1

Mobile Edge Computing for Efficient Task

Offloading

Department of Computer Science and

Engineering

Netaji Subhas University of Technology (NSUT)

Name of the student Student Roll No.

Om Tiwari (2020UCO1592)
Pavani Dubey (2020UCO1608)

Under the supervision of

Dr. Abhinav Tomar
(Assistant Professor NSUT)
 2

INDEX

Index Description Page No.

1 Introduction 3-4
1.1 Background 3
1.2 Motivation 3
2 Literature Survey 5-9
2.1 Literature Table 5
2.2 Previous work done 7
2.3 Research Gaps 8
3 Problem Statement 10-10
3.1 Objective 10
4 Methodology 11-11
5 Implementation 12-14
5.1 Work done till date 13
5.2 Future Work 14
6 References 15-16
 3

1. Introduction
1.1. Background

With the rapid advancement in development of more and more smart mobile devices
(MDs), computation hungry mobile applications have started to gain enormous
attractions. These applications have led to ever-increasing computation demands.
Some examples of these computation hungry applications are Image Recognition
Applications, Weather Forecasting, Virtual Reality Applications, Scientific and
Engineering Calculators, Cryptocurrency Mining Apps etc. The mobile devices on which
these applications execute have limited battery as well as limited computation resources
[1]. Therefore it becomes challenging for these mobile devices to support these
computation intensive mobile applications.

1.2. Motivation

Cloud computing is a computing paradigm that has been widely accepted by the
Internet services as it is used to offload these computation intensive tasks to a remote
cloud server, thereby liberating these mobile devices from heavy computation
workloads. Thus, cloud computing can dramatically save the power consumption of the
mobile devices from heavy computation. One of the primary challenges associated with
cloud computing is the geographical distribution of these remote cloud servers.
Generally these remote cloud servers are situated at larger distances from the MDs.
Consequently, these long range data transmissions between the user (mobile devices
here) to the remote cloud servers may introduce latency (or delay) in our cloud
computing systems. This delay can impact the quality of user experience that the mobile
application intends to provide to the user. Therefore, we can conclude that cloud
computing is helping to reduce the energy consumption in mobile devices but it is
inducing extra latency in our systems.

As introduced earlier, we need to find a solution that helps improve user experience as
well as reduces battery (energy consumption). To overcome the drawback of cloud
computing architecture we bring these remote servers closer to the edge devices so
that we can avoid long range data transmissions, thereby reducing the latency in the
system. Mobile Edge Computing (MEC) is a computing paradigm that provides cloud
computing functionalities within the Radio Access Network (RAN). Using MEC we can
liberate our devices from heavy computation tasks by offloading these tasks (or a part
of) to a nearby MEC server [2]. MEC is based on short distance communication
architecture. When we push the functionalities of cloud towards the network edges in
 4

the vicinity of the users (MDs) using MEC, it has the potential to dramatically reduce
battery consumption and reduce the latency of the MEC system. It has been shown that
when the tasks are executed closer to the users, the efficiency of task execution is
greatly improved [4].

Figure 1: Task Offloading Scenario in MEC server

 5

2. Literature Survey
2.1 Literature Table

We have explored a few researches related to the problem statement which we have
summarized in the below table.

Table 1: Task offloading algorithms

Referenc Paper Author Year Algorith Evaluate Objectiv Drawback

e of m d e
publi Performa function
catio nce
n Metrics

[5] Energy-efficient You, C., 2017 Offloading Energy, To The algorithm

resource allocation for Huang, algo QoE minimize is centralized
mobileedge K., based on the and has a
TDMA weighted very high
computation offloading. Chae,
and sum complexity.
IEEE Transactions on H., Kim, OFDMA mobile
Wireless B. H. energy
Communications consump
tion

[6] The partial computation Zhou, 2020 Offloading Latency, To May incur
offloading strategy S., algo Energy minimize high
based on game theory Jadoon, based on system computation
game computa overhead.
for multi-user in mobile W.
theory tion
edge computing overhea
environment. Computer d
Networks

[7] UAVs for traffic Alioua, 2020 Offloading Latency, To The UAVs
monitoring: A A., algo Energy, reduce inflight energy
sequential game-based Djeghri, based on Cost computa usage was
game tion neglected.
computation H. E.,
theory delay
 6

offloading/sharing Cherif while

approach. M. E. T., optimizin
Senouci, g the
energy
S. M.,
and
Sedjelm computa
aci, H. tion cost

[8] Meta-heuristic search Xu, Y., 2020 Offloading Throughp To Low

based model for task Wang, algo ut maximiz convergence
offloading and time Y., Yang, based on e the under high
J. Metaheuri computa mobility.
allocation in mobile
stic tion rate
edge computing search of user

[9] Decision tree based Rego, P. 2017 Make To Single user,

approaches for A. L., Offloading offloading handle single-server
handling offloading Cheong, decision decision offloadin model, Not
E., making g easily
decisions and
Coutinh algo decision scalable
performing adaptive o, E. F., based on s and
monitoring in MCC Trinta, F. decision monitor
systems. A. M., tree offloadin
Hasany, g metrics
M. Z. et
al.

[10] Machine learning-based Eom, H., 2013 Computat To tackle Single-user

runtime scheduler for Juste, P. Offloading ion cost the Single server
mobile offloading S., algo adaptive
Figueire based on scheduli
framework
do, R., instance ng
Tickoo, based problem
O., learning in
Illikkal, offloadin
R. et al. g
framewo
rk

[11] Latency optimization Ren, J., 2018 Heuristic Latency To High

for resource allocation Yu, G., algo base minimize Complexity
in mobile edge Cai, Y., on the
He, Y. segmenta overall
computation offloading
tion system
strategy delay of
all
 7

device

[12] Code offloading using Majeed, 2017 SVM Energy, To Long Training
support vector machine A. A., Latency reduce Time
Khan, A. respons
e time
U. R.,
and
Ulamin, energy
R.,
Muham
mad, J.,
Ayub, S.

[13] Federated learning for Wang, 2020 Energy To High

energy-efficient task S., Offloading minimize Complexity
computing in wireless Chen, algorithm the total
M., based on energy
networks
Saad, SVM and consump
W., Yin, federated tion
C. learning
(FL)

2.2. Previous Work Done

A lot of research has been done in regard to this problem statement. Various Algorithms
have been proposed by the researchers, some of them are
1. EECO (Energy-efficient computation offloading) algorithm [14] - In the past,
researchers have focused on the design of an efficient computation offloading
mechanism for the MEC system by making use of the EECO algorithm.
2. Game theory [15] - There have been made use of game theory to design an
efficient computation offloading mechanism by making the use of game theory.
3. GCGH (Gini coefficient-based greedy heuristic) algorithm [16] - This algorithm
has been designed to solve the energy consumption minimization problem.

These are a few of many centralized optimization algorithms that have been designed
by the researchers.
 8

2.3. Research Gaps

1. It has been observed that the majority of the work that has been done in regard
to this problem statement is based on centralized algorithms.
2. Centralized algorithms generally require a central scheduler/operator that collects
information from every device and server in the network and then schedules the
task allocation.
3. Complexity wise this can be exponential.

Figure 2: Centralized Architecture

4. In contrast to the centralized architecture we have a distributed architecture

where allocation decisions are made by edge users themselves locally in a
reactive manner.
 9

Figure 3: Distributed Architecture

 10

3. Problem Statement
A key problem that needs to be solved in MEC systems is the efficient resource
allocation during the task offloading. In MEC systems, mobile devices have to basically
perform some part of their tasks on their own and offload the remaining part to a nearby
MEC server. Therefore during offloading an extra communication overhead is
introduced between a MD and its MEC server. Therefore to improve the quality of user
experience that the mobile applications intend to provide we need to balance the
computational and communication tradeoffs for a MD.

3.1. Objective

The problem statement focuses on resource allocation problems for computation

offloading. We will be assuming a small-cell network which consists of I MDs and J
SBSs. Each SBS is equipped with a MEC server with a finite computation capability.
Our objective is to minimize the weighted sum energy and latency of all the MDs.

Let us denote the total time taken (latency) to execute a task of a i’th mobile device be
TiTOL.Now this TiTOL is equal to TiMD + TiUL +TiMEC. A MD will execute some part of its task
by itself and upload the rest to a MEC server for its execution.
TiMD is the time taken by a MD to execute a part of its task, TiUL is the time taken by a
MD to upload a part of its task to a MEC server, TiMEC is the time duration for executing
the offloaded task of Ui at the MEC server.
Let us denote the whole energy consumption of i’th MD as EiTOL . Now this EiTOL is equal
to EiUL+EiMD.Where EiMD is equal to energy consumption of a mobile device caused by
local computations. EiUL is the energy consumption of a MD due to the transmission of
the offloaded task.

There now we can formulate our objective function (F) to be minimized as follows:-
F = ΣαTiTOL + βEiTOL —-- (1)
where Σ is the sum over all the I mobile devices and where α, β ∈ [0, 1] denote the
weighting parameters balancing the energy and time consumption, and we have α+β =
1.
 11

4. Methodology
We will be using a distributed method for efficient task offloading in MEC systems. The
network will consist of multiple MDs and multiple SBSs (small-cell base stations). There
will be two constraints in our MEC system. The first constraint will be that each SBS will
be equipped with a single MEC server, and the MEC server will have limited (finite)
computational capacity. The second constraint is that each mobile device can associate
with only one MEC server.
Let us assume our MEC network to consist of I MD’s and J SBS where each SBS is
equipped with a MEC server.
Mathematically these constraints can be represented as follows:
1. Constraint C1: Σλi,jQiCi≤Kj —--(2)
Here Σ is sum over all the I mobile devices.
Qi is the number of bits in the task to be executed for an ith mobile device.
Ci is the number of CPU cycles required to execute 1 bit.
λi,j is the fraction of the task that a ith MD offloads at a jth MEC server.
Kj is the task upper bound of Bj ’s MEC server within a time slot.
Therefore, C1 states that in a given time slot jth MEC can handle at max Kj CPU
cycles

2. Constraint C2: λi,j*λi,l = 0 where j ≠ l, l is between 1 and J. —--(3)

C2 states that a MD can offload its task to at most one MEC server.

We first formulate an optimization problem and establish the objective function F as

defined in equation (1), i.e., the minimum value of weighted sum of mobile energy and
time consumption, subject to each SBS’s computing capacity. The parameters to be
optimized are the allocations of the MD’s tasks to be offloaded to the MEC servers.
Then we develop the factor graph according to the network topology and decompose
the object function into multiple local utilities to fit the factor graph. We transform local
utilities into the estimations of marginal distributions and propose a distributed BP
algorithm to solve the estimations. In our proposed BP, we enable the majority of the
computationally intensive calculations to be executed on the MEC servers of the SBSs
rather than on the MDs.
 12

5. Implementation
1. After defining our system and constraints now we develop the factor graph for the
system. The factor graph will consist of some variable nodes and some factor
nodes. For example a system with 2 SBS and 3 MD can be visualized as follows:

Figure 4: Connectivity between SBSs and devices

Now variable nodes in our graph will be the offloading row vectors (λi for ith MD)
of each MD’s. For formulation of factor nodes we will break down our objective
function F (obtained in eq 1 ) into J (number of SBS) local utility functions (F1,
F2, ……FJ). Now we define a new local utility function G such that if Fi obeys all
the constraints as mentioned in step2, then Gi = Fi else Gi = -∞. Thus G will
serve as our factor node.
So the factor graph for above depicted system is as follows:

Figure 5: Corresponding factor graph model

 13

2. Now we apply a belief propagation algorithm to solve the optimization problem.

The key components of the distributed belief propagation algorithm are as
depicted in the flowchart as follows.

3. Analysis of simulation results and comparing it with local computing model,

greedy algorithms etc.

5.1 Work done till date

The project was divided into 4 modules.

1. Module 1: Understanding of the problem statement and research on how
distributed algorithms are better when compared to centralized algorithms.
2. Module 2: Understanding the system model of the MEC system. Defining
objective and representing objective in terms of system parameters. This module
has been the major focus till date as calculating energy consumptions and
latency of the system involved a good understanding of mathematical concepts.
 14

This module also involved study of the belief propagation algorithm and how it is
better in calculating marginal probabilities as compared to other centralized
algorithms.
3. Module 3: Work on the implementation of the belief propagation algorithm to
solve the optimisation problem. Simulate and analyze the results.
4. Module 4: Working on publishing a research paper.

The progress made till date is as follows:

1. Module 1: Completed
2. Module 2: Completed
3. Module 3: In progress. We have implemented a basic python code that
calculates the marginal probabilities using BP algorithm. Now we will work on
inculcating the system parameters into this algorithm and analyzing the result.
(Expected to be completed by 10th of November.)
4. Module 4: Not started yet. We will work on our research paper once we are done
with the implementation.

5.2 Future work

1. Implementation and Analysis Part: We will be working now majorly on

implementing the belief propagation algorithm with respect to our system
parameters. We will be calculating and analyzing results of the algorithm on
small-sized networks (with 3 MD and 2 SBS) and medium sized networks (with 8
MD and 5SBS) respectively. We would be also analyzing and comparing results
with different scenarios like when the MD performs all computations on its own,
some greedy algorithms etc.
2. Research paper: After completing the implementation part and analyzing the
simulation results we would be then working on publishing our research paper.
 15

6. References
[1] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li, “Mobile-edge computing: Partial
computation offloading using dynamic voltage scaling,” IEEE Transactions on
Communications, vol. 64, no. 10, pp. 4268–4282, Oct. 2016.
[2] ETSI, Sophia Antipolis, France, “Mobile-edge computing-introductory technical white
paper,” http://portal.etsi.org/portals/0/tbpages/mec/ docs/mobile-edge computing
introductory technical white paper v1\ %2018-09-14.pdf, [Online].
[3] D. Calin, H. Claussen, and H. Uzunalioglu, “On femto deployment architectures and
macrocell offloading benefits in joint macro-femto deployments,” IEEE Comm. Mag., vol.
48, no. 1, pp. 26–32, Jan. 2010.
[4] M. Sstyanarayanan, P. Bahl, R. Caceres, and N. Davies, “The case for vm-based
cloudlets in mobile computing,” IEEE Pervasive Comput., vol. 8, no. 4, pp. 14–23,
Oct.-Dec. 2009.
[5] You, C., Huang, K., Chae, H., Kim, B. H. (2017). Energy-efficient resource allocation
for mobileedge computation offloading. IEEE Transactions on Wireless
Communications, 16(3), 1397–1411. DOI 10.1109/TWC.2016.2633522.
[6] Zhou, S., Jadoon, W. (2020). The partial computation offloading strategy based on
game theory for multi-user in mobile edge computing environment. Computer Networks,
178(2020), 107334. DOI 10.1016/j.comnet.2020.107334.
[7] Alioua, A., Djeghri, H. E., Cherif M. E. T., Senouci, S. M., Sedjelmaci, H. (2020).
UAVs for traffic monitoring: A sequential game-based computation offloading/sharing
approach. Computer Networks, 177, 107273. DOI 10.1016/j.comnet.2020.107273.
[8] Xu, Y., Wang, Y., Yang, J. (2020). Meta-heuristic search based model for task
offloading and time allocation in mobile edge computing. Proceedings of the 2020 6th
International Conference on Computing and Artificial Intelligence, pp. 117–122. Tianjin,
China.
[9] Rego, P. A. L., Cheong, E., Coutinho, E. F., Trinta, F. A. M., Hasany, M. Z. et al.
(2017). Decision tree based approaches for handling offloading decisions and
performing adaptive monitoring in MCC systems. Proceedings of 5th IEEE International
Conference on Mobile Cloud Computing, Services, and Engineering, pp. 74–81. San
Francisco, CA, USA.
[10] Eom, H., Juste, P. S., Figueiredo, R., Tickoo, O., Illikkal, R. et al. (2013). Machine
learning-based runtime scheduler for mobile offloading framework. Proceedings of 2013
IEEE/ACM 6th International Conference on Utility and Cloud Computing, pp. 17–25.
Dresden, Germany.
[11] Ren, J., Yu, G., Cai, Y., He, Y. (2018). Latency optimization for resource allocation
in mobileedge computation offloading. IEEE Transactions on Wireless Communications,
17(8), 5506–5519. DOI 10.1109/TWC.2018.2845360.
 16

[12] Majeed, A. A., Khan, A. U. R., Ulamin, R., Muhammad, J., Ayub, S. (2017). Code
offloading using support vector machine. 6th International Conference on Innovative
Computing Technology, pp. 98–103. Dublin, Ireland
[13] Wang, S., Chen, M., Saad, W., Yin, C. (2020). Federated learning for
energy-efficient task computing in wireless networks. IEEE International Conference on
Communications, pp. 20–25. Dublin, Ireland.
[14] K. Zhang, Y. Mao, S. Leng, Q. Zhao, L. Li, X. Peng, L. Pan, S. Maharjan, and Y.
Zhang, “Energy-efficient offloading for mobile edge computing in 5g heterogeneous
networks,” IEEE Access., vol. 4, pp. 5896–5907, Aug. 2016.
[15] ] X. Chen, L. Jiao, W. Li, and X. Fu, “Efficient multi-user computation offloading for
mobile-edge cloud computing,” IEEE/ACM Transactions on Networking, vol. 24, no. 5,
pp. 2795–2808, Oct. 2016.
[16] P. Zhao, H. Tian, C. Qin, and G. Nie, “Energy-saving offloading by jointly allocating
radio and computational resources for mobile edge computing,” IEEE Access., vol. 5,
pp. 11 255–11 268, Jun. 2017.