A computing platform resource integrated management method and system

By adjusting dynamic observation controls to adapt to changes in multi-dimensional computing time sequence data, the accuracy problem of traditional resource management methods in the face of rapid changing loads is solved, more accurate resource estimates and scheduling is achieved, and system efficiency is improved. Quality of service.

CN118760533BActive Publication Date: 2024-11-19CHENGDU SUPERCOMPUTING CENT OPERATION MANAGEMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411253974.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-09
Publication Date
2024-11-19
Estimated Expiration
2044-09-09

AI Technical Summary

Technical Problem

Traditional computing resource management methods are difficult to adapt to rapidly changing workload models, especially when facing periodic or sudden tasks, resulting in insufficient resource allocation, affecting service quality and system efficiency.

Method used

By obtaining a multi-dimensional computing time series data collection, defining and adjusting dynamic observation controls, extracting and extending the data range, new observation controls are built to capture resource usage changes more accurately and achieve more accurate resource estimates and scheduling.

Benefits of technology

It improves the accuracy of computing power resource estimates, enhances the inference ability of multi-dimensional computing power timing data, ensures the reasonable allocation and optimization of resources, and meets the ever-changing workload needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118760533B_ABST
    Figure CN118760533B_ABST
Patent Text Reader

Abstract

The present application provides a method and system for integrated management of computing power platform resources. For multiple first dynamic observation controls corresponding to a random data subset in a past multi-dimensional computing power time series data set, each first dynamic observation control is centrally adjusted and the extraction range is extended to obtain a corresponding second dynamic observation control. Data extraction is performed in each second dynamic observation control to obtain a multi-dimensional computing power time series data set after extraction. Reasoning tasks are performed based on the extracted multi-dimensional computing power time series data set. The present application is based on changing the position and coverage of a dynamic observation control that is locked in the range, and obtains a dynamic observation control with an extendable coverage range, which prevents ignoring the boundary data of the dynamic observation control and improves the detailed knowledge in the dynamic observation control. The extension of the coverage range makes the semantic consistency and fluency of the dynamic observation control consistent, and improves the reasoning ability of the multi-dimensional computing power time series data set.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a computing platform resource integration management method and system. Background Art

[0002] With the development of cloud computing and big data technologies, computing resource management in data centers and distributed computing environments has become increasingly important. Effective management and scheduling of computing resources are essential to ensuring quality of service (QoS), improving system efficiency, and reducing operating costs. Traditional computing resource management methods often rely on static resource allocation strategies, which perform poorly in the face of changing workloads. In addition, with the increasing application of machine learning and artificial intelligence technologies, the demand for computing resources has become more diverse and dynamic, which places higher demands on resource management. Existing computing resource management solutions mainly focus on how to optimize static resource allocation and adjust resource allocation through preset rules or simple feedback mechanisms. However, this management method is often difficult to adapt to rapidly changing workload patterns, especially when faced with periodic or bursty tasks. Traditional resource management algorithms usually lack sufficient flexibility to capture subtle changes in data streams, resulting in inaccurate resource allocation, which in turn affects the overall service quality and system operation efficiency. Summary of the invention

[0003] In view of this, the embodiment of the present application at least provides a computing platform resource integration management method and system. The technical solution of the present application is implemented as follows:

[0004] On the one hand, the present application provides a method for integrated management of computing power platform resources, the method comprising: obtaining multiple first dynamic observation controls corresponding to each data subset in a past multi-dimensional computing power time series data set, wherein the multiple first dynamic observation controls are used to extract each data subset, while the extraction range remains unchanged; wherein a data subset is a time series data subset consisting of computing power resource data corresponding to one dimension; each first dynamic observation control corresponding to each data subset is moved and the extraction range is extended to construct each second dynamic observation control corresponding to each data subset; in each second dynamic observation control corresponding to each data subset, data is extracted from the past multi-dimensional computing power time series data set to obtain a multi-dimensional computing power time series data set after extraction, and the number of extractions in each second dynamic observation control is consistent with the number of extractions in each first dynamic observation control; based on the multi-dimensional computing power time series data set after extraction, a subsequent multi-dimensional computing power time series data set is inferred to perform computing power planning based on the subsequent multi-dimensional computing power time series data set.

[0005] On the other hand, the present application provides a computer system, including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the above method when executing the program.

[0006] The beneficial effects of the present application include: for multiple first dynamic observation controls corresponding to a random data subset in a past multi-dimensional computing power time series data set, the present application performs center adjustment and extraction range extension on each first dynamic observation control to obtain a corresponding second dynamic observation control. Data extraction is performed in each second dynamic observation control to obtain a multi-dimensional computing power time series data set after extraction. Reasoning tasks are performed based on the extracted multi-dimensional computing power time series data set. The present application is based on changing the position and coverage of a dynamic observation control with a locked range to obtain a dynamic observation control with an extendable coverage range, thereby preventing the boundary data of the dynamic observation control from being ignored and improving the detailed knowledge in the dynamic observation control. The extension of the coverage range makes the semantic consistency and fluency of the dynamic observation control better, thereby improving the reasoning ability of the multi-dimensional computing power time series data set, thereby improving the accuracy of computing power resource estimation and helping to carry out accurate and reasonable computing power scheduling planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 A schematic diagram of the implementation flow of a computing power platform resource integration management method provided in an embodiment of the present application.

[0008] Figure 2 A hardware entity schematic diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0009] The embodiment of the present application provides a method for integrated management of computing platform resources, which can be executed by a processor of a computer system in a computing platform deployed in the cloud, wherein the computer system can refer to a server. Figure 1 A schematic diagram of the implementation process of a computing power platform resource integration management method provided in an embodiment of the present application is as follows: Figure 1 As shown, the method includes:

[0010] Step S100: Obtain multiple first dynamic observation controls corresponding to each data subset in the past multi-dimensional computing power time series data set. The multiple first dynamic observation controls are dynamic observation controls used to extract each data subset while keeping the extraction range unchanged; wherein, a data subset is a time series data subset consisting of computing power resource data corresponding to one dimension.

[0011] When the computer system executes step S100, it first identifies and obtains a set of past multi-dimensional computing power time series data. The set contains computing power resource data of multiple dimensions, and each dimension corresponds to a data subset. For example, on a computing power platform, the computer system collects data on multiple resources such as CPU utilization, memory utilization, storage I / O operation frequency, and each resource constitutes an independent data subset. For ease of explanation, it is assumed that there are three-dimensional data subsets: CPU utilization data subset, memory utilization data subset, and storage I / O operation frequency data subset. For each data subset, the computer system defines multiple first dynamic observation controls. These controls are similar to a movable observation window, which is used to extract data within a specific range from the corresponding data subset. For example, in the CPU utilization data subset, the computer system defines an initial first dynamic observation control, which specifies the data extraction range within a period of time, such as 9:00 to 10:00 in a day. At this time, the extraction range of the control remains unchanged, which means that it always covers data within the same time period. The specific form of the control can be an array, such as the array representing CPU utilization is [90, 92, 91, 93, 94, 92, 91,90, 89, 88], corresponding to the utilization data every 6 minutes from 9:00 to 10:00. The computer system continues to define multiple such first dynamic observation controls for each data subset. Each control covers a different time period, but the extraction range is the same. For example, in addition to the control covering 9:00 to 10:00, there is another control covering the time period from 10:00 to 11:00. These controls work together on their respective data subsets to ensure the consistency and effectiveness of the data extraction process. In this way, the computer system can efficiently extract useful information from the past multi-dimensional computing power time series data set, providing a solid data foundation for subsequent steps. For example, when performing computing power resource scheduling, by analyzing the resource usage in different time periods in these data subsets, the computer system can better predict the resource demand in a specific time period in the future, thereby optimizing resource allocation and improving the operating efficiency of the entire computing power platform.

[0012] Step S200: Move and extend the extraction range of each first dynamic observation control corresponding to each data subset to construct each second dynamic observation control corresponding to each data subset.

[0013] When the computer system executes step S200, each first dynamic observation control corresponding to each data subset is moved and the extraction range is extended to construct each second dynamic observation control corresponding to each data subset. This process includes multiple sub-steps. Specifically, the computer system first selects a data subset, such as a CPU utilization data subset, and then performs a move and extension operation on each first dynamic observation control in the subset.

[0014] For example, suppose that the computer system has an initial first dynamic observation control in the CPU utilization data subset, and the initial center position of the control is located in the middle of the time period from 9:00 to 10:00, that is, 9:30. The initial coverage range of the control is one hour, that is, 9:00 to 10:00. Next, the computer system adds an adjustment amount to the center position of the control to obtain a new center position. Assuming the adjustment amount is +15 minutes, the new center position becomes 9:45. After that, the computer system obtains half of the initial control coverage range, that is, half of 30 minutes, 15 minutes, as the initial control coverage range. The initial control coverage range is spliced ​​with the forward extension range of the control center. Assuming the forward extension range is +10 minutes, the new forward control coverage range is 25 minutes. Then, the computer system subtracts the forward control coverage range from the new center position to obtain a new starting position of 9:50. Similarly, the computer system splices the initial control coverage range with the control center backward extension range. Assuming the backward extension range is +10 minutes, the new backward control coverage range is also 25 minutes. Splicing the new center position with the backward control coverage range, the new end position is 10:40. Therefore, the coverage range of the new second dynamic observation control is 9:50 to 10:40, and the center position is 10:15.

[0015] In order to obtain the adjustment amount corresponding to the first dynamic observation control, the computer system performs implicit representation mining to obtain multiple first implicit array representations. For example, for the first control in the CPU utilization data subset, the computer system may obtain an implicit array representation [0.8, 0.9, 0.7, 0.85, 0.92]. Subsequently, the computer system multiplies these implicit array representations with their respective adjustment weights to obtain a second implicit array representation. The adjustment weights are parameters obtained by deep learning model training, such as the weight vector [0.1, 0.2, -0.1, 0.3, -0.2]. After multiplication, a new implicit array representation [0.08, 0.18, -0.07, 0.255, -0.184] is obtained. Based on these second implicit array representations, the computer system obtains the adjustment amount of the first dynamic observation control, for example, the adjustment amount is +15 minutes.

[0016] The computer system obtains an adjustment constraint based on the number of extractions in the first dynamic observation control. The adjustment constraint is used to limit the adjustment span of the center of the control. For example, if the number of extractions is 10, the adjustment span cannot exceed one tenth of the time length of the entire data subset. For each second implicit array representation, the computer system obtains a corresponding adjustment amount based on the array representation and the adjustment constraint.

[0017] Step S300: In each second dynamic observation control corresponding to each data subset, data is extracted from the past multi-dimensional computing power time series data set to obtain the extracted multi-dimensional computing power time series data set. The number of extractions in each second dynamic observation control is consistent with the number of extractions in each first dynamic observation control.

[0018] When the computer system executes step S300, for each second dynamic observation control corresponding to each data subset, the past multi-dimensional computing power time series data set is extracted, and finally the extracted multi-dimensional computing power time series data set is obtained. In this process, the number of extractions of each second dynamic observation control is consistent with the number of extractions of the previous first dynamic observation control.

[0019] Taking the CPU utilization data subset as an example, assume that the computer system has built several second dynamic observation controls. The starting position of one of the second dynamic observation controls is 9:50, the ending position is 10:40, and the number of extractions is 10. This means that the computer system needs to evenly extract 10 data points within this 50-minute time window. To this end, the computer system first calculates the extraction step, which is the control coverage divided by the number of extractions. In this example, the control coverage is 50 minutes, the number of extractions is 10, and the extraction step is 5 minutes.

[0020] The computer system extracts data in each second dynamic observation control according to the extraction step size. For the extraction position whose position identifier is an integer, the observation value corresponding to the position is directly obtained. For example, at the first position starting at 9:50, the CPU utilization data of 9:50 is obtained. For the extraction position whose position identifier is not an integer, the computer system constructs the observation value of the position through the upsampling method based on the observation values ​​of adjacent time points. For example, if the CPU utilization data of 9:52 needs to be obtained, and there are only data of 9:50 and 10:00 in the original data, the computer system uses the interpolation method to estimate the data of 9:52.

[0021] Specifically, if you need to obtain the CPU utilization data at 9:52, the computer system determines that the time points adjacent to 9:52 are 9:50 and 10:00. Get the CPU utilization observations at these two time points, assuming that the observation value at 9:50 is 85% and the observation value at 10:00 is 88%. Calculate the distance weight based on the interval between these two time points. In this example, 9:52 is 2 minutes away from 9:50 and 8 minutes away from 10:00, so the distance weights are 8 / 10=0.8 and / 10=0.2. Then, the distance weight is multiplied by the observation value at the corresponding time point to obtain the observation value of 9:52, which is 85%×0.8+88%×0.2=85.6.

[0022] After the computer system completes data extraction at all extraction locations, it combines multiple observation values ​​extracted from each second dynamic observation control to form a post-extraction multi-dimensional computing power time series data set. For example, for a subset of CPU utilization data, the extracted data set can be represented as a three-dimensional tensor, where the first dimension represents the observation value component (i.e., the specific CPU utilization value), the second dimension represents the time component (i.e., the extraction time point), and the third dimension represents the dynamic observation control component (i.e., different observation controls). This three-dimensional tensor can be used for further analysis and modeling to predict future computing power requirements.

[0023] In this way, the computer system can effectively extract meaningful data fragments from past multi-dimensional computing power time series data sets, providing a solid foundation for further data processing and analysis.

[0024] Step S400: Based on the extracted multi-dimensional computing power time series data set, inferring a subsequent multi-dimensional computing power time series data set, so as to perform computing power planning based on the subsequent multi-dimensional computing power time series data set.

[0025] When the computer system executes step S400, based on the extracted multi-dimensional computing power time series data set, it infers the subsequent multi-dimensional computing power time series data set so as to perform computing power planning based on the set. This process involves analyzing the correlation of the extracted data set and obtaining the subsequent data set for prediction through a series of transformations.

[0026] The computer system first processes the extracted multi-dimensional computing power time series data set to obtain a multi-dimensional computing power time series data set after correlation determination. This set contains correlation information in the extracted data set, including short-term correlation in dynamic observation controls, long-term correlation across controls, and correlation across observation values. For example, the computer system processes the extracted data set to obtain a new three-dimensional tensor, in which each dimension represents an observation value component, a time component, and a dynamic observation control component. To illustrate this, assume that the extracted data set contains data in three dimensions: CPU utilization, memory usage, and storage I / O operation frequency, and each dimension has data at multiple time points and multiple dynamic observation controls.

[0027] To process this three-dimensional tensor, the computer system uses multiple fusion components. Each fusion component contains three feedforward neural network layers: channel fusion unit, dynamic observation control fusion unit, and variable fusion unit. These three components are responsible for obtaining information about short-term correlations in dynamic observation controls, information about long-term correlations across controls, and information about correlations across observations. For example, the channel fusion unit processes the dynamic observation control component in the three-dimensional tensor to capture short-term correlations. Assume a three-dimensional tensor is T, where T ijkRepresents the value on the observation value component i, time component j, and dynamic observation control component k. The channel fusion unit processes T through a feedforward neural network layer to obtain a new three-dimensional tensor T', where T' ijk Reflects T ijk Short-term correlations on dynamically observed control components.

[0028] Next, the computer system re-dimensionalizes the processed three-dimensional tensor to expand the dynamic observation control component and the time component. For example, assuming that after T' is processed by the channel fusion unit, the computer system inputs it into the dimension re-dimensionalization operator to obtain a transitional multi-dimensional computing power time series data set T''. This transitional set T'' is further processed by the projection operator to infer the subsequent multi-dimensional computing power time series data set T''. The projection operator is a mapping function used to project the transitional multi-dimensional computing power time series data set T'' into a preset range. For example, the projection operator can be a linear function in the form of , where w and b are parameters obtained based on deep learning model training. Through the above steps, the computer system obtains the subsequent multi-dimensional computing power time series data set T''', which contains the forecast information of future computing power resource demand. For example, suppose the CPU utilization data forecast in T''' shows that the CPU utilization will fluctuate between 80% and 90% in the next hour. The memory utilization forecast shows that the memory utilization will fluctuate between 75% and 85%. The storage I / O operation frequency forecast shows that the operation frequency will fluctuate between 5000 times / second and 6000 times / second. Based on this subsequent multi-dimensional computing power time series data set, the computer system can perform computing power planning and reasonably allocate resources to meet future computing power needs. For example, if it is predicted that the CPU utilization will be high in the future, the computer system can increase the allocation of CPU resources in advance, or start a backup server to cope with the upcoming workload. In this way, the computer system ensures the efficient operation of the computing power platform and the optimal utilization of resources.

[0029] As an implementation mode, step S200, moving each first dynamic observation control corresponding to each data subset and extending the extraction range to construct each second dynamic observation control corresponding to each data subset, may include:

[0030] Step S210: For a random first dynamic observation control corresponding to each data subset, add the control center of the random first dynamic observation control and the adjustment amount corresponding to the random first dynamic observation control to obtain the control center of the second dynamic observation control.

[0031] When the computer system executes step S210, for a random first dynamic observation control corresponding to each data subset, the control center of the first dynamic observation control is added to the adjustment amount corresponding to the first dynamic observation control to obtain the control center of the second dynamic observation control. This process ensures that the new observation control can be moved on the original basis to meet a wider range of data extraction needs.

[0032] Taking the CPU utilization data subset as an example, the computer system first selects a first dynamic observation control. Assume that the initial center position of the control is at 9:30, and the time range covered is from 9:00 to 10:00. This means that the center position of the control is 9:30. Next, the computer system needs to determine the adjustment amount of the control. The adjustment amount is a parameter obtained through implicit representation mining and deep learning models, which is used to indicate the moving direction and distance of the center position of the control. For example, assume that the first implicit array representation obtained through implicit representation mining is [0.8, 0.9, 0.7, 0.85, 0.92], and the adjustment weight is [0.1, 0.2,-0.1, 0.3, -0.2]. The computer system performs a dot product operation on these two vectors to obtain a second implicit array representation [0.08, 0.18, -0.07, 0.255, -0.184]. Through further processing, the computer system finally determines that the adjustment amount is +15 minutes.

[0033] Specifically, the computer system adds the control center position of the first dynamic observation control to the adjustment amount. In this example, the control center position is 9:30 and the adjustment amount is +15 minutes. Therefore, the new control center position is 9:45. In this way, the computer system constructs the control center of the second dynamic observation control.

[0034] To further illustrate this process, consider a specific example. Assume that the timestamp of the CPU utilization data subset is recorded every ten minutes, starting at 9:00 and ending at 10:00. The initial first dynamic observation control covers the time period from 9:00 to 10:00, and the control center position is 9:30. The computer system obtains an adjustment amount of +15 minutes by implicitly representing mining and adjusting weight calculations. Add 15 minutes to 9:30, and the new control center position is 9:45. Through the above steps, the computer system adjusts the position of the first dynamic observation control and obtains a new control center of the second dynamic observation control. This adjustment enables the computer system to extract data more flexibly in subsequent steps, thereby better capturing changes in resource usage patterns and providing more accurate data support for the planning of computing resources. For example, in actual applications, the computer system can adjust the position of the observation control in this way to better track the changing trend of CPU utilization and allocate and schedule resources accordingly.

[0035] Step S220: Obtain half of the control coverage of the random first dynamic observation control to obtain the initial control coverage.

[0036] When the computer system executes step S220, it obtains half of the control coverage range of the first dynamic observation control to obtain the initial control coverage range. This step lays the foundation for the subsequent control range adjustment.

[0037] Taking the CPU utilization data subset as an example, assume that the computer system selects a first dynamic observation control whose initial coverage is one hour, that is, from 9:00 to 10:00. This means that the total time covered by the control is 60 minutes. In order to obtain the initial control coverage, the computer system needs to calculate half of the control coverage. Therefore, the initial control coverage is half of 60 minutes, that is, 30 minutes.

[0038] Specifically, the computer system first determines the coverage of the first dynamic observation control. In this example, the control covers the time period from 9:00 to 10:00, with a total time of 60 minutes. Next, the computer system calculates half of the control coverage, which is 30 minutes. This initial control coverage will be used to adjust the control range in the next step. To further illustrate this process, assume that the computer system is processing a data set containing multiple time points. For example, the timestamps of the CPU utilization data subset are recorded every ten minutes, starting at 9:00 and ending at 10:00. The initial first dynamic observation control covers the time period from 9:00 to 10:00, with a total time of 60 minutes. The computer system calculates half of the control coverage, which is 30 minutes, as the initial control coverage.

[0039] In this way, the computer system can provide a benchmark for the control range adjustment in the subsequent steps. For example, in the subsequent steps, the computer system will use this initial control coverage to determine the forward and backward control coverage of the new second dynamic observation control.

[0040] In specific implementation, the computer system determines the extension range of the control by implicit representation mining and extension weight calculation. For example, assume that the first implicit array representation obtained by implicit representation mining is [0.8, 0.9, 0.7, 0.85, 0.92], and the extension weight is [0.1, 0.2, -0.1, 0.3, -0.2]. The computer system performs a dot product operation on these two vectors to obtain a third implicit array representation [0.08, 0.18, -0.07, 0.255, -0.184]. Through further processing, the computer system finally determines that the extension range is +10 minutes. In summary, in step S220, the computer system obtains the initial control coverage range by calculating half of the control coverage range of the first dynamic observation control. This step provides the necessary benchmark for adjusting the control range in subsequent steps. In this way, the computer system can more flexibly adjust the control range in subsequent steps to adapt to different data extraction requirements, thereby better capturing changes in resource usage patterns and providing more accurate data support for the planning of computing resources.

[0041] Step S230: Concatenate the initial control coverage range with the center forward extension range corresponding to the random first dynamic observation control to obtain the forward control coverage range of the second dynamic observation control; subtract the forward control coverage range of the second dynamic observation control from the control center of the second dynamic observation control to obtain the starting position of the second dynamic observation control.

[0042] When the computer system executes step S230, the initial control coverage range is spliced ​​with the center forward extension range corresponding to the first dynamic observation control to obtain the forward control coverage range of the second dynamic observation control. Subsequently, the computer system subtracts the forward control coverage range from the control center of the second dynamic observation control to obtain the starting position of the second dynamic observation control.

[0043] Taking the CPU utilization data subset as an example, assume that the computer system selects a first dynamic observation control, and the initial coverage of the control is one hour, that is, from 9:00 to 10:00. The center of the control is located at 9:30. In step S220, the computer system has calculated that the initial control coverage is 30 minutes. Next, the computer system needs to determine the center forward extension range of the first dynamic observation control. Assume that through implicit representation mining and extension weight calculation, the center forward extension range is +10 minutes. Specifically, the computer system first splices the initial control coverage (30 minutes) with the center forward extension range (+10 minutes) to obtain the forward control coverage of the second dynamic observation control. Therefore, the new forward control coverage is 30 minutes plus 10 minutes, a total of 40 minutes. Next, the computer system calculates the starting position of the second dynamic observation control. Since the new control center position has been adjusted to 9:45 in step S210, the computer system subtracts the forward control coverage of 40 minutes from the new control center position 9:45 to obtain a new starting position of 9:05. Through the above steps, the computer system constructs a new forward control coverage range of the second dynamic observation control and determines a new starting position. This process ensures that the new observation control can be extended forward on the original basis to cover a larger data range. To further illustrate this process, consider the specific calculation steps. Assume that the initial control coverage range is 30 minutes and the center forward extension range is 10 minutes. The computer system splices the two and obtains a forward control coverage range of 40 minutes. Assuming that the new control center position is 9:45, the starting position is 9:45 minus 40 minutes, that is, 9:05.

[0044] In this way, the computer system can determine the starting position of the new observation control based on the adjusted center position of the control and the forward extension range. For example, in an actual application scenario, the computer system needs to monitor the CPU utilization data. The initial first dynamic observation control covers the time period from 9:00 to 10:00, with the center at 9:30. Through step S230, the computer system adjusts the position of the observation control and extends the observation range forward, so that the new second dynamic observation control covers the time period from 9:05 to a certain end position, with the center at 9:45. Through this method of dynamically adjusting and expanding the observation control, the computer system can more flexibly capture resource usage patterns in different time periods, thereby providing more accurate data support for the planning of computing resources.

[0045] Step S240: splice the initial control coverage range with the center backward extension range corresponding to the random first dynamic observation control to obtain the backward control coverage range of the second dynamic observation control; splice the control center of the second dynamic observation control with the backward control coverage range of the second dynamic observation control to obtain the end position of the second dynamic observation control.

[0046] When the computer system executes step S240, the initial control coverage is spliced ​​with the center backward extension range corresponding to the first dynamic observation control to obtain the backward control coverage of the second dynamic observation control. Subsequently, the computer system splices the control center of the second dynamic observation control with the backward control coverage to obtain the end position of the second dynamic observation control.

[0047] Taking the CPU utilization data subset as an example, assume that the computer system has determined that the new control center position is 9:45 and the initial control coverage is 30 minutes. In step S230, the computer system has determined the new starting position to be 9:05. Next, the computer system needs to determine the center backward extension range of the first dynamic observation control. Assume that through implicit representation mining and extension weight calculation, the center backward extension range is +10 minutes. Specifically, the computer system splices the initial control coverage (30 minutes) with the center backward extension range (+10 minutes) to obtain the backward control coverage of the second dynamic observation control. Therefore, the new backward control coverage is 30 minutes plus 10 minutes, a total of 40 minutes.

[0048] Next, the computer system calculates the end position of the second dynamic observation control. Since the new control center position has been adjusted to 9:45, the computer system adds the new control center position 9:45 to the backward control coverage range of 40 minutes to obtain a new end position of 10:25.

[0049] Through the above steps, the computer system constructs the backward control coverage of the new second dynamic observation control and determines the new end position. This process ensures that the new observation control can extend backward on the original basis to cover more data range.

[0050] To further illustrate this process, consider the specific calculation steps. Assume that the initial control coverage is 30 minutes and the center backward extension range is 10 minutes. The computer system splices the two and obtains a backward control coverage range of 40 minutes. Assume that the new control center position is 9:45, then the end position is 9:45 plus 40 minutes, that is, 10:25.

[0051] In this way, the computer system can determine the end position of the new observation control based on the adjusted center position of the control and the backward extension range. For example, in an actual application scenario, the computer system needs to monitor the CPU utilization data. The initial first dynamic observation control covers the time period from 9:00 to 10:00, with the center at 9:30. Through step S240, the computer system adjusts the position of the observation control and extends the observation range backward, so that the new second dynamic observation control covers the time period from 9:05 to 10:25, with the center at 9:45.

[0052] By dynamically adjusting and expanding observation controls, computer systems can more flexibly capture resource usage patterns in different time periods, thereby providing more accurate data support for computing resource planning. For example, by adjusting the coverage and center location of controls, computer systems can better track the changing trends of CPU utilization and allocate and schedule resources accordingly. In this way, computer systems can more effectively manage and optimize computing resources to meet changing workload requirements.

[0053] As an implementation mode, the method further includes a process of obtaining an adjustment amount corresponding to the first dynamic observation control, including:

[0054] Step S2101: performing implicit representation mining on the past multi-dimensional computing power time series data set to obtain a plurality of first implicit array representations, wherein the plurality of first implicit array representations have a corresponding relationship with the first dynamic observation control in the past multi-dimensional computing power time series data set;

[0055] Step S2102: Multiply the plurality of first implicit array representations by the corresponding adjustment weights to obtain a plurality of second implicit array representations, wherein the adjustment weights are parameters obtained based on deep training;

[0056] Step S2103: Obtain the adjustment amount of the first dynamic observation control in the past multi-dimensional computing power timing data set based on multiple second implicit array representations.

[0057] When obtaining the adjustment amount corresponding to the first dynamic observation control, the computer system performs the following steps: step S2101, step S2102 and step S2103. This process includes implicit representation mining, product calculation and determination of the adjustment amount to ensure that the first dynamic observation control can be properly adjusted according to historical data.

[0058] First, in step S2101, the computer system performs implicit representation mining on the past multi-dimensional computing power time series data set to obtain multiple first implicit array representations. These first implicit array representations have a one-to-one correspondence with the first dynamic observation controls in the past multi-dimensional computing power time series data set. For example, in a typical computing power platform resource integration management scenario, the computer system collects data in multiple dimensions such as CPU utilization, memory usage, and storage I / O operation frequency in the past month. The data of each dimension forms a data subset. For the CPU utilization data subset, the computer system defines multiple first dynamic observation controls, each of which covers data over a period of time. The computer system performs implicit representation mining on these data subsets to obtain multiple first implicit array representations. Assume that for a first dynamic observation control in the CPU utilization data subset, the computer system obtains a first implicit array representation, such as [0.8, 0.9, 0.7, 0.85, 0.92]. Next, in step S2102, the computer system multiplies the multiple first implicit array representations with the corresponding adjustment weights to obtain multiple second implicit array representations. These adjustment weights are parameters obtained based on deep learning model training. The computer system pre-trains a deep learning model that is used to learn how to adjust the position of the control based on the results of implicit representation mining. For example, assume that the adjustment weights are [0.1, 0.2, -0.1, 0.3, -0.2]. The computer system performs a dot product operation on the first implicit array representation [0.8, 0.9, 0.7, 0.85, 0.92] and the adjustment weights [0.1, 0.2, -0.1, 0.3, -0.2] to obtain the second implicit array representation [0.08, 0.18, -0.07, 0.255, -0.184].

[0059] Finally, in step S2103, the computer system obtains the adjustment amount of the first dynamic observation control in the past multi-dimensional computing power time series data set based on multiple second implicit array representations. Specifically, the computer system determines the adjustment amount by comprehensively analyzing the second implicit array representation. For example, for a first dynamic observation control in the CPU utilization data subset, the computer system determines the adjustment amount to be +15 minutes based on the second implicit array representation [0.08, 0.18, -0.07, 0.255, -0.184]. This means that the center position of the control will move forward 15 minutes.

[0060] Through the above steps, the computer system obtains the adjustment amount of the first dynamic observation control. This process ensures that the control can be appropriately adjusted according to historical data, so as to better capture changes in resource usage patterns. For example, in an actual application scenario, the computer system needs to monitor CPU utilization data. The initial first dynamic observation control covers the time period from 9:00 to 10:00, with the center at 9:30. Through steps S2101, S2102 and S2103, the computer system determines that the adjustment amount of the control is +15 minutes. Therefore, the new center position of the control will be adjusted to 9:45.

[0061] To further illustrate this process, consider the specific calculation steps. Assume that the computer system obtains the first implicit array representation [0.8, 0.9, 0.7, 0.85, 0.92] in implicit representation mining, and performs a dot product operation with the adjustment weights [0.1, 0.2, -0.1, 0.3, -0.2] obtained through deep learning model training to obtain the second implicit array representation [0.08, 0.18, -0.07, 0.255, -0.184]. Through further analysis, the computer system determines that the adjustment amount is +15 minutes. Therefore, the new control center position is 9:45.

[0062] In this way, the computer system can dynamically adjust the position of the observation control based on historical data to better capture the changes in resource usage patterns. For example, by adjusting the position of the control, the computer system can better track the changing trend of CPU utilization and allocate and schedule resources accordingly. In this way, the computer system can more effectively manage and optimize computing resources to meet the changing workload requirements.

[0063] As an implementation mode, the extraction times in the multiple first dynamic observation controls corresponding to each data subset are consistent. Based on this, step S2103, according to the multiple second implicit array representations, obtains the adjustment amount of the first dynamic observation control in the past multi-dimensional computing power time series data set, including:

[0064] Step S21031: obtaining an adjustment constraint according to the number of extractions in the random first dynamic observation control, where the adjustment constraint is used to constrain the adjustment span of the center of the first dynamic observation control in the past multi-dimensional computing power time series data set;

[0065] Step S21032: For each second implicit array representation, obtain the adjustment amount of the first dynamic observation control corresponding to the second implicit array representation based on the second implicit array representation and the adjustment constraint.

[0066] When the computer system executes an implementation of step S2103, for the case where the extraction times in the multiple first dynamic observation controls corresponding to each data subset are consistent, the adjustment amount of the first dynamic observation control in the past multi-dimensional computing power time series data set is obtained according to the multiple second implicit array representations. This process includes two steps: step S21031 and step S21032.

[0067] First, in step S21031, the computer system obtains an adjustment constraint based on the number of extractions in the first dynamic observation control. The adjustment constraint is used to limit the adjustment span of the center position of the control to ensure that the adjustment amount of all controls remains consistent. For example, assume that the number of extractions for each first dynamic observation control is 10 times. This means that within the time range covered by the control, the computer system will evenly extract 10 data points. Based on this number of extractions, the computer system determines the adjustment constraint to ensure that the adjustment amplitude of the center position of all controls does not exceed a certain range. Assume that the adjustment constraint stipulates that the maximum span of each adjustment is 15 minutes.

[0068] Specifically, the computer system first determines the time range covered by the control and calculates the interval between each data point based on the number of extractions. For example, assume that the time range covered by a first dynamic observation control is one hour, that is, 60 minutes. If the number of extractions is 10 times, the interval between each data point is 6 minutes. The computer system determines the adjustment constraint through this interval, combined with the control coverage. The calculation formula for the adjustment constraint can be expressed as:

[0069] ;

[0070] In order to ensure that the adjustment range of the center position of the control is not too large, the computer system sets the maximum span to 15 minutes. This means that the adjustment range of the center position of the control cannot exceed 15 minutes.

[0071] Next, in step S21032, the computer system obtains the adjustment amount of the first dynamic observation control corresponding to the second implicit array representation for each second implicit array representation according to the second implicit array representation and the adjustment constraint. Specifically, the computer system combines the second implicit array representation with the adjustment constraint to determine the final adjustment amount. For example, assuming that the second implicit array representation is [0.08, 0.18, -0.07, 0.255, -0.184]. Based on this array representation and the adjustment constraint, the computer system determines that the adjustment amount is +15 minutes.

[0072] In specific implementation, the computer system maps the second implicit array representation to the actual adjustment amount through a function. Assume that the mapping function is:

[0073] ;

[0074] In this example, assume that the adjustment weights are [0.1, 0.2, -0.1, 0.3, -0.2]. Then the adjustment amount is:

[0075] ;

[0076] Considering that the maximum span of the adjustment constraint is 15 minutes, the computer system limits the adjustment amount to this range. Therefore, the final adjustment amount is +15 minutes.

[0077] Through the above steps, the computer system determines the adjustment amount of the first dynamic observation control. This process ensures that the adjustment amounts of all controls remain consistent, making the data extraction process more unified and controllable. For example, in an actual application scenario, the computer system needs to monitor the CPU utilization data. The initial first dynamic observation control covers the time period from 9:00 to 10:00, with the center at 9:30. Through steps S21031 and S21032, the computer system determines that the adjustment amount of the control is +15 minutes. Therefore, the new center position of the control will be adjusted to 9:45. In this way, the computer system can dynamically adjust the position of the observation control according to historical data to better capture the pattern changes of resource usage. For example, by adjusting the position of the control, the computer system can better track the changing trend of CPU utilization and allocate and schedule resources accordingly. In this way, the computer system can more effectively manage and optimize computing resources to meet the changing workload requirements. In summary, the computer system realizes the adjustment of the first dynamic observation control through implicit representation mining, product operation, determination of adjustment constraints and calculation of adjustment amount. This process not only improves the flexibility of data extraction, but also ensures the consistency and accuracy of data extraction, providing a solid data foundation for subsequent computing resource planning.

[0078] As an embodiment, the method further includes a process of acquiring a center forward extension range and a center backward extension range of the random first dynamic observation control, including:

[0079] Step S21: performing implicit representation mining on the past multi-dimensional computing power time series data set to obtain a plurality of first implicit array representations, wherein the plurality of first implicit array representations have a corresponding relationship with the first dynamic observation control in the past multi-dimensional computing power time series data set;

[0080] Step S22: multiplying the plurality of first implicit array representations by the corresponding extension weights respectively to obtain a plurality of third implicit array representations, where the extension weights are parameters obtained based on deep training;

[0081] Step S23: Based on multiple third implicit array representations, obtain the center forward extension range and the center backward extension range of the first dynamic observation control in the past multi-dimensional computing power timing data set.

[0082] When obtaining the center forward extension range and the center backward extension range of the first dynamic observation control, the computer system performs the following steps: step S21, step S22 and step S23. This process includes implicit representation mining, quadrature operation and extension range determination, ensuring that the first dynamic observation control can be properly adjusted according to historical data.

[0083] First, in step S21, the computer system performs implicit representation mining on the past multi-dimensional computing power time series data set to obtain multiple first implicit array representations. These first implicit array representations have a one-to-one correspondence with the first dynamic observation control in the past multi-dimensional computing power time series data set. For example, in a typical computing power platform resource integration management scenario, the computer system collects data in multiple dimensions such as CPU utilization, memory usage, and storage I / O operation frequency in the past month. The data of each dimension forms a data subset. For the CPU utilization data subset, the computer system defines multiple first dynamic observation controls, each of which covers data within a period of time. The computer system performs implicit representation mining on these data subsets to obtain multiple first implicit array representations. Assume that for a first dynamic observation control in the CPU utilization data subset, the computer system obtains a first implicit array representation, such as [0.8, 0.9, 0.7,0.85, 0.92]. Next, in step S22, the computer system multiplies the multiple first implicit array representations by the corresponding extension weights to obtain multiple third implicit array representations. These extension weights are parameters obtained based on deep training. The computer system pre-trains a deep learning model that is used to learn how to adjust the extension range of the control based on the results of implicit representation mining. For example, assume that the extension weights are [0.1, 0.2, -0.1, 0.3, -0.2]. The computer system performs a dot product operation on the first implicit array representation [0.8, 0.9, 0.7, 0.85, 0.92] and the extension weights [0.1, 0.2, -0.1,0.3, -0.2] to obtain a third implicit array representation [0.08, 0.18, -0.07, 0.255, -0.184].

[0084] Finally, in step S23, the computer system obtains the center forward extension range and the center backward extension range of the first dynamic observation control in the past multi-dimensional computing power time series data set based on multiple third implicit array representations. Specifically, the computer system determines the extension range by comprehensively analyzing the third implicit array representation. For example, for a first dynamic observation control in the CPU utilization data subset, the computer system determines that the center forward extension range is +10 minutes and the center backward extension range is +10 minutes based on the third implicit array representation [0.08, 0.18, -0.07, 0.255, -0.184].

[0085] Through the above steps, the computer system obtains the center forward extension range and the center backward extension range of the first dynamic observation control. This process ensures that the control can be properly adjusted according to historical data, so as to better capture the changes in resource usage patterns.

[0086] To further illustrate this process, consider the specific calculation steps. Assume that the computer system obtains the first implicit array representation [0.8, 0.9, 0.7, 0.85, 0.92] in implicit representation mining, and performs a dot product operation with the extension weights [0.1, 0.2, -0.1, 0.3, -0.2] obtained through deep learning model training to obtain the third implicit array representation [0.08, 0.18, -0.07, 0.255, -0.184]. Through further analysis, the computer system determines that the center forward extension range is +10 minutes, and the center backward extension range is also +10 minutes.

[0087] In this way, the computer system can dynamically adjust the coverage of observation controls based on historical data to better capture changes in resource usage patterns. For example, by adjusting the coverage of controls, the computer system can better track the changing trends in CPU utilization and allocate and schedule resources accordingly. In this way, the computer system can more effectively manage and optimize computing resources to meet changing workload requirements.

[0088] Specifically, during the implicit representation mining process, the computer system processes each data subset and generates multiple first implicit array representations. These array representations reflect the characteristics of each observation control in the data subset. For example, for the CPU utilization data subset, assume that the computer system generates the first implicit array representation [0.8, 0.9, 0.7, 0.85, 0.92]. These array representations correspond one-to-one to the first dynamic observation control.

[0089] Next, the computer system performs dot product operations on these first implicit array representations and the extension weights. Assuming the extension weights are [0.1, 0.2, -0.1, 0.3, -0.2], the computer system obtains the third implicit array representations [0.08, 0.18, -0.07, 0.255, -0.184]. These third implicit array representations are used to determine the extension range.

[0090] To determine the stretch range, the computer system combines the third implicit array representation with the stretch constraint. Assume that the stretch constraint stipulates that the stretch range cannot exceed 15 minutes. The computer system determines that the center forward stretch range is +10 minutes and the center backward stretch range is also +10 minutes based on the third implicit array representation [0.08, 0.18, -0.07, 0.255, -0.184] and the stretch constraint.

[0091] Through the above steps, the computer system determines the center forward extension range and the center backward extension range of the first dynamic observation control. This process ensures that the extension range of all controls remains consistent, making the data extraction process more unified and controllable. For example, in an actual application scenario, the computer system needs to monitor the CPU utilization data. The initial first dynamic observation control covers the time period from 9:00 to 10:00, with the center at 9:30. Through steps S21, S22 and S23, the computer system determines that the center forward extension range of the control is +10 minutes, and the center backward extension range is also +10 minutes. Therefore, the new control coverage range is from 9:20 to 10:40.

[0092] In this way, the computer system can dynamically adjust the coverage of observation controls based on historical data to better capture changes in resource usage patterns. For example, by adjusting the coverage of controls, the computer system can better track the changing trends in CPU utilization and allocate and schedule resources accordingly. In this way, the computer system can more effectively manage and optimize computing resources to meet changing workload requirements.

[0093] In summary, the computer system adjusts the extension range of the first dynamic observation control through implicit representation mining, product calculation, extension constraint determination and extension range calculation. This process not only improves the flexibility of data extraction, but also ensures the consistency and accuracy of data extraction, providing a solid data foundation for subsequent computing resource planning.

[0094] As an implementation mode, the number of extractions in the multiple first dynamic observation controls corresponding to each data subset is consistent. Based on this, step S23, according to multiple third implicit array representations, obtains the center forward extension range and the center backward extension range of the first dynamic observation control in the past multi-dimensional computing power time series data set, including:

[0095] Step S231: obtaining an extension constraint according to the number of extractions in the random first dynamic observation control, where the extension constraint is used to constrain the center forward extension span and the center backward extension span of the first dynamic observation control in the past multi-dimensional computing power time series data set;

[0096] Step S232: For each third implicit array representation, according to the third implicit array representation and the extension constraint, obtain the center forward extension range and the center backward extension range of the first dynamic observation control corresponding to the third implicit array representation.

[0097] The computer system executes the process of step S23 according to the principle that the number of extractions in the multiple first dynamic observation controls corresponding to each data subset is consistent, which specifically includes step S231 and step S232. In step S231, the computer system obtains an extension constraint based on the number of extractions in the first dynamic observation control, and the extension constraint is used to limit the center forward extension span and the center backward extension span of the first dynamic observation control in the past multi-dimensional computing power time series data set. Subsequently, in step S232, the computer system obtains the corresponding center forward extension range and center backward extension range for each third implicit array representation according to the array representation and the extension constraint. In step S231, the computer system first determines the number of extractions of the first dynamic observation control. In the scenario of integrated management of computing power platform resources, the computer system sets a monitoring of resource usage (such as CPU utilization, memory occupancy, etc.) once every hour, which means that the number of extractions of the first dynamic observation control in each data subset is set to once per hour. Based on this, the computer system defines an extension constraint, which specifies the maximum extension span after each observation control is adjusted. For example, if the extension constraint stipulates that the forward extension span and the backward extension span of the center after any observation control adjustment cannot exceed 30 minutes, it means that if the center time point of the current control is 9:30 am, it can be extended forward to 9:00 and backward to 10:00.

[0098] After entering step S232, the computer system begins to process each third implicit array representation. Assume that the computer system has obtained multiple third implicit array representations from step S22, one of which is [0.08, 0.18, -0.07, 0.255, -0.184], which corresponds to a first dynamic observation control in the CPU utilization data subset. The computer system combines this third implicit array representation with the extension constraint to determine the center forward extension range and the center backward extension range of the control. In this example, the computer system may use a simple linear model to calculate the extension range. For example, the computer system may calculate the average value of the third implicit array representation as the extension coefficient, and then determine the final extension range based on the extension constraint. Assuming that the average value of the third implicit array representation is 0.076, and the extension constraint stipulates that the extension span does not exceed 30 minutes, the computer system may decide to set the center forward extension range to +10 minutes and the center backward extension range to +10 minutes to ensure that the total extension span meets the requirements of the extension constraint.

[0099] To further illustrate this process, the computer system may use a formula to calculate the stretch factor of each third implicit array representation. For example, let v be the third implicit array representation, c be the stretch factor, and n be the array length. The stretch factor c can be calculated by the following formula: In this case, v = [0.08, 0.18, -0.07, 0.255, -0.184], so c = 0.076. The computer system then adjusts the final extension range according to the extension constraint to ensure that the specified maximum span is not exceeded.

[0100] Through the above steps, the computer system determines the center forward extension range and the center backward extension range of the first dynamic observation control. This process not only ensures that the control can be dynamically adjusted according to historical data, but also ensures that the extension range of all controls remains consistent, making the data extraction process more unified and controllable. For example, in an actual application scenario, the computer system needs to monitor the CPU utilization data. The initial first dynamic observation control covers the time period from 9:00 to 10:00, with the center at 9:30. Through steps S231 and S232, the computer system determines that the center forward extension range of the control is +10 minutes, and the center backward extension range is also +10 minutes. Therefore, the new control covers a range from 9:20 to 10:40.

[0101] In this way, the computer system can dynamically adjust the coverage of observation controls based on historical data to better capture changes in resource usage patterns. For example, by adjusting the coverage of controls, the computer system can better track the changing trends in CPU utilization and allocate and schedule resources accordingly. In this way, the computer system can more effectively manage and optimize computing resources to meet changing workload requirements.

[0102] In summary, the computer system adjusts the extension range of the first dynamic observation control by determining the extension constraint and calculating the extension range based on the third implicit array representation. This process not only improves the flexibility of data extraction, but also ensures the consistency and accuracy of data extraction, providing a solid data foundation for subsequent computing power resource planning. For example, in a computing power platform resource integration management scenario, the computer system determines the extension range of the first dynamic observation control in the CPU utilization data subset through the above steps, ensuring more refined and comprehensive monitoring of resource usage.

[0103] As an implementation mode, step S300, in each second dynamic observation control corresponding to each data subset, extracts data from a past multi-dimensional computing power time series data set to obtain an extracted multi-dimensional computing power time series data set, including:

[0104] Step S310: for each second dynamic observation control corresponding to each data subset, divide the control coverage of each second dynamic observation control by the number of extractions to obtain an extraction step length;

[0105] Step S320: In each second dynamic observation control, average extraction is performed on the past multi-dimensional computing power time series data set according to the extraction step length; for the extraction position whose position identifier is not an integer, the observation value corresponding to the extraction position is obtained based on upsampling construction according to the observation value of the time point adjacent to the extraction position; for the extraction position whose position identifier is an integer, the observation value corresponding to the extraction position is obtained, wherein the position identifier represents the position of the extraction position in the time series dimension;

[0106] Step S330: Based on the multiple observation values ​​extracted from each second dynamic observation control, a multi-dimensional computing power time series data set is obtained after extraction.

[0107] When the computer system executes an implementation of step S300, in each second dynamic observation control corresponding to each data subset, data is extracted from the past multi-dimensional computing power time series data set to obtain the extracted multi-dimensional computing power time series data set. This process includes step S310, step S320 and step S330.

[0108] First, in step S310, the computer system divides the control coverage of each second dynamic observation control by the number of extractions for each second dynamic observation control corresponding to each data subset to obtain an extraction step. For example, in a typical computing platform resource integration management scenario, the computer system collects data in multiple dimensions such as CPU utilization, memory usage, and storage I / O operation frequency in the past month. The data of each dimension forms a data subset. For the CPU utilization data subset, the computer system defines multiple second dynamic observation controls, each of which covers data over a period of time. Assuming that the control coverage of a second dynamic observation control is 50 minutes and the number of extractions is 10 times, the extraction step is 5 minutes.

[0109] Specifically, the computer system first determines the control coverage of each second dynamic observation control. For example, assuming that the start time of a second dynamic observation control is 9:05 and the end time is 10:25, the control coverage is 50 minutes. Then, the computer system calculates the extraction step length according to the number of extractions. For example, if the number of extractions is 10 times, the extraction step length is 5 minutes.

[0110] Next, in step S320, the computer system performs an average extraction of the past multi-dimensional computing power time series data set in each second dynamic observation control according to the extraction step size. For extraction positions whose position identifiers are not integers, the computer system constructs the observation values ​​corresponding to the extraction positions based on upsampling (such as interpolation) based on the observation values ​​of the time points adjacent to the extraction positions. For extraction positions whose position identifiers are integers, the computer system obtains the observation values ​​corresponding to the extraction positions. Among them, the position identifier represents the position of the extraction position in the time series dimension.

[0111] For example, in the CPU utilization data subset, the computer system determines that the extraction step is 5 minutes. Assuming that the start time of the second dynamic observation control is 9:05 and the end time is 10:25, the computer system will extract data at 9:05, 9:10, 9:15, 9:20, 9:25, 9:30, 9:35, 9:40, 9:45, and 9:50. If there are corresponding observation values ​​at the ten time points of 9:05, 9:10, 9:15, 9:20, 9:25, 9:30, 9:35, 9:40, 9:45, and 9:50, the computer system directly obtains these observation values. If there are no observation values ​​at some time points, the computer system uses the interpolation method to upsample and obtain the observation values ​​at these time points.

[0112] Specifically, assume that there is no observation at 9:07, but there are observations at 9:05 and 9:10, which are 85% and 88% respectively. The computer system uses linear interpolation to calculate the observation at 9:07. Assuming that the time interval between 9:05 and 9:10 is 5 minutes, the time interval at 9:07 is 2 minutes. The computer system calculates the observation at 9:07 according to the following formula: .

[0113] Through the above steps, the computer system extracts data in each second dynamic observation control and fills in the missing data through the interpolation method.

[0114] Finally, in step S330, the computer system obtains a multi-dimensional computing power time series data set after extraction based on the multiple observation values ​​extracted from each second dynamic observation control. For example, in the CPU utilization data subset, the computer system obtains observation values ​​at 10 time points. The computer system combines these observation values ​​into an array to form a part of the multi-dimensional computing power time series data set after extraction.

[0115] Specifically, the computer system organizes the observation values ​​extracted from each second dynamic observation control into an array. Assuming that the observation values ​​extracted from a second dynamic observation control are [85%, 86.2%, 87%, 88%, 89%, 90%, 91%, 92%, 93%, 94%], the computer system organizes these observation values ​​into an array [85, 86.2, 87, 88, 89, 90, 91,92, 93, 94]. The computer system combines these arrays into a matrix to form a part of the multi-dimensional computing power time series data set after extraction.

[0116] Through the above steps, the computer system completes the data extraction process and obtains a multi-dimensional computing power time series data set after extraction. This process not only improves the flexibility of data extraction, but also ensures the consistency and accuracy of data extraction, and provides a solid data foundation for subsequent computing power resource planning. For example, in actual application scenarios, the computer system needs to monitor CPU utilization data. Through steps S310, S320 and S330, the computer system extracts data from each second dynamic observation control and forms a multi-dimensional computing power time series data set after extraction. In this way, the computer system can more effectively manage and optimize computing power resources to meet changing workload requirements.

[0117] In summary, the computer system realizes data extraction of the second dynamic observation control by determining the extraction step, performing data extraction and combining observation values. This process not only improves the flexibility of data extraction, but also ensures the consistency and accuracy of data extraction, providing a solid data foundation for subsequent computing resource planning.

[0118] As an implementation manner, in step S320, for an extraction position whose position identifier is not an integer, obtaining an observation value corresponding to the extraction position based on upsampling construction according to the observation value of a time point adjacent to the extraction position may include:

[0119] Step S321: for the extraction position whose position identifier is not an integer, in the xth round of loop, determine the xth time point in the second dynamic observation control, x>1;

[0120] Step S322: if the xth time point is an adjacent time point of the extraction position, obtain the interval between the position identifier of the extraction position and x, and obtain the distance weight according to the interval; if the xth time point is not an adjacent time point of the extraction position, set the distance weight to 0;

[0121] Step S323: Multiply the distance weight and the observation value corresponding to the x-th time point to obtain the x-th observation value; complete multiple rounds of cycles, the number of cycles is the number of time points in the second dynamic observation control, and add the multiple observation values ​​corresponding to the multiple rounds of cycles to obtain the observation value corresponding to the extraction position.

[0122] When the computer system executes an implementation of step S320, for the extraction position whose position identifier is not an integer, based on the observation value of the time point adjacent to the extraction position, the observation value corresponding to the extraction position is obtained based on upsampling. This process includes step S321, step S322 and step S323.

[0123] First, in step S321, the computer system determines the xth time point in the second dynamic observation control in the xth round of loop for the extraction position whose position identifier is not an integer, where x is greater than 1. For example, in a typical computing platform resource integration management scenario, the computer system collects data in multiple dimensions such as CPU utilization, memory usage, and storage I / O operation frequency in the past month. The data in each dimension forms a data subset. For the CPU utilization data subset, the computer system defines multiple second dynamic observation controls, each of which covers data over a period of time. Assuming that the control coverage range of a second dynamic observation control is 50 minutes and the number of extractions is 10 times, the extraction step length is 5 minutes.

[0124] Specifically, the computer system first determines the control coverage of each second dynamic observation control. For example, assuming that the start time of a second dynamic observation control is 9:05 and the end time is 10:25, the control coverage is 50 minutes. Then, the computer system calculates the extraction step length according to the number of extractions. For example, if the number of extractions is 10 times, the extraction step length is 5 minutes. This means that the computer system will extract data at ten time points: 9:05, 9:10, 9:15, 9:20, 9:25, 9:30, 9:35, 9:40, 9:45, and 9:50.

[0125] Next, in step S322, if the computer system finds that the x-th time point is an adjacent time point of the extraction location, it obtains the interval between the location identifier of the extraction location and x, and obtains the distance weight based on the interval. If the x-th time point is not an adjacent time point of the extraction location, the distance weight is set to 0. For example, suppose there is no observation at 9:07, but there are observations at 9:05 and 9:10, which are 85% and 88% respectively. The computer system calculates the observation at 9:07 using a linear interpolation method. Assuming that the time interval between 9:05 and 9:10 is 5 minutes, the time interval at 9:07 is 2 minutes. The computer system calculates the observation at 9:07 according to the following formula:

[0126] ;

[0127] ;

[0128] Through the above steps, the computer system calculates the distance weight.

[0129] Finally, in step S323, the computer system multiplies the distance weight with the observation value corresponding to the x-th time point to obtain the x-th observation value. After completing multiple cycles, the number of cycles is the number of time points in the second dynamic observation control, and multiple observation values ​​corresponding to the multiple cycles are added together to obtain the observation value corresponding to the extraction position. For example, in the CPU utilization data subset, the computer system determines that the distance weights are 0.6 and 0.4. Assuming that the observation values ​​at 9:05 and 9:10 are 85% and 88% respectively, the computer system calculates the observation value at 9:07 according to the following formula:

[0130] ;

[0131] Through the above steps, the computer system extracts data in each second dynamic observation control and fills in the missing data through the interpolation method.

[0132] Specifically, the computer system determines the time points in the second dynamic observation control in step S321. Assuming that the start time of the second dynamic observation control is 9:05 and the end time is 10:25, the control coverage is 50 minutes. If the number of extractions is 10, the extraction step is 5 minutes. This means that the computer system will extract data at the ten time points of 9:05, 9:10, 9:15, 9:20, 9:25, 9:30, 9:35, 9:40, 9:45, and 9:50.

[0133] In step S322, the computer system determines the interval between the location identifier of the extracted location and x, and obtains the distance weight according to the interval. Assume that there is no observation at 9:07, but there are observations at 9:05 and 9:10, which are 85% and 88% respectively. The computer system calculates the observation value at 9:07 using a linear interpolation method. Assume that the time interval between 9:05 and 9:10 is 5 minutes, then the time interval at 9:07 is 2 minutes. The computer system calculates the observation value at 9:07 according to the following formula:

[0134] ;

[0135] ;

[0136] In step S323, the computer system multiplies the distance weight with the observation value corresponding to the x-th time point to obtain the x-th observation value. After completing multiple cycles, the number of cycles is the number of time points in the second dynamic observation control, and multiple observation values ​​corresponding to the multiple cycles are added together to obtain the observation value corresponding to the extraction position. For example, in the CPU utilization data subset, the computer system determines that the distance weights are 0.6 and 0.4. Assuming that the observation values ​​at 9:05 and 9:10 are 85% and 88% respectively, the computer system calculates the observation value at 9:07 according to the following formula:

[0137] ;

[0138] Through the above steps, the computer system extracts data in each second dynamic observation control and fills in the missing data through the interpolation method.

[0139] In summary, the computer system realizes data extraction of the second dynamic observation control by determining the extraction step, performing data extraction and combining observation values. This process not only improves the flexibility of data extraction, but also ensures the consistency and accuracy of data extraction, providing a solid data foundation for subsequent computing power resource planning. For example, in actual application scenarios, the computer system needs to monitor CPU utilization data. Through steps S321, S322 and S323, the computer system extracts data from each second dynamic observation control and forms a multi-dimensional computing power time series data set after extraction. In this way, the computer system can more effectively manage and optimize computing power resources to meet changing workload requirements.

[0140] For example, in practical applications, the computer system needs to monitor CPU utilization data. Assuming that the start time of a second dynamic observation control is 9:05 and the end time is 10:25, the control coverage is 50 minutes. If the number of extractions is 10, the extraction step is 5 minutes. This means that the computer system will extract data at ten time points: 9:05, 9:10, 9:15, 9:20, 9:25, 9:30, 9:35, 9:40, 9:45, and 9:50. If there is no observation at 9:07, but there are observations at 9:05 and 9:10, which are 85% and 88% respectively, the computer system uses a linear interpolation method to calculate the observation at 9:07. Assuming that the time interval between 9:05 and 9:10 is 5 minutes, the time interval at 9:07 is 2 minutes. The computer system calculates the observation at 9:07 according to the following formula:

[0141] ;

[0142] ;

[0143] Through the above steps, the computer system extracts data in each second dynamic observation control and fills in the missing data through interpolation. In this way, the computer system can more effectively manage and optimize computing resources to meet the changing workload requirements.

[0144] As an embodiment, the method further includes:

[0145] Step S11: Process the extracted multidimensional computing power time series data set to obtain the multidimensional computing power time series data set after correlation determination, the multidimensional computing power time series data set after correlation determination includes information on correlations in the multidimensional computing power time series data set after extraction, the correlations include one or more of short-time correlations in dynamic observation controls, long-time correlations across controls, and cross-observation value correlations, the short-time correlations in the dynamic observation controls represent the correlations of observations in a second dynamic observation control corresponding to a data subset, the long-time correlations across controls represent the correlations of observations between multiple second dynamic observation controls corresponding to a data subset, and the cross-observation value correlations represent the correlations of observations between multiple data subsets.

[0146] Based on this, step S400, based on the extracted multi-dimensional computing power time series data set, infers a subsequent multi-dimensional computing power time series data set, including:

[0147] Step S410: Determine the multi-dimensional computing power time series data set based on the correlation, and infer a subsequent multi-dimensional computing power time series data set.

[0148] In the process of processing the multi-dimensional computing power time series data set after extraction, the computer system executes step S11 to obtain the multi-dimensional computing power time series data set after correlation determination. The set contains the correlation information in the extracted data set, including short-term correlation in the dynamic observation control, long-term correlation across controls, and correlation across observation values. Based on this, the computer system infers the subsequent multi-dimensional computing power time series data set based on the multi-dimensional computing power time series data set after correlation determination in step S400, specifically including step S410.

[0149] First, in step S11, the computer system processes the extracted multi-dimensional computing power time series data set to obtain a multi-dimensional computing power time series data set after correlation determination. The set contains the correlation information in the extracted data set. For example, in the scenario of integrated resource management of the computing power platform, the computer system collects data in multiple dimensions such as CPU utilization, memory usage, and storage I / O operation frequency in the past month. The data of each dimension forms a data subset. For the CPU utilization data subset, the computer system defines multiple second dynamic observation controls, each of which covers data within a period of time. The computer system obtains the extracted multi-dimensional computing power time series data set through step S300, and determines the correlation information on this basis.

[0150] Specifically, the computer system determines relevance information in the following manner:

[0151] 1. Short-term correlation in dynamic observation controls: Analyze the correlation of observations in a second dynamic observation control corresponding to a data subset. For example, in the CPU utilization data subset, the computer system analyzes the correlation between CPU utilization observations during the period from 9:05 to 10:25. Assuming that there are 10 observations during this period, which are 85%, 86.2%, 87%, 88%, 89%, 90%, 91%, 92%, 93%, and 94%, the computer system determines the correlation between these observations by calculating the correlation coefficients between them.

[0152] 2. Long-term correlation across controls: Analyze the correlation of observations between multiple second dynamic observation controls corresponding to a data subset. For example, in the CPU utilization data subset, the computer system analyzes the correlation between the CPU utilization observations from 9:05 to 10:25 and the CPU utilization observations from 10:30 to 11:50. Assuming that the observations in these two time periods are 85%, 86.2%, 87%, 88%, 89%, 90%, 91%, 92%, 93%, 94% and 86%, 87.2%, 88%, 89%, 90%, 91%, 92%, 93%, 94%, 95%, respectively, the computer system determines the correlation between these two groups of observations by calculating the correlation coefficient between them.

[0153] 3. Correlation across observations: The computer system analyzes the correlation between observations across multiple data subsets. For example, in the three data subsets of CPU utilization, memory usage, and storage I / O operation frequency, the computer system analyzes the correlation between the observations of these three dimensions within the same time period. Assuming that the observations of these three dimensions within the same time period are 85%, 75%, and 5,000 times / second, respectively, the computer system determines the correlation between these observations by calculating the correlation coefficients between them.

[0154] Through the above steps, the computer system determines the correlation information in the multi-dimensional computing power time series data set after extraction, and forms a multi-dimensional computing power time series data set after correlation determination.

[0155] Next, in step S410, the computer system infers the subsequent multi-dimensional computing power time series data set based on the multi-dimensional computing power time series data set after correlation determination. Specifically, the computer system infers the subsequent multi-dimensional computing power time series data set by analyzing the correlation information in the multi-dimensional computing power time series data set after correlation determination. For example, in the scenario of integrated management of computing power platform resources, the computer system infers the predicted values ​​of CPU utilization, memory usage, and storage I / O operation frequency in the future period of time by analyzing the correlation information in the multi-dimensional computing power time series data set after correlation determination.

[0156] Specifically, the computer system infers the subsequent multi-dimensional computing power time series data set in the following way:

[0157] 1. The computer system predicts the change trend of CPU utilization in the future by analyzing the short-term correlation in the dynamic observation control. For example, assuming that the correlation coefficient of the CPU utilization observation value in the current time period is 0.85, the computer system predicts the change trend of CPU utilization in the future by analyzing this correlation coefficient.

[0158] 2. The computer system predicts the correlation between CPU utilization in the future and other time periods by analyzing the long-term correlation across controls. For example, assuming that the correlation coefficient between the observed CPU utilization value in the current time period and the observed CPU utilization value in the next time period is 0.75, the computer system predicts the correlation between CPU utilization in the future and other time periods by analyzing this correlation coefficient.

[0159] 3. The computer system predicts the correlation between CPU utilization, memory usage and storage I / O operation frequency in the future by analyzing the correlation across observations. For example, assuming that the correlation coefficients of the observed values ​​of CPU utilization, memory usage and storage I / O operation frequency in the same time period are 0.8, 0.7 and 0.6 respectively, the computer system predicts the correlation between these three dimensions in the future by analyzing these correlation coefficients.

[0160] Through the above steps, the computer system infers the subsequent multi-dimensional computing power time series data set. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, and provides a solid data foundation for subsequent computing power resource planning. For example, in actual application scenarios, the computer system needs to monitor the CPU utilization data. Through steps S11 and S410, the computer system infers the subsequent multi-dimensional computing power time series data set from the multi-dimensional computing power time series data set after the correlation is determined. In this way, the computer system can more effectively manage and optimize computing power resources to meet the ever-changing workload requirements.

[0161] In summary, the computer system obtains the subsequent multi-dimensional computing power time series data set by determining the correlation information and reasoning, and realizes the processing and prediction of multi-dimensional computing power time series data. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, providing a solid data foundation for subsequent computing power resource planning.

[0162] As an implementation method, the extracted multi-dimensional computing power time series data set is based on a third-order tensor representation, and each order of the extracted multi-dimensional computing power time series data set includes an observation value component, a time component, and a dynamic observation control component; based on this, step S11, processing the extracted multi-dimensional computing power time series data set to obtain a multi-dimensional computing power time series data set after correlation determination, including:

[0163] Step S111: Process the dynamic observation control component of the first third-order tensor to obtain a second third-order tensor, where the second third-order tensor includes information on short-term correlation in the dynamic observation control;

[0164] Step S112: Process the time component of the second third-order tensor to obtain a third third-order tensor, where the third third-order tensor includes information on long-term correlation across controls;

[0165] Step S113: Process the observation value component of the third third-order tensor to obtain a fourth third-order tensor, which includes information on correlation across observation values; the first third-order tensor is used to characterize the multi-dimensional computing power time series data set after extraction, and the fourth third-order tensor is used to characterize the multi-dimensional computing power time series data set after correlation determination.

[0166] In the process of processing the extracted multi-dimensional computing power time series data set, the computer system executes step S11 to obtain the multi-dimensional computing power time series data set after correlation determination. This process includes step S111, step S112 and step S113. The computer system processes the third-order tensor of the extracted multi-dimensional computing power time series data set to gradually extract and determine the short-term correlation in the dynamic observation control, the long-term correlation across controls, and the correlation information across observation values.

[0167] First, in step S111, the dynamic observation control component of the first third-order tensor is processed to obtain a second third-order tensor, which includes information on short-term correlation in the dynamic observation control. For example, in the scenario of integrated resource management of the computing power platform, the computer system collects data in multiple dimensions such as CPU utilization, memory usage, and storage I / O operation frequency in the past month. The data of each dimension forms a data subset. For the CPU utilization data subset, the computer system defines multiple second dynamic observation controls, each of which covers data over a period of time. Assuming that the control coverage of a second dynamic observation control is 50 minutes and the number of extractions is 10 times, the extraction step is 5 minutes. This means that the computer system will extract data at ten time points: 9:05, 9:10, 9:15, 9:20, 9:25, 9:30, 9:35, 9:40, 9:45, and 9:50.

[0168] Specifically, the computer system organizes the extracted data into a third-order tensor, whose three dimensions are observation value component, time component and dynamic observation control component. Then, it is easy to understand that the first third-order tensor in the above steps is the initial multi-dimensional computing power time series data set to be processed after extraction. Assuming that the first third-order tensor is T1, its shape is (N, M, K), where N represents the dimension size of the observation value component (for example, different types of resources), M represents the dimension size of the time component (for example, the number of data points in each observation control), and K represents the dimension size of the dynamic observation control component (for example, the total number of observation controls). For example, the shape of T1 is (3, 10, 100), indicating that there are 3 types of resources, 10 data points in each observation control, and a total of 100 observation controls. Next, in step S111, the computer system processes the dynamic observation control component of the first third-order tensor T1 to obtain a second third-order tensor T2, which includes information on short-term correlation in the dynamic observation control. For example, the computer system can use a convolutional neural network (CNN) to extract short-term correlation. Assume that the observed values ​​of a certain observation control in T1 are [85%, 86.2%, 87%, 88%, 89%, 90%, 91%, 92%, 93%, 94%], and the computer system performs convolution operations on these observed values ​​through the convolution kernel to obtain the feature representation of short-term correlation. Assuming that the convolution kernel size is 3, the convolution operation is as follows: ;

[0169] Through the convolution operation, the computer system obtains a second third-order tensor T2, whose shape is still (3, 10, 100), but contains information about short-term correlations in the dynamic observation control.

[0170] Next, in step S112, the computer system processes the time component of the second third-order tensor T2 to obtain a third third-order tensor T3, which includes information about long-term correlations across controls. For example, the computer system can use a long short-term memory network (LSTM) to extract long-term correlations across controls. Assuming that the observed values ​​of a certain resource type in T2 are [85%, 86.2%, 87%, 88%, 89%, 90%, 91%, 92%, 93%, 94%], the computer system performs sequence modeling on these observed values ​​through LSTM to obtain feature representations of long-term correlations. Assuming that the time step of LSTM is 10, the LSTM operation is as follows:

[0171] ;

[0172] Through the LSTM operation, the computer system obtains the third third-order tensor T3, whose shape is still \( (3, 10,100) \), but contains information about the long-term correlation across controls.

[0173] Finally, in step S113, the computer system processes the observation value component of the third third-order tensor T3 to obtain a fourth third-order tensor T4, which includes information on the correlation across observation values. For example, the computer system can use a self-attention mechanism to extract the correlation across observation values. Assuming that the observation value at a certain time point in T3 is [85%, 75%, 5000], the computer system performs a weighted summation of these observation values ​​through the self-attention mechanism to obtain a feature representation of the correlation across observation values. Assuming that the feature dimension of the self-attention mechanism is 3, the self-attention operation is as follows:

[0174] ;

[0175] ;

[0176] ;

[0177] ;

[0178] Through the self-attention operation, the computer system obtains a fourth third-order tensor T4, whose shape is still (3, 10, 100) but contains information about the correlation across observations.

[0179] Through the above steps, the computer system determines the correlation information in the multi-dimensional computing power time series data set after extraction, and forms a multi-dimensional computing power time series data set after correlation determination. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, and provides a solid data foundation for subsequent computing power resource planning. For example, in actual application scenarios, the computer system needs to monitor CPU utilization data. Through steps S111, S112 and S113, the computer system determines the correlation information from the multi-dimensional computing power time series data set after extraction, and forms a multi-dimensional computing power time series data set after correlation determination. In this way, the computer system can more effectively manage and optimize computing power resources to meet the ever-changing workload requirements.

[0180] In summary, the computer system processes the third-order tensor of the extracted multi-dimensional computing power time series data set, gradually extracts and determines the correlation information therein, and realizes the processing of multi-dimensional computing power time series data. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, providing a solid data foundation for subsequent computing power resource planning.

[0181] As an implementation mode, step 111, processing the dynamic observation control component of the first third-order tensor to obtain a second third-order tensor, includes:

[0182] Step S1111: obtain a first transition third-order tensor based on the first normalization operator, the first mapping operator, the first nonlinear transformation operator and the second mapping operator connected in sequence; add the first transition third-order tensor to the first third-order tensor to obtain a second third-order tensor; the mapping parameters corresponding to the first mapping operator and the second mapping operator respectively are parameters obtained based on deep training and are used to obtain information on short-time correlation in dynamic observation controls.

[0183] In step S111, the computer system processes the dynamic observation control component of the first third-order tensor to obtain a second third-order tensor, which specifically includes step S1111. This process involves using a series of operators to extract information on short-term correlation in the dynamic observation control.

[0184] First, in step S1111, the computer system processes the first third-order tensor based on the first normalization operator, the first mapping operator, the first nonlinear transformation operator, and the second mapping operator connected in sequence to obtain a first transition third-order tensor. Afterwards, the first transition third-order tensor is added to the first third-order tensor to obtain a second third-order tensor. Here, the mapping parameters corresponding to the first mapping operator and the second mapping operator are parameters obtained based on deep training, which are used to obtain information on short-term correlation in dynamic observation controls.

[0185] Specifically, assume that the shape of the first third-order tensor T1 is (N, M, K), where N represents the dimension size of the observation value component (e.g., different types of resources), M represents the dimension size of the time component (e.g., the number of data points in each observation control), and K represents the dimension size of the dynamic observation control component (e.g., the total number of observation controls). For example, the shape of T1 is (3, 10, 100), which means there are 3 types of resources, 10 data points in each observation control, and a total of 100 observation controls.

[0186] The computer system first processes the first third-order tensor T1 using the first normalization operator. The normalization operator is usually used to standardize the input data so that it has zero mean and unit variance, which helps to improve the stability and convergence speed of subsequent processing. The normalization operator can use batch normalization or other normalization methods. Assuming that the batch normalization operator is BN, the result after processing is:

[0187] ;

[0188] Next, the computer system uses the first mapping operator to normalize the tensor T normThe first mapping operator can be a linear function, such as matrix multiplication. Assuming the first mapping operator is W1, the result after processing is:

[0189] ;

[0190] Then, the computer system uses the first nonlinear transformation operator to transform the tensor T processed by the first mapping operator. map1 Processing. Nonlinear transformation operators are usually used to introduce nonlinear characteristics so that the model can fit more complex patterns. Commonly used nonlinear transformation operators include ReLU, sigmoid, tanh, etc. Assuming that the first nonlinear transformation operator is ReLU, the result after processing is:

[0191] ;

[0192] Next, the computer system uses the second mapping operator to transform the tensor T after nonlinear transformation. nonlin The second mapping operator can also be a linear function. Assuming that the second mapping operator is W2, the result after processing is:

[0193] ;

[0194] At this point, the computer system obtains the first transition third-order tensor T trans Then, the computer system converts the first transition third-order tensor T trans Add it to the first third-order tensor T1 to obtain the second third-order tensor T2. This operation is usually called residual connection, which helps to alleviate the gradient vanishing problem and speed up the training process. The specific operation is:

[0195] ;

[0196] Through the above steps, the computer system obtains the second third-order tensor T2, which contains the information of short-term correlation in the dynamic observation control.

[0197] In specific implementation, the mapping parameters corresponding to the first mapping operator W1 and the second mapping operator W2 are parameters obtained based on deep training. These parameters are updated through the back propagation algorithm during the training process to minimize the loss function. The loss function can be mean square error (MSE), cross entropy, etc., depending on the requirements of the task. For example, assuming that the loss function is mean square error, the calculation formula is:

[0198] ;

[0199] Among them, y ijk is the true value, is the predicted value.

[0200] Through the above steps, the computer system processes the dynamic observation control component of the first and third order tensors to obtain the second and third order tensors containing short-term correlation information. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, providing a solid data foundation for subsequent computing resource planning.

[0201] For example, in an actual application scenario, the computer system needs to monitor the CPU utilization data. Through step S1111, the computer system extracts the information of short-term correlation in the dynamic observation control from the first third-order tensor and forms a second third-order tensor. In this way, the computer system can more effectively manage and optimize computing resources to meet the changing workload requirements.

[0202] In summary, the computer system uses a series of operators to process the first and third order tensors, extracts the information of short-term correlation in the dynamic observation control, and forms the second and third order tensors. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, providing a solid data foundation for subsequent computing resource planning.

[0203] As an implementation method, step S112 processes the time component of the second third-order tensor to obtain a third third-order tensor, including: obtaining a second transition third-order tensor based on the first flipping operator, the second normalization operator, the third mapping operator, the second nonlinear transformation operator, the fourth mapping operator and the second flipping operator connected in sequence; adding the second transition third-order tensor to the second third-order tensor to obtain a third third-order tensor; the mapping parameters corresponding to the third mapping operator and the fourth mapping operator are parameters obtained based on deep training and are used to obtain information on long-term correlation across controls.

[0204] In step S112, the computer system processes the time component of the second third-order tensor to obtain a third third-order tensor, specifically including the following steps: the second third-order tensor is processed based on the first flip operator, the second normalization operator, the third mapping operator, the second nonlinear transformation operator, the fourth mapping operator and the second flip operator connected in sequence to obtain a second transition third-order tensor; the second transition third-order tensor is added to the second third-order tensor to obtain a third third-order tensor. Among them, the mapping parameters corresponding to the third mapping operator and the fourth mapping operator are parameters obtained based on deep training, which are used to obtain information on long-term correlation across controls.

[0205] Specifically, assume that the shape of the second third-order tensor T2 is (N, M, K), where N represents the dimensional size of the observation value component (for example, different types of resources), M represents the dimensional size of the time component (for example, the number of data points in each observation control), and K represents the dimensional size of the dynamic observation control component (for example, the total number of observation controls). For example, the shape of T2 is (3, 10, 100), indicating that there are 3 types of resources, 10 data points in each observation control, and a total of 100 observation controls. First, the computer system uses the first flip operator to flip the dimension of the second third-order tensor T2 in order to focus on the time component. The flip operator is used to change the dimensional order of the tensor so that the time component becomes the first dimension. Assuming that the first flip operator is Flip, the processed result is:

[0206] ;

[0207] Next, the computer system uses the second normalization operator to transform the flipped tensor T flip1 Normalization operators are usually used to standardize input data so that it has zero mean and unit variance, which helps to improve the stability and convergence speed of subsequent processing. Assuming that the second normalization operator is BN, the result after processing is:

[0208] ;

[0209] The computer system then uses the third mapping operator to normalize the tensor T norm The third mapping operator can be a linear function, such as matrix multiplication. Assuming the third mapping operator is W3, the result after processing is:

[0210] ;

[0211] Next, the computer system uses the second nonlinear transformation operator to transform the tensor T after being processed by the third mapping operator. map3 Processing. Nonlinear transformation operators are usually used to introduce nonlinear characteristics so that the model can fit more complex patterns. Commonly used nonlinear transformation operators include ReLU, sigmoid, tanh, etc. Assuming that the second nonlinear transformation operator is ReLU, the result after processing is:

[0212] ;

[0213] Then, the computer system uses the fourth mapping operator to transform the tensor T after nonlinear transformation. nonlin The fourth mapping operator can also be a linear function. Assuming that the fourth mapping operator is W4, the result after processing is:

[0214] ;

[0215] At this point, the computer system obtains the second transition third-order tensor T trans2 The computer system then uses a second flip operator to transform the second transition third-order tensor T trans2 The dimension order is restored to the second and third order tensor T trans2 Assuming the second flip operator is Flip, the result after processing is:

[0216] ;

[0217] Finally, the computer system converts the second transition third-order tensor T flip2 Add it to the second third-order tensor T2 to obtain the third third-order tensor T3. This operation is usually called residual connection, which helps to alleviate the gradient vanishing problem and speed up the training process. The specific operation is:

[0218] ;

[0219] Through the above steps, the computer system obtains the third third-order tensor T3, which contains information on long-term correlation across controls.

[0220] In specific implementation, the mapping parameters corresponding to the third mapping operator W3 and the fourth mapping operator W4 are parameters obtained based on deep training. These parameters are updated by the back propagation algorithm during the training process to minimize the loss function. The loss function can be mean square error (MSE), cross entropy, etc., depending on the requirements of the task. For example, assuming that the loss function is mean square error, the calculation formula is:

[0221] ;

[0222] Among them, y ijk is the true value, is the predicted value.

[0223] Through the above steps, the computer system processes the time component of the second third-order tensor and obtains the third third-order tensor containing the long-term correlation information across controls. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, providing a solid data foundation for subsequent computing resource planning.

[0224] For example, in an actual application scenario, the computer system needs to monitor the CPU utilization data. Through step S112, the computer system extracts the information of the long-term correlation across controls from the second third-order tensor and forms a third third-order tensor. In this way, the computer system can more effectively manage and optimize computing resources to meet the changing workload requirements.

[0225] In summary, the computer system uses a series of operators to process the second and third order tensors, extracts information about long-term correlations across controls, and forms the third and third order tensors. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, providing a solid data foundation for subsequent computing resource planning.

[0226] As an implementation method, step S113 processes the observation value component of the third third-order tensor to obtain a fourth third-order tensor, including: obtaining a third transition third-order tensor based on the third third-order tensor connected in sequence based on a third flip operator, a third normalization operator, a fifth mapping operator, a third nonlinear transformation operator, a sixth mapping operator and a fourth flip operator; adding the third transition third-order tensor to the third third-order tensor to obtain a fourth third-order tensor; the mapping parameters corresponding to the fifth mapping operator and the sixth mapping operator are parameters obtained based on deep training and are used to obtain information on cross-observation value correlation.

[0227] In step S113, the computer system processes the observed value component of the third third-order tensor to obtain a fourth third-order tensor, specifically including the following steps: processing the third third-order tensor based on the third flip operator, the third normalization operator, the fifth mapping operator, the third nonlinear transformation operator, the sixth mapping operator and the fourth flip operator connected in sequence to obtain a third transition third-order tensor; adding the third transition third-order tensor to the third third-order tensor to obtain a fourth third-order tensor. The mapping parameters corresponding to the fifth mapping operator and the sixth mapping operator are parameters obtained based on deep training, which are used to obtain information on cross-observation correlation.

[0228] Specifically, assume that the shape of the third-order tensor T3 is (N, M, K), where N represents the dimension size of the observation value component (e.g., different types of resources), M represents the dimension size of the time component (e.g., the number of data points in each observation control), and K represents the dimension size of the dynamic observation control component (e.g., the total number of observation controls). For example, the shape of T3 is (3, 10, 100), which means there are 3 types of resources, 10 data points in each observation control, and a total of 100 observation controls.

[0229] First, the computer system uses the third flip operator to flip the dimension of the third third-order tensor T3 in order to focus on the observation value component. The flip operator is used to change the dimensional order of the tensor so that the observation value component becomes the first dimension. Assuming that the third flip operator is Flip, the result after processing is:

[0230] ;

[0231] Next, the computer system uses the third normalization operator to transform the flipped tensor T flip3 Normalization operators are usually used to standardize input data so that it has zero mean and unit variance, which helps to improve the stability and convergence speed of subsequent processing. Assuming that the third normalization operator is BN, the result after processing is:

[0232] ;

[0233] Then, the computer system uses the fifth mapping operator to normalize the tensor T norm The fifth mapping operator may be a linear function, such as matrix multiplication. Assuming the fifth mapping operator is W5, the result after processing is:

[0234] ;

[0235] Next, the computer system uses the third nonlinear transformation operator to transform the tensor T after the fifth mapping operator. map5 Processing. Nonlinear transformation operators are usually used to introduce nonlinear characteristics so that the model can fit more complex patterns. Commonly used nonlinear transformation operators include ReLU, sigmoid, tanh, etc. Assuming that the third nonlinear transformation operator is ReLU, the result after processing is:

[0236] ;

[0237] Then, the computer system uses the sixth mapping operator to transform the tensor T after nonlinear transformation. nonlin The sixth mapping operator can also be a linear function. Assuming that the sixth mapping operator is W6, the result after processing is:

[0238] ;

[0239] At this point, the computer system obtains the third transition third-order tensor T trans3 Then, the computer system uses the fourth flip operator to transform the third transition third-order tensor T trans3 The dimensional order of is restored to the dimensional order of the third-order tensor T3. Assuming that the fourth flip operator is Flip, the result after processing is:

[0240] ;

[0241] Finally, the third transition third-order tensor T flip4 Add it to the third-order tensor T3 to obtain the fourth-order tensor T4. This operation is usually called residual connection, which helps to alleviate the gradient vanishing problem and speed up the training process. The specific operation is:

[0242] ;

[0243] Through the above steps, the computer system obtains the fourth third-order tensor T4, which contains information on the correlation across observations.

[0244] In specific implementation, the mapping parameters corresponding to the fifth mapping operator W5 and the sixth mapping operator W6 are parameters obtained based on deep training. These parameters are updated by the back propagation algorithm during the training process to minimize the loss function. The loss function can be mean square error (MSE), cross entropy, etc., depending on the requirements of the task. Through the above steps, the computer system processes the observation value component of the third third-order tensor and obtains the fourth third-order tensor containing cross-observation correlation information. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, providing a solid data foundation for subsequent computing resource planning.

[0245] For example, in actual application scenarios, the computer system needs to monitor data of multiple resources such as CPU utilization, memory usage, and storage I / O operation frequency. Through step S113, the computer system extracts information about cross-observation correlation from the third third-order tensor and forms a fourth third-order tensor. In this way, the computer system can more effectively manage and optimize computing resources to meet the changing workload requirements.

[0246] In summary, the computer system processes the third-order tensor using a series of operators to extract information about the correlation between observations and form a fourth-order tensor. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, providing a solid data foundation for subsequent computing resource planning.

[0247] As another implementation of step S11, that is, processing the extracted multi-dimensional computing power time series data set to obtain the multi-dimensional computing power time series data set after correlation determination, it can include: inputting the extracted multi-dimensional computing power time series data set into multiple fusion components connected in sequence to obtain the multi-dimensional computing power time series data set after correlation determination, a random fusion component among the multiple fusion components includes a three-layer feedforward neural network, namely a channel fusion unit, a dynamic observation control fusion unit and a variable fusion unit; the channel fusion unit, the dynamic observation control fusion unit and the variable fusion unit are respectively used to obtain the information of short-time correlation in the dynamic observation control, the information of long-time correlation across controls and the information of correlation across observation values.

[0248] In another embodiment of step S11, the computer system processes the extracted multi-dimensional computing power time series data set to obtain a multi-dimensional computing power time series data set after correlation determination. This process includes inputting the extracted multi-dimensional computing power time series data set into multiple fusion components that are connected (i.e., stacked) in sequence to obtain a multi-dimensional computing power time series data set after correlation determination. Each fusion component includes a three-layer feedforward neural network, namely a channel fusion unit, a dynamic observation control fusion unit, and a variable fusion unit, which are respectively used to obtain information on short-term correlation in dynamic observation controls, information on long-term correlation across controls, and information on correlation across observation values.

[0249] Specifically, assume that the extracted multi-dimensional computing power time series data set is represented as a third-order tensor T, whose shape is (N, M, K), where N represents the dimension size of the observation value component (for example, different types of resources), M represents the dimension size of the time component (for example, the number of data points in each observation control), and K represents the dimension size of the dynamic observation control component (for example, the total number of observation controls). For example, the shape of T is (3, 10, 100), indicating that there are 3 types of resources, 10 data points in each observation control, and a total of 100 observation controls.

[0250] The computer system first inputs the extracted multi-dimensional computing power time series data set T into the first fusion component. The fusion component includes three layers of feedforward neural networks, namely channel fusion unit, dynamic observation control fusion unit and variable fusion unit. Each unit is composed of a feedforward neural network and is used to extract different types of correlation information.

[0251] The channel fusion unit is used to obtain information about short-term correlation in dynamic observation controls. The unit receives the third-order tensor T and processes it through a feedforward neural network. Assume that the feedforward neural network of the channel fusion unit consists of a fully connected layer and an activation function. The weight matrix of the fully connected layer is W c , the bias term is b c , the activation function is f c, then the output of the channel fusion unit T c It can be expressed as:

[0252] ;

[0253] The dynamic observation control fusion unit is used to obtain information on long-term correlation across controls. This unit also receives the third-order tensor T and processes it through a feedforward neural network. Assume that the feedforward neural network of the dynamic observation control fusion unit consists of a fully connected layer and an activation function. The weight matrix of the fully connected layer is W d , the bias term is b d , the activation function is f d , then the output T of the dynamic observation control fusion unit d It can be expressed as:

[0254] ;

[0255] The variable fusion unit is used to obtain information about the correlation between observations. The unit receives the third-order tensor T and processes it through a feedforward neural network. Assume that the feedforward neural network of the variable fusion unit consists of a fully connected layer and an activation function. The weight matrix of the fully connected layer is W v , the bias term is b v , the activation function is f v , then the output T of the variable fusion unit v It can be expressed as:

[0256] ;

[0257] In the first fusion component, the outputs of the channel fusion unit, the dynamic observation control fusion unit, and the variable fusion unit will be combined to form a new third-order tensor T', which contains information about three types of correlations. Assuming that the combination is element-by-element addition, T' can be expressed as:

[0258] ;

[0259] Next, the computer system passes T' as input to the next fusion component. This process is repeated until all fusion components have been traversed. Each fusion component further extracts and enhances the correlation information, ultimately forming a multi-dimensional computing power time series data set after the correlation is determined.

[0260] In actual application scenarios, the computer system needs to monitor the data of multiple resources. For example, for the data of multiple resources such as CPU utilization, memory usage, and storage I / O operation frequency, the computer system extracts the information of short-term correlation in dynamic observation controls, long-term correlation across controls, and correlation across observation values ​​from the extracted multi-dimensional computing power time series data set through this implementation method of step S11, and forms a multi-dimensional computing power time series data set after correlation determination. In this way, the computer system can more effectively manage and optimize computing power resources to meet the ever-changing workload requirements. In summary, the computer system processes the extracted multi-dimensional computing power time series data set by using multiple fusion components connected in sequence, extracts information of multiple correlations, and forms a multi-dimensional computing power time series data set after correlation determination. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, providing a solid data foundation for subsequent computing power resource planning.

[0261] As an implementation method, step S410, based on the multi-dimensional computing power time series data set after correlation determination, infers a subsequent multi-dimensional computing power time series data set, including: inputting the multi-dimensional computing power time series data set after correlation determination into a dimension reorganization operator to obtain a transition multi-dimensional computing power time series data set; inputting the transition multi-dimensional computing power time series data set into a projection operator, inferring a subsequent multi-dimensional computing power time series data set; the dimension reorganization operator is used to expand the dynamic observation control component and time component of the multi-dimensional computing power time series data set after correlation determination, and the projection operator is used to project the transition multi-dimensional computing power time series data set to a preset range.

[0262] In step S410, the computer system infers the subsequent multidimensional computing power time series data set based on the multidimensional computing power time series data set after correlation determination. This process includes inputting the multidimensional computing power time series data set after correlation determination into the dimension reorganization operator to obtain the transition multidimensional computing power time series data set; and then inputting the transition multidimensional computing power time series data set into the projection operator to infer the subsequent multidimensional computing power time series data set. The dimension reorganization operator is used to expand the dynamic observation control component and time component of the multidimensional computing power time series data set after correlation determination, and the projection operator is used to project the transition multidimensional computing power time series data set to a preset range.

[0263] Specifically, assuming that the correlation is determined, the multi-dimensional computing power time series data set is represented as a third-order tensor T, whose shape is (N, M, K), where N represents the dimension size of the observation value component (for example, different types of resources), M represents the dimension size of the time component (for example, the number of data points in each observation control), and K represents the dimension size of the dynamic observation control component (for example, the total number of observation controls). For example, the shape of T is (3, 10, 100), indicating that there are 3 types of resources, 10 data points in each observation control, and a total of 100 observation controls.

[0264] First, the computer system inputs the multi-dimensional computing power time series data set T after correlation is determined into the dimension reorganization operator. The dimension reorganization operator is used to expand the dynamic observation control component and time component of the tensor T, that is, to merge these two dimensions into one dimension to facilitate subsequent processing. Assuming that the dimension reorganization operator is Reshape, the result after processing is:

[0265] ;

[0266] Here, T reshape The shape of becomes (N, M×K), which means there are N types of resources and each resource type has data at M×K time points. reshape The shape is (3, 1000), which means there are 3 resource types and each resource type has data at 1000 time points.

[0267] Next, the computer system will transition the multi-dimensional computing power time series data set T reshape Input projection operator. The projection operator is a mapping function used to project the transition multi-dimensional computing power time series data set to a preset range. The projection operator can be implemented by one or more layers of neural network. Assuming the projection operator is Proj, the result after processing is:

[0268] ;

[0269] The projection operator Proj can contain one or more fully connected layers, each of which has its own weight matrix W and bias term b, and is nonlinearly transformed by the activation function f. Assuming that the projection operator contains two fully connected layers, the calculation process can be expressed as:

[0270] ;

[0271] Among them, W1 and b1 are the weight matrix and bias term of the first fully connected layer, respectively, and W2 and b2 are the weight matrix and bias term of the second fully connected layer, respectively. The activation function f can be ReLU, sigmoid, tanh, etc. Through the above steps, the computer system obtains the subsequent multi-dimensional computing power time series data set T proj , which contains the prediction of future resource usage. In specific implementation, the weight matrices W1, W2 and bias terms b1, b2 in the projection operator are parameters obtained based on deep training. These parameters are updated through the back propagation algorithm during the training process to minimize the loss function. The loss function can be mean square error (MSE), cross entropy, etc., depending on the requirements of the task.

[0272] Through step S410, the computer system infers a subsequent multi-dimensional computing power time series data set from the multi-dimensional computing power time series data set after correlation is determined, so that computing power resources can be managed and optimized more effectively to meet the ever-changing workload requirements.

[0273] In summary, the computer system uses the dimension reorganization operator and the projection operator to process the multi-dimensional computing power time series data set after the correlation is determined, and infers the subsequent multi-dimensional computing power time series data set. This process not only improves the flexibility of data processing, but also ensures the consistency and accuracy of data processing, providing a solid data foundation for subsequent computing power resource planning.

[0274] As an implementation method, the method may further include:

[0275] Step S500: for a random data subset in the past multi-dimensional computing power time series data set, determine the disorder of the data subset after the dynamic observation control is maintained after applying multiple first dynamic observation controls; and determine the disorder of the data subset after the extendable dynamic observation control is obtained after applying multiple second dynamic observation controls;

[0276] Step S600: subtract the degree of disorder of the data subset after the dynamic observation control is maintained from the degree of disorder of the data subset after the dynamic observation control can be extended to obtain the amount of information of the dynamic observation control;

[0277] Step S700: Calculate the mean of multiple dynamic observation control information amounts corresponding to the past multi-dimensional computing power time series data set to obtain the dynamic observation control information amount error.

[0278] The higher the chaos of a sequence, the more information is obtained. For example, the chaos of a data subset after a fixed dynamic observation control is calculated based on the following method: for a random maintenance subset among multiple maintenance subsets, calculate the similarity between the maintenance subset and each other maintenance subset, and the other maintenance subset is a random maintenance subset other than the maintenance subset; count the number of times the similarity is less than the similarity threshold; calculate the ratio of the number of times to the number of multiple maintenance subsets; calculate the logarithm of the ratio; calculate the average of multiple logarithmic values ​​corresponding to multiple maintenance subsets; and determine the average as the chaos of the data subset after the fixed dynamic observation control. Based on measuring the information gain of the multi-dimensional computing power time series data set brought by the application length extensible dynamic observation control compared to the application length fixed dynamic observation control, the information error of the dynamic observation control is constructed. The information error of the dynamic observation control is used to guide the division of the range extensible dynamic observation control, that is, to guide the generation of the adjustment amount and the extension range on both sides of the center. Based on the information error of the dynamic observation control, the complexity of the obtained extracted multi-dimensional computing power time series data set is increased to obtain more information suitable for the reasoning task of the multi-dimensional computing power time series data set. The computer system determines the degree of disorder of the data subset after the application of multiple first dynamic observation controls for a random data subset in the past multi-dimensional computing power time series data set; at the same time, it determines the degree of disorder of the data subset after the application of multiple second dynamic observation controls for the extended dynamic observation control. This step aims to evaluate the information gain by comparing the degree of disorder of the data subset under different dynamic observation controls, and then guide the design and optimization of subsequent dynamic observation controls.

[0279] Specifically, the computer system first defines chaos as an indicator to measure the amount of information. The higher the chaos, the richer the information. For the chaos of a data subset after a fixed dynamic observation control, the system calculates it based on a series of sub-steps. For a random maintenance subset among multiple maintenance subsets, the similarity between the maintenance subset and each other maintenance subset is calculated. The similarity here can be calculated in many ways, such as cosine similarity or Euclidean distance. Taking cosine similarity as an example, if two vectors are and , then the cosine similarity calculation formula between them is:

[0280] ;

[0281] The system then counts the number of times the similarity is less than the preset similarity threshold. Assuming the similarity threshold is set to 0.5, the number of times the similarity between all maintained subsets is less than 0.5 is counted. Next, the system calculates the ratio of the above number to the number of multiple maintained subsets, and takes the logarithm of the ratio. After that, the system calculates the average of the multiple logarithmic values ​​corresponding to the multiple maintained subsets, and determines this average as the chaos degree of the data subset after the dynamic observation control is fixed.

[0282] After completing the above steps, the computer system proceeds to the next step S600, which is to obtain the information amount of the dynamic observation control by subtracting the degree of disorder of the data subset after the dynamic observation control is maintained from the degree of disorder of the data subset after the dynamic observation control can be extended. This step is intended to quantify the information gain brought by applying dynamic observation controls of different lengths.

[0283] Finally, in step S700, the computer system averages the information of multiple dynamic observation controls corresponding to the past multi-dimensional computing power time series data set to obtain the information error of the dynamic observation control. This error value reflects the overall level of information gain of the data subset after applying different dynamic observation controls, which can be used to guide the design of subsequent dynamic observation controls, especially the determination of the adjustment amount and the extension range on both sides of the center.

[0284] To further illustrate this process, consider a specific example. Suppose a computer system is managing a computing power platform that has time series data for multiple resources such as CPU utilization, memory usage, and network traffic. The system decides to use a time window of length 5 (first dynamic observation control) and a time window of length 10 (second dynamic observation control) to analyze this data. For the time window of length 5, the system selects a data subset that contains the changes in CPU utilization in the recent period of time. The system calculates the similarity between this subset and other similar subsets and determines its chaos according to the aforementioned method. Then, the system repeats the same process, but this time using a time window of length 10. By comparing the two chaos values, the system calculates the amount of information of the dynamic observation control, which helps the system understand whether using a longer time window can obtain more useful information. In practical applications, the error in the amount of information of the dynamic observation control can be used to optimize the selection of the dynamic observation control, making the final extracted multi-dimensional computing power time series data set more complex and thus more suitable for reasoning tasks. For example, if it is found that using a longer time window does bring more information gain, the computer system may choose to use a longer observation control to capture more long-term trends and patterns. Conversely, if a shorter time window is sufficient, the system may choose to maintain the current setting to reduce unnecessary computational overhead. By accurately measuring and comparing the degree of chaos of data subsets under different dynamic observation controls, the computer system can effectively optimize its observation strategy to better serve the needs of integrated resource management of the computing power platform. This series of steps not only enhances the system's data analysis capabilities, but also provides it with more flexible and efficient data processing methods.

[0285] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

Claims

1. A computing platform resource integration management method, characterized in that: The method comprises: Acquire multiple first dynamic observation controls corresponding to each data subset in the past multi-dimensional computing power time series data set, where the multiple first dynamic observation controls are used to extract each data subset, while keeping the extraction range unchanged; wherein a data subset is a time series data subset consisting of computing power resource data corresponding to one dimension, and keeping the extraction range of the observation control unchanged means that the observation control always covers the same time period; Each first dynamic observation control corresponding to each data subset is moved and the extraction range is extended to construct each second dynamic observation control corresponding to each data subset; In each second dynamic observation control corresponding to each data subset, data extraction is performed on the past multi-dimensional computing power time series data set to obtain an extracted multi-dimensional computing power time series data set, and the number of extractions in each second dynamic observation control is consistent with the number of extractions in each first dynamic observation control; Inferring a subsequent multi-dimensional computing power time series data set based on the extracted multi-dimensional computing power time series data set, so as to perform computing power planning based on the subsequent multi-dimensional computing power time series data set; Wherein, in each second dynamic observation control corresponding to each data subset, data extraction is performed on the past multi-dimensional computing power time series data set to obtain the extracted multi-dimensional computing power time series data set, including: For each second dynamic observation control corresponding to each data subset, divide the control coverage of each second dynamic observation control by the number of extractions to obtain an extraction step length; In each of the second dynamic observation controls, the past multi-dimensional computing power time series data set is averagely extracted according to the extraction step; for an extraction position whose position identifier is not an integer, the observation value corresponding to the extraction position is obtained based on upsampling construction according to the observation value of the time point adjacent to the extraction position; for an extraction position whose position identifier is an integer, the observation value corresponding to the extraction position is obtained, wherein the position identifier represents the position of the extraction position in the time series dimension; According to the multiple observation values ​​extracted from each of the second dynamic observation controls, the extracted multi-dimensional computing power time series data set is obtained; For the extraction position whose position identifier is not an integer, obtaining the observation value corresponding to the extraction position based on upsampling construction according to the observation value of the time point adjacent to the extraction position includes: For the extraction position whose position identifier is not an integer, in the xth round of loop, the xth time point in the second dynamic observation control is determined, where x>1; If the x-th time point is an adjacent time point of the extraction position, obtain the interval between the position identifier of the extraction position and x, and obtain a distance weight according to the interval; if the x-th time point is not an adjacent time point of the extraction position, set the distance weight to 0; The distance weight is multiplied by the observation value corresponding to the x-th time point to obtain the x-th observation value; multiple cycles are completed, the number of cycles is the number of time points in the second dynamic observation control, and multiple observation values ​​corresponding to the multiple cycles are added to obtain the observation value corresponding to the extraction position; The method further comprises: Processing the extracted multidimensional computing power time series data set to obtain a multidimensional computing power time series data set after correlation determination, wherein the multidimensional computing power time series data set after correlation determination includes information on correlations in the extracted multidimensional computing power time series data set, wherein the correlations include one or more of short-term correlations in dynamic observation controls, long-term correlations across controls, and cross-observation value correlations, wherein the short-term correlations in the dynamic observation controls represent the correlations of observations in a second dynamic observation control corresponding to a data subset, the long-term correlations across controls represent the correlations of observations between multiple second dynamic observation controls corresponding to a data subset, and the cross-observation value correlations characterize the correlations of observations between multiple data subsets; The method of inferring a subsequent multi-dimensional computing power time series data set based on the extracted multi-dimensional computing power time series data set includes: Determine the multi-dimensional computing power time series data set based on the correlation, and infer the subsequent multi-dimensional computing power time series data set; The extracted multi-dimensional computing power time series data set is embodied based on a third-order tensor, and each order of the extracted multi-dimensional computing power time series data set includes an observation value component, a time component, and a dynamic observation control component; The step of processing the extracted multi-dimensional computing power time series data set to obtain a multi-dimensional computing power time series data set after correlation determination includes: Processing the dynamic observation control component of the first third-order tensor to obtain a second third-order tensor, wherein the second third-order tensor includes information on short-term correlation in the dynamic observation control; Processing the time component of the second third-order tensor to obtain a third third-order tensor, wherein the third third-order tensor includes information on the long-term correlation across controls; The observation value component of the third third-order tensor is processed to obtain a fourth third-order tensor, wherein the fourth third-order tensor includes information on the correlation across observation values; the first third-order tensor is used to characterize the extracted multi-dimensional computing power time series data set, and the first third-order tensor represents the initial extracted multi-dimensional computing power time series data set to be processed, and the fourth third-order tensor is used to characterize the multi-dimensional computing power time series data set after the correlation is determined.

2. The method according to claim 1, characterized in that The moving and extending the extraction range of each first dynamic observation control corresponding to each data subset to construct each second dynamic observation control corresponding to each data subset includes: For a random first dynamic observation control corresponding to each data subset, add the control center of the random first dynamic observation control and the adjustment amount corresponding to the random first dynamic observation control to obtain the control center of the second dynamic observation control; Obtain half of the control coverage of the random first dynamic observation control to obtain an initial control coverage; The initial control coverage range is spliced ​​with the center forward extension range corresponding to the random first dynamic observation control to obtain the forward control coverage range of the second dynamic observation control; the control center of the second dynamic observation control is subtracted from the forward control coverage range of the second dynamic observation control to obtain the starting position of the second dynamic observation control; Splice the initial control coverage range with the center backward extension range corresponding to the random first dynamic observation control to obtain the backward control coverage range of the second dynamic observation control; splice the control center of the second dynamic observation control with the backward control coverage range of the second dynamic observation control to obtain the end position of the second dynamic observation control.

3. The method according to claim 2, characterized in that The method further includes a process of obtaining an adjustment amount corresponding to the random first dynamic observation control, including: Performing implicit representation mining on the past multi-dimensional computing power time series data set to obtain a plurality of first implicit array representations, wherein the plurality of first implicit array representations have a corresponding relationship with a first dynamic observation control in the past multi-dimensional computing power time series data set; Multiplying the plurality of first implicit array representations with the respectively corresponding adjustment weights to obtain a plurality of second implicit array representations, wherein the adjustment weights are parameters obtained based on deep training; According to the plurality of second implicit array representations, the adjustment amount of the first dynamic observation control in the past multi-dimensional computing power timing data set is obtained.

4. The method according to claim 3, characterized in that The extraction times of the multiple first dynamic observation controls corresponding to each data subset are consistent; the adjustment amount of the first dynamic observation control in the past multi-dimensional computing power time series data set is obtained according to the multiple second implicit array representations, including: According to the number of extractions in the random first dynamic observation control, an adjustment constraint is obtained, where the adjustment constraint is used to constrain the adjustment span of the center of the first dynamic observation control in the past multi-dimensional computing power time series data set; For each second implicit array representation, an adjustment amount of the first dynamic observation control corresponding to the second implicit array representation is obtained according to the second implicit array representation and the adjustment constraint.

5. The method according to claim 2, characterized in that: The method also includes a process of acquiring a center forward extension range and a center backward extension range of the random first dynamic observation control, including: Performing implicit representation mining on the past multi-dimensional computing power time series data set to obtain a plurality of first implicit array representations, wherein the plurality of first implicit array representations have a corresponding relationship with a first dynamic observation control in the past multi-dimensional computing power time series data set; Multiplying the plurality of first implicit array representations by the corresponding extension weights respectively to obtain a plurality of third implicit array representations, wherein the extension weights are parameters obtained based on deep training; According to the plurality of third implicit array representations, the center forward extension range and the center backward extension range of the first dynamic observation control in the past multi-dimensional computing power timing data set are obtained.

6. The method according to claim 5, characterized in that The extraction times of the multiple first dynamic observation controls corresponding to each data subset are consistent; the obtaining of the center forward extension range and the center backward extension range of the first dynamic observation control in the past multi-dimensional computing power time series data set according to the multiple third implicit array representations includes: According to the number of extractions in the random first dynamic observation control, an extension constraint is obtained, wherein the extension constraint is used to constrain the center forward extension span and the center backward extension span of the first dynamic observation control in the past multi-dimensional computing power time series data set; For each third implicit array representation, according to the third implicit array representation and the extension constraint, a center forward extension range and a center backward extension range of the first dynamic observation control corresponding to the third implicit array representation are obtained.

7. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 6 are implemented.