Resource allocation strategies and task scheduling algorithms for cloud computing: A systematic literature review

5.1.1 Review of mathematical approach

This section reviews articles on resource allocation and task scheduling algorithms that are based on mathematical approach; Table 2 lists the algorithms used in cloud computing environments.

Zhu et al. [51] suggested a dynamic pricing system based on a model for the Stackelberg game to tackle the challenge of increasing the income for cloud computing providers of Infrastructure as a Service (IaaS) and Software as a Service (SaaS). The simulation results show that the suggested mechanism performs better than auction-based pricing and fixed pricing systems in terms of revenue maximization and resource usage.

A multi-resource fair scheduling (MRFS) algorithm for heterogeneous cloud computing environments was presented by Hamzeh et al. [52]. The study aims to maximize each user’s utility on a specific server by reducing the number of dominant resources. MRFS was introduced as a solution to the resource scheduling problem in cloud computing. It considers the fair and efficient allocation of resources to users. However, the study did not specify the specific data used. It focuses on introducing the MRFS algorithm and assessing its effectiveness.

Researchers presented a method for allocating resources to meet growing needs in cloud computing [56]. The study introduced a multi-objective optimization technique aimed at aligning resource performance with percentage distances and minimizing the number of physical servers utilized. The RAA-PI-NSGAII method not only maximizes resource utilization and reduces solving time but also improves the quality and uniformity of the solution set distribution. However, the study does not adequately address the specific challenges or limitations encountered during the implementation of the proposed resource allocation algorithm, particularly in the context of increasing cloud computing demands.

A mathematical model to enhance the performance and efficiency of cloud computing by scheduling tasks within containers was introduced by Swatthong and Aswakul [57]. The model employs ILP to increase the average compensation of tasks while adhering to resource and demand constraints. This study evaluated the model using two cloud scenarios: An edge-core cluster and a peer-to-peer federated cloud. Results demonstrated that the proposed model outperformed the standard RR scheduling method in terms of task completion time and reward level. Additionally, the study underscores the importance of flexible and fine-grained task separation in cloud architecture. However, it does not account for the time-series nature of the requests, which could impact the model’s ability to handle maximum load efficiently.

Yadav and Mishra [58] suggested an enhanced ordinal optimization method paired with linear regression to enhance scheduling of task in cloud computing to reduce the makespan. This method narrows down the search space for scheduling and efficiently assigns the workload to the best schedule. Additionally, it forecasts future dynamic workload to achieve a target minimum makespan. However, there is no comparison with existing approaches to validate the effectiveness of the proposed technique.

Tai et al. [59] proposed an enhanced algorithm for managing computing resources and energy consumption in heterogeneous cloud computing centers. The algorithm considered factors such as energy usage, task scheduling, execution time, employing Lagrangian relaxation and mathematical programming, which develop high-performance and energy-efficient cloud computing facilities. However, the study does not address the theoretical trade-offs between energy efficiency and other performance metrics, such as task execution time or resource utilization.

Al-Asaly et al. [60] presented a self-sufficient, intelligent workload prediction technique employing a diffusion convolutional recurrent neural network (DCRNN) model for cloud resource allocation. The objective is to enhance prediction precision and reduce the gap between forecasted and real workloads in cloud computing setups. Tested on real PlanetLab CPU usage data, the model surpassed other deep learning models, achieving a root-mean-square error of 2.40% and an average absolute percentage error of 0.18%. However, the study focuses specifically on CPU utilization as the input data for prediction model; it ignores other metrics or factors such as memory consumption or network traffic.

Ghobaei-Arani and Souri [61] introduced a linear programming method called LP-WSC for web service composition in geographically distributed cloud environments, intending to improve QoS parameters. This approach selects the most efficient service for each request and significantly reduces the cost of choosing and configuring services, while increasing the availability of services and reliability of servers compared to other methods. However, the suggested LP-WSC strategy is not compared to other cutting-edge methods for web service orchestration in geographically dispersed cloud environments in this study.

In order to provide variable job assignments on VMs, Rawat et al. [62] proposed a cost-effective model of Big-Bang Big-Crunch (BB-BC) for resource allocation in cloud setups. It targets a globally optimal outcome through an objective function that considers metrics such as cost and time, surpassing traditional static, dynamic, and bio-inspired provisioning methods. However, the authors did not discuss the limitations and the data source was not mentioned in the study.

Shi et al. [63] proposed a uniform model description of manufacturing resources using ontology and metadata modeling methods. They also outline a strategy for collaborative scheduling of manufacturing resources between enterprises using Mixed Set Programming, which enhances resource collaboration and allocation efficiency. In subsequent work, the authors suggest a model and scheduling mechanism that can be integrated with the study of inter-enterprise logistics to improve the efficiency of inter-enterprise resource collaboration further. However, the study does not compare the proposed model and scheduling strategy with the existing approaches for conducting large-scale experiments.

5.2.1 MH approach

MH are strategies that provide efficient and optimized solutions, helping to solve problems of complex optimization in various fields by providing acceptable answers in a reasonable amount of time. MH search techniques are general, sophisticated approaches that can be used to build basic heuristics to solve specific optimization problems [71]. Since scheduling is an NP-hard problem, most scheduling algorithms do not explore the entire problem space to find the optimal resource allocation; hence, MHs are the best choice for this problem [72].

1. Single-based optimization algorithms: Generate a single random solution and strive to enhance it through optimization processes [73].

2. Population-based optimization algorithms (POAs): POAs rely heavily on factors such as the choice of algorithms, strategies, parameter combinations, constraint handling methods, local search methods, surrogate models, and niching methods to determine solution quality for optimization problems, etc. [73].

3. Hybrid MH: A hybrid MH primarily distinguishes according to four criteria [74]:

• Hybridization based on algorithm: Hybridized algorithms include combinations such as MHs with MHs, MHs with Heuristics, MHs with Fuzzy logic techniques, or MHs with statistical techniques.

• Hybridization based on hybridization level: Combine algorithms based on their level of coupling strength.

• Hybridization based on the order of execution: The individual algorithms are executed either sequentially, intertwined, or even in parallel.

• Hybridization based on their control strategy: Hybrids are classified based on their control strategy, which can be either integrative (coercive) or collaborative (cooperative).

5.2.2 Review of heuristic and MH algorithms

This subsection reviews articles on resource allocation and task scheduling algorithms based on heuristic and MH algorithms, including single-based, population-based, and hybrid algorithms. Table 3 lists the algorithms used in cloud computing environments.

Chhabra et al. [29] addressed the optimization of bag-of-tasks scheduling in cloud data centers by developing the opposition learning enabled whale particle swarm optimization (OWPSO) algorithm, which combines the whale optimization algorithm (WOA) with PSO to improve scheduling efficiency. The study evaluates key performance metrics, including makespan and energy consumption, demonstrating that OWPSO significantly outperforms baseline algorithms in producing near-optimal scheduling solutions. However, the reliance on synthetic datasets for benchmarking and the need for thorough parameter tuning to achieve optimal performance limit the practical applicability of the suggested technique.

Mirmohseni et al. [39] presented the fuzzy PSO (FPSO)-GA approach, which combines GA and fuzzy PSO techniques to achieve effective load balancing in cloud networks. The proposed approach demonstrates significant improvements in load balancing and reducing energy, surpassing algorithms such as LBPSGORA, PSO, and GA in load balancing performance. While the authors highlight the effectiveness of the FPSO-GA algorithm in enhancing load distribution, the study lacks detailed numerical results or specific performance metrics to substantiate the claimed improvements.

Manavi et al. [68] proposed a novel hybrid approach for resource allocation and task scheduling in cloud computing, combining neural network classification with GA optimization. The method aims to improve execution time, cost, response time, and system utilization while considering fairness to prevent task starvation. Using a large Google dataset for simulation, the authors demonstrate improvements of 3.2% in execution time, 13.3% in cost, and 12.1% in response time compared to state-of-the-art methods. However, the study is limited to independent tasks without deadlines and uses a relatively small chromosome size in the GA, which may impact scalability for larger task sets.

Zhu et al. [70] proposed a heuristic multi-objective task scheduling framework for container-based clouds, exploiting actor-critic RIL to enhance task scheduling efficiency. The framework addresses the complexities of resource allocation and task management in dynamic environments, showing significant improvements in metrics such as resource utilization and execution time through extensive simulations. However, the study’s applicability is constrained by its focus on specific cloud environments, potentially limiting the generalizability of the findings. Additionally, the framework may face challenges with real-time adaptability in highly variable workloads.

Ibrahim et al. [75] presented a comparative evaluation of state-of-the-art static task scheduling algorithms in cloud computing, focusing on their performance in terms of resource utilization, makespan, throughput, and response time. The study compares methods such as PSSLB, MCT, min-min, max-avg, LBIMM, RASA, Sufferage, and TASA using CloudSim simulations with HCSP and GOCJ datasets. The results show that while MCT and Sufferage perform well in reaction time but poorly in resource utilization, TASA and PSSLB perform better across various metrics and datasets. The study emphasizes how crucial it is for cloud service providers and end users to balance resource usage and execution time. Nevertheless, the study is constrained by its emphasis on static scheduling techniques and dependence on virtualized settings instead of actual cloud configurations.

Hamid et al. [76] compared different task scheduling algorithms of cloud computing based on makespan using workflows as datasets, with makespan serving as the primary performance metric. Experimental results indicate that the FCFS algorithm outperforms RR, min-min, and max-min in the CyberShake workflow, while max-min outperforms FCFS, RR, and min-min in the Montage workflow. However, the study does not explore how makespan is affected as the number of data centers or hosts increases, leaving a critical aspect of scalability unaddressed.

Nayak et al. [77] intended to minimize the task rejection ratio and maximize the task acceptance ratio in a cloud-based scheduling mechanism for activities with deadlines. It improves upon the existing backfilling algorithm by addressing conflicts among similar leases and allowing the scheduling of new deadline-sensitive tasks during execution. An average lease acceptance ratio of 91.94% and a minimal average lease rejection ratio of 8.05% are attained by the suggested mechanism. It considers the current time and gap time as scheduling parameters, that are not addressed in the current studies. It allows the arrival of a new lease to be planned and does not require a decision maker such as analytic hierarchy process to resolve conflicts between similar leases. However, the performance analysis is based on limited workloads and parameters, which may not fully represent the diverse requirements of different cloud applications.

Mangalampalli et al. [78] proposed a deep RIL-based task scheduling algorithm in cloud computing, designed to enhance significant QoS parameters such as SLA violation, energy consumption, and time. The algorithm dynamically calculates task execution time based on a threshold value and the load on physical hosts, with host utilization determined using specific equations. While the approach effectively addresses key parameters such as makespan, SLA violations, and energy consumption, the study lacks a comprehensive discussion of the algorithm’s limitations. Including such insights could have highlighted potential drawbacks and opportunities for future refinement.

Mishra and Gupta [79] presented a study that compares heuristic algorithms for scheduling tasks in cloud computing, such as RALBA, DRALBA, DLBA, max-min, min-min, and RR, based on performance parameters such as throughput, makespan, and average resource utilization ratio (ARUR). The existing DRALBA approach outperforms other approaches in terms of performance parameters, both on realistic workloads and synthetic, making it an efficient and effective scheduling algorithm for cloud computing environments. However, there is no discussion regarding the suggested algorithms’ scalability and performance with larger workloads or in real-world cloud computing environments.

Yao et al. [80] proposed a task duplication-based scheduling algorithm (TDSA) to enhance the makespan for budget-limited workflows, utilizing idle slots on resources and reallocating the leftover budget. The TDSA algorithm shows notable enhancements in the makespan of workflows (up to 17.4%) and the utilization of cloud computing resources (up to 31.6%) compared to the four baseline algorithms, as evidenced by experiments on randomly generated and actual workflows. However, the study does not consider the running time of workflow tasks, and the amount of data transferred among workflow tasks is highly uncertain.

Ben Alla et al. [12] proposed a novel approach to address the problem of user requests and supplier resources not being prioritized. They proposed an efficient priority task scheduling scheme called MCPTS, in which four task parameters (length, waiting time, deadline, and burst time) are used to determine priority. The task priority, task queuing priority, and resource priority sub models make up the MCPTS scheme. To determine the priority of activities, they suggest using differential evolution (DE), a MH algorithm, and elimination and choice expressing reality version III, a new hybrid multi-criteria decision-making method. They also presented a queueing model-based dynamic priority-queue algorithm. Moreover, they created a productive and adaptable relationship between resource and task classes by dynamically adjusting resource priority depending on the task’s priority model. However, insufficient consideration of dynamic request characteristics and resource availability could affect scheduling choices.

To schedule tasks in cloud computing systems, Alhaidari and Balharith [81] proposed a novel method known as the dynamic RR heuristic algorithm (DRRHA). This method performs noticeably better than previous algorithms evaluated in terms of average waiting time, turnaround time, and response time. DRRHA employs the RR algorithm to increase task scheduling efficiency, modifying its time quantum based on the task’s remaining burst time and time quantum means. However, there has been no comparison between the proposed algorithm and other state-of-the-art task scheduling algorithms regarding resource usage or energy efficiency.

Mustapha and Gupta [82] introduced a fault-aware task scheduling algorithm in cloud computing using min-min and DBSCAN to enhance resource allocation efficiency fault tolerance, and enhance QoS in a dynamic and heterogeneous environment. The approach outperforms existing algorithms such as ant colony optimization (ACO), PSO, BB-BC, and whale harmony optimization (WHO) in terms of execution time, task completion rates, and defect reduction. However, the proposed technique uses DBSCAN to cluster resources, complicating the job scheduling procedure. Moreover, comparing the suggested algorithm with fewer existing algorithms may not provide as clear a picture of the algorithm’s performance as a more extensive comparison with a wider range of state-of-the-art methods.

Khan [83] released a power-efficient cloudlet scheduling (PACS) approach to reduce energy consumption, request processing time, and cloud computing setup costs. When compared to other popular cloudlet scheduling methods, PACS provides a noteworthy 3.80–23.82 speed boost. However, the study did not evaluate how well the suggested PACS technique scales in handling higher numbers of cloudlets and VMs.

Alsaidy et al. [84] reported that efficient task scheduling in cloud computing was vital for cost-effective execution and resource utilization, as addressed by the LJFP-PSO and MCT-PSO algorithms. These algorithms employ heuristic initialization to boost PSO performance, surpassing traditional PSO and comparative techniques in reducing makespan, execution time, imbalance, and energy consumption metrics. However, the study does not provide details about the specific dataset or real-world scenarios for evaluating the proposed algorithms.

Prathiba and Sankar [85] presented the multi-task wolf optimizer (MTWO) method, merging efficient task scheduling and secure resource allocation in cloud computing to tackle data and resource management challenges in a rapidly expanding cloud setting. This technique showcases a brief task schedule employing wolf optimization techniques to minimize makespan time and boost throughput. It further incorporates a deep neural network with cluster optimization for resource allocation efficiency within architectural limits, enhancing response time, power efficiency, and resource utilization. However, the study does not explicitly mention the specific data used in the simulation setup and analysis.

Yuvaraj et al. [86] introduced a machine-learning technique to optimize job allocation between the contributor and the event queue in the serverless framework. To increase the effectiveness of work distribution, they applied the gray wolf optimization (GWO) model. Moreover, the authors optimized GWO settings and improved work allocation using an RIL technique. According to simulation experiments, the suggested GWO-RIL technique significantly reduces runtimes and adapts to shifting load situations. However, the study assumes prior knowledge of the computational time period for each task and similar overheads prior to task scheduling, which may not always be realistic in real-world scenarios.

Nanjappan et al. [87] proposed a technique that integrates adaptive neuro fuzzy inference system (ANFIS) and black widow optimization (BWO) to enhance resource utilization and scheduling in a cloud computing. The ANFIS-BWO method determined which VM was best suited for each task. The BWO technique identifies the optimal solution for the ANFIS scheme. The proposed approach aims to minimize computation time, cost, and energy consumption while optimizing resource utilization. Nevertheless, it was not made clear which particular datasets were used in this investigation.

Tamilarasu and Singaravel [88] proposed the Coati Optimization Algorithm-Based Task Scheduling (ICOATS) to tackle challenges related to the scheduling of task in cloud environment. ICOATS aims to tackle issues such as long scheduling times, increased costs, and optimized VM workloads. A multi-objective fitness function is seamlessly integrated. This function aims to simultaneously reduce the makespan while enhancing the utilization of available resources. An exploitation strategy is incorporated to achieve this, which helps avoid premature convergence and improves the local search potential. ICOATS aims to find efficient solutions for cloud task scheduling by striking a harmony between exploitation and exploration. However, the study’s effectiveness in achieving optimal solutions may be limited due to the lack of consideration for a comprehensive set of QoS characteristics during the optimization process.

An improved WOA algorithm for cloud task scheduling (improved whale optimization [IWC]), using the WOA was described by the researchers [89] to improve task scheduling in cloud computing systems. Optimizing task scheduling plans and resource utilization to increase cloud system performance is one of the most important functions of IWC technology. The proposed IWC algorithm shows superior convergence speed and accuracy when compared to existing MH algorithms. It is versatile and can be applied to both small- and large-scale problems. However, the authors did not investigate parallel applications in cloud environments to reduce the scheduling overhead of this method when dealing with large workloads.

The researchers described an improved version of the ant colony optimization algorithm (ACOA) called Q-ACOAn [90] designed to fulfill predetermined time and cost objectives. This aims to enhance the algorithm for allocating resources and scheduling tasks in cloud computing. Results demonstrate that Q-ACOA outperforms alternative scheduling algorithms in job completion time, data migration time, and overall cost efficiency. Nonetheless, the study lacked explicit information regarding the dataset used in the research.

Alghamdi [91] introduced a novel resource allocation method for cloud computing environments using binary PSO (BPSO) and artificial neural networks (ANN). The proposed method aims to improve particle placements and thus reduce work completion times across VMs. Reducing response times, improving resource utilization, and ensuring good QoS for cloud computing applications are the project’s primary goals. Researchers worked to improve load and scheduling of the task in cloud environments using the proposed ANN-BPSO approach.

Shrichandran et al. [92] introduced a hybrid competitive swarm optimization algorithm for task scheduling (HCSOA-TS) in cloud environments. The method incorporates a Cauchy mutation operator into the Competitive Swarm Optimization algorithm to improve performance. The researchers assessed the HCSOA-TS technique across four situations with different numbers of tasks and edges, comparing it to GA, PSO, and GA-PSO algorithms. Performance metrics included makespan, execution cost, and load balance. Results showed that HCSOA-TS outperformed the other algorithms across all scenarios. However, the study is limited by its simulation-based nature, relatively small-scale scenarios (maximum 1,000 tasks), and lack of consideration for real-world cloud workloads.

The bird swarm optimization load balancing (BSO-LB) algorithm was the load balancing method suggested by Mishra and Majhi [93]. The algorithm views VMs as destination food patches and tasks as birds. The suggested approach seeks to reduce response time and optimize load balancing in cloud system. The load balancing technique is paired with the binary variation of the BSO algorithm. Experimental findings demonstrate that the suggested approach surpasses alternative algorithms. However, it is not as fast as other algorithms.

Paulraj et al. [94] covered the significance of task scheduling in cloud computing and suggested an effective hybrid job scheduling optimization method called efficient hybrid job scheduling optimization (EHJSO), which combines Cuckoo Search Optimization and GWO. Metrics such as makespan, computation time, fitness, iteration-based performance, and success rate are used to compare the suggested method to prior research and are determined to be superior. However, the authors do not specify the limitations of the proposed algorithm.

Hu et al. [95] presented well-organized scheduling algorithm based on energy for processing real-time applications in cloud computing, aiming to abate energy consumption while meeting real-time requirements. The method demonstrates an important decrease in energy consumption compared to existing algorithms, with a higher success rate in finding feasible schedules and comparable computation time. However, the study lacks specific details about the nature or source of the real-case benchmarks or the data used in the synthetic application.

Xu et al. [96] presented a reference vector-guided evolutionary algorithm (RVEA-NDAPD) and a many-objective scheduling strategy for cloud environments in order to solve the model. The performance of the suggested model in a cloud computing environment was effectively improved by the RVEA-NDAPD algorithm compared to the traditional many-objective evolutionary algorithms (MaOEAs) currently in use. However, the study lacks detailed information on the specific data used in the study, including task characteristics, system and user details, as well as performance parameters of VMs and tasks.

Bandaranayake et al. [97] proposed a TRETA method for cloud task scheduling to optimize the total execution time of computing resources. The study demonstrates that the suggested method enhances performance and resource usage for cloud jobs by comparing it with alternative heuristics on real-world workloads. The authors did not evaluate the algorithm when combined with other tactics or in relation to other measures.

Alsubai et al. [98] proposed an innovative swarm-based task scheduling approach that integrates the moth swarm algorithm and the chameleon swarm algorithm to optimize cloud scheduling in terms of efficiency and security. The objective is to enhance the tasks distribution using available resources and encode cloud information during task scheduling, with a focus on optimizing the available bandwidth for efficient scheduling of cloud computing tasks. Improvements in measures such as imbalance score, throughput, cost, average waiting time, reaction time, throughput, latency, execution time, speed, and bandwidth utilization are demonstrated by evaluating the performance of the approach. However, the specific details of the proposed algorithm and its technical implementation are not provided in the sources.

Kaur and Kaur [99] suggested a hybrid approach to improve load balancing in cloud systems by combining heuristic and MH methods. Within the HDD-PLB framework, two hybrid strategies were proposed and examined: hybrid heterogeneous early finishing time (HEFT) with ACO (hyper-heuristic algorithm [HHA]) and hybrid prediction early finishing time with ACO method (HPA). The authors assessed the efficiency of the suggested architecture using two key performance metrics: manufacturing scale and cost. Although the authors claim that the framework operates optimally in terms of cost and scale, no experimental evaluation or findings were offered to support their assertions.

Moazeni et al. [100] presented a new approach for cloud computing resource allocation, relying on the multi-objective-learning-based-optimization (AMO-TLBO). The AMO-TLBO algorithm includes features including the number of teachers, teaching factor adaptive, and self-motivated instructional learning to improve the skills of exploration and exploitation. Increasing utilization while reducing space and cost is one of the most important features of the proposed AMO-TLBO algorithm. Evaluation results show that the proposed method outperforms the TLBO, MOPSO, and NSGA-II algorithms in several performance metrics. However, the source of the dataset is not specified.

Zhao et al. [101] proposed an innovative cloud resource allocation strategy that prioritizes requests into two categories: high-priority primary requests (PRs) and low-priority secondary requests (SRs). This approach implements an entry threshold and probability mechanism to regulate SR admission, thereby enhancing service quality for PRs. The team developed a discrete-time queuing model to evaluate the strategy’s effectiveness and calculated various performance metrics. Experimental results were promising, showing a significant reduction in PR blocking rate by 47% and an impressive increase in PR throughput rate of up to 55% when compared to traditional strategies without entry control. However, the study has a notable limitation: it lacks comparative analysis with other contemporary resource allocation strategies. This omission makes it difficult to fully assess the relative performance and effectiveness of the proposed approach within the broader context of cloud resource management solutions.

Aktan and Bulut [102] discussed the advancement of MH and hybrid MH algorithms for task scheduling in cloud computing set to enhance completion time and load balancing of VMs. They evaluate various algorithms such as GA, DE, simulated annealing (SA), and their hybrid versions based on completion time and load balancing. The results indicate that the hybrid DE-SA algorithm outperforms the single DE and SA algorithms in completion time. Moreover, it improves the average completion time and standard deviation for specific task sets of VM loads. However, the study does not compare the proposed algorithms with existing state-of-the-art methods.

Albert and Nanjappan [103] presented a cloud task scheduling strategy using a hybrid WHO algorithm to balance the system load, minimize makespan, and reduce costs. Experimental results show that the algorithm achieves high load balance with decreased execution time, cost, and energy consumption, proving its effectiveness in cloud resource management. The proposed WHOA algorithm outperforms other algorithms such as OGWO, WOA, and HS in terms of cost efficiency and makespan optimization. However, there is no discussion of the challenges that may arise when implementing the proposed algorithm in a real-world cloud computing environment.

Nekooei-Joghdani and Safi-Esfahani [104] presented a new hybrid MH algorithm named Gabor opposition-based learning multi-verse optimizer for a dynamic cloud environment to solve the scheduling problem, demonstrating superior performance compared to baseline algorithms. GOMVO combines Gabor filter, multi-verse optimizer, opposition-based learning, and multi-tracker optimization algorithms to improve task allocation in cloud environments. The multi-verse optimizer-Gabor opposition-based learning multi-verse optimizer (MTOA-GOMVO) hybrid algorithm further enhances the resolution of premature convergence issues by leveraging the strengths of both the MTOA and GOMVO algorithms. While the proposed MTOA-GOMVO hybrid algorithm shows promising results in improving task scheduling in cloud computing environments, its scalability to larger and more complex scenarios could be a concern.

Pirozmand et al. [105] explored the application of a multi-objective hybrid GA for cloud computing task scheduling, focusing on optimizing both makespan and energy usage. They introduced the Genetic Algorithm and Energy-Conscious Scheduling Heuristic (GAECS) algorithm to tackle job scheduling in cloud environments. Combining GA and ECSH aims to reduce makespan, improve resource allocation efficiency, and enhance overall system performance. However, achieving optimal results requires careful tuning of the GA and ECSH algorithm parameters, which can be challenging for users unfamiliar with optimization techniques.

5.3.1 Review of hyper-heuristic approach

This section reviews articles on resource allocation and task scheduling algorithms that are based on hyper-heuristic approach; Table 4 lists the algorithms used in cloud computing environments.

Yin et al. [110] proposed a hyper-heuristic reinforcement learning (HHRL)-based approach to enhance task sequence completion time for complicated and dynamic cloud service scheduling tasks. This method integrates population diversity and maximum time to determine job scheduling and low-level heuristic strategies. It also introduces a linear regression-based approach to estimate task complexity, tracking the completion of 100 tasks in each category and examining their linear relationships. Nonetheless, the study does not address the potential limitations or constraints in the reward table setting phase of HHRL, particularly in the selection of low-level heuristic strategies and the integration of maximum time and population diversity.

Gupta et al. [111] proposed an algorithm to achieve balanced load distribution across VMs, aiming to reduce the time of makespan. The algorithm offers equitable scheduling frameworks by employing a honey bee load balancing and improvement detection operator to determine the appropriate low-level heuristic for enhanced candidate solutions. Experimental results demonstrate its efficiency compared to existing heuristic-based scheduling procedures. However, further analysis and evaluation are needed to identify potential limitations or areas for enhancement. Additionally, the comparison of the proposed algorithm was limited to specific metrics and did not address the full spectrum of cloud task scheduling optimization metrics.

A previous research [112] presented a new two-level container allocation issue in cloud computing and suggested a hybrid method utilizing genetic programming hyper-heuristics and human-designed rules to reduce energy usage notably. The proposed method surpasses current rules by considering various characteristics to choose and generate VMs, resulting in more effective resource allocation in container-based clouds, thereby reducing energy consumption. However, this study focuses on reducing power consumption as the primary metric while ignoring other important performance indicators in container-based cloud environments.

Kruekaew and Kimpan [8] developed a multi-objective approach known as MOABCQ, which uses the artificial bee colony algorithm, the Q-learning algorithm, and RIL technique, to enhance scheduling and resource utilization, maximize VM throughput, and create load balancing between VMs based on makespan, cost, and resource utilization. The MOABCQ approach demonstrated strong search capabilities and quick convergence. However, the MOABCQ method may not always be optimal and its performance may vary across different datasets.

Freire et al. [113] proposed a queue-priority algorithm that utilizes a new heuristic and PSO to optimize task scheduling in integration platforms, particularly for managing large data volumes in integration processes. The algorithm enhances the scalability of cloud computational resources and reduces business operational costs. The experimental results confirmed the efficacy of the proposed algorithm for improving the execution of integration processes with high data volumes. However, the study does not offer an exhaustive evaluation of the suggested algorithm’s performance about alternative methods in the sector, nor does it compare it with other task scheduling methods that are currently in use.

Tong et al. [114] proposed a novel intelligent HHA called HSQ-StudGA, which combines machine learning technology, specifically Q-learning, with GAs to solve optimization problems. It aims to enhance the performance and quality of arithmetic solutions for such problems. The HHA improves the search capability of the original algorithm solution and effectively enhances the overall quality of the solution. However, the study does not address the potential challenges or limitations of implementing the Q-learning method in the algorithm.

The mean completion time (MCT) was found to be minimized more effectively by Bouazza et al. [115] when they proposed a Product-Driven Control System (PDCS) based on smart products for the dynamic scheduling of production systems. The approach integrates a hyper-heuristic with a generic decisional strategy model, enabling efficient switching between scheduling rules and enhancing global performance and system reactivity. The proposed approach shows superior performance in minimizing mean completion time. However, the rapidly changing environment poses challenges in implementing conventional predictive methods. Table 4 shows the hyper-heuristic algorithms used in resource allocating and scheduling.

6.1.1 Hybrid and hyper-heuristic techniques for dynamic environments

Hybrid methods that integrate various algorithms or heuristics are increasingly popular for achieving near-optimal solutions under diverse scenarios. Trends include [114,116]:

• Hyper-heuristics: These techniques dynamically select and tune low-level heuristics to adapt to changing conditions, ensuring efficient resource allocation. Their potential to handle real-time changes in task scheduling offers a promising direction for future cloud applications.

• Hybrid MH: Combinations such as OWPSO and HCSOA-TS demonstrate enhanced performance across various metrics, suggesting that hybridization will remain a cornerstone of future algorithm development.

6.1.2 Multi-objective optimization

Future research will likely prioritize simultaneous achievement of multiple objectives. While traditional metrics such as cost and makespan remain essential, there is growing interest in incorporating the following:

• Energy efficiency: With sustainability becoming a priority, minimizing energy consumption will be critical. Approaches such as the hybrid genetic programming method [112] emphasize energy usage reduction, setting the stage for more energy-conscious scheduling algorithms.

• QoS considerations: Metrics such as latency, reliability, and availability are gaining prominence, especially in real-time applications.

• Load balancing and resource utilization: Ensuring optimal use of resources without overloading VMs or cloud instances will remain a core focus.

6.1.3 Integrating AI and machine learning

AI and ML are expected to play a transformative role in cloud computing:

• RIL approaches: Techniques such as HHRL [110] and MOABCQ [8] show potential for adapting to dynamic cloud environments, though simplifying their implementation complexities will be a future challenge.

• Predictive analytics: Employing linear regression or other predictive models to estimate task difficulty or resource needs can further enhance scheduling accuracy.

6.1.4 Emergence of edge-cloud collaboration

As edge computing expands, resource allocation and task scheduling need to tackle distributed and hierarchical environments.

• Algorithms must evenly distribute computation loads between cloud data centers and edge devices.

• Addressing latency-sensitive applications such as IoT and real-time analytics will drive innovations in hybrid edge-cloud scheduling models.

6.1.5 Focus on green and sustainable cloud computing

As energy consumption becomes a global concern, research will increasingly prioritize sustainable solutions. This includes

• Developing algorithms that optimize resource usage without compromising performance.

• Integrating renewable energy sources into cloud infrastructure and aligning scheduling tasks with energy availability patterns.

6.1.6 Scalability and real-world applicability

Scalability remains a critical challenge for most algorithms, especially in large-scale, heterogeneous cloud environments. Future efforts must focus on

• Scalable algorithms: Solutions capable of maintaining efficiency across increasing system sizes and complexity.

• Benchmarking and real-world testing: Moving beyond controlled simulations to test algorithms in diverse and realistic cloud scenarios.

Future research should prioritize developing hybrid models that combine various scheduling and allocation strategies for a comprehensive approach to these challenges. Fostering collaboration between academia and industry can accelerate the advancement of best practices and innovative solutions. Moreover, utilizing machine learning and predictive analytics can offer insights into usage patterns, enabling proactive resource management that anticipates demand [119].