A cloud-computing-based distributed node resource dynamic scheduling method

Abstract

The application discloses a kind of based on cloud computing's distributed node resource dynamic scheduling method, comprising: collecting node scheduling data and pre-processing, form standardized data set;Resource-task-link heterogeneous dynamic graph is constructed, input improved space-time graph neural rough differential equation network;The input of predicted state and measured state of MAPPO is constructed, and multi-agent scheduling state is generated;Determine candidate action set, generate action matching matrix;Calculate candidate action multi-objective advantage, update joint scheduling strategy;Sort candidate action set, form final scheduling strategy and issue;Collect execution feedback data, adjust next round prediction and strategy parameter.The application introduces resource-task-link heterogeneous dynamic graph, improves space-time graph neural rough differential equation network and MAPPO algorithm, realizes the dynamic evolution prediction of distributed node resource under cloud computing environment, multi-agent joint scheduling and closed-loop optimization management.

Description

Technical Field

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

[0002] With the development of cloud computing platforms, containerized deployments, virtualized resource pools, and distributed service architectures, an increasing number of business systems are running in cloud-based distributed environments composed of compute nodes, storage nodes, network links, and task instances. Existing cloud computing resource scheduling methods typically manage task deployment, instance scaling, resource quota adjustments, and node load balancing by monitoring node resources, task queues, and link status. Common approaches include rule-based scheduling, heuristic task allocation, reinforcement learning resource decision-making, and graph model node relationship modeling. These methods can alleviate problems such as single-node overload, task backlog, and uneven resource utilization, and update the cloud platform resource status by periodically collecting runtime data.

[0003] Existing distributed node resource scheduling methods still have shortcomings. Most existing methods treat node resource status, task execution status, and link transmission status separately, lacking a unified dynamic graph modeling of the heterogeneous relationships between resource nodes, task nodes, and link nodes. This makes it difficult to express the impact of communication links, service calls, task dependencies, resource contention, and migration candidate relationships on node resource evolution. Existing prediction models typically predict future resource states only based on historical load, rarely distinguishing resource evolution differences under different relational subgraphs and scheduling action conditions. They also lack early identification of nodes approaching resource boundaries and changes in pressure risk, leading to delayed scheduling decisions. Existing reinforcement learning scheduling methods typically generate scheduling policies using a single agent or a single comprehensive reward approach, making them difficult to adapt to multi-node collaborative scheduling scenarios. When multiple scheduling objectives conflict, they are prone to unstable action selection, insufficient inter-node coordination, and repeated resource migration.

[0004] Therefore, how to provide a dynamic scheduling method for distributed node resources based on cloud computing 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 distributed node resources based on cloud computing. This invention utilizes a heterogeneous dynamic graph of resources, tasks, and links, an improved spatiotemporal graph neural rough differential equation network, and the MAPPO algorithm. It details the implementation process of distributed node operation scheduling data acquisition, heterogeneous dynamic graph construction, node resource evolution prediction, node pressure and risk identification, candidate action matching, multi-agent joint scheduling strategy update, and execution feedback closed-loop adjustment. This achieves collaborative modeling and dynamic scheduling of distributed node resource states, task states, and link states in a cloud computing environment. Compared to traditional resource scheduling methods based on current load or single-agent decisions, this invention has the advantages of complete node relationship characterization, continuous resource evolution prediction, accurate scheduling action matching, strong multi-node collaborative capabilities, and high resource scheduling stability.

[0006] A method for dynamic scheduling of distributed node resources based on cloud computing according to an embodiment of the present invention includes: Collect distributed node operation scheduling data and perform preprocessing to form a standardized scheduling dataset; construct a resource-task-link heterogeneous dynamic graph based on the standardized scheduling dataset, and input it into an improved spatiotemporal graph neural coarse differential equation network. Through node relationship type separation coarse evolution layer, action condition coarse evolution layer and resource boundary crossing prediction layer, respectively, evolve various relationship subgraphs and action conditions to generate node resource evolution representation matrix and node pressure risk embedding; Based on the node resource evolution representation matrix, the predicted state input of each node agent in the MAPPO algorithm is constructed, and the measured state input is constructed based on the standardized scheduling dataset. The multi-agent scheduling state representation is generated by predictive-measured residual dual-branch state encoding. Based on the multi-agent scheduling state representation and the node resource evolution representation matrix, the candidate action set is determined, and the priority of the candidate actions is calculated by combining the node pressure risk embedding to generate the action matching matrix. Based on the candidate action set and action matching matrix, the multi-objective advantage of each node agent corresponding to the candidate action is calculated by the vector advantage decomposition and target conflict diversion update of the MAPPO algorithm, and the joint scheduling strategy is updated. The candidate action set is sorted according to the joint scheduling strategy to form the final scheduling strategy and then issued for execution; Collect node operation feedback data after the final scheduling strategy is executed, generate scheduling feedback records, adjust the next round of node resource evolution prediction and joint scheduling strategy parameters, and update the standardized scheduling dataset and resource-task-link heterogeneous dynamic graph.

[0007] Optionally, forming a standardized scheduling dataset includes: According to the scheduling and collection cycle, the node resource records, task operation records, link transmission records, service level records and historical scheduling records of each distributed node are read to form the original operation scheduling data table; Perform unified encoding on the node number, task number, link number, timestamp, and scheduling round identifier in the original operation scheduling data table to generate an indexed operation scheduling data table; Align the execution time of the indexed runtime scheduling data table according to the scheduling round, and write the node resource records, task execution records, link transmission records and service level records within the same scheduling round into the same data fragment; Perform padding, deduplication, and filtering on missing, duplicate, and out-of-bounds fields in the data fragments to generate a cleaned runtime scheduling data table; Normalization is performed on the numerical fields in the cleaned runtime scheduling data table, and encoding is performed on the category fields to form a standardized scheduling dataset.

[0008] Optionally, the construction of the resource-task-link heterogeneous dynamic graph includes: Resource node identifiers, task node identifiers, and link node identifiers are extracted from the standardized scheduling dataset, and resource nodes, task nodes, and link nodes are generated respectively to form a heterogeneous node set; Based on the node communication relationships, task deployment relationships, service call relationships, task dependency relationships, resource competition relationships, and migration candidate relationships in the heterogeneous node set, a heterogeneous edge set is formed; Configure node feature tables for heterogeneous node sets and edge feature tables for heterogeneous edge sets based on standardized scheduling datasets; According to the scheduling rounds, the heterogeneous node set, heterogeneous edge set, node feature table and edge feature table are combined to form a resource-task-link heterogeneous dynamic graph. The resource-task-link heterogeneous dynamic graph is split according to the edge type in the heterogeneous edge set, generating communication link relationship subgraph, service call relationship subgraph, task dependency relationship subgraph, resource competition relationship subgraph, and migration candidate relationship subgraph.

[0009] Optionally, the generation of the node resource evolution representation matrix and the node pressure risk embedding includes: An improved spatiotemporal graph neural rough differential equation network is constructed, which includes a node relationship type separation rough evolution layer, an action condition rough evolution layer, and a resource boundary crossing prediction layer. The node relationship type separation coarse evolution layer reads the communication link relationship subgraph, service call relationship subgraph, task dependency relationship subgraph, resource competition relationship subgraph, and migration candidate relationship subgraph. Based on the node characteristics and edge characteristics of each relationship subgraph, it generates coarse driving paths under the corresponding relationship type, performs node state evolution on each coarse driving path, and generates a node relationship separation evolution representation. An action condition vector is generated based on historical scheduling records and node action types in the standardized scheduling dataset. The node action types include task migration, instance expansion, instance shrinkage, resource adjustment, node wake-up, and node hibernation. The action condition coarse evolution layer reads the node relationship separation evolution representation and action condition vector, performs coarse evolution calculation on the resource state of each node under different action conditions, and generates action condition evolution representation and node pressure change representation; The resource boundary crossing prediction layer reads the action condition evolution representation and node pressure change representation, calculates the resource boundary approach state and boundary crossing time of each node in the next scheduling cycle, and generates node pressure risk embedding. The node relationship evolution representation, action condition evolution representation, and node pressure change representation are combined according to the node number to generate a node resource evolution representation matrix; The improved spatiotemporal graph neural rough differential equation network was trained, and the node resource prediction error, action response prediction error, boundary crossing prediction error and stress risk identification error were combined as joint optimization objectives. The network parameters of the node relationship type separation rough evolution layer, action condition rough evolution layer and resource boundary crossing prediction layer were continuously optimized. When the absolute value of the difference between the joint loss values ​​of two adjacent rounds in five consecutive training rounds is less than 0.001, the improved spatiotemporal graph neural rough differential equation network is considered to have completed convergence training.

[0010] Optionally, the generation of the multi-agent scheduling state representation includes: Construct a set of node agents based on the resource node numbers in the resource-task-link heterogeneous dynamic graph, and establish a correspondence between the node agent numbers and the resource node numbers. Extract the resource evolution representation of the corresponding node from the node resource evolution representation matrix according to the node agent number, and arrange them according to the state dimension of each node agent to form the predicted state input of each node agent. Extract the current running schedule data of the corresponding node from the standardized scheduling dataset according to the node agent number, and form the measured state input of each node agent; In the dual-branch state coding process of prediction-measured residual, the predicted state input is sent to the prediction state coding branch, and the measured state input is sent to the measured state coding branch to generate the prediction state coding vector and the measured state coding vector, respectively. The resource evolution residual vector is generated by calculating the difference between the predicted state coding vector and the measured state coding vector through the residual calculation branch. The predicted state encoding vector, the measured state encoding vector, and the resource evolution residual vector are combined according to the node agent number to generate a multi-agent scheduling state representation.

[0011] Optionally, determining the candidate action set and generating the action matching matrix includes: Read the multi-agent scheduling state representation according to the node agent number, and establish a node scheduling state table; Based on the node scheduling status table and the node resource evolution representation matrix, determine the nodes to be scheduled, the nodes that can be taken over, and the resource items to be adjusted, and generate the node scheduling object table; A set of candidate actions is generated based on the node scheduling object table. The set of candidate actions includes task migration, instance expansion, instance shrinkage, resource adjustment, node wake-up, and node hibernation. Each candidate action in the candidate action set is matched with the node pressure risk embedding of the corresponding node, and the priority of the candidate action is calculated. An action matching matrix is ​​constructed based on the candidate action number, node agent number, and candidate action priority. The rows of the action matching matrix correspond to the node agent number, the columns correspond to the candidate action number, and the elements record the corresponding candidate action priority.

[0012] Optionally, the updated joint scheduling policy includes: Input the candidate action set and action matching matrix into the MAPPO algorithm to generate the candidate action probability distribution of each node agent according to the node agent number; Based on the multi-agent scheduling state representation and candidate action probability distribution, calculate the target value vector of each node agent under the conditions of task latency, resource utilization, load balancing, migration cost, service default risk and energy consumption target; Based on the target value vector, action matching matrix, and node resource evolution representation matrix, calculate the multi-objective advantage vector of each node agent for the corresponding candidate action; In the process of vector dominance decomposition and target conflict diversion update, the multi-target dominance vector is divided into positive dominance components, negative dominance components and neutral dominance components to generate target conflict diversion results. Based on the target conflict diversion results, adjust the probability distribution of candidate actions of each node agent and update the joint scheduling strategy.

[0013] Optionally, forming the final scheduling strategy includes: Read the updated selection probability of each node agent in the joint scheduling strategy and establish a mapping relationship with the candidate action set to generate a candidate action ranking table; The candidate action set is sorted according to the candidate action sorting table, and mutually exclusive candidate actions corresponding to the same task, the same resource node, and the same link are filtered out to generate a list of actions that can be issued. Extract scheduling objects, scheduling action types, target nodes, and resource adjustment amounts from the list of actions that can be issued, and generate the final scheduling strategy; The final scheduling policy is converted into node scheduling instructions and distributed to the corresponding distributed nodes according to the resource node number; Receive the scheduling instruction reception status and execution start status returned by the distributed nodes, and generate the final scheduling policy distribution record.

[0014] Optionally, updating the standardized scheduling dataset and the resource-task-link heterogeneous dynamic graph includes: Collect node resource status, task execution status, link transmission status, and policy execution status after the final scheduling policy is executed to form node operation feedback data; The node operation feedback data is collected according to the scheduling round and node number to generate scheduling feedback records; Extract node resource changes, task migration results, instance adjustment results, link status changes, and service breach records from the scheduling feedback records to generate a feedback feature table; Adjust the next round of node resource evolution prediction parameters and joint scheduling strategy parameters based on the feedback feature table; Write the feedback feature table into the standardized scheduling dataset, and update the resource node status, task node status, link node status, and heterogeneous edge connection status according to the feedback feature table to form an updated resource-task-link heterogeneous graph.

[0015] The beneficial effects of this invention are: This invention proposes a cloud computing-based distributed node resource dynamic scheduling method. By constructing a heterogeneous dynamic graph of resources, tasks, links, and scheduling relationships in the distributed node operation scheduling data, it unifies the representation of resource status, task status, link status, and scheduling relationships. This enables collaborative modeling of communication links, service calls, task dependencies, resource contention, and migration candidate relationships among resource nodes, task nodes, and link nodes. Compared to traditional methods that schedule based solely on current node load or a single resource metric, this invention more comprehensively depicts the correlation between node resource changes and task execution and link transmission, reducing problems such as unreasonable task allocation, local node overload, and wasted link resources caused by insufficient representation of node relationships.

[0016] This invention improves the spatiotemporal graph neural coarse differential equation network to perform relational separation evolution, action condition evolution, and resource boundary crossing prediction on heterogeneous dynamic graphs of resources, tasks, and links. It generates a node resource evolution representation matrix and a node pressure risk embedding, enabling the prediction of node resource state changes and pressure risk changes before scheduling strategies are generated. Compared to methods that rely solely on historical load trends for prediction, this invention can distinguish resource evolution differences under different relational subgraphs and action conditions, and identify nodes approaching resource boundaries in advance. This reduces the risks of scheduling strategy lag, inaccurate resource pressure transfer, and re-congestion after task migration.

[0017] This invention combines the MAPPO algorithm to construct a multi-agent scheduling state representation. It fuses node resource evolution predictions and current measured states through a prediction-measured residual dual-branch state encoding. Furthermore, it calculates the multi-objective advantage of each node agent's candidate actions through vector advantage decomposition and target conflict diversion updates. This enables joint scheduling strategy updates across task latency, resource utilization, load balancing, migration costs, service default risk, and energy consumption targets. Compared to single-agent reinforcement learning or single comprehensive reward scheduling methods, this invention enhances the collaborative decision-making capabilities among multiple distributed nodes, reduces action selection oscillations and repeated resource migrations, and continuously adjusts the next round of node resource evolution predictions and joint scheduling strategy parameters through execution feedback. This improves the stability, balance, and adaptability of distributed node resource scheduling in a cloud computing environment. Attached Figure Description

[0018] 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: Figure 1 This is a flowchart of a distributed node resource dynamic scheduling method based on cloud computing proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved spatiotemporal graph neural rough differential equation network for a cloud computing-based distributed node resource dynamic scheduling method proposed in this invention. Figure 3 This is a schematic diagram of the MAPPO algorithm, which is a distributed node resource dynamic scheduling method based on cloud computing proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figure 1 , Figure 2 and Figure 3A method for dynamic scheduling of distributed node resources based on cloud computing, comprising: Collect distributed node operation scheduling data and perform preprocessing to form a standardized scheduling dataset; construct a resource-task-link heterogeneous dynamic graph based on the standardized scheduling dataset, and input it into an improved spatiotemporal graph neural coarse differential equation network. Through node relationship type separation coarse evolution layer, action condition coarse evolution layer and resource boundary crossing prediction layer, respectively, evolve various relationship subgraphs and action conditions to generate node resource evolution representation matrix and node pressure risk embedding; Based on the node resource evolution representation matrix, the predicted state input of each node agent in the MAPPO algorithm is constructed, and the measured state input is constructed based on the standardized scheduling dataset. The multi-agent scheduling state representation is generated by predictive-measured residual dual-branch state encoding. Based on the multi-agent scheduling state representation and the node resource evolution representation matrix, the candidate action set is determined, and the priority of the candidate actions is calculated by combining the node pressure risk embedding to generate the action matching matrix. Based on the candidate action set and action matching matrix, the multi-objective advantage of each node agent corresponding to the candidate action is calculated by the vector advantage decomposition and target conflict diversion update of the MAPPO algorithm, and the joint scheduling strategy is updated. The candidate action set is sorted according to the joint scheduling strategy to form the final scheduling strategy and then issued for execution; Collect node operation feedback data after the final scheduling strategy is executed, generate scheduling feedback records, adjust the next round of node resource evolution prediction and joint scheduling strategy parameters, and update the standardized scheduling dataset and resource-task-link heterogeneous dynamic graph.

[0021] In this embodiment, forming a standardized scheduling dataset includes: According to the scheduling collection cycle, the node resource records, task operation records, link transmission records, service level records and historical scheduling records of each distributed node are read to form the original operation scheduling data table. The scheduling collection cycle is specifically the time interval at which the system reads operation scheduling data from each distributed node, which is set to 10 seconds. Each collection is completed to form a scheduling round. Perform unified encoding on the node number, task number, link number, timestamp, and scheduling round identifier in the original operation scheduling data table to generate an indexed operation scheduling data table; Align the execution time of the indexed runtime scheduling data table according to the scheduling round, and write the node resource records, task execution records, link transmission records and service level records within the same scheduling round into the same data fragment; Perform padding, deduplication, and filtering on missing, duplicate, and out-of-bounds fields in the data fragments to generate a cleaned runtime scheduling data table; Normalization is performed on the numerical fields in the cleaned runtime scheduling data table, and encoding is performed on the category fields to form a standardized scheduling dataset.

[0022] In this embodiment, the construction of the resource-task-link heterogeneous dynamic graph includes: Resource node identifiers, task node identifiers, and link node identifiers are extracted from the standardized scheduling dataset, and resource nodes, task nodes, and link nodes are generated respectively to form a heterogeneous node set; Based on the standardized scheduling dataset, the node communication relationships, task deployment relationships, service invocation relationships, task dependency relationships, resource competition relationships, and migration candidate relationships among heterogeneous nodes are determined, forming a heterogeneous edge set, where: A heterogeneous edge set is formed, specifically as follows: The communication relationship between nodes is determined based on the source resource node number, target resource node number, link bandwidth, link delay and link packet loss rate in the link transmission record. The source resource node, target resource node and corresponding link node are connected to generate a communication link edge. The link packet loss rate is the ratio of the number of unsuccessfully arrived data packets to the number of sent data packets in the link transmission record. Based on the task node number, the current host resource node number and the task running status in the task running record, determine the task deployment relationship, connect the task node with the current host resource node, and generate a task deployment edge; The service call relationship is determined based on the task call record in the task execution record. The task node that initiates the call is connected to the task node that is called to generate a service call edge. The task call record includes the call initiating task number, the call receiving task number, and the number of calls. Based on the task execution order and the preceding task number in the task execution record, determine the task dependency relationship, connect the subsequent task node with the preceding task node, and generate task dependency edges. Based on node resource records and task execution records, resource competition relationships are determined. Task nodes with overlapping resource usage on the same resource node are connected to that resource node to generate resource competition edges. Here, overlapping resource usage means that multiple tasks simultaneously occupy any one of the resources of the same resource node, namely CPU, memory, storage, and network bandwidth, in the same scheduling round. Based on historical scheduling records, node resource records, task execution records, and link transmission records, migration candidate relationships are determined. Task nodes are connected to candidate resource nodes to generate migration candidate edges. Among them, candidate resource nodes are resource nodes that meet the task resource request amount and have a communication link edge with the resource node currently carried by the task. The communication link edge, task deployment edge, service call edge, task dependency edge, resource contention edge, and migration candidate edge are combined to form a heterogeneous edge set; Based on the standardized scheduling dataset, a node feature table is configured for the heterogeneous node set, and an edge feature table is configured for the heterogeneous edge set. The node feature table records the status fields of resource nodes, task nodes, and link nodes in each scheduling round, and the edge feature table records the connection relationship, edge type, and scheduling round identifier of various types of edges in the heterogeneous edge set. According to the scheduling rounds, the heterogeneous node set, heterogeneous edge set, node feature table and edge feature table are combined to form a resource-task-link heterogeneous dynamic graph. The resource-task-link heterogeneous dynamic graph is split according to the edge type in the heterogeneous edge set, generating communication link relationship subgraphs, service call relationship subgraphs, task dependency relationship subgraphs, resource contention relationship subgraphs, and migration candidate relationship subgraphs, among which: The resource-task-link heterogeneous dynamic graph is split according to the edge type in the heterogeneous edge set, specifically as follows: Extract the corresponding resource nodes and link nodes according to the communication link edges, and generate a communication link relationship subgraph; Extract task nodes with call relationships based on service call edges, and generate a service call relationship subgraph; Extract task nodes with dependencies based on task dependency edges, and generate a task dependency subgraph. Extract task nodes and resource nodes that have resource competition relationships according to the resource competition edges, and generate a resource competition relationship subgraph; Extract task nodes and candidate resource nodes based on migration candidate edges, and generate a migration candidate relationship subgraph; By binding the node number, edge number, node characteristics, and edge characteristics of each relational subgraph to the corresponding scheduling round, a set of relational subgraphs under each scheduling round is obtained.

[0023] In this embodiment, the generation of the node resource evolution representation matrix and the node pressure risk embedding includes: An improved spatiotemporal graph neural rough differential equation network is constructed, comprising a node relationship type separation rough evolution layer, an action condition rough evolution layer, and a resource boundary traversal prediction layer, wherein: Construct an improved spatiotemporal graph neural rough differential equation network, specifically as follows: The traditional spatiotemporal graph neural rough differential equation network includes a node temporal path encoding layer, a spatial graph rough evolution layer, and a resource state output layer. Based on the traditional spatial graph rough evolution layer, the node relationship evolution method based on a single graph structure is replaced with a node relationship type separation rough evolution layer. Rough evolution is performed on the communication link relationship subgraph, service call relationship subgraph, task dependency relationship subgraph, resource competition relationship subgraph, and migration candidate relationship subgraph respectively. After the node relationship type separation rough evolution layer, an action condition rough evolution layer is added. The action condition vector and the node relationship separation evolution representation are input together and the node resource evolution under the action condition is calculated. Before the resource state output layer, a resource boundary crossing prediction layer is added. Based on the action condition evolution results, the state of the node approaching the resource boundary is calculated and the node pressure risk embedding is generated. The three layers are connected to the original spatiotemporal graph neural rough differential equation network module in a sequential connection manner to obtain the improved spatiotemporal graph neural rough differential equation network. The node relationship type separation coarse evolution layer reads the communication link relationship subgraph, service call relationship subgraph, task dependency relationship subgraph, resource contention relationship subgraph, and migration candidate relationship subgraph. Based on the node features and edge features of each relationship subgraph, it generates coarse driving paths under the corresponding relationship types. It then performs node state evolution on each coarse driving path to generate a node relationship separation evolution representation, where: The coarse evolution layer for separating node relationship types includes: Relationship type index table: stores relation type identifier, subgraph number, and scheduling round identifier; Subgraph node cache: Stores the node numbers and node characteristics of each relational subgraph; Subgraph edge feature cache: Stores the edge number, edge type, and edge features of each relational subgraph; Coarse-driven path generator: Receives node features and edge features from consecutive scheduling rounds and generates coarse-driven paths; Relation Independent Evolution Calculator: Receives coarse driving paths corresponding to each relation type and calculates the node evolution state; Evolutionary Status Register: Stores node number, relation type identifier, and node evolution status; Relationship Separation and Convergence Table: Summarizes the node evolution states under different relationship types according to node number, forming a node relationship separation and evolution representation; In the node relation type separation coarse evolution layer, the relation type index table determines the relation type identifier and subgraph number of each relation subgraph. The subgraph node cache area and the subgraph edge feature cache area send the node features and edge features of the corresponding relation subgraph to the coarse driving path generator to generate the coarse driving path corresponding to each relation type. The relation independent evolution calculator receives the coarse driving path and calculates the node evolution state. The evolution state register stores the node evolution state according to the node number and relation type identifier. The relation separation aggregation table reads the node evolution state in the evolution state register and summarizes it according to the node number to form the node relation separation evolution representation. Based on the node and edge features of each relation subgraph, a coarse driving path is generated for the corresponding relation type, specifically as follows: Read the communication link relationship subgraph, service call relationship subgraph, task dependency relationship subgraph, resource contention relationship subgraph and migration candidate relationship subgraph in sequence according to the relationship type index table, and extract the node features and edge features of each relationship subgraph under consecutive scheduling rounds. According to the scheduling round order, the node features and edge features in the same relational subgraph are aligned and arranged according to the node number, edge number and feature dimension to form a relational feature sequence of the corresponding relational type. The numerical values ​​corresponding to each scheduling round, node number, edge number and feature dimension in the relational feature sequence are relational feature values. Read the relation feature sequence of the same relation subgraph under two adjacent scheduling rounds, subtract the node feature value in the later scheduling round from the node feature value in the earlier scheduling round item by item to obtain the change in node feature, and subtract the edge feature value in the later scheduling round from the edge feature value in the earlier scheduling round item by item to obtain the change in edge feature. The node feature changes and edge feature changes are arranged according to node number, edge number, and feature dimension to form single-round relationship feature increment segments. The single-round relationship feature increment segments corresponding to multiple consecutive scheduling rounds are spliced ​​together in chronological order to generate a relationship feature increment sequence. According to the same relation type identifier, node number, edge number and scheduling round identifier, the relation feature sequence and relation feature increment sequence are aligned in position, and the relation feature value and the corresponding relation feature increment value under each scheduling round are written into the same coarse path unit. Multiple coarse path units are connected according to the scheduling round order to form coarse path segments of corresponding relationship types, and a coarse path index table is generated based on the scheduling round identifier, relationship type identifier and node number in the coarse path segments. According to the relation type identifier, write the coarse path fragment into the coarse driving path generator to generate coarse driving paths for communication links, service calls, tasks, resources, and migration candidates, respectively. The node state evolution is performed for each coarse driving path, specifically as follows: Read the coarse driving paths of communication links, service calls, tasks, resources, and migration candidates according to the coarse path index table, and send them to the relation-independent evolution calculator according to the relation type identifier. Based on the node number in each coarse driving path, the initial node state is extracted from the node features of the corresponding subgraph and written into the evolution state register. Read the coarse path units in the coarse driving path according to the scheduling round order, input the relation feature value and relation feature increment value of the current scheduling round into the relation independent evolution calculator, and calculate the node state increment under the corresponding relation type. The node state increment is updated by comparing it with the node evolution state in the evolution state register from the previous scheduling round, and the node evolution state for the current scheduling round is generated. The node evolution states of multiple consecutive scheduling rounds under the same relation type are written sequentially to form a node evolution state sequence under the corresponding relation type; Write the node evolution state sequences corresponding to communication link relationships, service call relationships, task dependency relationships, resource competition relationships, and migration candidate relationships into the evolution state register, and store them according to the node number and relationship type identifier; The relation separation and aggregation table reads the node evolution state of the same node under different relation types from the evolution state register, and aligns and combines them according to the node number to generate a node relation separation evolution representation; An action condition vector is generated based on historical scheduling records and node executable action types in the standardized scheduling dataset. The node executable action types include task migration, instance expansion, instance reduction, resource adjustment, node wake-up, and node hibernation. Generate the action condition vector, specifically: Read the scheduling round identifier, resource node number, task node number, historical action type and action execution status from the historical scheduling records to form a historical action record table; Set action type identifiers according to the types of actions that nodes can execute, and map task migration, instance expansion, instance shrinkage, resource adjustment, node wake-up and node hibernation to the corresponding action type identifiers respectively; Based on the resource node number, task node number, action type identifier, and action execution status in the historical action record table, generate the action condition field for each resource node in the corresponding scheduling round. Arrange the action condition fields according to resource node number, task node number, and scheduling round identifier to form an action condition sequence; Encode the action type identifier and action execution state in the action condition sequence to generate an action condition vector; The action condition coarse evolution layer reads the node relationship separation evolution representation and action condition vector, performs coarse evolution calculations on the resource state of each node under different action conditions, and generates an action condition evolution representation and a node pressure change representation, where: The coarse evolutionary layer of action conditions includes: Action Type Index Table: Stores action type identifiers for task migration, instance expansion, instance shrinkage, resource adjustment, node wake-up, and node hibernation; Action condition vector cache: stores action condition vectors, resource node numbers, task node numbers, and scheduling round identifiers; Node relationship representation cache: stores the separated evolutionary representation of node relationships, node numbers, and relationship type identifiers; Action-Node Matching Table: Records the matching relationship between action type identifiers and corresponding resource nodes and task nodes; Action-driven path generator: Receives action condition vectors and node relationship separation evolutionary representations, and generates coarse-grained action condition-driven paths; Action Condition Evolution Calculator: Receives coarse driving paths for action conditions and calculates the evolution state of node resources under different action conditions; Pressure Change Register: Stores the difference in node resource status and the change in node pressure before and after the action is executed; Action evolution output buffer: Stores action condition evolution representations and node pressure change representations according to node number, action type identifier, and scheduling round identifier; In the action condition coarse evolution layer, the action type index table and the action condition vector buffer together form the action condition input, and establish a correspondence with the node relationship separation evolution representation output by the node relationship representation buffer in the action-node matching table. The action-node matching table passes the corresponding action condition input and node relationship separation evolution representation to the action-driven path generator. The action-driven path generator outputs the action condition coarse driving path to the action condition evolution calculator to generate the node resource evolution state of each node under different action conditions. The node resource evolution state is written into the action evolution output buffer to form the action condition evolution representation, and is also sent to the pressure change register to calculate the node pressure change value. The pressure change register writes the node pressure change value into the action evolution output buffer to form the node pressure change representation. A coarse evolution calculation is performed on the resource state of each node under different action conditions, specifically as follows: Read the node relationship separation evolution representation and action condition vector corresponding to the same resource node number, task node number and scheduling round identifier in the action-node matching table. Arrange the values ​​of each dimension in the node relationship separation evolution representation with the action type identifier, action execution status, resource node number and task node number in the action condition vector in a fixed field order to generate a joint action status vector. Read the joint vector of action state under two adjacent scheduling rounds in the order of scheduling rounds, subtract the value of each dimension of the later scheduling round from the corresponding value of the previous scheduling round item by item, and generate the action state increment vector. Write the joint vector of action state, the incremental vector of action state, and the scheduling round identifier into the same coarse path unit of action condition, and connect the coarse path units of action condition corresponding to multiple consecutive scheduling rounds in chronological order to form a coarse driving path of action condition. The action condition evolution calculator reads the node evolution state of the previous scheduling round, the joint vector of action states of the current scheduling round, and the action state increment vector. It then inputs these three vectors into the state mapping matrix, action mapping matrix, and increment mapping matrix, respectively, to obtain the state mapping vector, action mapping vector, and increment mapping vector. The state mapping vector, action mapping vector, and incremental mapping vector are added one by one to generate the node resource evolution state of the current scheduling round through nonlinear mapping; The node resource evolution states under the conditions of task migration, instance expansion, instance shrinkage, resource adjustment, node wake-up, and node dormancy are arranged according to the action type identifier to generate an action condition evolution representation. Read the baseline node state of the corresponding node in the node relationship separation evolution representation, and subtract the node resource evolution state under the action conditions from the baseline node state item by item to obtain the node pressure change value. Arrange the node pressure change values ​​according to the resource node number, action type identifier, and scheduling round identifier to generate a node pressure change representation; The resource boundary crossing prediction layer reads the action condition evolution representation and the node pressure change representation, calculates the resource boundary approach state and boundary crossing time of each node in the next scheduling cycle, and generates a node pressure risk embedding, where: Resource boundary crossing prediction layer includes: Resource Boundary Register Set: Stores the CPU boundary, memory boundary, storage boundary, network bandwidth boundary, task queue boundary, and response latency boundary corresponding to each resource node; Action evolution input buffer: stores action condition evolution representation, node pressure change representation, resource node number, action type identifier, and scheduling round identifier; Boundary Distance Calculator: Reads the boundary values ​​in the action condition evolution representation and resource boundary register group, and calculates the distance between the resource status of each node and the corresponding resource boundary; Pressure Change Corrector: Reads the node pressure change representation, corrects the boundary distance values ​​for pressure changes, and generates a corrected boundary distance sequence; Boundary crossing time lookup table: stores the correspondence between the corrected boundary distance sequence and the boundary crossing time; Risk Embedding Generator: Reads the corrected boundary distance sequence, boundary crossing time, and action type identifier to generate node stress risk embedding; Risk embedding in the output buffer: Storage node pressure risks are embedded according to resource node number, action type identifier, and scheduling round identifier; In the resource boundary crossing prediction layer, the action evolution input buffer passes the action condition evolution representation and the node pressure change representation to the boundary distance calculator. The boundary distance calculator reads various resource boundary values ​​from the resource boundary register group, generates the boundary distance value corresponding to each node, and inputs it together with the node pressure change representation into the pressure change corrector to generate a corrected boundary distance sequence. The corrected boundary distance sequence is input into the boundary crossing time lookup table to obtain the boundary crossing time. The corrected boundary distance sequence, the boundary crossing time, and the action type identifier are input together into the risk embedding generator to generate the node pressure risk embedding and write it to the risk embedding output buffer. Calculate the resource boundary approach state and boundary traversal time of each node in the next scheduling cycle, specifically as follows: Read the predicted resource status values ​​of each resource node in the next scheduling cycle from the action condition evolution representation, and read the resource boundary values ​​of the corresponding resource node in the resource boundary register group. The predicted resource status values ​​are the values ​​of the corresponding resource dimension in the action condition evolution representation. Subtract the resource boundary values ​​from the predicted resource status values ​​one by one to generate the boundary distance values ​​corresponding to each resource dimension; Read the pressure change value corresponding to the resource dimension in the node pressure change representation, match the boundary distance value with the pressure change value according to the resource node number, action type identifier, scheduling round identifier and resource dimension identifier, and mark the pressure change value with a sign. Among them, the increase in resource consumption, the growth of the task queue and the increase in response latency correspond to positive pressure changes, while the decrease in resource consumption, the shortening of the task queue and the decrease in response latency correspond to negative pressure changes. Under the same resource dimension, subtract the corresponding positive pressure change value from the boundary distance value to obtain the corrected boundary distance value after the pressure increases. Add the absolute value of the corresponding reverse pressure change value to the boundary distance value to obtain the corrected boundary distance value after the pressure is released. Arrange the corrected boundary distance values ​​of each resource dimension in the order of scheduling rounds to generate a corrected boundary distance sequence. Read the corrected boundary distance sequence according to the scheduling rounds. When the corrected boundary distance of any resource dimension is less than or equal to 0 for the first time, the corresponding scheduling round is determined as the boundary crossing time of that resource dimension. Based on the corrected boundary distance sequence and boundary crossing time, determine the resource boundary proximity status of each resource node in the next scheduling cycle: when there is a boundary crossing time in the corrected boundary distance sequence, mark the corresponding resource dimension as crossing state; when the minimum corrected boundary distance value in the corrected boundary distance sequence is greater than 0 and less than 10% of the corresponding resource boundary value, mark the corresponding resource dimension as approaching state; when the minimum corrected boundary distance value in the corrected boundary distance sequence is greater than or equal to 10% of the corresponding resource boundary value, mark the corresponding resource dimension as not approaching state. The resource boundary proximity status, boundary crossing time, corrected boundary distance sequence, and action type identifier are arranged according to the resource node number to generate a node pressure risk embedding. The node relationship evolution representation, action condition evolution representation, and node pressure change representation are combined according to the node number to generate a node resource evolution representation matrix; An improved spatiotemporal graph neural rough differential equation network was trained, using a combination of node resource prediction error, action response prediction error, boundary crossing prediction error, and stress risk identification error as the joint optimization objective. The network parameters of the node relationship type separation rough evolution layer, action condition rough evolution layer, and resource boundary crossing prediction layer were continuously optimized. The network was considered to have completed convergence training when the absolute value of the difference between the joint loss values ​​of two adjacent rounds in five consecutive training rounds was less than 0.001. The improved spatiotemporal graph neural rough differential equation network is trained as follows: Read the resource-task-link heterogeneous dynamic graph, various relational subgraphs, action condition vectors, real node resource status, real action response results, real boundary crossing time and real pressure risk labels from the training samples. Use the data from 12 consecutive scheduling rounds as the input window and the data from the next scheduling round as the prediction label. Set the sample size per batch to 64. Input the heterogeneous dynamic graph of resource-task-link and various relational subgraphs into the node relation type separation coarse evolution layer, and output the node relation separation evolution representation. Input the node relation separation evolution representation and action condition vector into the action condition coarse evolution layer, and output the action condition evolution representation and node pressure change representation. Input the action condition evolution representation and node pressure change representation into the resource boundary crossing prediction layer, and output the node pressure risk embedding and node resource evolution representation matrix. The node resource prediction error is obtained by averaging the squared difference between the node resource evolution representation matrix and the actual node resource state. The action response prediction error is obtained by averaging the squared difference between the action condition evolution representation and the actual action response result. The boundary crossing prediction error is obtained by averaging the squared difference between the boundary crossing time and the actual boundary crossing time. The stress risk identification error is obtained by averaging the squared difference between the node stress risk embedding and the actual stress risk label. Multiply the node resource prediction error by 0.35, the action response prediction error by 0.25, the boundary crossing prediction error by 0.20, and the pressure risk identification error by 0.20, and then sum them to generate a joint loss value. Read the rate of change of the joint loss value with respect to the network parameters to obtain the gradient value. Multiply the gradient value by 0.001 and subtract it from the current network parameters to update the network parameters of the node relationship type separation coarse evolution layer, the action condition coarse evolution layer, and the resource boundary crossing prediction layer. When the absolute value of the difference between the joint loss values ​​of two adjacent rounds in five consecutive training rounds is less than 0.001, it is determined that the improved spatiotemporal graph neural coarse differential equation network has completed convergence training.

[0024] In this embodiment, the generation of the multi-agent scheduling state representation includes: Construct a set of node agents based on the resource node numbers in the resource-task-link heterogeneous dynamic graph, and establish a correspondence between the node agent numbers and the resource node numbers. Extract the resource evolution representation of the corresponding node from the node resource evolution representation matrix according to the node agent number, and arrange them according to the state dimension of each node agent to form the predicted state input of each node agent. Extract the current running schedule data of the corresponding node from the standardized scheduling dataset according to the node agent number, and form the measured state input of each node agent; In the prediction-measured residual dual-branch state coding process, the predicted state input is fed into the prediction state coding branch, and the measured state input is fed into the measured state coding branch, generating the prediction state coding vector and the measured state coding vector respectively, where: The predicted state encoding vector and the measured state encoding vector are generated as follows: Based on the node agent number and state dimension identifier, the predicted state input and the measured state input are aligned with fields to generate the predicted state field sequence and the measured state field sequence. The predicted state field sequence and the measured state field sequence are input into the predicted state coding branch and the measured state coding branch, respectively. Dimension mapping and normalization of the same coding dimension are performed in the two coding branches to generate the predicted state coding vector and the measured state coding vector, respectively. The resource evolution residual vector is generated by calculating the difference between the predicted state-coded vector and the measured state-coded vector through the residual calculation branch, where: Generate resource evolution residual vectors, specifically as follows: Align the predicted state encoding vector with the measured state encoding vector according to the node agent number and state dimension identifier; For the same node agent and the same state dimension, the predicted state code value and the measured state code value are subtracted one by one to obtain the one-dimensional residual value; Multiple one-dimensional residual values ​​corresponding to the same node agent are arranged in order of state dimension to form a node-level resource evolution residual vector. The node-level resource evolution residual vectors corresponding to each node agent are combined according to the node agent number to generate resource evolution residual vectors. The predicted state encoding vector, the measured state encoding vector, and the resource evolution residual vector are combined according to the node agent number to generate a multi-agent scheduling state representation.

[0025] In this embodiment, determining the candidate action set and generating the action matching matrix includes: Read the multi-agent scheduling state representation according to the node agent number, and establish a node scheduling state table; Based on the node scheduling status table and the node resource evolution representation matrix, the nodes to be scheduled, the nodes that can be taken over, and the resource items to be adjusted are determined, and a node scheduling object table is generated, in which: Generate a node scheduling object table, specifically as follows: Based on the measured state encoding vector and resource evolution residual vector in the node scheduling state table, resource nodes with resource deviations are selected to generate a set of nodes to be scheduled, specifically: Based on the measured state encoding vector, determine the resource occupancy status, task queue status, and response latency status of each resource node in the current scheduling round. Based on the resource evolution residual vector, determine the deviation direction and deviation magnitude of the measured state of each resource node relative to the predicted state. When a resource node meets any of the following conditions, the resource node is written into the set of nodes to be scheduled: the resource occupancy status reaches the corresponding resource boundary, the task queue status reaches the corresponding resource boundary, the response latency status reaches the corresponding resource boundary, or the absolute value of the residual value of any dimension in the resource evolution residual vector is greater than the average of the absolute values ​​of the residual values ​​of the most recent 10 scheduling rounds of that dimension. Based on the resource evolution representation of each resource node in the node resource evolution representation matrix, resource nodes with task acceptance capabilities are selected to generate a set of acceptable nodes, specifically: Based on the resource evolution representation of each resource node in the next scheduling cycle, read the remaining resource amount, task queue change amount, response latency change amount and link carrying status, and write the resource nodes with remaining resource amount not lower than the average resource request amount of the task in the task running record in the same scheduling round, task queue change amount not reaching the queue boundary, response latency change amount not reaching the upper limit of service level requirement and link carrying status not reaching the link boundary into the set of acceptable nodes. Based on the resource evolution representation and resource evolution residual vector of each resource node in the set of nodes to be scheduled, the resource items to be adjusted for the corresponding nodes are determined, specifically as follows: The evolutionary state of each resource node in the next scheduling cycle is determined based on the resource evolution representation, including CPU, memory, storage, network bandwidth, task queue, and response latency. The deviation of the measured state from the predicted state for each resource dimension is determined based on the resource evolution residual vector. When a resource dimension meets any of the following conditions, the resource dimension is written into the resource item to be adjusted of the corresponding resource node: the evolution state of the resource dimension reaches the corresponding resource boundary, or the absolute value of the resource evolution residual value of the resource dimension is greater than the average absolute value of the residual value of the same resource dimension in the last 10 scheduling rounds. Write the set of nodes to be scheduled, the set of nodes that can be taken over, and the resource items to be adjusted into the node scheduling object table according to the node agent number; A candidate action set is generated based on the node scheduling object table. This candidate action set includes task migration, instance scaling up, instance scaling down, resource adjustment, node wake-up, and node hibernation, wherein: Generate a set of candidate actions, specifically as follows: Based on the set of nodes to be scheduled, the set of nodes that can be taken over, and the resource items to be adjusted in the node scheduling object table, an action generation list is created. When a node to be scheduled has a task to be migrated and the available resources, link status and service level requirements of the node to be accommodated meet the running requirements of the task to be migrated, a task migration action is generated. When the number of instances corresponding to the resource items to be adjusted of the node to be scheduled increases, an instance expansion action is generated; When the number of instances corresponding to the resource items to be adjusted of the node to be scheduled decreases, an instance scaling-down action is generated. When the resource item to be adjusted of the node to be scheduled corresponds to at least one of CPU, memory, storage, network bandwidth, task queue and response latency, a resource adjustment action is generated; When a node is in a dormant state and is written into the set of nodes that can be accepted, a node wake-up action is generated. When a resource node is neither written into the set of nodes to be scheduled nor into the set of nodes that can take over in the current scheduling round, a node dormancy action is generated. Each candidate action is recorded according to its action type, scheduling object, target node, and resource adjustment amount to form a candidate action set; Each candidate action in the candidate action set is matched with the node pressure risk embedding of the corresponding node, and the priority of the candidate action is calculated, where: Each candidate action in the candidate action set is matched with the node pressure risk embedding of the corresponding node, specifically: Read the candidate action number, action type, scheduling object, source resource node, target resource node and resource adjustment amount for each candidate action in the candidate action set, and determine the source resource node and target resource node as action-related nodes; According to the resource node number of the action-related node, read the resource boundary approach status, boundary crossing time, corrected boundary distance sequence and action type identifier of the corresponding node from the node pressure risk embedding. Among them, the crossing status, approach status and non-approach status in the resource boundary approach status are written with the values ​​2, 1 and 0 respectively, and the boundary crossing time is recorded according to the number of rounds from the current scheduling round. The action type of the candidate action is compared with the action type identifier in the node pressure risk embedding, and the resource adjustment amount of the action-related node is aligned with the corresponding resource dimension in the correction boundary distance sequence to generate the action-risk correspondence. When the action type identifier, action associated node number and affected resource dimension of a candidate action are consistent with the corresponding fields in the node pressure risk embedding, a matching relationship is established between the candidate action and the corresponding node pressure risk embedding. Write the candidate action number, node agent number, action type identifier, resource boundary proximity status value, boundary crossing time, and corresponding resource dimension corrected boundary distance value into the candidate action matching record. The priority of candidate actions is calculated as follows: Read the resource boundary proximity status value, boundary crossing time, corrected boundary distance value, and resource adjustment amount from the candidate action matching record, and multiply the resource boundary proximity status value by 0.40 to obtain the boundary status score; When the boundary crossing time is less than or equal to 2 scheduling rounds, the write time score is 0.30; when the boundary crossing time is greater than 2 scheduling rounds but less than or equal to 5 scheduling rounds, the write time score is 0.20; when the boundary crossing time is greater than 5 scheduling rounds, the write time score is 0.10; when the corresponding resource dimension is in an uncrossed state, the write time score is 0. Read the corrected boundary distance value, and use the ratio of the corrected boundary distance value to the corresponding resource boundary value as the boundary margin ratio. When the boundary margin ratio is less than 0.10, write a distance score of 0.20; when the boundary margin ratio is greater than or equal to 0.10 and less than 0.25, write a distance score of 0.10; when the boundary margin ratio is greater than or equal to 0.25, write a distance score of 0.05. Read the resource adjustment amount, and use the ratio of the resource adjustment amount to the corresponding resource boundary value as the action adjustment ratio. When the action adjustment ratio is greater than or equal to 0.15, write an adjustment score of 0.10; when the action adjustment ratio is less than 0.15, write an adjustment score of 0.05. The boundary state score, time score, distance score and adjustment score are added together to obtain the candidate action priority, and the candidate action priority, along with the candidate action number and the node agent number, are written into the candidate action priority record. An action matching matrix is ​​constructed based on the candidate action number, node agent number, and candidate action priority. The rows of the action matching matrix correspond to the node agent number, the columns correspond to the candidate action number, and the elements record the corresponding candidate action priority.

[0026] In this embodiment, the updating of the joint scheduling strategy includes: The candidate action set and action matching matrix are input into the MAPPO algorithm to generate the candidate action probability distribution for each node agent according to the node agent number, where: The probability distribution of candidate actions for each node agent is generated as follows: Read the candidate action set and action matching matrix according to the node agent number, and extract the candidate action number, action type identifier and candidate action priority for each node agent; Arrange the priority of candidate actions corresponding to the same node agent according to the candidate action number to generate a node action scoring sequence; Read the maximum score value in the node action score sequence, subtract the maximum score value from the priority of each candidate action in the node action score sequence, and generate a stable action score sequence. Perform an exponential mapping on each score value in the stable action score sequence to obtain the action index value, and divide each action index value by the sum of all action index values ​​under the same node agent to obtain the action probability value of the corresponding candidate action. Arrange the probability values ​​of each candidate action according to the node agent number and the candidate action number to generate the probability distribution of candidate actions for each node agent. Based on the multi-agent scheduling state representation and candidate action probability distribution, calculate the target value vector for each node agent under the objectives of task latency, resource utilization, load balancing, migration cost, service default risk, and energy consumption, where: The target value vector of each node agent is calculated as follows: Read the multi-agent scheduling state representation and candidate action probability distribution according to the node agent number, and arrange the scheduling state representation and candidate action probability value corresponding to the same node agent to form the node action state input; Calculate the target value of task latency based on the task completion time and service level requirements in the node action status input; calculate the target value of resource utilization based on resource occupancy and resource capacity; and calculate the target value of load balancing based on the load differences of each resource node. The migration cost target value is calculated based on the task migration data volume corresponding to the candidate action and the link transmission status between the source resource node and the target resource node. The service default risk target value is calculated based on the difference between the task response latency and the service level requirement. The energy consumption target value is calculated based on the resource node running power consumption and the action execution power consumption. The target values ​​are arranged into a target value vector according to the target order of task latency, resource utilization, load balancing, migration cost, service default risk and energy consumption. Write the target value vector into the target value table according to the node agent number and the candidate action number; Based on the target value vector, action matching matrix, and node resource evolution representation matrix, calculate the multi-objective advantage vector for each node agent's candidate actions, where: The multi-objective advantage vector for each node agent corresponding to the candidate action is calculated as follows: Read the target value vector in the target value table according to the node agent number and the candidate action number, and read the corresponding candidate action priority from the action matching matrix; The average value vector of all candidate actions under the same node agent is calculated according to the target dimension to generate a baseline target value vector. The target value vectors of candidate actions are compared with the baseline target value vectors dimension by dimension, based on the decreasing direction of the target values ​​of task latency, migration cost, service default risk and energy consumption, and the increasing direction of the target values ​​of resource utilization and load balance. Read the resource evolution representation of the corresponding resource node in the node resource evolution representation matrix, determine the resource change magnitude of the node corresponding to the candidate action based on the resource evolution representation, and multiply the resource change magnitude by the priority of the candidate action to generate the action adjustment factor; Multiply each target difference value in the target difference vector by the action adjustment factor to generate the advantages of task latency, resource utilization, load balancing, migration cost, service default risk, and energy consumption. The advantages of each objective are arranged in the order of task latency, resource utilization, load balancing, migration cost, service default risk and energy consumption, and a multi-objective advantage vector of candidate actions corresponding to each node agent is generated. In the process of vector dominance decomposition and target conflict splitting update, the multi-target dominance vector is divided into positive dominance components, negative dominance components, and neutral dominance components to generate target conflict splitting results, where: Vector advantage decomposition and target conflict splitting update are as follows: Read the multi-objective advantage vector of the candidate actions of each node agent, and extract the advantage value of each objective in the order of task latency, resource utilization, load balancing, migration cost, service default risk and energy consumption. The neutral dominance range is set to be greater than or equal to -0.02 and less than or equal to 0.02. Targets with a target dominance value greater than 0.02 are written into the positive dominance component, targets with a target dominance value less than -0.02 are written into the negative dominance component, and targets falling into the neutral dominance range are written into the neutral dominance component. The number of positive dominance components, negative dominance components, and neutral dominance components are counted according to the candidate action number, and the sum of the absolute values ​​of the positive dominance components, the negative dominance components, and the neutral dominance components is calculated. When the same candidate action has both positive and negative dominant components, the candidate action is marked as the target conflict action, and the target conflict direction is determined according to the relationship between the sum of the absolute values ​​of the positive dominant components and the sum of the absolute values ​​of the negative dominant components. When the sum of the absolute values ​​of the positive dominant components is greater than the sum of the absolute values ​​of the negative dominant components, the candidate action is written into the positive splitting result. When the sum of the absolute values ​​of the negative dominant components is greater than the sum of the absolute values ​​of the positive dominant components, the candidate action is written into the negative splitting result. When the number of neutral dominant components reaches 4, the candidate action is written into the neutral splitting result. The forward splitting results, reverse splitting results, and neutral splitting results are combined according to the node agent number and the candidate action number to generate the target conflict splitting results; Based on the target conflict diversion results, the probability distribution of candidate actions for each node agent is adjusted, and the joint scheduling strategy is updated, wherein: Update the joint scheduling policy as follows: Read the forward, reverse, and neutral splitting results of the candidate actions corresponding to each node agent in the target conflict splitting results, and read the probability distribution of the candidate actions of each node agent; When a candidate action is a positive flow result, increase the probability value of the current candidate action by 0.05; when a candidate action is a negative flow result, decrease the probability value of the current candidate action by 0.05; when a candidate action is a neutral flow result, keep the probability value of the current candidate action unchanged. Normalize the probability values ​​of all candidate actions under the same node agent after adjustment so that the sum of the probability values ​​of all candidate actions corresponding to the same node agent is 1. Write the normalized candidate action probability value into the joint policy probability table according to the node agent number and the candidate action number; The updated selection probability of each node agent for the corresponding candidate action is determined based on the joint policy probability table, and the updated selection probability is associated with the candidate action set to generate the updated joint scheduling policy.

[0027] In this embodiment, forming the final scheduling strategy includes: Read the updated selection probability of each node agent in the joint scheduling strategy and establish a mapping relationship with the candidate action set to generate a candidate action ranking table; The candidate action set is sorted according to the candidate action sorting table, and mutually exclusive candidate actions corresponding to the same task, the same resource node, and the same link are filtered out to generate a list of actions that can be issued, where: Generate a list of actions that can be issued, specifically: Arrange the candidate action set in descending order of the updated selection probability in the candidate action sorting table to generate the initial sorting sequence of candidate actions. Read the task number, source resource node number, target resource node number and link number in the initial sorting sequence of candidate actions, identify multiple migration actions corresponding to the same task, instance expansion and instance shrinkage actions corresponding to the same resource node, and multiple transmission occupation actions corresponding to the same link, and form a mutually exclusive candidate action group. In each mutually exclusive candidate action group, retain the candidate action with the highest sorting position, filter out the remaining candidate actions, and rearrange the filtered candidate actions according to their sorting position to generate a list of actions that can be issued. Extract scheduling objects, scheduling action types, target nodes, and resource adjustment amounts from the list of actions that can be issued, and generate the final scheduling strategy; The final scheduling policy is converted into node scheduling instructions and distributed to the corresponding distributed nodes according to the resource node number; Receive the scheduling instruction reception status and execution start status returned by the distributed nodes, and generate the final scheduling policy distribution record.

[0028] In this embodiment, updating the standardized scheduling dataset and the resource-task-link heterogeneous dynamic graph includes: Collect node resource status, task execution status, link transmission status, and policy execution status after the final scheduling policy is executed to form node operation feedback data; The node operation feedback data is collected according to the scheduling round and node number to generate scheduling feedback records; Extract node resource changes, task migration results, instance adjustment results, link status changes, and service breach records from the scheduling feedback records to generate a feedback feature table; Adjust the next round of node resource evolution prediction parameters and joint scheduling strategy parameters based on the feedback feature table; Write the feedback feature table into the standardized scheduling dataset, and update the resource node status, task node status, link node status, and heterogeneous edge connection status according to the feedback feature table to form an updated resource-task-link heterogeneous graph.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous resource scheduling cycle of a cloud computing platform, connecting to a resource pool consisting of 240 distributed computing nodes, 6800 task instances, and 420 inter-node links. This resource pool handles online querying, batch computing, caching services, and data synchronization tasks. The CPU request volume of task instances ranges from 0.2 to 4 cores, the memory request volume ranges from 256MB to 8192MB, and the service response latency requirements range from 80 to 500ms. Before scheduling began, 37 nodes on the platform had CPU utilization exceeding 85%, 22 nodes had task queue lengths exceeding 160, and the transmission latency of some links increased from an average of 18ms to 46ms. Traditional load balancing strategies tend to migrate tasks to seemingly idle but congested nodes, causing secondary queuing of tasks and repeated resource migration.

[0030] Distributed node operation scheduling data was collected at a 10-second interval for 180 consecutive scheduling rounds, yielding 43,200 node resource records, 6,800 task execution records, 420 link transmission records, and 12,560 historical scheduling records. After entering the preprocessing stage, node numbers, task numbers, link numbers, and scheduling round identifiers were uniformly encoded. Reporting times from different nodes were aligned to the same scheduling round, and missing, duplicate, and out-of-bounds fields were processed. The original data contained 913 missing records, 286 duplicate records, and 147 out-of-bounds records. After processing, 55,434 valid records were retained. Numerical fields were normalized to the 0–1 range, and categorical fields were converted to a unified code, forming a standardized scheduling dataset.

[0031] Based on a standardized scheduling dataset, a heterogeneous dynamic graph of resources, tasks, and links was constructed. The graph contains 240 resource nodes, 6800 task nodes, and 420 link nodes, and forms a heterogeneous edge set based on task deployment relationships, service call relationships, task dependency relationships, resource contention relationships, and migration candidate relationships. After graph construction, the communication link relationship subgraph contains 420 edges, the service call relationship subgraph contains 3180 edges, the task dependency relationship subgraph contains 2460 edges, the resource contention relationship subgraph contains 9120 edges, and the migration candidate relationship subgraph contains 15240 edges. In the 96th scheduling round, 18 links were found to have bandwidth utilization exceeding 0.82, and 31 task nodes were added to the migration candidate relationship subgraph.

[0032] An improved spatiotemporal graph neural coarse differential equation network is used as the input to a heterogeneous dynamic graph of resources, tasks, and links. A coarse evolution layer for separating node relationship types reads various relationship subgraphs and generates coarse driving paths under different relationship types. After processing by this layer, the dimension of the node relationship separation evolution representation is 240×128, where the average response value corresponding to resource competition relationships is 0.74, higher than the 0.51 corresponding to communication link relationships, indicating that the node pressure in this scheduling cycle mainly comes from resource competition. Subsequently, action condition vectors are generated based on historical scheduling records and the types of executable actions of nodes. The action condition coarse evolution layer performs evolution calculations on the resource state under the conditions of task migration, instance expansion, instance shrinkage, resource adjustment, node wake-up, and node hibernation. In the 112th scheduling round, node N037's predicted CPU utilization rate is 0.93 under no-action conditions, drops to 0.76 under instance expansion conditions, and drops to 0.71 under task migration conditions, corresponding to node pressure changes of 0.18 and 0.24, respectively. When a node in the prediction layer approaches the resource boundary, the task queue boundary crossing time for node N037 is 2 scheduling cycles, the response latency boundary crossing time is 3 scheduling cycles, and the pressure risk embedding value is 0.87. The pressure risk embedding value for node N144 is 0.29, and it is determined to be a node that can be accepted.

[0033] During the model training phase, the first 120 scheduling rounds were used as training samples, resulting in 28,800 node-level time-series samples and 118,000 task-node association samples. An additional 30 scheduling rounds were used as validation samples, and the remaining 30 rounds were used as test samples. The training batch size was set to 64, with an initial learning rate of 0.001. The joint optimization objective consisted of node resource prediction error, action response prediction error, boundary crossing prediction error, and stress risk identification error. By the 42nd round, the absolute value of the difference between adjacent joint loss values ​​in five consecutive rounds was less than 0.001, indicating that the model had achieved convergence. In the test set, the average absolute error of node CPU prediction was 0.031, the average absolute error of task queue prediction was 6.8, the average error of boundary crossing time was 0.42 scheduling cycles, and the accuracy of node stress risk identification was 94.6%.

[0034] In the MAPPO algorithm processing phase, 240 node agents are constructed based on resource node numbers. Each node agent extracts the resource evolution representation of its corresponding node from the node resource evolution representation matrix to form the predicted state input, and extracts the current running scheduling data from the standardized scheduling dataset to form the measured state input. Predictive-measured residual dual-branch state encoding generates predicted state encoding vectors and measured state encoding vectors respectively, and calculates the resource evolution residual vector. In the 112th scheduling round, the predicted CPU encoding value of node N037 is 0.91, the measured CPU encoding value is 0.95, and the residual is 0.04; the predicted CPU encoding value of node N144 is 0.48, the measured CPU encoding value is 0.46, and the residual is 0.02. The above encoding results are combined to generate a multi-agent scheduling state representation with a dimension of 240×192.

[0035] A set of candidate actions is determined based on the multi-agent scheduling state representation and node resource evolution representation matrix. In this scheduling round, a total of 782 candidate actions are generated, including 312 task migration actions, 96 instance expansion actions, 84 instance shrinkage actions, 178 resource adjustment actions, 42 node wake-up actions, and 70 node dormancy actions. Each candidate action is matched with its corresponding node pressure and risk embedding, and the priority of the candidate actions is calculated to generate an action matching matrix. The action priority for task T2156 migrating from node N037 to node N144 is 0.91, and the action priority for migrating to node N188 is 0.63. Since the link transmission latency for node N188 reaches 39ms, the action matching matrix lowers the ranking position of this action.

[0036] After the candidate action set and action matching matrix are input into the MAPPO algorithm, each node agent generates a probability distribution of candidate actions. Vector advantage decomposition and target conflict splitting update calculate the target value vector under the targets of task latency, resource utilization, load balancing, migration cost, service default risk, and energy consumption. For the action of migrating task T2156 to node N144, the task latency advantage is 0.26, the load balancing advantage is 0.31, the service default risk advantage is 0.22, the migration cost advantage is -0.08, and the energy consumption advantage is -0.03. It is divided into positive advantage components and negative advantage components, and the target conflict splitting result is generated. After the policy update, the probability of this candidate action increases from 0.38 to 0.57; the node sleep-related action shows a negative advantage under the service default risk target, and the probability decreases from 0.19 to 0.07. After sorting by the joint scheduling policy, 126 actions are finally selected for execution, including 48 task migrations, 21 instance expansions, 39 resource adjustments, 8 node wake-ups, and 10 node sleeps.

[0037] After the final scheduling strategy is executed, node operation feedback data is collected over the next six scheduling cycles. Upon completion, the number of nodes with CPU utilization exceeding 85% decreased from 37 to 11, the number of nodes with task queue lengths exceeding 160 decreased from 22 to 5, and the average link transmission latency decreased from 46ms to 24ms. Node N037's CPU utilization decreased from 0.95 to 0.68, its task queue length decreased from 184 to 93, and its response latency decreased from 418ms to 211ms. Node N144's CPU utilization increased from 0.46 to 0.62 after receiving the migration task, but remained within acceptable limits. Scheduling feedback records are generated, the feedback feature table is written to the standardized scheduling dataset, and the resource node status, task node status, link node status, and heterogeneous edge connection status are updated. In the next scheduling round, node N037's pressure risk embedding decreased from 0.87 to 0.32, and the probability of migration actions for this node in the joint scheduling strategy decreased accordingly.

[0038] On the same test set, the method of this invention was compared with traditional current load-based scheduling methods and single-agent PPO scheduling methods. The training samples for both methods consisted of 28,800 node-level time-series samples, and the test rounds were 30 scheduling rounds each. The average task response latency for the traditional method was 286ms, for the single-agent PPO method it was 231ms, and for the method of this invention it was 176ms. The average node resource utilization rate for the traditional method was 68.4%, for the single-agent PPO method it was 76.9%, and for the method of this invention it was 84.7%. The standard deviation of node load for the traditional method was 0.214, for the single-agent PPO method it was 0.158, and for the method of this invention it was 0.093. The number of task migrations per hour for the traditional method was 138, for the single-agent PPO method it was 92, and for the method of this invention it was 54. The service default rate for the traditional method was 5.8%, for the single-agent PPO method it was 3.1%, and for the method of this invention it was 1.4%. The average energy consumption of the traditional method is 412 kWh, the single-agent PPO method is 389 kWh, and the method of this invention is 354 kWh. The traditional method involves 76 tasks with repeated resource migrations during peak scheduling rounds, the single-agent PPO method involves 39, and the method of this invention involves 13.

[0039] As can be seen from Example 1, the present invention demonstrates a clear data change process in the formation of standardized scheduling datasets, construction of heterogeneous dynamic graphs, prediction of node resource evolution, matching of candidate actions, updating of multi-agent joint scheduling strategies, and feedback closed-loop adjustment. Compared with traditional methods, the present invention can more accurately identify node pressure risks in scenarios with complex node relationships, fluctuating link states, and sudden increases in task load, reducing scheduling lag and repeated resource migration, and improving task response efficiency, resource utilization, load balancing, and service stability.

[0040] The above description is only a preferred embodiment 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. A method for dynamic scheduling of distributed node resources based on cloud computing, characterized in that, include: Collect distributed node runtime scheduling data and perform preprocessing to form a standardized scheduling dataset; Based on a standardized scheduling dataset, a heterogeneous dynamic graph of resources, tasks, and links is constructed and input into an improved spatiotemporal graph neural coarse differential equation network. Through the node relationship type separation coarse evolution layer, action condition coarse evolution layer, and resource boundary crossing prediction layer, various relationship subgraphs and action conditions are evolved respectively, generating a node resource evolution representation matrix and a node pressure risk embedding. Based on the node resource evolution representation matrix, the predicted state input of each node agent in the MAPPO algorithm is constructed, and the measured state input is constructed based on the standardized scheduling dataset. The multi-agent scheduling state representation is generated by predictive-measured residual dual-branch state encoding. Based on the multi-agent scheduling state representation and the node resource evolution representation matrix, the candidate action set is determined, and the priority of the candidate actions is calculated by combining the node pressure risk embedding to generate the action matching matrix. Based on the candidate action set and action matching matrix, the multi-objective advantage of each node agent corresponding to the candidate action is calculated by the vector advantage decomposition and target conflict diversion update of the MAPPO algorithm, and the joint scheduling strategy is updated. The candidate action set is sorted according to the joint scheduling strategy to form the final scheduling strategy and then issued for execution; Collect node operation feedback data after the final scheduling strategy is executed, generate scheduling feedback records, adjust the next round of node resource evolution prediction and joint scheduling strategy parameters, and update the standardized scheduling dataset and resource-task-link heterogeneous dynamic graph.

2. The method for dynamic scheduling of distributed node resources based on cloud computing according to claim 1, characterized in that, The formation of the standardized scheduling dataset includes: According to the scheduling and collection cycle, the node resource records, task operation records, link transmission records, service level records and historical scheduling records of each distributed node are read to form the original operation scheduling data table; Perform unified encoding on the node number, task number, link number, timestamp, and scheduling round identifier in the original operation scheduling data table to generate an indexed operation scheduling data table; Align the execution time of the indexed runtime scheduling data table according to the scheduling round, and write the node resource records, task execution records, link transmission records and service level records within the same scheduling round into the same data fragment; Perform padding, deduplication, and filtering on missing, duplicate, and out-of-bounds fields in the data fragments to generate a cleaned runtime scheduling data table; Normalization is performed on the numerical fields in the cleaned runtime scheduling data table, and encoding is performed on the category fields to form a standardized scheduling dataset.

3. The method for dynamic scheduling of distributed node resources based on cloud computing according to claim 1, characterized in that, The construction of the resource-task-link heterogeneous dynamic graph includes: Resource node identifiers, task node identifiers, and link node identifiers are extracted from the standardized scheduling dataset, and resource nodes, task nodes, and link nodes are generated respectively to form a heterogeneous node set; Based on the node communication relationships, task deployment relationships, service call relationships, task dependency relationships, resource competition relationships, and migration candidate relationships in the heterogeneous node set, a heterogeneous edge set is formed; Configure node feature tables for heterogeneous node sets and edge feature tables for heterogeneous edge sets based on standardized scheduling datasets; According to the scheduling rounds, the heterogeneous node set, heterogeneous edge set, node feature table and edge feature table are combined to form a resource-task-link heterogeneous dynamic graph. The resource-task-link heterogeneous dynamic graph is split according to the edge type in the heterogeneous edge set, generating communication link relationship subgraph, service call relationship subgraph, task dependency relationship subgraph, resource competition relationship subgraph, and migration candidate relationship subgraph.

4. The method for dynamic scheduling of distributed node resources based on cloud computing according to claim 1, characterized in that, The generated node resource evolution representation matrix and node pressure risk embedding include: An improved spatiotemporal graph neural rough differential equation network is constructed, which includes a node relationship type separation rough evolution layer, an action condition rough evolution layer, and a resource boundary crossing prediction layer. The node relationship type separation coarse evolution layer reads the communication link relationship subgraph, service call relationship subgraph, task dependency relationship subgraph, resource competition relationship subgraph, and migration candidate relationship subgraph. Based on the node characteristics and edge characteristics of each relationship subgraph, it generates coarse driving paths under the corresponding relationship type, performs node state evolution on each coarse driving path, and generates a node relationship separation evolution representation. An action condition vector is generated based on historical scheduling records and node action types in the standardized scheduling dataset. The node action types include task migration, instance expansion, instance shrinkage, resource adjustment, node wake-up, and node hibernation. The action condition coarse evolution layer reads the node relationship separation evolution representation and action condition vector, performs coarse evolution calculation on the resource state of each node under different action conditions, and generates action condition evolution representation and node pressure change representation; The resource boundary crossing prediction layer reads the action condition evolution representation and node pressure change representation, calculates the resource boundary approach state and boundary crossing time of each node in the next scheduling cycle, and generates node pressure risk embedding. The node relationship evolution representation, action condition evolution representation, and node pressure change representation are combined according to the node number to generate a node resource evolution representation matrix; The improved spatiotemporal graph neural rough differential equation network was trained, and the node resource prediction error, action response prediction error, boundary crossing prediction error and stress risk identification error were combined as joint optimization objectives. The network parameters of the node relationship type separation rough evolution layer, action condition rough evolution layer and resource boundary crossing prediction layer were continuously optimized. When the absolute value of the difference between the joint loss values ​​of two adjacent rounds in five consecutive training rounds is less than 0.001, the improved spatiotemporal graph neural rough differential equation network is considered to have completed convergence training.

5. The method for dynamic scheduling of distributed node resources based on cloud computing according to claim 1, characterized in that, The generated multi-agent scheduling state representation includes: Construct a set of node agents based on the resource node numbers in the resource-task-link heterogeneous dynamic graph, and establish a correspondence between the node agent numbers and the resource node numbers. Extract the resource evolution representation of the corresponding node from the node resource evolution representation matrix according to the node agent number, and arrange them according to the state dimension of each node agent to form the predicted state input of each node agent. Extract the current running schedule data of the corresponding node from the standardized scheduling dataset according to the node agent number, and form the measured state input of each node agent; In the dual-branch state coding process of prediction-measured residual, the predicted state input is sent to the prediction state coding branch, and the measured state input is sent to the measured state coding branch to generate the prediction state coding vector and the measured state coding vector, respectively. The resource evolution residual vector is generated by calculating the difference between the predicted state coding vector and the measured state coding vector through the residual calculation branch. The predicted state encoding vector, the measured state encoding vector, and the resource evolution residual vector are combined according to the node agent number to generate a multi-agent scheduling state representation.

6. The method for dynamic scheduling of distributed node resources based on cloud computing according to claim 1, characterized in that, The process of determining the candidate action set and generating the action matching matrix includes: Read the multi-agent scheduling state representation according to the node agent number, and establish a node scheduling state table; Based on the node scheduling status table and the node resource evolution representation matrix, determine the nodes to be scheduled, the nodes that can be taken over, and the resource items to be adjusted, and generate the node scheduling object table; A set of candidate actions is generated based on the node scheduling object table. The set of candidate actions includes task migration, instance expansion, instance shrinkage, resource adjustment, node wake-up, and node hibernation. Each candidate action in the candidate action set is matched with the node pressure risk embedding of the corresponding node, and the priority of the candidate action is calculated. An action matching matrix is ​​constructed based on the candidate action number, node agent number, and candidate action priority. The rows of the action matching matrix correspond to the node agent number, the columns correspond to the candidate action number, and the elements record the corresponding candidate action priority.

7. The method for dynamic scheduling of distributed node resources based on cloud computing according to claim 1, characterized in that, The updated joint scheduling strategy includes: Input the candidate action set and action matching matrix into the MAPPO algorithm to generate the candidate action probability distribution of each node agent according to the node agent number; Based on the multi-agent scheduling state representation and candidate action probability distribution, calculate the target value vector of each node agent under the conditions of task latency, resource utilization, load balancing, migration cost, service default risk and energy consumption target; Based on the target value vector, action matching matrix, and node resource evolution representation matrix, calculate the multi-objective advantage vector of each node agent for the corresponding candidate action; In the process of vector dominance decomposition and target conflict diversion update, the multi-target dominance vector is divided into positive dominance components, negative dominance components and neutral dominance components to generate target conflict diversion results. Based on the target conflict diversion results, adjust the probability distribution of candidate actions of each node agent and update the joint scheduling strategy.

8. The method for dynamic scheduling of distributed node resources based on cloud computing according to claim 1, characterized in that, The formation of the final scheduling strategy includes: Read the updated selection probability of each node agent in the joint scheduling strategy and establish a mapping relationship with the candidate action set to generate a candidate action ranking table; The candidate action set is sorted according to the candidate action sorting table, and mutually exclusive candidate actions corresponding to the same task, the same resource node, and the same link are filtered out to generate a list of actions that can be issued. Extract scheduling objects, scheduling action types, target nodes, and resource adjustment amounts from the list of actions that can be issued, and generate the final scheduling strategy; The final scheduling policy is converted into node scheduling instructions and distributed to the corresponding distributed nodes according to the resource node number; Receive the scheduling instruction reception status and execution start status returned by the distributed nodes, and generate the final scheduling policy distribution record.

9. The method for dynamic scheduling of distributed node resources based on cloud computing according to claim 1, characterized in that, The update of the standardized scheduling dataset and the resource-task-link heterogeneous dynamic graph includes: Collect node resource status, task execution status, link transmission status, and policy execution status after the final scheduling policy is executed to form node operation feedback data; The node operation feedback data is collected according to the scheduling round and node number to generate scheduling feedback records; Extract node resource changes, task migration results, instance adjustment results, link status changes, and service breach records from the scheduling feedback records to generate a feedback feature table; Adjust the next round of node resource evolution prediction parameters and joint scheduling strategy parameters based on the feedback feature table; The feedback feature table is written into the standardized scheduling dataset, and the resource node status, task node status, link node status, and heterogeneous edge connection status are updated according to the feedback feature table to form an updated resource-task-link heterogeneous dynamic graph.

Images