Unloadable Computing Problem Based on Edge Distributed

Computing

Xiaoliang He

1,† a

, Jiazhen Xu

2,‡,* b

and Zhilin Yan

3,** c

College of Mathematics and Applied Mathematics, Guangdong University of Foreign and Studies, Guangzhou, China

College of Civil Engineering, Inner Mongolia University of Technology, Huhehaote, China

College of Data Science, City University of Macau, Macau, China

† ‡

Keywords: Edge Distributed Computing, Computation Offloading, Task Offloading, Resource Allocation.

Abstract: Due to limitations in power, computing power, and storage, mobile devices struggle to meet the demands of

rapidly developing applications. Offloading computing has emerged as a promising solution, becoming a

research focus. Edge distributed systems address this by offloading tasks from mobile devices to nearby edge

servers or networked devices. This paper focuses on edge distributed computing offloading and conducts the

following research: (1) An overview of common MEC mechanisms is provided, covering key technologies

like virtualization, SDN, CDN, SON, cloud computing, and collaborative computing. The characteristics of

edge distributed systems, such as scalability, location relevance, diversity, randomness, time-varying nature,

and autonomy, are discussed. Typical MEC implementation methods and future trends are explored. (2) A

shared bandwidth allocation algorithm using blockchain smart contracts is proposed to address wireless

resource scheduling challenges. This scheme eliminates the need for a central server, reducing latency and

enhancing scalability. (3) An optimized mobile edge collaborative learning algorithm is proposed, which

reduces data traffic and communication delays by shifting machine learning tasks to local devices. A compact

parameter set replaces redundant models to minimize excessive transmission.

1 INTRODUCTION

The rise of the Internet of Things (IoT) has led to a

surge in smart devices (Babatunde, Bodhaswar, 2016),

accompanied by an increase in compute-intensive and

latency-sensitive tasks. However, mobile devices

face significant challenges in processing these tasks

due to limitations in power consumption, computing

capability, and storage capacity. Against this

backdrop, edge computing, as a new computing

paradigm, has emerged. By deploying computing

resources at the network edge, it reduces task

processing latency, lowers energy consumption, and

enhances user experience. Computation offloading,

as one of the key technologies of edge distributed

computing, has become a current research hotspot.

Blockchain technology, with its characteristics of

decentralization, immutability, and transparency,

https://orcid.org/0009-0009-5025-2576

https://orcid.org/0009-0006-3452-0110

https://orcid.org/0009-0002-7818-9176

Corresponding author

holds great potential in the field of edge distributed

computing. It ensures task security and reliability

through a distributed ledger and achieves automated

management of computing tasks via smart contracts

(Baranwal, Kumar, Vidyarthi, 2023). Although

blockchain technology has been applied in task

scheduling optimization, enhancement of node

trustworthiness, and improvement of computing

result reliability, its challenges in performance,

scalability, and energy consumption limit its

widespread application (Chen, Wang, Liu, 2021).

This paper delves into the key technologies of

blockchain-based edge distributed computing

offloading, including trust foundation, elastic

offloading, fair scheduling, and security assurance.

The goal is to improve computational efficiency,

ensure security, and optimize user experience. The

research outcomes include a multi-chain

638

He, X., Xu, J. and Yan, Z.

Unloadable Computing Problem Based on Edge Distributed Computing.

DOI: 10.5220/0013703200004670

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Conference on Data Science and Engineering (ICDSE 2025), pages 638-644

ISBN: 978-989-758-765-8

collaborative trustworthy computing offloading

model that achieves data traceability, a collaborative

computing offloading method that constructs

standardized interaction processes, and a smart

contract-driven adaptive offloading strategy. By

transforming the task allocation problem into a

stochastic optimization problem, the strategy uses

stochastic optimization algorithms to reduce overall

costs and increase task completion rates. The

deployment of smart contracts with

reward/punishment mechanisms incentivizes

computing nodes to participate, ensuring successful

task completion and enhancing system robustness and

scheduling efficiency.

In related research, Li Yun et al. proposed

intelligent collaborative computing offloading and

resource allocation algorithms (Li, Zhang, Yao, 2024),

constructing a joint optimization problem to

minimize system energy consumption while meeting

user latency constraints. The Gat-HMARL algorithm

was employed, and simulations showed performance

improvements. In the joint optimization of intelligent

computing offloading and service caching (Li, Nan,

Yao, et al, 2025), a DAG task offloading and resource

optimization problem was established, and the

MADDPG algorithm was used to explore optimal

strategies. Experiments demonstrated that the

algorithm performed excellently in reducing energy

consumption and improving cache hit rates. Lin

Hongcai proposed the O2O-DRL and DR-DRL

methods based on DRL (Lin, 2024), which addressed

the cold start problem in reinforcement learning and

improved task success rates. In another study with the

same name, a solution for edge distributed computing

offloading was proposed in response to the limitations

of mobile devices. It introduced key technologies and

proposed a blockchain-based shared bandwidth

allocation algorithm driven by smart contracts to

reduce latency and enhance system scalability. An

optimized mobile edge collaborative learning

algorithm was also proposed.

This paper reviews the key technologies of edge

distributed computing offloading, including

computation offloading algorithms, resource

allocation strategies, the application of blockchain

technology in resource scheduling, and the

optimization of mobile edge collaborative learning

algorithms. It aims to summarize the current research

progress and challenges and to propose future

research directions, with the expectation of further

advancing the development of edge computing

technology.

2 DATASETS

The data in this paper presents the architecture of the

edge - distributed computing network, covering the

connections and data flow paths among components

such as edge devices, edge servers, and cloud

computing centers. It also provides symbols and

parameters related to computational offloading, such

as task volume and computing power, laying the

foundation for mathematical modeling. The

experimental parameters are diverse, involving

different computational tasks, network conditions,

and device performance scenarios, which effectively

verify the effectiveness and performance of the

computational offloading method and

comprehensively support the research work.

This paper trains Informer and double LSTM on

the ETTh1, ETTm1, and ECL datasets. Informer is a

prediction model adopting an encoder - decoder

architecture and has 11,330,055 parameters; while the

double - layer long short - term memory network

(LSTM) contains 4,233,735 parameters. ETTh1 and

ETTm1 represent two - year electricity data from two

cities in China. ETTh1 is applicable to hourly

readings, and ETTm1 is used for readings at 15 -

minute intervals. ECL shows the hourly electricity

consumption data of 321 customers in the United

States over two years.

3 EXPERIMENTAL METHODS

3.1 Distributed Reinforcement

Learning

In this paper, the task offloading method based on

distributed reinforcement learning demonstrates

unique research value and practical significance. It is

reflected in the following aspects:

1.Data Quality Assessment (DQ) Phase: Since

edge computing involves multi-source heterogeneous

data, these data may have low-quality issues such as

missing values and erroneous values due to device

differences and network fluctuations. By assessing

from multiple dimensions such as accuracy,

completeness, consistency, and timeliness, a

comprehensive understanding of the data status can

be achieved. For example, in terms of accuracy, data

verification algorithms can be used to compare with

known standards or historical data; completeness

assessment can check whether there is a lack of key

information in the data; consistency is achieved by

cross-checking the same type of data from different

data sources; and timeliness is judged based on the

data generation time and the time requirements of the

Unloadable Computing Problem Based on Edge Distributed Computing

639

application scenario. This step lays the foundation for

subsequent processing and ensures the reliability of

the data.

2.Data Repair (DR) Phase: The repair algorithm

based on a new repair consensus mechanism plays a

role, adopting corresponding technologies for

different types of low-quality data. For example, for

missing values, if the data has time series

characteristics, linear interpolation or spline

interpolation methods can be used; erroneous data can

be corrected according to the data distribution law

and business logic by setting thresholds or rules;

incomplete data blocks can be rebuilt using data

redundancy or backup. These means effectively

improve data quality and reduce its negative impact

on distributed computing.

3.Task Scheduling Distributed Reinforcement

Learning (DELTA) Step: Here, the distributed

reinforcement learning algorithm is used, regarding

edge devices as agents. Each agent makes decisions

based on its own computing resources, network

connection status, and the data quality assessed and

repaired, as well as the information obtained from the

surrounding environment. Agents continuously

interact with the environment during task offloading,

taking actions, such as executing tasks, offloading to

other devices or servers, and learn and optimize

strategies based on reward feedback such as the

timeliness of task completion and the degree of

energy consumption reduction, thereby achieving

efficient task offloading based on a new low-quality

data distribution strategy and ensuring data privacy

with the characteristics of blockchain.

From the perspective of theoretical basis,

distributed reinforcement learning fits the distributed

architecture characteristics of edge computing. Edge

devices have dispersed and limited resources, and

traditional centralized algorithms are powerless in

this environment. Distributed reinforcement learning

endows each edge device with the ability to learn and

make decisions independently, and continuously

optimizes system performance through interaction

and collaboration between devices. The data quality

assessment and repair phase directly hit the core of

the multi-source heterogeneous data problem,

effectively avoiding the disadvantages of computing

convergence delay and result deviation caused by

low-quality data, and laying a solid foundation for

accurate task offloading. The integration of

blockchain technology is the icing on the cake. Its

decentralized and tamper-proof attributes consolidate

the data security line of defense, effectively resist

malicious intrusion and data tampering, and

effectively ensure data privacy and the cornerstone of

system trust.

In the experimental design dimension, a simulated

edge computing scenario covering multiple edge

devices and servers is constructed, and variables such

as task types, data quality levels, and network

conditions are set multidimensionally. Key indicators

such as task completion time, system energy

consumption, data quality improvement, and privacy

protection effectiveness are comprehensively

considered to accurately judge the advantages and

disadvantages of this method compared with other

strategies. The experimental results show its

significant effectiveness. In terms of task completion

time, the offloading strategy that intelligently adapts

to data and network conditions can flexibly allocate

tasks to suitable devices or servers, greatly reducing

processing time. System energy consumption is

effectively controlled due to the reasonable task

allocation mechanism, avoiding device overload, and

distributed reinforcement learning continuously

optimizes strategies to achieve energy consumption

minimization. Data quality is significantly improved

after assessment and repair, effectively ensuring the

accuracy of application d computational resources,

and performance may be limited in complex edge

scenarios and large-scale task offloading. The

accuracy and efficiency of data quality assessment

and repair algorithms also need to be improved.

Compared with other related research, such as the

D2HM algorithm in "Distributed Heterogeneous

Task Offloading Algorithm in Mobile Edge

Computing," which uses distributed game theory and

Lyapunov optimization to achieve differentiated

control of heterogeneous tasks and elastic resource

allocation, reducing average delay; and the JORA

method in "JORA: Blockchain-based efficient joint

computing offloading and resource allocation for

edge video streaming systems," which uses

blockchain smart contract incentive mechanisms to

solve joint offloading, allocation, and video

compression optimization problems, achieving

efficient resource utilization and energy consumption

- accuracy balance. These achievements have opened

up new ideas for the research of blockchain-based

edge distributed computing offloading, and the task

offloading method based on distributed

reinforcement learning is expected to continuously

iterate and optimize in the process of integrating its

own advantages and drawing on the strengths of

others, deeply promoting the development of edge

computing offloading technology, and improving the

system's comprehensive performance and application

efficiency.

ICDSE 2025 - The International Conference on Data Science and Engineering

640

3.2 Test

3.2.1 Identity Information Encryption Test

In this study, in order to deeply evaluate the

performance of algorithms related to identity

information encryption, plaintext identity

information data in the range of ((0, 10]MB) was

specifically generated as the test dataset. The AES

algorithm, ECDH key agreement encryption and

decryption scheme, and ECIES scheme were

implemented using the BouncyCastle library in C#

for encryption and decryption performance testing. At

the same time, development packages such as

BitcoinSVCryptor and BsvSimpleLibrary were used

to better interact with the Bitcoin network.

3.2.2 Zero-Knowledge Identity

Authentication Test

Programming was carried out on two parts of the

crowdsourcing logistics platform: generating zero -

knowledge questions and verifying zero - knowledge

proofs. In the test, 1 to 250 users were simulated to

simultaneously initiate zero - knowledge

authentication challenges to the crowdsourcing

logistics platform.

3.3 Experimental design

In the algorithm comparison experiment of edge

distributed computing, in order to simulate the large-

scale multi-access edge computing (MEC) scenario,

the experimental environment is set to contain

multiple mobile devices (MD) and multiple edge

servers (ES). The amount of MD and ES can be

adjusted according to the specific needs of the

experiment. At the beginning of each time slot, each

MD randomly generates tasks of different sizes

according to the Bernoulli distribution, and the size of

the tasks is expressed in data volume (unit: Mbit). In

the experimental environment of this paper, several

algorithms are compared to evaluate their

performance in dealing with MEC task unloading.

MADDPG algorithm is a multi-agent depth

deterministic strategy gradient algorithm, which is a

classical algorithm to solve MEC task unloading

problem. The algorithm can deal with the cooperation

and competition between multiple agents, but the

computational complexity is high in large-scale

scenarios. In the experiment, MADDPG algorithm is

used as one of the comparison algorithms to evaluate

the performance of MF-MATO algorithm. DDPG

algorithm is a distributed single-agent algorithm,

which can not deal with the interaction between

agents and partial observability. Due to the partial

observability of MEC environments, the performance

of DDPG algorithms may be limited. In the

experiment, DDPG algorithm is used as one of the

comparison algorithms to highlight the advantages of

MF-MATO algorithm in dealing with partially

observable problems. LSTM-DQN algorithm

introduces LSTM network into DQN to solve part of

the observable problem, but it is still a single-agent

algorithm. The algorithm can deal with the partially

observable problem of a single agent, but it cannot

deal with the cooperative relationship between

multiple agents. In the experiment, LSTM-DQN

algorithm is used as one of the comparison algorithms

to evaluate the performance of MF-MATO algorithm

in collaborative task unloading. The Random

algorithm is a simple uninstall strategy in which MD

randomly generates uninstall decisions. The

algorithm does not rely on any optimization algorithm

or model, so it can be used as a benchmark

comparison algorithm to evaluate the performance

improvement of other algorithms. Then the

MADDPG, DDPG, LSTM-DQN and Random

experimental results were compared and analysed.

The performance of different algorithms is evaluated

by comparing the average delay and average task

dropout rate.

4 ANALYSIS OF

EXPERIMENTAL RESULTS

4.1 Distributed Reinforcement

Learning Analysis

In terms of task completion time, the experimental

results show that the method performs excellently.

Due to its ability to intelligently select task offloading

strategies based on data quality and network

conditions, for example, when a certain edge device

has limited resources or data processing capabilities,

it can quickly transfer tasks to other devices or servers

with sufficient resources. This allows tasks to be

processed in a more suitable environment, effectively

avoiding processing delays caused by device

performance bottlenecks or data issues, thereby

significantly reducing task completion time and

improving the system's overall response speed, which

is crucial for edge computing application scenarios

with high real-time requirements.

From the perspective of system energy

consumption, the method has also achieved good

Unloadable Computing Problem Based on Edge Distributed Computing

641

results. Through a reasonable task allocation

mechanism, it avoids the situation where some

devices are overloaded and consume too much energy.

The distributed reinforcement learning algorithm

continuously learns and optimizes task offloading

strategies, enabling tasks to be executed on devices or

servers with lower energy consumption, achieving the

minimization of system energy consumption. In edge

computing environments, devices usually rely on

limited energy supplies (such as battery-powered

mobile edge devices), and reducing energy

consumption helps to extend the device's endurance

time and improve the system's sustainable operation

capability.

Regarding the degree of data quality improvement,

after the data quality assessment and repair steps, the

reliability and availability of the data are significantly

enhanced. Experimental results show that low-quality

data repaired can produce more accurate results in

subsequent task offloading and processing processes.

In application fields with strict data quality

requirements, such as the Industrial Internet of Things

and the Energy Internet, accurate data processing

results are the key to ensuring the normal operation of

the system and the scientific nature of decision-

making. For example, in the monitoring and control

of industrial production processes, high-quality data

can ensure the precise operation of equipment and

timely warning of failures, avoiding production

accidents or efficiency losses caused by data errors.

In terms of privacy protection effects, the

introduction of blockchain technology provides a

solid guarantee for data privacy. Experimental results

have confirmed that the task offloading framework

based on blockchain effectively prevents data leakage

and tampering. In edge computing environments, data

may be transmitted and processed between multiple

devices and servers, facing many security risks. The

decentralized and tamper-proof characteristics of

blockchain ensure the security and integrity of data

throughout the computing offloading process,

protecting user privacy and enhancing user trust in

edge computing systems.

However, the experimental results also reveal

some limitations of the method. For example, the

training and convergence of distributed

reinforcement learning algorithms may require a

longer time and more computational resources. In

edge computing environments, the computing power

of edge devices is relatively weak, which may affect

the deployment and performance of the method in

practical applications. Especially in complex edge

computing environments and large-scale task

offloading scenarios, with a large number of devices

and complex task types, the difficulty of algorithm

training and convergence may further increase,

thereby limiting the application scope and

effectiveness of the method. In addition, the accuracy

and efficiency of data quality assessment and repair

algorithms still need to be further improved to better

cope with the complex and changing data quality

issues in edge computing.

4.2 Distributed Reinforcement

Learning Analysis

The ECDH key agreement scheme takes longer than

the AES algorithm. This is because it not only needs

to derive the shared key but also has to use the shared

key to derive the symmetric key for encryption and

decryption operations. The encryption and decryption

time of the ECDH key agreement encryption and

decryption scheme is shorter than that of the ECIES

scheme. This is because the identity information in

this paper is stored in Bitcoin SV (BSV) blockchain

transactions. During the process of uploading to the

blockchain, the trader signs the transaction with

OP_RETURN, which eliminates the computational

overhead of the MAC verification algorithm in the

ECIES scheme. Although the encryption and

decryption efficiency of the ECDH key agreement

scheme is lower compared to the AES algorithm, the

blockchain - based identity information encryption

and decryption scheme implemented in this study still

has a relatively fast encryption and decryption speed

and strong practicality.

As the number of users increases, when the

number of users is the same, the authentication delay

of the scheme proposed in this paper is shorter than

that of the SHAO scheme, and the authentication

speed is faster. Under the condition of an equal

number of users, the transaction - processing ability

of the scheme proposed in this paper is better than that

of the SHAO scheme. The scheme proposed in this

paper has significantly improved the authentication

delay and authentication efficiency of the

crowdsourcing logistics platform, effectively

enhancing the authentication performance of the

system.

4.3 Algorithm analysis

In Figure 1, the MF-MATO algorithm shows good

performance stability and optimization ability in

dealing with multi-agent cooperation problem and

partially observable problem. MADDPG algorithm

also shows some advantages in dealing with multi-

agent cooperation problem, but its performance

ICDSE 2025 - The International Conference on Data Science and Engineering

642

fluctuates greatly. As a single-agent algorithm,

DDPG algorithm may not be as effective as

MADDPG and MF-MATO in dealing with multi-

agent cooperation problems. LSTM-DQN algorithm

has some limitations in dealing with partially

observable problems, but its performance is relatively

stable (Lu, Li, 2024).

In Figure 2, two algorithms, MF-MATO and

LSTM-DQN, are selected for comparison. MF-

MATO may be an algorithm for multi-agent or edge

computing, while LSTM-DQN introduces LSTM

(Long short-term memory) networks in DQN (deep Q

network) to handle partially observable problems.

MF-MATO algorithm may have better performance

stability and optimization ability when dealing with

multi-agent or edge computing tasks. LSTM-DQN

algorithm is sensitive to the state transition mode

when dealing with partially observable problems, so

it needs to be carefully considered in specific

applications. Increasing the number of MD may

affect the performance of the algorithm, depending on

factors such as the algorithm type, state transition

method, and metrics. The average delay and MDR are

used as effective tools to evaluate the performance of

the algorithm (Lu, Li, 2024).

Finally, it is verified that MF-MATO algorithm

and MADDPG algorithm can make MD cooperative

decision in the system, and the utilization rate of

computing resources is higher.

Figure 3: Average latency curve of different algorithms

(Lu, Li, 2024).

Figure 4: Average latency and MDR vary with the number

of MDs (Lu, Li, 2024).

5 CONCLUSIONS

With the rise of Internet of Things (IoT) technology,

the number of smart devices has surged, and the

number of computing-intensive and latency-sensitive

tasks has gradually increased. Due to the limitations

of power consumption, computing power and storage

capacity, mobile devices face huge challenges in

processing these tasks. Edge distributed computing

offloading technology, as a solution, reduces task

processing latency and energy consumption by

deploying computing resources at the network edge.

Scalability, location correlation, diversity,

randomness, time variability and autonomy are the

characteristics of edge distributed systems, which

make edge distributed systems flexibly applied in

different scenarios. The application of blockchain

technology includes the use of the decentralization,

immutable and transparent characteristics of

blockchain technology to ensure the safety and

reliability of tasks in edge distributed computing, and

the automated management of computing tasks

through smart contracts. The paper proposes a smart

contract-based shared bandwidth allocation algorithm

combined with blockchain technology to reduce

delay and improve system scalability, optimize the

mobile edge collaborative learning algorithm, and

improve task unloading efficiency and system

performance. This paper verifies the effectiveness

and performance of the algorithm through

experiments, including task completion time, system

energy consumption, data quality improvement and

privacy protection effect.

The experimental results show that the proposed

method can significantly reduce the task completion

time, improve the overall response speed of the

system, reduce energy consumption and extend the

battery life of the equipment. Compared with other

research results such as D2HM algorithm and JORA

method, this paper shows the unique advantages of

the proposed method in task unloading, resource

allocation and system performance optimization.

In conclusion, this paper explores the key

technologies of blockchain-based edge distributed

computing offloading, proposes the optimization

algorithm method, and verifies its performance.

However, there are still some limitations and

challenges that need to be further explored and

addressed by future research.

Unloadable Computing Problem Based on Edge Distributed Computing

643

AUTHORS CONTRIBUTION

All the authors contributed equally and their names

were listed in alphabetical order.

REFERENCES

Babatunde, S., Bodhaswar, T. J., Attahiru, S. 2016. Solving

resource allocation problems in cognitive radio

networks: a survey. EURASIP J. Wireless Comm. and

Networking. Volume 2016 , Issue 1. PP 1-14.

Baranwal, G., Kumar, D., Vidyarthi, D. P. 2023.

Blockchain based resource allocation in cloud and

distributed edge computing: A survey. Computer

Communications. Volume 209, Issue. PP 469-498.

Chen, J., Wang, Y. L., Liu, T. 2021. A proactive resource

allocation method based on adaptive prediction of

resource requests in cloud computing. EURASIP

Journal on Wireless Communications and Networking.

Volume 2021, Issue 1.

Hisham, M. A. 2025. UAV-assisted mobile edge computing

model for cognitive radio-based IoT

networks. Computer Communications 108071-108071.

Gao, R. L., Zhang, W. Z., Mao, W. Y., Tan, J. J., Zhang, J.,

Huang, H. Y., ... & Huang, F. Y. 2025. Method towards

collaborative cloud and edge computing via RBC for

joint communication and computation resource

allocation. Journal of Industrial Information

Integration 100776-100776.

Li, Y., Nan, Z. Y., Yao, Z. X., et al. 2025. Distributed

intelligent computing offloading and service caching

joint optimization for DAG tasks. Journal of Sun Yat-

sen University (Natural Science Edition in Chinese and

English), 64(1): 71-82.

Li, Y., Zhang, J. X., Yao, Z. X., et al. 2024. Neighborhood-

aware distributed intelligent edge computing offloading

and resource allocation algorithm. Science China:

Information Sciences, 54: 413–429, doi: 10.1360/SSI-

2023-0177

Li, Z. Y., Wu, Y., Zhang, W. W., Li, Y. G. 2022. A Trusted

Identity Authentication Model for Crowdsourcing

Logistics Based on the UTXO - Model Blockchain.

Journal of Jiangsu University.

Lin, H. C., 2024. Research on distributed task offloading

technology under edge computing. Computer Science,

51(6): 1-10.

Lu, X. L. & Li, D. K. 2024. Task unloading algorithm for

large-scale multi-access edge computing scenarios.

Journal of Electronics and Information. 1-11.

ICDSE 2025 - The International Conference on Data Science and Engineering

644