                                                              Journal of Systems Architecture 160 (2025) 103360


                                                                  Contents lists available at ScienceDirect


                                                         Journal of Systems Architecture
                                                         journal homepage: www.elsevier.com/locate/sysarc




Collaborative optimization of offloading and pricing strategies in dynamic
MEC system via Stackelberg game
Jing Mei a , Cuibin Zeng a , Zhao Tong a ,1 ,∗, Longbao Dai a , Keqin Li b ,2
a College of Information Science and Engineering, Hunan Normal University, Changsha, Hunan, 410081, China
b
    Department of Computer Science, State University of New York, New Paltz, NY, 12561, USA



ARTICLE                 INFO                             ABSTRACT

Keywords:                                                The rapid advancement of 5G technology has indirectly propelled the growth of connected devices within the
Mobile edge computing                                    Internet of Things (IoT). Within the IoT domain, mobile edge computing (MEC) has demonstrated potential in
Energy harvesting                                        task processing. However, as computational services expand, the reliable determination of user offloading
Lyapunov optimization
                                                         strategies and the rational establishment of service prices offered by servers to users continue to present
Stackelberg game
                                                         challenging research directions. The primary focus of this paper revolves around task offloading in the MEC
                                                         system, encompassing numerous user terminal devices that support energy harvesting (EH), a MEC server and
                                                         a central cloud server. The optimization goals are to maximize the utilities for both users and the MEC server
                                                         by adjusting offloading and pricing strategies. To guarantee the task queue’s stability within the system and
                                                         achieve a reasonable allocation of system resources, we propose a dynamic task offloading approach rooted in
                                                         Lyapunov optimization theory and Stackelberg game theory. In this algorithm, the MEC server takes on the role
                                                         of the leader, while each user terminal device acts as the follower. Aiming at the game equilibrium existence of
                                                         the algorithm, a series of mathematical analysis is carried out. Additionally, we conduct extensive simulation
                                                         experiments to validate the proposed algorithm’s effectiveness. The proposed algorithm achieves improvements
                                                         in user utility, with a 6.43% increase compared to the average time-constrained task offloading (ATCTO)
                                                         scheme, a 61.80% improvement over the local-only processing (LOP) scheme, and a 23.97% enhancement
                                                         over the genetic algorithm (GA) scheme. Meanwhile, it achieves a task queue backlog reduction of 50.00%
                                                         compared to ATCTO, 70.00% compared to LOP and 15.28% compared to GA.



1. Introduction                                                                               data privacy and security.
                                                                                                  Energy consumption for local computing and wireless transmission
    With the proliferation of 5G technology, its high speed and low                           tasks on user terminal devices (e.g., wearable devices, tablets, and
latency enable real-time connectivity, and the IoT stands as one of                           smartphones) is sourced from their internal batteries. For convenience,
the beneficiaries in the era of 5G. According to [1], it is forecasted                        these user terminal devices will be referred to as ‘‘users’’. Due to limited
that the number of IoT devices will exceed 75 billions by 2025. This                          size of users, long lifetime battery might not be appropriate, while
explosive growth will significantly amplify the scale of mobile data                          smaller-capacity battery may require frequent replacements, causing
traffic. Effectively managing the surge in mobile data traffic can be                         significant inconvenience. Fortunately, in recent years, EH technologies
achieved through mobile data offloading [2]. However, relying solely                          have garnered significant attention. These technologies enable users
on cloud computing is difficult to improve user experience quality.                           to harvest energy from the nature (e.g., solar), and store it in bat-
It is the increasing user demand for quality of experience that drives                        teries. By integrating rechargeable batteries with energy harvesting
the development of MEC. MEC moves computing power and resources                               technologies, the frequency of battery replacements can be notably
closer to users, usually on edge devices of the network. This aids in
                                                                                              reduced. In the future, the IoT could potentially integrate various
reducing data transmission latency and supporting applications with
                                                                                              energy harvesting methods [3]. Therefore, the integration of energy
strong real-time requirements. Moreover, edge data processing reduces
                                                                                              harvesting technologies into MEC holds practical significance.
the need to transmit sensitive data to remote servers, thereby enhancing


    ∗ Corresponding author.
     E-mail addresses: jingmei1988@163.com (J. Mei), zeng1941183190@gmail.com (C. Zeng), tongzhao@hunnu.edu.cn (Z. Tong), awaken6758@gmail.com
(L. Dai), lik@newpaltz.edu (K. Li).
   1
     Member, IEEE.
   2
     Fellow, IEEE.

https://doi.org/10.1016/j.sysarc.2025.103360
Received 3 September 2024; Received in revised form 7 January 2025; Accepted 31 January 2025
Available online 12 February 2025
1383-7621/© 2025 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
J. Mei et al.                                                                                                   Journal of Systems Architecture 160 (2025) 103360


    In order to flexibly respond to the changing demands between user               • We account for the practical limitation of the MEC server’s com-
devices and MEC server, this study considers a dynamic MEC system.                    puting capacity within each time slot, determined by its CPU
In this system, factors such as the arrival of tasks and energy collection            frequency. If the server cannot process all accumulated tasks from
will change with the change of time. The system can dynamically adjust                users, it requests processing from the central cloud server at a
the task offloading strategy and the service pricing strategy of the                  cost.
MEC server according to real-time conditions, thereby improving user                Next, we will introduce the remaining structure of this paper.
utility and MEC server utility. The authors considered the dynamic MEC           Section 2 delves into the related work. Section 3 defines the system
system in [4], but did not set the services provided by the MEC server           model and presents the optimization problems. Section 4 describes
as paid services. In contrast, [5] explored the pricing issue in a dynamic       the analysis and solution of the objective optimization problems. Sub-
system, but the study implemented a unified service pricing for all user         sequently, we introduce a multi-device task offloading and pricing
devices and failed to consider the differences between user devices.             mechanism algorithm (MDTOPMA) in Section 5. Section 6 evaluates
Given the heterogeneity of user devices, we introduced a differentiated          the algorithm’s performance through experiments. The final section
pricing strategy.                                                                concludes by summarizing the contributions of this paper.
    In this paper, we explore a dynamic MEC system enabled with EH,
which is composed of multiple user terminal devices, a MEC server, and           2. Related work
a central cloud. Each user possesses the capability to harness energy
from the surrounding natural environment and store it in their built-in              Numerous investigations have delved into the domain of computa-
                                                                                 tion offloading within the context of MEC. Ning et al. [8] presented
rechargeable batteries. Users can offload tasks to the MEC server when
                                                                                 an energy-efficient scheduling framework for vehicle networking with
needed, and the MEC server can further offload tasks to the central
                                                                                 support for Multi-Access Edge Computing, aimed at minimizing the
cloud as required. In this study, the MEC server’s computational capac-
                                                                                 energy consumption of roadside units (RSUs) while meeting task la-
ity is limited, necessitating revenue generation through user charges.
                                                                                 tency requirements. The authors proposed a heuristic algorithm to
Generally, the energy consumed by users for wireless transmission tasks
                                                                                 address this issue. Mao et al. [9] focused on the joint optimization
is lower than that required for performing equivalent-sized local com-
                                                                                 of execution delay and device energy consumption, proposing a low-
puting tasks. In order to decrease energy consumption, users are more
                                                                                 complexity suboptimal algorithm based on an alternating minimization
likely to opt for the method of remote task offloading. Nevertheless, due        strategy. The algorithm minimized the weighted sum of execution
to the fact that the MEC server does not offer services free of charge,          delay and device energy consumption by adjusting task offloading
each user needs to strike a balance between task offloading and service          scheduling and transmission power allocation. Zhao et al. [10] pro-
pricing. The main purpose of our research is to determine the optimal            posed an energy-efficient offloading algorithm to save mobile device
distribution of tasks between local computing and remote offloading              energy while meeting application response time requirements. Chen
for users, while also determining suitable service pricing for individual        et al. [11] investigated energy-efficient task offloading in MEC and
users for the MEC server.                                                        proposed a dynamic offloading algorithm that guarantees the average
    Given that the optimization objective of this paper involves a dy-           queue length. Li et al. [12] proposed a computational offloading mech-
namic long-term MEC scenario, i.e., the parameters and conditions                anism based on a two-stage Stackelberg game and used two dynamic
(e.g., energy harvesting, task arrival) may change over time, mak-               iterative algorithms to solve the utility optimization problem in the
ing the solution of the problem extremely challenging. To simplify               game. Although the above studies have their own merits, they do not
the problem, the authors in [5–7] focused their target optimization              consider the computing power of the device itself.
on a short time slot, defined as a brief period during which sys-                    To overcome this deficiency, in recent years, many studies have
tem conditions are assumed to be relatively stable. However, focusing            begun to consider introducing local computing resources. Based on
only on a single time slot may cause system instability problems,                the size of the offloaded data, Hu et al. [13] determined the tasks
since it is difficult to cope with rapid changes. To overcome this               that necessitate local handling. They proposed a MEC system energy
difficulty, we employ Lyapunov optimization theory to transform the              consumption optimization problem and solved it in two stages. Wang
long-term optimization problem into a sequence of short-term optimiza-           et al. [14] divided the computing task into two parts, one of which
tion problems, enabling the achievement of long-term optimization                would be used for local computing. Their research intended to minimize
objectives through the resolution of multiple short-term optimization            the AP’s overall energy consumption, however they did not consider
                                                                                 the energy consumption resulting from the MEC server’s computational
tasks. To enhance the efficiency and rationality of resource allocation,
                                                                                 tasks. They omitted for the energy consumption consumed by the MEC
we introduce Stackelberg game theory. In this framework, the MEC
                                                                                 server during task processing.
server acts as a leader, formulating differentiated pricing strategies to
                                                                                     To more comprehensively address issues such as task offloading and
guide the offloading decisions of multiple user devices. By combin-
                                                                                 energy management, and to achieve optimal performance with limited
ing Lyapunov optimization theory with Stackelberg game theory, we
                                                                                 resources, game theory is an appropriate approach. There is currently
can effectively address the complexities in dynamic environments and
                                                                                 a lot of work taking game theory into MEC. Li et al. [6] described the
achieve long-term optimization objectives.                                       interaction between mobile device (MD) and edge cloud server (ECS)
    This research work makes the following main contributions:                   in the process of computing load as a Stackelberg game and confirmed
                                                                                 the equilibrium of this game. The authors in [5] proposed an optimal
     • We explore task offloading and pricing in a multi-user environ-           resource purchase strategy with a set price, and proposed the optimal
       ment with a single MEC server, incorporating energy harvesting            pricing for edge cloud computing resources utilizing the Stackelberg
       (EH) for users to utilize renewable energy for task computation.          game model. Liu et al. [7] proposed the problem of transmission
     • We utilize Lyapunov optimization theory to adapt to external              power offloading optimization and edge cloud pricing in mobile edge
       changes by breaking down long-term goals into short-term objec-           computing systems, and adopted the offloading strategy and price
       tives, stabilizing the task queue while optimizing performance.           control (OSPC) algorithm based on Stackelberg game to solve it. Bishoyi
                                                                                 et al. [15] presented a distributed algorithm utilizing the alternating
     • We propose a Stackelberg game model for task offloading and               direction multiplier method (ADMM) to solve the Stackelberg game
       pricing, where users consider energy use, queue length, and pric-         they proposed. Although Bishoyi et al. considered the Stackelberg
       ing, while the MEC server applies differentiated pricing based on         game, they only considered the optimization of the problem within one
       offloaded tasks.                                                          time slot. Zeng et al. [16] introduced a reward system to incentivize

                                                                             2
J. Mei et al.                                                                                                              Journal of Systems Architecture 160 (2025) 103360

                Table 1
                Difference between our scheme and the main related schemes.

                 Scheme                      Local computation         MEC energy consumption        Multiple time slots          Differentiated pricing

                 [5] S.-H.Kim et al.         ✓                         ×                             ×                            ×
                 [6] M.Li et al.             ✓                         ×                             ×                            ✓
                 [7] X.Liu et al.            ✓                         ×                             ×                            ✓
                 [8] Z.Ning et al.           ×                         ✓                             ×                            ×
                 [9] Y.Mao et al.            ×                         ×                             ×                            ×
                 [10] X.Zhao et al.          ×                         ✓                             ×                            ×
                 [11] Y.Chen et al.          ×                         ×                             ✓                            ×
                 [12] F.Li et al.            ×                         ✓                             ×                            ✓
                 [13] X.Hu et al.            ✓                         ×                             ×                            ×
                 [14] F.Wang et al.          ✓                         ×                             ✓                            ×
                 [15] P.K.Bishoyi et al.     ✓                         ✓                             ×                            ✓
                 [16] F.Zeng et al.          ✓                         ×                             ×                            ✓
                 Our Scheme                  ✓                         ✓                             ✓                            ✓



volunteer vehicles participating in Vehicular Edge Computing (VEC)
offloading and devised an enhanced genetic algorithm to explore the
optimal strategy for the VEC server. However, they also did not account
for changes in the number of tasks over time.
    Among most of the studies mentioned before, some studies failed to
fully leverage local computing resources, some did not account for the
energy consumption associated with the MEC server’s computations,
and some only focused on the optimization of single slot goals (see
Table 1). To cope with these situations, each user in this paper supports
local computing through EH technology, while the MEC server consid-
ers computing energy consumption as its own overhead. In order to
achieve a reasonable allocation of resources among users and the MEC
server and tackle the long-term optimization problem, we combine the
Stackelberg game with Lyapunov optimization.

3. System model and problem formulation

    In this section, we will provide an overview of the MEC system
architecture and various computation models. Based on these compu-
tational models, the optimization problems for both users and the MEC
server will be deduced.                                                                            Fig. 1. A mobile edge computing system architecture.


3.1. Mobile edge computing system architecture

     We investigate a system architecture consisting of three layers.                MEC server’s computational capacity is exceeded. We consider the cen-
The first layer, called the user terminal device layer, consists of 𝑛                tral cloud to possess formidable computational capabilities, enabling it
user devices (e.g., wearables, tablets, smartphones), indexed by 𝑁 =                 to handle a significant amount of tasks independently.
{1, 2, … , 𝑛}. Each user is equipped with an energy harvesting device                    Various uncertainties and interferences exist in practical application
that enables them to collect energy from the surrounding environment                 environments. To enhance system stability, we consider a time-slot-
to power their own operations. The second layer is known as the MEC                  based system and partition time into equidistant time slots. The system
service layer, consisting of a single MEC server co-located with a base              is indexed by 𝑡 ∈  = {0, 1, 2, … , 𝑇 } with slot length 𝜏.
station. Lastly, the third layer is denoted as the central cloud service
layer, consisting of a cloud server. The three-layer system architecture
is illustrated in Fig. 1.                                                            3.2. Computing task and task queue model
     In this architecture, users can transmit their computational tasks to
the base station via a wireless network using the time division multiple                 For each user 𝑖, at the start of the 𝑡th time slot, the user’s application
access (TDMA) protocol. The base station forwards the tasks to the MEC               requests a set of tasks for computation. The size of the tasks is the task
server for computation. Once tasks are completed, the MEC server sends               arrival rate 𝑎𝑖 (𝑡). The tasks received during the current time slot can
the results back to the base station, which then delivers them back to               only be handled in future time slots. In order to achieve more flexible
the users. If the cumulative tasks offloaded by users exceed the MEC                 task offloading, we assume that the computational tasks of the users
server’s computational capacity, the excess tasks are offloaded to the               follow a data partitioning model [4,17], where the data bits of the
central cloud via a wired network, and the results are subsequently re-              computational tasks are independent and can be arbitrarily partitioned
turned by the central cloud. Excess tasks can be generated by adjusting              into multiple independent subtasks. Each user’s tasks are stored in the
                                                                                                        {               }
the relevant parameters, such as the task arrival rate, and available                task queue 𝑸 = 𝑄1 , 𝑄2 , … , 𝑄𝑛 .
computational resources. By adjusting these parameters, the total of-                    Let 𝑄𝑖 (𝑡) represent the tasks remaining incomplete for the 𝑖th user
floading demand can be controlled, simulating the scenario where the                 in the preceding 𝑡 time slots. The amount of tasks in the queue can be

                                                                                 3
J. Mei et al.                                                                                                                     Journal of Systems Architecture 160 (2025) 103360


adjusted using the following formula.                                                3.3.4. Transmission rate model
𝑄𝑖 (𝑡 + 1) = max{0, 𝑄𝑖 (𝑡) − 𝑞𝑖 (𝑡)} + 𝑎𝑖 (𝑡),                            (1)            In the system architecture, data exchange between the users and
                                                                                     the MEC server occurs through wireless networks. During this data
where 𝑞𝑖 (𝑡) denotes the total tasks processed by user 𝑖 during the 𝑡th              exchange process, we consider Shannon’s formula as the calculation
time slot. 𝑞𝑖 (𝑡) can be modeled as 𝑞𝑖 (𝑡) = 𝑞𝑖0 (𝑡) + 𝑞𝑖1 (𝑡), where 𝑞𝑖0 (𝑡)        formula for the channel’s data transfer rate. Similar to [11], the trans-
denotes the quantity of tasks processed locally by user 𝑖, and 𝑞𝑖1 (𝑡)               mission rate can be modeled as
signifies the quantity of tasks offloaded remotely by user 𝑖. 𝑞𝑖 (𝑡) satisfies                   (          )
                                                                                                      𝑝𝑖 ℎ𝑖
                                                                                     𝑟𝑡𝑟
                                                                                      𝑖 = 𝑤 log2 1 + 𝑤𝑛      ,                                            (10)
0 ≤ 𝑞𝑖 (𝑡) ≤ 𝑄𝑖 (𝑡), 𝑡 ∈  .                                              (2)                             0

                                                                                     where 𝑤 signifies the base station’s bandwidth; 𝑛0 stands for the noise
    According to the definition of queue stability in [18], the constraint           power density; and ℎ𝑖 indicates the channel gain.
for task queue stability is given as                                                     According to the transmission time and transmission rate, the of-
        E{𝑄𝑖 (𝑡)}                                                                    floading task size 𝑞𝑖1 (𝑡) = 𝑇𝑖𝑡𝑟 (𝑡)𝑟𝑡𝑟
                                                                                                                           𝑖 can be deduced.
 lim              = 0, 𝑡 ∈  .                                            (3)
𝑡→+∞       𝑡
                                                                                     3.4. Mobile edge server processing model

3.3. Local execution and communication model
                                                                                         When users offload tasks to the MEC server, the server receives and
3.3.1. Local computing energy consumption model                                      processes these tasks, thereby resulting in energy consumption. The
    Local task execution involves users using their available processing             energy consumption of the MEC server can be modeled as
resources to handle tasks, leading to local energy consumption. This                                      ∑
                                                                                                          𝑛

local energy consumption is dynamic, as it varies based on offloading                𝐻 𝑒 (𝑡) = 𝑘𝑒 (𝑓 𝑒 )2    𝑐
                                                                                                            𝑞𝑖1 (𝑡),                                    (11)
                                                                                                           𝑖=1
strategy. Same as [19], we model it as
                                                                                     where 𝑘𝑒 is the capacitance switching coefficient of the MEC server;
 𝑢
𝐻𝑖0 (𝑡) = 𝑘𝑢𝑖 [𝑓𝑖𝑢 ]2 𝑞𝑖0 (𝑡)𝑑𝑖 ,                                         (4)                                                                             𝑐 (𝑡)
                                                                                     𝑓 𝑒 represents the CPU computing frequency of the MEC server; 𝑞𝑖1
where 𝑘𝑢𝑖 is the capacitance switching coefficient, which is dependent               denotes the cycle count of the offloaded task for user 𝑖 and can be
on the chip architecture [19], and the superscript 𝑢 is employed to                                 𝑐 (𝑡) = 𝑞 (𝑡)𝑑 .
                                                                                     described as 𝑞𝑖1        𝑖1   𝑖
denote user-specific parameters, distinguishing them from the parame-                    Considering the limited resources of the MEC server, it may not be
ters denoted by the same symbol for the MEC server; 𝑓𝑖𝑢 represents the               able to fully process all the tasks offloaded by users. Consequently, the
CPU computing frequency of user 𝑖, while 𝑑𝑖 represents the computing                 MEC server is required to upload the excess tasks to the central cloud
density of user 𝑖.                                                                   for synchronous processing. Let 𝑞 𝑒 represent the processing capacity of
    Given the computation frequency 𝑓𝑖𝑢 and the time slot length 𝜏, the              the MEC server within a time slot, i.e., the maximum number of task
local computing tasks are constrained as
                                                                                     cycles it can handle. When the amount of task cycles offloaded by users
                 𝑓𝑖𝑢 𝜏                                                               exceeds the MEC server’s computational capacity, the excess portion is
0 ≤ 𝑞𝑖0 (𝑡) ≤            ,𝑡 ∈  .                                         (5)
                  𝑑𝑖                                                                 transferred to the central cloud. Consequently, the MEC server’s energy
                                                                                     consumption model is expressed as

3.3.2. Transmission energy consumption model                                         𝐻 𝑒 (𝑡) = 𝑘𝑒 (𝑓 𝑒 )2 min {𝑞 𝑒 , 𝑞 𝑐 (𝑡)} ,                                              (12)
    Let 𝑝𝑖 represent the transmission power of user 𝑖. When users offload                                                  ∑𝑛      𝑐
                                                                                     where 𝑞 𝑒 = 𝑓 𝑒 𝜏 and 𝑞 𝑐 (𝑡) =          𝑖=1 𝑞𝑖1 (𝑡).
tasks to the MEC server, they incur transmission energy consump-
tion [19], which is modeled as
 𝑢                                                                                   3.5. The utility optimization problem
𝐻𝑖1 (𝑡) = 𝑇𝑖𝑡𝑟 (𝑡)𝑝𝑖 ,                                                    (6)

where 𝑇𝑖𝑡𝑟 (𝑡) represents the duration of task transmission for user 𝑖.              3.5.1. MEC server utility optimization problem
    The task transmission time does not exceed one time slot and
                                                                                         The MEC server generates costs while processing tasks. It is assumed
satisfies
                                                                                     that the MEC server obtains revenue by pricing data per cycle. We
0 ≤ 𝑇𝑖𝑡𝑟 (𝑡) ≤ 𝜏 , 𝑡 ∈  .                                                (7)        adopt a differential pricing approach. Let 𝑅𝑒𝑖 (𝑡) represent the fee that
                                                                                     user 𝑖 needs to pay to the MEC server for offloading data per cycle and
                                                                                                      {                          }
                                                                                     define 𝑹(𝒕) = 𝑅𝑒1 (𝑡), 𝑅𝑒2 (𝑡), … , 𝑅𝑒𝑛 (𝑡) . Let 𝜋 𝑒 (𝑡) represent the revenue
   Based on Eqs. (4) and (6), the overall energy consumption 𝐻𝑖𝑢 (𝑡) is
derived as 𝐻𝑖𝑢 (𝑡) = 𝐻𝑖0
                      𝑢 (𝑡) + 𝐻 𝑢 (𝑡).                                               obtained from processing all user tasks by the MEC server, which is
                               𝑖1                                                                          ∑
                                                                                     denoted as 𝜋 𝑒 (𝑡) = 𝑛𝑖=1 𝑅𝑒𝑖 (𝑡)𝑞𝑖1
                                                                                                                       𝑐 (𝑡). Let 𝑐 𝑒 represent the cost incurred by

3.3.3. Energy harvesting model                                                       the MEC server for each unit of energy consumption. When the amount
   We consider the user’s energy harvesting to follow the HUS strat-                 of tasks offloaded by all users exceeds the processing capacity of the
egy [20], where the energy collected during the current time slot is                 MEC server, the MEC server needs to pay a fee to the central cloud for
only available for use in subsequent time slots. Let 𝑒𝑖 (𝑡) represent the            handling. Let 𝑞 𝑟 (𝑡) = max {0, 𝑞 𝑐 (𝑡) − 𝑞 𝑒 } represent the tasks redirected to
energy harvested by user 𝑖, and 𝐵𝑖 (𝑡) represent the remaining energy in             the central cloud and 𝑅𝑐 represent the cost incurred by the MEC server
the battery. The battery energy update is modeled as:                                for offloading data to the central cloud per cycle. The optimization
𝐵𝑖 (𝑡 + 1) = min{max{𝐵𝑖 (𝑡) − 𝐻𝑖𝑢 (𝑡), 0} + 𝑒𝑖 (𝑡), 𝐵𝑖max },              (8)        problem 𝑬𝟏 of the MEC server can be modeled as
                                                                                                             1 ∑
                                                                                                                   𝑇 −1
where 𝐵𝑖max represents the maximum capacity of the battery.                          𝑬𝟏 ∶ max 𝑠 = lim              E{𝜋 𝑒 (𝑡) − 𝜖 𝑒 (𝑡)}                                      (13)
                                                                                             𝑹(𝒕)      𝑇 →+∞ 𝑇
   The overall energy consumption generated by user 𝑖 cannot surpass                                           𝑡=0
the remaining battery energy, i.e.,                                                           s.t.0 ≤ 𝑅𝑒𝑖 (𝑡) ≤ 𝑅𝑒_max
                                                                                                                 𝑖     (𝑡), 𝑖 ∈ 𝑁 , 𝑡 ∈  ,                                  (14)
0 ≤ 𝐻𝑖𝑢 (𝑡) ≤ 𝐵𝑖 (𝑡), 𝑡 ∈  .                                             (9)
                                                                                     where 𝜖 𝑒 (𝑡) = 𝑐 𝑒 𝐻 𝑒 (𝑡) + 𝑅𝑐 𝑞 𝑟 (𝑡); 𝑅𝑒_max
                                                                                                                                𝑖     (𝑡) represents the maximum price
                                                                                     per cycle of data charged by the MEC server to user 𝑖.

                                                                                 4
J. Mei et al.                                                                                                                              Journal of Systems Architecture 160 (2025) 103360


3.5.2. User utility optimization problem                                                            4.2. Problem transformation based on Lyapunov optimization
    A higher total task quantity processed by users in a time slot implies
a reduced task queue, thereby increasing user satisfaction. We adopt the                                 Since the initial optimization problem involves the situation in
logarithmic utility function [21], modeling it as                                                   the long-term range, and the parameters such as energy acquisition
              ∑
              1                                                                                     and task arrival will change over time, this makes it complicated to
𝜔𝑖 (𝑡) = 𝜒           log2 (1 + 𝑞𝑖𝑗 (𝑡)),                                               (15)         solve the initial problem directly. However, Lyapunov optimization
              𝑗=0                                                                                   can transform this long-term problem into multiple tractable short-
where 𝜒 is a weight parameter.                                                                      term problems, so that only these short-term problems need to be dealt
     In general, for a same task, the energy consumption resulting from                             with, and the original difficult problem can be solved more efficiently.
local execution exceeds that of transmitting the same quantity of of-                               Moreover, through Lyapunov optimization, the user’s task queue can
                                                                                                    remain stable (i.e., satisfying Eq. (3)). Therefore Lyapunov optimization
floaded tasks [22]. Therefore, we consider that offloading an appro-
                                                                                                    theory is employed.
priate amount of tasks remotely is advantageous for reducing energy
                                                                                                         The Lyapunov function for user 𝑖 can be defined as
consumption. Consequently, we incorporate the energy saved by of-
                                                                                                             1
floading tasks compared to local processing into the user utility model.                            𝐿𝑖 (𝑡) ≜ [𝑄𝑖 (𝑡)]2 .                                                 (19)
                                                                                                             2
The saved energy is modeled as
                                𝑞 (𝑡)𝑝
𝜓𝑖 (𝑡) = 𝑘𝑢𝑖 [𝑓𝑖𝑢 ]2 𝑞𝑖1 (𝑡)𝑑𝑖 − 𝑖1 𝑡𝑟 𝑖 .                          (16)                                 Based on [24], the Lyapunov drift for user 𝑖 can be defined as
                                   𝑟𝑖                                                                         {                          }
                                                                                                    𝛥𝑖 (𝑡) ≜ E 𝐿𝑖 (𝑡 + 1) − 𝐿𝑖 (𝑡)|𝑄𝑖 (𝑡) .                             (20)
    Considering that the MEC server does not provide services to users
for free, user 𝑖 is required to pay a certain cost to the MEC server,                                   To balance queue stability and user utility optimization, we intro-
determined by the offloaded task cycle count. The offloading cost is                                duce the concept of drift-plus-penalty function, which makes a trade-off
modeled as                                                                                          between task queue stability and user utility. By incorporating an
𝑐𝑖 (𝑡) = 𝑅𝑒𝑖 (𝑡)𝑞𝑖1
                 𝑐
                    (𝑡).                                                               (17)         additional penalty term into the Lyapunov function, we can optimize
                                                                                                    the user utility while satisfying the queue stability condition.
                                                                                                        To ensure queue stability, it is necessary to minimize Lyapunov
   The user’s utility is related to the total quantity of tasks handled by                          drift. However, the objective is to maximize user utility, we transform
the user and the energy savings achieved. The user’s costs are related to                           the maximization of the user utility function into the minimization of
the total energy consumption and the offloading costs. Therefore, the                               the user loss function (i.e., the negative of the user utility function).
user’s utility optimization problem 𝑷 𝟏 can be described as                                         The drift-plus-penalty function for user 𝑖 is represented as 𝛥𝑖 (𝑡) −
                                                                                                        {            }
                                                                                                    𝑉 E 𝑢𝑖 (𝑡)|𝑄𝑖 (𝑡) , where V is a non-negative weight parameter.
                                    1 ∑
                                           𝑇 −1
𝑷𝟏 ∶       max            𝑢𝑖 = lim        E{𝑢𝑖 (𝑡)}                                    (18)             When considering Lyapunov drift-plus-penalty function, the lack of
       𝑇𝑖𝑡𝑟 (𝑡),𝑞𝑖0 (𝑡)       𝑇 →+∞ 𝑇
                                      𝑡=0                                                           information about the following time slot (i.e., 𝐿𝑖 (𝑡 + 1)) makes direct
                     s.t.(2), (3), (5), (7) and (9),                                                solving challenging. To remove the reliance on future information, we
                                                                                                    use scaling to get an upper bound on this function.
where 𝑢𝑖 (𝑡) = 𝜋𝑖𝑢 (𝑡) − 𝜖𝑖𝑢 (𝑡), 𝜋𝑖𝑢 (𝑡) = 𝜔𝑖 (𝑡) + 𝜓𝑖 (𝑡), 𝜖𝑖𝑢 (𝑡) = 𝐻𝑖𝑢 (𝑡) + 𝜆𝑐𝑖 (𝑡), and
𝜆 is a weight parameter.                                                                            Theorem 1. When a control parameter 𝑉 > 0 is chosen, and considering
                                                                                                                        {              }
                                                                                                    that both 𝑞𝑖 (𝑡) ∈ 0, 1, … , 𝑄𝑖 (𝑡) and 𝑎𝑖 (𝑡) ∈ {0, 1, … , 𝑎max }, we obtain
4. Problem analysis and solution                                                                                {             }
                                                                                                    𝛥𝑖 (𝑡) − 𝑉 E 𝑢𝑖 (𝑡)|𝑄𝑖 (𝑡)
                                                                                                                {                            }
                                                                                                       ≤ 𝑧 − E 𝑞𝑖 (𝑡)[𝑄𝑖 (𝑡) + 𝑎𝑖 (𝑡)]|𝑄𝑖 (𝑡)
    In this section, the game relationship and related processes between                                        {             }
the users and the MEC server will be introduced in detail. To transform                                    − 𝑉 E 𝑢𝑖 (𝑡)|𝑄𝑖 (𝑡) ,                                                (21)
long-term optimization problem into multiple short-term optimization
                                                                                                    where 𝑧 = 21 {[𝑄𝑖 (𝑡)]2 + (𝑎max )2 } + 𝑄𝑖 (𝑡)𝑎𝑖 (𝑡).
problem and ensure the task queue’s stability, the Lyapunov optimiza-
tion theory will be adopted. The optimal strategies of both users and
MEC will be taken into consideration.                                                               Proof. Taking the square of both sides of Eq. (1) and we find that
                                                                                                    (𝑚𝑎𝑥[𝑥, 0])2 < 𝑥2 for any 𝑥 ∈ R. Therefore, the inequality can be
4.1. The game relationship between users and MEC server                                             calculated as
                                                                                                                   {                                }2
                                                                                                    [𝑄𝑖 (𝑡 + 1)]2 = max{0, 𝑄𝑖 (𝑡) − 𝑞𝑖 (𝑡)} + 𝑎𝑖 (𝑡)
     Without adequate incentive measures, the MEC server may be less
                                                                                                                  ≤ [𝑄𝑖 (𝑡) − 𝑞𝑖 (𝑡)]2 + [𝑎𝑖 (𝑡)]2
willing to participate in computation offloading [23]. To incentivize the
MEC server, we employ a strategy rooted in Stackelberg game theory                                               + 2𝑎𝑖 (𝑡)[𝑄𝑖 (𝑡) − 𝑞𝑖 (𝑡)]
to enable multiple users and the MEC server to both achieve their                                                 = [𝑄𝑖 (𝑡)]2 + [𝑞𝑖 (𝑡)]2 + [𝑎𝑖 (𝑡)]2
respective benefits. In this game process, the MEC server plays the role                                         − 2𝑄𝑖 (𝑡)𝑞𝑖 (𝑡) + 2𝑎𝑖 (𝑡)[𝑄𝑖 (𝑡) − 𝑞𝑖 (𝑡)]
of the leader, while users act as followers. In the 𝑡th time slot, firstly,
                                                                                                                  ≤ 2[𝑄𝑖 (𝑡)]2 + (𝑎max )2 − 2𝑄𝑖 (𝑡)𝑞𝑖 (𝑡)
the MEC server will provide each user with an initial service quotation
𝑅𝑒𝑖 (𝑡). Secondly, each user is required to make task offloading strategy                                        + 2𝑎𝑖 (𝑡)[𝑄𝑖 (𝑡) − 𝑞𝑖 (𝑡)].                                          (22)
(i.e., 𝑞𝑖0 (𝑡) and 𝑇𝑖𝑡𝑟 (𝑡)) based on the price. Subsequently, the MEC
server will update the corresponding prices based on the user’s remote
                                                                                                         Based on Definition (19) and the aforementioned inequality, the
offloading task 𝑞𝑖1 (𝑡) and its own computational capacity. Following
                                                                                                    Lagrangian function for the time slot 𝑡 + 1 is computed as
this, users will update their task offloading decisions. As this process
                                                                                                                  1
continues, through multiple iterative steps, until a balance is reached                             𝐿𝑖 (𝑡 + 1) = [𝑄𝑖 (𝑡 + 1)]2
                                                                                                                  2
between users’ task offloading decisions and the MEC server prices, the                                                        (𝑎max )2
iteration for the current time slot concludes. At this point, a Stackelberg                                    ≤ [𝑄𝑖 (𝑡)]2 +              − 𝑄𝑖 (𝑡)𝑞𝑖 (𝑡)            (23)
                                                                                                                                  2
equilibrium is achieved between users and the MEC server.                                                      + 𝑎𝑖 (𝑡)[𝑄𝑖 (𝑡) − 𝑞𝑖 (𝑡)].

                                                                                                5
J. Mei et al.                                                                                                                        Journal of Systems Architecture 160 (2025) 103360



     Based on Definitions (19) and (20), as well as Eq. (23), the inequal-                  The first-order partial derivative of 𝑖 with respect to 𝑞𝑖0 (𝑡) is given
ity can be derived as                                                                   by the following expression
         1                                                                                𝜕𝑖       𝜕𝑖 (𝑡)
𝛥𝑖 (𝑡) ≤ {[𝑄𝑖 (𝑡)]2 + (𝑎max )2 }                                                                  =           + 𝜇1 + 𝜇2 𝑘𝑢𝑖 (𝑓𝑖𝑢 )2 𝑑𝑖 + 𝜇5 − 𝜇6 .               (31)
         2                                                                              𝜕 𝑞𝑖0 (𝑡)   𝜕 𝑞𝑖0 (𝑡)
                           {                              }
       + 𝑄𝑖 (𝑡)𝑎𝑚𝑎𝑥
                𝑖   (𝑡) − E 𝑞𝑖 (𝑡)[𝑄𝑖 (𝑡) + 𝑎𝑖 (𝑡)]|𝑄𝑖 (𝑡) .          (24)
                                                                                                                                                               ∗ (𝑡)
                                                                                             Solving for 𝜕𝑖 ∕𝜕 𝑞𝑖0 (𝑡) = 0, the optimal local task workload 𝑞𝑖0
                                              }    {                                    can be calculated as
   By adding the expression −𝑉 E 𝑢𝑖 (𝑡)|𝑄𝑖 (𝑡) to both sides of Eq. (24),                          𝑉𝜒
                                                                                         ∗
we can deduce Eq. (21).                                                                 𝑞𝑖0 (𝑡) =        − 1,                                                 (32)
                                                                                                  𝑔 ln 2
    Taking into account the impact of task queue length on user expe-                   where 𝑔 = [𝜇1 + 𝜇5 − 𝜇6 − 𝑄𝑖 (𝑡) − 𝑎𝑖 (𝑡)] + 𝑘𝑢𝑖 (𝑓𝑖𝑢 )2 𝑑𝑖 [𝜇2 + 𝑉 ].
rience, we incorporate the task queue length into the consideration of
user utility through Lyapunov optimization. According to Theorem 1,                     4.3.2. Optimal offloading task strategy
the original problem 𝑷 𝟏 is transformed into 𝑷 𝟐, described as                               Compute the first-order partial derivative of 𝑖 with respect to 𝑇𝑖𝑡𝑟 (𝑡),
                       {                    }                                             𝜕𝑖      𝜕𝑖 (𝑡)
𝑷 𝟐 ∶ min 𝑖 (𝑡) = − 𝑞𝑖 (𝑡)[𝑄𝑖 (𝑡) + 𝑎𝑖 (𝑡)] − 𝑉 𝑢𝑖 (𝑡)           (25)                           =            + 𝜇1 𝑟𝑡𝑟                                          (33)
       𝑇𝑖𝑡𝑟 (𝑡),𝑞𝑖0 (𝑡)                                                                     𝑡𝑟                      𝑖 + 𝜇2 𝑝𝑖 + 𝜇3 − 𝜇4 .
                                                                                        𝜕 𝑇𝑖 (𝑡)   𝜕 𝑇𝑖𝑡𝑟 (𝑡)
                          s.t.(2), (5), (7) and (9).
                                                                                              By solving for 𝜕𝑖 ∕𝜕 𝑇𝑖𝑡𝑟 (𝑡) = 0, the optimal task transmission time
                                                                                        𝑇𝑖𝑡𝑟 (𝑡) can be obtained as
                                                                                                    𝑉𝜒     1
4.3. Optimal strategy for users                                                         𝑇𝑖∗ (𝑡) =        −    ,                                                 (34)
                                                                                                   𝑦 ln 2 𝑟𝑡𝑟
                                                                                                           𝑖

   Each user in the system can collect a certain amount of energy                       where 𝑦 = [𝑉 𝜆𝑅𝑒𝑖 (𝑡)𝑑𝑖 − 𝑄𝑖 (𝑡) − 𝑎𝑖 (𝑡) + 𝜇1 − 𝑉 𝑘𝑢𝑖 (𝑓𝑖𝑢 )2 𝑑𝑖 ]𝑟𝑡𝑟
                                                                                                                                                            𝑖 + [2𝑉 + 𝜇2 ]𝑝𝑖 +
during each time slot and store it in a battery. Users can use this                     𝜇3 − 𝜇4 .
battery energy for local task processing or task offloading to the MEC                     To ensure that the user 𝑖 maintains an transmission time 𝑇𝑖∗ (𝑡) ≥ 0,
server. Each user must address two essential inquiries: (i) How many                    we solve for 𝑇𝑖∗ (𝑡) = 0 to obtain the maximum price the user 𝑖 can
tasks should be processed locally in each time slot; (ii) How much task                 accept, denoted as 𝑅𝑒_max
                                                                                                                𝑖    (𝑡).
transfer time is required in each time slot.                                             𝑒_max         𝜒         𝜇4 − 𝜇3 − 𝑝𝑖 (2𝑉 + 𝜇2 )
                                                                                        𝑅𝑖     (𝑡) =          +
                                                                                                     𝜆𝑑𝑖 ln 2            𝑉 𝜆𝑑𝑖 𝑟𝑡𝑟
                                                                                                                                𝑖
4.3.1. Optimal local task strategy                                                                       𝑄𝑖 (𝑡) + 𝑎𝑖 (𝑡) − 𝜇1 + 𝑉 𝑘𝑢𝑖 (𝑓𝑖𝑢 )2 𝑑𝑖
    Referring to (25), we calculate the first-order partial derivative of                           +                                               .                           (35)
                                                                                                                           𝑉 𝜆𝑑𝑖
𝑖 (𝑡) with respect to 𝑞𝑖0 (𝑡) as
𝜕𝑖 (𝑡)                                                                                    By referring to Eq. (34), we can obtain the optimal amount of
          = −[𝑄𝑖 (𝑡) + 𝑎𝑖 (𝑡)]                                                                                          ∗ (𝑡)) for user 𝑖 within the 𝑡th time slot.
𝜕 𝑞𝑖0 (𝑡)                                                                               remote offloading tasks (i.e., 𝑞𝑖1
                      𝜒
          − 𝑉[                    − 𝑘𝑢𝑖 [𝑓𝑖𝑢 ]2 𝑑𝑖 ],                (26)                ∗
                                                                                        𝑞𝑖1 (𝑡) = 𝑇𝑖∗ (𝑡)𝑟𝑡𝑟
                                                                                                          𝑖 .                                                                   (36)
               (1 + 𝑞𝑖0 (𝑡)) ln 2
and the second-order partial derivative of 𝑖 (𝑡) with respect to 𝑞𝑖0 (𝑡) is
computed as
                                                                                        4.4. Optimal strategy for MEC server
𝜕 2 𝑖 (𝑡)         𝑉𝜒
           =                     .                                    (27)
𝜕 𝑞𝑖0 (𝑡)2   [1 + 𝑞𝑖0 (𝑡)]2 ln 2                                                           According to the dynamic programming theory, it breaks down the
                                                                                        problem into a sequence of sub-problems, and uses the properties of
     Since Eq. (27) is non-negative, 𝑖 (𝑡) exhibits convexity concerning               overlapping sub-problems to reduce the amount of computation. In the
𝑞𝑖0 (𝑡).                                                                                same way, in long-term optimization problem, each time slot can be
     Following Eq. (25), we calculate the first-order partial derivative of             considered a stage, and by making optimal strategies in each stage,
𝑖 (𝑡) with respect to 𝑇𝑖𝑡𝑟 (𝑡) as                                                      optimal result can be achieved in the long run. According to [17], the
 𝜕𝑖 (𝑡)                                                                                optimization problem 𝑬𝟏 can be transformed into 𝑬𝟐 as
           = −[𝑄𝑖 (𝑡) + 𝑎𝑖 (𝑡)]𝑟𝑡𝑟
                                𝑖
𝜕 𝑇𝑖𝑡𝑟 (𝑡)                                                                              𝑬𝟐 ∶ max (𝑡) = 𝜋 𝑒 (𝑡) − 𝜖 𝑒 (𝑡)                                                       (37)
                                                                                                𝑹(𝒕)
                           𝜒 𝑟𝑡𝑟
                              𝑖 (𝑡)
           − 𝑉[                           + 𝑘𝑢𝑖 [𝑓𝑖𝑢 ]2 𝑟𝑡𝑟
                                                         𝑖 (𝑡)𝑑𝑖 + 𝐶],      (28)                        s.t.(14).
                     (1 + 𝑞𝑖1   (𝑡)) ln 2
where 𝐶 = −2𝑝𝑖 − 𝜆𝑖 𝑅𝑒𝑖 (𝑡)𝑟𝑡𝑟     𝑖 𝑑𝑖 ; and the second-order partial derivative
of 𝑖 (𝑡) with respect to 𝑇𝑖𝑡𝑟 (𝑡) is calculated as                                        For the MEC server, the optimal price (i.e., 𝑅𝑒𝑖 (𝑡)) for each user
                                                                                        needs to be determined. By combining Eqs. (36) and (37), the first-order
𝜕 2 𝑖 (𝑡)     𝑉 𝜒(𝑟𝑡𝑟𝑖 (𝑡))
                            2
           =                     .                                           (29)       partial derivative of (𝑡) with respect to 𝑅𝑒𝑖 (𝑡) is calculated as
    𝑡𝑟
𝜕 𝑇𝑖 (𝑡)2    (1 + 𝑞𝑖1 (𝑡))2 ln 2
                                                                                                  ⎧𝑞 ∗ (𝑡)𝑑𝑖                                   if𝑞 𝑒 < 𝑞 𝑐 (𝑡);
                                                                                                  ⎪ 𝑖1
      Since Eq. (29) is non-negative, 𝑖 (𝑡) is also convex with respect to                       ⎪ 𝑉 𝜆𝑑𝑖 [𝑟𝑖 ] 𝜒 𝑒
                                                                                                        2   2   𝑡𝑟 2
                                                                                                                                𝑐
𝑇𝑖𝑡𝑟 (𝑡). In conclusion, 𝑖 (𝑡) is a convex function with respect to 𝑞𝑖0 (𝑡) and        𝜕(𝑡)     ⎪− 𝑦2 ln 2 [𝑅𝑖 (𝑡) − 𝑅 ],
                                                                                           𝑒    = ⎨                                                                             (38)
is also convex with respect to 𝑇𝑖𝑡𝑟 (𝑡), and the constraints of problem 𝑷 𝟐             𝜕 𝑅𝑖 (𝑡) ⎪𝑞 ∗ (𝑡)𝑑                                     if𝑞 𝑒 ≥ 𝑞 𝑐 (𝑡);
                                                                                                    𝑖1      𝑖
are affine. Therefore, 𝑷 𝟐 can be solved using the method of Lagrange                             ⎪ 2 2 𝑡𝑟 2
                                                                                                  ⎪− 𝑉 𝜆𝑑𝑖 [𝑟𝑖 ] 𝜒 [𝑅𝑒 (𝑡) − 𝑐 𝑒 𝑘𝑒 (𝑓 𝑒 )2 ],
multipliers. Let 𝝁 = {𝜇1 , 𝜇2 , … , 𝜇6 } denote the Lagrange multipliers.                         ⎩       𝑦2 ln 2    𝑖
Similar to [25], the Lagrangian function of Eq. (25) is expressed as                    and the second-order partial derivative of (𝑡) with respect to 𝑅𝑒𝑖 (𝑡) is
𝑖 (𝑞𝑖0 (𝑡), 𝑇𝑖𝑡𝑟 (𝑡), 𝝁) = 𝑖 (𝑡) + 𝜇1 [𝑞𝑖 (𝑡) − 𝑄𝑖 (𝑡)]                               calculated as
                                                                                                   ⎧ 2𝑉 2 𝜆𝑑𝑖2 [𝑟𝑡𝑟𝑖 ]2 𝜒 𝑧1
                          + 𝜇2 [𝐻𝑖𝑢 (𝑡) − 𝐵𝑖 (𝑡)] + 𝜇3 [𝑇𝑖𝑡𝑟 (𝑡) − 𝜏]                   𝜕 2 (𝑡)   ⎪                         ,   if𝑞 𝑒 < 𝑞 𝑐 (𝑡);
                                                                                                           𝑦3 ln 2
                                                    𝑓 𝑢𝜏                                          =⎨                                                                            (39)
                          − 𝜇4 𝑇𝑖𝑡𝑟 + 𝜇5 [𝑞𝑖0 (𝑡) − 𝑖 ] − 𝜇6 𝑞𝑖0 (𝑡).                   𝜕 𝑅𝑒𝑖 (𝑡)2 ⎪ 2𝑉 2 𝜆𝑑𝑖2 [𝑟𝑡𝑟𝑖 ]2 𝜒 𝑧2
                                                                            (30)                                             ,   if𝑞 𝑒 ≥ 𝑞 𝑐 (𝑡);
                                                     𝑑𝑖                                            ⎩       𝑦3 ln 2


                                                                                    6
J. Mei et al.                                                                                                                      Journal of Systems Architecture 160 (2025) 103360


where 𝑧1 = [𝑄𝑖 (𝑡) + 𝑎𝑖 (𝑡) − 𝜇1 + 𝑉 𝑘𝑢𝑖 (𝑓𝑖𝑢 )2 𝑑𝑖 − 𝑉 𝜆𝑅𝑐𝑖 (𝑡)𝑑𝑖 ]𝑟𝑡𝑟
                                                                     𝑖 − [2𝑉 + 𝜇2 ]𝑝𝑖 −       Proof. By taking the first-order partial derivative of Eq. (34) with
                                                                                                                                                                2
𝜇3 + 𝜇4 and 𝑧2 = [𝑄𝑖 (𝑡) + 𝑎𝑖 (𝑡) − 𝜇1 + 𝑉 𝑘𝑢𝑖 (𝑓𝑖𝑢 )2 𝑑𝑖 − 𝑉 𝜆𝑐 𝑒 𝑘𝑒 (𝑓 𝑒 )2 𝑑𝑖 ]𝑟𝑡𝑟
                                                                                   𝑖 −        respect to 𝑅𝑒𝑖 (𝑡), 𝜕 𝑇𝑖∗ (𝑡)∕𝜕 𝑅𝑒𝑖 (𝑡) = −𝑉 2 𝜒 𝜆𝑑𝑖 𝑟𝑡𝑟
                                                                                                                                                    𝑖 ∕ (𝑦 ln 2) is obtained. It is
[2𝑉 + 𝜇2 ]𝑝𝑖 − 𝜇3 + 𝜇4 .                                                                      evident that 𝜕 𝑇𝑖∗ (𝑡)∕𝜕 𝑅𝑒𝑖 (𝑡) < 0, indicating that 𝑇𝑖𝑡𝑟 (𝑡) decreases as 𝑅𝑒𝑖 (𝑡)
     Due to 𝑦 > 0 and the numerator of the second derivative of (𝑡)                          increases. This implies that as the price of the MEC server increases,
with respect to 𝑅𝑒𝑖 (𝑡) being constant when the MEC server makes game                         users are less willing to offload tasks.
decisions, the sign of the second derivative depends on the sign of the
numerator. Consequently, in any iteration, (𝑡) is either a concave or
                                                                                              Lemma 4. If the optimal task transmission time 𝑇𝑖∗ (𝑡) of user 𝑖 is fixed,
convex function relative to 𝑅𝑒𝑖 (𝑡). If (𝑡) exhibits convexity concerning                         (       )
                                                                                              the  𝑅𝑒𝑖 (𝑡) of the MEC server takes the maximum value at 𝑅∗𝑖 (𝑡).
𝑅𝑒𝑖 (𝑡), then the optimal solution of 𝑅𝑒𝑖 (𝑡) is at one of the endpoints of the
constraint range. Therefore, we can deduce that the optimal solution for
𝑅𝑒𝑖 (𝑡), denoted as 𝑅∗𝑖 (𝑡), is calculated as                                                 Proof. Based on the optimal strategy for the MEC server discussed
𝑅∗𝑖 (𝑡) = 𝑅𝑖𝑒_max (𝑡).                                                           (40)         in this section of the paper, it can be concluded that the optimization
                                                                                              function (i.e., (𝑡)) of the MEC server is either concave or convex with
                                                                                              respect to the variable 𝑅𝑒𝑖 (𝑡) in each iteration. When the optimization
    If (𝑡) exhibits concavity concerning 𝑅𝑒𝑖 (𝑡), and it is straightforward                  function (𝑡) is convex, the efficiency of the MEC server reaches its
to ascertain that the constraints of optimization problem 𝑬𝟐 are affine,                      maximum at 𝑅𝑒𝑖 (𝑡) = 𝑅∗𝑖 (𝑡), as shown in Eq. (40). Similarly, when (𝑡)
then we can employ the Lagrange multiplier method to obtain the                               is concave, the efficiency of the MEC server reaches its maximum at
solution for 𝑅∗𝑖 (𝑡). The Lagrangian function for Eq. (37) is expressed                       𝑅𝑒𝑖 (𝑡) = 𝑅∗𝑖 (𝑡), as indicated by Eq. (42). According to Definition 1,
as                                                                                            𝑅𝑆𝑖 𝐸 (𝑡) = 𝑅∗𝑖 (𝑡).
𝑒 (𝑹(𝒕), 𝜼) = (𝑡) − 𝜂1 𝑅𝑒𝑖 (𝑡) + 𝜂2 [𝑅𝑒𝑖 (𝑡) − 𝑅𝑒_max (𝑡)],                    (41)                              (              )
                                                  𝑖                                                In conclusion, 𝑇𝑖∗ (𝑡), 𝑅∗𝑖 (𝑡) represents the optimal decision for task
                                                                                              transmission time and price, and it is also the Stackelberg equilibrium
where 𝜼 = {𝜂1 , 𝜂2 } denotes the Lagrange multipliers, with each multi-                                  (                 )
plier being non-negative.                                                                     solution 𝑇𝑖𝑆 𝐸 (𝑡), 𝑅𝑆𝑖 𝐸 (𝑡) .
     By solving for 𝜕𝑒 ∕𝜕 𝑅𝑒𝑖 (𝑡) = 0, the optimal processing price 𝑅∗𝑖 (𝑡) is
calculated as                                                                                 5. Multi-device task offloading and pricing mechanism algorithm
          ⎧ ∗                          𝑦2 ln 2          𝑒   𝑐
          ⎪[𝑞𝑖1 (𝑡)𝑑𝑖 − 𝜂1 + 𝜂2 ] 𝑉 2 𝜆𝑑 2 (𝑟𝑡𝑟 )2 𝜒 if𝑞 < 𝑞 (𝑡);                                We will delve into the updates of Lagrange multipliers and MEC
          ⎪                              𝑖 𝑖
                                                                                              server prices in this section. At the end of this section, pseudocode for
          ⎪+𝑅𝑐 ,
𝑅∗𝑖 (𝑡) = ⎨                            𝑦2 ln 2
                                                                         (42)                 the algorithm that describes the game process between users and the
             ∗                                          𝑒   𝑐
          ⎪[𝑞𝑖1 (𝑡)𝑑𝑖 − 𝜂1 + 𝜂2 ] 𝑉 2 𝜆𝑑 2 (𝑟𝑡𝑟 )2 𝜒 if𝑞 ≥ 𝑞 (𝑡);                             MEC server is provided.
          ⎪                              𝑖 𝑖
          ⎪+𝑐 𝑒 𝑘𝑒 (𝑓 𝑒 )2 ,
          ⎩                                                                                   5.1. Lagrange multiplier update strategy

                                                                                                 According to Eqs. (32) and (34), the optimal local task workload and
4.5. Stackelberg equilibrium analysis                                                         task transmission time can be determined, respectively. As the results
                                                               (                              are obtained using the Lagrange multiplier method, it is necessary to
     In this section, we demonstrate that the optimal strategy 𝑇𝑖∗ (𝑡), 𝑅∗𝑖                   update these multipliers to ensure the satisfaction of constraints during
 (𝑡)), 𝑖 ∈ 𝑁 between the users and the MEC server is the Stakelberg                           the optimization process. A standard subgradient method is employed
equilibrium solution. We consider the MEC server as a leader in a                             to update the Lagrange multipliers (i.e., 𝝁). The update method is
Stackelberg game, and users as followers. For simplicity, the game equi-                      shown as
librium within a time slot will be analyzed. The Stackelberg equilibrium                           {       [                ]}+
is defined in the following.                                                                  𝜇1 = 𝜇1 + 𝛼 𝑞𝑖 (𝑡) − 𝑄𝑖 (𝑡)       ,
                                                                                                   {       [ 𝑢               ]}+
               ( 𝑆𝐸               )                                                           𝜇2 = 𝜇2 + 𝛼 𝐻𝑖 (𝑡) − 𝐵𝑖 (𝑡)         ,
Definition 1.    𝑇𝑖 (𝑡), 𝑅𝑆𝑖 𝐸 (𝑡) is a Stackelberg equilibrium solution                           {       [            ]}+
                                                                                              𝜇3 = 𝜇3 + 𝛼 𝑇𝑖∗ (𝑡) − 𝜏        ,
when the price 𝑅∗𝑖 (𝑡) of leader is determined, and 𝑇𝑖𝑆 𝐸 (𝑡) satisfies                            [              ]+
  (         )                     { ( 𝑡𝑟 )}                                                   𝜇4 = 𝜇4 − 𝛼 𝑇𝑖∗ (𝑡) ,                                                (45)
𝑖 𝑇𝑖𝑆 𝐸 (𝑡) =         inf          𝑖 𝑇𝑖 (𝑡) , ∀𝑡 ∈  ,                (43)                       {        [
                   𝑇𝑖min (𝑡)≤𝑇𝑖𝑡𝑟 (𝑡)≤𝑇𝑖max (𝑡)                                                                         𝑓 𝑢 𝜏 ]}+
                                                                                              𝜇5 = 𝜇5 + 𝛼 𝑞𝑖0   ∗
                                                                                                                   (𝑡) − 𝑖          ,
and when the task transmission time 𝑇𝑖∗ (𝑡) is determined, and 𝑅𝑆𝑖 𝐸 (𝑡)                                                 𝑑𝑖
                                                                                                   [       ∗
                                                                                                                  ] +
satisfies                                                                                     𝜇6 = 𝜇6 − 𝛼 𝑞𝑖0 (𝑡) ,
  (        )                { ( 𝑒 )}
 𝑅𝑆𝑖 𝐸 (𝑡) =     sup         𝑅𝑖 (𝑡) , ∀𝑡 ∈  .                  (44)                        where 𝛼 represents the iteration step size and [𝑥]+ = max{0, 𝑥}.
                 𝑅min    𝑒      max (𝑡)
                  𝑖 (𝑡)≤𝑅𝑖 (𝑡)≤𝑅𝑖                                                                 The local task workload and task transmission time for each user are
                                                            (              )                  computed by updating the Lagrange multipliers and the price in each
    Next, it will be verified whether the optimal solution 𝑇𝑖∗ (𝑡), 𝑅∗𝑖 (𝑡)                   iteration.
                                         ( 𝑆𝐸            )
is the Stackelberg equilibrium solution 𝑇𝑖 (𝑡), 𝑅𝑆𝑖 𝐸 (𝑡) .
                                                                                              5.2. Price update strategy
                                                                     (        )
Lemma 2. If the price 𝑅∗𝑖 (𝑡) of the leader is fixed, the function 𝑖 𝑇𝑖𝑡𝑟 (𝑡)
                                        ∗
of user 𝑖 takes the minimum value at 𝑇𝑖 .                                                         In one iteration, when the function (𝑡) exhibits convexity con-
                                                                                              cerning the price 𝑅𝑒𝑖 (𝑡), the optimal price 𝑅∗𝑖 (𝑡) = 𝑅𝑒_max
                                                                                                                                                       𝑖    (𝑡). However,
                                                                                              when the function (𝑡) is concave with respect to the price 𝑅𝑒𝑖 (𝑡), it is
Proof. Based on Eq. (29), 𝑖 (𝑡) exhibits convexity concerning 𝑇𝑖𝑡𝑟 (𝑡),                      difficult to compute the optimal price 𝑅∗𝑖 (𝑡) based on 𝜕(𝑡)∕𝜕 𝑅𝑒𝑖 (𝑡) = 0.
indicating that the function 𝑖 (𝑡) attains its minimum value at 𝑇𝑖∗ (𝑡).                     To address this, the gradient ascent method is employed to update the
According to Definition 1, 𝑇𝑖∗ (𝑡) is the 𝑇𝑖𝑆 𝐸 (𝑡).                                          price 𝑅𝑒𝑖 (𝑡), using the first-order partial derivative of the MEC server
                                                                                              utility function with respect to the price as the Marginal Utility [26] to
                                                                                              update the price. The update expression can be derived as
Lemma 3. For users, the optimal task transmission time 𝑇𝑖∗ (𝑡) decreases
                                                                                                                         𝜕(𝑡)
with the increased price 𝑅𝑒𝑖 (𝑡).                                                             𝑅𝑒_𝜅+1 (𝑡) = 𝑅𝑒_𝜅
                                                                                                             𝑖 (𝑡) + 𝛽          ,                                     (46)
                                                                                                𝑖
                                                                                                                       𝜕 𝑅𝑒_𝜅
                                                                                                                          𝑖 (𝑡)

                                                                                          7
J. Mei et al.                                                                                                                Journal of Systems Architecture 160 (2025) 103360


where 𝛽 represents the iteration step size and 𝜅 represents the number                Algorithm 2 Multi-Device Task Offloading and Pricing Mechanism
of iterations within the current time slot.                                           Algorithm (MDTOPMA)
    Due to the constraint (14), after one iteration, the service price of              1: Input:𝑘𝑢𝑖 , 𝑓𝑖𝑢 , 𝑑𝑖 , 𝜏, 𝑝𝑖 , 𝐵𝑖max , 𝑤, ℎ𝑖 , 𝑛0 , 𝑘𝑒 , 𝑓 𝑒 , 𝑐 𝑒 , 𝑅𝑐 , 𝑖 ∈ 𝑁;
the MEC server is calculated as                                                        2: Output:optimal solution 𝑞𝑖0            ∗ (𝑡), 𝑇 ∗ (𝑡), 𝑅∗ (𝑡), 𝑖 ∈ 𝑁, 𝑡 ∈  , 𝑡𝑜𝑡𝑎𝑙𝐼 𝑡;
                                                                                                                                         𝑖        𝑖
                 {               {            }}                                       3: initial 𝜆, 𝛼, 𝛽, 𝑄𝑖 (0), 𝐵𝑖 (0), 𝝁 ← [𝜇1 , ..., 𝜇6 ], 𝝈 ← [𝜎1 , ..., 𝜎6 ];
𝑅𝑖𝑒_𝜅+1 (𝑡) = min 𝑅𝑒_max
                   𝑖     (𝑡), max 0, 𝑅𝑖𝑒_𝜅 (𝑡) .                     (47)                   𝒑𝒓𝒆
                                                                                       4: 𝝁     ← 𝝁;
                                                                                       5: for all 𝑖 ∈ 𝑁 do
    Finally, the price of the (𝜅 + 1)-th iteration within the 𝑡th time slot            6:     Calculate 𝑟𝑡𝑟    𝑖 by Eq. (10);
can be expressed as                                                                    7: end for
              ⎧𝑅𝑒_𝜅 (𝑡),           2 (𝑡)
                            if 𝜕𝑅𝜕 𝑒_𝜅
                                                                                       8: for 𝑡 ∈  do
              ⎪ 𝑖                         2 ≤ 0;
  𝑒_𝜅+1                            𝑖 (𝑡)                                               9:     𝑖𝑡 ← 0;
𝑅𝑖      (𝑡) = ⎨                                                       (48)
                𝑒_max            𝜕 2 (𝑡)                                                     while 𝑖𝑡 ≤ 𝑡𝑜𝑡𝑎𝑙𝐼 𝑡 or 𝝁 − 𝝁𝒑𝒓𝒆 ≥ 𝝈 do
              ⎪𝑅𝑖      (𝑡), if 𝜕𝑅𝑒_𝜅 (𝑡)2 > 0;                                        10:
              ⎩                    𝑖                                                  11:          for all 𝑖 ∈ 𝑁 do
                                                                                      12:               Calculate 𝑞𝑖0 (𝑡) and 𝑇𝑖𝑡𝑟 (𝑡) according to
   To provide a clearer description of the price updating process, the
                                                                                      13:               Eq. (32) and Eq. (34) respectively;
pseudocode for price updating is presented in the Alg. 1.
                                                                                      14:               𝑡𝑖𝑚𝑒 ← 𝑡𝑖𝑚𝑒 + 𝑇𝑖𝑡𝑟 (𝑡);
Algorithm 1 Price Update Algorithm                                                    15:          end for
                                                                                      16:          if 𝑡𝑖𝑚𝑒 > 𝜏 then
 1: Input:𝑘𝑢𝑖 , 𝑓𝑖𝑢 , 𝑑𝑖 , 𝜏, 𝑝𝑖 , 𝝁, 𝑟𝑡𝑟   𝑒             𝑒   𝑒   𝑐   𝑐
                                       𝑖 , 𝑘 , 𝑉 , 𝜆, 𝜒, 𝑓 , 𝑐 , 𝑅 , 𝑞 , 𝑖 ∈ 𝑁;       17:               for all 𝑖 ∈ 𝑁 do
 2: Output:𝑅𝑒𝑖 (𝑡), 𝑖 ∈ 𝑁;                                                            18:                     𝑇𝑖𝑡𝑟 (𝑡) ← 𝜏 𝑇𝑖𝑡𝑟 (𝑡)∕𝑡𝑖𝑚𝑒;
 3: for all 𝑖 ∈ 𝑁 do                                                                  19:               end for
 4:    Calculate 𝜕(𝑡)∕𝜕 𝑅𝑒𝑖 (𝑡) based on Eq. (38);                                   20:          end if
 5:    Calculate 𝜕 2 (𝑡)∕𝜕 𝑅𝑒𝑖 (𝑡)2 based on Eq. (39);                               21:          for 𝑖 ∈ 𝑁 do
 6:    Calculate 𝑅𝑖𝑒_ max (𝑡) based on Eq. (40);                                      22:                 𝑐 (𝑡) ← 𝑇 𝑡𝑟 (𝑡)𝑟𝑡𝑟 ∕𝑑 ;
                                                                                                        𝑞𝑖1             𝑖     𝑖     𝑖
 7:    Calculate 𝑅𝑒_𝑖𝑡 𝑖 (𝑡) based on Eq. (47);                                       23:               𝑞 𝑐 ← 𝑞 𝑐 + 𝑞𝑖1   𝑐 (𝑡), 𝑖 ← 𝑖 + 1;
 8:    Calculate 𝑅𝑒𝑖 (𝑡) based on Eq. (48);                                           24:          end for
 9: end for                                                                           25:          Call Algorithm 1;
                                                                                      26:          𝝁𝒑𝒓𝒆 ← 𝝁, calculate 𝝁 by Eq. (45);
5.3. Multi-device task offloading and pricing mechanism algorithm                     27:          𝑖𝑡 ← 𝑖𝑡 + 1;
                                                                                      28:     end while
                                                                                                ∗ (𝑡) ← 𝑞 (𝑡), 𝑇 ∗ (𝑡) ← 𝑇 𝑡𝑟 (𝑡), 𝑅∗ (𝑡) ← 𝑅𝑒 (𝑡);
                                                                                              𝑞𝑖0
    By integrating the content in Section 4 and Section 5 of this pa-                 29:                   𝑖0        𝑖           𝑖        𝑖           𝑖
per, the optimization of 𝑞𝑖0 (𝑡), 𝑇𝑖𝑡𝑟 (𝑡), and 𝑅𝑒𝑖 (𝑡) for each time slot is         30: end for
formulated. The implementation process of the proposed Multi-Device
Task Offloading and Pricing Mechanism Algorithm is outlined in the
Alg. 2. The core of this algorithm lies in lines 8 to 22 of the pseu-
                                                                                      of [2000, 6000] bits [11]. The bandwidth of the base station 𝑤 is
docode. In pseudocode, lines 8 to 13 describe how each user calculates
                                                                                      set to 20 MHz [27]. The task computation density 𝑑𝑖 is uniformly
local processing tasks and transfer time based on price. Lines 14 to
                                                                                      distributed within [2,12]×103 cycles/bit. The time slot duration 𝜏 is
18 demonstrate the redistribution of transmission time for each user
                                                                                      configured as 0.1 s [28]. 𝜆 is set to 100. The noise power density
using time slot constraints when the total transmission time of all
                                                                                      𝑛0 is set to 10−9 W/Hz. The channel gain adheres to an exponential
users exceeds one time slot. Lines 19 to 22 illustrate how the MEC
                                                                                      distribution, denoted as E(1) [29]. The harvested energy 𝑒𝑖 (𝑡) is uni-
server updates the prices based on each user’s offloading tasks. The
                                                                                      formly distributed within [0, 20] mJ/s [30]. For the MEC server, the
time complexity of the algorithm primarily arises from the iterative
                                                                                      CPU computing frequency 𝑓 𝑒 is set to 3 GHz [27] and the capacitance
computation of task offloading and pricing for each user device. In each
                                                                                      switching coefficient 𝑘𝑒 is set to 10−28 [19]. All the simulations are
time slot, the time complexity is 𝑂(𝐷 ⋅ 𝑁), where 𝐷 is the number of
                                                                                      performed on a workstation PC with an Intel i5 12600KF processor,
iterations and 𝑁 is the number of users.
                                                                                      16 GB of memory, and Windows 10 operating system.

6. Performance evaluation                                                             6.2. Game convergence simulation experiments

    In order to assess the validity of MDTOPMA, we will organize three                    In this simulation, we primarily focus on the convergence of the
distinct sets of simulations. First of all, we demonstrate the convergence            Stackelberg game and the stability of the system. For ease of obser-
of the game through iterative experiments within a time slot. And the                 vation, there are 4 users considered in this simulation. In Fig. 2, we
stability of the queue is validated through experiments spanning mul-                 plot the differentiated prices of the MEC server, tasks offloaded by
tiple time slots. Second, under the premise of confirming the existence               users, tasks processed locally by users, and user utility as the number of
of game equilibrium, we will focus on parameter tuning experiments                    iterations accumulates. Fig. 2(a) shows that the prices of the MEC server
to evaluate the impact of different parameters on performance. Finally,               tend to stabilize after a certain number of iterations. In the figure, the
an empirical study is conducted to compare it with the benchmark                      final price obtained by each user through the game is different. This
schemes.                                                                              is attributed to the heterogeneity among users, for instance, a certain
                                                                                      user with higher energy consumption when processing tasks locally
6.1. Simulation setting                                                               tends to lean towards offloading tasks. By observing Eq. (42), it can be
                                                                                      seen that 𝜕 2 𝑅∗𝑖 (𝑡)∕𝜕 𝑞𝑖1
                                                                                                               ∗ (𝑡)2 > 0, the MEC server will increase the service

    For each user, the CPU computing frequency 𝑓𝑖𝑢 is uniformly dis-                  price of the user accordingly. In addition, since each user’s task arrival
tributed within [0.9, 1.2] GHz [15] and the capacitance switching coef-               and energy collection may be different, this will also result in varying
ficient 𝑘𝑢𝑖 is distributed within [10−28 , 10−27 ] [18,19]. The transmission          task size offloaded by each user, consequently leading to differentiated
power 𝑝𝑖 is uniformly distributed within [100,150] mW [11]. The task                  pricing by the MEC server. Fig. 2(b) and Fig. 2(c) demonstrate that
arrival quantity 𝑎𝑖 (𝑡) follows a uniform distribution within the range               the offloaded tasks and locally processed tasks for each user tend to

                                                                                  8
J. Mei et al.                                                                                           Journal of Systems Architecture 160 (2025) 103360




                Fig. 2. Price, offloaded tasks, locally processed tasks, and user utility versus iteration.




                Fig. 3. Price, offloaded tasks, locally processed tasks, and queue backlog versus time.



                                                            9
J. Mei et al.                                                                                                                   Journal of Systems Architecture 160 (2025) 103360




                                                                                                  Fig. 5. Average remote offloaded tasks with different values of 𝑎𝑖 (𝑡).




                                                                                           a state of continuous ups and downs. But overall, the task queue is
                                                                                           gradually stabilizing. This is consistent with what we expect to achieve
                                                                                           with Lyapunov optimization.
                                                                                              Observing both Fig. 2 and Fig. 3, it is evident that the game
                                                                                           converges in both short-term and long-term scenarios, ensuring the
                                                                                           long-term stability of the MEC system.

                                                                                           6.3. Parameter tuning simulation experiments

                                                                                               In this simulation, multiple experiments are conducted to demon-
                                                                                           strate the dynamic task offloading and update of prices with consid-
                                                                                           eration of 20 users. In Fig. 4, the average processed tasks and queue
                                                                                           backlog for users are plotted with varying parameter 𝑉 . Fig. 4(a)
                                                                                           depicts how the average processed tasks correlates with the parameter
Fig. 4. Average processed tasks and queue backlog vary with different values of 𝑉 .        𝑉 . The observed phenomenon shows that as 𝑉 decreases, the average
                                                                                           processing tasks present a downward trend. According to Eq. (25), as
                                                                                           the parameter 𝑉 decreases, the proportion of −𝑄𝑖 (𝑡) in the minimization
                                                                                           Eq. (25) increases, which means that the task queue backlog will show
                                                                                           an upward trend. Therefore, it can be deduced that as V decreases,
stabilize after a certain number of iterations. The trend in Fig. 2(b)                     the average processing tasks of users will show a downward trend.
exhibits a characteristic of initially increasing and then decreasing.                     This aligns with the findings exhibited in Fig. 4(a). Fig. 4(b) illustrates
This phenomenon stems from the relatively low initial prices depicted                      how the average queue backlog varies with the parameter 𝑉 . As 𝑉
in Fig. 2(a), resulting in a sharp increase in the tasks offloaded by                      increases, we can observe a decreasing trend in the queue backlog. This
users. However, as prices rise, the offloaded tasks gradually decrease,                    phenomenon aligns with the trend of increasing processed tasks shown
aligning with the anticipated trend. Additionally, Lemma 3 can be                          in Fig. 4(a).
used to substantiate this observation. If the MEC server offers a higher                       Fig. 5, the average remote offloaded tasks for users are plotted with
price to a particular user, that user typically reduces the amount of                      varying 𝑎𝑖 (𝑡). It can be observed that as the task arrival rate increases,
tasks offloaded, such as user 3. Based on the three previous figures in                    the average number of tasks offloaded by users also increases. This is
Fig. 2, it can be inferred that user utility also tends to stabilize with the              because with an increase in the task arrival rate, the backlog in the task
number of iterations, which aligns with the results shown in Fig. 2(d).                    queue also grows. To maintain queue stability, users tend to offload
According to the results in Fig. 2, it can be confirmed that there exists a                more tasks.
Stackelberg equilibrium between the users and the MEC server within                            In Fig. 6, we plot the offloaded tasks, prices, and the MEC server util-
a time slot.                                                                               ity with varying values of the parameter 𝜆. In Fig. 6(a), as 𝜆 increases,
    In Fig. 3, we present the variations of the MEC server prices, user                    the convergence price of the MEC server decreases. This is because
offloaded tasks, user local processing tasks, and the backlog in the user                  𝜆 is weighted on the offloading cost function (i.e., 𝑐𝑖 (𝑡)) of users.
task queue with varying time slots. Fig. 3(a) illustrates the gradual                      With increase of 𝜆, the proportion of user offloading costs increases,
stabilization of differential prices provided by the MEC server to every                   leading to a decreased inclination among users for task offloading.
user during long-term evolution. Due to the heterogeneity of users and                     Consequently, MEC server stimulate task offloading by reducing prices.
the varying task arrival rates and energy harvesting conditions in each                    In Fig. 6(b), as the value of 𝜆 decreases, users tend to offload more
time slot, the prices calculated through the game may also differ for                      tasks. This is because a lower 𝜆 results in a smaller proportion of user
each user. Fig. 3(b) depicts the increasing trend of remote offloaded                      offloading costs, leading to an increased desire for task offloading.
tasks for each user starting from zero and gradually decreasing there-                     Fig. 6(c) illustrates that the effectiveness of the MEC server increases
after. This pattern can be attributed to the initially low prices shown in                 with lower 𝜆. This experimental result can be calculated based on the
Fig. 3(a), which encourage users to offload more tasks. However, as the                    experimental data in Fig. 6(a) and Fig. 6(b).
prices subsequently rise, users’ inclination for offloading diminishes,
leading to a reduction in the amount of offloaded tasks. Fig. 3(c) and                     6.4. Comparison with benchmark schemes
Fig. 3(d) illustrate the steady state of user local processing tasks and
user task queue during long-term evolution, respectively. The arrival                         To further evaluate the MDTOPMA′ s performance, there are three
of tasks in each time slot is uncertain, so the queue in Fig. 3(d) shows                   schemes compared with MDTOPMA:

                                                                                      10
J. Mei et al.                                                                                                                 Journal of Systems Architecture 160 (2025) 103360




                                          Fig. 6. Price, offloaded tasks, and the MEC server utility with different values of 𝜆.




                                                Fig. 7. User utility and task queue backlog with four different schemes.


     • Local-Only Processing (LOP) scheme [19]: In each time slot,                      may still choose to remotely offload many tasks, resulting in reduced
       users solely process all tasks locally, i.e., 𝑇𝑖∗ (𝑡) = 0.                       utilities.
     • Average Time-Constrained Task Offloading (ATCTO) scheme:
       In each time slot, the channel transmission time for each user’s of-             7. Conclusion
       floaded tasks must not surpass a predefined threshold, i.e., 𝑇𝑖𝑡𝑟 (𝑡) ≤
       𝜏∕𝑛. The ATCTO scheme is inspired by the Equal Allocation Strat-                     In this paper, we study a task offloading problem for a MEC system
       egy from [11], which emphasizes fairly distributing the offloading               consisting of three layers of cloud–edge-device, where each user ter-
       time among users.                                                                minal device supports EH. While solving the optimization problem, the
     • Genetic Algorithm (GA) scheme: In this study, we use a genetic                   Lyapunov optimization theory is applied to convert the long-term prob-
       algorithm for comparison and initialize a population of 20 indi-                 lem into problem of each time slot and stabilize the task queue of each
       viduals, each representing offloading strategies for users and a                 user. In order to reasonably allocate resources between the users and
       pricing strategy for the MEC server. After multiple iterations, we               the MEC server, the Stackelberg game theory is employed to regulate
       select the individual with the highest fitness for comparison.                   resources. Combining the above two theories, we apply the MDTOPMA
                                                                                        to solve the optimal offloading strategy of each user and the pricing
    In Fig. 7, the average user utility and average queue backlog are                   strategy of the MEC server in each time slot. The simulation experiment
plotted alongside different algorithm schemes. In the experimental                      results indicate that, when compared to other algorithms, MDTOPMA
results, our MDTOPMA surpasses the other schemes in user utility                        not only enhances user benefits but also reduces the backlog of user
and task queue backlog. In Fig. 7(a), it can be observed that the                       task queues. In the simulation experiment of parameter performance
improvement in user utility by LOP is significantly smaller than other                  optimization, the adjustment of parameter value also leads to better
schemes. This is because LOP only allows users to process tasks locally,                choice of offloading decision and pricing decision.
neglecting the computing capacity of the MEC server, and the energy
consumption of remote offloading is generally lower than the energy                     CRediT authorship contribution statement
consumption generated by local computing tasks. Fig. 7(b) demon-
strates that the task queue backlog of LOP is significantly higher than                    Jing Mei: Writing – review & editing, Investigation, Funding ac-
that of our proposed algorithm. This is mainly due to the fact that LOP                 quisition, Formal analysis. Cuibin Zeng: Writing – original draft, Re-
does not utilize the MEC server’s computational resources. In Fig. 7,                   sources, Methodology. Zhao Tong: Writing – review & editing, Method-
the performance of ATCTO is less than our scheme. For instance, if a                    ology, Funding acquisition, Conceptualization. Longbao Dai: Writing –
user calculates the optimal transmission time T, such that 𝑇𝑖∗ (𝑡) > 𝜏∕𝑛,               review & editing, Investigation, Conceptualization. Keqin Li: Writing –
but is constrained by 𝑇𝑖𝑡𝑟 (𝑡) ≤ 𝜏∕𝑛, it results in the user being unable to            review & editing, Conceptualization.
achieve the optimal performance. Although ATCTO allows each user to
have a transmission time ranging from 0 to 𝜏∕𝑛, ensuring that each user                 Declaration of competing interest
has a fair opportunity to utilize network resources, achieving better
performance is challenging. In Fig. 7, GA’s task queue performance is                       The authors declare that they have no known competing finan-
better than ATCTO, but the user benefit is lower than ATCTO. This                       cial interests or personal relationships that could have appeared to
is due to the randomness of GA. Even when the price is high, users                      influence the work reported in this paper.

                                                                                   11
J. Mei et al.                                                                                                                       Journal of Systems Architecture 160 (2025) 103360


Acknowledgments                                                                                [19] M. Guo, W. Wang, X. Huang, Y. Chen, L. Zhang, L. Chen, Lyapunov-based partial
                                                                                                    computation offloading for multiple mobile devices enabled by harvested energy
                                                                                                    in mec, IEEE Int. Things J. 9 (11) (2022) 9025–9035, http://dx.doi.org/10.1109/
   The authors would like to thank the anonymous reviewers for
                                                                                                    JIOT.2021.3118016.
their valuable comments and suggestions. This work was supported by                            [20] C. Qiu, Y. Hu, Y. Chen, Lyapunov optimized cooperative communications with
the Program of National Natural Science Foundation of China (grant                                  stochastic energy harvesting relay, IEEE Int. Things J. 5 (2) (2018) 1323–1333,
No. 62072174, 61502165), Provincial Natural Science Foundation of                                   http://dx.doi.org/10.1109/JIOT.2018.2793850.
                                                                                               [21] B. Cao, S. Xia, J. Han, Y. Li, A distributed game methodology for crowdsensing
Hunan, China (grant No. 2020JJ5370, 2022JJ40278, 2023JJ30083),
                                                                                                    in uncertain wireless scenario, IEEE Trans. Mob. Comput. 19 (1) (2020) 15–28,
Scientific Research Fund of Hunan Provincial Education Department,                                  http://dx.doi.org/10.1109/TMC.2019.2892953.
China (grant No. 22A0026, 22A0592), Graduate Research Innovation-                              [22] M. Tao, K. Ota, M. Dong, H. Yuan, Stackelberg game-based pricing and offloading
Program of Hunan Province (Grant No. CX20240548).                                                   in mobile edge computing, IEEE Wirel. Commun. Lett. 11 (5) (2022) 883–887,
                                                                                                    http://dx.doi.org/10.1109/LWC.2021.3138938.
                                                                                               [23] Y. Liu, C. Xu, Y. Zhan, Z. Liu, J. Guan, H. Zhang, Incentive mechanism for
Data availability
                                                                                                    computation offloading using edge computing: A stackelberg game approach,
                                                                                                    Comput. Netw. 129 (DEC.24) (2017) 399–409.
    Data will be made available on request.                                                    [24] Y. Li, S. Xia, M. Zheng, B. Cao, Q. Liu, Lyapunov optimization-based trade-
                                                                                                    off policy for mobile cloud offloading in heterogeneous wireless networks, IEEE
                                                                                                    Trans. Cloud Comput. 10 (1) (2022) 491–505, http://dx.doi.org/10.1109/TCC.
References                                                                                          2019.2938504.
                                                                                               [25] X. Chen, Z. Lu, W. Ni, X. Wang, F. Wang, S. Zhang, S. Xu, Cooling-aware
                                                                                                    optimization of edge server configuration and edge computation offloading for
 [1] Ana Bera, 80 IoT statistics (infographic), 2019, https://safeatlast.co/blog/iot-
                                                                                                    wirelessly powered devices, IEEE Trans. Veh. Technol. 70 (5) (2021) 5043–5056,
     statistics/, Website.
                                                                                                    http://dx.doi.org/10.1109/TVT.2021.3076057.
 [2] K.P. Naveen, R. Sundaresan, Double-auction mechanisms for resource trading
                                                                                               [26] M.S.S. Rao, S.A. Soman, Marginal pricing of transmission services using min–
     markets, IEEE/ACM Trans. Netw. 29 (3) (2021) 1210–1223, http://dx.doi.org/
                                                                                                    max fairness policy, IEEE Trans. Power Syst. 30 (2) (2015) 573–584, http:
     10.1109/TNET.2021.3058251.
                                                                                                    //dx.doi.org/10.1109/TPWRS.2014.2331424.
 [3] D. Ma, G. Lan, M. Hassan, W. Hu, S.K. Das, Sensing, computing, and commu-
                                                                                               [27] L. Chen, S. Zhou, J. Xu, Computation peer offloading for energy-constrained
     nications for energy harvesting iots: A survey, IEEE Commun. Surv. & Tutorials
                                                                                                    mobile edge computing in small-cell networks, IEEE/ACM Trans. Netw. 26 (4)
     22 (2) (2020) 1222–1250, http://dx.doi.org/10.1109/COMST.2019.2962526.
                                                                                                    (2018) 1619–1632, http://dx.doi.org/10.1109/TNET.2018.2841758.
 [4] J. Peng, H. Qiu, J. Cai, W. Xu, J. Wang, D2d-assisted multi-user cooperative
                                                                                               [28] Y. Dai, K. Zhang, S. Maharjan, Y. Zhang, Deep reinforcement learning for
     partial offloading, transmission scheduling and computation allocating for mec,
                                                                                                    stochastic computation offloading in digital twin networks, IEEE Trans. Ind.
     IEEE Trans. Wirel. Commun. 20 (8) (2021) 4858–4873, http://dx.doi.org/10.
                                                                                                    Inform. 17 (7) (2021) 4968–4977, http://dx.doi.org/10.1109/TII.2020.3016320.
     1109/TWC.2021.3062616.                                                                    [29] J. Mei, L. Dai, Z. Tong, X. Deng, K. Li, Throughput-aware dynamic task offloading
 [5] S.-H. Kim, S. Park, M. Chen, C.-H. Youn, An optimal pricing scheme for                         under resource constant for mec with energy harvesting devices, IEEE Trans.
     the energy-efficient mobile edge computation offloading with ofdma, IEEE                       Netw. Serv. Manag. (2023) 1, http://dx.doi.org/10.1109/TNSM.2023.3243629.
     Commun. Lett. 22 (9) (2018) 1922–1925, http://dx.doi.org/10.1109/LCOMM.                   [30] X. Lyu, W. Ni, H. Tian, R.P. Liu, X. Wang, G.B. Giannakis, A. Paulraj, Op-
     2018.2849401.                                                                                  timal schedule of mobile edge computing for internet of things using partial
 [6] M. Li, Q. Wu, J. Zhu, R. Zheng, M. Zhang, A computing offloading game for                      information, IEEE J. Sel. Areas Commun. 35 (11) (2017) 2606–2615, http:
     mobile devices and edge cloud servers, Wirel. Commun. Mob. Comput. 2018                        //dx.doi.org/10.1109/JSAC.2017.2760186.
     (2018) 1–10.
 [7] Z. Liu, J. Fu, Y. Zhang, Computation offloading and pricing in mobile edge
     computing based on stackelberg game, Wirel. Netw. 27 (7) (2021) 4795–4806,                                          Jing Mei received the Ph.D degree in computer science
     http://dx.doi.org/10.1007/s11276-021-02767-z.                                                                       from Hunan University, China, in 2015. She is currently
 [8] Z. Ning, J. Huang, X. Wang, J.J.P.C. Rodrigues, L. Guo, Mobile edge computing-                                      an associate professor in the College of Information Sci-
     enabled internet of vehicles: Toward energy-efficient scheduling, IEEE Netw. 33                                     ence and Engineering in Hunan Normal University. Her
     (5) (2019) 198–205, http://dx.doi.org/10.1109/MNET.2019.1800309.                                                    research interests include cloud computing, fog computing
 [9] Y. Mao, J. Zhang, K.B. Letaief, Joint task offloading scheduling and transmit                                       and mobile edge computing, high performance computing,
     power allocation for mobile-edge computing systems, in: 2017 IEEE Wireless                                          task scheduling and resource management, etc. She has
     Communications and Networking Conference, WCNC, 2017, pp. 1–6, http://dx.                                           published more than 30 research articles in international
     doi.org/10.1109/WCNC.2017.7925615.                                                                                  conference and journals, such as IEEE Transactions on Com-
[10] X. Zhao, L. Zhao, K. Liang, An energy consumption oriented offloading algorithm                                     puters, IEEE Transactions on Parallel and Distributed System,
     for fog computing, in: Quality, Reliability, Security and Robustness in Heteroge-                                   IEEE Transactions on Service Computing, Cluster Computing,
     neous Networks: 12th International Conference, QShine 2016, Seoul, Korea, July                                      Journal of Grid Computing, Journal of Supercomputing.
     7–8, 2016, Proceedings 12, Springer, 2017, pp. 293–301.
[11] Y. Chen, N. Zhang, Y. Zhang, X. Chen, W. Wu, X. Shen, Energy efficient
                                                                                                                         Cuibin Zeng received the B.S. degree in computer science
     dynamic offloading in mobile edge computing for internet of things, IEEE Trans.
                                                                                                                         and technology from Jishou University, Jishou, China, in
     Cloud Comput. 9 (3) (2021) 1050–1060, http://dx.doi.org/10.1109/TCC.2019.
                                                                                                                         2022. He is currently pursuing the M.S. degree at the
     2898657.
[12] F. Li, H. Yao, J. Du, C. Jiang, Y. Qian, Stackelberg game-based computation                                         College of Information Science and Engineering, Hunan
     offloading in social and cognitive industrial internet of things, IEEE Trans. Ind.                                  Normal University, Changsha, China. His research focuses
     Inform. 16 (8) (2020) 5444–5455, http://dx.doi.org/10.1109/TII.2019.2961662.                                        on resource scheduling and price allocation in mobile edge
[13] X. Hu, K.-K. Wong, K. Yang, Wireless powered cooperation-assisted mobile                                            computing.
     edge computing, IEEE Trans. Wirel. Commun. 17 (4) (2018) 2375–2388, http:
     //dx.doi.org/10.1109/TWC.2018.2794345.
[14] F. Wang, J. Xu, X. Wang, S. Cui, Joint offloading and computing optimization in
     wireless powered mobile-edge computing systems, IEEE Trans. Wirel. Commun.
     17 (3) (2018) 1784–1797, http://dx.doi.org/10.1109/TWC.2017.2785305.                                                Zhao Tong received the Ph.D degree in computer science
[15] P.K. Bishoyi, S. Misra, Enabling green mobile-edge computing for 5g-based                                           from Hunan University, Changsha, China in 2014. He was a
     healthcare applications, IEEE Trans. Green Commun. Netw. 5 (3) (2021)                                               visiting scholar at the Georgia State University from 2017 to
     1623–1631, http://dx.doi.org/10.1109/TGCN.2021.3075903.                                                             2018. He is currently an associate professor at the College
[16] F. Zeng, Q. Chen, L. Meng, J. Wu, Volunteer assisted collaborative offloading                                       of Information Science and Engineering of Hunan Normal
     and resource allocation in vehicular edge computing, IEEE Trans. Intell. Transp.                                    University, the young backbone teacher of Hunan Province,
     Syst. 22 (6) (2021) 3247–3257, http://dx.doi.org/10.1109/TITS.2020.2980422.                                         China. His research interests include parallel and distributed
[17] S. Xia, Z. Yao, Y. Li, S. Mao, Online distributed offloading and computing                                          computing systems, resource management, big data and
     resource management with energy harvesting for heterogeneous mec-enabled iot,                                       machine learning algorithm. He has published more than 25
     IEEE Trans. Wirel. Commun. 20 (10) (2021) 6743–6757, http://dx.doi.org/10.                                          research papers in international conferences and journals,
     1109/TWC.2021.3076201.                                                                                              such as IEEE-TPDS, Information Sciences, FGCS, NCA, and
[18] Q. Zhang, L. Gui, F. Hou, J. Chen, S. Zhu, F. Tian, Dynamic task offloading                                         JPDC, PDCAT, etc. He is a senior member of the China
     and resource allocation for mobile-edge computing in dense cloud ran, IEEE                                          Computer Federation (CCF) and a Member of the IEEE.
     Int. Things J. 7 (4) (2020) 3282–3299, http://dx.doi.org/10.1109/JIOT.2020.
     2967502.


                                                                                          12
J. Mei et al.                                                                                  Journal of Systems Architecture 160 (2025) 103360


                Longbao Dai received the B.S. degree in computer science            machine learning, intelligent and soft computing. He has
                and technology from Hunan University of Science and                 authored or co-authored over 850 journal articles, book
                Engineering, Yongzhou, China, in 2021. He is currently              chapters, and refereed conference papers, and has received
                working toward the M.S. degree at the College of Infor-             several best paper awards. He holds over 70 patents an-
                mation Science and Engineering, Hunan Normal University,            nounced or authorized by the Chinese National Intellectual
                Changsha, China. His research interests focus on distributed        Property Administration. He is among the world’s top 5
                parallel computing, modeling and resource pricing and allo-         most influential scientists in parallel and distributed com-
                cation in mobile edge computing systems, and combinatorial          puting in terms of both single–year impact and career–long
                optimization.                                                       impact based on a composite indicator of Scopus citation
                                                                                    database. He has chaired many international conferences.
                                                                                    He is currently an associate editor of the ACM Comput-
                                                                                    ing Surveys and the CCF Transactions on High Performance
                Keqin Li is a SUNY Distinguished Professor of Computer
                                                                                    Computing. He has served on the editorial boards of the
                Science with the State University of New York. He is
                                                                                    IEEE Transactions on Parallel and Distributed Systems, the IEEE
                also a National Distinguished Professor with Hunan Uni-
                                                                                    Transactions on Computers, the IEEE Transactions on Cloud
                versity, China. His current research interests include cloud
                                                                                    Computing, the IEEE Transactions on Services Computing, and
                computing, fog computing and mobile edge computing,
                                                                                    the IEEE Transactions on Sustainable Computing. He is an
                energy–efficient computing and communication, embed-
                                                                                    IEEE Fellow and an AAIA Fellow. He is also a Member
                ded systems and cyber–physical systems, heterogeneous
                                                                                    of Academia Europaea (Academician of the Academy of
                computing systems, big data computing, high–performance
                                                                                    Europe).
                computing, CPU–GPU hybrid and cooperative computing,
                computer architectures and systems, computer networking,




                                                                               13
