD4
that
issue C70C
arises
0A8B4B
is
8B
that
C70C
these
C74B4
applications
0??;820C8>=B
require
A4@D8A4
resources
A4B>DA24B
0E08;018;8CH

availability.
14H>=3
the
beyond C74
computational
2><?DC0C8>=0;
capabilities
20?018;8C84B
of
>5
the
C74
devices
34E824B
that
C70C

4G42DC4
C74<
F7827
4E4=CD0;;H
;403B
C>
0
38?
8=
?4A5>A<0=24

execute them, which eventually leads to a dip in performance '78B
?0?4A
?A>E834B
0
34C08;43
4G?4A8<4=C0;
0=0;HB8B
>5
C74

This paper provides a detailed experimental analysis of the
0=3
4558284=2H

and efficiency .
B8<D;0C8>=B
performed
simulations ?4A5>A<43
using
DB8=6
various
E0A8>DB
machine
<0278=4
learning
;40A=8=6

<>34;B
C>
B>;E4
C74
>55;>03
3428B8>=<0:8=6
?A>1;4<
BD27

models to solve the offload decision-making problem, such

A4;801;4
B>;DC8>=
C>
BD27
0
?A>1;4<
F>D;3
14
C>
4G?;>8C
C74

A reliable solution to such a problem would be to exploit the 0B
0
2><?0A8B>=
>5
C748A
022DA0284B
0=3
45542C8E4=4BB
0=3
0

as a comparison of their accuracies and effectiveness, and a
?A4B4=24
of
presence >5
numerous
=D<4A>DB
edge
4364
devices,
34E824B
that
C70C
provide
?A>E834
advanced
03E0=243
@D82:
F0;:C7A>D67
>5
1D8;38=6
C74
?A>?>B43
64=4A82
<>34;

quick walkthrough of building the proposed generic model.
B4AE824B
such
services BD27
as
0B
high
7867
computing
2><?DC8=6
capability,
20?018;8CH
routing,
A>DC8=6
and
0=3
'78B
?0?4A
8B
>A60=8I43
0B
5>;;>FB
&42C

This paper is organized as follows: Sect. 2
1A845;H
BD<<0A8I4B

briefly summarizes
?A>24BB8=6
The
processing. '74
introduction
8=CA>3D2C8>=
of
>5
the
C74
concept
2>=24?C
of
>5
'offloading'
N>55;>038=6O
C74
A4;0C43
F>A:
&42C

the related work, Sect. 3
?A>E834B
0
2><?A474=B8E4
>E4AE84F

provides a comprehensive overview
70B
enabled
has 4=01;43
low-end
;>F4=3
portable
?>AC01;4
devices
34E824B
to
C>
take
C0:4
advantage
03E0=C064
of
>5
>5
our
of >DA
system,
BHBC4<
Sect.
&42C
4
delineates
34;8=40C4B
our
>DA
experiments
4G?4A8<4=CB
and
0=3

BD27

such 7867?>F4A43

high-powered 4364

edge 34E824B

devices 8=

in 4G42DC8=6

executing 38B2DBB4B
the
discusses C74
results
A4BD;CB
achieved,
02784E43
and
0=3
Sect.
&42C
5
summarizes
BD<<0A8I4B
this
C78B

2><?DC0C8>=0;;H
8=C4=B8E4
0??;820C8>=B
'7DB
>55;>038=6
20=

computationally intensive applications. Thus, offloading can F>A:
0=3
BD664BCB
5DCDA4
A4B40A27
work and suggests future research.

14
used
be DB43
to
C>
delegate
34;460C4
various
E0A8>DB
complicated
2><?;820C43
applications
0??;820C8>=B
to
C>

<0278=4B
that
machines C70C
can
20=
handle
70=3;4
those,
C7>B4
thereby
C74A41H
enhancing
4=70=28=6
the
C74

979-8-3503-3583-5/23/$3
979-8-3503-3583-5/23/$31.00 2023
c
1 .00 @2023 IEEE
IEEE 333
333

Authorized licensed use limited to: Oracle. Downloaded on February 20,2025 at 13:38:14 UTC from IEEE Xplore. Restrictions apply.
 
RELATED
II. %'
WORK
*"%
>18;4
Cloud
Mobile ;>D3
Computing
><?DC8=6
(MCC) 
./
08<B
to
[5] aims C>
alleviate
0;;4E80C4
the
C74

"
.[1]/
proposes
?A>?>B4B
a0
client-server
2;84=CB4AE4A
0A278C42CDA4
C>
0;;>F
A4B>DA24
?>E4ACH

resource poverty of >5
mobile
<>18;4
34E824B

devices by 1H
>55;>038=6

offloading
MOJA architecture to allow
0E0&2A8?C
0??;820C8>=B
C>
be
14
>55;>0343
C>
cloud
2;>D3
B4AE4AB
2><?DC0C8>=8=C4=B8E4
C>
2;>D310B43
30C0
24=C4AB
;C7>D67

computation-intensive to cloud-based data centers. Although
JavaScript applications to offloaded to servers
AD==8=6
Node.js.
!>349B
The'74
JavaScript
0E0&2A8?C
function
5D=2C8>=
8B
38BC0=C
?D1;82
2;>D3
30C0
24=C4AB
0A4
0
?>C4=C80;
34BC8=0C8>=
C>

distant public cloud data centers are a potential destination to
running is B4A80;8I43

serialized 0=3

and
CA0=B<8CC43
to C>
an
0=
offloading
>55;>038=6
B4AE4A
E80
WebSocket.
*41&>2:4C
'74
>55;>03
740EH
C0B:B
C78B
B>;DC8>=
BD554AB
5A><
8BBD4B
BD27
0B

offload heavy tasks, this solution suffers from issues such as
transmitted server via The
>55;>038=6
B4AE4A
AD=B
C74
5D=2C8>=B
5A><
C74
<>18;4
34E824B
;0A64
;0C4=2H
0=3
;>F
10=3F83C7

=BC403
>5
0
A4<>C4
?D1;82

large latency and low bandwidth. Instead of a remote public
offloading server runs the functions from the mobile devices
1H
Node.js
!>349B
and
0=3
returns
A4CDA=B
the
C74
results
A4BD;CB
to
C>
the
C74
requesting
A4@D4BC8=6
mobile
<>18;4
2;>D3
the
cloud, C74
Mobile
>18;4
Network
!4CF>A:
Operator
"?4A0C>A
;>D3

Cloud (MNOC)!"
can20=

by
34E824B

&74=6
4C
0;
. B4AE4
C74
>55;>0343
C0B:B
=
20B4B
F74A4
C74
A4B>DA24B
>5
C74

serve the offloaded tasks. In cases where the resources ofthe
devices. J. Sheng et al. [2]/
?A>E834
B>;DC8>=B
C>
@D4BC8>=B
BD27

provide solutions to questions such
0B
when
F74=
0=3
F74A4
toC>
>55;>03
C74
compute
2><?DC4
units
D=8CB
8=
<>18;4
!"
0A4

MNOC are 8=BD558284=C

insufficient to C>
handle
70=3;4
the
C74
user's
DB4AOB
request,
A4@D4BC
the
C74

as and where offload the in mobile
34E824B
2>=B834A8=6
C74
C8<4
0=3
4=4A6H
2>=BD<?C8>=
>5
;>20;
!"
forwards
MNOC 5>AF0A3B
it8C
to
C>
a0
public
?D1;82
2;>D3
C7A>D67
the
cloud through C74
WAN
*!

devices considering the time and energy consumption oflocal
2><?DC8=6
0=3
>55;>038=6
2><?DC0C8>=
C
BD664BCB
C74
DB4
>5
2>==42C8>=
'78B
C427=8@D4
8B
A0C74A
0=
0=0;HC820;
<4C7>3
5>A

connection. This technique is rather an analytical method for
computing and offloading computation. It suggests the use of
2><?DC0C8>=0;
A4B>DA24B
BD27
0B
#(
2H2;4B
0=3
4G42DC8>=
C74
?4A5>A<0=24
<>34;8=6
0=3
0=0;HB8B
>5

the performance modeling and analysis of MNOC, !"
10B43
>=

based on
computational resources such as CPU cycles, and execution
A0C4
C>
34C4A<8=4
C74
2><?DC8=6
?>F4A
0=3
4=4A6H
2>=BD<43
C74
Universal
the (=8E4AB0;
Mobile
>18;4
'4;42><<D=820C8>=

Telecommunication &4AE824
Service ( (UMTS)'&

rate, to determine the computing power and energy consumed
8=
C74
;>20;
0=3
A4<>C4
34E824
>=
4=01;8=6
>55;>038=6
C
0;B>
0A278C42CDA4
In
architecture. =
the
C74
proposed
?A>?>B43
0??A>027

approach, a 0
wide
F834
range
A0=64
of>5

in the local and remote device on enabling offloading. It also
8=CA>3D24B
a0
BC0C8BC820;
5>A<D;0
to C>
estimate
4BC8<0C4
the
C74
time
C8<4
>5
?0A0<4C4AB
BD27

parameters, such 0B
F>A:;>03
0=3

as workload 70=3>55
failure
and handoff 508;DA4
rate
A0C4
0A4

are
introduces statistical formula of
CA0=B54A
An
=
0D2C8>=
0;6>A8C7<
to C>
elect
4;42C
the
C74
>?C8<0;
4364
2>=B834A43
The
considered. '74
model
<>34;
captures
20?CDA4B
the
C74
4=3C>4=3
34;0HB
and
end-to-end delays 0=3

transfer. auction algorithm optimal edge
34E824
to
C>
attain
0CC08=
the
C74
maximum
<0G8<D<
possible
?>BB81;4
throughput
C7A>D67?DC
8B
0;B>
4E4=CB
0BB>280C43
F8C7
C74
B4AE828=6
BC4?B
>5
C74
0A278C42CDA4


events associated with the servicing steps of the architecture.
device is also
BD664BC43
>F4E4A
C74
DB4
>5
0
BC0C82
5>A<D;0
8=
0
3H=0<82
'74
>DC?DCB
0=3
A4BD;CB
0A4
>1C08=43
DB8=6
C74

The outputs and results are obtained using the Markov 0A:>E
A4F0A3

reward
suggested. However, the use of a static formula in a dynamic
4=E8A>=<4=C
F>D;3
;02:
C74
4558202H
C>
030?C
C>
5;D2CD0C8=6
0??A>027
with
approach F8C7
the
C74
closed-form
2;>B435>A<
B>;DC8>=B

solutions >5
C74
BD1<>34;B

of the sub-models
environment would lack the efficacy to adapt to fluctuating
502C>AB
BD27
0B
=4CF>A:
CA05582
0=3
#(
;>03
148=6
?A4B4=C43
'74
BC>270BC82
BD1<>34;B
0A4
B>;E43
1H
C74

being presented. The stochastic sub-models are solved by the
factors such as network traffic and CPU load.
&%#

SHARPE ./
[5] B>5CF0A4
?02:064
and
software package 0=3
=D<4A820;
A4BD;CB
are
numerical results 0A4

=
computation
2><?DC0C8>=
offloading,
>55;>038=6
theC74
actual
02CD0;
computation
2><?DC0C8>=
8B
>1C08=43
'74
0=0;HC820;
A4BD;CB
0A4
E4A85843
C7A>D67
38B2A4C4
obtained. The analytical results are verified through discrete­
In is
CA0=B54AA43
to C>
B4AE4ABBDAA>60C4B
F7827
have70E4
high
7867
resource
A4B>DA24
4E4=C
B8<D;0C8>=

event simulation.
transferred servers/surrogates which
BD??>AC
and
support 0=3
hence
74=24
the
C74
process
?A>24BB
8B
0;B>
known
is also :=>F=
as 0B
&DAA>60C4

Surrogate
><?DC8=6
.[3]. /
'74
>55;>038=6
?A>24BB
8=2;D34B
C74
34C42C8>=
H0?D=>E
>?C8<8I0C8>=
./
A454AB
C>
C74
DB4
>5
0
H0?D=>E

Lyapunov optimization [6] refers to the use of a Lyapunov
Computing The offloading process includes the detection
>5
&DAA>60C4B
<>=8C>A8=6
C74
A4B>DA24
0E08;018;8CH
0=3
=443
5D=2C8>=
C>
>?C8<0;;H
2>=CA>;
0
3H=0<820;
BHBC4<
H0?D=>E

function to optimally control a dynamical system. Lyapunov
of Surrogates, monitoring the resource availability and need,
?0AC8C8>=8=6
C74
C0B:B
B2743D;8=6
C74
C0B:B
>=
C74
0E08;01;4
5D=2C8>=
8B
DB43
C>
?A>E4
C74
BC018;8CH
>5
C74
4@D8;81A8D<
?>8=C

function is used to prove the stability of the equilibrium point.
partitioning the tasks, scheduling the tasks on the available
BDAA>60C4
0=3
2>=CA>;;8=6
C74
4G42DC8>=
0=3
64CC8=6
A4BD;CB
C
'74
0;6>A8C7<
?A>3D24B
C74
>55;>038=6
3428B8>=
0B
0
A4BD;C
>5

The algorithm produces the offloading decision as a result of
surrogate and controlling the execution and getting results. It
8B
4<?8A820;
C70C
C74
>55;>038=6
F8;;
14
5AD8C5D;
>=;H
F74=
C8<4
2>=B834A8=6
C74
CA0=B<8C
?>F4A
5>A
>55;>038=6
0=3
4G42DC8>=

considering the transmit power for offloading and execution
is empirical that the offloading will be fruitful only when time
C0:4=
to
C>
4G42DC4
C74
task
C0B:
;>20;;H
2>BC
that
cost C70C
includes
8=2;D34B
execution
4G42DC8>=
latency
;0C4=2H
and
0=3
task
C0B:
failure
508;DA4
as
0B

taken execute the locally 8B
6A40C4A
than
is greater C70=
that
C70C
of
>5
the
C74

C8<4
taken
C0:4=
to C>
execute
4G42DC4
DB8=6
?4A5>A<0=24
metrics.
performance <4CA82B
The
'74
implementation
8<?;4<4=C0C8>=
of >5
the
C74
algorithm
0;6>A8C7<

time using 2><?DC0C8>=
>55;>038=6
The
computation offloading. '74

0??A>027
computes
2><?DC4B
the C74
performance
?4A5>A<0=24
improvement
8<?A>E4<4=C
via E80
>=;H
A4@D8A4B
C>
B>;E4
0
34C4A<8=8BC82
?A>1;4<
8=
4027
C8<4

only requires to solve a deterministic problem in each time
approach
>55;>038=6
using
DB8=6
parameters
?0A0<4C4AB
BD27
0B
the
C74
B8I4
>5
transmitted
CA0=B<8CC43
B;>C
5>A
F7827
C74
>?C8<0;
B>;DC8>=
20=
14
>1C08=43
48C74A
8=

slot, for which the optimal solution can be obtained either in
offloading such as size of
30C0
30C0
C>
14
A4248E43
5A><
C74
BDAA>60C4
C748A
A4B?42C8E4
2;>B43
5>A<
>A
1H
18B42C8>=
B40A27
'74
0;6>A8C7<
2>=B834AB

closed form or by bisection search. The algorithm considers
data, data to be received from the surrogate, their respective
10=3F83C7B
0=3
C74
C8<4
C0:4=
C>
4G42DC4
C0B:
>=
C74
A4<>C4
>=;H
C74
2DAA4=C
BC0C4
>5
C74
3H=0<820;
BHBC4<
A0C74A
C70=
0;;

only the current state of the dynamical system rather than all
bandwidths and the time taken to execute task on the remote
B4AE4A
C
proposes
?A>?>B4B
an 0=
0BB4AC8>=
C70C
most
<>BC
methods
<4C7>3B
relyA4;H
on
>=
C74
?0BC

the past 2><?DC0C8>=

computation taskC0B:
request
A4@D4BC
and0=3
therefore
C74A45>A4
theC74

server. It assertion that
3H=0<82
strategy
BCA0C46H
with
F8C7
partitioning
?0AC8C8>=8=6
to C>
make
<0:4
offloading
>55;>038=6
2><?;4G8CH
>5
C74
0;6>A8C7<
8B
E4AH
;4BB

D0=6
4C
0;
./

complexity of the algorithm is very less. D. Huang et al. [7]
dynamic
3428B8>=B
=
8=E4BC860C8>=
F0B
2>=3D2C43
C>
8<?;4<4=C
C74
?A>?>B4
0
3H=0<82
>55;>038=6
0;6>A8C7<
10B43
>=
H0?D=>E

propose a dynamic offloading algorithm based on Lyapunov
decisions. An investigation was conducted to implement the
A4B>DA24
allocation
0;;>20C8>=
for5>A
0=
"?C8<8I0C8>=
'78B

Optimization. 0;6>A8C7<
0BBD<4B

This algorithm assumes thatC70C
0=

an 0??;820C8>=

application
resource an D?;8=:
!>="AC7>6>=0;
Multiple
uplink Non-Orthogonal D;C8?;4

224BB
.[4]/
based
10B43
Mobile
>18;4
Edge
364
Computing =4CF>A:
'74

><?DC8=6
network. 2>=B8BCB
of
consists >5
an
0=
un-offloadable 2><?>=4=C
and
D=>55;>0301;4
component 0=3
other
>C74A

Access The
10B82
8340
>5
>18;4
364
><?DC8=6
 
8B
C>
DC8;8I4
C74
>55;>0301;4
2><?>=4=CB
'74
3428B8>=
>5
F7827
2><?>=4=CB

offloadable components. The decision of which components
basic idea of Mobile Edge Computing (MEC) is to utilize the
?>F4A5D;
2><?DC8=6
5028;8C84B
within
F8C78=
the C74
radio
A038>
access
0224BB
C>
be
to 14
offloaded
>55;>0343
8B
<034
by
is made 1H
the
C74
0;6>A8C7<

algorithm 022>A38=6

according to C>
the
C74

powerful computing facilities
=4CF>A:
suchBD27
as0B
the
C74
MEC
B4AE4A
8=C46A0C43
8=C>
C74
base
10B4
=4CF>A:
connectivity,
network 2>==42C8E8CH
network
=4CF>A:
traffic
CA05582
and
0=3
data
30C0
rate.
A0C4
The
'74

network, server integrated into the
BC0C8>=
2><?;4G8CH
>5
C74
?A>?>B43
0;6>A8C7<
8B
2><?0A0C8E4;H
;>F

complexity of the proposed algorithm is comparatively low.
station &
%424=C;H
Non-Orthogonal
(BS). Recently, !>="AC7>6>=0;
Multiple D;C8?;4
Access
224BB

!" 
70B
144=
A42>6=8I43
0B
0
?>C4=C80;
C427=>;>6H
5>A
'74
0;6>A8C7<
CA84B
C>
<8=8<8I4
C74
0E4A064
4G42DC8>=
C8<4

The algorithm tries to minimize the average execution time
(NOMA) has been recognized as a potential technology for
C74
=4GC
64=4A0C8>=
<>18;4
2><<D=820C8>=
=4CF>A:B
C>
?DC
>5
C74
C0B:B
1H
>55;>038=6
C74
=424BB0AH
C0B:B
8=
BD27
0
F0H

of the tasks by offloading the necessary tasks in such a way
the next generation mobile communication networks to put
D?
F8C7
C74
4G?;>B8E4
6A>FC7
>5
30C0
CA05582
'74
C>C0;
4=4A6H
C70C
C74
0E4A064
4G42DC8>=
C8<4
8B
0;F0HB
;4BB4A
C70=
>A
4@D0;

that the average execution time is always lesser than or equal
up with the explosive growth of data traffic. The total energy
2>=BD<?C8>=
>5
0;;
DB4AB
8B
5>A<D;0C43
5>A
0=
D?;8=:
!"  C>
the
to C74
4G42DC8>=
C8<4
requirement
execution time A4@D8A4<4=C
of >5
the
C74
BHBC4<

system. '74

The
consumption of all users is formulated for an uplink NOMA­
10B43
MEC 
network
=4CF>A:
viaE80
>?C8<8I8=6
CA0=B<8BB8>=
power,?>F4A
?A>?>B43
algorithm
proposed 0;6>A8C7<
minimizes
<8=8<8I4B
the C74
average
0E4A064
energy
4=4A6H

based optimizing transmission
>55;>038=6
30C0
0=3
0;;>20C8>=
>5
time.
C8<4
Then
'74=
theC74
time
C8<4
2>=BD<?C8>=
BD27
C70C
C74
=D<14A
>5
4G42DC8>=B
F7827
3>
=>C

consumption such that the number of executions which do not
offloading data and allocation of
0;;>20C8>=
<44C
C74
4G42DC8>=
A4@D8A4<4=C
<DBC
14
;4BB
C70=
>A
4@D0;
C>

meet the execution requirement must be less than or equal to
allocation 5>A
38554A4=C
groups
for different 6A>D?B
8B
8=E4BC860C43
where
is investigated, F74A4
twoCF>

DB4AB
0A4
?08A43
8=
4027
6A>D?
C>
?4A5>A<
!"
users are paired in each group to perform NOMA. 
'74
C>C0;

The total 1 0
percent.
?4A24=C
A 
major
<09>A
drawback
3A0F102:
0BB>280C43
F8C7
the
associated with C74
existing
4G8BC8=6

4=4A6H
<8=8<8I0C8>=
?A>1;4<
8B
?A>E43
C>
14
0
2>=E4G
>=4
B>;DC8>=B
8B
C74
01B4=24
>5
0=
4558284=C
<0278=4
;40A=8=6
C>

solutions is the absence of an efficient machine learning to
energy minimization problem is proved to be a convex one.
4B834B
8C
8B
0;B>
B7>F=
C70C
CA0=B<8CC8=6
F8C7
<0G8<0;
C8<4
27>>B4
C74
>?C8<0;
4364
34E824
C>
>55;>03
C74
2><?DC4
D=8CB

choose the optimal edge device to offload the compute units
Besides, it is also shown that transmitting with maximal time
8B
>?C8<0;
8=
4=4A6H
B0E8=6
0B43
>=
C74
>?C8<0;
2>=38C8>=B
C0:8=6
into
taking 8=C>
022>D=C

account theC74
processing
?A>24BB8=6
?>F4A

power and 0=3
network
=4CF>A:

is optimal in energy saving. Based on the optimal conditions,
0=
8C4A0C8E4
0;6>A8C7<
8B
022>A38=6;H
?A>?>B43
F74A4
2;>B43 ?0A0<4C4AB

parameters.
an iterative algorithm is accordingly proposed, where closed­
5>A<
expressions
form 4G?A4BB8>=B
are 0A4
obtained
>1C08=43
in8=
each
4027
step
BC4?
5>A

for >?C8<8I8=6

optimizing 
PROPOSED *"%

#%"#"&
WORK
III.
C8<4
allocation
time 0;;>20C8>=
or >A
>55;>038=6
30C0
The
offloading data. '74
proposed
?A>?>B43
iterative
8C4A0C8E4

0;6>A8C7<
F8C7
;>F
2><?;4G8CH
8B
BD224BB5D;;H
?A>E43
C>
14
'74
>55;>03
3428B8>=<0:8=6
BHBC4<
0B
34?82C43
8=
586DA4

The offload decision-making system, as depicted in figure 1 
,
algorithm with low complexity is successfully proved to be
6;>10;;H
>?C8<0;

globally optimal. i8B

s a0
machine
<0278=4
learning
;40A=8=6
based
10B43
prediction
?A4382C8>=
model
<>34;
used
DB43
to
C>

34C4A<8=4
F74A4
to
determine where C>
execute
4G42DC4
the
C74
2><?DC4
D=8CB
8=

compute units in F41

web
0??;820C8>=B
locally
applications, ;>20;;H
or
>A
A4<>C4;H

remotely on>=
an
0=
4364
34E824
The
edge device. '74

334 2023 International Conference on Intelligent Systems for

for Communication, loT
IoT and Security (ICISCoIS)
(ICISColS)

Authorized licensed use limited to: Oracle. Downloaded on February 20,2025 at 13:38:14 UTC from IEEE Xplore. Restrictions apply.
?A>?>B43
BHBC4<

proposed system, F78;4

while <0:8=6

making ?A4382C8>=B

predictions 5>A

for 4G42DC8=6

executing 0

a >5

of A46A4BB>AB

regressors C>

to 14

be <08=C08=43

maintained 8=

in C74

the B4AE4A

server. DC

But C78B

this
2><?DC0C8>=0;;H

computationally 740EH

heavy 0??;820C8>=

application >=

on 0
a ;>F4=3

low-end 34E824

device, 0??A>027

approach =424BB8C0C4B

necessitates <08=C08=8=6

maintaining A46A4BB>AB

regressors
F>D;3

would ?A>101;H

probably BD664BC

suggest C74

the 34E824

device C>

to >55;>03

offload 8CB

its 2><?DC4

compute 3H=0<820;;H
8=
C74
B4AE4A
;B>
4E4AH
=4F
0??;820C8>=

dynamically in the server. Also, every new application
D=8CB
C>
C74
50BC4BC
0E08;01;4
4364
34E824
"=
C74
>C74A
70=3

units to the fastest available edge device. On the other hand, A4@D8A4B

requires 0
a =4F

new A46A4BB>A

regressor C>
to 14

be CA08=43

trained >E4A

over C74

the
F74=

when 0
a 34E824

device 8B

is 20?01;4

capable >5

of 70=3;8=6

handling BD27

such 2><?;820C43

complicated 2>AA4B?>=38=6

corresponding 30C0B4C

dataset 0=3

and 03343

added C>

to C74

the B4AE4A

server 0B

as C74

the
0??;820C8>=B

applications >=

on 8CB

its >F=

own >A

or F74=

when C74

the CA0=B54A

transfer C8<4

time C>

to 4364

edge A46A4BB>AB

regressors 0A4
are 1D8;C

built DB8=6

usmg 0??;820C8>=B?428582

application-specific
34E824B
8B
A4;0C8E4;H
7867
C74
?A>?>B43
BHBC4<
F>D;3
?A454A

devices is relatively high, the proposed system would prefer ?0A0<4C4AB

parameters.
C>
4G42DC4
C74
F41
0??;820C8>=
;>20;;H
to execute the web application locally.


3) "
$"#!


K-Means Clustering: A <>A4

more ?A>6A4BB8E4

progressive 0??A>027

approach
DB4B
C74
F4;;4BC01;8B743

uses the well-established K-Means 40=B
;DBC4A8=6
0;6>A8C7<

Clustering algorithm
'74

The <0278=4

machine ;40A=8=6

learning <>34;

model 5>A

for ?A4382C8=6

predicting C74

the C8<4

time >5

of 05C4A

after C74

the 4G8BC8=6

existing 30C0B4C

dataset 8B

is 2>=E4AC43

converted 8=C>

into 0

a BD8C01;4

suitable
4G42DC8>=

execution >=
on C74

the ;>20;

local 0=3

and A4<>C4

remote 34E824B

devices 8B

is 1D8;C

built DB8=6

using 18=0AH
5>A<0C
F74A4

binary format, where 0
0=3
and 1
A4?A4B4=C
C70C
C74
<0278=4

represent that the machine
#HC7>=
F8C7
C74
74;?
>5
0
54F
?A43458=43
<0278=4
;40A=8=6

Python with the help of a few pre-defined machine learning 8B

is 20?01;4

capable >5

of 70=3;8=6

handling C74

the 68E4=

given ;>03

load A4<>C4;H

remotely 0=3

and
0;6>A8C7<B
BD27
0B
;8=40A
;>68BC82
0=3
?>;H=><80;
A46A4BB>AB

algorithms such as linear, logistic, and polynomial regressors ;>20;;H

locally, A4B?42C8E4;H

respectively. ;C7>D67

Although C74

the <>34;

model 20=

can <0:4

make
0=3

and <40=B

K-means 2;DBC4A8=6

clustering. '74

The >E4A0;;

overall BHBC4<

system 0A278C42CDA4

architecture ?A4382C8>=B
01>DC
F74C74A
C>
4G42DC4
0
2><?DC4
D=8C
5>A

predictions about whether to execute a compute unit for
2>=B8BCB
>5
0
B8<?;4
F41
B4AE4A
1D8;C
F8C7
C74
74;?
>5
;0B:

consists of a simple web server built with the help of Flask 0
68E4=
B4C
>5
?0A0<4C4AB
;>20;;H
>A
A4<>C4;H
C74
<>34;

a given set of parameters, locally or remotely, the model
C70C
4=2;>B4B
C74
CA08=43
<0278=4
;40A=8=6
<>34;
0=3
8=5>A<B

that encloses the trained machine learning model and informs 8B

is 8=2><?;4C4

incomplete 0B
as 8C

it 3>4B

does =>C

not ?A4382C

predict C74

the >?C8<0;

optimal 4364

edge
C74
2;84=C
;>20;
34E824
01>DC
C74
?A4382C8>=
>5
C74
>55;>038=6

the client (local device), about the prediction of the offloading device to which the compute unit must be offloaded.

34E824
C>
F7827
C74
2><?DC4
D=8C
<DBC
14
>55;>0343
3428B8>=

decision.


B. *()(+
*#,#('
(%


Proposed Prediction Model


>=B834A8=6
C74
B4C102:B
50243
8=
C74
?A43458=43
0;6>A8C7<B

Considering the setbacks faced in the predefined algorithms,
F4
?A>?>B4
0
64=4A82
<0278=4
;40A=8=6
0;6>A8C7<
C70C
F>D;3

we propose a generic machine learning algorithm, that would
14

be 0??;820C8>=8=34?4=34=C

application-independent 0=3

and 20?01;4

capable >5

of 4BC8<0C8=6

estimating C74

the
C8<4
>5
4G42DC8>=
>5
0;;
A4<>C4
34E824B
DB8=6
0
B8=6;4
CA08=43

time of execution of all remote devices using a single trained
<>34;
BC>A43
8=
C74
B4AE4A

54F
34B86=
?A8=28?;4B
8;;DBCA0C43

model stored in the server. A few design principles illustrated
1H
C74
?A>?>B43
0;6>A8C7<
0A4
by the proposed algorithm are,



E>838=6
C74
<08=C4=0=24
>5
<D;C8?;4
CA08=43
<>34;B
8=
C74

• Avoiding the maintenance of multiple trained models in the

B4AE4A

server.


0?018;8CH
>5
4BC8<0C8=6
C74
C8<4
>5
4G42DC8>=
>5
0;;
A4<>C4

• Capability of estimating the time of execution of all remote

o � ·n �­
34E824B
10B43
>=
C74
C8<4
>5
4G42DC8>=
>=
C74
;>20;
34E824

devices based on the time of execution on the local device.


'74
<>34;
20=
14
DB43
C>
4BC8<0C4
C74
C8<4
>5
4G42DC8>=
8=

• The model can be used to estimate the time of execution in

0
E0A84CH
>5
0??;820C8>=B

a variety of applications.


'74
?A>?>B43
<>34;
?A>E834B
C74
5;4G818;8CH
8=
0338=6
=4F

• The proposed model provides the flexibility in adding new

34E824B
0=3
0??;820C8>=B
F8C7
40B4

devices and applications with ease.

Algorithm

Train_Model:

�
  )%

Input: Dataset
%

 
$
 

!



 

••
L, containing the execution time of all devices

+Li
J$



, represents the local


!


time of execution of devices 
&'


'j' for
$!

#

!%!


Figure !

1 : Client - Server Block Diagram 
&'
 

Output:
the 'i'th sample
%

Array "
NS,$
 


containing the normalized #



timings of each 
 

device
A.
/#+,#'!
*#,#('
(%+


Existing Prediction Models ,-
NSk,  $



represents the normalized #
summation 
!


execution time

 
&'

Function calculate_NS( ):
of device 'k'
8554A4=C

Different <0278=4

machine ;40A=8=6

learning <>34;B

models F4A4

were 8<?;4<4=C43

implemented 0=3
and 
  ) 3

L 4%

C4BC43

tested >E4A
over 0
a B0<?;4

sample 30C0B4C

dataset, 2DA0C43

curated 1H
by 4G42DC8=6

executing 0
a

initialize#
,-
NSk 

*


to 0 for all 
 

devices
0E0&2A8?C

JavaScript 5D=2C8>=

function >5
of B4;42C8>=

selection B>AC

sort 5>A

for 0

a 68E4=

given ;8BC

list >5

of


For each 
sample

+Li




in the dataset 
$
L,

?0A0<4C4AB

)


 


!



parameters 8=
in 0

a 54F

few 34E824B

devices F8C7

with 38554A4=C

different 2><?DC0C8>=0;

computational let min_L be the fastest time of execution in the
sample


20?018;8C84B

capabilities. '74
The <08=

main >1942C8E4

objective >5
of C74B4

these <>34;B

models F0B

was C>

to


 
!

+
For each execution time Li
J


 
&'$

for device 'j',
4BC8<0C4
C74
C8<4
>5
4G42DC8>=
>=
4027
8=38E83D0;
34E824
8=

estimate the time of execution on each individual device in


+Li
J
2
+
= Li
J
(
)




I min_L
C74
=4CF>A:
&><4
?A43458=43
<0278=4
;40A=8=6
<>34;B
C70C

the network. Some pre-defined machine learning models that


,-
NS1 
12
+
+= Li
J

F4A4
8<?;4<4=C43
0A4
0B
5>;;>FB

were implemented are as follows.




1) ! ' 
!""
Linear/Polynomial Regression:
'74

The <>BC

most ?A8<8C8E4

primitive

)
let min_NS
be 


the smallest


!

time of execution
 



in the NS

0??A>027

approach F0B
was C>

to <0:4

make DB4

use >5

of D=8@D4

unique ;8=40A

linear 0=3

and

 
,-
summation
For each NSk 

,-$

in NS,
?>;H=><80;
A46A4BB8>=
<>34;B
5>A
4027
34E824
C70C
C0:4

polynomial regression models for each device, that take

,-
NSk 
2
,-
= NSk 
(
)
I min_NS

8=

in C74

the 0??;820C8>=B?428582

application-specific ?0A0<4C4AB

parameters 0B
as 8=?DC

input 0=3

and

,-

return NS
A4CDA=B
C74
4BC8<0C43
C8<4
>5
4G42DC8>=

<09>A
502C>A

returns the estimated time of execution. A major factor
C70C
BD??>ACB
C74
DB064
>5
?>;H=><80;
A46A4BB8>=
8B
C74

that supports the usage of polynomial regression is the  
1:
><?DC8=6
=>A<0;8I43

Algo. Computing normalized
502C

fact C70C

that <>BC

most F41

web 0??;820C8>=B

applications AD=

run >=

on ?>;H=><80;

polynomial >A

or 4G42DC8>=
C8<4
>5
0;;
34E824B


execution time of all devices.
;>F4A

lower AD=C8<4

run-time 2><?;4G8C84B

complexities. >F4E4A

However, 0
a <09>A

major

3A0F102:

drawback >5
of C74

the 0??A>027

approach 8B
is C70C

that 0

a ;0A64

large =D<14A

number >5

of '74

The ?A>?>B43

proposed 64=4A82

generic <>34;

model ;6>

(Algo. 1

), 148=6

being 0

a ?0A0<4C4A

parameter
8=34?4=34=C
A46A4BB8>=
<>34;B
B7>D;3
14
<08=C08=43
8=

independent regression models should be maintained in 8=34?4=34=C

independent <>34;

model, 20=

can 14

be CA08=43

trained DB8=6

using 0=H

any B0<?;4

sample
the server.

C74
B4AE4A 0??;820C8>=
=
C78B
4G?4A8<4=C
0
B0<?;4
30C0B4C
8B
2DA0C43
1H

application. In this experiment, a sample dataset is curated by
2)
 "#
!""
Logistic Regression:
=
0;C4A=0C8E4
F0B
8<?;4<4=C8=6

An alternative was implementing A42>A38=6
C74
C8<4
>5
4G42DC8>=
>5
0
B4;42C8>=
B>AC
0E0&2A8?C

recording the time of execution of a selection sort JavaScript
0
;>68BC82
A46A4BB8>=
<>34;
>E4A
C74
BH=C74B8I43
CA08=8=6

a logistic regression model over the synthesized training 5D=2C8>=
5>A
C74
B0<4
2><18=0C8>=
>5
?0A0<4C4AB
>=
38554A4=C

function for the same combination of parameters on different
30C0B4C
'78B
0??A>027
3A0BC820;;H
342A40B4B
C74
=D<14A

dataset. This approach drastically decreases the number 34E824B

devices. '74

The 0;6>A8C7<

algorithm C>

to CA08=

train C74

the <>34;

model 1468=B

begins 1H

by

2023 International
International Conference
Conference on
on Intelligent
Intelligent Systems for
Systems for Communication,
Communication, IoT
loT and
and Security
Security (ICISCoIS)
(ICISCoiS) 335
335

Authorized licensed use limited to: Oracle. Downloaded on February 20,2025 at 13:38:14 UTC from IEEE Xplore. Restrictions apply.
 =>A<0;8I8=6

normalizing 4027

each A>F

row >5

of C74

the B0<?;4

sample 30C0B4C

dataset, 8=

in F7827

which, C74

the C74
2;84=C
<0:4B
DB4
>5
0
BD8C01;4
A46A4BB>A
C>
4BC8<0C4
C74

the client, makes use of a suitable regressor to estimate the
4G42DC8>=
C8<4
>5
C74
50BC4BC
34E824
8B
2>=B834A43
0B

execution time of the fastest device is considered as 1 
0=3

, and C8<4
C0:4=
C>
4G42DC4
C74
68E4=
2><?DC4
D=8CB
5>A
C74
4=;8BC43

time taken to execute the given compute units for the enlisted
C74
C8<4
>5
4G42DC8>=
>5
C74
A4BC
>5
C74
34E824B
8B
A4?A4B4=C43

the time of execution of the rest of the devices is represented ?0A0<4C4A
values, T1 
• &D1B4@D4=C;H

parameter E0;D4B
Subsequently, C74
the B4AE4A

server 4G?;>8CB

exploits C74

the
0B
0
<D;C8?;4
>5
C74
50BC4BC
34E824
'7DB
0
B8=6;4
A>F
8=
C74

as a multiple of the fastest device. Thus, a row in the ?A>?>B43
64=4A82
<>34;
4<143343
F8C78=
8C
C>
34C4A<8=4
C74

proposed generic model embedded within it to determine the
=4F;H

newly BH=C74B8I43

synthesized 30C0B4C

dataset, ?A>E834B

provides 0

a B8<?;4

simple 2><?0A8B>=

comparison 0??A>G8<0C4
C8<4
>5
4G42DC8>=
>5
C74
68E4=
2><?DC4
D=8C
>=

approximate time of execution of the given compute unit on
14CF44=
the
between C74
2><?DC0C8>=

computation ?>F4A

power >5

of C74

the E0A8>DB

various 34E824B

devices, 8=

in C74

the A4<>C4

remote 34E824B

devices, Tn  
DB8=6

using C74

the 4BC8<0C43

estimated ;>20;

local 4G42DC8>=

execution
C4A<B
>5
C74
50BC4BC
34E824


terms of the fastest device. C8<4

time.

>F4E4A

However, 8=
in >A34A

order C>

to 0E>83

avoid 4AA>AB

errors 3D4

due C>

to 0=><0;84B

anomalies 8=
in H

By C0:8=6

taking 8=C>

into 022>D=C

account, C74

the C8<4

time C0:4=

taken C>
transfer, T1,
to CA0=B54A


C74

the
8=38E83D0;
A4038=6B
C74
BD<<0C8>=
>5
C74
=4F;H
=>A<0;8I43

individual readings, the summation of the newly normalized 2><?DC4
D=8CB
C>
4027
A4<>C4
34E824
0;>=6
F8C7
C74
4BC8<0C43

compute units to each remote device along with the estimated
E0;D4
8B
20;2D;0C43
5>A
4027
34E824
D=34A
2>=B834A0C8>=
'74

value is calculated for each device under consideration. The A4<>C4

remote 4G42DC8>=

execution, C74

the 4364

edge 34E824B

devices 20=

can 14

be B>AC43

sorted 10B43

based >=
on
BD<<0C8>=
E0;D4B
5>D=3
8=
4027
2>;D<=
8;;DBCA0C4
C74
02CD0;

summation values found in each cohunn, illustrate the actual C748A
018;8CH
C>
70=3;4
C74
4G70DBC8E4
0E0&2A8?C
2><?DC4
D=8C

their ability to handle the exhaustive JavaScript compute unit
2><?DC0C8>=

computation 20?018;8CH

capability >5

of C74

the 34E824

device A4?A4B4=C43

represented 8=
in C74

the F8C7

with 40B4

ease. '74

The 34E824B

devices 20=

can C74=

then 14

be 2742:43

checked 5>A

for C74

the
2>;D<=

column. '74

The BD<<0C8>=B

summations >5
of C74

the E0A8>DB

various 34E824B

devices 20=
can >=24

once 0E08;018;8CH

availability C>

to C0:4

take ?0AC

part 8=

in C74

the >55;>038=6

offloading ?A>24BB

process, 0=3

and C74

the
0608=

again 14

be A4?A4B4=C43

represented 0B

as C74

the <D;C8?;4

multiple >5

of C74

the B<0;;4BC

smallest 50BC4BC
A4038;H
0E08;01;4
34E824
20=
14
?A><?C43
C>
4G42DC4

fastest readily available device can be prompted to execute
BD<<0C8>=
0<>=6
0;;
C74
34E824B
C>
0E>83
340;8=6
F8C7
;0A64

summation among all the devices, to avoid dealing with large C74
2><?DC4
D=8CB

the compute units.
=D<14AB

numbers. '74

The =4F;H

newly =>A<0;8I43

normalized BD<<0C8>=B

summations !&

(NS) >5
of C74

the

various devices can be represented as NS1,
!&
J
!&
E0A8>DB
34E824B
20=
14
A4?A4B4=C43
0B
!& . . . , NSn
0=3

and 0/ 
0/  ./

































(2)
C74B4
E0;D4B
2><?DC43
20=
14
DB43
C>
?A4382C
C74
4G42DC8>=

these values computed can be used to predict the execution F74A4

where,
C8<4
>5
0;;
C74
34E824B
F74=
C74
4G42DC8>=
C8<4
>5
>=4
34E824

time of all the devices, when the execution time of one device   
A4?A4B4=CB
C74
C>C0;
C8<4
C0:4=
C>
>55;>03
0=3
4G42DC4
0

Ert • represents the total time taken to offload and execute a
8B
?A>E8343
DB8=6
C74
B8<?;4
?A>?>AC8>=0;8CH
5>A<D;04


is provided, using the simple proportionality formulae, compute unit on remote device 'i'
2><?DC4
D=8C
>=
A4<>C4
34E824


 
A4?A4B4=CB
C74
C8<4
C0:4=
C>
4G42DC4
0
2><?DC4
D=8C
>=

Trt . represents the time taken to execute a compute unit on
T1 NS1

=




































(1)

the remote device 'i'
C74
A4<>C4
34E824


0 NSj
 T11,

 
A4?A4B4=CB
C74
C8<4
C0:4=
C>
CA0=B54A
0
2><?DC4
D=8C
C>

represents the time taken to transfer a compute unit to
F74A4

where, C74
A4<>C4
34E824


the remote device 'i'

T1,
 
A4?A4B4=CB
C74
:=>F=
C8<4
>5
4G42DC8>=
>5
;>20;
34E824
N8O

represents the known time of execution of local device 'i'
 
A4?A4B4=CB
C74


'0· the D=:=>F=

unknown C8<4

time >5

of 4G42DC8>=

execution >5

of A4<>C4

remote '74
?A>?>B43
>55;>03
3428B8>=<0:8=6
BHBC4<
20=
5DAC74A
14

The proposed offload decision-making system can further be
34E824
N9O

device 'j ' >?C8<8I43
1H
C74
8=2;DB8>=
>5
0
54F
540CDA4B
BD27
0B
5>;;>FB

optimized by the inclusion of a few features such as follows:

NS1,  
A4?A4B4=CB

represents C74

the =>A<0;8I43

normalized BD<<0C8>=

summation >5

of C74

the ;>20;

local


• 
C7A4B7>;3
E0;D4
20=
14
8=CA>3D243
C>
C74
BHBC4<
BD27

A threshold value can be introduced to the system, such
34E824
N8O

device 'i' C70C
>=;H
B86=85820=C
38554A4=24
8=
C8<4
C0:4=
C>
4G42DC4

that, only significant difference in time taken to execute

NSj, 
A4?A4B4=CB

represents C74

the =>A<0;8I43

normalized BD<<0C8>=

summation >5

of C74

the A4<>C4

remote a compute unit locally, T1 
0=3
C8<4
C0:4=
C>
CA0=B54A

0
2><?DC4
D=8C
;>20;;H
, and time taken to T
1

,
34E824
N9O

device 'j ' 0=3
4G42DC4
8C
8=
0
A4<>C4
34E824

and execute it in a remote device, Tr  
20=
14
2>=B834A43

, can be considered
0=3
C7>B4
F8C7
=46;8681;4
38554A4=24B
20=
14
0E>8343
0B

and those with negligible differences can be avoided as
'74

The C8<4

time >5

of 4G42DC8>=

execution >5
of C74

the 2><?DC4

compute D=8CB

units >=

on C74

the ;>20;

local C74H

they <0H

may 14
be ?A>=4

prone C>
to 4AA>AB

errors 8=

in 4BC8<0C8>=B

estimations <034

made 1H
by
34E824

device 20=

can 14

be 34C4A<8=43

determined 1H
by AD==8=6

nmning 0
a B8<?;4

simple ?>;H=><80;

polynomial <0278=4

machine ;40A=8=6

learning <>34;B

models 0=3

and D=?A42434=C43

unprecedented ;0C4=284B

latencies
A46A4BB8>=
5>A
>=4
0??;820C8>=
>=
C74
;>20;
34E824
8CB4;5
>A
1H

regression for one application on the local device itself or by C70C

that <0H

may >22DA

occur 0;>=6B834

alongside C74
the 20;2D;0C43

calculated CA0=B54A

transfer C8<4

time,
BC>A8=6
C74
C8<4
C0:4=
C>
2><?;4C4
4G42DC8>=
>5
C74
2><?DC4

storing the time taken to complete execution of the compute 84
 + T
1 
i.e., Tr »
T1 
•
D=8C

unit 8=

in C74

the ?0BC

past 5>A

for 5DCDA4

future A454A4=24B

references. '7DB

Thus, 0B

as <4=C8>=43

mentioned


• '74

The CA05582

traffic BC0CDB

status >5
of 4027

each A4<>C4

remote 34E824

device 20=

can 0;B>

also 14

be
01>E4

above, C74

the ?A>?>B43

proposed 64=4A82

generic <>34;

model ?A>3D24B

produces ?A><8B8=6

promising <>=8C>A43
5>A
C74
>55;>03
3428B8>=<0:8=6
C>
0E>83
>DC
monitored for the offload decision-making to avoid out­
A4BD;CB

results 8=

in ?A4382C8=6

predicting C74

the 4BC8<0C43

estimated C8<4

time >5

of 4G42DC8>=

execution 8=

in B>DA28=6

sourcing <D;C8?;4

multiple 2><?DC4

compute D=8CB

units C>

to 0
a B8=6;4

single A4<>C4

remote
<D;C8?;4
A4<>C4
34E824B
F8C7>DC
<0:8=6
0=H
2><?A><8B4B

multiple remote devices, without making any compromises, 34E824

device 0=3

and D=34ADC8;8I8=6

underutilizing C74
the >C74A

other 0E08;01;4

available A4<>C4

remote
D=;8:4

unlike C74

the ?A43458=43

predefined <0278=4

machine ;40A=8=6

learning <>34;B

models. '74

The 34E824B

devices.
?A>?>B43
<>34;
0;B>
20?CDA4B
3H=0<8B<
0B
2><?DC8=6
>=;H

proposed model also captures dynamism, as computing only
C74

the =>A<0;8I43

normalized BD<<0C8>=

summation >5
of 0

a =4F

new 34E824

device 8B
is @D8C4

quite )

IV. +#% !'
!
EXPERIMENT AND
%&('&

RESULTS
2>=E4=C8>=0;

conventional, 0=3
a:nd F>D;3

would 0;;>F

allow C74

the =4F

new 34E824

device C>

to 14

be '74

The ?A>?>B43

proposed <0278=4

machine ;40A=8=6

learning 10B43

based >55;>03

offload 3428B8>=
decision­
2>=B834A43
D=34A
C74
A030A
F78;4
<0:8=6
3428B8>=B
5>A
>C74A

considered under the radar, while making decisions for other <0:8=6
BHBC4<
F0B
4G?4A8<4=C43
D=34A
38554A4=C
B8CD0C8>=B

making system was experimented under different situations
34E824B
E4AH
B>>=
05C4A
20;2D;0C8=6
8CB
!&

devices, very soon after calculating its NS. 0=3
4364
2>=38C8>=B
C>
34C4A<8=4
8CB
4558284=2H
0=3
022DA02H

and edge conditions to determine its efficiency and accuracy.


C. *()(+
#+#('$#'!
0+,&

Proposed Decision-Making System '78B

This B42C8>=

section 38B2DBB4B

discusses C74

the ?4A5>A<0=24

performance >5
of C74

the E0A8>DB

various
'74
?A4382C8>=
<>34;B
C74
4G?4A8<4=C0;
B4CD?
0=3
C74
C4BC
A4BD;CB

prediction models, the experimental setup, and the test results
The >E4A0;;

overall 0A278C42CDA4

architecture >5
of C74

the ?A>?>B43

proposed >55;>03

offload 3428B8>=
decision­
<0:8=6
02784E43

achieved.
making BHBC4<

system, 2><?A8B4B

comprises 0
a B4AE4A

server, F8C7

with C74

the 01>E4
above­
<4=C8>=43
<>34;B
CA08=43
0=3
4<143343
F8C78=
8C
0=3
0
B4C

mentioned models trained and embedded within it, and a set A.
/)*#&',%
,-)

Experimental Setup
>5

of 4364

edge 34E824B

devices C70C

that 02C

act 0B

as 2;84=CB

clients, A4@D4BC8=6

requesting C74

the B4AE4A

server C>

to 

A 30C0B4C

dataset F0B

was 2>=BCAD2C43

constructed 5>A

for 0
a B0<?;4

sample 0E0&2A8?C

JavaScript
?A>E834

provide 0
a 3428B8>=

decision C>
to >55;>03

offload, C>

to C74

the >C74A

other 2><?DC0C8>=0;;H

computationally 2><?DC0C8>=
D=8C
>5
C74
B4;42C8>=
B>AC
F8C7
0
B8=6;4
?0A0<4C4A

computation unit of the selection sort with a single parameter
A4B>DA245D;
4364
34E824B
0E08;01;4
*74=4E4A
0=
0??;820C8>=

resourceful edge devices available. Whenever, an application ;4=6C7

(length >5

of C74

the A0=3><

random 0AA0H

array C>

to 14

be B>AC43

sorted). 027

Each A>F

row
8=
C74
2;84=C
CA84B
C>
4G42DC4
0
2><?DC4
D=8C
C74
2;84=C
A4@D4BCB

in the client tries to execute a compute unit, the client requests 4=20?BD;0C4B
C74
C8<4
C0:4=
C>
B>AC
0
A0=3><
0AA0H
>5
0
68E4=

encapsulates the time taken to sort a random array of a
C74

the B4AE4A

server C>

to ?A>E834

provide 8=5>A<0C8>=

information >=
on F74C74A

whether C74

the 2><?DC4

compute ;4=6C7

length 8=

in 58E4

five 38554A4=C

different 34E824B

devices F8C7

with 2><?;4C4;H

completely 38554A4=C

Ullltat:Hl
D=8CB
20=
14
4G42DC43
;>20;;H
>A
8=
0
B?428582
A4<>C4
34E824

units can be executed locally or in a specific remote device, 2>=586DA0C8>=B
0B
B7>F=
8=
'01;4

'74
34E824B
C0:4=
D=34A

configurations, as shown in Table I. The devices taken under
1H

by B4=38=6

sending C74
the B4C

set >5

of ?0A0<4C4AB

parameters C70C
that 34C4A<8=4

determine C74

the 2>=B834A0C8>=

consideration, 8=2;D34

include <>18;4

mobile ?7>=4B

phones 0=3

and ;0?C>?B

laptops, AD==8=6

running
2><?;4G8CH

complexity >5
of C74

the 2><?DC4

compute D=8C

unit D=34A

under 2>=B834A0C8>=

consideration. '74

The 38554A4=C

different >?4A0C8=6

operating BHBC4<B

systems BD27

such 0B

as =3A>83

Android, 8"&

iOS, 8=DG

Linux,
B4AE4A
>=
A4248E8=6
C74
?0A0<4C4AB
>5
C74
2><?DC4
D=8CB
5A><

server, on receiving the parameters of the compute units from <02"&

macOS 0=3

and *8=3>FB

Windows. 027

Each 34E824

device 703

had 0

a D=8@D4

unique <4<>AH

memory

336
336 2023 International
2023 /ntemationalConference
Conference on
on Intelligent
IntelligentSystems
Systems for Communication, IoT
for Communication, loTand
andSecurity
Security (ICISCoIS)
(ICISColS)

Authorized licensed use limited to: Oracle. Downloaded on February 20,2025 at 13:38:14 UTC from IEEE Xplore. Restrictions apply.
20?028CH

capacity 0=3

and 2;>2:

clock B?443

speed. '74

The C8<4

time C0:4=

taken C>

to B>AC

sort >=

on C74

the 34E824B
0=3
C74
2>AA4B?>=38=6
C8<4
C>
14
C0:4=
C>
>55;>03
C74

devices and the corresponding time to be taken to offload the
05>A4<4=C8>=43
34E824B
5>A
>E4A
 
D=8@D4
;4=6C7
E0;D4B

aforementioned devices for over 50,000 unique length values 2><?DC4
D=8CB
C>
C74<
0C
0=H
68E4=
8=BC0=C
B
B>>=
0B
C74

compute units to them, at any given instant. As soon as the
F0B
A42>A343
0=3
DB43
0B
C74
30C0B4C
'78B
30C0B4C
F0B
DB43

was recorded and used as the dataset. This dataset was used B4AE4A
A4248E4B
0
A4@D4BC
5A><
0
2;84=C
34E824
C>
?A>E834
0=

server receives a request from a client device to provide an
C>
CA08=
C74
<0278=4
;40A=8=6
<>34;B

to train the machine learning models. >55;>03
3428B8>=
8C
4BC8<0C4B
C74
0??A>G8<0C4
C8<4
C>
14
C0:4=

offload decision, it estimates the approximate time to be taken
C>

to 4G42DC4

execute C74

the 2><?DC4

compute D=8C

unit ;>20;;H

locally, 2>=B834A8=6

considering C74

the


$!#

%"
$"
!
#"#

Table I: Configuration of devices usedfor testing ?0A0<4C4AB

parameters B4=C

sent 1H

by C74

the 2;84=C

client, DB8=6

using 0

a BD8C01;4

suitable A46A4BB>A

regressor.
(B8=6
C74
;>20;
C8<4
>5
4G42DC8>=
?A4382C43
40A;84A
C74
B4AE4A

Using the local time of execution predicted earlier, the server


Device 

Chipset 

Operating 

RAM 

Sorting <0:4B

makes DB4

use >5

of C74

the 64=4A82

generic <>34;

model C>

to 34C4A<8=4

determine C74

the C8<4

time >5

of


no. 

System (GB)




Time 4G42DC8>=
>5
C74
A4<>C4
34E824B
0=3
B>ACB
C74<
10B43
>=
C74

execution of the remote devices and sorts them based on the


1 //+%


Apple M l !#

MacOS 8
25.6

,1


ms
4BC8<0C43
C8<4
>5
4G42DC8>=
0=3
CA0=B54A
C8<4
0B43
>=
C74

estimated time of execution and transfer time. Based on the


2 //+%


Apple A l 2 )

iOS 

3 25.2

,1

ms BC0CDB
5;06B
>5
C74
4364
34E824B
C74
B4AE4A
B44:B
5>A
0=
4364

status flags of the edge devices, the server seeks for an edge


3 Intel l O
th

-2%+
 )-$.51

Windows 8
35.5

,1

ms 34E824
C70C
8B
2><?DC0C8>=0;;H
<>A4
?>F4A5D;
C70=
C74
;>20;

device that is computationally more powerful than the local
%-
)
Gen i7
34E824

device, 2>=B834A8=6

considering C74

the 4BC8<0C43

estimated A4<>C4

remote 4G42DC8>=

execution C8<4

time,


4 3!+#.,,

Qualcomm -$0.)$

Android 6

,1

30.3 ms CA0=B54A

transfer C8<4

time, 0=3

and C74

the 0E08;018;8CH

availability >5

of C74

the A4<>C4

remote 34E824

device.
-!/$0!'.-

Snapdragon 

OS '74A45>A4
C74
B4AE4A
<0:4B
0
3428B8>=
C>
>55;>03
C74
2><?DC4

Therefore, the server makes a decision to offload the compute


750G D=8CB

units, 85

if 0C

at ;40BC

least >=4

one >5

of C74

the 78674A4=3

higher-end 4364

edge 34E824B

devices 8B

is
5
Intel 7
th

-2%+
)-36

Linux 8

,1

78.0 ms
0E08;01;4
5>A
>55;>038=6
4;B4
8C
BD664BCB
C74
2;84=C
C>
4G42DC4

available for offloading, else, it suggests the client to execute
%-
)
Gen iS

C74
0E0&2A8?C
2><?DC4
D=8C
>=
8CB
>F=

the JavaScript compute unit on its own.

8ABC;H
C74
AD38<4=C0AH
3428B8>=<0:8=6
BHBC4<
F0B
CA08=43

Firstly, the rudimentary decision-making system was trained 

B. +,
+-%,+

Test Results
0=3

and C4BC43

tested >E4A

over C74

the 01>E4<4=C8>=43

above-mentioned 30C0B4C

dataset, F8C7

with
?>;H=><80;
A46A4BB8>=
0B
C74
4;4<4=C0;
0;6>A8C7<
5>A
CA08=8=6

polynomial regression as the elemental algorithm for training '01;4

B7>FB
C74
C4BC
A4BD;CB
>5
DB8=6
0
?>;H=><80;
A46A4BB>A

Table II shows the test results of using a polynomial regressor
C74

the ?A4382C8>=

prediction <>34;

model. 
A B4C

set >5

of 8=34?4=34=C

independent <>34;B

models F4A4

were C>

to 4BC8<0C4

estimate C74

the C8<4

time >5

of 4G42DC8>=

execution >5
of 4027

each 34E824

device. '74

The
34E4;>?43
C>
4BC8<0C4
C74
C8<4
C0:4=
C>
B>AC
5>A
0
68E4=
;4=6C7

developed to estimate the time taken to sort for a given length 022DA02H
>5
C74
<>34;
2>=BCAD2C43
8B
<40BDA43
8=
C4A<B
>5
C74

accuracy ofthe model constructed is measured in terms of the
>=
4027
34E824
0=3
C74
3428B8>=
C>
>55;>03
C74
2><?DC4B
F0B

on each device and the decision to offload the computes was 40=
1B>;DC4
AA>A
F7827
8B
C74
0E4A064
34E80C8>=
5A><
C74

Mean Absolute Error, which is the average deviation from the
<034

made 10B43

based >=

on C74

the 4BC8<0C43

estimated 4G42DC8>=

execution C8<4

time. '01;4

Table 

II 4G?42C43

expected E0;D4

value. C

It 8B

is 4E834=C

evident 5A><

from C74

the C01;4

table C70C

that C74

the <>34;

model
A4?A4B4=CB

represents C74

the 022DA02H

accuracy >5

of C74

the ?>;H=><80;

polynomial A46A4BB>AB

regressors ?4A5>A<B
38554A4=C;H
8=
4027
34E824
0=3
C74
022DA02H
>5
C74

performs differently in each device and the accuracy of the
8<?;4<4=C43

implemented 5>A
for C74

the 58E4

five 34E824B

devices, 8=

in C4A<B

terms >5

of C74
40=

the Mean <>34;

model 8B

is =>F74A4

nowhere 2;>B4

close C>

to ?4A542C

perfect, 0B

as C74

the 0E4A064

average C8<4

time >5

of
1B>;DC4
AA>A
 
20;2D;0C43
DB8=6
C74
02CD0;
4G42DC8>=

Absolute Error (MAE), calculated using the actual execution 4G42DC8>=

execution >=

on C74

the 34E824B

devices 8B

is 01>DC

about 200ms <B
0=3

and 0=

an MAE 
>5

of
C8<4
0=3
C74
4BC8<0C43
4G42DC8>=
C8<4

time and the estimated execution time. <B
A4?A4B4=CB
0=
4AA>A
>5

26ms represents an error of 1 3%.

&42>=3;H

Secondly, <>34;B

models ;8:4

like ;>68BC82

logistic A46A4BB8>=

regression 0=3

and <40=B

K-means 

Table II:

"$#
!!!

#
%"

$"

Mean Absolute Error &

ofthe devices on using polynomial
2;DBC4A8=6
F4A4
8<?;4<4=C43
5>A
C74
>55;>03
3428B8>=<0:8=6

clustering were implemented for the offload decision-making !!""!

regressor
BHBC4<
F7827
A4@D8A43
0
B8=6;4
<>34;
C>
14
CA08=43
5>A
4027

system, which required a single model to be trained for each
0??;820C8>=

application. '74

The 8=8C80;;H

initially DB43

used 30C0B4C

dataset F0B

was BH=C74B8I43

synthesized 5>A

for %
$!

Device Number 
MAE
"

(ms)
1>C7
<>34;B
'74
30C0B4C
5>A
C74
;>68BC82
A46A4BB8>=
2>=C08=43

both models. The dataset for the logistic regression contained 4E824

Device 1
0. 1

5
0
=D<14A
14CF44=

a number between 1
0=3

A4?A4B4=C8=6
C74
34E824
8=
F7827

and 5, representing the device in which 4E824

Device 2
 

26.28
C74
2><?DC4
D=8CB
<DBC
14
4G42DC43
0=3
C74
30C0B4C
5>A

the compute units must be executed, and the dataset for K­ 4E824

Device 3

5 . 3

9
<40=B
2;DBC4A8=6
F0B
BH=C74B8I43
C>
0
18=0AH
30C0B4C
F74A4

means clustering was synthesized to a binary dataset, where 4E824

Device 4

8.90

1
34=>C4B
C70C
C74
2><?DC4
D=8CB
20=
14
4G42DC43
;>20;;H
0=3

denotes that the compute units can be executed locally and 4E824


Device 5 

1 2.66
0
34=>C4B

denotes C70C

that C74

the D=8CB

units <DBC

must 14

be >55;>0343

offloaded, 38B2>E4A43

discovered 1H

by
2><?0A8=6
C74
;>20;
4G42DC8>=
C8<4
0=3
C74
BD<
>5
0E4A064

comparing the local execution time and the sum of average '74
;>68BC82
A46A4BB8>=
BD8CB
F74=
C74
27>824
>5
C74
A4<>C4

The logistic regression suits when the choice of the remote
CA0=B54A

transfer C8<4

time 0=3

and A4<>C4

remote 4G42DC8>=

execution C8<4

time. '01;4B

Tables 

III 0=3

and )

IV 34E824

device C>
to >55;>03

offload C74

the 2><?DC4

compute D=8CB

units >A

or C74

the ?0A0<4C4A

parameter
A4?A4B4=C
C74
?4A5>A<0=24
<40BDA4
>5
C74
;>68BC82
A46A4BB>A

represent the performance measure of the logistic regressor C7A4B7>;3
5>A
C74
;>20;
4G42DC8>=
8B
C>
14
?A4382C43
'01;4


threshold for the local execution is to be predicted. Table III
0=3

and <40=

K-mean 2;DBC4A8=6

clustering <>34;

model, 8=

in C4A<B

terms >5

of C74

the =D<14A

number >5
of A4?A4B4=CB

represents C74
the C4BC

test A4BD;C

result >5

of C74

the <>34;

model 34E4;>?43

developed DB8=6

using
;4=6C7
E0;D4B
5>A
F7827
C74
<>34;B
<034
C74
2>AA42C
3428B8>=

length values for which the models made the correct decision. ;>68BC82
A46A4BB8>=
5>A
>55;>03
3428B8>=<0:8=6
>=
34E824


logistic regression for offload decision-making on device 5 .
C
8B
4E834=C
5A><
C74
C01;4
C70C
C74
<>34;
C4=3B
C>
34;460C4

I t i s evident from the table that the model tends to delegate
8=0;;H
C74
?A>?>B43
0??;820C8>=8=34?4=34=C
64=4A82
<>34;

Finally, the proposed application-independent generic model C74

the 2><?DC4

compute D=8CB

units C>

to >=;H

only 0

a B8=6;4

single 50BC4BC

fastest 34E824

device, D=34A

under 0;;

all
F0B
34E4;>?43
>E4A
C74
30C0B4C
>5
C74
B>AC8=6
0??;820C8>=
H

was developed over the dataset of the sorting application. By 28A2D<BC0=24B
85
8C
=443B
C>
14
>55;>0343
"=
C74
>C74A
70=3

circumstances, if it needs to be offloaded. On the other hand,
?A>E838=6

providing C74

the C8<4

time >5

of 4G42DC8>=

execution >5

of >=4

one A0=3><;H

randomly 27>B4=

chosen 5>A

for C74

the <>34;

model 2>=BCAD2C43

constructed DB8=6

using C74

the <40=B

K-means ;DBC4A8=6

Clustering
34E824
0<>=6
C74
58E4
C74
C8<4
>5
4G42DC8>=
F0B
4BC8<0C43

device among the five, the time of execution was estimated 0;6>A8C7<
0
?>BB81;4
022DA02H
<40BDA4
F>D;3
14
C>
<40BDA4

algorithm, a possible accuracy measure would be to measure
5>A
0;;
C74
>C74A
5>DA
34E824B
0=3
8C
F0B
2><?0A43
C>
C74
02CD0;

for all the other four devices and it was compared to the actual C74
=D<14A
>5
B0<?;4
20B4B
5>A
F7827
C74
<>34;
34C4A<8=4B

the number of sample cases for which the model determines
C8<4
C0:4=
C>
4G42DC4
C74
B>AC8=6
5D=2C8>=
'01;4
)
A4?A4B4=CB

time taken to execute the sorting function. Table V represents 2>AA42C;H

correctly C>
to >55;>03

offload >A

or 4G42DC4

execute ;>20;;H

locally. '01;4

Table 

III B7>FB

shows C74

the
C74

the ?4A5>A<0=24

performance 0=3
and 022DA02H

accuracy >5

of C74

the ?A>?>B43
model, 8=

proposed <>34;
m C4BC

test A4BD;CB

results >5

of C74

the <>34;

model 34E4;>?43

developed DB8=6

using C74

the <40=B

K-means
C4A<B
>5

terms of MAE 
5>A
4027
34E824
D=34A
2>=B834A0C8>=


for each device under consideration. 2;DBC4A8=6
0;6>A8C7<
=0;HB8B
>5
C74
C01;4
?A>E834B
DB
F8C7

clustering algorithm. Analysis of the table provides us with
C74

the 2>=2;DB8>=

conclusion C70C

that C74

the <>34;

model 20=

can 14
be D=A4;801;4

uureliable D=34A

under
'74

The ?A>?>B43

proposed <>34;

model >=24

once CA08=43

trained, F0B

was 4<143343

embedded F8C78=

within 0

a <D;C8?;4
28A2D<BC0=24B
0B
C74
?4A24=C064
>5
C74
2>AA42C=4BB

multiple circumstances as the percentage of the correctness
B4AE4A
1D8;C
DB8=6
;0B:
'74
B4AE4A
34E4;>?43
AD=B
C7A>D67

server built using Flask. The server developed, runs through >5
8CB
?A4382C8>=B
20=
50;;
D?
C>

of its predictions can fall up to 86.29%. 
'7DB
0
<>34;
F8C7
0

Thus, a model with a
0

a B4A84B

series >5

of ?A>24BB4B

processes, 0=3

and A4B?>=3B

responds C>

to 8CB

its 2;84=CB

clients 01>DC

about C74

the 14CC4A

better 022DA02H

accuracy F>D;3

would 14

be A4@D8A43

required C>
to B>;E4

solve C74

the >55;>03

offload
>55;>038=6

offloading 3428B8>=

decision, 10B43

based >=

on C74

the 2A8C820;

critical ?0A0<4C4AB

parameters >5

of C74

the 3428B8>=<0:8=6
?A>1;4<

decision-making problem.
0E0&2A8?C

JavaScript 2><?DC4

compute D=8C

unit C>

to 14

be >55;>0343

offloaded, F74=

when 0
a A4@D4BC

request
0A8B4B
'74
B4AE4A
<08=C08=B
C74
BC0CDB
>5
C74
0E08;01;4
4364

arises. The server maintains the status of the available edge

2023 International
2023 International Conference
Conference on
on Intelligent
Intelligent Systems
Systems for Communication,
for Communication, IoT
loT and
and Security
Security (ICISCoIS)
(ICISColS) 337
337

Authorized licensed use limited to: Oracle. Downloaded on February 20,2025 at 13:38:14 UTC from IEEE Xplore. Restrictions apply.
 

Table III:
Decision
"

!#"
$"
"#
!!""

makingpredictions using logistic regression C74
execution
the 4G42DC8>=
of >5
5D=2C8>=B

functions 8=
in ;>F4=3

low-end 34E824B
<>A4

devices more
!
%


for device 5 A4B?>=B8E4
Future
responsive. DCDA4
works
F>A:B
in
8=
the
C74
0A40

area 8=E>;E4
2>=B834A8=6

involve, considering
<>A4
540CDA4B
>5
C74
34E824B
C70C
05542C
C74
C8<4
>5
4G42DC8>=

more features of the devices that affect the time of execution
%

Device #$

Actual BD27

such 0B
2DAA4=C
CPU
#(
load,
;>03
device
34E824
battery,
10CC4AH
and
0=3
the
C74
network
=4CF>A:

Predicted &$# "

!#
Executions as current
$!

Number &$# "

Executions BC0CDB
C74A41H
<0:8=6
C74
?A>?>B43
0;6>A8C7<
<>A4
030?C8E4

status, thereby making the proposed algorithm more adaptive.
4E824
1

Device 406 
0
338C8>=0;;H
344?
;40A=8=6
0;6>A8C7<B
<0H
14
8<?;4<4=C43

Additionally, deep learning algorithms may be implemented,
4E824
2

Device 0
0
C>
8=2;D34
C74
F4867C064
>5
C74
?0A0<4C4AB
>5
C74
0E0&2A8?C

to include the weightage of the parameters of the JavaScript
4E824
3

Device 43 1 6
4484
2><?DC4
D=8CB
8=
C74
>55;>03
3428B8>=<0:8=6

compute units, in the offload decision-making.
4E824
4

Device 0
0

4E824
5

Device 80

1 53  1 
5618

!"* !'

ACKNOWLEDGEMENT
'78B
project
This ?A>942C
was
F0B
carried
20AA843
out
>DC
under
D=34A
the
C74
Samsung
&0<BD=6
PRISM
#%&



Table I
"


V Decision !#"
$"

making predictions "
$"#!

using K-means clustering =C4A=B78?
?A>6A0<
0=3
F4
F>D;3
;8:4
C>
4G?A4BB
>DA
B8=24A4

Internship program and we would like to express our sincere
6A0C8CD34
to
gratitude C>
the
C74
institution
8=BC8CDC8>=
mentor
<4=C>A
Mr. A
&DH0<1D;860<

Suyambuligam
%

Device !!#

Correct  !!#

Incorrect $!'

Accuracy DC7D?0=38
&4=8>A

%0C78=0B0<H
Muthupandi,
Rathinasamy Senior =68=44A

Engineer 0C

at &0<BD=6

Samsung
$!

Number (%)

%4B40A27
=BC8CDC4

Research 0=60;>A4
5>A

Institute, Bangalore, 4GC4=38=6
his
for extending 78B
constant
2>=BC0=C

4E824
1

Device 20 1 02
0


1 00.0 BD??>AC
0=3
6D830=24
C7A>D67>DC
C74
?A>942C
F>A:

support and guidance throughout the project work.
4E824
2

Device  

17347 

2755  

86.29
4E824
3

%%!&

REFERENCES
Device 20072 30


99.85
4E824
4

Device 18049


2053 

89.78 .[ 1 /

] 
+D
Y.
C. Xu, ,
Qiao,
$80>

44
and
B. Lee, 0=3
N.
!
Murray,
DAA0H
L"MOJA
"

-
4E824

>18;4
>55;>038=6

Mobile offloading 5>A
0E0&2A8?C
applications,"
for JavaScript 0??;820C8>=BM
8=

in
Device 5 1 
9832


270 

98.65

('
JET *'
Publications,
Conference -%#,#('+
20 1 4,
vol.
E>;
2014,
no.
=>



# 
??
K
CP639, pp. 59-63 
3>8

. doi: 10. 1 049/cp.20
2? 1 4.0659.
 

'01;4
)
B7>FB
C74
C4BC
A4BD;CB
>5
C74
<>34;
34E4;>?43
DB8=6

Table V shows the test results of the model developed using
C74
?A>?>B43
64=4A82
<>34;
C>
4BC8<0C4
C74
C8<4
>5
4G42DC8>=

the proposed generic model to estimate the time of execution


>5
4027

of each 34E824

device. A 
B8<?;4
2><?0A8B>=
of
simple comparison >5
its
8CB
022DA02H

accuracy, .[2]/

&74=6

J. Sheng, 
D
X.
J. Hu, +
'4=6

Teng, 

B. *0=6

Wang, 0=3

and X.+
Pan,
#0=

<40BDA43
8=
C4A<B
>5

measured 
B7>FB
C70C
C74
?A>?>B43
<>34;

in terms of MAE, shows that the proposed model L><?DC0C8>=
offloading
"Computation >55;>038=6
strategy
BCA0C46H
8=

in <>18;4

mobile 4364

edge
C0:4B
0
680=C
;40?
0608=BC
C74
>C74A
<>34;
2>=B834A43
C>
B>;E4

takes a giant leap against the other model considered to solve
2><?DC8=6M
'
computing," Inf(*&,#('
.#,1*%'
E>;

ormation (Switzerland), vol. 10,
=>
no.
C74
offload
the >55;>03
3428B8>=<0:8=6
?A>1;4<
Interpretation
decision-making problem. =C4A?A4C0C8>=
of
>5
the
C74

D=

3>8

6, Jun. 20 19,   8=5> 1 0060
doi: 10.3390/info  1 9 1 
.
C01;4
0;B>
?A>E834B
DB
F8C7
0
B7A43
>5
4E834=24
C70C
C74
<>34;

table also provides us with a shred of evidence that the model

?4A5>A<B
4@D0;;H
6A40C
8=
?A4382C8=6
C74
0??A>G8<0C4
C8<4
>5

performs equally great in predicting the approximate time of .[3]/
M 
. W*
. Nabeel,
!0144;
A 
. Embong,
<1>=6
0=3

and M 
. Ali,;8

4G42DC8>=
of>5
the
C74
compute
2><?DC4
units
D=8CB
on
>=
all
0;;
the
C74
34E824B
D=34A
L><?DC0C8>=
>55;>038=6
5>A
B<0AC
8=C4A=4C
34E824BM

"Computation offloading for smart internet devices,"
8=
2015
execution devices under
2>=B834A0C8>=
0;;>F8=6
0
E4AH
B<0;;
4AA>A
>5
01>DC

allowing a very small error of about 3 . 1% 
0C
in 
4th ',*',#('%
('
,"
International *'
on
Conference ('
So
(f,.*

tware
and Computer Systems, ICSECS 2015:
consideration, at
F>ABC

worst. '!#'*#'!
'
(&)-,*
0+,&+


Engineering 


#*,-(-+
(
Virtuous Sof,.*
(%-,#('+

tware Solutionsfor (*
#!
,
!>E

Big Data, Nov. 20 1 5
,



"$#
!!!

#
%"

$"

Table V: Mean Absolute Error ! "

of the devices on usingproposed ??
223-228.
pp. K 
3>8

doi: 1 0. 1 1 09/ICSECS.20
&& 1 5 .7333 1 1 

5.
!

generic

.[4]/

,>D

D0=6

704
0=3


8<
L=4A6H
C. You, K. Huang, H. Chae, and B. H. Kim, "Energy­
!
 

Generic Model 
"

MAE (ms) 558284=C
Resource
Efficient %4B>DA24
Allocation
;;>20C8>=
5>A
>18;4364

for Mobile-Edge
4E824

Device 1
    

1 . 1 5 1 4097076902787 ><?DC0C8>=
"55;>038=6M
8=

*'+,#('+
('

Computation Offloading," in IEEE Transactions on
4E824
2

Device  1 0380372
2.552   1 5364

#*%++
(&&-'#,#('+

Wireless Communications, Mar. 0A
20 1 
E>;

7, vol. 1 
=>

6, no. 3 

,
4E824
3
     
??
1 3 K
pp. 97-1 4 1 1 
3>8

. doi: 10. 1 1 09/TWC.20
'* 1    

6.2633522.
Device 1 . 1 8325 1 04970 1 7548


4E824
4

Device 
6. 1   

859298685428790
./

[5] 
%048

H. 0=3
N.
Raei and !
Yazdani,
,0I30=8
L=0;HC820;
?4A5>A<0=24

"Analytical performance
4E824


Device 5   6309700
5.078575891 

<>34;
5>A
<>18;4
=4CF>A:
>?4A0C>A
2;>D3M
(-*'%

model for mobile network operator cloud," Journal
o Supercomputing, vol. 7 1 
=>

(f
-)*(&)-,#'!
E>;
 , no. 12,
??
K 

pp. 4555-4577,
)
"!(&"!

V. CONCLUSION 42
20 1 5
3>8

Dec. , doi: 1 0 . 1007/s
B 1 1 227-0
 1 
5- 1 
55 1 -4.

*8C7
C74
03E0=24<4=CB
8=
C74
584;3
>5
>'
>55;>038=6
B4AE4B

With the advancements in the field of loT, offloading serves

0B
0
<09>A
2><?>=4=C
8=
4=70=28=6
C74
DB4A
4G?4A84=24
1H

as a major component in enhancing the user experience by ./

[6] ,
Mao,
Y. 0>
J.

-70=6
0=3
K.
Zhang, and 

4C0845
"Dynamic
B . Letaief, LH=0<82

34;460C8=6
exhaustive
delegating 4G70DBC8E4
computational
2><?DC0C8>=0;
workloads
F>A:;>03B
to C>
the
C74
><?DC0C8>=
Offloading
Computation "55;>038=6
5>A
for >18;4364

Mobile-Edge
34E824B
C70C
20=
70=3;4
C74<
F8C7
40B4
>F4E4A
C74
02CD0;

devices that can handle them with ease. However, the actual ><?DC8=6
F8C7
=4A6H
0AE4BC8=6
4E824BM


Computing with Energy Harvesting Devices," IEEE
?4A5>A<0=24
1>>BC
608=43
C7A>D67
>55;>038=6
7D64;H
;84B
>=

performance boost gained through offloading, hugely lies on (-*'%
('
%,

Journal on Selected Areas*+
#'
(&&-'#,#('+
E>;

in Communications, vol.
27>>B8=6
C74
2>AA42C
A4<>C4
34E824B
C>
F7827
C74
2><?DC4

choosing the correct remote devices, to which the compute 34,
no.
=>
12,
pp.
??
3590-3605,
 K  
Dec. 42
20 1 

6, 3>8

doi:
D=8CB
are
units 0A4
>55;>0343

offloaded. '78B
?0?4A
?A4B4=CB

This paper presents a 0
platform­
?;0C5>A< 10. 1 109/JSAC.20
& 1   1 1 
6 .26 964.

8=34?4=34=C
>55;>03
3428B8>=<0:8=6
BHBC4<
1D8;C
DB8=6
0=

independent offload decision-making system, built using an

0??;820C8>=8=34?4=34=C
64=4A82

application-independent generic <0278=4

machine ;40A=8=6
<>34;

learning model. ./

[7] 
D0=6

D. Huang, P. #
*0=6

Wang, and 0=3

!8H0C>
"A
D. Niyato, L
3H=0<82

dynamic
'74
?A4B4=C43
<>34;
BC0=3B
0?0AC
5A><
C74
4G8BC8=6
<0278=4

The presented model stands apart from the existing machine >55;>038=6
algorithm
offloading 0;6>A8C7<
5>A
<>18;4
computing,"
for mobile 2><?DC8=6M
IEEE 

;40A=8=6
models,
learning <>34;B
being
148=6
highly
7867;H
accurate
022DA0C4
in
8=
34C4A<8=8=6
C74

determining the *'+
#*%
(&&-'
E>;

Trans Wire! Commun, vol. 1 1 
=>

??

, no. 6, pp. 1 
99 1-1K 

995,
C8<4
>5

time 4G42DC8>=
on
of execution >=
remote
A4<>C4
BHBC4<B
C>
?A>2443

systems, to F8C7
the
proceed with C74
D=
20 12,
3>8

Jun. doi: 10. 1 1 09/TWC.20
'* 1 2.04  1 9 1 2. 1 10912.
 

>55;>03
decision-making.
offload 3428B8>=<0:8=6
'7DB

Thus, >=
C74
10B8B

on the >5
the
basis of C74



4G?4A8<4=C0;
A4BD;CB
it
experimental results, 8C
can
20=
14
8=54AA43
that
be inferred C70C
0E0&2A8?C

JavaScript
>55;>038=6
done
offloading 3>=4
10B43

based >=
C74
predictions
on the ?A4382C8>=B
made
<034
by
1H
the
C74

?A>?>B43
3428B8>=<0:8=6
BHBC4<
8B
A4;801;4
0=3
F>D;3
<0:4

proposed decision-making system is reliable and would make

338 2023 International Conference on Intelligent Systems for

2023 for Communication, JoT
IoT and Security (ICISCoIS)
(ICISCoiS)

Authorized licensed use limited to: Oracle. Downloaded on February 20,2025 at 13:38:14 UTC from IEEE Xplore. Restrictions apply.