Governance and sustainability of distributed continuum systems: a big data approach

The self-healing power of human beings creates a dynamic equilibrium for the deprived body. With the self-healing mechanism, one can endure for a long time The human body has a brilliant self-repair system that kills cancer cells, attacks infectious agents, repairs broken proteins, keeps coronary arteries open, and prevents aging naturally [16]. Consider the simple case of how the human body heals a cut on a specific part of our body, such as a finger [17]. The human body heals the wound in four stages which include rapid hemostasis (bleeding control), appropriate inflammation, proliferation (repairing tissues), and reshaping (maturation of tissues) [18]. Each stage is associated with key molecular and cellular events as well as a multitude of secreted factors that coordinate them.

During the initial phase of hemostasis, blood cells are restricted from moving through the vessels. When a blood vessel breaks, platelets adhere to each other to seal the opening. During the inflammatory phase, the body’s repair and healing cell moves to the site of the wound to protect it from infection. Bacteria, pathogens, and damaged cells are eliminated during this phase. As a result of inflammation, swelling, pain, and redness are caused by several white blood cells, enzymes, nutrients, and growth factors. A third phase of the process is proliferation, which is responsible for ensuring that newly formed blood vessels provide enough nutrients and oxygen to the granulated tissue. In the final reshaping process, the wound is closed and the collagen is reshaped as it was originally. Additionally, this phase removes dead skin cells and thickens the skin where the wound was inflicted [19, 20]. DCCS can also be modeled as a self-healing system, similar to the human body, by acting autonomously [21].

This paper models a self-governing and sustainability conceptual framework with action capabilities that are similar to how human bodies heal themselves. Governing DCCS must consider their current state and behavior. Hence, this is achieved through continuous monitoring and state forecasting based on the analysis of data collected from the devices. However, analyzing DCCS data is challenging due to the complexity associated with its data structure, and the heterogeneity of the devices. This paper presents the challenges associated with DCCS’ data and provides the simplest ways to minimize downtime through governance and sustainability. The contributions of the paper are summarized in Sect. "Contribution" .

The major contributions of this article are summarized as follows:

The remaining sections of this article are arranged as follows: In section "Background", we present information about the data sources, data types, and formats in DCCS. Furthermore, we present various challenges that arise due to the data heterogeneity and differences between manufacturers. In addition, we describe how DCCS’ data matches the characteristics of Big Data. Section "Problem definition" provides the problem definition and the importance of the proposed governance and sustainable conceptual framework for the near future. Section "Existing solutions or tools for IoT monitoring" discusses the most common tools available in the literature to monitor IoT networks, we examine these tools and summarize their benefits and limitations. Section "Proposed governance and sustainability framework" provides the proposed general model for the governance and sustainability of DCCS. Future research directions are covered in Sect. "Future research directions" . Finally, we conclude the paper in Sect. "Conclusion" .

DCCS are composed of a large number of heterogeneous computing devices, which perform large computations by coordinating their resources. The overall infrastructure of DCCS is highly scalable, flexible, and resilient to perform these computations effectively [22, 23]. During these computations, each device produces a vast amount of data, which is discussed in the following subsections.

A vast number of machines and humans interacting with them continuously generate structured, semi-structured, or unstructured data. These massive amounts of data cannot be handled through conventional storage or processing approaches[29]. Big data is a term used to describe an immense amount of data that is too complex and too large to be processed and analyzed using traditional data processing tools [30]. Big data has become increasingly important in recent years because it allows organizations to gain insights from this data. By analyzing large volumes of data, organizations can uncover hidden patterns, correlations, and trends that can help them make better decisions and improve their operations [31]. It is generated by a variety of sources, such as sensors, social media, agriculture [32] and transactions, and it is typically stored in distributed systems such as Hadoop or NoSQL databases. Because of its size and complexity, big data requires specialized tools and techniques to process and analyze the data, such as machine learning (ML) and distributed computing. Big data is often characterized by the 5Vs: volume, velocity, variety, value, and veracity [33, 34], which are summarized as follows:

Characteristics such as Value and Veracity help to gain insights and achieve the quality of the big data.

In the current digital age, data plays a crucial role in several fields e.g., Business, Healthcare, Industry automation, Education, and DCCS data is not an exception. DCCS applications use business data analytics, but the data associated with the devices is usually ignored. There is a lot of data available or generated by each device, which will be discussed in Subsection  and it comes in different formats. These data can be used as input to the governance and sustainability model to perform efficient monitoring and predictions about the DCCS condition. Since DCCS data is in different formats (unstructured), various methods are mentioned in the literature to convert it to structured data such as data modeling, text mining [53], data wrangling [54, 55], natural language processing, etc. Evaluation metrics such as detection accuracy or false alarm rates require ground truth tables that assign labels unambiguously to all events. Moreover, big data analytics tools can perform them without having to convert data from unstructured to structured form [56, 57].

Because DCCS data poses several challenges (subsection ), it leads to regular issues in ML datasets such as missing values, high dimensionality, label imbalance, lack of generalization, and noisy data. Several methods have been proposed in the literature to mitigate these challenges, including feature engineering, data preprocessing, data augmentation, and dimensionality reduction [58]. Nevertheless, addressing these challenges requires additional computing resources, which is a complex task for constrained edge devices in DCCS (since we are distributing the loads among various parts of the systems). In addition, it requires knowledge of data that is being converted from unstructured to structured. Here, representation learning plays a significant role in extracting the underlying features from unstructured and non-linear data, which helps to mitigate the challenges discussed in subsection .

Many preprocessing approaches have been proposed in the literature for mitigating missing, noisy values, outliers, and inconsistent details within the given data. They are efficient at performing these tasks but they do not ease the extraction of knowledge from them. So, learning data representation makes the knowledge extraction process easier and helps to build efficient predictors and classifiers. The ReL strategies consider a vast amount of data and perform multiple transformations to extract useful information from it [59]. They also represent it in a way that is easily accessible, such as through graphs. There are several benefits of using representation learning including scalability, structure preservation, information preservation, data sparsity, and easy-to-identify explanatory factors among the data [60, 61]. Due to these reasons, learning representation approaches are used in several applications such as Natural Language Processing (NLP), speech recognition, and signal processing [62]. To simplify the process of knowledge extraction, we should also consider learning representation for resource-constrained devices, such as DCCS [2].

DCCS data contain several complexities which are discussed in Sect. "Challenges of DCCS’s data", while representation learning can handle all of those complexities. In addition, ReL can minimize manual preprocessing or data modeling time. It reduces the dimensionality of data and compares large amounts of data to useful features. ReL can remove inconsistencies, noisy or outlier values, and recover missing values from the data during this process. Because the dimension is lowered, learning algorithms require less effort to analyze and produce knowledge. On the other hand, the best useful features help to improve the accuracy of the results. There are several ReL algorithms in the literature, and the most popular are Graph Representation Learning (GRL), Contrastive Representation Learning (CRL), Bayesian Network Structure Learning (BNSL), Matrix factorization, Random walk learning (RWL), Deep walk, etc [23]. We present an illustrative example to better understand how ReL (through BNSL) can generate structures, which further supports knowledge discovery.

The term “knowledge” in the DCCS governance and sustainability model refers to information that can be extracted from data through analysis or insight. An imperative determinant of the quality of the knowledge that has been acquired from learning data is its ability to be generalized. The accumulation of knowledge is the result of interactions with a pool of information. It is easier and faster to monitor, predict, and make decisions when knowledge is extracted. There are several ways to represent knowledge, but graph representation is an easy way to store and visualize [68]. The knowledge graph is a comprehensive view of the nodes, edges, and labels. On ingestion of data, they are capable of identifying individual objects and their relationships with one another [69]. An extended form of knowledge representation is the use of DAGs as illustrated in Fig. 9, which are easy to query and provide decisions or answers.

DCCS involves a wide variety of devices (from the edge to the cloud), each with its capabilities. Hence, testing the performance of each device is essential to achieve maximum benefit. As there are so many devices present and their locations also vary, it is impossible to keep track of them manually. However, there are several monitoring tools in the literature (discussed in Sect. "Existing solutions or tools for IoT monitoring" ), but they are not useful for monitoring the entire DCCS. The governance and sustainability model is used to monitor the condition and efficiency of these devices autonomously, to maximize their benefit from them. Thus, the knowledge extracted from DCCS data helps identify the root cause of the problem, such as manufacturer’s defects, network issues, inefficient energy use, etc. The ready knowledge helps to identify or predict the issues in the devices and notify the administrator or try to self-heal to achieve the best usage of devices. As a result of readily available knowledge, we can identify or predict issues with devices and report them to the administration or attempt to self-heal to achieve maximum effectiveness. In addition to showing the issues, predictions and monitoring variables are also useful in improving the efficiency of resource utilization. Due to the limitations of the computational resources of edge devices, it is difficult to determine their availability for performing tasks [70, 71]. In addition to determining the most effective use of resources through prediction, the knowledge extracted from the data also assists in identifying the most appropriate future needs for the devices and their resources. However, it is necessary to verify the reliability of the suggestions made by governance and sustainable models, and this is a challenging issue. This challenge can be mitigated by leveraging SLOs [72, 73]. Whenever monitoring or predicting issues are identified through querying, their reliability must be verified by mapping them to SLOs before a decision is made.

In the governance process, notification plays a vital role in resolving errors in DCCS. However, our model monitors and predicts several aspects such as networking, computation efficiency, device health, etc., so it is crucial to decide how these notifications are delivered to administrators. These notifications are usually in text format and can be received via text messages, push notes, emails, or sent to visualization tools. In any case, it is necessary to set up who can receive these messages (privacy concerns), what frequency of notifications is appropriate, and what level of information to transmit. Multiple operations are running on DCCS simultaneously, and it is important to decide whether notifications should be delivered to relevant people who monitor specific aspects of DCCS. Privacy concerns make it necessary to restrict notifications to unauthorized entities.

Another significant metric to take into account is the frequency of notifications since more frequent notifications require more computational resources. However, continuous monitoring is a requirement of DCCS, so determining the frequency is crucial. It is imperative to know what level of detail should be notified to the administrator. Consequently, if the administrator needs to solve the issue, they must know the details of where the problem arises, or what the root cause of the issue is. Since the environment is dynamic, we need an adaptive framework, it will be decided dynamically depending on the issues raised. When there are no issues, report the performance metrics to the administrator without providing detailed analysis. When a device fails to respond, there is usually a detailed problem such as either a network failure, power supply failure or any other problem. Once the problem is identified, the administrator can solve it. To establish dynamic DCCS governance, it is necessary to make autonomous decisions based on these notifications. These notifications are further analyzed or compared using SLOs to decide if the notified problem can be rectified, or if it truly caused a problem or not.

The visualization can be a dashboard that provides analytical results in a graphical interface. Visualization provides a powerful means of monitoring DCCS performance in terms of various metrics, enabling the user to understand patterns and malfunctions more easily. These visualization metrics include network performance, computational efficiency, resource usage, etc. Visualization tools can alert administrators by highlighting the fault. Graphite and Grafana [74] are open-source tools that provide a visualization with flexible and custom dashboards. In the literature, these tools are used in MetricFire [41] for efficient visualizations. These tools are popular for visualization because of efficient storage in databases, having a web-based interface to interact with data and easy scripting for customization. Recent advances such as digital twins [75‐77], AR/VR [78], and metaverse technologies [79‐81] can also help improve DCCS visualization effectively. These tools are very efficient at identifying the fault and visualizing them to the user/administrator, but they are not able to solve the problem autonomously.

In the governance and sustainability model, the decision-making phase shows that some actions will be taken to correct problems identified in previous phases. These decisions are categorized into autonomous and manual decision-making. Autonomous decision-making in this context refers to the DCCS being able to rectify issues without human intervention. This can be achieved through predefined rules or using learning-driven approaches. When a rule-based approach is used, the notifications match the SLOs and make the necessary decisions. A learning-driven approach monitors DCCS behavior continuously while making dynamic decisions. Among these two approaches, SLOs provide quick decisions and balance time and computation, whereas learning-based methods are computationally intensive. However, combining both scenarios leads to better results and helps with dynamic decision-making. Manual decisions are made by the administrator based on identified problems that are not possible to address autonomously. E.g., network failure, hardware damage, being disconnected from the network through physical failures, etc. This means that problems that are not addressed through software instructions must be handled by humans or robots, but not autonomously by the system.

Once the decision-making system recommends an autonomously solvable problem, it will be further received by the action stage to cure the failure or fault. Other issues are handled manually by the appropriate administrator. The autonomously solvable problems are solved similarly to the self-healing process of the human body (this paper considered wound healing). The human body evolved over millions of years to develop a complex and highly effective self-healing mechanism. The human body and DCCS have many similarities, and using this self-healing system not only helps improve performance but also minimizes managing costs and performance improvement. Simulating an similar system is extremely difficult but not impossible. However, it can be achieved through continuous monitoring and analysis of collected DCCS data.

The mechanism of self-healing in the Action phase of the governance and sustainable DCCS model is illustrated in Figure 10. The four stages of the wound healing process described in motivation (subsection ) are described concerning DCCS as follows.

DCCS are designed to provide efficient computations for business data. But, it is necessary to govern (through monitoring tools) the devices to get longer sustainability with minimal downtime. These monitoring tools have to do additional analyses of the information from various sources. So, it is necessary to trade off the computations between the business and device data, which imposes several challenges. They are summarized as follows:

It is possible to perform monitoring tasks anywhere in DCCS, such as cloud, fog, or edge. But, performing this task at the Edge is more beneficial in terms of shortening the delay in providing services. As AI and ML strategies have become more widespread and beneficial, most edge devices are using them to some extent. Edge intelligence has become popular because of its minimal latency and quick decision-making. Since AI/ML approaches are computationally hungry, they consume more computation and use more energy and memory [82]. Due to the rapid power drain of edge nodes, they can only sustain for a short time. Hence, it is pertinent to determine whether AI/ML is needed for edge devices and minimize unnecessary computations as much as possible. It does not mean stopping the use of AI/ML techniques at the edge, but choosing the most appropriate algorithms based on the application, which can be efficient in the trade-off between efficiency and long-term sustainability. There are several ways to minimize edge computations, mainly when learning algorithms are run. This includes 1) distribution of learning among multiple edge devices [83, 84], 2) skipping some parts of the interactions that are unnecessary, e.g. frame skipping methods in deep neural networks [85], 3) provide input as knowledge or features [86], 4) decide on an optimal number of layers sufficient to produce maximum accuracy [87], etc. The majority of learning models extract correlations between variables but do not extract causal relationships. So when comparing causality with correlation, causality results in better predictions. In addition, causality approaches are computationally more efficient than correlation approaches. So, there is a chance to minimize computations at the edge through causality.

DCCS are an example of a system in which multiple related but not identical tasks are performed simultaneously. E.g., the information on a device configuration and its network configuration is not identical but related. Learning such tasks can simultaneously improve the system’s overall performance and also accuracy [88]. Multi-task Learning (MTL) strategies can optimize multiple objective functions simultaneously through learning different tasks. These approaches initially classify the jobs from the given data and perform learning separately. In the computing continuum, MTL is used to identify inference attacks [89], traffic predictions in the IoT [90], computational offloading at Edge [91], among others. Especially through MTL, different tasks, such as device monitoring, network monitoring, and privacy preservation, can simultaneously perform operations from the available system data. Since we are working with resource-constrained devices in DCCS, distributed MTL can be more promising [92]. It can minimize the load on a single machine, and the tasks/learning models are distributed among multiple devices.

Zero-touch Provisioning (ZTP) is an end-to-end network service provisioning model to configure network devices [93, 94]. The devices in DCCS software and configurations need to be updated in a timely manner. There are several heterogeneous devices that make it difficult for humans to perform it. In order to fill this gap, software services like ZTP can be used in governance and sustainable frameworks. The ZTP is used for configuration management in the literature for several network-based services. However, the current functionalities of the ZTP could be improved [95]. Extending these services can help the user or administrators rectify issues diagnosed through the proposed reference model. Currently, the ZTP services are running through centralized cloud servers, but it is necessary to decentralize them. When working with a large to a vast number of computing devices, the distributed ZTP increases its services and rapidly completes its assigned tasks. Consequently, by using ZTP, DCCS becomes more autonomous and is able to reduce failures or response delays without the involvement of humans.

Efficient DCCS monitoring requires the use of diversified information from each device that includes sensitive configuration details. The configuration files are sensitive because they contain information such as passwords, access keys, or several authors’ information that cannot be shared with unauthorized parties. When the models we propose are considered to learn or monitor devices, each piece of information can determine the quality of the monitoring. So, an efficient monitoring tool can also protect our data and must follow privacy rules. Misusing this information can cause a device or network to malfunction. An efficient way to access the configuration details of any device is to use encryption mechanisms [96]. There are several secure data or file transmission approaches available in the literature, such as Secure FTP (SFTP/FTPS), and Secure Shell (SSH), which enable authentication mechanisms before accessing the information for learning, version control, etc. The learning process performs only read operations to learn from the given input data to monitor the systems, but there is a chance that secret information can be misused when privacy is not enabled. Zero-trust architecture is a solution that can protect accessing such sensitive information [97, 98].

From Subsection , we understand how the human body as a complex system can heal itself to sustain the human body longer. Since the behavior and complex functionality of DCCS resemble the human body [21], there is a chance to reflect similar healing features in DCCS. However, understanding the human body is a challenging task compared to the complex behavior of DCCS. Therefore, the information from the DCCS can help identify the root cause of the problem in the systems and further use this information (see Sect. "Source of DCCS’s data") to heal a problem without external instructions or human interventions. Using efficient, lightweight, and accurate learning algorithms that can understand the underlying information from DCCS’s data will further help in understanding the root cause. Understanding the root cause of any problem helps to fix the problems easily. However, deciding the accuracy of the identified cause is a challenging issue.