A Multi-Workflow Scheduling Optimization Method Based on Containers in a Cloud-Edge Environment

Abstract

This invention discloses a multi-workflow scheduling optimization method based on containers in a cloud-edge environment, belonging to the field of intelligent technology scheduling optimization. The method includes: establishing a multi-workflow scheduling optimization model; determining the model's optimization objectives, establishing an encoding scheme, and generating initialization codes; creating a dynamic allocation algorithm for containers and virtual machines; creating an enhanced multi-objective African Vulture algorithm based on the above steps; and applying the enhanced multi-objective African Vulture algorithm to solve the model. This invention, by establishing a multi-workflow scheduling optimization model and proposing a novel enhanced multi-objective African Vulture algorithm for solving the model, effectively addresses multiple optimization objectives in a two-layer architecture composed of virtual machines and containers. The algorithm provides an appropriate encoding scheme and proposes a dynamic allocation algorithm for containers and virtual machines based on it, optimizing resource allocation and service deployment for containers and virtual machines. The algorithm's archiving mechanism and adaptive grid mechanism can solve the problems of slow convergence speed and susceptibility to local optima in heuristic algorithms.

Description

Technical Field

[0001] This invention relates to the field of intelligent computing scheduling optimization technology, and more specifically, to a multi-workflow scheduling optimization method in a container-based cloud-edge environment. Background Technology

[0002] Cloud computing resources offer scalability and flexibility, providing application service providers with greater business agility and more significant economic benefits. Edge computing is gaining popularity due to its ability to guarantee Quality of Service (QoS) for latency-sensitive applications. On the one hand, cloud servers possess powerful computing capabilities but are geographically distant and experience higher latency; on the other hand, while edge servers offer the advantage of short distances, they cannot provide sufficient computing power for certain compute-intensive microservices. Deploying microservice-based applications in edge cloud environments promises to reduce service latency. Compared to traditional architectures, the computing power provided by cloud computing and microservice architectures can shorten development cycles, simplify development, and enhance scalability by decoupling applications and adjusting service granularity.

[0003] In task scheduling, most microservice applications are described as workflows. A workflow consists of multiple tasks that follow a given execution order. Workflow scheduling maps interdependent tasks to different geographically distributed resources to run on, enabling rapid workflow completion. When managing application workflows in a cloud environment, the workflow task scheduler controller must coordinate distributed resources and schedule tasks to execute across multiple instances. When instances are insufficient to meet performance requirements, the algorithm must leverage cloud resources to create new instances to achieve a trade-off between performance and cost. Specifically, containers, due to their fast startup speed and low overhead, well support the deployment of microservice instances in a microservice architecture. Virtual machines, with their strong isolation capabilities, can accommodate containers in security-critical scenarios, forming a two-tier architecture. Cloud providers can offer computing resources through virtual machines with different configurations.

[0004] However, having multiple virtual machines and containers with various configurations can complicate workflow task scheduling. First, with a two-tiered resource architecture, it's necessary to consider resource allocation at both the container and virtual machine levels. Furthermore, resource allocation at both levels is constrained by resource vectors. Second, unbalanced resource utilization among virtual machines can lead to system underload or overload. This can cause various system failures and significant issues related to power consumption, execution time, and machine reliability. Finally, because containers within the same virtual machine share images, newly added containers can be deployed on virtual machines with the corresponding images. This helps reduce image pull time and container startup time.

[0005] Therefore, it is necessary to consider how to fully leverage existing images to improve the overall performance of task scheduling. Scheduling multiple workflows in a two-tier architecture with virtual machines and containers is a cumbersome task. Improper scheduling can lead to low resource utilization and may increase completion time, energy consumption, and operating costs.

[0006] In cloud-edge environments, service level agreements (SLAs) must be reached between users and cloud service providers to maintain the required quality of service. It is expected that task scheduling algorithms can strike a good balance between the conflicting demands of time and cost. Efficiently scheduling concurrent workflows with multiple objectives is a fundamental problem in cloud-edge environments. Furthermore, global environmental concerns have made the high energy consumption and carbon emissions of data centers a significant issue. Renewable energy sources such as wind power have lower carbon emissions. Cloud providers can reduce carbon emissions by placing workloads in data centers that utilize more renewable energy.

[0007] Two-tier architectures based on virtual machines and containers typically involve dynamic resource allocation, load balancing, and elastic scaling. These issues are highly dynamic and complex. Existing heuristic algorithms generally suffer from the following problems: (1) Many heuristic algorithms aim to optimize a single objective, and are therefore insufficient to effectively solve multiple optimization objectives in a two-tier architecture composed of virtual machines and containers. (2) Resource allocation and service deployment in a two-tier architecture composed of virtual machines and containers are very complex. In addition, existing heuristic algorithms lack appropriate coding schemes to effectively address these challenges. (3) Some existing algorithms converge slowly when the solution space is large or the complexity is high. Furthermore, these algorithms are prone to getting trapped in local optima, thus affecting their ability to identify the global optimum. Summary of the Invention

[0008] The purpose of this invention is to overcome the problems of the above-mentioned background technology, namely (1) many heuristic algorithms are designed to optimize a single objective, and therefore are insufficient to effectively solve multiple optimization objectives in a two-tier architecture composed of virtual machines and containers; (2) in a two-tier architecture composed of virtual machines and containers, resource allocation and service deployment are very complex. In addition, existing heuristic algorithms lack appropriate coding schemes to effectively deal with these challenges; (3) some existing algorithms have slow convergence speed and are prone to getting trapped in local optima when the solution space is large or the complexity is high, thus affecting their ability to identify the global optimal solution. This invention provides a multi-workflow scheduling optimization method based on containers in a cloud-edge environment.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows:

[0010] This invention provides a multi-workflow scheduling optimization method based on container cloud-edge environments, comprising the following steps:

[0011] Establish a multi-workflow scheduling optimization model based on the execution relationship of workflows in a container-based cloud-edge environment;

[0012] The optimization objective is determined based on the multi-workflow scheduling optimization model, and an encoding scheme is established. Initialization codes are generated according to the encoding scheme.

[0013] A dynamic allocation algorithm for containers and virtual machines is created based on the optimization objectives and coding scheme;

[0014] An enhanced multi-objective African vulture algorithm, including an archiving mechanism and an adaptive grid mechanism, is created based on the optimization objective, the encoding scheme, and the container and virtual machine dynamic allocation algorithm.

[0015] Based on the initialization encoding, the enhanced multi-objective African vulture algorithm is applied to solve the multi-workflow scheduling optimization model.

[0016] Preferably, the multi-workflow scheduling optimization model includes: a user layer, a workflow layer, a microservice instance layer, a virtual machine layer, an edge layer, and a cloud layer;

[0017] The user layer includes multiple user devices that generate various task requests;

[0018] The workflow layer includes multiple scheduled workflows formed by task requests from multiple user devices, and each scheduled workflow includes several sub-tasks;

[0019] The microservice instance layer includes container instances, i.e. microservice instances, created on virtual machines according to the container type assigned to the subtask;

[0020] The virtual machine layer includes all virtual machines;

[0021] The edge layer is abstracted from virtual machines deployed on edge computing nodes and is used for edge computing;

[0022] The cloud layer is abstracted from virtual machines deployed on cloud computing nodes and is used for cloud computing.

[0023] Preferably, the structure of the workflow to be scheduled is represented by a directed acyclic graph, expressed by the formula: in This represents the task set of the workflow. Represents task t i and t j Priority constraints and data transmission direction between tasks are defined by a tuple for each task set. express, Represents task t i The set of all prerequisite tasks, Represents task t i The set of all subsequent tasks, d i Indicates t i To t j The size of the output data, with tasks having no preceding tasks and tasks having no succeeding tasks respectively denoted as entry task t. O and export task t F .

[0024] Preferably, the optimization objectives include workflow completion time, total workflow energy consumption, and carbon emissions;

[0025] The completion time of the workflow includes the completion time of the export task;

[0026] The total energy consumption of the workflow includes the energy consumption of the virtual machine when it is working, the energy consumption of the virtual machine when it is idle, the transmission energy consumption between two computing nodes, and the computing energy consumption of the user device.

[0027] The carbon emissions include carbon emissions from both renewable and non-renewable energy sources.

[0028] Preferably, the container-based cloud-edge environment includes one cloud computing node, multiple edge computing nodes, and multiple user computing nodes. Let m represent the set of all nodes, where m is the number of all computing nodes. All edge computing nodes and cloud computing nodes form a fully connected topology. User computing nodes are special edge computing nodes. The multi-workflow scheduling optimization model only considers the data transmission time between nodes, not the data transmission time between individual virtual machines within a node, and does not consider bandwidth capacity allocation, only the impact of bandwidth size on data transmission time. In the multi-workflow scheduling optimization model, computing resources are mainly provided by virtual machines. Each computing node can provide multiple types of virtual machines, and virtual machines are represented by a set. This indicates that h represents the number of virtual machines, and each virtual machine is configured with two types of resources to be allocated: CPU and memory.

[0029] Preferably, the encoding scheme is as follows: to express the scheduling scheme of multi-workflow tasks, containers and virtual machines are represented by two integer strings.<U,V> The encoding is performed such that the length of each string is equal to the sum of the number of subtasks in all workflows. The index of each bit in the string corresponds to each subtask. The value of each bit in U represents the container type corresponding to the subtask. The value of each element in V represents the index of the virtual machine executing the subtask. Each encoding represents a solution of the multi-workflow scheduling optimization model.

[0030] Preferably, a fragmented linear chaotic graph is used to generate the initialization code, so that the initial solution is uniformly distributed in the solution space.

[0031] Preferably, the container and virtual machine dynamic allocation algorithm first traverses all containers in the virtual machine and adds all containers that meet the judgment conditions to the list. Then, it calculates the running time of the current task under each container configuration and selects the container configuration corresponding to the earliest completion time.

[0032] Preferably, the container and virtual machine dynamic allocation algorithm uses the Topology-Topology-Simulation-Symmetry (TOPSIS) technique to select the most suitable virtual machine to meet the optimization objective.

[0033] Preferably, the enhanced multi-objective African vulture algorithm calls the container and virtual machine dynamic allocation algorithm to evaluate the optimization objective value of each vulture individual, and uses an adaptive grid mechanism to delete redundant solutions in the archive set during each iteration. After reaching the maximum number of iterations, it outputs an archive set storing the solutions of the multi-workflow scheduling optimization model.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] This invention proposes a multi-workflow scheduling optimization method based on containers in a cloud-edge environment. By establishing a multi-workflow scheduling optimization model in the cloud-edge environment, and proposing a novel enhanced multi-objective African vulture algorithm to solve the model, it effectively addresses multiple optimization objectives in a two-layer architecture composed of virtual machines and containers. In this algorithm, an appropriate encoding scheme is given, and based on this encoding scheme, a dynamic allocation algorithm for containers and virtual machines is proposed to optimize the resource allocation and service deployment of containers and virtual machines. The archiving mechanism and adaptive grid mechanism of this algorithm can, to some extent, solve the problems of slow convergence speed and easy getting trapped in local optima in heuristic algorithms. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating a multi-workflow scheduling optimization method based on containers in a cloud-edge environment, as described in this application.

[0037] Figure 2 This is a schematic diagram of a multi-workflow scheduling optimization model based on a container in a cloud-edge environment, according to Embodiment 2 of this application;

[0038] Figure 3 This is a comparison of the Pareto front results of different algorithms on Montage in Embodiment 3 of this application;

[0039] Figure 4 This is a comparison of the Pareto front results of different algorithms on CyberShake in Embodiment 3 of this application;

[0040] Figure 5 This is a comparison of the Pareto front results of different algorithms on the Inspiral platform in Embodiment 3 of this application. Detailed Implementation

[0041] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0043] Example 1

[0044] Please see Figure 1 Embodiment 1 of the present invention provides a multi-workflow scheduling optimization method based on container cloud edge environment, including the following steps:

[0045] S1: Establish a multi-workflow scheduling optimization model based on the execution relationship of workflows in a container-based cloud-edge environment.

[0046] Microservice task loading in a cloud-edge environment can be viewed as loading tasks from multiple workflows onto two layers of computing resources with different processing capabilities: containers and virtual machines.

[0047] Multiple workflows can be treated as a single workflow by adding two virtual tasks with zero execution and communication costs, and then connecting the inlet and outlet tasks of independent workflows. Multiple workflows consist of many individual workflows, using... It means that among them This represents the total number of individual workflows. The workflow structure is represented by a directed acyclic graph (DAG), and the formula is as follows: in This represents the task set of the workflow. Represents task t i and t j Priority constraints and data transmission direction between tasks. Typically, tasks are represented as nodes, and the relationships between them are represented as directed edges. Each task can be represented by a tuple. express. Represents task t i The set of all prerequisite tasks, Represents task t i The set of all subsequent tasks, d i Indicates t i To t j The size of the output data. Tasks without predecessor tasks and tasks without successor tasks are respectively represented as entry tasks t. O and export task tF .

[0048] This invention takes into account the scenario where the task runs on a user device. This represents the set of user devices, where y is the number of user device units. In this embodiment 1, the cloud-edge environment consists of one cloud computing node, multiple edge computing nodes, and multiple user computing nodes. Let m represent the set of all nodes, where m is the number of nodes. In this embodiment, all edge computing nodes and cloud computing nodes form a fully connected topology. A user node is a special edge node. We only consider the data transmission time between nodes, not the data transmission time between individual virtual machines within a node. We do not consider bandwidth capacity allocation, only the impact of bandwidth size on data transmission time. Therefore, let B represent the bandwidth between two different computing nodes. In the model, computing resources are mainly provided by virtual machines, and containers on virtual machines deploy microservice instances. Each computing node can provide multiple types of virtual machines. The set of virtual machines used by the system is... This represents the number of virtual machines, where h represents the number of virtual machines. Each virtual machine is configured with two types of resources to be allocated: CPU and memory. Vector R l,j Indicates virtual machine v l,j All available resources on the virtual machine, including occupied and remaining resources. l,j The occupied and remaining resources are represented by vectors. and Indicated. Using ms i,j This represents the j-th instance of the i-th type of microservice.

[0049] S2: Determine the optimization objective based on the multi-workflow scheduling optimization model, establish a coding scheme, and generate initialization codes according to the coding scheme.

[0050] In order to better address real-world needs, optimization objectives were determined by referencing key considerations in real-world cloud computing task scheduling.

[0051] The optimization objectives include workflow completion time, total workflow energy consumption, and carbon emissions.

[0052] The completion time of the workflow is calculated, expressed as formula (1):

[0053]

[0054] in This indicates the completion time of the workflow's exit task.

[0055] The total energy consumption of the workflow is calculated and expressed as formula (2):

[0056]

[0057] in This indicates the energy consumption of the virtual machine while it is working. This represents the energy consumption of the virtual machine when it is idle. E represents the energy consumption for transmission between two computing nodes, while E represents the computing energy consumption of the user equipment.

[0058] Carbon emissions are calculated using formula (3).

[0059] CarbonEmission = Energy × Q(τ) (3)

[0060] Where Q(τ) represents the carbon emissions per unit of different energy sources, which are divided into carbon emissions per unit of renewable energy and non-renewable energy.

[0061] The final overall optimization objective is expressed as formula (4).

[0062] Minimize:Makespan,CarbonEmission(4)

[0063] Encoding Scheme Determination: Due to the complexity of the two-layer structure consisting of containers and virtual machines in a cloud-edge environment, this invention decides to adopt the multi-objective African Vulture algorithm to achieve the optimization objective. Using the multi-objective African Vulture algorithm requires encoding the solution to the problem, but existing encoding schemes cannot fully represent the solution of the model established in this embodiment. Therefore, the encoding scheme of this invention is proposed:

[0064] Containers and virtual machines are represented by two integer strings.<U,V> Encoding is performed where the length of each string equals the sum of the number of subtasks in all workflows. Each bit in the string corresponds to a subtask, the value of each bit in U represents the container type of the subtask, and the value of each element in V represents the index of the virtual machine executing the subtask. Each encoding represents a solution to the multi-workflow scheduling optimization model. Therefore, the quality of the initial solution affects the algorithm's performance. A uniform distribution of initial solutions in the solution space is beneficial for subsequent development and exploration of the solution space (i.e., the vulture population). Therefore, a fragmented linear chaotic graph is used to generate the initial encoding, i.e., the initial solution, ensuring a uniform distribution of initial solutions in the solution space.

[0065] S3: Create a dynamic allocation algorithm for containers and virtual machines based on the optimization objectives and coding scheme.

[0066] In a cloud-edge environment with a two-tier architecture consisting of containers and virtual machines, multi-workflow scheduling presents two main challenges: First, there's the issue of container reuse and allocation. Since containers come in various categories, both image retrieval and creation take time. Reusing containers of the same category, while meeting constraints, can effectively save retrieval and creation time. However, container reuse may incur resource waiting time, necessitating dynamic container reuse. Second, in a two-tier architecture, differing resource configurations at each layer mean that initially allocated virtual machine resources may not meet container resource requirements, leading to infeasible solutions. In such cases, it's necessary to reselect resource-sufficient virtual machines for these containers. To address these two challenges in the multi-objective African Vulture algorithm, a dynamic container and virtual machine allocation algorithm is proposed as an improvement.

[0067] Deploying containers on virtual machines takes time to bring up and start the container image. Therefore, a container reuse mechanism is needed. First, the algorithm checks if a container of the same scheduling type exists in the virtual machine. If so, their task completion times are compared. This has two advantages. First, it saves container startup time. Second, from a global perspective, the algorithm considers the impact of the current container configuration on subsequent tasks when making decisions. This strikes a good balance between resource usage and computational speed. The algorithm first iterates through all containers in the virtual machine and adds all eligible containers to a list. Then, it calculates the runtime of the current task under each container configuration and selects the container configuration corresponding to the earliest completion time.

[0068] In a two-tier architecture, due to the different resource configurations of the two tiers, virtual machine resources may be insufficient compared to the resource requirements of containers. This necessitates reselecting virtual machines with sufficient resources for these containers. Furthermore, to quickly find a high-quality solution, the Top-Order Similarity Ranked Preference (TOPSIS) technique is employed to select the most suitable virtual machine to achieve the optimal optimization objective. TOPSIS is a multi-criteria decision-making method. It can find the solution closest to the ideal solution and furthest from the worst solution. First, a decision matrix is ​​defined containing h candidate virtual machines and q optimization objectives, where x... ij Let represent the j-th optimization objective value when the task is deployed to candidate i. The decision matrix is ​​represented by formula (5).

[0069]

[0070] The decision matrix is ​​then normalized, and the ideal and negative solutions are selected. The Euclidean distance between each alternative and the ideal and negative ideal solutions is calculated. Finally, the relative proximity of each alternative to the ideal solution is obtained according to formula (6). The optimal solution is selected based on this proximity.

[0071]

[0072] After the above steps, the dynamic matching algorithm for containers and virtual machines is created:

[0073] The algorithm's input parameter is the container resource requirement. Container type requirements Ctype and initial virtual machine v j A new container Cnew is created based on the input parameters, if the task is scheduled to use the total virtual machine resources R. l,j Greater than Cnew's resource requirements Then, iterate through all the containers that have been created on the virtual machine and create a container list CON_LIST. Check if the type of the container being iterated is the required container type Ctype; if so, add it to the container list CON_LIST. If the container list CON_LIST is not empty after the above iteration, it means that the current virtual machine has created containers of the same type. Then, iterate through all containers in the CON_LIST list, calculate and record the earliest completion time of the task after deployment to the corresponding container, and the corresponding container number. After the iteration, select the container with the smallest earliest completion time value as the final container deployment scheme. If the task is scheduled to use the total resources R of the virtual machine... l,j Resource requirements less than Cnew Then create a new list VM_LIST containing the total amount of all resources R. l,j Greater than container resource requirements The virtual machines are added to the list VM_LIST. The virtual machines in VM_LIST are iterated through, and the target value is calculated for each virtual machine according to formulas (1) and (3). After the iteration is completed, the TOPSIS algorithm is used to select the optimal virtual machine. Finally, the algorithm returns two values: one is the final container deployment scheme for the current task, and the other is the final virtual machine deployment scheme, which reallocates the containers and virtual machines.

[0074] The pseudocode flow of the dynamic matching algorithm for containers and virtual machines is as follows:

[0075]

[0076]

[0077] S4: Based on the optimization objective, the encoding scheme, and the container and virtual machine dynamic allocation algorithm, create an enhanced multi-objective African vulture algorithm that includes an archiving mechanism and an adaptive grid mechanism.

[0078] Based on steps S1-S3, the enhanced multi-target African vulture algorithm RMOAVOA is created:

[0079] The algorithm's inputs are: population size N, maximum iterations (maxiteration), leader selection pressure parameter beta, archive overflow member selection pressure parameter gamma, and iteration count g = 0. First, a set of candidate solutions (called "vultures") is randomly generated, with each vulture representing a possible solution. Simultaneously, parameters such as population size Vi (i = 1, 2, ..., N), maximum iterations (maxiteration), and the dimension of the objective function need to be set. Algorithm 1 (a dynamic allocation algorithm for containers and virtual machines) is called to evaluate the objective value of each vulture, obtaining non-dominated solutions, establishing an archive set, and finding the leader. Next, iteration begins. In each iteration, the optimal solution in the population, i.e., the optimal vulture, is assigned to BestVulture1; the suboptimal solution is assigned to BestVulture1. tVulture2; Then, iterate through each individual in the population, updating the hunger level of each individual. If the hunger level of an individual is greater than or equal to one, perform an exploration operation; if the hunger level of an individual is less than one, perform an exploitation operation. In one iteration, after traversing the entire population, it is necessary to evaluate the target value of all individuals and obtain all non-dominated solutions in the population, and add the newly found non-dominated solutions to the archive. If the number of solutions stored in the archive reaches the limit (the archive is full), the adaptive grid mechanism is called to delete redundant solutions in the archive and add new solutions to the archive. At the same time, if any newly added solution is outside the range of the hypercube, the grid update operation is performed. After a specified number of iterations, the algorithm finally returns the final archive.

[0080] The pseudocode flow of the enhanced multi-target African vulture algorithm is as follows:

[0081]

[0082]

[0083] S5: Based on the initialization encoding, the enhanced multi-objective African vulture algorithm is applied to solve the multi-workflow scheduling optimization model.

[0084] Example 2

[0085] This embodiment 2, based on embodiment 1, further explains the multi-workflow scheduling optimization model in a container-based cloud-edge environment, as follows:

[0086] Please see Figure 2 , Figure 2 This is a schematic diagram of a multi-workflow scheduling optimization model based on a container in a cloud-edge environment, as described in Embodiment 2.

[0087] The first layer, the user layer, is the user device layer, where user devices generate various task requests.

[0088] In this invention, each user device's task request is modeled as a workflow, and the requests from multiple user devices form multiple scheduled workflows. Each scheduled workflow includes several sub-tasks, which is the second workflow layer.

[0089] Each workflow subtask can be executed locally by the user or scheduled to be executed in a container-based cloud-edge environment. If the task is scheduled to be executed in a container-based cloud-edge environment, a container category will be assigned to each task. After the task is scheduled, a container instance (microservice instance) will be created on the virtual machine according to the container category. These container instances constitute the microservice instance layer (container layer).

[0090] Container instances are all deployed on virtual machines, and all virtual machines constitute the virtual machine layer.

[0091] Virtual machines are divided into virtual machines deployed at cloud computing nodes and virtual machines deployed at edge computing nodes, which further abstracts into the edge layer and the cloud layer.

[0092] Example 3

[0093] This embodiment 3, based on embodiment 1, applies a container-based multi-workflow scheduling optimization method in a cloud-edge environment to evaluate the performance of the enhanced multi-objective African vulture algorithm RMOAVOA of this invention, and compares it with three benchmark algorithms, as follows:

[0094] Three real-world workflow applications were selected to simulate multiple microservice workflows, including Montage, CyberShake, and Inspiral.

[0095] For the edge-cloud computing scenario discussed, a small-scale workflow was selected as the workflow instance, with one cloud computing node and seven edge computing nodes configured. In this work, the bandwidth between nodes was randomly set from {500, 800, 1000, 1500} Mbps. The communication distance latency between edge computing nodes was then set from {100, 200, 300, 400} ms. Since the cloud computing node is farther away, the communication latency between the cloud computing node and the edge computing nodes was randomly selected from {500, 600, 700, 800} ms. Users were treated as special edge nodes for latency calculation.

[0096] In addition, each computing node is equipped with green energy harvesting and storage devices. Cloud computing nodes can harvest 2 to 20 joules of green energy per second, while edge computing nodes can harvest 1 to 10 joules of green energy per second. The carbon emissions of green energy and grid energy are set at 53 and 968 grams of CO2 equivalent per kilowatt-hour, respectively. See Table 1 for details on the virtual machine configurations in the cloud-edge environment. Table 2 provides the container configuration parameters to simulate resource scaling at the container level.

[0097] Table 1: Virtual Machine Settings in Edge Environments

[0098]

[0099] Table 2: Container Settings in Edge Environments

[0100] Container Type vCPU Memory (GB) cc(.) 1 4 16 1.0 2 2 8 1.6 3 1 4 2.2 4 0.5 2 2.8 5 0.25 1 3.4

[0101] Here, cc(.) represents the execution time weight of different container rankings, with containers having more resources having a lower weight. Furthermore, the task runtime recorded by workflow tracking is considered a benchmark for task execution time. Table 3 shows the configuration parameters for RMOAVOA.

[0102] Table 3: Algorithm-related configuration parameter settings

[0103]

[0104] To measure the performance of the RMOAVOA algorithm, we compared it with the following algorithms.

[0105] The MOGWO algorithm is inspired by the social hierarchy and hunting behavior of gray wolves. This algorithm utilizes position updates and a dynamic search mechanism to efficiently converge to the optimal solution.

[0106] The MOBAN algorithm is inspired by the echolocation behavior of bats. Bats dynamically adjust their positions based on the best solution discovered so far. This algorithm combines randomness and local search techniques to effectively optimize the objective function.

[0107] The MOSPO algorithm is built upon the principles of color theory, color wheels, and color combination methods. It achieves effective exploration and utilization through four simple color combination rules, eliminating the need for internal parameters.

[0108] To more intuitively compare the performance of different algorithms, we generated different directed acyclic graph applications and ran different algorithms to generate Pareto fronts. Figures 3 to 5The comparison results of the Pareto fronts of various algorithms are presented. The x-axis represents completion time, and the y-axis represents carbon emissions. As shown in the figure, the Pareto front generated by RMOAVOA in this invention is superior to other algorithms. The RMOAVOA algorithm can find a wide range of non-dominated solutions, satisfying various requirements. Compared with the Pareto front generated by RMOAVOA, other algorithms failed to explore solutions when solving the complex structure problem presented in this paper. However, the solutions obtained by these three compared algorithms are concentrated in a specific region, indicating their diversity. Overall, the RMOAVOA algorithm proposed in this invention has better performance.

[0109] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A multi-workflow scheduling optimization method based on container cloud-edge environment, characterized in that, Includes the following steps: Establish a multi-workflow scheduling optimization model based on the execution relationship of workflows in a container-based cloud-edge environment; Based on the multi-workflow scheduling optimization model, the optimization objective is determined, and an encoding scheme is established. The encoding scheme is as follows: containers and virtual machines are represented by two integer strings.<U, V> The encoding is performed such that the length of each string is equal to the sum of the number of all subtasks in all workflows, the index of each bit in the string corresponds to each subtask, the value of each bit in U represents the container type corresponding to the subtask, and the value of each element in V represents the index of the virtual machine that executes the subtask. Generate initialization code according to the encoding scheme; A dynamic allocation algorithm for containers and virtual machines is created based on the optimization objective and coding scheme. The algorithm includes: traversing the containers already created in the virtual machine, adding containers whose type matches the current task to the candidate list, and selecting the container configuration based on the earliest completion time; if the current virtual machine resources are insufficient, the similarity ranking preference technique TOPSIS of the ideal solution is adopted to reselect a virtual machine with sufficient resources based on the optimization objective. An enhanced multi-objective African vulture algorithm is created based on the optimization objective, the coding scheme, and the container and virtual machine dynamic allocation algorithm. This algorithm invokes the container and virtual machine dynamic allocation algorithm when evaluating each individual and includes an archiving mechanism and an adaptive grid mechanism. Based on the initialization encoding, the enhanced multi-objective African vulture algorithm is applied to solve the multi-workflow scheduling optimization model.

2. The multi-workflow scheduling optimization method based on container cloud-edge environment according to claim 1, characterized in that, The multi-workflow scheduling optimization model includes: user layer, workflow layer, microservice instance layer, virtual machine layer, edge layer, and cloud layer; The user layer includes multiple user devices that generate various task requests; The workflow layer includes multiple scheduled workflows formed by task requests from multiple user devices, and each scheduled workflow includes several sub-tasks; The microservice instance layer includes container instances, i.e. microservice instances, created on virtual machines according to the container type assigned to the subtask; The virtual machine layer includes all virtual machines; The edge layer is abstracted from virtual machines deployed on edge computing nodes and is used for edge computing; The cloud layer is abstracted from virtual machines deployed on cloud computing nodes and is used for cloud computing.

3. The multi-workflow scheduling optimization method based on container cloud-edge environment according to claim 2, characterized in that, The structure of the workflow to be scheduled is represented by a directed acyclic graph, and the formula is as follows: W x = ( , ),in = { t i | i n } represents the task set of the workflow. = { e i,j = ( t i ,t j )| t i , t j } represents a task t i and t j Priority constraints and data transmission direction between them, each of the task sets is represented by a tuple { , , d i }express, Indicates task t i The set of all prerequisite tasks Indicates task t i The set of all subsequent tasks, d i express t i Towards t j The size of the output data, with tasks having no preceding tasks and tasks having no succeeding tasks respectively representing the entry task. and export mission .

4. The multi-workflow scheduling optimization method based on container cloud-edge environment according to claim 3, characterized in that, The optimization objectives include workflow completion time, total workflow energy consumption, and carbon emissions; The completion time of the workflow includes the completion time of the export task; The total energy consumption of the workflow includes the energy consumption of the virtual machine when it is working, the energy consumption of the virtual machine when it is idle, the transmission energy consumption between two computing nodes, and the computing energy consumption of the user device. The carbon emissions include carbon emissions from both renewable and non-renewable energy sources.

5. The multi-workflow scheduling optimization method based on container cloud-edge environment according to claim 1, characterized in that, The container-based cloud-edge environment includes one cloud computing node, multiple edge computing nodes, and multiple user computing nodes. = { cn l | l m } represents the set of all nodes, where m This refers to the total number of computing nodes. All edge computing nodes and cloud computing nodes form a fully connected topology. User computing nodes are special edge computing nodes. The multi-workflow scheduling optimization model only considers the data transmission time between nodes, not the data transmission time between individual virtual machines within a node, and does not consider bandwidth capacity allocation, only the impact of bandwidth size on data transmission time. In this multi-workflow scheduling optimization model, computing resources are mainly provided by virtual machines. Each computing node can provide multiple types of virtual machines, and virtual machines are used in sets. = { v j | j h } indicates that, among which h This indicates the number of virtual machines. Each virtual machine is configured with two types of resources to be allocated: CPU and memory.

6. The multi-workflow scheduling optimization method based on container cloud-edge environment according to claim 1, characterized in that, The initialization code is generated using a fragmented linear chaotic graph, so that the initial solution is uniformly distributed in the solution space.

7. The multi-workflow scheduling optimization method based on container cloud-edge environment according to claim 1, characterized in that, The dynamic allocation algorithm for containers and virtual machines first iterates through all containers in the virtual machine and adds all containers that meet the judgment criteria to the list. Then, it calculates the running time of the current task under each container configuration and selects the container configuration corresponding to the earliest completion time.

8. The multi-workflow scheduling optimization method based on container cloud-edge environment according to claim 7, characterized in that, The dynamic allocation algorithm for containers and virtual machines uses the similarity ranking preference technique TOPSIS (Topological Solution Similarity Ranking Preference) to select the most suitable virtual machine to meet the optimization objective.

9. The multi-workflow scheduling optimization method based on container cloud-edge environment according to claim 1, characterized in that, The enhanced multi-objective African vulture algorithm calls the container and virtual machine dynamic allocation algorithm to evaluate the optimization objective value of each vulture individual, and uses an adaptive grid mechanism to delete redundant solutions in the archive set during each iteration. After reaching the maximum number of iterations, it outputs an archive set storing the solutions of the multi-workflow scheduling optimization model.

Images