1. Introduction

Today, with the rapid development of internet technology, cloud computing processing capability has gradually matured, and its understanding has deepened. It continuously improves task-scheduling algorithms, work efficiency, and people's life quality ^[1,2]. The maturity of cloud computing capabilities means that more high-quality resources can be scheduled through a unified cloud platform, thereby optimizing the resource structure. The advent of edge computing has brought about new demands for scheduling computations on cloud computing platforms. Consequently, it has become a pressing issue for cloud computing platforms to discover an accurate and high-quality scheduling algorithm that can meet these demands effectively ^[3,4].

With the continuous promotion and application of cloud computing technology, the user scale of cloud computing platforms continues to expand. The massive user data and parallel user tasks bring great challenges to the cloud computing backend, requiring optimization of task scheduling and allocation in cloud computing and finding efficient scheduling optimization solutions for cloud computing platform tasks ^[5,6]. However, resource scheduling algorithms are difficult to adapt to the huge computational workload of cloud platforms and have poor learning ability ^[7]. Therefore, heuristic algorithms have received significant attention from researchers, and their powerful learning and adaptation capabilities are exactly what cloud computing platforms require.

This study explores an improved task scheduling method of cloud computing platforms for customer-oriented online training applications based on an ant colony optimization (ACO) algorithm. The parameters of other ACO algorithms are set based on past experience, so they are prone to fall into local optima and have low robustness. Therefore, based on the classic ant colony algorithm, a virtual machine evaluation factor and pheromone correction coefficient can be used to determine the optimal value of the algorithm parameters and effectively improve the task execution efficiency of the algorithm.

The study utilized an improved ant colony algorithm to optimize and schedule task resources on cloud computing platforms, effectively achieving a relative balance of system load and reducing system energy consumption. On the basis of considering overall coordination, the optimization of task scheduling on cloud computing platforms was effectively improved.

2. Related Work

To promote a customer online training system to connect with the intelligent era, the study of an intelligent cloud-computing scheduling strategy has become a focal issue nowadays. Wu et al. believe that cloud computing systems are dynamic and diverse, and their task and resource scheduling is very challenging. They introduced a novel cloud-computing task-scheduling algorithm that leveraged particle swarm optimization to model the resource scheduling problem in cloud computing systems. They formulated an objective function of the task execution time and adopted the particle swarm optimization algorithm to simulate the outcomes. Their proposed method had high utilization rates of cloud computing resources, leading to a significant reduction in cloud computing task time ^[8].

Smith believes that online training will be one of the most important learning methods in the future. When people are faced with a threat from the environment and viruses, they have to stay at home to study and work. Therefore, online training could become an inevitable way of learning in the future. In addition, online training is also the latest form of online education evolution and is the future of online education. Online training covers the dual attributes of education and training, which can meet people's learning and practical needs ^[9].

Majumder et al. proposed that online training is an important technical means of education and training in imaging medicine. They believe that practice is an extremely important part of medical imaging education and training and plays an irreplaceable role in improving clinical experience. The emergence of online training platforms helped them solve the difficulty of opening practical teaching after the COVID-19 pandemic and provided great help for medical students. Therefore, actively promoting online training platforms has become an inevitable trend of future development ^[10].

To solve the problem of cloud computing resource transmission, Lu et al. Proposed mobile edge computing (MEC). For the problem of executing task scheduling in MEC servers, they proposed a task scheduling-based queuing algorithm (TSBQ) that took into account data transfer latency and server load and implemented a reasonable task allocation policy. Experimental results showed that the MS-CE architecture outperformed other architectures, and TSBQ was more efficient than Corral and Greedy ^[11]. Oliveros upgraded the online training system of a university and developed an interactive online training system. With the help of this system, students and teachers had good interaction. At the same time, schools and teachers could correctly help students learn with the help of the system. This research provides a good reference for the extension of an online training system ^[12].

As an excellent metaheuristic algorithm, the ant colony algorithm is widely used in cloud computing. Many researchers use the ant colony algorithm to solve combinatorial optimization problems in cloud computing. Ge et al. proposed a multidimensional quality of service (QoS) cloud-computing task-scheduling algorithm based on an improved ant colony algorithm while considering the QoS needs of users and the load balancing of cloud platforms. They defined a QoS model composed of task completion time and execution cost and defined a cloud platform load balancing constraint function.

They also addressed the shortcomings of slow convergence speed and easily falling into local optima of the ant colony algorithm. The pheromone update method and expectation heuristic function were improved, and the pheromone strength was dynamically changed. The simulation results showed that the proposed algorithm was superior to the ACO algorithm and max-min ant system (MMAS) algorithm in terms of user satisfaction and cloud platform load ^[13].

Dahan proposed a quality-aware ant colony optimization algorithm to solve the problem of cloud service composition in cloud computing. The ant colony algorithm was used to optimize and perceive the service quality attributes of different suppliers, thereby achieving cloud service composition to meet the complex and personalized cloud-service needs of users. The study conducted algorithm-comparison experiments using 25 real datasets, and the results showed that the ACO algorithm was the most competitive ^[14].

Liu et al. proposed an integrated model based on extreme learning machines and ACO algorithms for virtual machine integration in cloud computing environments. The model utilized extreme learning machines to predict host states and then migrated virtual machines on overloaded hosts. The model combined ACO algorithms for local search and group migration plan optimization to avoid excessive virtual machine integration. The research results indicated that the integrated model based on extreme learning machines and ACO algorithms can effectively reduce model energy consumption and improve transfer speed ^[15].

Karmakar et al. introduced ACO algorithms to the optimization problem of virtual machine placement to find the best solution for virtual machine integration. They explored the application of ACO algorithms in solving multi-objective problems. They utilized this approach to optimize the placement problem of virtual machines to satisfy the dual requirements of service providers and users' service quality while also enhancing the quality of virtual machine services and reducing operational costs for suppliers. The study compared the ACO algorithm with multi-objective and single-objective problem-solving algorithms, and the results showed that the ACO algorithm had the best solution performance and convergence ^[16].

The optimization and convergence of intelligent task scheduling methods on cloud computing platforms still need to be improved. At present, cloud computing platforms have poor performance in task scheduling, poor load performance, and communication delays in parallel multitasking processing, The ability and efficiency to solve high task volume problems are not high. The ant colony algorithm performs well in solving cloud computing optimization problems. Therefore, an improved ant colony algorithm was used for adaptive optimization research to improve the task processing ability and efficiency of cloud-computing task-scheduling algorithms. This study proposes using virtual machine evaluation factors and pheromone correction coefficients to remedy the shortcomings of ACO in solving optimization problems and optimize cloud-computing task scheduling.

3. Task Scheduling Method based on Ant Colony Algorithm

A cloud platform for a customer-oriented online practical training system was established based on a general customer-oriented cloud platform, and its task scheduling model follows the cloud platform rules and uses cloud computing for task scheduling ^[17]. The core processing power of cloud-computing task scheduling comes from the scheduling algorithm. Through the deployment of the scheduling algorithm, the limited resources of the cloud platform are reasonably allocated to each running task and the tasks that will be run in the future. The structure of cloud-computing task scheduling is shown in Fig. 1.

Fig. 1. Cloud-computing task-scheduling structure.

Cloud scheduling is a comprehensive scheduling method that combines static scheduling and dynamic scheduling. Static scheduling is mainly for task scheduling for consumption minimization optimization ^[18]. Dynamic scheduling focuses on resource queuing and load balancing issues for resource provisioning. In addition, cloud-computing task scheduling requires high heterogeneity, scalability, and dynamic adaptability of the scheduling system ^[19]. The goal of task scheduling is to ensure optimal span, guaranteed quality of service, and load balance, and preferably to maintain a certain level of affordability. The choice of scheduling algorithm for cloud computing is extremely demanding and is an NP-complete problem.

The study used the ant colony algorithm for cloud computing scheduling using the traversal city shortest path problem as an entry point. Assuming that $b_{i}(t)$ is the number of ants presenting in the region $i$ at the moment $t$ , $\tau _{ij}(t)$ is the concentration of pheromones left on the path ($i,j$) at the moment t. The number of cities is denoted by $n$, and the total number of ants by is denoted $m$ ($m=\sum _{i=1}^{n}b_{i}(t)$). $\chi $ is used to denote the set of information concentrations of any two cities in the city set $C$ at the moment $t$, and $\chi $ =$\left\{\tau _{ij}(t)\left| c_{i},c_{i}\subset C\right.\right\}$. The initial settings of the relevant parameters and the road-energy pheromone are made before the start of the algorithm. $\eta _{ij}$ represents a heuristic factor of the path ($i,j$), which is calculated as shown in Eq. (1).

In Eq. (1), $d_{ij}$ denotes the distance from node $i$ to $j$. The path length is inversely proportional to the heuristic factor. When the path is longer, the heuristic factor is smaller, indicating that it is less attractive to the ants. During the travel of ants, the selection of paths is closely related to the concentration of pheromones on path $\tau _{ij}(t)$. ``Tabu'' is used to refer to a taboo table to prevent the ants from passing through duplicate paths, and ``$allowed$''refers to the next target city of the ants, which divides the taboo table (i.e., $allowed$=$\left\{C-tabu\right\}$). Then, the state transfer probability $P_{ij}$ of an ant moving from node $i$ to node $j$ is calculated as shown in Eq. (2).

In Eq. (2), $\alpha $ is the influence coefficient of the pheromone, which indicates the degree of guidance of the pheromone on the path, and $\beta $ is the influence coefficient of the heuristic factor, representing its influence on the path selection. The ant forms a complete path after traversing the nodes following the selection method in Eq. (2). Every time it passes through a node, the ant updates the local pheromone, as shown in Eqs. (3) and (4).

After traversing all nodes, the routes generated by the iterations are updated globally with pheromones as shown in Eqs. (5) and (6).

$\rho $ in Eqs. (3) to (6) denotes the pheromone volatility factor, $\Delta \tau _{ij}(t)$ denotes the pheromone increment on the path ($i,j$) at each cycle, and $\Delta \tau _{ij}^{k}(t)$ denotes the pheromone increment generated by the ant $k$ when it passes through the path ($i,j$).

When applying ACO to practical problems, there are three common models: perimeter, volume, and density. The method of the ant perimeter model is shown in Eq. (7).

In Eq. (7), $Q$ denotes the intensity constant of the pheromone, $T^{k}\left(t\right)$ denotes the allocation scheme of the ant $k$, at the completion of the cycle, and $L_{k}$ denotes the length of the path passed by ant $k$ . The method of the ant volume model is shown in Eq. (8).

The method of the anthropomorphic model is shown in Eq. (9).

The three algorithmic models are compared in detail in Table 1.

Table 1. Comparison of three ant colony algorithm models.

The research on task scheduling tends to consider the whole situation, so it is more appropriate to choose the ant-week model for global information updates. The running process of the ant colony algorithm is shown in Fig. 2. The ant colony algorithm has strong stability because the ants are simple in foraging and they communicate with each other using only pheromones without interference from other factors. By distributing the ant colony to different nodes, the path can be found spontaneously. However, the ant colony algorithm also has problems such as slow solution speed in the initial stage, easily falling into local optimal solutions, computationally tedious selection of nodes, and large dependence on initial parameters. In the beginning stage, the pheromone concentration on the paths is low, and the selection of paths relies almost entirely on heuristic factors. This may lead to more ants choosing longer paths, and the pheromones released during crawling may mislead subsequent ants, leading to an ineffective search.

Fig. 2. Flow chart of ant colony algorithm.

In the early stage of ant search, ants may also rotate in place due to the few nodes recorded in the taboo table, which may also reduce search efficiency. The accumulation of key pheromones in the search process takes some time, so the search efficiency of the ant colony algorithm in the initial stage is relatively low. The possible shortest path is easily ignored by ants because the pheromone concentration on it is too low or even 0. For the paths with high accumulated pheromone concentration, even if they are not the shortest paths, the excessive pheromone accumulated on the paths affects the path selection of ants and leads to locally optimal solutions because the pheromone cannot be volatilized in time. When ants select the next node, the ant colony algorithm needs to calculate the probability of state transfer between all unselected nodes. When the problem size is large, the number of paths recorded in the forbidden table is very large, and the final path plan of each ant grows exponentially.

3.2 The Improvement of Ant Colony Algorithm and Cloud Platform Optimization Adaptation Research

During the crawling process, the probability of ants choosing path $\left(i,j\right)$ is related to the heuristic factor and the concentration of path retention information. If $tabu$ is used to represent the taboo table, and only the cities that the ants have already walked through are included, it can prevent ants from repeatedly selecting paths. If allowed is used to represent cities outside the taboo table, the state transition probability formula for ants transitioning from node $i$ to node $j$ can be represented by Eq. (10).

(10)

$ P_{ij}=\left\{\begin{array}{l} \frac{\left(\tau _{ij}\right)^{\alpha }\left(\eta _{ij}\right)^{\beta }}{\sum _{s\in allowed}\left[\left(\tau _{is}\right)^{\alpha }\left(\eta _{is}\right)^{\beta }\right]},j\in allowed\\ 0,\begin{array}{ll} & \end{array}otherwise \end{array}\right. $

In Eq. (10), the state transfer probability $P_{ij}$ has the characteristics of stochasticity combined with determinism. According to its dynamic and static characteristics, the optimization of the heuristic factor and the pheromone concentration on the path of the ant colony algorithm can effectively improve the performance of the algorithm. Combined with the operation resource scheduling, in addition to the efficiency of the algorithm, the load of the virtual machine should be taken into account. The computational capacity of the virtual machine is denoted by $E_{ij}$ , which is calculated as shown in Eq. (11).

(11)

$ E_{ij}=\frac{length_{i}}{mips_{j}}+\frac{size_{i}}{bw_{j}} $

In the operation process, the heuristic factor is influenced by the computing power of the VM $E_{ij}$ in addition to the direct impact of the performance of the VM itself, which is required to add the VM evaluation factor $\sigma _{i}$ to evaluate the operation status of the VM $VM_{i}$. $\sigma _{i}$ is calculated as shown in Eq. (12).

(12)

$ \sigma _{i}=\sqrt{\sum _{{T_{i}}\in Tas{k_{j}}}\left(E_{ij}-\frac{E_{j}}{m}\right)^{2}/m} $

In Eq. (12), $E_{i}$ represents the overall ability of the virtual machine $VM_{i}$ to perform tasks in the task set $Task_{i}$. The virtual machine evaluation factor $\sigma _{i}$ represents the overall stability of the execution, and a smaller value indicates better stability. The optimized path heuristic factor $\eta _{ij}$ is calculated as shown in Eq. (13).

(13)

$ \eta _{ij}=\frac{1}{\sigma _{j}\ast E_{ij}} $

In the ant colony algorithm, ants are influenced not only by heuristic factors when searching for a path, but also by the concentration of remaining pheromones on the path. However, since the ant's search path process is spontaneous, it is highly stochastic. In an environment where the remaining pheromone concentration on the path is low in the initial stage, ants may choose longer paths and also leave pheromones behind, which may mislead ants, so all ants will choose longer paths. This makes the ant colony algorithm too slow and inefficient in an early stage. In order to minimize the pheromone interference with the ants on long paths during the search process, the global pheromone update rule in the ant cycle model of the classical ant colony algorithm was improved, and a pheromone correction factor $w$ ($0<w<1$) is introduced to solve this problem. The pheromone update method was optimized using the correction coefficients, and the results are shown in Eq. (14).

(14)

$ \Delta \tau _{ij}^{k}\left(t\right)=\left\{\begin{array}{l} \left(1+w\right)\frac{Q}{L_{k}},L_{k}<L_{\min }\\ \left(1-w^{2}\right)\frac{Q}{L_{k}},L_{\min }<L_{k}<L_{\max }\\ \left(1-w\right)\frac{Q}{L_{k}},L_{k}>L_{\max } \end{array}\right. $

In Eq. (14), $w$ represents the correction factor of the pheromone, $L_{k}$ represents the total path length currently traversed by an ant, $L_{\min }$ represents the length of the shortest path in the traversal, and $L_{\max }$ represents the length of the longest path in the traversal. If the ideal route reached by the ant during the search is shorter than the shortest route found by all ants so far, the pheromone concentration on the path is enhanced according to the coefficient, and more ants are attracted. During a given search, if an ant discovers an ideal path that is longer than the shortest foraging path that it has previously found, but it is still shorter than the longest path, the pheromone on that path will slightly decrease. The reduction amount is based on a coefficient subtracted by the ratio of 1 to the small proportion, which helps minimize other ants mistakenly selecting a longer route.

On the other hand, if the ideal path discovered by this ant during the search is longer than the longest path found by any of the ants so far, then the pheromone on the worst path will be significantly decreased. The reduction amount is determined by the corresponding coefficient, which helps reduce the likelihood of following ants incorrectly choosing the worst path. The flow of the optimized algorithm is shown in Fig. 3.

Fig. 3. Flow chart of optimized ant colony algorithm.

As shown in Fig. 3, after the environment parameters are initialized, the ants begin to search and continuously perform resource node searches according to the optimized state transition probability method. As ants continue to move in new nodes, they constantly modify the tabu list. Once an ant acquires the ideal route, it compares the feasible routes of all other ants and records the best and worst routes. The global pheromone is then updated accordingly. This process is repeated until the optimal route is achieved. The optimized ant colony algorithm still needs a suitable initial setting to avoid problems such as falling into local optimum solutions or wasting resources. The optimal solution is obtained for the information correction coefficient $w$ , and the updated image of the pheromone is drawn according to the ant-optimized shortest path, as shown in Fig. 4(a).

Fig. 4. Information correction coefficient correlation function image.

The iterative difference between the optimal solution and the worst solution can be obtained from Fig. 4 and Eq. (14), as shown in Eq. (15).

(15)

$ \left(1+w\right)-\left(1-w\right)=2w $

The iterative difference between the optimal solution and the general solution is shown in Eq. (16).

(16)

$ \left(1+w\right)-\left(1-w^{2}\right)=w+w^{2} $

To achieve maximum optimization, the difference between the optimal solution and the worst solution should be widened as much as possible. Conversely, to reduce interference, the gap between the optimal solution and the general solution should be minimized, as illustrated in Fig. 4(b). The difference derivation between Eqs. (15) and (16) is performed, and the result is shown in Eq. (17).

(17)

$ f'\left(w-w^{2}\right)=1-2w $

According to the calculation result, the maximum value of the difference is obtained by making$1-2w$ equal to 0, (i.e.,$w=0.5$) and setting the pheromone correction factor to 0.5.

4. Experiment and Analysis

The performance of the optimized ant colony cloud scheduling algorithm (OACC) proposed in the study was simulated and experimentally analyzed using the platform CloudSim 4.0 in the Windows 10 operating system. The virtual machine parameters and task parameters are shown in Table 2. The pheromone correction parameter w was introduced into the optimized ant colony algorithm. To improve the accuracy of the experiment, OACC needed to be simulated with the default parameter settings to determine the optimal value of the parameters. The experimental results are shown in Fig. 5.

Fig. 5. Simulation results of determining parameters by control variable method.

From Fig. 5, it can be seen that the parameter value with the shortest completion time was selected as the optimal parameter value based on the control variable method. In the control variables of the pheromone weighting factor, the shortest completion time was optimal with $\alpha =1.4$. In the control variables of the pheromone weighting factor, the shortest completion time is optimal with $\beta =2.6$. In the control variables of the pheromone volatility factor $\rho $ , the shortest completion time is optimal with $\rho =0.59$. In the control variables of the pheromone correction factor, the shortest completion time was optimal when assigned a value of 0.59. In the control variable of the pheromone correction factor, the shortest completion time was optimal with $w=0.52$.

After setting parameters, comparative experiments were carried out on the simulation platform, and the OACC proposed in this research was compared with the discrete firefly algorithm (DFA), improved group search optimization (IGSO), and improved differential evolution (MODE). Experiments were conducted on the four algorithms in the same experimental environment, and the average of the results of 10 experiments was recorded. The comparison results are shown in Fig. 6.

Fig. 6. Comparison of simulation results of different algorithms performing different tasks.

As shown in Fig. 6, when the number of tasks was small, there was not much difference in the performance of each algorithm. As the number of tasks increased, there was a significant gap between the OACC, DFA, IGSO, and MODE algorithms. When the number of tasks was 300, the makespan value of OACC was 340, that of DFA was 350, that of MODE was 380, and that of IGSO was 409. The overall performance of OACC was 20.3% higher than that of IGSO, which met the expectations and reflected the effectiveness of the optimization algorithm.

Fig. 7. Load unbalance comparison of four algorithms executing different tasks.

The degree of load imbalance (DI) of the four algorithms executing different task volumes is shown in Fig. 7. Fig. 7, it can be seen that OACC and MODE maintain low and stable DI values under different task count tests. However, IGSO and DFA performed poorly. This result indicates that the OACC proposed in this study outperformed IGSO in system load balancing, and the research results are in line with expectations. The good performance of MODE indicates that the virtual machine evaluation factor improved load balancing in the cloud computing system. The OACC algorithm was applied to cloud computing task instances, and the overall satisfaction of customer service was compared using the overall utility function F. The overall utility comparison of the OACC proposed in this study with AOC, TAOC, and LB-AACO is shown in Fig. 8.

Fig. 8. Load unbalance comparison of four algorithms executing different tasks.

As can be seen from Fig. 8, OACC performed the best for all different task quantities, followed by LB-AACO. The other two algorithms performed relatively poorly. At a task quantity of 300, the overall utility of OACC was rated as 146, which is 31.5% higher than ACO, 18.7% higher than TACO, and 8.1% higher than LB-AACO. The results of this study indicated that the overall utility of OACC was better than that of similar algorithms, and it can perform well for cloud-computing task scheduling.

5. Conclusion

This research focused on the problem of task scheduling methods for cloud computing platforms in customer-oriented online training systems. Based on the optimization of the ant colony algorithm, an optimized ant colony cloud-computing task-scheduling algorithm was proposed. The research results indicated that when the number of tasks was 300, the makespan value of OACC was 340, that of DFA was 350, that of MODE was 380, and that of IGSO was 409. The overall performance of OACC was 20.3% higher than that of IGSO. OACC maintained low and stable DI values under different task count tests. At a task volume of 300, the overall utility evaluation of OACC was 146, which was 31.5% higher than ACO, 18.7% higher than TACO, and 8.1% higher than LB-AACO.

The experimental results met expectations, indicating that the OACC cloud-computing task-scheduling algorithm proposed in this study had high task processing ability and efficiency and was capable of scheduling tasks on cloud computing platforms for customer-oriented online training systems. However, there are also some shortcomings in this research, such as the cloud computing scheduling tasks in the experiment not being comprehensive enough. It is expected that in future research, more experiments with different types of tasks can be conducted to promote the method to become more comprehensive and accurate.