Enhanced Particle Swarm Optimisation and Multi Objective

Optimization for the Orchestration of Edge Cloud Clusters

Haﬁz Faheem Shahid and Claus Pahl

Free University of Bozen-Bolzano, Bolzano, Italy

Keywords:

Swarm Intelligence, Particle Swarm Optimization, Distributed Systems, Load Balancing, Edge Cloud Cluster.

Abstract:

Load balancing and workload distribution cause challenges for the management of IoT and distributed systems

in the edge computing environment. Swarm intelligence is a technology suitable for the management of dis-

tributed systems, networks, communication and routing protocols. Swarm intelligence-based PSO algorithms

(particle swarm optimization) can be applied for load balancing and task scheduling in cloud computing en-

vironments operating through a broker agent. In distributed cloud environments, data is collected and then

processed at the center of the cloud, rather than making decision at edge nodes closer to IoT infrastructures.

Here, we develop an automated orchestration technique for clustered cloud architectures. An Autonomous

Particle Swarm Optimization, called the A-PSO algorithm, is implemented that enables an edge node, such as

a remote storage, to work as part of a decentralized, self-adaptive intelligent task scheduling and load balancing

agant between resources in distributed systems. Using Multi Objective Optimization (MOO), complementing

the A-PSO algorithm, we also include metrics such as Actual Round-Trip Time (ARTT) of tasks assignments

to the remote storage to reduce the execution cost. Our A-PSO algorithm can orchestrate the distribution of

large volumes of data to remote storage and back in cluster, i.e., coordinated distributed cloud environments.

1 INTRODUCTION

Task scheduling and allocation of tasks to cloud re-

sources is a load balancing problem that needs to be

optimized in cloud environments. Resource manage-

ment and load balancing are an important part of any

cloud computing environment in a distributed context.

With the advancement of hardware and software ap-

plication, a remaining problem is the optimization of

distributing (i.e., storing and fetching) data in cloud

clusters (1; 1; 1). The problem we are discussing here

is to balance the load between the remote data stor-

ages as a sample type of a cloud resource. We can of-

ten experience that with limited resources of remote

data storage in edge cloud applications, sometimes

applications show performance degradations (1) re-

sulting from may uncertainties in the environment (1).

This can happen when the number of transactions in-

creases dramatically, possibly causing downtime.

In recent years, swarm intelligence algorithms

have been widely used, speciﬁcally using swarm in-

telligence techniques for the development of dis-

tributed systems, networks, communications and

routing protocols. The two basic swarm intelli-

gence algorithms, that are used for the development

of collective behavior of applications and software

are Stochastic Diffusion Search (SDS) and Particle

Swarm Optimization (PSO). SDS is often seen as

representing the collective behavior of social insects

(such as ants) and PSO the collective behavior of so-

cial animals (birds and ﬁsh) (2). In SDS and PSO

algorithms, social interaction and exchange of infor-

mation is the signiﬁcant key role that is practiced by

the social animal and social insects in swarm intelli-

gence. A basic strategy behind these two algorithms

of SDS and PSO is to work in a swarm to achieve the

target instead of working individually (3).

The aim of this paper is to explore the suitabil-

ity of bio-inspired algorithms for autonomous cloud

resource management. The solution shall be based

on applying a swarm intelligence approach to coordi-

nate workloads between cloud resource clusters and

to manage them in dispersed settings. Swarm intelli-

gence as a form of distributed intelligence is about im-

plementing algorithms for distributed systems man-

agement using Particle Swarm Optimization (PSO) or

Ant Colony Optimization (ACO), and adapt them to

schedule task fairly to the cloud resources (storage in

our case) and balance the load correspondingly be-

tween them (16). Cloud technology is moving to-

Shahid, H. and Pahl, C.

Enhanced Particle Swarm Optimisation and Multi Objective Optimization for the Orchestration of Edge Cloud Clusters.

DOI: 10.5220/0008019201550162

In Proceedings of the 11th International Joint Conference on Computational Intelligence (IJCCI 2019), pages 155-162

ISBN: 978-989-758-384-1

155

wards multi-cloud environments with the inclusion

of various devices, data storages and sensors at the

edges (0). This requires the integration of data cen-

ter technologies with much more constrained devices,

but still using virtualized solutions to deal with scala-

bility, ﬂexibility and multi-tenancy concerns.

We aim at an automated orchestration technique

for edge cloud PaaS (Platform as a Service) architec-

tures based on workloads that are deployed in several

local edge clusters. For edge clouds, application and

service orchestration can help to manage and orches-

trate applications through containers (17). In this way,

computation can be brought to the edge through an

orchestration technique, rather than transferring large

volumes of data from the edge to the cloud.

The paper is organised as follows. We start with

background and related work in Section 2. Section 3

introduces our extended PSO solution in the context

of related work. Section 4 describes the experimental

evaluation. In Section 5, we discuss the results, before

concluding in Section 6.

2 BACKGROUND

With the proliferation of internet-capable devices,

workload distribution in cloud computing is becom-

ing major problem. Here, we propose a method of

load balancing at storage resource level based on Par-

ticle Swarm Optimization (PSO) algorithm. In many

existing load balancing solutions, a broker node plays

an important role for task scheduling to the resources.

In our proposed autonomous PSO algorithm, there is

no central broker node to assign and executes to the

virtual machines. The data computation is held in be-

tween the edge nodes (resource hosts) instead of send-

ing information for decisions to the central broker or

controller node of the cloud computing environment.

We focus on Particle Swarm Optimization (PSO)

for the development of our orchestration solution.

PSO is inspired by the social behavior of birds ﬂock-

ing and ﬁsh schooling (4), (5). A particle in PSO re-

ﬂects to a bird or ﬁsh that moves from one position to

another to reach its best position to ﬁnd for instance

food and a safe place. The movement of the particle is

determined by the velocity, which has both direction

and magnitude. At any instance of time the position

of the particle is inﬂuenced by its best position (pbest)

and the position of the best particle in entire swarm

(gbest). To calculate the performance of each particle

in the swarm, we must apply an objective function.

The velocity of ﬂocking birds is continuously chang-

ing and thus the positions of birds in swarm will also

need to change continuously (6), (7), (8), (9), (10).

In order to provide some background, we intro-

duce two basic equations of PSO algorithms that up-

date the velocity and position of a swarm particle:

k+1

= wv

+C1r1x(pbest

− x

) +C2r2x(gbest − x

)

(1)

k+1

= x

+ v

k+1

(2)

PSO algorithm (1) describes the updated velocity

of the particle and equation (2) describes the updated

position of that particle (18). The parameters of the

equations are summarised in Table 1.

Table 1: PSO – Parameters.

pbest

best position of particle i

gbest position of best particle in swarm

current velocity of particle i at k iteration

k+1

velocity of particle i at k+1 iteration

current position of particle i at k iteration

k+1

position of particle i at k+1 iteration

w inertia

c1,c2 acceleration coefﬁcient factor

r1,r2 random number between 0 to 1

3 PSO+MOO LOAD BALANCING

Our overall solution combines Particle Swarm Op-

timization (PSO) and Multi Objective Optimization

(MOO), which we will introduce in sequence.

3.1 Load Balancing with PSO

We now map the PSO algorithm to our problem of

load balancing in edge clouds as follows:

• we choose the cloud storage resources (RES) as

particles;

• we deﬁne the velocity v

of particle i at iteration k

as change of load below; and

• we interpret the direction d as the number of stor-

age resources connected.

We consider in our adapted PSO algorithm the cur-

rent load CL

on the RES

at iteration k for our load

balancing problem. The load on the RES

will up-

date to CL

k+1

at iteration k + 1, after every iteration

of incoming tasks, see Equation (3). The objective

function in Equation (4) calculates the ﬁtness value

of each RES to get Global best (gbest) and Personal

best (pbest) values of the RESs. In load balancing,

we use the objective function min f

min

(x) to obtain

the best minimum ﬁtness value of the resource.

k+1

= 1(pbest) + (CL

–pbest) (3)

ECTA 2019 - 11th International Conference on Evolutionary Computation Theory and Applications

156

min

k+1

) = (CL

k+1

/MaxStorage

) × d (4)

After each Iteration, the load on the resources RES

is updated by Equation (3). Incoming tasks to the next

iteration could be different in size and varying in their

numbers. We can set the limit of incoming tasks be-

tween any two numbers. After updating the load on

resources, each RES resource needs to change its po-

sition in a queue for accepting the incoming load. In

Equation (4), CL

k+1

is the current load on the cloud

storage resournce i in any iteration k+1. MaxStorage

is the maximum storage capacity or upper limit of

storing data at the same iteration, where d denotes the

direction of the particle.

We interpret the direction d as the number of stor-

age resources connected under control of the PSO al-

gorithm. Thus, for our load balancing, we use the

min

k+1

) objective function, because we need to ﬁnd

the best resource (data storage) with the minimum

memory load in the entire swarm of resources RES

(19). The cloud storage resource that has the mini-

mum ﬁtness value in the cluster of RESs is the global

best position of RES in the cloud.

In order to illustrate this, we use an example of

three cloud resources that provide 3 different storage

services with a maximum memory capacity of 100,

150, or 200 MBs in them. After 1, 2 and 3 iterations,

we get the current load on each RES as follows:

Cloud Storage Resource 1 (RES 1):

1. CLi

k+1

: current load on RES 1 = 25 MBs

2. MaxStorage

: max capacity on RES 1 = 100 MBs

which we add to the objective function (4) as follows:

min

(RES 1) = (CL

k+1

/MaxStorage

) d

min

(RES 1) = (25/100) 3

with d as the number of resources in the RES swarm:

(RES 1) = 0.75

Cloud Storage Resource 2 (RES 2):

1. CL

k+1

: current load on RES 2 = 47 MBs

2. MaxStorage

: max capacity on RES 2 = 150 MBs

which we add to the objective function (4) as follows:

min

(RES 2) = (CL

k+1

/MaxStorage

) d

min

(RES 2) = (47/150) 3

with d as above and:

min

(RES 2) = 0.94

Cloud Storage Resource 3 (RES 3):

1. CL

k+1

: current load on RES 3 = 39 MBs

2. MaxStorage

: max capacity on RES 3 = 200 MBs

which we add to the objective function (4) as follows:

min

(RES 3) = (CL

k+1

/MaxStorage

) d

min

(RES 3) = (51/200) 3

with d as above and:

min

(RES 3) = 0.76

After this, we have obtained the personal best values

for the 3 resources, which are 0.75, 0.94 and 0.76 re-

spectively. The global best for the entire swarm is

0.75, which corresponds to resource 1.

After each iteration of incoming load on re-

sources, the objective function (4) calculates pbest

values for each RES. As we are using the objec-

tive minimum function f

min

k+1

) that reﬂects the re-

source with minimum load on it among all resources.

This is said to be the gbest resource of the swarm.

Table 2: A-PSO Load Balancing Algorithm – Parameters.

MaxStorage

: total storage capacity on RESi

Dimension: d number of RESs

pbest

personal best position of RESi at iteration k

gbest

global best position of RESi at iteration k

current load on RESi at iteration k

(initial velocity of particle i in PSO)

k+1

updated load on RESi at iteration k +1

(updated velocity of particle i in PSO)

current position of RESi at iteration k

k+1

updated position of RESi at iteration k +1

After updating the load on RES i, the position of

the resource at iteration k,x

will also change. The

gbest RES then moves forward in the priority queue

and lead the rest of the RESs by means of balancing

load. The positions of the RESs in the cloud will be

updated by the load of incoming tasks at iteration k +

1. Then x

k+1

is the position of RES

at iteration k + 1.

When a resource reaches a threshold load of 60%

or higher, then it shall be made temporarily unavail-

able for further tasks. In that case, the tasks in the

incoming task queue will need to be dealt with by

other resources in the swarm that are still in a work-

able state. A backup resource shall also be provided in

case all primary resources have reached their thresh-

old load of 60%.

3.2 PSO Load Balancing with MOO

As an enhancement of the A-PSO algorithm, we com-

bine our load balancing PSO solution with multi ob-

jective optimization. When we have a higher num-

ber of iterations for the incoming load for resources,

we naturally observe that task execution takes longer.

For the management of distributed systems, our A-

PSO algorithms does distribute load equally to re-

sources and balances the load between edge nodes

correspondingly, but on the other hand its cost for task

Enhanced Particle Swarm Optimisation and Multi Objective Optimization for the Orchestration of Edge Cloud Clusters

157

assignment cycles per seconds becomes higher. The

challenges we are facing is to minimize the execu-

tion cost along with balancing load for the distributed

management in the edge cloud environment. To over-

come the problems, we decided to apply Multi Ob-

jective Optimization (MOO) (13) in conjunction with

the PSO solution. For MOO we aim to minimize the

execution cost in task scheduling along with the A-

PSO algorithm to balance the load. Time costing is

based on the execution of tasks as cycles per millisec-

ond. At every iteration with new tasks, the execution

of tasks at the resource can be calculated in millisec-

onds. A single cycle is assumed to consist here of

100 milliseconds as a default, i.e., 200 milliseconds

to execute a storage request corresponds to 2 cycles.

For each iteration we know the execution cost for a

swarm resource. For the next iteration, we can reduce

the execution cost by assigning the task to the virtual

machine that takes the minimum number of cycles to

execute. We prioritize resource requests dealt with

by A-PSO additionally with respect to the minimum

number of cycles they take during every iteration.

Table 3: Variables for Time Costing in Load Balancing.

N: Number of tasks

M: Number of virtual machines

EET : Expected Execution Time

ET T : Expected Transmission Time

ART T : Actual Round-Trip Time

Load

: Load on RES

i j

: Decision to allocate task i to RES

or not [0,1]

In order to reduce the number of cycles of task

execution at the resources, we adopt the time costing

deﬁnitions (5), (6) and (7) (11).

Min

cost

∑

.EET

i j

∗ x

i j

(5)

Min

cost

∑

.ET T

i j

∗ x

i j

(6)

ART T

i j

= EET

i j

+ ET T

i j

(7)

In the following, we consider cost with respect

to the Expected Execution Time (EET), Expected

Transmission Time (ET R) and Expected Round-Trip

Time (ERT T ) factors.

• Expected Execution Time (EET ) is deﬁned as

EET

i j

of Task

on resource RES

denoting the

time of a task from initialization to the eventual

termination at the execution point (11). Our goal

is to minimize the EET to reduce the time costing

in the load balancing process (12).

• Expected Transmission Time (ET T ): ET T

i j

Task

on resource RES

is the time from the as-

signment of the task to the resource to the com-

plete transmission of the actual task (6).

• Actual Round-Trip Time (ART T ): we can cal-

culate the ART T

i j

of Task

on resource RES

equation (7). We sum up the time of execution

EET

i j

with the transmission time ET T

i j

to min-

imize the cost of load balancing (7) (13). Divid-

ing the ART T by 100, we get the number of exe-

cution cycles per milliseconds. The number cy-

cles that each task management consumes indi-

cates the cost of the load balancing.

Task scheduling and load balancing is calculated with

respect to the execution cycle of tasks in milliseconds:

Execution Cycle = ART T

i j

/100

which deﬁnes the Actual Round-Trip Time of tasks

to RES

, i.e., ART T

i j

The prioritization of the resources based on the

time consumption considers:

• the current load on the storage resource (minimum

value of ﬁtness by object function Equation (4) ),

• that the resource takes minimum cycle per mil-

liseconds to execute tasks based on Equation (7).

We set each iteration weight for an automated solution

of PSO for task scheduling and load balancing from 0

to 10. 0 is the minimum number and 10 is the max-

imum weight of each iteration. If the weight of the

iteration is between 0-7 (random value generated at

each iteration between 0-10), then the system assigns

tasks to storage resources with respect to the ﬁtness

value of the resources. A global best RES(gbest) will

be one with the minimum f .v following equation (4).

If the iteration weight is between 8-10, then the so-

lution assigns the incoming task to the resource that

takes the minimum number of cycles to execute them.

In this case, the global best RES(gbest) is the one that

costs the minimum number of cycles to execute.

Below, we present the enhanced A-PSO algorithm

for task scheduling and load balancing.

————————————————————–

A-PSO Load Balancing and Scheduling

1. For all tasks (ti) ∈ T do

Determine incoming tasks

Determine size of each task

2. End for

3. For all resources (RESi) ∈ RES do

Determine current load on RESi by Eq. (4)

//Velocity of particle i in PSO algorithm

Determine real load on RESi

//Max storage of RES

4. End for

5. Initialize RES

ECTA 2019 - 11th International Conference on Evolutionary Computation Theory and Applications

158

6. While (current load on RESi < 60% ) stop condition

7. For all resources (RESi) do

Calculate ﬁtness value of each RES by using

objective function, Eq. (4)

(current f .vi > pbest

)

pbest

= current f .v

End if

(pbest

< ∀ RESs personalbest)

gbest = pbest

End if

8. End for

9. For all resources RES do

Calculate execution time cost (7) in msec

Calculate task assignment to RES in msec

10. End for

11. Prioritize all tasks according to size

12. Prioritize all RESs with respect to minimum f .v(4) ||

Prioritize all RES according to min task exec cy-

cles/msec

13. For all Prioritized tasks (t

) ∈ T do

For all Prioritized resources (RESj) ∈ RES do

If (iteration weight between 0-7 &&

current load in RES

<= 60%)

assign task(t

) to resource (RES

)

remove task from sorted tasks queue

current load RES

+ +

End if

Else If (iteration weight between 8-10 &

current load in RES

<= 60%)

assign task (t

) to resource (RES

)

remove task from sorted task queue

current load RES

+ +

End if

Else

Break // (assign next task (t

) in queue

to next (RES

))

End for

14. End for

15. Update current load on RESs // Equation (3)

16. Update position of RES // objective function (4)

17. End for

—————————————————————-

We distinguish non-prioritized requests of incom-

ing tasks with different sizes that will receive a cloud

resource storage position after updating the overall

load, and prioritized ones. The latter are incoming

tasks sorted with respect size in descending order for

execution. Also, resources are sorted wrt. their posi-

tion, considering minimum ﬁtness value in descend-

ing order for receiving incoming task load and also

minimum cycles to execute incoming tasks.

4 IMPLEMENTATION

In order to implement an autonomous solution for a

decentralised architecture, we need to avoid an often

used central broker node from the system that assigns

the task to remote storage. For instance, many tools

for the experimentation of load balancing and task

scheduling exist, like CloudSim (7), (11), (14) as a

simulator. CloudSim for instance works with a cen-

tral broker node that has complete information of re-

sources as well as complete information of the tasks.

We implemented our autonomous A-PSO algorithm

in Java that works without centralised control.

In a sample experiment, we illustrate that how our

A-PSO algorithm works. For this scenario, we as-

sume three storage resources with different capacity

sizes. The tasks are also different in size. Three input

tasks are assigned to resources.

In the 1st iteration, we start with the following

sample task queue (Table 4):

Table 4: 1st Iteration – unsorted task queue.

Tasks T1 T2 T3

Size 9 6 13

The algorithm groups the incoming tasks (this is

the task information phase of general load balancing

systems) according to storage size. The task with the

maximum size will execute ﬁrst (Table 5):

Table 5: 1st Iteration – sorted task queue.

Tasks T1 T2 T3

Size 6 9 13

Now the solution assigns these tasks to the re-

sources in a way that they should balance the load us-

ing the PSO algorithm that calculates the gbest value

of the resources using the objective function in Equa-

tion (4). In a group of prioritized resources, the gbest

resource with the least load is moved to the front of

the task allocation process. The personal best posi-

tions of the resources x

are then calculated. After-

wards, we move the tasks to resources, i.e., map from

Table 5 to Table 6, which is illustrated in Fig. 1.

Table 6: 1st Iteration – initial position of storage resources.

Resource (RES) Current Load on RES

c undeﬁned

b undeﬁned

a undeﬁned

After updating the load on the resources (veloc-

ity of particles in PSO) CL

at iteration step k + 1,

the resources change their positions x

k+1

accordingly.

Here, the objective function will calculate the ﬁt-

ness value of each resource to ﬁnd the pbest

k+1

and

gbest

k+1

at iteration step k +1.

As discussed before, the task with the maximum

size will be assigned to the resource positioned at the

Enhanced Particle Swarm Optimisation and Multi Objective Optimization for the Orchestration of Edge Cloud Clusters

159

Figure 1: 1st Iteration mapping of tasks to resources.

Table 7: 1st Iteration – updated positions of resources.

Resource (RES) Current Load on RES

a 6

b 9

c 13

Figure 2: 4th Iteration mapping of tasks to resources.

top of the queue with the least load on it. Table 7

shows the updated positions of the resources x

k+1

After the 2nd and 3rd iteration, the updated load on

the resources is shown in Table 8.

Table 8: 3rd Iteration – updated load/position of resources.

Resource (RES) Current Load on RES

c 30

b 33

a 36

We continue this experiment with a fourth itera-

tion with a different size of input, but using the same

method as in iteration 1, see Tables 9 and 10.

Table 9: 4th Iteration – unsorted task queue.

Tasks T1 T2 T3

Size 13 7 11

Table 10: 4th Iteration – sorted task queue.

Tasks T1 T2 T3

Size 7 11 13

The sorted task requests are then as shown in Ta-

ble 10. Assigning tasks to resources is a mapping

Figure 3: Results – for each of the resources 1, 2 and 3 and

each of the 4 iterations, allocated task loads are shown.

from Table 10 to Table 11, illustrated in Fig. 2. Re-

sources with incoming inputs of tasks update their

load CL

k+1

and change position x

k+1

by calculating

the ﬁtness value (objective function in Equation (4)).

Table 11: 4th-Iteration – updated load/postion of resources.

Resource (RES) Current Load on RES)

b 44

a 43

c 43

Based on the resulting stage in Table 11, we can

see that after the 4th iteration the load is balanced

among all resources connected in cluster.

Fig. 3 visualizes the task load allocated to the re-

sources for all iterations and resources.

5 EXPERIMENTAL VALIDATION

This section illustrates the experiments and their re-

sults that were obtained from evaluating the imple-

mentation of the proposed method.

In order to evaluate the research questions around

the load balancing problem for particularly edge

nodes in a distributed cloud environment, there are

some parameters that need to be deﬁned. We de-

ﬁne the following key variables in the proposed au-

tonomous A-PSO algorithm along with the cost esti-

mation cycles for task scheduling and load balancing

in cloud storage resources:

• Resources RES (cloud storage): range (2-5)

• Number of tasks in each iteration: range (2-6)

• Size (memory) of tasks randomly in each itera-

tion: range (0-100)

• Number of iterations: range (10-20)

• Iteration weight: a.) position of RESs by ﬁtness

value: range (0-7); b.) position RESs by execu-

tions time cycles/ms: range (8-10)

• Max storage of RES1, RES2, RES3, RES4, RES

set to 300, 500, 200, 250, 400 MBs, resp.

ECTA 2019 - 11th International Conference on Evolutionary Computation Theory and Applications

160

Table 12: Experimental Evaluation: Test cases and Results.

Sr. RES Tasks Task No Iter- Iteration Round-Trip Time (RT T ) Autonomous-PSO Algorithm BK

Size ation weight RES1 RES2 RES3 RES4 RES5 RES1 RES2 RES3 RES4 RES5

1. 5 3 0-15 15 0-7 14 18 16 1 7 16 14 32 23 15 N

8-10

2. 5 4 0-30 10 0-7 19 24 18 21 20 48 32 43 38 44 N

8-10

3. 4 3 0-15 15 0-7 26 28 21 4 - 27 25 29 25 - N

8-10

4. 3 3 0-15 15 0-7 15 16 14 - - 38 38 39 - - N

8-10

5. 3 3 0-30 10 0-7 13 16 15 - - 38 35 43 - - N

8-10

6. 3 5 0-30 10 0-7 21 21 19 3 - 67 31 71 16 - Y

8-10

7. 3 4 0-25 12 0-7 25 26 25 - - 57 57 57 - - N

8-10

8. 4 5 0-25 10 0-7 22 25 25 19 - 58 55 58 60 - N

8-10

9. 4 5 0-25 15 0-7 27 24 26 24 - 62 63 64 62 - N

8-10

10. 2 3 0-20 10 8-10 18 20 - - - 34 34 - - - N

8-10

11. 2 3 0-50 10 0-7 13 12 1 - - 72 61 17 - - Y

8-10

12. 2 5 0-30 15 0-7 23 23 17 - - 65 93 59 - - Y

8-10

13. 2 6 0-25 15 0-7 31 33 27 - - 63 65 67 - - Y

8-10

14. 3 2 0-100 10 0-7 13 14 1 1 - 72 67 80 10 - Y

8-10

15. 4 2 0-70 20 0-7 13 15 5 2 7 81 63 78 63 54 Y

8-10

Figure 4: Experimental Results – here the load for resources

RES hosting the storage tasks.

• Backup (BK) RES storage set to 1000 MBs

These are applied and shown in Table 12 below.

The test cases have been executed with both sug-

gested methods, thus combining the autonomous PSO

algorithm with the MOO-based execution cost cycles

for task scheduling and load balancing. The results

are presented in Table 12.

For instance, in the experimental test case 3, we

took 4 resources and 3 tasks in each iteration, and we

limited the size of tasks to between 0 to 15 (MBs).

The total number of iterations was 15. In the results,

we can see that the tasks are assigned to the cloud

storages equally by the autonomous PSO algorithm

and execution cycle management approach. Effec-

tively, the load is balanced between resources (cloud

storages) with minor difference in usage percentage.

Similarly, in experiment 4, we took 3 storage re-

sources, 3 tasks in each iteration, 15 as the number of

iterations and the size limits of incoming tasks were

0 to 15 MBs. As a results, we found that the total

number of execution cycles per milliseconds were 15,

16, 14 in RES1, RES2 and RES3, respectively. The

memory consumption was 38, 38 and 39 percent in

RES1, RES2 and RES3, respectively.

If we look at the results of test cases 7 and 10,

the tasks are scheduled to cloud storage resources

equally, with an exactly balanced corresponding load.

In test case 7, all 3 resources used up to 57 % of the

total memory. In test case 10, 2 resources with 3 tasks

in each iteration consumed 34 % of the total memory

to execute the tasks until their termination.

A summary view of these test cases is provided

in Fig. 4 that shows the continuous operation of the

selected use cases and the resulting successfully bal-

anced task load.

6 CONCLUSIONS

Our aim was to investigate the use of swarm intelli-

gence for task scheduling and load balancing in dis-

tributed systems, speciﬁcally considering distributed

cloud and edge computing environments. For this,

we introduced an autonomous Particle Swarm Opti-

mization (A-PSO) algorithm that we combined with

Multi-Objective Optimization (MOO) .

In many investigations and applications of PSO al-

gorithms for distributed systems management, for in-

stance simulation techniques include the involvement

of a central manager entity. Our work adopts and ex-

tends the standard PSO algorithm, but with the ex-

tension of s decentralized autonomous aspect. In ad-

dition, we also examined the cost of time to execute

iteratively incoming task allocation requests to stor-

age resources using multi objective optimization. By

referring to Actual Round-Trip Time, we can calcu-

Enhanced Particle Swarm Optimisation and Multi Objective Optimization for the Orchestration of Edge Cloud Clusters

161

late the execution time of one single task allocated

to a remote storage in cycles per milliseconds. The

experimental results show that our enhanced A-PSO

algorithm effectively balanced the load across the re-

sources available. Thus, the autonomous coordination

of task scheduling and load balancing here is realized

by combining the two approaches – ﬁrstly, the A-PSO

algorithm and secondly, the time costing together as

a form of Multi Objective Optimization (MOO) for

task scheduling, load balancing and reducing the exe-

cution cost of the incoming iterations of tasks for dis-

tributed system in the cloud computing environment.

In the future, we plan to extend the applicability

of the solution to a wider range of distrbuted systems

architecture, taking also other resource types such as

compute and network into account.

REFERENCES

Mishra, S. K., Sahoo, B., and Parida, P. P. (2018). Load

balancing in cloud computing: A big picture. Jrnl of

King Saud University - Comp and Inf Sciences.

Pahl, C., Jamshidi, P., and Zimmermann, O. (2018). Archi-

tectural principles for cloud software. ACM Transac-

tions on Internet Technology (TOIT) 18 (2), 17.

al-Rifaie, M. M., Bishop, J. M., and Caines, S. (2012). Cre-

ativity and autonomy in swarm intelligence systems.

Cognitive computation 4.3: 320-331.

von Leon, D., Miori, L., Sanin, J., El Ioini, N., Helmer, S.,

and Pahl, C. (2018). A performance exploration of ar-

chitectural options for a middleware for decentralised

lightweight edge cloud architectures. Intl Conf on In-

ternet of Things, Big Data and Security.

Jamshidi, P., Pahl, C., and Mendonca, N. C. (2016). Man-

aging uncertainty in autonomic cloud elasticity con-

trollers. IEEE Cloud Computing 3 (3), 50-60.

Tan, Y., Shi, Y., and Ji, X. (2012). Advances in Swarm In-

telligence: Third International Conference ICSI.

Eberhart, R. and Kennedy, J. (1995). A new optimizer us-

ing particle swarm theory. International Symposium

on Micro Machine and Human Science. IEEE.

von Leon, D., Miori, L., Sanin, J., El Ioini, N., Helmer, S.,

and Pahl, C. (2019). A Lightweight Container Mid-

dleware for Edge Cloud Architectures. Fog and Edge

Computing: Principles and Paradigms, 145-170.

Scolati, R., Fronza, I., El Ioini, N., Samir, A., and Pahl,

C. (2019). A Containerized Big Data Streaming Ar-

chitecture for Edge Cloud Computing on Clustered

Single-Board Devices. CLOSER.

Kennedy, J. (2010). Particle swarm optimization. Encyclo-

pedia of machine learning: 760-766.

Visalakshi, P. and Sivanandam, S. N. (2009). Dynamic task

scheduling with load balancing using hybrid particle

swarm optimization. Int. J. Open Problems Compt.

Math 2.3:475-488.

Al-Maamari, A. and Omara, F.A. (2015). Task schedul-

ing using PSO algorithm in cloud computing environ-

ments. Intl Journal of Grid and Distributed Computing

8.5:245-256.

Zhang, L. et al. (2008). A task scheduling algorithm based

on PSO for grid computing. Intl Journal of Computa-

tional Intelligence Research 4.1:37-43.

Sharma, S. and Agnihotri, M. (2016). A Particle Swarm

Optimization based Technique for Scheduling Work-

ﬂow in Cloud DataCenter. Intl Journal of Engineering

Trends and Applications 3.4.

Pandey, S., Wu, L., Guru, S. M., and Buyya, R. (2010).

A particle swarm optimization-based heuristic for

scheduling workﬂow applications in cloud computing

environments. Intl Conference on Advanced Informa-

tion Networking and Applications.

Awad, A. I., El-Hefnawy, N. A., and Abdel kader, H. M.

(2015). Enhanced particle swarm optimization for task

scheduling in cloud computing environments. Proce-

dia Computer Science 65:920-929.

Selvarani, S., and Sudha Sadhasivam, G. (2010). Im-

proved cost-based algorithm for task scheduling in

cloud computing. International Conference on Com-

putational Intelligence and Computing Research.

Awad, A.I., El-Hefnawy, N.A., and Abdel kader, H.M.

(2015). Dynamic Multi-objective task scheduling in

Cloud Computing based on Modiﬁed particle swarm

optimization. Advances in Computer Science: an In-

ternational Journal 4.5:110-117.

Katyal, M. and Mishra, A. (2014). A comparative study of

load balancing algorithms in cloud computing envi-

ronment. arXiv preprint arXiv:1403.6918.

Acharya, J., Mehta, M., and Saini, B. (2016). Particle

swarm optimization based load balancing in cloud

computing. Intl Conf on Communication and Elec-

tronics Syst.

Mishra, R. and Jaiswal, A., (2012). Ant colony optimiza-

tion: A solution of load balancing in cloud. Intl Jour-

nal of Web & Semantic Technology 3.2:33.

Pahl, C. and Lee, B. (2015). Containers and clusters for

edge cloud architectures - a technology review. Intl

Conf on Future Internet of Things and Cloud.

Shi. Y. (2001). Particle swarm optimization: developments,

applications and resources. Proceedings Congress on

Evolutionary Computation Vol. 1.

Kalpana, C., Karthick Kumar, U., and Gogulan, R. (2012).

Max-Min Particle Swarm Optimization Algorithm

with Load Balancing for Distributed Task Scheduling

on the Grid Environment. Intl Journal of Computer

Science Issues 9.3.

ECTA 2019 - 11th International Conference on Evolutionary Computation Theory and Applications

162