Scheduling in cloud manufacturing: state-of-the-art and research challenges

Abstract

For the past eight years, cloud manufacturing as a new manufacturing paradigm has attracted a large amount of research interest worldwide. The aim of cloud manufacturing is to deliver on-demand manufacturing services to consumers over the Internet. Scheduling is one of the critical means for achieving the aim of cloud manufacturing. Thus far, about 158 articles have been published on scheduling in cloud manufacturing. However, research on scheduling in cloud manufacturing faces numerous challenges. Thus, there is an urgent need to ascertain the current status and identify issues and challenges to be addressed in the future. Covering articles published on the subject over the past eight years, this article aims to provide a state-of-the-art literature survey on scheduling issues in cloud manufacturing. A detailed statistical analysis of the literature is provided based on the data gathered from the Elsevier’s Scopus abstract and citation database. Typical characteristics of scheduling issues in cloud manufacturing are systematically summarised. A comparative analysis of scheduling issues in cloud manufacturing and other scheduling issues such as cloud computing scheduling, workshop scheduling and supply chain scheduling is also carried out. Finally, future research issues and challenges are identified.

Keywords:

• cloud manufacturing
• task decomposition
• service selection
• service composition
• scheduling

View correction statement:

Corrigendum

1. Introduction

During the past years, driven by new industrial manufacturing trends and requirements such as globalisation, individualisation, digitalisation, cloud, collaboration, integration and enabled by newly emerging technologies such as cloud computing, Internet of Things (IoT), Cyber-Physical Systems (CPS), big data analytics and artificial intelligence (AI), a new manufacturing paradigm known as cloud manufacturing was proposed (Li et al. 2010; Zhang et al. 2011; Xu 2012). To date, cloud manufacturing has attracted a large amount of research interest worldwide (Zhang et al. 2013). More than 800 articles have been published in this particular area (Adamson et al. 2017; Liu and Xu 2017). The aim of cloud manufacturing is to deliver on-demand manufacturing services to consumers over the Internet. Scheduling is a critical means for achieving the aim of cloud manufacturing. Scheduling has been a research topic in various domains for decades such as process and threat scheduling in operating systems (Tanenbaum and Woodhull 1987), job shop scheduling (Chaudhry and Khan 2016; Sharma and Jain 2016) and flow shop scheduling (Yenisey and Yagmahan 2014) in production environments, and task scheduling in computing and manufacturing systems such as in computing grid (Rahman et al. 2013), cloud computing (Singh and Chana 2016) and manufacturing grid (Tao et al. 2009).

In manufacturing or production, scheduling can be defined as a process of arranging, controlling and optimising work or workloads (Pinedo 2016). In the context of cloud manufacturing, scheduling can be defined narrowly or broadly. In the narrow sense, scheduling refers only to the process of allocating resources/services to tasks (or dispatching tasks to resources/services), monitoring, controlling and optimising resource/service status and task execution so as to satisfy consumers’ individualised requirements. In the broad sense, scheduling encompasses not only the scheduling process in the narrow sense, but also many other activities such as task processing (especially task decomposition), service discovery, matching, selection and composition that are either involved in or indispensable for the scheduling process (Tao et al. 2015; Cheng et al. 2017). Thus far, scheduling in the broad sense is a research topic that has attracted most attention in cloud manufacturing (Liu et al. 2016), and approximately 158 papers on this topic have been published, amongst which dozens of papers focused on scheduling in the narrow sense. In this case, there is an urgent need to ascertain the current status and identify issues and challenges to be addressed in the future. Currently, only Zhou, Zhang, and Liu (2017) discussed some issues pertinent to scheduling in cloud manufacturing. In contrast, the current work presents a more comprehensive literature survey and gives a more in-depth discussion and analysis. In addition, different from the work of (Zhou, Zhang, and Liu 2017), typical characteristics of scheduling in cloud manufacturing are systematically analysed and summarised, and a comparative analysis of scheduling issues in cloud manufacturing and other scheduling issues such as cloud computing scheduling, workshop scheduling, and supply chain scheduling is carried out as well.

The rest of this paper is structured as follows. Section 2 presents the fundamentals of scheduling in cloud manufacturing, including its operation model, scheduling procedure and typical characteristics. Section 3 gives a state-of-the-art literature survey on topics relevant to scheduling in cloud manufacturing, including task decomposition, service discovery, matching and selection and service configuration, allocation and composition, as well as scheduling. Section 4 discusses some related scheduling issues, and identifies the similarities and differences between these scheduling issues and that in cloud manufacturing. Section 5 analyses and points out future research issues and challenges, and discussed associated promising approaches. Section 6 concludes this paper with an outlook for future research.

2. Fundamentals of scheduling in cloud manufacturing

This section presents the fundamentals of scheduling in cloud manufacturing, including its operation model, scheduling procedure and typical characteristics.

2.1 Operation model of cloud manufacturing

Scheduling in cloud manufacturing has much to do with its operation model. It is therefore first necessary to give a brief introduction to its operation model (Zhang et al. 2014) (Figure 1). A complete cloud manufacturing system overall consists of three categories of stakeholders, namely, operator(s), providers and consumers. It is their cooperation with each other that maintains sustainable operation of a cloud manufacturing system.

•	Operator(s). One or several operators are introduced to manage and operate a cloud manufacturing platform. With the introduction of the operator, consumers can obtain sustainable, stable, and high-quality manufacturing services from the cloud platform in an on-demand manner, and providers are also allowed to publish their resources and capabilities conveniently using tools (e.g. virtualisation and service encapsulation tools) provided by the cloud platform.
•	Providers. Under the unified management of the operator, providers, on the one hand, publish manufacturing resources (including physical manufacturing resources and manufacturing capabilities encompassed in the entire product lifecycle) (Luo et al. 2013) to the cloud manufacturing platform for the sharing purpose, and on the other hand, receive manufacturing tasks dispatched from the cloud platform. All manufacturing resources from different providers are transformed into services, which are then clustered into different manufacturing clouds (e.g. design cloud, manufacturing cloud, and logistics cloud) according to the pre-defined rules.
•	Consumers. Under the unified management of the operator, consumers, including enterprises consumers and individual consumers (Wang and Xu 2013), on the one hand, submit their requirement tasks (e.g. design tasks, manufacturing tasks, test tasks, simulation tasks) to the cloud manufacturing platform, and on the other hand, receive the execution results of their orders.
•	Knowledge. Cloud manufacturing is a knowledge-based manufacturing paradigm. Knowledge (e.g. models, rules, standards, protocols, and algorithms) plays key roles in many activities and processes in the entire lifecycle of services including service generation (e.g. perception, connection, virtualisation and encapsulation), service management (e.g. description, cloud construction, search, matching, aggregation, clustering, composition, allocation, and scheduling), and service applications (e.g. service access, invocation, fault-tolerance, task migration, and business process management).

Figure 1. Operation model of cloud manufacturing.

2.2 Procedure of scheduling in cloud manufacturing

There are overall five phases for the entire scheduling process in cloud manufacturing, including order/task submission, preliminary order/task processing, scheduling, result delivery and service assessment (Figure 2).

(1)	Order/task submission: The entire scheduling process begins with submission of consumers’ orders/tasks. In terms of functional requirements, tasks can be classified into design tasks, manufacturing tasks, test tasks, etc., or their combination.
(2)	Preliminary order/task processing: After orders/tasks are submitted to a cloud platform, they first need to be preliminarily processed, including mainly classification, description, analysis, decomposition, etc. After the preliminary processing, each task’s requirements, including functional requirements and non-functional requirements, are made clear. The former refer to the function that needs to be realised for completing the task (such as a part or a product with a specific function). The realisation of the function requires invocation and execution of required types of services. The latter refer usually to some criteria (e.g. time, cost, quality) and associated constraints.
(3)	Scheduling: After preliminary order processing, the scheduling management module carries out task scheduling under the support of the scheduling supporting module, service management module and monitoring management module. The core scheduling module is responsible for generating optimised schedules and managing task execution processes. The scheduling supporting module is responsible for managing scheduling metrics, rules, methods, and algorithms and providing supports to the scheduling management module (e.g. assisting the scheduling management module in determining scheduling methods). For the service management module, its role is to manage service-related activities that are necessary for scheduling, including service classification, search and matching, composition, assessment, etc. The monitoring managing module plays an important role in monitoring the status of resources and orders on factory floors (Li et al. 2015) and provides real-time status information (e.g. machine availability) of resources and orders that is necessary for achieving optimised scheduling (Wang 2013; Mourtzis, Vlachou, Milas, and Dimitrakopoulos 2016; Mourtzis, Vlachou, Milas, and Xanthopoulos 2016; Liu et al. 2017). The scheduling execution process is as follows. Firstly, an optimised schedule is generated, and then tasks are dispatched to different providers for execution. During the execution process, resource and task status is monitored in real time. Sometimes remote control over resources from the cloud platform is needed (Wang, Gao, and Ragai 2014; Adamson et al. 2015). During the scheduling process, enterprises undertaking different subtasks of a task interact and communicate with each other to ensure smooth execution of the tasks.
(4)	Result delivery: After completing a task, associated resources are released, and the ultimate execution results (e.g. parts, components or end-products) are delivered to consumers via logistics or the Internet.
(5)	Service assessment: After consumers receive the execution results, they are endowed with the chance to assess the services they have used (Feng and Huang 2017). The evaluation results reflect their overall degree of satisfaction towards the results, and can also provide important reference for subsequent consumers to select services (Cui, Ren, and Zhang 2016).

Figure 2. Procedure of scheduling in cloud manufacturing.

2.3 Typical characteristics of scheduling in cloud manufacturing

Cloud manufacturing is a manufacturing paradigm that evolves from but differs from preceding manufacturing models such as agile manufacturing, networked manufacturing and manufacturing grid in terms of its operation model, system scale and technologies integrated (Zhang et al. 2011). Scheduling in cloud manufacturing bears a number of typical characteristics, which can overall be divided into operation mode-, resource/service and requirement-, and technology- and knowledge-related characteristics.

2.3.1 Operation mode-related characteristics

2.3.1.1 Scheduling involving multiple categories of stakeholders (CH1, where CH is a shorthand of characteristics)

There are multiple categories of stakeholders in a cloud manufacturing system, including mainly providers, operator(s), and consumers, and each category of stakeholder involves numerous individuals. The introduction of the operator(s) for the unified management and operation of a cloud manufacturing system distinguishes cloud manufacturing from previous manufacturing models. Moreover, compared with previous manufacturing models, far more individuals are involved in a cloud manufacturing system than those involved in previous manufacturing models.

Individuals in each category of the stakeholders are autonomous decision-making and interest-independent entities (Tai, Xu, and Hu 2012). They each have their own objectives and preferences (Tao et al. 2012), and thus the scheduling objectives are more diverse than other smaller scale manufacturing systems.

Collaboration, including collaboration of multiple providers (e.g. collaboration between multiple geographically distributed job shops) (Li et al. 2012) and collaboration between scheduling of cloud tasks (i.e. tasks dispatched from the cloud platform) and local tasks (i.e. tasks from other channels) (Lu et al. 2017; Wang, Zhang, and Qi 2017) is also an important characteristic for scheduling in cloud manufacturing.

2.3.1.2 Many resources/services-to-many orders/tasks scheduling (CH2)

Due to the introduction of the operator, cloud manufacturing enables not only integration of distributed resources but also distribution of integrated resources. Because of the integration and centralised management of manufacturing resources from different providers, cloud manufacturing enables multiple orders/tasks to be processed at the same time, giving rise to ‘many resources/services-to-many orders/tasks scheduling’. This is an important characteristic that differentiates cloud manufacturing from preceding manufacturing models which usually can only achieve many resources-to-one order/task scheduling. Many-to-many scheduling can also be referred to as multi-task-oriented scheduling (i.e. multiple tasks from consumers are processed simultaneously) (Lin and Chong 2017). Multi-task-oriented scheduling requires scheduling multiple tasks as a whole instead of dealing with each of them separately. Coupling relationships amongst multiple tasks in terms of required resources/services and execution flows are important factors that need to be taken into account (Cheng et al. 2014; Liu et al. 2017).

2.3.2 Resource/service- and requirement-related characteristics

2.3.2.1 Scheduling with larger-scale resources/services (CH3)

In cloud manufacturing, massive manufacturing resources from numerous providers are aggregated in the cloud platform and encapsulated into cloud services. In comparison with those in previous manufacturing systems, a much larger scale of resources/services are involved in cloud manufacturing. Scheduling of these large-scale resources/services requires more efficient scheduling methods and algorithms. During the scheduling process, large volumes of resource/service usage (e.g. composition and scheduling) data will be generated (Zhu et al. 2014; Yuan et al. 2016). Collection and utilisation of these data can effectively boost the scheduling efficiency and improve scheduling performance.

2.3.2.2 Scheduling of multi-level and multi-granularity resources/services and tasks (CH4)

From the perspective of an entire cloud manufacturing system, resources in cloud manufacturing exist at different levels: inter-enterprise level, enterprise level, workshop level, cell level, machine level (Wang, Zhu, and Kang 2016; Zhang et al. 2016). The quantity and variety of resources at different levels are different, and can be encapsulated into services with different granularities.

On the other hand, tasks are also of different granularities. Tasks in cloud manufacturing tasks are hierarchical and can be decomposed iteratively, i.e. a high-level task can be decomposed into a series of small tasks with a small granularity, and each of the small task can further be decomposed into a number of smaller tasks with a smaller granularity, and so on (Liu, Ma, and Liu 2013). For example, there are four levels for a product-type task: product, component, part and process (Li 2013; Yu et al. 2015; Wang, Zhu, and Kang 2016). Due to the multi-granularity characteristic of both manufacturing resources and manufacturing tasks, scheduling in cloud manufacturing can take place at multiple levels and multiple granularities instead of always at the lowest level or smallest granularity.

2.3.2.3 Individualised requirement-driven scheduling (CH5)

Scheduling in cloud manufacturing is driven by consumers’ requirements in the sense that the major driver for scheduling in cloud manufacturing is to satisfy requirements of consumers. Compared with previous manufacturing models, performance criteria in cloud manufacturing are more diverse, which include not only time and cost but also quality (e.g. machining precision and pass rate), energy consumption, service assessment, and enterprise reputation and location, etc. Different consumers have different preferences towards them. With larger scale manufacturing services and their configurability, consumers’ requirements in cloud manufacturing could be more individualised (Huang et al. 2013). Hence, scheduling in cloud manufacturing is more individualised requirement-driven (Tzafilkou, Protogeros, and Koumpis 2017).

2.3.2.4 Scheduling with more complexities and dynamics (CH6)

Scheduling in cloud manufacturing is more complex and dynamic compared with preceding manufacturing models. The complexity of scheduling in cloud manufacturing comes from the involvement of wide-area logistics and the complexity of resources (Qu et al. 2014). Manufacturing enterprises registered in a cloud manufacturing system are distributed in wide areas, which makes logistics indispensable for scheduling in cloud manufacturing (Lartigau et al. 2015; Zhong et al. 2016). Resource complexity results from different nature of resources, which can be either renewable or non-renewable, quantifiable or quantifiable, structured or unstructured, etc. (Hu et al. 2012). Computing resources such as servers, storage, network, bandwidth, software are also regarded as manufacturing resources in the broad sense in cloud manufacturing (Figure 2). All these make scheduling in cloud manufacturing more complex.

On the other hand, cloud manufacturing environments are full of dynamics (Yadekar, Shehab, and Mehnen 2016). In addition to traditional dynamics concerning resources (e.g. machine and software failures, tool damage, shortage of materials and worker absenteeism) and orders (e.g. rush orders, order cancellation, due date change, priority change, task processing time change and more importantly, requirement change), many other dynamics exist as well. For example, enterprises in cloud manufacturing are allowed to publish or withdraw their manufacturing resources freely, which can lead to dynamic resource publication and withdrawal. Service attributes such as quality, quantity, cost and availability may also change dynamically. Another source of dynamics comes from dynamic task arrivals over time (Li et al. 2012; Tai et al. 2013). All these can lead to changes in resource availability and task priority.

2.3.3 Technology-enabled and knowledge-related characteristics

2.3.3.1 Cloud service-based scheduling (CH7)

In cloud manufacturing, all manufacturing resources are encapsulated into cloud services by means of technologies of virtualisation and servitisation, which can effectively shield the heterogeneity of manufacturing resources and overcome the inconvenience resulting from geographical distributions of manufacturing resources with the features of interoperability and platform-independence. Encapsulation of manufacturing resources into cloud services is a core characteristic that makes cloud manufacturing different from previous manufacturing models. Scheduling in cloud manufacturing is thus based on cloud services, which can be conveniently orchestrated and configured to adapt to requirement tasks with different granularities (Li et al. 2012; Ren et al. 2011).

2.3.3.2 Scheduling with real-time perception and connection of resources (CH8)

Cloud manufacturing emphasises intelligent perception and connection of manufacturing resources into the cloud platform using IoT technologies (e.g. RFID, wired and wireless sensor networks and embedded systems) (Yang et al. 2017). Resource status on factory floors can thus be monitored (Caggiano, Segreto, and Teti 2016) and associated data can be collected into the cloud platform in real time. This creates a transparent environment for scheduling in cloud manufacturing. Owing to the real-time perception and connection, large volumes of data about resource status can be collected. Due to larger scale resources are involved in a cloud manufacturing system, larger volumes of resource perception data will be gathered in comparison with other manufacturing systems (Su, Xu, and Li 2016). Scheduling in cloud manufacturing relies closely on these data (Mourtzis et al. 2015; Zhang et al. 2016, 2017).

2.3.3.3 Scheduling with collaboration between computing resources and manufacturing resources (CH9)

Cloud computing is one of the key underpinning technologies for cloud manufacturing. At the bottom of the infrastructure of a cloud manufacturing platform is cloud computing infrastructure (e.g. CPU, storage, servers, virtual machines and networks). These cloud computing infrastructure resources provide the computing and storage capabilities needed for scheduling of manufacturing resources in cloud manufacturing. Optimal scheduling of cloud computing infrastructure resources is critical for the success of scheduling of manufacturing resources in cloud manufacturing. As a result, scheduling of computing resources constitutes the foundation of scheduling of manufacturing resources, and therefore needs to be conducted collaboratively (Lin and Chong 2017).

2.3.3.4 Knowledge-reliant scheduling (CH10)

Cloud manufacturing is a knowledge-based manufacturing paradigm. Various types of knowledge such as models, rules, standards, protocols, algorithms play key roles in scheduling in cloud manufacturing. Due to the high complexity of a cloud manufacturing system, cloud manufacturing is more knowledge-reliant in comparison with other manufacturing systems or models.

It should be pointed out that some of the characteristics are unique to cloud manufacturing whereas some of the characteristics are shared with previous manufacturing models (cloud manufacturing is a manufacturing paradigm that evolves from previous manufacturing models and is therefore inevitable to share some common characteristics of manufacturing scheduling with them). Specifically, the former include CH2, CH3, CH6, CH7 and CH9, and the latter include CH1, CH4, CH5, CH8 and CH10. Although there are some common characteristics between scheduling issues in cloud manufacturing and in previous manufacturing models, the characteristics are, to a great extent, enhanced in the context of cloud manufacturing. For example, more individuals are involved in a cloud manufacturing system; cloud manufacturing provides better supports for satisfying consumers’ highly individualised requirements with its large-scale virtualised, dynamically scalable and reconfigurable resource/services.

3. State-of-the-art

This section presents a state-of-the-art literature survey on scheduling in cloud manufacturing. As mentioned in the ‘Introduction’ section, scheduling in cloud manufacturing can be understood narrowly or broadly. Understanding scheduling in cloud manufacturing broadly enables generation of more effective and feasible schedules. This is because scheduling in cloud manufacturing is a complicated process involving many interdependent activities, and if we focus only on scheduling and neglect other processes and activities, the aim of scheduling in cloud manufacturing actually cannot be achieved. For example, scheduling without taking into account task decomposition cannot guarantee generation of practical schedules as task decomposition needs to be performed in combination with the status of services (e.g. granularity and level) instead of in isolation; scheduling of composite tasks (a composite task consists of a series of subtasks) in cloud manufacturing needs to employ service composition techniques for orchestrating services or evaluating the performance of scheduling. In the following, an up-to-date literature survey on scheduling in cloud manufacturing is presented from the broad-sense perspective.

3.1 Task decomposition

Task decomposition is a critical procedure as well as a prerequisite for scheduling of tasks in cloud manufacturing. However, task decomposition is a challenging issue in cloud manufacturing because it is a process that needs to be carried out taking into account services’ status (such as service granularity and correlations) instead of a process that can be done separately. Due to the significant challenge of task decomposition in cloud manufacturing, only a couple of researchers paid their attention to this issue. Yi et al. (2015) proposed a clustering algorithm-based optimisation method for task decomposition, in which tasks are firstly decomposed into atomic tasks according to a certain predefined decomposition rule, and then the atomic tasks are reorganised by means of a clustering algorithm taking into account task correlations, task-resource matching degrees, as well as competition between resources. Liu, Wang, and Ling (2017) proposed an ordered task decomposition approach (i.e. a hierarchical task network-based task decomposition method) for cloud manufacturing with the consideration of task granularity, cohesion and correlations.

3.2 Matching, discovery and selection

Task-resource/service matching can be understood broadly or narrowly. In the narrow sense, matching is a process of determining whether resources/services are suitable for tasks in terms of functional and non-functional requirements, while in the broad sense, matching is a process of connecting services and tasks for performing the function of a cloud manufacturing system. Task- resource/service matching has been extensively studied, and many methods and mechanisms have been proposed, including semantics-based hierarchical matching (Li, Dong, and Song 2012), hypernetwork-based supply demand matching (Tao et al. 2017) and QoS-based two-sided matching (Zhao and Wang 2016). Typically, Li, Dong, and Song (2012) proposed an intelligent service searching and matching method based on a two-process strategy. Firstly, services are filtered according to their type and status information to obtain a preliminary set of candidate services. Secondly, requests and services are matched according to their functional and non-functional information to finally obtain the services with a maximum matching degree. It has been proved that relatively to previous keyword- and semantics-based matching methods, this method can achieve more efficient and accurate matching. Tao et al. (2017) proposed a manufacturing supply demand matching simulator based on hypernetworks, which consists of seven key functions and subsystems, including manufacturing service management, manufacturing task management, matching and scheduling algorithms/strategies selection and design, etc. A prominent feature of this paper is that a service network and a task network are established for facilitating service and task management, matching and scheduling.

Resource/service discovery is a process of finding resources/services that are matched with specified tasks. It is easy to see that discovery involves and also depends on the matching process. Various resource/service discovery methods have been proposed, such as ontology-based resource discovery (Kang et al. 2013), task-driven service proactive discovery (Zhang et al. 2016), and agent-based service discovery (Guo et al. 2015). In particular, Zhang et al. (2016) proposed a task-driven manufacturing cloud service proactive discovery mechanism where services can respond to tasks proactively. A semantics-based intelligent matching method consisting of a product library and function matching module was proposed. Guo et al. (2015) proposed an agent-based manufacturing service discovery method for part-level tasks in cloud manufacturing. An agent-based manufacturing service discovery framework is proposed, which consists of four layers: manufacturing service information input layer, manufacturing service information management layer, manufacturing task information input layer and manufacturing service decision system. Five kinds of agents are involved, including task agent, task interface agent, service agent, service interface agent and decision agent. A structural matching method is proposed to implement the static parameters matching of task agents and service agents. A multi-agent system bid mechanism is built to accomplish the dynamic parameters matching of the two agents.

Resource/service selection is a process of determining and selecting suitable resources/services for tasks in terms of functional and non-functional requirements. Resource/service selection as an important issue in cloud manufacturing has also drawn some attention, such as multi-agent-based machine tool selection (Yin et al. 2016), fuzzy QoS-aware resource/service selection (Zheng, Feng, and Tan 2016), workflow-based resource selection (Zhang and Li 2015) and IoT-enabled dynamic service selection (Yang et al. 2016). Typically, Yin et al. (2016) proposed an optimal selection model for machine tools based on the multi-agent technologies, which effectively took advantages of the intelligent interactions and autonomous responses of agents. Zhang and Li (2015) proposed a workflow-based resource selection method for cloud manufacturing. In this approach, a business process is divided into the resource-dependent sub-processes and non-resource-dependent sub-processes. All optimal resource sets of different sub-processes are then merged to solve resource optimal set of the business process.

3.3 Resource/service configuration

Resource/service configuration refers to the process of allocating manufacturing resources/services to manufacturing tasks for achieving optimal system performance. Broadly speaking, service configuration encompasses service allocation, composition, and scheduling. In this section, the focus is on resource/service allocation and composition problems. Resource/service allocation usually includes the service composition process, especially for allocation of resources/services to tasks consisting of a series of subtasks.

As the name implies, service allocation is a process of allocating resources/services to tasks for optimally completing them. Service allocation has attracted some attention of researchers. It should be noted that some articles with the term ‘scheduling’ in the title deal actually with resource/service allocation instead of scheduling because no dynamic changes of resources/services or tasks over time are considered. For example, Zhou et al. (2017) addressed the issue of allocating services to individualised requirement tasks in cloud manufacturing. Cheng et al. (2013) investigated four types of resource service scheduling modes (i.e. provider-cantered, consumer-cantered, operator-cantered and system-cantered modes) and found that the system-cantered cooperative scheduling method can maximise overall utilities of the whole system and all users at the same time.

3.3.1 Service composition

Service composition refers to the process of combining multiple services, atomic or composite, into value-added services to fulfil a task or a set of tasks. Service composition in cloud manufacturing is a typical multi-objective, multi-constraint NP-hard combinatorial optimisation problem. The issue of service composition has been extensively studied in cloud manufacturing. Most research work adopts the QoS-aware method (Tao et al. 2013).

Due to the NP-hard nature of the service composition problem in cloud manufacturing, a variety of meta-heuristic algorithms and methods have been proposed to find optimal or near-optimal solutions in a reasonable time (Table 1). Apart from algorithms and methods, a more important issue is to establish appropriate service composition models. QoS criteria, objectives and constraints are core elements for building mathematical models of service composition. Regarding objective functions, most papers simply adopted the additive weighting method (Jin, Yao, and Chen 2015), i.e. assigning simply different weights to different QoS criteria. One exception is that Xu and Sun (2016) introduced the fuzzy analytic hierarchy process to determine the weights of different indices.

Table 1. Typical work on service composition in cloud manufacturing.

References	Algorithm	Nature of tasks	Performance criteria
Tao et al. (2013)	Chaos optimization	Single composite task	Time, cost, energy, reliability
Lartigau et al. (2015)	Artificial Bee Colony	Single sequentially executed task	Price, time, availability, reliability, maintainability, eco-impact
Jin, Yao, and Chen (2015)	Genetic algorithm	Single sequentially executed task	Time, cost, reliability, availability
Liu, Liu, and Sun (2013)	Genetic algorithm	Multiple homogeneous tasks	Time, cost, availability
Liu et al. (2013)	Genetic algorithm	Multiple heterogeneous tasks	Time, cost, availability
Liu and Zhang (2017)	Genetic algorithm	Single composite task	Time, cost, reliability
Zhou and Yao (2017d)	Teaching–learning-based optimization	Single composite task	Time, price, reliability, reputation
Zhou and Yao (2017c)	Artificial bee colony algorithm	Single composite task	Time, cost, availability, reliability
Zhou and Yao (2017a)	Artificial bee colony algorithm	Single composite task	Time, price, availability, reputation
Huang, Li, and Tao (2014)	Chaos optimization	Single composite task	Time, energy consumption, cost, reliability
Li et al. (2017)	Genetic algorithm	Multiple composite tasks	–
Chen et al. (2016)	Multi-objective evolutionary algorithm	Single composite task	Execution time, latency time, cost, reliability, availability
Xu and Sun (2016)	Bat algorithm	Single composite task	Time, availability, reliability
Zhou and Yao (2017b)	Bee colony algorithm	Single composite task	Time, price, availability, reputation, energy consumption
Xu et al. (2016)	Bees algorithm based on Pareto	Single composite task	Execution cost, response time
Liu, Ma, and Liu (2013)	Simulated annealing algorithm	Single hierarchical composite task	Time, cost, reliability, reputation
Xiang et al. (2016)	Pareto group leader algorithm	Single composite task	Service cost, service execution time span, service provider trust, service maintenance
Xiang et al. (2014)	Group leader algorithm	Single composite task	Time, cost, reliability
Zhang et al. (2016)	–	Single composite task	Cost, delivery time, delay time, reliability, credit, energy

All of current research on service composition in cloud manufacturing evades task decomposition by assuming that tasks have already been decomposed into subtasks that can thus be mapped to existing services directly. However, due to the complexity, diversity, and multi-level characteristic of services in cloud manufacturing, services in cloud manufacturing are of different granularities. As a result, the matching between tasks and services does not necessarily always occur at the lowest level (i.e. atomic task or atomic service level).

Different from most papers that dealt with single-task-oriented service composition, some researchers focused on multi-task-oriented service composition. Liu et al. (2013) addressed multi-task-oriented manufacturing cloud service composition and optimisation. In order to overcome the difficulty of satisfying users’ high QoS requirements using the ‘Each Composition for Each Task’ pattern, they proposed a ‘Multi-Composition for Each Task’ pattern-based global approach for combining incompetent composite services, which is able to increase the success rate of QoS requirement fulfilment. Li et al. (2017) proposed a clustering network-based approach to service composition in cloud manufacturing, which is able to tackle multi-task-oriented service composition.

Cloud manufacturing creates an open and dynamic environment where many ingredients that can impact QoS of services constantly change, such as service availability and assessment, which have rarely been considered adequately. Lartigau et al. (2015) investigated QoS-based service composition in cloud manufacturing with the consideration of geo-perspective transportation. Dynamical variations of service availability are considered. Zhou and Yao (2017a) proposed a context-aware artificial bee colony (ABC) algorithm based on the principle of ABC and service features in cloud manufacturing, which takes into account dynamics of trust QoS.

Logistics has also been considered in some research work. As mentioned above, Lartigau et al. (2015) investigated the service composition problem in cloud manufacturing with the consideration of transportation. However, in (Lartigau et al. 2015) only transportation was considered without taking into account other essential ingredients of logistics such as inventory (Cachon and Fisher 2000). In the future, more elaborate models of logistics need to be established for service composition in cloud manufacturing.

Both services and tasks in cloud manufacturing are of different granularities. Service composition in cloud manufacturing is essentially a dynamic matching process between multi-granularity tasks and multi-granularity services. Most of current papers simply assumed that tasks have been decomposed into a number of subtasks, each of which can be performed by a service without mentioning granularities of tasks and services. Current research on multi-granularity resource virtualisation and composition can provide some insights into multi-granularity service composition in cloud manufacturing (Li 2013; Liu, Li, and Shen 2014).

Manufacturing processes in cloud manufacturing are hierarchical, which means that tasks in cloud manufacturing are of different hierarchies (i.e. a task can be decomposed iteratively into a series of simple tasks until atomic tasks are reached). Most of previous work ignores this characteristic of tasks in cloud manufacturing. Liu, Ma, and Liu (2013) addressed hierarchical manufacturing cloud service composition based on the hierarchical manufacturing implementation processes, and developed a simulated annealing algorithm for solving that issue.

Other particular aspects of tasks, services and their matching method have also been taken into account, including energy consumption (Xiang et al. 2014; Xiang, Xu, and Jiang 2016), correlations (Li, Jiang, and Ge 2014; Jin, Yao, and Chen 2015; Li et al. 2016; Xu et al. 2016; Zhou and Yao 2017d), execution reliability (Jing et al. 2014), etc. In addition, Liu and Zhang (Liu and Zhang 2017) proposed a synergistic elementary service group-based service composition method, which allows free combination of multi-function, equivalent elementary services into a synergistic elementary service group to perform each subtask collectively. Xiang et al. (2016) pointed out the enormous opportunities and challenges of manufacturing big data to service composition and optimal selection and designed a case library-based method.

Research on service composition can provide many inspirations for scheduling in cloud manufacturing in terms of models (including task and service modelling, scheduling objective modelling, etc.), algorithms and approaches. For example, energy consumption, service correlations are important factors that need also to be considered in cloud manufacturing scheduling. However, there are great differences between service composition and service scheduling in cloud manufacturing. Firstly, service composition emphasises composition, which is more about how to orchestrate a number of services with relatively small granularities into a greater-granularity service for fulfilling a compound task, while service scheduling is usually about when and where tasks should be arranged and executed. That is, time as an important dimension has to be considered for scheduling in cloud manufacturing, but it is not the case for service composition. As a result of this, dynamical aspects of services (e.g. availability) and tasks (e.g. task insertion, cancellation and changes of attributes) over time that will frequently encounter in scheduling have rarely been considered in service composition. Secondly, service composition can focus on a single task or multiple tasks, while scheduling, which aims to optimise the arrangement and execution of multiple tasks over time, is usually concerned with multiple tasks. These differences should be considered when applying the models, algorithms and approaches of service composition to service scheduling.

3.4 Service and task scheduling

Currently, dozens of papers focusing on the scheduling process in the narrow sense have been published (Table 2).

Table 2. Typical work on scheduling in cloud manufacturing.

References	Scheduling algorithms	Performance criteria	Scheduling methods	Nature of tasks	Logistics
Laili, Zhang, and Tao (2011)	Genetic algorithm	Time	Static	Multiple sequentially executed design task	–
Li, Song, and Huang (2016)	Max percentages algorithm	Makespan, resource utilisation, load balancing	Static	single scientific workflow task	–
Lin and Chong (2017)	Genetic algorithm	Time	–	Single composite computing tasks	–
Li et al. (2012)	–	Cost, capability	Dynamic	Multiple composite tasks	✓
Tai et al. (2013)	Genetic ant colony algorithm	Time, capability, quality, cost	Dynamic	Multiple composite manufacturing tasks	✓
Cheng et al. (2014)	Genetic algorithm	Profit, task execution success rate	Static	Multiple composite, homogeneous manufacturing tasks	–
Jian and Wang (2014)	Particle swarm optimization	Time, cost	Static	Single composite manufacturing tasks	–
Ma et al. (2014)	Contract net	Time	Adaptive	Multiple tasks with different numbers of operations	–
Wang, Zhang, and Si (2014)	Particle swarm optimization	Time	Static	Multiple atomic manufacturing tasks	✓
Lartigau, Xu, and Zhan (2015)	Artificial bee colony	Availability, reliability	–	Single composite task	–
Mourtzis et al. (2015)	Intelligent search algorithm	Cost, time, quality	Adaptive	Multiple tasks	–
Cao et al. (2016)	Ant colony optimization	Time, cost, quality, service	Static	Single composite manufacturing task	✓
Li et al. (2017)	Genetic algorithm	Time, cost, load balancing	Static	Multiple composite, heterogeneous manufacturing tasks	✓
Jiang et al. (2016)	Genetic algorithm	Makespan, cost	Static	Multiple products to be disassembled	–
Li et al. (2016)	Specifically designed algorithms	Makespan	Static	Multiple homogeneous jobs	–
Liu et al. (2016), (2017)	Workload-based heuristic	Time, cost, quality	Static	Multiple heterogeneous manufacturing tasks	✓
Zhang et al. (2017)	Game theory	Time	Static	Multiple composite, heterogeneous manufacturing tasks	–
Lu et al. (2017)	Biogeography-based optimization	Time, load balancing, machine utilisation	Static	Multiple composite, heterogeneous manufacturing tasks	–
Xiao et al. (2015)	Shuffled frog leaping algorithm	Cost, time, capability utilisation, tardiness	Static	Manufacturing tasks	✓
Yuan et al. (2017)	Particle swarm optimization	Load balancing, reconstruction cost, delayed workload minimisation	Static	Multiple different kinds of products	–
Wang, Zhang, and Qi (2017)	Particle swarm optimization	Makespan	Static	Multiple jobs	–
Zhang et al. (2017)	Genetic algorithm	Cost, priority, reliability, energy consumption, customer satisfaction	Dynamic	Multiple independent orders	–
Zhou and Zhang (2016)	–	Time	Dynamic	Multiple composite tasks	✓
Zhou and Zhang (2016)	Event-triggered method	Time	Dynamic	Dynamically arriving composite tasks	✓

A couple of researchers dealt with scheduling of computing resources and design tasks in cloud manufacturing. Lin and Chong (2017) addressed resource constraint project scheduling for solving computing resource allocation problems in a cloud manufacturing system by proposing a genetic algorithm (GA) incorporating a number of new ideas (enhancements and local search). Laili, Zhang, and Tao (2011) dealt with collaborative design task scheduling in cloud manufacturing by designing a new energy adaptive immune GA. Li, Song, and Huang (2016) presented an architecture of scientific workflow management system based on a cloud manufacturing service platform and proposed a novel workflow scheduling algorithm called max percentages.

Most of current research focuses on scheduling of manufacturing tasks and manufacturing resources. Li et al. (2012) presented a framework for cloud manufacturing resource scheduling and proposed a queue balancing cutover strategy for solving the request dispatching issue based on stochastic advanced Petri nets. Lartigau et al. (2012) proposed a scheduling methodology for production services in cloud manufacturing in which orders are decomposed into several batches. In another work, they (Lartigau, Xu, and Zhan 2015) proposed a scheduling framework for cloud manufacturing with the specific consideration of resource service availability. Cao et al. (2016) addressed the service selection and scheduling issue in cloud manufacturing. Although being entitled ‘scheduling’, the problem under consideration is actually a single-task-oriented service composition issue taking explicitly into account service occupancy over time. Akbaripour et al. (2017) proposed mixed-integer programming models for solving the service selection optimisation and scheduling problem in cloud manufacturing where all composition structures including sequential, parallel, loop and selective are all incorporated. Moreover, in the models optimised routing decisions are made within a given hybrid hub-and-spoke transportation network.

A number of authors focused on the multi-task scheduling scenario. Cheng et al. (2014) studied multi-task-oriented scheduling in cloud manufacturing taking virtual resource correlations into account. The tasks they considered have completely identical subtask execution flows and thus require the same candidate resource sets. Li et al. (2017) investigated subtask scheduling of distributed robots in the context of cloud manufacturing. The scheduling addressed in this work is multi-task-oriented and the multiple tasks they considered are heterogeneous in terms of subtask type, execution flow and required robot resources. Task scheduling in this work occurs at the subtask level and subtasks are processed in a descending order of their average layer position amongst all processes. Inventory and transportation are also explicitly considered. Liu et al. (2016) proposed a multi-task-oriented service composition and scheduling model for cloud manufacturing, which combines service composition and scheduling in a reasonable manner. Moreover, different from most of previous work in which task execution times are known in advance, the execution time of a task is calculated in real time according to a task’s workload and an enterprise’s capability. Based on this model, they (Liu et al. 2017) further investigated the workload-based multi-task scheduling in cloud manufacturing, and discovered that scheduling larger workload tasks with a higher priority can usually lead to better system performance no matter whether time constraint exists. Wang, Zhang, and Si (2014) dealt with resource assignment and scheduling issue in cloud manufacturing environments. In their model, operations were regarded as the minimum task unit that can be executed by a service. Jiang et al. (2016) addressed the issue of cloud-based product disassembly task scheduling. A mathematical model that considered the uncertainty nature of the disassembly process and precedence relationships of disassembly tasks was built with the objectives to minimise the expected total makespan and the expected total cost. All of the above work adopts the static scheduling method.

Some researchers paid their attention to dynamic scheduling in cloud manufacturing. Tai et al. (2013) dealt with multi-objective dynamic scheduling in cloud manufacturing. In their model, rescheduling is triggered when the utility discrepancy between the current schedule and the new schedule (the generation of the new schedule is triggered by disturbances) exceeds a threshold in the time-domain period. Zhou and Zhang (2016) presented a dynamic task scheduling method based on real-time simulation. A system framework consisting of three layers, namely, task processing layer, core scheduling layer and resource service layer is proposed. Zhang et al. (2017) developed a real-time order dispatching mechanism to provide an optimal scheduling plan for cloud services encapsulated from virtual machining service of injection moulding machines. Zhou et al. (2018) addressed scheduling of dynamically arriving tasks in cloud manufacturing by proposing an event-triggered, subtask-oriented dynamic task scheduling method taking the average execution time of all tasks as the scheduling objective. In particular, Ma et al. (2014) proposed the concept of cloud agent, based on which adaptive management and scheduling of cloud manufacturing services can be achieved using the improved contract net mechanism.

There are also some researchers paying their attention to workshop scheduling problems in the context of cloud manufacturing. Jian and Wang (2014) dealt with batch task scheduling in cloud manufacturing workshops (a batch of tasks resemble a single composite task with a specific execution flow). Lu et al. (2017) dealt with mixed-flow, hybrid job shop scheduling problem in cloud manufacturing, which took into account integrated optimisation of mixed-flow assembly and part processing as well as collaborative scheduling of cloud service tasks and self-made tasks. To achieve real-time, data-driven optimised decision-making, Zhang et al. (2017) proposed a dynamic optimisation model for flexible job shop scheduling based on game theory, in which each machine is an active entity that can request task processing and tasks were assigned to optimal machines based on their real-time status. Yuan et al. (2017) considered the problem of multi-objective optimisation scheduling of a reconfigurable assembly line with the aim to minimise the cost of assembly line reconstruction, achieve the production load equalisation and minimise the delayed workload. Li et al. (2016) considered two uniform parallel machine scheduling problems with fixed machine cost under the background of cloud manufacturing. The goal is to minimise the makespan with a given budget of total cost. All the jobs are homogeneous in that processing times of all jobs are identical. Both non-pre-emptive and pre-emptive problems are considered.

Aiming to make use of surplus capacities of enterprises in cloud manufacturing. Wang, Zhang, and Qi (2017) proposed a job shop scheduling method taking idle times into account. A scheduling framework of job shops incorporating idle times of processing units, the method of determining processing time series and update strategy of idle times are proposed with the objective to minimise the makespan. Similarly, some researchers addressed the issue of dynamic or adaptive scheduling in cloud manufacturing workshops. Li et al. (2012) investigated collaborative scheduling technologies between multiple geographically distributed job shops based on dynamical resource capability services. They dealt with job scheduling with specified resource occupancy, as well as the issue of dynamic coordination of scheduling of distributed jobs. In particular, the time tolerance degree technology and dynamical adjustment technology were proposed to solve the problem. The technique for analysing mutual influence of job chains has been shown to be very useful for dynamic scheduling in cloud manufacturing. Mourtzis et al. (2015) addressed cloud-based adaptive shop-floor scheduling considering machine tool availability based on gathering of data from a multi-sensory system and machine tool operators. Machine tools’ status and available time windows can be obtained through an information fusion procedure.

Supply chain scheduling in the context of cloud manufacturing has also attracted some research interest. Xiao et al. (2015) addressed distributed supply chain scheduling for customisation of multiple products, considering the manufacturing and delivery stages in the supply chain and incorporating metrics of time, cost, production idle rate, and order tardiness into the scheduling objectives. In another work (Xiao et al. 2016), they reviewed planning and scheduling technologies of supply chain management in smart cloud manufacturing, encompassing the short-term production planning and scheduling, medium- and long-term plan management problem, and multi-dimensional integration planning and scheduling problems. Intelligent technologies for solving complicated planning and scheduling problems were also reviewed, including evolutionary algorithms, swarm intelligence, bio-inspired algorithms, etc.

3.5 A statistical analysis

This section provides a statistical analysis of existing articles on the aforementioned research topics, namely, task decomposition, resource/service discovery, matching and selection, service allocation and composition, and service and task scheduling. The method for the literature search and statistical analysis is as follows. Instead of collecting literature on scheduling in cloud manufacturing directly, we first collect research articles on cloud manufacturing and then analyse them one by one to determine the research issues they truly dealt with. The reason is that the issues addressed by some papers are not as that described in their titles. For example, some papers with the word ‘scheduling’ in their titles actually addressed service allocation or service composition instead of scheduling (only papers that focus on scheduling with specific consideration of dynamic changes of status (e.g. availability) of resources/services over time are perceived as addressing scheduling issues). The keywords used for collecting the literature on cloud manufacturing include ‘cloud manufacturing’, ‘cloud-based manufacturing’ or ‘cloud-based design and manufacturing’. Any article with the keywords in the title, abstract and keywords will be hit in our literature retrieval. As cloud manufacturing has been given different names by different researchers, the keywords above enable us to collect literature on cloud manufacturing as comprehensively as possible. The publications cited in this article and discussed in this section are mainly sought from the world’s largest abstract and citation database of peer-reviewed literature – Elsevier’s Scopus abstract and citation database (https://www.elsevier.com/solutions/scopus). The statistics presented in this section are based on data obtained from the Scopus database on the date of 5 February 2018. Our literature search and analysis found approximately 158 articles (including 127 journal papers and 31 conference papers) in total.

Figure 3 shows the number of articles published on different topics across different years (2010–2018). As shown in the figure, (1) task decomposition as an important prerequisite for scheduling in cloud manufacturing has largely been evaded, which is due to its high complexity, and (2) other issues such as service composition and scheduling have overall attracted more and more attention of researchers during the past years.

Figure 3. Number of articles published on different topics across different years (2010–2018).

There are more than 45 journals publishing articles on topics pertinent to scheduling. The typical journals that publish two or more articles are shown in Figure 4. Seventy-four (74) papers out of 127 journal papers are published in these journals, accounting for approximately 58%. Computer Integrated Manufacturing Systems, International Journal of Advanced Manufacturing Technology and China Mechanical Engineering published more papers than other journals did.

Figure 4. Top ten journals publishing most papers on scheduling in cloud manufacturing (2010–2018).

Figure 5 shows the algorithms used in the articles for solving the corresponding problems. As shown in the figure, various meta-heuristics are used, amongst which GA, PSO algorithm and Ant Colony Optimisation (ACO) algorithm are most frequently used. It should be noted that agent-based and game-based approaches are also used in a couple of articles.

Figure 5. Number of articles using different algorithms (or methods) from 2010 to 2018 on (a) service configuration, allocation and composition, and (b) service/task scheduling.

4. Related scheduling issues

This section discusses a number of related scheduling issues that have overlaps with scheduling issues in cloud manufacturing, including scheduling in cloud computing, workshop scheduling, and supply chain scheduling, focusing on their similarities and especially differences. Research on these scheduling issues has been ongoing for decades, and associated research outcomes can provide important references and inspirations for solving scheduling issues in cloud manufacturing. However, in order to apply the research outcomes to scheduling in cloud manufacturing, it is first necessary to identify the differences between these scheduling issues and that in cloud manufacturing. In the following, we discuss the similarities and, in particular, differences. Due to space limitation, we focus on the most important aspects instead of giving a comprehensive discussion.

4.1 Scheduling in cloud computing

Currently, a vast number of papers on scheduling in cloud computing have been published and a variety of scheduling approaches have been investigated, including bargaining-based scheduling, cost-based scheduling, dynamic and adaptive scheduling, energy-aware scheduling, service level agreement- and QoS-based scheduling, nature inspired and bio-inspired-based scheduling, profit-based scheduling, priority-based scheduling, etc. (Singh and Chana 2016). Many meta-heuristic, heuristic and hybrid algorithms have been proposed, including ACO, GA, PSO, etc. (Tsai and Rodrigues 2014; Kalra and Singh 2015). In particular, workflow scheduling (Masdari et al. 2016) has also been investigated, and corresponding scheduling schemes including meta-heuristic-based scheduling, heuristic workflow scheduling, and their hybrid have been extensively studied (Xu et al. 2009). Scheduling in cloud computing and scheduling in cloud manufacturing have many things in common. For example, both cloud computing and cloud manufacturing aim to deliver on-demand services to consumers using cloud technologies, and many common issues have been considered by both of them such as QoS, energy consumption, deadline and budget constraints, profit, task priority, etc. Therefore, the research outcomes on scheduling in cloud computing can provide important insights into scheduling in cloud manufacturing in terms of modelling approaches of scheduling problems and scheduling methods and algorithms. However, it should be noted that there are a number of fundamental differences between the scheduling issues in these two areas:

(1)	Operation mode. The stakeholders involved in a cloud computing system and a cloud manufacturing system are different. The former include a cloud provider and cloud service consumers, while the latter involve resource providers, platform operator(s) and cloud service consumers. In comparison with cloud computing where no infrastructure resource (e.g. servers, storage and routers) providers are needed, cloud manufacturing relies on providers to offer their manufacturing resources. This is because manufacturing resources in cloud manufacturing are far more expensive, complex, and diverse than infrastructure computing resources in cloud computing, it is almost impossible for the operator to purchase all manufacturing resources necessary for implementing a cloud manufacturing platform (Zhang, Cheng, and Boutaba 2010; Buyya et al. 2009). As manufacturing resources in cloud manufacturing are owned by autonomous and interest-independent providers, the difficulty of scheduling of manufacturing resources in cloud manufacturing significantly increases because resources cannot be scheduled for granted as they are owned by the operator. Instead, because different individuals have different objectives and preferences, their interest should also be properly coordinated during scheduling processes. In contrast, such an issue does not exist in cloud computing as the cloud provider (also the operator) owns all cloud infrastructure resources.
(2)	Scheduling process. Scheduling in cloud computing belongs to computing scheduling, while scheduling in cloud manufacturing is essentially production scheduling (or manufacturing scheduling). Because production scheduling is usually a much longer process than computing scheduling, scheduling in cloud manufacturing usually takes a much longer time and is also far more complex than scheduling in cloud computing. For example, during the scheduling process of cloud manufacturing, various dynamic, real-time events about enterprises, resources and orders and logistics (e.g. enterprise exit, machine breakdown, rush orders and logistics jam and stop) may occur frequently. These disruptions usually render established schedules infeasible, and thus require scheduling adjustments or rescheduling.
(3)	Logistics. Scheduling in cloud manufacturing usually takes place in the cloud (virtual space) and within and across different enterprises (physical space) at the same time (e.g. scheduling of machine and equipment resources). This is because schedules generated in the cloud platform eventually need to be executed on shop floors of different factories, and there is a need for logistics between cooperative enterprises. As a result, logistics is an important part of scheduling in cloud manufacturing. In contrast, due to scheduling in cloud computing concerns computing resources where only transmission of information flow is needed, no logistics is involved.
(4)	Performance criteria and scheduling objectives. Due to the differences between computing resources and manufacturing resources, some criteria considered for scheduling in cloud computing and cloud manufacturing and scheduling objectives are different. For example, commonly considered criteria in both cloud computing and cloud manufacturing usually include time, cost, energy consumption and resource utilisation (Kalra and Singh 2015), etc. However, manufacturing resources (such as machine tools, robots) are highly heterogeneous and due to this, resource quality is an important criterion used for resource selection for scheduling in cloud manufacturing. However, due to computing resources such as CPU, storage, servers are relatively not as heterogeneous as manufacturing resources, quality for scheduling in cloud computing is not a metric as important as for scheduling in cloud manufacturing. Additionally, due to consumers’ highly individualised requirements, scheduling objectives in cloud manufacturing are far more complex. For example, load-balancing is an important objective for scheduling in cloud computing. However, the case in cloud manufacturing could be different. Depending on consumers’ requirements, sometimes the scheduling objective should be focused on balancing workloads across different resources, while sometimes the focus should be on increasing the utilisation of high-quality manufacturing resources to achieve aims such as reduced energy consumption and high task fulfilment quality instead of load-balancing.

4.2 Workshop scheduling

Workshop scheduling represents the most classic scheduling problem and has been studied for decades. A great variety of issues have been addressed, including single machine scheduling, parallel machine scheduling, flow shop and flexible flow shop scheduling, job shop scheduling, open shop scheduling and many different models, deterministic or stochastic, have been built (Allahverdi 2015). Many approaches have been proposed, from the early mathematical programming (Blazewicz, Dror, and Weglarz 1991) and branch-and-bound technique (Brucker, Jurisch, and Sievers 1994) for small-size problems to various approximation methods such as local search (Vaessens, Aarts, and Lenstra 1996) and AI (Wang 2005; Çaliş and Bulkan 2015), etc. Especially, because of the dynamic nature of manufacturing systems, dynamic scheduling, which can effectively tackle various disturbances, has received much attention since the 1950s (Johnson 1954). Various dynamic scheduling techniques have been proposed, including heuristics, meta-heuristics, knowledge-based systems, fuzzy logic, neural networks and multi-agent systems (Ouelhadj and Petrovic 2009). Various scheduling rules have also been proposed, including typically earliest due date rule, shortest process time rule, longest processing rule, etc. (Blackstone, Phillips, and Hogg 1982). Workshop scheduling is an important part of scheduling in cloud manufacturing as schedules produced in the cloud ultimately need to be executed on shop floors of different factories. Consequently, workshop scheduling can provide important insights for scheduling in cloud manufacturing at the shop level. However, workshops in the context of cloud manufacturing are different from the traditional workshop environment, which should be identified before applying workshop scheduling techniques to cloud manufacturing.

(1)	Scheduling environment. The workshop scheduling environment in the context of cloud manufacturing is more efficient, flexible and transparent than the traditional scheduling environment (Liu et al. 2017). High efficiency makes workshops in cloud manufacturing environments able to adapt to the highly efficient scheduling in the cloud. High flexibility makes workshops in cloud manufacturing able to effectively collaborate scheduling of local tasks with cloud tasks. In addition, high transparency (which means that resource status can be comprehensively monitored and associated data can be collected in real time) is also required to facilitate resource perception and connection to the cloud platform. However, traditional workshops do not explicitly emphasise high efficiency, flexibility and transparency.
(2)	Scheduling object. Scheduling objects in traditional workshops are usually machines. In contrast, resources in cloud manufacturing are far more diverse than machines (Wang and Xu 2014). Various manufacturing resources and capabilities are transformed into services in cloud manufacturing. Several services can be conveniently orchestrated into a service with a greater granularity. As a result, in cloud manufacturing resources are often scheduled in the form of resource packages instead of a single type of resource. This will make resource scheduling in cloud manufacturing far more complicated than traditional workshop scheduling in terms of resource description, scheduling model establishment, etc.
(3)	Performance criteria and scheduling objectives. In the traditional workshop, the scenario is usually that factories schedule their own machines, and thus the scheduling objective is usually how to minimise the makespan and flow times without concerning cost and quality (Allahverdi 2015). However, in cloud manufacturing, due to resources are owned by different enterprises, many other metrics such as cost, quality, service evaluation, and enterprise reputation need also to be considered. Hence, the objective for workshop scheduling in the context of cloud manufacturing is more diverse and complex.

4.3 Supply chain scheduling

Supply chain emphasises the flow of goods and services involving the movement and storage of raw materials, of work-in-process inventory, and of finished goods from point of origin to point of consumption (Stadtler 2015). Supply chain management as an important issue has been widely studied (Werner 2000). Research on supply chain management can provide important insights for scheduling in cloud manufacturing in terms of business process integration and management and logistics modelling (Cooper, Lambert, and Pagh 1997; Bowersox, Closs, and Cooper 2002). The problem of supply chain management and scheduling exists in cloud manufacturing as well. However, there are some differences between supply chain management and scheduling in cloud manufacturing and the traditional supply chain management and scheduling.

(1)

Operation mode. In cloud manufacturing, partners (e.g. suppliers, manufacturers, retailers and logistics enterprises) of a supply chain are selected by the operator according to the services they provide. The driver for the formation of a supply chain is consumers’ requirements. The operation of a supply chain is managed by the operator. During the management and scheduling process, partners interact and collaborate with each other through the cloud platform, which is different from the traditional supply chains where partners interact and collaborate with each other directly. Due to partners of a supply chain are selected according to their services, their status is equal. However, in a traditional supply chain, the status of partners is not necessarily equal (e.g. dominating enterprises exist in some supply chains). As a result, their operation modes are different.

(2)

Characteristics. As mentioned above, owing to a supply chain in cloud manufacturing is generated through orchestration of services (such as manufacturing-as-a-service and logistics-as-a-service), it is thus more dynamic, agile, and flexible (Akbaripour, Houshmand, and Valilai 2015). Compared with traditional supply chains where the cooperative relationship amongst enterprises is usually stable and long-lasting, the cooperative relationship of partners in supply chains of cloud manufacturing is more dynamic and temporal. The cloud manufacturing platform provides a conveniently accessible infrastructure that allows integration of information between all partners across a supply chain so as to facilitate the decision-making process (Jassbi et al. 2014).

5. Research challenges and methodologies

This section discusses issues and challenges to be addressed in the future for scheduling in cloud manufacturing, which can overall be categorised into operation model-related issues and challenges, resource-related issues and challenges, requirement-related issues and challenges, as well as other issues and challenges. Major methodologies that can potentially be used for addressing the challenges have also been presented.

5.1 Research challenges

5.1.1 Operation model-related issues and challenges

The tri-party operation model of cloud manufacturing, on the one hand, enables sustainable, stable and high-quality manufacturing services, and on the other hand, brings many issues and challenges for scheduling in cloud manufacturing. The most challenging issue is how to guarantee the continuous participation of each individual in every category of the stakeholders (i.e. provider, operator, and consumer). As each individual is an autonomous entity that has its own objective and preference, the interest of them should be balanced and their objectives should be satisfied in order to ensure their continuous participation in cloud manufacturing. However, the interest and objectives of different individuals are sometimes in conflict with each other. In this case, an important issue is how to balance their interest and satisfy their objectives. It they fail to reach their objectives or aspirations, they may withdraw from the system.

Some research has already been carried out on this issue (Cheng et al. 2013, 2010). However, they dealt with only the static scenario where requirements and resources are fixed, and the utility function of each party considered only limited metrics. In the future, more comprehensive and in-depth research is needed. First of all, preferences and aspirations of providers and consumers need to be modelled. Providers are autonomous entities which have different preferences and aspirations with respect to transaction of their resources. For example, different providers have different degrees of preference towards historical cooperative relationships. Revenue is an important factor of resource providers, and different providers have different aspirations towards profit margin of their resources. Price is an important aspect that impacts revenues, and thus the different ways of pricing models should be considered such as fixed pricing, bargaining-based pricing (Peng, Guo, and Shao 2017), and auction-based pricing (Lin, Lin, and Wei 2010). Consumers in cloud manufacturing have highly individualised requirements, in which they have different preferences towards different metrics (such as time, cost, and quality). These preferences and aspirations of providers and consumers should be incorporated into the modelling of their utility. For utility modelling of the operator, the main issue is to determine the charge model for providing services to providers and consumers. Secondly, cloud manufacturing environments are highly dynamic, including, for example, dynamic task arrivals and dynamic cooperative relationships amongst different providers. Research on equilibrium of interest of different individuals should be conducted taking these dynamic ingredients into account. Thirdly, due to the complexity of cloud manufacturing environments, the selection of criteria and the determination of scheduling objectives should be determined in combination with concrete situations.

5.1.2 Resource-related issues and challenges

As mentioned in Section 2.3, resources in cloud manufacturing are complex, diverse and dynamic. These characteristics pose great challenges to resource scheduling in cloud manufacturing.

(1)	The diversity and complexity of resources/services in cloud manufacturing increase the difficulty of modelling and description, which is a prerequisite for scheduling them. As a result, appropriate modelling and description approaches of resources/services need to be explored. Hu et al. (2012) proposed a method for classifying virtual resources according to whether they are quantifiable and addable, which provides a feasible way for service modelling and description. However, more sophisticated methods that can satisfy the matching and scheduling requirements need to be developed.
(2)	Matching between multi-granularity resources/services and multi-granularity tasks is an important research issue. In order to increase the efficiency, matching should occur at an appropriate granularity level – neither at the atomic service level nor at a highest level (Hu et al. 2012; Yi et al. 2015).
(3)	High dynamics of manufacturing resources pose great challenges for their scheduling in cloud manufacturing (refer to Section 3.2 for details). High dynamics of resources require dynamic and adaptive scheduling methods (Ouelhadj and Petrovic 2009) because these dynamics may lead to the unavailability of resources at any time. Hence, dynamic scheduling (also called real-time scheduling) techniques which uses real-time information about resources for generating schedules are needed instead of generating a schedule in advance using static scheduling techniques. Although using resource information at the time of dispatching tasks can increase the adaptability of the generated schedules, it is inevitable that resources become unavailable during the execution process. In this case, reactive scheduling or rescheduling techniques are needed (Lou et al. 2012; Zhang and Wong 2017).
(4)	Logistics, including logistics within (Lee and Chen 2001; Zhong et al. 2016) and across enterprises, is an important issue for scheduling in cloud manufacturing. Logistics time and cost account for a large proportion of the total time and cost (Liu et al. 2017). The challenge concerning logistics for scheduling in cloud manufacturing is logistics modelling. In most of current research, logistics is usually simply modelled as a transportation issue. In fact, other issues such as warehousing and inventory need also to be considered for comprehensively modelling logistics in cloud manufacturing. The involvement of logistics complicates service scheduling in cloud manufacturing. Logistics monitoring and synchronisation based on IoT technologies such as RFID are also necessary for smooth scheduling in cloud manufacturing (Qu et al. 2016).

5.1.3 Requirement-related issues and challenges

As resources, tasks in cloud manufacturing are also diverse, complex and dynamic. Tasks in cloud manufacturing can be at different phases of the product lifecycle (e.g. design, manufacturing and test) and can be submitted by consumers from various different industries (diverse). In addition, tasks in cloud manufacturing are highly individualised and of different granularities, and are therefore very complex. The dynamics of cloud manufacturing tasks include changes of tasks (e.g. changes of tasks in terms of functional and non-functional requirements, tasks cancellation) and continuous task arrivals (including rush orders/tasks). The main research issues and challenges with scheduling in cloud manufacturing are as follows.

(1)	Consumers’ requirements first need to be modelled appropriately, which should consider task type and granularity. Different types of tasks have different execution flows. In addition, different types of tasks have different requirements for logistics. For example, no logistics is required for design tasks but required for machining tasks. Manufacturing tasks are also of different granularities, which refer to the fact that they are hierarchical, i.e. manufacturing task can be decomposed iteratively until the smallest granularity is reached. The characteristics above pose some challenges for modelling, which, however, is a prerequisite for scheduling.
(2)	Another significant challenge for scheduling in cloud manufacturing is task decomposition (Yi et al. 2015). Only with appropriate task decomposition can scheduling proceed. Most of current papers on service composition and scheduling assume that tasks have already been decomposed into subtasks, and thus evade this problem. Traditionally, task decomposition is a thorny issue. In cloud manufacturing, the high diversity, complexity, and dynamics increase the difficulty of task decomposition. Moreover, the most difficult issue for task decomposition in cloud manufacturing is that it is not an isolated process but a process that needs to be considered in combination with the status of services. That is, the extent to which tasks are decomposed depends on service granularities and does not always necessarily need to be decomposed into atomic tasks (i.e. tasks with the smallest granularity).
(3)	Scheduling of individualised tasks is an important research issue in cloud manufacturing. Individualisation means that different requirements from different consumers are different in terms of objectives, constraints and execution flows. When it comes to scheduling objectives, the main issue is that different consumers have different preferences towards criteria such as time, cost, quality and service rating. Also, consumers may raise different constraints for their requirements. The execution flows of different requirement tasks may also be different. Consequently, it is important to uncover the effects of different individualised requirements on scheduling. To do this, it is first needed to model these requirements properly, including the preferences, constraints, and execution flows (e.g. the distribution of preferences of different consumers). Zhou et al. (2017 addressed the issue of individualised requirement-oriented scheduling in cloud manufacturing. However, only four different types of individualised requirements in terms of execution flow and subtask type were considered.
(4)	Scheduling of multiple heterogeneous compound tasks is also an important issue. Multiple tasks can be completely identical, partly identical or completely different in terms of required resources and their execution flows. Completely identical tasks’ required resources and execution flows are completely the same, and vice versa for completely different tasks. Partly identical tasks lie between these two extreme scenarios (i.e. they are partly coupled with each other). Completely identical tasks or completely different tasks are relatively easy to be scheduled. The most challenging problem is to schedule the partly coupled tasks. Most of the current research focuses on scheduling of completely identical tasks (Cheng et al. 2014; Wu, Zhang, and Li 2015). Li et al. (2017) addressed the issue of scheduling of multiple heterogeneous, compound tasks at the subtask level in which the executed order to different subtasks was determined by their average layer positions in the execution flow. Liu et al. (2016, 2017) dealt with the issue of scheduling of multiple heterogeneous compound tasks at the task level, focusing on the effects of the execution order of the tasks. However, these researches considered only some special factors such as subtask layer and task workload. In the future, more factors (such as task priority, due date) and more reasonable methods (such as dynamic real-time or batch scheduling) should be investigated (Liu et al. 2016).
(5)	High dynamics of tasks in cloud manufacturing pose another challenge for scheduling in cloud manufacturing. The dynamics of tasks entails dynamic scheduling techniques (Nie et al. 2013). Given that tasks arrive dynamically, task arriving patterns should be properly modelled. Given the fact that generated schedules may no long feasible due to changes of tasks, reactive scheduling or rescheduling techniques are needed (Lou et al. 2012).

5.1.4 Other issues and challenges

(1) Due to the real-time perception and connection of resources on shop floors in the context of cloud manufacturing, large amounts of data (big data) will be obtained. The historical data can be used for forecasting resource (e.g. machines) performance in the next period of time. These real-time data reflects the current resource operation status, and thus can enable us to know machine availability (Lee et al. 2013). These data thus plays a critical role in the maintaining smooth scheduling in cloud manufacturing. Despite these data provides great supports for achieving optimal scheduling in cloud manufacturing, they also bring challenges regarding gathering, storage, analysis and utilisation of these data (e.g. fault diagnosis) (Kumar et al. 2016). A core issue is how to extract critical information from the big data according to the requirement of scheduling in cloud manufacturing.

(2) Collaborative scheduling-related issues and challenges

Collaborative scheduling, including collaborative scheduling of cloud task and local tasks and collaborative scheduling of cloud computing resources and manufacturing resources, is another challenge for scheduling in cloud manufacturing. The essence of collaborative scheduling of cloud tasks and local tasks is scheduling under dynamic resources occupancy. Due to task uncertainties, there must be some disturbances or disruptions that cause frequent adjustments of the collaborative schedules. In order to cope with this scenario, effective dynamical scheduling approaches are needed, such as the technique for analysing the degree for tolerating dynamic disturbances, the technique for analysing chain of influences between different tasks, as well as the movement-based dynamical adjustment technique (Li et al. 2012). Collaborative scheduling of computing resources and manufacturing resources in the cloud manufacturing platform is another challenge.

(3) Large-scale service-related scheduling issues and challenges

Large-scale services in the cloud manufacturing pose a major challenge for scheduling in cloud manufacturing in terms of efficiency. In order to enhance the efficiency, highly efficient algorithms and methods are required. Currently, various heuristic and meta-heuristics algorithms have been employed to solve the efficiency issue. However, in most case studies of the current research work, the number of services are usually dozens or hundreds at most, which is not the case in a cloud manufacturing platform where there are tens of thousands or hundreds of thousands services. As a result, most of the proposed algorithms will fail in the presence of large-scale services. In this case, new algorithms and approaches need to be developed (Laili et al. 2016). For example, parallel algorithms are needed (Tao et al. 2013). New approaches such as service classification and filtering which can be used to increase efficiency of scheduling should also be considered. In fact, preliminary service screening and filtering according to service requirements are effective methods for enhancing the scheduling efficiency.

5.2 Methodologies and techniques

Due to the high dynamics of cloud manufacturing environments where a variety of unexpected disruptions are inevitable to occur, it is almost impossible to find an optimal or near-optimal predictive schedules in advance by assuming all elements (e.g. services and tasks) are known and unchanged over time. Therefore, dynamic scheduling methods are required for cloud manufacturing. A number of dynamic scheduling methods such as heuristics, meta-heuristics (e.g. tabu search, simulated annealing and GA), knowledge-based systems, fuzzy logic and neural networks have been proposed. However, most of these scheduling techniques are centralised and hierarchical, which cannot effectively adapt to the distributed nature of cloud manufacturing systems (Ouelhadj and Petrovic 2009). For example, centralised and hierarchical scheduling approaches can produce globally better schedules, which, however, have problems with rapidly responding to disturbances and generating schedules efficiently in cloud manufacturing.

Multi-agent technologies provide an effective method for addressing scheduling issues in cloud manufacturing. By modelling different types of stakeholders as well as services and tasks as autonomous agents with different objectives and preferences, schedules can be generated through communication, cooperation, coordination, and negotiation amongst different types of agents. Distributing computing tasks to different agents can effectively increase the efficiency of producing schedules in the presence of large-scale services in cloud manufacturing. Multi-agent systems are open and dynamic, which means that any agent can join or exist a system dynamically without disrupting the system. This feature makes multi-agent technologies well suited for scheduling in cloud manufacturing, in which enterprises and resources are allowed to join or exist a cloud manufacturing system freely. Furthermore, local autonomy of agents allows rapid responses to various local variations, thus increasing robustness and flexibility of scheduling in cloud manufacturing. Consequently, multi-agent technologies are a very promising approach for solving scheduling issues in cloud manufacturing (Shen 2002; Zhao et al. 2016). Game theory is an effective theory tool for building multi-agent systems, and thus can also be used for scheduling in cloud manufacturing (Wei et al. 2010; Wu et al. 2013; Zhang et al. 2017). In addition, in order to develop more optimised schedules for cloud manufacturing, it is also necessary to combine multi-agent technologies with other theory, methods and techniques such as meta-heuristics, big data analytics and complex network theory (Tao et al. 2011, 2017), etc.

6. Conclusions

Thus far, research on cloud manufacturing has been ongoing for about eight years. Scheduling as one of the critical means for achieving the aim of providing on-demand services for cloud manufacturing has attracted a large amount of research interest. In order to ascertain the current research status on scheduling in cloud manufacturing and identify associated issues and challenges, this paper summarised the typical characteristics of scheduling in cloud manufacturing, presented a state-of-the-art literature survey, and discussed and analysed issues and challenges with scheduling in cloud manufacturing comprehensively and systematically. Moreover, in order to further promote the understanding of scheduling in cloud manufacturing, related scheduling issues such as cloud computing scheduling, workflow scheduling and supply chain scheduling were analysed and compared with that in cloud manufacturing.

Due to the high complexity of scheduling issues in cloud manufacturing, it is very difficult and even impossible to give a comprehensive and in-depth presentation and discussion of all associated issues. The major limitations of this paper are as follows. Firstly, the state-of-the-art literature survey includes only the most relevant topics concerning scheduling without including topics such as resource/service evaluation, monitoring and control, which are actually also encompassed in the scheduling process. In addition, the statistical analysis includes only documents (including journal and conference papers) written in English and Chinese. Secondly, due to the space limitation, the discussion of related scheduling issues focuses on the most important aspects instead of giving a comprehensive comparison. Thirdly, the identified research issues and challenges and presented associated approaches for addressing the issues and challenges are probably not comprehensive. However, it is believed that this paper can effectively deepen people’s understanding of scheduling issues in cloud manufacturing and accelerate research in this particular area so as to ultimately promote the development and implementation of cloud manufacturing.

As mentioned in Section 5, multi-agent technologies are a very promising approach for scheduling in cloud manufacturing. More importantly, using multi-agent technologies to solve scheduling issues in cloud manufacturing is just at the beginning and much more research is to be conducted. In the future, we will use multi-agent technologies in combination with other techniques such as meta-heuristics and big data analytics to address scheduling issues and challenges in cloud manufacturing.

Disclosure statement

No potential conflict of interest was reported by the authors.