Method and system for scheduling optimization of computing resource allocation

Abstract

The application discloses a scheduling optimization method and system for computing power resource allocation, relates to the technical field of computing power resource allocation scheduling, and comprises the following steps: collecting multi-dimensional computing power data through a global computing power situation intelligent research and judgment analysis platform, extracting node dynamic load characteristics through a graph convolution node load perception model, inputting a progressive computing power adaptive decision model to generate an initial computing power allocation scheme set, screening a candidate scheme with the optimal fault resistance capability through a node fault game equilibrium scheduling algorithm, combining real-time situation feedback to complete dynamic adjustment and iterative optimization of the scheme, and finally outputting a scheduling scheme and driving execution. The application integrates innovative configurations such as an edge computing gateway cluster and an FPGA heterogeneous computing module at the hardware level, optimizes model parameters at the software level, realizes accurate matching and collaborative scheduling of global computing power resources in the time and space dimensions, and improves scheduling accuracy, fault resistance capability and real-time performance.

Description

Technical Field

[0001] This invention relates to the field of computing resource allocation and scheduling technology, and in particular to a scheduling optimization method and system for computing resource allocation. Background Technology

[0002] In the digital economy era, computing power has become a critical infrastructure, and its uneven distribution across regions and time periods is becoming increasingly prominent. Efficient collaborative scheduling of cross-domain and heterogeneous computing resources has become a core industry demand. With increasing task complexity and expanding computing network scale, traditional scheduling models struggle to cope with dynamically fluctuating demands and complex node topologies. There is an urgent need to build a computing resource allocation system that balances real-time performance, reliability, and accuracy to address issues such as supply-demand mismatch and insufficient resource utilization, supporting the large-scale deployment and efficient operation of computing networks. Existing technologies primarily achieve computing power allocation by constructing scheduling middleware, dynamic resource adjustment algorithms, or neural network-based prediction models. These mainly rely on single-dimensional resource status data and employ static or semi-dynamic scheduling strategies to complete task allocation and resource adaptation. Some solutions introduce game theory or distributed frameworks to optimize cross-domain collaboration, improving scheduling flexibility through resource pool abstraction and network topology optimization. However, they generally lack comprehensive perception and dynamic adaptation capabilities for the overall computing power situation, relying heavily on preset rules or simple models for decision-making, failing to fully integrate multi-dimensional factors such as node load correlation characteristics and fault risks for comprehensive optimization.

[0003] Existing technologies have two main drawbacks: First, the accuracy of load perception and decision-making adaptation is insufficient. Traditional methods struggle to effectively capture complex topological relationships and dynamic load changes between nodes, resulting in low matching between computing power allocation schemes and actual needs, and easily leading to resource waste or task delays. Second, their fault tolerance and dynamic optimization capabilities are weak. The scheduling process does not fully consider the game equilibrium of node failure risks and lacks an iterative optimization mechanism based on real-time situational feedback. When faced with node failures or sudden changes in demand, it is difficult to quickly adjust the scheduling strategy, affecting the stability and continuity of computing power services. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of existing technologies, this invention provides a scheduling optimization method and system for computing power resource allocation.

[0005] The technical solution adopted in this invention is a scheduling optimization method for computing power resource allocation, comprising the following steps: S1, collecting information on the computing power occupancy status, task queue length, data transmission latency, and node connection topology of each node within the entire domain through a global computing power situation intelligent analysis platform, and constructing a multi-dimensional computing power situation awareness dataset; S2, performing feature extraction and correlation analysis on the dataset based on a graph convolutional node load awareness model, and generating dynamic load assessment results for each node through node neighborhood information aggregation and high-order feature mapping; S3, inputting the dynamic load assessment results into a progressive computing power adaptation decision model, combining task computing power demand priority and resource supply... S4. Based on the elasticity coefficient and node response sensitivity, a multi-objective optimization decision space is constructed and an initial set of computing power allocation schemes is generated; S5. The initial set of computing power allocation schemes is risk-verified using a node failure game-theoretic equilibrium scheduling algorithm, and the candidate scheme with the best fault resistance is selected through game analysis of node failure probability and scheduling benefits; S6. Based on the real-time situation feedback of the intelligent judgment and analysis platform for the global computing power situation, the candidate schemes are dynamically adjusted and iteratively optimized to achieve precise matching of computing power resources in the spatiotemporal dimension; S7. The optimized computing power resource allocation and scheduling scheme is output, driving each node to execute computing power scheduling operations according to the scheme, and completing the dynamic allocation and collaborative scheduling of global computing power resources.

[0006] Furthermore, the expression for the graph convolutional node load-aware model is: ;in, For nodes The load feature vector, For activation function, For nodes The set of neighboring nodes, For nodes The degree, For the first The weight matrix of the layer, Neighboring nodes No. The feature vector of the layer, For the first Layer bias terms; For nodes The final load assessment value, For the feature fusion order, For the first Weighting coefficients for load characteristics of the first order. For the first The weighting coefficients of the first-order load characteristic gradient. for of Partial derivatives of order 1.

[0007] Furthermore, the expression for the incremental computing power adaptation decision model is: , in, For the first Initial computing power allocation decision at time step For decision-making space, These are the weighting coefficients. For decision making The corresponding computing power supply, The computing power required for the task. For decision making Execution delay, For decision making Resource consumption costs; For the first Optimized computing power allocation decisions at any given moment The length of the historical decision-making window. For the first The weight of each historical decision The attenuation coefficient is... For the first Historical decisions at every moment.

[0008] Furthermore, the expression for the node failure game-theoretic equilibrium scheduling algorithm is as follows: in, The evaluation value of the scheduling strategy corresponding to the Bochian equilibrium solution. The total number of nodes participating in the game. For nodes Fault risk weighting coefficient For nodes The probability of failure, For nodes The benefit value after implementing the scheduling strategy This represents the maximum value of the revenue of all nodes.

[0009] Furthermore, the intelligent analysis platform for the overall computing power situation includes an edge computing gateway cluster, a distributed storage array, an FPGA heterogeneous computing module, and a quantum key encryption module. The edge computing gateway cluster adopts a 5G+WiFi6 dual-mode communication architecture, with each gateway integrating 8 high-speed data acquisition interfaces. The distributed storage array adopts a distributed hash table storage structure, supporting millions of data read and write operations per second. The FPGA heterogeneous computing module is equipped with 4 computing core groups, each core group including 32 computing units. The quantum key encryption module encrypts data transmission based on the BB84 protocol. At the software level, the feature extraction layer of the graph convolutional node load perception model is set to 6 layers, with the number of convolutional kernels in each layer being 64, 128, 256, 256, 128, and 64, respectively. The decision window length of the progressive computing power adaptation decision model is set to 10, the number of game iterations of the node failure game equilibrium scheduling algorithm is set to 20, and the overall computing power situation data acquisition cycle is set to 50 milliseconds to ensure the real-time performance of computing power situation perception and the accuracy of scheduling optimization.

[0010] Further, S2 includes the following sub-steps: S21, based on the input layer of the graph convolutional node load perception model, feature mapping is performed on the collected multi-dimensional computing power situation perception dataset to convert different types of situation data into feature vectors of a unified dimension, and the topological relationship between nodes is constructed through the adjacency matrix; S22, the first layer of graph convolution is started, and the initial feature vector of each node and the feature vector of the neighboring nodes are weighted and summed, and a linear transformation is performed in combination with the bias term to obtain the first layer feature output; S23, the subsequent layers of graph convolution are executed in sequence, and each layer performs feature aggregation based on the output features of the previous layer and the node neighborhood information, and performs nonlinear feature transformation through the activation function to enhance the expressive ability of node load-related features; S24, the output features of the last layer are normalized and compressed in dimension, core load feature indicators are extracted, and dynamic load evaluation results of each node are generated to provide data support for subsequent computing power adaptation decisions.

[0011] Further, S3 includes the following sub-steps: S31, associating and aligning the dynamic load assessment results generated in S2 with the task computing power demand priority data to construct a decision input dataset including node load status and task demand parameters, and clarifying the constraints and optimization objectives of computing power adaptation decisions; S32, constructing a multi-objective optimization function based on the incremental computing power adaptation decision model, taking the matching degree between computing power supply and demand, scheduling delay, and resource consumption as optimization objectives, and setting the weight coefficients and constraint thresholds for each objective; S33, traversing the decision space using a grid search algorithm to generate multiple initial computing power allocation schemes, each scheme including node computing power allocation ratio, task allocation path, and resource scheduling timing parameters; S34, performing feasibility verification on the initial computing power allocation schemes, eliminating schemes that violate constraints, retaining schemes that meet basic scheduling requirements to form an initial computing power allocation scheme set, and submitting it to the subsequent risk verification stage.

[0012] Further, S4 includes the following sub-steps: S41, acquiring historical fault data, hardware operating status parameters, and network connection stability indicators for each node; calculating the fault probability and fault impact range parameters for each node based on statistical analysis methods; and constructing a node fault risk assessment dataset; S42, inputting the initial computing power allocation scheme set into the node fault game equilibrium scheduling algorithm; using the fault probability of each node as the risk factor and the scheduling benefit as the benefit factor; and constructing the payoff matrix of the game model; S43, solving the game equilibrium solution through iterative calculation; analyzing the stability and benefit performance of each initial scheme under different fault scenarios; and selecting candidate schemes with strong fault resistance and good scheduling benefits; S44, conducting fault simulation tests on the candidate schemes to verify the computing power scheduling adjustment capability of the schemes under sudden node failures, ensuring that the schemes have good fault tolerance performance.

[0013] Further, S5 includes the following sub-steps: S51, real-time collection of computing power usage status, task execution progress, and network transmission status data of each node through the global computing power status intelligent analysis platform, generating real-time computing power status feedback information, and comparing and analyzing it with the expected execution effect of the candidate scheme; S52, based on the feedback information, identifying deviations in the execution process of the candidate scheme, including uneven computing power allocation and excessive scheduling latency, and determining the types and directions of parameters that need to be adjusted; S53, combining the real-time load evaluation results of the graph convolution node load perception model, dynamically adjusting the computing power allocation ratio and task scheduling path parameters of the candidate scheme to optimize resource allocation efficiency; S54, repeating the status collection, deviation identification, and parameter adjustment process until the scheme execution effect meets the preset index requirements, completing the iterative optimization of the computing power allocation scheme, and forming the final scheduling scheme.

[0014] A scheduling optimization system for computing power resource allocation, applied to a scheduling optimization method for computing power resource allocation, includes: a multi-dimensional data acquisition unit for the overall computing power status, a graph convolutional node load intelligent perception unit, a progressive computing power adaptation decision generation unit, a node fault game-theoretic equilibrium scheduling verification unit, a dynamic optimization and adjustment unit for the overall computing power, and a scheduling scheme execution driving unit. The multi-dimensional data acquisition unit for the overall computing power status is bidirectionally connected to the graph convolutional node load intelligent perception unit, used to collect computing power-related data for all nodes and transmit it to the graph convolutional node load intelligent perception unit. The graph convolutional node load intelligent perception unit is electrically connected to the progressive computing power adaptation decision generation unit, and based on the collected data, completes node load evaluation and outputs it to the progressive computing power. The system includes an adaptation decision generation unit; a progressive computing power adaptation decision generation unit connected to a node fault game-theoretic equilibrium scheduling verification unit, generating an initial computing power allocation scheme set and transmitting it to the node fault game-theoretic equilibrium scheduling verification unit for risk verification; a node fault game-theoretic equilibrium scheduling verification unit connected to a global computing power dynamic optimization and adjustment unit, outputting candidate schemes to the global computing power dynamic optimization and adjustment unit; a global computing power dynamic optimization and adjustment unit connected to a scheduling scheme execution driving unit, outputting the schemes to the scheduling scheme execution driving unit after iterative optimization; and a scheduling scheme execution driving unit connected to each computing power node, driving each node to perform computing power scheduling operations, with each unit working collaboratively to efficiently schedule and optimize global computing power resources.

[0015] Beneficial Effects: This invention proposes a scheduling optimization method and system for computing power resource allocation. Utilizing the comprehensive data acquisition capabilities of a full-domain computing power situation intelligent analysis platform and the neighborhood information aggregation and high-order feature mapping functions of a graph convolutional node load perception model, it captures the dynamic load and topological association features of nodes. This addresses the shortcomings of traditional technologies in terms of insufficient accuracy in load perception and decision-making adaptation, improving the matching degree between computing power allocation schemes and actual needs, and reducing resource waste and task delays. Through a node failure game-theoretic equilibrium scheduling algorithm, it analyzes the game between failure probability and scheduling benefits, combined with multi-objective optimization of a progressive computing power adaptation decision model and a full-domain computing power dynamic optimization adjustment mechanism. This enhances the fault resistance and dynamic adaptability of the scheduling scheme, solving the problem of weak fault resistance and dynamic optimization capabilities in traditional technologies, and ensuring the stability and continuity of computing power services. Simultaneously, the collaborative architecture of each unit of the system and the optimized configuration of hardware and software parameters further improve the real-time performance, reliability, and efficiency of computing power scheduling, achieving accurate matching and collaborative scheduling of full-domain computing power resources in the spatiotemporal dimension, supporting the large-scale deployment and efficient operation of computing power networks. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0017] Figure 2This is a flowchart of method step S2 of the present invention;

[0018] Figure 3 This is a flowchart of method step S3 of the present invention;

[0019] Figure 4 This is a flowchart of method step S4 of the present invention;

[0020] Figure 5 This is a flowchart of step S5 of the method of the present invention;

[0021] Figure 6 This is a diagram showing the system unit composition of the present invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 1 As shown, the scheduling optimization method for computing power resource allocation includes the following steps: S1, collecting information on the computing power occupancy status, task queue length, data transmission latency, and node connection topology of each node across the entire domain through a global computing power situation intelligent analysis platform, and constructing a multi-dimensional computing power situation awareness dataset; S2, performing feature extraction and correlation analysis on the dataset based on a graph convolutional node load awareness model, and generating dynamic load assessment results for each node through node neighborhood information aggregation and high-order feature mapping; S3, inputting the dynamic load assessment results into a progressive computing power adaptation decision model, combining task computing power demand priority, resource supply elasticity coefficient, and... S4. Node response sensitivity is assessed, a multi-objective optimization decision space is constructed, and an initial set of computing power allocation schemes is generated. The initial set of computing power allocation schemes is then risk-verified using a node failure game-theoretic equilibrium scheduling algorithm. The candidate scheme with the best fault tolerance is selected through game analysis of node failure probability and scheduling benefits. S5. Based on real-time situational feedback from the intelligent analysis platform for the overall computing power situation, the candidate schemes are dynamically adjusted and iteratively optimized to achieve precise matching of computing power resources in the spatiotemporal dimension. S6. The optimized computing power resource allocation and scheduling scheme is output, driving each node to execute computing power scheduling operations according to the scheme, thus completing the dynamic allocation and collaborative scheduling of overall computing power resources.

[0024] Step S1 is implemented through the hardware architecture and software acquisition mechanism of the intelligent analysis platform for the overall computing power situation. Specifically, it involves the edge computing gateway cluster to collect data from each node across the entire domain. This cluster adopts a dual-mode communication architecture of 5G and WiFi 6, with each gateway integrating eight high-speed data acquisition interfaces. It can simultaneously access the status data of different types of nodes. The collected data includes the computing power occupancy status of each node, task queue length, data transmission latency, and node connection topology information. The acquisition process strictly adheres to a fixed acquisition cycle of 50 milliseconds to ensure the real-time and continuous nature of the data. During the acquisition process, a distributed storage array synchronously stores the data. This storage array adopts a distributed hash table storage structure, supporting millions of data read and write operations per second. It can quickly receive and store multi-dimensional data transmitted by the gateway cluster, avoiding data loss or delay. The collected data are encrypted using a quantum key encryption module based on the BB84 protocol. Then, they are organized and categorized according to a preset data classification standard to form a multi-dimensional computing power situation awareness dataset that includes basic node status, task operation parameters, network connection characteristics, and topology association information. This dataset provides comprehensive and accurate data support for subsequent steps such as load assessment and decision generation. Its data integrity and real-time performance directly affect the effectiveness of subsequent scheduling optimization. Through standardized data collection and storage processes, it is ensured that the dataset can accurately reflect the dynamic changes in computing power across the entire domain, laying a solid data foundation for the entire computing power resource allocation, scheduling, and optimization process.

[0025] Step S2 performs data processing and load assessment based on a graph convolutional node load awareness model. The model has 6 feature extraction layers, with 64, 128, 256, 256, 128, and 64 convolutional kernels per layer, respectively. This multi-layer convolutional structure enables deep feature mining of the multi-dimensional computing power situation awareness dataset. In implementation, the dataset is first input into the model's input layer for feature standardization, converting data of different types and ranges into a unified format that the model can process. Then, the first layer of convolutional computation is initiated, combining node neighborhood information for preliminary feature aggregation. Weighted calculations are performed on the node's own features and neighboring node features using a weight matrix, followed by a bias term to complete a linear transformation, yielding the first layer's feature output. Subsequent layers of convolutional computation are then executed sequentially. Each layer, based on the feature vector output from the previous layer, further aggregates node features from a broader neighborhood. Non-linear feature transformation is achieved through activation functions, gradually strengthening the core feature representation related to node load and suppressing interference from irrelevant features. After the sixth layer of convolution is completed, the output features are normalized and compressed to remove redundant information and extract core feature indicators that can accurately reflect the load status of nodes. Finally, dynamic load assessment results for each node are generated. These results include key information such as the current load intensity, load change trend, and load carrying capacity of the node. They can comprehensively and accurately present the load status of each node, providing accurate input data for the subsequent progressive computing power adaptation decision model and ensuring that the decision-making process can be based on the real node load status.

[0026] Step S3 combines the dynamic load assessment results with task-related parameters and inputs them into the incremental computing power adaptation decision model to generate an initial computing power allocation scheme set. The decision window length of the model is set to 10, which controls the reference range of historical decision data to improve the rationality and adaptability of the decision. During implementation, the dynamic load assessment results are first correlated and aligned with the task computing power demand priority data to clarify parameters such as the computing power demand intensity and execution urgency for each task. Simultaneously, key indicators such as the resource supply elasticity coefficient and node response sensitivity of each node are collected to construct a decision input dataset including node status, task requirements, and resource characteristics. Then, a multi-objective optimization decision space is constructed based on the model, with computing power supply and demand matching degree, scheduling latency, and resource consumption as the core optimization objectives. A corresponding weight coefficient is set for each objective, and different optimization directions are emphasized through weight adjustment. Afterwards, a grid search algorithm is used to traverse the decision space, combining historical decision data within the decision window to generate multiple initial computing power allocation schemes. Each scheme clearly defines key parameters such as the node computing power allocation ratio, task allocation path, and resource scheduling sequence. After generating the schemes, each scheme undergoes a feasibility check. The check includes whether the scheme meets the node load capacity limit, whether it meets the task execution time requirements, and whether there are resource conflicts. Schemes that violate the constraints are eliminated, and schemes that meet the basic scheduling requirements are retained to form the initial computing power allocation scheme set. This scheme set provides a rich set of alternative schemes for subsequent risk checks. Through the generation and screening of multiple schemes, the diversity and feasibility of the initial schemes are ensured, laying the foundation for subsequent optimization.

[0027] Step S4 employs a node failure game-theoretic equilibrium scheduling algorithm to perform risk verification on the initial computing power allocation scheme set. The algorithm's game iteration count is set to 20, ensuring the accuracy and stability of the game equilibrium solution through multiple iterations. During implementation, historical failure data, hardware operating status parameters, and network connection stability indicators for each node are first acquired. These data are then processed using statistical analysis methods to calculate the failure probability and impact range parameters for each node, constructing a node failure risk assessment dataset that includes node failure risk-related information. Subsequently, the initial computing power allocation scheme set is input into the node failure game-theoretic equilibrium scheduling algorithm. Using the failure probability of each node as the risk factor and the scheduling reward as the reward factor, a payoff matrix for the game model is constructed, clarifying the risk-reward relationship for each node under different schemes. The algorithm then initiates its iterative calculation process, progressively solving for the game equilibrium solution according to the preset 20 iterations. In each iteration, based on the strategy selection and payoff matrix of each node, the evaluation weights of the schemes are adjusted, and the stability and reward performance of each initial scheme under different failure scenarios are analyzed. After iteration, candidate solutions with strong fault resistance and good scheduling benefits are selected. To further verify the fault tolerance performance of the solutions, fault simulation tests are conducted on the candidate solutions to simulate scenarios of sudden failures of different nodes, observe the computing power scheduling and adjustment capabilities of the solutions, and verify whether the solutions can maintain normal computing power services when failures occur. This ensures that the selected candidate solutions have good fault resistance and fault tolerance performance, providing a reliable solution foundation for subsequent dynamic optimization.

[0028] Step S5 leverages the real-time situational feedback function of the intelligent analysis platform for the overall computing power situation to dynamically adjust and iteratively optimize candidate solutions, achieving precise matching of computing resources in the spatiotemporal dimensions. During implementation, the edge computing gateway cluster of the intelligent analysis platform for the overall computing power situation first collects real-time data on the computing power usage status, task execution progress, and network transmission status of each node at a 50-millisecond acquisition cycle. This data is encrypted and then transmitted to the platform processing module to generate real-time computing power situational feedback information. Subsequently, the feedback information is compared and analyzed with the expected execution effects of the candidate solutions. The comparison includes the difference between the actual and expected computing power allocation ratios, the deviation between the actual and expected task execution delays, and the degree of fit between the actual and expected node loads. Based on the comparison results, problems in the execution of candidate solutions are identified, and the types and directions of parameters requiring adjustment are determined. Subsequently, based on the real-time load assessment results of the graph convolutional node load awareness model, key parameters such as the computing power allocation ratio and task scheduling path of the candidate schemes are dynamically adjusted. During the adjustment process, the multi-objective optimization logic of the incremental computing power adaptation decision model is referenced to ensure that the adjusted schemes still meet the optimization objectives of computing power supply and demand matching, low latency, and low consumption. Then, the above-mentioned situation acquisition, deviation identification, and parameter adjustment process is repeated. After each round of adjustment, the implementation effect of the scheme is evaluated until the implementation effect of the scheme meets the preset index requirements, completing the iterative optimization of the computing power allocation scheme. Through continuous dynamic adjustment, the scheme can adapt to the real-time changes in the computing power situation across the entire domain, achieving accurate matching of computing power resources in the spatiotemporal dimension and improving the actual effect of scheduling optimization.

[0029] Step S6 outputs the optimized computing power resource allocation and scheduling scheme, and drives each node to execute computing power scheduling operations according to the scheme, completing the dynamic allocation and collaborative scheduling of computing power resources across the entire domain. During implementation, the scheme after dynamic optimization and adjustment is first subjected to a final integrity and consistency check to ensure that all key parameters in the scheme, such as node computing power allocation ratios, task allocation paths, and resource scheduling sequences, are accurate and error-free, without logical conflicts or parameter omissions. After passing the check, the scheme is converted into an execution instruction format recognizable by each node. Subsequently, the execution instructions are transmitted to each computing power node through the communication module of the intelligent analysis platform for the overall computing power situation. During transmission, the quantum key encryption module ensures the security of instruction transmission, preventing instructions from being tampered with or stolen. After receiving the instructions, each node initiates computing power scheduling operations according to the instructions, adjusts its own computing power resource allocation status, and performs operations such as task reception, data transmission, and computing power output. The edge computing gateway cluster monitors the instruction execution status of each node in real time, collects status data during node execution, and feeds it back to the intelligent analysis platform for the overall computing power situation. The platform tracks the execution progress and results of each node in real time to ensure that each node strictly follows the optimized plan. It coordinates the computing power collaboration between nodes to avoid resource conflicts or scheduling chaos. Through the collaborative execution of each node, the platform ultimately achieves dynamic allocation and collaborative scheduling of computing power resources across the entire domain, realizing efficient utilization of computing power resources and smooth execution of tasks, thus achieving the core goal of scheduling optimization.

[0030] Preferably, the expression for the graph convolutional node load-aware model is: ;in, For nodes The load feature vector, For activation function, For nodes The set of neighboring nodes, For nodes The degree, For the first The weight matrix of the layer, Neighboring nodes No. The feature vector of the layer, For the first Layer bias terms; For nodes The final load assessment value, For the feature fusion order, For the first Weighting coefficients for load characteristics of the first order. For the first The weighting coefficients of the first-order load characteristic gradient. for of Partial derivatives of order 1.

[0031] Specifically, the graph convolutional node load perception model, based on the hierarchical extraction and multi-order fusion requirements of node load features, addresses the characteristic that node load is influenced by both itself and its neighboring nodes. It derives a first-layer formula to achieve neighborhood feature aggregation, normalizing and weighting the features of neighboring nodes using the inverse of the square root product of node degrees to prevent high-degree nodes from excessively dominating feature expression. A linear transformation is then performed using the weight matrix and bias term, followed by a non-linear feature mapping through an activation function. This formula is based on the neighborhood association modeling principle of graph convolution, effectively uncovering topological dependencies between nodes. Based on the load feature vector output from the first layer, a second-layer formula is further derived, introducing a multi-order feature fusion mechanism. By integrating load features of different orders and corresponding gradient information through product operations, it achieves a comprehensive representation of static load features and dynamic trends. This formula solves the problem of incomplete single-order feature expression, improving the accuracy of load assessment. The feature fusion order was set to 6, matching the number of feature extraction layers in the model. The weight coefficients for each order ranged from 0.1 to 0.8, and the gradient weight coefficients ranged from 0.05 to 0.3. The optimal combination was determined through multiple experimental calibrations. During implementation, the standardized node features were first input into the first-layer formula to calculate the load feature vector for each node. This vector was then substituted into the second-layer formula, and multi-level feature fusion was completed through product operations, outputting the final load evaluation value. The entire process leveraged the parallel computing capabilities of the FPGA heterogeneous computing module to ensure model computational efficiency and provide accurate load data support for computing power adaptation decisions.

[0032] Preferably, the expression for the incremental computing power adaptation decision model is:

[0033]

[0034]

[0035] in, For the first Initial computing power allocation decision at time step For decision-making space, These are the weighting coefficients. For decision making The corresponding computing power supply, The computing power required for the task. For decision making Execution delay, For decision making Resource consumption costs; For the first Optimized computing power allocation decisions at any given moment The length of the historical decision-making window. For the first The weight of each historical decision The attenuation coefficient is... For the first Historical decisions at every moment.

[0036] Specifically, the incremental computing power adaptation decision model is based on the multi-objective optimization characteristics of computing power allocation. The first-level formula is derived to construct the initial decision model, selecting the mismatch between computing power supply and demand, execution latency, and resource consumption as optimization objectives. Weight coefficients are used to balance the priorities of each objective. This formula is based on multi-objective optimization theory and comprehensively considers both the accuracy and economy of scheduling. Considering the dynamic changes in computing power status, based on the first-level decision results, the second-level formula introduces a historical decision decay and fusion mechanism. Historical decision data is integrated through summation, and an exponential decay factor is used to reduce the influence weight of long-term historical decisions, achieving incremental optimization of the decision. This formula solves the problem that static decisions cannot adapt to changes in status and improves the dynamic adaptability of the decision. The sum of the weight coefficients of the three objectives is 1, where the matching deviation weight coefficient is set to 0.5, the execution latency weight coefficient is set to 0.3, the resource consumption weight coefficient is set to 0.2, the historical decision window length is set to 10, consistent with the model's decision window parameters, and the decay coefficient is set to 0.15. During implementation, the node load assessment results and task requirement parameters are first input into the first-level formula to solve for the initial computing power allocation decision. Then, historical decision data from the last 10 time points are extracted and substituted into the second-level formula. The attenuation fusion is completed through summation, and the optimized decision result is output. The entire process ensures the speed and accuracy of decision response through the real-time data processing capability of the intelligent judgment and analysis platform for the global computing power situation, and generates an initial computing power allocation scheme set that conforms to the current situation.

[0037] Preferably, the expression for the node failure game-theoretic equilibrium scheduling algorithm is: in, The evaluation value of the scheduling strategy corresponding to the Bochian equilibrium solution. The total number of nodes participating in the game. For nodes Fault risk weighting coefficient For nodes The probability of failure, For nodes The benefit value after implementing the scheduling strategy This represents the maximum value of the revenue of all nodes.

[0038] Specifically, the node failure game-theoretic equilibrium scheduling algorithm is based on the equilibrium optimization principle of game theory. Addressing the trade-off between node failure risk and scheduling benefits, it constructs a single formula to solve for the game equilibrium solution. This formula introduces a product operation to integrate the failure risk factor and benefit factor of each node. The failure risk factor is calculated as the ratio of the failure probability to the failure probability plus 1, achieving normalization of the failure probability. The benefit factor is calculated as the ratio of the node's actual benefit to its maximum benefit. This formula is based on game equilibrium theory and can find a balance between resilience and benefit through the strategic interactions of each node. The node failure risk weight coefficient ranges from 0.3 to 0.7, adjusted according to the node's importance level. The number of game iterations is set to 20 to ensure the convergence of the equilibrium solution. During implementation, historical fault data and benefit data of each node are first collected to calculate the fault probability and benefit value of each node and determine the maximum benefit value. Then, the risk weight coefficient, fault probability, and benefit value of each node are substituted into the formula, and the evaluation value of each scheduling scheme is calculated by multiplication. The scheme with the highest evaluation value is the optimal scheme corresponding to the game equilibrium solution. During implementation, a distributed computing architecture is combined to process the evaluation calculation of multiple schemes in parallel, improve the algorithm execution efficiency, and screen out the candidate scheme with the best fault resistance.

[0039] Preferably, the intelligent analysis platform for the overall computing power situation includes an edge computing gateway cluster, a distributed storage array, an FPGA heterogeneous computing module, and a quantum key encryption module. The edge computing gateway cluster adopts a 5G+WiFi6 dual-mode communication architecture, with each gateway integrating 8 high-speed data acquisition interfaces. The distributed storage array adopts a distributed hash table storage structure, supporting millions of data read and write operations per second. The FPGA heterogeneous computing module is equipped with 4 computing core groups, each core group including 32 computing units. The quantum key encryption module encrypts data transmission based on the BB84 protocol. At the software level, the feature extraction layer of the graph convolutional node load perception model is set to 6 layers, with the number of convolutional kernels in each layer being 64, 128, 256, 256, 128, and 64, respectively. The decision window length of the progressive computing power adaptation decision model is set to 10, the number of game iterations of the node failure game equilibrium scheduling algorithm is set to 20, and the overall computing power situation data acquisition cycle is set to 50 milliseconds to ensure the real-time performance of computing power situation perception and the accuracy of scheduling optimization.

[0040] Preferred, such as Figure 2As shown, S2 includes the following sub-steps: S21, based on the input layer of the graph convolutional node load perception model, feature mapping is performed on the collected multi-dimensional computing power situation perception dataset to convert different types of situation data into feature vectors of a unified dimension, and the topological relationship between nodes is constructed through the adjacency matrix; S22, the first layer operation of graph convolution is started, the initial feature vector of each node and the feature vector of the neighboring nodes are weighted and summed, and a linear transformation is performed in combination with the bias term to obtain the first layer feature output; S23, the subsequent layers of graph convolution operations are executed in sequence, and each layer performs feature aggregation based on the output features of the previous layer and the node neighborhood information, and performs nonlinear feature transformation through the activation function to enhance the expressive ability of node load-related features; S24, the output features of the last layer are normalized and dimensionally compressed, the core load feature indicators are extracted, and the dynamic load evaluation results of each node are generated to provide data support for subsequent computing power adaptation decisions.

[0041] Specifically, step S2 performs feature extraction and load assessment for the graph convolution node load perception model. During implementation, the model feature extraction layer consists of 6 layers, with the number of convolution kernels in each layer being 64, 128, 256, 256, 128, and 64 respectively. The bias term for each layer ranges from 0.01 to 0.05. Step S21 standardizes the four types of core data collected by the intelligent analysis platform for the overall computing power situation, converting them into a 128-dimensional unified feature vector. A 1024×1024 adjacency matrix is ​​constructed based on node connection relationships, clarifying the association weights to provide structured input for graph convolution operations, ensuring consistency and accuracy in the calculations. Step S22 initiates the first layer of graph convolution operations, calling 64 3×3 convolution kernels to perform a weighted summation of the initial node features and neighborhood features. The normalization coefficient is the reciprocal of the square root product of the node degree, and a 0.02 bias term is added to complete the linear transformation, outputting a 64-dimensional feature vector, achieving initial feature fusion. S23 executes the subsequent five layers of operations sequentially, with each layer employing a corresponding number of 3×3 convolutional kernels. Based on the aggregation of features from the previous layer and neighborhood information, a non-linear transformation using the Sigmoid activation function is applied to enhance the expression of load-related features and improve discriminative power. S24 performs L2 normalization and principal component analysis on the 64-dimensional feature vector of the last layer, retaining 32 principal components as core indicators to generate dynamic load evaluation results. The data dimensionality is reduced by 50%, and parallel processing is achieved through four processing core groups in the FPGA heterogeneous computing module. Each core group simultaneously computes 32 nodes, ensuring the entire process is completed within 100 milliseconds, providing efficient and accurate data support for subsequent decision-making.

[0042] Preferred, such as Figure 3As shown, step S3 includes the following sub-steps: S31, associating and aligning the dynamic load assessment results generated in S2 with the task computing power demand priority data to construct a decision input dataset including node load status and task demand parameters, clarifying the constraints and optimization objectives of computing power adaptation decisions; S32, constructing a multi-objective optimization function based on the incremental computing power adaptation decision model, taking the matching degree between computing power supply and demand, scheduling delay, and resource consumption as optimization objectives, and setting the weight coefficients and constraint thresholds for each objective; S33, using a grid search algorithm to traverse the decision space and generate multiple initial computing power allocation schemes, each scheme including the node computing power allocation ratio, task allocation path, and resource scheduling timing parameters; S34, performing feasibility verification on the initial computing power allocation schemes, eliminating schemes that violate the constraints, retaining schemes that meet the basic scheduling requirements to form an initial computing power allocation scheme set, and submitting it to the risk verification stage.

[0043] Specifically, step S3 generates the initial scheme for the incremental computing power adaptation decision model, adhering to a decision window length of 10, a total multi-objective weight coefficient sum of 1, a grid search step size of 0.05, and matching degree deviation, execution latency, and resource consumption weights of 0.5, 0.3, and 0.2, respectively. S31 aligns the 32-dimensional dynamic load assessment results generated in step S2 with the task computing power demand priority data, collects node resource supply elasticity coefficients and response sensitivity, constructs a decision input dataset of over 1000 data points, sets constraints such as a node load capacity limit of 80%, a task execution latency of 500 milliseconds, and resource consumption not exceeding 120% of the baseline value, and clarifies the three major optimization objectives. S32 constructs a multi-objective optimization function, weighted sums according to weight coefficients, sets the optimization objective as the minimum value of the function, and sets a matching degree deviation threshold of 10%. S33 uses a grid search algorithm to traverse the decision space, including computing power allocation ratios of 0-100%, more than 10 task allocation paths, and scheduling time intervals of 10-100 milliseconds, combining historical decision data from 10 time points to generate multiple initial schemes. S34 performs feasibility verification on the initial schemes, eliminates schemes that violate the constraints, and retains the initial scheme set that meets the requirements. It promotes the process efficiently through the platform's computing resources, ensuring that the scheme set has both diversity and feasibility, thus laying the foundation for risk verification.

[0044] Preferred, such as Figure 4As shown, step S4 includes the following sub-steps: S41, acquiring historical fault data, hardware operating status parameters, and network connection stability indicators for each node; calculating the fault probability and fault impact range parameters for each node based on statistical analysis methods; and constructing a node fault risk assessment dataset; S42, inputting the initial computing power allocation scheme set into the node fault game equilibrium scheduling algorithm; using the fault probability of each node as the risk factor and the scheduling benefit as the benefit factor; and constructing the payoff matrix of the game model; S43, solving the game equilibrium solution through iterative calculation; analyzing the stability and benefit performance of each initial scheme under different fault scenarios; and selecting candidate schemes with strong fault resistance and excellent scheduling benefits; S44, conducting fault simulation tests on the candidate schemes to verify the computing power scheduling adjustment capability of the schemes under sudden node failures, ensuring that the schemes have good fault tolerance performance.

[0045] Specifically, step S4 verifies the risk of the node failure game-theoretic equilibrium scheduling algorithm, adhering to a game iteration count of 20 and a node failure risk weight coefficient range of 0.3 to 0.7. S41 collects historical failure data, hardware operating status parameters, and network connection stability indicators for each node. Through statistical analysis, it calculates the node failure probability and failure impact range parameters, constructing a node failure risk assessment dataset to provide basic risk data for game analysis. S42 inputs the initial computing power allocation scheme set into the algorithm, using node failure probability as the risk factor and scheduling revenue as the revenue factor. Risk weight coefficients are allocated based on node importance to construct the game model payoff matrix, clarifying the risk-reward relationship for each node under different schemes. S43 initiates calculation with a preset 20 iterations. Each iteration adjusts the scheme evaluation weights based on node strategy selection and the payoff matrix, analyzing the stability and revenue performance of each scheme under different failure scenarios, gradually converging to the game equilibrium state. S44 conducts fault simulation tests on the selected candidate solutions, simulating different node failure scenarios, observing the solution's computing power scheduling and adjustment capabilities, verifying the stability of the computing power service when a fault occurs, improving the execution efficiency of each step through a distributed computing architecture, ensuring that the selected candidate solutions meet the standards for fault resistance and fault tolerance, and providing a reliable foundation for subsequent dynamic optimization.

[0046] Preferred, such as Figure 5As shown, S5 includes the following sub-steps: S51, real-time collection of computing power usage status, task execution progress, and network transmission status data of each node through the global computing power status intelligent analysis platform, generating real-time computing power status feedback information, and comparing and analyzing it with the expected execution effect of the candidate scheme; S52, based on the feedback information, identifying deviations in the execution process of the candidate scheme, including uneven computing power allocation and excessive scheduling latency, and determining the types and directions of parameters that need to be adjusted; S53, combining the real-time load evaluation results of the graph convolution node load perception model, dynamically adjusting the computing power allocation ratio and task scheduling path parameters of the candidate scheme to optimize resource allocation efficiency; S54, repeating the status collection, deviation identification, and parameter adjustment process until the scheme execution effect meets the preset index requirements, completing the iterative optimization of the computing power allocation scheme, and forming the final scheduling scheme.

[0047] Specifically, step S5 performs global computing power status feedback and dynamic optimization of the solution, adhering to a computing power status collection cycle of 50 milliseconds, a preset computing power matching threshold of 90%, a latency threshold of 500 milliseconds, and a resource consumption threshold of 120%. S51 uses an edge computing gateway cluster to collect real-time data on the computing power usage status, task execution progress, and network transmission status of each node at 50-millisecond intervals. This data is encrypted by a quantum key encryption module and transmitted to the platform processing module to generate real-time computing power status feedback information, ensuring data real-time performance and security. S52 compares the feedback information with the expected execution effect of candidate solutions, including three core dimensions: allocation ratio, execution latency, and node load. This accurately identifies problems such as uneven computing power allocation and excessive scheduling latency, clarifying the type and direction of parameter adjustments. S53 combines the real-time evaluation results of the graph convolution node load perception model with the multi-objective optimization logic of the progressive computing power adaptation decision model to dynamically adjust key parameters such as the computing power allocation ratio and task scheduling path, ensuring that the adjustments still meet the three major optimization objectives. The S54 repeatedly collects situation data, identifies deviations, and adjusts parameters. After each round of adjustments, the effectiveness of the solution is evaluated until the preset index requirements are met, completing iterative optimization. Through the platform's real-time data processing capabilities and model computation support, the efficiency of each step is ensured, enabling the solution to adapt to dynamic changes in computing power and achieve precise matching in the spatiotemporal dimensions.

[0048] like Figure 6As shown, a scheduling optimization system for computing power resource allocation is applied to a scheduling optimization method for computing power resource allocation. This system includes: a multi-dimensional data acquisition unit for the overall computing power situation, a graph convolutional node load intelligent sensing unit, a progressive computing power adaptation decision generation unit, a node fault game-theoretic equilibrium scheduling verification unit, an overall computing power dynamic optimization and adjustment unit, and a scheduling scheme execution driving unit. The multi-dimensional data acquisition unit for the overall computing power situation is bidirectionally connected to the graph convolutional node load intelligent sensing unit, used to collect computing power-related data for all nodes and transmit it to the graph convolutional node load intelligent sensing unit. The graph convolutional node load intelligent sensing unit is electrically connected to the progressive computing power adaptation decision generation unit, and based on the collected data, completes node load evaluation and outputs it to the progressive computing power adaptation decision generation unit. The system comprises a computing power adaptation decision generation unit; a progressive computing power adaptation decision generation unit connected to a node fault game-theoretic equilibrium scheduling verification unit, generating an initial computing power allocation scheme set and transmitting it to the node fault game-theoretic equilibrium scheduling verification unit for risk verification; a node fault game-theoretic equilibrium scheduling verification unit connected to a global computing power dynamic optimization and adjustment unit, outputting candidate schemes to the global computing power dynamic optimization and adjustment unit; a global computing power dynamic optimization and adjustment unit electrically connected to a scheduling scheme execution driving unit, outputting the schemes to the scheduling scheme execution driving unit after iterative optimization; and a scheduling scheme execution driving unit connected to each computing power node, driving each node to perform computing power scheduling operations, with each unit working collaboratively to achieve efficient scheduling and optimization of global computing power resources.

[0049] This invention relates to a scheduling optimization method and system for computing power resource allocation. Addressing the shortcomings of traditional methods in terms of load perception and decision-making accuracy, it utilizes a comprehensive intelligent analysis platform for full-domain computing power status to collect multi-dimensional computing power data. Leveraging the neighborhood information aggregation and high-order feature mapping capabilities of a graph convolutional node load perception model, it deeply mines the dynamic load and topological correlation features of nodes. Combined with a multi-objective optimization mechanism of a progressive computing power adaptation decision model, it generates computing power allocation schemes highly matched to actual needs. This approach abandons traditional single-dimensional data dependence and static decision-making models, improving resource allocation accuracy and reducing resource waste and task latency. Addressing the weaknesses of traditional technologies in terms of fault tolerance and dynamic optimization capabilities, it employs a node fault game-theoretic equilibrium scheduling algorithm to analyze the probability of faults and scheduling benefits, selecting candidate schemes with optimal fault tolerance. Combined with real-time feedback and dynamic optimization adjustment mechanisms across the entire computing power status, it achieves iterative optimization of the scheme, ensuring rapid response and adjustment in the face of node faults or sudden changes in demand, and guaranteeing the stability and continuity of computing power services.

[0050] This method and system employ innovative hardware configurations, including edge computing gateway clusters and FPGA heterogeneous computing modules, coupled with a 5G+WiFi6 dual-mode communication architecture and distributed storage arrays, to achieve high-speed acquisition, secure storage, and efficient computation of computing power data. At the software level, optimization of key parameters such as the number of model feature extraction layers and the decision window length further enhances the real-time performance and reliability of scheduling. The collaborative architecture of each functional unit and the step-by-step, refined execution process ensure accurate matching and efficient collaboration of computing power resources across the entire domain in the spatiotemporal dimensions. This not only addresses the core pain points of traditional scheduling models but also provides comprehensive support for the large-scale deployment and high-quality operation of computing power networks.

[0051] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A scheduling optimization method for computing power resource allocation, characterized in that, Includes the following steps: S1. Collect information on the computing power occupancy status, task queue length, data transmission latency, and node connection topology of each node across the entire domain through a comprehensive intelligent computing power situation assessment and analysis platform to construct a multi-dimensional computing power situation awareness dataset. S2. Perform feature extraction and correlation analysis on the dataset based on a graph convolutional node load awareness model. Generate dynamic load assessment results for each node through node neighborhood information aggregation and high-order feature mapping. S3. Input the dynamic load assessment results into a progressive computing power adaptation decision model, combining task computing power demand priority, resource supply elasticity coefficient, and node response sensitivity to construct a multi-dimensional computing power situation awareness dataset. The objective is to optimize the decision space and generate an initial set of computing power allocation schemes; S4, a node failure game-theoretic equilibrium scheduling algorithm is used to perform risk verification on the initial set of computing power allocation schemes, and the candidate scheme with the best fault resistance is selected through game analysis of node failure probability and scheduling benefits; S5, based on the real-time situation feedback of the intelligent judgment and analysis platform for the global computing power situation, the candidate schemes are dynamically adjusted and iteratively optimized to match computing power resources in the spatiotemporal dimension; S6, the optimized computing power resource allocation and scheduling scheme is output, driving each node to execute computing power scheduling operations according to the scheme, and completing the dynamic allocation and collaborative scheduling of global computing power resources.

2. The scheduling optimization method for computing resource allocation according to claim 1, characterized in that, The expression for the graph convolutional node load-aware model is: ;in, For nodes The load feature vector, For activation function, For nodes The set of neighboring nodes, For nodes The degree, For the first The weight matrix of the layer, Neighboring nodes No. The feature vector of the layer, For the first Layer bias terms; For nodes The final load assessment value, For the feature fusion order, For the first Weighting coefficients for load characteristics of the first order. For the first The weighting coefficients of the first-order load characteristic gradient. for of Partial derivatives of order 1.

3. The scheduling optimization method for computing resource allocation according to claim 1, characterized in that, The expression for the incremental computing power adaptation decision model is: , in, For the first The initial computing power allocation decision at time step. For decision-making space, These are the weighting coefficients. For decision making The corresponding computing power supply, The computing power required for the task. For decision making Execution delay, For decision making Resource consumption costs; For the first Optimized computing power allocation decisions at each moment The length of the historical decision-making window. For the first The weight of each historical decision The attenuation coefficient is... For the first Historical decisions at every moment.

4. The scheduling optimization method for computing resource allocation according to claim 1, characterized in that, The expression for the node failure game-theoretic equilibrium scheduling algorithm is: in, The evaluation value of the scheduling strategy corresponding to the Bochian equilibrium solution. The total number of nodes participating in the game. For nodes Fault risk weighting coefficient For nodes The probability of failure, For nodes The benefit value after implementing the scheduling strategy This represents the maximum value of the revenue of all nodes.

5. The scheduling optimization method for computing resource allocation according to claim 1, characterized in that, The intelligent analysis platform for the overall computing power situation includes an edge computing gateway cluster, a distributed storage array, an FPGA heterogeneous computing module, and a quantum key encryption module. The edge computing gateway cluster adopts a 5G+WiFi6 dual-mode communication architecture, with each gateway integrating 8 high-speed data acquisition interfaces. The distributed storage array adopts a distributed hash table storage structure, supporting millions of data read and write operations per second. The FPGA heterogeneous computing module is equipped with 4 computing core groups, each core group including 32 computing units. The quantum key encryption module encrypts data transmission based on the BB84 protocol. At the software level, the feature extraction layer of the graph convolutional node load perception model is set to 6 layers, with the number of convolutional kernels in each layer being 64, 128, 256, 256, 128, and 64, respectively. The decision window length of the progressive computing power adaptation decision model is set to 10, the number of game iterations of the node failure game equilibrium scheduling algorithm is set to 20, and the overall computing power situation data acquisition cycle is set to 50 milliseconds to ensure the real-time performance of computing power situation perception and the accuracy of scheduling optimization.

6. The scheduling optimization method for computing resource allocation according to claim 1, characterized in that, S2 includes the following steps: S21, based on the input layer of the graph convolutional node load perception model, feature mapping is performed on the collected multi-dimensional computing power situation perception dataset to convert different types of situation data into feature vectors of a unified dimension, and the topological relationship between nodes is constructed through the adjacency matrix; S22, the first layer of graph convolution is started, and the initial feature vector of each node and the feature vector of the neighboring nodes are weighted and summed, and a linear transformation is performed in combination with the bias term to obtain the first layer feature output; S23, the graph convolution operations of each layer are executed in sequence, and each layer performs feature aggregation based on the output features of the previous layer and the node neighborhood information, and performs nonlinear feature transformation through the activation function to enhance the expressive ability of node load-related features; S24, the output features of the last layer are normalized and compressed in dimension, core load feature indicators are extracted, and dynamic load evaluation results of each node are generated to provide data support for computing power adaptation decisions.

7. The scheduling optimization method for computing resource allocation according to claim 1, characterized in that, S3 includes the following steps: S31, associating and aligning the dynamic load assessment results generated in S2 with the task computing power demand priority data to construct a decision input dataset including node load status and task demand parameters, clarifying the constraints and optimization objectives of computing power adaptation decisions; S32, constructing a multi-objective optimization function based on the incremental computing power adaptation decision model, taking the matching degree between computing power supply and demand, scheduling delay, and resource consumption as optimization objectives, and setting the weight coefficients and constraint thresholds for each objective; S33, using a grid search algorithm to traverse the decision space and generate multiple initial computing power allocation schemes, each scheme including the node computing power allocation ratio, task allocation path, and resource scheduling timing parameters; S34, performing feasibility verification on the initial computing power allocation schemes, eliminating schemes that violate the constraints, retaining schemes that meet the basic scheduling requirements to form an initial computing power allocation scheme set, and submitting it to the risk verification stage.

8. The scheduling optimization method for computing resource allocation according to claim 1, characterized in that, The S4 includes the following sub-steps: S41, acquiring historical fault data, hardware operating status parameters and network connection stability indicators of each node, calculating the fault probability and fault impact range parameters of each node based on statistical analysis methods, and constructing a node fault risk assessment dataset; S42, input the initial computing power allocation scheme set into the node failure game equilibrium scheduling algorithm, using the failure probability of each node as the risk factor and the scheduling revenue as the revenue factor, to construct the payoff matrix of the game model; S43, solve the game equilibrium solution through iterative calculation, analyze the stability and revenue performance of each initial scheme under different failure scenarios, and screen out candidate schemes with strong fault resistance and good scheduling revenue; S44, conduct failure simulation tests on the candidate schemes to verify the computing power scheduling adjustment capability of the schemes under sudden node failures, and ensure that the schemes have good fault tolerance performance.

9. The scheduling optimization method for computing resource allocation according to claim 1, characterized in that, S5 includes the following steps: S51, real-time collection of computing power usage status, task execution progress, and network transmission status data of each node through the global computing power status intelligent analysis platform, generating real-time computing power status feedback information, and comparing and analyzing it with the expected execution effect of the candidate scheme; S52, based on the feedback information, identifying deviations in the execution process of the candidate scheme, including uneven computing power allocation and excessive scheduling latency, and determining the types and directions of parameters that need to be adjusted; S53, combining the real-time load evaluation results of the graph convolution node load perception model, dynamically adjusting the computing power allocation ratio and task scheduling path parameters of the candidate scheme to optimize resource allocation efficiency; S54, repeating the status collection, deviation identification, and parameter adjustment process until the scheme execution effect meets the preset index requirements, completing the iterative optimization of the computing power allocation scheme, and forming the final scheduling scheme.

10. A scheduling and optimization system for computing power resource allocation, characterized in that, This system is applied to the scheduling optimization method for computing power resource allocation as described in claim 1, comprising: a multi-dimensional data acquisition unit for the overall computing power situation, a graph convolutional node load intelligent sensing unit, a progressive computing power adaptation decision generation unit, a node fault game-theoretic balanced scheduling verification unit, an overall computing power dynamic optimization and adjustment unit, and a scheduling scheme execution driving unit; the multi-dimensional data acquisition unit for the overall computing power situation is bidirectionally connected to the graph convolutional node load intelligent sensing unit, used to collect computing power-related data of the overall nodes and transmit it to the graph convolutional node load intelligent sensing unit; the graph convolutional node load intelligent sensing unit is electrically connected to the progressive computing power adaptation decision generation unit, and completes node load evaluation based on the collected data and outputs it to the progressive computing power adaptation decision generation unit. The system comprises a strategy generation unit; a progressive computing power adaptation decision generation unit connected to a node fault game-theoretic equilibrium scheduling verification unit, generating an initial computing power allocation scheme set and transmitting it to the node fault game-theoretic equilibrium scheduling verification unit for risk verification; a node fault game-theoretic equilibrium scheduling verification unit connected to a global computing power dynamic optimization and adjustment unit, outputting candidate schemes to the global computing power dynamic optimization and adjustment unit; a global computing power dynamic optimization and adjustment unit electrically connected to a scheduling scheme execution driving unit, outputting the schemes to the scheduling scheme execution driving unit after iterative optimization; and a scheduling scheme execution driving unit connected to each computing power node, driving each node to perform computing power scheduling operations, with each unit working collaboratively to achieve efficient scheduling and optimization of global computing power resources.

Images