Application Migrations in Edge Computing Using Real Time

Workload Ranking Algorithm

Saravanan Velrajan and Thangakumar Jeyaprakash

Department of Computer Science, Hindustan Institute of Technology and Science Chennai, India

Keywords: 5G Networks, Edge Computing, Application Performance Management, Workload Migrations.

Abstract: Service Providers host a multitude of diverse applications in edge clouds in 5G networks for improving the

user experience. The demand for automation of workflows in different industries has created the need for 5G-

based private wireless networks with edge infrastructure. The varied nature of the workloads hosted at the

edge clouds, lead to increased focus on service quality and performance. Application workload migrations are

done in edge infrastructure to improve the SLA adherence of services and to improve user experience. This

research proposes a new adaptive ranking approach for application workload migrations in edge computing.

The proposed Adaptive Workload Ranking Algorithm (AWRA) predicts the available time for workload

migrations and triggers migration of application workloads. The selection of workloads to be migrated is

made by balancing the resource utilisation of application instances and the estimated time taken for migration.

Our experiments demonstrate that AWRA effectively manages overload conditions of varied applications

hosted in edge infrastructure. AWRA outperformed state-of-the art algorithms such as ASMWA and EWMA3

in minimising SLA violations during workload migrations.

1 INTRODUCTION

The growth of 5G networks has created a paradigm

shift in wireless communication capabilities.

Characterised by a significant bandwidth surge and a

dramatic latency reduction, 5G has unlocked the

potential for applications such as online gaming, live

video streaming, holograms and smart communities.

Wireless service providers deploy edge cloud at the

network edges to guarantee QoS for applications

accessed by mobile subscribers, as given in Figure 1

(Velrajan and Sharmila, 2023).

Figure 1: Edge Cloud Deployment at Service Provider Edge

Edge computing provides new monetization

opportunities for service providers. However, the

diverse applications hosted at the edge servers in 5G

networks have strict Quality of Service (QoS) needs

(Velrajan and Sharmila, 2023). The diverse

applications deployed in the edge have varying QoS

characteristics. For example, a video streaming

application would expect the 5G network

infrastructure to support high bandwidth; an

autonomous car application would need support for

lower latencies; and an Industry 4.0 deployment

would require higher reliability from the network.

Managing the QoS of users at the edge is complex

because of the distributed deployment of the edge

infrastructure and the dynamic nature of the

workload.

Edge infrastructure continuously monitors the

availability and performance of applications running

on the edge servers (Velrajan and Sharmila, 2023).

When an application instance in edge server becomes

overloaded, it impacts the Service Level Agreements

(SLAs) of users. Thus, wireless service providers

perform workload migrations in edge to prevent

applications from becoming overloaded (Velrajan

and Sharmila, 2023). However, migrating a workload

from one edge host to another often results in

disruption of services to users, impacting their quality

of experience (Velrajan and Sharmila, 2023).

Existing research primarily focuses on workload

migrations in centralised cloud environments. A new

320

Velrajan, S. and Jeyaprakash, T.

Application Migrations in Edge Computing Using Real-Time Workload Ranking Algorithm.

DOI: 10.5220/0013591600004664

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

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 320-327

ISBN: 978-989-758-763-4

approach is necessary to handle the challenges

created by dynamic workloads, diverse applications,

and the distributed deployment architecture of edge

computing in 5G networks.

Our research focuses on workload migrations to

manage application overload scenarios in edge

infrastructure in 5G networks. We adopt a proactive

approach to initiate workload migrations that

minimise the impact on QoS experienced by users.

This work introduces two key advancements to

enable QoS-aware migration of workloads in edge

clouds:

1. Adaptive Workload Ranking Algorithm

(AWRA) for proactive workload migrations

within edge clouds.

2. We evaluate the performance of AWRA

against other workload migration methods

through a comprehensive analysis.

AWRA ranks workloads to migrate, considering

the user sessions handled and the migration time

needed. Furthermore, AWRA uses a variable

application performance limit to initiate workload

migrations. Our simulations reveal that AWRA

outperforms existing approaches in minimising QoS

impact during workload migrations.

2 LITERATURE REVIEW

Ensuring optimal application performance in edge

cloud necessitates efficient workload migration

strategies. This review focuses on recent workload

migration advancements, specifically geared towards

resource optimisation, user mobility, energy savings

and application performance management.

2.1 Optimisation of system resources

Optimising system resource allocation is important

for delivering high-quality workloads to users in edge

clouds running resource hungry applications like

gaming and live video delivery. Several existing

studies highlight the importance of workload

migration in a timely fashion to optimise resource

utilisation and ensure workload continuity for users.

The experiments done by Chen et al. focus on

continuity of user workloads during migrations while

optimising resource utilisation. The authors aim to

perform migrations without impacting the service

quality for users (Chen et al., 2024). Through their

simulation, the authors show that initiating workload

migrations during appropriate time and identifying

the best node for migrating workloads are important

for achieving outstanding service quality.

The studies done by Wan et al. use federated

mechanisms to assign resources to user services, with

a goal of lowering energy consumption and

minimising the impact to service quality (Wan,

2024). The authors use resource-based and task-based

methods to optimise compute and power resources in

an edge computing infrastructure deployed in IoT

networks. The authors claim that the proposed

solution improves the performance of edge

infrastructure and lowers power consumption, in

addition to optimising resource management.

2.2 Management of user mobility

Quite often user services get migrated in edge clouds

due to the mobility of subscribers, as their sessions

get relocated across the different mobile towers.

Autonomous vehicles connecting to 5G networks

want uninterrupted and low-latency connectivity to

the workloads hosted in the edge clouds.

Cao et al. introduce a dynamic path selection

algorithm to manage workload migrations during

subscriber mobility scenarios (Cao et al., 2024). The

authors claim to reduce the impact of migrations on

user session latency, and at the same time, operating

within the constraints of computing cost.

Shokouhi et al. present a hierarchical architecture

based framework for user tasks offloading in edge

clouds during mobility situations (Shokouhi et al.,

2024). The framework introduced by the authors uses

a predictive approach to decide the appropriate edge

server to deliver service to the user. The authors claim

that the presented approach is best suited to reduce

the cost of offloading user tasks during mobility

scenarios.

2.3 Optimise energy consumption

Zolfaghari adopts a predictive mechanism to migrate

workloads in cloud data centers, to achieve energy

and performance tradeoff (Zolfaghari, 2024). The

proposed VM selection method for migrations in the

cloud, is claimed to achieve energy savings,

complying with SLAs to the best extent possible.

Ouhame et al. use self-optimising machine

learning approach to migrate VMs in a cloud

environment with workloads that are dynamic in

nature (Ouhame et al., 2024). The proposed algorithm

uses an intelligent learning approach that disables

VM nodes in cloud that are not in use. The proposed

solution claims to achieve quicker workload

migration decisions, improved SLA adherence and

reduced energy consumption in datacenter clouds.

Machine learning based intelligent techniques are

Application Migrations in Edge Computing Using Real-Time Workload Ranking Algorithm

321

used often in cloud to optimise energy consumption,

as they directly reflect on operational cost savings.

However, it is equally important to honour the SLAs

for user sessions and avoid delays.

2.4 VM migrations in centralised cloud

Workload migrations in the cloud has been gaining

attention in the recent times with researchers

exploring techniques such as migration of workloads

using VMs or containers. However, the distributed

nature of edge infrastructure prevents the seamless

extension of cloud workload migration techniques

without compromising on the user’s service quality

(Velrajan and Sharmila, 2023).

Kulshrestha et al. use a method based on

Exponential Weighted Moving Average (EWMA3)

to migrate VMs hosted in the centralised datacenters

(Kulshrestha and Patel, 2021). The authors compare

the behaviour of EWMA3 with other popular VM

selection algorithms and demonstrate that EWMA3

reduces SLA violations in the cloud during workload

migrations.

A proactive method is introduced by Patil et al. for

overload detection in the cloud to migrate VMs,

especially during burst load conditions (Patil and

Patil, 2023). The authors claim that the proposed

algorithm efficiently manages recurrent VM

migrations and instant overload conditions in the

cloud environment.

The studies on VM migration in cloud are mostly

restricted to host overload management and energy

optimisation and do not consider application-specific

characteristics during migration. Also, the VM

migration techniques adopted in the centralised cloud

environments are not extensible for the distributed

edge cloud architectures.

2.5 Application Performance

Management

Velrajan et al. have introduced a distributed

mechanism to migrate workloads at the edge

minimising the impact on user’s QoS (Velrajan and

Sharmila, 2023). The proposed approach decides the

best application performance limit for initiating

workload migrations using an enhanced Particle

Swarm Optimization (PSO) algorithm. This approach

is suitable in environments where latency-sensitive

workload migrations must be performed.

Velrajan et al. have proposed a dynamic migration

approach in edge computing using a continuously

adapting workload migration window (Velrajan and

Sharmila, 2023). The proposed Adaptive Service

Migration Window Algorithm (ASWMA)

significantly minimises SLA violations in a QoS-

sensitive environment while performing workload

migrations. However, it doesn't differentiate

workloads during migration, impacting critical

workloads that handle user sessions.

While significant progress has been made in

workload migrations in edge cloud environments,

there's still room for improvement, especially in

studying the behaviour of workload migrations in

QoS-constrained environments. Existing workload

migration approaches do not consider the time taken

for workload migrations. However, in the real world,

there is a need for time-bound proactive workload

migrations for improved application performance

management.

Furthermore, existing approaches do not

differentiate the workloads of an application during

workload migrations. For example, a user facing

workload that streams video to users may have a

higher rank than the workload that generates

dashboards in a video streaming server. Thus, a

migration approach that ranks workloads would help

improve QoS and reduce SLA violations in QoS-

sensitive applications such as on-demand or live

video streaming, smart surveillance and AR/VR.

3 ADAPTIVE WORKLOAD

MIGRATIONS

This work introduces an Adaptive Workload

Ranking Algorithm (AWRA) to address the growing

need for efficient and QoS-aware workload

migrations in edge clouds. The algorithm uses the

application load, number of services handled by the

application and workload migration time to decide

which workloads to prioritise during migrations.

AWRA’s proactive approach based on workload

priority minimises users’ QoS impact during

workload migration. Table 1 captures the

abbreviations and symbols used in this paper.

Table 1: Abbreviations and Symbols

Parameter

Description

appPerfLimit

Application performance limit at

which workload migrations are

initiate

𝐴𝑃𝑆

Application Performance Score

representing application’s overall

resource utilisation and user’s QoE

𝐴𝑃𝑆



Application Performance Score before

initiating workload migration

INCOFT 2025 - International Conference on Futuristic Technology

322

Parameter

Description

𝐴𝑃𝑆



Application Performance Score after

completing workload migration

𝐴𝑃𝑆



Adjustment done to the application

erformance limit for migrations

Application’s Aggregated Resource

utilisation

QoE Quality of Experience for the users

𝑀𝑇



Predicted Migration Time for

Workloads

𝑀𝑇



Actual Time taken to migrate

workloads

𝑊𝑘𝑙𝑑𝑠



Workloads list with the user sessions

count for an application instance

𝑊𝑘𝑙𝑑𝑠



Ranked workload list for an

application instance

slaViolations

SLA Violations during the migration

of workloads

Figure. 2 visually depicts the critical steps

involved in workload migrations based on real-time

application performance and user experience metrics

using the AWRA algorithm.

AWRA monitors the performance of applications

running on the edge hosts using Application

Performance Score (APS) (Velrajan and Sharmila,

2023). Unlike static performance limits configured in

a centralised cloud environment (VMWare, 2024),

the APS metric considers overall system resource

utilisation and user-experienced QoS for each

application instance. Thus, APS uniquely represents

the characteristics of diverse applications hosted in

edge cloud.

Figure 2: Flowchart for AWRA Algorithm.

The calculation of APS is detailed in equation (1).

Resource utilisation of an application represented by

RU(app_id) aggregates the utilisation of system

resources including CPU, disk storage, network ports

and RAM; and QoE is the users’ quality of experience

from the workload (Velrajan and Sharmila, 2023).

𝐴𝑃𝑆 (𝑎𝑝𝑝_𝑖𝑑) = 𝑅𝑈(𝑎𝑝𝑝_𝑖𝑑) + (𝑄𝑜𝐸/𝑅𝑈(𝑎𝑝𝑝_𝑖𝑑))

(1)

At the beginning of the execution, AWRA

initialises the application performance limit

(appPerfLimit) to trigger migrations to

DEFAULT_LOAD_LIMIT, which is 70% of

resource utilisation. However, as the algorithm

continuously performs workload migrations, the

appPerfLimit value gets dynamically adjusted to

represent the most optimal application migration

performance limit.

AWRA invokes Linear_Regression algorithm to

detect whether the application’s overall load

increases or decreases. Linear_Regression helps

identify the application instance’s performance trend.

When the application’s performance score (APS)

grows and exceeds the application’s performance

limit for migration (appPerfLimit), AWRA decides to

migrate the workload to prevent SLA violations.

Before initiating migrations, AWRA invokes a

simple moving average algorithm to forecast the

available time to complete workload migrations

(𝑀𝑇



AWRA strives to complete application migration

by ranking workloads and selecting workloads that

can fit the forecasted time interval. AWRA creates a

prioritised workloads list (𝑊𝑘𝑙𝑑𝑠



) considering two

factors – user sessions handled and the migration

time. While dynamic algorithms such as ASMWA

assume that all workloads consume the same time for

migration, AWRA considers the migration time of

individual workloads. Thus, AWRA is well-

positioned to migrate an optimal number of

workloads without violating SLAs.

AWRA then iterates through the ranked

workloads list and invokes the

MigrateAppWorkloads() procedure to trigger

migrations from heavily loaded instances to

application running in target host with available

resources. AWRA attempts to migrate as many

workloads as possible within the forecasted migration

time.

The migration procedure is stopped when the

forecasted migration time is exceeded or when all the

workloads belonging to an application are migrated.

When the workloads of an application instance

cannot be migrated in the estimated time duration, it

is considered an SLA Violation. AWRA decides

whether to adjust the application performance limit

for migrations based on the SLA Violation incurred

during the workload migration.

AWRA updates the application performance limit

when the actual migration time exceeds the

forecasted migration time or when all the workloads

part of the application are not migrated (Velrajan and

Sharmila, 2023). AWRA calculates 𝐴𝑃𝑆



and

Application Migrations in Edge Computing Using Real-Time Workload Ranking Algorithm

323

adjusts the application performance limit for

migrations as given in (2).

𝐴𝑃𝑆



(



 



) ∗ (



 



)





(2)

By continuously adjusting the application

performance limit used to trigger migration, AWRA

dynamically adapts to the indeterministic workloads

served by the edge cloud infrastructure. Dynamic

adaptation of AWRA to the changing workloads

contributes to the improved user experience and

dramatic reduction in workload quality violations.

Algorithm: Adaptive Workload Migration

Algorithm (AWRA)

Input: Edge Network Topology with

Applications, Edge Hosts and Edge Clusters,

Resource Utilisation of

Applications, user sessions’ QoE.

Output: No return value

proc AWRA (topology, RU, QoE)

{

appPerfLimit = DEFAULT_LOAD_LIMIT

foreach app_id in

topology(app_instances) do

appPerf =run

Linear_Regression(RU(app_id))

APS



= ComputeAPS(RU(app_id), QoE)

if ((appPerf != LOAD_INCREASE) or

(APS



< appPerfLimit))

goto next app instance

closeif



= MovingAverage(RU(app_id))



= GetTimestamp (app_id)

Wkld



= GetWorkloadsList (app_id)

sort Wklds



BY reverse=True

Wklds



=NULL



= 0

foreach wkld in Wklds



if (MT(wkld) + MT



> MT



)

break

Wklds



.append(wkld)



= MT



+ MT(wkld)

endfor

slaViolations = MigrateAppWorkloads

(Wklds



)



= GetTimestamp (app_id)



= (TS



− TS



)

APS



= ComputeAPS(RU(app_id), QoE))

if ((slaViolations) or

(MT



< MT



))

APS



= (APS



- APS



) *

(MT



− MT



)/MT



appPerfLimit = appPerfLimit-APS



closeif

endfor

return

}

4 EXPERIMENTS & RESULTS

AWRA was validated in a simulated environment

using a edge cloud topology built using NetworkX, a

popular network topology simulation toolkit. We

leveraged SciPy, Numpy and Pandas Python libraries

for data analysis and reporting. We used a video

streaming application profile, to validate the behavior

of AWRA.

A topology similar to the real-world service

provider networks is simulated using 20 edge

clusters, each with 6 hosts. The tests were performed

with 100 application instances running across the

different edges in edge cloud. Testing was done using

25,000 user sessions accessing the video streaming

server.

User tasks modelling video streaming users were

created and mapped to various application instances

running on the edges, randomly. The compute,

storage, network and memory utilisation of each

application instance was continuously monitored and

recorded. AWRA keeps track of application

instances’ resource utilisation at the workload level

granularity. Furthermore, user sessions’ QoS metrics

are periodically collected. APS is calculated from the

application’s aggregated resource consumption and

users’ quality of experience metrics collected

periodically from the edge cloud.

INCOFT 2025 - International Conference on Futuristic Technology

324

As time progressed, the tasks handled by an

application instance increase, resulting in increased

resource consumption at the edge hosts. Our

experimental findings in Fig 3 reveal a direct

correlation between an application instance's

hardware resource consumption and the user sessions

handled. Fig 3 also shows the application instance's

aggregated resource utilisation (App. RU), which

grows with a rise in user sessions.

Figure 3. Overall Resource Utilisation of Application

Our experiments demonstrated AWRA’s ability

to prevent application overload by proactively

triggering workload migrations. As AWRA migrates

workloads by ranking them based on the user session

count and the load reduction achieved, it reduces SLA

violations and improves the QoS experienced by

users, as represented in Fig 4.

Figure 4. QoS-aware Migration of Workloads

The performance of AWRA was compared with 3

different algorithms for workload migrations –

ASMWA, EWMA3 and static performance limit

algorithms. We evaluated these algorithms

objectively based on their ability to minimise SLA

violations during the migration of application

workloads in edge cloud.

In our analysis, 17% of workload migrations

exceeded the allocated migration time for AWRA,

26% for ASMWA, 41% for EWMA3, and 42% for

migrations done using static performance limit, as

given in Figure 5. AWRA’s ability to predict the

available workload migration time and rank

workloads substantially decreases the workload

migration time violations.

Our experiments prove that a dynamic application

performance limit is more suitable for handling

workload migrations in environments with

indeterministic workloads. Using a greedy approach

to pack workloads based on migration interval helps

minimise workload migration time-related violations

in AWRA. Migrations done using static performance

limit could not rapidly adapt to the changing

performance characteristics of application instances,

resulting in more SLA violations. ASMWA and

EWMA3 algorithms could perform relatively better,

because of their ability to adapt to real-time load

conditions as given in Figure 5.

Application Migrations in Edge Computing Using Real-Time Workload Ranking Algorithm

325

Figure 5. Workload Migration Interval – Allotted vs.

Actual

We compared AWRA's performance with state-

of-the-art algorithms in minimising QoS degradation.

Our simulations demonstrated that AWRA reduces

QoS violations by 13% compared to ASMWA and

33% compared to EWMA3. As represented in Figure

6, AWRA achieves a 67% improvement over

migrations using a static performance limit.

Prioritising workloads for migration and using an

adaptable mechanism to decide the application

performance limit enables AWRA to migrate an

optimal number of workloads within the forecasted

migration time.

Figure 6. Comparison of QoS Violations of algorithms

The experiments showed that AWRA

outperforms previously established algorithms,

including ASMWA, EWMA3, and static

performance limit, in adhering to SLAs and

enhancing the user experience while migrating

application workloads in edge cloud. Our study

demonstrates that AWRA proactively adapts to

changing load conditions in the edge infrastructure,

compared to the existing methods used for workload

migrations.

5 CONCLUSION

We proposed an Adaptive Workload Ranking

Algorithm (AWRA) that performs application

migrations in edge cloud by dynamically ranking

workloads based on the load and migration time.

AWRA distinctly uses a dynamic application

performance limit for migrations, compared to the

existing static approaches. This enables AWRA to

quickly adapt to dynamic workloads. We studied the

effectiveness of AWRA in migrating workloads at the

edge and compared it with the existing algorithms -

static performance limit, ASMWA and EWMA3 by

running simulations. Our research established that

AWRA significantly outperformed static

performance limit, ASMWA and EWMA3 methods

in reducing user QoS violations and improving the

service quality. Compared to ASMWA, AWRA

achieved a 13% improvement, while it delivered a

33% reduction in violations compared to EWMA3.

AWRA performed 67% better than migrations done

using static performance limit.

We observed that AWRA performs better in

managing application overload conditions in edge

topologies where there is gradual increase in

workloads. When the workloads are unpredictable or

when there are traffic bursts, 17% of workload

migrations exceeded the allocated migration time

using AWRA. Thus, future research can look at

AWRA's efficiency under regular load conditions and

at the same time, its adaptability to traffic bursts. In

addition, there is opportunity to compare the

performance of AWRA with other algorithms that

rely on distributed decision-making methods such as

PSO.

REFERENCES

Wan, X. (2024). Dynamic Resource Management in MEC

IoT. Journal of Grid Computing.

INCOFT 2025 - International Conference on Futuristic Technology

326

Chen, C., Song, Y., Jiang, Y., & Zhang, M. (2024).

Optimizing Network Service Continuity with Quality-

Driven Resource Migration. Electronics, 13(9), 1666.

Patil, A., & Patil, R. (2023). Proactive and dynamic load

balancing model for workload spike detection in cloud.

Measurement: Sensors.

Cao, B., Ye, H., Liu, J., Tang, B., Tao, Z., & Deng, S.

(2024). SMART: Cost-Aware Service Migration Path

Selection Based on Deep Reinforcement Learning.

IEEE Transactions on Intelligent Transportation

Systems.

Velrajan, S., & Sharmila, V. C. (2023). Service Migrations

in Multi-Access Edge Computing Using Adaptive

Migration Window. In 2023 6th International

Conference on Contemporary Computing and

Informatics (IC3I) (Vol. 6). IEEE.

Velrajan, S., & Sharmila, V. (2023). QoS-aware service

migration in MEC using closed-loop adaptive PSO

algorithm. Journal of Network and Systems

Management.

Shokouhi, M. H., Hadi, M., & Pakravan, M. R. (2024).

Mobility-Aware Computation Offloading for

Hierarchical Mobile Edge Computing. IEEE

Transactions on Network and Service Management.

Zolfaghari, R. (2024). Energy-performance-aware VM

migration in cloud by using prediction and fuzzy

approaches. Engineering Applications of Artificial

Intelligence.

Ouhame, S., Hadi, M. Y., Mrhari, A., & Laassar, I. (2024).

Artificial intelligence-based cloud-IoT resource

management for energy conservation. International

Journal of Advances in Applied Sciences.

VMWare DRS Migration Threshold (2024).

docs.vmware.com. Retrieved from

https://docs.vmware.com/en/VMware-

vSphere/8.0/vsphere-resource-management/GUID-

8ACF3502-5314-469F-8CC9-4A9BD5925BC2.html

Kulshrestha,S., & Patel,S.(2021). An efficient host

overload detection algorithm for cloud data center

based on exponential weighted moving average.

International Journal of Communication Systems.

Application Migrations in Edge Computing Using Real-Time Workload Ranking Algorithm

327