An online transaction task scheduling method and system based on a storage system

Abstract

The application discloses an online transaction task scheduling method and system based on a storage system, comprising the following steps: obtaining scheduling data and preprocessing to generate a node feature set and an associated feature set; constructing a heterogeneous graph to represent entities and relationships; setting MPNN propagation rounds and a leading node to form a propagation configuration; performing role rotation propagation to generate node representation; calculating a differential representation based on adjacent round node representation; generating neighbor contribution results through message masking and local recalculation; screening connection relationships according to the contribution results and reconstructing a propagation graph, and iteratively updating until convergence, generating scheduling parameters based on the final node representation and completing task scheduling. The application realizes efficient online transaction task scheduling, improves resource utilization and system stability.

Description

Technical Field

[0001] This invention relates to the field of distributed computing technology, and in particular to an online transaction task scheduling method and system based on a storage system. Background Technology

[0002] In existing technologies, online transaction task scheduling typically relies on rule-based scheduling strategies or simple queue scheduling mechanisms, allocating and executing tasks based on preset priorities, time sequences, or resource utilization. Some improved solutions utilize distributed storage systems for persistent task management, combined with multi-node collaborative processing to enhance system throughput. With the development of machine learning technology, some systems have attempted to introduce neural network models to optimize scheduling strategies; however, these often focus on single-dimensional feature modeling, lacking a unified modeling capability for the complex relationships between tasks, execution resources, and storage resources.

[0003] Existing technologies still have significant shortcomings. Traditional scheduling methods mainly rely on static rules or local states for decision-making, making it difficult to dynamically reflect the dependencies, resource competition, and data access relationships between tasks. This leads to problems such as uneven resource allocation, local congestion, and decreased processing efficiency in high-concurrency scenarios. Existing machine learning-based scheduling methods mostly treat tasks as independent entities, lacking modeling of the interaction relationships between various types of nodes, and failing to effectively characterize the complex relationships between tasks, execution nodes, and storage nodes.

[0004] Therefore, how to provide an online transaction task scheduling method and system based on a storage system 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 an online transaction task scheduling method and system based on a storage system. This invention constructs a heterogeneous graph and combines role rotation propagation and path convergence mechanisms to achieve joint modeling and dynamic optimization of the relationship between tasks, resources and storage, which has the advantages of high scheduling accuracy, high resource utilization and strong system stability.

[0006] According to an embodiment of the present invention, an online transaction task scheduling method and system based on a storage system includes the following steps: Acquire basic scheduling data and scheduling status data, perform preprocessing, and generate node feature sets and associated feature sets; Construct a heterogeneous graph based on the node feature set and the association feature set, map the various entities participating in scheduling to different types of nodes, and map the associations between entities to corresponding connection relationships; Roles are assigned and configured. During the MPNN propagation process, the dominant node category and corresponding update order are assigned to the current round of computation according to the preset propagation sequence, thus forming the current round of propagation configuration. Role-based propagation involves inputting the heterogeneous graph into the MPNN, controlling the propagation control rules during the current round's message passing process based on the current round's propagation configuration, completing the current round's message passing calculation, and generating the current round's node representation. State difference extraction generates node state change results based on the current round node representation and the previous round node representation, forming a difference representation; Neighbor contribution calculation involves performing message masking and local recalculation on each of the target node's associated neighbors, and generating contribution results for each neighbor based on the local recalculation results and differential representation. The path convergence scheduling determines the effective propagation path based on the contribution results and reconstructs the propagation graph for the next round. It repeats the role arrangement setting, role rotation propagation, state difference extraction and neighbor contribution calculation until the preset termination condition is met. It generates scheduling parameters based on the final node representation and writes them into the task record. The online transaction task scheduling is completed based on the task record.

[0007] Optionally, constructing the heterogeneous graph specifically includes: The node feature set is divided into node type sets according to the entity category participating in scheduling, and a unique node identifier is assigned to each node. Classify the relationships in the set of associated features, divide the relationships in the set of associated features into different relationship sets according to the relationship type, and assign a corresponding relationship identifier to each relationship. Establish connections between nodes based on node identifiers and relationship identifiers, create connecting edges between nodes with related relationships, and assign corresponding related feature vectors to the connecting edges; Heterogeneous graph structures are constructed based on node type sets and connecting edges, and node type information, node feature vectors, and associated feature vectors of connecting edges are preserved in the heterogeneous graph. Perform structural verification on the heterogeneous graph, delete invalid nodes and isolated connections, and obtain a standardized heterogeneous graph.

[0008] Optionally, the character arrangement settings specifically include: Based on heterogeneous graphs, nodes in heterogeneous graphs are divided into different node category sets according to node type information; Construct a propagation round sequence, set several propagation rounds, assign a dominant node category to each propagation round, and rotate the dominant node category in a preset order between different propagation rounds; For each propagation round, a propagation direction rule is defined. The propagation direction rule is used to limit the types of nodes that are allowed to initiate messages and the types of nodes that are allowed to receive messages in the current round. For each propagation round, a node update order rule is defined. The node update order rule is used to limit the update order of each node category in the current round. Based on the propagation direction rules and node update order rules, propagation control parameters for the corresponding propagation rounds are generated; The dominant node category, propagation direction rules, node update order rules, and propagation control parameters of each propagation round are combined to form a propagation configuration set for the corresponding propagation round. The propagation configuration set is arranged in order according to the propagation round sequence to generate a rotating propagation configuration sequence.

[0009] Optionally, the role rotation propagation specifically includes: Input the heterogeneous graph into the MPNN and perform message passing computation round by round according to the round-robin propagation configuration sequence; In the current propagation round, the dominant node category and propagation control parameters are determined according to the corresponding propagation configuration; When the dominant node category is the task node category, the task node is set as the dominant node set, and the execution node category and storage node category are set as auxiliary node sets. Control is that only the task node category sends messages outward, and the execution node category and storage node category generate response messages. The execution node representation of the task node category is updated based on the received response messages. When the dominant node category is the execution node category, the execution node is set as the dominant node set. The execution node category is controlled to obtain input information from the associated task node category and the storage node category, update the execution node representation of the execution node category, and send the updated execution node representation to the associated task node category. When the dominant node category is the storage node category, the storage node is set as the dominant node set. The storage node category is controlled to update the execution node representation based on the input information of the associated task node category and the execution node category. The updated storage node representation is then sent to the associated task node category and the execution node category. Each round of propagation is executed sequentially according to the rotation propagation configuration sequence to generate the current round node representation.

[0010] Optionally, the state difference extraction specifically includes: After completing the node representation update for the current propagation round, obtain the node representation for the current round and the node representation for the previous round, and align the node representations for the two rounds according to the node identifier to form the state pair of the corresponding node; For each node, perform difference calculation based on the node representation of the current round and the node representation of the previous round to obtain the node state change vector; The magnitude of the node state change vector is calculated to generate the node state change amount, which represents the degree of change of the node in the current propagation round; The node state change vector and the node state change amount are combined to form a difference representation vector; The differential representation vectors corresponding to each node are summarized according to the node identifier to generate a differential representation.

[0011] Optionally, the neighbor contribution calculation specifically includes: For each target node in the differential representation set, obtain the set of associated neighbors of the target node based on the connection relationship in the heterogeneous graph; For each neighbor node in the associated neighbor set, message masking is performed sequentially. Message masking includes replacing the message input generated by the neighbor node to the target node in the current propagation round with a preset baseline vector. After completing the message masking process, local recalculation is performed on the target node. Local recalculation includes recalculating the node representation only for the target node in the MPNN-based propagation calculation process while keeping the message input of other neighboring nodes unchanged, so as to obtain the node representation under the masking condition. A difference calculation is performed between the node representation under the shielding condition and the current round node representation under the unshielded condition to generate a single neighbor influence difference vector; The contribution results of the neighbor nodes are obtained by combining the single neighbor influence difference vector with the difference representation of the corresponding target node. Repeat message masking and local recalculation for all neighbor nodes in the associated neighbor set to generate a set of contribution results for each neighbor node.

[0012] Optionally, the path convergence scheduling specifically includes: Based on the contribution result set corresponding to each neighbor node, the associated neighbors of the target node are filtered, and the neighbor nodes whose contribution results meet the preset conditions are determined as valid neighbor nodes. Based on the connection relationship between the valid neighbor nodes and the target node, construct a set of valid connection edges, delete connection edges that do not meet the preset conditions, and generate an updated set of connection relationships; The heterogeneous graph is reconstructed based on the updated set of connections to obtain the next round of propagation graph. Input the next round of propagation graph into MPNN, and repeat the role arrangement setting, role rotation propagation, state difference extraction and neighbor contribution calculation to generate a new set of node representations and contribution results; After each round of propagation, the set of valid connection edges in the current round is compared with that in the previous round. When the change in the set of valid connection edges is less than a preset threshold, the termination condition is determined to be met. After the termination condition is met, scheduling parameters are generated based on the final node representation. The scheduling parameters include task priority and execution node allocation identifier. Write the scheduling parameters into the task record, and execute the online transaction task scheduling based on the task record.

[0013] An online transaction task scheduling system based on a storage system according to an embodiment of the present invention includes: The storage module is used to store basic scheduling data, scheduling status data, node feature sets, associated feature sets, and task records. The data processing module is used to acquire basic scheduling data and scheduling status data, perform preprocessing, and generate node feature sets and associated feature sets. The heterogeneous graph construction module is used to construct heterogeneous graphs based on node feature sets and associated feature sets, and to establish connection relationships between different types of nodes. The role orchestration module is used to set the propagation round sequence in the MPNN propagation process and generate a rotation propagation configuration sequence corresponding to the dominant node category, propagation direction rules and node update order rules. The MPNN computation module is used to input heterogeneous graphs into the MPNN, perform role rotation propagation according to the rotation propagation configuration sequence, and generate node representations. The neighbor contribution calculation module is used to perform state difference extraction, message masking and local recalculation based on node representation, and generate contribution results corresponding to neighbor nodes. The scheduling decision module is used to filter effective propagation paths based on contribution results, generate scheduling parameters, and write them into task records. The task records provided by the storage module complete the online transaction task scheduling.

[0014] The beneficial effects of this invention are: (1) This invention constructs a heterogeneous graph containing multiple types of nodes and their relationships, and introduces a message passing mechanism based on MPNN to achieve unified modeling of the complex relationships between online transaction tasks, execution resources and storage resources. Compared with the existing technology that only schedules based on a single task attribute or local state, it can more comprehensively reflect the overall operating status of the system and improve the accuracy of scheduling decisions and global optimization capabilities.

[0015] (2) This invention introduces a node role rotation mechanism in the MPNN propagation process, with different types of nodes taking the lead in information interaction and representation update in different propagation rounds. This enables task-side information, computing resource-side information and storage-side information to participate in scheduling modeling from multiple perspectives, thereby enhancing the model's ability to express multi-dimensional factors such as resource load, data access and task dependence. This alleviates the problem of uneven resource allocation and local congestion, and improves the stability of system operation.

[0016] (3) This invention uses a neighbor contribution calculation and path convergence scheduling mechanism to dynamically filter and reconstruct message transmission paths, eliminate low-contribution connections, and retain key propagation paths. This reduces computational redundancy and improves the effectiveness of information propagation. Compared with the graph neural network propagation method with a fixed topology, it can reduce invalid computational overhead and improve the response speed and execution efficiency of scheduling results. Attached Figure Description

[0017] 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 Here is a flowchart of an online transaction task scheduling method based on a storage system proposed in this invention; Figure 2 This is a structural diagram of an online transaction task scheduling system based on a storage system proposed in this invention; Figure 3 This is a flowchart of the propagation graph reconstruction of an online transaction task scheduling method based on a storage system proposed in this invention. Detailed Implementation

[0018] 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.

[0019] refer to Figures 1-3 A method and system for scheduling online transaction tasks based on a storage system, comprising the following steps: Acquire basic scheduling data and scheduling status data, perform preprocessing, and generate node feature sets and associated feature sets; Construct a heterogeneous graph based on the node feature set and the association feature set, map the various entities participating in scheduling to different types of nodes, and map the associations between entities to corresponding connection relationships; Roles are assigned and configured. During the MPNN propagation process, the dominant node category and corresponding update order are assigned to the current round of computation according to the preset propagation sequence, thus forming the current round of propagation configuration. Role-based propagation involves inputting the heterogeneous graph into the MPNN, controlling the propagation control rules during the current round's message passing process based on the current round's propagation configuration, completing the current round's message passing calculation, and generating the current round's node representation. State difference extraction generates node state change results based on the current round node representation and the previous round node representation, forming a difference representation; Neighbor contribution calculation involves performing message masking and local recalculation on each of the target node's associated neighbors, and generating contribution results for each neighbor based on the local recalculation results and differential representation. The path convergence scheduling determines the effective propagation path based on the contribution results and reconstructs the propagation graph for the next round. It repeats the role arrangement setting, role rotation propagation, state difference extraction and neighbor contribution calculation until the preset termination condition is met. It generates scheduling parameters based on the final node representation and writes them into the task record. The online transaction task scheduling is completed based on the task record.

[0020] In this embodiment, scheduling basic data and scheduling status data are collected through system log interfaces, resource monitoring interfaces, and storage access interfaces. Preprocessing operations are performed on the scheduling basic data and scheduling status data, including data cleaning, time alignment, and standardization, to obtain a unified dataset. Feature construction is performed based on the unified dataset, mapping the scheduling basic data and scheduling status data to node features respectively. Feature concatenation processing is performed on the node features, connecting the multi-dimensional features of corresponding entities in a preset order to form a unified feature vector, generating a node feature set. The association information between entities is extracted, and the association information is encoded to generate an association feature set.

[0021] In this embodiment, constructing the heterogeneous graph specifically includes: The node feature set is divided into node type sets according to the entity category participating in scheduling, and a unique node identifier is assigned to each node. Classify the relationships in the set of associated features, divide the relationships in the set of associated features into different relationship sets according to the relationship type, and assign a corresponding relationship identifier to each relationship. Establish connections between nodes based on node identifiers and relationship identifiers, create connecting edges between nodes with related relationships, and assign corresponding related feature vectors to the connecting edges; Heterogeneous graph structures are constructed based on node type sets and connecting edges, and node type information, node feature vectors, and associated feature vectors of connecting edges are preserved in the heterogeneous graph. Perform structural verification on the heterogeneous graph, delete invalid nodes and isolated edges to obtain a standardized heterogeneous graph. Invalid nodes are defined as nodes with missing node feature vectors, and isolated edges are defined as edges that connect only a single node.

[0022] In this embodiment, the specific steps for setting up character arrangements include: Based on heterogeneous graphs, the nodes in the heterogeneous graphs are divided into different node category sets according to node type information; the node category sets include task node categories, execution node categories, and storage node categories; Construct a propagation round sequence, set several propagation rounds, assign a dominant node category to each propagation round, and rotate the dominant node category in different propagation rounds according to a preset order; the dominant node category is cyclically assigned in each propagation round in the order of task node category, execution node category, and storage node category; For each propagation round, a propagation direction rule is defined. The propagation direction rule is used to limit the types of nodes that are allowed to initiate messages and the types of nodes that are allowed to receive messages in the current round. For each propagation round, a node update order rule is defined. The node update order rule is used to limit the update order of each node category in the current round. Based on the propagation direction rules and node update order rules, propagation control parameters for the corresponding propagation rounds are generated; the propagation control parameters include node category identifier, propagation direction identifier, and update order identifier; The dominant node category, propagation direction rules, node update order rules, and propagation control parameters of each propagation round are combined to form a propagation configuration set for the corresponding propagation round. The propagation configuration set is arranged in order according to the propagation round sequence to generate a rotating propagation configuration sequence.

[0023] In this embodiment, role rotation propagation specifically includes: Input the heterogeneous graph into the MPNN and perform message passing computation round by round according to the round-robin propagation configuration sequence; In the current propagation round, the dominant node category and propagation control parameters are determined according to the corresponding propagation configuration; When the dominant node category is the task node category, the task node is set as the dominant node set, and the execution node category and storage node category are set as auxiliary node sets. Only the task node category initiates messages, while the execution node category and storage node category generate response messages. The execution node representation of the task node category is updated based on the received response messages. The current node representations and corresponding associated feature vectors of the associated execution and storage nodes are obtained based on the connection relationships in the heterogeneous graph. The task node initiates a message, which is generated by combining the task node's current node representation with the associated feature vector. The auxiliary nodes generate response messages based on their own current node representations and return them to the corresponding task node. The MPNN propagation calculation process is performed on the response messages received by the task node, and an updated task node representation is generated by combining the task node's current node representation. When the dominant node category is the execution node category, the execution node is set as the dominant node set. The execution node category obtains input information from associated task node categories and storage node categories, updates the execution node representation, and sends the updated execution node representation to the associated task node categories. Based on the connection relationships in the heterogeneous graph, the current node representations and corresponding associated feature vectors of the associated task nodes and storage nodes are obtained to form the execution node input information. The MPNN propagation calculation process is performed on the input information, and an updated execution node representation is generated by combining the current node representation of the execution node. A propagation message is generated based on the updated execution node representation and sent to the associated task nodes according to the connection relationships. After receiving the propagation message, the task node updates its corresponding task node representation based on the MPNN propagation calculation process. When the dominant node category is the storage node category, the storage node is set as the dominant node set. The storage node category is controlled to update the execution node representation based on the input information of the associated task node category and execution node category. The updated storage node representation is then sent to the associated task node category and execution node category. For each storage node, the current node representation and corresponding associated feature vector of its associated task node and execution node are obtained based on the connection relationship in the heterogeneous graph. The input information is used to characterize data access association and resource status association. The MPNN propagation calculation process is performed on the input information, and the updated storage node representation is generated by combining the current node representation of the storage node. A propagation message is generated based on the updated storage node representation and sent to the associated task node and execution node according to the connection relationship. After receiving the propagation message, the task node and execution node update the corresponding node representation based on the MPNN propagation calculation process. Each round of propagation is executed sequentially according to the rotation propagation configuration sequence to generate the current round node representation.

[0024] In this embodiment, state difference extraction specifically includes: After completing the node representation update for the current propagation round, obtain the node representation for the current round and the node representation for the previous round, and align the node representations for the two rounds according to the node identifier to form the state pair of the corresponding node; For each node, perform difference calculation based on the node representation of the current round and the node representation of the previous round to obtain the node state change vector; The magnitude of the node state change vector is calculated to generate the node state change amount, which represents the degree of change of the node in the current propagation round; The node state change vector and the node state change amount are combined to form a difference representation vector; The differential representation vectors corresponding to each node are summarized according to the node identifier to generate a differential representation.

[0025] In this embodiment, the neighbor contribution calculation specifically includes: For each target node in the differential representation set, the set of associated neighbors of the target node is obtained based on the connection relationship in the heterogeneous graph; the set of associated neighbors is obtained by traversing the connection edges corresponding to the target node. For each neighbor node in the associated neighbor set, message masking is performed sequentially. Message masking includes replacing the message input generated by the neighbor node to the target node in the current propagation round with a preset baseline vector. The preset baseline vector has the same dimension as the node feature vector and is an all-zero vector. After completing the message masking process, a local recalculation is performed on the target node. The local recalculation includes recalculating the node representation only for the target node while keeping the message input of other neighboring nodes unchanged, so as to obtain the node representation under the masking condition. The local recalculation only performs the propagation calculation once for the target node and does not update other nodes. The difference calculation is performed between the node representation under the shielding condition and the current round node representation under the unshielded condition to generate a single neighbor influence difference vector; the difference calculation is an element-wise subtraction operation of the corresponding dimension. The contribution results of the neighbor nodes are obtained by combining the difference vector of the single neighbor influence with the difference representation of the corresponding target node; the combination calculation is implemented by linear mapping after vector concatenation. Repeat message masking and local recalculation for all neighbor nodes in the associated neighbor set to generate a set of contribution results for each neighbor node.

[0026] In this embodiment, path convergence scheduling specifically includes: Based on the contribution result set corresponding to each neighbor node, the associated neighbors of the target node are filtered, and neighbor nodes whose contribution results meet the preset conditions are determined as valid neighbor nodes; the preset condition is that the contribution result is greater than the preset threshold, and the threshold is 0.5. Based on the connection relationship between the valid neighbor nodes and the target node, construct a set of valid connection edges, delete connection edges that do not meet the preset conditions, and generate an updated set of connection relationships; The heterogeneous graph is reconstructed based on the updated set of connections to obtain the next round of propagation graph. Input the next round of propagation graph into MPNN, and repeat the role arrangement setting, role rotation propagation, state difference extraction and neighbor contribution calculation to generate a new set of node representations and contribution results; After each round of propagation, the set of valid connecting edges in the current round is compared with that in the previous round. When the change in the set of valid connecting edges is less than a preset threshold, the termination condition is met. The change is obtained by calculating the proportion of the number of different edges in the set of valid connecting edges in the current round and the previous round. Here, the preset threshold is set to 0.01. After the termination condition is met, scheduling parameters are generated based on the final node representation. The scheduling parameters include task priority and execution node allocation identifier; the task priority is a scalar value with a value range of 0 to 1. Write the scheduling parameters into the task record, and execute the online transaction task scheduling based on the task record.

[0027] According to an online transaction task scheduling system based on a storage system in this embodiment, it includes: The storage module is used to store basic scheduling data, scheduling status data, node feature sets, associated feature sets, and task records. The data processing module is used to acquire basic scheduling data and scheduling status data, perform preprocessing, and generate node feature sets and associated feature sets. The heterogeneous graph construction module is used to construct heterogeneous graphs based on node feature sets and associated feature sets, and to establish connection relationships between different types of nodes. The role orchestration module is used to set the propagation round sequence in the MPNN propagation process and generate a rotation propagation configuration sequence corresponding to the dominant node category, propagation direction rules and node update order rules. The MPNN computation module is used to input heterogeneous graphs into the MPNN, perform role rotation propagation according to the rotation propagation configuration sequence, and generate node representations. The neighbor contribution calculation module is used to perform state difference extraction, message masking and local recalculation based on node representation, and generate contribution results corresponding to neighbor nodes. The scheduling decision module is used to filter effective propagation paths based on contribution results, generate scheduling parameters, and write them into task records. The task records provided by the storage module complete the online transaction task scheduling.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to an online transaction processing scenario. In this scenario, a large number of users initiate transaction requests in a short period of time, and the system needs to complete task scheduling, resource allocation, and data access processing under limited resource conditions. As the concurrency scale continues to increase, traditional scheduling methods gradually reveal problems, such as severe task queuing, uneven load on execution nodes, frequent data access conflicts, and increased system response latency, resulting in a decrease in overall processing efficiency and a deterioration in user experience.

[0029] In this scenario, the system's basic scheduling data and scheduling status data are uniformly collected and processed. This data includes task request information, resource running status, and storage access status. Through data cleaning, time alignment, and standardization, data from different sources is unified into a computable form, further generating node feature sets and association feature sets. Based on this, a heterogeneous graph structure is constructed, mapping tasks, execution resources, and storage resources to different types of nodes. Connections between nodes are established through association features, enabling unified modeling of various entity relationships within the system.

[0030] During scheduling, the constructed heterogeneous graph is input into the MPNN model, and role-based propagation is executed through a pre-defined propagation sequence. During propagation, different types of nodes participate in information transmission as dominant nodes in different rounds, allowing task-side information, execution resource status, and storage access status to alternately influence the propagation process, thus forming a more comprehensive node representation. After each round of propagation, the difference between the node representation in the current round and the previous round is calculated to obtain the node state change results, which reflect the degree of impact of information propagation on the nodes.

[0031] Based on this, the neighbor contribution calculation process is performed. For each target node, message masking is performed on each of its associated neighbors, and the node representation is recalculated locally. By comparing the differences before and after masking, the actual contribution of each neighbor node to the target node is quantified. Based on this contribution result, the connection relationships are filtered, retaining high-contribution paths and eliminating low-contribution paths, thereby dynamically reconstructing the propagation graph structure. As the propagation rounds continue, the graph structure gradually converges, eventually forming a stable and effective propagation path.

[0032] After propagation convergence, scheduling parameters, including task priority and execution node allocation results, are generated based on the final node representation. These scheduling parameters are then written into the task record, and the storage system drives task execution. This approach enables dynamic optimization of task scheduling, allowing the system to maintain stable operation under high concurrency conditions.

[0033] To verify the effectiveness of this invention, under the same load conditions, the method of this invention was compared with the traditional scheduling method, and the following results were obtained. The average response time was statistically analyzed and rounded down to a minimum unit of 10ms, the average queue length was statistically analyzed and rounded down to a minimum unit of 10, and the throughput was statistically analyzed and rounded down to a minimum unit of 50 tasks / second.

[0034] Table 1: Comparison of Scheduling Effects

[0035] As shown in Table 1, this invention outperforms traditional scheduling methods in several key performance indicators. Regarding average response time, the method of this invention optimizes the propagation path and scheduling strategy, enabling tasks to acquire execution resources faster, reducing the response time from 320 milliseconds to 180 milliseconds, thus improving system real-time performance. In terms of task success rate, due to reduced resource conflicts and duplicate execution issues, the success rate increases from 92.1% to 98.6%, indicating enhanced system stability.

[0036] Regarding resource utilization, traditional methods suffer from uneven resource allocation. This invention, by modeling the relationships between execution nodes and tasks, achieves a more balanced resource allocation, increasing utilization from 65% to 82%. As for conflict rate, the neighbor contribution filtering mechanism reduces invalid propagation paths and hotspot access, lowering the conflict rate from 12.5% ​​to 4.2%, thus reducing system rollback and retry overhead.

[0037] Regarding queue length, the method of this invention reduces task backlog through a dynamic scheduling strategy, decreasing the average queue length from 1500 to 820, indicating that system congestion is alleviated. In terms of throughput, due to improved scheduling efficiency, the number of tasks processed per second increases from 1800 to 2650, demonstrating enhanced overall processing capacity.

[0038] In summary, this invention solves the problems of uneven resource allocation, frequent conflicts, and response delays in traditional scheduling methods under high-concurrency scenarios by introducing MPNN modeling and path convergence mechanism. It improves scheduling efficiency and resource utilization while ensuring system stability, and has good application value.

[0039] 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 scheduling online transaction tasks based on a storage system, characterized in that, Includes the following steps: Acquire basic scheduling data and scheduling status data, perform preprocessing, and generate node feature sets and associated feature sets; Construct a heterogeneous graph based on the node feature set and the association feature set, map the various entities participating in scheduling to different types of nodes, and map the associations between entities to corresponding connection relationships; Roles are assigned and configured. During the MPNN propagation process, the dominant node category and corresponding update order are assigned to the current round of computation according to the preset propagation sequence, thus forming the current round of propagation configuration. Role-based propagation involves inputting the heterogeneous graph into the MPNN, controlling the propagation control rules during the current round's message passing process based on the current round's propagation configuration, completing the current round's message passing calculation, and generating the current round's node representation. State difference extraction generates node state change results based on the current round node representation and the previous round node representation, forming a difference representation; Neighbor contribution calculation involves performing message masking and local recalculation on each of the target node's associated neighbors, and generating contribution results for each neighbor based on the local recalculation results and differential representation. The path convergence scheduling determines the effective propagation path based on the contribution results and reconstructs the propagation graph for the next round. It repeats the role arrangement setting, role rotation propagation, state difference extraction and neighbor contribution calculation until the preset termination condition is met. It generates scheduling parameters based on the final node representation and writes them into the task record. The online transaction task scheduling is completed based on the task record.

2. The online transaction task scheduling method based on a storage system according to claim 1, characterized in that, The construction of the heterogeneous graph specifically includes: The node feature set is divided into node types according to the entity category participating in scheduling, and a unique node identifier is assigned to each node. The association feature set is classified into different relationship sets according to the association type, and a corresponding relationship identifier is assigned to each association relationship. Establish connections between nodes based on node identifiers and relationship identifiers, create connecting edges between nodes with related relationships, and assign corresponding related feature vectors to the connecting edges; Heterogeneous graph structures are constructed based on node type sets and connecting edges, and node type information, node feature vectors, and associated feature vectors of connecting edges are preserved in the heterogeneous graph. Perform structural verification on the heterogeneous graph, delete invalid nodes and isolated connections, and obtain a standardized heterogeneous graph.

3. The online transaction task scheduling method based on a storage system according to claim 2, characterized in that, The specific steps of setting up and arranging roles include: Based on heterogeneous graphs, nodes in heterogeneous graphs are divided into different node category sets according to node type information; Construct a propagation round sequence, set several propagation rounds, assign a dominant node category to each propagation round, and rotate the dominant node category in a preset order between different propagation rounds; For each propagation round, a propagation direction rule is defined. The propagation direction rule is used to limit the types of nodes that are allowed to initiate messages and the types of nodes that are allowed to receive messages in the current round. For each propagation round, a node update order rule is defined. The node update order rule is used to limit the update order of each node category in the current round. Based on the propagation direction rules and node update order rules, propagation control parameters for the corresponding propagation rounds are generated; The dominant node category, propagation direction rules, node update order rules, and propagation control parameters of each propagation round are combined to form a propagation configuration set for the corresponding propagation round. The propagation configuration set is arranged in order according to the propagation round sequence to generate a rotating propagation configuration sequence.

4. The online transaction task scheduling method based on a storage system according to claim 3, characterized in that, The role rotation propagation specifically includes: Input the heterogeneous graph into the MPNN and perform message passing computation round by round according to the round-robin propagation configuration sequence; In the current propagation round, the dominant node category and propagation control parameters are determined according to the corresponding propagation configuration; When the dominant node category is the task node category, the task node is set as the dominant node set, and the execution node category and storage node category are set as auxiliary node sets. Control is that only the task node category sends messages outward, and the execution node category and storage node category generate response messages. The execution node representation of the task node category is updated based on the received response messages. When the dominant node category is the execution node category, the execution node is set as the dominant node set. The execution node category is controlled to obtain input information from the associated task node category and the storage node category, update the execution node representation of the execution node category, and send the updated execution node representation to the associated task node category. When the dominant node category is the storage node category, the storage node is set as the dominant node set. The storage node category is controlled to update the execution node representation based on the input information of the associated task node category and the execution node category. The updated storage node representation is then sent to the associated task node category and the execution node category. Each round of propagation is executed sequentially according to the rotation propagation configuration sequence to generate the current round node representation.

5. The online transaction task scheduling method based on a storage system according to claim 4, characterized in that, The state difference extraction specifically includes: After completing the node representation update for the current propagation round, obtain the node representation for the current round and the node representation for the previous round, and align the node representations for the two rounds according to the node identifier to form the state pair of the corresponding node; For each node, perform difference calculation based on the node representation of the current round and the node representation of the previous round to obtain the node state change vector; The magnitude of the node state change vector is calculated to generate the node state change amount, which represents the degree of change of the node in the current propagation round; The node state change vector and the node state change amount are combined to form a difference representation vector; The differential representation vectors corresponding to each node are summarized according to the node identifier to generate a differential representation.

6. The online transaction task scheduling method based on a storage system according to claim 5, characterized in that, The neighbor contribution calculation specifically includes: For each target node in the differential representation set, obtain the set of associated neighbors of the target node based on the connection relationship in the heterogeneous graph; For each neighbor node in the associated neighbor set, message masking is performed sequentially. Message masking includes replacing the message input generated by the neighbor node to the target node in the current propagation round with a preset baseline vector. After completing the message masking process, local recalculation is performed on the target node. Local recalculation includes recalculating the node representation only for the target node in the MPNN-based propagation calculation process while keeping the message input of other neighboring nodes unchanged, so as to obtain the node representation under the masking condition. A difference calculation is performed between the node representation under the shielding condition and the current round node representation under the unshielded condition to generate a single neighbor influence difference vector; The contribution results of the neighbor nodes are obtained by combining the single neighbor influence difference vector with the difference representation of the corresponding target node. Repeat message masking and local recalculation for all neighbor nodes in the associated neighbor set to generate a set of contribution results for each neighbor node.

7. The online transaction task scheduling method based on a storage system according to claim 6, characterized in that, The path convergence scheduling specifically includes: Based on the contribution result set corresponding to each neighbor node, the associated neighbors of the target node are filtered, and the neighbor nodes whose contribution results meet the preset conditions are determined as valid neighbor nodes. Based on the connection relationship between the valid neighbor nodes and the target node, construct a set of valid connection edges, delete connection edges that do not meet the preset conditions, and generate an updated set of connection relationships; The heterogeneous graph is reconstructed based on the updated set of connections to obtain the next round of propagation graph. Input the next round of propagation graph into MPNN, and repeat the role arrangement setting, role rotation propagation, state difference extraction and neighbor contribution calculation to generate a new set of node representations and contribution results; After each round of propagation, the set of valid connection edges in the current round is compared with that in the previous round. When the change in the set of valid connection edges is less than a preset threshold, the termination condition is determined to be met. After the termination condition is met, scheduling parameters are generated based on the final node representation. The scheduling parameters include task priority and execution node allocation identifier. Write the scheduling parameters into the task record, and execute the online transaction task scheduling based on the task record.

8. An online transaction task scheduling system based on a storage system, applied to the online transaction task scheduling method based on a storage system as described in any one of claims 1 to 7, characterized in that, include: . The storage module is used to store basic scheduling data, scheduling status data, node feature sets, associated feature sets, and task records. The data processing module is used to acquire basic scheduling data and scheduling status data, perform preprocessing, and generate node feature sets and associated feature sets. The heterogeneous graph construction module is used to construct heterogeneous graphs based on node feature sets and associated feature sets, and to establish connection relationships between different types of nodes. The role orchestration module is used to set the propagation round sequence in the MPNN propagation process and generate a rotation propagation configuration sequence corresponding to the dominant node category, propagation direction rules and node update order rules. The MPNN computation module is used to input heterogeneous graphs into the MPNN, perform role rotation propagation according to the rotation propagation configuration sequence, and generate node representations. The neighbor contribution calculation module is used to perform state difference extraction, message masking and local recalculation based on node representation, and generate contribution results corresponding to neighbor nodes. The scheduling decision module is used to filter effective propagation paths based on contribution results, generate scheduling parameters, and write them into task records. The task records provided by the storage module complete the online transaction task scheduling.

Images