See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/360232899

Scheduling IoT Applications in Edge and Fog Computing

Environments: A Taxonomy and Future Directions

Preprint in ACM Computing Surveys · April 2022

DOI: 10.1145/3544836

CITATIONS READS

59 281

3 authors, including:

Marimuthu Palaniswami
University of Melbourne
648 PUBLICATIONS 31,208 CITATIONS

SEE PROFILE

All content following this page was uploaded by Mohammad Goudarzi on 30 April 2022.

The user has requested enhancement of the downloaded file.

 1

Scheduling IoT Applications in Edge and Fog Computing

Environments: A Taxonomy and Future Directions
MOHAMMAD GOUDARZI, MARIMUTHU PALANISWAMI, and RAJKUMAR BUYYA,
arXiv:2204.12580v1 [cs.DC] 26 Apr 2022

The University of Melbourne, Australia

Fog computing, as a distributed paradigm, offers cloud-like services at the edge of the network with low
latency and high-access bandwidth to support a diverse range of IoT application scenarios. To fully utilize the
potential of this computing paradigm, scalable, adaptive, and accurate scheduling mechanisms and algorithms
are required to efficiently capture the dynamics and requirements of users, IoT applications, environmental
properties, and optimization targets. This paper presents a taxonomy of recent literature on scheduling IoT
applications in Fog computing. Based on our new classification schemes, current works in the literature are
analyzed, research gaps of each category are identified, and respective future directions are described.
CCS Concepts: • General and reference → General literature; • Computer systems organization →
Distributed architectures.
Additional Key Words and Phrases: Fog computing, Internet of Things, Scheduling Taxonomy, Application
Structure, Environmental Architecture, Optimization Characteristics, Performance Evaluation
ACM Reference Format:
Mohammad Goudarzi, Marimuthu Palaniswami, and Rajkumar Buyya. 2022. Scheduling IoT Applications
in Edge and Fog Computing Environments: A Taxonomy and Future Directions. ACM Comput. Surv. 1, 1,
Article 1 (January 2022), 43 pages. https://doi.org/nn.nnnn/nnnnnnn.nnnnnnn

1 INTRODUCTION
The Internet of Things (IoT) paradigm has become an integral part of our daily life, thanks to the
continuous advancements of hardware and software technologies and ubiquitous access to the
Internet. The IoT concept spans a diverse range of application areas such as smart city, industry,
transportation, smart home, entertainment, and healthcare, in which context-aware entities (e.g.,
sensors) can communicate together without any temporal or spatial constraints [41, 52]. Thus, it
has shaped a new interaction model among different real-world entities, bringing forward new
challenges and opportunities. According to Business Insider [1] and International Data Corporation
(IDC) [2], 41 Billion active IoT devices will be connected to the Internet by 2027, generating more than
73 Zettabytes of data. The real power of IoT resides in collecting and analyzing the data circulating
in the environment [63], while the majority of IoT devices are equipped with a constrained battery,
computing, storage, and communication units, preventing the efficient execution of IoT applications
and data analysis on time. Thus, data should be forwarded to surrogate servers for processing
and storage. The processing, storage, and transmission of this gigantic amount of IoT data require
special considerations while considering a diverse range of IoT applications.
Authors’ addresses: M. Goudarzi, M. Palaniswami, and R. Buyya, The Cloud Computing and Distributed Systems
(CLOUDS) Laboratory, School of Computing and Information Systems, The University of Melbourne, Australia,
[email protected].

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires
prior specific permission and/or a fee. Request permissions from [email protected].
© 2022 Association for Computing Machinery.
0360-0300/2022/1-ART1 $15.00
https://doi.org/nn.nnnn/nnnnnnn.nnnnnnn

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:2 M. Goudarzi et al.

Cloud 1

Cloud 2

Cloud 3

Edge Layer

IoT Layer
Fog Layer Cloud Layer

Fog Computing Paradigm

Fig. 1. An illustration of Fog computing environments

1.1 Edge and Fog Computing Paradigms

Cloud computing is one of the main enablers of IoT that offers on-demand services to process,
analyze, and store the data generated from IoT devices in a simplified, scalable, and affordable man-
ner [36, 63, 106]. However, the current cloud infrastructure cannot solely satisfy the requirements
of a wide range of IoT applications. First, Cloud Data Centers (CDCs) are located at a multi-hop
distance from IoT devices, incurring high access latency and data transmission between IoT devices
and CDCs. Thus, it poses an important barrier to the efficient service delivery of real-time and
latency-sensitive IoT applications. Besides, the service startup time of IoT applications can be
negatively affected due to the extended transmission time required for sending commands to
distant Cloud Servers (CSs). Also, the extended transmission period and higher latency lead to
higher energy consumption for battery-constrained IoT devices. Second, when myriad IoT devices
initiate data-driven interactions with applications deployed on remote CSs, it incurs substantial
loads on the network and may lead to severe congestion. Third, it increases the computational
overhead on CDCs and may reduce their computing efficiency [76]. Fourth, the transmission of
sensitive raw data over the Internet is not feasible for some IoT applications due to privacy and
security concerns [87, 126]. To address these limitations, Edge and Fog computing paradigms have
emerged, bringing Cloud-like services to the proximity of IoT devices [15, 23].
In the Fog computing paradigm, a large number of geo-distributed and heterogeneous Fog Servers
(FSs) are located in an intermediate layer between CSs and IoT devices [37, 152]. Distributed FSs
(e.g., Raspberry Pis (Rpi), Nvidia Jetson platform, small-cell base stations, nano servers, femtocells,
regional servers, core routers, and switches) offer heterogeneous computational and storage re-
sources for IoT devices running various applications, as depicted in Fig. 1. Since FSs are located in
the proximity of IoT devices, compared to CDCs, they can offer Cloud-like services with less latency,
which effectively address the requirements of real-time and latency-sensitive IoT applications
[150]. Moreover, FSs can be accessed with higher bandwidth (i.e., data rate), reducing the required
transmission time. Furthermore, Fog computing can help reduce the energy consumption of IoT
devices, which is an important parameter, especially for battery-constrained IoT devices. Also, it
conserves network bandwidth that decreases the scope of network congestion [85, 107]. Besides,
Fog computing helps finer computational load distribution, reducing the massive load on CDCs.
Finally, Fog computing enables on-premises pre-processing/processing and storage of raw data,
which minimizes the requirement of transmitting raw data to distant servers. In this way, we could
mitigate the risks of security breaches and preserve data privacy compared to utilizing raw data.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:3

Compared to CSs, FSs usually have limited resources (e.g., CPU, RAM) while they can be ac-
cessed more efficiently. Thus, Fog computing does not compete with Cloud computing, but they
complement each other to satisfy diverse requirements of heterogeneous IoT applications. In our
view, Edge computing harnesses only distributed Edge resources at the closest layer to IoT devices,
while Fog computing harnesses distributed resources located in different layers and also Cloud
resources (although some works use these terms interchangeably [75, 113]), as shown in Fig. 1.

1.2 Scheduling IoT Applications in Fog Computing Environments

Fog computing paradigm provides a scalable solution for integrating a diverse range of hard-
ware and software technologies to offer a wide variety of services for end-users. Fog computing
environment is highly heterogeneous in terms of end-users’ devices, IoT applications, infrastruc-
tures, communication protocols, and deployed frameworks. Hence, the smooth execution of IoT
applications in this highly heterogeneous computing environment depends on a large number of
contributing parameters, making the efficient scheduling of IoT applications an important and yet
a challenging problem in Fog computing environments. To effectively utilize the potential of Fog
computing paradigm, these challenges should be thoroughly identified.
1.2.1 Challenges of Scheduling IoT Applications. The important challenges of scheduling IoT
applications are listed below:
• Challenges related to IoT devices: The IoT contains a large variety of devices with het-
erogeneous resource capabilities such as CPU, RAM, storage, and networking [63]. While
IoT devices can be used for the implementation of an unprecedented number of innovative
services, their resources are insufficient to host such applications in most scenarios [107].
Besides, a large category of IoT devices are battery-powered and the direct execution of IoT
applications on these devices negatively affects their lifetime. Also, many IoT application
scenarios consider moving IoT devices (e.g., real-time patient monitoring systems, drones
[77]), which further intensify the problem of energy consumption of IoT devices due to
increased interference and communication overhead [37].
• Challenges related to IoT applications: The number of IoT applications is rapidly increas-
ing due to recent advancements in technology, emerging business models, and end-users’
expectations. The design factors of IoT applications (e.g., granularity level, workload type,
and interaction model) heavily depend on application scenario and targeted end-users, neces-
sitating a specific amount of resources for smooth execution and different Quality of Service
(QoS) requirements. To illustrate, even in one IoT application domain such as healthcare,
there are significantly latency-sensitive applications (e.g., sleep apnea analysis [130]) while
there are other applications that are more computation-intensive (e.g., radiography [151]).
However, the number of computing and communication resources is usually limited. Besides,
when several applications are simultaneously requested by the same or different users, the
admission control for arriving concurrent requests should be considered as well, making the
design factors of an ideal application scheduler even more complex.
• Challenges related to Edge, Fog, and Cloud resources: In Fog computing environments,
heterogeneous FSs are situated between IoT devices and CDCs through several layers. Usually,
it is assumed that FSs are resource-constrained compared to CSs, where FSs at the bottom-most
layer (i.e., Edge) have the least amount of resources while providing better access latency and
communication bandwidth. Hence, several resource-specific parameters should be considered
for scheduling even one IoT application. Besides, considering the ever-increasing number
of IoT applications and their diverse resource requirements, some applications cannot be
executed on one FS. Thus, FSs require cooperation and resource-sharing with other FSs or

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:4 M. Goudarzi et al.

CSs for the execution of IoT applications, making the scheduling problem even more complex.
However, resource-sharing among FSs is less resilient than CSs due to spatial constraints and
high heterogeneity in deployed operating systems, standards, and protocols, just to mention a
few [67, 76, 79]. In addition, FSs are more exposed to end-users which makes them potentially
less-secured compared to CSs [126]. Besides, the geo-distribution of resources and the IoT
data hosted and shared on different servers may also impose privacy implications. As many
IoT users may share personal information in Fog computing environments, adversaries can
gain access to this shared information [121].
• Challenges related to optimizing parameters Optimizing the performance of IoT appli-
cations running in Fog computing environments depends on numerous parameters such
as the main goal of each IoT application, the capabilities of IoT devices, servers properties,
networking characteristics, and the imposed constraints. Optimizing the performance of
even one IoT application in such a heterogeneous environment with numerous contributing
parameters is complex, while multiple IoT applications with different parameters and goals
further complicate the problem.
• Challenges related to decision engines: Decision engines are responsible to collect all
contextual information and schedule IoT applications. Based on the context of IoT applica-
tions and environmental parameters, these decision engines may use different optimization
modeling [68]. Besides, there are several placement techniques to solve these optimization
problems. However, considering the types of IoT application scenarios and the number and
types of contributing parameters, different placement techniques lead to completely different
performance gain [39, 129]. For example, some placement techniques result in high-accuracy
decisions while their decision time takes a long time. However, some other techniques find ac-
ceptable solutions with shorter decision time. Moreover, the decision engines can be equipped
with several advanced features such as mobility support and failure recovery, enabling them
to work in more complex environments.
• Challenges related to real-world performance evaluation: The lack of global Fog ser-
vice providers offering infrastructure on pay-as-you-go models like commercial Cloud plat-
forms such as Microsoft Azure and Amazon Web Services (AWS) pushes researchers to set
up small-scale Edge/Fog computing environments [77]. The real-world performance eval-
uation of IoT applications and different placement techniques in Fog computing is not as
straightforward as Cloud computing since the management of distributed FSs incurs extra
effort and cost, specifically in large-scale scenarios. Besides, the modification and tuning of
system parameters during the experiments are time-consuming. Hence, while real-world
implementations are the most accurate approach for the performance evaluation of the
systems, it is not always feasible, specifically in large-scale scenarios.

1.2.2 Motivation of Research. Numerous techniques for scheduling IoT applications in Fog com-
puting environments have been developed to address the above-mentioned challenges. Several
works focused on the structure of IoT applications and how these parameters affect the sched-
uling [29, 79, 101] while some other techniques mainly focus on environmental parameters of
Fog computing, such as the effect of hierarchical Fog layers on the scheduling of IoT applications
[38, 62]. Besides, several techniques focus on defining specific optimization models to formulate
the effect of different parameters such as FSs’ computing resources, networking protocols, and IoT
devices characteristics, just to mention a few [68]. Moreover, several works have proposed different
placement techniques to find an acceptable solution for the optimization problem [5, 51, 78] while
some other techniques consider mobility management [26, 38, 131] and failure recovery [8, 40].
All these perspectives directly affect the scheduling problem, especially, when designing decision

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:5

in Fog Computing Environments

Application Structure

Scheduling IoT Applications

Environmental Architecture

Optimization Characteristics

Decision Engine Characteristics

Performance Evaluation

Fig. 2. Different perspectives of scheduling IoT applications in Fog computing

Application
Structure

Optimization Decision Engine Performance

Characteristics Characteristics Evaluation

Environmental
Architecture

Fig. 3. The relation among different identified perspectives

engines. These perspectives should be simultaneously considered when studying and evaluating
each proposal. However, only a few works in the literature have identified the scheduling challenges
that directly affect the designing and evaluation of decision engines in Fog computing environments
and accordingly categorized proposed works in the literature. Thus, we identify five important
perspectives regarding scheduling IoT applications in Fog computing environments, as shown in
Fig. 2, namely application structure, environmental architecture, optimization modeling, decision
engines’ characteristics, and performance evaluation.
Fig. 3 depicts the relationships among identified perspectives. The features of application structure
and environmental architecture help define the optimization characteristic and formulate the
problem. Then, an efficient decision engine is required to effectively solve the optimization problem.
Besides, the performance of the decision engine should be monitored and evaluated based on
the main goal of optimization for the target applications and environment. Considering each
perspective, we present a taxonomy and review the existing proposals. Finally, based on the studied
works, we identify the research gaps in each perspective and discuss possible solutions. The main
contributions of this work are:
• We review the recent literature on scheduling IoT applications in Fog computing from
application structure, environmental architecture, optimization modeling, decision engine
characteristics, and performance evaluation perspectives and propose separate taxonomies.
• We identify research gaps of scheduling IoT applications in Fog computing considering each
perspective.
• We present several guidelines for designing a scheduling technique in Fog computing para-
digm.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:6 M. Goudarzi et al.

• We identify and discuss several future directions to help researchers advance Fog computing
paradigm.

1.3 Paper Organization

The rest of the paper is organized as follows. The existing related surveys and taxonomies on
scheduling IoT applications in Fog computing environments are studied and compared with ours in
Section 2. Section 3 presents a taxonomy and overview of the IoT applications’ structure. Section 4
introduces a taxonomy on environmental properties of resources in Fog computing environments
and studies the existing works accordingly. In Section 5, a taxonomy of optimization characteristics
of problems in Fog computing environments is introduced. Section 6 identifies the important aspects
of decision engines and presents a taxonomy of decision engines for scheduling IoT applications.
Section 7 demonstrates the approaches and metrics used for the evaluation of scheduling strategies
in Fog computing. Section 8 presents a guideline for designing a scheduling technique. According
to identified research gaps, Section 9 provides several future research directions. Finally, Section 10
concludes this survey.

2 RELATED SURVEYS
In the context of Fog computing, surveys targeted different aspects of Fog computing, such as
security [112, 121, 126, 154], smart cities [105], live migration techniques [99], existing software and
hardware [107], deep learning applications [139], healthcare [66], and general surveys studied the
Fog computing paradigm, its scope, architectures, and recent trends [10, 44, 52, 91–93, 96, 122]. Also,
some surveys mainly discussed resource management, application management, and scheduling
in the context of Fog computing, such as [3, 34, 60, 68, 79, 86, 113, 118, 152], that we discuss and
compare them with ours.
Aazam et al. [3] reviewed enabling technologies and research opportunities in Fog computing
environments alongside studying computation offloading techniques in different domains such
as Fog, Cloud, and IoT. Hong et al. [48] and Ghobaei-Arani [34] studied resource management
approaches in Fog computing environments and discussed the main challenges for resource man-
agement. Yousefpour et al. [152] discussed the main features of the Fog computing paradigm and
compared it with other related computing paradigms such as Edge and Cloud computing. Besides,
it studied the foundations, frameworks, resource management, software, and tools proposed in Fog
computing. Mahmud et al.[79] mainly discussed the application management and maintenance
in Fog computing and proposed a taxonomy accordingly. Salaht et al. [113] presented a survey
of current research conducted on service placement problems in Fog Computing and categorized
these techniques. Shakarami et al. [118] studied machine learning-based computation offloading
approaches while Adhikari et al. presented the type and applications of nature-inspired algorithms
in the Edge computing paradigm. Martinez et al. [86] mainly focused on designing and evaluating
Fog computing systems and frameworks. Lin et al. [68] and Sonkoly et al. [123] mainly studied
and categorized different approaches for modeling the resources and communication types for
computation offloading in Edge computing. Finally, Islam et al. [60] proposed a taxonomy for
context-aware scheduling in Fog computing and surveyed the related techniques in terms of
contextual information such as user and networking characteristics.
Table 1 summarizes the characteristics of related surveys and provides a qualitative comparison
with our work. The proper scheduling of IoT applications in Fog computing environments can
be viewed from different perspectives, such as application structure, environmental architecture,
optimization modeling, and the features of decision engines. Besides, the performance of scheduling
techniques should be continuously evaluated to offer the best performance. As depicted in Table 1,
the existing surveys barely study and provide comprehensive taxonomy for the above-mentioned

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:7

Table 1. Related surveys on scheduling in Fog computing

Application Structure Environmental Architecture Optimization Characteristics Decision Engine Performance Evaluation Conceptualize Research
Survey Scheduling Gap
Research Research Research Research Research
Taxonomy Taxonomy Taxonomy Taxonomy Taxonomy Framework (in Years)
Gaps Gaps Gaps Gaps Gaps
[3] # G
# # # # # # # # # 4
[48] # # G
# G
# G
# # G
# # # # # 3.5
[152] # # G
# G
# # # # G
# # # # 3
[34] # # # # # # G
# G
# # # # 2.5
[113] # # # # G
# G
# G
# G
# # 2
[79] # G
# # # G
# G
# # # 1.5
[118] # # # # # # G
# G
# # # # 1.5
[86] # # # # # # # G
# G
# G
# # 1.5
[68] # # G
# # G
# # G
# # # # 1.5
[60] # G
# G
# # # # # G
# # # # 1
[123] G
# # # # G
# G
# # # # # # 0.5
[7] # # # # # G
# # G
# # # # 0.5
This Survey Current
: Full Cover, G
#: Partial Cover, # : Does Not Cover

perspectives. In this work, we identify the main parameters of each perspective and present a
taxonomy accordingly. Moreover, we identify related research gaps and provide future directions
to improve the Fog computing paradigm.

3 APPLICATION STRUCTURE
The primary goal of Fog computing is to offer efficient and high-quality service to users with
heterogeneous applications and requirements. Hence, service providers require a comprehensive
understanding of each IoT application structure (e.g., workload model and latency requirements) to
better capture its complexities, perform efficient scheduling and resource management, and offer
high-quality service to the users. Also, when designing the architecture of each IoT application,
dynamics, constraints, and complexities of resources in Fog computing should be carefully consid-
ered to exploit the potential of this paradigm. Fig. 4 presents a taxonomy and main elements of IoT
application structure, described below.

3.1 Architectural Design

The logic of IoT applications can be implemented in different ways. To illustrate, operations of a
Video Optical Character Recognition (VideoOCR) such as capturing frames, similarity check, and
text extraction can be implemented as a single encapsulated program or as a set of interdependent
components [24]. Hence, according to the granularity level of applications, their distribution, and
coupling intensity, the architectural design of applications can be classified into four types:
3.1.1 Monolithic. It encapsulates the complete logic of an application as a single component or
program. The parallel execution of these applications can be obtained using multi processing
approaches [79]. In the context of Fog computing, several works such as [13, 20, 84, 89] have
considered monolithic applications.
3.1.2 Independent. These applications require the execution of a set of independent tasks or com-
ponents for the complete execution of the application. The constituent parts of these applications
can be simultaneously executed on different FSs or CSs. Several works such as [9, 46, 55, 104]
discuss applications with independent components or tasks in the Fog literature.
3.1.3 Modular. Each modular application is composed of a set of dependent tasks or components.
While constituent parts of each application can be distributed over several FSs or CSs for parallel
execution, there are some constraints for the execution of tasks based on their data-flow dependency
model. Several works in the literature such as [39, 70, 72, 103] discuss modular applications.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:8 M. Goudarzi et al.

Monolithic

Independent
Architectural Design
Loosely-Coupled

Application Structure
Modular

Homogeneous
Granularity-based Heterogeneity
Heterogeneous

Stream/Realtime
Workload Model
Batch

Communication-Intensive

Communication-Computation Ratio (CCR) Computation-Intensive

Hybrid

Fig. 4. Application structure taxonomy

3.1.4 Loosely-coupled. Components of loosely-coupled applications (i.e., microservices) can be

distributed over several CSs or FSs. Besides, due to service-level isolation, components of applica-
tions can be shared among different applications, providing high application extendability. Several
works such as [24, 25, 38, 137] have considered loosely-coupled applications.

3.2 Granularity-based Heterogeneity

Tasks within an IoT application have different properties such as computation size, input size,
output size, and deadline. These features affect the scheduling complexity, where identifying the
dynamics of applications with heterogeneous task properties requires further effort. Accordingly,
we categorize IoT applications based on their granularity-level specifications to 1) heterogeneous
such as [47, 70] or 2) homogeneous such as [55, 153].

3.3 Workload Model

The workload model specifies the data architecture of an application, which can be broadly divided
into two categories for IoT applications:

3.3.1 Stream/Realtime. In this category, the data should be processed by the servers as soon as it
was generated (i.e., real-time), and hence, the data usually require relatively simple transformation
or computation. Several works such as [22, 65, 82] discuss stream workload for IoT applications.

3.3.2 Batch. In batch processing, the input data of an application is usually bundled for processing.
However, contrary to heavy batch processing models, IoT applications often use micro-batches to
provide a near-realtime experience. In the literature, several works such as [19, 45, 138] consider
batch workload for the applications.

3.4 Communication-Computation Ratio (CCR)

Each IoT application, regardless of its architecture, contains some amount of input data for transmis-
sion and computational load for processing. These properties can significantly affect the scheduling
decision to find proper FSs or CSs for an application. The CCR defines whether an application
on average is more 1) computation-intensive [42, 101, 132] or 2) communication-intensive
[13, 49, 116]. Besides, some works consider a range of applications to cover both computation-
intensive and communication-intensive applications, to which we refer as 3) hybrid [16, 40].

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:9

3.5 Discussion
In this section, we discuss the effects of identified application structure’s elements on the decision
engine and describe the lessons that we have learned. Besides, we identify several research gaps
accordingly. Table 2 summarizes the characteristics related to IoT application structure in Fog
computing.

3.5.1 Effects on the decision engine. The application structure properties affect the decision
engine in various aspects, as briefly described below.

1. Architectural design: It defines the number of tasks/modules and their respective dependency
within a single application. Hence, as the number of tasks/modules per application increases, the
problem space significantly increases. Considering an application with 𝑛 number of tasks and 𝑚
possible candidate configuration per task, the Time Complexity (TC) of finding the optimal solution
is 𝑂 (𝑚𝑛 ). Besides, the dependency of tasks within an application imposes hard constraints on
the problem, which further increases the complexity. Thus, finding the optimal solution for the
scheduling of applications becomes very time-consuming, and the design of an efficient placement
technique to serve applications in a timely manner remain an important yet challenging problem.

2. Granularity-based heterogeneity: It shows the corresponding properties of each task/module

within an application and plays a principal role in identifying the dynamics of applications. One of
the most important features of decision engines is their adaptability and their capability to extract
the complex dynamics of applications so that the decision engine can receive diverse types of
applications’ requests. Since applications with heterogeneous granularity-based properties have
higher dynamics’ complexity, the decision engines designed for this application category should
support high adaptability.

3. Workload model and CCR: These elements provide insightful information regarding the input
data architecture of the application and its behavior in the runtime. Accordingly, the decision
engine may define different priority queues for incoming requests based on their workload model
and CCR to provide higher QoS for the users. For example, applications with real-time workload
types and communication-intensive CCR may have higher priority for the placement on servers
closer to the IoT devices than computation-intensive applications that are not real-time.

3.5.2 Lessons learned. Our findings regarding the IoT application structure in the surveyed
works are briefly described in what follows:

1. Almost 70% of the surveyed works have overlooked studying the dependency model of tasks
within an application and selected either the independent or monolithic design. The rest of the works
consider dependency among tasks of an application in different models (i.e., sequential, parallel,
or hybrid dependency). Moreover, only about 10% of the studied works consider microservices in
their application design.

2. The most realistic assumption for the granular properties of each task/module is heterogeneous
(i.e., heterogeneous input size, output size, and computation size). Almost 85% of the studied works
consider heterogeneous properties for each task/module, while around 15% of the works consider
the homogeneous properties for the tasks/modules.

3. The workload model and CCR in each proposal depend on the targeted application scenarios.
Almost 55% of the works did not study the CCR, or the required information to obtain the CCR
(i.e., computation size of tasks, average data size) was not mentioned. Among the rest of the works,

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:10 M. Goudarzi et al.

computation-intensive, communication-intensive, and hybrid CCR form roughly 25%, 15%, and 5%
of proposals respectively.
3.5.3 Research Gaps. We have identified several open issues for further investigation, that are
discussed below:
1. According to Alibaba’s data of 4 million applications, more than 75% of the applications consist
of dependent tasks [155]. However, only around 30% of the recent works surveyed in this study
consider applications with dependent tasks (i.e., modular or loosely-coupled), showing further
investigation is required to identify the dynamics of these types of complex applications.
2. Although the microservice-based applications can significantly benefit the IoT scenarios,
only a few works such as [38, 137] have studied the scheduling and migration of microservices in
Edge/Fog computing environments. So, further investigation is required to study the behavior of
microservice-based applications with different resource management techniques.
3. Modular or loosely-coupled IoT applications can be distributed over different FSs or CSs
for parallel execution. However, several works such as [78] statically assign components of an
application on pre-defined FSs or CSs and only schedule one or two remaining components. Hence,
the best placement configuration of applications based on the current dynamics of the system
cannot be investigated, leading to diminished performance gain.
4. When the number of IoT applications increases, there is a high probability that application
requests with different workload models are submitted to the system. However, none of the studied
works in the literature consider applications with different workload models and how they may
mutually affect each other in terms of performance.
5. Due to the high heterogeneity of IoT applications in Fog, applications with diverse CCR may
be submitted for processing, requiring special consideration such as networking and prioritization.
Although there are only a few recent works such as [22, 39] that consider hybrid CCR, most of the
recent works target one of the computation-intensive or communication-intensive applications.
Table 2. Summary of existing works considering application structure taxonomy

Application Structure Application Structure

Ref Ref
Granular Workload Granular Workload
Design CCR Design CCR
Hetero Model Hetero Model
[9] Independent Hetero Batch Comp [81] Modular Hetero Stream ND
[13] Monolithic Hetero Batch Comm [27] Monolithic Hetero Batch Comp
[18] Independent Hetero Batch Comm [55] Independent Homo Batch ND
[25] Loosely-coupled Hetero Batch ND [89] Monolithic Homo ND ND
[39] Modular Hetero Batch Hybrid [20] Independent Hetero Batch Comm
[40] Modular Hetero Batch Hybrid [56] Independent Hetero Batch ND
[84] Monolithic Homo Batch ND [73] Independent Hetero Batch Comm
[70] Modular Hetero Batch Comp [144] Independent Hetero ND ND
[129] Independent Hetero Stream Comp [31] Independent Hetero Batch ND
[156] Independent Hetero Stream ND [72] modular Hetero Stream ND
[47] Monolithic Hetero Batch Comm [38] Loosely-coupled Hetero Stream Comp
[54] Monolithic ND ND ND [141] Independent Hetero Batch Comp
[147] Monolithic Homo Batch Comp [100] Monolithic Hetero ND ND
[137] Loosely-coupled Hetero Stream Comm [119] Loosely-coupled Hetero Stream ND
[140] Independent Hetero Stream Comm [133] Monolithic Homo Batch ND
[97] Modular Hetero Batch Comp [71] Monolithic Homo Batch Comp
[125] Independent Hetero Stream Comm [148] Independent Hetero Batch ND
[127] ND ND Batch ND [109] Modular Hetero Batch ND
[32] Independent Hetero Stream ND [157] Independent Hetero Batch ND
[146] Independent Hetero ND Comp [153] Monolithic Homo ND ND
[136] Monolithic Homo ND Comp [111] Independent Hetero Batch ND

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:11

[28] Modular Hetero Batch ND [120] Monolithic Hetero Batch Comp

[132] Loosely-coupled Hetero Stream Comp [143] Independent Hetero Batch ND
[138] Independent Hetero Batch ND [83] Independent Hetero ND ND
[42] Independent Hetero Batch Comp [117] Loosely-coupled Hetero ND ND
[45] Independent Hetero Batch ND [90] Modular Hetero Batch ND
[149] Loosely-coupled Hetero Stream Comp [14] Independent ND Stream ND
[53] Independent Hetero Stream Comp [58] Independent Hetero Stream ND
[64] Loosely-coupled Hetero Batch Comp [114] Loosely-coupled Hetero Batch ND
[128] Monolithic Hetero Stream Comp [6] Independent Hetero Batch ND
[33] Independent Hetero Batch Comp [16] Modular Hetero Batch Hybrid
[19] Independent Hetero Batch ND [22] Independent Hetero Stream Hybrid
[11] Monolithic Homo Batch ND [4] Independent Hetero Batch Comp
[12] Independent Hetero ND ND [50] Independent Hetero Batch ND
[21] Monolithic Homo Batch ND [59] Modular Hetero Batch ND
[74] Monolithic Homo Batch Comp [61] Independent Hetero Batch ND
[30] Loosely-coupled Hetero ND Comm [65] Independent Hetero Stream ND
[155] Modular Hetero Batch Comm [69] Modular Hetero Batch ND
[135] Modular Hetero Batch Comp [82] Independent Hetero Stream ND
[110] Independent Hetero Stream ND [95] Independent Hetero Stream ND
[88] Independent Hetero Stream ND [101] Loosely-coupled Hetero Stream Comp
[142] Modular Hetero Batch Comm [98] Independent Hetero Stream ND
[24] Loosely-coupled Hetero Stream ND [103] Modular Hetero Stream ND
[49] Monolithic Homo Batch Comm [104] Independent Hetero Batch ND
[145] Modular Homo Batch ND [108] Loosely-coupled Hetero Stream ND
[57] Independent Hetero ND Comp [115] Independent Hetero Batch ND
[80] Independent Hetero Stream ND [116] Independent Hetero Batch Comm
[94] Independent Hetero Batch ND [134] Modular Hetero Batch Comp
[124] Modular Hetero Batch Hybrid [46] Independent Hetero Batch ND
[78] Modular Hetero ND ND [102] Modular Hetero Batch Comp
ND: Not Defined, Hetero: Heterogeneity/heterogeneous, Comp: Computation-Intensive, Comm: Communication-Intensive
Homo: Homogeneous

4 ENVIRONMENTAL ARCHITECTURE
The configuration and properties of IoT devices and resource providers directly affect the complexity
and dynamics of scheduling IoT applications. To illustrate, as the number of resource providers
increases, heterogeneity in the system also grows as a positive factor, while the complexity of
making a decision also increases that may negatively affect the process of making decisions. In
this section, we classify the environmental architecture properties, as depicted in Fig. 5, into the
following categories:

4.1 Tiering Model

IoT devices and resource providers can be conceptually organized in different tiers based on their
proximity to users and resources, described below:

4.1.1 Two-Tier. In this resource organization, IoT devices are situated at the bottom-most layer
and resource providers are placed at the edge of the network in the proximity of IoT devices (i.e.,
Edge computing). Several works use two-tier resource organization such as [55, 69, 84, 125].

4.1.2 Three-Tier. Compared to two-tier model, this model also uses CSs at the highest-most layer
to support edge resources (i.e., Fog computing). Several works considered three-tier model in the
literature such as [59, 65, 108, 116].

4.1.3 Many-Tier. In many-tier resource organizations, IoT devices and CSs are situated at the
bottom-most and highest-most tiers respectively, while FSs are placed in between through several
tiers (i.e., hierarchical Fog computing). In the literature, several works have considered many-tier
model such as [31, 38, 83, 111].

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:12 M. Goudarzi et al.

Two-Tier
Tiering Three-Tier
Many-Tier
Single
Number
Multiple
Mobile Device
IoT Device Type Vehicles
General
Environmental
Architecture

Homogeneous
Heterogeneity
Heterogeneous
Single
Number
Multiple
Type
Homogeneous
Heterogeneity
Fog Servers Heterogeneous
Intra-Tier
Cooperation
Inter-Tier
Single
Number
Cloud Servers Multiple
Cooperation

Fig. 5. Environmental architecture taxonomy

4.2 IoT Devices

IoT devices can play two roles; as service requester and/or resource provider. When they act as
service requesters, still they can execute a portion of their tasks or components based on their
available resources. Moreover, IoT devices can simultaneously play these different roles. We study
the properties of IoT devices from the following perspectives:

4.2.1 Number. The higher number of IoT devices (either as service requester or service provider),
the higher complexity of the scheduling problem. Some works only consider single IoT device in the
environment such as [70, 97, 138] while other works consider multiple IoT devices simultaneously
such as [50, 100, 133].

4.2.2 Type. The type of IoT devices help us understand the amount of resources, capabilities, and
constraints of these devices. The IoT devices used in the current literature can be broadly classified
into three categories, namely 1) mobile devices (MD) which are mostly considered as smartphones
or tablets [25, 102, 119], 2) Vehicles [137, 140, 148], and 3) General devices containing a set of IoT
devices, ranging from small sensors to drones [39, 84, 120].

4.2.3 Heterogeneity. We also study the resources of IoT devices and their request types, and classify
proposals into 1) heterogeneous where IoT devices have various resources and different request
types and sizes such as [31, 72, 145] or 2) homogeneous where the resources of IoT devices are
the same or they have the same request type and size such as [70, 138, 149].

4.3 Fog Servers (FSs)

FSs usually act as resource providers for IoT devices. The environmental properties of FSs can be
classified based on the following criteria:

4.3.1 Number. Similar to IoT devices, we classify the number of FSs in the environment into 1)
Single and 2) Multiple. The complexity and dynamics of system in surveyed works that have
considered only single FS such as [53, 54, 110] is simpler to the works that have considered multiple
FSs such as [28, 127, 146].

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:13

4.3.2 Type. The type of FSs acting as service provider in the Fog computing environment ranges
from IoT devices with additional resources to resource-rich data centers. Several works have
considered a specific type of FS and discuss their properties in their works such as 1) Base Station
(BS) and Macro-cell Station (MS) [18, 127, 138], 2) femtocells [19, 32, 37], 3) Rpi [30] and 4)
access points (AP) [27, 88]. Moreover, several works consider 5) general FSs containing a set of
FSs with different types such as [31, 94].
4.3.3 Heterogeneity. We study the FSs’ resources and classified works based on their heterogeneity
into 1) heterogeneous and 2) homogeneous accordingly. Many works have considered hetero-
geneous resources for FSs [9, 39, 46, 102] while some works consider homogeneous resources for
FSs [54, 73, 134].
4.3.4 Cooperation. Compared to CSs, each FS has fewer resources and it may not be able to
satisfy the requirements of IoT applications. Cooperation among FSs helps augmenting their
resources and providing service for demanding IoT applications. We classify proposals based on
their cooperation among FSs into 1) intra-tier where FSs of same tier collaborate to satisfy users’
requests [24, 25, 25, 109] and 2) inter-tier where FSs of different layers also collaborate for the
execution of IoT one application [38, 39].

4.4 Cloud Servers (CSs)

The environmental properties of CSs can be classified based on the following criteria:
4.4.1 Number. The current literature based on the CSs’ number can be divided into 1) single and
2) multiple categories. Majority of works only consider one CDC as resource provider (either as
a central entity with aggregated resources or different number of VMs) to support FSs such as
[14, 50, 59, 65]. In real-world environment, different CDCs are available which can provide services
with different QoS for multiple applications. Some works such as [35, 39, 40, 132] have considered
multiple CDCs with heterogeneous CSs in the literature.
4.4.2 Cooperation. Among the works considered multi CDCs, we study either CSs from different
CDCs are configured to collaboratively execute an IoT application or not. In the literature, some
works such as [35, 40, 104] have considered collaborative multi CDCs scenarios.

4.5 Discussion
In this section, we discuss the effects of identified environmental architecture’s elements on the
decision engine and describe the lessons that we have learned. Besides, we identify several research
gaps accordingly. Table 3 provides a summary of properties related to environmental architecture
in Fog computing.
4.5.1 Effects on the decision engine. The elements of environmental architecture affect the
decision engine in various aspects, as briefly described below.
1. Tiering: It represents the organization of end-users’ devices and resources in the computing
environment. Considering the properties of resources in different tiers, it helps find the most
suitable deployment layer for the decision engine to efficiently serve IoT applications’ requests with
a wide variety of requirements. For example, to serve real-time IoT applications with low startup
time requirements, the most suitable deployment layer in the three-tier model is the lowest-level
Fog layer.
2. IoT devices: The number of IoT devices directly relates to the number of incoming requests
to decision engines. It affects the admission control of decision engines. The type of IoT devices

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:14 M. Goudarzi et al.

provides contextual information about the number of resources and intrinsic properties of the IoT
devices that are important for the decision engine. For example, the MD type not only states that
the IoT device does not have significant computing resources, but also presents that the device has
mobility features. Thus, the IoT device type affects the advanced features of the decision engine,
such as mobility, and also specifies whether the IoT devices can serve one or several tasks/modules
of IoT applications or not.

3. Fog and Cloud servers: The number of available servers directly affects the TC of the scheduling
problem. Considering an application with 𝑛 number of tasks and 𝑚 possible candidate configuration
per task, the TC of finding the optimal solution is 𝑂 (𝑚𝑛 ). Hence, it directly affects the choice of
placement technique and scalability feature of the decision engine. As the problem space increases,
a suitable decision engine should be selected to solve the scheduling problem. Moreover, the type
and heterogeneity of resources provide further contextual information for the decision-making,
such as the number of resources, networking characteristics, and resource constraints, just to
mention a few.

4.5.2 Lessons learned. Our findings regarding the environmental architecture in the surveyed
works are briefly described in what follows:

1. Almost 60% of works consider the three-tier model and many-tier models for the organization
of end-users and resources. Not only do these works consider real-time applications, but also some
of them assume both real-time and computation-intensive applications, such as [16, 39, 124]. This is
mainly because these works use CSs as a backup plan for more computation-intensive applications
or when the number of incoming IoT requests increases and the FSs cannot solely manage the
incoming requests. Moreover, nearly 40% of surveyed works assume a two-tier model for the
organization of end-users and resources. These works mostly assume real-time workload type and
communication-intensive applications for the deployment on Edge servers, such as [13, 22, 53, 110].

2. In the surveyed works, almost 90% of the works considered an environment with multiple
IoT devices, while 10% of works only focused on a single IoT device. When the number of IoT
devices increases, the diversity of IoT applications and heterogeneity of their tasks also increase
accordingly. Moreover, the greater number of works assume IoT devices as general devices with
sensors, actuators, and diverse application requests. In contrast, some works targeted a specific IoT
devices such as mobile devices and vehicles with almost 30% and 10% of proposals, respectively.
These proposals studied other contextual information of targeted IoT devices in detail such as
mobility [138], energy consumption [140], and networking characteristics [109, 137]. Finally, about
90% of works have studied IoT devices with heterogeneous properties and diverse application
request types, which are the closest scenario to real-world computing environments.

3. Regarding Fog resources, almost 90% of the proposals consider multiple FSs in the environment.
However, only 40% of the current literature has considered any cooperation model among FSs.
There is a high probability that a single FS cannot solely manage the execution of several incoming
resources due to its limited resources. Also, sending partial/complete applications’ tasks to the
Cloud may negatively affect IoT devices’ response time and energy consumption, especially for
real-time IoT applications. Thus, cooperation among FSs is of paramount importance that can lead
to the execution of IoT applications with better performance and QoS. Considering the type of the
FSs, about 60% of the studied literature considered general FSs. The rest of the works studied a
specific type of FSs and tried to involve their contextual information in the scheduling process of IoT
applications, such as networking characteristics [9]. Moreover, some works considered IoT devices

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:15

can simultaneously play different roles in the computing environments (i.e., service requester and
service provider) such as [132, 140, 157].
4. In the current literature, around 60% of the works consider CSs as computing resources
in the environment. However, only in 8% of these works multiple Cloud service providers (i.e.,
multi-Cloud), their communication, and interactions are studied, such as [39, 40, 104, 132].
4.5.3 Research Gaps. We have identified several open issues for further investigation, as dis-
cussed below:
1. Hierarchical Fog computing (i.e., multi-tier) has not been thoroughly considered by researchers.
Only a few works (almost 5%) consider the organization of resources in the multi-tier environment,
and most have focused on the heterogeneity of resources among different tiers. However, these
works have not considered the heterogeneity of communication protocols and latency in multi-tier
environments.
2. In the literature, several works have considered abstract CDC as a central unit with huge
computing capacity [80], while in reality, CDCs contain several CSs hosting computing instances.
Such assumptions affect the computing and communication time in simulation studies.
3. One of the main advantages of Fog computing is providing heterogeneous FSs in IoT devices’
vicinity to collaboratively serve applications. However, many works have not considered coopera-
tion among FSs. In this case, due to the limited computing and communication resources of each
FS and a large number of IoT requests, the serving FS may become a bottleneck which negatively
affects the response time and QoS [37]. Besides, in uncooperative scenarios, the overloaded FS
forwards requests to CSs, incurring higher latency. Hence, cooperative Fog computing, associated
protocols, and constraints require further investigation for different IoT application scenarios.
Table 3. Summary of existing works considering environmental architecture taxonomy

Environmental Architecture
Ref IoT Device Fog Servers Cloud Servers
Tiering
Number Type Hetero Number Type Hetero Coop Number Coop
[9] Two-Tier Multiple MD Hetero Multiple BS, MS Hetero # # #
[13] Two-Tier Multiple Vehicle Hetero Multiple RSU ND # # #
[18] Three-Tier Multiple MD ND Multiple BS Hetero Intra Single #
[25] Three-Tier Multiple MD Hetero Multiple BS Hetero Intra Single #
[39] Many-Tier Multiple General Hetero Multiple General Hetero Intra, Inter Multiple
[40] Three-Tier Multiple General Hetero Multiple General Hetero Intra Multiple
[84] Two-Tier Multiple General Hetero Multiple Cloudet Hetero # # #
[70] Three-Tier Single General Homo Multiple General Hetero Intra Single #
[129] Three-Tier Multiple General Homo Multiple General Hetero ND Single #
[156] Two-Tier Multiple General ND Multiple BS, MS Hetero # # #
[47] Three-Tier Multiple MD Hetero Multiple BS Hetero Intra Single #
[54] Two-Tier Multiple MD Hetero Single BS Homo # # #
[147] Two-Tier Multiple MD Hetero Multiple BS Hetero Intra # #
[137] Three-Tier Multiple Vehicle Hetero Multiple Hybrid Hetero Intra Single #
[140] Three-Tier Multiple Vehicle Hetero Multiple BS, Vehicle Homo # Single #
[97] Three-Tier Single MD Hetero Single General Homo # Single #
[125] Two-Tier Multiple MD Hetero Multiple General Hetero # # #
[127] Three-Tier Multiple MD Hetero Multiple BS Hetero Intra Single #
[32] Three-Tier Multiple MD Hetero Multiple Femto Hetero Intra Single #
[146] Three-Tier Multiple MD Hetero Multiple BS Hetero Intra Single #
[136] Two-Tier Multiple MD Hetero Multiple BS, MD Hetero Intra # #
[28] Three-Tier Multiple ND ND Multiple General Hetero Intra Single #
[132] Three-Tier Multiple MD Hetero Multiple General, MD Hetero Intra Multiple
[138] Two-Tier Single MD Homo Single BS Homo # # #
[42] Two-Tier Multiple MD Hetero Multiple BS, MS Hetero # # #
[45] Two-Tier Multiple MD Hetero Multiple General Hetero # # #
[149] Two-Tier Multiple MD Homo Single General Homo # # #

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:16 M. Goudarzi et al.

[53] Two-Tier Multiple General Hetero Single General Homo # # #

[64] Three-Tier Multiple General Hetero Multiple General Hetero Intra Single #
[128] Three-Tier Multiple General Hetero Multiple General Hetero Intra Single #
[33] Three-Tier Multiple General Hetero Multiple General Homo Intra Single #
[19] Two-Tier Multiple MD Hetero Multiple BS, Femto Hetero # # #
[11] Two-Tier Multiple MD Hetero Multiple BS Hetero # # #
[12] Three-Tier Multiple General ND Multiple General Hetero # Single #
[21] Two-Tier Multiple General Hetero Multiple General Hetero # # #
[74] Two-Tier Multiple General Hetero Multiple BS Hetero # # #
[30] Three-Tier Single General Hetero Single Rpi Homo Intra Single #
[155] Two-Tier Multiple General Hetero Multiple Hybrid Homo ND # #
[135] Two-Tier Multiple General Hetero Single General Homo # # #
[110] Two-Tier Multiple General Hetero Single General Homo # # #
[88] Two-Tier Multiple General Hetero Multiple AP Hetero # Single #
[142] Three-Tier Single ND ND Single General Homo # Single #
[24] Three-Tier Multiple General Hetero Multiple General Hetero Intra Single #
[49] Three-Tier Multiple General Hetero Multiple BS Hetero Intra Single #
[145] Three-Tier Multiple General Hetero Single Cloudlet Homo # Single #
[57] Three-Tier Multiple MD Hetero Multiple General Homo # Single #
[80] Three-Tier Multiple General Hetero Multiple General Hetero # Single #
[94] Three-Tier Multiple General Hetero Multiple General Hetero # Single #
[124] Three-Tier Single General Hetero Multiple General Hetero Intra Single #
[78] Three-Tier Multiple General Hetero Multiple General Hetero Intra Single #
[81] Three-Tier Multiple General Hetero Multiple General Hetero # # #
[27] Three-Tier Multiple MD Hetero Multiple AP Hetero # Single #
[55] Two-Tier Multiple MD Homo Single AP Homo # # #
[89] Two-Tier Single General Homo Multiple General Hetero # # #
[20] Two-Tier Single MD Hetero Multiple BS Homo # # #
[56] Two-Tier Multiple MD Hetero Single General Homo # # #
[73] Two-Tier Multiple MD Hetero Single General Homo # # #
[144] Three-Tier Multiple MD Hetero Multiple General Hetero # Single #
[31] Many-Tier Multiple General Hetero Multiple General Hetero # Single #
[72] Three-Tier Multiple General Hetero Multiple General Hetero Intra Single #
[38] Many-Tier Multiple General Hetero Multiple General Hetero Intra, Inter Single #
[141] Two-Tier Multiple MD Hetero Multiple BS Hetero # # #
[100] Two-Tier Multiple MD Hetero Multiple BS Hetero Intra # #
[119] Three-Tier Single MD Homo Multiple Hybrid Hetero # Single #
[133] Three-Tier Multiple MD Hetero Multiple AP Hetero # Single #
[71] Two-Tier Multiple MD Hetero Multiple BS Hetero Intra # #
[148] Two-Tier Multiple Vehicle ND Multiple AP Hetero Intra # #
[109] Two-Tier Multiple Vehicle Hetero Multiple BS Hetero Intra # #
[157] Two-Tier Multiple Vehicle Hetero Multiple BS, Vehicle Hetero Intra # #
[153] Two-Tier Single MD Homo Multiple General Hetero ND # #
[111] Many-Tier Multiple MD Hetero Multiple General Hetero ND Single #
[120] Two-Tier Multiple General Hetero Multiple General Hetero # # #
[143] Two-Tier Single Vehicle Hetero Multiple AP, Vehicle Hetero Intra # #
[83] Many-Tier Multiple General Hetero Multiple General Hetero ND Single #
[117] Two-Tier Multiple General Hetero Multiple General Hetero ND ND #
[90] Three-Tier Multiple General Hetero Multiple General Hetero ND Multiple #
[14] Three-Tier Multiple General Hetero Multiple General Hetero # Single #
[58] Three-Tier Multiple General Hetero Multiple General Hetero # Single #
[114] Two-Tier Multiple Vehicle Hetero Multiple General, Vehicle Hetero Intra # #
[6] Three-Tier Multiple General Hetero Multiple General Hetero # Single #
[16] Three-Tier Multiple General ND Multiple General Hetero Intra Single #
[22] Two-Tier Multiple General Homo Multiple General Hetero Intra # #
[4] Three-Tier Multiple General Hetero Multiple General Hetero # Single #
[50] Three-Tier Multiple General ND Multiple General Hetero # Single #
[59] Three-Tier Multiple General Hetero Multiple General Hetero # Single #
[61] Two-Tier Multiple General ND Multiple General Hetero Intra # #
[65] Three-Tier Multiple General Hetero Multiple General Hetero # Single #
[69] Two-Tier Multiple General Hetero Multiple General Hetero Intra # #
[82] Three-Tier Multiple General Hetero Multiple BS Hetero # Single #
[95] Three-Tier Multiple General Hetero Multipe General Hetero # Single #
[101] Many-Tier Multiple General Hetero Multipe General Hetero Intra Single #
[98] Three-Tier Multiple General ND Multipe General Hetero # Single #
[103] Many-Tier Multiple General Hetero Multipe General Hetero Intra Single #
[104] Three-Tier Multiple General Hetero Multipe General Hetero Intra Multiple

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:17

[108] Three-Tier Multiple General ND Multipe General Hetero Intra Single #

[115] Three-Tier Multiple ND ND Multipe General Hetero ND Single #
[116] Three-Tier Multiple General Hetero Multipe General Hetero ND Single #
[134] Three-Tier Multiple MD Hetero Single General Homo # Single #
[46] Three-Tier Multiple General Hetero Multiple General Hetero Intra Single #
[102] Three-Tier Single MD Homo Multiple General Hetero Intra Single #
Hetero: Heterogeneity/Heterogeneous, Homo: Homogeneous, Coop: Cooperation, MD: Mobile Device, BS: Base Station,
MS: Macrocell Station, AP: Access Point, RSU: Roadside Units, Femto: Femtocell, : Yes, # : No

5 OPTIMIZATION CHARACTERISTICS
Considering the application structure, environmental parameters, and the target objectives, each
proposal formulates the problem of scheduling IoT applications in Fog computing. Optimization
parameters directly affect the selection process and properties of suitable decision engines. Fig. 6
presents the principal elements in optimization characteristics, as described in what follows:

5.1 Main Perspective

The proposals in the literature can be divided into three categories based on their main optimization
goal, namely IoT devices/users, system, and hybrid, which are described in the following:
5.1.1 IoT devices/users. The main perspective of several proposals is to satisfy the requirements of
IoT applications such as minimizing their execution time and energy consumption of IoT devices,
or improving user experience in terms of QoS and Quality of Experinece (QoE). Several works have
considered IoT perspective for the optimization such as [13, 73, 114, 114, 124, 134].
5.1.2 System. The main perspective of this category is to improve the efficiency of resource
providers such as minimizing their energy consumption, improving resource utilization, and
maximizing the monetary profit [12, 19, 28, 50]. Hence, these works often assume IoT devices with
very limited limited computational resources that transfer sensed data to the surrogate servers for
processing and storage.
5.1.3 Hybrid. Some proposals targeted optimizing the parameters of both IoT devices/users and
resource providers, referred to as hybrid optimization [55, 82, 98, 144]. In these works, IoT devices
have some computational resources to serve their partial/complete tasks. However, they may send
their requests to other surrogate servers if overall global parameters of IoT devices and systems
can be optimized.

5.2 Objective Number

According to the number of main optimization objectives of each proposal, we classify the current
literature into 1) single objective and 2) multi objective proposals. Multi objective proposals
consider several parameters to simultaneously optimize them, incurring higher complexity. In the
literature, several proposals targeted single objective optimization such as [56, 69, 114, 157], while
other proposals try to optimize several parameters such as [38, 46, 95, 145].

5.3 Parameters
According to the main objectives and the nature of optimization parameters in the literature, we
categorize these parameters into the following categories:
5.3.1 Time. One of the most important optimization parameters is the execution time of IoT
applications. Minimizing the execution time of IoT applications provides users with a better QoS
and QoE. This category contains any metrics related to time such as response time, execution time,
and makespan used in the literature such as [24, 45, 103, 135].

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:18 M. Goudarzi et al.

IoT
Main Perspective System
Hybrid
Single
Objective Number
Multi
Time
Energy

Characteristics
Optimization
Parameters Monetary Cost
Hybrid
Other
ILP
MILP
Problem Modeling MINLP
MDP
Other
Deadline
QoS Constraints Energy
Hybrid

Fig. 6. Optimization characteristics taxonomy

5.3.2 Energy. IoT devices are usually considered as battery-limited devices. Hence, minimizing
their energy consumption is one of the most important optimization parameters. Besides, energy
consumption from FSs’ perspective is two-fold. First, some FSs, similar to IoT devices, are battery-
constrained, making optimizing the energy consumption of FSs an important challenge. Second,
from the system perspective, the energy consumption of FSs should be minimized to reduce carbon
emissions. This category contains any proposals considered energy consumption as an optimization
parameter either from IoT devices or system perspectives such as [13, 54, 59, 110].

5.3.3 Monetary Cost. This category studies the proposals that have considered the monetary
aspects either from IoT users (i.e., minimizing monetary cost) or system perspectives (i.e., increasing
monetary profit) [12, 74, 94, 94, 108].

5.3.4 Other. Some works have considered other optimization parameters such as the number of
served requests, system utility, and resource utilization, just to mention a few, such as [19, 28, 37,
115].

5.3.5 Hybrid. Several works also have considered a set of optimization parameters, referred
as hybrid. These works use any combination of above-mentioned parameters simultaneously
[39, 40, 46, 134].

5.4 Problem Modeling

Considering the main goal and optimization parameters, the optimization problem can be mod-
eled/formulated. Considering surveyed literature in terms of the problem modeling approach, we
classify the works into the following categories:

5.4.1 Integer Linear Programming (ILP). It is a problem type where the variables and constraints
are all integer values, and the objective function and equations are linear. Several works have used
ILP for problem modeling such as [74, 90, 97, 136].

5.4.2 Mixed Integer Linear Programming (MILP). In these problems, only some of the variables are
constrained to be integers, while other variables are allowed to be non-integers. Also, the objective
function and equations are linear. Several works have modelled their problem as an MILP such as
[55, 70, 104, 145].

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:19

5.4.3 Mixed Integer Non-Linear Programming (MINLP). It refers to problems with integer and
non-integer variables and non-linear functions in the objective function and/or the constraints.
Several works such as [22, 84, 147, 156] have used MINLP to present their optimization problems.
5.4.4 Markov Decision Process (MDP). It provides a mathematical framework to model and analyzes
problems with stochastic and dynamic systems. Several works have used MDP to model scheduling
problem in Fog computing such as [9, 39, 115, 134].
5.4.5 Other. There are also some other optimization modeling approaches in the literature of
scheduling applications in Fog computing such as game theory [49], lyapunov [19, 100], and mixed
integer programming [6, 116].

5.5 QoS Constraints

The formulated optimization problem usually contains several constraints, incurring higher com-
plexity compared to unconstrained problems. In this work, we classify techniques based on the QoS-
related constraints applied to the main formulated problem into 1) deadline such as [32, 128, 140],
2) energy such as [11, 146], and 3) hybrid (i.e., any combination of deadline, energy, and cost)
such as [22, 46, 100].

5.6 Discussion
In this section, we discuss the effects of identified optimization characteristics’ elements on the
decision engine and describe the lessons that we have learned. Besides, we identify several research
gaps accordingly. Table 4 provides a summary of characteristics related to optimization problems
in Fog computing.
5.6.1 Effects on the decision engine. The optimization characteristics affect the decision engine
in various aspects, as briefly described below.
1. Objective number and parameters: Simultaneous optimization of multi-objective problems
usually incur higher complexity for the decision engine. Also, when the number of key parameters
in a multi-objective scheduling problem increases, finding the best parameters’ coefficients becomes
a critical yet challenging step.
2. Problem modeling: It can affect the choice of placement technique as some specific algorithms
and techniques can be used to solve the scheduling problem. For example, several traditional LP
and ILP tools and libraries exist to solve LP and ILP scheduling problems.
3. QoS constraints: They incur hard or soft constraints and limitations on the main objec-
tive/objectives of the scheduling problem, which intensify the complexity of the scheduling problem.
The decision engine should satisfy these constraints either using classic Constraint Satisfaction
Problem (CSP) techniques or using customized approaches.
5.6.2 Lessons learned. Our findings regarding the optimization characteristics in the surveyed
works are briefly described in what follows:
1. The main perspective of optimization for almost 75% of works is IoT, while for the rest of the
works is hybrid and system by 15% and 10%, respectively. The main perspective element affects
how some metrics are evaluated. For example, when evaluating the energy consumption in the IoT
perspective, the energy consumed by the surrogate servers for the execution of tasks is overlooked.
However, in the system and hybrid perspectives, the energy consumption of all resource providers
and all entities in the systems are evaluated, respectively.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:20 M. Goudarzi et al.

2. Considering objective numbers in the optimization problem, the works are almost equally
divided into single and multiple objective numbers. Overall, the majority of works studied time
and/or energy as their main optimization metrics. While the works with an IoT perspective follow
the same trend for the optimization metrics, the proposals with a system perspective almost consider
the cost as their main optimization parameter. Also, the hybrid perspective proposals often consider
a combination of time, energy, and/or cost as their main optimization metrics.
3. In problem modeling, the greater number of works have used either MDP or MINLP (each
with roughly 25% of proposals) to formulate their problem. Also, some works initially had modeled
their work as MINLP and then defined the MDP accordingly, such as [22, 56]. The rest of the works
have used MILP (almost 15%), ILP (almost 15 %), and other optimization modeling approaches.
4. Almost 25% of works defined single or multiple QoS constraints for their problem, among
which 90% have considered a single constraint, and the rest went for two QoS constraints. Among
the QoS constraints, the deadline by 90% is the most used constraint in all works.
5.6.3 Research Gaps. We have identified several open issues for further investigation that are
discussed below:
1. The main part of works in the literature either consider optimization problems from IoT
devices/users. However, only a few works have considered IoT and system perspectives simultane-
ously (i.e., hybrid). Optimizing either of these perspectives can negatively affect other perspectives.
To illustrate, when the principal target is minimizing the energy consumption of IoT devices, the
majority of components or tasks are placed at FSs or CSs. However, it may negatively affect the
energy consumption of resource providers and even increase the aggregated energy consumption
in the environment. Hence, further investigation on hybrid optimization perspectives and mutual
effects of different perspectives is required.
2. The cooperation among the resource providers (i.e., FSs, CSs) is an essential factor in offering
higher-quality services. Proposals in the system and hybrid perspectives can also consider other
metrics such as trust and privacy index for resource providers and study how they affect the overall
performance.
3. QoS constraints are set to guarantee a minimum service level for end-users. In current literature,
most of proposals have focused on the deadline as the constraint. However, several other parameters
such as privacy, security, and monetary cost and their combination as hybrid QoS constraints are
not studied.

6 DECISION ENGINE CHARACTERISTICS

The requirements of IoT applications in Fog computing can be satisfied if incoming IoT requests
can be accurately scheduled based on the characteristics of application structure, environmental
architecture, and optimization problems by the decision engine. The main responsibilities of the
decision engine are organizing received IoT requests and solving the optimization problem through
a placement decision while considering contextual information. Fig. 7 presents the main elements
in decision engine, as described in what follows:

6.1 Deployment Layer

Decision engines can be deployed on servers at different layers unless the servers do not have
sufficient resources to host them. Based on the deployment layer of the decision engine, the current
literature can be classified into:

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:21

Table 4. Summary of existing works considering optimization Characteristics taxonomy

Optimization Characteristics Optimization Characteristics

Ref Ref
Main Object Prob QoS Main Object Prob QoS
Metrics Metrics
Persp Number Model Const Persp Number Model Const
[9] System Single Cost MDP # [81] IoT Single Time ILP Deadline
Time,
[13] IoT Single Energy MINLP # [27] IoT Multiple
Cost
MINLP Deadline
Time,
Energy,
[18] IoT Multiple Energy, IP # [55] Hybrid Multiple
Time
MILP #
Cost
Enery,
[25] IoT Single Cost MINLP Deadline [89] IoT Multiple
Time
MDP #
Time,
Energy, Energy,
[39] IoT Multiple
Weighted
MDP # [20] IoT Multiple
Time
MDP #
Cost
Time,
Energy, MINLP,
[40] IoT Multiple
Weighted
MILP # [56] IoT Single Energy
MDP
#
Cost
Energy,
[84] IoT Single Time MINLP # [73] IoT Multiple
Time
MDP #
Time,
[70] IoT Single Time MILP # [144] Hybrid Multiple Resource MDP #
Utilization
Time,
Time,
[129] Hybrid Multiple Energy, MDP # [31] Hybrid Multiple
Cost
MDP Deadline
Cost
Energy,
[156] IoT Single Time MINLP # [72] Hybrid Multiple
Cost
MDP #
Time,
[47] IoT Single Time MDP # [38] IoT Multiple
Eneegy
ILP #
Time,
[54] IoT Multiple Comput MDP # [141] IoT Single Time ND #
Ratio
Time,
Energy, Cost,
[147] IoT Multiple MINLP Deadline [100] Hybrid Single Time Lyapu
Weighted Deadline
Cost
[137] IoT Single Time MDP # [119] IoT Single Time ND #
Time,
[140] Hybrid Single Energy MINLP Deadline [133] Hybrid Multiple MINLP Deadline
Energy
Time,
[97] IoT Single Energy ILP # [71] IoT Multiple
Energy
MINLP Deadline
[125] IoT Single Time MDP # [148] IoT Single Time MINLP Deadline
Time,
[127] IoT Multiple
Cost
ND # [109] IoT Single Time MDP #
Time/ Quality
[32] Hybrid Single MILP Deadline [157] IoT Single Time MILP
Enegy Loss
Time,
[146] IoT Multiple
Energy
MINLP Energy [153] IoT Single QoS ND #
Cost,
[136] IoT Multiple
Time
ILP # [111] IoT Multiple ND ND #
Served Time,
[28] System Single
Requests
MILP # [120] IoT Multiple
Energy
ND #
[132] IoT Single Time MINLP # [143] IoT Single Time MDP #
[138] IoT Single Time MDP # [83] IoT Single Time ND #
[42] IoT Single Energy MINLP Deadline [117] IoT Single Bandwidth ND #
Time,
Cost,
[45] IoT Single Deadline LP # [90] IoT Multiple
Weighted
ILP #
Cost
[149] IoT Single Time MINLP # [14] IoT ND Time ND #
Time, MINLP,
[53] IoT Multiple
Energy MDP
# [58] IoT Single Time ND #
Time,
[64] Hybrid Multiple Enery, ND # [114] IoT Single Time ND #
Cost

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:22 M. Goudarzi et al.

Offloaded
[128] IoT Single
Task
MILP Deadline [6] IoT Single Time MIP #
Time,
[33] Hybrid Multiple
Energy
MILP # [16] IoT Single Time ND #
Stystem Time, MINLP, Deadline,
[19] System Single
Utility
Lyapu # [22] IoT Multiple
Energy MDP Energy
[11] Hybrid Single Time IP Energy [4] IoT Single Time ND #
Time,
[12] System Single Cost MILP # [50] System Multiple
Energy
ILP #
[21] IoT Single Cost MILP Deadline [59] System Single Energy ND Deadline
[74] System Single Cost ILP # [61] IoT Single Time ND #
[30] IoT Single Time ND # [65] IoT Single Time ND #
[155] IoT Single Time ND # [69] IoT Single Time MINLP #
Time,
[135] IoT Single Time MDP # [82] Hybrid Multiple
Cost
MINLP #
Time,
[110] IoT Single Energy MDP Deadline [95] Hybrid Multiple Energy, ND Deadline
Cost
[88] IoT Single Time ND Deadline [101] IoT Single Time ND #
Time,
Time,
[142] IoT Multiple
Energy
ILP # [98] hybrid Multiple Energy, ND Deadline
Cost
[24] IoT Single Time ND # [103] IoT Single Time ND #
Time, Game Time,
[49] IoT Multiple
Energy Theory
# [104] IoT Multiple
Energy
MILP Deadline
Time,
Energy,
[145] IoT Multiple
Weighted
MILP # [108] IoT Single Cost ILP Deadline
Cost
Served
Time, Requests,
[57] IoT Multiple
Energy
MINLP # [115] Hybrid Multiple
Fog
MDP #
Numbers
Cost,
[80] IoT Single QoS ILP
Deadline
[116] IoT Single Time MIP #
Time,
Energy,
[94] IoT Single Cost Lyapu Deadline [134] IoT Multiple
Weighted
MDP #
Cost
Time,
Energy, Deadline,
[124] IoT Single Time ND # [46] IoT Multiple
Weighted
MDP
Energy
Cost
Time, Cost,
[78] System Multiple
Resource
ILP # [102] IoT Multiple
Energy
ND Deadline
Main Pers: Main Perspective, Object number: Objective Number, Prob Model: Problem Model, QoS Const: QoS Constraints
Cost: Monetary Cost, MDP: Markov Decision Process, ILP: Integer Linear Programming, MINLP: Mixed Integer Non-Linear Programming,
MIP: Mixed Integer Programming, MILP: Mixed Integer Linear Programming, ND: Not Defined, Lyapu: Lyapunov, #: No

6.1.1 IoT Layer. The IoT devices usually are considered as resource-limited and battery-constrained
devices. Hence, decision engines running on IoT devices should be very lightweight even with
compromising the accuracy. In the literature, several works such as [56, 88, 97, 125] deployed
decision engines at the IoT layer.
6.1.2 Fog/Edge Layer. Distributed FSs with sufficient resources situated in the proximity of IoT
devices are the main deployment targets for the decision engines. They provide low-latency and
high-bandwidth access to decision engines for IoT devices. Majority of works such as [39, 64, 89, 102]
deployed the decision engines in Edge/Fog Layer.
6.1.3 Cloud Layer. CSs are potential targets for the deployment of decision engines. Although the
access latency to CSs is higher, they provide high availability, making them a suitable deployment
target where FSs are not available or when IoT applications are insensitive to higher startup time.
Some works such as [27, 119] considered cloud layer for the deployment of decision engines.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:23

IoT
Deployment Layer Edge/Fog
Cloud
FIFO
Queuing Priority-based
Admission Control Single
Dispatch Mode
Batch
Exact
Decision Engine

Direct Optimization
Characteristics

Traditional Heuristics Approximation

Meta-heuristics
Placement Technique
Supervised
Machine Learning Unsupervised Multi-Arm Bandit
Reinforcement Learning Deep RL
Mobility Support
Distributed Deep RL
Failure Recovery
Advanced Features
Scalability
Adaptability Java
Python
Language
C++
Other
Implementation Source Code
Low
Time Complexity Medium
High

Fig. 7. Decision engine taxonomy

6.2 Admission Control

The admission control presents the behavior of decision engines when new requests arrive. It
denotes how the new requests are queued and organized by the dispatching module for placement.

6.2.1 Queuing. Decision engine may use different queuing policies when incoming IoT requests
arrives. Based on queuing policy, we classify works into 1) First-in-First-Out (FIFO) such as
[32, 39, 125] and 2) Priority-based where incoming requests are sorted based on their priority
(e.g., deadline) [45, 50, 128].

6.2.2 Dispatching Mode. The dispatching module forwards requests from input queue to the
placement module. Based on the selection policy of dispatching module, current literature can be
classified to 1) single model where only one task at a time is dispatched for placement [22, 47, 143]
and 2) batch model where a set of tasks are forwarded to placement module [4, 40, 98].

6.3 Placement Technique

Placement technique is the actual algorithm used to solve the optimization problem. Each placement
algorithm has its advantages and disadvantages. Hence, it should be carefully selected based on the
properties and dynamics of applications, users, environment, and deployment layer. We classify
placement techniques based on their approach to find the solution into two broad categories:

6.3.1 Traditional. In this approach, the programmer/designer defines the required logic of policies
for the placement technique. The traditional placement technique can be further divided into three
subcategories:

1. Direct Optimization: In this category, the optimization problem will be solved using classical
optimization tools either using 1) Exact approach to find the optimal solution such as [18, 25] or 2)
Approximation approach to find a near optimal solution such as [11, 84, 136].

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:24 M. Goudarzi et al.

2. Heuristics: These algorithms are a set of typically problem-dependent algorithmic steps to

find a feasible solution for the problem. Heuristics usually scale well as their TC is low while they
do not guarantee finding the optimal solution of the problem [13, 14, 28, 70].
3. Meta-heuristics: Meta-heuristics are composed of several advanced heuristics and typically
are problem-independent, such as Genetic Algorithm (GA) and Simulated Annealing (SA). Although
these algorithms usually perform better than heuristics, similarly they cannot guarantee to find the
optimal solution. Several works such as [33, 40, 90, 132] used meta-heuristics.
6.3.2 Machine Learning (ML). ML is a family of algorithms that can learn the required policies
for placement techniques from historic data. ML algorithms scale reasonably well, however, they
require accurate and ideally large samples of historic data. The ML-based placement techniques
can be further divided into three subcategories:
1. Supervised Learning: These algorithms learn by using labeled data as its input. Type of
problems are regression and classification, and some of the algorithms are linear regression and
logistic regression. Some works in the current literature used supervised ML for the placement
technique such as [57, 146, 147].
2. Unsupervised Learning: These algorithms are trained using unlabelled data, contrary to
supervised ML, without any guidance. Some unsupervised algorithms are K–Means and fuzzy
C–Means. Some works in the current literature used unsupervised ML for the placement technique
such as [117, 120, 143].
3. Reinforcement Learning (RL): In these algorithms, agent/agents learn the required policy
for placement technique by interaction with an uncertain and potentially complex environment. It
does not require pre-defined data, and type of problems are exploitation or exploration. The current
RL-based literature in scheduling IoT applications can be divided into 1) Multi-Armed Bandit
(MAB) which are among the simplest RL problems such as [127, 156], 2) Deep RL (DRL) where
deep learning is used in RL such as [9, 89, 125], and 3) Distributed DRL where several agents
works collaboratively in a distributed manner for efficient learning such as [22, 39, 129].

6.4 Advanced Features

To fully utilize the potential of the Fog computing paradigm, several advanced features can be
augmented with decision engines to capture high dynamics of this paradigm, described below:
6.4.1 Mobility Support. A significant number of IoT devices are moving entities (e.g., vehicles),
requiring connected service through their path. So, decision engines should manage the migra-
tion process of application components and find suitable surrogate servers accordingly. Several
works such as [11, 38, 90, 138] address mobility and migration management challenges alongside
scheduling IoT applications.
6.4.2 Failure Recovery. In highly dynamic systems such as Fog computing, failure may happen
due to software or hardware-related issues. So, application components faced with failure should
be re-executed. Some works consider failure recovery mechanisms in their decision engines such
as [16, 38, 40].
6.4.3 High Scalability. As a large number of IoT devices and servers exist in the Fog environment,
mechanisms and algorithms used in the decision engine should be highly scalable and provide
well-suited performance when the system size grows. Several works have studied the scalability
feature of their techniques when the number of IoT applications and servers increases or discussed
how their distributed techniques work efficiently in large systems such as [30, 53, 73].

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:25

6.4.4 High Adaptability. This feature ensures that the decision engine dynamically captures the
contextual information (i.e., application, environment, etc), and updates the policies of placement
techniques accordingly. In Fog literature, several works such as [31, 39, 108, 153] offer solutions
with high adaptability.

6.5 Implementation
The implementation characteristics of decision engines are studied based on the following criteria:
6.5.1 Language. Different programming languages are used for the implementation of decision
engines, while the majority have used Python [39, 74], Java [40, 119], and C++ [12, 124].
6.5.2 Source Code. Open-source decision engines help researchers to understand the detailed
implementation specifications of each work, and minimize the reproducibility effort of decision
engines, especially for comparison purposes. Some works such as [24, 110, 134] have provided the
source code repository of their decision engines.
6.5.3 Time Complexity (TC). TC of each placement technique presents the required time to solve
the optimization problem in the worst-case scenario. It directly affect the service startup time and
the decision overhead of each technique. Based on the current literature, we classify the TC into 1)
Low the solution of optimization problem can be obtained in polynomial time where the maximum
power of variable is equal or less than two (i.e., 𝑂 (𝑛 2 )) [9, 16, 39], 2) medium where the time
complexity is polynomial and the maximum power is less than or equal to 3 (i.e., 𝑂 (𝑛 3 )) [13, 70, 104],
and 3) High for exponential TC and polynomials with high maximum power [18, 25, 128].

6.6 Discussion
In this section, we describe the lessons that we have learned regarding identified elements in
decision engine characteristics of the current literature. Besides, we identify several research
gaps accordingly. Table 5 provides a summary of decision engines-related characteristics in Fog
computing.
6.6.1 Lessons learned. Our findings regarding the decision engine characteristics in the surveyed
works are briefly described in what follows:
1. Almost 85% of surveyed works deployed the decision engine at the Edge layer in the proximity
of IoT devices. Since the Edge servers can be accessed with lower latency and higher access
bandwidth, deployment of decision engines at the Edge can reduce the startup of IoT applications.
However, the Edge devices should have sufficient resources to run the decision engine. Some
proposals (about 10%) also deployed the decision engine on IoT devices. Deployment of a decision
engine on IoT devices provides more control for IoT devices, especially mobile ones. It eliminates
the extra overhead of communication with surrogate servers for making a decision. However, IoT
devices often have very limited resources that are incapable of running powerful decision engines.
2. The queuing is an important element in the admission control that almost 80% of the works
have not studied. Since most of works have considered several IoT devices in the environment,
several IoT requests may arrive in each decision time-slot with a high probability. Hence, different
queuing models can dramatically affect the decision engine performance and the QoS of end-
users. FIFO and priority queue share the same proportion of proposals among the works that
mentioned their queuing policy. Also, in priority-based queuing, almost all works have considered
the deadline of applications or tasks as their main priority metric. Moreover, for the policy of
dispatching module, about 75% of works selected single dispatching while 25% of works studied
batch dispatching policy. Since different IoT requests may arrive in the same decision time-slot, the

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:26 M. Goudarzi et al.

batch dispatching policy helps study the mutual effects of IoT applications with diverse resource
requirements in the placement decision.

3. The traditional placement techniques are used in almost 60% of the proposals, while the
ML-based placement techniques are studied in the rest of the works (almost 40%). However, the
number of ML-based placement techniques has significantly increased in recent years. In traditional
placement techniques, direct optimization, heuristics, and meta-heuristics share the same proportion
of proposals. Also, in meta-heuristics techniques, the majority of works used population-based
meta-heuristics, especially different variations of the GA. In the ML-based techniques, the majority
of proposals have used RL-based techniques (almost 70%), specially DRL. Moreover, in the DRL
techniques, the larger number of works used centralized DRL techniques such as DQN. However,
the exploration and convergence rate of centralized DRL techniques are very slow. Thus, several
studies have recently been conducted to adapt distributed DRL (i.e., DDRL) techniques for resource
management in Edge/Fog computing environments, such as [39, 53, 73], to improve the exploration
cost and convergence rate of the DRL techniques.

4. In advanced features, almost 25% of proposals embedded different mechanisms (i.e., traditional
or ML-based) for the mobility management of IoT devices and migration of applications’ constituent
parts. Also, about 25% of studied works, mostly ML-based techniques, offer high adaptability features
in their decision engine. However, traditional works often neglect to provide different mechanisms
to support high adaptability. This is mainly because the scheduling policies are not statically defined
in ML-based techniques. Hence, as the environmental or application properties change, the policies
can be learned and updated accordingly. However, in the traditional scheduling techniques, updating
the scheduling policies according to dynamic changes in environmental or application properties is
very costly and time-consuming. Almost 20% of proposals studied different mechanisms to support
high scalability feature, either using ML-based techniques or traditional approaches. In advanced
features, the failure recovery mechanisms and techniques in scheduling are not well-studied and
only a few works embedded these mechanisms in their scheduling techniques.

5. Considering the implementation of the techniques, almost 50% of the works mentioned their
employed programming language. Java and python programming language are the most-employed
programming language and are almost equally used in different proposals. However, Python is
mainly used for ML-based techniques and direct optimization techniques, while Java is mostly used
to implement traditional decision engines. Moreover, only about 10% of proposals shared their
open-source repositories with researchers and developers among the surveyed works. Finally, about
65% of proposals discussed the TC of their works, among which almost 80% proposed decision
engines with low TC while some proposals (almost 10%) went for medium TC and few works
(almost 10%) proposed decision engines with high TC. The high TC proposals are among the direct
optimization category of traditional approaches. While these high TC proposals cannot be currently
adapted to large-scale Edge and Fog computing environments, they can find the optimal solution in
small-scale problems. Hence, they can be used as a reference for the evaluation of other proposals.

6.6.2 Research Gaps. We have identified several open issues for further investigation that are
discussed below:

1. The admission control concept in terms of different queuing, dispatching, and their mutual
effect is not well studied in the current literature. Also, the greater number of works consider a single
task dispatching model and overlook batch placement of applications, especially for applications
with dependent tasks.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:27

2. While traditional placement techniques (e.g., heuristics, meta-heuristics) are studied well in
the literature, ML-based techniques are still in their infancy. Due to the lack of a large number of
datasets, supervised and unsupervised ML have not been thoroughly considered. Also, the majority
of employed RL techniques are centralized approaches, neglecting collaborative learning of multiple
distributed agents for better efficiency and lower exploration costs.
3. Although all servers and devices are prone to failures, among advanced features, failure
recovery mechanisms, algorithms, and their integration with the placement technique is the least-
studied concept. Even the best placement techniques cannot complete their process in real-world
scenarios unless a suitable failure recovery mechanism is embedded.
4. In the surveyed works, there is no proposal to study all the four identified elements in the
advanced features (i.e., mobility, failure recovery, scalability, and adaptability) and describe the
behavior and mutual effects of these elements on each other and decision engine.
5. Among the studied literature, none of the works has studied the privacy problem from different
perspectives, such as end-users’ data privacy, resource providers’ privacy, and the decision engine’s
mechanisms for improving privacy.

7 PERFORMANCE EVALUATION
Different approaches and metrics have been used by the research community to evaluate the
performance of their techniques. Identifying and studying these parameters helps to select the best
approach and metrics for the implementation of new proposals and fair comparisons with other
techniques in the literature. Fig. 8 presents a taxonomy and the main elements of performance
evaluation, described below:

7.1 Approaches
The performance evaluation approaches can be divided into four categories, namely analytical,
simulation, practical, and hybrid. There are different important aspects to consider when selecting
an approach for the evaluation of proposals, such as credibility, implementation time, monetary
cost, reproducibility time, and scalability. Fig. 9 presents a qualitative comparison of different
approaches used in performance evaluation.
7.1.1 Analytical. One of the popular approaches for the evaluation of different proposals is analyt-
ical tools. Usually, the implementation time, reproducibility time, and monetary cost of analytical
tools are low, and scalable experiments can be executed. However, the credibility of such ex-
periments is low since the dynamics of resources, applications, and environment cannot be fully
captured and tested. Matlab is among the most popular tools that is either used directly [28, 147, 147]
or integrated with some other libraries such as Sedumi1 [116]. Also, C++ based analytical tools
have been used in the literature, such as [12, 84].

1 https://github.com/sqlp/sedumi

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
 Table 5. Summary of existing works considering decision engine taxonomy
ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.

1:28
Decision Engine Characteristics
Ref Deployment Admission Control Advanced Features Implementation
Placement Technique
Layer
Mobility Failure High High Source
Queuing Dispatch Language Time Complexity
Support Recovery Scalability Adaptability Code
[9] Edge ND Single ML, RL, DRL, (DQN) # # # ND # Low (MP 2)
[13] Edge ND ND Tr, H, Greedy # # # ND # Medium (MP 3)
[18] Edge ND ND Tr, DO, Exact # # # ND # High (Exp)
[25] Edge FIFO Single Tr, DO, Exact, (BB) # # # # Java # High (Exp)
[39] Edge FIFO Single ML, RL, DDRL, (IMPALA) # # Python # Low (MP 2)
[40] Edge FIFO Batch Tr, MetaH, (MA) # # Java # Low (MP 2)
[84] Edge ND Batch Tr, DO, Approx # # # C++ # ND
[70] Edge FIFO Single Tr, H # # # # Java # Medium (MP 3)
[129] Edge ND ND ML, RL, DDRL, (A3C) # Java # Low
[156] Edge ND Batch ML, RL, MAB # # # ND # Low
[47] Edge ND Single ML, RL, DRL, (DQN) # # # ND # Low
[54] Edge ND ND ML, RL, DRL # # # ND Low
[147] Edge ND Single ML, Sup, (DeepL) # # # # ND # Low
[137] Edge ND Single ML, RL, (Q-learning) # # ND # Low
[140] Edge ND Single ML, Sup, (Imitation) # # # ND # Medium
[97] IoT ND Single ND # # # # Android/Java ND
[125] IoT FIFO Single ML, RL, DRL (Double DQN) # # ND # Low
[127] IoT ND Single ML, RL, MAB # ND # Low
[32] Edge FIFO Single Tr, H # # # # Android/Java # Low
[146] Edge ND Single ML, Sup, (Imitation) # # # # ND # Low
[136] Edge ND Single Tr, DO, Approx # # # ND # Low (MP 2)
[28] Edge ND Single Tr, H, Greedy # # # # ND # High
[132] Edge ND Batch Tr, MetaH, (SA) # # # # ND # ND
[138] Edge ND ND Tr, DO, Approx # # # ND # Medium
[42] Edge ND Single Tr, MetaH (GA-PSO) # # # # ND # Low (MP 2)
[45] Edge Priority Single Tr, DO, Approx # # # ND # ND
[149] Edge ND Single Tr, H # # # ND # ND
[53] Edge ND Single ML, RL, DRL, DDRL # # ND # Low
[64] Edge ND Single Tr, MetaH, (NSGA2) # # # Java # ND
[128] Edge Priority Single Tr, H # # # # ND # High (MP 5)
[33] Edge ND Batch Tr, MetaH, (Ant Mating) # # # # ND # ND
[19] Edge ND ND Tr, H # # # # ND # ND
[11] Edge ND Single Tr, DO, Approx, (SAA) # # # ND # ND

M. Goudarzi et al.
[12] ND ND Batch Tr, H # # # # C++ # Low
[21] Edge ND Single Tr, DO, Approx # # # # ND # Low (MP 2)
[74] Edge ND Batch Tr, DO, Approx # # # # Python # Low (MP 2)
[30] Edge ND Single Tr, DO, Approx # # # Python # ND
[155] Edge ND Single Tr, DO, Approx # # # # ND # ND
[135] Edge ND Single ML, RL, DRL, (PPO) # # # Python # Low
 Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions
[110] Hybrid ND Single ML, RL, DDRL # # Python Low
[88] IoT ND Single Tr, DO, Approx # # # # ND # ND
[142] Edge ND Single Tr, Other, (Min-cut) # # # # Java Low (MP 2)
[24] Edge FIFO Single Tr, MetaH, (GA) # # Python Low
[49] Edge ND Single Tr, DO, Approx # # # # ND ND ND
[145] Edge ND Batch Tr, MetaH, (NSGA3) # # # # Java # ND
[57] IoT ND Single ML, Sup, DDeepL # # # Python Low
[80] Edge ND Single Tr, DO, Exact # # # # Java # high
[94] Edge ND Single Tr, DO, Approx # # # # Python # ND
[124] Edge ND Single Tr, H # # # # C++ # Low
[78] Edge ND Single Tr, H # # # # Java # ND
[81] Edge ND Single Tr, H # # # # Java # Low
[27] Cloud ND Single Tr, DO, Approx # # # # ND # ND
[55] IoT ND Single ML, Sup, (DDeepL) # # # Python # Low
[89] Edge ND Single ML, RL, DRL, (DQN) # # # ND # Low
[20] Edge FIFO Single ML, RL, DRL, (DoubleDQN) # # # ND # Low
[56] IoT ND Batch ML, RL, DRL, (DQN) # # # ND # Low
ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.

[73] Edge ND Single ML, RL, DDRL, (D3PG) # # ND # Low

[144] Edge FIFO Single ML, RL, DRL, (DQN) # # # Python # Low
[31] Edge ND Single ML, RL, DRL, (DoubleDQN) # # # Python # Low
[72] ND ND Single ML, RL, DRL, (DQN) # # # Java # Low
[38] Edge FIFO Single Tr, H # # Java Low (MP 2)
[141] Edge ND Single Tr, H # # # ND # ND
[100] Edge ND Single Tr, DO, Approx # # # Java # ND
[119] Cloud ND Single ML, Sup, (Gradient Tree Boosting) # # # Java Low
[133] Edge ND Single Tr, MetaH, (GA) # # # ND # Low (MP 2)
[71] ND ND Batch Tr, MetaH, (GA) # # # ND # ND
[148] IoT FIFO Batch Tr, DO, Approx # # ND # Low
[109] IoT ND Batch ML, RL, DDRL, (A3C) # ND # ND
[157] Edge ND Batch Tr, DO, Approx # # # ND # ND
[153] Edge ND Single ML, RL, DRL, (DQN) # # Python # Low
[111] Edge ND Single ML, Sup, (Regression Tree) # # # # Java # ND
[120] ND ND Single ML, Unsup, (AHP) # # # # ND # ND
[143] Edge ND Single ML, Unsup, (Baum-Welch Algorithm), # # # ND # ND
[83] Edge ND Batch ML, Unsup, (K-means) # # # # ND # ND
[117] ND ND Single ML, Unsup, (K-means) # # # # ND # ND
[90] Edge ND Single Tr, MetaH, (Tabu) # # # ND # ND
[14] Edge Priority Single Tr, H, (Spring Algorithm) # # # # ND # ND
[58] Edge ND Batch Tr, MetaH, (ACO) # # # # ND # ND
[114] Edge ND Batch Tr, MetaH, (MA) # # # ND # Low (MP 2)
[6] Edge Priority Batch Tr, MetaH, (GA) # # # ND # ND
[16] Edge Priority Single Tr, H # # # Python # Low (MP 2)
[22] IoT ND Single ML, RL, DDRL, (DDPG) # # Python # Low
[4] Edge ND Batch Tr, MetaH, (AEO) # # # # ND # ND
[50] Edge Priority Batch Tr, MetaH, (GA) # # # # ND # ND

1:29
ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.

1:30
[59] Edge ND Single Tr, H # # # # ND # Low (MP 2)
[61] Edge Priority Batch Tr, MetaH, (PSO) # # # # ND # Low
[65] Edge ND Single Tr, MetaH, (ACO) # # # # ND # Low
[69] Edge ND Batch Tr, MetaH # # # # Python # Low (MP 2)
[82] Edge ND Batch Tr, MetaH, (GA) # # # # Python # Low
[95] Edge ND Batch Tr, MetaH, (GA) # # # # Python # Low
[101] Edge FIFO Single Tr, H # # # # Java # Medium (MP of 3)
[98] Edge ND Batch Tr, MetaH, (SSA) # # # # Python ND
[103] Edge Priority Single Tr, H # # # # Java # ND
[104] IoT ND Single Tr, MetaH, (GA) # # # ND # Medium (MP 3)
[108] Edge Priority Single Tr, H, Greedy # # # Go Low
[115] Edge Priority Single ML, RL, DRL, (DQN) # # # Python # Low
[116] IoT ND Batch Tr, DO, Approx # # # ND # Low
[134] IoT ND Single ML, RL, DRL # # Python Low
[46] Edge ND Single ML, RL, DRL, (DQN) # # # Python # ND
[102] Edge ND Single Tr, MetaH, (SPEA) # # # # ND # Low (MP 2)
ND: Not Defined, ML: Machine Learning, Tr: Traditional, RL, Reinforcement Learning, DRL: Deep Reinforcement Learning, DDRL: Distributed Deep Reinforcement Learning,
MP: Max Power, H: Heuristics, MetaH: Metaheuristics, DO: Direct Optimization, BB: Branch and Bound, MA: Memetic Algorithm, FIFO: First-In-First-Out, Approx: Approximation,
MAB: Multi-Arm Bandit, A3C: Asynchronous Actor-Critic Agents, DeepL: Deep Learning, DDeepL: Distributed Deep Learning, Imitation: Imitation Learning, SA: Simulated Annealing,
GA: Genetic Algorithm, SAA: Sample Average Approximation, PPO: Proximal Policy Optimization, D3PG: Double-Dueling-Deterministic Policy Gradients, AHP: Analytic Hierarchy Process,
Sup: Supervised, Unsup: Unsupervised, ACO: Ant Colony Optimization, AEO: Artificial Ecosystem-based Optimization, Tabu: Tabu Search, PSO: Particle Swarm Optimization,
SSA: Sparrow Search Algorithm, SPEA: Strength Pareto Evolutionary Algorithms
: Yes, #: No

M. Goudarzi et al.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:31

Implementation
Analytical Time

Simulation
Approaches
Practical
Performance
Evaluation

Hybrid Monetary Cost Credibility

Decision Overhead
Time
Metrics Energy
Reproducibility
Monetary Cost Time
Scalability

Other Analytical Simulation Practical Hybrid

Fig. 8. Performance evaluation taxonomy Fig. 9. Performance evaluation approaches

7.1.2 Simulation. Simulators keep the advantages of analytical tools while improving the credibility
of evaluations by simulating the dynamics of resources, applications, and environments. In the
literature, iFogSim [43, 77] is among the most popular simulators for Fog computing [40, 64, 72, 129].
Besides, several researchers have used Cloudsim [17] such as [25, 145] or SimPy2 such as [31, 94]
to simulate their scenarios in Fog computing.
7.1.3 Practical. The most credible approach for the evaluation of proposals is practical implemen-
tation. However, due to high monetary cost, implementation time, and reproducibility time, it is not
the most efficient approach for different scenarios, especially evaluations requiring high scalability.
In the literature, few works such as [11, 32, 117, 119] evaluated their proposals using small-scale
practical implementations.
7.1.4 Hybrid. In this approach, researchers evaluate their proposals using practical implemen-
tations in small-scale and simulators or analytical tools in large-scale. Although implementation
and reproducibility time of this approach is high, it provides high scalability and credibility. In the
literature, few works such as [39, 74, 110] follow the hybrid approach.

7.2 Metrics
The metrics used in performance evaluation in Fog computing are directly or indirectly related to
the optimization parameters and system properties. Based on the nature and popularity of these
metrics in the literature, we categorize them into 1) time (e.g., deadline, response time, execution
time, makespan) [24, 25, 70, 89], 2) energy (e.g., battery percentage, saved energy) [13, 20, 55, 55],
3) monetary cost (e.g., service cost, switching cost) [9, 90, 94, 100], and 4) other metrics (e.g.,
number of interrupted tasks, resource utilization, throughput, deadline miss ratio) [14, 16, 38, 49].
Also, we consider 5) decision overhead as an important evaluation metric to study the overhead
of proposals (often in terms of time and energy), used in some works such as [39, 64, 102, 142].

7.3 Discussion
In this section, we describe the lessons that we have learned regarding identified elements in
the performance evaluation of the current literature. Besides, we identify several research gaps
accordingly. Table 6 provides a summary of characteristics related to performance evaluation in
Fog computing.
7.3.1 Lessons learned. Our findings regarding the performance evaluation in the surveyed works
are briefly described in what follows:
2 https://simpy.readthedocs.io/en/latest/

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:32 M. Goudarzi et al.

1. More than half of the works used the simulation as their performance evaluation approach
while 30% of the proposals used an analytical approach. The practical and hybrid approaches equally
share the rest of 20% of the works. For the analytical approach, the most of works used Matlab
or Python programming languages, while Java and Python are mostly used for the simulation
approach. In practical and hybrid approaches, Java and Python are equally employed in proposals.
2. As the performance evaluation metric, time and its variations (e.g., response time, makespan)
are used in more than 80% of the works. The second-highest-used metric is energy at 35%. However,
the decision overhead and cost are only studied in 15% of the works. Besides, less than 5% of the
proposals studied the performance of their scheduling technique using all the identified metrics.
7.3.2 Research Gaps. We have identified several open issues for further investigation that are
discussed below:
1. Although the monetary costs of sensors and edge devices (e.g., Rpi, Jetson Platform) have
reduced and they are highly available in different configurations, compared to a few years ago,
the majority of proposals still consider analytical tools and simulators as their only approach for
performance evaluation. While some works have considered practical and hybrid approaches for
the performance evaluation of their work, further efforts are required to study the dynamics of the
system, resource contention, and collaborative execution of the application in real environments,
especially considering new machine learning techniques such as DRL and DDRL [39, 110].
2. The decision overhead of proposals has direct effects on users and resources in terms of the
startup time of requested services and resource utilization. To illustrate, not only do healthcare
applications require low response time, but they also need low startup time, especially for critical
applications such as emergency-related applications (e.g., heart-attack prediction and detection).
Also, the overhead of proposals can severely affect the resource usage and energy consumption of
servers, especially battery-constrained ones. Among the techniques considered decision overhead as
a metric, they mostly focus on time while other metrics (e.g., energy, cost) need further investigation.

8 SCHEDULING TECHNIQUE: IMPORTANT DESIGN OPTIONS

In this section, we discuss the real-world characteristics of application structure and environmental
architecture and accordingly present several guidelines for designing a scheduling technique.
1. The number of IoT applications is constantly increasing in different application domains. The
majority of these applications are defined as a set of dependent modules/services [155]. Besides,
sharing and reusing modules/services for faster development and better management of applications
is of paramount importance. Moreover, dependent IoT applications are usually modeled as a graph of
tasks and their respective invocations. In this case, IoT applications with monolithic and independent
design can also be defined as an application graph with only one module and an application graph
with several modules where the size of invocations is zero, respectively. Hence, we consider IoT
applications with dependent modules/services (i.e., modular and loosely-coupled categories) as the
main architectural design choices in the application structure. Accordingly, the decision engine
requires a component for identifying and satisfying the constraint among modules/services.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:33

Table 6. Summary of existing works considering performance evaluation taxonomy

Performance Evaluation Performance Evaluation

Metrics Metrics
Ref Ref
Approach DO T E C O Approach DO T E C O
Hybrid
Sim Resource
[9]
(OPNET)
# # # # [81] (Sim, Prac) # # # Overhead
(iFogSim)
Weighted
[13] ND # # # # [27] Analytic # # # # Cost
[18] Analytic # # [55] Sim # #
Sim Task Drop
[25]
(Cloudsim)
# # # [89] Sim # # Rate
Hybrid
Weighted Task Drop
[39] (Sim, Prac), # Cost
[20] Sim # # Rate
(iFogSim)
Sim Weighted
[40]
(iFogSim)
# Cost
[56] Sim # # # #
Resource Success
[84] Analytic # # # Utilization
[73] Sim # # Rate
Sim Resource
[70]
(Cloudsim)
# # # # [144] Sim # # # Utilization
Sim Sim
[129]
(iFogSim)
# [31]
(SimPy)
# #
Sim Network Usage,
[156] Sim # # # # [72]
(iFogSim)
# Weighted Score
Weighted Cost,
Sim
[47] Sim # # # # [38]
(iFogSim)
# Interrupted Tasks,
Migrated Tasks
Computation
[54] Sim # # # ratio
[141] Sim # # # #
Analytic Weighted Sim (One
[147]
(Matlab)
# # # # Cost
[100]
Simulator)
# #
Analytic Migration
[137]
(Matlab)
# # # cost
[119] Prac # # # #
Weighted
[140] Analytic # # # [133] Sim # # # # Cost
Hybrid Weighted Cost,
[97]
(Sim+Prac)
# # [71] Analytic # # # # Offloaded Tasks
Dropped
[125] Sim # # # Tasks
[148] Analytic # # # #
[127] Sim # # Switching [109] Sim # # # #
[32] Prac # # Throughput [157] Sim # # # # #
[146] Sim # # # # [153] Sim # # # # QoS
Weighted Sim
[136] Analytic # # # # Cost
[111]
(Cloudsim)
# # #
Offloaded Tasks,
Analytic Satisfied Analytic
[28]
(Matlab)
# # # # Requests
[120]
(Matlab)
# # Failed Tasks,
Server Load
Hybrid Finished
[132]
(Sim+Prac)
# # # [143] Sim # # # Tasks
Analytic Analytic
[138]
(Matlab)
# # # # [83]
(Matlab)
# # # #
[42] Sim # # # # [117] Prac # # # Bandwidth
Resource
[45] Analytic # # # # [90] Sim # Usage
Analytic Failed
[149] Sim # # # # [14]
(Matlab)
# # Transmission
Weighted Analytic
[53] Sim # # # Cost
[58]
(Matlab)
# # # #
Hybrid
Sim
[64] (Sim, Prac) # [114]
(Sumo)
# # # #
(iFogSim)
Average
[128] Analytic # # # Offloaded [6] Sim # # # #
Tasks
Analytic Sim (Edge Success rate,
[33]
(Matlab)
# # # [16]
SimDAG)
# # Utilization

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:34 M. Goudarzi et al.

System Weighted
[19] Analytic # # # # Utility
[22] Sim # # Cost
Analytic
[11] Prac # # # # [4]
(Matlab)
# # # Throughput
Analytic Weighted
[12] Analytic # # # [50]
(Matlab)
# # Cost
Analytic
[21] Prac # # # # [59]
(Matlab)
# # #
Hybrid (Sim, Missed
[74]
Prac)
# # # # [61] ND # # # Deadline
Analytic
[30] Prac # # # # [65]
(Matlab)
# # # #
Hybrid Resource
[155]
(Sim, Prac)
# # # # [69] Analytic # # # Utilization
[135] Sim # # # # [82] Sim # # Availability
Hybrid
[110]
(Sim, Prac)
# # # # [95] Prac # Utilization
Hybrid Sim Network
[88]
(Sim, Prac)
# # # # [101]
(iFogSim)
# # Usage
Utility
[142] Analytic # # [98] Sim # # # # Function
Startup Time, Sim Network
[24] Prac # # # Ram Usage
[103]
(iFogSim)
# # # Usage
Performance System
[49] Analytic # # # # Gain
[104] ND # # Gain
Sim Deployed
[145]
(Cloudsim)
# # # [108] Prac # # # Instances
Weighted Weighted
[57] Analytic # # # # Reward
[115] Sim # # # Cost
Sim Analytic
[80]
(iFogSim)
# # QoS [116]
(Matlab)
# # # Throughput
Sim Application Weighted
[94]
(SimPy)
# # Loss
[134] Sim # # Cost
Deadline Weighted
[124] Analytic # # # Miss Ratio
[46] Sim # # Cost
Sim Deadline Sim (Fog
[78]
(iFogSim)
# # # Miss Ratio
[102]
WorkflowSim)
#
DO: Decision Overhead, T: Time, E: Energy, C: Monetary Cost, O: Other, Analytic: Analytical, Sim: Simulation, Prac: Practical, : Yes, #: No

2. Besides, in a real-world scenario, application modules have different characteristics (e.g.,

computation size, input size, ram usage). Thus, the best assumption for application modules
is applications with heterogeneous granularity specifications. As the number of contributing
parameters and the dynamicity of the application elements increases, capturing the application
parameters with temporal patterns for efficient scheduling decisions becomes more complex
[39, 129]. Although traditional-based placement techniques (e.g., heuristic, meta-heuristic) often
work well in general scenarios, they fail to adapt to continuous changes and dynamic contexts.
ML-based decision engines, such as RL, can more efficiently work in a dynamic context and provide
higher adaptability.
3. In large-scale Fog computing environments, numerous IoT applications with different workload
models and hybrid CCR may exist. Hence, the decision engine requires an admission control
component with an appropriate queuing mechanism (based on application requirements) to manage
diverse incoming requests and prioritize them for making the decision.
4. Regarding the environmental architecture, the most generalized scenario is when the envi-
ronment consists of several heterogeneous IoT devices, several heterogeneous FSs, and multiple
heterogeneous CSs. Also, the required mechanisms for intra-tier and inter-tier cooperation among
servers should be embedded to support diverse IoT application scenarios, such as mobility. Besides,
multiple distributed servers can collaboratively provide better performance for the execution of
IoT applications. Moreover, different fault domains can be prepared to improve the availability of

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:35

services. However, as the number of IoT applications and available servers in the environment
increase, the complexity of making decisions increases. Hence, the optimal scheduling decision
cannot be obtained in a timely manner. Consequently, other placement techniques such as heuristics
and ML-based techniques should be employed to obtain the scheduling decision in a reasonable
time.

5. The decision engine can be implemented as a set of distributed services/microservices. A

decision engine developed as a monolithic application may not be able to be deployed on a single
server, especially on resource-limited FSs. Hence, distributed deployment of decision engine com-
ponents on several distributed servers can provide several benefits: 1) more efficient deployment of
resource-limited devices, 2) provides better fault tolerance 3) offers better scalability 4) support
different deployment models (e.g., deployment of decision engine on FSs, CSs, or hybrid on both
FSs and CSs). Hybrid deployment of decision engine components on both FSs and CSs can lead to a
better user experience for end-users. To illustrate, applications requiring low latency and startup
time can be managed at the low-level FSs (i.e., at the Edge), and then be scheduled based on the
decision engine deployed at the Edge. However, application requests that are insensitive to latency
or startup time can be forwarded to CSs for scheduling.

6. Regardless of application and environmental characteristics, failure recovery mechanisms

and policies should be integrated into any decision engine. Independent failures and the non-
deterministic nature of any components (either hardware or software) in distributed systems
cause the most impactful issues in distributed systems. If the decision engine, which manages
the scheduling and execution of incoming IoT application requests, does not have an appropriate
failure recovery mechanism, the smooth execution of the whole system stalls.

9 FUTURE RESEARCH DIRECTIONS

This section presents future research directions, guiding researchers to further progress in the field
of Fog computing.

Microservices-based applications. The popularity of microservices for the deployment of IoT

applications is due to their loosely-coupled design, modularity, and the capability of microservices
to be shared among multiple IoT applications. But, it may incur data consistency and data privacy
challenges. To overcome these challenges, the placement techniques should consider the context of
applications and data before sharing microservices.

Practical Container orchestration in Fog computing. Orchestrating container-based IoT

applications is well studied in the cloud computing paradigm. However, in Fog computing, in
which CSs and FSs collaborate to run an application, several deployment models of orchestration
techniques are available. To illustrate, the master node can either be deployed on a FS or CS.
When the master node runs on a FS, the communication overhead and latency for end-users will
be reduced. However, the master node will use the most of resources on the FS for the cluster
management, especially for resource-limited FSs. Also, when the master runs on a CS, the startup
time and application latency will be negatively affected. Thus, based on the application structure
and its goal, different container orchestration models should be studied to find the best deployment
model according the application scenario. Several practical studies can be conducted to find which
deployment model is suitable for each IoT application scenario in terms of communication overhead,
the startup time of services, memory footprint, failure management, load balancing, and scheduling.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:36 M. Goudarzi et al.

Hybrid scheduling decision engines. Usually, decision engines only use one placement tech-
nique for different IoT applications. However, the requirements of IoT applications are heteroge-
neous, where one application is sensitive to startup time while the extremely high accuracy is not
important, or vice versa. Besides, decision engines should be adapted to work with either single
or batch placement approaches. Hence, context-aware decision engines with a suite of placement
techniques can be implemented to address the requirements of different IoT applications.

Systems for ML. Due to advancements in ML techniques and their rapid adoptions across
many IoT applications, it creates new demand for specialized hardware resources and software
frameworks (e.g., Nvidia GPU-powered Jetson, Google Coral Edge Tensor Processing Unit (Edge
TPU)) for Fog computing. New systems and software frameworks should be built to support the
massive computational requirement of these AI workloads. Besides, these systems can be a potential
target for the deployment of decision engines due to their high computational capacity.

ML for systems. While ML systems themselves are becoming mature and adopted into many
critical application domains, it is equally important to use these ML techniques to design and
operate large-scale systems. Adopting the ML techniques to solve different resource management
problems in Edge/Fog and Cloud is crucial to managing these complex infrastructures and workloads.
Moreover, majority of ML techniques are not optimized to run on resource-constrained devices.
To illustrate, consider an efficient ML model trained for resource management. Many resource-
constrained devices require full integer quantization to run the trained model. However, post
quantization of trained models is not always possible and in some cases they cannot be efficiently
converted. As a result, a study on requirements for the efficient execution of resource management
ML models on resource-limited FSs should be conducted.

Thermal management. The temperature of FSs (e.g., racks of Rpi or Nvidia Jetson platform),
especially those executing large workloads, increases significantly. So, the cooling systems should
be embedded to avoid system breakdown. Hence, a study on the temperature of these devices based
on their main processing and communication modules can be conducted to find the respective
temperature dynamics in different application scenarios and workloads. Moreover, lightweight
thermal management software systems for FSs can be designed to control the temperature dynamics
of devices. Also, the thermal index can be added as an important optimization/decision parameter
alongside other currently available parameters (e.g., time, energy, cost) for the placement techniques.

Trade-off between execution cost of IoT devices and resource providers. The goal of sched-
uling algorithms is to minimize the execution cost of applications either from IoT or resource
providers’ perspectives. However, some parameters such as energy consumption or carbon footprint
should be considered from both perspectives. Hence, not only is minimizing these parameters from
either perspective critical to reducing total energy consumption, but a trade-off parameter between
the execution cost of IoT devices and resource providers can be designed, aiming at total energy or
carbon footprint minimization.

Privacy aware and adaptive decision engines. Data-driven and distributed scheduling ap-
proaches are gaining popularity due to their high adaptability and scalability. However, sharing raw
data of users or systems incurs privacy issues. To illustrate, in DDRL-based scheduling techniques,
sharing experiences of multiple agents significantly reduce the exploration costs and improve
convergence time of DDRL agents while incurring privacy concerns when raw agents’ experiences
are shared. Accordingly, privacy-aware mechanisms for sharing such data (e.g., agents’ experiences)
can be integrated with these highly adaptive distributed scheduling techniques.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:37

10 SUMMARY
In this paper, we mainly focused on scheduling IoT applications in Fog computing environments.
We identified several main perspectives that play an important roles in scheduling IoT applications,
namely application structure, environmental architecture, optimization characteristics, decision
engines properties, and performance evaluation. Next, we separately identified and discussed the
main elements of each perspective and provided a taxonomy and research gaps in the recent
literature. Finally, we highlighted several future research directions for further improvement of
Fog computing.

REFERENCES
[1] Business Insider. 2020. THE INTERNET OF THINGS 2020. https://www.businessinsider.com/internet-of-things-report.
[Online; accessed 20-October-2021].
[2] IDC. 2020. IoT Growth Demands Rethink of Long-Term Storage Strategies. https://www.idc.com/getdoc.jsp?
containerId=prAP46737220. [Online; accessed 20-October-2021].
[3] Mohammad Aazam, Sherali Zeadally, and Khaled A Harras. 2018. Offloading in fog computing for IoT: Review,
enabling technologies, and research opportunities. Future Generation Computer Systems 87 (2018), 278–289.
[4] Mohamed Abd Elaziz, Laith Abualigah, and Ibrahim Attiya. 2021. Advanced optimization technique for scheduling
IoT tasks in cloud-fog computing environments. Future Generation Computer Systems (2021).
[5] Mohamed Abdel-Basset, Reda Mohamed, Mohamed Elhoseny, Ali Kashif Bashir, Alireza Jolfaei, and Neeraj Kumar.
2020. Energy-aware marine predators algorithm for task scheduling in IoT-based fog computing applications. IEEE
Transactions on Industrial Informatics 17, 7 (2020), 5068–5076.
[6] Raafat O Aburukba, Taha Landolsi, and Dalia Omer. 2021. A heuristic scheduling approach for fog-cloud computing
environment with stationary IoT devices. Journal of Network and Computer Applications 180 (2021), 102994.
[7] Mainak Adhikari, Satish Narayana Srirama, and Tarachand Amgoth. 2021. A comprehensive survey on nature-
inspired algorithms and their applications in edge computing: Challenges and future directions. Software: Practice
and Experience (2021).
[8] Cosimo Anglano, Massimo Canonico, Paolo Castagno, Marco Guazzone, and Matteo Sereno. 2020. Profit-aware
coalition formation in fog computing providers: A game-theoretic approach. Concurrency and Computation: Practice
and Experience 32, 21 (2020), e5220.
[9] Alia Asheralieva and Dusit Niyato. 2021. Learning-based mobile edge computing resource management to support
public blockchain networks. IEEE Transactions on Mobile Computing 20, 3 (2021), 1092–1109.
[10] Enzo Baccarelli, Paola G Vinueza Naranjo, Michele Scarpiniti, Mohammad Shojafar, and Jemal H Abawajy. 2017. Fog
of everything: Energy-efficient networked computing architectures, research challenges, and a case study. IEEE access
5 (2017), 9882–9910.
[11] Hossein Badri, Tayebeh Bahreini, Daniel Grosu, and Kai Yang. 2020. Energy-aware application placement in mobile
edge computing: A stochastic optimization approach. IEEE Transactions on Parallel and Distributed Systems 31, 4
(2020), 909–922.
[12] Tayebeh Bahreini, Hossein Badri, and Daniel Grosu. 2022. Mechanisms for resource allocation and pricing in mobile
edge computing systems. IEEE Transactions on Parallel and Distributed Systems 33, 3 (2022), 667–682.
[13] Tayebeh Bahreini, Marco Brocanelli, and Daniel Grosu. 2021. VECMAN: A Framework for Energy-Aware Resource
Management in Vehicular Edge Computing Systems. IEEE Transactions on Mobile Computing (2021). (in press).
[14] Hayat Bashir, Seonah Lee, and Kyong Hoon Kim. 2019. Resource allocation through logistic regression and multicriteria
decision making method in IoT fog computing. Transactions on Emerging Telecommunications Technologies (2019),
e3824.
[15] Flavio Bonomi, Rodolfo Milito, Jiang Zhu, and Sateesh Addepalli. 2012. Fog computing and its role in the internet of
things. In Proceedings of the first edition of the MCC workshop on Mobile cloud computing. ACM, 13–16.
[16] Lingfeng Cai, Xianglin Wei, Changyou Xing, Xia Zou, Guomin Zhang, and Xiulei Wang. 2021. Failure-resilient DAG
task scheduling in edge computing. Computer Networks 198 (2021), 108361.
[17] Rodrigo N Calheiros, Rajiv Ranjan, Anton Beloglazov, César AF De Rose, and Rajkumar Buyya. 2011. CloudSim:
a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning
algorithms. Software: Practice and experience 41, 1 (2011), 23–50.
[18] Lixing Chen, Cong Shen, Pan Zhou, and Jie Xu. 2021. Collaborative service placement for edge computing in dense
small cell networks. IEEE Transactions on Mobile Computing 20, 2 (2021), 377–390.
[19] Weiwei Chen, Dong Wang, and Keqin Li. 2019. Multi-user multi-task computation offloading in green mobile edge
cloud computing. IEEE Transactions on Services Computing 12, 5 (2019), 726–738.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:38 M. Goudarzi et al.

[20] Xianfu Chen, Honggang Zhang, Celimuge Wu, Shiwen Mao, Yusheng Ji, and Medhi Bennis. 2019. Optimized
computation offloading performance in virtual edge computing systems via deep reinforcement learning. IEEE
Internet of Things Journal 6, 3 (2019), 4005–4018.
[21] Yu Chen, Sheng Zhang, Yibo Jin, Zhuzhong Qian, Mingjun Xiao, Jidong Ge, and Sanglu Lu. 2022. LOCUS: User-
Perceived Delay-Aware Service Placement and User Allocation in MEC Environment. IEEE Transactions on Parallel
and Distributed Systems 33, 7 (2022), 1581–1592.
[22] Zhipeng Cheng, Minghui Min, Minghui Liwang, Lianfen Huang, and Zhibin Gao. 2021. Multi-Agent DDPG-Based
Joint Task Partitioning and Power Control in Fog Computing Networks. IEEE Internet of Things Journal 9, 1 (2021),
104–116.
[23] A.V. Dastjerdi, H. Gupta, R.N. Calheiros, S.K. Ghosh, and R. Buyya. 2016. Fog Computing: principles, architectures,
and applications. In Internet of Things: Principles and Paradigms, Rajkumar Buyya and Amir Vahid Dastjerdi (Eds.).
Morgan Kaufmann, 61 – 75. https://doi.org/10.1016/B978-0-12-805395-9.00004-6
[24] Qifan Deng, Mohammad Goudarzi, and Rajkumar Buyya. 2021. FogBus2: a lightweight and distributed container-based
framework for integration of IoT-enabled systems with edge and cloud computing. In Proceedings of the International
Workshop on Big Data in Emergent Distributed Environments. 1–8.
[25] Shuiguang Deng, Zhengzhe Xiang, Javid Taheri, Mohammad Ali Khoshkholghi, Jianwei Yin, Albert Y Zomaya, and
Schahram Dustdar. 2020. Optimal application deployment in resource constrained distributed edges. IEEE Transactions
on Mobile Computing 20, 5 (2020), 1907–1923.
[26] Wanchun Dou, Wenda Tang, Bowen Liu, Xiaolong Xu, and Qiang Ni. 2020. Blockchain-based Mobility-aware
Offloading mechanism for Fog computing services. Computer Communications 164 (2020), 261–273.
[27] Elie El Haber, Tri Minh Nguyen, Dariush Ebrahimi, and Chadi Assi. 2018. Computational cost and energy efficient
task offloading in hierarchical edge-clouds. In 2018 IEEE 29th Annual International Symposium on Personal, Indoor and
Mobile Radio Communications (PIMRC). IEEE, 1–6.
[28] Vajiheh Farhadi, Fidan Mehmeti, Ting He, Thomas F La Porta, Hana Khamfroush, Shiqiang Wang, Kevin S Chan, and
Konstantinos Poularakis. 2021. Service placement and request scheduling for data-intensive applications in edge
clouds. IEEE/ACM Transactions on Networking 29, 2 (2021), 779–792.
[29] Claudio Fiandrino, Nicholas Allio, Dzmitry Kliazovich, Paolo Giaccone, and Pascal Bouvry. 2019. Profiling performance
of application partitioning for wearable devices in mobile cloud and fog computing. Ieee access 7 (2019), 12156–12166.
[30] Kaihua Fu, Wei Zhang, Quan Chen, Deze Zeng, and Minyi Guo. 2022. Adaptive Resource Efficient Microservice
Deployment in Cloud-Edge Continuum. IEEE Transactions on Parallel and Distributed Systems 33, 8 (2022), 1825–1840.
[31] Pegah Gazori, Dadmehr Rahbari, and Mohsen Nickray. 2020. Saving time and cost on the scheduling of fog-based IoT
applications using deep reinforcement learning approach. Future Generation Computer Systems 110 (2020), 1098–1115.
[32] Hend Kamal Gedawy, Karim Habak, Khaled Harras, and Mounir Hamdi. 2021. Ramos: A resource-aware multi-
objective system for edge computing. IEEE Transactions on Mobile Computing 20, 8 (2021), 2654–2670.
[33] Sara Ghanavati, Jemal H Abawajy, and Davood Izadi. 2020. An energy aware task scheduling model using ant-mating
optimization in fog computing environment. IEEE Transactions on Services Computing (2020). (in press).
[34] Mostafa Ghobaei-Arani, Alireza Souri, and Ali A Rahmanian. 2020. Resource management approaches in fog
computing: a comprehensive review. Journal of Grid Computing 18, 1 (2020), 1–42.
[35] Mohammad Goudarzi, Qifan Deng, and Rajkumar Buyya. 2021. Resource Management in Edge and Fog Computing
using FogBus2 Framework. arXiv preprint arXiv:2108.00591 (2021).
[36] Mohammad Goudarzi, Zeinab Movahedi, and Masoud Nazari. 2016. Mobile cloud computing: a multisite computation
offloading. In Proceedings of the 8th International Symposium on Telecommunications (IST). IEEE, 660–665.
[37] Mohammad Goudarzi, Marimuthu Palaniswami, and Rajkumar Buyya. 2019. A fog-driven dynamic resource allocation
technique in ultra dense femtocell networks. Journal of Network and Computer Applications 145 (2019), 102407.
[38] Mohammad Goudarzi, Marimuthu Palaniswami, and Rajkumar Buyya. 2021. A distributed application placement and
migration management techniques for edge and fog computing environments. In Proceedings of the 16th Conference
on Computer Science and Intelligence Systems (FedCSIS). IEEE, 37–56.
[39] Mohammad Goudarzi, Marimuthu S Palaniswami, and Rajkumar Buyya. 2021. A Distributed Deep Reinforcement
Learning Technique for Application Placement in Edge and Fog Computing Environments. IEEE Transactions on
Mobile Computing (2021). (in press).
[40] Mohammad Goudarzi, Huaming Wu, Marimuthu Palaniswami, and Rajkumar Buyya. 2021. An application placement
technique for concurrent IoT applications in edge and fog computing environments. IEEE Transactions on Mobile
Computing 20, 4 (2021), 1298–1311.
[41] Jayavardhana Gubbi, Rajkumar Buyya, Slaven Marusic, and Marimuthu Palaniswami. 2013. Internet of Things (IoT):
A vision, architectural elements, and future directions. Future generation computer systems 29, 7 (2013), 1645–1660.
[42] Fengxian Guo, Heli Zhang, Hong Ji, Xi Li, and Victor CM Leung. 2019. An efficient computation offloading management
scheme in the densely deployed small cell networks with mobile edge computing. IEEE/ACM Transactions on

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:39

Networking 26, 6 (2019), 2651–2664.

[43] Harshit Gupta, Amir Vahid Dastjerdi, Soumya K Ghosh, and Rajkumar Buyya. 2017. iFogSim: A toolkit for modeling
and simulation of resource management techniques in the Internet of Things, Edge and Fog computing environments.
Software: Practice and Experience 47, 9 (2017), 1275–1296.
[44] Pooyan Habibi, Mohammad Farhoudi, Sepehr Kazemian, Siavash Khorsandi, and Alberto Leon-Garcia. 2020. Fog
computing: a comprehensive architectural survey. IEEE Access 8 (2020), 69105–69133.
[45] Zhenhua Han, Haisheng Tan, Xiang-Yang Li, Shaofeng H-C Jiang, Yupeng Li, and Francis CM Lau. 2019. Ondisc:
Online latency-sensitive job dispatching and scheduling in heterogeneous edge-clouds. IEEE/ACM Transactions on
Networking 27, 6 (2019), 2472–2485.
[46] Abhishek Hazra and Tarachand Amgoth. 2021. CeCO: Cost-efficient Computation Offloading of IoT Applications in
Green Industrial Fog Networks. IEEE Transactions on Industrial Informatics (2021). (in press).
[47] Tai Manh Ho and Kim-Khoa Nguyen. 2020. Joint server selection, cooperative offloading and handover in multi-access
edge computing wireless network: A deep reinforcement learning approach. IEEE Transactions on Mobile Computing
(2020). (in press).
[48] Cheol-Ho Hong and Blesson Varghese. 2019. Resource management in fog/edge computing: a survey on architectures,
infrastructure, and algorithms. ACM Computing Surveys (CSUR) 52, 5 (2019), 1–37.
[49] Zicong Hong, Wuhui Chen, Huawei Huang, Song Guo, and Zibin Zheng. 2019. Multi-hop cooperative computation
offloading for industrial IoT–edge–cloud computing environments. IEEE Transactions on Parallel and Distributed
Systems 30, 12 (2019), 2759–2774.
[50] Farooq Hoseiny, Sadoon Azizi, Mohammad Shojafar, Fardin Ahmadiazar, and Rahim Tafazolli. 2021. PGA: a priority-
aware genetic algorithm for task scheduling in heterogeneous fog-cloud computing. In Proceedings of the IEEE
INFOCOM 2021-IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS). IEEE, 1–6.
[51] Miao Hu, Lei Zhuang, Di Wu, Yipeng Zhou, Xu Chen, and Liang Xiao. 2019. Learning driven computation offloading
for asymmetrically informed edge computing. IEEE Transactions on Parallel and Distributed Systems 30, 8 (2019),
1802–1815.
[52] Pengfei Hu, Sahraoui Dhelim, Huansheng Ning, and Tie Qiu. 2017. Survey on fog computing: architecture, key
technologies, applications and open issues. Journal of Network and Computer Applications 98 (2017), 27–42.
[53] Zheyuan Hu, Jianwei Niu, Tao Ren, Bin Dai, Qingfeng Li, Mingliang Xu, and Sajal K Das. 2021. An Efficient Online
Computation Offloading Approach for Large-scale Mobile Edge Computing via Deep Reinforcement Learning. IEEE
Transactions on Services Computing (2021). (in press).
[54] Liang Huang, Suzhi Bi, and Ying-Jun Angela Zhang. 2020. Deep reinforcement learning for online computation
offloading in wireless powered mobile-edge computing networks. IEEE Transactions on Mobile Computing 19, 11
(2020), 2581–2593.
[55] Liang Huang, Xu Feng, Anqi Feng, Yupin Huang, and Li Ping Qian. 2018. Distributed deep learning-based offloading
for mobile edge computing networks. Mobile networks and applications (2018), 1–8.
[56] Liang Huang, Xu Feng, Cheng Zhang, Liping Qian, and Yuan Wu. 2019. Deep reinforcement learning-based joint task
offloading and bandwidth allocation for multi-user mobile edge computing. Digital Communications and Networks 5,
1 (2019), 10–17.
[57] Liang Huang, Xu Feng, Luxin Zhang, Liping Qian, and Yuan Wu. 2019. Multi-server multi-user multi-task computation
offloading for mobile edge computing networks. Sensors 19, 6 (2019), 1446.
[58] Mohamed K Hussein and Mohamed H Mousa. 2020. Efficient task offloading for IoT-based applications in fog
computing using ant colony optimization. IEEE Access 8 (2020), 37191–37201.
[59] Samia Ijaz, Ehsan Ullah Munir, Saima Gulzar Ahmad, M Mustafa Rafique, and Omer F Rana. 2021. Energy-makespan
optimization of workflow scheduling in fog–cloud computing. Computing (2021), 1–27.
[60] Mir Salim Ul Islam, Ashok Kumar, and Yu-Chen Hu. 2021. Context-aware scheduling in Fog computing: A survey,
taxonomy, challenges and future directions. Journal of Network and Computer Applications (2021), 103008.
[61] Chengen Ju, Yue Ma, Zhenyu Yin, and Feiqing Zhang. 2021. An Request Offloading and Scheduling Approach Base
on Particle Swarm Optimization Algorithm in IoT-Fog Networks. In Proceedings of the 13th International Conference
on Communication Software and Networks (ICCSN). IEEE, 185–188.
[62] Vasileios Karagiannis and Stefan Schulte. 2020. Comparison of alternative architectures in fog computing. In
Proceedings of the 4th IEEE International Conference on Fog and Edge Computing (ICFEC). IEEE, 19–28.
[63] F Khodadadi, A V Dastjerdi, and R Buyya. 2016. Internet of things: an overview. Internet of Things (2016).
[64] Dragi Kimovski, Narges Mehran, Christopher Emanuel Kerth, and Radu Prodan. 2021. Mobility-Aware IoT Applications
Placement in the Cloud Edge Continuum. IEEE Transactions on Services Computing (2021). (in press).
[65] Amit Kishor and Chinmay Chakarbarty. 2021. Task offloading in fog computing for using smart ant colony optimization.
Wireless Personal Communications (2021), 1–22.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:40 M. Goudarzi et al.

[66] Frank Alexander Kraemer, Anders Eivind Braten, Nattachart Tamkittikhun, and David Palma. 2017. Fog computing
in healthcare–a review and discussion. IEEE Access 5 (2017), 9206–9222.
[67] Gilsoo Lee, Walid Saad, and Mehdi Bennis. 2019. An online optimization framework for distributed fog network
formation with minimal latency. IEEE Transactions on Wireless Communications 18, 4 (2019), 2244–2258.
[68] Hai Lin, Sherali Zeadally, Zhihong Chen, Houda Labiod, and Lusheng Wang. 2020. A survey on computation offloading
modeling for edge computing. Journal of Network and Computer Applications (2020), 102781.
[69] Bowen Liu, Xiaolong Xu, Lianyong Qi, Qiang Ni, and Wanchun Dou. 2021. Task scheduling with precedence and
placement constraints for resource utilization improvement in multi-user MEC environment. Journal of Systems
Architecture 114 (2021), 101970.
[70] Jiagang Liu, Ju Ren, Yongmin Zhang, Xuhong Peng, Yaoxue Zhang, and Yuanyuan Yang. 2021. Efficient Dependent
Task Offloading for Multiple Applications in MEC-Cloud System. IEEE Transactions on Mobile Computing (2021). (in
press).
[71] Zhaolin Liu, Xiaoxiang Wang, Dongyu Wang, Yanwen Lan, and Junxu Hou. 2019. Mobility-aware task offloading and
migration schemes in SCNs with mobile edge computing. In Proceedings of the IEEE Wireless Communications and
Networking Conference (WCNC). IEEE, 1–6.
[72] Haifeng Lu, Chunhua Gu, Fei Luo, Weichao Ding, and Xinping Liu. 2020. Optimization of lightweight task offloading
strategy for mobile edge computing based on deep reinforcement learning. Future Generation Computer Systems 102
(2020), 847–861.
[73] Haodong Lu, Xiaoming He, Miao Du, Xiukai Ruan, Yanfei Sun, and Kun Wang. 2020. Edge QoE: Computation
offloading with deep reinforcement learning for Internet of Things. IEEE Internet of Things Journal 7, 10 (2020),
9255–9265.
[74] Zhi Ma, Sheng Zhang, Zhiqi Chen, Tao Han, Zhuzhong Qian, Mingjun Xiao, Ning Chen, Jie Wu, and Sanglu Lu. 2022.
Towards Revenue-Driven Multi-User Online Task Offloading in Edge Computing. IEEE Transactions on Parallel and
Distributed Systems 33, 5 (2022), 1185–1198.
[75] Pavel Mach and Zdenek Becvar. 2017. Mobile edge computing: A survey on architecture and computation offloading.
IEEE Communications Surveys & Tutorials 19, 3 (2017), 1628–1656.
[76] Henrik Madsen, Bernard Burtschy, G Albeanu, and FL Popentiu-Vladicescu. 2013. Reliability in the utility computing
era: Towards reliable fog computing. In Proceedings of the 20th International Conference on Systems, Signals and Image
Processing (IWSSIP). IEEE, 43–46.
[77] Redowan Mahmud, Samodha Pallewatta, Mohammad Goudarzi, and Rajkumar Buyya. 2021. IFogSim2: An Extended
iFogSim Simulator for Mobility, Clustering, and Microservice Management in Edge and Fog Computing Environments.
arXiv preprint arXiv:2109.05636 (2021).
[78] Redowan Mahmud, Kotagiri Ramamohanarao, and Rajkumar Buyya. 2018. Latency-aware application module
management for fog computing environments. ACM Transactions on Internet Technology (TOIT) 19, 1 (2018), 1–21.
[79] Redowan Mahmud, Kotagiri Ramamohanarao, and Rajkumar Buyya. 2020. Application management in fog computing
environments: A taxonomy, review and future directions. ACM Computing Surveys (CSUR) 53, 4 (2020), 1–43.
[80] Redowan Mahmud, Satish Narayana Srirama, Kotagiri Ramamohanarao, and Rajkumar Buyya. 2019. Quality of
Experience (QoE)-aware placement of applications in Fog computing environments. J. Parallel and Distrib. Comput.
132 (2019), 190–203.
[81] Redowan Mahmud, Adel N Toosi, Kotagiri Ramamohanarao, and Rajkumar Buyya. 2020. Context-aware placement
of Industry 4.0 applications in fog computing environments. IEEE Transactions on Industrial Informatics 16, 11 (2020),
7004–7013.
[82] Adyson M Maia, Yacine Ghamri-Doudane, Dario Vieira, and Miguel Franklin de Castro. 2021. An improved multi-
objective genetic algorithm with heuristic initialization for service placement and load distribution in edge computing.
Computer Networks 194 (2021), 108146.
[83] Prasenjit Maiti, Hemant Kumar Apat, Bibhudatta Sahoo, and Ashok Kumar Turuk. 2019. An effective approach of
latency-aware fog smart gateways deployment for IoT services. Internet of Things 8 (2019), 100091.
[84] Erfan Farhangi Maleki, Lena Mashayekhy, and Seyed Morteza Nabavinejad. 2021. Mobility-Aware Computation
Offloading in Edge Computing using Machine Learning. IEEE Transactions on Mobile Computing (2021). (in press).
[85] Yuyi Mao, Changsheng You, Jun Zhang, Kaibin Huang, and Khaled B Letaief. 2017. A survey on mobile edge computing:
The communication perspective. IEEE Communications Surveys & Tutorials 19, 4 (2017), 2322–2358.
[86] Ismael Martinez, Abdelhakim Senhaji Hafid, and Abdallah Jarray. 2020. Design, Resource Management, and Evaluation
of Fog Computing Systems: A Survey. IEEE Internet of Things Journal 8, 4 (2020), 2494–2516.
[87] Francesca Meneghello, Matteo Calore, Daniel Zucchetto, Michele Polese, and Andrea Zanella. 2019. IoT: Internet of
threats? A survey of practical security vulnerabilities in real IoT devices. IEEE Internet of Things Journal 6, 5 (2019),
8182–8201.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:41

[88] Jiaying Meng, Haisheng Tan, Xiang-Yang Li, Zhenhua Han, and Bojie Li. 2020. Online deadline-aware task dispatching
and scheduling in edge computing. IEEE Transactions on Parallel and Distributed Systems 31, 6 (2020), 1270–1286.
[89] Minghui Min, Liang Xiao, Ye Chen, Peng Cheng, Di Wu, and Weihua Zhuang. 2019. Learning-based computation
offloading for IoT devices with energy harvesting. IEEE Transactions on Vehicular Technology 68, 2 (2019), 1930–1941.
[90] Carla Mouradian, Somayeh Kianpisheh, Mohammad Abu-Lebdeh, Fereshteh Ebrahimnezhad, Narjes Tahghigh Jahromi,
and Roch H Glitho. 2019. Application component placement in NFV-based hybrid cloud/fog systems with mobile fog
nodes. IEEE Journal on Selected Areas in Communications 37, 5 (2019), 1130–1143.
[91] Carla Mouradian, Diala Naboulsi, Sami Yangui, Roch H Glitho, Monique J Morrow, and Paul A Polakos. 2017. A
comprehensive survey on fog computing: State-of-the-art and research challenges. IEEE communications surveys &
tutorials 20, 1 (2017), 416–464.
[92] Mithun Mukherjee, Lei Shu, and Di Wang. 2018. Survey of fog computing: Fundamental, network applications, and
research challenges. IEEE Communications Surveys & Tutorials 20, 3 (2018), 1826–1857.
[93] Ranesh Kumar Naha, Saurabh Garg, Dimitrios Georgakopoulos, Prem Prakash Jayaraman, Longxiang Gao, Yong
Xiang, and Rajiv Ranjan. 2018. Fog computing: Survey of trends, architectures, requirements, and research directions.
IEEE access 6 (2018), 47980–48009.
[94] Yucen Nan, Wei Li, Wei Bao, Flavia C Delicato, Paulo F Pires, and Albert Y Zomaya. 2018. A dynamic tradeoff data
processing framework for delay-sensitive applications in cloud of things systems. J. Parallel and Distrib. Comput. 112
(2018), 53–66.
[95] BV Natesha and Ram Mohana Reddy Guddeti. 2021. Adopting elitism-based Genetic Algorithm for minimizing
multi-objective problems of IoT service placement in fog computing environment. Journal of Network and Computer
Applications 178 (2021), 102972.
[96] Shubha Brata Nath, Harshit Gupta, Sandip Chakraborty, and Soumya K Ghosh. 2018. A survey of fog computing and
communication: current researches and future directions. arXiv preprint arXiv:1804.04365 (2018).
[97] José Leal D Neto, Se-Young Yu, Daniel F Macedo, José Marcos S Nogueira, Rami Langar, and Stefano Secci. 2018.
ULOOF: A user level online offloading framework for mobile edge computing. IEEE Transactions on Mobile Computing
17, 11 (2018), 2660–2674.
[98] Thieu Nguyen, Thang Nguyen, Quoc-Hien Vu, Thi Thanh Binh Huynh, and Binh Minh Nguyen. 2021. Multi-objective
Sparrow Search Optimization for Task Scheduling in Fog-Cloud-Blockchain Systems. In Proceedings of the IEEE
International Conference on Services Computing (SCC). IEEE, 450–455.
[99] Opeyemi Osanaiye, Shuo Chen, Zheng Yan, Rongxing Lu, Kim-Kwang Raymond Choo, and Mqhele Dlodlo. 2017. From
cloud to fog computing: A review and a conceptual live VM migration framework. IEEE Access 5 (2017), 8284–8300.
[100] Tao Ouyang, Zhi Zhou, and Xu Chen. 2018. Follow me at the edge: Mobility-aware dynamic service placement for
mobile edge computing. IEEE Journal on Selected Areas in Communications 36, 10 (2018), 2333–2345.
[101] Samodha Pallewatta, Vassilis Kostakos, and Rajkumar Buyya. 2019. Microservices-based IoT application placement
within heterogeneous and resource constrained fog computing environments. In Proceedings of the 12th IEEE/ACM
International Conference on Utility and Cloud Computing. 71–81.
[102] Lei Pan, Xiao Liu, Zhaohong Jia, Jia Xu, and Xuejun Li. 2021. A Multi-objective Clustering Evolutionary Algorithm for
Multi-workflow Computation Offloading in Mobile Edge Computing. IEEE Transactions on Cloud Computing (2021).
[103] Maycon Peixoto, Thiago Genez, and Luiz Fernando Bittencourt. 2021. Hierarchical Scheduling Mechanisms in
Multi-Level Fog Computing. IEEE Transactions on Services Computing (2021). (in press).
[104] Guang Peng, Huaming Wu, Han Wu, and Katinka Wolter. 2021. Constrained Multi-objective Optimization for
IoT-enabled Computation Offloading in Collaborative Edge and Cloud Computing. IEEE Internet of Things Journal 8,
17 (2021), 13723–13736.
[105] Charith Perera, Yongrui Qin, Julio C Estrella, Stephan Reiff-Marganiec, and Athanasios V Vasilakos. 2017. Fog
computing for sustainable smart cities: A survey. ACM Computing Surveys (CSUR) 50, 3 (2017), 1–43.
[106] Euripides GM Petrakis, Stelios Sotiriadis, Theodoros Soultanopoulos, Pelagia Tsiachri Renta, Rajkumar Buyya, and
Nik Bessis. 2018. Internet of things as a service (itaas): Challenges and solutions for management of sensor data on
the cloud and the fog. Internet of Things 3 (2018), 156–174.
[107] Carlo Puliafito, Enzo Mingozzi, Francesco Longo, Antonio Puliafito, and Omer Rana. 2019. Fog computing for the
internet of things: A survey. ACM Transactions on Internet Technology (TOIT) 19, 2 (2019), 1–41.
[108] Thomas Pusztai, Fabiana Rossi, and Schahram Dustdar. 2021. Pogonip: Scheduling asynchronous applications on the
edge. In Proceedings of the IEEE 14th International Conference on Cloud Computing (CLOUD). IEEE, 660–670.
[109] Qi Qi, Jingyu Wang, Zhanyu Ma, Haifeng Sun, Yufei Cao, Lingxin Zhang, and Jianxin Liao. 2019. Knowledge-driven
service offloading decision for vehicular edge computing: A deep reinforcement learning approach. IEEE Transactions
on Vehicular Technology 68, 5 (2019), 4192–4203.
[110] Xiaoyu Qiu, Weikun Zhang, Wuhui Chen, and Zibin Zheng. 2021. Distributed and collective deep reinforcement
learning for computation offloading: A practical perspective. IEEE Transactions on Parallel and Distributed Systems 32,

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
1:42 M. Goudarzi et al.

5 (2021), 1085–1101.
[111] Dadmehr Rahbari and Mohsen Nickray. 2020. Task offloading in mobile fog computing by classification and regression
tree. Peer-to-Peer Networking and Applications 13, 1 (2020), 104–122.
[112] Rodrigo Roman, Javier Lopez, and Masahiro Mambo. 2018. Mobile edge computing, fog et al.: A survey and analysis
of security threats and challenges. Future Generation Computer Systems 78 (2018), 680–698.
[113] Farah Ait Salaht, Frédéric Desprez, and Adrien Lebre. 2020. An overview of service placement problem in fog and
edge computing. ACM Computing Surveys (CSUR) 53, 3 (2020), 1–35.
[114] Hani Sami, Azzam Mourad, and Wassim El-Hajj. 2020. Vehicular-OBUs-as-on-demand-fogs: Resource and context
aware deployment of containerized micro-services. IEEE/ACM Transactions on Networking 28, 2 (2020), 778–790.
[115] Hani Sami, Azzam Mourad, Hadi Otrok, and Jamal Bentahar. 2021. Demand-Driven Deep Reinforcement Learning for
Scalable Fog and Service Placement. IEEE Transactions on Services Computing (2021). (in press).
[116] Indranil Sarkar, Mainak Adhikari, Neeraj Kumar, and Sanjay Kumar. 2021. A Collaborative Computational Offloading
Strategy for Latency-sensitive Applications in Fog Networks. IEEE Internet of Things Journal (2021). (in press).
[117] Mennan Selimi, Llorenç Cerdà Alabern, Felix Freitag, Luís Veiga, Arjuna Sathiaseelan, and Jon Crowcroft. 2019. A
lightweight service placement approach for community network micro-clouds. Journal of Grid Computing 17, 1
(2019), 169–189.
[118] Ali Shakarami, Mostafa Ghobaei-Arani, and Ali Shahidinejad. 2020. A survey on the computation offloading approaches
in mobile edge computing: A machine learning-based perspective. Computer Networks (2020), 107496.
[119] Shashank Shekhar, Ajay Chhokra, Hongyang Sun, Aniruddha Gokhale, Abhishek Dubey, and Xenofon Koutsoukos.
2019. Urmila: A performance and mobility-aware fog/edge resource management middleware. In Proceedings of the
IEEE 22nd International Symposium on Real-Time Distributed Computing (ISORC). IEEE, 118–125.
[120] Jinfang Sheng, Jie Hu, Xiaoyu Teng, Bin Wang, and Xiaoxia Pan. 2019. Computation offloading strategy in mobile
edge computing. Information 10, 6 (2019), 191.
[121] Syed Noorulhassan Shirazi, Antonios Gouglidis, Arsham Farshad, and David Hutchison. 2017. The extended cloud:
Review and analysis of mobile edge computing and fog from a security and resilience perspective. IEEE Journal on
Selected Areas in Communications 35, 11 (2017), 2586–2595.
[122] Jagdeep Singh, Parminder Singh, and Sukhpal Singh Gill. 2021. Fog computing: A taxonomy, systematic review,
current trends and research challenges. J. Parallel and Distrib. Comput. (2021).
[123] Balázs Sonkoly, János Czentye, Márk Szalay, Balázs Németh, and László Toka. 2021. Survey on Placement Methods in
the Edge and Beyond. IEEE Communications Surveys & Tutorials (2021).
[124] Georgios L Stavrinides and Helen D Karatza. 2019. A hybrid approach to scheduling real-time IoT workflows in fog
and cloud environments. Multimedia Tools and Applications 78, 17 (2019), 24639–24655.
[125] Ming Tang and Vincent WS Wong. 2020. Deep reinforcement learning for task offloading in mobile edge computing
systems. IEEE Transactions on Mobile Computing (2020). (in press).
[126] Koen Tange, Michele De Donno, Xenofon Fafoutis, and Nicola Dragoni. 2020. A systematic survey of industrial
Internet of Things security: Requirements and fog computing opportunities. IEEE Communications Surveys & Tutorials
22, 4 (2020), 2489–2520.
[127] Ouyang Tao, Xu Chen, Zhi Zhou, Lirui Li, and Xin Tan. 2021. Adaptive User-managed Service Placement for Mobile
Edge Computing via Contextual Multi-armed Bandit Learning. IEEE Transactions on Mobile Computing (2021). (in
press).
[128] Shujuan Tian, Chang Chi, Saiqin Long, Sangyoon Oh, Zhetao Li, and Jun Long. 2021. User Preference-Based
Hierarchical Offloading for Collaborative Cloud-Edge Computing. IEEE Transactions on Services Computing (2021).
(in press).
[129] Shreshth Tuli, Shashikant Ilager, Kotagiri Ramamohanarao, and Rajkumar Buyya. 2022. Dynamic scheduling for
stochastic edge-cloud computing environments using a3c learning and residual recurrent neural networks. IEEE
Transactions on Mobile Computing 21, 3 (2022), 940–954.
[130] Shreshth Tuli, Redowan Mahmud, Shikhar Tuli, and Rajkumar Buyya. 2019. Fogbus: A blockchain-based lightweight
framework for edge and fog computing. Journal of Systems and Software 154 (2019), 22–36.
[131] Kanupriya Verma, Ashok Kumar, Mir Salim Ul Islam, Tulika Kanwar, and Megha Bhushan. 2021. Rank based
mobility-aware scheduling in Fog computing. Informatics in Medicine Unlocked (2021), 100619.
[132] Can Wang, Sheng Zhang, Zhuzhong Qian, Mingjun Xiao, Jie Wu, Baoliu Ye, and Sanglu Lu. 2020. Joint server
assignment and resource management for edge-based mar system. IEEE/ACM Transactions on Networking 28, 5 (2020),
2378–2391.
[133] Dongyu Wang, Zhaolin Liu, Xiaoxiang Wang, and Yanwen Lan. 2019. Mobility-aware task offloading and migration
schemes in fog computing networks. IEEE Access 7 (2019), 43356–43368.
[134] Jin Wang, Jia Hu, Geyong Min, Wenhan Zhan, Albert Zomaya, and Nektarios Georgalas. 2021. Dependent Task
Offloading for Edge Computing based on Deep Reinforcement Learning. IEEE Trans. Comput. (2021). (in press).

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.
Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions 1:43

[135] Jin Wang, Jia Hu, Geyong Min, Albert Y Zomaya, and Nektarios Georgalas. 2021. Fast adaptive task offloading in
edge computing based on meta reinforcement learning. IEEE Transactions on Parallel and Distributed Systems 32, 1
(2021), 242–253.
[136] Lin Wang, Lei Jiao, Ting He, Jun Li, and Henri Bal. 2021. Service Placement for Collaborative Edge Applications.
IEEE/ACM Transactions on Networking 29, 1 (2021), 34–47.
[137] Shangguang Wang, Yan Guo, Ning Zhang, Peng Yang, Ao Zhou, and Xuemin Sherman Shen. 2021. Delay-aware
microservice coordination in mobile edge computing: A reinforcement learning approach. IEEE Transactions on
Mobile Computing 20, 3 (2021), 939–951.
[138] Shiqiang Wang, Rahul Urgaonkar, Murtaza Zafer, Ting He, Kevin Chan, and Kin K Leung. 2019. Dynamic service
migration in mobile edge computing based on Markov decision process. IEEE/ACM Transactions on Networking 27, 3
(2019), 1272–1288.
[139] Xiaofei Wang, Yiwen Han, Victor CM Leung, Dusit Niyato, Xueqiang Yan, and Xu Chen. 2020. Convergence of
edge computing and deep learning: A comprehensive survey. IEEE Communications Surveys & Tutorials 22, 2 (2020),
869–904.
[140] Xiaojie Wang, Zhaolong Ning, Song Guo, and Lei Wang. 2020. Imitation learning enabled task scheduling for online
vehicular edge computing. IEEE Transactions on Mobile Computing (2020). (in press).
[141] Zi Wang, Zhiwei Zhao, Geyong Min, Xinyuan Huang, Qiang Ni, and Rong Wang. 2018. User mobility aware task
assignment for mobile edge computing. Future Generation Computer Systems 85 (2018), 1–8.
[142] Huaming Wu, William J Knottenbelt, and Katinka Wolter. 2019. An efficient application partitioning algorithm in
mobile environments. IEEE Transactions on Parallel and Distributed Systems 30, 7 (2019), 1464–1480.
[143] Jindou Xie, Yunjian Jia, Zhengchuan Chen, and Liang Liang. 2019. Mobility-aware task parallel offloading for vehicle
fog computing. In Proceedings of the International Conference on Artificial Intelligence for Communications and Networks.
Springer, 367–379.
[144] Xiong Xiong, Kan Zheng, Lei Lei, and Lu Hou. 2020. Resource allocation based on deep reinforcement learning in IoT
edge computing. IEEE Journal on Selected Areas in Communications 38, 6 (2020), 1133–1146.
[145] Xiaolong Xu, Qingxiang Liu, Yun Luo, Kai Peng, Xuyun Zhang, Shunmei Meng, and Lianyong Qi. 2019. A computation
offloading method over big data for IoT-enabled cloud-edge computing. Future Generation Computer Systems 95
(2019), 522–533.
[146] Zun Yan, Peng Cheng, Zhuo Chen, Branka Vucetic, and Yonghui Li. 2021. Two-Dimensional Task Offloading for
Mobile Networks: An Imitation Learning Framework. IEEE/ACM Transactions on Networking 29, 6 (2021), 2494–2507.
[147] Bo Yang, Xuelin Cao, Joshua Bassey, Xiangfang Li, and Lijun Qian. 2021. Computation offloading in multi-access
edge computing: A multi-task learning approach. IEEE transactions on mobile computing 20, 9 (2021), 2581–2593.
[148] Chao Yang, Yi Liu, Xin Chen, Weifeng Zhong, and Shengli Xie. 2019. Efficient mobility-aware task offloading for
vehicular edge computing networks. IEEE Access 7 (2019), 26652–26664.
[149] Lei Yang, Bo Liu, Jiannong Cao, Yuvraj Sahni, and Zhenyu Wang. 2021. Joint computation partitioning and resource
allocation for latency sensitive applications in mobile edge clouds. IEEE Transactions on Services Computing 14, 5
(2021), 1439–1452.
[150] Jingjing Yao and Nirwan Ansari. 2018. QoS-aware fog resource provisioning and mobile device power control in IoT
networks. IEEE Transactions on Network and Service Management 16, 1 (2018), 167–175.
[151] Ibrahim Yasser, Abeer Twakol, Abd El-Khalek, Ahmed Samrah, AA Salama, et al. 2020. COVID-X: novel health-fog
framework based on neutrosophic classifier for confrontation covid-19. Neutrosophic Sets and Systems 35, 1 (2020), 1.
[152] Ashkan Yousefpour, Caleb Fung, Tam Nguyen, Krishna Kadiyala, Fatemeh Jalali, Amirreza Niakanlahiji, Jian Kong,
and Jason P Jue. 2019. All one needs to know about fog computing and related edge computing paradigms: A complete
survey. Journal of Systems Architecture 98 (2019), 289–330.
[153] Cheng Zhang and Zixuan Zheng. 2019. Task migration for mobile edge computing using deep reinforcement learning.
Future Generation Computer Systems 96 (2019), 111–118.
[154] PeiYun Zhang, MengChu Zhou, and Giancarlo Fortino. 2018. Security and trust issues in Fog computing: A survey.
Future Generation Computer Systems 88 (2018), 16–27.
[155] Gongming Zhao, Hongli Xu, Yangming Zhao, Chunming Qiao, and Liusheng Huang. 2021. Offloading Tasks With
Dependency and Service Caching in Mobile Edge Computing. IEEE Transactions on Parallel and Distributed Systems
32, 11 (2021), 2777–2792.
[156] Ruiting Zhou, Xueying Zhang, Shixin Qin, John CS Lui, Zhi Zhou, Hao Huang, and Zongpeng Li. 2020. Online Task
Offloading for 5G Small Cell Networks. IEEE Transactions on Mobile Computing (2020). (in press).
[157] Chao Zhu, Giancarlo Pastor, Yu Xiao, Yong Li, and Antti Ylae-Jaeaeski. 2018. Fog following me: Latency and quality
balanced task allocation in vehicular fog computing. In Proceedings of the 15th Annual IEEE International Conference
on Sensing, Communication, and Networking (SECON). IEEE, 1–9.

ACM Comput. Surv., Vol. 1, No. 1, Article 1. Publication date: January 2022.

View publication stats