A computing power resource scheduling method and system

Abstract

This invention provides a computing resource scheduling method and system, belonging to the field of computing resource scheduling technology. The scheduling method includes: collecting multi-source heterogeneous data to generate a full-dimensional fused dataset; using a pre-defined unsupervised learning technique to transform the full-dimensional fused dataset into a standardized unsupervised feature vector set; unifying the dispersed multi-dimensional potential energy components and using a pre-defined unsupervised clustering algorithm to perform comprehensive calculations of potential energy values; selecting unsupervised auxiliary causal variables and causal relationships to generate a preliminary system causal graph, and using a pre-defined quantization method to calculate the causal effect strength of the preliminary system causal graph; transforming the target task of the target computing facility into quantifiable sub-objectives, and using the do-calculus operation of the system causal graph model to generate multiple sets of candidate parameter tuning schemes. This invention achieves a balance between global optimization and local efficiency, proactively avoids potential risks, dynamically balances multiple objectives and continuously evolves, and maximizes overall benefits.

Description

Technical Field

[0001] This invention relates to the field of computing resource scheduling technology, and specifically to a computing resource scheduling method and system. Background Technology

[0002] With the rapid development of cloud computing, big data, and artificial intelligence technologies, computing power has become a key infrastructure supporting the development of the digital economy. As a core technology for the operation and management of data centers and computing power centers, computing resource scheduling methods aim to efficiently and rationally allocate computing tasks to heterogeneous computing nodes to meet performance, cost, and reliability requirements. The development of green computing centers emphasizes the efficient utilization of fluctuating renewable energy sources such as wind and solar power, which places higher demands on computing resource scheduling. This requires achieving multi-objective synergistic optimization, including minimum energy consumption, optimal cost, and green and low-carbon development, while ensuring the quality of computing services. Traditional scheduling methods, such as static rule-based strategies or simple heuristic algorithms, are no longer sufficient to cope with such complex dynamic environments. In recent years, although unsupervised learning and other artificial intelligence technologies have been introduced to mine energy consumption patterns from massive amounts of data, intelligent scheduling of computing resources still faces a series of severe challenges in practical applications.

[0003] Currently, computing resource scheduling methods still suffer from the following shortcomings: The architecture of the scheduling system has inherent contradictions. While centralized scheduling can perform global planning, it faces risks such as high decision-making latency, poor scalability, and single points of failure. While fully distributed scheduling offers rapid response and high reliability, decisions made by individual nodes based on local information are often short-sighted and lack a global perspective, easily leading the system into local optima and hindering the achievement of global energy efficiency goals. The scheduling decision-making process is opaque and lacks interpretability. Scheduling systems based on complex models such as deep neural networks are like "black boxes." When they make unconventional decisions, operations personnel struggle to understand their underlying logic, leading to a lack of trust. Furthermore, the system lacks the ability to foresee the long-term impact of decisions, hindering risk avoidance and strategy optimization. The system struggles to achieve dynamic collaborative optimization of multiple objectives and lacks adaptability. Faced with multiple potentially conflicting objectives such as energy efficiency, economy, and low carbon emissions, existing methods often employ fixed weights for weighted summation, failing to flexibly respond to continuous fluctuations in the external environment and internal states, resulting in long-term performance degradation. Summary of the Invention

[0004] The purpose of this invention is to provide a computing resource scheduling method and system to solve the contradiction between the difficulty of global optimization in centralized scheduling and the short-sightedness of local decision-making in distributed systems in existing computing resource scheduling methods, the lack of interpretability and predictability of scheduling decisions due to their black box nature, and the difficulty of coordinating multi-objective optimization.

[0005] To achieve the above objectives, this invention provides a computing resource scheduling method, comprising: collecting multi-source heterogeneous data from data acquisition devices deployed on the target computing facility, and generating a full-dimensional fused dataset through standardization cleaning, spatiotemporal alignment, and structured fusion; using a preset unsupervised learning technique to perform data purification, feature deep mining, and optimization, transforming the full-dimensional fused dataset into a standardized unsupervised feature vector set; defining and mapping multi-dimensional potential energy components based on the standardized unsupervised feature vector set, unifying the dispersed multi-dimensional potential energy components using a Min-Max normalization method, and utilizing a preset unsupervised clustering algorithm; Unsupervised adaptive weights are set to perform comprehensive calculation of potential energy values. Based on the full-dimensional fusion dataset and the standardized unsupervised feature vector set, unsupervised auxiliary causal variables and causal relationships are selected to generate a preliminary system causal graph. Combined with the historical fusion dataset, the causal effect strength of the preliminary system causal graph is calculated using a preset quantization method to obtain the system causal graph model. The target task of the target computing power facility is transformed into quantifiable sub-objectives. The association mapping between the sub-objectives and the target variables in the system causal graph model is established, a counterfactual problem set is constructed, and the prediction results are quantified using the do-calculus operation of the system causal graph model to generate multiple sets of candidate parameter tuning schemes.

[0006] Optionally, the step of utilizing a preset unsupervised learning technique to perform data purification, in-depth feature mining, and optimization, transforming the full-dimensional fusion dataset into a standardized unsupervised feature vector set, includes: using the K-means clustering algorithm to group data samples in the full-dimensional fusion dataset and removing abnormal data clusters that deviate from the mainstream clusters; using principal component analysis to perform correlation analysis on high-dimensional variables in the full-dimensional fusion dataset, identifying and removing redundant variables; using an unsupervised clustering algorithm to perform in-depth analysis on time-series data in the full-dimensional fusion dataset to uncover potential operational patterns; using time-series feature mining techniques to extract trend features, periodic features, and abrupt change features from the time-series data in the full-dimensional fusion dataset, and combining operational patterns with the target task to extract derived features and generate an initial feature set; and using the Min-Max normalization method to standardize all features in the initial feature set, fusing and optimizing to generate a standardized unsupervised feature vector set.

[0007] Optionally, the definition and feature mapping of multidimensional potential energy components, using the Min-Max normalization method, unifies the dispersed multidimensional potential energy components, including: defining the multidimensional potential energy components according to the target task of the target computing power facility; establishing the mapping relationship between each potential energy component and the core features in the standardized unsupervised feature vector set; quantifying each potential energy component through a weighted combination of corresponding features; using a unified Min-Max normalization method to map all potential energy components to the standard interval [0, 1]; combining the historical operating data of the target computing power facility with industry preset thresholds to determine the preset value range of each potential energy component; and pruning and calibrating abnormal potential energy components that exceed the range.

[0008] Optionally, the step of combining historical fusion datasets and using a preset quantization method to calculate the causal effect strength of the system causal graph prototype to obtain the system causal graph model includes: quantitatively calculating the causal effect strength of each directed edge in the system causal graph prototype using a preset quantization method to determine the degree and direction of influence between causal variables; grouping the quantified causal effects using a preset unsupervised learning algorithm to identify the differences in causal effects under different operating modes; and adjusting the topological structure and causal effect strength of the system causal graph prototype based on prior knowledge to eliminate false causal relationships and supplement missing target causal edges to obtain the system causal graph model.

[0009] Optionally, the step of using the do-calculus operation of the system causal graph model to quantify the prediction results and generate multiple candidate parameter tuning schemes includes: using the do-calculus operation of the system causal graph model to simulate the intervention effect of different potential energy component weight adjustment schemes on each sub-target and quantify the prediction results; using a preset unsupervised learning algorithm to analyze the latest data of the full-dimensional fusion dataset, identify the current system operation mode, and optimize the intervention prediction results in a targeted manner; comparing the benefits of the prediction results of different potential energy component weight adjustment schemes and generating multiple candidate parameter tuning schemes.

[0010] Optionally, the scheduling method further includes: establishing a multi-objective optimization decision model, combining the sub-objective prediction results corresponding to multiple candidate parameter tuning schemes, using a weighted summation method to calculate the comprehensive benefit score of each candidate parameter tuning scheme, selecting the optimal parameter tuning scheme, and dynamically updating the calculated potential energy according to the weight of the most advantageous energy component of the optimal parameter tuning scheme.

[0011] Optionally, the step of using a weighted summation method to calculate the comprehensive benefit score of each candidate parameter tuning scheme, selecting the optimal parameter tuning scheme, and dynamically updating the calculated potential energy according to the weight of the most advantageous energy component of the optimal parameter tuning scheme includes: setting weights according to the importance of the target task, calculating the comprehensive benefit score of each candidate parameter tuning scheme using a weighted summation method, selecting the weight combination with the highest comprehensive benefit score and meeting the constraint target requirements as the optimal parameter tuning scheme; verifying the rationality of the optimal parameter tuning scheme through a preset unsupervised anomaly detection algorithm, checking whether the weight of each potential energy component exceeds a preset safety threshold, and if the check fails, reselecting candidate parameter tuning schemes or adjusting the weight combination until the constraint target requirements are met.

[0012] Optionally, the scheduling method further includes: based on dynamically updated computational potential energy, using an unsupervised sorting algorithm to prioritize the running target tasks, prioritizing the migration of high-priority tasks, and migrating the target tasks without loss through local interaction and autonomous decision-making of distributed nodes, so that the target computing resources converge to nodes with lower computational potential energy.

[0013] Optionally, the unsupervised sorting algorithm is used to prioritize the running target tasks, prioritizing the migration of high-priority tasks. Through local interaction and autonomous decision-making by distributed nodes, the target tasks are migrated without loss. This includes: embedding a lightweight multicast communication module and service discovery mechanism in the local agent of each computing node to build a dynamically updated list of local neighbor nodes; each node comparing its own computational potential with that of all neighbor nodes, triggering a task migration evaluation process when the difference between its own computational potential and that of the target neighbor node reaches or exceeds a preset computational potential difference threshold; when a node identifies multiple qualified target migration nodes, prioritizing all currently running target tasks using an unsupervised sorting algorithm, prioritizing the migration of high-priority tasks to ensure the rapid achievement of core target tasks; and developing a personalized migration plan for each high-priority task, considering the resource status and task operation characteristics of the source and target nodes.

[0014] On the other hand, the present invention provides a computing resource scheduling system, the scheduling system including a control module, the control module including a memory, a processor and a computer program stored in the memory and executable on the processor, the processor executing the computer program to implement the computing resource scheduling method described in any one of the above.

[0015] Through the above technical solutions, this invention, by employing a distributed biomimetic executor and a system causal graph model, balances global optimization with local efficiency, effectively avoiding short-sighted behavior in distributed systems. The distributed decision-making mechanism allows the system to easily scale horizontally, and the failure of a single node does not affect overall operation. Through causal inference technology, an interpretable and predictable scheduling cognitive core is constructed, making the decision-making process interpretable and trustworthy, moving from passive response to proactive planning and actively avoiding potential risks. Through a dynamic potential energy field and a closed-loop learning mechanism, multi-objective dynamic trade-offs and continuous evolution are achieved, dynamically adjusting strategic priorities and flexibly balancing multiple objectives such as energy efficiency, economy, and low carbon emissions to maximize comprehensive benefits.

[0016] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 This is a flowchart illustrating the computing resource scheduling method of the present invention. Figure 1 ;

[0019] Figure 2 This is a schematic diagram of the process of converting into a standardized unsupervised feature vector set in this invention;

[0020] Figure 3 This is a flowchart illustrating the process of defining and mapping multidimensional potential energy components in this invention.

[0021] Figure 4 This is a schematic diagram of the process of obtaining the system cause-effect graph model in this invention;

[0022] Figure 5 This is a flowchart illustrating the process of generating multiple candidate parameter tuning schemes in this invention.

[0023] Figure 6 This is a schematic diagram of the process for dynamically updating and calculating potential energy in this invention;

[0024] Figure 7 This is a schematic diagram of the process of lossless migration of the target task in this invention;

[0025] Figure 8 This is a flowchart illustrating the computing resource scheduling method of the present invention. Figure 2 . Detailed Implementation

[0026] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0027] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0028] Please refer to Figure 1 and Figure 8 This invention provides a computing resource scheduling method, which may include:

[0029] Step S100: Collect multi-source heterogeneous data using data acquisition devices deployed on the target computing facility, and generate a full-dimensional fused dataset through standardized cleaning, spatiotemporal alignment, and structured fusion.

[0030] In a preferred embodiment of the present invention, the target computing facility may include key equipment and environmental locations such as server clusters, power supply cabinets, precision cooling units, and temperature and humidity monitoring points in the computer room. Lightweight, low-power data acquisition agents are deployed to ensure comprehensive data collection coverage. At the logical resource level, data acquisition agents adapted to the corresponding virtualization environment are deployed for virtual machine clusters, container scheduling platforms, task queue management systems, etc., to achieve real-time awareness of the operational status of logical resources. Data acquisition devices uniformly adopt industry-standard protocols such as IPMI, SNMP, and Modbus, acquiring raw data streams (e.g., server internal core temperature, cooling fan speed, C...) by real-time retrieval of device status data or receiving proactive pushes from devices. The system collects real-time data on PU / GPU / memory utilization, virtual machine instance start / stop status and resource usage, rack-level total power consumption, chiller unit operating power consumption, cooling tower inlet and outlet water temperatures, real-time output power of photovoltaic inverters, and real-time readings of grid interface electricity meters. Simultaneously, it establishes a stable data exchange channel with external systems through a standardized API gateway to acquire in batches macro-decision data such as weather forecasts for the coming period (e.g., sunshine duration, wind speed, temperature trends), grid time-of-use electricity pricing tables, regional carbon intensity factors, and green electricity consumption quota indicators. It also collects historical operational data from the computing center over the past three years (e.g., computing load fluctuation records, energy consumption statistics reports, equipment fault alarm logs).

[0031] In a preferred embodiment of the present invention, a sliding window filtering algorithm can be used to filter noise from continuous data (e.g., server temperature, energy output). An unsupervised anomaly detection algorithm (e.g., the Isolation Forest algorithm) can be used to identify and remove outliers in the dataset (e.g., abnormal data such as sudden changes in photovoltaic output, abnormal temperature peaks caused by server failures). Missing data can be filled in using a similarity-based interpolation method to ensure data integrity. A unified timeline based on UTC time is established, and the timestamps of all data streams are precisely aligned to a preset unified sampling time point. For data streams with sampling frequencies higher than the unified standard, a time aggregation algorithm is used for downsampling to retain key change features. For data streams with sampling frequencies lower than the unified standard, linear interpolation or polynomial interpolation methods are used for data synchronization to ensure the consistency and comparability of all data in the time dimension.

[0032] In a preferred embodiment of the invention, the cleaned and spatiotemporally aligned raw data undergoes systematic feature engineering processing, and specific feature extraction rules are designed for the characteristics of different types of data. For example, from time-series data of server power consumption and CPU utilization, derived features such as the moving average, variance, peak value, valley value, and ratio to cooling unit power consumption within a specified time window are calculated; from multi-energy output data, features such as output stability, peak-valley difference, and matching degree with load demand are extracted; from external macroeconomic data, features such as electricity price fluctuation coefficient, carbon intensity change trend, and influence coefficient of meteorological conditions on green electricity output are extracted. All extracted raw and derived features are horizontally concatenated along the timestamp dimension to construct a fusion dataset with a wide table structure. Each row of this dataset fully records all dimensions of information of the computing center at a specific time, including "equipment operating status - logical resource occupancy - multi-energy output - environmental parameters - economic costs - carbon emission indicators - historical load".

[0033] For example, at the Qinghai Provincial Green Computing Center, lightweight data acquisition agents are deployed across all physical equipment in the data center (e.g., 500 servers, 80 power cabinets, and 30 cooling units), 2000 virtual machine instances, and 50 container clusters. The SNMP protocol is used to acquire real-time equipment operation data, and the Qinghai Provincial Power Grid's time-of-use electricity price, regional carbon intensity factor, and 72-hour weather forecast data are synchronized via an API gateway. Simultaneously, the center's load fluctuation and energy consumption historical data from the past three years are imported. The isolated forest algorithm is used to remove outliers in photovoltaic output and server temperature noise data. Linear interpolation is used to supplement some missing wind power output data. After aligning all data streams to a 1-minute sampling time point, a fused dataset with over 120 dimensions, including timestamp, server core temperature, CPU / GPU utilization, multi-energy output value, data center temperature and humidity, real-time electricity price, carbon intensity, and historical load matching degree, is generated. This provides complete and standardized data support for subsequent unsupervised feature engineering and computational potential quantification.

[0034] Step S200: Using preset unsupervised learning techniques, perform data purification, in-depth feature mining and optimization, and transform the full-dimensional fusion dataset into a standardized unsupervised feature vector set.

[0035] Please refer to Figure 2 In this embodiment of the invention, a preset unsupervised learning technique is used to perform data purification, feature deep mining and optimization, transforming the full-dimensional fused dataset into a standardized unsupervised feature vector set, which may include:

[0036] Step S210: Use the K-means clustering algorithm to group the data samples in the full-dimensional fusion dataset and remove the abnormal data clusters that deviate from the mainstream clusters.

[0037] Step S220: Use principal component analysis to perform correlation analysis on high-dimensional variables in the full-dimensional fusion dataset, identify and remove redundant variables (e.g., highly correlated variables such as server temperature and data center temperature).

[0038] In a preferred embodiment of the present invention, after unsupervised data cleaning and enhancement and redundancy removal, core data dimensions directly related to energy consumption optimization, computing power scheduling, and green electricity consumption can be retained, reducing the complexity of subsequent feature processing and model calculation, while avoiding model overfitting problems caused by data redundancy.

[0039] Step S230: Use an unsupervised clustering algorithm to perform in-depth analysis on the time series data in the full-dimensional fusion dataset to uncover potential operational patterns.

[0040] In a preferred embodiment of the present invention, an unsupervised clustering algorithm (e.g., the DBSCAN algorithm) is used to perform in-depth analysis on the time series data in the fused dataset to identify energy consumption feature clusters (e.g., "high load-high energy consumption cluster" and "low load-green electricity surplus cluster") and computing power load pattern clusters (e.g., "continuous high load mode" and "fluctuating load mode") under different operating conditions, thereby mining the potential operating patterns in the dataset.

[0041] Step S240: Using time series feature mining technology, extract trend features, periodic features and abrupt change features from the time series data in the full-dimensional fusion dataset, and combine the operating rules and target tasks to extract derived features and generate an initial feature set.

[0042] In a preferred embodiment of the present invention, by using time-series feature mining techniques (e.g., periodic feature extraction based on Fourier transform, trend feature analysis based on sliding window), trend features (e.g., load growth / decline trend), periodic features (e.g., the period of peak daily load, the periodic peak of wind power output at night), and abrupt change features (e.g., nodes of sudden load increase / decrease) are extracted from time-series data such as computing power load fluctuations and changes in multi-energy output. Combined with the core requirements of green computing power optimization, the derived features such as green electricity consumption rate, carbon emission intensity, energy utilization efficiency, and equipment operation reliability are extracted to form an initial feature set covering multiple dimensions such as equipment status, energy utilization, economic cost, and carbon emission control.

[0043] Step S250: Use the Min-Max normalization method to standardize all features in the initial feature set, and then fuse and optimize to generate a standardized unsupervised feature vector set.

[0044] In a preferred embodiment of the present invention, the Min-Max normalization method is used to standardize all features in the initial feature set, mapping features with different dimensions and value ranges to the [0,1] interval, eliminating the impact of dimensional differences on subsequent calculations, and ensuring that each feature has a fair weight in the comprehensive evaluation. An unsupervised feature selection algorithm (e.g., feature selection based on mutual information, feature screening based on variance threshold) is used to calculate the contribution of each feature to the scheduling decision objective (e.g., maximizing green energy consumption, minimizing costs, etc.), selecting the top 80% of core features by contribution, eliminating invalid and weakly correlated features, and reducing the feature vector dimension. Finally, the selected core features are combined in a certain order to form a standardized unsupervised feature vector set, where each feature vector corresponds to key state information at a specific time point or calculation node.

[0045] For example, the DBSCAN unsupervised clustering algorithm was used to identify six typical operating mode clusters, including "high load - high green electricity", "low load - low electricity price", and "fluctuating load - wind power-dominated". 30 redundant variables, such as server surface temperature and rack ambient temperature, were removed using PCA dimensionality reduction technology. Features such as intraday fluctuation amplitude, peak occurrence time, and matching coefficient with load demand were extracted from photovoltaic output time series data. Features such as energy consumption-load correlation coefficient, equipment continuous operating time, and fault warning probability were extracted from server operation data. Combined with carbon intensity data, derived features such as "green electricity adaptability", "carbon emission correlation", and "energy utilization efficiency" were generated. All features were mapped to the [0, 1] interval through Min-Max normalization. Forty core features were selected using a feature selection algorithm based on mutual information, and finally a standardized unsupervised feature vector set was formed, which included key indicators such as energy utilization efficiency, green electricity consumption potential, carbon emission correlation, and equipment reliability.

[0046] Step S300: Based on the standardized unsupervised feature vector set, define and map the multidimensional potential energy components, use the Min-Max normalization method to unify the dispersed multidimensional potential energy components, and use the preset unsupervised clustering algorithm to set unsupervised adaptive weights to perform comprehensive calculation of potential energy values.

[0047] Please refer to Figure 3 In this embodiment of the invention, the definition and feature mapping of multidimensional potential energy components are performed, and the Min-Max normalization method is used to unify the dispersed multidimensional potential energy components, which may include:

[0048] Step S310: Based on the target task of the target computing power facility, clarify the definition of the multi-dimensional potential energy components, establish the mapping relationship between each potential energy component and the core features in the standardized unsupervised feature vector set, and quantify each potential energy component through the weighted combination of the corresponding features.

[0049] In a preferred embodiment of the present invention, the core optimization requirements of the green computing center can be combined to clarify the definition of the five potential energy components and establish the mapping relationship between each component and the core features in the standardized unsupervised feature vector set. First, thermal potential energy, which characterizes the temperature state and heat dissipation pressure of computing nodes, and maps to the features "normalized value of server core temperature" and "data center temperature and humidity matching degree" in the feature vector set. Second, resource potential energy, which reflects the occupancy and remaining carrying capacity of node computing resources (CPU / GPU / memory), and maps to the features "normalized value of CPU / GPU utilization", "remaining memory ratio", and "virtual machine / container deployment density". Third, economic potential energy, which is related to economic indicators such as electricity price cost and energy consumption cost, and maps to the features "normalized value of real-time electricity price" and "node energy consumption cost coefficient". Fourth, carbon potential energy, which reflects the carbon emission intensity and green electricity utilization level of node operation, and maps to the features "green electricity adaptability", "carbon emission correlation", and "green electricity absorption rate". Fifth, reliability potential energy, which characterizes the stability of node equipment and the reliability of task operation, and maps to the features "continuous operation time of equipment", "probability of fault warning", and "historical operation stability coefficient". Each potential energy component is quantitatively defined through a weighted combination of corresponding features.

[0050] Step S320: Using a unified Min-Max normalization method, all potential energy components are mapped to the standard range of [0, 1]. Combining the historical operating data of the target computing facility with the industry's preset threshold, the preset value range of each potential energy component is determined, and abnormal potential energy components that exceed the range are clipped and calibrated.

[0051] In a preferred embodiment of the present invention, since each potential energy component is calculated based on different types of features, their initial value ranges differ. A unified Min-Max normalization method is needed to map all potential energy components to the standard interval [0, 1] to eliminate the impact of differences in dimensions and value ranges, ensuring that each component has a fair weight in the comprehensive evaluation of calculated potential energy. During the normalization process, combined with historical operating data from the green computing center and industry standard thresholds, a reasonable value range for each potential energy component is determined (e.g., the lowest threshold for thermal potential energy corresponds to the lowest allowable operating temperature of the equipment, and the highest threshold corresponds to the alarm temperature of the equipment). Abnormal component values ​​exceeding the reasonable range are pruned and calibrated to ensure the consistency of the physical meaning of each component and the validity of the data.

[0052] In a preferred embodiment of the present invention, when performing comprehensive calculation of the potential energy value, the typical operating modes of the green computing center (e.g., "green electricity surplus mode", "peak electricity price mode", "high load operation mode") can be identified based on the analysis results of the fused dataset using an unsupervised clustering algorithm. For the differences in the contribution of each potential energy component to the scheduling optimization objective under different operating modes, the initial weight coefficients of each component are calculated using the entropy weight method to achieve adaptive weight allocation. Simultaneously, combined with domain expert knowledge, the initial weights calculated by the entropy weight method are fine-tuned and calibrated to form the final potential energy component weight set (e.g., Wheat, Wresource, Weconomic, Wcarbon, Wreliable). Finally, according to the formula "Calculated potential energy = Wheat × thermal potential energy + Wresource × resource potential energy + Weconomic × economic potential energy + Wcarbon × carbon potential energy + Wreliable × reliable potential energy", the five major potential energy components of each computing node are weighted and summed to obtain the comprehensive calculated potential energy value of each node, forming an initial calculated potential energy field covering all computing nodes in the entire network, constituting the initial topographic map for system scheduling decisions.

[0053] For example, for a core server in the Qinghai Green Computing Center, its thermal potential energy corresponding features (e.g., server core temperature normalization value 0.6, data center temperature and humidity matching degree 0.7) are extracted using a standardized unsupervised feature vector set, and the weighted calculation yields a thermal potential energy of 0.65; resource potential energy corresponding features (e.g., CPU utilization normalization value 0.5, memory remaining ratio 0.3, container deployment density 0.4) yields a resource potential energy of 0.42; economic potential energy corresponding features (e.g., real-time electricity price normalization value 0.3, energy consumption cost coefficient 0.2) yields an economic potential energy of 0.25; carbon potential energy corresponding features (e.g., green electricity adaptability 0.8, carbon emission correlation degree 0.1, green electricity consumption rate 0.9) yields a carbon potential energy of 0.82; and reliability potential energy corresponding features (e.g., equipment continuous running time normalization value 0.2, fault warning probability 0.1, historical stability coefficient 0.9) yields a reliability potential energy of 0.75. The weight set was determined by combining entropy weight method with expert fine-tuning as [Wheat=0.25, WCapacity=0.3, WEconomy=0.15, WCarbon=0.2, WCapacity=0.1]. Finally, the computational potential energy of the server was calculated as 0.25×0.65+0.3×0.42+0.15×0.25+0.2×0.82+0.1×0.75=0.1625+0.126+0.0375+0.164+0.075=0.565. This value intuitively reflects the current comprehensive adaptation status of the node and provides a quantitative basis for subsequent task migration decisions.

[0054] Step S400: Based on the full-dimensional fusion dataset and the standardized unsupervised feature vector set, select unsupervised auxiliary causal variables and causal relationships, generate a preliminary system causal graph, and combine it with the historical fusion dataset to calculate the causal effect strength of the preliminary system causal graph using a preset quantization method, thereby obtaining the system causal graph model.

[0055] In a preferred embodiment of the present invention, core variables directly related to each component of the calculated potential energy and the global optimization objective (e.g., green electricity consumption rate, operating cost, carbon emission intensity, etc.) are selected from the fused dataset and the standardized unsupervised feature vector set. These include multi-energy output (e.g., photovoltaic, wind power, traditional power), equipment operating status (e.g., temperature, utilization rate), external environment (e.g., weather, electricity price, carbon intensity), potential energy component weights, calculated potential energy values, and other key variables, forming a causal analysis variable set. Using the PC algorithm or Fast Causal Inference algorithm, combined with unsupervised clustering results (e.g., differences in variable distribution under different operating modes), the potential causal relationships between variables are automatically explored, identifying the topological structure of "who influences whom" among variables, and outputting a directed acyclic graph or a directed partially acyclic graph, forming a preliminary system causal graph. An unsupervised anomaly detection algorithm is used to remove abnormal samples from the variable set, avoiding interference from abnormal data in the discovery of causal relationships and improving the reliability of the preliminary system causal graph.

[0056] Please refer to Figure 4 In this embodiment of the invention, by combining historical fusion datasets and using a preset quantization method, the causal effect strength of the system causal graph prototype is calculated to obtain the system causal graph model, which may include:

[0057] Step S410: Using a preset quantification method, the causal effect strength of each directed edge in the prototype of the system causal graph is quantitatively calculated to determine the degree and direction of the influence between causal variables.

[0058] In a preferred embodiment of the present invention, based on the prototype of the system causal graph, historical data from the fusion dataset can be used to quantitatively calculate the causal effect strength of each directed edge in the causal graph through quantitative methods such as dual machine learning and linear regression, thereby clarifying the specific degree and direction of the influence between variables, such as "the specific extent of the reduction in carbon potential energy for every unit increase in photovoltaic output" and "the proportion of the decrease in operating costs for every 0.1 increase in economic potential energy weight".

[0059] Step S420: Using a preset unsupervised learning algorithm, the quantified causal effects are grouped to identify the differences in causal effects under different operating modes.

[0060] In a preferred embodiment of the present invention, an unsupervised learning algorithm (e.g., a density-based clustering algorithm) is introduced to group the quantified causal effects, identify the differences in causal effects under different operating modes (e.g., the intensity of the causal effect of "ambient temperature-thermal potential energy" is different in winter and summer), and calibrate the intensity of the effect accordingly.

[0061] Step S430: Based on prior knowledge, adjust the topological structure and causal effect strength of the prototype system causal graph, eliminate false causal relationships, and supplement missing target causal edges to obtain the system causal graph model.

[0062] In a preferred embodiment of the present invention, the topological structure and causal effect strength of the prototype system causal graph are fine-tuned and improved by combining the prior knowledge of domain experts, eliminating false causal relationships (e.g., variable pairs with high correlation but no actual causal relationship), and supplementing missing key causal edges (e.g., the causal relationship of "wind power output fluctuation - reliable potential energy"), and finally forming a system causal graph model with stable structure and accurate effect quantification.

[0063] For example, based on the fusion dataset and standardized unsupervised feature vector set of the Qinghai Provincial Green Computing Center, 20 core causal variables were selected, including photovoltaic output, wind power output, ambient temperature, electricity price, carbon potential weight, economic potential weight, green electricity consumption rate, and operating costs. Using the FCI algorithm combined with unsupervised clustering results, key causal relationships were discovered, such as "increased photovoltaic output → decreased carbon potential," "increased ambient temperature → increased thermal potential," and "increased economic potential weight → reduced energy consumption during high electricity price periods," forming a preliminary system causal graph. Through dual... Machine learning quantifies the causal effects, such as "for every 10% increase in photovoltaic output, carbon potential energy decreases by 8%" and "for every 5°C increase in ambient temperature, thermal potential energy increases by 0.2." Unsupervised clustering identifies three operating modes: "green electricity surplus," "peak electricity price," and "high load." The causal effects under different modes are calibrated (e.g., the effect strength of "photovoltaic output - carbon potential energy" is higher in the "green electricity surplus mode" than in other modes). Expert knowledge is used to supplement the causal edge of "wind power output fluctuation → reduced reliable potential energy," ultimately forming a system causal graph model. This model can accurately predict the impact of "adjusting carbon potential energy weights" on the green electricity absorption rate, providing logical support for subsequent calculations of dynamic parameter adjustment of the potential energy field.

[0064] Step S500: Transform the target task of the target computing facility into quantifiable sub-objectives, establish the correlation mapping between the sub-objectives and the target variables in the system causal graph model, construct a counterfactual problem set, use the do-calculus operation of the system causal graph model to quantify the prediction results, and generate multiple sets of candidate parameter tuning schemes.

[0065] In a preferred embodiment of the present invention, a global optimization objective system is defined in conjunction with the core operational needs of the green computing center. This system includes core objectives (e.g., increasing green electricity consumption rate, reducing total operating costs, and decreasing carbon emission intensity) and constraint objectives (e.g., system operational stability and ensuring equipment lifespan). These macro-level objectives are broken down into quantifiable and calculable sub-objectives, such as "increasing the green electricity consumption rate to a specified percentage," "reducing operating costs during peak hours by a specified amount," "controlling carbon emission intensity below industry standards," and "keeping equipment failure rate below a threshold." A mapping relationship is established between these sub-objectives and key variables in the system's causal graph model. This transforms the macro-level objectives into optimization problems solvable by the causal model, forming a chain of "objective-variable-potential energy components," for example, "increasing the green electricity consumption rate → increasing the carbon potential energy weight → adjusting the weight ratio of carbon potential energy and economic potential energy."

[0066] In a preferred embodiment of the present invention, based on the system causal graph model, a set of counterfactual questions is constructed for the target-variable correlation mapping relationship, such as "If the carbon potential weight is increased by ΔW, how much will the green electricity consumption rate increase?" and "If the economic potential weight is adjusted before the peak electricity price, how much can the operating cost be reduced?"

[0067] Please refer to Figure 5 In this embodiment of the invention, the prediction results are quantified using the do-calculus operation of the system causal graph model, and multiple candidate parameter tuning schemes are generated, which may include:

[0068] Step S510: Using the do-calculus operation of the system causal graph model, simulate the intervention effect of different potential energy component weight adjustment schemes on each sub-target, and quantify the prediction results.

[0069] In a preferred embodiment of the present invention, the do-calculus operation of the system causal graph model is used to simulate the intervention effect of different potential energy component weight adjustment schemes on each sub-objective, and the prediction results are quantified (e.g., "increasing the carbon potential energy weight by 0.1 will increase the green electricity consumption rate by 5%"). Unsupervised learning algorithms (e.g., online clustering algorithms) are introduced to analyze the latest data of the fused dataset in real time, identify the current system operation mode (e.g., "green electricity surplus mode", "peak electricity price mode", "low load mode"), and optimize the intervention prediction results in a targeted manner—for example, under the "green electricity surplus mode", the prediction accuracy of carbon potential energy weight adjustment on green electricity consumption rate is enhanced; under the "peak electricity price mode", the prediction effect of economic potential energy weight adjustment on operating costs is optimized. By comparing the prediction benefits of different weight adjustment schemes, multiple sets of candidate parameter adjustment schemes are generated.

[0070] Step S520: Using a pre-defined unsupervised learning algorithm (e.g., online clustering algorithm), analyze the latest data of the full-dimensional fusion dataset, identify the current system operation mode (e.g., "green electricity surplus mode", "peak electricity price mode", "low load mode"), and optimize the intervention prediction results accordingly.

[0071] In a preferred embodiment of the present invention, when optimizing the intervention prediction results in a targeted manner, for example, under the "green electricity surplus mode", the prediction accuracy of carbon potential energy weight adjustment on green electricity consumption rate is enhanced; under the "electricity price peak mode", the prediction effect of economic potential energy weight adjustment on operating costs is optimized.

[0072] Step S530: Compare the benefits of different potential energy component weight adjustment schemes in the prediction results, and generate multiple sets of candidate parameter tuning schemes.

[0073] Step S600: Establish a multi-objective optimization decision model, combine the sub-objective prediction results corresponding to multiple candidate parameter tuning schemes, use the weighted summation method to calculate the comprehensive benefit score of each candidate parameter tuning scheme, select the optimal parameter tuning scheme, and dynamically update the calculated potential energy according to the weight of the most advantageous energy component of the optimal parameter tuning scheme.

[0074] Please refer to Figure 6 In this embodiment of the invention, a weighted summation method is used to calculate the comprehensive benefit score of each candidate parameter tuning scheme, select the optimal parameter tuning scheme, and dynamically update the calculated potential energy according to the weight of the optimal optimal energy component of the optimal parameter tuning scheme. This may include:

[0075] Step S610: Set weights according to the importance of the target task, and use the weighted summation method to calculate the comprehensive benefit score of each candidate parameter tuning scheme (for example, the weights are set according to the importance of the target). Select the weight combination with the highest comprehensive benefit score and that meets the constraint target requirements as the optimal parameter tuning scheme.

[0076] Step S620: Verify the rationality of the optimal parameter tuning scheme through a preset unsupervised anomaly detection algorithm, and check whether the weights of each potential energy component exceed the preset safety threshold. If the verification fails, reselect a candidate parameter tuning scheme or adjust the weight combination until the constraint target requirements are met.

[0077] In a preferred embodiment of the present invention, to avoid drastic fluctuations in the calculated potential energy field caused by weight adjustments, a safety verification mechanism for weight distribution is established: The rationality of the optimal weight combination is verified using an unsupervised anomaly detection algorithm (e.g., anomaly detection based on statistical distribution), checking whether the weights of each potential energy component exceed the safe range and whether the weight changes are too drastic; the impact of the optimal weight combination on system stability is predicted using a system causal graph model, ensuring that the adjusted device operating status and task execution efficiency do not fluctuate significantly. If the verification fails, the system returns to the candidate solution pool to reselect or fine-tune the weight combination until the constraint target requirements are met.

[0078] In a preferred embodiment of the present invention, the weighted combination of the most advantageous potential energy components that have passed security verification is distributed in batches to the local agents of all computing nodes in the network through an encrypted communication channel. After receiving the weight update instruction, each node's local agent first performs local verification (e.g., verifying the integrity of the instruction and the legality of the weights). After the verification is passed, the local stored potential energy component weight set is updated, and the calculated potential energy value of the node is recalculated according to the new weights. After all nodes have completed the updated calculated potential energy, the updated calculated potential energy value is synchronized to the entire network through a multicast mechanism, realizing the unified reshaping of the network's calculated potential energy field and providing an updated "topography map" for subsequent task migration decisions.

[0079] For example, when the overall optimization goal of the Qinghai Provincial Green Computing Center is set as "increasing the green electricity consumption rate to 85% and controlling operating costs," it is broken down into three sub-goals: "increasing the green electricity consumption rate by 10%," "reducing peak-hour operating costs by 15%," and "reducing equipment failure rate to below 0.5%." A correlation mapping between "green electricity consumption rate - carbon potential weight" and "operating cost - economic potential weight" is established. Based on a system causal graph model, five candidate schemes are simulated, including "increasing the carbon potential weight from 0.2 to 0.3" and "adjusting the economic potential weight from 0.15 to 0.1." Online unsupervised clustering identifies the current "green electricity surplus" operating mode, optimizes the prediction results, and finds that the combination of "increasing the carbon potential weight by 0.3, adjusting the economic potential weight to 0.1, maintaining the resource potential weight at 0.3, the thermal potential weight at 0.25, and the reliability potential weight at 0.05" can increase the green electricity consumption rate by 12%, reduce peak-hour costs by 18%, and meet equipment stability constraints, resulting in the optimal overall benefit. After the unsupervised anomaly detection verifies that the weight change range is reasonable, the weight combination is distributed to all nodes. Each node recalculates its potential energy, and the updated potential energy field tilts towards nodes with sufficient green energy utilization, guiding the task to migrate to these nodes.

[0080] Step S700: Based on dynamically updated computational potential, an unsupervised sorting algorithm is used to prioritize the running target tasks, migrate high-priority tasks first, and migrate the target tasks without loss through local interaction and autonomous decision-making of distributed nodes, so that the target computing resources can be concentrated on nodes with lower computational potential.

[0081] Please refer to Figure 7 In this embodiment of the invention, an unsupervised sorting algorithm is used to prioritize the running target tasks, prioritizing the migration of high-priority tasks. Through local interaction and autonomous decision-making of distributed nodes, the target tasks are migrated without loss. This may include:

[0082] Step S710: Integrate a lightweight multicast communication module and service discovery mechanism into the local agent of each computing node to build a dynamically updated list of local neighbor nodes.

[0083] In a preferred embodiment of the present invention, each computing node's local agent has a built-in lightweight multicast communication module and service discovery mechanism. It sends status broadcast packets to the network at preset intervals (e.g., every 10 seconds), containing key information such as the node's latest computing potential energy value, remaining computing resources (CPU / GPU / memory balance), green energy utilization ratio, and device operating status. Simultaneously, it receives broadcast packets from other nodes in the network and uses an unsupervised clustering algorithm to quickly group the received node information, identifying "neighboring nodes" that are close to the node both in network and physical distance (e.g., avoiding cross-regional, high-latency node communication), and constructing a local neighboring node list. The neighboring node list is updated in real time. When a node's computing potential energy changes significantly or its device status becomes abnormal, an instant synchronization mechanism is triggered to ensure that each node's perception of its surrounding "potential energy terrain" is real-time and accurate.

[0084] Step S720: Each node compares its own calculated potential energy value with that of all neighboring nodes. When the difference between the calculated potential energy of the node and that of the target neighboring node reaches or exceeds the preset calculated potential energy difference threshold, the task migration evaluation process is triggered.

[0085] In a preferred embodiment of the present invention, during the evaluation process, a cost-benefit accounting model is constructed using an unsupervised learning algorithm: on the one hand, based on historical migration data and real-time network status, migration costs are estimated, including network bandwidth usage costs, data transmission latency costs, performance jitter costs of task pause and recovery, and target node resource reservation costs; on the other hand, based on computational potential difference and global optimization objectives, migration benefits are quantified, including benefits from improved green energy consumption, reduced operating costs, reduced carbon emissions, and reduced equipment wear and tear. By calculating "net benefit = migration benefit - migration cost", if the net benefit is positive and reaches a preset threshold, the neighboring node is determined to be a qualified target migration node, and the migration preparation stage begins; if the net benefit is negative or does not reach the threshold, the neighboring node is abandoned, and other neighboring nodes continue to be monitored.

[0086] Step S730: When a node identifies multiple qualified target migration nodes, it prioritizes all currently running target tasks using an unsupervised sorting algorithm, prioritizing the migration of high-priority tasks (e.g., core business tasks, high-energy-consumption tasks, and tasks with low green electricity compatibility) to ensure that core target tasks are achieved quickly.

[0087] Step S740: For each high-priority task, formulate a personalized migration plan based on the resource status and task operation characteristics of the source and target nodes.

[0088] In a preferred embodiment of the present invention, a personalized migration plan can be formulated by combining the resource status of the source node and the target node (e.g., number of CPU cores, memory size, network bandwidth) and task running characteristics (e.g., memory usage, data transmission volume, real-time requirements): for tasks with low real-time requirements and small data volume, a fast migration mode (e.g., direct serialization transmission) is adopted; for tasks with high real-time requirements and large data volume, a breakpoint resume migration mode (e.g., block transmission + state synchronization) is adopted; for critical core tasks, a dual-active migration mode is adopted (e.g., the target node first loads the task image, runs it synchronously, and then switches traffic) to ensure that the migration process does not affect business continuity.

[0089] In a preferred embodiment of the present invention, according to the established migration scheme, the source node uses Checkpointing technology (e.g., CRIU tool for container tasks, snapshot technology for virtual machine tasks) to serialize the complete running state of the task (e.g., including memory data, register information, open file handles, network connection status, etc.) to generate a task state file. The state file is transmitted to the target node via a high-speed network channel, employing data compression and verification technologies during transmission to ensure efficient and complete data transmission. Upon receiving the state file, the target node immediately performs deserialization processing to restore the task's running environment and state. After verifying that the task can run normally, it sends an acknowledgment signal to the source node. Upon receiving the acknowledgment, the source node suspends local task execution, switches business traffic to the target node, and completes the migration. During the migration process, an unsupervised anomaly detection algorithm monitors the transmission status and task recovery status in real time. If transmission failures or abnormal state recovery occur, a fault tolerance mechanism is immediately triggered (e.g., switching to a backup target node or rolling back to the source node) to ensure a smooth, lossless migration process without user awareness.

[0090] For example, node A (computational potential energy 0.75) of the Qinghai Provincial Green Computing Center discovers its neighbor node B (computational potential energy 0.5) during state synchronization. The difference in computational potential energy between the two is 0.25, reaching the preset threshold ΔP=0.15, triggering a migration assessment. Through a cost-benefit analysis model, the migration cost of the high-energy-consuming batch processing task on node A includes a 0.5-second network transmission delay and 100Mbps bandwidth usage. The migration benefits include an 8% increase in green electricity consumption and a 20 yuan reduction in hourly cost. The net benefit is positive and meets the target, thus node B is deemed a qualified target. Tasks on node A are prioritized, with the high-energy-consuming batch processing task being migrated first. A block-based migration plan is developed: the source node uses the CRIU tool to serialize the task status and transmits it to node B in three blocks, verifying each block after transmission. After receiving the blocks, node B deserializes and restores the status, verifies that the task can run normally, and then sends an acknowledgment signal to node A. Node A pauses the task and switches traffic to node B. The entire migration process takes 3 seconds, with no interruption to user services. After the migration, the calculated potential energy of node A decreased to 0.6, while the calculated potential energy of node B increased to 0.58, resulting in a better distribution of the calculated potential energy field.

[0091] This invention also provides a computing resource scheduling system. The scheduling system includes a control module, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the computing resource scheduling method described above.

[0092] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0093] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0094] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0096] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0097] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0098] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0099] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0100] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for scheduling computing resources, characterized in that, The scheduling method includes: Based on the data acquisition equipment deployed on the target computing facility, multi-source heterogeneous data is collected, and through standardized cleaning, spatiotemporal alignment and structured fusion, a full-dimensional fused dataset is generated; By utilizing pre-defined unsupervised learning techniques, data purification, in-depth feature mining and optimization are performed to transform the full-dimensional fusion dataset into a standardized unsupervised feature vector set; Based on a standardized unsupervised feature vector set, multidimensional potential energy components are defined and feature mapped. The Min-Max normalization method is used to unify the dispersed multidimensional potential energy components. A preset unsupervised clustering algorithm is used to set unsupervised adaptive weights and perform comprehensive calculation of potential energy values. Based on the full-dimensional fusion dataset and the standardized unsupervised feature vector set, unsupervised auxiliary causal variables and causal relationships are selected to generate a preliminary system causal graph. Combined with the historical fusion dataset, the causal effect strength of the preliminary system causal graph is calculated using a preset quantization method to obtain the system causal graph model. The target task of the target computing power facility is transformed into a quantifiable sub-objective. The association mapping between the sub-objective and the target variable in the system causal graph model is established, a counterfactual problem set is constructed, and the prediction results are quantified by using the do-calculus operation of the system causal graph model to generate multiple sets of candidate parameter tuning schemes.

2. The computing resource scheduling method according to claim 1, characterized in that, The process involves using pre-defined unsupervised learning techniques to perform data purification, in-depth feature mining, and optimization, transforming the full-dimensional fused dataset into a standardized unsupervised feature vector set, including: The K-means clustering algorithm is used to group the data samples in the full-dimensional fused dataset and remove outlier clusters that deviate from the mainstream clusters. Principal component analysis was used to perform correlation analysis on high-dimensional variables in the full-dimensional fused dataset to identify and remove redundant variables. An unsupervised clustering algorithm is used to perform in-depth analysis on time series data in a multi-dimensional fused dataset to uncover potential operational patterns. By using time-series feature mining technology, trend features, periodic features, and abrupt change features are extracted from time-series data in a multi-dimensional fusion dataset. Combined with operational patterns and target tasks, derived features are extracted to generate an initial feature set. The Min-Max normalization method is used to standardize all features in the initial feature set and then fuse and optimize them to generate a standardized unsupervised feature vector set.

3. The computing resource scheduling method according to claim 1, characterized in that, The process of defining and mapping multidimensional potential energy components, and using the Min-Max normalization method to unify the dispersed multidimensional potential energy components, includes: Based on the target tasks of the target computing power facility, the definition of multi-dimensional potential energy components is clarified, and the mapping relationship between each potential energy component and the core features in the standardized unsupervised feature vector set is established. Each potential energy component is quantitatively defined by the weighted combination of the corresponding features. A unified Min-Max normalization method is adopted to map all potential energy components to the standard range of [0, 1]. Combined with the historical operation data of the target computing facility and the industry preset threshold, the preset value range of each potential energy component is determined, and abnormal potential energy components that exceed the range are clipped and calibrated.

4. The computing resource scheduling method according to claim 1, characterized in that, The process of combining historical fusion datasets and using a pre-defined quantization method to calculate the causal effect strength of the system causal graph prototype, thereby obtaining the system causal graph model, includes: By using a pre-defined quantification method, the causal effect strength of each directed edge in the prototype of the system's causal graph is quantitatively calculated to determine the degree and direction of the influence between causal variables. Using a pre-defined unsupervised learning algorithm, the quantified causal effects are grouped to identify the differences in causal effects under different operating modes; By combining prior knowledge, the topological structure and causal effect strength of the prototype system causal graph are adjusted, false causal relationships are eliminated, and missing target causal edges are added to obtain the system causal graph model.

5. The computing resource scheduling method according to claim 1, characterized in that, The method utilizes the do-calculus operation of the system's causal graph model to quantify the prediction results and generate multiple sets of candidate parameter tuning schemes, including: Using the do-calculus operation of the system causal graph model, the intervention effect of different potential energy component weight adjustment schemes on each sub-objective is simulated, and the prediction results are quantified. Using a pre-defined unsupervised learning algorithm, the latest data from the full-dimensional fusion dataset is analyzed to identify the current system's operating mode and optimize the intervention prediction results accordingly. By comparing the predictive benefits of different potential energy component weight adjustment schemes, multiple candidate parameter tuning schemes are generated.

6. The computing resource scheduling method according to claim 1, characterized in that, The scheduling method further includes: A multi-objective optimization decision model is established. Combining the sub-objective prediction results of multiple candidate parameter tuning schemes, the weighted summation method is used to calculate the comprehensive benefit score of each candidate parameter tuning scheme, select the optimal parameter tuning scheme, and dynamically update the calculated potential energy according to the weight of the most advantageous energy component of the optimal parameter tuning scheme.

7. The computing resource scheduling method according to claim 6, characterized in that, The process employs a weighted summation method to calculate the comprehensive benefit score of each candidate parameter tuning scheme, selects the optimal parameter tuning scheme, and dynamically updates the calculated potential energy based on the weight of the optimal optimal energy component. This includes: Weights are assigned based on the importance of the target task. The weighted summation method is used to calculate the comprehensive benefit score of each candidate parameter tuning scheme. The weight combination with the highest comprehensive benefit score and that meets the constraint target requirements is selected as the optimal parameter tuning scheme. The rationality of the optimal parameter tuning scheme is verified by a pre-defined unsupervised anomaly detection algorithm. The weights of each potential energy component are checked to see if they exceed the preset safety threshold. If the check fails, a new candidate parameter tuning scheme is selected or the weight combination is adjusted until the constraint target requirements are met.

8. The computing resource scheduling method according to claim 6, characterized in that, The scheduling method further includes: Based on dynamically updated computational potential, an unsupervised sorting algorithm is used to prioritize the running target tasks, and high-priority tasks are migrated first. Through local interaction and autonomous decision-making of distributed nodes, the target tasks are migrated without loss, so that the target computing resources are concentrated on nodes with lower computational potential.

9. The computing resource scheduling method according to claim 8, characterized in that, The method of using an unsupervised sorting algorithm to prioritize the target tasks and migrate high-priority tasks first, and then migrating the target tasks without loss through local interaction and autonomous decision-making of distributed nodes, includes: A lightweight multicast communication module and service discovery mechanism are built into the local agent of each computing node to build a dynamically updated list of local neighbor nodes. Each node compares its own calculated potential energy value with that of all neighboring nodes. When the difference between the calculated potential energy of the node and that of the target neighboring node reaches or exceeds the preset calculated potential energy difference threshold, the task migration evaluation process is triggered. When a node identifies multiple qualified target migration nodes, it prioritizes all currently running target tasks using an unsupervised sorting algorithm, prioritizing the migration of high-priority tasks to ensure that core target tasks are achieved quickly. For each high-priority task, a personalized migration plan is developed by combining the resource status of the source and target nodes with the task's operational characteristics.

10. A computing resource scheduling system, characterized in that, The scheduling system includes a control module, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the computing resource scheduling method according to any one of claims 1-9.

Images