A method and system for identifying and correcting energy consumption of multi-task computing power tasks in data centers

By collecting and aligning resource usage and power data in data centers, constructing pre- and post-event analysis windows for differential analysis, identifying and correcting interference, and generating highly reliable task-level energy consumption labels, this method solves the problem of identifying the energy consumption contribution of a single computing task in a multi-task environment in data centers, and improves the accuracy of energy consumption assessment and the reliability of green scheduling.

CN123114881APending Publication Date: 2026-11-17GUIZHOU POWER GRID CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611026195.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In a data center multi-task shared operation environment, existing methods are unable to accurately identify the energy consumption contribution of a single computing task, resulting in inaccurate energy consumption assessment results and affecting green scheduling decisions.

Method used

By collecting resource usage data and server power data of the target computing task and other computing tasks on the same server, and constructing an analysis window before and after the event after time alignment, performing dual-window baseline difference, identifying interference feature information, and performing interference correction and energy increment analysis, a credible task-level energy consumption label is generated.

Benefits of technology

It enables accurate identification of the energy consumption contribution of a single computing task in a multi-task sharing environment, improves the reliability of task-level energy consumption assessment results, reduces the interference of background task fluctuations on energy consumption management, and supports more accurate energy consumption prediction and green scheduling.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114881A_ABST
    Figure CN123114881A_ABST
Patent Text Reader

Abstract

This application provides a method and system for identifying and correcting energy consumption of multi-task computing power tasks in data centers, relating to the field of data center energy consumption monitoring technology. It obtains resource sequences and power sequences by time-aligning collected resource usage data and server power data; constructs pre-event analysis windows and post-event analysis windows; performs dual-window baseline differencing on the power sequences to obtain the power change characteristics of the target computing power task under critical events; identifies interference characteristics of other computing power tasks on server power changes; further performs interference correction and energy increment analysis on the target computing power task to obtain incremental power and incremental energy consumption results; and generates a credible task-level energy consumption label based on the incremental power and incremental energy consumption results. This application can accurately identify the energy consumption contribution of a single computing power task in a multi-task shared operation environment in a data center, improving the credibility of task-level energy consumption assessment results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of data center energy consumption monitoring technology, and more specifically, to a method and system for identifying and correcting energy consumption of multi-task computing power tasks in data centers. Background Technology

[0002] With the continuous growth of artificial intelligence training, model inference, video processing, scientific computing, and big data analysis tasks in data centers, data center energy consumption monitoring technology has become an important supporting technology for the construction of green computing infrastructure. This technology can continuously collect and manage power data at different levels such as servers, processors, graphics cards, and racks, providing basic data for energy consumption accounting, resource scheduling, task migration, and computing resource optimization. Especially in a multi-task shared operation environment, the energy consumption monitoring results are not only related to the overall energy consumption management level of the data center, but also directly affect the energy consumption assessment and green scheduling decisions of different computing tasks.

[0003] In existing technologies, data centers can typically obtain data on server power consumption, processor power consumption, graphics card power consumption, or rack power consumption. However, this power data is mostly equipment-level or component-level monitoring results, which are difficult to directly reflect the actual energy consumption contribution of a single computing task. To obtain task-level energy consumption, existing methods often use resource allocation, power consumption prediction models, or aggregated power decomposition for estimation. However, when multiple containers, processes, or GPU tasks run simultaneously on the same server, the start-up, termination, or load changes of the target task often overlap with the resource fluctuations of other tasks, making it difficult to distinguish the source of server power changes. If allocation is directly based on resource allocation, the differences in power changes between different tasks are easily overlooked. If only the overall power curve is split, it is easily affected by background task changes, missing monitoring data, and competition for shared resources, resulting in overestimation or underestimation of task energy consumption, which in turn affects subsequent energy consumption prediction and green scheduling judgment. Therefore, how to accurately identify the energy consumption contribution of a single computing task in a multi-task shared operation environment in a data center, thereby improving the credibility of task-level energy consumption assessment results, has become a challenge for the industry. Summary of the Invention

[0004] This application provides a method and system for identifying and correcting the energy consumption of multi-task computing power tasks in data centers. It can accurately identify the energy consumption contribution of a single computing power task in a multi-task shared operation environment in a data center, thereby improving the reliability of task-level energy consumption assessment results.

[0005] In a first aspect, this application provides a method for identifying and correcting energy consumption of multi-task computing power tasks in a data center, comprising the following steps: Collect resource usage data and server power data of the target computing power task and other computing power tasks on the same server; The collected resource usage data and server power data are time-aligned to obtain resource sequences and power sequences; Based on the running events of the target computing power task, a pre-event analysis window and a post-event analysis window are constructed. Based on the pre-event analysis window and the post-event analysis window, a dual-window baseline difference is performed on the power sequence to obtain the power change characteristics of the target computing power task under key events. The interference characteristics of other computing power tasks on server power changes can be identified by the resource changes of other computing power tasks in the resource sequence. Interference correction and energy increment analysis are performed on the target computing power task using the interference feature information and the power change feature to obtain the incremental power and incremental energy consumption results of the target computing power task. Generate a credible task-level energy consumption label using the incremental power and the incremental energy consumption results.

[0006] In this embodiment, time alignment is performed on the collected resource usage data and server power data to obtain the resource sequence and power sequence, specifically including: Set a globally uniform time window length to divide the continuous system running time axis into a series of non-overlapping discrete time windows; Using the discrete time window as the alignment reference, the collected server power data is aggregated and calculated to obtain the average server power value of each time window in the discrete time window sequence, and a power sequence is formed. Using the discrete time window as the alignment reference, the collected resource usage data is aggregated and calculated to obtain the average resource usage vector of each computing power task in each time window within the discrete time window sequence, and a resource sequence is formed.

[0007] In this embodiment, constructing a pre-event analysis window and a post-event analysis window based on the running events of the target computing power task specifically includes: Set a preset skip time based on the time point when the key event of the target computing power task occurs; Before the skipped time, a first preset time period is captured forward as the pre-event analysis window; After the skipped time, a second preset time period is extracted as the post-event analysis window.

[0008] In this embodiment, the power sequence is subjected to dual-window baseline differencing based on the pre-event analysis window and the post-event analysis window to obtain the power change characteristics of the target computing power task under critical events, specifically including: The power sample values ​​within the pre-event analysis window are extracted from the power sequence to obtain the average power of the pre-event analysis window; The power sample values ​​within the post-event analysis window are extracted from the power sequence to obtain the average power of the post-event analysis window. The power change characteristics of the target computing power task under critical events are determined based on the difference between the average power in the post-event analysis window and the average power in the pre-event analysis window.

[0009] In this embodiment, determining the power change characteristics of the target computing power task under critical events based on the difference between the average power of the post-event analysis window and the average power of the pre-event analysis window specifically includes: Read the average server power values ​​in the pre-event analysis window and the post-event analysis window from the power sequence, and perform effective sampling and filtering on the average server power values ​​to obtain effective power sampling points; A preset deviation range is set based on the hardware power reading error and the power deviation within the historical stable operation window, and the effective power sampling points are filtered for steady-state sampling based on the preset deviation range to obtain steady-state power sampling points; The validity weight is determined based on the number of valid power sampling points in the steady-state power sampling points and the number of valid power sampling points in the pre-event analysis window and the post-event analysis window. The difference between the average power of the post-event analysis window and the average power of the pre-event analysis window is corrected by the validity weight, thereby obtaining the power change characteristics of the target computing power task under critical events.

[0010] In this embodiment, identifying the interference characteristics of other computing power tasks on server power changes through resource changes in the resource sequence specifically includes: Extract the resource usage of other computing power tasks within the pre-event analysis window and the post-event analysis window from the resource sequence, and then determine the resource changes of other computing power tasks in the resource sequence; The degree of resource variation in the background environment is determined by combining the resource changes of other computing power tasks with the weights of the impact of different resource types on power changes. When the resource change rate is greater than the stability threshold, it is determined that other computing power tasks have changed significantly, and the interference characteristics of other computing power tasks on server power changes are determined based on the resource change amount of other computing power tasks and the pre-built resource power mapping model.

[0011] In this embodiment, interference correction and energy increment analysis are performed on the target computing task using the interference feature information and the power change feature to obtain the incremental power and incremental energy consumption results of the target computing task, specifically including: The event direction coefficient is set according to the key event type of the target computing power task, and the basic slowly varying power of the server is evaluated; The power change feature is subjected to interference subtraction processing using the interference feature information to obtain the power change value after interference elimination; The incremental power of the target computing power task is determined based on the power change value after interference elimination, the event direction coefficient, and the basic slowly varying power. The incremental energy consumption value within the post-event analysis window is determined based on the incremental power and the duration of the post-event analysis window. Throughout the entire lifecycle of the target computing power task, the incremental energy consumption value within the post-event analysis window is accumulated to obtain the incremental energy consumption result of the target computing power task.

[0012] In this embodiment, determining the incremental power of the target computing task based on the power change value after interference elimination, the event direction coefficient, and the basic slowly varying power specifically includes: The power change value after direction adjustment is determined based on the power change value after interference elimination and the event direction coefficient. The power value after deducting the effect of gradual change is determined based on the power change value after direction adjustment and the basic gradually changing power; The power value after deducting the effects of gradual variation is subjected to non-negative truncation to obtain the incremental power of the target computing power task.

[0013] In this embodiment, generating a reliable task-level energy consumption label based on the incremental power and the incremental energy consumption result specifically includes: During energy consumption calculation, evaluation context quality parameters are collected simultaneously. Penalty weights are set for the evaluation context quality parameters based on historical calibration data; The reliability scores of the incremental power and incremental energy consumption results are determined based on the evaluation context quality parameters and the penalty weights. The incremental power, the incremental energy consumption result, and the credibility score are encapsulated in a standardized format to generate a structured task-level energy consumption label.

[0014] Secondly, this application provides a data center multi-task computing power task energy consumption identification and correction system, used to execute a data center multi-task computing power task energy consumption identification and correction method, the energy consumption identification and correction system comprising: The resource power acquisition module is used to collect resource usage data and server power data of the target computing power task and other computing power tasks on the same server; The time-aligned sequence module is used to time-align the collected resource usage data and server power data to obtain resource sequences and power sequences. The dual-window baseline difference module is used to construct a pre-event analysis window and a post-event analysis window based on the running events of the target computing power task, and to perform dual-window baseline difference on the power sequence based on the pre-event analysis window and the post-event analysis window to obtain the power change characteristics of the target computing power task under key events. The interference feature identification module is used to identify the interference feature information of other computing power tasks on the server power change by the resource change amount of other computing power tasks in the resource sequence; The interference correction increment module is used to perform interference correction and energy increment analysis on the target computing power task through the interference feature information and the power change feature, so as to obtain the incremental power and incremental energy consumption results of the target computing power task. An energy consumption tag generation module is used to generate a reliable task-level energy consumption tag based on the incremental power and the incremental energy consumption result.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: Resource usage data and server power data of the target computing task and other computing tasks on the same server are collected. The collected resource usage data and server power data are time-aligned to obtain resource sequences and power sequences. A pre-event analysis window and a post-event analysis window are constructed based on the running events of the target computing task. Dual-window baseline differencing is performed on the power sequences based on the pre-event analysis window and the post-event analysis window to obtain the power change characteristics of the target computing task under critical events. Interference characteristic information of other computing tasks on server power changes is identified through the resource change amounts of other computing tasks in the resource sequences. Interference correction and energy increment analysis are performed on the target computing task using the interference characteristic information and the power change characteristics to obtain the incremental power and incremental energy consumption results of the target computing task. A credible task-level energy consumption label is generated based on the incremental power and incremental energy consumption results.

[0016] Therefore, this application demonstrates that a reliable task-level energy consumption label can be generated using the incremental power and the incremental energy consumption results. Firstly, by collecting resource usage data and server power data of the target computing task and other computing tasks on the same server, the operating status of the target computing task itself, resource fluctuations of other tasks on the same machine, and overall server power changes can be incorporated into the same energy consumption identification basis, avoiding attribution bias caused by directly inferring the energy consumption of a single task based solely on device-level power data. Secondly, by aligning the resource usage data and server power data over time, resource sequences and power sequences are obtained, allowing data from different sources and with different sampling frequencies to be compared under a unified time reference, reducing energy consumption estimation errors caused by sampling misalignment. Furthermore, by constructing pre-event and post-event analysis windows based on the running events of the target computing task, and performing dual-window baseline differencing on the power sequence, the power difference before and after the start, end, or load change of the target computing task can be used as the identification basis. This method extracts power change characteristics of the target computing task under critical events from the overall server power change, reducing the reliance on fixed-ratio assumptions in traditional resource allocation methods. Furthermore, by identifying interference characteristics through resource changes in other computing tasks, it can explicitly distinguish the impact of sudden increases, decreases, or shared resource competition of other tasks on server power changes, avoiding the misattribution of power fluctuations caused by other tasks to the target computing task. Finally, interference correction and energy increment analysis are performed on the target computing task using the interference characteristics and power change characteristics to obtain the incremental power and energy consumption results of the target computing task. Furthermore, it generates a credible task-level energy consumption label, enabling the task energy consumption results to more clearly reflect the actual contribution of the target computing task to the server power increment. This also facilitates the priority use of high-quality energy consumption labels in subsequent energy consumption prediction, green scheduling, and energy consumption accounting processes, reducing the interference of background task fluctuations on task-level energy consumption management results.

[0017] In summary, the technical solution adopted in this application can accurately identify the energy consumption contribution of a single computing task in a multi-task shared operation environment in a data center, thereby improving the reliability of task-level energy consumption assessment results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1This is an exemplary flowchart of a data center multi-task computing power task energy consumption identification and correction method provided in this application; Figure 2 This is a module structure diagram of a data center multi-task computing power task energy consumption identification and correction system provided in this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] This application provides a method and system for identifying and correcting energy consumption of multi-task computing power tasks in a data center. The core of this method involves collecting resource usage data and server power data of a target computing power task and other computing power tasks on the same server; aligning the collected resource usage data and server power data over time to obtain resource sequences and power sequences; constructing a pre-event analysis window and a post-event analysis window based on the running events of the target computing power task; performing dual-window baseline differencing on the power sequences based on the pre-event analysis window and the post-event analysis window to obtain the power change characteristics of the target computing power task under critical events; identifying interference characteristics of other computing power tasks on server power changes through resource changes in the resource sequences; performing interference correction and energy increment analysis on the target computing power task using the interference characteristics and the power change characteristics to obtain incremental power and incremental energy consumption results for the target computing power task; and generating a credible task-level energy consumption label based on the incremental power and incremental energy consumption results.

[0022] Example 1 To better understand the above technical solutions, a detailed explanation will be provided below with reference to the accompanying drawings and specific implementation methods. Figure 1 As shown in the figure, this is an exemplary flowchart of a data center multi-task computing power task energy consumption identification and correction method according to this embodiment of the present application, including the following steps: In step S1, resource usage data and server power data of the target computing task and other computing tasks on the same server are collected.

[0023] It should be noted that, in this application, the target computing task refers to a computing task that requires task-level energy consumption identification and correction; the other computing tasks refer to computing tasks that run in parallel on the same server as the target computing task and may affect the server power changes; the resource usage data refers to data reflecting the server's computing, storage, transmission and acceleration resources occupied by the computing tasks during operation; and the server power data refers to data reflecting the changes in the server's overall power when the target computing task and other computing tasks run together.

[0024] In specific implementation, firstly, the task execution list on the same server is read through the data center task scheduling platform, container orchestration platform, or server process monitoring interface. The user-specified task identifier is matched to the corresponding container, process, or GPU computing process, and the matched task is taken as the target computing power task. Simultaneously, other running tasks on the same server besides the target computing power task are taken as other computing power tasks. Then, the start, end, restart, or runtime phase switch time of the target computing power task is read from the scheduling log or container lifecycle events, and the read event times are taken as the runtime event information of the target computing power task. Secondly, existing server monitoring interfaces are called according to the container identifier, process identifier, or GPU process identifier of the target computing power task and other computing tasks. This is achieved through container cgroups (Control... The Groups interface reads CPU usage and memory usage, the operating system performance counter interface reads disk read / write and network transmit / receive, and the GPU management interface reads GPU utilization and GPU memory usage. Various resource metrics collected under the same task identifier are aggregated into multiple resource records at the task dimension, and these multiple resource records are used as the resource usage data. Next, the server's overall power is read through the server baseboard management controller, and the read server's overall power is used as the server power data.

[0025] In step S2, the collected resource usage data and server power data are time-aligned to obtain resource sequences and power sequences.

[0026] In this embodiment, the resource usage data and server power data are time-aligned to obtain the resource sequence and power sequence, which can be achieved by the following steps: Set a globally uniform time window length to divide the continuous system running time axis into a series of non-overlapping discrete time windows; Using the discrete time window as the alignment reference, the collected server power data is aggregated and calculated to obtain the average server power value of each time window in the discrete time window sequence, thereby forming a power sequence; Using the discrete time window as the alignment reference, the collected resource usage data is aggregated and calculated to obtain the average resource usage vector of each computing power task in each time window within the discrete time window sequence, thereby forming a resource sequence.

[0027] It should be noted that, in this application, the globally unified time window length represents the time scale used to constrain the mapping granularity of resource usage data and server power data; the system running time axis represents the continuous time reference for the sampling times of the target computing power task, other computing power tasks, and server power data; the discrete time window sequence represents the set of time slices obtained by dividing the system running time axis according to a fixed time scale; the server average power value represents the power characterization value obtained by aggregating and calculating the server power data within the discrete time window; the power sequence represents the set of server average power values ​​arranged according to the discrete time window sequence; the average resource usage vector of each computing power task represents the aggregated characterization data of each computing power task in the resource usage dimension within the discrete time window; and the resource sequence represents the set of average resource usage vectors of each computing power task arranged according to the discrete time window sequence.

[0028] In specific implementation, firstly, the sampling timestamps carried in the collected resource usage data and server power data are read. The earliest valid sampling time of the system is used as the starting point, and the last valid sampling time of the system is used as the ending point. The system running time axis between the two is divided into equal intervals according to a pre-set globally unified time window length. Each segment is treated as a discrete time window, and all discrete time windows arranged chronologically are treated as a discrete time window sequence. Secondly, the server power data is assigned to the corresponding discrete time window according to the sampling timestamps. Valid sampling points are filtered for server power data within the same discrete time window, and invalid readings with missing timestamps or exceeding the hardware sampling range are removed. The remaining server power data is then averaged to obtain the average server power value. The average server power values ​​arranged sequentially according to the discrete time window sequence are treated as a power sequence. The average server power value is calculated using the following formula: in, Indicates the first Average server power within a time window This indicates the number of valid sampling points within the time window. Indicates a point in time The raw power readings are obtained; finally, the resource usage data are assigned to the corresponding discrete time windows according to the sampling time markers. The resource usage data belonging to the same computing power task within the same discrete time window are aggregated using window integration or mean calculation. The aggregated values ​​of CPU usage, memory usage, disk read / write, network transmit / receive, GPU utilization, and GPU memory usage within the discrete time window are combined into the average resource usage vector of each computing power task. The average resource usage vectors of each computing power task arranged in the discrete time window sequence are taken as the resource sequence.

[0029] In step S3, a pre-event analysis window and a post-event analysis window are constructed based on the running events of the target computing power task. Based on the pre-event analysis window and the post-event analysis window, the power sequence is subjected to dual-window baseline difference to obtain the power change characteristics of the target computing power task under key events.

[0030] In this embodiment, constructing the pre-event analysis window and the post-event analysis window based on the running events of the target computing power task can be achieved through the following steps: Set a preset skip time based on the time point when the key event of the target computing power task occurs; Before the skipped time, a first preset time period is captured forward as the pre-event analysis window; After the skipped time, a second preset time period is extracted as the post-event analysis window.

[0031] It should be noted that, in this application, the "running event" refers to an event record used to mark changes in the running state of the target computing power task; the "critical event" refers to a state change event generated when the target computing power task starts, ends, restarts, or switches running phases; the "time point" refers to the moment the critical event occurs on the system running timeline; the "anchor point" refers to the time reference used to construct the pre-event analysis window and the post-event analysis window; the "skip time" refers to the time interval used to avoid instantaneous jitter data after the occurrence of the critical event; the "first preset time period" refers to the time range used to extract power baseline data before the occurrence of the critical event; the "second preset time period" refers to the time range used to extract power observation data after the occurrence of the critical event; the "pre-event analysis window" refers to the time window used to carry the resource sequence and power sequence before the occurrence of the critical event; and the "post-event analysis window" refers to the time window used to carry the resource sequence and power sequence after the occurrence of the critical event.

[0032] In specific implementation, firstly, the time points corresponding to key events are read from the running events of the target computing power task. The event timestamps of task start, task end, task restart, or running phase switching are mapped to the system running time axis, and the mapped event timestamps are used as anchor points. Secondly, skip time is set before and after the anchor point according to a pre-configured preset duration to prevent the drastic fluctuation range caused by container startup, monitoring sampling delay, or server power response lag from entering the subsequent power sequence sampling range, and the time interval excluded near the anchor point is used as the skip time. Next, a first preset period is intercepted along the system running time axis before the skip time, and the time segment that continuously covers the resource sequence and power sequence within this period is used as the pre-event analysis window. Finally, a second preset period is intercepted along the system running time axis after the skip time, and the time segment that continuously covers the resource sequence and power sequence within this period is used as the post-event analysis window.

[0033] In this embodiment, the power sequence is subjected to dual-window baseline differencing based on the pre-event analysis window and the post-event analysis window to obtain the power change characteristics of the target computing power task under critical events. This can be achieved by the following steps: The power sample values ​​within the pre-event analysis window are extracted from the power sequence to obtain the average power of the pre-event analysis window; The power sample values ​​within the post-event analysis window are extracted from the power sequence to obtain the average power of the post-event analysis window. The power change characteristics of the target computing power task under critical events are determined based on the difference between the average power in the post-event analysis window and the average power in the pre-event analysis window.

[0034] It should be noted that, in this application, the dual-window baseline difference refers to the processing method of extracting the power change of the target computing power task by utilizing the power baseline difference between the pre-event analysis window and the post-event analysis window; the power sample value represents the average power value of the server falling into the corresponding analysis window in the power sequence; the average power of the pre-event analysis window represents the power mean used to characterize the baseline power level of the server before the occurrence of the critical event; and the average power of the post-event analysis window represents the power mean used to characterize the observed power level of the server after the occurrence of the critical event.

[0035] In specific implementation, firstly, time index matching is performed in the power sequence according to the start and end times of the pre-event analysis window. The average power value of the servers falling within the window is read, and after removing missing sample values, a set of continuous and valid power sample values ​​is retained. The average of the average power values ​​of each server in this power sample value set is used as the average power of the pre-event analysis window. Secondly, the corresponding average power value of the servers in the power sequence is read according to the start and end times of the post-event analysis window. The set of power sample values ​​is retained using the same valid sample point filtering method as the pre-event analysis window. The average of the average power values ​​of each server in this power sample value set is used as the average power of the post-event analysis window. Finally, the power change characteristics of the target computing power task under the critical event are determined based on the difference characteristics between the average power of the post-event analysis window and the average power of the pre-event analysis window.

[0036] In this embodiment, determining the power change characteristics of the target computing power task under critical events based on the difference between the average power of the post-event analysis window and the average power of the pre-event analysis window can be achieved through the following steps: Read the average server power values ​​in the pre-event analysis window and the post-event analysis window from the power sequence, and perform effective sampling and filtering on the average server power values ​​to obtain effective power sampling points; A preset deviation range is set based on the hardware power reading error and the power deviation within the historical stable operation window, and the effective power sampling points are filtered for steady-state sampling based on the preset deviation range to obtain steady-state power sampling points; The validity weight is determined based on the number of valid power sampling points in the steady-state power sampling points and the number of valid power sampling points in the pre-event analysis window and the post-event analysis window. The difference between the average power of the post-event analysis window and the average power of the pre-event analysis window is corrected by the validity weight, thereby obtaining the power change characteristics of the target computing power task under critical events.

[0037] It should be noted that, in this application, the average server power value refers to the power characterization value obtained by aggregating server power data within a discrete time window; the effective sampling screening refers to the sampling and processing process that removes missing, out-of-bounds, and abnormally read power data; the effective power sampling point refers to the average server power value that can participate in the window power steady-state judgment after effective sampling screening; the hardware power reading error refers to the allowable reading deviation range of the server power monitoring module under rated sampling conditions; the historical stable operation window refers to the baseline time window in which no critical events occur in the target computing power task and the resource changes of other computing power tasks do not exceed the stability threshold; and the power deviation magnitude refers to the relative value of the average server power value. The analysis window's average power generates a power difference; the preset deviation range represents the allowable power fluctuation range centered on the power benchmark and formed according to the upper deviation limit; the steady-state sampling screening represents the process of screening out stable power sampling points based on the preset deviation range; the steady-state power sampling point represents a valid power sampling point whose power deviation amplitude is within the preset deviation range; the validity weight represents the weight value used to constrain the degree to which the difference feature participates in the generation of power change features; the difference feature represents the amount of power change generated by the average power of the analysis window after the event relative to the average power of the analysis window before the event; the power change feature represents the feature reflecting the server power change caused by key events of the target computing power task.

[0038] In specific implementation, firstly, according to the start and end times of the pre-event analysis window and the post-event analysis window, the average server power value falling within the pre-event analysis window or the post-event analysis window is read from the power sequence, and a window power sampling sequence is generated according to the sampling time order. Server average power values ​​with missing sampling times, exceeding the server power reading range, or failing power reading verification in the window power sampling sequence are removed, and the remaining server average power values ​​are used as valid power sampling points within the pre-event analysis window or the post-event analysis window. Secondly, the hardware power reading error of the server power monitoring module is read, and a historical stable operation window is selected from historical operating data where no critical events have occurred in the target computing power task and the resource changes of other computing power tasks do not exceed the stability threshold. The deviation of the average server power within the historical stable operation window from the window average power is calculated. The deviation that can cover the main power fluctuations during the stable operation phase is taken as the historical stable fluctuation upper limit. Then, the larger value between the hardware power reading error and the historical stable fluctuation upper limit is taken as the deviation upper limit. The power range formed by adding the deviation upper limit upwards and subtracting the deviation upper limit downwards from the power benchmark of the pre-event analysis window or the post-event analysis window is used as the power range. To define a preset deviation range, effective power sampling points whose deviation falls within this range are designated as steady-state power sampling points. Next, the steady-state power sampling points are divided into continuous segments according to their sampling time sequence. Steady-state power sampling points continuously falling within the preset deviation range are grouped into the same steady-state sampling segment. The proportion of steady-state power sampling points within the pre-event analysis window to the total number of effective power sampling points within that window is calculated, and this proportion is used as the window steady-state ratio for the pre-event analysis window. Similarly, the proportion of steady-state power sampling points within the post-event analysis window to the total number of effective power sampling points within that window is calculated. The proportion of the number of sampling points is used as the steady-state proportion of the post-event analysis window. The smaller of the steady-state proportions of the pre-event analysis window and the post-event analysis window is used as the effectiveness weight of the target computing power task under the critical event. Finally, the average power of the post-event analysis window is subtracted from the average power of the pre-event analysis window to obtain the power difference between the two analysis windows. This power difference is used as the difference feature, and the difference feature is then weighted and corrected using the effectiveness weight. The corrected difference feature is used as the power change feature of the target computing power task under the critical event.

[0039] In step S4, the interference characteristics of other computing power tasks on server power changes are identified by the resource changes of other computing power tasks in the resource sequence.

[0040] In this embodiment, identifying the interference characteristics of other computing power tasks on server power changes by measuring the resource changes of other computing power tasks in the resource sequence can be achieved through the following steps: Extract the resource usage of other computing power tasks within the pre-event analysis window and the post-event analysis window from the resource sequence, and then determine the resource changes of other computing power tasks in the resource sequence; The degree of resource variation in the background environment is determined by combining the resource changes of other computing power tasks with the weights of the impact of different resource types on power changes. When the resource change rate is greater than the stability threshold, it is determined that other computing power tasks have changed significantly, and the interference characteristics of other computing power tasks on server power changes are determined based on the resource change amount of other computing power tasks and the pre-built resource power mapping model.

[0041] It should be noted that, in this application, the resource usage of other computing tasks refers to the server resource usage data occupied by other computing tasks within the pre-event analysis window and the post-event analysis window; the resource change of other computing tasks refers to the change in resource usage of other computing tasks between the pre-event analysis window and the post-event analysis window; the influence weight of different resource types on power change refers to the weight parameters of the contribution of different server resource changes to server power change; the resource change degree refers to the quantitative result of resource change used to measure the degree of fluctuation in the background environment; the stability threshold refers to the judgment criterion used to distinguish whether other computing tasks have undergone significant changes; the resource power mapping model refers to the mapping relationship used to map the resource change of other computing tasks to the server power interference amount; and the interference feature information refers to the characterization information of the interference effect of resource changes of other computing tasks on server power change.

[0042] In specific implementation, firstly, according to the start and end times of the pre-event analysis window and the post-event analysis window, the resource usage of other computing tasks besides the target computing task is read from the resource sequence. The average resource usage vector of the same other computing task in the pre-event analysis window is used as the resource usage of the front window, and the average resource usage vector of the same other computing task in the post-event analysis window is used as the resource usage of the back window. The change vector of the resource usage of the back window relative to the resource usage of the front window is used as the resource change of other computing tasks in the resource sequence. Secondly, the influence weights of different resource types on power changes are read. The influence weights of different resource types on power changes are pre-calibrated through historical operation windows. Specifically, historical operation windows are selected from the server's historical monitoring data where no critical events have occurred for the target computing task, the server power data is complete, and other computing tasks have resource fluctuations. The resource changes of other computing tasks and the average power changes of the server in adjacent historical operation windows are read respectively. The resource changes of other computing tasks are used as input samples of the resource power mapping model, and the average power changes of the server are used as output samples of the resource power mapping model. Then, the least squares fitting method is used to obtain the resource power mapping model. The mapping coefficients corresponding to CPU usage, memory usage, disk read / write, network transmit / receive, GPU utilization, and GPU memory usage are obtained, and each mapping coefficient is used as the weight of the impact of different resource types on power changes. These mapping coefficients are then written into the resource power mapping model in order of resource type. Next, the resource changes of other computing tasks are weighted and summarized according to the impact weights of different resource types on power changes. The summation result is used as the resource change degree of the background environment fluctuation, and the resource change degree is compared with a stability threshold. When the resource change degree is less than or equal to the stability threshold, the interference power corresponding to other computing tasks is counted as zero, and the zero-value interference is recorded as interference feature information. Finally, when the resource change degree is greater than the stability threshold, it is determined that other computing tasks have undergone significant changes. The resource changes of other computing tasks are input into the resource power mapping model composed of the mapping coefficients. The resource power mapping model outputs an interference power estimate based on the mapping relationship between the overall change feature vector of the resource changes of other computing tasks and the mapping coefficients. This interference power estimate is used as the interference feature information generated by other computing tasks on server power changes. The interference power estimate is calculated according to the following formula: in, This represents the estimated interference power caused by changes in computing power tasks other than the target computing power task. This represents the feature vector indicating the overall change in resource usage for other computing tasks before and after the event. This represents the pre-established mapping coefficient between resource changes and power changes.

[0043] In step S5, interference correction and energy increment analysis are performed on the target computing task using the interference feature information and the power change feature to obtain the incremental power and incremental energy consumption results of the target computing task.

[0044] In this embodiment, the interference correction and energy increment analysis of the target computing power task using the interference feature information and the power change feature, to obtain the incremental power and incremental energy consumption results of the target computing power task, can be achieved through the following steps: The event direction coefficient is set according to the key event type of the target computing power task, and the basic slowly varying power of the server is evaluated; The power change feature is subjected to interference subtraction processing using the interference feature information to obtain the power change value after interference elimination; The incremental power of the target computing power task is determined based on the power change value after interference elimination, the event direction coefficient, and the basic slowly varying power. The incremental energy consumption value within the post-event analysis window is determined based on the incremental power and the duration of the post-event analysis window. Throughout the entire lifecycle of the target computing power task, the incremental energy consumption value within the post-event analysis window is accumulated to obtain the incremental energy consumption result of the target computing power task.

[0045] It should be noted that, in this application, interference correction refers to the process of subtracting the interference effects of other computing power tasks on server power changes from the power change characteristics; energy increment analysis refers to the process of calculating the energy consumption increment of the target computing power task based on the incremental power of the target computing power task and the duration of the post-event analysis window; key event type refers to the event category of the target computing power task starting, ending, entering a high-load phase, or exiting a high-load phase; event direction coefficient refers to the coefficient used to distinguish the direction of power increase or decrease caused by key events; and base slowly varying power refers to server temperature changes and fan speed. The power change is caused by slow changes in power consumption or low-frequency background operation. The interference deduction process refers to the process of removing power interference from other computing tasks from the power change characteristics using interference feature information. The power change value after interference elimination represents the amount of power change retained after deducting the influence of interference from other computing tasks. The incremental power represents the additional server power increment introduced by the target computing task under critical events. The incremental energy consumption value represents the energy consumption value converted from the incremental power of the target computing task in the post-event analysis window. The incremental energy consumption result represents the cumulative result of the incremental energy consumption value corresponding to the target computing task throughout its entire life cycle.

[0046] In specific implementation, firstly, the key event types of the target computing power task are read. When the key event type is the start-up or entry into the high-load phase of the target computing power task, the event direction coefficient is set to a positive coefficient. When the key event type is the end or exit of the high-load phase of the target computing power task, the event direction coefficient is set to a negative coefficient. At the same time, the low-frequency change portion not triggered by the event is extracted from the server power data between the pre-event analysis window and the post-event analysis window. The power change amount corresponding to the temperature, fan speed, or low-frequency background operation changes is used as the basic slowly varying power of the server. Secondly, the power change characteristics and interference power estimates in the interference characteristic information are read from the power change characteristics. The interference is subtracted from the power change characteristics. The power estimate is used as the power change value after interference elimination. Next, the incremental power of the target computing task is determined based on the power change value after interference elimination, the event direction coefficient, and the basic slowly varying power. Subsequently, the start and end times of the post-event analysis window are read and the duration of the post-event analysis window is obtained. The incremental power of the target computing task is multiplied by the duration of the post-event analysis window, and the result is used as the incremental energy consumption value within the post-event analysis window. Finally, the incremental energy consumption values ​​corresponding to each key event in the entire lifecycle are read according to the start time to end time of the target computing task. The incremental energy consumption values ​​within each post-event analysis window are accumulated, and the accumulated result is used as the incremental energy consumption result of the target computing task.

[0047] In this embodiment, determining the incremental power of the target computing task based on the power change value after interference elimination, the event direction coefficient, and the basic slowly varying power can be achieved through the following steps: The power change value after direction adjustment is determined based on the power change value after interference elimination and the event direction coefficient. The power value after deducting the effect of gradual change is determined based on the power change value after direction adjustment and the basic gradually changing power; The power value after deducting the effects of gradual variation is subjected to non-negative truncation to obtain the incremental power of the target computing power task.

[0048] It should be noted that, in this application, the power change value after eliminating interference represents the amount of power change retained after deducting the interference from other computing power tasks; the event direction coefficient represents a coefficient used to distinguish the direction of power increase or decrease caused by critical events; the power change value after direction adjustment represents the amount of power change after unifying the power change direction according to the type of critical event; the basic slowly varying power represents the amount of slow power change caused by changes in server temperature, fan speed, or low-frequency background operation; the power value after deducting the effect of slowly varying power represents the power value obtained by deducting the server's basic slowly varying power from the power change value after direction adjustment; the non-negative truncation processing represents a truncation processing method that sets power values ​​less than zero to zero and retains non-negative power values; and the incremental power represents the additional server power increment introduced by the target computing power task under critical events.

[0049] In specific implementation, firstly, the power change value and event direction coefficient after interference elimination are read. A positive event direction coefficient is used when the critical event type is the start of the target computing task or entering a high-load phase, and a negative event direction coefficient is used when the critical event type is the end of the target computing task or exiting a high-load phase. The product of the power change value after interference elimination and the event direction coefficient is used as the power change value after direction adjustment. Secondly, the server's basic slowly varying power is read. The slowly varying portion of the power change value after direction adjustment caused by temperature, fan speed, or low-frequency background operation changes is subtracted, and the remaining power value after subtraction is used as the power value after subtracting the effect of slowly varying power. Finally, the power value after subtracting the effect of slowly varying power is input into the non-negative truncation stage. When the power value after subtracting the effect of slowly varying power is less than zero, a zero value is output; when the power value after subtracting the effect of slowly varying power is greater than or equal to zero, the original power value is retained. The output value after non-negative truncation is used as the incremental power of the target computing task, whereby the incremental power of the target computing task is calculated according to the following formula: in, This represents the incremental power of the target computing task. Indicates the event direction coefficient. This represents the power change value after interference is eliminated. Indicates the basic slowly varying power. This indicates non-negative truncation.

[0050] In step S6, a credible task-level energy consumption label is generated using the incremental power and the incremental energy consumption result.

[0051] In this embodiment, generating a reliable task-level energy consumption label based on the incremental power and the incremental energy consumption result can be achieved through the following steps: During energy consumption calculation, evaluation context quality parameters are collected simultaneously. Penalty weights are set for the evaluation context quality parameters based on historical calibration data; The reliability scores of the incremental power and incremental energy consumption results are determined based on the evaluation context quality parameters and the penalty weights. The incremental power, the incremental energy consumption result, and the credibility score are encapsulated in a standardized format to generate a structured task-level energy consumption label.

[0052] It should be noted that, in this application, the evaluation context quality parameters refer to parameters used to reflect the quality of the calculation process of incremental power and incremental energy consumption results; the historical calibration data refer to historical comparison data used to configure the penalty weights corresponding to the evaluation context quality parameters; the penalty weights refer to the weight parameters that reduce the credibility score of different evaluation context quality parameters; the credibility score refers to the numerical value used to measure the credibility of incremental power and incremental energy consumption results; the standardized format refers to the data format used to uniformly encapsulate incremental power, incremental energy consumption results, and credibility scores; and the structured task-level energy consumption label refers to the task-level energy consumption output result carrying incremental power, incremental energy consumption results, and credibility scores.

[0053] In specific implementation, firstly, during the generation of incremental power and incremental energy consumption results, the degree of missing monitoring data, time alignment error, degree of background task change, residual after interference correction, and shared resource complexity are synchronously read and normalized to the same value range. The normalized parameter set is used as the evaluation context quality parameter. Secondly, the energy consumption estimation results and calibration results already saved in the historical calibration data are called, and the correlation ratio between each evaluation context quality parameter and the energy consumption estimation error is read. When historical calibration data is insufficient, the system default settings are used, and the obtained penalty coefficients of each parameter are used as penalty weights. Then, the evaluation context quality parameters and penalty weights are substituted into the inverse proportional credibility evaluation function to calculate the credibility score of the incremental power and incremental energy consumption results. The credibility score is calculated according to the following formula: in, This indicates the credibility score, and The higher the value, the more reliable the task-level energy consumption label; Indicates the degree of data missing; Indicates time alignment error; Indicates the degree of change in the background task; This represents the residual after interference correction; Indicates the complexity of shared resources; , , , and These represent the penalty weights for the corresponding evaluation context quality parameters. Finally, the incremental power, incremental energy consumption results, and credibility scores are written into the same data record according to a standardized format, and the task identifier of the target computing power task and the corresponding event analysis window timestamp are written simultaneously. The encapsulated data record is used as a structured task-level energy consumption label.

[0054] Therefore, this application demonstrates that a reliable task-level energy consumption label can be generated using the incremental power and the incremental energy consumption results. Firstly, by collecting resource usage data and server power data of the target computing task and other computing tasks on the same server, the operating status of the target computing task itself, resource fluctuations of other tasks on the same machine, and overall server power changes can be incorporated into the same energy consumption identification basis, avoiding attribution bias caused by directly inferring the energy consumption of a single task based solely on device-level power data. Secondly, by aligning the resource usage data and server power data over time, resource sequences and power sequences are obtained, allowing data from different sources and with different sampling frequencies to be compared under a unified time reference, reducing energy consumption estimation errors caused by sampling misalignment. Furthermore, by constructing pre-event and post-event analysis windows based on the running events of the target computing task, and performing dual-window baseline differencing on the power sequence, the power difference before and after the start, end, or load change of the target computing task can be used as the identification basis. This method extracts power change characteristics of the target computing task under critical events from the overall server power change, reducing the reliance on fixed-ratio assumptions in traditional resource allocation methods. Furthermore, by identifying interference characteristics through resource changes in other computing tasks, it can explicitly distinguish the impact of sudden increases, decreases, or shared resource competition of other tasks on server power changes, avoiding the misattribution of power fluctuations caused by other tasks to the target computing task. Finally, interference correction and energy increment analysis are performed on the target computing task using the interference characteristics and power change characteristics to obtain the incremental power and energy consumption results of the target computing task. Furthermore, it generates a credible task-level energy consumption label, enabling the task energy consumption results to more clearly reflect the actual contribution of the target computing task to the server power increment. This also facilitates the priority use of high-quality energy consumption labels in subsequent energy consumption prediction, green scheduling, and energy consumption accounting processes, reducing the interference of background task fluctuations on task-level energy consumption management results.

[0055] In summary, the technical solution adopted in this application can accurately identify the energy consumption contribution of a single computing task in a multi-task shared operation environment in a data center, thereby improving the reliability of task-level energy consumption assessment results.

[0056] Example 2 This application provides a data center multi-task computing power task energy consumption identification and correction system, referencing... Figure 2As shown in the figure, this is a block structure diagram of a data center multi-task computing power task energy consumption identification and correction system according to this embodiment of the application. The energy consumption identification and correction system includes: The resource power acquisition module 100 is used to collect resource usage data and server power data of the target computing power task and other computing power tasks on the same server; The time alignment sequence module 200 is used to perform time alignment on the collected resource usage data and server power data to obtain resource sequences and power sequences; The dual-window baseline difference module 300 is used to construct a pre-event analysis window and a post-event analysis window based on the running events of the target computing power task, and to perform dual-window baseline difference on the power sequence based on the pre-event analysis window and the post-event analysis window to obtain the power change characteristics of the target computing power task under key events. The interference feature identification module 400 is used to identify the interference feature information of other computing power tasks on the server power change by the resource change amount of other computing power tasks in the resource sequence; The interference correction increment module 500 is used to perform interference correction and energy increment analysis on the target computing power task through the interference feature information and the power change feature, so as to obtain the incremental power and incremental energy consumption results of the target computing power task. The energy consumption tag generation module 600 is used to generate a reliable task-level energy consumption tag based on the incremental power and the incremental energy consumption result.

[0057] 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 process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0058] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0059] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A method for identifying and correcting energy consumption of multi-task computing power tasks in a data center, characterized in that, Includes the following steps: Collect resource usage data and server power data of the target computing power task and other computing power tasks on the same server; The collected resource usage data and server power data are time-aligned to obtain resource sequences and power sequences; Based on the running events of the target computing power task, a pre-event analysis window and a post-event analysis window are constructed. Based on the pre-event analysis window and the post-event analysis window, a dual-window baseline difference is performed on the power sequence to obtain the power change characteristics of the target computing power task under key events. The interference characteristics of other computing power tasks on server power changes can be identified by the resource changes of other computing power tasks in the resource sequence. Interference correction and energy increment analysis are performed on the target computing power task using the interference feature information and the power change feature to obtain the incremental power and incremental energy consumption results of the target computing power task. Generate a credible task-level energy consumption label using the incremental power and the incremental energy consumption results.

2. The data center multi-task computing power task energy consumption identification and correction method as described in claim 1, characterized in that, Time alignment is performed on the collected resource usage data and server power data to obtain the resource sequence and power sequence, which specifically include: Set a globally uniform time window length to divide the continuous system running time axis into a series of non-overlapping discrete time windows; Using the discrete time window as the alignment reference, the collected server power data is aggregated and calculated to obtain the average server power value of each time window in the discrete time window sequence, and a power sequence is formed. Using the discrete time window as the alignment reference, the collected resource usage data is aggregated and calculated to obtain the average resource usage vector of each computing power task in each time window within the discrete time window sequence, and a resource sequence is formed.

3. The data center multi-task computing power task energy consumption identification and correction method as described in claim 1, characterized in that, The pre-event analysis window and post-event analysis window are constructed based on the runtime events of the target computing power task, specifically including: Set a preset skip time based on the time point when the key event of the target computing power task occurs; Before the skipped time, a first preset time period is captured forward as the pre-event analysis window; After the skipped time, a second preset time period is extracted as the post-event analysis window.

4. The data center multi-task computing power task energy consumption identification and correction method as described in claim 1, characterized in that, Based on the pre-event analysis window and the post-event analysis window, a dual-window baseline difference is performed on the power sequence to obtain the power change characteristics of the target computing power task under critical events, specifically including: The power sample values ​​within the pre-event analysis window are extracted from the power sequence to obtain the average power of the pre-event analysis window; The power sample values ​​within the post-event analysis window are extracted from the power sequence to obtain the average power of the post-event analysis window. The power change characteristics of the target computing power task under critical events are determined based on the difference between the average power in the post-event analysis window and the average power in the pre-event analysis window.

5. The data center multi-task computing power task energy consumption identification and correction method as described in claim 4, characterized in that, The power change characteristics of the target computing power task under critical events are determined based on the difference between the average power in the post-event analysis window and the average power in the pre-event analysis window. Specifically, this includes: Read the average server power values ​​in the pre-event analysis window and the post-event analysis window from the power sequence, and perform effective sampling and filtering on the average server power values ​​to obtain effective power sampling points; A preset deviation range is set based on the hardware power reading error and the power deviation within the historical stable operation window, and the effective power sampling points are filtered for steady-state sampling based on the preset deviation range to obtain steady-state power sampling points; The validity weight is determined based on the number of valid power sampling points in the steady-state power sampling points and the number of valid power sampling points in the pre-event analysis window and the post-event analysis window. The difference between the average power of the post-event analysis window and the average power of the pre-event analysis window is corrected by the validity weight, thereby obtaining the power change characteristics of the target computing power task under critical events.

6. The data center multi-task computing power task energy consumption identification and correction method as described in claim 1, characterized in that, The interference characteristics of other computing power tasks on server power changes can be identified by the resource changes of other computing power tasks in the resource sequence. Specifically, this includes: Extract the resource usage of other computing power tasks within the pre-event analysis window and the post-event analysis window from the resource sequence, and then determine the resource changes of other computing power tasks in the resource sequence; The degree of resource variation in the background environment is determined by combining the resource changes of other computing power tasks with the weights of the impact of different resource types on power changes. When the resource change rate is greater than the stability threshold, it is determined that other computing power tasks have changed significantly, and the interference characteristics of other computing power tasks on server power changes are determined based on the resource change amount of other computing power tasks and the pre-built resource power mapping model.

7. The data center multi-task computing power task energy consumption identification and correction method as described in claim 1, characterized in that, By performing interference correction and energy increment analysis on the target computing task using the interference characteristic information and the power change characteristics, the incremental power and incremental energy consumption results of the target computing task are obtained, specifically including: The event direction coefficient is set according to the key event type of the target computing power task, and the basic slowly varying power of the server is evaluated; The power change feature is subjected to interference subtraction processing using the interference feature information to obtain the power change value after interference elimination; The incremental power of the target computing power task is determined based on the power change value after interference elimination, the event direction coefficient, and the basic slowly varying power. The incremental energy consumption value within the post-event analysis window is determined based on the incremental power and the duration of the post-event analysis window. Throughout the entire lifecycle of the target computing power task, the incremental energy consumption value within the post-event analysis window is accumulated to obtain the incremental energy consumption result of the target computing power task.

8. The data center multi-task computing power task energy consumption identification and correction method as described in claim 7, characterized in that, Determining the incremental power of the target computing task based on the power change value after interference elimination, the event direction coefficient, and the basic slowly varying power specifically includes: The power change value after direction adjustment is determined based on the power change value after interference elimination and the event direction coefficient. The power value after deducting the effect of gradual change is determined based on the power change value after direction adjustment and the basic gradually changing power; The power value after deducting the effects of gradual variation is subjected to non-negative truncation to obtain the incremental power of the target computing power task.

9. The data center multi-task computing power task energy consumption identification and correction method as described in claim 1, characterized in that, Generating a reliable task-level energy consumption label based on the incremental power and the incremental energy consumption results specifically includes: During energy consumption calculation, evaluation context quality parameters are collected simultaneously. Penalty weights are set for the evaluation context quality parameters based on historical calibration data; The reliability scores of the incremental power and incremental energy consumption results are determined based on the evaluation context quality parameters and the penalty weights. The incremental power, the incremental energy consumption result, and the credibility score are encapsulated in a standardized format to generate a structured task-level energy consumption label.

10. A data center multi-task computing power task energy consumption identification and correction system, used to execute the data center multi-task computing power task energy consumption identification and correction method as described in any one of claims 1 to 9, characterized in that, The energy consumption identification and correction system includes: The resource power acquisition module is used to collect resource usage data and server power data of the target computing power task and other computing power tasks on the same server; The time-aligned sequence module is used to time-align the collected resource usage data and server power data to obtain resource sequences and power sequences. The dual-window baseline difference module is used to construct a pre-event analysis window and a post-event analysis window based on the running events of the target computing power task, and to perform dual-window baseline difference on the power sequence based on the pre-event analysis window and the post-event analysis window to obtain the power change characteristics of the target computing power task under key events. The interference feature identification module is used to identify the interference feature information of other computing power tasks on the server power change by the resource change amount of other computing power tasks in the resource sequence; The interference correction increment module is used to perform interference correction and energy increment analysis on the target computing power task through the interference feature information and the power change feature, so as to obtain the incremental power and incremental energy consumption results of the target computing power task. An energy consumption tag generation module is used to generate a reliable task-level energy consumption tag based on the incremental power and the incremental energy consumption result.