A Dynamic Cloud Computing Resource Scheduling Method Based on Artificial Intelligence

Abstract

This invention discloses a dynamic scheduling method for cloud computing resources based on artificial intelligence, comprising the following steps: S1, obtaining the original scheduling dataset; S2, preprocessing to generate a task cluster set and constructing a local state vector; S3, constructing a mean field state vector; S4, constructing a benefit evaluation model and establishing a mean field game model; S5, outputting candidate scheduling actions and value evaluation results; S6, updating the mapping relationship between task clusters and resource nodes and the mean field state vector, iteratively executing candidate scheduling action updates and mean field state vector updates to obtain a mean field equilibrium scheduling strategy; S7, obtaining the scheduling execution result; S8, updating the conditional generation flow network scheduling strategy model and outputting the dynamic scheduling result of cloud computing resources. This invention can realize dynamic scheduling and online optimization of cloud computing resources, enhancing the stability, adaptability, and reliability of the system under continuous operating conditions.

Description

Technical Field

[0001] This invention relates to the field of cloud computing resource management, and in particular to a dynamic scheduling method for cloud computing resources based on artificial intelligence. Background Technology

[0002] With the widespread application of cloud computing platforms in scenarios such as medical image analysis, industrial inspection, and big data processing, these platforms need to continuously handle a large number of concurrent task requests. Since different tasks have significantly different requirements for resources such as CPU, memory, storage, and network bandwidth, and the computing nodes, storage nodes, and network links typically have multi-level topologies, how to achieve efficient dynamic scheduling of cloud platform resources has become a key issue affecting task execution efficiency, service quality, and system stability.

[0003] Existing cloud resource scheduling methods mainly include fixed threshold scheduling, minimum load scheduling, and intelligent scheduling methods based on reinforcement learning or deep learning. While these methods can accomplish task allocation and resource adjustment to a certain extent, they typically rely on local load or global average state for decision-making, making it difficult to reflect the interrelationships between resource domains at different topology levels. They also lack unified modeling of the contention states of various resources, including computing, storage, network, and task queues. When tasks arrive in bursts or high-priority tasks enter the system in a concentrated manner, existing methods are prone to problems such as local resource domain overload, excessive cross-domain migrations, link congestion propagation, and increased service level agreement (SLA) default rates.

[0004] Therefore, how to provide a dynamic scheduling method for cloud computing resources based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a dynamic scheduling method for cloud computing resources based on artificial intelligence. This invention takes the mean field game model and the conditional generation flow network scheduling strategy model as its core, and combines node topology association data, task cluster local state vectors and mean field state vectors to uniformly model computing resources, storage resources, network resources and queue risks in the cloud platform, generate candidate scheduling actions and evaluate their value, and realize the dynamic scheduling and online optimization of cloud computing resources through iterative updates of the mapping relationship between task clusters and resource nodes.

[0006] According to an embodiment of the present invention, a dynamic scheduling method for cloud computing resources based on artificial intelligence includes the following steps:

[0007] S1. Collect operational monitoring data, task request data, task operation feedback data, and node topology association data of computing nodes, storage nodes, and network links in the cloud platform to form the original scheduling dataset;

[0008] S2. Perform preprocessing on the original scheduling dataset, group the tasks, generate a set of task clusters, and construct a corresponding local state vector for each task cluster.

[0009] S3. Based on the node topology association data and the original scheduling dataset, construct the mean field state vector;

[0010] S4. Map each task cluster to a corresponding representative agent. Based on the local state vector of each representative agent and the mean field state vector, define the action space, construct the profit evaluation model, and establish the mean field game model.

[0011] S5. Input the local state vectors and mean field state vectors of each representative agent into the conditions to generate a flow network scheduling strategy model, and output the candidate scheduling actions of each representative agent in the current scheduling cycle and the value evaluation results corresponding to the candidate scheduling actions.

[0012] S6. Update the mapping relationship between task clusters and resource nodes based on the candidate scheduling actions of each representative agent, and update the mean field state vector. Iterate the candidate scheduling action update and mean field state vector update to obtain the mean field balanced scheduling strategy.

[0013] S7. Output resource scheduling control instructions to the cloud platform according to the mean field equilibrium scheduling strategy to form a scheduling execution result;

[0014] S8. Construct scheduling feedback data based on the scheduling execution results, and update the conditional generation flow network scheduling strategy model online based on the scheduling feedback data, and output the dynamic scheduling results of cloud computing resources.

[0015] Optionally, S1 specifically includes:

[0016] S11. Collect the CPU utilization, memory usage, local storage read / write speed, task queue length and node energy consumption of each computing node to obtain computing node operation monitoring data;

[0017] S12. Collect the total storage capacity, remaining storage capacity, data read / write rate, and storage access latency of each storage node to obtain storage node operation monitoring data;

[0018] S13. Collect the bandwidth utilization, packet loss rate, transmission delay and delay jitter of each network link to obtain network link operation monitoring data;

[0019] S14. Collect the task arrival time, task type, task priority, CPU requirements, memory requirements, storage requirements, network bandwidth requirements, service level agreement level, task deadline and execution duration for each task to obtain task request data.

[0020] S15. Collect the task response latency and service level agreement satisfaction rate of each task to obtain task operation feedback data;

[0021] S16. Collect the connection relationships between each computing node, each storage node and each network link to obtain node topology association data;

[0022] S17. Collect computing node operation monitoring data, storage node operation monitoring data, network link operation monitoring data, task request data, task operation feedback data, and node topology association data, and organize them according to the collection time sequence to form an original scheduling dataset containing time sequence evolution information.

[0023] Optionally, S2 specifically includes:

[0024] S21. Perform outlier removal, missing value completion, and timestamp alignment on the original scheduling dataset to obtain an aligned dataset;

[0025] S22. Perform min-max normalization on the CPU requirements, memory requirements, storage requirements, network bandwidth requirements, and execution time in the aligned dataset to obtain a normalized dataset.

[0026] S23. Extract the CPU requirements, memory requirements, storage requirements, network bandwidth requirements and service level protocol level corresponding to each task from the normalized dataset, and arrange them in the order of CPU requirements, memory requirements, storage requirements, network bandwidth requirements and service level protocol level to form a task feature vector, thereby obtaining a task feature set.

[0027] S24. Based on the task feature set, cluster and group each task to obtain a task cluster set;

[0028] S25. For each task cluster, calculate the number of tasks within the cluster, average CPU requirements, average memory requirements, average storage requirements, average network bandwidth requirements, average execution time, and service level protocol distribution, and construct the corresponding local state vector.

[0029] Optionally, S3 specifically includes:

[0030] S31. Based on the node topology association data, the computing nodes, storage nodes and network links are divided into multiple topology-level resource domains according to the node connection level and the number of link hops.

[0031] S32. Within each topology level resource domain, the CPU utilization, memory usage, local storage read / write speed, task queue length, and node energy consumption of computing nodes are aggregated to generate a computing resource state subfield.

[0032] S33. Within each topology level resource domain, aggregate the total storage capacity, remaining storage capacity, data read / write rate, and storage access latency of storage nodes to generate a storage resource state subfield.

[0033] S34. Within each topology level resource domain, the bandwidth utilization, packet loss rate, transmission delay, and delay jitter of network links are aggregated to generate network resource state subfields.

[0034] S35. Based on task response latency, service level agreement satisfaction rate, task priority and task deadline, aggregate the task queuing status and service risk status in each topology level resource domain to generate a queue risk status subfield.

[0035] S36. Based on the resource type correspondence, the computing resource state subfield, storage resource state subfield, network resource state subfield and queue risk state subfield in each topology level resource domain are associated and coupled to generate a basic mean field state vector.

[0036] S37. Extract the changes in central processing unit load, memory load, storage load, link congestion, task queue, and service level agreement risk of each topology level resource domain within a continuous historical scheduling period, determine the direction, magnitude, and rate of change of each type of status, and generate status trend information.

[0037] S38. The basic mean field state vector and the state trend information are fused according to the topological hierarchy correspondence and resource type correspondence to generate the mean field state vector.

[0038] Optionally, S36 specifically includes:

[0039] S361. Extract link latency, link bandwidth utilization, node hop count, cross-domain migration cost of task clusters, and resource load differences between any two topology-level resource domains to form inter-level correlation impact parameters.

[0040] S362. Perform reverse conversion on the link bandwidth utilization rate to obtain the remaining bandwidth ratio of the link, and perform unified dimensional processing on the link latency, remaining bandwidth ratio of the link, node hop count, cross-domain migration cost of task clusters, and resource load differences to generate standardized correlation parameters between layers.

[0041] S363. Based on the standardized correlation parameters between the layers, and in accordance with the principle that the higher the proportion of remaining bandwidth of the link, the lower the link latency, the fewer the number of node hops, the lower the cost of cross-domain migration of task clusters, and the smaller the difference in resource load, the higher the correlation degree, determine the correlation weight between any two topological layer resource domains, and normalize each correlation weight corresponding to the same current resource domain proportionally to generate a set of topological layer resource domain correlation weights.

[0042] S364. Taking each topological level resource domain as the current resource domain, and using the associated weights corresponding to the current resource domain, the same type of resource state subfields corresponding to the other topological level resource domains are weighted and aggregated to generate the computing resource coupling state, storage resource coupling state, network resource coupling state and queue risk coupling state corresponding to the current resource domain.

[0043] S365. The computing resource state subfield, storage resource state subfield, network resource state subfield and queue risk state subfield corresponding to the current resource domain are weighted and merged with the corresponding coupling state according to the degree of influence to generate the mean field component of each resource domain corresponding to the current resource domain.

[0044] S366. According to the hierarchical order of the topological resource domains and the arrangement order of computing resources, storage resources, network resources and queue risks, the mean field components in each resource domain are spliced ​​together to generate the basic mean field state vector for the current scheduling cycle.

[0045] Optionally, S4 specifically includes:

[0046] S41. Map each task cluster to a corresponding representative agent, and extract the local state vector and mean field state vector of each representative agent.

[0047] S42. Based on the resource demand characteristics, service level characteristics, and task time limit characteristics of the task clusters corresponding to each representative intelligent agent, and combined with the load status, link congestion status, and service risk status of each topology level resource domain represented by the mean field state vector, constrain and screen the reachable topology level resource domains, resource nodes, and network paths to obtain candidate resource domains, candidate resource nodes, and candidate network paths.

[0048] S43. Based on the candidate resource domain, candidate resource node and candidate network path, combine the deployment relationship between the task cluster and the target computing node, the migration relationship between the task cluster and the resource node, the resource adjustment relationship corresponding to the task cluster and the network path switching relationship corresponding to the task cluster to generate a candidate action set consisting of deployment action, migration action, scaling up and down action and rerouting action.

[0049] S44. Calculate the changes in node load balance, link congestion, task response latency, node energy consumption, migration overhead, and service level agreement default risk after each candidate action is executed.

[0050] S45. Determine the resource adjustment effect based on the changes in node load balancing and link congestion after each candidate action is executed; determine the service guarantee effect based on the changes in task response latency and service level agreement default risk; determine the action execution cost based on the changes in node energy consumption and migration overhead; and construct a benefit evaluation model corresponding to each candidate action based on the resource adjustment effect, service guarantee effect, and action execution cost.

[0051] S46. Based on the representative agents, local state vectors, mean field state vectors, candidate action sets, and benefit evaluation models, determine the action response relationships of each representative agent within the target scheduling period and construct a mean field game model.

[0052] Optionally, S5 specifically includes:

[0053] S51. Concatenate and encode the local state vector and mean field state vector of each representative agent to generate the conditional state features of each representative agent.

[0054] S52. Input the conditional state features into the conditional generation flow network scheduling strategy model, perform conditional control encoding on the conditional state features, and generate an action generation condition vector.

[0055] S53. Based on the action generation condition vector, starting from the empty action state, generate action components that match the current action type in the order of action type, target topology level resource domain, target resource node, target network path and resource adjustment parameters, and input the generated action components back into the condition generation flow network scheduling strategy model until the end marker is output, and generate complete candidate scheduling actions.

[0056] S54. Repeat S53 to generate multiple sets of candidate scheduling actions for the same representative agent, and deduplicate the generated results to form a set of candidate scheduling actions corresponding to each representative agent.

[0057] S55. Input each candidate scheduling action and its corresponding local state vector and mean field state vector into the conditional generation flow network scheduling strategy model, evaluate the resource adjustment effect, service guarantee effect and action execution cost after each candidate scheduling action is executed, and generate the value evaluation result corresponding to each candidate scheduling action.

[0058] S56. Output the candidate scheduling actions of each representative agent in the current scheduling cycle and the value evaluation results corresponding to the candidate scheduling actions.

[0059] Optionally, S6 specifically includes:

[0060] S61. Based on the value evaluation results of the candidate scheduling actions corresponding to each representative agent, determine the target scheduling action of each representative agent in the current scheduling cycle.

[0061] S62. Based on the target scheduling actions corresponding to each representative agent, update the deployment relationship, migration relationship, resource allocation relationship and network path relationship between task clusters and resource nodes to obtain the mapping relationship between task clusters and resource nodes corresponding to the current scheduling cycle.

[0062] S63. Based on the updated mapping relationship between task clusters and resource nodes, re-statistically analyze the computing resource load status, storage resource load status, network link congestion status, task queue status, and service risk status within each topology level resource domain, and generate an updated basic mean field state vector.

[0063] S64. Based on the updated base mean field state vector and the state change information within the continuous historical scheduling period, generate the updated mean field state vector.

[0064] S65. Re-input the updated mean field state vector into the conditional generation flow network scheduling strategy model to generate the next round of candidate scheduling actions and corresponding value evaluation results, and repeat the target scheduling action determination, mapping relationship update and mean field state vector update.

[0065] S66. Compare the changes in the mapping relationship between task clusters and resource nodes and the changes in the mean field state vector during two adjacent iterations. When the changes in the mapping relationship are less than the first convergence threshold and the changes in the mean field state vector are less than the second convergence threshold, stop the iteration to obtain the mean field balanced scheduling strategy.

[0066] Optionally, S8 specifically includes:

[0067] S81. Collect the node central processing unit utilization, memory usage, storage usage, network link bandwidth usage, task response latency, node energy consumption, migration overhead, and service level protocol satisfaction rate after each representative intelligent agent executes the target scheduling action, and generate scheduling execution status data.

[0068] S82. Associate the scheduling execution status data with the value assessment results corresponding to the target scheduling action, the mapping relationship between task clusters and resource nodes, and the mean field state vector to generate scheduling feedback data;

[0069] S83. Based on the scheduling feedback data, extract the changes in resource adjustment effect, service guarantee effect, and action execution cost before and after the execution of the target scheduling action, and generate model update samples.

[0070] S84. Update the model of the sample input condition generation flow network scheduling strategy model, and update the condition encoding parameters, action generation parameters and value evaluation parameters online.

[0071] S85. Reuse the updated conditional generation flow network scheduling strategy model for the next scheduling cycle's candidate scheduling action generation and value evaluation results output.

[0072] S86. Repeat the above-mentioned scheduling feedback data construction, model update sample generation and condition generation flow network scheduling strategy model online update, and output the dynamic scheduling results of cloud computing resources.

[0073] The beneficial effects of this invention are:

[0074] This invention combines node topology association data, task cluster local state vectors, and mean field state vectors to uniformly model computing resources, storage resources, network resources, and queue risks in cloud platforms. This enables a more accurate reflection of resource competition relationships and their changing trends between resource domains at different topological levels. By establishing a mean field game model and a conditional generation flow network scheduling strategy model, candidate scheduling actions are generated and their value is evaluated, achieving dynamic optimization of the mapping relationship between task clusters and resource nodes. Compared with existing scheduling methods, this invention reduces high-cost cross-domain migration and link congestion propagation, improves resource utilization and task completion efficiency, reduces service level agreement (SSPA) default rates, and allows for online updates to the scheduling strategy model based on scheduling execution results, thereby enhancing the system's stability, adaptability, and reliability under continuous operating conditions. Attached Figure Description

[0075] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0076] Figure 1 This is a flowchart of a cloud computing resource dynamic scheduling method based on artificial intelligence proposed in this invention;

[0077] Figure 2 This is a schematic diagram illustrating the construction of the mean field state vector for a cloud computing resource dynamic scheduling method based on artificial intelligence proposed in this invention.

[0078] Figure 3 This diagram illustrates the generation and iterative update of candidate scheduling actions in a cloud computing resource dynamic scheduling method based on artificial intelligence proposed in this invention. Detailed Implementation

[0079] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0080] refer to Figure 1-3 A dynamic scheduling method for cloud computing resources based on artificial intelligence includes the following steps:

[0081] S1. Collect operational monitoring data, task request data, task operation feedback data, and node topology association data of computing nodes, storage nodes, and network links in the cloud platform to form the original scheduling dataset;

[0082] S2. Perform preprocessing on the original scheduling dataset, group the tasks, generate a set of task clusters, and construct a corresponding local state vector for each task cluster.

[0083] S3. Construct a mean field state vector based on node topology association data and the original scheduling dataset;

[0084] S4. Map each task cluster to a corresponding representative agent. Based on the local state vector and mean field state vector of each representative agent, define the action space, construct the benefit evaluation model, and establish a mean field game model.

[0085] S5. Input the local state vector and mean field state vector of each representative agent into the conditions to generate a flow network scheduling strategy model, and output the candidate scheduling actions of each representative agent in the current scheduling cycle and the value evaluation results corresponding to the candidate scheduling actions.

[0086] S6. Update the mapping relationship between task clusters and resource nodes based on the candidate scheduling actions of each representative agent, and update the mean field state vector. Iterate the candidate scheduling action update and mean field state vector update to obtain the mean field balanced scheduling strategy.

[0087] S7. Output resource scheduling control instructions to the cloud platform according to the mean field equilibrium scheduling strategy, and form scheduling execution results;

[0088] S8. Construct scheduling feedback data based on the scheduling execution results, and update the condition-generated flow network scheduling strategy model online based on the scheduling feedback data, and output the dynamic scheduling results of cloud computing resources.

[0089] In this embodiment, S1 specifically includes:

[0090] S11. Collect the CPU utilization, memory usage, local storage read / write speed, task queue length and node energy consumption of each computing node to obtain computing node operation monitoring data;

[0091] S12. Collect the total storage capacity, remaining storage capacity, data read / write rate, and storage access latency of each storage node to obtain storage node operation monitoring data;

[0092] S13. Collect the bandwidth utilization, packet loss rate, transmission delay and delay jitter of each network link to obtain network link operation monitoring data;

[0093] S14. Collect the task arrival time, task type, task priority, CPU requirements, memory requirements, storage requirements, network bandwidth requirements, service level agreement level, task deadline and execution duration for each task to obtain task request data.

[0094] S15. Collect the task response latency and service level agreement satisfaction rate of each task to obtain task operation feedback data;

[0095] S16. Collect the connection relationships between each computing node, each storage node and each network link to obtain node topology association data;

[0096] S17. Collect computing node operation monitoring data, storage node operation monitoring data, network link operation monitoring data, task request data, task operation feedback data, and node topology association data, and organize them according to the collection time sequence to form an original scheduling dataset containing time sequence evolution information.

[0097] In this embodiment, S2 specifically includes:

[0098] S21. Perform outlier removal, missing value completion, and timestamp alignment on the original scheduling dataset to obtain an aligned dataset;

[0099] S22. Perform min-max normalization on the CPU requirements, memory requirements, storage requirements, network bandwidth requirements, and execution time in the aligned dataset to obtain a normalized dataset.

[0100] S23. Extract the CPU requirements, memory requirements, storage requirements, network bandwidth requirements, and service level protocol level corresponding to each task from the normalized dataset, and arrange them in the order of CPU requirements, memory requirements, storage requirements, network bandwidth requirements, and service level protocol level to form a task feature vector, thus obtaining the task feature set.

[0101] S24. Cluster and group each task according to the task feature set to obtain a task cluster set;

[0102] S25. For each task cluster, calculate the number of tasks within the cluster, average CPU requirements, average memory requirements, average storage requirements, average network bandwidth requirements, average execution time, and service level protocol distribution, and construct the corresponding local state vector.

[0103] In this embodiment, S3 specifically includes:

[0104] S31. Based on the node topology association data, the computing nodes, storage nodes and network links are divided into layers according to the node connection level and the number of link hops, forming multiple topology-level resource domains.

[0105] S32. Within each topology level resource domain, the CPU utilization, memory usage, local storage read / write speed, task queue length, and node energy consumption of computing nodes are aggregated to generate a computing resource state subfield.

[0106] S33. Within each topology level resource domain, aggregate the total storage capacity, remaining storage capacity, data read / write rate, and storage access latency of storage nodes to generate a storage resource state subfield.

[0107] S34. Within each topology level resource domain, the bandwidth utilization, packet loss rate, transmission delay, and delay jitter of network links are aggregated to generate network resource state subfields.

[0108] S35. Based on task response latency, service level agreement satisfaction rate, task priority and task deadline, aggregate the task queuing status and service risk status in each topology level resource domain to generate a queue risk status subfield.

[0109] S36. Based on the resource type correspondence, the computing resource state subfield, storage resource state subfield, network resource state subfield and queue risk state subfield in each topology level resource domain are associated and coupled to generate a basic mean field state vector.

[0110] S37. Extract the changes in central processing unit load, memory load, storage load, link congestion, task queue, and service level agreement risk of each topology level resource domain within a continuous historical scheduling period, determine the direction, magnitude, and rate of change of each type of status, and generate status trend information.

[0111] S38. The basic mean field state vector and the state trend information are fused according to the topological hierarchy correspondence and resource type correspondence to generate the mean field state vector.

[0112] In this embodiment, S36 specifically includes:

[0113] S361. Extract link latency, link bandwidth utilization, node hop count, cross-domain migration cost of task clusters, and resource load differences between any two topology-level resource domains to form inter-level correlation impact parameters.

[0114] S362. Perform reverse conversion on the link bandwidth utilization rate to obtain the remaining bandwidth ratio of the link, and perform unified dimensional processing on the link latency, remaining bandwidth ratio of the link, node hop count, cross-domain migration cost of task clusters, and resource load differences to generate standardized correlation parameters between layers.

[0115] S363. Based on the standardized correlation parameters between layers, and in accordance with the principle that the higher the proportion of remaining bandwidth of the link, the lower the link latency, the fewer the number of node hops, the lower the cost of cross-domain migration of task clusters, and the smaller the difference in resource load, the higher the correlation degree, determine the correlation weight between any two topological layer resource domains, and normalize each correlation weight corresponding to the same current resource domain proportionally to generate a set of topological layer resource domain correlation weights.

[0116] S364. Taking each topology level resource domain as the current resource domain, and using the associated weights corresponding to the current resource domain, the subfields of the same type of resource status corresponding to the other topology level resource domains are weighted and aggregated to generate the computing resource coupling status, storage resource coupling status, network resource coupling status and queue risk coupling status corresponding to the current resource domain.

[0117] S365. The computing resource state subfield, storage resource state subfield, network resource state subfield and queue risk state subfield corresponding to the current resource domain are weighted and merged with the corresponding coupling state according to the degree of influence to generate the mean field component of each resource domain corresponding to the current resource domain.

[0118] S366. According to the hierarchical order of the topological resource domains and the order of computing resources, storage resources, network resources and queue risks, the mean field components in each resource domain are spliced ​​together to generate the basic mean field state vector for the current scheduling cycle.

[0119] In this implementation, based on node topology association data, compute nodes, storage nodes, and network links are hierarchically divided according to node connection level and link hop count, forming multiple topology-level resource domains. Node connection levels can be divided according to same-node domain, same-rack domain, same-cluster domain, and cross-cluster domain, or they can be divided into several resource domains with topological proximity relationships according to the actual deployment structure of the cloud platform. This division method allows for separate characterization of intra-domain resource contention and inter-domain resource contention propagation states, thereby avoiding the masking of local congestion and cross-domain migration costs caused by using only a global average state.

[0120] Within each topology-level resource domain, subfields for computing resource status, storage resource status, network resource status, and queue risk status are constructed. The computing resource status subfield is determined by CPU utilization, memory usage, local storage read / write speed, task queue length, and node energy consumption, characterizing the computing load level and task carrying capacity within this resource domain. The storage resource status subfield is determined by total storage capacity, remaining storage capacity, data read / write speed, and storage access latency, characterizing the storage supply capacity and storage access pressure within this resource domain. The network resource status subfield is determined by bandwidth utilization, packet loss rate, transmission latency, and latency jitter, characterizing the link congestion level and data transmission capacity within this resource domain. The queue risk status subfield is determined by task response latency, service level agreement (SLA) satisfaction rate, task priority, and task deadline, characterizing the task backlog and service default risk within this resource domain.

[0121] After forming the aforementioned resource state subfields, the correlation between resource domains at different topological levels is further quantified. Specifically, link latency, link bandwidth utilization, node hop count, task cluster cross-domain migration cost, and resource load difference are extracted between any two topological level resource domains to form inter-level correlation impact parameters. The task cluster cross-domain migration cost is determined by the data migration amount corresponding to the task cluster, the link bandwidth occupied during migration, link transmission latency, node hop count, and task interruption time during migration. The resource load difference is determined by the differences in CPU load, memory load, storage load, and task queue length. The link bandwidth utilization is converted inversely to obtain the remaining link bandwidth ratio, which is then processed with link latency, node hop count, task cluster cross-domain migration cost, and resource load difference to generate standardized inter-level correlation parameters.

[0122] Based on the standardized correlation parameters mentioned above, and following the principle that a higher correlation is indicated by a larger proportion of remaining link bandwidth, lower link latency, fewer node hops, lower cross-domain migration cost, and smaller resource load differences, the correlation weight between any two topology-level resource domains is determined. The correlation weights corresponding to the same current resource domain are proportionally normalized to serve as coefficients for weighted aggregation of similar resource state subfields in other resource domains. Subsequently, taking each topology-level resource domain as the current resource domain, the weighted aggregation of similar resource state subfields corresponding to other resource domains is performed to generate the computational resource coupling state, storage resource coupling state, network resource coupling state, and queue risk coupling state of the current resource domain. Then, the original resource state subfields of the current resource domain are weighted and merged with their corresponding coupling states to generate the mean field components within the current resource domain. Finally, according to the hierarchical order of the topology-level resource domains and the order of computational resources, storage resources, network resources, and queue risks, the mean field components within each resource domain are concatenated to generate the basic mean field state vector for the current scheduling cycle.

[0123] To enable the mean field state vector to be used for action generation and load balancing in the next scheduling cycle, temporal evolution information needs to be incorporated into the basic mean field state vector. Specifically, changes in CPU load, memory load, storage load, link congestion, task queues, and service level agreement (SLA) risks for each topology level resource domain within consecutive historical scheduling cycles are extracted. The direction, magnitude, and rate of change of each state type within adjacent scheduling cycles are determined, forming state trend information. The basic mean field state vector and the state trend information are then fused according to topology level and resource type correspondences to generate the mean field state vector. The mean field state vector obtained in this way not only reflects the current resource competition landscape but also characterizes the evolution direction of the resource competition situation in the next scheduling cycle.

[0124] In this embodiment, S4 specifically includes:

[0125] S41. Map each task cluster to a corresponding representative agent, and extract the local state vector and mean field state vector of each representative agent.

[0126] S42. Based on the resource demand characteristics, service level characteristics, and task time limit characteristics of the task clusters corresponding to each representative intelligent agent, and combined with the load status, link congestion status, and service risk status of each topology level resource domain represented by the mean field state vector, constrain and screen the reachable topology level resource domains, resource nodes, and network paths to obtain candidate resource domains, candidate resource nodes, and candidate network paths.

[0127] S43. Based on the candidate resource domain, candidate resource node and candidate network path, combine the deployment relationship between the task cluster and the target computing node, the migration relationship between the task cluster and the resource node, the resource adjustment relationship corresponding to the task cluster and the network path switching relationship corresponding to the task cluster to generate a candidate action set consisting of deployment action, migration action, scaling up and down action and rerouting action.

[0128] S44. Calculate the changes in node load balance, link congestion, task response latency, node energy consumption, migration overhead, and service level agreement default risk after each candidate action is executed.

[0129] S45. Determine the resource adjustment effect based on the changes in node load balancing and link congestion after each candidate action is executed; determine the service assurance effect based on the changes in task response latency and service level agreement default risk; determine the action execution cost based on the changes in node energy consumption and migration overhead; and construct a benefit evaluation model for each candidate action based on the resource adjustment effect, service assurance effect, and action execution cost.

[0130] S46. Based on the representative agents, local state vectors, mean field state vectors, candidate action sets, and benefit evaluation models, determine the action response relationships of each representative agent within the target scheduling period and construct a mean field game model.

[0131] In this implementation, each task cluster is mapped to a representative agent. Each representative agent corresponds to a local state vector, which characterizes the resource demand characteristics, service level characteristics, and time limit characteristics of that task cluster. The mean field state vector serves as the globally shared state, characterizing the resource competition situation, link congestion situation, and service risk situation in each topology level resource domain. Based on the aforementioned local and globally shared states, the scheduling actions of each representative agent are subject to constraint screening and benefit evaluation.

[0132] Specifically, based on the CPU requirements, memory requirements, storage requirements, network bandwidth requirements, service level agreement level, task deadlines and execution durations of the task clusters, and combined with the load status, link congestion status and service risk status of each topology level resource domain represented by the mean field state vector, reachable topology level resource domains, resource nodes and network paths are screened, and resource domains, resource nodes and network paths that do not meet the resource capacity, link reachability, latency constraints and service constraints are eliminated, resulting in candidate resource domains, candidate resource nodes and candidate network paths.

[0133] Based on candidate resource domains, candidate resource nodes, and candidate network paths, an action space for representative agents is constructed. The action space includes deployment actions, migration actions, scaling actions, and rerouting actions. Deployment actions are determined by the deployment relationship between the task cluster and candidate computing nodes; migration actions are determined by the migration relationship between the task cluster and candidate migration target nodes; scaling actions are determined by the correspondence between the currently allocated resources of the task cluster and the adjustment range of the target resources; and rerouting actions are determined by the switching relationship between the current network path of the task cluster and candidate network paths. All of these actions take the task cluster as the executing entity and the candidate resource domain, candidate resource nodes, and candidate network paths as the constraint objects of action parameters.

[0134] For each candidate action, the changes in node load balance, link congestion, task response latency, service level agreement (SLA) default risk, node energy consumption, and migration overhead after the action are calculated. Specifically, the change in node load balance reflects whether the load distribution among resource nodes tends towards equilibrium; the change in link congestion reflects whether network bandwidth usage and transmission congestion have decreased; the change in task response latency and SLA default risk reflects whether the task service quality has improved after the action is executed; and the changes in node energy consumption and migration overhead reflect whether the implementation cost of the action has increased. A benefit evaluation model is constructed based on these state changes, enabling it to simultaneously reflect the resource adjustment effect, service guarantee effect, and action execution cost. Based on representative agents, the action space, and the benefit evaluation model, a mean-field game model is established to characterize the impact of group competition in large-scale task cluster parallel scheduling scenarios through a mean-field approximation.

[0135] In this embodiment, S5 specifically includes:

[0136] S51. Concatenate and encode the local state vector and mean field state vector of each representative agent to generate the conditional state features of each representative agent.

[0137] S52. Input the conditional state features into the conditional generation flow network scheduling strategy model, perform conditional control encoding on the conditional state features, and generate action generation condition vectors.

[0138] S53. Based on the action generation condition vector, starting from the empty action state, generate action components that match the current action type in the order of action type, target topology level resource domain, target resource node, target network path and resource adjustment parameters, and input the generated action components back to the input condition generation flow network scheduling strategy model until the output end marker is generated, thus generating a complete candidate scheduling action.

[0139] S54. Repeat S53 to generate multiple sets of candidate scheduling actions for the same representative agent, and deduplicate the generated results to form a set of candidate scheduling actions corresponding to each representative agent.

[0140] S55. Generate a flow network scheduling strategy model by inputting each candidate scheduling action with the corresponding local state vector and mean field state vector. Evaluate the resource adjustment effect, service guarantee effect and action execution cost after each candidate scheduling action is executed, and generate the value evaluation result corresponding to each candidate scheduling action.

[0141] S56. Output the candidate scheduling actions of each representative agent in the current scheduling cycle and the corresponding value evaluation results of the candidate scheduling actions.

[0142] In this embodiment, the conditional generation flow network scheduling strategy model is used to generate candidate scheduling actions under the joint constraints of local state and mean field state, and output the value evaluation results of each candidate scheduling action. Specifically, firstly, the local state vector and the mean field state vector of each representative agent are concatenated and encoded to generate conditional state features; then, conditional control encoding is performed on the conditional state features to generate action generation conditional vectors.

[0143] The condition-generated flow network scheduling strategy model includes a condition encoding module, an action generation module, and a value evaluation module. The condition encoding module converts local and mean field states into conditional representations that can be invoked by the action generation module. The action generation module progressively generates action parameters under conditional constraints. The value evaluation module evaluates the generated actions and outputs the value evaluation results. Starting from an empty action state, the action generation module progressively generates action type parameters and corresponding action parameters based on the action generation condition vector. Action parameters include, but are not limited to, target topology-level resource domain parameters, target resource node parameters, target network path parameters, and resource adjustment parameters. After generating each action parameter, the generated action parameter and the action generation condition vector are fed back into the action generation module to update the current action state, and then the next action parameter is generated, until a complete candidate scheduling action is formed.

[0144] In this process, the generation flow intensity in the conditional generation flow network characterizes the tendency of the current action state to extend to the next action parameter. The generation flow intensity is jointly determined by the current action state, the set of action generation condition parameters, and the already generated action parameters; the larger the value, the higher the priority of the current action state extending to the corresponding candidate action parameter. Based on the generation flow intensity, the next action parameter can be determined from multiple optional action parameters, thus forming a complete candidate scheduling action that satisfies the conditional constraints. After the complete candidate scheduling action is generated, resource constraint verification, path reachability verification, and service constraint verification are still required; only candidate scheduling actions that pass the verification are retained.

[0145] After candidate scheduling actions are generated, the value assessment module evaluates the resource adjustment effect, service guarantee effect, and execution cost of the candidate scheduling actions, and outputs the value assessment results in conjunction with the corresponding generation flow intensity. In this way, candidate scheduling actions are not only executable, but also have comparable value assessment results, which facilitates the selection of subsequent target scheduling actions.

[0146] In this embodiment, S6 specifically includes:

[0147] S61. Based on the value evaluation results of the candidate scheduling actions corresponding to each representative agent, determine the target scheduling action of each representative agent in the current scheduling cycle.

[0148] S62. Based on the target scheduling actions corresponding to each representative agent, update the deployment relationship, migration relationship, resource allocation relationship and network path relationship between task clusters and resource nodes to obtain the mapping relationship between task clusters and resource nodes corresponding to the current scheduling cycle.

[0149] S63. Based on the updated mapping relationship between task clusters and resource nodes, re-statistically analyze the computing resource load status, storage resource load status, network link congestion status, task queue status, and service risk status within each topology level resource domain, and generate an updated basic mean field state vector.

[0150] S64. Based on the updated base mean field state vector and the state change information within the continuous historical scheduling period, generate the updated mean field state vector.

[0151] S65. Re-input the updated mean field state vector into the conditional generation flow network scheduling strategy model to generate the next round of candidate scheduling actions and corresponding value evaluation results, and repeat the target scheduling action determination, mapping relationship update and mean field state vector update.

[0152] S66. Compare the changes in the mapping relationship between task clusters and resource nodes and the changes in the mean field state vector during two adjacent iterations. Stop the iteration when the changes in the mapping relationship are less than the first convergence threshold and the changes in the mean field state vector are less than the second convergence threshold, and obtain the mean field balanced scheduling strategy.

[0153] In this implementation, within each scheduling cycle, the target scheduling action for the current scheduling cycle is determined based on the value evaluation results of the candidate scheduling actions corresponding to each representative agent. According to the target scheduling action, the deployment relationships, migration relationships, resource allocation relationships, and network path relationships between task clusters and resource nodes are updated to obtain the mapping relationship between task clusters and resource nodes for the current scheduling cycle.

[0154] After updating the mapping relationship, the computing resource load status, storage resource load status, network link congestion status, task queue status, and service risk status in each topology level resource domain are re-statistically analyzed to generate an updated basic mean field state vector. Then, based on the updated basic mean field state vector and the state change information within consecutive historical scheduling cycles, an updated mean field state vector is generated. Subsequently, the updated mean field state vector is re-input into the conditional flow network scheduling strategy model to generate the next round of candidate scheduling actions and value assessment results. This process of determining the target scheduling action, updating the mapping relationship, and updating the mean field state vector is repeated.

[0155] To determine whether the iteration process has ended, the changes in the mapping relationship between task clusters and resource nodes and the changes in the mean field state vector are statistically analyzed between adjacent iterations. The changes in the mapping relationship between task clusters and resource nodes can be determined by changes in deployment relationships, migration relationships, resource allocation relationships, and network path relationships; the changes in the mean field state vector can be determined by changes in resource load state, link congestion state, task queue state, and service risk state. The iteration stops when the change in the mapping relationship is less than a first convergence threshold and the change in the mean field state vector is less than a second convergence threshold, resulting in the mean field balanced scheduling strategy.

[0156] In this embodiment, the first convergence threshold is determined based on the degree of change in the mapping relationship between task clusters and resource nodes in two adjacent iterations. This degree of change is jointly determined by changes in deployment relationships, migration relationships, resource allocation relationships, and network path relationships. The second convergence threshold is determined based on the degree of change in the mean field state vector in two adjacent iterations. This degree of change is jointly determined by changes in resource load state, link congestion state, task queue state, and service risk state. The first and second convergence thresholds can be set according to the historical fluctuation range of various state variables, scheduling cycle length, and system stability requirements during the continuous operation of the cloud platform.

[0157] In this embodiment, S8 specifically includes:

[0158] S81. Collect the node central processing unit utilization, memory usage, storage usage, network link bandwidth usage, task response latency, node energy consumption, migration overhead, and service level protocol satisfaction rate after each representative intelligent agent executes the target scheduling action, and generate scheduling execution status data.

[0159] S82. Associate the scheduling execution status data with the value assessment results corresponding to the target scheduling action, the mapping relationship between task clusters and resource nodes, and the mean field state vector to generate scheduling feedback data;

[0160] S83. Based on the scheduling feedback data, extract the changes in resource adjustment effect, service guarantee effect, and action execution cost before and after the execution of the target scheduling action, and generate model update samples.

[0161] S84. Update the model of the sample input condition generation flow network scheduling strategy model, and update the condition coding parameters, action generation parameters and value evaluation parameters online.

[0162] S85. Reuse the updated conditional generation flow network scheduling strategy model for the next scheduling cycle's candidate scheduling action generation and value evaluation results output.

[0163] S86, repeatedly executes scheduling feedback data construction, model update sample generation and condition generation flow network scheduling strategy model online update, and outputs dynamic scheduling results of cloud computing resources.

[0164] After the mean-field equilibrium scheduling strategy is executed, the following data are collected for each representative agent: CPU utilization, memory usage, storage usage, network link bandwidth usage, task response latency, node energy consumption, migration overhead, and service level agreement satisfaction rate. This data generates scheduling execution status data. The scheduling execution status data is then correlated with the value assessment results corresponding to the target scheduling actions, the mapping relationship between task clusters and resource nodes, and the mean-field state vector to generate scheduling feedback data.

[0165] Based on scheduling feedback data, changes in resource state, service state, and action execution cost before and after the execution of the target scheduling action are extracted to form model update samples. These model update samples are used to update the conditional generation flow network scheduling strategy model online. Specifically, conditional encoding parameters are used to update the mapping relationship between the local state vector and the mean field state vector in the conditional encoding module and the conditional representation; action generation parameters are used to update the action type parameters and the extended weights in the step-by-step generation process of action parameters in the action generation module; and value evaluation parameters are used to update the evaluation relationship between candidate scheduling actions and their corresponding value evaluation results in the value evaluation module. By updating these parameters online, the model can continuously adjust the action generation strategy and action evaluation results based on real-time resource state changes and task feedback results from the cloud platform, thereby outputting candidate scheduling actions and value evaluation results that are more suitable for the next scheduling cycle.

[0166] Example 1: To verify the feasibility of this invention in practice, it was applied to a medical image-assisted analysis cloud platform. This cloud platform is used to handle image preprocessing tasks, lesion identification tasks, report generation tasks, and archiving and storage tasks. Because different tasks have different requirements in terms of CPU, memory, storage, and network bandwidth, and because there are topological differences between computing nodes, storage nodes, and network links, traditional scheduling methods are prone to problems such as local resource domain overload, frequent cross-domain migrations, and increased service level agreement (SLA) default rates.

[0167] In this embodiment, the method of the present invention is deployed on a cloud platform consisting of 64 computing nodes, 16 storage nodes and 48 core network links. The scheduling cycle is set to 30 seconds, and it runs continuously for 120 scheduling cycles, processing a total of 9,600 task requests.

[0168] The system first collects monitoring data on the operation of computing nodes, storage nodes, network links, task requests, task feedback, and node topology associations to form a raw scheduling dataset. Then, it performs outlier removal, missing value completion, and timestamp alignment on the raw scheduling dataset, normalizes the task resource requirements and execution time, clusters tasks according to their characteristics to generate a set of task clusters, and constructs a local state vector for each task cluster.

[0169] Subsequently, based on node topology association data, the system divides compute nodes, storage nodes, and network links into multiple topology-level resource domains. Within each topology-level resource domain, it constructs compute resource state subfields, storage resource state subfields, network resource state subfields, and queue risk state subfields. Then, it extracts link latency, link bandwidth utilization, node hop count, task cluster cross-domain migration cost, and resource load differences between any two topology-level resource domains to determine the association weights between resource domains. It then couples similar resource state subfields to generate a basic mean field state vector for the current scheduling cycle. Finally, it combines resource load changes, link congestion changes, task queue changes, and service level agreement risk changes over continuous historical scheduling cycles to generate a mean field state vector.

[0170] During the scheduling decision-making phase, the system maps each task cluster to a corresponding representative agent and inputs the local state vector and mean field state vector into the conditional generation flow network scheduling strategy model. Based on resource demand characteristics, service level characteristics, and task time limit characteristics, as well as the resource domain load state, link congestion state, and service risk state represented by the mean field state vector, the system filters reachable topology-level resource domains, resource nodes, and network paths to generate candidate resource domains, candidate resource nodes, and candidate network paths. Then, the conditional generation flow network scheduling strategy model generates candidate scheduling actions and outputs the corresponding value assessment results.

[0171] The system determines the target scheduling action based on the value assessment results of candidate scheduling actions, updates the mapping relationship between task clusters and resource nodes, and re-statistically analyzes the resource status of resource domains at each topology level, updating the mean field state vector. Then, it repeats the processes of candidate action generation, target action determination, and mapping relationship updating until the changes in the mapping relationship and the mean field state vector in two adjacent iterations are simultaneously less than the corresponding convergence thresholds, thus obtaining a mean field equilibrium scheduling strategy. Finally, the system outputs resource scheduling control instructions based on the mean field equilibrium scheduling strategy and constructs scheduling feedback data based on the execution results to update the conditional generation flow network scheduling strategy model online.

[0172] In this embodiment, the average time for data acquisition and preprocessing within a single scheduling cycle is 0.28 seconds, the average time for constructing the mean field state vector is 0.31 seconds, the average time for generating candidate scheduling actions is 0.37 seconds, the average time for solving the mean field equilibrium is 0.26 seconds, and the average total time for scheduling decision-making per cycle is 1.22 seconds. 113 out of 120 scheduling cycles enter a stable mapping interval after one iteration, resulting in a stability rate of 94.17% after one iteration. The average task completion latency is 6.21 seconds, the service level agreement (SLA) breach rate is 1.54%, the average resource utilization rate is 84.73%, and the number of cross-domain migrations is 52 per 1000 tasks.

[0173] To further verify the actual performance of the method of the present invention, the method of the present invention was compared with the fixed threshold scheduling method, minimum load scheduling method, DQN scheduling method and PPO scheduling method. The results are shown in Table 1 and Table 2.

[0174] Table 1. Comparison of Overall Performance of Different Cloud Resource Dynamic Scheduling Schemes

[0175]

[0176] As shown in Table 1, the method of this invention outperforms the fixed threshold scheduling method, minimum load scheduling method, DQN scheduling method, and PPO scheduling method in key indicators such as average task completion latency, service level agreement (SLA) default rate, average resource utilization, number of cross-domain migrations, and task completion rate. Specifically, the average task completion latency of the method of this invention is 6.21 seconds, significantly lower than the comparative methods, indicating that it can more effectively utilize available resources within and between resource domains; the SLA default rate is reduced to 1.54%, indicating that the method can better suppress timeouts of high-priority tasks under congestion conditions; the average resource utilization rate is increased to 84.73%, indicating that it improves resource scheduling efficiency while maintaining high service quality; the number of cross-domain migrations is significantly reduced, indicating that the method reduces high-cost migrations under topology-aware constraints; and the task completion rate reaches 98.63%, further demonstrating that the method of this invention has higher scheduling reliability in complex heterogeneous task environments.

[0177] Table 2 Comparison of Continuous Operation Stability of Different Cloud Resource Scheduling Schemes

[0178]

[0179] As shown in Table 2, the method of the present invention also has significant advantages in terms of continuous operational stability. Under the condition of the same total number of scheduling cycles, the number of abnormal rollbacks of the method of the present invention is 0, which is better than the comparative methods; the proportion of resource hotspot duration cycles is only 2.14%, indicating that the method can alleviate the overload phenomenon of local resource domains more quickly; the proportion of repeated rescheduling is only 1.27%, indicating that the mean field balanced scheduling strategy generated by the method of the present invention has better stability and consistency. In summary, Table 2 shows that the method of the present invention not only improves scheduling accuracy and efficiency, but also enhances the reliability, robustness and service continuity of the system during long-term continuous operation.

[0180] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based cloud computing resource dynamic scheduling method, characterized in that, Includes the following steps: S1. Collect operational monitoring data, task request data, task operation feedback data, and node topology association data of computing nodes, storage nodes, and network links in the cloud platform to form the original scheduling dataset; S2. Perform preprocessing on the original scheduling dataset, group the tasks, generate a set of task clusters, and construct a corresponding local state vector for each task cluster. S3. Based on the node topology association data and the original scheduling dataset, construct the mean field state vector; S4. Map each task cluster to a corresponding representative agent. Based on the local state vector of each representative agent and the mean field state vector, define the action space, construct the profit evaluation model, and establish the mean field game model. S5. Input the local state vectors and mean field state vectors of each representative agent into the conditions to generate a flow network scheduling strategy model, and output the candidate scheduling actions of each representative agent in the current scheduling cycle and the value evaluation results corresponding to the candidate scheduling actions. S6. Update the mapping relationship between task clusters and resource nodes based on the candidate scheduling actions of each representative agent, and update the mean field state vector. Iterate the candidate scheduling action update and mean field state vector update to obtain the mean field balanced scheduling strategy. S7. Output resource scheduling control instructions to the cloud platform according to the mean field equilibrium scheduling strategy to form a scheduling execution result; S8. Construct scheduling feedback data based on the scheduling execution results, and update the conditional generation flow network scheduling strategy model online based on the scheduling feedback data, and output the dynamic scheduling results of cloud computing resources. 2.The method of claim 1, wherein, S1 specifically includes: S11. Collect the CPU utilization, memory usage, local storage read / write speed, task queue length and node energy consumption of each computing node to obtain computing node operation monitoring data; S12. Collect the total storage capacity, remaining storage capacity, data read / write rate, and storage access latency of each storage node to obtain storage node operation monitoring data; S13. Collect the bandwidth utilization, packet loss rate, transmission delay and delay jitter of each network link to obtain network link operation monitoring data; S14. Collect the task arrival time, task type, task priority, CPU requirements, memory requirements, storage requirements, network bandwidth requirements, service level agreement level, task deadline and execution duration for each task to obtain task request data. S15. Collect the task response latency and service level agreement satisfaction rate of each task to obtain task operation feedback data; S16. Collect the connection relationships between each computing node, each storage node and each network link to obtain node topology association data; S17. Collect computing node operation monitoring data, storage node operation monitoring data, network link operation monitoring data, task request data, task operation feedback data, and node topology association data, and organize them according to the collection time sequence to form an original scheduling dataset containing time sequence evolution information. 3.The method of claim 1, wherein, S2 specifically includes: S21. Perform outlier removal, missing value completion, and timestamp alignment on the original scheduling dataset to obtain an aligned dataset; S22. Perform min-max normalization on the CPU requirements, memory requirements, storage requirements, network bandwidth requirements, and execution time in the aligned dataset to obtain a normalized dataset. S23. Extract the CPU requirements, memory requirements, storage requirements, network bandwidth requirements and service level protocol level corresponding to each task from the normalized dataset, and arrange them in the order of CPU requirements, memory requirements, storage requirements, network bandwidth requirements and service level protocol level to form a task feature vector, thereby obtaining a task feature set. S24. Based on the task feature set, cluster and group each task to obtain a task cluster set; S25. For each task cluster, calculate the number of tasks within the cluster, average CPU requirements, average memory requirements, average storage requirements, average network bandwidth requirements, average execution time, and service level protocol distribution, and construct the corresponding local state vector.

4. The method for dynamic scheduling of cloud computing resources based on artificial intelligence according to claim 1, characterized in that, S3 specifically includes: S31. Based on the node topology association data, the computing nodes, storage nodes and network links are divided into multiple topology-level resource domains according to the node connection level and the number of link hops. S32. Within each topology level resource domain, the CPU utilization, memory usage, local storage read / write speed, task queue length, and node energy consumption of computing nodes are aggregated to generate a computing resource state subfield. S33. Within each topology level resource domain, aggregate the total storage capacity, remaining storage capacity, data read / write rate, and storage access latency of storage nodes to generate a storage resource state subfield. S34. Within each topology level resource domain, the bandwidth utilization, packet loss rate, transmission delay, and delay jitter of network links are aggregated to generate network resource state subfields. S35. Based on task response latency, service level agreement satisfaction rate, task priority and task deadline, aggregate the task queuing status and service risk status in each topology level resource domain to generate a queue risk status subfield. S36. Based on the resource type correspondence, the computing resource state subfield, storage resource state subfield, network resource state subfield and queue risk state subfield in each topology level resource domain are associated and coupled to generate a basic mean field state vector. S37. Extract the changes in central processing unit load, memory load, storage load, link congestion, task queue, and service level agreement risk of each topology level resource domain within a continuous historical scheduling period, determine the direction, magnitude, and rate of change of each type of status, and generate status trend information. S38. The basic mean field state vector and the state trend information are fused according to the topological hierarchy correspondence and resource type correspondence to generate the mean field state vector.

5. The method for dynamic scheduling of cloud computing resources based on artificial intelligence according to claim 4, characterized in that, Specifically, S36 includes: S361. Extract link latency, link bandwidth utilization, node hop count, cross-domain migration cost of task clusters, and resource load differences between any two topology-level resource domains to form inter-level correlation impact parameters. S362. Perform reverse conversion on the link bandwidth utilization rate to obtain the remaining bandwidth ratio of the link, and perform unified dimensional processing on the link latency, remaining bandwidth ratio of the link, node hop count, cross-domain migration cost of task clusters, and resource load differences to generate standardized correlation parameters between layers. S363. Based on the standardized correlation parameters between the layers, and in accordance with the principle that the higher the proportion of remaining bandwidth of the link, the lower the link latency, the fewer the number of node hops, the lower the cost of cross-domain migration of task clusters, and the smaller the difference in resource load, the higher the correlation degree, determine the correlation weight between any two topological layer resource domains, and normalize each correlation weight corresponding to the same current resource domain proportionally to generate a set of topological layer resource domain correlation weights. S364. Taking each topological level resource domain as the current resource domain, and using the associated weights corresponding to the current resource domain, the same type of resource state subfields corresponding to the other topological level resource domains are weighted and aggregated to generate the computing resource coupling state, storage resource coupling state, network resource coupling state and queue risk coupling state corresponding to the current resource domain. S365. The computing resource state subfield, storage resource state subfield, network resource state subfield and queue risk state subfield corresponding to the current resource domain are weighted and merged with the corresponding coupling state according to the degree of influence to generate the mean field component of each resource domain corresponding to the current resource domain. S366. According to the hierarchical order of the topological resource domains and the arrangement order of computing resources, storage resources, network resources and queue risks, the mean field components in each resource domain are spliced ​​together to generate the basic mean field state vector for the current scheduling cycle.

6. The method for dynamic scheduling of cloud computing resources based on artificial intelligence according to claim 1, characterized in that, S4 specifically includes: S41. Map each task cluster to a corresponding representative agent, and extract the local state vector and mean field state vector of each representative agent. S42. Based on the resource demand characteristics, service level characteristics, and task time limit characteristics of the task clusters corresponding to each representative intelligent agent, and combined with the load status, link congestion status, and service risk status of each topology level resource domain represented by the mean field state vector, constrain and screen the reachable topology level resource domains, resource nodes, and network paths to obtain candidate resource domains, candidate resource nodes, and candidate network paths. S43. Based on the candidate resource domain, candidate resource node and candidate network path, combine the deployment relationship between the task cluster and the target computing node, the migration relationship between the task cluster and the resource node, the resource adjustment relationship corresponding to the task cluster and the network path switching relationship corresponding to the task cluster to generate a candidate action set consisting of deployment action, migration action, scaling up and down action and rerouting action. S44. Calculate the changes in node load balance, link congestion, task response latency, node energy consumption, migration overhead, and service level agreement default risk after each candidate action is executed. S45. Determine the resource adjustment effect based on the changes in node load balancing and link congestion after each candidate action is executed; determine the service guarantee effect based on the changes in task response latency and service level agreement default risk; determine the action execution cost based on the changes in node energy consumption and migration overhead; and construct a benefit evaluation model corresponding to each candidate action based on the resource adjustment effect, service guarantee effect, and action execution cost. S46. Based on the representative agents, local state vectors, mean field state vectors, candidate action sets, and benefit evaluation models, determine the action response relationships of each representative agent within the target scheduling period and construct a mean field game model.

7. The method for dynamic scheduling of cloud computing resources based on artificial intelligence according to claim 1, characterized in that, S5 specifically includes: S51. Concatenate and encode the local state vector and mean field state vector of each representative agent to generate the conditional state features of each representative agent. S52. Input the conditional state features into the conditional generation flow network scheduling strategy model, perform conditional control encoding on the conditional state features, and generate an action generation condition vector. S53. Based on the action generation condition vector, starting from the empty action state, generate action components that match the current action type in the order of action type, target topology level resource domain, target resource node, target network path and resource adjustment parameters, and input the generated action components back into the condition generation flow network scheduling strategy model until the end marker is output, and generate complete candidate scheduling actions. S54. Repeat S53 to generate multiple sets of candidate scheduling actions for the same representative agent, and deduplicate the generated results to form a set of candidate scheduling actions corresponding to each representative agent. S55. Input each candidate scheduling action and its corresponding local state vector and mean field state vector into the conditional generation flow network scheduling strategy model, evaluate the resource adjustment effect, service guarantee effect and action execution cost after each candidate scheduling action is executed, and generate the value evaluation result corresponding to each candidate scheduling action. S56. Output the candidate scheduling actions of each representative agent in the current scheduling cycle and the value evaluation results corresponding to the candidate scheduling actions.

8. The method for dynamic scheduling of cloud computing resources based on artificial intelligence according to claim 1, characterized in that, S6 specifically includes: S61. Based on the value evaluation results of the candidate scheduling actions corresponding to each representative agent, determine the target scheduling action of each representative agent in the current scheduling cycle. S62. Based on the target scheduling actions corresponding to each representative agent, update the deployment relationship, migration relationship, resource allocation relationship and network path relationship between task clusters and resource nodes to obtain the mapping relationship between task clusters and resource nodes corresponding to the current scheduling cycle. S63. Based on the updated mapping relationship between task clusters and resource nodes, re-statistically analyze the computing resource load status, storage resource load status, network link congestion status, task queue status, and service risk status within each topology level resource domain, and generate an updated basic mean field state vector. S64. Based on the updated base mean field state vector and the state change information within the continuous historical scheduling period, generate the updated mean field state vector. S65. Re-input the updated mean field state vector into the conditional generation flow network scheduling strategy model to generate the next round of candidate scheduling actions and corresponding value evaluation results, and repeat the target scheduling action determination, mapping relationship update and mean field state vector update. S66. Compare the changes in the mapping relationship between task clusters and resource nodes and the changes in the mean field state vector during two adjacent iterations. When the changes in the mapping relationship are less than the first convergence threshold and the changes in the mean field state vector are less than the second convergence threshold, stop the iteration to obtain the mean field balanced scheduling strategy.

9. The method for dynamic scheduling of cloud computing resources based on artificial intelligence according to claim 1, characterized in that, S8 specifically includes: S81. Collect the node central processing unit utilization, memory usage, storage usage, network link bandwidth usage, task response latency, node energy consumption, migration overhead, and service level protocol satisfaction rate after each representative intelligent agent executes the target scheduling action, and generate scheduling execution status data. S82. Associate the scheduling execution status data with the value assessment results corresponding to the target scheduling action, the mapping relationship between task clusters and resource nodes, and the mean field state vector to generate scheduling feedback data; S83. Based on the scheduling feedback data, extract the changes in resource adjustment effect, service guarantee effect, and action execution cost before and after the execution of the target scheduling action, and generate model update samples. S84. Update the model of the sample input condition generation flow network scheduling strategy model, and update the condition encoding parameters, action generation parameters and value evaluation parameters online. S85. Reuse the updated conditional generation flow network scheduling strategy model for the next scheduling cycle's candidate scheduling action generation and value evaluation results output. S86. Repeat the above-mentioned scheduling feedback data construction, model update sample generation and condition generation flow network scheduling strategy model online update, and output the dynamic scheduling results of cloud computing resources.

Images