Dynamic resource provisioning for cyber-physical systems in cloud-fog-edge computing

As is shown in Table 1, Cyber Physical Systems(CPS) is widely known as the integration of our physical world with all kinds of communication and the computing capability along with the storage. CPS is figuratively explained as a feedback loop influenced by cyber communications, physical acturators and human utilization [11]. In the conclusion of the article, security issues of the whole system is put forward, and [12] raised a quantitative hierarchized prediction model to answer the previous question, which also points out that as a real-time and autonomous system, CPS is a robust and complex system with multi-scale computation and physical components integrated with each other under the fifth-generation wireless network.

In terms of system composition, CPS includes information systems, sensors and mechanical equipment. All these parts connect and interact with each other under standard Internet Protocols [13]. Furthermore, CPS is a hybrid including both dynamic continuous processes of changes and discrete state transition, featuring the deep integration of computing and physical process and heterogeneous composition [14]. It might also be noted that CPS is beyond software and control domains and integrates both software components and plant components tightly [15]. What’s more, in [16], CPS is a kind of system supplying the integration of physical processes, computation and networks. It can also be viewed as a system with physical subsystems and software bounded together and each of them interacts with each other in various time and space dimensions [17, 18]. So it is apparent to reach the conclusion that it is system composition that distinguishes CPS from diverse data processing paradigms.

A cloud data center is a seemingly infinite pool of resources. Cloud computing opens up new horizons for distributed computing methods and it also creates an established, dynamic computing mode, which supplies information technology(IT) services as public computing resources. It decomposes huge data computing and processing programs into enormous small programs through the cyber ‘cloud’, and then processes and analyzes them through multiple servers to get results which will be sent back to the users [19]. Concludingly, in the standard definition document published by the National Institute of Standards and Technology(NIST), cloud computing is a mode which offers an available, convenient and on-demand network access to the shared pool, owning rich and powerful configurable computing resources [20].

Moreover, all these resources can be used immediately and take up few general commands along with the interaction with service providers [21, 22]. Hence, the cloud computing can be viewed as a new mode which can create the next-generation data centers in a dynamic manner by assembling the services supplied by cyber virtual machines [23].

Proposed by Cisco, fog computing platform extends the core of cloud data centers to its edge with the benefits of placing fog data centers and edge servers closer to mobile devices so that both better computing capacity and less transmission delay come to a win-win end. Furthermore, by establishing itself as a neighboring controller being able to provide immediate attention, it offers Internet of Things(IoT) applications with time-sensitive and privacy-concern services. It should also be noted that fog computing is a highly visible service platform which can integrate resources such as the CPU, bandwidth and storage from different network devices into the fog nodes to supply services to terminal devices. By placing these light-weight facilities at the proximity of IoT devices, fog computing can overcome the downsides of limited computation or storage capacities and ultra-high transmission latency by exploiting the opportunities of locality [24].

Edge computing refers to an open platform on the side close to the object or data source. It integrates network, computing, storage and the core capabilities of applications, providing the nearest service nearby, and generating faster network service responses [25]. Edge Computing is a dispatched computing mode that can address some inherent issues in cloud computing like resource centralization by withdrawing computing and storage resources from centralized points, which in other words, can be viewed as pushing applications, data, and services closer to those request senders geographically, perfectly overcoming intolerable transmission delay [26].

What’s more, as an essential method of IoT networks, edge computing is a platform with the ability to ease the burden of centralized data processing along with personal privacy issue, compared to cloud computing which may be restricted by data transferring and system efficiency [27].

Dynamic resource provisioning is a multi-agent system managing the providers’ resources along with the customers’ requests simultaneously or sequentially, as determined by the service-level agreement. [28] describes the concept concentrating on cloud computing while [29] focuses on edge computing. In spite of the revolutionary role cloud-fog-edge computing has displayed, dynamic resource provisioning is fast becoming a key instrument in the process of large scale and heterogeneous distributed data flow. Unproductive tasks scheduling may induce excessive energy consumption and unsuccessful task response.

The fundamental idea of dynamic resource provisioning is to supply the resources based on users’ or the applications’ real-time needs and network bandwidth requests [42], which is to identify the amount of resources that should be used on the execution of workflow at different moments and then schedule activities based on the allocated resources for the sake of networking reusability and flexibility.

In [30] and [31], a system involving a signal monitor, a decision-controller and an execution gadget is put forward. In addition, a feedback loop controlling principle is adopted, accessible to each layer in the system and serving as a interweaving thread that connects three layers together as a whole. [32] illustrates the scene where end clients have access to information via any person computer(PC) linked to the Internet and meanwhile, programming level is not open to client users as it objects to cloud centers, achieving the sharing and protecting of data distinctively. [33] argues that the behaviors of physical processors is the joint result of cyber and physical level of the system in the large. [34] and [35] highlight the indiscerptible role of the computing core and communication core which dominate both cyber and physical system. [36] maintains that event is the clue to the Cyber-Physical System and concentrates on the timestamp of each event. In [37], a novel paradigm called Dynamic Social Structure of Things(DSSoT) is carried out, taking full advantage of users’ past experience and preference settings to provide accurate predictions for future better QoE. AcOP is a combination of mobile apps and cloud-based CPS, flexibly corresponding to surroundings and certain conditions. Edge computing in AcOP and the associated framework has high security performance because trusted entities can work with each other in a variety of CPS applications and act as backup options. Given one device fails, instantly other devices will replace it [38].

In light of the defense for cyber security, the privacy computing framework proposed in [39] allows private information exchanges, extensive permission of private information flow in a multi-interface scenario. The competition between different data centers is no less intense than that of track-and-field events [40]. Thus more reliable and stable guarantee of data assets make for both more clients for service providers and less risks for service clients [41].

[43] and [44] propose a framework where part of the tasks are migrated from cloud servers to edge nodes to arrange services and resources in proximity to users, in which way lower transmission delay and network bandwidth requirements become a reality. Additionally, the paper illustrates the framework with a typical application scenario: smart cities. Smart cities are convincing examples of an effective combination between digital level and physical level. All aforementioned architectures are concluded in Table 2.

As illustrated in Fig. 2, we introduced the holistic integration of cloud computing, fog computing and edge computing. Edge computing offers edge-server-oriented service provision in a bid to achieve an optimized method for processing with low latency the computation offloading and data offloading tasks generated from mobile users. Additionally, fog computing paradigm employs multiple cloudlets collaborations to meet the demanding requirements from multimedia applications without the shortcomings of transmission delay. The aforementioned two architecture lies in proximity to end devices while cloud computing servers are far away from mobile users. Therefore, the cloud servers are capable of offering constant computing power and coping with challenging applications and services.

In this subsection, the question of how to build, employ, manage and control cloud computing in Cyber-physical-social system is resolved in a comprehensive and all-round slope as is demonstrated in Table 3.

Indeed, reducing energy costs for data center operations accounts for less waste, but the QoS of running applications can also dramatically decrease due to overcommitment to resources. Thus with multi-core homogeneous processors, an algorithm for virtual machine placements is proposed in [46] and can fulfill energy cost reduction by over 60 per cent, estimated by experimental results. In [47], an automotive system and a quadrotor, a small helicopter with four propellers, are employed to elaborate on the complexity and heterogeneity of maintaining data consistency and guaranteeing verification. A QoS-aware VM scheduling method in cloud-based CPS is put forward to figure out optimal energy consumption scheme, assisted by the non-dominated sorting genetic algorithm III, simple additive weighting and multiple criteria decision making. Genetic algorithm is suitable in large scale problems including searching and optimization problems as well as under the circumstances where branch and bound schemes cannot make ideal results [48].

Modern factories overflow the boundaries of the traditional mills. From the standpoint of customers, not only does participating in the designing is favored but tracing and tracking the production along assembly line would be an extra bonus. CPS platform is considered in [49] as the bridge that connects CPS nodes and stakeholders in manufacturing. However, this study also points out that the solitude CPS cannot undertake future personalized and heterogeneous requirements, thus social tier should be added to the original pattern, forming integrated cyber-physical-social systems. Under the trend of market-of-one mode, [50] and [51] merge the utilization and architectures of manufacturing resources of diversified scale into an integrated resource fractals on the theoretical evidence of fractal theory. It is stated that human-friendly manufacturing requirements should be translated into a formalized and systematic machine language for high-level data sharing and more intelligent communications. The transformation is in conformity with customized production process. Shu, Zhaogang, et al present in [52] an orchestration of three entities, i.e., CIA layer, virtual cluster and physical host. At the bottom, physical hosts are responsible for manage VM asynchronous lifecycle, optimize network I/O performance and disk I/O performance, with middleware providing workflow-oriented resources arrangement and the top layer targeting at collect QoE information.

[53] proposes a novel paradigm called Cyber infrastructure as a Service (CaaS) and interpret the concept with a use case. The essence of cloud computing lies in cloud data centers that offer a mighty and scalable computing and storage environment away from end users. Transmitting resource-demanding and computing-intensive to the remote cloud considerably bring down the objective requirements for device performances [54]. For those settling down in high-rises and skyscrapers, evacuation in cases of fires or other emergent conditions can be intelligently scheduled and dispatched.

[55] focuses on bettering the long existing fluctuations of electricity infrastructure with CRBM and FCRBM employing deep learning algorithm. In recent years, deep learning has been introduced in a wide range of use cases. It is based on learning applications embarking on modeling the question, categorizing the task and estimating potential results to achieve an optimized computation accuracy. [56] regards the combination of technology strategy and policy strategy a stimulator to the development of survival of small-and medium-sized enterprises in the times of CPS and automated driving.

Via a use case of smart driving, [57] and [58] elaborate on the steps from raw sensor data sampling to MLAs in multi-way, half-duplex, or large-antenna networks. The whole process can be divided into two stages. On the one hand, smartphone sensors collect raw data material and then translate all of them to earth coordinate system. On the other hand, computers first generate attribute vectors and the dataset is transmitted for training, testing and accessing MLAs before the final results come out. [59] proposes a transportation featuring distribution, flexibility and scalability. In a gesture to achieve traffic efficiency including commanding emergency situations and predicting possible congestions and accidents, the paper utilizes C4.5 algorithm and artificial intelligence techniques.

A lightweight design is raised in [60], in which bridge bond archive storage and three levels together. At the bottom of the architecture lies physical level with sensors, actuators and zigbee wireless mesh, followed by middleware’s publishing and subscribing managers. Through the application bridge, middleware can readily transmit data with the top level which is responsible for activity recognition, activity discovery and energy stuff. [61] presents an architecture and its implementation for data sharing across organizations and healthcare workers to confront with curious cloud providers.

Empowered by data, digital cities evolve into smart cities where requirements from citizens, as well as residents, workers, and visitors are supposed to be better met [62]. We envision a livable, comfortable and sustainable city with physical facets: traffic signals, sewage works, air quality measures being controlled algorithmically [63].

As an overview of this subsection, we propose Fig. 3 to shed light on the relationship among above-mentioned articles. Mainstream academic schools devide cloud computing architecture into five categories including Infrastucture as a Service(IaaS), Platform as a Servcie(PaaS), Software as a Servcie(SaaS), Communications as a Service(CaaS) and Machine as a Service(MaaS). Further, the use cases of each one enjoy promising future, for instance, manufacturing, smart grid, automated driving, transport, smart home, healthcare, smart city. Also, cloud computing is supported by powerful algorithms like evolutionary algorithms and multi-objective optimization.

In this subsection, we review the existing researches on fog computing. A summary of literatures about CPS with the basis of fog computing are listed in Table 4, where key points in each article are concluded and the architecture is further analyzed is Fig. 4. We review on CPS which is simply supported by fog computing in “FC-CPS” Section and methods which integrate CPS, fog computing along with other kinds of technology in “FC-CPS with other technology” Section.

In this subsection, we dive into the application of edge computing on the development of CPS. We first give a brief review on what has been concluded before, as can be readily available in Table 5 and then we probe into the current situation of edge computing for CPS as well as some closely relevant algorithms and methods.

The edge computing-assisted CPS framework consists of the following three planes: edge plane, inference plane and application plane. Edge planes are responsible for local data processing, including representation and filtering of local data. After the pre-processing stage, an all-round compromise is achieved where the noise and redundancy are reduced considerably while the communication cost is saved ideally.

The edge plane receives workflow data stream from the cloud plane and then provides users better QoE. The edge plane and the cloud plane complement each other to ensure real-time processing.

The multi-vendor Mobile Edge Computing platforms proposed in [76] motivates a brand new value chain, novel business opportunities and use cases that is totally out of imagination before. Such edge computing platform greatly improves the ability and real-time performance of data transmission. Edge computing saves CPU computing resources based on filtering data transmission. MEC is Mobile Edge Computing, in a hostile environment, MEC can overcome the problems originateing from the limited resource of mobile devices. Under hostile environment and high mobility, MEC can also provide sound services to end users. MEC layer can fulfill high-demanding computing tasks from mobile devices.

Additionally, MEC promotes the development of 5G for meeting the high criterion of 5G in delay, programmability, scalability and so on. The edge computing node is a logical node and the Explicit Congestion Notification(ECN) integrates network, computing, storage, control, application functions and security. ECN is blessed with high liquidity, and when taking off from an access point(AP), the ECN can safely connect with another AP Network.

There exists no doubt that end users and edge server nodes are two basic components in edge computing. Guan, Peiyuan, et al model their bidirectional relationship via game theory where end users pay for different QoS while edge server nodes choose either which end user to provide service or what task to operate on [77].

In [78], an embedded digital signal processor stands out for its impressive scalability and is adopted as a case study. What’s more, network Bit Error Rate(BER) is contrasted under both through power line and under wireless circumstances. [79] raises a novel architecture where front-haul and back-haul links are inset in unmanned aerial vehicle technologies so that the supply-demand contradiction resolves each other. The proposed the UAV-mounted infrastructure can also display their inherent capacity when it comes to the recover of a small region like the villages around mountain ranges after a horrific landslide.

[80] targets at the hostile environment where the amount of data is more than overwhelming plus some tasks call for extra computation resources and at the same time, time constraints cannot be jumped over the line. This paper puts forward an extensive edge computing platform based upon Elijah platform, taking full advantage of the active participation of human.

[81] uses simulation experiments to back up the proposed failure prediction model and communication saving strategy, and the paper also elaborates on the particular functions on edge computing of smart sensors. CEC includes edge plane, cloud plane and application plane. The edge plane collects data and extracts the corresponding high quality data. The edge planes provide these high-quality data and services directly to users. In the local CPS, some simple computing tasks, such as matching or sorting, can be implemented on the edge plane. CEC framework can effectively integrate cloud computing and edge computing. IoT/CPS data processing pattern is based on flow processing technology, which distributes the workload in the cluster of edge devices. It can solve big data problems on mobile devices.

[82] and [83] elucidate the functions of the sensing plane, cloud plane, and application plane respectively and capitalize on smart living and smart space as a qualitative use case. With increasing pressures on city resources as a motivation, smart city data ecosystems embrace environment of diversified hierarchies and cultures. The paper regards user satisfaction as the best policy taking locality contentment, energy consumption contentment, QoS contentment, and human computer interaction contentment into measurement.

Edge cloud-assisted CPS framework facilitates cities to be more humanized and reasonable with elevated efficiency, sanitation and security. By applying traffic rules and legislations, reduced city cruise time, eased traffic congestion, lowed carbon emissions and improved the efficiency of traffic management can be readily achieved as well as optimized parking space allocation. With embedded smart sensors, the system’s current state report and potential failure prediction can be get, and the data can be filtered by event generator [84].

Under the common sense that conventional RF communications is not a possible solution to IoT devices, Xie, Yutong, et al. raised deep reinforcement learning scheme concentrating on the optimized power allocation strategies despite uncertainties and locations [85]. FPGA can implement high-performance embedding requirements for applications. In the implementation of the binary neural network, the FPGA provides far more accuracy than the result of the optimized software running in the GPU. ARTICo can realize the large-scale of data and the clarity of data. ARTICo provides a good basis for edge computing in CPS [86].

[87] focuses on security provisioning in the context of mobile environment, and therefore comes up with a software-defined networking(SDN)-based verification management. Moreover, edge analytics can be applied to CPS and IoT. ENA has effective real-time collaborative computing. Processing elements are placed in the edge layer and can run powerful computing intensive software such as R and Matlab at the edge. ES can share computing tasks from mobile users and reduce delays in data transmission to the cloud, concluded in Fig. 5.

Although transmitting to data centers saves the power and storage of end devices, the consequent latency cannot be underestimated and the major challenge of that is the strategy for computation offloading in essence. As edge computing and fog computing are implemented into mobile scenarios, dynamic physical and logical resource allocation is allowed in accordance with catering to low latency requirements and high quality computing experience [88]. However, there is no doubt that blemish still exist. The difference of data types generated by four categories, i.e., Infrastructure as a Service(IaaS), Platform as a Service(PaaS), Function as a Service(FaaS) and Software as a Service(SaaS) poses difficulties when it comes to data transmission. As a result, failures happen from time to time when the operation time of computing and transmitting scheme comes in excess of a threshold, and that is what is defined as a task failure. Hence, better computation offloading algorithms and strategies will be the key to the realization of completing the computation of tasks smoothly [89].

The rich and powerful computing enabled CPS is an optimized fault tolerance mechanism where better services can be fulfilled when faults come across both in hardware and software [90]. Accordingly, tasks are fulfilled before hard deadlines and at the same time, the Quality-of-Service(QoS) is optimized to the maximum extent. Because of the difference of data processing among edge computing, fog computing and cloud computing, it is also important to improve the data filtering rules between different levels of the algorithm [91].

Therefore, with next fifth-generation(5G) wireless networks era just around the corner, heterogeneous data convergence worth weighted attention in the future [92]. Besides, in a gesture to avert resource waste in cloud data centers, the future resource allocation needs to be estimated in real-time. Thus further creative algorithms and strategies are to be proposed.

Under Privacy-Preserving Protocol based on ISO 11770-3, privacy violations have declined to a certain extend since its implementation. In contrast with traditional cloud data centers where the level of concentration cannot even be higher, the emergence of distributed computing dispatches workflow in means of allocating packs of data to a series of edge servers by pushing them geographically closer to mobile users.

However, it should be confessed that entities involved are still prone to be harnessed when privacy-sensitive information comes into the limelight. Undeniably, with data encryption and Blockchain, for any intruder it is tough to alter the stored data at majority places on the basis of the distributed authentication and trust among the peer nodes of the IoT system [89]. But the price-performance ratio may not worth it. Under current framework, information services are confronted with severe privacy-leakage challenges including unknown system vulnerabilities and backdoors during the entire lifetime of stream flow, embarking upon information aggregation, memory, processing, printing, and release in turn [93].

Research on achieving broad IoT data sharing within the spectrum of information security enjoys promising prospects.

Application Programming Interfaces(API) are responsible for satisfying user needs and meanwhile, sustaining data interoperability. Current data managing strategy highlights the emergence of new challenges: normalized bandwidth and transmitting capacity. Transparent and seamless data sharing across ecosystems allowed by technological openness and open value exchanges between government and SME enabled by organizational openness call for a better tomorrow big data applications [94].

The future researches should involve establishment of an opportunistic prototype for CPS, just as we have mentioned in “Dynamic resource provisioning for CPS in cloud computing” Section, better optimized scenarios along with use cases will play a pivotal role in the maintenance of big data applications, like smart city, smart grid, automated driving, etc. Besides, how to integrate data from previous scenarios with the possible behavior predictions of a certain user based on his preference deserves critical attention, which calls for the accuracy and applicability of data centres.

On the basis of CPS, Cyber-Physical-Social system(CPSS) is a system which furthers the integration with social information and extends the research to social network system, enabling individual or organizations to control the physical entities in a reliable, secure and collaborative way through the network. CPSS is also thought of as a hybrid world in which humans, objects and cyber participants’ community detection can be obtained. CPSS is also thought of as a hybrid world in which humans, objects and cyber participants’ community detection can be obtained.

In the future, owing to higher level of data sharing and transmitting, it would be a promising aspect to explore on the optimized operation of data exchanges taking into account task offloading strategies, energy consumptions and privacy concerns, especially when applying and extending those research applications into realistic life scenarios.