A software-defined resource management and scheduling method based on a server cluster platform

Abstract

The application provides a software-defined resource management and scheduling method based on a server cluster platform, carries out software and hardware resource modeling on the server cluster, forms a static resource demand baseline through the foreign key association of a hardware resource table, a scheduling state table and a software resource demand table; a three-level deep feature extraction architecture can mine complex nonlinear resource use rules, a multi-task learning framework simultaneously outputs prediction results in three dimensions of resource demand, node matching degree and task execution risk; combined with a comprehensive matching degree calculation formula, accurate alignment of resource demand fluctuation curves and node available resource fluctuation curves is realized. The application is simultaneously adapted to space vehicle on-orbit management earth-space integrated computing power scheduling scenes of a satellite measurement and control center, a low-orbit satellite constellation, a manned space station, a deep space probe and the like, and solves problems such as space measurement and control task time window constraints, multi-satellite heterogeneous resource coupling, large migration delay in abnormal service emergency disposal and the like.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and specifically to a software-defined resource management and scheduling method based on a server cluster platform. Background Technology

[0002] Server clusters are the core hardware carriers of complex and large-scale software systems. In traditional software systems, the deployment location and resource usage configuration of software are fixed, which has obvious technical defects: On the one hand, the software is strongly bound to the physical server, and its operation is highly dependent on the hardware devices and operating system environment of the nodes. It cannot achieve dynamic migration across devices, and the CPU, memory, storage and other resources required for the software to run are pre-allocated and "exclusively used" by the software. Even if the server is under low load, idle resources cannot be released for other software to use. There is no resource scheduling mechanism on demand, resulting in extremely low cluster resource utilization. On the other hand, when the server node to which the software is bound experiences hardware failure or resource abnormality, the software cannot run normally, and the business is interrupted. There is no automatic fault handling mechanism, and manual intervention is required for troubleshooting and recovery. The system has poor reliability and high operation and maintenance costs. Traditional server cluster scheduling solutions are only designed for general IT business and cannot adapt to the specific business constraints of satellite telemetry, tracking and command (TT&C) and spacecraft on-orbit management. When the GPU / network card hardware of the ground TT&C server fails, the traditional system can only manually migrate the TT&C business, which takes hours and is prone to missing the golden window for handling business failures. There is a risk of on-orbit loss of control. There is a lack of an automated TT&C task seamless migration mechanism based on node health prediction.

[0003] In existing technologies, some solutions have introduced container technology and machine learning scheduling algorithms, which have achieved dynamic resource allocation to a certain extent. However, they still have the following key shortcomings: First, they ignore the heterogeneity of cluster resources. Different types of resources, such as CPU, GPU, memory, and network, have drastically different usage characteristics and interaction relationships. Existing models use a unified feature extraction method, which cannot effectively capture the differentiated patterns of heterogeneous resources. Second, they lack modeling of the spatiotemporal correlation between nodes. Clusters are distributed systems, and there are network communication, load transmission, and resource competition relationships between nodes. Resource usage not only has temporal dependence but also spatial correlation. Existing models only consider the temporal characteristics of a single node, leading to inaccurate global scheduling decisions. Third, the feature extraction level is relatively shallow, which cannot uncover complex nonlinear resource usage patterns. For dynamic load tasks and large-scale cluster scenarios, the accuracy of resource demand prediction and node matching degree are low. Fourth, they do not consider task execution risks and only schedule based on resource matching degree, which can easily lead to task failure due to node load fluctuations or hidden faults.

[0004] Therefore, there is an urgent need for an intelligent resource management and scheduling scheme that can effectively model the interaction relationships of heterogeneous resources and the spatiotemporal correlation of nodes. Through deeper and more comprehensive deep feature extraction technology, the cluster resource utilization rate and system operation reliability can be further improved to solve the industry pain points of the existing aerospace operation and control scheduling system, such as satellite transit computing power burst conflict, lack of space-ground topology spatial correlation modeling, low prediction accuracy of aerospace multimodal telemetry and control data resources, lack of resource preemption mechanism for high-priority emergency telemetry and control tasks, and excessively long migration delay of abnormal services. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies, such as neglecting resource heterogeneity and spatiotemporal correlation of nodes, shallow feature extraction, inaccurate resource demand prediction, low node matching accuracy, failure to consider task execution risks, inability to adapt to the operating environment of large-scale complex clusters, and existing aerospace operation and control scheduling systems suffering from satellite transit computing power burst conflicts, low accuracy of aerospace multimodal telemetry and control data resource prediction, lack of resource preemption mechanisms for high-priority emergency telemetry and control tasks, and excessively long latency in abnormal business migration. It provides a software-defined resource management and scheduling method based on a server cluster platform. This method performs dual-dimensional modeling of server cluster resources (hardware and software). The hardware dimension covers node brand, CPU / GPU architecture, number of cores, memory, disk, network bandwidth, and scheduling attributes such as online status, health status, isolation status, and load level. The software dimension configures and saves runtime resource requirements by type, and resource management and scheduling are carried out entirely based on this model. The three-level deep feature extraction architecture can uncover complex nonlinear resource usage patterns. The multi-task learning framework simultaneously outputs prediction results across three dimensions: resource demand, node matching degree, and task execution risk. Combined with the comprehensive matching degree calculation formula, it achieves precise alignment between the resource demand fluctuation curve and the node's available resource fluctuation curve, reducing resource demand prediction errors in dynamic load scenarios and improving node resource utilization. It is also suitable for integrated space-ground computing power scheduling scenarios for on-orbit management of spacecraft such as satellite telemetry and control centers, low-Earth orbit satellite constellations, manned space stations, and deep space probes. It addresses industry-specific pain points such as time window constraints for aerospace telemetry and control missions, multi-satellite heterogeneous resource coupling, and large latency in emergency response migration for abnormal services, enabling intelligent and highly reliable cluster resource scheduling for on-orbit telemetry and control missions of tens of thousands of low-Earth orbit satellites, manned space stations, and deep space probes.

[0006] This invention provides a software-defined resource management and scheduling method based on a server cluster platform, comprising the following steps: S1. A resource model is obtained by performing full-dimensional modeling of the software and hardware resources of the server cluster. The resource model is based on relational database tables and forms a static resource requirement baseline by linking the hardware resource table, scheduling status table and software resource requirement table through foreign keys. S2. Construct a scheduling model and configure a time series database. The time series database automatically collects and stores time series data from the server cluster. The scheduling model performs multi-dimensional preprocessing on the time series data and non-time series static data from the resource model, and then extracts the first-level features, second-level features, and third-level features in sequence. By mining non-linear resource usage patterns and combining multi-dimensional prediction and comprehensive matching, dynamic resource demand is predicted. The spatiotemporal consistency loss function is used for training and optimization. S3. Build a container cloud management platform and deploy a scheduling model that achieves the optimization goal. The container cloud management platform performs manual and automatic dual scheduling of tasks to be scheduled, and supports automatic resource allocation, manual node specification, GPU core range and CPU core range. S4. Perform task resource requirement prediction, submit the tasks to be scheduled to the container cloud management platform, and the container cloud management platform, in combination with the start and end time, software type and quantity, data volume and task priority of the tasks to be scheduled, predicts the overall resource requirement and fluctuation curve of the tasks to be scheduled through resource model and scheduling model. S5. For the time range of the task to be scheduled, calculate the available resources and fluctuation curves of each node, match the available resources of the node with the requirements of the task to be scheduled, select the optimal node scheduling container based on the resource matching degree and the predicted value of task execution risk, and start the task to be scheduled. S6. Each node reports its health status periodically. After receiving abnormal node status data, the container cloud management platform triggers resource isolation operations.

[0007] The software-defined resource management and scheduling method based on a server cluster platform described in this invention, as a preferred method, involves the container cloud management platform directly querying the software resource requirement table to obtain the default resource requirements when performing static task scheduling, and then linearly adjusting the requirements based on the input data volume and scaling factor to obtain the final resource requirements. Static tasks include periodic data backup and fixed report generation. When scheduling dynamic load tasks or new tasks without a preset baseline, the container cloud management platform inputs the task attributes and the current running status of the server cluster into the scheduling model. The scheduling model outputs dynamic resource demand values ​​and fluctuation curves by learning the mapping relationship between the attributes of historical tasks and resource consumption. Dynamic load tasks include big data analysis and AI model training. Task attributes include software type, input data volume, runtime, and priority. When scheduling tasks in mixed task scenarios, the container cloud management platform uses the static resource requirement baseline as the lower limit and the dynamic resource requirement prediction value as the upper limit, and takes the weighted average of the static resource requirement baseline and the dynamic resource requirement prediction as the final resource requirement.

[0008] The software-defined resource management and scheduling method based on a server cluster platform described in this invention, as a preferred embodiment, in step S1, stores the static hardware attributes of nodes in the hardware resource table. The static hardware attributes include CPU architecture, GPU model, number of cores, memory size, disk capacity and IOPS, network bandwidth, network latency, and scheduling attributes such as online, health, isolation, and load level. The scheduling status table stores dynamic attributes related to resource scheduling, including node online status, health status, isolation status, load level, and last heartbeat time. The software resource requirements table stores software-level resource requirements information, including software type, default CPU / GPU / memory / network / disk requirements, and preset resource requirement scaling factors for different data volumes. In step S2, the time series database automatically stores time series data and supports querying, aggregation, downsampling, and correlation analysis of time series data for resource conflict judgment and model iterative training. The time-series data includes: multimodal heterogeneous resource usage data, node status data, and inter-node network communication data. The multimodal heterogeneous resource usage data includes historical data on CPU utilization, GPU utilization, GPU memory utilization, memory utilization, and disk utilization of each heterogeneous node. Non-time-series static data includes node hardware attributes, basic software information, and basic task attributes; Non-time-series static data and time-series data are concatenated and then input into the scheduling model; The time series database supports fast queries by time range, node ID, and task ID, and supports aggregation calculations on time series data, including calculating the average, maximum, and minimum values. The downsampling method involves converting high-frequency data into low-frequency data in a time-series database. Multi-dimensional preprocessing includes data cleaning, missing value imputation, outlier handling, data standardization, and data alignment; Data cleaning: Removing duplicate, invalid, and noisy data; Missing value imputation: Time series-based interpolation is used to imput continuous missing data, and mode imputation is used to imput discrete missing data; Outlier handling: The Isolation Forest algorithm is used to identify and remove outliers; Data standardization: The Z-score standardization method is used to map resource data of different dimensions to a standard normal distribution with a mean of 0 and a variance of 1; Data alignment: Aligning resource data with different sampling frequencies to a unified time granularity.

[0009] In the software-defined resource management and scheduling method based on a server cluster platform described in this invention, as a preferred embodiment, in step S2, the scheduling model is a multi-branch heterogeneous convolutional neural network. The feature extraction method for the scheduling model is as follows: A multi-branch heterogeneous convolutional neural network is constructed, with each branch extracting features from time-series data of a different basic resource type. Each branch uses at least two stacked one-dimensional convolutional layers with different kernel sizes (kernel sizes of 3, 7, and 15, stride of 1, and Swish activation function). Each convolutional layer is followed by a batch normalization layer and a max-pooling layer. Multi-dimensional preprocessed data is input, and each branch performs feature extraction. The max-pooling layer outputs the local features extracted by each branch. The local features are then concatenated to obtain the first-level features. Construct a cross-modal attention-gated recurrent unit (CM-AGRU); input the first-level features, and CM-AGRU calculates the interaction weights between different resource type features in the first-level features through a cross-modal attention mechanism to generate a cross-modal interaction feature matrix; Cross-modal interaction feature matrix The input bidirectional gated recurrent unit (Bi-GRU) captures the bidirectional temporal dependencies of resource usage, resulting in the second-level features. ; in, For local features of different resource types, Here is a learnable attention weight matrix, and N is the number of resource types; To construct the spatiotemporal graph attention network ST-GAT, each server node in the server cluster platform is first modeled as a vertex of the spatiotemporal graph, with network bandwidth and communication latency between nodes serving as edge weights to obtain the cluster topology graph. The second-level features are then input, and the spatial association weights between nodes are calculated through the spatial graph attention layer of ST-GAT to capture load propagation and resource competition relationships between different nodes, thus obtaining spatial features. Temporal convolutional layers are then used to perform temporal modeling on the features output by the spatial attention layer, capturing the long-term evolution trend of overall cluster resource usage to obtain temporal features. Finally, the spatial and temporal features are fused to obtain the third-level features. The spatiotemporal consistency loss function is the sum of temporal consistency loss and spatial consistency loss. Temporal consistency loss constrains the difference in feature vectors of the same node at adjacent time steps to make the features change smoothly over time. Spatial consistency loss constrains the difference in feature vectors of adjacent nodes in the server cluster at the same time step to reflect the load transmission and resource competition relationship between nodes.

[0010] In the software-defined resource management and scheduling method based on a server cluster platform described in this invention, as a preferred embodiment, in step S211, the multi-branch heterogeneous convolutional neural network is a CNN model, and the branches include CPU branch, GPU branch, memory branch, disk branch and network branch. The cross-modal attention-gated recurrent unit CM-AGRU is constructed using a multi-task learning framework, and resource demand prediction head, node matching degree prediction head, and task execution risk prediction head are set. The scheduling model was trained using the AdamW optimization algorithm; The loss function of the scheduling model is a weighted sum of resource demand prediction loss, node matching degree prediction loss, task execution risk prediction loss, and spatiotemporal consistency loss.

[0011] In the preferred embodiment of the software-defined resource management and scheduling method based on a server cluster platform described in this invention, step S5 calculates the available resources of each node by subtracting the occupied resources from the total resource capacity of the node. The optimal node n* is: ; in, To find the function that minimizes the value of the expression inside the parentheses; This is the availability status indicator for the nth server node, with a value of 0 or 1, where 1 indicates availability and 0 indicates unavailability. The overall resource matching degree of server node n; The optimal node is selected based on the principle of minimizing the proportion of available resources to total resources while minimizing task risk. ; in, The predicted risk of server node n executing tasks is given by the scheduling model, and the value range is [0,1]. λ is the balance coefficient between resource matching degree and task execution risk, which is a constant and the default value is 0.7. ; in, These are the weighting coefficients for CPU, GPU, memory, network, and disk resources, respectively. And adjust according to business scenarios; , , , , These are the CPU core requirements, GPU quantity requirements, memory size requirements, network bandwidth requirements, and disk IOPS requirements for the tasks to be scheduled. , , , , These represent the total number of CPU cores, total number of GPUs, total memory size, total network bandwidth, and total disk IOPS for the nth node.

[0012] , , , , .

[0013] In the software-defined resource management and scheduling method based on a server cluster platform described in this invention, as a preferred embodiment, in step S6, each node periodically reports its health status. The node's periodic self-detection cycle is 5-30 seconds, and the detection content includes CPU temperature, GPU memory usage, memory error rate, network packet loss rate, number of bad sectors on the disk, and container running status. After receiving the abnormal node status data, the container cloud management platform triggers resource isolation operations, prohibits new tasks from being scheduled to abnormal nodes, and completes the seamless migration of all running services on the node according to task priority. The method for determining the availability of a node is as follows: a node is considered available only if it is online, healthy, and not isolated. When applied to the integrated space-ground computing power scheduling scenario, if an anomaly occurs in a node within the satellite telemetry and control window, the resource management and scheduling system will prioritize ensuring the seamless migration of satellite tasks within the telemetry and control window to avoid missing the telemetry and control time window; ; in, This is the online status indicator for the nth node, where 1 indicates online and 0 indicates offline; This is the health status identifier for the nth node, where 1 indicates healthy and 0 indicates abnormal. This is the isolation status identifier for the nth node, where 1 indicates that it is isolated and 0 indicates that it is not isolated. This is a Boolean logical AND operator, which returns 1 only if all operands are 1; This is a Boolean NOT operator; the result is 0 when the operand is 1, and 1 when the operand is 0.

[0014] The software-defined resource management and scheduling method based on a server cluster platform described in this invention, as a preferred embodiment, involves a resource management and scheduling system performing software-defined resource management and scheduling steps S1 to S6. The resource management and scheduling system includes: a hardware and software resource modeling module, a scheduling model construction module, a container cloud management platform, a refined resource demand decomposition module, a storage module, a node available resource calculation module, a resource precise matching and scheduling module, and a node anomaly automatic handling module. The hardware and software resource modeling module generates resource models; The scheduling model building module constructs the scheduling model and completes model training, optimization, and deployment; The container cloud management platform manages pending tasks, performs both manual and automatic resource scheduling, monitors resources, issues alarms, and provides a visual interface. The refined resource demand decomposition module calculates the overall resource demand and resource demand fluctuation curve of the task to be scheduled; The storage module stores all resource usage information from historical tasks into the time-series database, providing data support for resource conflict judgment and model iteration; The node available resource calculation module calculates the actual available resources and available resource fluctuation curves of each server cluster node based on the time range of the task to be scheduled. The resource matching and scheduling module matches the available resources of nodes with the resource requirements of tasks, and selects the optimal node for container scheduling based on the task execution risk assessment results. The automatic node anomaly handling module receives health status data reported by nodes, performs resource isolation operations on abnormal nodes, and seamlessly migrates the services on the abnormal nodes to healthy nodes.

[0015] The software-defined resource management and scheduling method based on a server cluster platform described in this invention, as a preferred embodiment, includes a scheduling model construction module comprising: a data acquisition and preprocessing unit, a three-level feature extraction unit, a model training and optimization unit, and a model deployment and iteration unit; The data acquisition and preprocessing unit collects multimodal historical operational data from the server cluster and performs multi-dimensional preprocessing. The three-level feature extraction unit sequentially extracts the first-level features, second-level features, and third-level features of the overall resource requirements and fluctuation curves of the multimodal historical operation data or predicted tasks to be scheduled. The model training and optimization unit trains the scheduling model based on a multi-task learning framework and introduces a spatiotemporal consistency loss function to optimize the model parameters. The model deployment and iteration unit deploys the trained scheduling model to the container cloud management platform and periodically uses newly generated runtime data to incrementally update the model.

[0016] The software-defined resource management and scheduling method based on a server cluster platform described in this invention is preferably applied to resource management and scheduling scenarios of ground systems, or to integrated space-ground computing power scheduling scenarios where aerospace operation and control scheduling systems manage spacecraft on orbit. In the scenario of integrated space-ground computing power scheduling, the scheduling objects include any one or more of the following: satellite tracking and control centers, low-Earth orbit satellite constellations, manned space stations, and deep space probes; In step S1, the hardware resource table includes the number of baseband processing boards, the number of demodulation channels, radio frequency link resources and communication port numbers, and the software resource requirement table includes satellite telemetry demodulation software, payload data interpretation software, orbit calculation software, spacecraft fault analysis software and orbit type field. The orbit type field distinguishes between low Earth orbit, medium Earth orbit and deep space missions. In step S2, the time-series database supports quick querying by satellite number. The time-series data includes the time-series data of the telemetry and control hardware load and the high-frequency sampling data of the satellite transit window. The non-time-series static data includes the baseband board configuration, satellite number and orbit type. The scheduling model includes a telemetry and control hardware branch. It extracts features from the timing data of the baseband board and demodulation channel and splices them into the first-level features. The second-level features capture the coupling relationship between the surge in baseband load and the surge in CPU and network synchronization during the satellite's passage. When extracting the third-level features, the ground telemetry and control gateway and server cluster nodes are included in the spatiotemporal graph vertices to model the relationship between the satellite-to-ground link and the ground computing power nodes. The training dataset for the scheduling model includes historical low-Earth orbit constellations, manned space stations, and deep space probe tracking and control mission samples. In step S3, the container cloud management platform supports manual labeling of satellite missions at multiple levels of priority; the tasks to be scheduled include satellite historical telemetry and control data archiving tasks, low-orbit satellite constellation transit telemetry and control tasks, deep space probe telemetry data real-time interpretation tasks, and aerospace multi-satellite concurrent telemetry and control scenarios. The task attributes include the telemetry and control transit window duration and the satellite orbit type. In step S4, the resource management and scheduling system also predicts resource demand based on orbit type. The start and end times of the task to be scheduled are the transit start and end control windows, the number of software types is the amount of payload data, and the orbit types include low Earth orbit, medium Earth orbit, and deep space orbit. In step S5, when calculating the available resources of each node, the remaining demodulation channels and baseband hardware resources of the node are checked simultaneously for the time window of the satellite mission's transit.

[0017] The loss function is a weighted sum of the resource demand prediction loss, node matching degree prediction loss, task execution risk prediction loss, and spatiotemporal consistency loss, and the calculation formula is: ; in, This represents the mean squared error of resource demand forecasting. The mean squared error of the node matching degree prediction. The binary cross-entropy loss is used for task execution risk prediction. For spatiotemporal consistency loss, .

[0018] The model training of this invention uses the AdamW optimization algorithm, with a learning rate of 0.0005, a batch size of 64, and 200 training epochs. The default weights of the loss function are .

[0019] This invention employs a structured relational data model to construct a resource model. Specifically, the model uses relational database tables as its carrier, and achieves a unified description of resources through three interrelated tables: The first is a hardware resource base table, used to store the static hardware attributes of nodes, such as CPU architecture, GPU model, number of cores, memory size, disk capacity and IOPS, network bandwidth and latency, etc.; the second is a scheduling status table, used to store dynamic attributes related to resource scheduling, such as node online status, health status, isolation status, load level, last heartbeat time, etc.; the third is a software resource requirement table, used to store software-level resource requirement information, including software type, default CPU / GPU / memory / network / disk requirements, and preset resource requirement scaling factors for different data volumes. The three tables are linked by a foreign key through node IDs, forming a standardized resource description system. Regarding software-level resource requirement prediction, this invention adopts a dual-track method of "static baseline + dynamic prediction". For routine static tasks (such as periodic data backups and fixed report generation), the system directly queries the default resource requirement baseline from the software resource requirement table and then makes linear adjustments based on the actual input data volume and preset scaling factors (e.g., for every 10GB increase in data volume, the CPU requirement increases by 1 core). For dynamic load tasks (such as big data analysis and AI model training) or new tasks without preset baselines, the system inputs task attributes (software type, input data volume, runtime, priority) and the current cluster running status into a scheduling model based on three-level deep feature extraction. By learning the mapping relationship between historical task "attributes → resource consumption", the system outputs accurate dynamic resource requirement values ​​and fluctuation curves. In mixed task scenarios, the system uses the static baseline as the lower limit and the dynamic prediction value as the upper limit, taking the weighted average of the two as the final resource requirement, balancing stability and accuracy.

[0020] Regarding the role of the time-series database, in this invention, the time-series database (InfluxDB) is not merely used for storage, but undertakes several core functions. First, it is responsible for writing time-series data with timestamps, such as task resource usage curves, node resource load curves, and node status data, in real time. Second, it supports fast querying, allowing data retrieval by time range, node ID, task ID, etc., providing historical occupancy data for calculating available node resources. Third, it supports aggregation calculations, such as calculating averages, maximums, and minimums, providing preprocessed batch data for model training. Fourth, it supports data downsampling, converting high-frequency collected data into low-frequency data to reduce storage and computational overhead. Furthermore, the full resource usage information stored in the time-series database also serves two core purposes: resource conflict judgment and model iterative training. Therefore, in this invention, the time-series database is both a data storage center and a support platform for data querying and preprocessing. Regarding the relationship between historical data and time-series data, the document clearly distinguishes the composition of historical operational data: historical data is not entirely equal to time-series data, but time-series data is the main body of historical data. Over 90% of the historical data is time-series data, including multimodal heterogeneous resource usage data (CPU / GPU / memory / disk / network usage), node status data, and inter-node network communication data. This data is all timestamped to the second and generated continuously in chronological order, serving as the primary input to the three-level deep feature extraction model. In addition, the historical data also contains a small amount of non-time-series static data, such as node hardware attributes (CPU architecture, number of GPUs), basic software information (software type, version), and basic task attributes (task ID, software ID). This type of data does not change over time and is used as auxiliary features, concatenated with the time-series data, to be input into the model. Therefore, time-series data is a core component of historical data, but the scope of historical data is broader, also encompassing a small amount of static auxiliary feature data.

[0021] This invention includes the following steps: A dual-dimensional model of server cluster resources (hardware and software) is performed. The hardware dimension covers node brand, CPU / GPU architecture, number of cores, memory, disk, network bandwidth, and scheduling attributes such as online status, health status, isolation status, and load level. The software dimension configures and saves runtime resource requirements by type, and resource management and scheduling are carried out entirely based on this model. A three-level deep feature extraction architecture can uncover complex nonlinear resource usage patterns. A multi-task learning framework simultaneously outputs prediction results for three dimensions: resource demand, node matching degree, and task execution risk. Combined with a comprehensive matching degree calculation formula, the resource demand fluctuation curve is accurately aligned with the node's available resource fluctuation curve, reducing resource demand prediction errors in dynamic load scenarios and improving node resource utilization. It is also suitable for integrated space-ground computing power scheduling scenarios for on-orbit management of spacecraft such as satellite telemetry and control centers, low-Earth orbit satellite constellations, manned space stations, and deep space probes, addressing industry-specific pain points such as time window constraints for aerospace telemetry and control missions, coupling of heterogeneous resources across multiple satellites, and large latency in emergency response migration for abnormal services.

[0022] The scheduling model is deployed to a container cloud platform for resource demand prediction, matching degree calculation, and task risk assessment.

[0023] The present invention has the following advantages: (1) This invention fully covers hardware attributes and software resource requirements through a resource model; by designing independent feature extraction branches for different resource types such as CPU, GPU, and memory through a multi-branch heterogeneous convolutional neural network, it solves the problem of loss of heterogeneous resource differentiation patterns caused by the unified feature extraction of existing technologies, and effectively improves feature representation capabilities.

[0024] (2) This invention constructs a cross-modal attention-gated recurrent unit (CM-AGRU) to capture the interaction relationship and temporal dependency between resources, models the spatial correlation (load transmission, resource competition) of cluster topology through a spatiotemporal graph attention network (ST-GAT), and introduces a spatiotemporal consistency loss function to constrain model training, which completely solves the problem that traditional solutions only consider the temporal characteristics of a single node and have large global scheduling decision deviations, and effectively improves the scheduling accuracy of large-scale clusters.

[0025] (3) The three-level deep feature extraction architecture of this invention can mine complex nonlinear resource usage patterns. The multi-task learning framework simultaneously outputs prediction results in three dimensions: resource demand, node matching degree, and task execution risk. Combined with the comprehensive matching degree calculation formula, it can achieve accurate alignment between the resource demand fluctuation curve and the node available resource fluctuation curve, reduce the resource demand prediction error in dynamic load scenarios, and improve the node resource utilization rate.

[0026] (4) This invention implements a manual + automatic dual scheduling mode, supports fine-grained resource allocation (specifying CPU cores and GPU card numbers); high-frequency health detection of nodes, automatic isolation of abnormal nodes and seamless migration of services according to priority, ensuring data consistency and service continuity during the migration process, reducing service interruption time from hours to seconds, and reducing maintenance manpower costs.

[0027] (5) The task resources of this invention use full data to store in a time-series database, support regular incremental updates of the model, enable the scheduling system to adapt to changes in cluster load, hardware upgrades and business growth, maintain optimal scheduling performance in the long term, and avoid the problem of performance degradation of traditional static scheduling algorithms over time. Attached Figure Description

[0028] Figure 1 A flowchart illustrating a software-defined resource management and scheduling method based on a server cluster platform; Figure 2 This is a system architecture diagram for a software-defined resource management and scheduling method based on a server cluster platform. Detailed Implementation

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0030] Example 1 like Figure 1 As shown, a software-defined resource management and scheduling method based on a server cluster platform includes the following steps: S1. Unified resource modeling of both hardware and software resources: This invention performs full-dimensional system modeling of the hardware and software resources of a server cluster system. It adopts a structured relational unified resource model, which uses relational database tables as carriers and forms a standardized resource description system by linking hardware resource tables, scheduling status tables, and software resource requirement tables through foreign keys.

[0031] Hardware resource modeling: Constructing a basic hardware resource table and a scheduling status table. The basic hardware resource table stores static node attributes, including node brand, CPU architecture, GPU architecture, number of CPU cores, number of GPUs, GPU memory size, memory size, memory bandwidth, disk size, disk IOPS, network bandwidth, and network latency. For aerospace-ground integrated scheduling scenarios, the hardware resource table additionally labels hardware attributes specific to telemetry and control services: number of baseband processing boards, number of demodulation channels, RF link resources, and communication port numbers. The scheduling status table stores dynamic node attributes, including online status, health status, isolation status, load level, and last heartbeat time. The two tables are linked by node ID.

[0032] Software resource modeling: Construct a software resource requirements table to store the static resource requirement baselines for various software types, including software ID, software name, software type, default number of CPU cores, default number of GPUs, default memory size, default network bandwidth, default disk IOPS, and resource requirement scaling factors corresponding to different input data volumes (e.g., for every 10GB increase in data volume, the CPU requirement increases by 1 core). For satellite telemetry and control services, add telemetry and control software entries to the software resource requirements table: satellite telemetry demodulation software, payload data interpretation software, orbit calculation software, and spacecraft fault analysis software; also add an orbit type field to distinguish between low Earth orbit, medium Earth orbit, and deep space missions; Resource demand forecasting mechanism: This invention employs a dual-track resource demand forecasting method combining a static baseline and dynamic prediction. The resource model itself provides a static resource demand baseline, while the dynamic resource demand is achieved through a scheduling model based on three-level deep feature extraction. For routine static tasks (such as periodic data backup, fixed report generation, and archiving of historical satellite tracking and control data): directly query the software resource requirements table to obtain the default resource requirements, and then make linear adjustments based on the input data volume and scaling factor to obtain the final resource requirements.

[0033] For dynamic workload tasks (such as big data analysis, AI model training, low-Earth orbit satellite constellation transit telemetry and control, and real-time interpretation of deep space probe telemetry data) or new tasks without a preset baseline: input the task attributes (software type, input data volume, runtime, priority, transit window duration, satellite orbit type) and the current operating status of the cluster into the scheduling model. The model learns the "attribute-resource consumption" mapping relationship of historical tasks and outputs accurate dynamic resource demand values ​​and fluctuation curves.

[0034] Hybrid mission scenarios: Using a static baseline as the lower limit and dynamic predictions as the upper limit, the weighted average of the two is taken as the final resource requirement, balancing stability and accuracy. In multi-satellite concurrent telemetry and control scenarios, when the telemetry and control windows of multiple satellites overlap, the resource requirements of multiple missions are automatically superimposed, and the risk of computing power conflicts is predicted.

[0035] S2 Scheduling Model Construction and Deployment The core function of the time-series database InfluxDB is to store timestamped time-series data, but it also supports querying, aggregation, downsampling, and correlation analysis of time-series data, and is not limited to storage. Its specific functions include: Real-time writing of time-series data such as task resource usage curves, node resource load curves, and node status data; It supports quick queries by time range, node ID, task ID, and satellite number, providing historical occupancy data for calculating available node resources; It supports aggregation calculations on time-series data (such as calculating the average, maximum, and minimum values) to provide preprocessed batch data for model training. It supports data downsampling, converting high-frequency data into low-frequency data to reduce storage usage and computation.

[0036] Composition of historical data: Of the historical operational data collected, over 90% is time-series data, but it also includes a small amount of non-time-series static data, which are categorized as follows: Time-series data (core training data): This includes multimodal heterogeneous resource usage data, node status data, inter-node network communication data, and telemetry and control hardware load time series. All data are timestamped to the second and generated continuously in chronological order, serving as the primary input for level three deep feature extraction. In aerospace scenarios, high-frequency sampling data during satellite transit windows is prioritized (sampling period shortened to 10 seconds) to capture instantaneous computing power peaks during telemetry and control missions.

[0037] Non-time-series static data includes node hardware attributes (such as CPU architecture, number of GPUs, baseband card configuration), basic software information (such as software type and version), and basic task attributes (such as task ID, software ID, satellite number, and orbit type). This type of data does not change over time and is used as auxiliary features to be concatenated with time-series data and input into the model.

[0038] Data Acquisition and Preprocessing: Collect operational data from the cluster over the past 3 months, with a sampling frequency of 1 minute. During the telemetry and control transit period, the sampling frequency is switched to 10 seconds, resulting in a total of 129,600 data entries. Perform the following preprocessing operations: Data cleaning: 1200 duplicate records and 320 invalid records were removed; Missing value imputation: Linear interpolation is used to imput continuous missing values ​​such as CPU / GPU utilization, and mode is used to imput discrete missing values ​​such as node online status and demodulation channel occupancy status. Outlier handling: The Isolation Forest algorithm identified and removed 350 outliers; Data standardization: Z-score standardization maps all resource data to a distribution with a mean of 0 and a variance of 1; Data alignment: Unify all data to a 1-minute time granularity, and retain 10-second fine-grained timing during satellite telemetry and control windows; Level 3 feature extraction: Level 1: Multi-branch heterogeneous CNN. Each branch uses 3 stacked convolutional kernels (3, 7, 15) with a stride of 1, the Swish activation function, followed by a BN layer and a max pooling layer. In addition to the CPU, GPU, memory, disk, and network branches, a new measurement and control hardware branch is added to extract features from the time-series data of the baseband board and demodulation channel. The outputs of all branches are concatenated to obtain the first-level features. Level 2: CM-AGRU calculates the cross-modal interaction weights between CPU / GPU / memory / disk / network and telemetry and control hardware resources, inputs them into Bi-GRU to capture bidirectional temporal dependencies, and obtains Level 2 features; adapts to aerospace scenarios, and captures the coupling relationship between the surge in baseband load and the synchronous surge in CPU and network load during satellite transit.

[0039] Level 3: Construct a cluster topology graph (edge ​​weights are the network bandwidth between nodes), ST-GAT's spatial graph attention layer calculates the spatial correlation of nodes, and the temporal convolutional layer captures long-term trends, and the fusion is used to obtain the third-level features; In the space-ground integrated scenario, the ground telemetry and control gateway and server cluster nodes are included in the spatiotemporal graph vertices to model the correlation between the space-ground link and the ground computing power nodes.

[0040] Model Training: A multi-task learning framework is employed, including three output heads: resource demand prediction, node matching degree prediction, and task execution risk prediction. Training parameters: AdamW optimizer, learning rate 0.0005, batch size 64, training epochs 200; loss function weights ω1=0.4, ω2=0.3, ω3=0.2, ω4=0.1. The training dataset includes historical low-Earth orbit constellations, manned space stations, and deep space probe tracking and control mission samples. After training, the model is deployed to a container cloud platform built on Kubernetes.

[0041] The loss function is a weighted sum of resource demand prediction loss, node matching degree prediction loss, task execution risk prediction loss, and spatiotemporal consistency loss, where the spatiotemporal consistency loss is... Used to constrain the continuity of features learned by the model in the time and space dimensions, and to avoid prediction errors caused by abrupt feature changes, its specific structure and physical meaning are as follows: ; in: Temporal consistency loss To constrain the difference in feature vectors of the same node at adjacent time steps and ensure smooth feature changes over time, the calculation formula is as follows: ; : The time step of the training data, the default setting is 1440 (corresponding to 1 minute granularity data for 1 day). The total number of nodes in the cluster; : Spatiotemporal correlation global feature vector of node i at time t (dimension 2048). : The spatiotemporal correlation global feature vector of node i at time t-1; It is the square of the L2 norm, used to measure the Euclidean distance between two vectors.

[0042] Physical meaning: If the resource load of the same node does not change abruptly in adjacent minutes, the feature vectors extracted by the model should also remain similar to avoid the model being overly sensitive to noisy data; for satellite telemetry and control services, reasonable feature jumps are allowed at the start of satellite transit to distinguish between real service computing power spikes and fault noise.

[0043] Spatial consistency loss The constraint determines the difference in feature vectors between adjacent nodes in the cluster at the same time step, reflecting the load propagation and resource competition relationships between nodes. The calculation formula is as follows: ; The spatial association weight between node i and node j, with a value range of [0,1], is calculated using the following formula: ; in, The network bandwidth between node i and node j. For inter-node communication delay, and These represent the maximum network bandwidth and maximum communication latency within the cluster, respectively. In integrated space-ground scheduling scenarios, a higher association weight is set between the gateway and the computing nodes.

[0044] Physical meaning: The higher the network bandwidth and the lower the latency, the stronger the spatial correlation between two nodes, and the more similar their feature vectors should be; if a node is overloaded, the load of its neighboring nodes will also be affected, and their feature vectors will show similar changing trends.

[0045] The role of spatiotemporal consistency loss: by introducing The model can not only learn the resource usage patterns of individual nodes, but also the spatiotemporal evolution patterns of the entire cluster, significantly improving the prediction accuracy of the model in abnormal scenarios such as node failure, load surge, and satellite overpass computing power burst, and reducing scheduling decision errors.

[0046] S3, Container Cloud Management Platform Setup: It integrates the native Kubernetes scheduler and extends custom scheduling plugins to implement the scheduling logic of this invention; it provides a web-based visual interface that supports management of tasks to be scheduled, manual / automatic scheduling switching, resource monitoring and anomaly alarms, and allows manual specification of GPU card number and CPU core range.

[0047] S4. Task Resource Requirement Forecasting: Submit a satellite telemetry, telemetry and demodulation task with the following parameters: start and end time, real-time telemetry demodulation + payload data processing, and a real-time payload data volume of 500GB; task priority = high.

[0048] The following is calculated using the preset resource model and the scheduling model: Basic resource requirements: 8-core CPU, 1-card GPU (A100), 32GB RAM, 150GB disk space Resource demand fluctuation curve: At the beginning of the telemetry and control window, the instantaneous computing power of telemetry demodulation surges, with a peak CPU core of 12; during the data loading and interpretation phase, the peak memory size is 48GB; the time granularity is 10 seconds, precisely matching the satellite transit window.

[0049] Resource usage information storage configuration: Configure the InfluxDB time series database to automatically collect all data per second during task execution, including CPU / GPU usage, memory usage, disk I / O, network traffic, baseband board load, demodulation channel occupancy status, etc., and retain all measurement and control business data for 6 months.

[0050] S5. Calculation of available node resources: For the mission transit time window, calculate the available resource fluctuation curve for each node, and simultaneously verify the remaining demodulation channels and baseband hardware resources of the nodes: Node 5: Total resources: 64 CPU cores, 2 GPUs, 256GB RAM, 2TB disk, 25Gbps network; 4 available demodulation channels; Currently occupied: 16 CPU cores, 0 GPUs, 64GB RAM, 500GB disk, 5Gbps network; 2 currently occupied demodulation channels; Available resources: 48 CPU cores, 2 GPUs, 192GB RAM, 1.5TB disk, 20Gbps network; 2 available demodulation channels; Stable fluctuations. Node 6: GPU1 card is available, but all demodulation channels are occupied by other tasks, and the measurement and control hardware resources do not meet the task requirements; Nodes 7 and 8: Each node has 1 available GPU, with slightly lower CPU / memory resources than node 5; each node has 1 available demodulation channel, which is insufficient to meet the 2-channel demodulation requirement of this task. Resource matching and container scheduling: Calculate the overall matching degree of all valid nodes (online, healthy, not isolated, and with sufficient monitoring and control resources): Node 5: Resource matching degree M(5) = 0.2 × (8 / 64) + 0.4 × (1 / 2) + 0.2 × (32 / 192) + 0.1 × (2 / 20) + 0.1 × (150 / 1500) = 0.278; Execution risk R(5) = 0.05; Overall matching degree M_total(5) = 0.7 × 0.278 + 0.3 × 0.05 = 0.2096.

[0051] Node 6 is filtered directly due to insufficient measurement and control hardware resources; nodes 7 and 8 are filtered due to insufficient demodulation channels.

[0052] Select node 5, which has the lowest overall matching degree, as the scheduling node. Create a Docker container on node 5, allocate 8 CPU cores, GPU card 0, 32GB of memory, 150GB of disk, and a network bandwidth limit of 3Gbps. Bind 2 idle demodulation channels. The task will be ready to start before the satellite enters the telemetry and control window.

[0053] S6, Node Health Monitoring: The node reports its health status every 10 seconds. After this scheduling, the CPU temperature, GPU memory usage, and memory error rate of node 5 are all within the normal range, and there are no abnormal alarms. If a node experiences an anomaly within the satellite tracking and control window, the system will prioritize ensuring the seamless migration of the high-priority satellite mission to avoid missing the tracking and control time window.

[0054] like Figure 2 As shown, the resource management and scheduling system includes: a hardware and software resource modeling module, a scheduling model construction module, a container cloud management platform, a refined resource demand decomposition module, a storage module, a node available resource calculation module, a resource precise matching and scheduling module, and a node anomaly automatic handling module.

[0055] Example 2 like Figure 1 As shown, a software-defined resource management and scheduling method based on a server cluster platform is presented. This embodiment is based on the cluster environment and the deployed scheduling system of Embodiment 1, and focuses on verifying the automatic handling of abnormal nodes and the adaptive iteration capability of the model.

[0056] 1. Node anomaly detection and automatic isolation Node 5 periodically reports its health status and has detected the following abnormal indicators: GPU0 card's video memory usage remained above 98% for 15 seconds; The memory error rate reached 0.05% (preset threshold 0.01%). Disk I / O latency exceeds 200ms (preset threshold 50ms); The container cloud platform immediately triggers resource isolation operations: Mark node 5 as "isolated", update the cluster resource pool, and prevent new tasks from being scheduled to this node; Send voice alerts to operations and maintenance personnel, including abnormal indicators, the time of occurrence, and the actions already taken.

[0057] 2. Seamless migration of services based on priority The platform retrieves all tasks running on node 5 and sorts them by priority from highest to lowest: High priority: Measurement and control tasks (tasks scheduled in Example 1); Medium priority: Background data preprocessing tasks; Low priority: Log collection task; Execute the migration process: High-priority task migration: Pause task execution and synchronize intermediate task data (12GB in total) to node 6 through the distributed storage system (verify in advance that the available resources of node 6 meet the requirements); create a container on node 6 with the exact same configuration as the original container and resume task execution; the migration process takes 12 seconds, the data consistency check passes, the training task is uninterrupted, and there is no loss of accuracy; Medium-priority task migration: The data preprocessing task was migrated to node 7, taking 28 seconds; Low-priority task migration: Migrate the log collection task to node 8, taking 5 seconds; The migration of all services was completed in 45 seconds, and node 5 entered the maintenance pending state.

[0058] 3. Recovery of abnormal nodes Remote troubleshooting by operations and maintenance personnel revealed a hardware failure in GPU 0 of node 5. After replacing the GPU, a manual node health check was performed, and all indicators returned to normal. Node 5's status was then changed to "healthy," and it was re-added to the cluster resource pool, restoring its scheduling capabilities.

[0059] 4. Incremental Iterative Update of the Model Every Monday at 2:00 AM, the platform automatically executes the model iteration process: Extract 100,800 records of runtime data from the past 7 days from the InfluxDB time series database; Perform the same multi-dimensional preprocessing operations as in Example 1; Incremental training: Freeze the parameters of the first two feature extraction layers, and only update the parameters of the third feature extraction layer and the three output heads; Training parameters: AdamW optimizer, learning rate 0.0001, batch size 32, training epochs 20; Performance validation: Compare the metrics of the new model with those of the original model: Mean squared error of resource demand forecast: 0.082 → 0.075; Node matching degree prediction mean square error: 0.065→0.058; Accuracy of task execution risk prediction: 92.3% → 94.1%; Gray-scale deployment: Deploy the new model to the test environment, run it for 24 hours without any abnormalities, and then replace the original model in the production environment.

[0060] 5. Iteration effect verification Submit the same measurement and control task as in Example 1: The new model improved the fit between the resource demand fluctuation curve calculated and the actual usage curve by 8.7%. The overall matching degree of the selected scheduling node is 0.062 lower than that of the original model; The total task execution time was reduced from 12 hours to 11 hours and 22 minutes, a reduction of 5.2%. The peak node resource utilization rate reached 72%, which is 6 percentage points higher than the original model. This embodiment verifies the model's adaptive iteration capability and long-term scheduling performance advantages of the present invention.

[0061] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A software-defined resource management and scheduling method based on a server cluster platform, characterized in that: Includes the following steps: S1. A resource model is obtained by performing full-dimensional modeling of the software and hardware resources of the server cluster. The resource model is based on a relational database table and forms a static resource requirement baseline by linking the hardware resource table, scheduling status table and software resource requirement table through foreign keys. S2. Construct a scheduling model and configure a time series database. The time series database automatically collects and stores time series data from the server cluster. The scheduling model preprocesses the time series data and the non-time series static data from the resource model and extracts the first-level features, the second-level features and the third-level features in sequence. It predicts dynamic resource demand by mining nonlinear resource usage patterns and combining multi-dimensional prediction and comprehensive matching. It uses the spatiotemporal consistency loss function for training and optimization. S3. Build a container cloud management platform and deploy the scheduling model. The container cloud management platform performs both manual and automatic scheduling of the tasks to be scheduled. S4. Submit the task to be scheduled to the container cloud management platform. The container cloud management platform combines the start and end time, software type and quantity, data volume and task priority of the task to be scheduled, and predicts the overall resource demand and fluctuation curve of the task to be scheduled through the resource model and the scheduling model. S5. For the time range of the task to be scheduled, calculate the available resources and fluctuation curves of each node, match the available resources of the nodes with the requirements of the task to be scheduled, select the optimal node scheduling container based on the resource matching degree and the predicted value of task execution risk, and start the task to be scheduled. S6. Each node reports its health status periodically. After receiving abnormal node status data, the container cloud management platform triggers resource isolation operations.

2. The software-defined resource management and scheduling method based on a server cluster platform according to claim 1, characterized in that: When the container cloud management platform performs static task scheduling, it directly queries the software resource requirement table to obtain the default resource requirements and makes linear adjustments based on the input data volume and scaling factor to obtain the final resource requirements. The static tasks include periodic data backup and fixed report generation.

3. The software-defined resource management and scheduling method based on a server cluster platform according to claim 1, characterized in that: The hardware resource table stores the static hardware attributes of the nodes, including CPU architecture, GPU model, number of cores, memory size, disk capacity and IOPS, network bandwidth and network latency; The scheduling status table stores dynamic attributes related to resource scheduling, including node online status, health status, isolation status, load level, and last heartbeat time. The software resource requirement table stores software-dimensional resource requirement information, which includes software type, default CPU / GPU / memory / network / disk requirements, and preset resource requirement scaling factors for different data volumes. The time-series data includes: multimodal heterogeneous resource usage data, node status data, and inter-node network communication data. The multimodal heterogeneous resource usage data includes historical data on the CPU utilization, GPU utilization, GPU memory utilization, memory utilization, and disk utilization of each node.

4. The software-defined resource management and scheduling method based on a server cluster platform according to claim 1, characterized in that: In step S2, the scheduling model is a multi-branch heterogeneous convolutional neural network; The feature extraction method of the scheduling model is as follows: each branch in the multi-branch heterogeneous convolutional neural network extracts features from time-series data of a basic resource type. Each branch uses at least two stacked one-dimensional convolutional layers with different kernel sizes. After each convolutional layer, a batch normalization layer and max pooling are performed, and the local features extracted by each branch are output. The local features are concatenated to obtain the first-level features. The cross-modal attention-gated recurrent unit CM-AGRU calculates the interaction weights between different resource type features in the first-level features, generates a cross-modal interaction feature matrix, and inputs it into the bidirectional gated recurrent unit Bi-GRU to capture the bidirectional temporal dependency of resource usage, thus obtaining the second-level features; The Spatiotemporal Graph Attention Network (ST-GAT) models each server node in the server cluster platform as a vertex of a spatiotemporal graph, using network bandwidth and communication latency between nodes as edge weights to obtain the cluster topology graph. The second-level features are input, and the spatial association weights between nodes are calculated through the ST-GAT spatial graph attention layer to capture load propagation and resource competition relationships between different nodes, thus obtaining spatial features. Temporal convolutional layers are used to perform temporal modeling on the features output by spatial attention, capturing the long-term evolution trend of overall cluster resource usage, thus obtaining temporal features. The spatial features and the temporal features are then fused to obtain the third-level features. The spatiotemporal consistency loss function is the sum of temporal consistency loss and spatial consistency loss. The temporal consistency loss constrains the difference in feature vectors of the same node at adjacent time steps to make the features change smoothly over time. The spatial consistency loss constrains the difference in feature vectors of adjacent nodes in the server cluster at the same time step to reflect the load transmission and resource competition relationship between nodes.

5. The software-defined resource management and scheduling method based on a server cluster platform according to claim 4, characterized in that: The multi-branch heterogeneous convolutional neural network is a CNN model, including CPU branch, GPU branch, memory branch, disk branch and network branch; The cross-modal attention-gated recurrent unit CM-AGRU is constructed using a multi-task learning framework, and resource requirement prediction head, node matching degree prediction head, and task execution risk prediction head are set.

6. The software-defined resource management and scheduling method based on a server cluster platform according to claim 1, characterized in that: In step S5, the optimal node n* is: ； in, To find the function that minimizes the value of the expression inside the parentheses; This is the availability status indicator for the nth server node, with a value of 0 or 1, where 1 indicates availability and 0 indicates unavailability. This represents the overall resource matching degree of server node n.

7. The software-defined resource management and scheduling method based on a server cluster platform according to claim 6, characterized in that: ； in, The predicted risk of server node n executing tasks is given by the scheduling model, with a value range of [0,1]. λ is the balance coefficient between resource matching degree and task execution risk, which is a constant. ； in, These are the weighting coefficients for CPU, GPU, memory, network, and disk resources, respectively. And adjust according to business scenarios; , , , , These are the CPU core requirements, GPU quantity requirements, memory size requirements, network bandwidth requirements, and disk IOPS requirements for the tasks to be scheduled. , , , , These represent the total number of CPU cores, total number of GPUs, total memory size, total network bandwidth, and total disk IOPS for the nth node, respectively. , , , , .

8. The software-defined resource management and scheduling method based on a server cluster platform according to claim 1, characterized in that: In step S6, the node autonomously detects the following: CPU temperature, GPU memory usage, memory error rate, network packet loss rate, number of bad sectors on disk, and container running status. When the container cloud management platform performs resource isolation operations, it prohibits new tasks from being scheduled to abnormal nodes and completes the seamless migration of all running services on the node according to task priority. A node is considered usable only if it is online, healthy, and not isolated.

9. The software-defined resource management and scheduling method based on a server cluster platform according to claim 1, characterized in that: The resource management and scheduling system performs resource management and scheduling in steps S1 to S6. The resource management and scheduling system includes: a hardware and software resource modeling module, a scheduling model construction module, the container cloud management platform, a refined resource demand decomposition module, a storage module, a node available resource calculation module, a resource precise matching and scheduling module, and a node anomaly automatic handling module. The scheduling model construction module includes a data acquisition and preprocessing unit, a three-level feature extraction unit, a model training and optimization unit, and a model deployment and iteration unit.

10. A software-defined resource management and scheduling method based on a server cluster platform according to any one of claims 1 to 9, characterized in that: The software-defined resource management and scheduling method is applied to resource management and scheduling scenarios of ground systems, or to integrated space-ground computing power scheduling scenarios where aerospace operation and control scheduling systems manage spacecraft on-orbit. In the scenario of integrated space-ground computing power scheduling, the scheduling objects include any one or more of the following: satellite tracking and control centers, low-Earth orbit satellite constellations, manned space stations, and deep space probes; In step S1, the hardware resource table includes the number of baseband processing boards, the number of demodulation channels, and radio frequency link resources. The software resource requirement table includes satellite number, satellite telemetry demodulation software, spacecraft fault analysis software, and orbit type field. The orbit type field distinguishes between low Earth orbit, medium Earth orbit, and deep space missions. In step S2, the time-series database supports fast querying by satellite number, and the time-series data includes time-series data of telemetry and control hardware load and high-frequency sampling data of satellite transit windows; The scheduling model includes a telemetry and control hardware branch, which extracts features from the baseband board and demodulation channel timing data and splices them into the first-level features. The second-level features capture the coupling relationship between the surge in baseband load and the synchronous surge in CPU and network at the moment of satellite passage. When extracting the third-level features, the ground telemetry and control gateway and server cluster nodes are included in the spatiotemporal graph vertices to model the relationship between the satellite-to-ground link and the ground computing power nodes. The training dataset of the scheduling model includes historical low-Earth orbit constellations, manned space stations, and deep space probe telemetry and control mission samples. In step S4, the container cloud management platform predicts resource demand based on the track type, and the start and end times of the task to be scheduled are the transit start and end monitoring and control windows.

Images