DM2-ECOP: An Efficient Computation Offloading Policy

for Multi-user Multi-cloudlet Mobile Edge Computing

Environment

HOUSSEMEDDINE MAZOUZI and NADJIB ACHIR, L2TI, Institut Galilée, Université Paris 13,
Sorbonne Paris Cité, France
KHALED BOUSSETTA, L2TI, Institut Galilée, Université Paris 13, Sorbonne Paris Cité,
Agora/INRIA, France

Mobile Edge Computing is a promising paradigm that can provide cloud computing capabilities at the edge of
the network to support low latency mobile services. The fundamental concept relies on bringing cloud com-
putation closer to users by deploying cloudlets or edge servers, which are small clusters of servers that are
mainly located on existing wireless Access Points (APs), set-top boxes, or Base Stations (BSs). In this article,
we focus on computation offloading over a heterogeneous cloudlet environment. We consider several users
with different energy—and latency-constrained tasks that can be offloaded over cloudlets with differentiated
system and network resources capacities. We investigate offloading policies that decide which tasks should
be offloaded and select the assigned cloudlet, accordingly with network and system resources. The objective
is to minimize an offloading cost function, which we defined as a combination of tasks’ execution time and
mobiles’ energy consumption. We formulate this problem as a Mixed-Binary Programming. Since the cen-
tralized optimal solution is NP-hard, we propose a distributed linear relaxation-based heuristic approach that
relies on the Lagrangian decomposition method. To solve the subproblems, we also propose a greedy heuris-
tic algorithm that computes the best cloudlet selection and bandwidth allocation following tasks’ offloading
costs. Numerical results show that our offloading policy achieves a good solution quickly. We also discuss the
performances of our approach for large-scale scenarios and compare it to state-of-the-art approaches from
the literature.
CCS Concepts: • Networks → Cloud computing; • Human-centered computing → Ubiquitous and
mobile computing systems and tools; • Theory of computation → Network optimization;
Additional Key Words and Phrases: Computation offloading, mobile cloud computing, mobile edge comput-
ing, cloudlet, Lagrangian decomposition
ACM Reference format:
Houssemeddine Mazouzi, Nadjib Achir, and Khaled Boussetta. 2019. DM2-ECOP: An Efficient Computation
Offloading Policy for Multi-user Multi-cloudlet Mobile Edge Computing Environment. ACM Trans. Internet
Technol. 19, 2, Article 24 (April 2019), 24 pages.
https://doi.org/10.1145/3241666

Authors’ addresses: H. Mazouzi and N. Achir, L2TI, Institut Galilée, Université Paris 13, Sorbonne Paris Cité, 99 Avenue J-B
Clement, Villetaneuse, 93430, France; emails: {mazouzi.houssemeddine, nadjib.achir}@univ-paris13.fr; K. Boussetta, L2TI,
Institut Galilée, Université Paris 13, Sorbonne Paris Cité, Agora/INRIA, 99 Avenue J-B Clement, Villetaneuse, 93430, France;
email: [email protected].
Publication rights licensed to ACM. ACM acknowledges that this contribution was authored or co-authored by an em-
ployee, contractor or affiliate of a national government. As such, the Government retains a nonexclusive, royalty-free right
to publish or reproduce this article, or to allow others to do so, for Government purposes only.
© 2019 Copyright held by the owner/author(s). Publication rights licensed to ACM.
1533-5399/2019/04-ART24 $15.00
https://doi.org/10.1145/3241666

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019. 24
24:2 H. Mazouzi et al.

1 INTRODUCTION
The Mobile Cloud Computing paradigm has been proposed to allow remote execution of resource-
hungry mobile applications in the cloud. The application’s computation is then transmitted to
the remote cloud to be performed. The latter operation is known as computation offloading
[6, 7]. Unfortunately, the geographical distance between the cloud and user can introduce large
and variable latency. That can significantly degrade the quality of experience of delay-sensitive
applications, such as mobile gaming, augmented-reality, and face and speech recognition [8, 30].
To overcome such problems, Mobile Edge Computing (MEC) has emerged as a main paradigm
that aims to provide cloud computing capabilities at the edge of the network to support latency-
sensitive mobile applications. The main concept relies on deploying small clusters of servers, called
cloudlets, at the edge of the network [11, 18]. Users can then offload their computation to closer
cloudlets.
In multi-user context, several mobile devices can compete to offload their computations to the
cloudlets. Hence, the performances of offloading policies are strongly dependent on the cloudlets’
computational resources sharing and on the wireless bandwidth allocation strategies [4, 5, 12]. In
addition, in a multi-cloudlet MEC environment, where many cloudlets are available around users,
the performance of the computation offloading depends on the cloudlet selection [14, 33, 34].
Many recent works have investigated cloudlet selection problems [21, 27, 34]. Most of the pro-
posed offloading policies rely on user density to statically assign each region to a cloudlet [15, 27].
Therefore, as shown in Figure 1, users within a region will always offload to the same cloudlet.
Nevertheless, the dynamic density of users may imbalance the load between the cloudlets, leading
to suboptimal MEC capacities usage and longer offloading delays. Therefore, to achieve high per-
formance, an offloading policy must jointly consider bandwidth allocation, computation resource
allocation, and cloudlet selection.
To tackle this problem, we explore, in this article, computation offloading in multi-user, multi-
cloudlet MEC. Our aim is to provide an efficient offloading policy, which determines the best of-
floading decision and cloudlet selection for each user, with the aim of reducing the total offloading
cost.
This work presents a new computation offloading policy named Distributed Multi-user Multi-
cloudlet Efficient Computation Offloading Policy (DM2-ECOP), which aims to improve the per-
formance of offloading in an MEC environment. It extends the offloading strategies presented in
References [4, 5, 12]. As in previous works, we assume that each user executes only one applica-
tion at a time. However, unlike these works, we define two categories of applications that can be
supported by the MEC: (1) applications that must be performed remotely and (2) applications that
can be performed either locally or remotely, accordingly with the conditions at execution time.
DM2-ECOP tries to select the best cloudlet according to network and system resource availability,
while minimizing the offloading cost. The offloading cost is defined as a combination of the energy
consumed by the mobile devices and the total applications’ completion times.
We formulate this computation offloading problem as a Mixed Binary Programming. Then, we
solve it using a distributed linear relaxation-based heuristic that follows the Lagrangian decompo-
sition approach. DM2-ECOP is composed of two decision levels: (1) The local offloading manager
handles the users associated within the same AP and solves the offloading subproblem related to
this AP; the local offloading manager uses our proposed algorithm, named Greedy Best Cloudlet
Selection First Heuristic (GBC-SFH), which selects the cloudlet to which each application will be
offloaded to minimize the energy consumption and completion times. (2) At a second level, a global
offloading manager ensures that the cloudlets’ resources allocated by each local offloading man-
ager satisfy the capacity constraints of each cloudlet.

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
Efficient Computation Offloading for Mobile Edge Computing 24:3

Fig. 1. Multiuser computation offloading in multi-cloudlet MEC environment.

The remainder of this article is organized as follows: Section 2 presents existing works in com-
putation offloading. Section 3 introduces the system modeling. The multi-user, multi-cloudlet of-
floading problem is formulated in Section 4. Our offloading policy, named DM2-ECOP, is explained
in Section 5. Performance evaluation is detailed and analyzed in Section 6. Finally, a conclusion is
drawn in Section 7.

2 RELATED WORK
Many works were proposed to explore computation offloading to improve the performance of
mobile devices. However, not all of the proposed offloading policies have the same goals. In the
following, we distinguish between three main goals: (i) offloading decision, (ii) cloudlets placement,
and (iii) cloudlet selection.
Some of the proposed works investigate the offloading decision to decide which computation
should be offloaded to the remote cloud, such as: Meng-Hsi Chen et al. [4], Xu Chen et al. [5],
Songtao Guo et al. [12], Keke Gai et al. [10], Yuyi Mao et al. [22], and Dong Huang et al. [13].
Meng-Hsi et al. are one of the first to work on multi-user computation offloading in mobile cloud
computing. The proposed offloading policy determines which computation must be performed in
the remote cloud and which one must be performed locally by the mobile device. Then, it allocates
the wireless bandwidth to each user to reduce the energy consumption of the mobile device. The
Xu Chen et al. offloading policy was designed for a single cloudlet MEC environment. Each user
tries to offload its computation, accordingly with the available wireless bandwidth to reduce the
energy consumption. Another offloading approach for multi-user was presented by Songtao Guo
et al. Similar to DM2-ECOP, this work minimizes an offloading cost defined as a combination of
energy consumption and processing time. The offloading policy decides which computation can
be offloaded and allocates the wireless bandwidth and the processor frequency to each offloaded

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
24:4 H. Mazouzi et al.

computation. In similar way, Keke Gai et al. proposed a scheduler to assign the tasks between the
local mobile device and the remote cloud to save energy consumption. Yuyi Mao et al. presented an
offloading policy that tries to offload the computation in a multi-user scenario to an edge server,
cloudlet. The proposed policy allocates the CPU frequency and the bandwidth to each user to
reduce the energy consumption of the mobile device. Lastly, Dong Huang et al. designed a compu-
tation offloading policy for a single-user scenario to reduce the energy consumption of the mobile
device. They focused on partial offloading, where the offloading policy partitions the application at
runtime to determine which computation must be performed locally and which must be offloaded
to a remote server. Although all these policies improve the performances of the mobile device,
they rely on an unlimited capacity of the cloud. Consequently, they need some enhancements to
be applied for the MEC, where cloudlets have limited computing resources.
Yucen Nan et al. [25, 26] and Chongyu Zhou et al. [35] proposed computation offloading poli-
cies to reduce the energy consumption of fog nodes. They introduced an offloading policy where
the fog nodes try to offload their computation to the remote cloud. For each fog node, the policy
decides which computation must be offloaded to the remote cloud and which one must be per-
formed locally by the node. In Reference [26], the offloading policy has been extended to reduce
the completion time of the applications. Similarly, Chongyu Zhou et al. introduced an online of-
floading policy. It can select the computations that should be performed by the nearest cloudlet
in order to minimize a system-wide utility, which is the execution time. Contrary to these poli-
cies, which reduce the energy consumption of the fog server, DM2-ECOP focuses on the reduction
of the offloading cost on the mobile device side. In addition, the IoT device has a tiny comput-
ing capacity that cannot perform any application. However, the mobile device has considerable
computing capacity that performs complex applications.
Cloudlets placement is also a challenging issue for MEC, and many recent works propose some
cloudlets placement heuristics in MEC environment. Mike Jia et al. [14, 33], Hong Yao et al. [34],
and Longjie Ma et al. [21] introduced cloudlets placement and selection algorithms in a multi-user,
multi-cloudlet MEC environment. The Mike Jia et al. offloading approach is one of the first heuris-
tics on cloudlets placement in a large-scale environment. Its main goal is to find the best cloudlets
placement in a large network, then select a cloudlet to perform the computation of each AP. The
K-median clustering based on user density is used to place the cloudlets. Then each AP is stati-
cally assigned to a cloudlet. Similarly, Hong Yao et al. have been designing heuristics to support
heterogeneous cloudlets environment. Finally, Longjie Ma et al. have been introducing a heuris-
tic to find the minimal number of cloudlets that must be placed to improve the user experience
quality in a large-scale network. In a multi-user MEC environment, the density of mobile users is
dynamic and changes over time. So, static assignment of the APs to cloudlets may decrease the
performance of the computation offloading. To avoid this problem, our DM2-ECOP approach con-
siders dynamic cloudlet selection and wireless bandwidth allocation with the aim of minimizing
energy consumption and improving the performance of mobile devices.
Other works try to find a dynamic cloudlet selection in a multi-cloudlet MEC environment.
Anwesha Mukherjee et al. [24, 27], Mike Jia et al. [15], Qiliang Zhu et al. [36], and Arash
Bozorgchenani et al. [2] have proposed to support the dynamic cloudlet selection to reduce the of-
floading cost. Anwesha Mukherjee et al. designed a multilevel offloading policy to optimize energy
consumption. The users offload to the nearest cloudlet in the first step. According to the amount of
resources available in this cloudlet, it can perform the task or offload it to another cloudlet. Mike
Jia et al. introduced a heuristic to balance the load between the cloudlet. Its main goal is to migrate
some computations from overloaded cloudlets to underloaded cloudlets to reduce the execution
time. Similarly, Qiliang Zhu et al. developed a two-tier offloading policy, where the mobile device
offloads its computation to an offloading server based on the resource availability. They used an

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
Efficient Computation Offloading for Mobile Edge Computing 24:5

agent that decides to perform the computation in the local cloudlet or to offload it to the remote
cloud. Arash Bozorgchenani et al. offloading policy tries to select a nearby fog node to offload
some computation of a busy fog node to save energy consumption and completion time. Even
these works proposed dynamic cloudlet selection heuristics, the tasks still always offloaded to the
nearest cloudlet that decides to perform them locally or transmit them to other cloudlets. Thus, an
additional offloading cost is induced; consequently, the performance of offloading will decrease. In
our proposal, the cloudlet that performs each offloaded task will be determined at the offloading
decision time without any additional cost.
All the offloading policies presented above have been focusing on reducing the offloading cost.
They offload computations to a predetermined remote server (the remote cloud or a local cloudlet).
The selection of the remote server is done statically at the development time based on metrics such
as user density, despite the fact that the density of users can change dynamically over the time.
In addition, the computing capacity of the cloudlet is limited and cannot perform all the offloaded
computation. To avoid this situation, the most adopted strategy in the literature was to consider
a two-tier approach. Basically, the tasks are offloaded to the nearest cloudlet, and this cloudlet
offloads some computation to another cloudlet or to the remote cloud when it is overloaded.
Although two-tier offloading policies can improve the performance of the offloading approach,
they engender an additional offloading cost. Moreover, in a multi-cloudlet scenario, where many
cloudlets are available around the user, selecting the same cloudlet always is not the best strategy.
Therefore, in this article, we propose a new offloading policy to improve the efficiency of com-
putation offloading in MEC. The new policy must consider many cloudlets for which a user can
offload its computation and compute optimal computation placements to optimize the offloading
cost.

3 SYSTEM DESCRIPTION
In this section, we describe our system modeling. We first introduce the MEC model, then we
present the communication and computation offloading models. Finally, an offloading cost is pro-
posed as an objective function for our optimization problem. Table 1 presents variables and nota-
tions used, in this article, to model our multi-user, multi-cloudlet computation offloading problem.

3.1 MEC Environment Model

Let us consider an MEC environment composed of M APs and K cloudlets, as illustrated in Figure 2.
We suppose that the number of cloudlets is less than the number of APs (K ≤ M ). In this article,
we assume that the cloudlets have already been deployed and are co-located with the APs. We
also consider that the users can communicate with the cloudlets through their APs. We denote
in the following the set of APs by M = {1, 2, . . . , M }, and we assume that each AP i is associated
with Ni users. Let us consider Nm = {1, 2, . . . , Nm } as the set of users associated with the mt h AP
and K = {1, 2, . . . , K } as the set of the cloudlets. We also define um,n as to nt h user of the mt h AP.
Similar to existing works [14, 22, 33], every user runs one application on his mobile device. The
application is characterized by its: (i) computational resource requirement in terms of CPU cycles,
denoted by γum, n , (ii) the amount of data uploaded to MEC, denoted by upum, n , (iii) the amount of
data that must be downloaded by the user from MEC at the end of execution on the MEC, denoted
by dw um, n , and (iv) finally, the maximum tolerated delay according to the Quality of Service (QoS)
required by the application, denoted by tum, n .
As considered in previous works [11, 19], we distinguish between two categories of computation
offloading applications: (1) static offloading decision task and (2) dynamic offloading decision task. In
the first category, the application is partitioned in advance at the design time between: a local part
(task) that should always be executed in the mobile and a remote part (task) that should always

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
24:6 H. Mazouzi et al.

Table 1. Notation

Symbol Description
γum, n the computational resource required by the task of the user um,n .
λm the Lagrangian multiplier of the subproblem m.
ck the computing resource allocation on cloudlet k.
dwum, n the amount of data to download by the user um,n from the MEC.
Eul m, n the local energy consumption for user um,n .
Eue m, n ,k the remote energy consumption for user um,n in cloudlet k.
Fk the computing capacity of the cloudlet k.
fum, n the local computing capacity of the user um,n .
K the number of cloudlets available in the network.
M the number of APs in network.
Nm the number of users associated to the AP m.
Put xm,/rn x power consumption when the Wi-Fi interface is transforming or receiving data.
PuI dl e
m, n
power consumption when the Wi-Fi interface is not transforming or receiving data.
rum, n the task of the user um,n .
tum, n the maximum tolerated delay according the QoS of the task of the user um,n .
Tutm, n ,k the communication time when user um,n offload to cloudlet k.
Tul m, n the local processing time for user um,n .
Tuem, n ,k the remote processing time for user um,n in cloudlet k.
um,n the nth user for the mth AP.
upum, n the amount of data uploaded to the MEC from the user um,n .
Wm the wireless data rate at the AP m.
wum, n the allocated data rate for the user um,n .
xum, n ,k the offloading decision variable for the task of user um,n in the cloudlet k.
yum, n the category to which belong the tasks (1 static offloading decision task and 0 otherwise).
Zul m, n the local offloading cost for user um,n .
Zuem, n ,k the remote offloading cost for user um,n on cloudlet k.

be executed remotely. As illustrated in Figure 3(a), the task’s source code is already in the remote
server, so the mobile device needs to transmit only the input data to the remote server. A typical
example of static offloading decision task is the FLUID application on Android [16] that is used
for particle simulations. The thin client side of FLUID is executed on the mobile, while the server
part must be performed remotely in MEC, because it requires high-performance GPU computing
processors that are not commonly available in mobile devices.
In the second category, the application needs to be partitioned at runtime accordingly with the
network and MEC resource availability. Basically, the mobile terminal needs to decide if it is useful
to execute the task on the mobile or to offload all or part of the task. In this case, as illustrated in
Figure 3(b), the mobile device must transmit its source code and the input data when the task
is offloaded. An example of this kind of application is the Linpack benchmarks [32] for Android,
which aims to measure the performances of Android devices. This application can be either totally
executed on the mobile terminal or partially offloaded to a cloudlet.
To simplify the analysis, we model in this article both static offloading decision tasks and dy-
namic offloading decision tasks as tasks with an offloading computation ratio noted as aum, n . Basi-
cally, aum, n denotes the computation ratio of the application that should be executed remotely. To

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
Efficient Computation Offloading for Mobile Edge Computing 24:7

Fig. 2. Example of a multi-user (MU), multi-cloudlet MEC environment with 4 APs and 2 cloudlets (M = 4,
K = 2).

Fig. 3. Illustration of the two categories of tasks.

distinguish between static offloading decision tasks and dynamic offloading decision tasks, we con-
sider that for static offloading decision tasks the offloading ratio is always equal to 1, which means
that the tasks that belong to that category are always offloaded. The local part is considered equal
to zero, because it will not affect the performance of the system, since it should always be executed
at the terminal. However, for dynamic offloading decision tasks, this offloading computation ratio
can take any value between 0 and 1 (i.e., aum, n ∈ [0, 1]). In addition, for simplicity (but without the
loss of generality), we also assume that when a task is offloaded with a given offloading computa-
tion ratio, then the amount of data that should be transmitted are also proportional to that ratio.
However, the output of the task, noted as dw um, n , does not change whatever the value of aum, n .
To indicate which category the application of user um,n belongs, we introduce the binary vari-
able yum, n , which is equal to 1 for static offloading decision tasks and 0 for dynamic offloading
decision tasks.

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
24:8 H. Mazouzi et al.

Finally, due to hardware and software constraints required by the task, we assume that some
cloudlets cannot perform some tasks. In this case, we define a second binary variable, дum, n ,k , to
indicate if the cloudlet k can perform the task. Thus, дum, n ,k is equal to 1 if the k t h cloudlet can
execute the task, 0 otherwise.

3.2 Communication Model

Let G = (V , E) be a weighted graph, where V (M ∪ K ) is a finite set of vertices corresponding to
the sets of APs and cloudlets and E is a set of connections (edges) denoting a possible communica-
tion between any two vertices. We also consider a weight of the edge, noted as ei, j , that represents
the network delay between each two vertices i and j. Thus, using the Dijkstra algorithm, we can
compute a delay matrix, noted as Dm,k , that represents the delay between the mt h AP and the k t h
cloudlet. According to the last considerations, we can estimate the bandwidth allocated to each
user as follows:
Wm (πm )
wum, n = , (1)
πm
where πm is the offloading capacity that represents the number of users that offload their ap-
plications to MEC, as defined in Definition 5.1. Wm (πm ) is the bandwidth shared by the mt h AP
between the πm users associated with it. To estimate this bandwidth, we consider the Bianchi
model [1, 20].
According to the last assumptions, we can compute the communication time of any task. This
time, noted as Tutm, n ,k , is composed of time needed to upload the data from the user terminal to the
cloudlet, plus the time needed to download the results from the cloudlet once the task is completed.
In this case, Tutm, n ,k can be written as:

upum, n dwum, n
Tutm, n ,k = aum, n · + Dm,k + + Dm,k
wum, n wum, n
aum, n · upum, n + dwum, n
= + 2Dm,k . (2)
wum, n

3.3 Computation Processing Model

a) Local processing: We assume that the user’s device has a local computational capability
of fum, n used for the task computation. When a user offloads a percentage aum, n of its
computation, the remaining part must be performed locally. So, the local processing time
for the local part can be estimated by:
γum, n
Tul m, n = (1 − aum, n ) · . (3)
fum, n
From the above equation, we can notice that the local processing time of a Static offload de-
cision task is equal to zero (Tul m, n = 0), since aum, n for static offload decision task is always
equal to 1.
b) Remote processing: For remote computation, we consider that each cloudlet has a compu-
tational capability of Fk . Every cloudlet allocates some of its computational resource to
perform the offloaded tasks. So, the remote execution time of the user’s um,n task can be
estimated as the ratio of the offloaded computational resources and the allocated compu-
tation resource, as illustrated by the following equation:
γum, n
Tuem, n ,k = aum, n · , (4)
ck

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
Efficient Computation Offloading for Mobile Edge Computing 24:9

where c k is the amount of computational resource allocated to perform the task at the k t h
cloudlet. To simplify our model, we consider that the allocated resource at each cloudlet is
fixed and does not change during the computation [4, 5, 12].

3.4 Offloading Cost Model

We define the offloading cost as a combination of the energy consumption and the execution time
of the application. In the following, we present both local offloading cost and remote offloading
cost.
a) Local offloading cost: According to the forgoing considerations, the local offloading cost is
expressed as a combination of the energy consumption and the execution time of the tasks,
as in the following:

Zul m, n = βum, n · Eul m, n + (1 − βum, n ) · Tul m, n , (5)

where Eul m, n and Tul m, n are, respectively, the total amount of energy and processing time of
the local part of the application of the user um,n . βum, n denotes the weighting parameter of
execution time and energy consumption of the user’s offloading decision. When the battery
of the user’s device is at a low state, and the user needs to reduce the energy consumption,
then it can set βum, n = 1. However, when a delay-sensitive task is running, the user can set
βum, n = 0 to give more priority to the execution time. To get a more flexible cost model,
we allow a multi-criteria offloading policy by considering energy consumption, execution
time, or a combination of both of them.
Considering the energy-consumption model, we use in this article the model proposed
in References [3, 23]. Using this model, we can compute Eul m, n as following:

Eul m, n = κ · ( fum, n ) 3 · Tul m, n (6)

where κ is the effective switched capacitance, which depends on the chip architecture and
is used to adjust the processor frequency. In the following, we set κ = 10−9 as shown in
[3, 23].
b) Remote offloading costs: The total amount of energy consumed by the user’s terminal to
perform the task remotely is equal to the energy used when the device turns the radio in
the transmission mode to send the data to the remote server, plus the energy used when
the device turns the radio in idle mode to wait the task completion plus the energy used
by the device when it turn again the radio in the reception mode to receive the result data
from the remote server. This consumed energy can be expressed as follows:
   
Eue m, n ,k = Put xm,/rn x · Tutm, n ,k − 2Dm,k + Puidl e
m, n
· Tuem, n ,k + 2Dm,k , (7)

where Put xm,/rn x is the power consumption when the radio interface is set to transmission or
reception mode, and Puidl e is the power consumption in the case when the radio interface
m, n
is set to idle mode [3, 23, 29].
Finally, we can define the remote offloading costs as follows:
 
Zuem, n ,k = βum, n · Eue m, n ,k + (1 − βum, n ) · Tutm, n ,k + Tuem, n ,k (8)

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
24:10 H. Mazouzi et al.

4 MULTI-USER, MULTI-CLOUDLET COMPUTATION OFFLOADING PROBLEM

FORMULATION AND DECOMPOSITION
To propose an efficient offloading policy, we formulate the problem as an optimization problem.
Then, we use Lagrangian relaxation to decompose the problem into subproblems and solve each
one separately.

4.1 Problem Formulation

As introduced earlier, the objective of this article is to propose an efficient offloading policy. It de-
cides which users should offload their tasks, determines the amount of computation to offload,
and selects a cloudlet for each user, while minimizing the total offloading cost. Let us denote
xum, n ,k to the offloading decision for the task of the user um,n on the cloudlet k, which means that
xum, n ,k = 1 if the user um,n offloads its task to the cloudlet k, 0 otherwise. Given the system descrip-
tion and according to the QoS and cloudlets’ resource-capability constraints, the can be formulated
as follows:

M 
Nm
Minimize Zum, n
m n
Subject to:

K
C1 : xum, n ,k ≤ 1, ∀m ∈ M, um,n ∈ Nm
k=1

K
C2 : yum, n − xum, n ,k ≤ 0, ∀m ∈ M, um,n ∈ Nm
k=1
C3 : Tum, n ≤ tum, n , ∀m ∈ M, um,n ∈ Nm
C4 : xum, n ,k ≤ дum, n ,k , ∀m ∈ M, um,n ∈ Nm , k ∈ K
M  Nm
C5 :  xum, n ,k · c k  ≤ Fk , ∀k ∈ K
m  n 
C6 : xum, n ,k ∈ {0, 1}, ∀m ∈ M, um,n ∈ Nm , k ∈ K
C7 : aum, n ∈ [0, 1], ∀m ∈ M, um,n ∈ Nm
C8 : aum, n ≥ yum, n , ∀m ∈ M, um,n ∈ Nm , (9)

where Zum, n is the offloading cost of the user’s um,n task. Zum, n can be expressed by the following
formula:

K
Zum, n = Zul m, n + xum, n ,k · Zuem, n ,k (10)
k=1

As indicated in the problem formulation, our objective is to minimize the total offloading cost of
the users of the network. The first constraint (C1) ensures that each task is assigned to one cloudlet
at most. Constraints (C2) guarantee that any static offloading decision task must be assigned to
exactly one cloudlet and a dynamic offloading decision task may be assigned to one cloudlet at
most. The next constraint (C3) shows that the QoS required by the task, in terms of completion
time, must be less than a given threshold. The threshold is obtained based on the characteristics of
the mobile application [5, 7]. For example, for an interactive application, the user’s perception—the
duration of the submission of the task until receiving the response—is a well-used technique to

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
Efficient Computation Offloading for Mobile Edge Computing 24:11

determine the threshold [8]. Completion time can be expressed as follows:


K  
Tum, n = Tul m, n + xum, n ,k · Tutm, n ,k + Tuem, n ,k (11)
k=1

The next constraint (C4) ensures that every offloaded task must be performed by a cloudlet
that meets the hardware and software required by the task. The constraint C5 shows that it is
not possible to exceed the computing capacity of each cloudlet. Constraint C6 ensures that any
decision variable is a binary variable.
Finally, constraints (C7) indicate that the ratio of the offloaded computation must be a real value
between 0 and 1. And, (C8) ensures that each static offloading decision task must always be entirely
offloaded.
Theorem 4.1. Equation (9) is a Non-Linear Mixed Binary Problem (NLMBP) with an exponential
function and constraints. It is an NP-hard problem.
Proof. Let us consider a special case where the same number of users are associated to each AP
and all tasks are static offloading decisions. So, all the tasks must be offloaded to the cloudlets and
the bandwidth allocated to each user is known in advance. Thus, the special case is a Linear Binary
Integer Problem (LBIP). In fact, this special case can be easily reduced to the General Assignment
Problem (GAP) with assignment restrictions, which is NP-hard, as shown in Reference [17]. Since
the special case is NP-hard, Equation (9) is also NP-hard. 
Considering the NP-hardness of the problem, it is difficult to achieve an optimal solution. Next,
we propose a simplification version of Equation (9) using Lagrangian relaxation and decomposition
approaches.

4.2 Problem Decomposition

To solve the above problem, we need a decomposition approach. Decomposing a complex opti-
mization problem consists of breaking it up into smaller ones, called subproblems, and solving
each of the smaller ones separately. Unfortunately, the constraint C5 is considered a complicat-
ing constraint [9, 31], since it involves the local variables of more than one subproblem. Conse-
quently, the decomposition of Equation (9) does not work in one step and the subproblem cannot
be solved independently. For these kinds of complex problems, there are advanced decomposition
techniques that solve the problem by iteratively solving a sequence of subproblems. In this article,
we use one of the most popular decomposition techniques, Lagrangian relaxation [9, 31]. The idea
of Lagrangian relaxation comes up in the context of using Lagrangian multipliers to decompose
the problem; thus, we introduce the Lagrangian multipliers λ = [λk , k ∈ K ]T on the constraint C5,
where λk denotes the price of all the tasks performed by the k t h cloudlet. X and A are the set of
the offloading decision variables and the set of the offloading ratio, respectively. The Lagrangian
function is given by:

M 
Nm 
K 
M 
Nm
L(X , A, λ) = Zum, n + λk (xum, n ,k · c k − Fk ),
m n k m n

M 
Nm 
M 
Nm 
K 
K
= Zum, n + λk xum, n ,k · c k − Fk ,
m n m n k k

M 
Nm 
K 
K
= Zu λk xum, n ,k · c k  −
m, n + λk Fk .
m n  k  k

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
24:12 H. Mazouzi et al.

Fig. 4. DM2-ECOP architecture overview.

The Lagrangian dual problem for primal Equation (9) is then given by:
max min L(X , A, λ).
λ X,A

We can see that the Lagrangian dual problem is separable into two levels: Level 1 is the inner
minimizing, which consists of M subproblems, each one concerning only one AP. Level 2 is outer
maximization, which is the master problem that considers the global variables and constraint C5.
Focusing on this observation, we introduce a new offloading policy named Distributed Multi-
user Multi-cloudlet Efficient Computation Offloading Policy (DM2-ECOP). In the following, we
describe the proposed computation offloading policy.

5 DM2-ECOP: DISTRIBUTED MULTI-USER, MULTI-CLOUDLET EFFICIENT

COMPUTATION OFFLOADING POLICY
As introduced in the last section, the Lagrangian dual problem is decomposable into M subprob-
lems. Each subproblem tries to find an optimal offloading decision and cloudlet selection to the
users associated with one AP. Considering this characteristic, we design a new offloading policy,
DM2-ECOP. As shown in Figure 4, DM2-ECOP has two levels of decision. The local offloading
manager is responsible for the offloading decision and cloudlet selection of an AP; it solves the
associated subproblem, then sends the offloading decision and the cloudlet selection to the cen-
tralized decision level. The MEC computation offloading manager receives the solution of all
subproblems, then it ensures that the obtained offloading solution is feasible and respects all the
constraints. After, it updates the Lagrangian multipliers and transmits the new values to every
local offloading manager to improve the local solutions.

5.1 Local Offloading Manager: A Greedy Best Cloudlet Selection First

Heuristic(GBC-SFH)
The local offloading manager tries to solve the subproblem of one AP to decide which user can
offload. Then, it selects the appropriate cloudlet to perform the task of each user. According to
the previous considerations, we can formulate the subproblem of a local offloading manager as

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
Efficient Computation Offloading for Mobile Edge Computing 24:13

follows:

Nm 
K
Zu λk xum, n ,k · c k 
Minimize m, n +
n  k 
Subject to:
constraints C1 – C4 and C6–C8. (12)
To solve the subproblem, we need to know the bandwidth allocated to each user. Unfortunately,
this bandwidth depends on the number of users that offload their tasks. Therefore, we need to know
the bandwidth allocation to decide whether a user should offload its task or not. To overcome this
dependency problem, we use a branching heuristic. The key idea is that for any AP m, the number
of users that can offload their tasks is between an upper and a lower bound. The lower bound
corresponds to the number of users with static offloading decision tasks, and the upper bound
corresponds to the maximum number of users associated with the AP m, Nm .
Definition 5.1. The offloading capacity of the AP m is defined as the number of tasks that have
been accepted being performed by the MEC environment. We note it by πm , and it is given by:

Nm 
K
πm = xum, n ,k .
n k

The strategy of solving the subproblem is very simple: It consists of finding the optimal πm that
gives the minimal offloading cost. We add the constraint C9 to the subproblem (12):

Nm 
K
C9 : xum, n ,k = πm . (13)
n k
To achieve a good and fast offloading decision, the local offloading manager uses greedy heuris-
tics to solve the subproblem (12). The Greedy Best Cloudlet Selection First Heuristic (GBC-SFH)
offers heuristics to determine which users offload, to determine the computation to offload, and
selects the cloudlet to perform each offloaded task. GBC-SFH iterates all possible values of the
offloading capacity, πm , in an increasing order, as illustrated in Algorithm 1. In brief, the idea is to
find the best cloudlet selection for all static offloading tasks in the first step by minimizing the La-
grangian cost Zuem, n ,k + λk c k under the constraints C1 − C4 and C6 − C8. So, each static offloading
task is offloaded to the cloudlet that minimizes the Lagrangian cost.
For each dynamic offloading decision task, GBC-SFH tries to select the best cloudlet and compute
the optimal ratio aum, n of the computation to offload. According to the resource availability, GBC-
SFH can offload the task or perform it locally, aum, n = 0, by the user’s device. Since the wireless
bandwidth at the AP may not be enough to offload all the dynamic offloading decision tasks, we
need to define an order to determine which dynamic offloading decision task is preferred for the
offloading. To this purpose, we define an offloading priority for each task according to the following
formula:
ξum, n = Zul m, n − min (Zuem, n ,k ); under aum, n = 1.
k ∈K
Here, the offloading priority is the local cost minus the cost when all the computation is offloaded
to the best cloudlet. The idea is that where ξum, n is going higher, the user um,n is more preferred to
offload its task. Unlike the static offloading decision tasks, for dynamic offloading decision tasks,
we need to compute the computation to offload (aum, n ). To this end, GBC-SFH uses a two-step
method. In the first step, it selects the best cloudlet to offload the computation of the user um,n . At
this step, GBC-SFH chooses the cloudlet that minimizes the Lagrangian cost Zuem, n ,k + λk c k under

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
24:14 H. Mazouzi et al.

the constraints C1 − C4, C6 − C7, and aum, n = 1. After the selection of the best cloudlet, GBC-SFH
computes the optimal value of aum, n for the current user. The optimal value of aum, n is the solution
to the following problem:
 
min Zuem, n ,k + Zul m, n
Subject to: constraint C7. (14)
Equation (14) is a simple problem with one variable. Its optimal solution can be achieved by deriv-
ative sign rules [28]. Theorem 5.2 shows when the minimum of this problem will be achieved.
Theorem 5.2. Let us define ψum, n and μum, n the upload data-computing ratio of the dynamic of-
floading decision task, and the local-remote offloading cost ratio of the user um,n , respectively. They
are given as follows:
upum, n
ψum, n = ,
γum, n
wum, n · [κ · fu3m, n · c k · βum, n + (1 − βum, n ) · (c k − fum, n ) − βum, n · Puidl e · f
um, n ]
μum, n =
m, n

c k · fum, n · (Put xm,/rn x · βum, n + 1 − βum, n )

The minimum of Equation (14) is achieved when:
• aum, n = 1, if and only if :
ψum, n < μum, n
• aum, n = 0, if and only if:
ψum, n > μum, n
• the problem is constant, if and only if :
ψum, n = μum, n .
Proof. The proof of Theorem 5.2 is detailed in the appendix. 
Using Theorem 5.2, we have three possible scenarios for dynamic offloading decision tasks:
(1) when aum, n = 1, the whole computation must be offloaded; (2) when aum, n = 0, there is no
offloading; and 3) when the Equation (14) is constant, all possible values of aum, n give the same
performance. As the computation offloading is a solution to improve the performance, we choose
to not offload, aum, n = 0 when this case occurs. Once the number of the offloading tasks is equal
to the current offloading capacity (πm ), the remaining tasks are assigned to be performed locally
by the user’s device.
Consequently, in the worst case, the GBC-SFH iterates Nm in the outer loop when there is
no static offloading decision task, which means a complexity of Nm · loд(Nm ) to sort the tasks
and Nm · K to assign to task at the inner loop. Thus, the maximum total number of iterations is
Nm2 · K + Nm2 · loд(Nm ). Therefore, the complexity of GBC-SFH is O(Nm2 · loд(Nm )), which is fast
especially when the number of users associated to each AP is small [14, 21, 33] (≤20).

5.2 MEC Computation Offloading Manager

The outer level of the Lagrangian dual problem is the master problem. It ensures a feasible offload-
ing solution of the primal Equation (9). Finding the optimal solution of the Lagrangian dual prob-
lem requires an exhaustive search of all solutions’ space and Lagrangian multiplier values, which
is a difficult task in general [9]. Consequently, we need to adopt a faster approach. In this work, we
use the Subgradient-based heuristic [31]. The proposed heuristic used in the MEC computation
offloading manager has three steps, as illustrated in Algorithm 2. First, it solves the subproblems

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
Efficient Computation Offloading for Mobile Edge Computing 24:15

ALGORITHM 1: The local offloading manager: GBC-SFH

Input:
1: Πm : Set of offloading capacity;
Output: Output the offloading decision X, ratio A, and cost Z;
2: Sort Πm in increasing order;
3: for πm ∈ Πm , do
4: allocate bandwidth using Equation (1);
5: offload each static offloading decision task to the cloudlet k that minimizes Zuem, n ,k + λk c k
under constraints C1 − C4 and C6 − C8;
6: nboffloaded_t ask = number of static offloading decision tasks;
7: compute ξum, n for every dynamic offloading decision task;
8: Sort dynamic offloading decision tasks in decreasing order of ξ nm ;
9: while nboffloaded_t ask ≤ πm , do
10: Select the cloudlet k that minimizes Zuem, n ,k + λk c k under constraints C1 − C4, C6 − C7,
and aum, n = 1;
11: Compute the optimal value of aum, n using Theorem 5.2;
12: if aum, n == 0, then
13: there is no offloading. This dynamic Offloading task must be performed locally;
14: else
15: Offload this dynamic Offloading task to the cloudlet k;
16: nboffloaded_t ask + +;
17: end if
18: if (there is no more task) and (nboffloaded_t ask < πm ), then
19: break the while-loop. There is no feasible solution for this value of πm ;
20: end if
21: end while
22: all the remaining tasks must be performed locally;
23: update the best offloading cost Z, ratio A and decision X;
24: end for

in the local offloading manager by the GBC-SFH for the current Lagrangian multipliers λ. Next,
the MEC computation offloading manager checks if they obtained an offloading solution that
is not feasible. If so, the Lagrangian Adjustment Heuristic (LAH) will be used the get a feasible
solution using a local search. The idea of LAH heuristics is to check if every cloudlet respects
the constraint C5. When a cloudlet does not respect this constraint, LAH heuristic reassigns some
tasks offloaded from this cloudlet to another cloudlet that respects all constraints.
At the end, the MEC computation offloading manager updates the Lagrangian multipliers
by the following formula:

M  Nm
λk (t + 1) = λk (t ) + θ (t ) ·   xum, n ,k · c k  − Fk  , (15)
m  n  
where θ (t ) is the update step. In this work, we use the Held and Karp formula [9, 31] to update
this step as follows:
Z ∗ − Z(t )
θ (t ) = η(t ) · K M  Nm , (16)
n xum, n ,k · c k − F k )
2
k=1 ( m

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
24:16 H. Mazouzi et al.

ALGORITHM 2: MEC computation offloading manager

Input:
1: Itmax : maximum number of iterations;
2: ε: an infinitesimal number;
Output: offloading decision and cloudlet selection for all users;
3: Initialize λ k randomly;
4: Zmax = −∞;
5: while (t < Itmax and θ (t ) > ε), do
6: for (m ∈ M), do
7: Zm (t )= get the solution of subproblem m from the local offloading manager m;
8: end for
 
9: Z(t ) = m ∈M Z(t )m − k ∈K λk · Fk ;
10: if (Z(t ) > Zmax ), then
11: Zfeasible = use Heuristic LAH to find a feasible solution;
12: if (Zfeasible < Z ∗ ), then
13: Z ∗ = Zfeasible ;
14: update the best solution of the primal problem;
15: end if
16: Zmax = Z(t );
17: end if
18: update the Lagrangian multipliers and η using Equations (15), (16), and (17);
19: end while

where η(t ) is a decreasing adaptation parameter 0 < η(0) ≤ 2, Z ∗ is the best obtained solution of
Equation (9), and Z(t ) refers to the current solution of the Lagrangian dual problem. η(t ) can be
expressed by the following formula:

⎪ϑ · η(t )
⎧
⎪ if Z(t ) did not increase
η(t + 1) = ⎨
⎪ . (17)
⎪η(t ) otherwise
⎩
As suggested in References [9, 31], we set the values of ϑ = 0.9 and η(0) = 2. The master problem
repeats these steps until the stop conditions, which are the maximum number of iterations Itmax
and the maximum tolerated error of the update step ε (θ ≤ ε).

6 NUMERICAL RESULTS
In this section, we evaluate the performance of DM2-ECOP using the characteristics of realistic
system configuration. We use an MEC environment consisting of a metropolitan area, which is
composed of 20 APs forming a ring topology. The delay between any two APs is 3ms and the delay
between every AP and the remote cloud is 100ms [14, 33]. We suppose that four cloudlets are
equidistantly deployed among the network, i.e., cloudlet 1 is collocated with the AP 1, cloudlets 2
with the AP 6, cloudlet 3 with the AP 11, and cloudlet 4 with the AP 16. To study the performance
of our offloading policy, we consider four cloudlet configurations. Table 2 illustrates the list of
the cloudlets’ configurations considered for our tests. We consider real configurations used by
public cloud providers, such as, Amazon Web Services (AWS) and Microsoft Azure [5, 12, 21], to
simulate the behavior of DM2-ECOP policy for real-world scenarios.
The wireless bandwidth of each AP is 150Mbps. The bandwidth allocated to each user is esti-
mated using the parameter settings used in the Bianchi model [1]. Similar to Reference [33], we

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
Efficient Computation Offloading for Mobile Edge Computing 24:17

Table 2. List of the Cloudlets’ Configurations Used for the Tests

Computing capacity Fk and allocation c k in Giga CPU cycles/s

Configuration cloudlet 1 cloudlet 2 cloudlet 3 cloudlet 4
c1 F1 c2 F2 c3 F3 c4 F4
configuration 1 10 1,000 10 1,000 10 1,000 10 1,000
configuration 2 15 1,000 10 1,000 15 1,000 10 1,000
configuration 3 10 500 10 500 10 1,500 10 1,500
configuration 4 15 500 10 500 15 1,500 10 1,500

Table 3. The Characteristic of the Real-world Applications

Used for Our Tests

γum, n upum, n dwum, n tum, n

Application
(Giga CPU cycles) (Kilobyte) (Byte) (Second)
static offloading decision tasks
FACE 12.3 62 60 5
SPEECH 15 243 50 5.1
OBJECT 44.6 73 50 13
dynamic offloading decision tasks
Linpack 50 10,240 120 62.5
CPUBENCH 3.36 80 80 4.21
PI BENCH 130 10,240 200 163

assume that the number of users connected to every AP, Nm is not greater than 20. Precisely,
Nm takes values from {5, 10, 15, 20}. Each user runs one application from Table 3. The first three
applications are static offloading decision tasks; the others are dynamic offloading decision tasks
[11]. We assume that Put xm,/rn x = 10 ∗ Puidl e and P idl e = 100mW, as shown in Reference [3]. The local
m, n um, n
computing capability of each user was randomly chosen from fum, n ∈ [0.8, 1, 1.2] gigacycles.
The performances of DM2-ECOP are compared to two offloading policies from the literature:
• Nearest Cloudlet Offloading (NCO) [14, 33]: in which each AP is associated with the nearest
cloudlet. So, all the users connected to this AP offload their tasks to the same cloudlet. When
a cloudlet is overloaded, the tasks are migrated to another cloudlet.
• Full Cloud Offloading (FCO) [4, 5]: In this case, the users offload their tasks to the remote
cloud. To make sense of the performances comparison of the offloading policies DM2-ECOP,
NCO, and FCO, we assume that the computing capacity allocated to perform each offloading
task in the remote cloud is 10 gigacycles.
In the following, the default cloudlet configuration is the first configuration (configuration 1),
and the density of users at each AP is considered as the same, i.e., the same number of users at each
AP. Furthermore, the stop criteria of the MEC computation offloading manager for DM2-ECOP are
Itmax = 100 for the maximal number of iterations, and ε = 10−20 for the maximum tolerated error
of update steps.

6.1 Convergence of DM2-ECOP

To evaluate the performance of DM2-ECOP and its convergence to a feasible solution, Table 4
depicts the required number of iterations to get a feasible solution and the last value of update
steps (θ ). As expected, the required number of iterations increases as the number of users in the

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
24:18 H. Mazouzi et al.

Table 4. Number of Iterations and Update Step Taken by

DM2-ECOP to Converge to a Feasible Solution

Number of users Number of iterations Update step θ

100 15 0.0
200 20 9.12 × 10−22
300 29 5.09 × 10−21
400 43 4.46 × 10−21

Fig. 5. Comparison of offloading policies DM2-ECOP, NCO, and FCO where the cost parameter β = 1.

network increases. Moreover, the update step converges to the maximum tolerated error ε with a
few number of iterations; this convergence changes while the number of users increases. Thanks
to the Held and Karp formula used in our work to update the Lagrangian update step, for the rapid
convergence of DM2-ECOP offloading policy.

6.2 Offloading Performance Comparison

Figure 5 plots the offloading performances of DM2-ECOP, NCO, and FCO if we set the cost param-
eter β = 1. We also distinguish between the costs related to the network access, network backhaul,
and processing. We note that DM2-ECOP reduces the total offloading cost compared to NCO and
FCO. More precisely, we can see that the access cost of DM2-ECOP is the lowest, but the processing
cost is the highest. This is due to the bandwidth allocation heuristic used by DM2-ECOP, which
tries to maximize the bandwidth allocated to each user by minimizing the offloading capacity (π )
of each AP. So, less users can offload their tasks to the MEC with DM2-ECOP compared to NCO
and FCO. However, where the wireless bandwidth is enough to offload all tasks, DM2-ECOP and
NCO are equivalent, as shown in Table 4 where 100 users are in the network.
To understand the effect of user density on offloading performances, we investigate in Figure 6
the offloading gain of DM2-ECOP compared to NCO under different user density. We consider four
scenarios where the topology is divided into two regions, each one containing 10 APs. In Scenario
1, the regions have the same user density. In Scenario 2, user density in Region 1 is twice the
user density in Region 2. In Scenario 3, user density in Region 1 is three times the user density in
Region 2. Finally, in Scenario 4, user density in Region 1 is four times the user density in Region 2.
In Figure 6, we note that the offloading gain of DM2-ECOP compared to NCO goes up when user
density goes high. For example, where 200 users are in the network, the gain is 6.1% for Scenario 1
and 10.5% for Scenario 4. This is because the cloudlet selection in NCO is static, so this policy needs
to migrate some tasks where the cloudlet is overloaded. Consequently, it adds an extra offloading

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
Efficient Computation Offloading for Mobile Edge Computing 24:19

Fig. 6. Comparison of DM2-ECOP and NCO for different user density in the network.

Fig. 7. Comparison of offloading policies DM2-ECOP, NCO, and FCO over the parameter β, where 200 users
are in the mobile edge computing environment.

cost. However, DM2-ECOP tries to find the best cloudlet selections dynamically at the offloading
decision, according to the system and network resource availability.

6.3 Impact of the Cost Parameter β on the Offloading Performance

Figure 7 studies the effect of the offloading cost parameter β on the performance of the policies
DM2-ECOP, NCO, and FCO. As we can see in Figure 7(a), the energy consumption of DM2-ECOP
is better than NCO and FCO for all possible values of β. Indeed, even if we set β to 0, which
means that we give a complete priority to the tasks’ completion time, DM2-ECOP obtains better
performances. Moreover, when we increase the value of β, the obtained performances are even
better. Consequently, DM2-ECOP achieves better performance, in terms of energy consumption,
whatever the offloading cost: energy consumption, completion time, or a combination of energy
and time. This is because of the dynamic cloudlet selection adopted in DM2-ECOP.
In Figure 7(b), we investigate the effect of β on the performance in terms of completion time. We
note that the completion time of DM2-ECOP and NCO are better than FCO, because the cloudlets

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
24:20 H. Mazouzi et al.

Fig. 8. Offloading policies’ performances over different cloudlet configurations, where β = 1 (the offloading
cost is energy consumption) and 200 users are in the MEC environment.

are close to users. Moreover, the completion time of DM2-ECOP is the lowest where β = 0. How-
ever, NCO achieves better completion time than DM2-ECOP where β closes to 1. Consequently,
NCO has the best performance in terms of completion time. In fact, when β is close to 1, the energy
consumption becomes more important than the completion time in the expression of the offload-
ing cost. DM2-ECOP reduces the offloading cost by offloading less tasks to the MEC, as shown in
Figure 5, to minimize the energy consumed by the wireless access level. As a result, more tasks are
executed locally, which increases the completion time.

6.4 Impact of the Cloudlets’ Configurations

In the following, we study the performance of the offloading policies over heterogeneous cloudlet
configurations. In Figure 8, we investigate the performance of the offloading policies DM2-ECOP
and NCO over four cloudlet configurations presented at the beginning of this section. We observe
that the DM2-ECOP and NCO are equivalent when the cloudlets have exactly the same configura-
tions, such as Configuration 1. However, where the computing resources allocated to each task are
heterogeneous, which corresponds to a more realistic scenario, DM2-ECOP achieves better per-
formance in terms of energy consumption and completion time. This can be explained by the fact
that DM2-ECOP offloads to the best cloudlet, but NCO offloads to the nearest cloudlet. We also can
notice that DM2-ECOP gets better performance for Configuration 2 than Configuration 4, even if
the resource allocated for each task are the same in the two configurations. This is because the
total computing capacity in Configuration 4 is not homogeneous. Thus, NCO needs to migrate
some tasks from Cloudlets 1 and 2 to Cloudlets 3 and 4. To summarize, these results show that
DM2-ECOP can achieve a good offloading performance under different cloudlet configurations.

6.5 Impact of the Applications Characteristics

As shown previously in our analytical model, the characteristics of the application have a crucial
role in the efficiency of the offloading performance. To understand this role, we investigate the
impact of the application on the offloading cost and on the amount of the offloaded computation,
aum, n .
Figure 9 depicts the performances of the offloading policies for each application under the
cloudlets’ Configuration 3. We note that for static offloading decision tasks, DM2-ECOP has better
energy consumption and completion time followed by NCO. Indeed, the static offloading decision
tasks must be offloaded, and DM2-ECOP tries to find the best cloudlet selection at the decision

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
Efficient Computation Offloading for Mobile Edge Computing 24:21

Fig. 9. Offloading policies’ performances for each application for Configuration 3, where β = 1 (the offload-
ing cost is energy consumption) and 200 users are in the MEC environment.

Fig. 10. The effect of the number of users, wum, n , the considered offloading cost βum, n , and the allocation
computing resource c k on the application partition decision, where the wireless bandwidth at AP is 150mbps,
local cpu power is fum, n = 1 gigacycle.

time. However, FCO offloads to the nearest cloudlet, which induces an additional offloading cost,
since some tasks must be migrated to other cloudlets. However, for the dynamic offloading de-
cision tasks, we note that DM2-ECOP has the lowest energy consumption and completion time
where the application does not need to upload lots of data to the cloudlet, such as CPUBENCH.
However, where the application requires lots of data, e.g., Linpack and PIBENCH, the completion
time of NCO and FCO are better than that of DM2-ECOP. This is because DM2-ECOP tries to
perform the application locally when it uploads a lot of data to minimize the access cost of the
users.
Finally, to deeply analyze the performance of DM2-ECOP with the dynamic offloading decision
tasks, we study the effect of the number of users and the ratio between the remote and local
processing capacity on the offloading decision. In Figure 10(a), we plot the effect of the number of
users per access point and, thus, the amount of bandwidth allocated to each user on the offloading
decision. The red area corresponds to the case where we execute a task only locally (i.e., aum, n = 0),
and the green area corresponds to the case where a task is totally offloaded (i.e., aum, n = 1). In
addition to the offloading decision, we also plot the upload data-computing ratio ψum, n for three
dynamic offloading decision tasks, namely Linpack, CPUBENCH, and PI BENCH. As we can see,

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
24:22 H. Mazouzi et al.

the three applications do not behave the same way when we increase the number of users per AP.
Indeed, Linpack is the most sensitive application and stops offloading when the number of users
is greater than 14. However, CPUBENCH is the less-sensitive application, since it stops offloading
when the number of users reaches 37. This is due to the fact that Linpack sends much more data
when it offloads compared to CPUBENCH. In Figure 10(b), we investigate the impact of the ratio
between the local and remote processing capacity on the offloading decision. As in the Figure 10(a),
we also plot the upload data-computing ratio ψum, n for Linpack, CPUBENCH, and PI BENCH.
As we can see, when the ratio between the remote processing capacity and the local processing
capacity increases, our proposal tends to more offload tasks. However, in contrast to Figure 10(a),
Linpack is less sensitive to that increase compared to CPUBENCH and PI BENCH. Indeed, since
the total amount of data that should be offloaded for Linpack is important, the offloading becomes
beneficial only if the remote processing capability is very important in comparison to the local
processing capability.

7 CONCLUSION
Computation offloading in a multi-user, multi-cloudlet mobile edge computing environment is a
challenging issue. In this article, we propose a new computation offloading policy to decide which
users should offload and to which cloudlet. First, we formulate the problem as a Non-Linear Mixed
Binary Integer Program. Then, we propose an efficient distributed heuristic to solve the problem
using the Lagrangian decomposition approach. The proposed heuristic uses a branching algorithm
to maximize the bandwidth allocation and minimize the offloading cost.
In addition, compared to other works, our proposal (DM2-ECOP) considers two categories of
offloadable tasks: the static offloading decision tasks that must be performed remotely and the
dynamic offloading decision tasks that can be performed both locally and remotely. We also add
an offloading computation ratio associated with both static and dynamic decision tasks. This ratio
denotes the portion of the application that is executed locally in the terminal and the portion of
the application that should be offloaded to the cloudlet.
The obtained numerical results show performance improvements in terms of the offloading
cost compared to existing offloading policies under different scenarios and cloudlet configurations.
Moreover, because we consider that all the tasks have the same priority and they are not sharing
resources (same CPU) at the cloudlet, we demonstrate that the best possible value of the offloading
computation ratio is either 0 or 1.
For future work, we will consider an adaptive offloading policy, where the offloaded tasks must
be determined at runtime. Moreover, in this article, we assume that each mobile is executing only
one task at a time. In future work, we propose to explore the case where an application is charac-
terized by a tasks dependency graph. In this case, more than one task can be offloaded at the same
time to the remote cloudlet or cloud.

APPENDIX: PROOF OF THEOREM 5.2

Given the objective function of Equation (14), which is a function with the variable aum, n , we
can find the minimum following the derivative sign rules [28]. Let Fum, n be the derivative of the
objective function, Zuem, n ,k + Zul m, n . Fum, n is given as follows:

Put xm,/rn x · βum, n + 1 − βum, n

Fum, n = upum, n ·
wum, n
βum, n · Puidl e
1 − βum, n 1 − βum, n
+ γum, n · + − − κ · fu2m, n · βum, n  .
m, n

 ck ck fum, n 

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
Efficient Computation Offloading for Mobile Edge Computing 24:23

The derivative function Fum, n is a constant and does not change when the variable aum, n does.
According to the derivative sign rules [28], we have three cases:
Case 1: Fum, n < 0, there the objective function of Equation (14) is monotonically decreasing.
Consequently, its minimum is achieved at aum, n = 1. This case occurs when:

upum, n wum, n · κ · fu3m, n · c k · βum, n + (1 − βum, n ) · (c k − fum, n ) − βum, n · Puidl e · f

m, n um, n
<  t x /r x  .
γum, n c k · fum, n · Pum, n · βum, n + 1 − βum, n
Case 2: Fum, n > 0, there the objective function of Equation (14) is monotonically increasing.
Consequently, its minimum is achieved at aum, n = 0. This case occurs when:

upum, n wum, n · κ · fu3m, n · c k · βum, n + (1 − βum, n ) · (c k − fum, n ) − βum, n · Puidl e · f

m, n um, n
>  t x /r x  .
γum, n c k · fum, n · Pum, n · βum, n + 1 − βum, n
Case 3: Fum, n = 0, there the objective function of Equation (14) is constant and does not change.
So, at the values of aum, n give the same cost. This case occurs when:

upum, n wum, n · κ · fu3m, n · c k · βum, n + (1 − βum, n ) · (c k − fum, n ) − βum, n · Puidl e · f

m, n um, n
=  t x /r x  .
γum, n c k · fum, n · Pum, n · βum, n + 1 − βum, n
Ending of the proof. 

REFERENCES
[1] Giuseppe Bianchi. 2000. Performance analysis of the IEEE 802.11 distributed coordination function. IEEE J. Select.
Areas Comm. 18, 3 (2000), 535–547.
[2] Arash Bozorgchenani, Daniele Tarchi, and Giovanni Emanuele Corazza. 2017. An energy and delay-efficient partial
offloading technique for fog computing architectures. In Proceedings of the IEEE Global Communications Conference
(GLOBECOM’17). IEEE, 1–6.
[3] Aaron Carroll, Gernot Heiser. 2010. An analysis of power consumption in a smartphone. In Proceedings of the USENIX
Annual Technical Conference (USENIXATC’10), Vol. 14. Boston, MA, 21–21.
[4] Meng-Hsi Chen, Ben Liang, and Min Dong. 2016. Joint offloading decision and resource allocation for multi-user
multi-task mobile cloud. In Proceedings of the IEEE International Conference on Communications (ICC’16). IEEE, 1–6.
[5] Xu Chen, Lei Jiao, Wenzhong Li, and Xiaoming Fu. 2016. Efficient multi-user computation offloading for mobile-edge
cloud computing. IEEE/ACM Trans. Netw. 24, 5 (2016), 2795–2808.
[6] Byung-Gon Chun, Sunghwan Ihm, Petros Maniatis, Mayur Naik, and Ashwin Patti. 2011. Clonecloud: Elastic execu-
tion between mobile device and cloud. In Proceedings of the 6th European Conference on Computer Systems (EuroSys’11).
ACM, 301–314.
[7] Eduardo Cuervo, Aruna Balasubramanian, Dae-ki Cho, Alec Wolman, Stefan Saroiu, Ranveer Chandra, and Paramvir
Bahl. 2010. MAUI: Making smartphones last longer with code offload. In Proceedings of the 8th International Conference
on Mobile Systems, Applications, and Services (MobiSys’10). ACM, 49–62.
[8] Debessay Fesehaye, Yunlong Gao, Klara Nahrstedt, and Guijun Wang. 2012. Impact of cloudlets on interactive mobile
cloud applications. In Proceedings of the 16th IEEE International Enterprise Distributed Object Computing Conference
(EDOC’12). IEEE, 123–132.
[9] Marshall L. Fisher. 2004. The Lagrangian relaxation method for solving integer programming problems. Manag. Sci.
50, 12-supplement (2004), 1861–1871.
[10] Keke Gai, Meikang Qiu, and Hui Zhao. 2018. Energy-aware task assignment for mobile cyber-enabled applications
in heterogeneous cloud computing. J. Parallel and Distrib. Comput. 111 (2018), 126–135.
[11] Ying Gao, Wenlu Hu, Kiryong Ha, Brandon Amos, Padmanabhan Pillai, and Mahadev Satyanarayanan. 2015. Are
cloudlets necessary? Technical Report CMU-CS-15-139. School of Computer Science, Carnegie Mellon University,
Pittsburgh, PA.
[12] Songtao Guo, Bin Xiao, Yuanyuan Yang, and Yang Yang. 2016. Energy-efficient dynamic offloading and resource
scheduling in mobile cloud computing. In Proceedings of the 35th IEEE International Conference on Computer Commu-
nications (INFOCOM’16), Vol. 2016-July. IEEE, 1–9.

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.
24:24 H. Mazouzi et al.

[13] Dong Huang, Ping Wang, and Dusit Niyato. 2012. A dynamic offloading algorithm for mobile computing. IEEE Trans.
Wireless Comm. 11, 6 (2012). IEEE, 1991–1995.
[14] Mike Jia, Jiannong Cao, and Weifa Liang. 2015. Optimal cloudlet placement and user to cloudlet allocation in wireless
metropolitan area networks. IEEE Trans. Cloud Comput. 99 (2015).
[15] Mike Jia, Weifa Liang, Zichuan Xu, and Meitian Huang. 2016. Cloudlet load balancing in wireless metropolitan
area networks. In Proceedings of the 35th IEEE International Conference on Computer Communications (INFOCOM’16),
Vol. 2016-July. IEEE, 1–9.
[16] Doyub Kim, Woojong Koh, Rahul Narain, Kayvon Fatahalian, Adrien Treuille, and James F. O’Brien. 2013. Near-
exhaustive precomputation of secondary cloth effects. ACM Trans. Graph. 32, 4 (July 2013), 87:1–87:8.
[17] Sven O. Krumke and Clemens Thielen. 2013. The generalized assignment problem with minimum quantities. Euro. J.
Op. Res. 228, 1 (2013), 46–55.
[18] Grace Lewis, Sebastián Echeverría, Soumya Simanta, Ben Bradshaw, and James Root. 2014. Tactical cloudlets: Moving
cloud computing to the edge. In Proceedings of the IEEE Military Communications Conference. IEEE, 1440–1446.
[19] Grace Alexandra Lewis. 2016. Software Architecture Strategies for Cyber-foraging Systems. Ph.D Dissertation, Carnegie
Mellon University, Pittsburgh, PA.
[20] Jia-Liang Lu and Fabrice Valois. 2006. Performance evaluation of 802.11 WLAN in a real indoor environment. In
Proceedings of the IEEE International Conference on Wireless and Mobile Computing, Networking, and Communications
(WiMob’06). IEEE, 140–147.
[21] Longjie Ma, Jigang Wu, and Long Chen. 2017. DOTA: Delay bounded optimal cloudlet deployment and user associa-
tion in WMANs. In Proceedings of the 17th IEEE/ACM International Symposium on Cluster, Cloud, and Grid Computing.
IEEE Press, 196–203.
[22] Yuyi Mao, Jun Zhang, S. H. Song, and Khaled Ben Letaief. 2016. Power-delay tradeoff in multi-user mobile-edge
computing systems. In Proceedings of the IEEE Global Communications Conference (GLOBECOM’16). IEEE, 1–6.
[23] Antti P. Miettinen and Jukka K. Nurminen. 2010. Energy efficiency of mobile clients in cloud computing. HotCloud
10 (2010), 4–4.
[24] Anwesha Mukherjee, Debashis De, and Deepsubhra Guha Roy. 2016. A power and latency aware cloudlet selection
strategy for multi-cloudlet environment. IEEE Trans. Cloud Comput. 99 (2016), 1–14.
[25] Yucen Nan, Wei Li, Wei Bao, Flavia C. Delicato, Paulo F. Pires, Yong Dou, and Albert Y. Zomaya. 2017. Adaptive
energy-aware computation offloading for cloud of things systems. IEEE Access 5 (2017), 23,947–23,957.
[26] Yucen Nan, Wei Li, Wei Bao, Flavia C. Delicato, Paulo F. Pires, and Albert Y. Zomaya. 2018. A dynamic tradeoff data
processing framework for delay-sensitive applications in cloud of things systems. J. Parallel and Distrib. Comput. 112
(2018), 53–66.
[27] Deepsubhra Guha Roy, Debashis De, Anwesha Mukherjee, and Rajkumar Buyya. 2016. Application-aware cloudlet
selection for computation offloading in multi-cloudlet environment. J. Supercomput. (2016), 1–19.
[28] Mark Ryan. 2005. Calculus Workbook for Dummies. Wiley Publishing, Inc.
[29] Swetank Kumar Saha, Pratham Malik, Selvaganesh Dharmeswaran, and Dimitrios Koutsonikolas. 2016. Revisiting
802.11 power consumption modeling in smartphones. In Proceedings of the 17th IEEE International Symposium on
World of Wireless, Mobile, and Multimedia Networks (WoWMoM’16). IEEE, 1–10.
[30] Mahadev Satyanarayanan, Grace Lewis, Edwin Morris, Soumya Simanta, Jeff Boleng, and Kiryong Ha. 2013. The role
of cloudlets in hostile environments. IEEE Pervas. Comput. 12, 4 (2013), 40–49.
[31] Jiafu Tang, Chongjun Yan, Xiaoqing Wang, and Chengkuan Zeng. 2014. Using Lagrangian relaxation decomposition
with heuristic to integrate the decisions of cell formation and parts scheduling considering intercell moves. IEEE
Trans. Automat. Sci. Eng. 11, 4 (2014), 1,110–1,121.
[32] Song Wu, Chao Niu, Jia Rao, Hai Jin, and Xiaohai Dai. 2017. Container-based cloud platform for mobile computation
offloading. In Proceedings of the IEEE International Parallel and Distributed Processing Symposium (IPDPS’17). IEEE,
123–132.
[33] Zichuan Xu, Weifa Liang, Wenzheng Xu, Mike Jia, and Song Guo. 2016. Efficient algorithms for capacitated cloudlet
placements. IEEE Trans. Parallel Distrib. Systems 27, 10 (2016), 2,866–2,880.
[34] Hong Yao, Changmin Bai, Muzhou Xiong, Deze Zeng, and Zhangjie Fu. 2017. Heterogeneous cloudlet deployment
and user-cloudlet association toward cost-effective fog computing. Concurr. Comput.: Practice and Exper. 29, 16 (2017).
[35] Chongyu Zhou, Chen-Khong Tham, and Mehul Motani. 2017. Online auction for truthful stochastic offloading in
mobile cloud computing. In Proceedings of the IEEE Global Communications Conference (GLOBECOM’17). IEEE, 1–6.
[36] Qiliang Zhu, Baojiang Si, Feifan Yang, and You Ma. 2017. Task offloading decision in fog computing system. China
Comm. 14, 11 (2017), 59–68.

Received December 2017; revised July 2018; accepted July 2018

ACM Transactions on Internet Technology, Vol. 19, No. 2, Article 24. Publication date: April 2019.