A multi-constraint scheduling method for network flow in a heterogeneous cloud environment based on immune particle swarm optimization

Abstract

This invention relates to a multi-constraint scheduling method for network flows in heterogeneous cloud environments based on immune particle swarm optimization, belonging to the field of cloud computing resource scheduling and optimization technology. The method first abstracts the network flow as a directed acyclic graph and encodes the scheduling scheme as particles; then, it constructs a multi-objective fitness function that considers both execution cost and completion time, and introduces antibody concentration calculation and incentive mechanisms; finally, iterative optimization is performed through immune operations such as antibody cloning, crossover mutation, memory cell update, and population refresh to output the optimal scheduling scheme. This invention effectively balances population diversity and convergence speed through an immune mechanism, and can significantly optimize the completion time of the scheduling scheme while meeting user deadline constraints, providing a solution for efficient scheduling of complex network flows in cloud computing environments.

Description

Technical Field

[0001] This invention belongs to the field of cloud computing resource scheduling and optimization technology, specifically involving a multi-constraint scheduling method for network flows in a heterogeneous cloud environment based on immune particle swarm optimization. Background Technology

[0002] In the current context of accelerated digital transformation, cloud computing, with its core advantages of elastic scaling, resource sharing, and pay-as-you-go pricing, has become a critical infrastructure supporting the operation of data-intensive and computationally intensive network streaming tasks. Whether in scientific computing scenarios such as earthquake hazard analysis and genome sequencing, or in commercial service scenarios such as large-scale data processing and distributed application deployment, network traffic optimization directly impacts resource utilization efficiency, task execution costs, and service quality. Network streaming tasks typically have complex dependencies, requiring abstract description in the form of a directed acyclic graph (DAG). Furthermore, their traffic characteristics exhibit significant heterogeneity, temporal correlation, and dynamic fluctuation: traffic heterogeneity is reflected in the large differences in data transmission volume and computational load across different tasks; for example, scientific computing tasks are often high-computation, low-transmission, while data processing tasks are high-transmission, low-computation. Temporal correlation in traffic means that traffic transmission between dependent tasks follows a strict sequence; the output data of the preceding task must be transmitted to the subsequent task in a timely manner, otherwise, task blocking will occur. Dynamic fluctuation in traffic refers to the fact that task load and data transmission rate change over time in real-world scenarios, further increasing scheduling difficulty.

[0003] From a computational perspective, network traffic also exhibits significant characteristics of computational load coupling, scalability, and quality of service (QoS) sensitivity. Computational load coupling manifests as a strong correlation between traffic volume and task computational complexity; high-complexity computational tasks involve the exchange and transmission of a large amount of intermediate result data, with traffic peaks highly synchronized with the execution phase of the computational task. Scalability is reflected in the dynamic change of traffic volume with the parallelism of computational tasks; when cloud nodes are expanded to improve computational parallelism, cross-node data exchange traffic grows exponentially. QoS sensitivity refers to the strict requirements of computational tasks on traffic transmission latency and packet loss rate. For example, real-time data analysis tasks require low-latency traffic transmission; otherwise, the timeliness and accuracy of computational results will be directly affected. Network flow scheduling not only needs to meet the timing constraints and data transmission requirements between tasks but also needs to achieve multi-objective optimization in heterogeneous cloud resource environments. This characteristic makes network flow scheduling in cloud computing environments a research hotspot with both theoretical and practical significance.

[0004] Currently, the core challenges of network flow scheduling focus on two levels: first, the rationality of bandwidth allocation, which requires selecting suitable cloud node types and determining the lease duration and number of cloud nodes based on task characteristics; second, the optimality of task mapping, which requires efficiently allocating dependent tasks to cloud node resources, achieving dual optimization of execution cost and completion time while meeting user-defined deadlines. Traditional scheduling algorithms, such as greedy algorithms and heuristic algorithms, while simple to implement and computationally inexpensive, are prone to local optima when dealing with complex scheduling problems with multiple constraints and objectives, making it difficult to balance the contradiction between cost and efficiency.

[0005] Particle Swarm Optimization (PSO), a stochastic optimization technique based on swarm intelligence, is widely used in scheduling optimization due to its fast convergence and robustness. This algorithm simulates the foraging behavior of bird flocks, abstracting candidate solutions into particles and using the particle's velocity and position update mechanism to search for the optimal solution. However, PSO has significant drawbacks during the iteration process: particles tend to cluster towards the current global optimum, leading to a decrease in population diversity; when the current optimum is a local optimum, the algorithm struggles to escape the search trap, ultimately affecting the quality of the scheduling scheme. This problem is particularly prominent in high-dimensional and complex optimization problems such as network flow scheduling, severely limiting the practical application effectiveness of PSO.

[0006] The antigen recognition, antibody diversity maintenance, immune memory, and self-regulation characteristics of the biological immune system offer novel approaches to solving the local optima problem in swarm intelligence algorithms. Artificial immune systems, by simulating this biological mechanism and introducing operations such as affinity calculation, clonal mutation, and population refresh, can both screen for high-quality candidate solutions to accelerate convergence and maintain population diversity by regulating antibody concentration, effectively avoiding local optima traps.

[0007] To address the network flow scheduling problem in heterogeneous cloud environments, existing technologies mainly fall into three core categories:

[0008] The first category is traditional greedy and heuristic algorithms. Their core logic is to make local optimal decisions based on task priority or resource attributes. They allocate cloud node resources by calculating the earliest completion time of the task. They are simple to implement and have low computational overhead. However, when dealing with problems involving multiple constraints (QoS) and multiple objectives (cost and time optimization), they are prone to getting stuck in local optimal solutions and it is difficult to balance resource utilization and scheduling efficiency.

[0009] The second category is particle swarm optimization algorithms, which abstract the scheduling scheme into particles by simulating swarm intelligence and search for the optimal solution using a velocity-position update mechanism. These algorithms have fast convergence speed and strong robustness. However, they have inherent drawbacks: particles tend to cluster towards the global optimum during iteration, leading to a decrease in population diversity. In scenarios with complex network flow task dependencies and heterogeneous traffic characteristics, they struggle to escape local optima traps. Furthermore, they do not fully consider the adaptability to QoS constraints and dynamic traffic characteristics, limiting the practicality and stability of the scheduling scheme.

[0010] The third type is multi-objective evolutionary algorithms, which decompose multiple objectives into sub-problems and optimize them in parallel to adapt to heterogeneous cloud resource allocation scenarios. However, these algorithms have high computational complexity, insufficient convergence efficiency under strict deadline constraints, and no targeted optimization mechanism designed for the temporal correlation and dynamic fluctuation of network flows, making it difficult to meet the requirements of real-time scheduling and dynamic resource adaptation.

[0011] In summary, existing technologies generally suffer from problems such as "weak ability to balance multiple constraints, easy to get trapped in local optima, insufficient QoS guarantee, and limited ability to adapt to dynamic traffic characteristics," making it difficult to efficiently solve the multi-objective collaborative scheduling needs of network flows in heterogeneous cloud environments. Summary of the Invention

[0012] The purpose of this invention is to address the challenges of complex task dependencies, difficult multi-objective optimization, and the tendency of traditional algorithms to get trapped in local optima in network flow scheduling under heterogeneous cloud computing environments. This invention provides a multi-constraint scheduling method for network flows in heterogeneous cloud environments based on Immune Mechanism-based Particle Swarm Optimization (IMPSO). By mapping the scheduling scheme to particles (antibodies) through a reasonable encoding pattern, a multi-objective fitness function that balances cost and time is designed. Antibody concentration and incentive mechanisms are introduced to balance population diversity and convergence efficiency. The search process is optimized through operations such as cloning, crossover mutation, and memory cell updates. To verify the effectiveness of IMPSO, this invention uses three typical network flows—HEFT, IC-PSO, and MOEA / D-DAG—as research objects, and conducts comparative experiments with traditional PSO and IC-PCP algorithms in a simulated Baidu cloud computing environment. Experimental results show that IMPSO outperforms in terms of deadline satisfaction, completion time optimization, and convergence speed. Under user-defined deadline constraints, it can achieve higher task execution efficiency with controllable cost increments. It provides an efficient and feasible solution for multi-constraint, multi-objective network flow scheduling in cloud computing environments, and has important practical value for improving cloud resource utilization efficiency and reducing user operating costs.

[0013] To achieve the above objectives, the technical solution of the present invention is: a multi-constraint scheduling method for network flows in a heterogeneous cloud environment based on immune particle swarm optimization, comprising:

[0014] Obtain the network flow to be scheduled, wherein the network flow is represented as a directed acyclic graph. Where T is the set of n tasks. E is the set of dependencies between tasks. Dependency Indicates task and tasks Dependency constraints, and add virtual entry tasks to the network flow. and export mission ;

[0015] Obtain cloud node resource set Each cloud node With computing power Rental price per unit time And define the unit data transmission price trans and the average bandwidth between nodes bw;

[0016] Define scheduling scheme ,in It is the set of resources used by the scheduling solution, for each cloud node. It has a corresponding cloud node type, the earliest start time (LST), and the earliest end time (LFT); A is a set of mappings between tasks and cloud nodes, each mapping... Indicates task Assigned to cloud nodes From the start time Start to end time Finish;

[0017] The scheduling scheme is encoded as an N-dimensional particle, where N is the total number of tasks in the network flow, and the coordinate value of each dimension of the particle is mapped to a specific cloud node type.

[0018] The encoded scheduling scheme is used as antibody-particle input to the particle swarm optimization (IMPSO) algorithm for fusion immune mechanism for iterative optimization.

[0019] Output the global optimal scheduling scheme globalBestSol obtained after iterative optimization of the IMPSO algorithm, including the task-cloud node mapping relationship, cloud node leasing parameters, total execution cost, and total completion time.

[0020] Furthermore, before encoding the scheduling scheme as N-dimensional particles, the fitness function is also constructed. The formula used to evaluate the quality of a scheduling scheme is as follows:

[0021]

[0022] in, and These are fixed parameters: COST is the total execution cost of the scheduling plan, and Makespan is the total completion time of the scheduling plan. The distance between the current i-th particle and the best particle in the current iteration is represented by the following formula:

[0023]

[0024] in, gbest and gbest are N-dimensional particles and the globally optimal particle, respectively.

[0025] Furthermore, the total execution cost (COST) and total completion time (Makespan) of the scheduling scheme are calculated using the following formulas:

[0026] Task The execution time on a certain cloud node is:

[0027] Task The execution cost on a certain cloud node is:

[0028] Task With his father's mission The data transmission time between them is:

[0029] Task The total cost of incoming data is:

[0030] The total cost of the scheduling scheme is:

[0031] Task start time and end time for:

[0032]

[0033]

[0034] The total completion time is the end time of the export task: .

[0035] Furthermore, the IMPSO algorithm, during iterative optimization, includes calculating antibody affinity, antibody concentration, and activation level.

[0036] Antibody With antibodies Affinity between The similarity is represented by the following formula:

[0037]

[0038] Where non represents antibody and The Euclidean distance between them;

[0039] Antibody concentration Defined using affinity, the calculation formula is:

[0040]

[0041] in,

[0042]

[0043] It is the antibody similarity coefficient;

[0044] Antibody Incentive level The calculation formula is:

[0045] .

[0046] Furthermore, the IMPSO algorithm includes immune selection, antibody cloning, crossover mutation, memory cell update, and population refresh operations.

[0047] Furthermore, the IMPSO algorithm also includes a memory unit update operation. After each iteration, the M antibodies with the highest excitation intensity are selected from the current population and stored in the memory unit according to a preset ratio. The capacity of the memory unit is 1 / 2 to 2 / 3 of the population size.

[0048] Furthermore, the IMPSO algorithm also includes a population refresh operation, in which d new antibodies are randomly generated after each iteration to replace the d antibodies with the worst motivation in the current population.

[0049] This invention also provides a multi-constraint scheduling system for network flows in a heterogeneous cloud environment based on immune particle swarm optimization, used to implement the above-described method, the system comprising:

[0050] The network flow and resource acquisition module is used to acquire the directed acyclic graph of network flows to be scheduled. and cloud node resource collection and its parameters, where T is the set of n tasks. E is the set of dependencies between tasks. Dependency Indicates task and tasks Dependency constraints, each cloud node With computing power Rental price per unit time ;

[0051] The encoding module is used to encode the scheduling scheme. Encoded as N-dimensional particles, with particle dimension coordinates mapped to cloud node types;

[0052] The IMPSO algorithm module for immune particle swarm optimization includes:

[0053] Fitness calculation unit, used to calculate the fitness of each antibody-particle. Total cost and total completion time (Makespan);

[0054] The immune optimization unit is used to calculate the affinity between antibodies, antibody concentration and activation level, and to perform immune selection, antibody cloning and cross-mutation operations based on the calculation results;

[0055] The memory and update unit is used to manage the updating of the memory cell bank and perform population refresh operations;

[0056] The scheduling scheme output module is used to output the globally optimal scheduling scheme globalBestSol obtained by the IMPSO algorithm module.

[0057] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the heterogeneous cloud environment network flow scheduling method based on immune particle swarm optimization as described above.

[0058] The present invention also provides a computer-readable storage medium storing computer program instructions that can be executed by a processor. When the processor executes the computer program instructions, it can implement the heterogeneous cloud environment network flow scheduling method based on immune particle swarm optimization as described above.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. This invention introduces an artificial immune mechanism and designs an antibody concentration regulation and incentive mechanism, which can quantify and suppress the excessive proliferation of similar scheduling schemes (antibodies) in the population. At the same time, it actively introduces diversity through operations such as cloning mutation and population refresh, which fundamentally solves the defect of the traditional PSO algorithm that is prone to getting trapped in local optima due to particle aggregation. It allows for more thorough exploration in the solution space, thereby finding a higher quality global or near-global optimal scheduling scheme.

[0061] 2. This invention constructs a multi-objective fitness function that integrates execution cost, completion time, and deadline satisfaction, directly incorporating user QoS constraints into the optimization objective. By fusing immune mechanisms with particle swarm optimization, it can effectively and synergistically optimize cost and time while satisfying strict deadlines (e.g., under the constraint ε < 0.002). Experiments show that, compared to comparative algorithms, the method of this invention can achieve an average reduction in completion time of approximately 22% with a reasonable cost increase of approximately 10%, achieving a better balance between cost and efficiency.

[0062] 3. The immune memory mechanism introduced in this invention can retain high-quality scheduling schemes during the iteration process and utilize them in subsequent optimizations, thus accelerating the algorithm's convergence process. Experimental convergence curve comparisons show that the method of this invention (IMPSO) can converge to a stable optimal solution faster than the traditional PSO algorithm on different types of network flow tasks, making it more suitable for dynamic or real-time scheduling scenarios that require rapid response.

[0063] 4. The coding mode designed in this invention can flexibly map tasks to heterogeneous cloud node types. The parameters in the immune mechanism (such as antibody similarity coefficient, memory unit capacity, population refresh ratio, etc.) can be adjusted according to the characteristics of different network flow tasks (such as heterogeneity and dynamic fluctuation), which enhances the algorithm's adaptability to different application scenarios and its overall robustness.

[0064] 5. This invention combines the intelligent recognition, memory, and regulation mechanisms of the biological immune system with the swarm optimization capability of particle swarm optimization, and creatively applies it to the complex problem of heterogeneous cloud network flow scheduling. It provides a new, efficient, and feasible technical path for solving the long-standing problem of multi-objective balance and local optima in this field, and has significant practical value for improving the overall utilization efficiency of cloud resources and reducing user operating costs. Attached Figure Description

[0065] Figure 1 This is the IMPSO scheduling model.

[0066] Figure 2 Optimize the directed acyclic graph for network flow.

[0067] Figure 3 This is the IMPSO flowchart.

[0068] Figure 4 To change This causes the network flow cutoff time to change.

[0069] Figure 5 The completion times are for the three network flows under different methods.

[0070] Figure 6 The results show the convergence rates of PSO and IMPSO. Detailed Implementation

[0071] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0072] This invention provides a multi-constraint scheduling method for network flows in a heterogeneous cloud environment based on immune particle swarm optimization, comprising:

[0073] Obtain the network flow to be scheduled, wherein the network flow is represented as a directed acyclic graph. Where T is the set of n tasks. E is the set of dependencies between tasks. Dependency Indicates task and tasks Dependency constraints, and add virtual entry tasks to the network flow. and export mission ;

[0074] Obtain cloud node resource set Each cloud node With computing power Rental price per unit time And define the unit data transmission price trans and the average bandwidth between nodes bw;

[0075] Define scheduling scheme ,in It is the set of resources used by the scheduling solution, for each cloud node. It has a corresponding cloud node type, the earliest start time (LST), and the earliest end time (LFT); A is a set of mappings between tasks and cloud nodes, each mapping... Indicates task Assigned to cloud nodes From the start time Start to end time Finish;

[0076] The scheduling scheme is encoded as an N-dimensional particle, where N is the total number of tasks in the network flow, and the coordinate value of each dimension of the particle is mapped to a specific cloud node type.

[0077] The encoded scheduling scheme is used as antibody-particle input to the particle swarm optimization (IMPSO) algorithm for fusion immune mechanism for iterative optimization.

[0078] Output the global optimal scheduling scheme globalBestSol obtained after iterative optimization of the IMPSO algorithm, including the task-cloud node mapping relationship, cloud node leasing parameters, total execution cost, and total completion time.

[0079] This invention also provides a multi-constraint scheduling system for network flows in a heterogeneous cloud environment based on immune particle swarm optimization, used to implement the above-described method, the system comprising:

[0080] The network flow and resource acquisition module is used to acquire the directed acyclic graph of network flows to be scheduled. and cloud node resource collection and its parameters, where T is the set of n tasks. E is the set of dependencies between tasks. Dependency Indicates task and tasks Dependency constraints, each cloud node With computing power Rental price per unit time ;

[0081] The encoding module is used to encode the scheduling scheme. Encoded as N-dimensional particles, with particle dimension coordinates mapped to cloud node types;

[0082] The IMPSO algorithm module for immune particle swarm optimization includes:

[0083] Fitness calculation unit, used to calculate the fitness of each antibody-particle. Total cost and total completion time (Makespan);

[0084] The immune optimization unit is used to calculate the affinity between antibodies, antibody concentration and activation level, and to perform immune selection, antibody cloning and cross-mutation operations based on the calculation results;

[0085] The memory and update unit is used to manage the updating of the memory cell bank and perform population refresh operations;

[0086] The scheduling scheme output module is used to output the globally optimal scheduling scheme globalBestSol obtained by the IMPSO algorithm module.

[0087] The following is a detailed implementation process of the present invention.

[0088] This invention addresses the challenges of complex task dependencies, difficult multi-objective optimization, and the tendency of traditional algorithms to get trapped in local optima in network flow scheduling within heterogeneous cloud computing environments. It proposes a multi-constraint scheduling method for network flows in heterogeneous cloud environments based on Immune Particle Swarm Optimization (IMPSO). Network flows are abstractly described using a directed acyclic graph (DAG), and scheduling must achieve dual optimization of execution cost and completion time while satisfying user deadline constraints. IMPSO maps the scheduling scheme to "antibody-particle," associating tasks with cloud node types through N-dimensional encoding. It designs a multi-objective fitness function that balances cost and time, introducing mechanisms such as antibody concentration regulation, clonal mutation, and immune memory to balance population diversity and optimization efficiency, thus overcoming the population aggregation deficiency of traditional Particle Swarm Optimization (PSO) algorithms. Experiments are conducted using three typical network flows—HEFT, IC-PSO, and MOEA / D-DAG—in a simulated Baidu Cloud environment for comparative verification. The results show that IMPSO performs better in terms of deadline satisfaction and can stably output feasible solutions under the strict constraint of ε<0.002. The completion time is reduced by an average of 22% compared with the comparison algorithm, and the convergence speed is significantly improved. It provides an efficient and feasible solution for multi-constraint and multi-objective network flow scheduling in cloud computing environment.

[0089] The IMPSO scheduling model primarily considers the high concurrency of multiple network requests in cloud environments. It uses multiple network flows (including multiple sets of network flows containing task sets and throughput) as input, and encapsulates the task-cloud node mapping strategy into IMPSO antibodies (particles, i.e., candidate scheduling schemes) through encoding patterns combined with cloud node attributes. Subsequently, the network flow scheduling module relies on an immune mechanism (using cost-time fitness as affinity, performing immune selection, clonal mutation, etc.) and particle swarm optimization logic to generate an optimized scheduling scheme. This scheme is deployed to cloud servers to provide cloud services. Simultaneously, through a "population update" link, high-quality or new particles are replenished to the encoding layer, balancing population diversity and convergence efficiency, forming a closed-loop scheduling logic of "input-encoding-optimization-execution-iteration" to improve the QoS of cloud services. The IMPSO scheduling model is as follows: Figure 1 As shown.

[0090] 1. Model Definition

[0091] For ease of description, such as Figure 2 As shown, network flow is typically defined as a directed acyclic graph. Where T is the set of n tasks. Each task is an independent and indivisible entity, and each has a corresponding workload. E is the set of dependencies between tasks. Dependency Indicates task and tasks Dependency constraints, which mean the task We must wait until the mission Execution can only begin after all previous executions are complete. If there is data transfer between them, the task... We need to wait for the data transmission to complete before we can begin. Therefore, the task... Called a task The parent or predecessor task, the task Called a task The child task or the subsequent task. Used to represent tasks The set of parent tasks, Indicates task A DAG is a set of subtasks. Each task can have 0, 1, or more parent tasks. In a DAG, tasks without parent tasks are called entry tasks, and tasks without subtasks are called ending tasks. , To facilitate the design of scheduling algorithms, a virtual task is typically added at the beginning and end of the network flow. and Their execution time and data transmission time are both 0.

[0092] In daily network operations, operator P provides users with "unlimited" bandwidth with varying capacity, price, and other parameters. .price Represents cloud nodes The fee is charged for a given time interval. Users are charged based on the number of time intervals they use, with any fractional intervals rounded up to the nearest whole time interval. Used to represent the processing power of a cloud node.

[0093] When the task When assigned to a cloud node, the execution time can be calculated using the following formula:

[0094] (1)

[0095] Task Execution cost

[0096] (2)

[0097] In addition, if the task It is a task The parent task, there is between them Data is sent from the parent task to the child task. This invention assumes that all cloud nodes are provided by operators in the same region, therefore, the bandwidth bw between all cloud nodes is approximately equal. If tasks are assigned to the same cloud node, the transmission time between them is 0. Equation (3) represents the transmission time between tasks.

[0098] (3)

[0099] For the task In summary, the cost of all incoming data is as follows:

[0100] (4)

[0101] Where trans is the price per unit of transmission.

[0102] The total cost of the entire network flow can be calculated using the following formula.

[0103] (5)

[0104] make For the task The start time, The task's end time is calculated using formulas (6) and (7) respectively.

[0105] (6)

[0106] (7)

[0107] For a network flow, ending the task End time This is the completion time of the entire network flow, as shown in formula (8). Furthermore, it constitutes the completion time of the task. All tasks on the path with the longest end time constitute the critical path of the network flow.

[0108] (8)

[0109] 2. Problem Modeling

[0110] In scheduling network flows in a cloud computing environment, two issues are considered: bandwidth allocation and the actual scheduling or mapping between cloud nodes. The first is bandwidth allocation, which means the scheduling solution needs to determine how many cloud node leases to use, their types, and their start and end times. The second is mapping each task to the most suitable cloud node resource. Typical network flow scheduling problems are a combination of these two issues.

[0111] A complete schedule can be represented as . It is the set of resources used by the scheduling solution, for each cloud node. Each task has a corresponding cloud node type, earliest start time (LST), and earliest end time (LFT). A is a set of mappings between tasks and cloud nodes. Each mapping... Indicates task Assigned to cloud nodes ,from Beginning, to End. The total cost and total completion time of the network flow can be calculated using formulas (5) to (8). Let D represent the user-defined deadline. The cost and completion time optimization problem of network flow scheduling under deadline constraints in a cloud computing environment can be modeled as follows:

[0112] min COST, Makespan (9)

[0113] st Makespan≤D

[0114] 3. Particle Swarm Optimization Algorithm Based on Immune Mechanism

[0115] 3.1 Overview of Particle Swarm Optimization Algorithm

[0116] Particle Swarm Optimization (PSO) is an evolutionary computational technique derived from animal behavior simulation. Since its inception in 1995, it has been widely researched and applied in various fields such as function optimization, engineering design, and resource scheduling due to its advantages of simple principle, strong robustness, and fast convergence speed. The core inspiration for this algorithm comes from the foraging behavior of flocks of birds in nature. Its modeling logic is intuitive and easy to understand: there is a unique food source in a specific area. Although each bird in the flock does not know the exact location of the food, it knows its own distance from the food source and has the ability to fly autonomously. During foraging, each bird refers to its previously found best location and moves towards the optimal location already discovered by the group. Through the synergy of individual experience and collective wisdom, they gradually approach the food source. This collaborative optimization mechanism is the core idea of ​​PSO.

[0117] In the algorithm model, the "bird flock" corresponds to the set of candidate solutions (population) for the optimization problem. Each "bird" is abstracted as a "particle," and each particle is an individual in the N-dimensional search space, with its position vector directly corresponding to a set of candidate solutions to the optimization problem. The core attributes of a particle include velocity and position: velocity is a vector that defines both the speed (amplitude) and direction of the particle's movement, determining its movement trend in the search space; position represents the particle's specific coordinates in the N-dimensional coordinate system, corresponding to the specific value of a candidate solution. The particle's flight process is essentially the search process of the optimization problem. Its velocity is not fixed but dynamically adjusted based on two key pieces of information—first, the particle's own historical optimal position (individual optimal); and second, the optimal position of the entire population so far (global optimal). In each iteration, the algorithm calculates the fitness value of all particles, selects the current globally optimal particle, and then all particles update their velocity and position based on the individual optimal and global optimal information. Through multiple iterations, the globally optimal solution is continuously updated, gradually converging to the optimal or near-optimal solution that satisfies the problem constraints. The particle swarm optimization algorithm uses a learning factor and inertia weight to dynamically adjust the particle velocity. Equations (10) and (11) represent the particle velocity and position. and These are the velocity and position of particle i at time k, respectively; both are vectors. and These are the optimal positions for the i-th particle and the global particle, respectively.

[0118] (10)

[0119] (11)

[0120] The optimization process of PSO is as follows: ① First, the population is initialized and each particle is given a randomly generated speed and position, with the best one being the current global optimal particle; ② Next, the speed and position for the next move are calculated according to formulas (10) and (11); ③ Each particle calculates its fitness value, evaluates the fitness value of each particle, and updates its historical optimal position; ④ Updates the global optimal position of the population; ⑤ Determines whether the termination condition is met, and if not, continues the loop from ② to ④. In the process of particle swarm optimization, if a particle finds a current optimal position, another particle quickly moves closer to it. Aggregation is likely to occur during the optimization process, leading to a decrease in population diversity. If the current optimal position is a local optimum, the particle swarm optimization algorithm will find it difficult to search again in the solution space and may get stuck in a local optimum. This work introduces an immune mechanism into the particle swarm algorithm when performing network flow scheduling in a cloud computing environment, thereby optimizing the problems existing in the particle swarm algorithm and accelerating the optimization speed.

[0121] 3.2 Overview of Immune Mechanisms

[0122] The biological immune system is a complex defense system composed of immune organs, immune cells, and immune molecules. Its core functions are to accurately identify "self" and "non-self" antigens, initiate specific immune responses, and maintain homeostasis. It possesses three core characteristics: antigen recognition, immune memory, and dynamic regulation. This biological mechanism provides key inspiration for addressing the weakness of Particle Swarm Optimization (PSO) algorithms in network flow scheduling, which are prone to getting trapped in local optima. Artificial immune systems, by simulating the core logic of this mechanism, have become the core support for the IMPSO algorithm in optimizing network flow scheduling in heterogeneous cloud environments.

[0123] In the design of the IMPSO algorithm, the artificial immune system and the network flow scheduling problem form a deeply adapted mapping relationship: the multi-constraint scheduling problem of network flow in the heterogeneous cloud environment to be solved corresponds to the antigen in the biological immune system, and its core requirement (satisfying the dual optimization of cost and completion time under the deadline constraint) constitutes the "immune stimulation signal" of the antigen; the "candidate scheduling scheme" (i.e., the task-cloud node mapping strategy after particle encoding) in the algorithm corresponds to the "antibody", and each antibody carries specific resource allocation logic and scheduling path design; "affinity" is used as the core indicator for quantifying antibody quality. In IMPSO, it is refined into two dimensions: the affinity between antibody and antigen is calculated through a multi-objective fitness function, which directly reflects the degree to which the candidate scheme meets the cost and time constraints; the affinity between antibodies is quantified through Euclidean distance, which characterizes the similarity of different scheduling schemes and provides a basis for population diversity regulation.

[0124] The three core characteristics of the immune mechanism are specifically applied in IMPSO: First, the antigen recognition characteristic corresponds to the fitness calculation logic of IMPSO. The algorithm quickly selects high-quality antibodies (efficient scheduling schemes) by accurately quantifying the fit between candidate schemes and scheduling targets, avoiding ineffective searches. Second, the immune memory characteristic is reflected in IMPSO through the construction and updating of the memory cell bank. The algorithm retains the scheduling scheme with the best affinity during the iteration process in the memory unit, which can be directly called in subsequent iterations. This not only speeds up the convergence speed but also ensures the stability of critical path task scheduling, especially adapting to the strong temporal correlation of network flows. Third, the dynamic regulation characteristic is achieved through antibody concentration regulation and population refresh mechanism. IMPSO effectively maintains population diversity by inhibiting the excessive proliferation of high-concentration similar antibodies and introducing random new antibodies to replace inferior schemes. This breaks the local optimum dilemma caused by particle aggregation in traditional PSO, enabling it to better cope with complex characteristics such as network flow heterogeneity and dynamic fluctuations. The core correspondence between the biological immune system and the IMPSO algorithm (in conjunction with network flow scheduling scenarios) is shown in Table 1.

[0125] Table 1. Core Correspondence between Biological Immune System and IMPSO Algorithm

[0126]

[0127] The introduction of the immune mechanism enables the IMPSO algorithm to retain the advantage of fast convergence speed of PSO, and significantly improve its global optimization ability in complex network flow scheduling scenarios through population diversity regulation and high-quality solution memory, providing key technical support for multi-constraint, multi-objective heterogeneous cloud network flow scheduling.

[0128] 3.3 Particle Swarm Optimization Algorithm for Immune Mechanisms

[0129] The IMPSO algorithm flowchart is as follows: Figure 3 As shown in Table 2, particles are treated as antibodies in this algorithm. After initializing the antibody population and obtaining the initial global optimum, each antibody has affinity and concentration. Based on this, an incentive value for evaluating the antibody is given. This is followed by a series of optimization and diversity-enhancing operations, such as antibody cloning, crossover mutation, memory cell formation, and replacing the worst-case solution. In the final stage, the population is updated, and the algorithm is checked to see if it meets the termination condition. If it does, the result is output; otherwise, the iteration continues.

[0130] Table 2 IMPSO Algorithm

[0131]

[0132] (1) Encoding Mode.

[0133] When modeling the network flow scheduling problem in a cloud computing environment as a solution model using the IMPSO method, two problems arise. First, how to define the encoding scheme for this problem? Second, how to evaluate the excellence of the particles, i.e., the fitness function?

[0134] A particle represents a scheduling solution. Therefore, the dimension of a particle is equal to the number of tasks in a given network flow. Assuming the algorithm comprises N tasks, the particle in the algorithm is an N-dimensional particle. The particle's size determines its position in the search space coordinate system. Since cloud computing involves an infinite number of cloud nodes, it's impossible for the particle's coordinates to take arbitrary values ​​within the real-world range, as this hinders fast searching. Therefore, the particle's movement range is determined by the available cloud node types. Assuming there are M cloud node types, the maximum value of the particle's position is M. Based on this, the integer part of the particle's coordinate value represents the selected cloud node type. The fitness function, used to determine the validity of candidate solutions, needs to reflect the objectives of the scheduling problem: cost and time.

[0135] (2) Affinity Calculation.

[0136] The affinity between antibody and antigen indicates that the antibody has the ability to resolve the antigen. Since the scheduling problem we aim to solve in this chapter is a multi-objective optimization problem, the antigen corresponds to the objective function and constraints being optimized. The antibody is a candidate solution. Therefore, the affinity between antibody and antigen is expressed as:

[0137] (12)

[0138] in, and These are fixed parameters. The distance between the current i-th particle and the best particle in the current iteration is represented by the following formula.

[0139] (13)

[0140] in, Let gbest be an N-dimensional particle and the globally optimal particle. It can be seen that the more similar a particle is to the globally optimal particle, the smaller their distance. The smaller the distance, the greater the affinity. In the affinity calculation process, to accelerate the algorithm's convergence and ensure particles move in better directions, this chapter stipulates that the individual with the highest affinity directly enters the next generation without subsequent cloning and mutation operations.

[0141] The affinity between particles represents their similarity, as shown in Equation (14). "non" indicates antibody. and The Euclidean distance between particles. The smaller the distance between particles, the greater their affinity, indicating that they are more similar. Too many similar particles can cause the population to easily get trapped in local optima, so appropriate measures are needed to control the concentration of similar particles.

[0142] (14)

[0143] (3) Antibody concentration and incentive function.

[0144] When a large number of similar particles exist in a population, it is easy to get trapped in local optimization. This needs to be suppressed, and the concentration of the antibody is defined by its affinity, as shown in Equation (15). The total dimension of the particles is N.

[0145] (15)

[0146] in

[0147] (16)

[0148] in, This is the antibody similarity coefficient, a constant. It is known that antibodies with high affinity and low concentration are more desirable. Particle diversity and immature convergence are also achievable. Therefore, the antibody excitation degree is defined as...

[0149] (17)

[0150] (4) Antibody cloning, crossover and mutation.

[0151] Based on the degree of motivation, the number of clones of each individual in a population can be defined as...

[0152] (18)

[0153] The stronger the affinity between individuals, the more offspring individuals are cloned. This protects good genes and accelerates the convergence speed of the algorithm. To prevent getting trapped in local optimization, this invention employs a real-number-based intermediate crossover and mutation method to expand the population, increasing the diversity of each individual except the parent. The crossover method is an intermediate crossover method, and offspring individuals are generated according to the following formula.

[0154] (19)

[0155] in, It is by The scaling factor is generated by uniformly distributed random numbers. The value of each dimension is calculated according to formula (19), and each dimension has a new value. In addition, if the memory cell is not empty, the parent particle is selected from it; otherwise, the parent particles are selected from the entire population. After all particles have undergone crossover, a roulette wheel selection method is used to determine whether each particle needs to be mutated, where particles with low excitation are more likely to be selected.

[0156] (5) Update Population.

[0157] After all the operations, all particles are sorted according to the excitation mechanism, and the first one is selected. The particles serve as the next generation population. It can be seen that the new population consists of highly motivated individuals, which are obtained from their parent populations. A certain proportion of the best particles are selected to update the memory units. Furthermore, to filter out the worst particles, some newly generated random particles will replace them. Then, the velocities and positions of all particles are updated to obtain a new global optimum. At this point, the iteration ends.

[0158] 4. Experiment and Analysis

[0159] This invention describes the experimental environment and parameters used, and then compares the experimental results obtained by other methods.

[0160] 4.1 Experimental Setup

[0161] This invention uses three network processes for evaluation: HEFT (Heterogeneous Earliest FinishTime), IC-PSO (Improved Clone Particle Swarm Optimization), and MOEA / D-DAG (Multi-Objective Evolutionary Algorithm based on Decomposition for DAG Scheduling). Each has different concerns regarding the network resource requirements. HEFT, a classic greedy algorithm for DAG network flow scheduling in heterogeneous environments, prioritizes scheduling critical path tasks and allocates cloud node resources based on the earliest task completion time. It is simple to implement and has low computational overhead, making it a fundamental benchmark for evaluating scheduling algorithm performance. IC-PSO is an improved particle swarm optimization algorithm adapted to DAG network flow scheduling. It integrates an immune cloning mechanism with traditional PSO, optimizing both cost and completion time by maintaining population diversity. Like IMPSO, it belongs to the "PSO + immune" technical framework, highlighting the latter's innovation in coding design and optimization mechanism. MOEA / D-DAG is a multi-objective evolutionary algorithm specifically for DAG network flows. It decomposes the "cost-time-deadline" multi-objective into sub-problems for parallel optimization, adapting to heterogeneous cloud node resource allocation scenarios. It can be compared with IMPSO across frameworks, fully validating its advantages in multi-objective convergence efficiency and solution superiority.

[0162] Considering that the experimental environment of a real data center is limited by factors such as resource scheduling and access control, and it is difficult to accurately reproduce the directed acyclic graph (DAG) structure, task dependencies, and cross-node data transmission characteristics of network flows, this invention uses a simulated cloud environment to conduct experiments to ensure the reproducibility and relevance of the experiments. The core parameters of cloud nodes, such as computing power and rental costs, are set with reference to the actual product specifications of a certain cloud provider (see Table 3), and are adapted to the resource requirements of different types of network flows, such as data-intensive, CPU-intensive, and compute-intensive flows. The average network bandwidth for data transmission between cloud nodes is uniformly set to 20 Mbit / s to meet the real-time data interaction requirements caused by dependencies between network flow tasks. The time metering interval for cloud node rental is set to 1 hour, with any portion less than one metering interval billed as a complete interval, which aligns with the continuous execution characteristics of network flow tasks. In the experiment, ECUs (Elastic Compute Units) are used as a quantitative indicator of the computing power of cloud nodes to ensure that the evaluation of cloud node processing performance matches the workload characteristics and execution timing constraints of network flow tasks, providing an experimental basis that fits the network flow scheduling scenario for verifying the effectiveness of the scheduling algorithm.

[0163] Table 3 Cloud Node Type Parameters

[0164]

[0165] The experimental environment was configured as follows: Core™ i7 5.60GHz, 32GB RAM, Windows 10, Java 2 Standard Edition V1.8.0.

[0166] 4.2 Experimental Results and Analysis

[0167] The experimental results of IMPSO will be compared with those of Particle Swarm Optimization (PSO) and the Integral Clone Particle Swarm Optimization (IC-PCP), both of which are classic optimization methods widely used in network flow scheduling. All methods were run independently 250 times in each experimental scenario, and the average of the 25 valid results was used as the final comparison data to reduce the impact of random factors on the experimental conclusions.

[0168] The core parameter settings for the Particle Swarm Optimization (PSO) algorithm refer to mainstream configuration standards in the field: the learning factor and the inertia factor for velocity updates in PSO are set as follows: , , The initial value of the inertia factor for speed updates is set to 0.9, which decreases linearly to 0.4 as the iteration progresses, balancing global exploration and local development capabilities. The population size and number of iterations are uniformly set to 100 to ensure that different algorithms can be compared under the same computational overhead. Based on the research characteristics of immune mechanisms and multiple sets of pre-experiment verification, the immune-related parameters in IMPSO are configured as follows: the memory unit capacity is set to half the population size to retain high-quality scheduling schemes; the number of randomly generated new particles is 1 / 10 of the population to maintain population diversity; the clone size is controlled at about twice the population size to ensure the inheritance of high-quality genes while avoiding excessive increase in computational overhead.

[0169] Furthermore, to objectively evaluate the performance boundaries of the scheduling algorithm in a heterogeneous cloud environment, two extreme scheduling scenarios were designed as benchmarks. Both scenarios used a single cloud node and combined with the HEFT algorithm to generate scheduling schemes: The first scenario involved all tasks selecting the cloud node type with the lowest cost and running on the same node. This scheme aimed for the lowest execution cost but had the longest completion time, and was denoted as "The slowest scheduling." The second scenario involved all tasks selecting the cloud node type with the strongest computing power and deploying and running independently. This scheme aimed for the shortest completion time but had the highest execution cost, and was denoted as "The fastest scheduling." The results of the two extreme scenarios were used to define the reasonable range of deadline values. The network flow size selected in the experiment was uniformly set to 50 tasks to ensure consistency in the comparison scenarios.

[0170] To evaluate the degree to which IMPSO meets deadlines within a certain range, a varying deadline D is used for verification. The definition is as follows: To control the looseness of the deadline, as shown in formula (20), where This is the completion time obtained by the fastest Scheduling. This is the completion time obtained by The slowest Scheduling.

[0171] (20)

[0172] make The value changes from 0.001 to 0.05, increasing by 0.001 each time. Figure 4 As shown, IMPSO outperforms PSO for all network flows. When the value is less than 0.002, IMPSO may miss a solution during a single optimization round, while PSO... This problem occurs when the value is less than 0.006. ICPCP exhibits inconsistent performance across different network flows, performing worst in HEFT, and when... Increasing the value from 0.006 to 0.034 may result in the inability to find a solution that meets the requirements. For network flow IC-PSO, all methods find solutions at each stage.

[0173] Regarding the scheduling objective of completion time, the operational results of different network flows are as follows: Figure 5 As shown in the figure, for each network flow, the experiment compared the completion times obtained by the three methods. Since the IC-PCP algorithm yielded basically consistent results each time, the average of 25 runs was taken and compared with IMPSO and PSO. It can be seen that IMPSO outperforms IC-PCP and PSO in most cases. The difference in running time for the same method is caused by the different combinations of solutions sought by the heuristic algorithm each time. When the completion time is relatively long, the cost is correspondingly low. That is, a low-cost, high-completion-time solution has been found.

[0174] In terms of execution cost, IMPSO's cost advantage is not particularly prominent; in most experimental scenarios, its execution cost is no lower than that of PSO and IC-PCP algorithms. The core reason for this phenomenon is that when designing the affinity calculation model, IMPSO treats cost and completion time as equally important optimization objectives. The algorithm's optimization direction tends to be a balance between cost-effectiveness and time efficiency. Therefore, it achieves an average 22% reduction in completion time by exchanging approximately 10% of the cost increase. This "efficiency-first, cost-controllable" optimization result has high acceptance in most practical application scenarios.

[0175] In terms of convergence speed, IMPSO exhibits a significant advantage. Figure 6 compares the convergence curves of PSO and IMPSO algorithms after 50 runs each. It can be seen that, regardless of the network flow, IMPSO converges to a stable optimal solution faster: in the early stages of iteration, IMPSO quickly locks onto high-quality search directions through an immune memory mechanism, resulting in a significantly faster convergence speed than traditional PSO; as the number of iterations increases, PSO is prone to getting trapped in local optima due to decreased population diversity, and its convergence curve gradually flattens out, while IMPSO, through antibody concentration regulation and population refresh mechanisms, continuously maintains population vitality and can still slowly approach the global optimum. This characteristic gives IMPSO a significant advantage in dynamic scheduling or online real-time scheduling scenarios, enabling it to output a satisfactory scheduling scheme in a shorter time.

[0176] In summary, this invention addresses the core problems of network flow scheduling in heterogeneous cloud environments, such as the difficulty of balancing multiple constraints and the tendency of traditional algorithms to get trapped in local optima. It proposes an IMPSO algorithm that integrates immune mechanisms. This algorithm maps scheduling schemes to "antibody-particle" pairs, constructs a multi-objective fitness function that considers cost, time, and QoS, and introduces mechanisms such as immune memory, antibody concentration regulation, and clonal mutation to effectively balance population diversity and optimization efficiency, overcoming the population aggregation defect of traditional PSO algorithms. Experimental results show that IMPSO performs well under strict deadline constraints (…). Even with a value of <0.002), it can still stably output feasible solutions, achieving an average reduction of 22% in completion time with a reasonable cost increment of about 10%, and its convergence speed is significantly better than PSO and IC-PCP algorithms. At the same time, it has the ability to flexibly optimize to adapt to the heterogeneity of network flow and temporal correlation, providing an efficient and feasible solution for multi-constraint network flow scheduling in heterogeneous cloud environments. It has important practical value for improving cloud resource utilization, reducing user operating costs, and ensuring stable service quality.

[0177] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the heterogeneous cloud environment network flow scheduling method based on immune particle swarm optimization as described above.

[0178] The present invention also provides a computer-readable storage medium storing computer program instructions that can be executed by a processor. When the processor executes the computer program instructions, it can implement the heterogeneous cloud environment network flow scheduling method based on immune particle swarm optimization as described above.

[0179] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A multi-constraint scheduling method for network flows in a heterogeneous cloud environment based on immune particle swarm optimization, characterized in that, include: Obtain the network flow to be scheduled, wherein the network flow is represented as a directed acyclic graph. Where T is the set of n tasks. E is the set of dependencies between tasks. Dependency Indicates task and tasks Dependency constraints, and add virtual entry tasks to the network flow. and export mission ; Obtain cloud node resource set Each cloud node With computing power Rental price per unit time And define the unit data transmission price trans and the average bandwidth between nodes bw; Define scheduling scheme ,in It is the set of resources used by the scheduling solution, for each cloud node. It has a corresponding cloud node type, the earliest start time (LST), and the earliest end time (LFT); A is a set of mappings between tasks and cloud nodes, each mapping... Indicates task Assigned to cloud nodes From the start time Start to end time Finish; The scheduling scheme is encoded as an N-dimensional particle, where N is the total number of tasks in the network flow, and the coordinate value of each dimension of the particle is mapped to a specific cloud node type. The encoded scheduling scheme is used as antibody-particle input to the particle swarm optimization (IMPSO) algorithm for fusion immune mechanism for iterative optimization. Output the global optimal scheduling scheme globalBestSol obtained after iterative optimization of the IMPSO algorithm, including the task-cloud node mapping relationship, cloud node leasing parameters, total execution cost, and total completion time.

2. The multi-constraint scheduling method for network flows in a heterogeneous cloud environment based on immune particle swarm optimization as described in claim 1, characterized in that, Before encoding the scheduling scheme as N-dimensional particles, the fitness function is also constructed. The formula used to evaluate the quality of a scheduling scheme is as follows: in, and These are fixed parameters: COST is the total execution cost of the scheduling plan, and Makespan is the total completion time of the scheduling plan. The distance between the current i-th particle and the best particle in the current iteration is represented by the following formula: in, gbest and gbest are N-dimensional particles and the globally optimal particle, respectively.

3. The multi-constraint scheduling method for network flows in a heterogeneous cloud environment based on immune particle swarm optimization as described in claim 2, characterized in that, Makespan calculates the total execution cost (COST) and total completion time of the scheduling scheme using the following formulas: Task The execution time on a certain cloud node is: Task The execution cost on a certain cloud node is: Task With his father's mission The data transmission time between them is: Task The total cost of incoming data is: The total cost of the scheduling scheme is: Task start time and end time for: The total completion time is the end time of the export task: .

4. The multi-constraint scheduling method for network flows in a heterogeneous cloud environment based on immune particle swarm optimization as described in claim 2, characterized in that, The IMPSO algorithm performs iterative optimization, including calculating antibody affinity, antibody concentration, and activation level. Antibody With antibodies Affinity between The similarity is represented by the following formula: Where non represents antibody and The Euclidean distance between them; Antibody concentration Defined using affinity, the calculation formula is: in, It is the antibody similarity coefficient; Antibody Incentive level The calculation formula is: 。 5. A multi-constraint scheduling method for network flows in a heterogeneous cloud environment based on immune particle swarm optimization, as described in claim 4, is characterized in that... The IMPSO algorithm includes immune selection, antibody cloning, crossover mutation, memory cell update, and population refresh operations.

6. The multi-constraint scheduling method for network flows in a heterogeneous cloud environment based on immune particle swarm optimization according to claim 5, characterized in that, The IMPSO algorithm also includes a memory unit update operation. After each iteration, the M antibodies with the highest excitation intensity are selected from the current population and stored in the memory unit according to a preset ratio. The capacity of the memory unit is 1 / 2 to 2 / 3 of the population size.

7. A multi-constraint scheduling method for network flows in a heterogeneous cloud environment based on immune particle swarm optimization, as described in claim 5, is characterized in that... The IMPSO algorithm also includes a population refresh operation, in which d new antibodies are randomly generated after each iteration to replace the d antibodies with the worst motivation in the current population.

8. A multi-constraint scheduling system for network flows in a heterogeneous cloud environment based on immune particle swarm optimization, characterized in that, The system for implementing the method of any one of claims 1-7 comprises: The network flow and resource acquisition module is used to acquire the directed acyclic graph of network flows to be scheduled. and cloud node resource collection and its parameters, where T is the set of n tasks. E is the set of dependencies between tasks. Dependency Indicates task and tasks Dependency constraints, each cloud node With computing power Rental price per unit time ; The encoding module is used to encode the scheduling scheme. Encoded as N-dimensional particles, with particle dimension coordinates mapped to cloud node types; The IMPSO algorithm module for immune particle swarm optimization includes: Fitness calculation unit, used to calculate the fitness of each antibody-particle. Total cost and total completion time (Makespan); The immune optimization unit is used to calculate the affinity between antibodies, antibody concentration and activation level, and to perform immune selection, antibody cloning and cross-mutation operations based on the calculation results; The memory and update unit is used to manage the updating of the memory cell bank and perform population refresh operations; The scheduling scheme output module is used to output the globally optimal scheduling scheme globalBestSol obtained by the IMPSO algorithm module.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the heterogeneous cloud environment network flow scheduling method based on immune particle swarm optimization as described in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon computer program instructions executable by a processor, wherein when the processor executes the computer program instructions, it is able to implement the heterogeneous cloud environment network flow scheduling method based on immune particle swarm optimization as described in any one of claims 1-7.

Images