Hindawi

Scientific Programming
Volume 2021, Article ID 7216795, 11 pages
https://doi.org/10.1155/2021/7216795

Research Article
Makespan Optimisation in Cloudlet Scheduling with Improved
DQN Algorithm in Cloud Computing

Amine Chraibi , Said Ben Alla , and Abdellah Ezzati

Hassan First University of Settat, Faculty of Science and Technology, Mathematics and Computer Science Department,
LAVETE Laboratory, 26000, Settat, Morocco

Correspondence should be addressed to Amine Chraibi; [email protected]

Received 6 July 2021; Revised 28 September 2021; Accepted 1 October 2021; Published 21 October 2021

Academic Editor: Pengwei Wang

Copyright © 2021 Amine Chraibi et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Despite increased cloud service providers following advanced cloud infrastructure management, substantial execution time is lost due to
minimal server usage. Given the importance of reducing total execution time (makespan) for cloud service providers (as a vital metric)
during sustaining Quality-of-Service (QoS), this study established an enhanced scheduling algorithm for minimal cloudlet scheduling
(CS) makespan with the deep Q-network (DQN) algorithm under MCS-DQN. A novel reward function was recommended to enhance
the DQN model convergence. Additionally, an open-source simulator (CloudSim) was employed to assess the suggested work per-
formance. Resultantly, the recommended MCS-DQN scheduler revealed optimal outcomes to minimise the makespan metric and other
counterparts (task waiting period, resource usage of virtual machines, and the extent of incongruence against the algorithms).

1. Introduction (1) The TS on a cloud computing platform is an ac-

knowledged NP-hard problem
Cloud computing denotes an established shared-computing (2) Multiple TS optimisation objectives are evident:
technology that dynamically conveys measurable on-demand completion time reduction and high resource usage
services over the global network [1]. Essentially, cloud for the entire task queue
computing offered users limitless and diverse virtual re-
sources that could be obtained on-demand and with different (3) Cloud resource dynamics, measurability, and het-
billing standards (subscription and static-oriented) [2]. The erogeneity resulted in high complexity
CS (task scheduling or TS) also outlined independent task Recent research has been performed to enhance TS in
mapping processes on a set of obtainable resources within a cloud environment through artificial intelligence algorithms
cloud context (for workflow applications) for execution (particularly metaheuristics involving particle swarm opti-
within users’ specified QoS restrictions (makespan and cost). misation or PSO [3], ant colony, and genetic algorithm or
Workflows (common applications associated with empirical GA) with TS capacities. However, this article does not rely
studies involving astronomy, earthquake, and biology) were on these algorithms but instead proposes a viable alternative
migrated or shifted to the cloud for execution. Although to them and compares it to one of the metaheuristic algo-
optimal resource identification for every workflow task (to rithms, such as PSO, a widely used technique in the task
fulfil user-defined QoS) was widely studied over the years, scheduling area. The proposed method primary objective
substantial intricacies required further research: was recommending a novel DQN scheduler for optimal
2 Scientific Programming

outcomes by comparing TS measures (waiting time, Sun and Qi [15] proposed a hybrid tasks scheduler based on
makespan reduction, and enhanced resource usage). Local search and differential evolution (DE) to enhance the
The remaining sections are arranged as follows: Section 2 makespan and the cost metrics. The authors in the paper [16]
outlines pertinent literary works, Section 3 presents the DQN presented a parallel optimized relay selection protocol to
algorithm, Section 4 highlights the recommended work, minimise latency, collision, and energy for wake-up radio-
Section 5 explains the research experiment setup and sim- enabled WSNs.
ulation outcomes, and Section 6 offers the study conclusion.

2.2. Reinforcement Learning-Based Research. Reinforcement

2. Related Work
learning (RL) is a machine-learning category that primarily
In cloud computing, TS, jobs scheduling, or resource se- communicated with the specified context using consecutive
lection is one of the most substantial complexities that has trials for an optimal TS method. Recently, RL has garnered
garnered cloud service providers’ and customers’ attention. much attention in cloud computing. For example, a higher
Additionally, specific study types on TS intricacies reflected TS success rate and minimal delay and energy consumption
positive outcomes. The literature’s research accomplish- were attained in [17] by recommending a Q-learning-ori-
ments in cloud resources scheduling can be divided into the ented and flexible TS from a global viewpoint (QFTS-GV). In
following categories based on the used techniques. [18], Ding et al. recommended a task scheduler using Q-
learning for energy-efficient cloud computing (QEEC).
Resultantly, QEEC was the most energy-efficient task
2.1. Heuristic-Based Research. The heuristic algorithms, scheduler compared to other counterparts (primarily cata-
including the metaheuristic ones [4] following intuition or lysed by the M/M/S queueing model and Q-learning
experimental development, offered a potential alternative for method). In [19], a TS algorithm was proposed with
the affordable resolution of every optimisation occurrence. Q-learning for wireless sensor network (WSN) to establish
Following the unpredictable degree of variance between Q-learning scheduling on time division multiple access (QS-
optimal and viable alternatives, past studies selected meta- TDMA). The algorithm implied QS-TDMA to be an
heuristic algorithms, such as PSO [5], GA [6], and ACO [7] approaching optimal TS algorithm that potentially enhanced
to solve optimal TS policy in cloud computing using met- real-time WSN performance. In [20], Che et al. recom-
aheuristics algorithms. Based on Huang et al.’s [8] recom- mended a novel TS model with the deep RL (DRL) algorithm
mendation of a PSO-based scheduler with a logarithm- that incorporated TS into resource-utilisation (RU) opti-
reducing approach for makespan optimisation, higher misation. The recommended scheduling model that was
performance was achieved against other heuristics algo- evaluated against conventional TS algorithms (on real
rithms. Meanwhile, Liang et al. [9] suggested a TS approach datasets) in experiments demonstrated a higher model
following PSO in cloud computing by omitting some in- performance of the defined metrics. Another task scheduler
ferior particles (to accelerate the convergence rate and dy- under the DRL architecture (task scheduling algorithm based
namically adjust the PSO parameters). The experimental on a deep reinforcement learning architecture or RLTS) was
findings revealed that the PSO algorithm obtained improved suggested by Dong et al. [21] for minimal task execution time
outcomes compared to other counterparts. The proposed with a preceding dynamic link to cloud servers. The RLTS
shift in genetic algorithm crossovers and mutation operators algorithm (compared against four heuristic counterparts)
implied flexible genetic algorithm operators (FGAO) [10]. reflected that RLTS could efficiently resolve TS in a cloud
For example, the FGAO algorithm minimised execution manufacturing setting. In [22], a cloud-edge collaboration
time and iterations compared to GA. Furthermore, Musa scheduler was constructed following the asynchronous ad-
et al. [11] recommended an improved GA-PSO hybrid with vantage actor-critic (CECS-A3C). The simulation outcomes
small position value (SPV) applications (for the initial demonstrated that the CECS-A3C algorithm decreased the
population) to diverge from arbitrariness and enhance task processing period compared to the current DQN and
convergence speed. Consequently, the improved GA-PSO RL-G algorithms. The authors of the article [23] suggest a
hybrid reflected more valuable outcomes than the con- learning-based approach based on the deep deterministic
ventional GA-PSO algorithm in resource usage and policy gradient algorithm to improve the performance of
makespan. Yi et al. [12] recommended a task scheduler mobile devices’ fog resource provisioning. Wang et al. [24]
model following an enhanced ant colony algorithm under introduced an adaptive data placement architecture that can
the cyber-physical system (CPS). The numerical simulation modify the data placement strategy based on LSTM and
implied that the model resolved local searchability and TS Q-learning to maximise data availability while minimising
quality concerns. Peng et al. [13] proposed a scheduling overall cost. Authors in [25] presented a hybrid deep neural
algorithm based on cloud computing’s two-phase best network scheduler to solve task scheduling issues in order to
heuristic scheduling to reduce the makespan and energy minimise the makespan metric. Wu et al. [26] utilised DRL
storage metrics. The authors of the paper [14] suggested a to address scheduling in edge computing for enhancing the
VM clustering technique for allocating VMs based on the quality of the services offered in IoT apps to consumers. The
duration of the requested task and the bandwidth level in authors in the paper [27] applied a DQN model with a
order to improve efficiency, availability, and other factors multiagent reinforcement learning setting to control the task
such as VM utilisation, bucket size, and task execution time. scheduling over cloud computing.
Scientific Programming 3

3. Background Rewardt

3.1. The RL. The RL theory was inspired by the psychological

and neuroscientific viewpoints of human behaviour [28] to
Statet
contextually select a pertinent action (from a set of actions) Environment Agent
for optimal cumulative rewards. Although the trial-and-
error approach was initially utilised for goal attainment (RL
was not offered a direct path), the experience was eventually Actiont
employed towards an optimal path. An agent only deter- Figure 1: Representation of reinforcement learning.
mined the most appropriate action in the problem following
the current condition, such as the Markov decision process
[29]. Figure 1 presents a pictorial RL representation where
the RL model encompassed the following elements [30]: Action 1

(1) A set of environment and agent states (S) Action 2

State
(2) A set of actions (A) of the agent

......
(3) Policies of transitioning from states to actions
(4) Rules that identified the immediate reward scalar of a Action N
transition
Figure 2: Representation of deep Q-network.
(5) Rules that outlined agent perception

agent experiences: st implied the current state, a denoted

3.2. The Q-Learning. One of the solutions for the rein-
action, r implied reward, st+1 reflected the next state, and
forcement problem in polynomial time is Q-learning. As
done denoted the Boolean value to identify goal attainment.
Q-learning could manage problems involving stochastic
As all experiences were stored in fixed-size memory, none
transitions and rewards without action adaptions or prob-
were linked to values (raw data input for neural network).
abilities at a specific point, the technique was also known as
Once the memory reached a saturation point during the
the “model-free” approach. Although RL proved successful
training process, arbitrary batches of a specific size were
in different domains (game playing), it was previously re-
chosen from the fixed memory. Regarding the insertion of
stricted to low dimensional state space or domains for
novel experiences, old experiences were eliminated once the
manual feature assignation. Equation (1) presents Q-value
memory became full. In this vein, experience relay deterred
computations where S denoted an actual and immediate
overfitting problems. Notably, the same data could be uti-
agent situation, α implied learning rate, c reflected a dis-
lised multiple times for network training to resolve insuf-
count factor, and Q(St , at ) denoted Q value to attain the “S”
ficient training data.
state by acting (a). Specifically, reinforcement began with
trial and error followed by posttraining experience (the
decisions corresponded to policy values that resulted in high 4. Proposed DQN Algorithm
reward counterparts).
4.1. The TS Problem. The TS protocol in cloud computing
Q St , at 􏼁 � Q St , at 􏼁 + α × 􏼢rt+1 + c × max􏼐St+1,a′ 􏼑 − Q St , at 􏼁􏼣. implies one of the vital problem-solving mechanisms on the
a′ significant overlap between cloud provider and user needs,
(1) including QoS and high profit [32]. Cloud service providers
strived to attain optimal virtual machine (VM) group usage
through reduced makespan and waiting time. Following
3.3. The DQN Architecture. Training encompassed specific Figure 3, a large set of autonomous work with varying
parameters [st , a, r, st+1 , done] that were stored as agent parameters was submitted by multiple users (to be managed
experiences: st implied the current state, a implied action, r by cloud providers in a cloud computing setting). For ex-
reflected the reward, st+1 denoted the following state, and ample, the cloud broker performed task delegations to the
done implied a Boolean value to identify goal attainment. current VMs [33]. Different optimisation algorithms were
The initial idea served to ascertain state and action as the also employed to attain optimal VM utilisation. Notably,
neural network input. Meanwhile, the output should denote equation (2) was incorporated to compute the overall exe-
the value representing how the aforementioned action cution time (makespan) as follows:
would reflect within the given state (see Figure 2). makespan � max􏼒Exvm1 , Exvm2 , . . . , Exvmj , . . . , Exvmm 􏼓.
(2)
3.4. Experience Replay. Experience replay [31] highlights the
capacity to learn from mistakes and adjust rather than re- Specifically, Exvmj � 􏽐ni�0 Exj (Cloudleti ), Exj (Cloudleti )
peating the same errors. Essentially, training encompassed demonstrated the cloudlet i execution time on vmj [34], n
several parameters [st , a, r, st+1 , done] that were stored as implied the total number of cloudlets, and Exvmj reflected the
4 Scientific Programming

Cloud provider
Task 1
Cloud Broker
Task 2
Data Centers
Task 3
Tasks scheduler
Task waiting
Task n
queue

Figure 3: Cloud scheduling architecture.

complete execution time of a set of cloudlets on vmj exe- 4.3. Model Training. The MCS-DQN model was retrained
cution. Figure 4 presents an example of the first-come first- for each episode in line with the workflow in Figure 6 as
served (FCFS) scheduling process where the number of follows:
virtual machines was 2 and the number of tasks was 7. Every
Step 1: the environment and agent contexts were
task encompassed varied time unit lengths. Notably,
established, including server, virtual machine, and
makespan denoted the most considerable execution time
cloudlet attributes.
between the aforementioned VMs. The makespan (com-
puted in VM2) was 45. Step 2: the environment state and cloudlet queues were
reset.
Step 3: the next cloudlet was selected from cloudlet
4.2. Environment Definition. This study regarded a system queues.
with multiple virtual machines and cloudlets. Every VM
encompassed specific attributes (processing power in MIPS, Step 4: the agent selected the following action in line with
memory in GB, and bandwidth in GB/s). As users submitted the existing environment state under ε factor. Essen-
distinct cloudlets that arrived in a queue, the broker tially, the ε factor (exploration rate) influenced the
implemented the defined scheduling algorithm to assign choice between exploration and exploitation in every
every cloudlet to an adequate VM. As the broker scheduling iteration. The possibility of an agent arbitrarily choosing
algorithm needed to make an assignment decision in every a VM (exploration) was 1 − ε while the possibility of the
cloudlet input from the queue, the system state was changed agent choosing a VM under the model (exploitation) was
in line with the decision. Figure 5 presents CS with a length ε. The ε factor (initialized by one) would reduce in every
of 3 to VM2 . iteration following a decay factor.
Step 5: the environment state was updated by adding
the cloudlet execution time to the chosen VM.
4.2.1. State Space. Only the time taken for each virtual
machine during a set of task execution was regarded in this Step 6: the environment produced a reward under the
study to support the defined system state identification recommended reward function in the following
process. The time counted on every virtual machine implied subsection.
the total cloudlet time running on VM. The virtual machine Step 7: the agent saved the played experience into the
running time facilitated makespan computation to enhance experience replay queue.
each novel cloudlet delegation where the system state Step 8: upon experience storage, the algorithm iden-
changed. In this vein, the t system state with n VMs was tified more cloudlets to schedule (to be repeated from
provided by 􏼈βt (VM1 ), . . . , βt (VMi ), . . . , βt (VMn )􏼉. Specif- Step 3 if more cloudlets were determined).
ki
ically, βt (VMi ) � 􏽐j�1 Exi (Cloudletj ), Exi (Cloudletj ) Step 9: the model was retrained in every episode
denoted the cloudlet j run time in VMi while ki implied the (completing all cloudlet queues) with a batch of defined
total cloudlets in VMi . Figure 5 presents the t state as {9, 7, 11} cloudlets from the experience queue. The experience
and t + 1 state as {9, 10, 11}. replay queue was applied as a FIFO queue. The oldest
experience was omitted when the queue reached a limit.
4.2.2. Action Space. Available agent actions were defined in Step 10: the algorithm was repeated from Step 2 if the
the action space. The broker scheduling algorithm was re- number of iterations was yet to reach the predefined
quired to choose a VM from all current VMs to schedule the episode limits.
existing task from the queue. For example, the agent would Step 11: the trained MCS-DQN model was saved and
make an action in the space with the same dimension as the exited.
number of VMs so that the action space denoted all VMs in
the system. The action space was outlined with n VMs by
{1, . . . , i, . . . , n}, wherein i denoted the VM index conceded 4.4. Reward Function. The recommended reward function
by the scheduler for cloudlet assignment. In Figure 5, action was utilised with the MCS-DQN model in Algorithm 1. The
space denotes {1, 2, 3}, while the chosen action implied is 2. makespan of every potential scheduling was first computed.
Scientific Programming 5

Burst time of
Serial Number VM1 1 (7) 3 (9) 5 (13) 7 (16)
task
1 7 0 7 16 29 45
2 3
VM2 2 (3) 4 (15) 6 (21)
3 9
0 3 18 39
4 15
5 13 Ex = 7 + 9 + 13 + 16 = 45
VM1

6 21 Ex = 3 + 15 + 21 =39
VM2
7 16 Makespan = Max (ExVM1, ExVM2) = 45

Figure 4: Makespan calculation example.

VM1 1 5 3 VM1 1 5 3

VM2 2 4 1 VM2 2 4 1 3

VM3 3 5 3 VM3 3 5 3
State t State t + 1
Figure 5: Example of system state change.

Start

Reset Cloudlet
queue

Input cloudlet

rand () < ε
Exploration
Exploitation
End
Predict next VM
Select next
using MCS-DQN
VM randomly
model

[false]

[true] Update VMs state

Save MCS-DQN
Episodes limit?
model

MCS-DQN reward

Remember current
experience (old state,
reward, new state)

Experience replay [true] [false]

and learning rate last task in queue ?
decay

Figure 6: The workflow of MCS-DQN training.

6 Scientific Programming

Input: vms_state//The current state of VMs

Input: vms_mips//VMs speeds in MIPS
Input: cloudlet//The cloudlet to be scheduled in MIPS
Input: action//The selected virtual machine
Output: reward//The reward
Function rewardCurrentState(vms_state, vms_mips, cloudlet, action):
vms_length⟵length(vms_state)//Get the number of VMs
makespans⟵(0, . . . , 0)//List of makespans initialized with zeros with a length of vms_length
for i � 0; i < vms_length; i � i + 1Iterate over VMs
do
vms_state_copy ⟵ copy(vms_state)//Make a copy of the current VMs state
task_run_time ⟵ runtime(cloudlet, vms_mips[i])//Calculate the run time of the given task in VMi
vms_state_copy[i] ⟵ vms_state_copy[i] + task_run_time//Add task’s execution time to VMi
makespans[i] ⟵ max(vms_state_copy)//Calculate the makespan of the copied state
end
makespans_sorted ⟵ descSort(makespans)//Sort VMs executions times in descending order
makespans_unique_length ⟵ unique_length(makespans)//Get the unique length of makespans to exclude makespans with the
same value
ranks_dict ⟵ emptyDict//Create an empty dictionary to store each VM rank.
start_rank ⟵ makespans_unique_length/2//The highest rank is given to a VM
for i � 0; i < vms_length; i � i + 1//Iterate over VMs
do
makespan ⟵ makespans_sorted[i]//makespan’s value from sorted makespans
makespan not in ranks_dict//Check the makespan’s value if it is not in the rank dictionary
then
ranks_dict[makespan] ⟵ start_rank//Assign the rank’s start value to makespani
start_rank←start_rank − 1//Decrement rank’s start value
end
end
vm_scores ⟵ (0, . . . , 0)//List of scores initialized with zeros with a length of vms_length
for index � 0; index < vms length; index � index + 1//Iterate over makespans
do
makespan ⟵ makespans[i]//Get makespani
vm_scores[index] ⟵ ranks_dict[makespan]//Score the VM based on makespan’s rank
end
return vm_scores[action]//Select the score based on the given action
End Function

ALGORITHM 1: Pseudocode of the proposed reward function.

Every VM was subsequently ranked following the makespan makespan (to decrease the highest score to be delivered to
computation during CS. A simple example was provided to the following makespan) (see Figure 7(e)). Lastly, the cor-
present the recommended MCS-DQN reward function (see responding reward was identified following the makespan
Figure 7). The example encompassed the reward compu- index and corresponding VM to be scheduled. In the study
tation for a specific VM state (elaborated following the total context, VM2 reflected the reward as 2.
execution time in every VMi ). Based on five VMs, every
VMi involved a set of cloudlet execution times {9, 7, 11, 8, 8}. 5. Results and Discussion
Specifically, a newly-arrived cloudlet was scheduled with a
length of five to VM2 in the example (see Figure 7(a)) by 5.1. Experimental Setup. The recommended trained model
iterating over VMs, creating a copy of VM state in every under deep Q-learning was assessed against FCFS and PSO
iteration, adding the cloudlet to the chosen VM in iteration, algorithms with the CloudSim simulator.
and computing the makespan following the added cloudlet.
Figure 7(b) presents the first iteration where the arrived
cloudlet was added to VM1. Figure 7(c) presents the 5.1.1. CloudSim Parameters. CloudSim is a modular sim-
computed makespans. For example, the makespan would be ulation toolkit for modelling and simulating cloud com-
14 when the cloudlet was added to VM1 in the first iteration, puting systems and application provisioning environments
13 when added to VM2, and so on. In every VMi ranking, [33]. It enables the modelling of cloud system components
the computed makespans were ranked following the lowest such as data centres, virtual machines (VMs), and resource
value by sorting the aforementioned makespans (see provisioning rules on both a system and behavioural level
Figure 7(d)) and providing the highest score to the lowest [33].
Scientific Programming 7

VM1 1 5 3 VM1 1 5 3 5

VM2 2 4 1 VM2 2 4 1

• VMs state VM3 3 5 3 VM3 3 5 3

VM4 5 1 2 VM4 5 1 2

VM4 3 5 VM4 3 5

• Cloudlet to be scheduled 5

• Schedule cloudlet to VM 2
(a) (b)

makespan 1 14 makespan 2 12

makespan 2 12 makespan 5 13

makespan 3 16 makespan 4 13

makespan 4 13 makespan 1 14

makespan 5 13 makespan 3 16

(c) (d)

VM 2 2

VM 5 1

Schedule cloudlet toVM 2

VM 4 1
means that the reward equals 2

VM 1 0

VM 3 -1

(e)

Figure 7: Running example of the proposed MCS-DQN reward function. (a) The example initialisation. (b) Adding cloudlet to VM1. (c) The
calculated makespans. (d) The sorted makespans. (e) VM ranks.

The CloudSim simulator configuration in the imple- Center North (HPC2N) in Sweden(the HPC2N data:
mentation began with establishing one data center, two https://www.cse.huji.ac.il/labs/parallel/workload/
hosts, and five VMs with subsequent parameters (see l_hpc2n/). The data contain information about tasks such
Table 1). This configuration setup is taken from the ex- as the number of processors, the average CPU time, the
ample 6 of CloudSim code source available on GitHub used memory, and other task specifications. The utilised
(CloudSim codebase: https://github.com/Cloudslab/ tasks from the workload completely differ from the in-
cloudsim), which is based on real servers and VMs in- dependent counterparts employed in the trained model.
formation. At the VM level, a time-shared policy (one of
the two different scheduling algorithms utilised in
CloudSim) was selected. The time-shared policy facilitated 5.1.2. The MCS-DQN Model Parameters. The MCS-DQN
VMs and cloudlets towards immediate multitasks and model application employs a neural network with five fully
progress within the host. Moreover, the tasks data used in connected layers (see Figure 8): an input layer (for state),
the experiments are real-world workloads of real computer three hidden layers (64 × 128 × 128), and an output layer
systems recorded by the High-Performance Computing (for actions). The network was taken from an original Keras
8 Scientific Programming

Table 1: Simulation parameters. Input Hidden Hidden Hidden Output

layer layer layer layer layer
Entity Parameter Value
Exvm1 VM1
Data center Size 1
1 1 1
Size 2
Amount of storage 1 terabyte Exvm2 VM2
RAM 2 gigabytes
Host
Scheduling’s algorithm Time-shared policy
Exvm3 VM3
Amount of bandwidth 10 gigabytes per second
Processing power 27079, 177730 MIPS
Cores 2, 6 cores Exvm4 64 128 128 VM4
Size 5
Size of the image 10 gigabytes
VM Exvm5 VM5
Amount of the bandwidth 1 gigabyte per second
RAM 0.5 gigabyte
Processing power 9726 MIPS
Source The HPC2N Seth log Figure 8: The MCS-DQN network setup.
Workload
Cloudlets’s number From 20 to 200
Table 2: Training parameters.
Parameter Value
RL tutorial [35] and modified to fit our defined environ-
Episodes 1000
ment. The training was executed following the parameters
Discount rate 0.9
in Table 2. The aforementioned parameters were obtained Exploration rate max 1.0
following specific training process execution (for a high Exploration rate decay 0.996
score in queue scheduling). Exploration rate min 0.05
Experience replay queue size 5000
Input size (state size) 5
5.1.3. PSO Parameters. The PSO algorithm was applied Output size (actions size) 5
following the recommended version in [5] with several it-
erations (equal to 1000), particles (equal to 500), local
weights (c1 and c2) with the same value of 1.49445, and a
1.0
fixed inertia weight with a value of 0.9.
460
0.8
5.2. Experimental Results and Analysis. Following Figure 9, 440
0.6
Epsilon

the MCS-DQN agent average assessment score reflected over

Score
800 episodes. Perceivably, learning remained steady despite 420
approximately 800 training iterations. The ε parameter 0.4
evolution was also incorporated into the ε-greedy explora- 400
tion method during training. Following increased agent 0.2
scores when ε began decaying, MCS-DQN could already 380
generate sufficiently good Q-value estimates for more
0 200 400 600 800
thoughtful state and action explorations to accelerate the
agent learning process. Training Steps
After the training process, various cloudlet sets were Epsilon
executed with the MCS-DQN scheduler saved model, Score
FCFS, and PSO algorithms for every metric assessment. Figure 9: Training score.
As every cloudlet of the same set was simultaneously
executed, this study essentially emphasised the makespan
metric (the elapsed time when simultaneously executing congruent system. Equation (3) was employed in this
cloudlet groups on available VMs). Figure 10 presents the research to calculate the DI metric.
reduced research makespan compared to other Emax − Emin
algorithms. DI � . (3)
Eavg
The makespan metric (employed as the primary model
training objective) impacted other performance metrics:
Specifically, Eavg , Emin , and Emax implied the average,
(1) The degree of imbalance (DI) metric demonstrated minimum, and maximum total execution time of all
load-balancing between VMs. Specifically, DI was VMs [34]. Figure 11 presents the recommended
utilised to compute the incongruence between VMs MCS-DQN scheduler that minimised the DI metric
when simultaneously executing a set of cloudlets. in every utilised set of cloudlets for an enhanced
The DI metric reduction was attempted for a more load-balancing system.
Scientific Programming 9

1600
1.75

Average Degree of Imbalance (DI)

1400
1.50
1200
Makespan (seconds)

1.25
1000
1.00
800

600 0.75

400 0.50

200 0.25

0 0.00
20 60 100 140 180 200 20 60 100 140 180 200
Numbers of cloudlets Numbers of cloudlets
FCFS FCFS
PSO PSO
MCS-DQN MCS-DQN
Figure 10: Makespan. Figure 11: Average DI.

(2) In the waiting time (WT) metric, the cloudlets ar-

rived in the queue and executed following the
800
scheduling algorithm. For example, the waiting time
Average waiting time (seconds)

algorithm was applied to compute all cloudlet se-

quence and average waiting time measures (see 600
equation (4)). Specifically, WTi denoted the cloudlet
waiting time while i and n reflected the queue length:
􏽐ni�1 WTi 400
WTavg � . (4)
n
200
Following Figure 12, the recommended MCS-DQN
scheduler could efficiently provide an optimal al-
ternative to heighten the cloudlet queue manage- 0
ment speed and effectiveness by reducing cloudlet 20 60 100 140 180 200
waiting time and queue length. Numbers of cloudlets
(3) The RU metric proved vital for elevated RU in the CS FCFS
process. PSO
MCS-DQN
Equation (5) is employed to compute the average RU
[34]. Figure 12: Average waiting time.
􏽐ni�1
ETVMi
ARU � . (5) set of VMs, we scheduled a number of tasks equal to 60, 140,
Makespan × N
and 200, respectively. These experiments were done to the
Specifically, ETVMi denoted the VMi duration to com- makespan since it is our main metric and compared with the
plete all cloudlets while N reflected the number of resources. PSO, the chosen scheduling algorithm in this work. We
In Figure 13, the recommended MCS-DQN scheduler was notice that our proposed MCS-DQN algorithm performance
more improved than PSO and FCFS regarding RU. Spe- is still better than the PSO scheduler even when adding more
experiments.
cifically, the MCS-DQN scheduler ensured busy resources
while CS (as service providers) intended to earn high profits However, our suggested approach is restricted to a set
by renting restricted resources. number of virtual machines, and any change in the number
Furthermore, to prove the effectiveness of our proposed of virtual machines requires a new model training. We
work, more executions based on the same previous VMs intend to concentrate on variable-length output prediction
configurations were conducted. Figure 14 illustrates the in the future such that the number of VMs does not impact
results of these executions where we increased the number of the model and no training is necessary for every change in
virtual machines to 10, 15, 20, and 30, respectively. In each VMs.
10 Scientific Programming

1.0 approaches, taking into account more efficiency metrics

such as task priority, VM migration, and energy con-
sumption. Furthermore, assuming that (n) tasks are
0.8
Average Resource Utilization

scheduled to (m) fog computing resource, we can apply

adjustments into the proposed algorithm to work on the
0.6 edge computing; this may also be an idea for future work.

0.4
Data Availability
The data for this research are available in the “Parallel
0.2 Workloads Archive: HPC2N Seth”: https://www.cse.huji.ac.
il/labs/parallel/workload/l_hpc2n/.
0.0
20 60 100 140 180 200 Conflicts of Interest
Numbers of cloudlets
The authors declare no conflicts of interest.
FCFS
PSO
MCS-DQN Authors’ Contributions
Figure 13: Average resource utilisation. All of the authors participated to the article’s development,
including information gathering, editing, modelling, and
reviewing. The final manuscript was reviewed and approved
1600
by all of the authors.
1400

1200 Acknowledgments
Makespan (seconds)

1000 Laboratory of Emerging Technologies Monitoring in Settat,

800
Morocco, provided assistance for this study.

600
References
400
[1] I. Foster, Y. Zhao, I. Raicu, and S. Lu, “Cloud computing and
200 grid computing 360-degree compared,” in Proceedings of the
2008 grid computing environments workshop, pp. 1–10, IEEE,
0
60 140 200 60 140 200 60 140 200 60 140 200 Austin, Texas, November 2008.
[2] M. A. Rodriguez and R. Buyya, “A taxonomy and survey on
10 VMs 15 VMs 20 VMs 30 VMs scheduling algorithms for scientific workflows in iaas cloud
Number of tasks computing environments,” Concurrency and Computation:
Practice and Experience, vol. 29, no. 8, Article ID e4041, 2017.
PSO
[3] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in
MCS-DQN
Proceedings of the ICNN’95-International Conference on
Figure 14: Increasing the number of virtual machines. Neural Networks, IEEE, Perth, WA, Australia, December
1995.
[4] W. Wong and C. I. Ming, “A review on metaheuristic al-
6. Conclusion gorithms: recent trends, benchmarking and applications,” in
Proceedings of the 2019 7th International Conference on Smart
This study encompassed effective CS application using deep Computing & Communications (ICSCC), pp. 1–5, IEEE, Miri,
Q-learning in cloud computing. Additionally, the MCS- Malaysia, June 2019.
DQN scheduler recommended TS problem enhancement [5] H. S. Al-Olimat, M. Alam, R. Green, and J. K. Lee, “Cloudlet
and metric optimisation. The simulation outcomes revealed scheduling with particle swarm optimization,” in Proceedings
that the presented work attained optimal performance for of the 2015 Fifth International Conference on Communication
minimal waiting time and makespan and maximum re- Systems and Network Technologies, pp. 991–995, IEEE,
source employment. Additionally, the recommended algo- Gwalior, India, April 2015.
rithm regarded load-balancing during cloudlet distribution [6] N. Zhang, X. Yang, M. Zhang, Y. Sun, and K. Long, “A genetic
to current resources beyond PSO and FCFS algorithms. This algorithm-based task scheduling for cloud resource crowd-
funding model,” International Journal of Communication
proposed model can be applied to solve task scheduling
Systems, vol. 31, no. 1, Article ID e3394, 2018.
problems in cloud computing, specifically in cloud broker. [7] Q. Guo, “Task scheduling based on ant colony optimization in
To solve the limitation of fixed VMs, we plan in the future to cloud environment,” AIP conference proceedings, vol. 1834,
enhance our work by relying on variable-length output p. 040039, 2017.
prediction using dynamic neural networks to include var- [8] X. Huang, C. Li, H. Chen, and D. An, “Task scheduling in
ious VM sizes, as well as adding other optimisation cloud computing using particle swarm optimization with time
Scientific Programming 11

varying inertia weight strategies,” Cluster Computing, vol. 23, Communications in China (ICCC), pp. 835–840, IEEE,
no. 13, pp. 1–11, 2019. Chongqing, China, August 2020.
[9] Y. Liang, Q. Cui, L. Gan, Z. Xie, and S. Zhai, “A cloud [23] M. Chen, T. Wang, S. Zhang, and A. Liu, “Deep reinforcement
computing task scheduling strategy based on improved learning for computation offloading in mobile edge com-
particle swarm optimization,” in Proceedings of the 2020 2nd puting environment,” Computer Communications, vol. 175,
International Conference on Big Data and Artificial Intelli- pp. 1–12, 2021.
gence, pp. 543–549, New York, NY, USA, April 2020. [24] P. Wang, C. Zhao, Y. Wei, D. Wang, and Z. Zhang, “An
[10] M. S. Ajmal, Z. Iqbal, M. B. Umair, and M. S. Arif, “Flexible adaptive data placement architecture in multicloud envi-
genetic algorithm operators for task scheduling in cloud ronments,” Scientific Programming, vol. 2020, Article ID
datacenters,” in Proceedings of the 2020 14th International 1704258, 2020.
Conference on Open Source Systems and Technologies [25] Z. Zang, W. Wang, Y. Song et al., “Hybrid deep neural
(ICOSST), pp. 1–6, IEEE, Lahore, Pakistan, December 2020. network scheduler for job-shop problem based on convolu-
[11] N. Musa, A. Y. Gital, F. U. Zambuk, A. M. Usman, tion two-dimensional transformation,” Computational In-
M. Almutairi, and H. Chiroma, “An enhanced hybrid genetic telligence and Neuroscience, vol. 2019, Article ID 7172842,
algorithm and particle swarm optimization based on small 2019.
position values for tasks scheduling in cloud,” in Proceedings [26] J. Wu, G. Zhang, J. Nie, Y. Peng, and Y. Zhang, “Deep re-
inforcement learning for scheduling in an edge computing-
of the 2020 2nd International Conference on Computer and
based industrial internet of things,” Wireless Communications
Information Sciences (ICCIS), pp. 1–5, IEEE, Sakaka, Saudi
and Mobile Computing, vol. 2021, Article ID 8017334, 2021.
Arabia, October 2020.
[27] Y. Wang, H. Liu, W. Zheng et al., “Multi-objective workflow
[12] N. Yi, J. Xu, L. Yan, and L. Huang, “Task optimization and
scheduling with deep-q-network-based multi-agent rein-
scheduling of distributed cyber-physical system based on
forcement learning,” IEEE access, vol. 7, pp. 39974–39982,
improved ant colony algorithm,” Future Generation Com-
2019.
puter Systems, vol. 109, pp. 134–148, 2020. [28] V. Mnih, K. Kavukcuoglu, D. Silver et al. “Human-level
[13] Z. Peng, B. Barzegar, M. Yarahmadi, H. Motameni, and control through deep reinforcement learning,” Nature,
P. Pirouzmand, “Energy-aware scheduling of workflow using vol. 518, no. 7540, pp. 529–533, 2015.
a heuristic method on green cloud,” Scientific Programming, [29] S. Jain, P. Sharma, J. Bhoiwala et al., “Deep q-learning for
vol. 2020, Article ID 8898059, 2020. navigation of robotic arm for tokamak inspection,” in Pro-
[14] C. Saravanakumar, M. Geetha, S. Manoj Kumar, ceedings of the International Conference on Algorithms and
S. Manikandan, C. Arun, and K. Srivatsan, “An efficient Architectures for Parallel Processing, pp. 62–71, Springer,
technique for virtual machine clustering and communications Xiamen, China, December 2018.
using task-based scheduling in cloud computing,” Scientific [30] S. He and W. Wang, “Pricing qoe with reinforcement learning
Programming, vol. 2021, Article ID 5586521, 2021. for intelligent wireless multimedia communications,” in
[15] Y. Sun and X. Qi, “A de-ls metaheuristic algorithm for hybrid Proceedings of the ICC 2020-2020 IEEE International Con-
flow-shop scheduling problem considering multiple re- ference on Communications (ICC), pp. 1–6, IEEE, Dublin,
quirements of customers,” Scientific Programming, vol. 2020, Ireland, June 2020.
Article ID 8811391, 2020. [31] L. J. Lin, “Reinforcement learning for robots using neural
[16] C. Huang, G. Huang, W. Liu, R. Wang, and M. Xie, “A parallel networks,” Tech. rep., Carnegie-Mellon Univ Pittsburgh PA
joint optimized relay selection protocol for wake-up radio School of Computer Science, Pittsburgh, 1993.
enabled wsns,” Physical Communication, vol. 47, Article ID [32] A. Karthick, E. Ramaraj, and R. G. Subramanian, “An efficient
101320, 2021. multi queue job scheduling for cloud computing,” in Pro-
[17] J. Ge, B. Liu, T. Wang, Q. Yang, A. Liu, and A. Li, “Q-learning ceedings of the 2014 World Congress on Computing and
based flexible task scheduling in a global view for the internet Communication Technologies, pp. 164–166, IEEE, Trichir-
of things,” Transactions on Emerging Telecommunications appalli, India, March 2014.
Technologies, vol. 32, Article ID e4111, 2020. [33] R. N. Calheiros, R. Ranjan, A. Beloglazov, C. A. F. De Rose,
[18] D. Ding, X. Fan, Y. Zhao, K. Kang, Q. Yin, and J. Zeng, “Q- and R. Buyya, “Cloudsim: a toolkit for modeling and simu-
learning based dynamic task scheduling for energy-efficient lation of cloud computing environments and evaluation of
cloud computing,” Future Generation Computer Systems, resource provisioning algorithms,” Software: Practice and
vol. 108, pp. 361–371, 2020. Experience, vol. 41, no. 1, pp. 23–50, 2011.
[19] B. Zhang, W. Wu, X. Bi, and Y. Wang, “A task scheduling [34] M. Kalra and S. Singh, “A review of metaheuristic scheduling
algorithm based on q-learning for wsns,” in Proceedings of the techniques in cloud computing,” Egyptian informatics journal,
International Conference on Communications and Networking vol. 16, no. 3, pp. 275–295, 2015.
in China, pp. 521–530, Springer, Shanghai, China, December [35] M. Plappert: keras-rl, 2016, https://github.com/keras-rl/keras-
2018. rl.
[20] H. Che, Z. Bai, R. Zuo, and H. Li, “A deep reinforcement
learning approach to the optimization of data center task
scheduling,” Complexity, vol. 2020, Article ID 3046769, 2020.
[21] T. Dong, F. Xue, C. Xiao, and J. Li, “Task scheduling based on
deep reinforcement learning in a cloud manufacturing en-
vironment,” Concurrency and Computation: Practice and
Experience, vol. 32, no. 11, Article ID e5654, 2020.
[22] F. Qi, L. Zhuo, and C. Xin, “Deep reinforcement learning
based task scheduling in edge computing networks,” in
Proceedings of the 2020 IEEE/CIC International Conference on