Swarm Intelligence-Based Algorithm for Workload Placement in

Edge-Fog-Cloud Continuum

Kefan Wu

1 a

, Abdorasoul Ghasemi

2 b

and Melanie Schranz

1 c

Lakeside Labs, Klagenfurt, Austria

Research Centre for Computational Sciences and Mathematical Modelling, Coventry University, Coventry, U.K.

{wu, schranz}@lakeside-labs.com, ae4959@coventry.ac.uk

Keywords:

Swarm Intelligence, Edge-Fog-Cloud Continuum, Ant Colony Optimization.

Abstract:

This paper addresses the workload placement problem in the edge-fog-cloud continuum. We model the edge-

fog-cloud computing continuum as a multi-agent framework consisting of networked resource supply and

demand agents. Inspired by the swarm intelligence behavior of the ant colony optimization, we propose a

workload scheduler for the arriving demand agents to increase local resource utilization and reduce com-

munication costs without relying on a centralized scheduler. Like the ants, the demand agents will release

pheromones on the resource agent to indicate the available resources. The next arriving demand agent will

most probably choose a neighbor, following the pheromone value and communication cost. The framework’s

performance is evaluated in terms of local resource utilization, dependency on fog and cloud, and commu-

nication cost. We compare these metrics for the ant-inspired algorithm with random and greedy algorithms.

The simulation results reveal that the proposed algorithm inspired by swarm intelligence can increase resource

utilization at the edge and reduce the dependency on higher layers, while also decreasing the communication

cost for the task of resource allocation.

1 INTRODUCTION

Data management and computing are key research ar-

eas in the Internet of Things (IoT). The initial solu-

tion is based on cloud computing due to its scalability

and efﬁciency. However, transmitting a large amount

of data to the cloud servers suffers from unpredictable

delays (Jiao et al., 2013). Storing privacy-critical data

in the cloud server may also trigger security issues

(Parikh et al., 2019). Hence, data processing closer to

the user, on the edge, should be considered.

Edge computing facilitates data processing at the

network’s edge (IoT devices or local servers) (Satya-

narayanan, 2017), This approach brings computation

closer to the data source, allowing for faster response

times and improved data security (Simi

c et al., 2021).

Nevertheless, edge devices have limited computing

and storage capabilities compared to cloud servers.

To bridge this gap, fog computing acts as an inter-

mediary layer to extend cloud services closer to the

edge devices (Bonomi et al., 2012), materializing the

https://orcid.org/0000-0001-7612-581X

https://orcid.org/0000-0002-4432-4504

https://orcid.org/0000-0002-0714-6569

edge-fog-cloud continuum. When the request, e.g., a

pod, the smallest and most basic deployable unit in

Kubernetes (Kim et al., 2021), arrives in the contin-

uum step by step randomly, they will be placed based

on their speciﬁc requirements such as resource de-

mand, latency, data security, etc. Hence, the workload

scheduler should be designed carefully to minimize

latency and maximize resource utilization.

Inspired by the collective behaviors in nature such

as schools of ﬁsh and ﬂocks of birds, swarm intelli-

gence operates through decentralized, self-organizing

systems where the agents follow local rules with-

out centralized controllers leading to scalable, robust

and adaptive solutions (Schranz et al., 2021). These

characteristics make agent-based, swarm intelligence

methods particularly well-suited for optimizing re-

source allocation and task distribution in the hetero-

geneous and dynamic edge-fog-cloud environment.

Exactly this approach presents the focus of our pa-

per: We model the edge-fog-cloud computing contin-

uum as an agent-based system with resource agents

(edge, fog, cloud agents) and demand agents (pods).

Resource agents are characterized by different pa-

rameters such as available resources, communication

cost, etc. Demand agents decide the optimal resource

310

Wu, K., Ghasemi, A. and Schranz, M.

Swarm Intelligence-Based Algorithm for Workload Placement in Edge-Fog-Cloud Continuum.

DOI: 10.5220/0013140800003890

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

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 310-317

ISBN: 978-989-758-737-5; ISSN: 2184-433X

agent for processing based on these parameters. We

propose a swarm intelligence workload scheduler mo-

tivated by the ant colony optimization (ACO) algo-

rithm (Dorigo and St

utzle, 2003). Pods are modeled

as ants. They can lay down the pheromones at each

resource agent with a designed update rule based on

the available resources when pods leave the agent.

We also deﬁne the probability of the pods moving

to the neighbors based on the pheromone values and

communication costs if pods are not served in the

current resource agent. Simulation results illustrate

that the ACO-based scheduler outperforms baseline

schedulers by employing random and greedy strate-

gies. Summarized, the main contributions of the pa-

per are the following:

• We use agent-based modelling to map the re-

source and demand agents, presenting the com-

plex interplay in the edge-fog-cloud continuum.

• We propose a swarm intelligence-based workload

scheduler inspired by ACO. Speciﬁcally, we intro-

duce the pheromone update to indicate the avail-

able resources, joined with the communication

cost for a reallocation on each resource agent.

• We develop a new simulation framework to evalu-

ate the ACO algorithm against random and greedy

methods through different case studies. The re-

sults show that ACO algorithm outperforms the

baseline algorithms in terms of higher local re-

source utilization, lower communication cost, and

a reduced dependency on fog and cloud nodes.

The remainder of the paper proceeds as follows:

Section 2 introduces the current state-of-the-art in the

edge-fog-cloud continuum, swarm intelligence, and

ACO. Section 3 presents the model of the edge-fog-

cloud continuum. The workload scheduler based on

ACO is proposed in Section 4. Section 5 provides the

simulation case studies compared with two baseline

algorithms. Section 6 concludes this work.

2 RELATED WORK

This section reviews the current state-of-the-art in

the edge-fog-cloud continuum workload placement

mechanism, general and with a special focus on

swarm intelligence.

Workload allocation in the edge-fog-cloud contin-

uum presents signiﬁcant challenges due to the het-

erogeneous nature of resources and varying compu-

tational capabilities across different layers (Shi et al.,

2016). In recent years, several methods worked on

strategies for workload ofﬂoading in edge computing:

For instance, federated learning was implemented to

tackle the ofﬂoading problem in vehicular edge com-

puting in (Hasan et al., 2024). In another work, game

theory was considered for the task ofﬂoading in edge

computing (Chen et al., ). In (Tong et al., 2020), an

adaptive task ofﬂoading and resource allocation al-

gorithm is proposed for mobile edge computing en-

vironments, utilizing deep reinforcement learning to

make ofﬂoading decisions and allocate computing re-

sources efﬁciently. However, in the works mentioned

above, the paradigms are considered isolated. To inte-

grate them as a continuum seamlessly over all layers

on the edge, fog and cloud, analytical and hierarchical

availability models were provided for the edge-fog-

cloud computing continuum in (Pereira et al., 2022).

A workload scheduling based on an adaptive con-

tainer managed by Kubernetes was proposed for the

continuum in (Robles-Enciso and Skarmeta, 2024). It

is noticeable that how to manage the edge-fog-cloud

continuum more effectively is still a challenge for

workload placement design (Gkonis et al., 2023).

This brings us to swarm intelligence, a promis-

ing approach that provides adaptive methods for op-

timizing workload allocation in the edge-fog-cloud

continuum, leveraging self-organization to man-

age distributed resources efﬁciently (Schranz et al.,

2021).In (Rajesh et al., 2020), a swarm intelligence

algorithm was applied in IoT routing to extend net-

work life and cut energy costs. The authors in (Saba

et al., 2023) proposed a load-balancing method us-

ing swarm intelligence for data management to re-

duce the response time of the cloud servers. In an-

other work, an innovative hybrid swarm intelligence

algorithm was designed in (Attiya et al., 2022) for

task scheduling in cloud computing to increase the

rate of convergence. The most innovative work comes

from (Ghasemi and Schranz, 2024), where the au-

thors tackle the dynamic nature of the available re-

sources and workloads from the bottom-up. The

delay-sensitive and insensitive pod agents are co-

located in the modelled multi-agent system to min-

imize slack resources while improving resource uti-

lization for edge micro data centers.

To tackle the high complexity behind workload

allocation in the continuum, we implement the Ant

Colony Optimization (ACO) based on the swarm be-

haviors of real ant colonies (Dorigo et al., 2006). The

authors in (Corominas et al., 2023) proposed an ACO-

based method to deal with the network alignment is-

sue. In (Wu et al., 2023), an adaptive ACO algorithm

was designed for path planning in robotics. In another

work, the resource allocation problem was considered

in (Du et al., 2023) for avionics systems. They pro-

vided an improved ACO strategy to minimize the en-

ergy cost during the allocation task. Moreover, the

Swarm Intelligence-Based Algorithm for Workload Placement in Edge-Fog-Cloud Continuum

311

MEA 𝑣

MFA 𝑣

MCA 𝑣

IoT Devices

EMDCs

Cloud Server

Figure 1: Exemplary architecture of the considered Edge-

Fog-Cloud Continuum.

authors in (Chandrashekar et al., 2023) proposed a

hybrid weight ACO method for task scheduling in

cloud computing. This novel scheduler was able to

enhance the performance compared with some tradi-

tional algorithms. Summarized, ACO is a powerful

tool for optimization problems in network routing and

scheduling (Bai et al., 2024). Thus, the ACO seems

to be a candidate algorithm for the ﬂexible workload

placement we require in the highly dynamic edge-fog-

cloud continuum.

3 SYSTEM MODEL AND MAIN

OBJECTIVES

In this Section, we model the edge-fog-cloud contin-

uum as a multi-agent swarm system and describe the

features of the swarm agents. Three metrics are pro-

vided to evaluate the performance of the scheduler.

3.1 System Model of the Continuum

We introduce two types of swarm agents in the

continuum: resource agents and demand agents.

The resource agents are located in each comput-

ing node with available CPU and Memory (MEM)

resources. They are embodied by a Multi-layer

360° dynamIc RunTime Orchestration (MIRTO) en-

gine (related to the EU-Horizon project MYRTUS,

see ACKNOWLEDGEMENT), so we call them

MIRTO agents (Palumbo et al., 2024). Thus,

the MIRTO agents in the edge/fog/cloud comput-

ing paradigm are the MIRTO edge/fog/cloud agents

(MEA/MFA/MCA) in this paper. The type and capac-

ity of the resources depend on their layer. Each MEA

includes some IoT devices such as robots, smart-

phones, etc. MFA is comprised of a set of edge micro

data centers (EMDC), and MCA is run by big servers

with unlimited resource capacities. The main prin-

ciple of the considered edge-fog-cloud continuum is

sketched in Figure 1. The MEA/MFA/MCA are de-

noted by blue/green/orange blocks. The solid black

lines represent the connections in the continuum. The

demand agents are represented by pods in the Kuber-

netes context. Each pod has a speciﬁc demand for

resources and execution steps. So they are divided

into small, medium, and large pods. Pods arrive in

the continuum with speciﬁc steps generated by an ex-

ponential random variable with a parameter µ ∈ (0,1],

which reveals the frequency of arriving pods.

The different resource agents are also connected

to ensure that pods can move freely through the

links. This enables all resource agents to jointly

form a seamless edge-fog-cloud continuum. Ac-

cording to the concepts in graph theory (Godsil and

Royle, 2013), we can deﬁne the interaction topology

of the resource agents in the continuum as a graph

G = (V ,E). V = V

∪ V

= {v

,..., v

} is

the vertex set of each resource swarm agent, where

= {v

,..., v

}, V

= {v

,..., v

}, and

= {v

,..., v

} represent the vertex set of

the MEA, MFA, and MCA, respectively. N

, N

, and

denote the number of the MEA, MFA, and MCA.

N = N

+ N

stands for the number of all the re-

source swarm agents. The set E ⊆ V × V reveals the

connection of the continuum. ∀v

∈ V , (v

) ∈ E

means that there are connection between the resource

agent i and j, and pods can move between the i

and j

resource agent, otherwise not. Since the con-

nections in the continuum can be inter-layer or intra-

layer, the edge set E can be rewritten as

E = E

∪ E

e f

∪ E

f f

∪ E

f c

(1)

where E

and E

f f

stand for the set of the connec-

tions in edge and fog computing paradigm. E

e f

f c

)

is the set of connections between the edge layer and

fog layer (fog layer and cloud layer). It can be ob-

served from (1) that there are no connections between

the cloud servers because they manage their resources

in centralized manner. There is also no direct connec-

tions between the edge to the cloud layer. Then we

can deﬁne the neighbor set of the resource agent v

= {v

∈ V : (v

) ∈ E}. (2)

The available resources in v

are denoted as

= (R

CPU

MEM

), where R

CPU

and R

MEM

repre-

sent the available CPU and MEM resources. The de-

mand resource for the k

pod is denoted by D

CPU

MEM

). When the k

pod arrives at the re-

source agent v

, this MIRTO agent will provide infor-

mation about the resource available to the pod. If

CPU

≤ R

CPU

and D

MEM

≤ R

MEM

(3)

there are enough resources that can be assigned and

pods can ﬁnish execution here. Otherwise, the pod

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

312

can receive the available resource information from

the neighbor set N

, chooses one, and moves to it.

Each pod constantly rechecks the available resources

until it ﬁnds the resource agent with adequate re-

sources. Once pod k processes in the MIRTO agent

, the available resources in v

will be reduced to

− D

until pod ﬁnishes execution. After process-

ing, the resources are released again.

3.2 Main Objectives

This section focuses on developing the workload

scheduler for the edge-fog-cloud continuum. The

main objective for each pod is to ﬁnd the resource

agent to meet its demand with the least latency. To

evaluate the performance of the workload scheduler,

we introduce three metrics as key performance indi-

cators (KPI): resource utilization (RU), dependency

on fog/cloud (DF/DC) and communication cost (CC).

The RU of CPU/MEM for v

is deﬁned as

CPU

− R

CPU

, RU

MEM

− R

MEM

(4)

where

CPU

and

MEM

are the capability of CPU and

MEM resources of v

CPU

− R

CPU

and

MEM

−

MEM

are the CPU and MEM resource used by v

These resources are assigned to pods for execution

and will be released when pods leave the continuum.

Assume that there are Q pods arriving in the con-

tinuum. For the k

arriving pod, denote the visited

resource agent set as M

= {v

,··· ,v

} ⊆ V . Let I

and I

be the indicators to reﬂect whether the k

pod

visited MFA and MCA. They are shown as follows



1, if M

∩ V

̸= ∅

0, if M

∩ V

= ∅

(5)

and



1, if M

∩ V

̸= ∅

0, if M

∩ V

= ∅.

(6)

Hence, DF and DC are deﬁned as

DF =

∑

k=1

, DC =

∑

k=1

. (7)

The path of the k

pod is deﬁned as

= {(v

),··· ,(v

m−1

)} ⊆ E. Let c

i j

be the

CC to transmit pod from resource agent v

to resource

agent v

, thus, the average CC for Q pods that arrive

in the continuum is deﬁned as

avg

∑

k=1

∑

)∈P

i j

(8)

where

∑

)∈P

i j

denotes the CC of the k

pod

(CC

). In summary, the ultimate goal of this work

is to design a workload scheduler with high local RU,

low DF/DF, and low CC.

4 METHODOLOGY

In this Section, we will introduce the swarm intelli-

gence workload placement inspired by the ACO al-

gorithm, where we map the pods to the role of ants

that can release pheromones on each resource agent.

Therefore, we deﬁne the pheromone value on the re-

source agent v

as τ

. The initial pheromone value

on each resource agent depends on the layer they are

located at. We design the update scheme τ

on the

MIRTO agent v

:= τ

+ γ∆τ

(9)

if the k

pod ﬁnished execution on v

. ∆τ

is the re-

leased pheromone value by pod k on v

. This value is

related to the demand of the k

pod. γ is a parameter

which represents the pheromone deposit factor.

In contrast, if the k

pod arrives at v

, but cannot

be served here, we introduce the evaporation for the

pheromone value on v

. This can be regarded as the

reduction of the pheromone value if the pods are re-

jected to be served. The updated rule is designed as

:= (1 − ρ

)τ

(10)

where ρ

denotes the evaporation rate which is also

related to the demand of the k

pod. Then, the pod

should decide where to go next from the neighbor set

. It can be observed from (9) and (10) that the

pheromone values reveal the available resources on

each MIRTO agent. When it comes to how to select

the neighbors, we ﬁrst implement two baseline algo-

rithms (random and greedy). After that, we design a

swarm intelligence algorithm for the selection based

on the ACO method to make a decision for the next

MIRTO agent from the set N

. The algorithms are

presented as follows.

• Random. The pods randomly choose a resource

agent from the neighbor set N

and check the

available resource again until they ﬁnd the re-

source agent meeting their demand. Although the

random algorithm exhibits excellent scalability,

often requires less computational resources and

can be easier to implement, it is hard to ﬁnd the

appropriate resource agent at an appropriate time

in many cases.

• Greedy. The pods move to a resource agent with

the maximum pheromone value from the neigh-

bor set N

(if there are multiple MIRTO agents

Swarm Intelligence-Based Algorithm for Workload Placement in Edge-Fog-Cloud Continuum

313

MEA 𝑣

MFA 𝑣

MCA 𝑣

D=(16,16)

R=(8,4)

R=(16,8)

R=(32,32)

R=(8,64)

R=(2,4)

R=(8,2)

R=(32,4)

ACO

Random

Greedy

Figure 2: Three possible paths of a pod to ﬁnd a suit-

able resource agent in the continuum under random (green),

greedy (blue), and ACO (red) algorithms.

with the same maximum pheromone value, one is

randomly chosen). However, privacy issues may

not be guaranteed since the pods are inclined to go

to the higher layer (edge to fog, or fog to cloud)

under the greedy algorithm. Additionally, the

inter-layer transmission cost is much higher than

the intra-layer, which is not suitable for latency-

sensitive tasks.

• ACO. As the transmission cost is also an impor-

tant factor for workload management, it should be

taken into consideration when pods make their de-

cision. Motivated by the ACO method, the pods

move to a resource agent v

∈ N

with the proba-

bility p

deﬁned as follows

i j

∑

∈N

i j

(11)

where η

i j

= 1/c

i j

denotes the heuristic desirabil-

ity which is the inverse of the transmission cost

between v

and v

. α and β are two weighting pa-

rameters, one for the pheromone value, the other

for the transmission cost. It can be observed from

(11) that pods are favorable to move to the neigh-

bor with a large number of available resources

and low communication costs. Hence, more pods

can be served locally on the edge layer instead of

moving to the fog or cloud layer. Algorithm 1

summarizes the behavior of pods by implement-

ing the ACO-based workload scheduling.

Figure 2 illustrates three possible paths of a pod n

with an exemplary demand of D

= (D

CPU

MEM

) =

(16,16) to ﬁnd the resource agent. The green path

is generated under the random algorithm. Since

the neighbor is selected randomly, the pod visits six

MIRTO agents without enough available resources.

Finally, it is transmitted to the cloud server. The blue

path stands for the greedy algorithm, the pod moves

directly to the MFA v

and then goes to the cloud

server. The red one represents the most likely path

Algorithm 1: ACO Based Workload Placement.

Input: Pods enter the continuum

Output: The path to the best resource agent

1 Initialization phrase;

2 for k ← 1 to Q do

3 Start randomly from a MEA v

;

4 Check the available resources ;

5 while Resource condition (3) is not

satisﬁed do

6 Update the pheromone value by (10) ;

7 Choose a neighbor based on (11);

8 Move to the selected resource agent;

9 Recheck the available resources.

10 Finishing execution;

11 Update the pheromone value by (9);

12 Leave the continuum.

Figure 3: Resource utilization (RU) on edge layer of the

three algorithms with µ = 0.5 (a) and µ = 0.8 (b).

of the pod under ACO algorithm. It can be served on

fog layer without visiting too many MIRTO agents.

5 SIMULATION CASE STUDIES

In this Section, the edge-fog-cloud continuum is sim-

ulated using the MESA library (Masad et al., 2015) in

a Python environment. We ﬁrst set up the parameters

in the continuum. After that, the analysis and discus-

sion of the simulation results are presented to com-

pare the ACO algorithm with two baseline algorithms

in three aspects (RU, DF/DC, and CC), as mentioned

in Section 3.2.

5.1 Setup of the Continuum

In this simulation framework, we exemplary consider

ten MEA, ﬁve MFA, and two MCA in the contin-

uum. The capacity is deﬁned as 16 or 32 CPU/MEM

resource units for each MEA. MFA is comprised of

a set of EMDC with a capacity equal to 64 or 128

CPU/MEM resource units, and the capacity MCA is

assumed unlimited. Their connections are generated

randomly by NetworkX (Platt, 2019) to ensure that

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

314

Figure 4: Dependency on fog (a) and cloud (b) (DF/DC) of the three algorithms with µ = 0.8, and box-plot of DF (c) and DC

(d) with different µ under three algorithms in 50 simulations.

the continuum is connected according to our model

described in Section 3. The communication cost c

i j

deﬁned as follows

i j











1, if i, j ∈ V

2, if i, j ∈ V

5, if i ∈ V

, j ∈ V

30, if i ∈ V

, j ∈ V

(12)

The initial pheromone values τ

for the MEA are

deﬁned as τ

= 4, if the capacity is (16,16), τ

= 6 if

the capacity is (32,32), and τ

= 5 otherwise. Simi-

larly, for the MFA, τ

= 12, if the capacity is (64,64),

= 18 if the capacity is (128,128), and τ

= 15 oth-

erwise. The initial pheromone values on MCA equal

to 50 and will not be updated during the process.

The demand agents are represented by pods in the

Kubernetes context. Small (Medium) pods require 1

or 2 (4 or 6) CPU/MEM resource units with the ex-

ecution steps distributed from 20 to 30 (50 to 130)

uniformly and randomly. Their frequency of arrival is

0.4. For large pods, they need 8 or 16 CPU/MEM re-

source units with the execution steps distributed from

200 to 400 and the frequency of arrival is 0.2. Table 1

lists all the features of the arriving pods.

The parameters in the algorithms are deﬁned

based on following criteria: As γ indicates the proba-

bility of considering pheromone values for a resource

Table 1: Pod proﬁles.

CPU/MEM Demand Execution Steps

Small 1 or 2 / 1 or 2 [20,30]

Medium 4 or 6 / 4 or 6 [50,130]

Large 8 or 16 / 8 or 16 [200,400]

allocation decision, we set γ = 1 on edge layer, γ = 0.6

on fog layer, and γ = 0 on cloud servers. As we

want to achieve a low dependency on the cloud, the

values are chosen in descending order. The released

pheromone ∆τ

and ρ

are designed related to the de-

mand of the pods. We set ∆τ

as 1, 2, and 3 for small,

medium, and large pods. ρ

is set as 0.15, 0.1, and

0.05 for small, medium, and large pods. The weight

parameters are designed as α = 1.4 and β = 1. We set

the iteration steps to 2,000 for each simulation.

5.2 Results and Discussion

Figure 3 demonstrates the resource utilization (RU)

on the edge layer ﬁrst, as lightly loaded edge-fog-

cloud continuum with µ = 0.5 and second, a heav-

ily loaded one using µ = 0.8. It can be obtained that

RU on the edge layer under the ACO algorithm is

increased compared with the other two baseline al-

gorithms. More pods can be served locally instead

of moving to higher layers under ACO algorithm for

Swarm Intelligence-Based Algorithm for Workload Placement in Edge-Fog-Cloud Continuum

315

Figure 5: Communication costs (CC) of the three algorithms with µ = 0.8 (a), and box-plot of CC with different µ in 50

simulations (b).

both lightly and heavily loaded continuums.

Figure 4 (a) and (b) illustrate the dependency on

fog and cloud layer (DF/DC) with µ = 0.8. We

can observe that DF/DC under the ACO algorithm is

lower than another algorithm after 2,000 steps if the

load is heavy. Moreover, we explore the DF/DC with

different µ. The results are shown by box-plot in Fig-

ure 4 (c) and (d) by running 50 simulations. It is no-

ticeable that the greedy algorithm relies much more

on the cloud (as it has much more resources available

and the algorithm goes always for the highest avail-

ability). The DF/DC under the ACO algorithm keeps

the lowest for the different loaded continuums.

Figure 5 (a) reﬂects the communication costs (CC)

with µ = 0.8. The baseline algorithms are still higher

than the ACO algorithm as the iteration steps increase.

Moreover, the CC with different µ is presented in Fig-

ure 5 (b) in 50 simulations. ACO algorithms can also

save the CC compared with the baseline algorithms.

To sum up, compared with the two baseline algo-

rithms, the ACO algorithm increases the local RU and

decreases the DF/DC. This is because the scheduler

based on the ACO algorithm considers both available

resources and transmission costs. The pods are in-

clined to go to the matched MIRTO agent without too

many costs. Implementing the ACO algorithm also

reduces the CC. In conclusion, the ACO algorithm

demonstrates improved performance in optimizing lo-

cal resource utilization and minimizing data trans-

fer and computational costs across the edge-fog-cloud

continuum. This approach not only enhances the ef-

ﬁciency of workload distribution but also addresses

critical concerns related to latency and privacy, mak-

ing it particularly suitable for applications with strin-

gent requirements in these areas, where computing on

the edge is favored over cloud computing.

6 CONCLUSION

This paper investigates the workload placement prob-

lem for the edge-fog-cloud continuum. A multi-

agent framework is established to model the contin-

uum with arriving pods as demand agents and contin-

uum nodes as resource agents. We design a swarm-

based workload scheduler to place each demand agent

at a proper resource agent considering the available

resources and communication cost. The proposed al-

gorithm is inspired by ACO, where pods mimic the

behavior of ants by laying down pheromones on each

resource agent to reveal the available resources. The

probability of the pods moving to neighbors is also

deﬁned based on the pheromone value and commu-

nication costs. Compared with random and greedy

algorithms, numerical studies illustrate that the work-

load scheduler based on swarm intelligence shows

higher local resource utilization with lower depen-

dency on fog and cloud. Future works will try to mod-

ify the algorithm to consider more factors, such as pri-

vacy issues. Moreover, other approaches, such as evo-

lutionary algorithms and approaches from reinforce-

ment learning, will be explored to design adaptable

rules for distributed resource allocation in dynamic

edge-fog-cloud continuum.

ACKNOWLEDGEMENT

Funded by the European Union, project MYRTUS, by

grant No. 101135183. Views and opinions expressed

are however those of the author(s) only and do not

necessarily reﬂect those of the European Union. Nei-

ther the European Union nor the granting authority

can be held responsible for them.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

316

REFERENCES

Attiya, I., Abd Elaziz, M., Abualigah, L., Nguyen, T. N.,

and Abd El-Latif, A. A. (2022). An improved hybrid

swarm intelligence for scheduling iot application tasks

in the cloud. IEEE Transactions on Industrial Infor-

matics, 18(9):6264–6272.

Bai, X., Liu, D., and Xu, X. (2024). A review of improved

methods for ant colony optimization in path planning.

Journal of Ship Research, pages 1–16.

Bonomi, F., Milito, R., Zhu, J., and Addepalli, S. (2012).

Fog computing and its role in the internet of things. In

Proceedings of the 1st Edition of the MCC workshop

on Mobile Cloud Computing, pages 13–16.

Chandrashekar, C., Krishnadoss, P., Kedalu Poornachary,

V., Ananthakrishnan, B., and Rangasamy, K. (2023).

Hwacoa scheduler: Hybrid weighted ant colony opti-

mization algorithm for task scheduling in cloud com-

puting. Applied Sciences, 13(6):3433.

Chen, Y., Zhao, J., Hu, J., Wan, S., and Huang, J. Dis-

tributed task ofﬂoading and resource purchasing in

noma-enabled mobile edge computing: Hierarchical

game theoretical approaches. ACM Transactions on

Embedded Computing Systems, 23(1):1–28.

Corominas, G. R., Blesa, M. J., and Blum, C. (2023).

Antnetalign: Ant colony optimization for network

alignment. Applied Soft Computing, 132:109832.

Dorigo, M., Birattari, M., and Stutzle, T. (2006). Ant colony

optimization. IEEE Computational Intelligence Mag-

azine, 1(4):28–39.

Dorigo, M. and St

utzle, T. (2003). The ant colony opti-

mization metaheuristic: Algorithms, applications, and

advances. Handbook of Metaheuristics, pages 250–

285.

Du, X., Du, C., Chen, J., and Liu, Y. (2023). An energy-

aware resource allocation method for avionics systems

based on improved ant colony optimization algorithm.

Computers and Electrical Engineering, 105:108515.

Ghasemi, A. and Schranz, M. (2024). Bottom-up re-

source orchestration in edge computing: An agent-

based modeling approach. In 12th IEEE International

Conference on Intelligent Systems, pages 1–7.

Gkonis, P., Giannopoulos, A., Trakadas, P., Masip-Bruin,

X., and D’Andria, F. (2023). A survey on iot-edge-

cloud continuum systems: status, challenges, use

cases, and open issues. Future Internet, 15(12):383.

Godsil, C. and Royle, G. F. (2013). Algebraic graph theory,

volume 207. Springer Science & Business Media.

Hasan, M. K., Jahan, N., Nazri, M. Z. A., Islam, S., Khan,

M. A., Alzahrani, A. I., Alalwan, N., and Nam, Y.

(2024). Federated learning for computational ofﬂoad-

ing and resource management of vehicular edge com-

puting in 6g-v2x network. IEEE Transactions on Con-

sumer Electronics, 70(1):3827–3847.

Jiao, L., Friedman, R., Fu, X., Secci, S., Smoreda, Z., and

Tschofenig, H. (2013). Cloud-based computation of-

ﬂoading for mobile devices: State of the art, chal-

lenges and opportunities. 2013 Future Network & Mo-

bile Summit, pages 1–11.

Kim, E., Lee, K., and Yoo, C. (2021). On the resource man-

agement of kubernetes. In 2021 International Confer-

ence on Information Networking, pages 154–158.

Masad, D., Kazil, J. L., et al. (2015). Mesa: An agent-based

modeling framework. In SciPy, pages 51–58. Citeseer.

Palumbo, F., Zedda, M. K., Fanni, T., Bagnato, A., Castello,

L., Castrillon, J., Ponte, R. D., Deng, Y., Driessen, B.,

Fadda, M., et al. (2024). Myrtus: Multi-layer 360 dy-

namic orchestration and interoperable design environ-

ment for compute-continuum systems. In Proceedings

of the 21st ACM International Conference on Comput-

ing Frontiers: Workshops and Special Sessions, pages

101–106.

Parikh, S., Dave, D., Patel, R., and Doshi, N. (2019). Secu-

rity and privacy issues in cloud, fog and edge comput-

ing. Procedia Computer Science, 160:734–739.

Pereira, P., Melo, C., Araujo, J., Dantas, J., Santos, V., and

Maciel, P. (2022). Availability model for edge-fog-

cloud continuum: an evaluation of an end-to-end in-

frastructure of intelligent trafﬁc management service.

The Journal of Supercomputing, pages 1–28.

Platt, E. L. (2019). Network science with Python and Net-

workX quick start guide: explore and visualize net-

work data effectively. Packt Publishing Ltd.

Rajesh, G., Mercilin Raajini, X., Ashoka Rajan, R.,

Gokuldhev, M., and Swetha, C. (2020). A multi-

objective routing optimization using swarm intelli-

gence in iot networks. In In Proceedings of the In-

telligent Computing and Innovation on Data Science

Conference, pages 603–613. Springer.

Robles-Enciso, A. and Skarmeta, A. F. (2024). Adapting

containerized workloads for the continuum comput-

ing. IEEE Access, 12:104102–104114.

Saba, T., Rehman, A., Haseeb, K., Alam, T., and Jeon, G.

(2023). Cloud-edge load balancing distributed proto-

col for ioe services using swarm intelligence. Cluster

Computing, 26(5):2921–2931.

Satyanarayanan, M. (2017). The emergence of edge com-

puting. Computer, 50(1):30–39.

Schranz, M., Di Caro, G. A., Schmickl, T., Elmenreich,

W., Arvin, F., S¸ekercio

glu, A., and Sende, M. (2021).

Swarm intelligence and cyber-physical systems: con-

cepts, challenges and future trends. Swarm and Evo-

lutionary Computation, 60:100762.

Shi, W., Cao, J., Zhang, Q., Li, Y., and Xu, L. (2016). Edge

computing: Vision and challenges. IEEE Internet of

Things Journal, 3(5):637–646.

Simi

c, M., Proki

c, I., Dedei

c, J., Sladi

c, G., and Milosavl-

jevi

c, B. (2021). Towards edge computing as a ser-

vice: Dynamic formation of the micro data-centers.

IEEE Access, 9:114468–114484.

Tong, Z., Deng, X., Ye, F., Basodi, S., Xiao, X., and Pan,

Y. (2020). Adaptive computation ofﬂoading and re-

source allocation strategy in a mobile edge computing

environment. Information Sciences, 537:116–131.

Wu, L., Huang, X., Cui, J., Liu, C., and Xiao, W. (2023).

Modiﬁed adaptive ant colony optimization algorithm

and its application for solving path planning of mobile

robot. Expert Systems with Applications, 215:119410.

Swarm Intelligence-Based Algorithm for Workload Placement in Edge-Fog-Cloud Continuum

317