A method and system for joint sensing and transferability profiling of computing centers

By deploying a combined computer and computing sensor in the computing center, dual-source data acquisition and preprocessing of heterogeneous systems are achieved, generating standardized transferability profiles. Combined with the M/M/c queuing theory model and ensemble learning model, the problem of incomplete data perception in the computing center is solved, improving the accuracy of power consumption prediction and the flexibility of system scheduling.

CN123114754APending Publication Date: 2026-11-17CHINA ELECTRIC POWER RESEARCH INSTITUTE CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611306864.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing computing center scheduling systems suffer from incomplete data perception, lack of multi-dimensional quantification of the portability of computing tasks, and inaccurate power consumption prediction due to ignoring the equivalent load of queuing during system congestion.

Method used

By deploying a combined computer-computer sensor on computing nodes, dual-source data acquisition and preprocessing of heterogeneous systems are achieved, generating standardized transferability profiles. Combined with the M/M/c queuing theory model and ensemble learning model, load forecasting is corrected, improving the accuracy of power consumption forecasting.

Benefits of technology

It achieves comprehensive data awareness of heterogeneous systems, provides a multi-dimensional quantitative basis for task portability, corrects insufficient load forecasting during system congestion, and improves the accuracy of power consumption forecasting and the flexibility of system scheduling.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114754A_ABST
    Figure CN123114754A_ABST
Patent Text Reader

Abstract

This invention relates to the field of resource scheduling and discloses a method and system for joint sensing and portability profiling of computing power centers. It performs dual-source data acquisition from heterogeneous systems, writes the converted standard-format data into a time-series database, extracts multi-dimensional demand attributes of computing tasks, generates standardized representation vectors, calculates a comprehensive portability index, and outputs a standardized portability profile snapshot. By deploying joint sensing devices on computing nodes and simultaneously enabling platform API call channels and local monitoring channels, comprehensive acquisition of multi-source operational data from heterogeneous systems is achieved. Simultaneously, the cleaned and standardized time-series data is written into a time-series database with an independently established physical isolation mechanism, effectively eliminating the risk of data read / write conflicts in high-concurrency acquisition scenarios and ensuring the stable operation of the underlying sensing system during the data collection phase. This invention can effectively quantify task migration flexibility and improve the accuracy of power consumption prediction for computing nodes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, specifically to a method and system for combined computer-based sensing and portability profiling for computing centers. Background Technology

[0002] With the rapid development of computing networks and the construction of large-scale computing centers, the computing tasks carried by data centers are showing a trend of massiveization and heterogeneity. In the daily operation of computing centers, how to efficiently coordinate and schedule computing resources and achieve precise energy consumption management has become crucial to ensuring the stable operation and green energy conservation of computing infrastructure. As an important means to improve resource utilization and reduce overall operating costs, computing-power collaborative scheduling places increasingly higher demands on the perception accuracy of underlying operational data, the quantitative assessment of task characteristics, and the accuracy of power consumption prediction.

[0003] In existing computing cluster scheduling and energy management processes, the scheduling system typically obtains task status and resource allocation information by calling the upper-layer interface of the heterogeneous computing cluster management platform, using this as the basic data for business scheduling. Regarding task allocation, the system mostly executes scheduling actions based on real-time available resource capacity. However, in the power consumption prediction stage for computing nodes, traditional load modeling methods mainly rely on extracting intuitive operational indicators such as the hardware resource utilization rate of tasks currently running on the computing node, and directly importing these indicators as input features into a conventional prediction model to output the node's expected power consumption value.

[0004] However, the aforementioned conventional methods have several limitations in practical applications. First, relying solely on platform-level interfaces to obtain data lacks deep integration with the underlying physical node hardware operation, resulting in incomplete perception data in complex environments. Second, existing solutions often treat computing tasks as rigid loads, failing to objectively quantify the portability flexibility of tasks from dimensions such as time latency tolerance, spatial transmission costs, and interrupt recovery, leading to a lack of effective elastic reference for cross-domain scheduling. Furthermore, traditional power consumption prediction models directly use the utilization rate of currently running resources as a feature, failing to incorporate the underlying system scheduling overhead and resource reservation consumption caused by task queuing. When the system is in a high-concurrency congestion state, this neglect of the equivalent queuing load will bias the input data, causing the model to systematically underestimate actual power consumption during peak business periods, failing to meet the actual needs of accurate energy consumption management. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for joint sensing and portability profiling of computing centers, which solves the problems of incomplete data sensing of heterogeneous systems, lack of multi-dimensional quantification of the portability of computing tasks, and inaccurate prediction of node power consumption due to ignoring the equivalent load of queuing when the system is congested.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for joint sensing and transferable profiling of computing centers, comprising the following steps: Perform dual-source data acquisition and preprocessing for heterogeneous systems: acquire platform-level running data and node-level time-series data through parallel platform API call channels and local monitoring channels respectively; clean the acquired platform-level running data and node-level time-series data, convert them into a unified standard format, and write them into the time-series database; establish physical isolation between the time-series database and the computing power node local monitoring program database. Multi-dimensional requirement attributes of computing power tasks are extracted from the time-series database, and standardized representation vectors of the computing power tasks are generated. Based on the standardized representation vectors, time portability, spatial portability, and interruption recoverability are calculated and weighted to obtain a comprehensive portability index. Based on the comprehensive portability index, flexibility categories are defined, and a standardized portability profile snapshot of the computing power tasks containing flexibility category labels is output.

[0007] Preferably, the dual-source data acquisition and preprocessing of the heterogeneous system specifically includes: Deploy the computer-computer joint sensor in a containerized manner on the computing power node and perform hardware baseline acquisition to extract the static configuration information of the computing power node; Start a periodic data collection cycle and simultaneously enable the platform API call channel and the local monitoring channel; The platform-level operation data is obtained by calling the heterogeneous computing power cluster management platform interface through the multi-source API unified adaptation layer, and the computer-connected sensor is granted read-only monitoring permission through the role-based access control (RBAC) model. Node-level time-series data is obtained by reading the underlying operating status of the device through the local monitoring channel; The heterogeneous data acquired in parallel from two sources undergoes outlier removal, missing value filling, timestamp alignment, and unit standardization, and is then converted into a unified standard schema structured format and written into the time series database.

[0008] Preferably, the multi-dimensional requirement attributes consist of task type, service level agreement constraint set, computing resource requirement set, network transmission requirement set, and task migration attribute set; The set of service level agreement constraints includes deadline, maximum delay duration, checkpoint interval, and data localization requirements. The set of computing resource requirements includes the number of graphics processing units (GPUs), the number of CPU cores, the memory capacity, and the network bandwidth. The set of network transmission requirements includes end-to-end latency requirements, bandwidth requirements, and network jitter tolerance. The set of task migration attributes includes the amount of data migrated across nodes, task recovery time, and recalculation cost.

[0009] Preferably, the step of calculating temporal portability, spatial portability, and interruption recoverability based on standardized representation vectors and then weighting them to obtain a comprehensive portability index specifically includes: The time portability is calculated based on the ratio of the maximum delay duration to the expected execution duration; The spatial portability is calculated based on the ratio of cross-node migration data volume to bandwidth requirement, adjusted by a network cost coefficient. The interruption recoverability is calculated based on an indicator function that checks whether the checkpoint interval is greater than zero and the proportion of recalculation cost. The overall portability index, calculated by weighting the temporal portability, spatial portability, and interruptibility recoverability, satisfies the following: ; In the formula, It is a comprehensive portability index; This is the time weighting coefficient; For time portability; Spatial weight coefficient; For spatial portability; This is the interruption weighting coefficient; For interruption recoverability; weighting coefficient , , The sum of the values ​​must be 1.

[0010] Preferably, the step of classifying flexibility categories based on the comprehensive transferability index and outputting a standardized transferability profile snapshot specifically includes: The comprehensive portability index value is limited to between 0 and 1, and a preset value range is defined. The flexibility category is automatically determined based on the specific value range in which the comprehensive portability index falls. The flexibility category includes rigidity, time elasticity, spatial elasticity, and full elasticity. The task's runtime context is saved by serialization, generating a lightweight, standardized, and portable profile snapshot. The standardized and portable profile snapshot includes a comprehensive portability index, flexibility category, task scheduling priority, multidimensional migration cost, and power consumption sensitivity. The standardized multidimensional portable cost profile, which includes a flexibility category label, is then encapsulated and output.

[0011] Preferably, after outputting the standardized and transferable profile snapshot, the method further includes the following steps: extracting task arrival rate, task service rate, and number of parallel service nodes from the time-series database; establishing an M / M / c queuing theory model for multi-node parallel scheduling scenarios of computing power tasks; calculating the system idle probability and average task queuing delay; multiplying the task arrival rate, the average task queuing delay, and the preset average equivalent resource occupancy coefficient of a single queued task to obtain the queuing equivalent load; and adding the queuing equivalent load to the current running task resource occupancy to generate a corrected comprehensive effective load sequence.

[0012] Preferably, the step of establishing the M / M / c queuing theory model and calculating the system idle probability and average task queuing delay specifically includes: Calculate the current workload of the computing system:

[0013] In the formula, For business intensity; For task completion rate; This refers to the number of parallel service nodes. Service rate for tasks; Determine whether the business intensity ρ is strictly less than 1 to confirm that the system meets the conditions for stable operation; After confirming that the system meets the stability conditions, the system idle probability is calculated, and the average queuing delay of the task is calculated based on the system idle probability.

[0014] Preferably, the queuing equivalent load satisfies: ; In the formula, For queuing equivalent load; For task completion rate; Average queuing time for tasks; This represents the average equivalent resource utilization coefficient for a single queued task. The overall effective load satisfies: ; In the formula, To comprehensively assess effective load; Index of the set of tasks that are currently running; For the first Resource utilization coefficient of each running task; This represents the sum of resource usage for all currently running tasks.

[0015] Preferably, after generating the corrected comprehensive effective load sequence, the method further includes the following steps: extracting the comprehensive effective load sequence and synchronously reading the hardware runtime timing features in the time series database, concatenating and combining the two to generate a multi-scale feature matrix, inputting the multi-scale feature matrix into a pre-trained ensemble learning model, and outputting the predicted power consumption value corresponding to the computing power node.

[0016] Preferably, the pre-trained ensemble learning model adopts a network architecture that combines the LightGBM model with the Sequential neural network; The LightGBM model is used to process the discrete features and structured data in the multi-scale feature matrix, and the Sequential neural network is used to extract the dynamic features of the time series in the multi-scale feature matrix. The hardware runtime characteristics include CPU utilization, GPU utilization, and memory usage data.

[0017] Preferably, after outputting the predicted power consumption value, it also includes: Receive and record the predicted power consumption value and the standardized transferable profile snapshot; The power consumption prediction results and task elasticity parameters are fed back to the external scheduling system through the standard data output interface to perform computing power task migration scheduling and system operation optimization.

[0018] In a second aspect, the present invention provides a computer-integrated sensing and transferability profiling system for computing centers, comprising a data acquisition layer and a modeling and analysis layer. The data acquisition layer acquires platform-level operational data and node-level time-series data through parallel platform API call channels and local monitoring channels, respectively. After cleaning, the acquired platform-level operational data and node-level time-series data are converted into a unified standard format and written into the time-series database. The time-series database is physically isolated from the local monitoring program database of the computing power node. The modeling and analysis layer is used to extract multi-dimensional demand attributes of computing power tasks from the time-series database, generate standardized representation vectors of the computing power tasks, calculate time portability, spatial portability and interruption recoverability based on the standardized representation vectors and weight them to obtain a comprehensive portability index, classify flexibility categories based on the comprehensive portability index, and output a standardized portability profile snapshot of the computing power tasks containing flexibility category labels.

[0019] Preferably, the data acquisition layer is configured with a protocol conversion middleware and a computer-integrated sensor; The protocol conversion middleware is used to receive RESTful protocol requests, gRPC protocol requests, and system private remote procedure call protocol requests issued by the heterogeneous computing power cluster management platform, and to uniformly parse and reassemble them into the system's internal standard message format. The combined computer and sensor is deployed on the computing node in a containerized manner and is used to call the local monitoring script to read the underlying operating status of the device and obtain node-level time-series data. The data acquisition layer embeds a role-based access control (RBAC) model, which is used to grant read-only monitoring permissions to the computer-connected sensor.

[0020] Preferably, the modeling and analysis layer includes a profile generation module and a load modeling module; The profile generation module is used to extract multi-dimensional requirement attributes of computing power tasks to construct standardized representation vectors, calculate time portability, spatial portability and interruption recoverability respectively, and weight them to obtain a comprehensive portability index. Based on the comprehensive portability index, the flexibility category is defined and a standardized portability profile snapshot is output.

[0021] Preferably, it also includes an application service layer; the application service layer is used to extract task arrival rate, task service rate and number of parallel service nodes from the time-series database, establish an M / M / c queuing theory model for multi-node parallel scheduling scenarios of computing power tasks, calculate the system idle probability and the average queuing delay of tasks, multiply the task arrival rate and the average queuing delay of tasks with a preset average equivalent resource occupancy coefficient of a single queued task to obtain the queuing equivalent load, and add the queuing equivalent load to the current running task resource occupancy to generate a corrected comprehensive effective load sequence.

[0022] Preferably, it also includes an output terminal layer; the output terminal layer is used to extract the comprehensive effective load sequence and synchronously read the hardware runtime sequence features in the time series database, concatenate and combine the two to generate a multi-scale feature matrix, input the multi-scale feature matrix into a pre-trained ensemble learning model, and output the power consumption prediction value corresponding to the computing power node.

[0023] Preferably, the application service layer is configured with a standard data output interface for providing data query and read services to the computing power scheduling terminal and the energy management terminal; The application service layer records the predicted power consumption value and the standardized portable profile snapshot in the application database, and feeds back the power consumption prediction results and task elasticity parameters to the external scheduling system through the standard data output interface, providing data basis for cross-domain collaborative scheduling of computing power and green computing energy trading.

[0024] Thirdly, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the computing center-oriented computer-integrated sensing and portable profiling method.

[0025] Fourthly, a computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the computer program implements the steps of the computer-computer joint sensing and transferable profiling method for computing power centers.

[0026] This invention provides a method and system for joint sensing and transferability profiling of computing centers. It has the following beneficial effects: This invention achieves comprehensive collection of multi-source operational data from heterogeneous systems by deploying a combined computer-computer sensor on computing nodes and simultaneously enabling platform API call channels and local monitoring channels. Simultaneously, the cleaned and standardized time-series data is written into a time-series database with an independently established physical isolation mechanism, effectively eliminating the risk of data read / write conflicts in high-concurrency acquisition scenarios and ensuring the stable operation of the underlying sensing system during the data collection phase.

[0027] This invention constructs a standardized representation vector by extracting multi-dimensional requirement attributes of computing tasks, such as service level agreement constraints, computing resources, and network transmission. Based on this vector, it calculates temporal portability, spatial portability, and interruptibility recoverability. This approach transforms complex task operation characteristics into specific comprehensive portability indices and defines corresponding flexibility categories to generate profile snapshots, providing an objective and quantitative data foundation for subsequent cross-domain task scheduling in the system.

[0028] This invention establishes an M / M / c queuing theory model for multi-node parallel scheduling scenarios of computing tasks. It adds the equivalent load of the queued tasks under congestion to the resource occupancy of the currently running tasks to generate a corrected comprehensive effective load sequence. This modeling approach corrects the shortcomings of traditional load models in insufficiently calculating the underlying scheduling overhead and resource reservation consumption during system congestion. Combined with an ensemble learning model, it effectively improves the accuracy of power consumption prediction for computing nodes. Attached Figure Description

[0029] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see the appendix Figure 1 This invention provides a computer-based joint sensing and portability profiling system for computing centers, including a data acquisition layer and a modeling and analysis layer; in embodiments with extended functions, the system may also include an application service layer and an output terminal layer.

[0032] The data acquisition layer is used for unified management of awareness agents. It is configured with protocol conversion middleware to connect to the open interface of the heterogeneous computing power cluster management platform. The data acquisition layer connects the local log system and real-time monitoring port of the computing power nodes. It performs dual-source parallel acquisition operations to obtain task data, queue data, resource usage data, load data, network latency data, and power consumption data.

[0033] Specifically, platform-level operational data and node-level time-series data are obtained through parallel platform API call channels and local monitoring channels, respectively. The collected platform-level operational data and node-level time-series data are cleaned, converted into a unified standard format, and written into the time-series database. The time-series database is physically isolated from the local monitoring program database of the computing power node.

[0034] The modeling and analysis layer connects to the data acquisition layer, receiving and processing the time-series data transmitted from the data acquisition layer. The modeling and analysis layer includes a profile generation module and a load modeling module. The profile generation module constructs a standardized representation vector of the computing power task based on the acquired task parameters and resource data, and calculates the comprehensive portability index of the task. The load modeling module establishes an effective load correction model based on the M / M / c queuing theory algorithm, integrates multi-scale time-series features into the effective load correction model, and performs mapping prediction modeling of node computing power load and power consumption.

[0035] Specifically, multi-dimensional requirement attributes of computing power tasks are extracted from the time-series database, and standardized representation vectors of the computing power tasks are generated. Based on the standardized representation vectors, time portability, spatial portability, and interruption recoverability are calculated and weighted to obtain a comprehensive portability index. Based on the comprehensive portability index, flexibility categories are defined, and a standardized portability profile snapshot of the computing power tasks containing flexibility category labels is output.

[0036] The application service layer connects to the modeling and analysis layer. It receives standardized task profiles, comprehensive effective load sequences, and power consumption prediction results from the modeling and analysis layer. The application service layer is configured with standard data output interfaces to provide data query and read services to the output terminal layers (computing power scheduling terminals and energy management terminals), and to perform computing power task migration scheduling and system operation optimization. The output terminal layer is used to extract the comprehensive effective load sequence and synchronously read the hardware runtime sequence features in the time series database, concatenate the two to generate a multi-scale feature matrix, input the multi-scale feature matrix into the trained ensemble learning model, and output the power consumption prediction value corresponding to the computing node.

[0037] Please see the appendix Figure 2 This invention provides a method for joint sensing and transferability profiling of computing centers, comprising the following steps: S1, the data acquisition layer performs dual-source data acquisition and preprocessing of heterogeneous systems: a combined computer-computer sensor is deployed on the computing node to establish a platform API call channel and a local monitoring channel; task scheduling status, resource allocation information and queue depth data are obtained through the platform API call channel; real-time power consumption, temperature, hardware utilization and network traffic data of the GPU are synchronously obtained through the local monitoring channel; outlier removal, timestamp alignment and unit standardization are performed on the acquired multi-source heterogeneous data, which is converted into a standard format and written into a time-series database, and the time-series database is physically isolated from the local monitoring program database of the computing node. S2, the modeling and analysis layer generates a quantitative portability profile of the computing task; extracts the SLA constraint parameters, resource requirement parameters, and network requirement parameters of the computing task, and generates a standardized representation vector; calculates the time portability, spatial portability, and interruption recoverability of the computing task based on the standardized representation vector; calculates a comprehensive portability index by weighting the three types of portability; and classifies the flexibility of the computing task into rigid, time-elastic, spatial-elastic, or fully elastic categories based on the numerical range of the comprehensive portability index, and outputs a standardized portability profile of the task.

[0038] In an implementation that includes extended functionality, the method may further include the following steps: S3, the modeling and analysis layer establishes an effective load correction model that integrates queuing theory; for the multi-node parallel scheduling and operation mechanism of computing power tasks, an M / M / c queuing theory model is constructed; the task arrival rate, service rate, and number of service nodes are read to calculate the system idle probability and the average queuing delay of tasks; the average equivalent resource occupancy coefficient of a single queuing task is multiplied by the average queuing delay of the task to obtain the queuing equivalent load under queuing congestion; the queuing equivalent load is added to the current running task resource occupancy to generate a corrected comprehensive effective load sequence; S4, the output terminal layer performs multi-scale feature fusion for power consumption prediction modeling; extracts the comprehensive effective load sequence and collected hardware runtime sequence features to construct a multi-scale feature matrix; inputs the multi-scale feature matrix into the trained ensemble learning model; the ensemble learning model performs feature fitting processing and outputs the power consumption prediction value corresponding to the computing power node; the application service layer receives and records the power consumption prediction value and the standardized transferable profile, and feeds it back to the external calling system.

[0039] The technical solution of this invention will be described in detail below, with reference to each module and specific operating steps.

[0040] The data acquisition layer performs dual-source data acquisition and preprocessing for heterogeneous systems, specifically including the following sub-steps: S101 deploys a combined computing and computing sensor on the computing node and performs hardware baseline data acquisition. The data acquisition layer deploys the combined computing and computing sensor to the target computing node in a containerized manner. The container base image integrates an interpreted program runtime environment, network protocol dependency libraries, and system-level monitoring and acquisition tools. When the combined computing and computing sensor starts for the first time, it triggers the hardware baseline data acquisition logic, extracting the static configuration information of the computing node in parallel through both platform API and local monitoring scripts.

[0041] Static configuration information includes the device's interconnection network IP address, hostname, device unique identifier (UUID), number of CPU cores, GPU model, and network communication interface type. The computer-integrated sensor writes this static configuration information into an independently configured relational database, serving as static baseline data for subsequent dynamic comparisons of device status.

[0042] For the specific operations of building the container base image and packaging the dependent libraries, those skilled in the art can use general container engine instructions to complete them. The image building process is a well-known technology in this field and will not be described in detail here.

[0043] S102, initiates periodic data acquisition loop, performing dual-source parallel data acquisition. When the computer-controlled sensor is not running for the first time, it directly enters the periodic acquisition loop, simultaneously activating the platform API call channel and the local monitoring channel.

[0044] Within the platform's API call channel, the data acquisition layer is configured with a unified multi-source API adaptation layer. This unified multi-source API adaptation layer incorporates a protocol conversion middleware. The protocol conversion middleware receives RESTful protocol requests for descriptive state transitions, open-source gRPC protocol requests, and system-specific remote procedure call protocol requests from the heterogeneous computing cluster management platform, and uniformly parses and reassembles them into the system's internal standard message format.

[0045] The multi-source API unified adaptation layer calls the heterogeneous computing power cluster management platform interface to obtain platform-level operational data. Platform-level operational data includes task scheduling status, computing resource allocation information, queue depth parameters, task lifecycle logs, and scheduling strategy parameters.

[0046] The multi-source API unified adaptation layer embeds a role-based access control (RBAC) model. The RBAC model grants read-only monitoring permissions to the computer-connected sensor, blocking write operation requests that modify cluster resource configurations or interfere with task scheduling.

[0047] In the local monitoring channel, the computer-controlled sensor calls the local monitoring script to execute system hardware monitoring commands. The local monitoring script directly reads the underlying operating status of the device and obtains node-level time-series data. The node-level time-series data includes computing load operating status data, cross-node communication data, and task power consumption data.

[0048] The computing load operation status data includes real-time server power, CPU and GPU computing chip utilization, memory space utilization, disk read / write utilization, task queue length, and task throughput.

[0049] Cross-node communication data includes network port communicability, communication bandwidth utilization, end-to-end physical latency, data packet loss rate, and packet retransmission rate. Task power consumption data includes real-time power consumption curves for the task's execution cycle.

[0050] The data acquisition layer sets the data acquisition time granularity to be between 1 and 10 seconds, and the spatial granularity level covers the entire computing power cluster level, the independent business domain level, and the individual device node level.

[0051] For the specific syntax of system hardware monitoring commands and the method of reading hardware parameters, those skilled in the art can call the system's built-in commands and open-source hardware interaction toolsets to implement them. The underlying command calling mechanism is a well-known technology in this field and will not be elaborated here.

[0052] S103 performs heterogeneous data cleaning, isolation, standardization, and persistent storage. The data acquisition layer extracts multi-source heterogeneous data acquired in parallel from two sources and performs data cleaning on it.

[0053] Data cleaning operations include setting hardware physical thresholds to remove outlier values ​​that exceed the limits, filling missing data items with historical smoothed values ​​from adjacent time windows, aligning the timestamps of multi-source data with the system standard clock as a reference, and standardizing the engineering units of measurement for various types of parameters.

[0054] The data acquisition layer converts the cleaned data sequence into a unified standard schema structured format. The data acquisition layer then writes the time-series data in the standard schema structured format into a dedicated time-series database. A physical isolation mechanism is established between the time-series database and the local monitoring program database on the computing nodes to eliminate the risk of data read / write conflicts in high-frequency concurrent acquisition scenarios.

[0055] Based on analysis requirements, the data acquisition layer performs aggregation and grouping operations on records in the time-series database according to time spans of minutes, hours, and days. The application service layer generates a structured query language (SQL) interface based on the aggregated and grouped underlying data, extracts target feature data according to the call requests of downstream data analysis modules, and completes the standardized on-demand output of multi-dimensional perception data.

[0056] The portrait generation module within the modeling and analysis layer performs multi-dimensional attribute extraction and quantification portrait generation for computing power tasks, specifically including the following sub-steps: S201, the profile generation module quantifies the performance data of computing power nodes and receives user input and task parameters transmitted from the data acquisition layer. The profile generation module extracts multi-dimensional requirement attributes of the computing power task and constructs a standardized representation vector for the computing power task. The calculation formula for this standardized representation vector is as follows: ; In the formula, This is the standardized representation vector for the computing power task; Task type; For the set of service level agreement constraints; For the set of computing resource requirements; A set of network transmission requirements; This is a set of migration attributes for the task.

[0057] In the above standardized representation vectors, task type The range of values ​​includes model training, inference services, data processing, and graphics rendering.

[0058] Service Level Agreement Constraints Specifically, the deadline is... Maximum delay duration Checkpoint interval and data localization requirements constitute.

[0059] Set of computing resource requirements Specifically, it depends on the number of graphics processing units required. Central Processing Unit Core Requirement Memory capacity requirements and network bandwidth requirements constitute.

[0060] Network transmission demand set Specifically, end-to-end latency requirements Bandwidth requirements and network jitter tolerance constitute.

[0061] Task migration attribute set Specifically, it depends on the amount of data migrated across nodes. Task recovery time and recalculation of costs constitute.

[0062] S202, the profile generation module constructs a cross-domain system-level evaluation index. Based on the underlying parameters in the standardized representation vector, the profile generation module calculates the temporal portability, spatial portability, and interruptibility recoverability of the computing task. The profile generation module then calculates a comprehensive portability index based on these three indicators, using the following formula: ; In the formula, It is a comprehensive portability index; This is the time weighting coefficient; For time portability; Spatial weight coefficient; For spatial portability; This is the interruption weighting coefficient; For interruption recoverability. Weighting coefficients. , , The sum of the values ​​must be 1.

[0063] The specific mathematical steps for determining the weight coefficients of each item using the analytic hierarchy process can be completed by those skilled in the art based on the judgment matrix construction and consistency verification process. This is a well-known technique in the field and will not be elaborated here.

[0064] The calculation formulas for each sub-indicator in the above comprehensive portability index are as follows: ; ; ; In the formula, This is the maximum delay duration; This represents the estimated execution time of the computing task. This is the network cost coefficient; To migrate data across nodes; For bandwidth requirements; An indicator function to determine whether the checkpoint interval is greater than zero; To recalculate costs. Temporal portability reflects the degree to which a task can be delayed; spatial portability reflects the physical network feasibility of cross-domain node migration; interruption recoverability reflects the computational efficiency of resuming computational tasks after an interruption.

[0065] S203, the profile generation module outputs a standardized, transferable profile snapshot of the computing power task. The profile generation module reads the calculated comprehensive transferability index and limits its value range to between 0 and 1. The profile generation module presets a value range and automatically classifies the flexibility category of the computing power task based on the specific value range in which the comprehensive transferability index falls.

[0066] The profile generation module saves the task's runtime context through serialization, generating lightweight, standardized, and transferable profile snapshots, expressed in the following formula: ; In the formula, Standardize and transferable profile snapshots; It is a comprehensive portability index; For the flexibility category; Prioritize task scheduling; For multidimensional migration costs; For power consumption sensitivity.

[0067] In the above expression, flexibility category The set of values ​​includes four independent states: rigidity, temporal elasticity, spatial elasticity, and full elasticity. The profile generation module encapsulates and outputs a standardized multidimensional portable cost profile containing flexibility category labels, providing characterization data for task elasticity in cross-domain scheduling of the system.

[0068] The load modeling module within the modeling and analysis layer establishes an effective load correction model that incorporates queuing theory, specifically including the following sub-steps: S301, the load modeling module establishes an M / M / c queuing theory model for multi-node parallel scheduling scenarios of computing power tasks. The load modeling module extracts task arrival rate, task service rate, and the number of parallel service nodes from the time-series database. The load modeling module calculates the current workload of the computing power system and determines whether the system meets the conditions for stable operation. The formula for calculating workload is as follows: ; In the formula, For business intensity; For task completion rate; This refers to the number of parallel service nodes. Service rate for tasks. The system's stable operation condition is based on workload. Strictly less than 1.

[0069] After confirming that the system meets the stability conditions, the load modeling module calculates the system idle probability and, based on this probability, calculates the average queuing delay and total average delay for tasks. The corresponding calculation formulas are as follows: ; ; ; In the formula, The system is idle probability; For iterative accumulation of variables; This refers to the number of parallel service nodes. For business intensity; Average queuing time for tasks; For task completion rate; This represents the total average latency; Service rate for tasks.

[0070] For the historical data distribution statistics and verification calculation of computing task arrival rate and service rate, those skilled in the art can use general statistical model parameter estimation methods to complete it. The underlying data fitting process is a well-known technology in this field and will not be described in detail here.

[0071] S302, the load modeling module calculates the queuing equivalent load of the system under congestion. Traditional load models use the running computing power load index to calculate power consumption, but do not include the underlying scheduling overhead and resource reservation consumption caused by tasks during the queuing waiting phase in the calculation scope. The load modeling module sets the average equivalent resource occupancy coefficient for a single queuing task. The load modeling module multiplies the task arrival rate, the average queuing delay of the task, and the average equivalent resource occupancy coefficient to obtain the queuing equivalent load of the computing power task. The calculation formula for the queuing equivalent load is as follows: ; In the formula, For queuing equivalent load; For task completion rate; Average queuing time for tasks; This represents the average equivalent resource utilization coefficient for a single queued task.

[0072] S303, the load modeling module generates a corrected comprehensive effective load sequence. The load modeling module extracts resource usage data for all running tasks within the current computing node. The load modeling module adds the resource usage of running tasks to the queued equivalent load to uniformly represent the true physical load of the computing system. The formula for calculating the corrected comprehensive effective load is as follows: ; In the formula, To comprehensively assess effective load; Index of the set of tasks that are currently running; For the first Resource utilization coefficient of each running task; For task completion rate; Average queuing time for tasks; This represents the average equivalent resource utilization coefficient for a single queued task.

[0073] The first term on the right side of the above formula represents the sum of resource usage of currently running tasks, and the second term represents the equivalent load of queued tasks. The load modeling module stores the generated comprehensive effective load sequence in a designated storage area, using it as the basic underlying feature for high-concurrency operation scenarios, providing data support for the subsequent ensemble learning model to output high-precision power consumption prediction values.

[0074] The load modeling module within the modeling and analysis layer performs multi-scale feature fusion-based power consumption prediction modeling, specifically including the following sub-steps: S401, the load modeling module constructs a multi-scale feature matrix and an ensemble learning model architecture. The load modeling module extracts the comprehensive effective load sequence generated in the preceding steps and simultaneously reads the hardware runtime timing features acquired by the data acquisition layer. These hardware runtime timing features include CPU utilization, GPU utilization, and memory usage data.

[0075] The load modeling module combines the effective load sequence with hardware runtime timing features to generate a multi-scale feature matrix. The load modeling module also constructs an ensemble learning model within the system. This ensemble learning model employs a network architecture that integrates the LightGBM model and a Sequential neural network.

[0076] The load modeling module uses the LightGBM model to process discrete features and structured data in the multi-scale feature matrix, and uses the Sequential neural network to extract the dynamic features of the time series in the multi-scale feature matrix.

[0077] S402, the load modeling module acquires historical data to train and optimize the parameters of the ensemble learning model. The load modeling module accesses a historical dataset of dynamic workload scheduling in the data center. This historical dataset covers continuous operation logs and includes fields such as task arrival time, computing resource requirements, server configuration information, task execution duration, queue length, and real-time power consumption. The data time granularity is uniformly set to 15 minutes.

[0078] The load modeling module adopts a time-series data partitioning strategy, using the first 80% of the data sequences as the training set to input into the ensemble learning model for iterative fitting, and the last 20% of the data sequences as the test set to evaluate the model's generalization performance.

[0079] For the hyperparameter optimization and gradient calculation of the gradient boosting framework of the decision tree model and the backpropagation of the neural network, those skilled in the art can use grid search and conventional optimizer algorithms to complete the task. The underlying parameter iteration process is a well-known technology in this field and will not be described in detail here.

[0080] S403, the load modeling module establishes model performance evaluation indicators and performs error calculations. The load modeling module uses two calculation standards: root mean square error (RMSE) and mean absolute percentage error (MAS). RMSE measures the absolute deviation between the model's predicted power consumption and the actual power consumption. MAS measures the relative error of the model's predictions. The corresponding error calculation formulas are as follows: ; ; In the formula, This is the root mean square error; This represents the number of samples in the test set. The sample sequence number; For the first The actual power consumption value of each sample; For the first The model predicts the power consumption value for each sample; The mean absolute percentage error.

[0081] S404, the load modeling module outputs the prediction results and hands them over to the application service layer for scheduling. The load modeling module inputs the multi-scale feature matrix of the current moment into the trained and validated ensemble learning model. The ensemble learning model combines the effective load features corrected by queuing theory to perform data fitting processing and outputs the predicted power consumption value corresponding to the current computing node.

[0082] The effective load model based on task queuing delay correction corrects the systematic underestimation defect of traditional models in congested scenarios such as task arrival rate approaching service capacity. It avoids the fitting bias introduced by directly inputting the original high-dimensional features into the Sequential neural network, reduces the root mean square error to 177.3550, and reduces the mean absolute percentage error to 0.6450%.

[0083] The application service layer receives power consumption predictions and standardized, transferable profile snapshots from the load modeling module. It records this data in the application database and feeds back the power consumption prediction results and task elasticity parameters to the external scheduling system through a standard data output interface, providing data support for cross-domain collaborative scheduling of computing power and green computing energy trading in actual business operations.

[0084] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve corresponding method flows or corresponding functions. The processor described in this embodiment can be used for the operation of a computer-based joint sensing and portable profiling method for computing power centers, including the following steps: performing dual-source data acquisition and preprocessing of heterogeneous systems: acquiring platform-level operating data and node data through parallel-enabled platform API call channels and local monitoring channels, respectively. The platform-level and node-level time-series data are cleaned, transformed into a unified standard format, and written into a time-series database. This time-series database is physically isolated from the local monitoring program database of the computing power nodes. Multi-dimensional requirement attributes of computing power tasks are extracted from the time-series database to generate standardized representation vectors for the computing power tasks. Based on these standardized representation vectors, time portability, spatial portability, and interruption recoverability are calculated and weighted to obtain a comprehensive portability index. Based on the comprehensive portability index, flexibility categories are defined, and a standardized portability profile snapshot of the computing power task, including a flexibility category label, is output.

[0085] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory). This computer-readable storage medium is a memory device in a terminal device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device.

[0086] One or more instructions stored in a computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the computer-computer joint sensing and portability profiling method for computing power centers in the above embodiments. One or more instructions in the computer-readable storage medium are loaded by the processor and executed as follows: Perform dual-source data acquisition and preprocessing of heterogeneous systems: acquire platform-level running data and node-level time-series data through parallel platform API call channels and local monitoring channels respectively; clean the acquired platform-level running data and node-level time-series data, convert them into a unified standard format, and write them into a time-series database; establish physical isolation between the time-series database and the local monitoring program database of computing power nodes; extract multi-dimensional demand attributes of computing power tasks from the time-series database, generate standardized representation vectors of computing power tasks, calculate time portability, spatial portability, and interrupt recoveryability based on the standardized representation vectors, and obtain a comprehensive portability index by weighting; classify flexibility categories based on the comprehensive portability index, and output a standardized portability profile snapshot of the computing power task containing flexibility category labels.

[0087] 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.

[0088] 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.

[0089] 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.

[0090] 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.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for joint sensing and transferability profiling of computing centers, characterized in that, Includes the following steps: Platform-level operational data and node-level time-series data are obtained through parallel platform API call channels and local monitoring channels, respectively. The collected platform-level operational data and node-level time-series data are cleaned, converted into a unified standard format, and written into the time-series database. The time-series database is physically isolated from the computing power node local monitoring program database. Multi-dimensional requirement attributes of computing power tasks are extracted from the time-series database, and standardized representation vectors of the computing power tasks are generated. Based on the standardized representation vectors, time portability, spatial portability, and interruption recoverability are calculated and weighted to obtain a comprehensive portability index. Based on the comprehensive portability index, flexibility categories are defined, and a standardized portability profile snapshot of the computing power tasks containing flexibility category labels is output.

2. The method for joint sensing and transferability profiling of computing centers according to claim 1, characterized in that, The dual-source data acquisition and preprocessing of the heterogeneous system specifically includes: Deploy the computer-computer joint sensor in a containerized manner on the computing power node and perform hardware baseline acquisition to extract the static configuration information of the computing power node; Start a periodic data collection cycle and simultaneously enable the platform API call channel and the local monitoring channel; The platform-level operation data is obtained by calling the heterogeneous computing power cluster management platform interface through the multi-source API unified adaptation layer, and the computer-connected sensor is granted read-only monitoring permission through the role-based access control (RBAC) model. Node-level time-series data is obtained by reading the underlying operating status of the device through the local monitoring channel; The heterogeneous data acquired in parallel from two sources undergoes outlier removal, missing value filling, timestamp alignment, and unit standardization, and is then converted into a unified standard schema structured format and written into the time series database.

3. The method for joint sensing and transferability profiling of computing centers according to claim 1, characterized in that, The multi-dimensional requirement attributes consist of task type, service level agreement constraint set, computing resource requirement set, network transmission requirement set, and task migration attribute set; The set of service level agreement constraints includes deadline, maximum delay duration, checkpoint interval, and data localization requirements. The set of computing resource requirements includes the number of graphics processing units (GPUs), the number of CPU cores, the memory capacity, and the network bandwidth. The set of network transmission requirements includes end-to-end latency requirements, bandwidth requirements, and network jitter tolerance. The set of task migration attributes includes the amount of data migrated across nodes, task recovery time, and recalculation cost.

4. The method for joint sensing and transferability profiling of computing centers according to claim 1, characterized in that, The process involves calculating temporal portability, spatial portability, and interruption recoverability based on standardized representation vectors, and then weighting them to obtain a comprehensive portability index. Specifically, this includes: The time portability is calculated based on the ratio of the maximum delay duration to the expected execution duration; The spatial portability is calculated based on the ratio of cross-node migration data volume to bandwidth requirement, adjusted by a network cost coefficient. The interruption recoverability is calculated based on an indicator function that checks whether the checkpoint interval is greater than zero and the proportion of recalculation cost. The overall portability index, calculated by weighting the temporal portability, spatial portability, and interruptibility recoverability, satisfies the following: ; In the formula, It is a comprehensive portability index; This is the time weighting coefficient; For time portability; Spatial weight coefficient; For spatial portability; This is the interruption weighting coefficient; For interruption recoverability; weighting coefficient , , The sum of the values ​​must be 1.

5. The method for joint sensing and transferability profiling of computing centers according to claim 1, characterized in that, The process of classifying flexibility categories based on a comprehensive transferability index and outputting a standardized transferability profile snapshot specifically includes: The comprehensive portability index value is limited to between 0 and 1, and a preset value range is defined. The flexibility category is automatically determined based on the specific value range in which the comprehensive portability index falls. The flexibility category includes rigidity, time elasticity, spatial elasticity, and full elasticity. The task's runtime context is saved by serialization, generating a lightweight, standardized, and portable profile snapshot. The standardized and portable profile snapshot includes a comprehensive portability index, flexibility category, task scheduling priority, multidimensional migration cost, and power consumption sensitivity. The standardized multidimensional portable cost profile, which includes a flexibility category label, is then encapsulated and output.

6. The computing center-based computer-based joint sensing and transferability profiling method according to claim 1, characterized in that, After outputting the standardized and transferable profile snapshot, the following steps are also included: extracting the task arrival rate, task service rate, and number of parallel service nodes from the time-series database; establishing an M / M / c queuing theory model for multi-node parallel scheduling scenarios of computing power tasks; calculating the system idle probability and the average queuing delay of tasks; multiplying the task arrival rate, the average queuing delay of tasks, and the preset average equivalent resource occupancy coefficient of a single queued task to obtain the queuing equivalent load; adding the queuing equivalent load to the current running task resource occupancy to generate a corrected comprehensive effective load sequence.

7. The method for joint sensing and transferability profiling of computing centers according to claim 6, characterized in that, The establishment of the M / M / c queuing theory model, and the calculation of the system idle probability and the average queuing delay of tasks, specifically includes: Calculate the current workload of the computing system: In the formula, For business intensity; For task completion rate; This refers to the number of parallel service nodes. Service rate for tasks; Determine whether the business intensity ρ is strictly less than 1 to confirm that the system meets the conditions for stable operation; After confirming that the system meets the stability conditions, the system idle probability is calculated, and the average queuing delay of the task is calculated based on the system idle probability.

8. The method for joint sensing and transferability profiling of computing centers according to claim 6, characterized in that, The queuing equivalent load satisfies: ; In the formula, For queuing equivalent load; For task completion rate; Average queuing time for tasks; This represents the average equivalent resource utilization coefficient for a single queued task. The overall effective load satisfies: ; In the formula, To comprehensively assess effective load; Index of the set of tasks that are currently running; For the first Resource utilization coefficient of each running task; This represents the sum of resource usage for all currently running tasks.

9. The computing center-based computer-based joint sensing and transferability profiling method according to claim 6, characterized in that, After generating the corrected comprehensive effective load sequence, the following steps are also included: extracting the comprehensive effective load sequence and synchronously reading the hardware runtime timing features in the time series database, concatenating and combining the two to generate a multi-scale feature matrix, inputting the multi-scale feature matrix into a pre-trained ensemble learning model, and outputting the predicted power consumption value corresponding to the computing power node.

10. The method for joint sensing and transferability profiling of computing centers according to claim 1, characterized in that, The pre-trained ensemble learning model adopts a network architecture that combines the LightGBM model with the Sequential neural network; The LightGBM model is used to process the discrete features and structured data in the multi-scale feature matrix, and the Sequential neural network is used to extract the dynamic features of the time series in the multi-scale feature matrix. The hardware runtime characteristics include CPU utilization, GPU utilization, and memory usage data.

11. The method for joint sensing and transferability profiling of computing centers according to claim 1, characterized in that, After outputting the predicted power consumption value, it also includes: Receive and record the predicted power consumption value and the standardized transferable profile snapshot; The power consumption prediction results and task elasticity parameters are fed back to the external scheduling system through the standard data output interface to perform computing power task migration scheduling and system operation optimization.

12. A computer-based joint sensing and transferable profiling system for computing power centers, characterized in that, It includes a data acquisition layer and a modeling and analysis layer; The data acquisition layer acquires platform-level operational data and node-level time-series data through parallel platform API call channels and local monitoring channels, respectively. After cleaning, the acquired platform-level operational data and node-level time-series data are converted into a unified standard format and written into the time-series database. The time-series database is physically isolated from the local monitoring program database of the computing power node. The modeling and analysis layer is used to extract multi-dimensional demand attributes of computing power tasks from the time-series database, generate standardized representation vectors of the computing power tasks, calculate time portability, spatial portability and interruption recoverability based on the standardized representation vectors and weight them to obtain a comprehensive portability index, classify flexibility categories based on the comprehensive portability index, and output a standardized portability profile snapshot of the computing power tasks containing flexibility category labels.

13. The system according to claim 12, characterized in that, The data acquisition layer is configured with protocol conversion middleware and computer-integrated sensing; The protocol conversion middleware is used to receive RESTful protocol requests, gRPC protocol requests, and system private remote procedure call protocol requests issued by the heterogeneous computing power cluster management platform, and to uniformly parse and reassemble them into the system's internal standard message format. The combined computer and sensor is deployed on the computing node in a containerized manner and is used to call the local monitoring script to read the underlying operating status of the device and obtain node-level time-series data. The data acquisition layer embeds a role-based access control (RBAC) model, which is used to grant read-only monitoring permissions to the computer-connected sensor.

14. The system according to claim 12, characterized in that, The modeling and analysis layer includes a profile generation module and a load modeling module; The profile generation module is used to extract multi-dimensional requirement attributes of computing power tasks to construct standardized representation vectors, calculate time portability, spatial portability and interruption recoverability respectively, and weight them to obtain a comprehensive portability index. Based on the comprehensive portability index, the flexibility category is defined and a standardized portability profile snapshot is output.

15. The computing center-based computer-integrated sensing and transferable profiling system according to claim 12, characterized in that, It also includes an application service layer; the application service layer is used to extract task arrival rate, task service rate and number of parallel service nodes from the time-series database, establish an M / M / c queuing theory model for multi-node parallel scheduling scenarios of computing power tasks, calculate the system idle probability and the average queuing delay of tasks, multiply the task arrival rate and the average queuing delay of tasks with the preset average equivalent resource occupancy coefficient of a single queued task to obtain the queuing equivalent load, and add the queuing equivalent load to the current running task resource occupancy to generate a corrected comprehensive effective load sequence.

16. The computing center-based computer-integrated sensing and transferable profiling system according to claim 15, characterized in that, It also includes an output terminal layer; the output terminal layer is used to extract the comprehensive effective load sequence and synchronously read the hardware runtime sequence features in the time series database, concatenate the two to generate a multi-scale feature matrix, input the multi-scale feature matrix into a pre-trained ensemble learning model, and output the power consumption prediction value corresponding to the computing power node.

17. The system according to claim 16, characterized in that, The application service layer is configured with a standard data output interface, which is used to provide data query and read services to the computing power scheduling terminal and the energy management terminal; The application service layer records the predicted power consumption value and the standardized portable profile snapshot in the application database, and feeds back the power consumption prediction results and task elasticity parameters to the external scheduling system through the standard data output interface, providing data basis for cross-domain collaborative scheduling of computing power and green computing energy trading.

18. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the computer-based joint sensing and transferability profiling method for computing centers as described in any one of claims 1 to 11.

19. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the computer-based joint sensing and transferability profiling method for computing centers as described in any one of claims 1 to 11.