A Multi-Factor Scoring-Based Algorithm and System for Matching Resource Demand in Rail Transit Cloud

By constructing a dynamic influence subgraph to optimize the scheduling of rail transit cloud resources, the problems of low resource utilization and supply-demand mismatch have been solved, achieving more efficient resource scheduling and service quality assurance.

CN123123196APending Publication Date: 2026-11-17GUANGZHOU METRO GRP CO LTD +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611620934.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The existing cloud-based resource scheduling methods for rail transit suffer from problems such as low resource utilization, response delays, mismatch between resource supply and demand, and unstable scheduling, making it impossible to meet the service quality requirements of different businesses.

Method used

A multi-factor scoring-based algorithm for matching the demand of rail transit cloud resources is adopted. By constructing a dynamic influence subgraph, potential and acquired matching gains are analyzed to accurately determine the range of nodes that need to be scheduled and adjusted, thereby optimizing the matching relationship between tasks and resources.

Benefits of technology

It improves the resource utilization and scheduling accuracy of cloud resources, ensures the service quality requirements of different services, and reduces scheduling instability and response latency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123196A_ABST
    Figure CN123123196A_ABST
Patent Text Reader

Abstract

This application provides a multi-factor scoring-based algorithm and system for matching the demand of cloud resources in rail transit, belonging to the fields of cloud computing and rail transit information technology. The method obtains the set of scheduling nodes of the cloud resource platform in the current scheduling cycle, further determines the set of feasible edges and the set of matched edges, and then constructs an initial influence subgraph based on the matching status of the matching pairs formed by the nodes involved in the point set and edge set. The initial influence subgraph is then alternately expanded along feasible edges and matched edges, and the matching gain inside and outside the subgraph boundary is analyzed to determine the influence subgraph used to determine the platform scheduling scheme. This method analyzes the matching status of task nodes and resource nodes within the platform from multiple levels (points, edges, and surfaces), and uses the interpretability judgment of the boundary matching gain algorithm during the alternating expansion process to finally determine the influence subgraph used to determine the platform scheduling scheme, effectively improving the resource utilization and scheduling accuracy of cloud resources.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of cloud computing and rail transit information technology, and in particular to a rail transit cloud resource demand matching algorithm and system based on multi-factor scoring. Background Technology

[0002] As urban rail transit develops towards digitalization and intelligence, the massive amounts of data and complex computing demands generated during operation are increasing daily. Cloud computing technology is gradually being introduced into key systems such as rail transit signal control and operation scheduling, providing them with centralized computing and storage capabilities to improve overall operational efficiency.

[0003] However, existing resource scheduling methods for rail transit cloud platforms have shortcomings: First, traditional rail transit cloud platforms often use pre-configured static strategies or simple rules to allocate resources, making it difficult to optimize in a timely manner according to system changes, resulting in low resource utilization or response delays for critical tasks. Second, in a cloud architecture, numerous services compete for limited resources. If scheduling is based on only a single factor (such as first-come, first-served or fixed priority), resource supply and demand mismatches can easily occur, failing to meet the service quality requirements of different services.

[0004] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0005] The main purpose of this application is to propose a multi-factor scoring-based algorithm and system for matching the demand of rail transit cloud resources. This method dynamically constructs an influence subgraph for the corresponding scheduling cycle based on the matching relationship between nodes. By analyzing the potential matching gain inside and outside the subgraph boundary and the relationship between the obtained matching gain, the range of nodes that need to be scheduled and adjusted can be accurately determined, thereby effectively improving the resource utilization rate and scheduling accuracy of cloud resources.

[0006] To achieve the above objectives, one aspect of this application proposes a method for matching the demand for rail transit cloud resources based on multi-factor scoring, the method comprising: Obtain the set of scheduling nodes of the target cloud resource platform; the set of scheduling nodes includes several task nodes and several resource nodes; determine the set of feasible edges and the set of matched edges based on the matching situation between nodes in the set of scheduling nodes; the feasible edges indicate matching pairs that are feasible in the current scheduling period; the matched edges indicate matching pairs that are still effective from the previous scheduling period to the current scheduling period; the matching pairs are used to indicate the matching situation between task nodes and resource nodes; An initial influence subgraph is determined based on the set of scheduling nodes, the set of feasible edges, the set of matched edges, and the matching status of each matching pair in the edge set during the corresponding scheduling period. Starting from the set of scheduling nodes, the initial influence subgraph is expanded alternately along feasible edges and matched edges. After each expansion, the potential matching gain outside the subgraph boundary and the obtained matching gain inside the subgraph boundary are calculated until the potential matching gain outside the subgraph boundary is no greater than the obtained matching gain inside the subgraph boundary. The influence subgraph for the current scheduling period is obtained based on the expansion results. The influence subgraph is used to characterize the range of nodes in the set of scheduling nodes affected by scheduling changes. The matching gain characterizes the matching optimization between nodes within the node range after the scheduling change. Based on the influence subgraph of the current scheduling period, the matching relationship between task nodes and resource nodes in the scheduling node set is adjusted, and the cloud resource scheduling status of the target cloud resource platform under the current scheduling period is determined according to the adjustment result.

[0007] Optionally, the matching status of the matching pair in the corresponding scheduling period is characterized by the matching value of the matching pair in the corresponding scheduling period; the matching value of the matching pair in the corresponding scheduling period is calculated in the following way: Obtain the scoring factors of the target cloud resource platform in the corresponding scheduling period and the weight of each scoring factor; the types of scoring factors include task scoring factors, resource scoring factors and comprehensive scoring factors; Based on the task requirement information of the task node and the task rating factors of the corresponding scheduling period, the values ​​of the matching pair corresponding to each task rating factor in the corresponding scheduling period are calculated and determined; based on the available resource status information of the resource node and the resource rating factors of the corresponding scheduling period, the values ​​of the matching pair corresponding to each resource rating factor in the corresponding scheduling period are calculated and determined; based on the task requirement information of the task node, the available resource status information of the resource node, and the comprehensive rating factors of the corresponding scheduling period, the values ​​of the matching pair corresponding to each comprehensive rating factor in the corresponding scheduling period are calculated and determined. The matching value of the matching pair in the corresponding scheduling period is obtained by weighting and normalizing the values ​​of each scoring factor corresponding to the matching pair in the corresponding scheduling period and the weight of each scoring factor.

[0008] Optionally, the set of scheduling nodes further includes capacity nodes; the capacity nodes are obtained by expanding the corresponding resource nodes according to the number of task instances they can support; the capacity nodes inherit the node attributes and matching status of the corresponding resource nodes, and form a one-to-one matching pair with the task nodes; the influence subgraph of the current scheduling cycle is obtained in the following way: From the set of matched edges, select matching pairs whose upper bound of the matching value drift is greater than the degradation threshold. Obtain a first initial node set based on the task nodes and capacity nodes involved in the selected matching pairs. The upper bound of the matching value drift is calculated based on the actual adjustment amount of each scoring factor weight from the previous scheduling cycle to the current scheduling cycle and the value range of each scoring factor. The degradation threshold is calculated based on the matching value of the matching pair in the previous scheduling cycle and a preset ratio. From the set of scheduling nodes, select unmatched task nodes and capacity nodes that can be carried. Obtain a second initial node set based on the selected nodes. The first initial node set and the second initial node set are incorporated into the initial influence subgraph, and the potential matching gain outside the subgraph boundary and the already obtained matching gain inside the subgraph boundary are calculated. If the potential matching gain outside the boundary of the initial influence subgraph is greater than the already obtained matching gain inside the subgraph boundary, the feasible capacity nodes corresponding to each task node in the initial influence subgraph are sequentially included in the initial influence subgraph based on the feasible edge set, and the task nodes carried by the included feasible capacity nodes in the previous scheduling cycle are included in the initial influence subgraph based on the already matched edge set; the potential matching gain outside the subgraph boundary and the already obtained matching gain inside the subgraph boundary of the initial influence subgraph are recalculated. If the potential matching gain outside the boundary of the initial influence subgraph is not greater than the already obtained matching gain inside the subgraph boundary, stop expanding the initial influence subgraph; output the influence subgraph for the current scheduling period based on the expanded initial influence subgraph.

[0009] Optionally, the upper bound of the matching value drift is calculated in the following way: The actual adjustment amount and maximum value range of each scoring factor are determined based on the weight of each scoring factor in the previous scheduling cycle and the current scheduling cycle of the target cloud resource platform. The maximum impact of each of the scoring factors on the matching value in the current scheduling period is calculated based on the actual adjustment amount and the maximum range of each value of each scoring factor. The upper bound of the matching value drift is obtained by calculating the sum of the maximum influence of each of the scoring factors on the matching value in the current scheduling period.

[0010] Optionally, the potential matching gain outside the subgraph boundary of the initial influence subgraph is calculated in the following way: Based on the initial influence subgraph and the set of feasible edges, a set of unexplored feasible edges on the subgraph boundary of the initial influence subgraph and a set of task nodes within the subgraph boundary are selected; the two ends of the unexplored feasible edges are located inside and outside the subgraph boundary, respectively. Calculate the matching value gain corresponding to each feasible edge in the set of unexplored feasible edges on the boundary of the initial influence subgraph; filter out the maximum value among the matching value gains corresponding to each feasible edge in the set of unexplored feasible edges to confirm the maximum exploration gain; the matching value gain is the difference between the matching values ​​of the matching pairs formed by the corresponding feasible edge in the set of unexplored feasible edges and the task nodes involved in the corresponding feasible edge within the boundary of the initial influence subgraph; The potential matching gain outside the subgraph boundary of the initial influencing subgraph is calculated based on the maximum exploration gain and the number of task nodes within the subgraph boundary.

[0011] Optionally, the acquired matching gain within the subgraph boundary of the initial influence subgraph is calculated in the following manner: Based on the task nodes and capacity nodes involved in the initial influence subgraph, the set of feasible edges, and the scoring factors of the target cloud resource platform in the current scheduling cycle, as well as the weight of each scoring factor, a task and capacity node matching value matrix table is constructed within the initial influence subgraph; the task and capacity node matching value matrix table indicates all feasible matching situations of tasks and capacity nodes within the corresponding node range; The calculation is performed based on the preset algorithm and the task and capacity node matching value matrix table. The sum of the matching values ​​of the task and capacity node pairing scheme with the largest matching value in the current scheduling period is calculated to obtain the first matching value sum. The preset algorithm is the maximum weight matching algorithm with capacity or the minimum cost flow algorithm. Based on the task and capacity node matching value matrix table, the set of matched edges, and the scoring factors and their weights of the target cloud resource platform in the previous scheduling period, the sum of matching values ​​corresponding to the task and capacity node pairing schemes involved in the task and capacity node matching value matrix table in the previous scheduling period is calculated to obtain the second matching value sum. The obtained matching gain within the subgraph boundary of the initial influence subgraph is calculated based on the sum of the first matching value and the sum of the second matching value.

[0012] Optionally, the method further includes: After executing the cloud resource scheduling of the target cloud resource platform in the current scheduling period, obtain the task execution feedback data of the current scheduling period; Based on the task execution feedback data, a differentiable proxy loss function is constructed for the executed matching pairs; the differentiable proxy loss function includes the feedback quantity of each dimension, the weighted sum of squares of the differences between the target values ​​of the feedback quantities of each dimension, and a regularization term for the difference between the matching value of the corresponding matching pair and the observed matching quality; The gradient of each rating factor is obtained by calculating the gradient of the weight change of each rating factor based on the differentiable proxy loss function. The candidate adjustment amount of the weight of each of the scoring factors in the next scheduling cycle is determined by multiplying the gradient solution result of each scoring factor with the preset learning rate.

[0013] Optionally, determining the candidate adjustment amount of the weight of each of the rating factors in the next scheduling cycle based on the product of the gradient solution result of each of the rating factors and the preset learning rate includes: The candidate adjustment amount of the weight of each of the rating factors is determined by multiplying the gradient solution result of each rating factor with the preset learning rate; Based on the candidate adjustment amount of the weight of each of the scoring factors, the matching value corresponding to each matching pair composed of task nodes and resource nodes in the target cloud resource platform in the current scheduling period is recalculated, and it is determined whether the order of the matching values ​​corresponding to each matching pair has changed. If the order of the matching values ​​corresponding to each matching pair has not changed, the candidate adjustment amount of the weight of each scoring factor is used as the candidate adjustment amount of the weight of each scoring factor in the next scheduling period. If the order of the matching values ​​corresponding to each matching pair changes, the candidate adjustment amount of the scoring factors that are affected by the changed matching pair is truncated according to the upper limit of the single-cycle change range. Based on the truncation result, the candidate adjustment amount of the weight of each scoring factor in the next scheduling cycle is determined. The upper limit of the single-cycle change range is determined based on the suboptimal scheduling scheme of the current scheduling cycle.

[0014] To achieve the above objectives, another aspect of this application proposes a rail transit cloud resource demand matching system based on multi-factor scoring, the system comprising: The data acquisition module is used to acquire the set of scheduling nodes of the target cloud resource platform; the set of scheduling nodes includes several task nodes and several resource nodes; based on the matching situation between nodes in the set of scheduling nodes, a set of feasible edges and a set of matched edges are determined; the feasible edges indicate matching pairs that are feasible in the current scheduling cycle; the matched edges indicate matching pairs that are still effective from the previous scheduling cycle to the current scheduling cycle; the matching pairs are used to indicate the matching situation between task nodes and resource nodes; An influence subgraph construction module is used to determine an initial influence subgraph based on the set of scheduling nodes, the set of feasible edges, the set of matched edges, and the matching status of each matching pair in the edge set during the corresponding scheduling period. Starting from the set of scheduling nodes, the initial influence subgraph is extended alternately along feasible edges and matched edges. After each extension, the potential matching gain outside the subgraph boundary and the obtained matching gain inside the subgraph boundary are calculated until the potential matching gain outside the subgraph boundary is no greater than the obtained matching gain inside the subgraph boundary. The influence subgraph for the current scheduling period is obtained based on the extension results. The influence subgraph is used to characterize the range of nodes in the set of scheduling nodes affected by scheduling changes. The matching gain characterizes the matching optimization between nodes within the node range after the scheduling change. The scheduling module is used to adjust the matching relationship between task nodes and resource nodes in the scheduling node set based on the influence subgraph of the current scheduling period, and determine the cloud resource scheduling status of the target cloud resource platform under the current scheduling period based on the adjustment result.

[0015] To achieve the above objectives, another aspect of this application proposes an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0016] The proposed solution includes at least the following beneficial effects: This application provides a method, system, and electronic device for matching the demand of rail transit cloud resources based on multi-factor scoring. The solution includes: first, obtaining the set of scheduling nodes for the target cloud resource platform; then, determining the set of feasible edges and the set of matched edges based on the matching situation between nodes; next, constructing an influence subgraph that dynamically changes with the scheduling cycle based on the determined set of nodes and edges, and the matching situation between each matching pair; finally, determining the range of nodes affected by scheduling changes based on the influence subgraph, and then adjusting the matching relationship between nodes in the set of scheduling nodes to complete the cloud resource scheduling of the target cloud resource platform under the current scheduling cycle. On the one hand, this application method schedules the cloud resource platform's resources from three progressive levels: points, edges, and surfaces. The method analyzes the matching situation between nodes, the matching status of each matching edge in the corresponding scheduling cycle, and the matching gain inside and outside the subgraph boundary of the influencing subgraph. This analysis ultimately determines the range of nodes requiring scheduling adjustments, providing a multi-faceted and multi-dimensional approach to cloud resource scheduling. This effectively improves the resource utilization and scheduling accuracy of cloud resources. Furthermore, unlike traditional methods that rely on simple scheduling rules that consider only a single factor, this method is based on a boundary matching gain algorithm. It alternately expands the initial influencing subgraph along feasible and matched edges, considering the changes in matching gain inside and outside the subgraph boundary after each expansion. This process ultimately constructs an influencing subgraph used to determine the range of scheduling nodes, providing a more interpretable way to determine the cloud resource scheduling scheme and effectively improving the scheduling accuracy of cloud resources. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for matching the demand of rail transit cloud resources based on multi-factor scoring, provided in an embodiment of this application; Figure 2 This application provides a structural block diagram of a method for calculating matching values ​​based on a scoring model. Figure 3 This is a structural block diagram of another rail transit cloud resource demand matching system based on multi-factor scoring provided in this application embodiment; Figure 4 This is a structural block diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0020] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0021] 1) Service Level Agreement (SLA): This is a formal commitment between the service provider and the customer that clearly defines the quality, availability, performance metrics of the services provided, as well as the responsibilities of both parties.

[0022] 2) Hungarian Algorithm: A classic combinatorial optimization algorithm for solving maximum (or minimum) weighted matching problems in bipartite graphs in polynomial time. Its core idea is to gradually find a perfect matching by adjusting labels and finding augmenting paths, while preserving the property of equal subgraphs. This algorithm is only applicable to assignment problems where the number of tasks on both sides is equal and there is a one-to-one correspondence. In cases where a resource node can host multiple task instances, or where the number of tasks is unequal to the number of capacity nodes, it must be generalized to a maximum weighted matching problem with capacity, or equivalently transformed into a minimum cost flow problem.

[0023] 3) ATS: In the subway signaling and control system, ATS refers to the Automatic Train Monitoring System, which is the "brain" and command center of the automatic train control system. The main function of this system is to centrally monitor the real-time operating status of trains throughout the line (such as position and speed), and automatically or in assistance to dispatchers to direct trains according to the predetermined timetable, including automatic arrangement of train routes, adjustment of train intervals, timetable management, and automatic adjustment of train operation.

[0024] 4) Matching Boundary: In the value space of each scoring factor weight, the critical surface that causes the optimal feasible edge of a task to switch, that is, the point where the matching values ​​of two feasible edges of the task are equal. When the weight changes on one side of the matching boundary, the matching result remains unchanged, and the derivative of the hard matching result with respect to the weight is always zero; crossing the matching boundary, the matching result changes abruptly.

[0025] As urban rail transit develops towards digitalization and intelligence, the massive amounts of data and complex computing demands generated during operation are increasing daily. Cloud computing technology is gradually being introduced into key systems such as rail transit signal control and operation scheduling, providing them with centralized computing and storage capabilities to improve overall operational efficiency. However, existing resource scheduling methods still have many shortcomings: First, traditional rail transit cloud platforms often use pre-configured static strategies or simple rules to allocate resources, making it difficult to optimize in a timely manner according to real-time load changes, resulting in low resource utilization or response delays for critical tasks. Second, in a cloud architecture, numerous services compete for limited resources. If scheduling is based on only a single factor (such as first-come, first-served or fixed priority), resource supply and demand mismatches are likely to occur, failing to meet the service quality requirements of different services. Third, feedback-driven weight tuning, if unconstrained, is prone to instability in long-term operation: weights may exceed a reasonable range, making the scores between different tasks incomparable, or they may swing back and forth between adjacent cycles due to abnormal feedback in individual scheduling cycles, repeatedly changing the direction of resource allocation, thus causing repeated migration of tasks between nodes. Fourth, the optimal matching solution for tasks and resource nodes typically has computational complexity proportional to the cube of the scale. Resolving all tasks and all nodes in each scheduling cycle will exceed the scheduling cycle time budget as the number of services handled by the platform increases. On the other hand, resolving based solely on experience by selecting a local range cannot explain the difference between the obtained result and the global optimum. Fifth, the final task allocation is generated by discrete optimal matching solutions. Small changes in the scoring weights do not change the matching result in most cases, but crossing the matching boundary will cause the result to jump suddenly. Therefore, it cannot be directly assumed that the actual performance of task execution has an analytical gradient with respect to the scoring weights in the usual sense.

[0026] In view of this, this application provides a method, system, and electronic device for matching the demand of rail transit cloud resources based on multi-factor scoring. The scheme includes: first, obtaining the set of scheduling nodes of the target cloud resource platform; then, determining the set of feasible edges and the set of matched edges based on the matching situation between nodes; next, constructing an influence subgraph that dynamically changes with the scheduling cycle based on the determined set of nodes and edges and the matching situation between each matching pair; finally, determining the range of nodes affected by scheduling changes based on the influence subgraph, and then adjusting the matching relationship between nodes in the set of scheduling nodes to complete the cloud resource scheduling of the target cloud resource platform under the current scheduling cycle. On the one hand, this application method addresses the resource scheduling situation of the cloud resource platform from three progressive levels: points, edges, and surfaces. This paper analyzes the matching situation between nodes, the matching situation of each matching edge in the corresponding scheduling cycle, and the matching gain inside and outside the boundary of the subgraph that affects the subgraph. This analysis ultimately determines the range of nodes that need to be adjusted for scheduling, thus determining the cloud resource scheduling scheme from multiple perspectives and dimensions. This approach effectively improves the resource utilization and scheduling accuracy of cloud resources. Furthermore, unlike traditional methods that rely on simple scheduling rules that only consider a single factor, this method is based on a boundary matching gain algorithm. It alternately expands along feasible edges and matched edges, considering the changes in matching gain inside and outside the boundary after each expansion. This process ultimately constructs an influence subgraph used to determine the range of scheduling nodes, providing a more interpretable way to determine the cloud resource scheduling scheme and effectively improving the scheduling accuracy of cloud resources.

[0027] The rail transit cloud resource demand matching method based on multi-factor scoring provided in this application relates to the fields of cloud computing and rail transit information technology. This method can be applied to terminals, servers, or software running on either terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc., but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the rail transit cloud resource demand matching method based on multi-factor scoring, etc., but is not limited to the above forms.

[0028] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0029] Figure 1 This is an optional flowchart of a multi-factor scoring-based method for matching the demand of rail transit cloud resources, provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S101 to S103.

[0030] Step S101: Obtain the set of scheduling nodes of the target cloud resource platform; the set of scheduling nodes includes several task nodes and several resource nodes; determine the set of feasible edges and the set of matched edges based on the matching situation between nodes in the set of scheduling nodes; feasible edges indicate matching pairs that are feasible in the current scheduling period; matched edges indicate matching pairs that are still effective from the previous scheduling period to the current scheduling period; matching pairs are used to indicate the matching situation between task nodes and resource nodes.

[0031] Obtain the set of scheduling nodes, feasible edges, and matched edges of the target cloud resource platform as the data foundation for subsequent construction of the influence subgraph.

[0032] Step S102: Determine the initial influence subgraph based on the set of scheduling nodes, the set of feasible edges, the set of matched edges, and the matching status of each matching pair in the edge set during the corresponding scheduling period; starting from the set of scheduling nodes, alternately expand the initial influence subgraph along feasible edges and matched edges. After each expansion, calculate the potential matching gain outside the subgraph boundary and the obtained matching gain inside the subgraph boundary of the initial influence subgraph until the potential matching gain outside the subgraph boundary of the initial influence subgraph is no greater than the obtained matching gain inside the subgraph boundary. Obtain the influence subgraph for the current scheduling period based on the expansion results; the influence subgraph is used to characterize the range of nodes affected in the set of scheduling nodes when the scheduling changes; the matching gain characterizes the matching optimization between nodes within the node range after the scheduling change.

[0033] An initial influence subgraph is determined by analyzing the set of scheduling nodes, the matching edges that can be formed between nodes, and the matching status of each matching pair. Then, the initial influence subgraph is expanded sequentially along feasible edges and matched edges. After each expansion, the relationship between the potential matching gain inside and outside the subgraph boundary and the already obtained matching gain is analyzed, ultimately yielding the influence subgraph for the current scheduling, which serves as a reference basis for subsequent cloud resource scheduling. As defined by feasible edges and matched edges, expanding along feasible edges extends the initial influence subgraph by finding feasible matching schemes, while expanding along matched edges optimizes the initial influence subgraph by finding matched schemes. By alternately expanding the influence subgraph along feasible and matched edges, the influence subgraph is updated during the expansion and optimization process, adjusting the range of affected nodes in the set of scheduling nodes when scheduling changes occur. Ultimately, the potential matching gain outside the subgraph boundary is no greater than the already obtained matching gain inside the subgraph boundary; that is, further expansion of the influence subgraph cannot further optimize the matching between nodes within the node range after scheduling changes. In response, the method in this application is based on the boundary matching gain algorithm, which considers the changes in matching gain inside and outside the boundary, and finally obtains the influence subgraph used to determine the range of scheduling nodes. This results in a cloud resource scheduling scheme in a more interpretable way, which can effectively improve the scheduling accuracy of cloud resource scheduling.

[0034] Step S103: Adjust the matching relationship between task nodes and resource nodes in the scheduling node set based on the influence subgraph of the current scheduling cycle, and determine the cloud resource scheduling status of the target cloud resource platform under the current scheduling cycle based on the adjustment results.

[0035] The influence subgraph is used to characterize the range of nodes in the set of scheduling nodes affected when scheduling changes occur. Based on this subgraph, the task nodes and resource nodes in the set of scheduling nodes that need to be adjusted are determined. The cloud resource scheduling of the target cloud resource platform in the current scheduling cycle is executed according to the cloud resource scheduling situation, and the cloud resource scheduling of the target cloud resource platform is finally realized.

[0036] Specifically, for the task nodes and resource nodes in the scheduling node set mentioned in step S101, the node attributes corresponding to the task nodes are determined by their task requirement information, and the node attributes corresponding to the resource nodes are determined by their available resource status information; wherein, Task requirements include: task priority, urgency, required resources, deadline, latency limit, jitter limit, minimum resource guarantee, security integrity level, and fault domain of the backup instance. Available resource status information includes: the proportion of idle CPU, the proportion of idle memory, available bandwidth, current load, network latency, fault domain and security integrity level of the resource node, etc. Specifically, for each task node in the scheduling nodes, its corresponding set of feasible resource nodes is determined according to preset hard security constraints; the aforementioned hard security constraints may include: The network latency from the resource node to the task does not exceed the latency limit of the task; the latency jitter of the resource node does not exceed the jitter limit of the task; the available amount corresponding to the idle CPU ratio and idle memory ratio of the resource node is not lower than the minimum resource guarantee of the task; the fault domain to which the resource node belongs is different from the fault domain where the backup instance of the task is located, and the security integrity level of the resource node is not lower than the security integrity level of the task.

[0037] In some embodiments, the process of determining the set of feasible resource nodes for each task according to preset security hard constraints may include: The network latency from the resource node to the task is calculated by adding half of the round-trip latency from the resource node to the task access gateway, plus the product of the network hop count and the single-hop latency coefficient between the resource node and the task. The difference between the high quantile and median values ​​of the network latency of the resource node in the most recent scheduling cycles is taken as the latency jitter of the resource node. If the security integrity level of a task is higher than the set level, it is further required that the resource node is not currently carrying other tasks with a security integrity level lower than that task; the security integrity level is a static attribute registered when the service goes live.

[0038] Furthermore, feasible edges are formed only between tasks and resource nodes within the set of feasible resource nodes corresponding to those tasks; infeasible edges do not participate in subsequent matching value calculations and matching solutions.

[0039] In step S102 of some embodiments, the matching status of a matching pair in the corresponding scheduling period is characterized by the matching value of the matching pair in the corresponding scheduling period; the matching value of the matching pair in the corresponding scheduling period is calculated in the following way: Step S201: Obtain the scoring factors of the target cloud resource platform in the corresponding scheduling period and the weight of each scoring factor; the types of scoring factors include task scoring factors, resource scoring factors and comprehensive scoring factors.

[0040] Obtain the scoring factors and their weights for the corresponding scheduling period, and use them to calculate the score of the node. The types and number of scoring factors are fixed within the same scheduling lifecycle, and only the weights of the scoring factors are modified in each scheduling period. Among them, the task scoring factor is determined only by the task requirement information corresponding to the task node, the resource scoring factor is determined only by the available resource status information corresponding to the resource node, and the comprehensive scoring factor is jointly determined by the node attributes of both the task and the resource node.

[0041] Step S202: Based on the task requirement information of the task node and the task rating factors of the corresponding scheduling period, calculate and determine the value of the matching pair for each task rating factor in the corresponding scheduling period; based on the available resource status information of the resource node and the resource rating factors of the corresponding scheduling period, calculate and determine the value of the matching pair for each resource rating factor in the corresponding scheduling period; based on the task requirement information of the task node, the available resource status information of the resource node, and the comprehensive rating factors of the corresponding scheduling period, calculate and determine the value of the matching pair for each comprehensive rating factor in the corresponding scheduling period.

[0042] The matching value of a matching pair is calculated from the values ​​of each rating factor corresponding to that matching pair. The values ​​of each rating factor are calculated based on the corresponding node attributes and the rating factors corresponding to the node, which serve as the data basis for subsequent calculations of the matching value of the matching pair.

[0043] Step S203: Based on the values ​​of each scoring factor corresponding to the matching pair in the corresponding scheduling period and the weight of each scoring factor, perform weighted calculation and normalization to obtain the matching value of the matching pair in the corresponding scheduling period.

[0044] The values ​​of each scoring factor are weighted and normalized to obtain the matching value of the matching pair in the corresponding scheduling period; wherein the values ​​of each scoring factor are normalized to between zero and one.

[0045] In some embodiments, such as Figure 2 As shown, the matching value corresponding to the matching pair can be obtained by inputting the task requirement information data corresponding to the task node and the available resource status information data corresponding to the resource node into the preset scoring model for calculation; and the model parameters of the preset scoring model can be adaptively updated, including the adjustment of the weight of the scoring factors.

[0046] In some embodiments, the matching value of the matching pair in the corresponding scheduling period. The corresponding matching value model can be shown in equation (1):

[0047] in For the first Each rating factor in the matching pair The value on, For its weight, The total number of rating factors; where the matching pair corresponds to the task node. The sum of the values ​​of each task rating factor is the task requirement score for that task node. This matching corresponds to the resource node. The sum of the values ​​of each resource rating factor is the resource supply rating of the resource node, and the matching pair The sum of the values ​​of each comprehensive scoring factor is the comprehensive score of the matching pair; when a scoring factor is only related to a task node or resource node, the other part does not participate in the scoring.

[0048] In some embodiments, task requirement scoring An example of its expanded form is shown in equation (2):

[0049] in, The weights of the corresponding rating factors; In some embodiments, resource supply scoring The expanded form is shown in equation (3):

[0050] in, The weights of the corresponding rating factors; In some embodiments, the process of calculating the matching value for a matching pair can be illustrated by the following example: Taking the request for cloud platform resources by three business tasks at a certain railway station as an example, this paper introduces the complete execution process of the method in this application within a single scheduling cycle. Specifically, this embodiment uses five scoring factors: For task priority, Both are task urgency factors and depend solely on the task itself; they are task scoring factors. The idle computing power of a node is the smaller of the idle CPU ratio and the idle memory ratio of that node. It depends only on the resource node and is a resource scoring factor. To determine resource fit, the ratio of the number of CPU cores required for the task to the number of currently idle CPU cores on the node is used. For latency adaptability, take one minus the ratio of the measured network latency of the node to the upper limit of the latency of the task; and This also depends on the task and resource nodes, and the edge adaptation factor. All factor values ​​are normalized to between 0 and 1. Task J1 is for signal system data synchronization, with a security integrity level requirement of 2, an end-to-end latency limit of 10 milliseconds, a jitter limit of 5 milliseconds, and requires a 4-core CPU and 8 GiB of memory. and Use 1.00 and 1.00; Task J2 is station video analysis, with a security integrity level requirement of Level 2, a latency limit of 30 milliseconds, a jitter limit of 10 milliseconds, and requires a 6-core CPU and 12 GiB of memory. and Use 0.60 and 0.60; Task J3 generates operational reports, with a security integrity level requirement of Level 1, a latency cap of 200 milliseconds, a jitter cap of 50 milliseconds, and requires a 2-core CPU and 4 GiB of memory. and Take 0.40 and 0.30.

[0051] The resource pool contains two resource nodes: Node N1 has a concurrent capacity of 2, a measured network latency of 6 milliseconds, a measured jitter of 2 milliseconds, a security integrity level of 4, is located in fault domain FD-A, and has 20 cores and 48 GiB of idle memory, with an idle CPU ratio of 0.50 and an idle memory ratio of 0.60. The value is set to 0.50; the concurrent capacity of node N2 is 1, the measured network latency is 12 milliseconds, the measured jitter is 3 milliseconds, the security integrity level is level 2, it is located in the fault domain FD-B, the idle capacity is 10 cores and 32 GiB, the idle CPU ratio is 0.80, and the idle memory ratio is 0.80. The value is set to 0.80. The standard resource consumption for a single task instance is 8 cores and 12 GiB.

[0052] The weights of the five scoring factors in this scheduling cycle are 0.35, 0.24, 0.25, 0.06, and 0.10, respectively, while those in the previous scheduling cycle were 0.33, 0.23, 0.24, 0.10, and 0.10, respectively. The sum of the two sets is one; the difference represents the actual adjustment amount of each scoring factor weight, which is +0.02, +0.01, +0.01, and... 0.04 and 0.00.

[0053] First, infeasible edges are eliminated one by one according to the five hard security constraints. The measured network latency of node N2 is 12 milliseconds, which exceeds the latency limit of 10 milliseconds for task J1. Therefore, the combination of task J1 and node N2 is infeasible and is removed. The remaining five combinations all satisfy the five constraints of latency, jitter, security integrity level, fault domain mutual exclusion, and resource margin, thus forming feasible edges. There are a total of 6 candidate edges at the node level. One is removed, leaving 5 feasible edges.

[0054] According to formula (1), the weighted summation based on the weight of this period yields the matching value between task J1 and node N1: 0.35×1.00+0.24×1.00+0.25×0.50+0.06×0.20+0.10×0.40=0.7670, where Take 4 ÷ 20 = 0.20. Take 1 6 ÷ 10 = 0.40.

[0055] The five hard constraints are independent criteria; if any one of them is not met, the edge is pruned. For example: if the measured jitter of node N2 is 8 milliseconds, it also exceeds the jitter limit of task J1 by 5 milliseconds and is therefore pruned; if the security integrity level requirement of task J2 is level 3, then node N2's level 2 requirement is insufficient, causing the combination to be pruned; if the backup instance of task J1 is deployed in fault domain FD-A, then its combination with node N1 is pruned due to fault domain mutual exclusion; if the idle amount of node N2 is less than the 6 cores or 12 GiB required by task J2, then the combination is pruned due to insufficient resource margin. Infeasible edges are not included in the scoring, which saves the corresponding five-factor weighted evaluation and fundamentally excludes combinations with high scores but violating security constraints.

[0056] The matching values ​​for the remaining four feasible edges are calculated similarly: 0.5770 for task J2 and node N1, 0.6500 for task J2 and node N2, 0.4400 for task J3 and node N1, and 0.5180 for task J3 and node N2. The matching value for task J3 and node N2 is 0.35×0.40+0.24×0.30+0.25×0.80+0.06×0.20+0.10×0.94=0.5180. Using the weights from the previous scheduling cycle, the matching values ​​for the five feasible edges are calculated as follows: 0.7400, 0.5660, 0.6480, 0.4280, and 0.5070, respectively.

[0057] In some embodiments, the scheduling node set further includes capacity nodes; capacity nodes are obtained by expanding the corresponding resource nodes according to the number of task instances they can support; capacity nodes inherit the node attributes and matching status of the corresponding resource nodes, and correspond one-to-one with task nodes when forming matching pairs. In actual scheduling, a resource node can simultaneously carry tasks from multiple task nodes. To more accurately explain scheduling allocation, this application proposes expanding a resource node into several capacity nodes based on the number of task instances it can carry. The expanded capacity nodes inherit the node attributes and matching status of the corresponding resource nodes. This allows for discussion of the matching relationship between the traced resource nodes and task nodes, or the matching status of unique corresponding pairs formed by capacity nodes and task nodes, based on the capacity nodes. For this purpose, capacity nodes inherit at least the resource score, fault domain, and security integrity level of the resource nodes, and participate in subsequent matching pair analysis with the same status. When analyzing the matching status between task nodes and resource nodes, the analysis is based on the matching status between task nodes and capacity nodes. Furthermore, the matching status between task nodes and resource nodes, or between task nodes and capacity nodes, is uniformly described using the matching values ​​corresponding to the matching pairs formed by the nodes.

[0058] In some embodiments, after each resource node is expanded into a corresponding number of capacity nodes, the matching relationship between each capacity node and the task node can be described using a capacity-bearing bipartite graph. Here, the capacity-bearing bipartite graph is constructed by matching pairs consisting of task nodes and capacity nodes, with all task nodes on one side and all capacity nodes on the other side. In fact, if a task node and a capacity node cannot form a feasible edge, it will not be included in the bipartite graph for description. Therefore, the capacity-bearing bipartite graph is constructed by feasible edges consisting of task nodes and capacity nodes.

[0059] Following the previous example, the process of expanding a resource node into several capacity nodes and the process of determining the optimal and suboptimal scheduling schemes will be introduced: The nodes are expanded into capacity nodes based on the number of task instances that each resource node can currently support. These numbers are min(20÷8,48÷12)=2 and min(10÷8,32÷12)=1 respectively: Node N1 is expanded into capacity nodes N1a and N1b, and node N2 is expanded into capacity node N2a, for a total of 3 capacity nodes. There are 8 feasible edges at the capacity node level.

[0060] In this regard, the one-to-one assignment model can only accept one task at most per resource node, which cannot provide this result. However, the method in this application expands the resource node into capacity nodes, which can obtain a more interpretable scheduling scheme. Specifically, the maximum weight matching is found on the expanded capacity bipartite graph, and the total matching value of task J1 to capacity node N1a, task J2 to capacity node N1b, and task J3 to capacity node N2a is 0.7670 + 0.5770 + 0.5180 = 1.8620. Node N1 thus simultaneously carries task J1 and task J2, and node N2 carries task J3. Another suboptimal scheme with complete allocation at the node level is task J1 to node N1, task J2 to node N2, and task J3 to node N1, with a total matching value of 1.8570. The difference between the optimal and suboptimal schemes is 0.0050. The matching obtained by solving according to the weight of the previous scheduling cycle is task J1 to capacity node N1a, task J2 to capacity node N2a, and task J3 to capacity node N1b, with a total matching value of 1.8160.

[0061] Furthermore, based on the matching values ​​corresponding to each matching pair consisting of capacity nodes and task nodes, the influence subgraph of the current scheduling cycle mentioned in step S102 is obtained in the following way: Step S301: Select matching pairs from the set of matched edges whose upper bound of the matching value drift is greater than the degradation threshold, and obtain a first initial node set based on the task nodes and capacity nodes involved in the selected matching pairs; the upper bound of the matching value drift is calculated based on the actual adjustment amount of each scoring factor weight from the previous scheduling cycle to the current scheduling cycle and the value range of each scoring factor; the degradation threshold is calculated based on the matching value of the matching pair in the previous scheduling cycle and the preset ratio; select unmatched task nodes and capacity nodes that can be carried from the set of scheduling nodes, and obtain a second initial node set based on the selected nodes.

[0062] Select the task nodes and capacity nodes involved in the matching pairs that meet the scheduling requirements from the matched edges, as well as the newly arrived tasks and newly released capacity nodes in this scheduling cycle, as the starting point of the initial change set and the initial influence subgraph.

[0063] Step S302: Incorporate the first initial node set and the second initial node set into the initial influence subgraph, and calculate the potential matching gain outside the subgraph boundary and the already obtained matching gain inside the subgraph boundary of the initial influence subgraph.

[0064] Using the first initial node set and the second initial node set as the initial change set, the task nodes and capacity nodes involved in the initial influence subgraph and the matching situation between nodes are determined. Based on the matching situation of each matching pair formed by the nodes, the gain situation inside and outside the subgraph boundary is calculated, which serves as the basis for judging whether the subgraph should continue to expand, and also as the basis for judging whether the influence subgraph needs to continue to expand the scope of the scheduling influence nodes.

[0065] Step S303: If the potential matching gain outside the boundary of the initial influence subgraph is greater than the already obtained matching gain inside the subgraph boundary, the feasible capacity nodes corresponding to each task node in the initial influence subgraph are sequentially included in the initial influence subgraph based on the feasible edge set, and the task nodes carried by the included feasible capacity nodes in the previous scheduling cycle are included in the initial influence subgraph based on the already matched edge set; the potential matching gain outside the subgraph boundary and the already obtained matching gain inside the subgraph boundary of the initial influence subgraph are recalculated. If the potential matching gain outside the boundary of the initial influence subgraph is not greater than the already obtained matching gain inside the subgraph boundary, stop expanding the initial influence subgraph; output the influence subgraph for the current scheduling period based on the expanded initial influence subgraph.

[0066] In response, if the potential matching gain outside the initial influence subgraph boundary is greater than the already obtained matching gain inside the subgraph boundary, it proves that the scope of the scheduling influence nodes of the influence subgraph can continue to expand, thereby further optimizing the matching relationship between nodes inside and outside the subgraph and realizing the scheduling optimization of the cloud resource platform; conversely, it proves that the current scope of the scheduling influence nodes no longer needs to be expanded, and the current matching relationship inside and outside the subgraph is already the optimal solution under the current scheduling cycle. In this regard, the stopping condition for judging the matching gain inside and outside the subgraph boundary is different from the approach of "repairing existing assignments after a given cost change": the latter does not involve the determination of how large the expansion range should be, while the method of this application provides a boundary that can be used for determination.

[0067] In some embodiments, the upper bound of the matching value drift mentioned in step S301 is calculated in the following manner: Step S401: Calculate the actual adjustment amount and maximum value range of each scoring factor based on the weight of each scoring factor in the previous scheduling cycle and the current scheduling cycle of the target cloud resource platform.

[0068] First, determine the actual adjustment amount and the maximum range of weight values ​​for each scoring factor after the scheduling cycle changes, as the basis for subsequent calculations.

[0069] Step S402: Calculate the maximum impact of each rating factor on the matching value in the current scheduling period based on the actual adjustment amount and the maximum range of each value.

[0070] Based on the actual adjustment amount and maximum range of each rating factor, the maximum impact that each rating factor can have on the matching value is calculated.

[0071] Step S403: Calculate the sum of the maximum influence of each scoring factor on the matching value in the current scheduling period to obtain the upper bound of the matching value drift.

[0072] The sum of the maximum influence that each scoring factor can have on the matching value is used as the upper bound of the matching value drift, that is, the upper limit of the possible adjustment of the matching value, and serves as the basis for judging whether the nodes involved in the matched pair can be included in the initial influence subgraph.

[0073] In some embodiments, the potential matching gain outside the subgraph boundary of the initial influence subgraph mentioned above is calculated in the following manner: Step S501: Based on the initial influence subgraph and the set of feasible edges, select the set of feasible edges that have not yet been explored on the subgraph boundary of the initial influence subgraph and the set of task nodes within the subgraph boundary; the two ends of the unexplored feasible edges are located inside and outside the subgraph boundary, respectively.

[0074] The potential matching gain outside the subgraph boundary that initially affects the subgraph is based on the set of unexplored feasible edges on the subgraph boundary and the matching situation of the tasks inside the boundary.

[0075] Step S502: Calculate the matching value gain corresponding to each feasible edge in the set of unexplored feasible edges on the boundary of the initial influence subgraph; filter out the maximum value among the matching value gains corresponding to each feasible edge in the set of unexplored feasible edges to confirm the maximum exploration gain; the matching value gain is the difference between the matching values ​​of the corresponding feasible edge and the matching pairs formed by the corresponding feasible edge and the task nodes involved in the corresponding feasible edge within the boundary of the initial influence subgraph in the set of unexplored feasible edges.

[0076] Calculate the maximum matching value gain of feasible edges in the set of unexplored feasible edges on the subgraph boundary, and use this as a reference to assume the maximum matching value gain that feasible edges outside the subgraph boundary can achieve.

[0077] Step S503: Calculate the potential matching gain outside the subgraph boundary that initially affects the subgraph based on the maximum exploration gain and the number of task nodes within the subgraph boundary.

[0078] Furthermore, the expansion process of the influencing subgraph involves incorporating the feasible capacity nodes corresponding to the task nodes, and then calculating based on the maximum exploration gain and the number of task nodes within the subgraph boundary. The purpose is to evaluate and analyze how much matching gain the matching pairs formed by the task nodes within the subgraph boundary and their corresponding incorporated capacity nodes can bring to the influencing subgraph. Among them, the product of the maximum exploration gain and the number of task nodes within the subgraph boundary can be used as the potential matching gain outside the subgraph boundary of the initial influencing subgraph.

[0079] In some embodiments, the acquired matching gain within the subgraph boundary of the initial influence subgraph is calculated as follows: Step S601: Based on the task nodes and capacity nodes involved in the initial influence subgraph, the set of feasible edges, and the scoring factors of the target cloud resource platform in the current scheduling cycle and the weight of each scoring factor, construct a task and capacity node matching value matrix table in the initial influence subgraph; the task and capacity node matching value matrix table indicates all feasible matching situations of tasks and capacity nodes within the corresponding node range.

[0080] Based on the influence subgraph and the set of feasible edges, we analyze all feasible matching situations between tasks and capacity nodes within the influence subgraph, and calculate the corresponding matching values ​​based on all feasible matching situations between tasks and capacity nodes. This constructs a task-capacity node matching value matrix table corresponding to the influence subgraph, which serves as the basis for subsequent analysis of scheduling schemes.

[0081] Step S602: Based on the preset algorithm and the task and capacity node matching value matrix table, calculate the sum of matching values ​​of the task and capacity node pairing scheme with the largest matching value in the current scheduling period, and obtain the first matching value sum; the preset algorithm is the maximum weight matching algorithm with capacity or the minimum cost flow algorithm.

[0082] Analyze the task and capacity node matching value matrix table corresponding to the influence subgraph. The node pairing scheme with the largest matching value that each task node and capacity node can form under the current scheduling cycle corresponds to the optimal matching scheme that the tasks and capacity nodes involved in the currently expanded influence subgraph can achieve.

[0083] Step S603: Based on the task and capacity node matching value matrix table, the set of matched edges, and the scoring factors and their weights of the target cloud resource platform in the previous scheduling period, calculate the sum of matching values ​​corresponding to the task and capacity node pairing schemes involved in the task and capacity node matching value matrix table of the previous scheduling period, and obtain the second matching value sum.

[0084] Furthermore, based on the dispatch scheme of the previous scheduling cycle, that is, the scheduling scheme or node pairing scheme of the previous scheduling cycle, the matching situation of the tasks and capacity nodes involved in the currently expanded influence subgraph boundary is calculated under the scheduling scheme of the previous scheduling cycle, and is determined by the sum of the matching values ​​of each matching pair corresponding to the dispatch scheme of the previous cycle.

[0085] Step S604: Calculate the obtained matching gain within the subgraph boundary of the initial influence subgraph based on the sum of the first matching value and the sum of the second matching value.

[0086] Finally, the sum of matching values ​​corresponding to the optimal scheduling scheme of the currently expanded influence subgraph is compared with the sum of matching values ​​corresponding to the dispatch scheme of the previous scheduling cycle. This yields the optimal matching situation of tasks and capacity nodes within the boundary of the currently expanded influence subgraph compared to the matching optimization situation obtained by the dispatch scheme of the previous scheduling cycle, which is represented as the obtained matching gain within the subgraph boundary of the initial influence subgraph. When calculating the obtained matching gain within the subgraph boundary of the initial influence subgraph, an expansion tolerance can be added, taking one percent to five percent of the obtained matching gain within the subgraph boundary of the initial influence subgraph as an adaptive adjustment consideration.

[0087] Following the previous example, the process of expanding the influence subgraph will be introduced: The weights of the five scoring factors in this scheduling cycle are 0.35, 0.24, 0.25, 0.06, and 0.10, respectively, while those in the previous scheduling cycle were 0.33, 0.23, 0.24, 0.10, and 0.10, respectively. The sum of the two sets is one; the difference represents the actual adjustment amount of each scoring factor weight, which is +0.02, +0.01, +0.01, and... 0.04 and 0.00; the sum of the absolute values ​​of the actual adjustments of the weights of each scoring factor in this scheduling cycle is 0.08, and the maximum range of each factor value is 1, so the upper bound of the drift of the matching value of each existing matching pair is 0.08. The degradation threshold is taken as 12% of the matching value of the matching pair in the previous cycle.

[0088] Analysis of the inclusion of matched pairs in the initial impact subgraph: Matching pair (J1, N1a) had a matching value of 0.7400 in the previous period, with a threshold of 0.0888. The upper drift limit did not exceed this threshold, so it was not included. Matching pair (J2, N2a) had a matching value of 0.6480 in the previous period, with a threshold of 0.0778. The upper drift limit exceeded this threshold, so it was included. Matching pair (J3, N1b) had a matching value of 0.4280 in the previous period, with a threshold of 0.0514. The upper drift limit exceeded this threshold, so it was included. Therefore, the initial change set corresponding to the initial impact subgraph is tasks {J2, J3} and their occupied capacity nodes {N2a, N1b}. Task J1 is outside the change set, retaining its original match and continuing to occupy capacity node N1a. Therefore, the available capacity nodes in the impact subgraph are N1b and N2a. The matching value matrix of tasks and capacity nodes involved in the impact subgraph is shown in Table 1. Table 1. Matrix of matching task and capacity nodes corresponding to the influence subgraph

[0089] The maximum weighted match with capacity is calculated from the feasible matching cases of task nodes and capacity nodes involved in the task and capacity node matching value matrix table corresponding to the subgraph. If task J2 is assigned to capacity node N1b and task J3 is assigned to capacity node N2a, the total matching value in the subgraph is 0.5770 + 0.5180, which is 1.0950; if task J2 is assigned to capacity node N2a and task J3 is assigned to capacity node N1b, it is 0.6500 + 0.4400, which is 1.0900. The former is better, so the result of rematching in the subgraph is task J2 to capacity node N1b and task J3 to capacity node N2a. The total matching value of the matching situation of these two tasks and these two capacity nodes in the previous scheduling cycle is 1.0900 under the values ​​of this cycle, so the matching gain obtained in the subgraph is 1.0950 minus 1.0900, which is 0.0050; at this point, the matching of task J2 and task J3 has changed.

[0090] Before proceeding with the second round of expansion, the stopping condition of the boundary gain algorithm is used to determine the following: There are two unexplored feasible edges on the subgraph boundary: one from task J2 to capacity node N1a and the other from task J3 to capacity node N1a. Their matching value gains are 0.5770 minus 0.5770 (i.e., 0) and 0.4400 minus 0.5180 (i.e., 0). The maximum value is 0.0780; the number of tasks within the subgraph is 2, so the upper bound of the maximum potential matching gain is 0 multiplied by 2, which is 0. The expansion tolerance is three percent of the gain already obtained within the subgraph, i.e., 0.0050 multiplied by 0.03 equals 0.00015. This upper bound of 0 is not greater than the sum of the obtained gain of 0.0050 and the expansion tolerance of 0.00015, so expansion stops. The final matching is from task J1 to capacity node N1a, task J2 to capacity node N1b, and task J3 to capacity node N2a, with a total matching value of 1.8620; the optimal total matching value obtained by resolving all 3 tasks and all 3 capacity nodes is also 1.8620, with a deviation of 0. The incremental method only enumerates 2 schemes within the subgraph consisting of 2 tasks and 2 capacity nodes, thus obtaining the same optimal result as the full recalculation; matching outside the subgraph does not need to be resolved.

[0091] In step S103 of some embodiments, adjusting the matching relationship between task nodes and resource nodes in the scheduling node set based on the influence subgraph of the current scheduling period includes: Based on the impact subgraph of the current scheduling cycle, the matching relationship between task nodes and resource nodes within the affected node range in the scheduling node set is adjusted, while nodes within the unaffected node range maintain their original scheduling.

[0092] The scope of task nodes and resource nodes whose matching relationships need to be adjusted in the current scheduling cycle is determined based on the influence subgraph. Nodes outside the scope can maintain their original scheduling.

[0093] In some embodiments, the method further includes: Step S701: After executing the cloud resource scheduling of the target cloud resource platform in the current scheduling period, obtain the task execution feedback data of the current scheduling period.

[0094] Obtain task execution feedback data as the data basis for adjusting the weight of the scoring factor in the next scheduling cycle.

[0095] Step S702: Construct a differentiable proxy loss function for the executed matching pairs based on the task execution feedback data; the differentiable proxy loss function includes the weighted sum of squares of the differences between the feedback quantities of each dimension, the target values ​​of the feedback quantities of each dimension, and a regularization term for the difference between the matching value of the corresponding matching pair and the observed matching quality.

[0096] By constructing a differentiable proxy loss function, the proxy loss of each matching pair is reflected. For each executed matching pair, the proxy loss is calculated by adding the weighted sum of the squared differences between its feedback dimensions (task execution latency, service level agreement achievement rate, CPU and memory resource utilization ratio, and task execution success rate) and the corresponding target value, along with a regularization term representing the difference between the matching value and the observed matching quality. The sum of the proxy losses of all executed matching pairs constitutes the differentiable proxy loss function. The observed matching quality is a scalar, normalized to between 0 and 1, calculated from the execution feedback of the matching pair and having the same dimensions as the matching value.

[0097] Step S703: Calculate the gradient of the weight changes of each rating factor based on the differentiable surrogate loss function to obtain the gradient solution results of each rating factor.

[0098] Furthermore, the gradient of the rating factor weights is given by a differentiable proxy loss function, without going through the matching solution process.

[0099] Step S704: Determine the candidate adjustment amount of the weight of each scoring factor in the next scheduling cycle based on the product of the gradient solution result of each scoring factor and the preset learning rate.

[0100] Finally, based on the gradient solution of the weight of each scoring factor and the product of the preset learning rate, the candidate adjustment amount of the corresponding scoring factor in the next scheduling cycle is determined.

[0101] The task execution feedback data obtained from the platform is used to construct a multidimensional feedback vector. , which serves as the data basis for calculating the loss of a differentiable agent; its expression is shown in equation (4):

[0102] in, For the task In the The actual delay time of round-robin scheduling execution. For the task In the Service level agreement (SLA) achievement rate during round-robin scheduling. For the task In the The percentage of CPU resources actually used during round execution reflects resource consumption. For the task In the The proportion of memory resources actually used during round execution. For the task In the The success rate of round execution.

[0103] In some embodiments, step S704, the process of determining the candidate adjustment amount of the weight of each scoring factor in the next scheduling cycle based on the product of the gradient solution result of each scoring factor and the preset learning rate, includes, but is not limited to, the following steps: Step S801: Determine the candidate adjustment amount of the weight of each scoring factor based on the product of the gradient solution result of each scoring factor and the preset learning rate.

[0104] Step S802: Based on the candidate adjustment amount of each scoring factor weight, recalculate the matching value corresponding to each matching pair composed of task nodes and resource nodes in the target cloud resource platform in the current scheduling cycle, and determine whether the sorting of the matching values ​​corresponding to each matching pair has changed.

[0105] Step S803: If the order of the matching values ​​corresponding to each matching pair has not changed, the candidate adjustment amount of the weight of each scoring factor is used as the candidate adjustment amount of the weight of each scoring factor in the next scheduling period.

[0106] If the order of the matching values ​​corresponding to each matching pair changes, the candidate adjustment amount of the scoring factors that are affected by the changed matching pair is truncated according to the upper limit of the single-cycle change range. Based on the truncation result, the candidate adjustment amount of the weight of each scoring factor in the next scheduling cycle is determined. The upper limit of the single-cycle change range is determined based on the suboptimal scheduling scheme of the current scheduling cycle.

[0107] For steps S801 to S803, the aforementioned differentiable proxy loss function takes only continuous scoring functions as independent variables, and its partial derivatives with respect to the weights of each scoring factor exist and can be calculated. After calculating the candidate adjustment amount for the weights of each scoring factor, it can be handled according to two boundary cases: First, if the candidate adjustment amount does not change the ranking of the matching values ​​of any matching pair, the gradient component generated by the discrete matching result is zero. In this case, the gradient component provided by the regularization term is used as the candidate adjustment amount for the scoring factor, so that the weight can still move in the direction of reducing the scoring error. Second, if the ranking of the matching values ​​corresponding to each matching pair changes, which is manifested as the candidate adjustment amount causing at least one matching pair to cross the matching boundary, the candidate adjustment amount is truncated according to the upper limit of the single-cycle change amplitude, and the matching pair that crosses the boundary is included in the initial change set of the next scheduling cycle to avoid the matching result from repeatedly jumping between adjacent cycles.

[0108] The upper limit of the single-cycle variation amplitude is determined based on the suboptimal scheduling scheme of the current scheduling cycle, specifically as follows: Step S901: Calculate the sum of matching values ​​corresponding to the optimal scheduling scheme and the suboptimal scheduling scheme of the task nodes and capacity nodes involved in the influence subgraph of the current scheduling cycle to obtain the suboptimal tolerance.

[0109] Step S902: Divide the suboptimal tolerance by the number of matching pairs in the influence subgraph not included in the current scheduling cycle to obtain the allowable drift of the matching value for a single matching pair.

[0110] Step S903: Solve inversely according to the calculation method of the upper limit of the matching value drift to obtain the weight change range of each scoring factor that ensures that the upper limit of the matching value drift of all matching pairs does not exceed the allowable drift amount, and take the minimum value among them as the upper limit of the single-cycle change range; wherein, the value range of the upper limit of the single-cycle change range is 0.02 to 0.08.

[0111] For steps S901 to S903, the suboptimal tolerance is calculated by comparing the suboptimal scheduling scheme with the optimal scheduling scheme, and the upper limit of the single-cycle weight change amplitude of each matching pair is further solved in reverse. This prevents the weight from being skewed by the abnormal feedback of individual cycles and repeatedly swinging between adjacent cycles, and also ensures that the deviation caused by using the old matching outside the subgraph does not exceed the suboptimal tolerance.

[0112] The scoring factor weight update rule can be shown in Equation (5), where a single-cycle change range upper limit determined by the inverse of the suboptimal tolerance is applied to the candidate adjustment amount, and the weights after applying the upper limit are non-negatively truncated and normalized:

[0113] in, For the updated rating factor weights, The weights of the scoring factors in the previous scheduling period. This represents the candidate adjustment amount for the scoring factor; Indicates the upper limit of the change range in a single period. Cut it off proportionally. Determined in reverse by suboptimal tolerance; This means that the weights after applying the upper limit are non-negatively truncated and normalized so that the sum of the weights of each scoring factor is one; the difference between the normalized weights and the weights of the previous period is the actual adjustment amount, and its absolute value is generally less than the candidate adjustment amount. When calculating the upper limit of the matching value drift in the next scheduling period, the actual adjustment amount should be used as the input.

[0114] And, for each rating factor Define its candidate adjustment amount For equation (6):

[0115] in, The learning rate is used to adjust the step size. Indicates task In the The differentiable proxy loss function established in round scheduling is the weighted sum of squares of the differences between each dimension of feedback quantity and its target value, plus the regularization term of the difference between the matching value of the matching pair and its observed matching quality. Its independent variable is only a continuous scoring function and does not pass through the discrete matching solution process. loss function For the Weight of each scoring factor The partial derivative of the weight reflects the sensitivity of the weight to the loss.

[0116] Following the above example, the process of determining the candidate adjustment amount and the actual candidate amount of the weights of each scoring factor will be introduced: Based on the matching results of the above example, candidate adjustment amounts for the weights of the scoring factors are determined. Hard matching results in discrete values; when the weights change within a range that does not alter the matching process, their derivative with respect to the weights remains zero, making weight updates impossible. Therefore, this application's method establishes a differentiable surrogate loss function that does not pass through the matching solution process: the surrogate loss for the matching pair is constructed by adding the weighted sum of the squares of the differences between the feedback quantities of each dimension and their target values, plus a regularization term representing the difference between the matching value and the observed matching quality of the matching pair. Its independent variable is only the continuous scoring function. Assuming the regularization coefficient is 1, the observed matching quality calculated from the feedback of this cycle for the three executed matching pairs is 0.72, 0.54, and 0.55, respectively, while their matching values ​​from the previous cycle are 0.7400, 0.6480, and 0.4280, respectively. Therefore, the differences between the two are +0.0200, +0.1080, and... 0.1220; based on this, the partial derivatives of the agency loss with respect to the weights of the five rating factors are calculated to be 0.0720, 0.0964, 0.0708, 0.113, and... 0.0911; The analytical gradient is compared component-by-componently with the numerical gradient obtained by finite difference. The deviation does not exceed 3 times 10 to the power of -8, indicating that the analytical gradient is correct. Using the negative of the product of the preset learning rate 0.35 and this gradient as the candidate adjustment amount, the candidate adjustment amounts are as follows: 0.0252 0.0337 0.0248 0.0396 and +0.0319.

[0117] Next, the upper limit of the single-step change in weight is determined by reverse calculation using the suboptimal tolerance. This period's suboptimal tolerance is set to 0.08 based on operational tolerance, approximately 4.3% of the total matching value of 1.8620 for this period. The number of matching pairs not included in the influence subgraph is 1, i.e., matching pair (J1, N1a). Therefore, the allowable drift of a single matching pair is 0.08 divided by 1, which is 0.08. Following the reverse calculation method for the upper bound of drift, the maximum range of each scoring factor's value is 1. Therefore, the weight change amplitude of each scoring factor must not exceed 0.08. Taking the smallest value, the upper limit of the single-period change amplitude is 0.08, falling within the range of 0.02 to 0.08. The sum of the absolute values ​​of the above candidate adjustment amounts is 0.1552, exceeding this upper limit. Therefore, it is truncated proportionally according to this upper limit. The truncated values ​​are as follows: 0.0130 0.0174 0.0128 0.0204 and +0.0164; then, after applying the upper limit, the weights are truncated to non-negative values ​​and normalized, resulting in weights for the next scheduling cycle of 0.3537, 0.2336, 0.2490, 0.0416, and 0.1221, which sum to one; the difference between these and the current cycle weights is the actual adjustment amount, which is +0.0037, ... 0.0064 0.0010 The sum of the absolute values ​​of 0.0184 and +0.0221 is 0.0516, which is still within the upper limit. This shows that the candidate adjustment amount is not equal to the actual adjustment amount. When calculating the upper bound of drift in the next cycle, the actual adjustment amount should be used as the input. The matching is resolved with the updated weights. The result is still from task J1 to capacity node N1a, task J2 to capacity node N1b, and task J3 to capacity node N2a. The matching has not changed. However, the surrogate loss has decreased. It can be seen that the weights have changed but the hard matching has not changed. At this time, the derivative of the hard matching with respect to the weights is zero. If the gradient of the hard matching is directly calculated, no update direction will be obtained. However, the differentiable surrogate can still provide an effective descent direction. By constructing the matching value matrix only for feasible edges and solving the maximum weighted matching with capacity in the influence subgraph, tasks with high demand scores can be assigned as much as possible to capacity nodes with high supply scores and high edge fitness, so as to achieve the best matching and balance between supply and demand.

[0118] In this regard, there is a two-way coupling relationship between weight tuning and incremental matching: the positive direction involves calculating the upper bound of the matching value's drift based on the actual adjustment of the weights of each scoring factor in the current period; the negative direction involves inferring the upper limit of the single-period change range of the weights based on the permissible suboptimal degree. The combined effect of the two directions is that as long as the single-period change of the weights does not exceed this upper limit, the deviation of the matching pairs not included in the influence subgraph from the optimal solution obtained by full re-solution will not exceed the suboptimal tolerance. Thus, the upper limit of the weight change range is no longer merely a stability measure to prevent weight divergence, but becomes a necessary condition for the establishment of local re-solution; and the scope of local re-solution is no longer selected based on experience, but is determined by the weight drift boundary.

[0119] In some embodiments, the method of this application also provides a task and resource node classification mechanism, which classifies tasks into high, medium, and low levels according to their real-time / criticality, and resource nodes into three categories according to their performance capabilities; and introduces a "level offset penalty" item when calculating the matching value. When the task level and the resource node level are not compatible, the matching value will be reduced. It should be noted that the level offset penalty is an optional correction item, and the matching values ​​used in the aforementioned steps are the basic matching values ​​without the application of this correction item; if the correction item is enabled, the aforementioned steps of feasible edge filtering, capacity node expansion, and capacity matching should be re-executed with the corrected matching value; by introducing a penalty mechanism, high-level tasks are guaranteed to be preferentially allocated to high-performance nodes, thereby improving resource utilization efficiency and business reliability.

[0120] In summary, the method embodiments of this application have at least the following beneficial effects: This application provides a multi-factor scoring-based method and related equipment for matching rail transit cloud resource demands. It employs safety hard constraint feasibility judgment, capacity-based maximum weight matching solution, influence subgraph boundary gain judgment, and feedback-driven weight tuning to construct a mutually supportive closed-loop optimization system. The multi-factor scoring rules, constructed from task scoring factors, resource scoring factors, and edge adaptation factors jointly determined by tasks and resources, provide comprehensive quantitative data for matching solutions. Safety hard constraints and hierarchical pruning ensure that critical tasks receive priority support from high-quality nodes in a resource-competitive environment. The feedback-driven weight correction algorithm dynamically adjusts the scoring factor parameters based on the running results, making subsequent matching decisions more accurate. The algorithms in this scheme form a progressive relationship and complement each other. The upper limit of the single-cycle change amplitude of the weights and the screening range of the influence subgraph are mutually causal. Under the constraint that the matching relationship remains unchanged outside the influence subgraph, the optimal solution to the matching problem within the influence subgraph is obtained. The difference between this solution and the optimal solution obtained by resolving all factors does not exceed the suboptimal tolerance, achieving real-time, adaptive, and controllable suboptimal rail transit cloud resource scheduling effects.

[0121] The technical effects of the method embodiments of this application are verified based on the following calculation examples: A comparative experiment was conducted on a cloud platform of a city rail transit operation command center. The resource pool consisted of 100 computing nodes, each configured with a 16-core CPU and 64GB of memory. Each node could support an average of 5 task instances, resulting in a total capacity of 500 nodes and 500 concurrent tasks. Considering the response requirements of real-time services, the experiment set the target scheduling cycle to 1 second, meaning that the matching solution and instruction issuance within each scheduling cycle must be completed within 1000 milliseconds. Actual measurements showed that the solution time with capacity matching was approximately proportional to the cube of the size of the data involved in the solution, with the scaling factor resulting in a solution time of 2115 milliseconds for a size of 500. The fixed overhead for incrementally updating the score and feasible edges each cycle was approximately 25 milliseconds. To avoid overestimating the effectiveness due to comparison with a self-defined benchmark, the control group selected three existing practices: Baseline 1 re-solves the maximum weighted matching with capacity for all tasks and all capacity nodes for each scheduling cycle; Baseline 2 performs local repairs on existing matching after the cost changes, but does not determine the scope that should be expanded; Baseline 3 directly and dynamically adjusts the weights of each dimension based on historical matching data and execution feedback, without imposing amplitude constraints or projections.

[0122] Regarding the matching solution, baseline 1 re-solves all 500 tasks and 500 capacity nodes, taking 2115 milliseconds to solve, plus a fixed overhead of 25 milliseconds, totaling approximately 2140 milliseconds, exceeding the 1000 millisecond cycle budget. Baseline 2 uses a fixed scale of one-quarter of the total platform size, i.e., 125, for local repair based on experience, taking 2115 milliseconds multiplied by 0.25 cubed, i.e., approximately 33 milliseconds, totaling approximately 58 milliseconds. However, because the expansion range is selected based on experience rather than provable criteria, some matching pairs that should have been re-solved were missed in some scheduling cycles, resulting in a maximum decrease of 8.3% in the total matching score compared to baseline 1. In the embodiment of this application, the influence subgraph size screened by the upper bound of the matching value drift accounts for an average of 9% of the total platform size, i.e., approximately 45, with a solution time of 2115 milliseconds multiplied by 0.09 cubed, i.e., approximately 1.5 milliseconds, totaling approximately 27 milliseconds. Moreover, the deviation of the total matching score from baseline 1 never exceeds 1.8%, which is consistent with the suboptimal tolerance setting of 2%. It can be seen that the subgraph size of the embodiment of this application is smaller than that of baseline two, but the quality of the solution is better. The reason is that the screening basis is the upper bound of the matching value drift derived from the weight change, rather than a fixed ratio selected based on experience.

[0123] Regarding weight tuning, in baseline 3, the update direction of each scoring factor weight reversed 63 times over 200 consecutive scheduling cycles, with a direction alternation rate of 31.5%. One scoring factor's weight rose to 0.71, while the combined weights of the other four scoring factors were only 0.29. The scoring result was almost entirely determined by a single factor, thus losing horizontal comparability between different tasks and causing repeated task migrations between nodes, averaging 11.4 migrations per scheduling cycle. In the embodiment of this application, under the same conditions, the number of direction reversals decreased to 9, with a direction alternation rate of 4.5%. The weights of each scoring factor stabilized within a band of ±0.03 after the 40th scheduling cycle, and the sum of the weights remained constant at one. The average task migration was 4.1 times per scheduling cycle, a reduction of approximately 64% compared to baseline 3. In terms of business metrics, the average task response time was 2.1 seconds, cloud resource utilization was 91.2%, and the service level agreement achievement rate was 98.1%.

[0124] Please see Figure 3 This application also provides a rail transit cloud resource demand matching system based on multi-factor scoring, which can implement the above method. The system includes: The data acquisition module is used to acquire the set of scheduling nodes of the target cloud resource platform. The set of scheduling nodes includes several task nodes and several resource nodes. Based on the matching situation between nodes in the set of scheduling nodes, the feasible edge set and the matched edge set are determined. Feasible edges indicate matching pairs that are feasible in the current scheduling period. Matched edges indicate matching pairs that are still effective from the previous scheduling period to the current scheduling period. Matching pairs are used to indicate the matching situation between task nodes and resource nodes. The influence subgraph construction module is used to determine the initial influence subgraph based on the set of scheduling nodes, the set of feasible edges, the set of matched edges, and the matching status of each matching pair in the edge set during the corresponding scheduling period. Starting from the set of scheduling nodes, the initial influence subgraph is expanded alternately along feasible edges and matched edges. After each expansion, the potential matching gain outside the subgraph boundary and the obtained matching gain inside the subgraph boundary are calculated until the potential matching gain outside the subgraph boundary is no greater than the obtained matching gain inside the subgraph boundary. The influence subgraph for the current scheduling period is obtained based on the expansion results. The influence subgraph is used to characterize the range of nodes in the set of scheduling nodes affected when scheduling changes. The matching gain characterizes the matching optimization between nodes within the node range after scheduling changes. The scheduling module is used to adjust the matching relationship between task nodes and resource nodes in the scheduling node set based on the influence subgraph of the current scheduling period, and determine the cloud resource scheduling status of the target cloud resource platform under the current scheduling period based on the adjustment results.

[0125] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0126] Please see Figure 4 This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0127] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0128] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0129] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0130] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0131] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0132] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0133] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of this application, and do not constitute a limitation on the technical solutions provided in this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in this application are also applicable to similar technical problems.

[0134] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0135] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0136] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0137] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0138] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0139] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0140] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0141] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0142] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0143] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A rail transit cloud resource demand matching method based on a multi-factor score, characterized in that, The method includes the following steps: Obtain the set of scheduling nodes of the target cloud resource platform; the set of scheduling nodes includes several task nodes and several resource nodes; determine the set of feasible edges and the set of matched edges based on the matching situation between nodes in the set of scheduling nodes; the feasible edges indicate matching pairs that are feasible in the current scheduling period; the matched edges indicate matching pairs that are still effective from the previous scheduling period to the current scheduling period; the matching pairs are used to indicate the matching situation between task nodes and resource nodes; An initial influence subgraph is determined based on the set of scheduling nodes, the set of feasible edges, the set of matched edges, and the matching status of each matching pair in the edge set during the corresponding scheduling period. Starting from the set of scheduling nodes, the initial influence subgraph is expanded alternately along feasible edges and matched edges. After each expansion, the potential matching gain outside the subgraph boundary and the obtained matching gain inside the subgraph boundary are calculated until the potential matching gain outside the subgraph boundary is no greater than the obtained matching gain inside the subgraph boundary. The influence subgraph for the current scheduling period is obtained based on the expansion results. The influence subgraph is used to characterize the range of nodes in the set of scheduling nodes affected by scheduling changes. The matching gain characterizes the matching optimization between nodes within the node range after the scheduling change. Based on the influence subgraph of the current scheduling period, the matching relationship between task nodes and resource nodes in the scheduling node set is adjusted, and the cloud resource scheduling status of the target cloud resource platform under the current scheduling period is determined according to the adjustment result.

2. The method of claim 1, wherein, The matching status of a matching pair in a corresponding scheduling period is characterized by the matching value of the matching pair in the corresponding scheduling period; the matching value of the matching pair in the corresponding scheduling period is calculated in the following way: Obtain the scoring factors of the target cloud resource platform in the corresponding scheduling period and the weight of each scoring factor; the types of scoring factors include task scoring factors, resource scoring factors and comprehensive scoring factors; Based on the task requirement information of the task node and the task rating factors of the corresponding scheduling period, the values ​​of the matching pair corresponding to each task rating factor in the corresponding scheduling period are calculated and determined; based on the available resource status information of the resource node and the resource rating factors of the corresponding scheduling period, the values ​​of the matching pair corresponding to each resource rating factor in the corresponding scheduling period are calculated and determined; based on the task requirement information of the task node, the available resource status information of the resource node, and the comprehensive rating factors of the corresponding scheduling period, the values ​​of the matching pair corresponding to each comprehensive rating factor in the corresponding scheduling period are calculated and determined. The matching value of the matching pair in the corresponding scheduling period is obtained by weighting and normalizing the values ​​of each scoring factor corresponding to the matching pair in the corresponding scheduling period and the weight of each scoring factor.

3. The method of claim 2, wherein, The set of scheduling nodes also includes capacity nodes; the capacity nodes are obtained by expanding the corresponding resource nodes according to the number of task instances they can support; the capacity nodes inherit the node attributes and matching status of the corresponding resource nodes, and form a one-to-one matching pair with the task nodes; the influence subgraph of the current scheduling cycle is obtained in the following way: From the set of matched edges, select matching pairs whose upper bound of the matching value drift is greater than the degradation threshold, and obtain the first initial node set based on the task nodes and capacity nodes involved in the selected matching pairs; The upper bound of the matching value drift is calculated based on the actual adjustment amount of the weight of each scoring factor from the previous scheduling period to the current scheduling period and the value range of each scoring factor; the degradation threshold is calculated based on the matching value of the matching pair in the previous scheduling period and a preset ratio. Unmatched task nodes and capacity nodes are selected from the set of scheduling nodes, and a second initial node set is obtained based on the selected nodes. The first initial node set and the second initial node set are incorporated into the initial influence subgraph, and the potential matching gain outside the subgraph boundary and the already obtained matching gain inside the subgraph boundary are calculated. If the potential matching gain outside the boundary of the initial influence subgraph is greater than the already obtained matching gain inside the subgraph boundary, the feasible capacity nodes corresponding to each task node in the initial influence subgraph are sequentially included in the initial influence subgraph based on the feasible edge set, and the task nodes carried by the included feasible capacity nodes in the previous scheduling cycle are included in the initial influence subgraph based on the already matched edge set; the potential matching gain outside the subgraph boundary and the already obtained matching gain inside the subgraph boundary of the initial influence subgraph are recalculated. If the potential matching gain outside the boundary of the initial influence subgraph is not greater than the already obtained matching gain inside the subgraph boundary, the expansion of the initial influence subgraph is stopped. Output the influence subgraph for the current scheduling period based on the expanded initial influence subgraph.

4. The method of claim 3, wherein, The upper bound of the matching value drift is calculated in the following way: The actual adjustment amount and maximum value range of each scoring factor are determined based on the weight of each scoring factor in the previous scheduling cycle and the current scheduling cycle of the target cloud resource platform. The maximum impact of each of the scoring factors on the matching value in the current scheduling period is calculated based on the actual adjustment amount and the maximum range of each value of each scoring factor. The upper bound of the matching value drift is obtained by calculating the sum of the maximum influence of each of the scoring factors on the matching value in the current scheduling period.

5. The method of claim 3, wherein, The potential matching gain outside the subgraph boundary of the initial influence subgraph is calculated in the following way: Based on the initial influence subgraph and the set of feasible edges, a set of unexplored feasible edges on the subgraph boundary of the initial influence subgraph and a set of task nodes within the subgraph boundary are selected; the two ends of the unexplored feasible edges are located inside and outside the subgraph boundary, respectively. Calculate the matching value gain corresponding to each feasible edge in the set of unexplored feasible edges on the boundary of the initial influence subgraph; filter out the maximum value among the matching value gains corresponding to each feasible edge in the set of unexplored feasible edges to confirm the maximum exploration gain; the matching value gain is the difference between the matching values ​​of the matching pairs formed by the corresponding feasible edge in the set of unexplored feasible edges and the task nodes involved in the corresponding feasible edge within the boundary of the initial influence subgraph; The potential matching gain outside the subgraph boundary of the initial influencing subgraph is calculated based on the maximum exploration gain and the number of task nodes within the subgraph boundary.

6. The method of claim 3, wherein, The acquired matching gain within the subgraph boundary of the initial influence subgraph is calculated in the following way: Based on the task nodes and capacity nodes involved in the initial influence subgraph, the set of feasible edges, and the scoring factors of the target cloud resource platform in the current scheduling cycle, as well as the weight of each scoring factor, a task and capacity node matching value matrix table is constructed within the initial influence subgraph; the task and capacity node matching value matrix table indicates all feasible matching situations of tasks and capacity nodes within the corresponding node range; The calculation is performed based on the preset algorithm and the task and capacity node matching value matrix table. The sum of the matching values ​​of the task and capacity node pairing scheme with the largest matching value in the current scheduling period is calculated to obtain the first matching value sum. The preset algorithm is the maximum weight matching algorithm with capacity or the minimum cost flow algorithm. Based on the task and capacity node matching value matrix table, the set of matched edges, and the scoring factors and their weights of the target cloud resource platform in the previous scheduling period, the sum of matching values ​​corresponding to the task and capacity node pairing schemes involved in the task and capacity node matching value matrix table in the previous scheduling period is calculated to obtain the second matching value sum. The obtained matching gain within the subgraph boundary of the initial influence subgraph is calculated based on the sum of the first matching value and the sum of the second matching value.

7. The method according to claim 2, characterized in that, The method further includes: After executing the cloud resource scheduling of the target cloud resource platform in the current scheduling period, obtain the task execution feedback data of the current scheduling period; Based on the task execution feedback data, a differentiable proxy loss function is constructed for the executed matching pairs; the differentiable proxy loss function includes the feedback quantity of each dimension, the weighted sum of squares of the differences between the target values ​​of the feedback quantities of each dimension, and a regularization term for the difference between the matching value of the corresponding matching pair and the observed matching quality; The gradient of each rating factor is obtained by calculating the gradient of the weight change of each rating factor based on the differentiable proxy loss function. The candidate adjustment amount of the weight of each of the scoring factors in the next scheduling cycle is determined by multiplying the gradient solution result of each scoring factor with the preset learning rate.

8. The method according to claim 7, characterized in that, The step of determining the candidate adjustment amount of the weight of each of the rating factors in the next scheduling cycle based on the product of the gradient solution result of each of the rating factors and the preset learning rate includes: The candidate adjustment amount of the weight of each of the rating factors is determined by multiplying the gradient solution result of each rating factor with the preset learning rate; Based on the candidate adjustment amount of the weight of each of the scoring factors, the matching value corresponding to each matching pair composed of task nodes and resource nodes in the target cloud resource platform in the current scheduling period is recalculated, and it is determined whether the order of the matching values ​​corresponding to each matching pair has changed. If the order of the matching values ​​corresponding to each matching pair has not changed, the candidate adjustment amount of the weight of each scoring factor is used as the candidate adjustment amount of the weight of each scoring factor in the next scheduling period. If the order of the matching values ​​corresponding to each matching pair changes, the candidate adjustment amount of the scoring factors that are affected by the changed matching pair is truncated according to the upper limit of the single-cycle change range. Based on the truncation result, the candidate adjustment amount of the weight of each scoring factor in the next scheduling cycle is determined. The upper limit of the single-cycle change range is determined based on the suboptimal scheduling scheme of the current scheduling cycle.

9. A rail transit cloud resource demand matching system based on multi-factor scoring, characterized in that, The system includes: The data acquisition module is used to acquire the set of scheduling nodes of the target cloud resource platform; the set of scheduling nodes includes several task nodes and several resource nodes; based on the matching situation between nodes in the set of scheduling nodes, a set of feasible edges and a set of matched edges are determined; the feasible edges indicate matching pairs that are feasible in the current scheduling cycle; the matched edges indicate matching pairs that are still effective from the previous scheduling cycle to the current scheduling cycle; the matching pairs are used to indicate the matching situation between task nodes and resource nodes; An influence subgraph construction module is used to determine an initial influence subgraph based on the set of scheduling nodes, the set of feasible edges, the set of matched edges, and the matching status of each matching pair in the edge set during the corresponding scheduling period. Starting from the set of scheduling nodes, the initial influence subgraph is extended alternately along feasible edges and matched edges. After each extension, the potential matching gain outside the subgraph boundary and the obtained matching gain inside the subgraph boundary are calculated until the potential matching gain outside the subgraph boundary is no greater than the obtained matching gain inside the subgraph boundary. The influence subgraph for the current scheduling period is obtained based on the extension results. The influence subgraph is used to characterize the range of nodes in the set of scheduling nodes affected by scheduling changes. The matching gain characterizes the matching optimization between nodes within the node range after the scheduling change. The scheduling module is used to adjust the matching relationship between task nodes and resource nodes in the scheduling node set based on the influence subgraph of the current scheduling period, and determine the cloud resource scheduling status of the target cloud resource platform under the current scheduling period based on the adjustment result.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 8.