Scientific workflows offer a compelling way to represent the complex orchestration of interdependent computations and have become widely adopted across various scientific fields [1] . They enable users to define the different steps required to process the large volumes of data typically generated by scientific experiments and to produce original scientific outcomes. The execution of such data-intensive applications, which often involve hundreds of computational tasks on large-scale distributed infrastructures, is generally managed by a Workflow Management System (WMS) [2] . Tasks such as resource selection, data management, and computation scheduling are handled by the WMS, thus shielding the end user from the complexity of these operations.

For a long time, commodity clusters and computing grids were the preferred infrastructures for running scientific workflows. Clusters, typically hosted and managed by the institution owning the workflow, facilitated resource access, while grids enabled scientists to scale their workflows by pooling resources from multiple institutions. However, with the rise of major providers like Amazon [3] , Google [4] , and Microsoft [5] , Infrastructure as a Service (IaaS) clouds have emerged as strong competitors to clusters and grids. IaaS clouds combine the benefits of both by offering easy access to a virtually unlimited pool of resources. By carefully planning the execution of a workflow, a WMS can dynamically build a compute and storage infrastructure tailored to the workflow’s specific needs, utilizing a customized set of virtual machine instances.

The description of a scientific workflow is typically independent of the characteristics of the infrastructure on which it will be executed. This provides users with greater flexibility, allowing them to run the same workflow on different infrastructures via the WMS without needing to modify their application. A direct consequence of this flexibility is that dependencies between computational tasks (where data produced by one task is consumed by another) are generally managed through files. Intermediate data is written to disk, and the file may then be transferred over the network to another storage device, where the consuming task will eventually read it.

In this paper, we propose to take advantage of two key features of a specific family of virtual machine instances provided by Amazon Web Services: a large number of cores and dedicated storage on high-speed SSD drives. By improving data locality, this approach aims to reduce the amount of data transferred over the network during workflow execution, which should have a direct and positive effect on the workflow’s execution time. Additionally, we propose using realistic simulations to design a cloud infrastructure that strikes a good balance between cost and performance. To achieve this, we developed a simulator built on the WRENCH framework [6] . WRENCH enables the construction of Workflow Management System (WMS) simulators that are accurate, fast, scalable on a single machine, and require minimal software development effort.

The main contributions of this article are therefore:

This paper is organized as follows. Section 2 reviews the related work on scheduling of scientific workflows on IaaS clouds. In Section 3, we describe the platform and application models used in this work. Then, in Section 4, we detail the proposed approach for Multi-Objective Scheduling while Section 5 explains results and discussion. Finally, we conclude this paper and present future work directions in Section 6.

The rise of cloud computing has transformed the management of scientific workflows, enabling the processing of large volumes of data with flexible and scalable resources. However, task scheduling in a cloud environment remains a major challenge, particularly due to multi-objective constraints such as minimizing makespan (total execution time) and resource usage costs. This review explores recent approaches proposed to address this problem, focusing on heuristic, meta-heuristic, and hybrid algorithms.

IaaS cloud environments offer flexible and scalable infrastructure but pose significant challenges for the efficient deployment of workflows. Shahid et al. [7] developed a multi-objective allocation strategy aimed at balancing execution time (makespan) and resource usage costs. Their approach relies on heuristic algorithms, demonstrating a significant improvement in efficiency compared to classical methods. However, despite its performance, this method does not account for dynamic workload variations, limiting its adaptability. Zhang et al. [8] introduced EHEFT-R, an enhancement of the Heterogeneous Earliest Finish Time (HEFT) algorithm designed for multi-objective scheduling in heterogeneous cloud environments. The algorithm incorporates an exploration strategy based on evolutionary techniques, achieving a trade-off between minimizing makespan and optimizing resource utilization. This method stands out for its efficiency in dynamic and complex environments often observed in cloud computing [8] . Hussain et al. [9] introduce a cost-aware and deadline-constrained scheduling algorithm in a hybrid cloud environment. The focus is on meeting deadlines while minimizing expenses, which is particularly useful for workflows with strict temporal requirements. Simulations show that this approach outperforms traditional algorithms by balancing deadlines and costs. Mangalampalli et al. [10] introduced an innovative method leveraging deep reinforcement learning to prioritize and schedule workflows in cloud computing. Their approach combines dynamic modeling with task prioritization based on multi-objective goals. This method demonstrates significant improvements in scheduling accuracy and resource utilization, particularly in complex scenarios requiring rapid adaptability [10] .

Hybrid algorithms combine multiple approaches to leverage their respective strengths. Malti et al. [11] proposed a hybrid optimization algorithm integrating evolutionary and metaheuristic techniques for task scheduling in cloud environments. Their solution demonstrated optimal task allocation while reducing operational costs. Similarly, Abualigah and Diabat [12] introduced a bio-inspired optimization algorithm based on ant-lion behavior, which excels in complex environments due to its efficient exploration of the search space. These approaches highlight the power of hybrid solutions for scheduling problems, though they often require manual parameter tuning, which may limit their generalizability. Kruekaew and Kimpan [13] propose a hybrid approach combining the Artificial Bee Colony (ABC) optimization algorithm and reinforcement learning. The objective is to address the load balancing problem in cloud environments while meeting the goals of minimizing costs and makespan. This hybrid method demonstrates increased efficiency by optimally distributing workloads. The study by Doostali et al. [14] presents the CP-PGWO algorithm, combining Critical Path concepts with the Grey Wolf Optimizer to address scheduling challenges. This approach underscores the ability of hybrid algorithms to exploit the structural features of workflows while adapting to workload variations in cloud environments [14] . The DE-GWO algorithm [15] combines Differential Evolution (DE) optimization and Grey Wolf Optimization (GWO) for scheduling in heterogeneous fog-cloud environments. It aims to simultaneously minimize makespan and cost while optimizing resource allocation between fog and cloud nodes. The results show significant improvements compared to existing algorithms, particularly in environments with limited resources. The Enhanced Artificial Bee Colony (ABC) algorithm, introduced by Zeedan et al. [16] , exemplifies the effectiveness of nature-inspired algorithms in workflow scheduling problems. This hybrid approach combines the advantages of bee colony behavior with local search mechanisms, improving convergence toward optimal solutions. ABC is particularly suited for cloud computing environments due to its ability to balance conflicting objectives while ensuring efficient resource utilization [16] . Mohammadzadeh and Masdari [17] proposed a hybrid approach that combines multi-objective algorithms to optimize scientific workflows in multi-cloud environments. Their model leverages the synergy between local and global search heuristics, yielding robust solutions despite the complexity of tasks distributed across multiple clouds. Calzarossa et al. [18] tackled workflow planning under uncertainty, integrating parameters such as deadlines and budgets into a multi-objective approach. Their algorithm showed substantial performance improvements in unpredictable resource conditions, emphasizing the importance of resilience in cloud computing environments. Konjaang and Xu proposed the MOWOS (Multi-Objective Workflow Optimization Strategy), which applies optimization based on probabilistic models [19] . Their study demonstrated that MOWOS effectively minimizes gaps between theoretical and practical workflow performance by better modeling task and resource uncertainties. Qin et al. [20] introduced a reliability-focused multi-objective memetic algorithm for multi-cloud systems. Their approach balances the trade-off between makespan minimization and reliability maximization by combining global exploration mechanisms with local adjustments to enhance solution quality. Belgacem and Beghdad-Bey [21] focused on the trade-off between makespan and execution costs of workflows in the cloud. Their model simultaneously optimizes these objectives using a heuristic method tailored for highly heterogeneous environments. Modified genetic algorithms integrate adaptive techniques to address uncertainties in cloud environments. Rizvi et al. [22] proposed an approach incorporating fuzzy logic to dynamically adjust the parameters of a genetic algorithm. This improves makespan and reduces costs under changing conditions. A key strength of this method lies in its ability to adapt to unforeseen workload variations, though its increased algorithmic complexity may pose challenges for implementation in large-scale systems.

The integration of multiple metaheuristics into a single optimization method represents an emerging trend. Thekkepuryil et al. [23] developed an approach that combines Genetic Algorithms (GA) with Particle Swarm Optimization (PSO). This hybrid method enhances the exploration and exploitation of the search space, leading to more efficient resource utilization and reduced workflow execution times. Despite these advantages, implementing such algorithms can require significant effort in terms of design and computational complexity. Cai et al. [24] introduce a bi-level evolutionary algorithm for data-intensive scientific workflows. This multitasking algorithm allocates resources while considering task priorities and optimizing costs and deadlines. The results demonstrate increased efficiency for complex workflows requiring intensive data management.

Workflow scheduling in cloud computing remains an active area of research, with diverse approaches ranging from heuristics to hybrid algorithms incorporating learning techniques. However, challenges related to the dynamic nature of cloud environments and multi-objective trade-offs demand new perspectives. The proposed work aims to address these gaps by leveraging innovative techniques and expanded criteria, thereby enhancing the capabilities of modern cloud systems.

The works analyzed in the literature review show a diversity of approaches to solve the multi-objective optimization problem in workflow scheduling. Each study proposes specific algorithms adapted to particular contexts, such as cloud, multi-cloud, or fog-cloud environments, focusing on various objectives, such as minimizing makespan, costs, or load balancing. However, these contributions also present recurring limitations, particularly in terms of dynamic adaptability, scalability, or management of complex environments. To better understand the strengths and weaknesses of these methods, the following synthesis summarizes the main characteristics of the approaches studied in Table 1 .

Table 1. Comparative multi-objective workflow scheduling in cloud environments.

In this study, we design our platform model based on a standard IaaS cloud setup, deploying various virtual machine (VM) instances across physical servers within a single datacenter. Our focus is on VMs comparable to Amazon EC2 M5 instances, particularly the M5d series, which, unlike standard M5 instances relying on Amazon Elastic Block Storage (EBS), feature local NVMe SSD storage. Table 2 provides the specifications of the M5d instances used.

The M5d instance series offers virtual cores (vCPUs) ranging from 2 to 96, with a consistent memory allocation of 4 GiB per core. These instances are typically deployed on nodes equipped with Intel Xeon Platinum 8000 series processors. A unique feature of M5d instances is the inclusion of high-speed, block-level SSD storage directly linked to the instance’s lifecycle. Our study leverages this fast storage, shared among but exclusive to the instance’s vCPUs, to store intermediate files generated during workflow execution. This approach minimizes data transfers across the network for tasks assigned to the same VM, with only the workflow’s input and output files stored on an external storage node.

Network bandwidth between instances and EBS varies by instance size. We assume that only the largest instances (48, 64, and 96 vCPUs), which can occupy a full node, are assured bandwidths of 10, 20, and 25 Gbps, respectively. For smaller instances (2 to 32 cores), the bandwidth is proportional to the number of cores, allocated at 208.33 Mbps per core. All VMs deployed for a specific workflow are interconnected through a single switch.

For M5d instances, the VM-to-EBS connection is established via a dedicated network link, integrated into our simulation environment. We assume that the network bandwidth between a VM and EBS is proportional to the number of cores for VMs with up to 32 cores, estimated at 218.75 Mbps per core.

The costs indicated, in dollars per hour, correspond to on-demand Linux instances in the US-East (Ohio) region as of the time of writing this article.

Table 2. Characteristics of the AWS M5d instances.

The scientific workflows we aim to schedule are modeled as Directed Acyclic Graphs (DAGs), denoted as $G=\left\{T,ℰ\right\}$, where $T=\left\{t^i|i=1,\cdots ,T\right\}$ represents the set of vertices corresponding to the computational tasks of the workflow, and $ℰ=\left\{e^{i,j}|\left(i,j\right)\in \left\{1,\cdots ,T\right\}\times \left\{1,\cdots ,T\right\}\right\}$ is the set of edges between these vertices, indicating either data dependencies (i.e., file transfers) or control flows between tasks. We focus on workflows consisting of numerous sequential tasks that run on a single core, reflecting the structure of real-world scientific applications. Each task within the workflow has an estimated execution time, requires specific input files to start, and produces output files upon completion. To evaluate our contributions, we consider three workflow applications from the Pegasus Workflow Gallery [25] . Specifically, we use synthetic workflows generated by the Workflow Generator Toolkit [26] , which resemble those used in real-world scientific applications but with a higher task count. The primary characteristics of these applications are presented in Table 3 , and their structures are illustrated in Figure 1 .

Table 3. Some characteristics of used workflows.

To generate the different platforms and avoid wasting resources, i.e. avoid leasing resources that will not be used, the number of tasks that can be executed in parallel for each of the workflows is determined. To do this, the platform is oversized, i.e. this platform has as many cores as there are tasks in the workflow. In the case of the workflows used in this study, there are 1000 tasks, so the platform with 1000 cores is considered. This allowed to determine the total number of cores that could be used in parallel for each of the workflows. Thus, for CyberShake, 374 cores are used in parallel while for Epigenomics and Montage it is respectively 246 and 662 cores used in parallel. This preliminary study will be used in following section in order to determine the limit of platforms to be used for each workflow in experiments. For each workflow, we consider infrastructures where we vary (i.e. increase) the maximum number of cores per VM (from 2 cores to 96 cores maximum). For a total number of cores to be used, we generate platforms with 2 cores per VM, 4 cores per VM, ..., 96 cores per VM. The platforms composed of 2 cores per VM provide poor execution times compared to platforms of 32 cores per VM and they themselves provide poor execution times compared to platforms of 96 cores per VM. The increasing of the total number of cores per VM provides good execution times. It is for this reason that this study favor large VMs (i.e. VMs with several cores) because these VMs can execute several tasks in parallel and considerably reducing the makespan.

Figure 2 highlights the inverse relationship between the two objectives under study: makespan (total execution time) and cost, as a function of the number of allocated cores. Initially, as the number of cores increases, the makespan decreases significantly. This is due to the enhanced parallelism capability, allowing more tasks to be executed simultaneously. However, this reduction in execution time is accompanied by a progressive increase in cost, as resource usage prices in a cloud environment are directly tied to the number of cores utilized. This inverse relationship becomes even more pronounced beyond a certain threshold (approximately 50 - 100 cores), where adding additional resources results in only marginal reductions in makespan, while costs continue to rise almost linearly. In other words, beyond a certain point, each additional core contributes little to performance improvement but significantly inflates costs. This observation underscores the challenge of achieving a perfect balance between these two objectives, as further reducing makespan entails exponential cost increases.

In the Epigenomics workflow ( Figure 3 ), a trend similar to that observed in other workflows such as CyberShake ( Figure 2 ) becomes evident: the two objectives (makespan and cloud resource usage cost) exhibit opposing behaviors as the number of available cores increases. When the number of cores increases, the makespan (depicted by the blue curve) initially decreases drastically. This improvement is due to the enhanced ability to parallelize more tasks within the workflow, thereby reducing the total execution time. Simultaneously, the red curve representing the cost shows a progressive and non-linear increase with the number of cores. This is because the cost calculation formula depends on both the makespan and the total number of cores. Although the makespan decreases, the addition of more resources drives up expenses, particularly in configurations with a large number of cores. An important observation in such graphs is the presence of a zone where a trade-off between minimal makespan and acceptable cost is achievable. In the case of the Epigenomics workflow, this zone appears to be in the range of 40 to 100 cores, where the makespan reduction is significant, yet the cost increase remains moderate. Beyond a certain threshold (approximately 200 cores), the cost becomes excessively high without yielding a substantial reduction in makespan. This underscores the importance of an appropriate resource-sizing strategy to avoid unnecessary costs.

Figure 4 for the Montage workflow illustrates an inversely proportional relationship between the makespan and cost as a function of the total number of cores used. When the number of cores increases, the workflow execution time (makespan) decreases rapidly in the initial stages. This effect is particularly pronounced for smaller numbers of cores, where adding additional resources significantly enhances performance. However, as the number of cores continues to grow, the improvement becomes marginal, and the curve flattens, indicating a saturation effect where additional resources have a limited impact on the makespan. Conversely, the cost increases almost linearly with the total number of cores. This phenomenon is explained by the cost calculation formula, where increased resource usage directly translates into higher expenses, even as the makespan improvements diminish. This highlights the inherent trade-off in cloud resource allocation between minimizing execution time and managing costs effectively.

The main purpose of such a study lies in identifying an optimal trade-off between the conflicting objectives of makespan and cost. This trade-off is represented by a Pareto front, which encompasses all non-dominated solutions, where an improvement in one objective can only be achieved at the expense of the other. Analyzing the behavior of these curves not only allows us to visualize the complex relationship between these two objectives but also to identify strategic solutions tailored to the specific needs of users or applications. Below, we present the Pareto fronts obtained for the different workflows studied, highlighting the optimal solutions in this multi-objective space.

The main objective of our algorithm is to help IaaS cloud users find a good trade-off between the execution time and cost of their workflow by selecting sets of VM instances on a Pareto front. In this section, we evaluate the quality of the Pareto front solutions produced by our algorithm. In the previous sections, we have shown that to minimize the execution time for a fixed number of cores, priority should be given to large VM instances. To obtain the Pareto front given by our algorithm, it is therefore sufficient to perform a simulation for each total number of cores and discard all dominated solutions. Two solutions, S_i and S_j, are considered equivalent if the same tasks are assigned to different but equivalent VMs with identical start times. During each iteration, only one representative of such equivalent solutions is retained. Furthermore, in all iterations except the final one, only strictly dominated solutions are eliminated. A solution S_i is strictly dominated by a solution S_j if both the execution time and cost of S_i are strictly higher than those of S_j. This approach ensures that intermediate solutions, which could lead to better schedules, are not prematurely discarded. In the final iteration, however, all simply dominated solutions (i.e., solutions where one metric is worse) are removed. The three algorithms exhibit similar behaviors, with a rapid decrease in cost as makespan increases from the initial points, followed by stabilization. The MOS-MWMC algorithm (red points) demonstrates superior dominance over the entire set of solutions. Its curve is generally closer to the origin, indicating a better trade-off between makespan and cost. Compared to HEFT and MAX-MIN, MOS-MWMC is able to provide solutions with slightly lower costs while maintaining competitive makespans. The HEFT algorithm (dashed green line) follows a similar trajectory but is slightly less efficient than MOS-MWMC for most trade-offs. HEFT solutions tend to have slightly higher costs, especially in the low makespan region. However, for very high makespans, HEFT reaches the performance of MOS-MWMC, indicating convergence. The MAX-MIN algorithm (dashed blue line) is generally less effective. Its curve is often above those of MOS-MWMC and HEFT, indicating higher costs for similar makespans. This difference is more pronounced in the region where the makespan is low, suggesting that MAX-MIN is not optimal for minimizing both objectives simultaneously. The superiority of MOS-MWMC is evident in the critical areas where the trade-off between makespan and cost is most important (low makespans). This demonstrates the algorithm’s ability to effectively balance these two conflicting objectives for the scientific workflow CyberShake ( Figure 5 ).

The Epigenomics workflow ( Figure 6 ) reveals a markedly superior performance of the MOS-MWMC algorithm compared to the HEFT and Max-Min methods. The points generated by MOS-MWMC stand out due to their proximity to the Pareto front, demonstrating an enhanced ability to provide optimal solutions by simultaneously minimizing both cost and makespan. This proximity reflects the effectiveness of MOS-MWMC in balancing the two objectives, a crucial aspect for complex workflows. In contrast, the HEFT and Max-Min algorithms, while following similar trends, struggle to achieve the same levels of performance. For equivalent makespan values, the costs associated with the solutions of HEFT and Max-Min are significantly higher, highlighting their relative inadequacy for this type of workflow. Furthermore, the solutions from these two algorithms exhibit greater dispersion, suggesting variability in their ability to effectively optimize the objectives.

In the case of the Montage workflow ( Figure 7 ), MOS-MWMC confirms its superiority by generating solutions that also approach the Pareto front, even in scenarios with more pronounced trade-offs between cost and makespan. The robustness of this algorithm is particularly evident in its ability to maintain an optimal balance between the two objectives, even for extreme points where cost or execution time becomes the priority. In contrast, the HEFT and Max-Min algorithms, while displaying similar trajectories to those observed for Epigenomics, remain inferior in terms of the quality of the solutions produced. HEFT, while slightly more effective than Max-Min in this case, fails to reach the optimal trade-offs proposed by MOS-MWMC. The increased complexity of the Montage workflow is reflected in the variability of the solutions obtained, further highlighting the adaptive capacity of MOS-MWMC in the face of diverse workflow profiles.

The analysis of the results for the CyberShake, Epigenomics, and Montage workflows highlights the consistent superiority of the MOS-MWMC approach over the HEFT and Max-Min algorithms. MOS-MWMC excels in its ability to balance the conflicting objectives of cost and makespan, demonstrating remarkable robustness and adaptability in complex and varied workflow environments. The solutions generated by this algorithm are characterized by their proximity to the Pareto front, offering trade-offs that effectively meet the demands of cloud environments. In summary, MOS-MWMC emerges as the preferred solution for scheduling scientific workflows in cloud computing environments, especially in scenarios where multi-objective optimization is critical. Its consistent performance and flexibility make it an essential approach for addressing challenges related to resource efficiency and time constraints.

Table 4 represents a comparison of the performance parameters for the three algorithms MOS-MWMC, HEFT and Max-Min, based on typical results from our simulations. The MOS-MWMC algorithm dominates on most criteria, notably due to its ability to simultaneously optimize makespan and costs, while efficiently exploiting virtual machine (VM) resources. As for HEFT, this algorithm offers acceptable performances but remains inferior to MOS-MWMC on multi-objective trade-offs, particularly for complex workflows. On the other hand, the performances of the Max-Min algorithm are generally inferior, notably in terms of costs and makespan minimization, which makes it less suitable for demanding cloud environments.

Table 4. Performance criteria for MOS-MWMC, HEFT, and Max-Min.

Infrastructure as a Service (IaaS) clouds now enable scientists to run data-intensive workflows on customized infrastructures designed to meet the specific computing and storage needs of these applications. Selecting the appropriate set of virtual machine instances to form these infrastructures is a challenging task, typically handled by Workflow Management Systems. Maximizing performance hinges on effectively utilizing the unique characteristics of these virtual machine instances.

In this paper, we proposed an innovative multi-objective approach called MOS-MWMC. The MOS-MWMC algorithm establishes itself as a benchmark solution for multi-objective workflow scheduling in IaaS cloud environments. Unlike the HEFT and Max-Min algorithms, MOS-MWMC stands out for its ability to simultaneously optimize makespan and costs while maintaining robust and adaptable performance. This efficiency is attributed to several key strengths. Firstly, the algorithm generates solutions that are close to the Pareto front, ensuring an optimal trade-off between conflicting objectives. Secondly, it significantly reduces data transfers through a strategy of local storage for intermediate files, thereby minimizing network communication, a major advantage over HEFT and Max-Min.

Moreover, MOS-MWMC makes optimal use of virtual machine resources by leveraging their features, such as a high number of cores and fast storage, resulting in better task distribution and more efficient utilization of cloud infrastructures. Additionally, its robustness in handling complex workflows, as demonstrated by its performance on CyberShake, Epigenomics, and Montage, makes it particularly suitable for diverse and demanding scenarios. Experimental results show that MOS-MWMC consistently outperforms HEFT and Max-Min in minimizing makespan and optimizing costs, further validating its relevance in addressing the challenges of modern cloud environments.

Finally, this study opens up promising future prospects, including the evaluation of MOS-MWMC in real cloud environments and the integration of additional objectives, such as energy consumption reduction. These potential advancements would further enhance the relevance of this algorithm for industrial applications, where performance, flexibility, and cost control are critical.