1 Introduction

Nowadays, many traditional businesses are moving their workloads to Cloud Data Centers (CDCs) [1]. These businesses are experiencing the advantages offered by Cloud Service Providers (CSPs), like cost savings, reliability, performance improvement, scalability, and flexibility. CSPs run the CDC and provide pay-per-use access to computing resources [2, 3] as well as a variety of other services to customers. They operate data centers (DCs) in multiple geographic locations [4], allowing them to provide redundant infrastructure [5] and backup systems [6] to ensure that services remain available even in the event of a failure or outage [7]. Many processes exist to secure data and applications from unauthorized access [6] and other security risks [8] inside the CDCs. These include firewalls [9], access controls [10], encryption [11], and intrusion detection/prevention systems [12]. These advantages have led to a meteoric rise in the number of CSPs and the range of services accessible in the cloud [13].

This rapid growth generates problems like fluctuating demand for resources [14], a lack of Quality of Service (QoS) [15], an uneven distribution of work [16,17,18], energy efficiency [19, 20], and many others. The dynamism and unpredictability of the resource's demand influence their accessibility during scheduling [21]. Hence, effective resource management is required so that work can be planned according to their execution demands [22, 23]. The QoS is the assessment of an entire service's performance. It is directly affected by the increase in service requests and the exponential growth of cloud users. To tackle these challenges, efficient prediction-based scheduling is required [24, 25]. Irregular load distribution across Virtual Machines (VMs) leads to inefficient utilization of resources, which affects the scheduling. A prediction-based scheduling scheme that handles the unbalancing of data is required [26, 27].

The scheduling of resources in a Cloud Workload (CW) is not only one of the fundamental difficulties in computing, but it also impacts the primary interests of cloud regulars and facility providers [28]. In fact, this issue is an NP-hard multi-constraint and multi-objective optimization problem [29, 30]. This signifies that finding the optimal solution is very difficult. To overcome these challenges, many researchers are working on integration of prediction methods with heuristic, meta-heuristic and hybrid algorithms [8, 27, 31]. The prediction method helps to estimate the future workload placed on a data center and provides the possible resource consumption patterns. And these patterns provide us appropriate resources and a significant step to building a resource-efficient scheduling method. Therefore, accurate forecasting is essential for preventing performance decline and reducing resource wastage, both of which improve revenue [32]. In CW, managing resources properly allows activities to be scheduled based on execution situations [23].

Many business sectors currently rely on cloud-based workloads [33, 34] and here optimizing resource consumption [33, 35, 36], is an essential component in this context. Hence, it is vital to understand which workloads are appropriate for customers and CSPs to make optimal use of available resources. Next, the delivery of computer resources (such as CPU, memory, network, virtualized servers, etc.) via the internet using cloud computing has become the trend [25, 37, 38]. For the purpose of offering these resources as a service in the cloud market, the CSPs have their own optimization goals. These objectives include limiting maintenance costs and boosting the money produced by underlying computer resources. Furthermore, in order to achieve these objectives, individual resource utilization needs to be predicted and the demanding resources must be deployed properly in the operational environment. However, in this perspective, workload prediction gives more definite or deterministic future knowledge about resource demands. Also, it allows for making appropriate choices to optimize resources in cloud data centers.

The above paragraphs conclude that accurate prediction and optimum scheduling are the most important aspects in achieving the optimization goal. A lot of researches [23, 39,40,41,42,43,44,45,46,47] have been done to work on these two techniques that help cloud users and providers to make better decisions for accurate load distribution. Also, researchers need to have a complete knowledge of workload and techniques that helps to enhance its performance. In the literature, various intelligent methods and algorithms are reviewed for the selection of best techniques from a large pool of options. Various researchers come up with different ways to solve the problem, and each suggests that their approach is the best. For further clarity, in this survey a Systematic Review Study (SRS) has been conducted. The current research aim is to compares several existing methods for making decisions to choose CW and its suitable techniques. The work also lists the pros and cons of each method and future research directions. Consequently, our research examines the robust specifications of several Cloud Workload Applications (CWAs). This study discusses the theoretical background, current methodologies, and performance metrics of the research topic. This research is anticipated to benefit both scholars and business professionals.

1.1 Motivation

In the present time, many researchers are continuously working to find an optimal solution for their respective fields and coming up with new ideas. For CW, the approaches such as scheduling and prediction plays very important role by ensuring that resources are efficiently allocated to execution requirements of cloud users [25, 40, 48, 49]. Also, these techniques helps to solve the problem of resource allocation [50], inaccurate prediction [51], improper load distribution [52], also improvement in cost [34] and performance optimization of the system [53].

An example is discussed considering a CWDC having multiple virtual machines (VMs) that execute different workloads. The CWDC is equipped with the following resources: 64 GB of RAM, 16 CPU cores, 1 TB of storage, and 1 Gbps network bandwidth. Now, let's suppose there are four VMs running in the DC, and each of them needs the following resources:

  • VM1: RAM = 16 GB, CORE = 4, STORAGE = 250 GB, 200 Mbps network bandwidth.

  • VM2: RAM = 8 GB, CORE = 2, STORAGE = 100 GB, 50 Mbps network bandwidth.

  • VM3: RAM = 32 GB, CORE = 6, STORAGE = 500 GB, 500 Mbps network bandwidth.

  • VM4: RAM = 8 GB, CORE = 2, STORAGE = 150 GB, 250 Mbps network bandwidth.

To allocate resources to these VMs, CW can use a resource management mechanism such as a hypervisor or container orchestrator. It can be configured to optimize efficiency, reliability, and affordability while allocating resources. Here scheduling technique is required for allocation of resources to the particular task and prediction helps to predict the resources to improve system performance. In the above example during allocation process, if VM1 and VM3 are given the resources they need to perform their workloads, while VM2 and VM4 are given fewer resources as they have lower resource requirements. This allocation also ensures that there is enough network bandwidth available for all VMs to communicate with each other and the outside world. Therefore, resource allocation can be a complex process in a CWDC, especially when there are many VMs running different workloads with varying resource requirements. A good resource management technique such as prediction-based scheduling can help to ensure that resources are allocated in a way that optimizes performance, availability, and cost.

From the previous studies found that limited surveys are conducted for CWAs. Also, there is a lack of information related to its dataset, resources, prediction techniques, scheduling algorithms, and their associated factors. Therefore, this study addresses challenges and motivates us to conduct the survey. The current work covers information related to similar types of CWA and details using various approaches with their appropriate solutions. It also includes the research questions related to the CW. In existing study, prediction models and scheduling techniques have been implemented in different applications of the cloud for proper utilization of resources. Hence the current research aims to address the requirement for a SRS for prediction-based scheduling techniques in CW.

1.2 Contribution

The contribution of current SRS is mentioned as follows:

  • Review existing survey papers related to prediction-based scheduling techniques for CW and highlighted their issues.Provides research questions related to the problem and gives appropriate solutions.

  • Various criteria for extraction and evaluation of data in CW are discussed and analyzed.

  • Selected workload is analyzed for prediction and scheduling-based strategies with associated metrics.

  • The research challenges are discussed with future directions in prediction-based scheduling techniques.

The remaining parts of this research are organized into the following categories: The methodology of the study is discussed in Sect. 2. The description of the theoretical background of the topic can be found in Sect. 3. The research questions are answered in Sect. 4 by doing an analysis of the chosen publications, and the material that is necessary for this study is supplied. The research problems and possible future directions are presented in Sect. 5.

2 Research methodology

The purpose of a systematic literature review (SLR) is to search, examine, and draw conclusions from all published works in a certain field of study. The term SRS is a kind of SLR in which latest research material is compiled and organized to provide a comprehensive description of a particular domain through a uniform and efficient research methodology. In this survey, the objective of SRS is to gather and evaluate past research on prediction-based scheduling techniques in CW. Inspired by survey paper [21], the Research Questions (RQ) are formulated at the beginning to serve as a framework for the creation of search strings. An appropriate examination of digital libraries is performed based on a search string to provide an answer to the RQ stated in the first step.

2.1 Research questions (RQ)

In SRS, preparation of RQ is one of the most essential steps, as research findings are examined in the consideration of these questions. This study is intended to answer some RQ. To obtain clarification, a comprehensive literature review has been conducted. This section is devoted to identifying six survey questions. The aim is to find answers to these RQ by examining related papers. The RQs generated with the help of SRS are given below:

RQ1: How to identify dataset and appropriate resources related to CW? (Answered in Sect. 4.1)

RQ2: What are the existing studies that address the issues of CW using predictive models and also mention its parameters? (Answered in Sect. 4.2)

RQ3: What are the current scheduling techniques and parameters that are working to increase performance of the system? (Answered in Sect. 4.3)

RQ4: What are the recent works of prediction-based scheduling related to CWs? (Answered in Sect. 4.4)

2.2 Article search strategy

This section describes the SRS article search strategy. The criteria for articles selection is choosing those studies which are related to CW. Finding relevant studies in digital libraries serves as a preliminary step in conducting a systematic review of the available literature. In SRS, the search technique is essential and has an impact on general performance. This investigation searches published publications from 2015 to the present in two phases: generalized and targeted. As illustrated in Fig. 1, the article search strategy is separated into three stages.

Fig. 1
figure 1

Stages of article search strategy

Full size image

2.2.1 STAGE 1: Start searching by entering relevant strings and keywords

The search for strings and keywords is carried out in two stages: generalized and targeted. Google Scholar is used to identifying papers pertinent to the study issue throughout the generalized research phase. The targeted phase identified relevant research articles in four digital libraries (IEEEXplore, SpringerLink, Wiley, and Elsevier). The main actions are carried out during this phase:

Step 1::

Building strings and keywords from RQ is the initial stage in the article search strategy. Strings are generated by concatenating the keywords workload prediction, cloud scheduling, resource allocation, resource management, prediction-based scheduling, cloud resources, load balancing, resource overutilization, resource underutilization, and quality of service.

Step 2::

Second, we performed both generalized and targeted searches for keywords and strings to identify relevant articles. The search returned 15,639 journal, book, conference paper, note, and chapter, etc. publications that included the previously defined keywords or phrases.

2.2.2 STAGE 2: Eliminating unnecessary and redundant contents

Considering problems related to research, only called high-research articles for our analysis. Also, we place a significant amount of importance on works that have been published in English and have appeared in reputable journals or conferences. Furthermore, unpublished work, reports, publication notes, and book articles are not added in the SRS. Cloud and similar principles include papers selected based on their titles in prestigious journals and conferences that deal with prediction-based scheduling. The following actions are performed during this phase:

Step 3::

Third, we filter out results that have already been chosen in the first two steps (generalized and targeted). (6478 papers removed)

Step 4::

The fourth step involves sorting the chosen publications by their titles and removing those that don't belong. Finally, it is concluded that 8024 studies have been collected in Stage 2.

2.2.3 STAGE 3: Paper selection based on quality and relevance of data

At this step, the entire texts and abstracts of the chosen papers were reviewed to determine their relevance to the RQ. The participation of each paper was determined by the research problem's relevance and the year of publication. For additional analysis, the abstracts and keywords of the chosen publications were entered into spreadsheets.

Step 5::

At this phase, papers are rejected if their abstracts don't have any relevance to the subject under study. (Removed 3997 articles).

Step 6::

In this stage, the author studied the whole text of each article. The rejection of articles that failed to address the RQ (described in Section 2.1). (Excluded 3886 articles).

In conclusion, the selected papers are submitted to three phases of review, and only relevant research studies capable of addressing the research questions are selected for further examination. Through this procedure, 15,497 publications were eliminated. In the end, 141 research papers were chosen for the SRS from a total of 15,639 research articles.

3 Theoretical background

This section provides the information related to CW which includes definition, importance, different reasons to adopt CW, types, and challenges.

3.1 Cloud workload (CW)

It is refers to the total amount of work to be performed, whether it in real-time by users interacting with cloud services, or in batches [54]. Also it is a collection of application resources that support a common business goal with multiple services, such as data stores and APIs, functioning together to deliver exact end-to-end functionality [55].

3.2 Importance of cloud workload

Today, a lot of businesses are transferring traditional workload to the cloud [8]. The Cloud Security Alliance (CSA) published its study on January 11, 2019, which examined the condition of cloud adoption and the effect of the cloud on enterprise resource planning applications [56]. Almost 70% of organizations asked questions associated to move, are in the process of enterprise resource planning (ERP) data and analyzing the time of workloads to cloud environments, although most of them are concerned to move [39, 57]. Experts mentioned the main advantages of migration toward ERP schemes to cloud paradigms:

  • Scalability: (65% of respondents) the primary advantage of cloud migration for new technologies.

  • Lower cost: (61% of respondents) lower ownership costs are a significant benefit of cloud computing.

  • Security: (49% of respondents) update and regular service upgrades are a strong reason for migrating to the cloud.

3.3 Reason to adopt cloud workload

In article [58] Gartner suggested some key points why enterprise workloads moving to the public cloud given below:

  • Mobility: Mobile technologies and apps make remote work much easier. The adaptable cloud approach is an excellent choice for mobile solutions.

  • Disaster recovery: Cloud-based disaster recovery is both cost-effective and safe. It also saves business costs and efforts of maintaining redundant production-quality infrastructure in a different location.

  • Web conferencing: Due to the pandemic, this virtual meeting facilitator has become a significant operational role. With the varying networking bandwidth needs, large enterprise cloud providers can reliably supply videoconferencing solutions.

  • Collaborative effort and information management: Collaboration is an effective way whereby persons collectively work on a common problem to achieve business benefit. This demonstrates the cloud’s usefulness for business productivity software

  • Remote workstation management and virtual desktops: A depend-able Virtual Desktop Infrastructure (VDI) is essential for allowing remote work. Cloud-based virtualization and Desktop as a Service (DaaS) are now commonplace, offering a scalable and secure alternative to traditional DC based solutions.

3.4 Challenges in cloud workload

With the advancements and development of cloud computing services, businesses confront a variety of cloud computing issues. The cloud provides businesses with advantages, but resource management is not a simple procedure. The dynamic workload may lead following issues with associated to cloud-hosted services. The points mentioned below are the hurdles that must be overcome to use the cloud benefits efficiently.

  • Dynamic environment: The business environment is quickly changing that impact the market [23, 59, 60], businesses have to incorporate these changes and provide new ideas, solutions, and services to retain technology and new developments.

  • Understanding of resource demands: It may be dangerous to distribute resources arbitrarily without first analyzing incoming requirements, identifying priorities, or considering the organization’s goals [61, 62]. Before distributing resources, it is essential to have full awareness of coming demand and resource availability according to matching skill sets and responsibilities.

  • Over-provisioning: It can be a host or computational node that has assigned unused computing resources namely CPU, memory, I/O, disk, or network at peak time [25, 63, 64].

  • Under-provisioning: The issue of resource under-provisioning occurs when service gives out fewer resources than are needed, it won’t be able to provide a good service to its users [51, 63, 64]. If there aren’t enough resources for the website, it might seem slow or out of reach. People stop using the Web, which means the service provider loses customers.

  • Oscillation: When oscillation (auto-scaling) is used, it leads to both over-provisioning and under-provisioning issues [65]. It is a combination of under-provisioning and over-provisioning.

  • Inefficient prediction: This type of prediction leads to unnecessary outcomes. Inaccurate prediction in the cloud raises issues of resource allocation, load balancing, energy efficiency, and many more difficulties. When training models formulate predictions, the sequence establishes an order on the data that must be maintained. In general, prediction issues involve a sequence of data which raises prediction problems [26, 27].

  • Efficient Scheduling: This type of scheduling working on these factors such as to maximize the quality [18, 32, 66], service provider productivity [67], and efficiency of the system [24, 68]. The objective of these three factors helps to explore many businesses with alternative methods. Inefficient scheduling can affect each of the three goals.

3.5 Types of cloud workload

Depending on the purpose of the application, various types of workloads are classified. They are:

3.5.1 Business-critical workload applications

These applications are required for survival and long-term activities, although its failure does not necessarily result in an instant disaster [61, 69]. During normal operations, enterprises are using a broad range of apps, but not all of them are necessary for immediate survival during power failures and other disasters. Failure of a business-critical application suffers from decreased user productivity and a worse overall experience for those using the application [39]. Although, the organization is required to perform basic operation for limited hours, without causing serious harm to operations and profits. In general, the organization may continue the task with current resources or through alternative ways to the unsuccessful system. Most of the time, the applications on the list below are considered business-critical, but this can be changed.

  • Financial applications It gives a business a way to handle money and information [8, 70]. Banks offer a wide range of financial services, each of which is made to meet a specific need. Each organization chooses the financial application that best fits its needs. Then, each application is categorized and given a priority based on how it affects the organization. Most organizations need a financial app to make sure they have a steady flow of money coming in, but how the organization gets paid is a key factor in how important these apps are. For instance, a short interruption of service might not have a big effect on a company that processes subscriptions once a month at a certain time. If an ecommerce site that needs to process purchases during the holidays goes down, even for a short time, it could lose a lot of money. Also, organizations must follow the rules when using financial applications to make sure that the transactions are safe and that sensitive and personal information stays private.

  • Messaging systems It provides information between workers, business partners, customers, and other stakeholders. Organizations make use of a broad range of communications technologies, such as email programs, text messaging, and cross-functional platforms [71]. Messages often include critical information required for typical corporate operations. Private and sensitive information, proprietary data, and trade secrets may also be included in messages. All of these interactions and information exchanges are often necessary to keep routine corporate activities running. These systems also have security issues, for example, an Email system can put the firm at risk of security. If threat actors acquire access to email accounts, they may use them to steal information, trick people into disclosing information, use them as a gateway into the business network, and engage in other criminal actions.

  • Legacy systems A legacy system is usually found for a long time within an organization’s ecosystem [72]. It is a known methodology that has possibly been modified to meet the company’s specific needs. These technologies come with a cost—the initial setup cost, followed by the cost of continuous maintenance. Because many legacy systems have not been designed to work in the cloud, they must be significantly updated when transferred. When legacy systems are designated as business-critical, they must be managed carefully to minimize interruptions. As a result, many businesses are still relying on old systems and are unwilling to shift to the cloud.

3.5.2 Mission-critical workload applications

In this workload gathering of application resources are described, which need to be extremely dependable on the platform [73]. The workload is consecutively available, quick recovery of failures, and quick operations. The organizations depend on instance operations based on mission-critical models and machines. With the condition mission critical resource experience downtime, further in brief, maybe a reason for the huge disruption and negative impacts are shown at once in the short and long term. Mission-critical machines and workloads deal with a maximum priority that needs to be continued to guarantee operations stay viable.

3.5.3 Low-priority workload applications

Organizations categorize applications as low-priority or non-critical, when they proceed with similar operations for a large duration without making use of the application [74]. This application highlights the organization suffering minimal consequences but alternatively performing the entire necessary work. These are frequently adopted because the system development simplifies operations and is highly productive.

3.5.4 Scientific computing workload applications

Scientific computing workload applications are software programs meant to do complex scientific computations and simulations using advanced mathematical models and methods [75,76,77]. These applications are often used in domains such as physics, chemistry, biology, engineering, finance, and weather forecasting, and need a huge amount of computing power. Solving systems of equations, numerical optimization, data analysis, and data visualization are often involved. They also need access to massive datasets and be able to effectively manage and modify enormous volumes of data.

  • High-performance computing applications Generally, high-performance computing refers to the practice of grouping computing power in a single technology [78]. It provides considerably higher performance than other systems such as desktop computers or workstations to resolve large issues in the field of engineering, science or business [79]. It is the capacity to rapidly assimilate data and do complex computations.

In the above paragraphs, various types of applications are mentioned whose computing criteria are different from each other. The applications such as databases and web servers are referred to as business-critical workloads [31] and can be recovered from any losses. A site transfer is possible because the data for the business-critical job is mirrored. When determining what should be prioritized, it is common practice to use a relative measure that is established and relevant to the specific demands of the industry. For instance, one corporation may consider a communications system to be mission crucial, but another may consider it to be business-critical or even low priority.

4 Reporting of literature

This section provides the solution of RQ by analyzing the selected papers with their discussions to resolve the issues.

4.1 Identify dataset and appropriate resources for CW (RQ1)

In this section, datasets for CW from selected literature are analyzed. Next, the resources used in previous studies are examined based on their maximum utilization. The dataset and resources both are analyzed; this process will help the researcher for selection of dataset and choose appropriate resources for their study.

Earlier research has demonstrated that the dataset is generated either synthetic approach or real time approach. Workload generators are used to produce synthetic workloads, though real workloads are collected via benchmark datasets namely NASA dataset [80], Google Cluster Trace [33, 60], etc., or recovered from existent cloud platforms. To provide a better understanding, Table 1 lists the various datasets in relationship to the CW categories also provides its description.

Table 1 Cloud workload dataset description
Full size table

Based on the data in the table, we can conclude that category of business-critical workload has many traces which comprise the same type of features. For example, Materna, Bitbrain, Alibaba, and Google Cluster have cloud resources-related features. Dynamic prediction and scheduling can be computed by integrating these datasets. The next category is a scientific workflow which is widely used in a variety of fields, including physics, chemistry, biology, and engineering. Last, the mission-critical workload is essential to the operation of a business, organization, or system.

Also, the description given by this section helps the researchers to have a clear picture of the problem area. The information provided here helps them in the initial phase of investigation and also gives direction to the designing phase. Table 1 helps the researchers to know which dataset falls under the category of CW.

Next, the survey analyzes cloud resources which are also an essential part in CW. The DC provides internet-based supply of computational resources, such as Artificial Intelligence (AI), processing power, storage, networking, databases, analytic, and software applications over the internet. By outsourcing these services, businesses may obtain computing assets on-demand basis without having to acquire and maintain an on-premise IT infrastructure. This allows and permits for adaptable resources, quick innovation, and fast scaling of economies. Table 2, presents the list of cloud resources this analysis helps the researcher finds the most appropriate resources for resolving the various issues in CW.

Table 2 Cloud resources description
Full size table

Table 2 concludes that both CPU and memory are highly occupied resources in previous studies. In future work, the researcher may include other resources to test and check the effect on the performance of the system.

4.2 Predictive models and their effective parameters (RQ2)

Workload prediction is an important for intelligent resource scaling and load balancing that maximizes cloud service provider’s economic development and user’s Quality of Experience (QoE) [18]. Predicting workload helps cloud-end cluster maintainers determine whether the resource allocation method is appropriate or not [85]. To properly manage cloud resources, it’s essential to accurately forecast VM workload for resource provisioning [17]. Scheduling of resources is a set of rules and policies used to distribute jobs to appropriate resources (bandwidth, memory, and CPU) to maximize performance and resource usage. The proper planning of resources is a critical issue in cloud computing to improve the overall performance of the system. The present study describes the benefit of algorithms that are used most frequently in the past and suggests new algorithms as per the demand.

In a CDC load prediction is required for the allocation of resources that depend on demand applications [107]. According to [60] for efficient resource utilization in the cloud, ML approaches are used in the development that produces reliable prediction solutions. Many scholars focus on workload prediction and use different methodologies. Generally, there are two common methods of prediction: history-based and homeostatic. History-based models are easy and popular. Models examine past workloads to forecast future needs this technique involves prior workload patterns, whereas the homeostatic model uses the mean. The homeostatic prediction, estimate the future workload by adding or removing the present workload from the mean of prior workloads here values can be static or dynamic.

Previous studies in different areas such as global solar radiation [108], the stock market [109], the sports industry [110], and rainfall forecasting [111] state that prediction with time series data is not an easy task. In time series group data points are studied by consecutive points at constant intervals of time. Classical methods used in time series forecasting [88] include Auto Regression (AR) model, Exponential Smoothing (ES) model, Autoregressive Integrated Moving Average (ARIMA) model, Moving Average (MA) model, etc.

ML approach helps organizations to generate precise predictions of a query based on past data to the desired solutions. It can be about anything from a user point of view to suspected dishonest conduct. These give insights to the business that result in genuine financial revenues. For instance, if a model predicts client is likely to switch business, user personal chat can be targeted to avoid losing that customer. The prediction-based selection makes it possible to prevent the upcoming inactive virtual machine (VMs) foremost to decrease the inactive request with retransmission from inactive VMs to active ones [67].

The fast advancement of AI has drawn significant attention to DL techniques [91]. Latest years have seen a grow-up attention in short-term time-series prediction using deep learning [112].When analyzing complicated nonlinear patterns in data, deep neural networks have an advantage over traditional ML structures because DL [113] can examine hierarchical and distributed characteristics. In the paper [79], the authors designed BG-LSTM an integrated method for time series data using a DL approach. The proposed method helps to increase the accuracy of DC predictions and examine patterns in the workload and resource consumption data.

Cloud computing has scalable and flexible sharing of resource services through resource management. In a cloud environment the foundation is laid on a resource that depends on monitoring and prediction to obtain resource automation and manage high performance. The issues associated in [114] is that with monitoring and prediction of resources are addressed in the cloud computing model, execution and design of flexible resource monitoring schemes for cloud computing, and current resource prediction technique founded on VAR through the relationship among different resources.

Table 3 displays the available and newly generated prediction algorithms selected from the recent year’s studies. It also gives idea related to its respective parameters.

Table 3 Comparison between prediction models and its parameters
Full size table

To accurately predict the workload in the CDC the regression technique is considered the most favorable and significant method. Time feature is associated with data therefore time series forecasting methods are applicable for this type of workload. The previous studies show that various prediction techniques such as statistical methods (i.e. AR, MA, ARIMA, ES, Holt’s method, etc.), ML methods (i.e. LR, SVM, NN, etc.), DL methods (i.e. LSTM, DBN, etc.), ensemble techniques (i.e. stacking, voting, boosting, etc.) nature inspired algorithms (i.e. ANN, DE, etc.) are used to solve various cloud workload issues. Many authors tried to combine two or more techniques to resolve the issues.

The above table shows that the parameters such as accuracy and error metrics have been the maximum pick-up ratio. And error metrics include MAE, MSE, MAPE, and RMSE measure values. The researcher can be considering these parameters for their future work.

4.3 Scheduling techniques and its optimize parameters (RQ 3)

In the cloud, each user task uses many virtualized resources and scheduling plays an important role to manage computing resources [121]. The task scheduling method developed in the study [23] aims to minimize service-level agreement (SLA) violations, overall execution time, and costs while distributing m number of tasks of a particular application among a collection of diverse VMs. As a result, an effective scheduling algorithm is required that can balance competing priorities.

In a cloud paradigm, tasks are sorted into two types: they are computing intensity and data intensity. Though the task scheduling requires computing intensity, the data is migrated to the scheduler having high output resources; hence it minimizes the implementation time of the tasks. Alternatively, task scheduling requires data intensity, which helps to decrease the number of data migrations which results to reduce data transfer time [29].

Currently, cloud service numbers are increasing, which in exchange increases the load on cloud nodes for processing. Hence, needs an efficient method to schedule tasks and resources are managed in the cloud environments [122]. Many researchers have enhanced heuristic techniques to achieve better performance in scheduling and others are working on the meta-heuristic algorithm as well as hybrid approaches too. Table 4 gives the recent scheduling techniques and their description.

Table 4 Scheduling techniques
Full size table

Scheduling problem is an NP-hard; means it cannot be solved in a polynomial amount of time. It require finding best scheduling methods that relay on various factors such as work characteristics, different goals, and multiple machine conditions. Therefore, optimization techniques are used to provide efficient scheduling. From the Table 4 we have conclude that the scheduling algorithms categorize in such a way i.e. classical optimization algorithms (i.e. FCFS, Round Robin, Min–Max, Min–Min, etc.), nature-inspired optimization algorithms (i.e. GA, PSO, ACO, etc.), Fuzzy theory based optimization algorithms, and many more are used to schedule the task to the VM.

In scheduling, evaluation parameters refer to the criteria or metrics used to assess the performance or efficacy of a timetable. These metrics are used to analyze how effectively a schedule meets its goals and objectives, as well as to suggest opportunities for improvement. Metrics are significant because it allow schedulers to evaluate a current schedule performance and make accurate decisions about schedule modifications. Schedulers may increase schedule efficiency, accuracy, flexibility, and overall quality by identifying areas where a schedule may be running low. Table 5 represents the list of effective parameters used in scheduling techniques.

Table 5 Scheduling techniques effective parameters
Full size table

From Table 5 we can see that cost, energy consumption, times are the most useful parameters in the previous study. Researchers may choose other metric values for their future work.

4.4 Prediction based scheduling techniques (RQ 4)

In this section several articles are presented in the domain of workload prediction and resource scheduling. Also it discusses the research gaps, importance and need of the current work.

Prediction based scheduling is an important technique to optimize the allocation of resources based on predictions about future demand [40]. This technique involves predicting future workloads and scheduling resources accordingly to ensure efficient and effective utilization. It is important because it can help improve system performance and reduce costs. The CSP can optimize resource utilization, reduce idle time, and avoid over provisioning resources, which can result in cost savings for the provider and its customers. In addition, prediction-based scheduling can help improve user experience by ensuring that resources are available when needed.

For instance, if a website is expecting a sudden surge in traffic, prediction-based scheduling can allocate more resources to the website to ensure that it remains responsive and does not crash due to overload. Overall, prediction-based scheduling is important because it can help improve system performance, reduce costs, and improve user experience by efficiently allocating resources based on predictions about future demand.

The research gaps are discusses here that are found from the previous studies. It helps to design the relevant questions for doing this survey. According to the studies [25, 57, 60, 86] there is a need for more accurate prediction models that can effectively predict the workload of cloud applications. Existing models may not account for certain variables that can affect the performance of the system. Many prediction-based scheduling techniques rely on heuristics [17, 18, 34, 90] and rule-based approaches [42, 133,134,135]. There is a need for more sophisticated techniques, such as DL, hybrid and ensemble to improve the accuracy of predictions. Cloud applications are often composed of multiple heterogeneous workloads that have different resource requirements [46, 61, 63, 98, 136]. There is a need for prediction-based scheduling techniques that can effectively manage and schedule heterogeneous workloads. Prediction-based scheduling techniques must be scalable to handle large numbers of requests and data [137]. There is a need for techniques that can handle large-scale prediction and scheduling tasks in real-time.

Table 6 shows that comparison between previous studies related to the prediction and scheduling approaches which contain eight cells fine the information which includes: the references, problems in workload, what are requirements for the problem, the proposed technique, the solution given by the proposed approach, the related environment in which the experiment performed, experiment setup and future works.

Table 6 Existing prediction based scheduling approaches
Full size table

5 Research challenges and future directions

The research covered the challenges in CW solved by prediction models and scheduling techniques. The survey [39, 41, 47, 49], stated that accurately predicting workload demands is vital for effective scheduling. But the prediction of resource allocation in the CW is difficult due to unpredictable demand of resources. Also, this changing workload patterns in real time makes difficult to scheduler to schedule tasks. In this survey, the possible solutions related to these challenges have been discussed. Apart from these, there are another challenges are required to outlook. Hence, future directions addressed that challenges driven by the research communities can be further investigated are given below:

  • It is necessary to enhance data quality before utilizing ML and DL methods. Irrelevant features lower model performance. Preprocessing technique will help the researcher find the right domain by removing unnecessary features.

  • Compliance requirements can vary based on industry, geography, and other factors, making it challenging to maintain compliance in the cloud. CSPs must ensure that their services comply with regulations and standards to avoid penalties and legal issues.

  • Moving workloads from one cloud provider to another can be challenging due to the differences in technology and infrastructure. This can lead to vendor lock-in, where businesses find it challenging to switch providers, limiting their options and potentially increasing costs.

  • High computation capabilities are required due to the large size of datasets in the cloud environment.

  • Researchers could explore new prediction techniques that leverage ML and AI to analyze large datasets and identify patterns in workload demands. This could involve the use of DL methods to automatically extract features and learn complex relationships between different variables, or the use of ensemble methods that combine multiple prediction models to improve accuracy.

  • Increased focus on the use of prediction-based scheduling for edge computing, which involves processing data closer to where it is generated, rather than in a centralized data center.

  • By using prediction algorithms to anticipate workload demands at the edge, cloud providers can optimize the allocation of resources and ensure that workloads are processed efficiently and cost-effectively.

  • CWs store and process sensitive data, making them a prime target for cyber attacks. Ensuring data security and privacy in the cloud is a critical issue that needs to be addressed in future.

6 Conclusion

Currently, traditional businesses are rapidly migrating to cloud environments and the increasing demand for multiple resources to perform a single task is the main challenges for cloud service providers. This paper deals with different CWs and their applications where virtual resources are provided with specific features. To maximize the management of diverse resources, efficient scheduling is needed. From the literature, it is also observed that prediction plays an important role to improve scheduling performance. Therefore, an efficient prediction-based scheduling framework is needed to address the challenges in CWs. The present study examined various research articles with their issues, research gaps, and research challenges. Also it gives solutions of related RQ and future directions for the upcoming problem. The research paper is reviewed from the year 2015 to the current and presented in tabular forms with their comparative analysis. The information provided by each table helps the researchers to reduce the time to resolve the issues related to the various scenarios. Here, the data is classified into various categories including workload datasets, resources used in the cloud, prediction models, scheduling algorithms used in previous studies, and their respective evaluation parameters. Finally, the survey paper concludes that designing a framework for prediction-based scheduling with a hybrid deep learning method is the best suite for workload prediction, and integrating with this for scheduling a nature-inspired optimization algorithm can perform excellent work.