Full-stack resource big data analysis dynamic optimization method for computing power server

Abstract

The application discloses a full-stack resource big data analysis dynamic optimization method for computing power servers, relates to the technical field of resource management, and comprises the following steps: standardizing and aggregating a full-stack multi-dimensional resource element structured data vector, obtaining resource standard indexes, and obtaining computing power and data correlation comprehensive indexes by calculating the coupling relationship among computing power, data and energy efficiency; based on the computing power and data correlation comprehensive indexes, a dynamic feedback window function is constructed, performance feedback data of each node is collected, a feedback package is obtained, runtime sequence normalization processing is performed on the feedback package, a window feedback dataset is obtained; the correlation of performance indexes in the resource standard indexes and the window feedback dataset is calculated, the weight factor is dynamically corrected by analyzing the influence degree of different resource dimensions, a resource contribution degree vector is obtained, and an executable optimization strategy package is generated. The application constructs a dynamic feedback window function, realizes real-time synchronization and time sequence alignment of full-stack resource states.

Description

Technical Field

[0001] This invention relates to the field of resource management technology, and in particular to a dynamic optimization method for full-stack resource big data analysis of computing servers. Background Technology

[0002] With the rapid development of cloud computing, artificial intelligence, and massive parallel computing, efficient scheduling and optimization of full-stack resources are crucial to ensuring performance and stability in the management of computing server clusters. Conventional computing optimization methods are usually based on resource monitoring and static scheduling, using static resource configuration and monitoring mechanisms based on fixed thresholds to periodically collect and analyze multi-dimensional indicators such as CPU, memory, and storage to achieve resource allocation and adjustment.

[0003] Existing methods have two shortcomings when dealing with dynamic workloads: First, in multi-node parallel scenarios, conventional computing power optimization methods still mainly rely on fixed time windows and static weight update mechanisms, which are difficult to reflect the coupling relationship between computing power, data and energy efficiency in real time; in addition, static resource configuration is difficult to adapt to fluctuations in resource demand in real time, resulting in low resource utilization. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a dynamic optimization method for full-stack resource big data analysis for computing servers to solve the problems of insufficient collaborative optimization capabilities and lack of time-adaptive resource optimization in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a full-stack resource big data analysis and dynamic optimization method for computing servers, comprising:

[0008] Collect full-stack multi-dimensional resource elements of computing power servers and integrate them into a structured data vector of full-stack multi-dimensional resource elements;

[0009] Standardize and aggregate the structured data vectors of full-stack multi-dimensional resource elements to obtain resource standard indicators. By calculating the coupling relationship between computing power, data, and energy efficiency, obtain a comprehensive indicator of the correlation between computing power and data.

[0010] Based on the comprehensive index of computing power and data correlation, a dynamic feedback window function is constructed to collect the performance feedback data of each node, obtain the feedback package, perform runtime sequence normalization processing on the feedback package, and obtain the window feedback dataset.

[0011] The correlation between resource standard indicators and performance indicators in the window feedback dataset is calculated. By analyzing the degree of influence of different resource dimensions, the weight factors are dynamically adjusted to obtain the resource contribution vector and generate an executable tuning strategy package.

[0012] After the execution parameters in the tuning strategy package are subject to safety limits, they are sent to each node. The tuning strategy is executed and the resource status is verified in real time. The comprehensive performance improvement value is calculated, and the multi-dimensional resource elements of the whole stack are adjusted according to the comprehensive performance improvement value.

[0013] Collect execution feedback information and write it back to the control plane, generating a receipt message.

[0014] As a preferred embodiment of the dynamic optimization method for full-stack resource big data analysis of computing servers described in this invention, the steps of collecting full-stack multi-dimensional resource elements of the computing server and integrating them into a structured data vector of full-stack multi-dimensional resource elements are as follows.

[0015] Collect full-stack multi-dimensional resource elements, perform time synchronization processing on the full-stack multi-dimensional resource elements, and perform interpolation correction on the full-stack multi-dimensional resource elements with missing time points to obtain continuous full-stack multi-dimensional resource elements.

[0016] Based on continuous full-stack multi-dimensional resource elements, an energy response balance acquisition algorithm is used to calculate the energy response balance coefficient and generate a structured data vector of full-stack multi-dimensional resource elements.

[0017] As a preferred embodiment of the full-stack resource big data analysis and dynamic optimization method for computing servers described in this invention, the steps for standardizing and aggregating the structured data vectors of full-stack multi-dimensional resource elements to obtain resource standard indicators are as follows:

[0018] Perform linear standardization on each resource element in the full-stack multi-dimensional resource element structured data vector according to a unified interval to obtain full-stack multi-dimensional resource element standardized data.

[0019] A sliding window mean correction method is used to smooth and correct the standardized data of full-stack multidimensional resource elements, generating corrected full-stack multidimensional resource element data.

[0020] The corrected full-stack multidimensional resource element data is divided into a set of computing power elements, a set of communication elements, and a set of energy efficiency elements. A three-dimensional coupled aggregation calculation of computing power, data, and energy efficiency is performed to obtain resource standard indicators.

[0021] As a preferred embodiment of the full-stack resource big data analysis and dynamic optimization method for computing power servers described in this invention, the step of obtaining a comprehensive index relating computing power and data by calculating the coupling relationship between computing power, data, and energy efficiency is as follows:

[0022] Based on resource standard indicators, and combined with sets of computing power elements, communication elements, and energy efficiency elements, a coupled input matrix of computing power, data, and energy efficiency is constructed.

[0023] Based on the coupling input matrix of computing power, data, and energy efficiency, the coupling coefficient is obtained by calculating the coordinated change rate;

[0024] Based on the coupling coefficient, a comprehensive index of the correlation between computing power and data is obtained by calculating the standard deviation of the set of energy efficiency factors.

[0025] As a preferred embodiment of the full-stack resource big data analysis and dynamic optimization method for computing power servers described in this invention, the steps of constructing a dynamic feedback window function based on a comprehensive index linking computing power and data, collecting performance feedback data from each node, and obtaining feedback packets are as follows:

[0026] A dynamic feedback window function is constructed by calculating the relative rate of change of the comprehensive index of computing power and data correlation.

[0027] Based on the time length of the dynamic feedback window function, performance feedback data is collected on each computing server node to obtain node performance feedback data.

[0028] Perform time normalization processing on the node performance feedback data to obtain the node time-series normalized feedback value;

[0029] The comprehensive index of computing power and data association, the time length of the dynamic feedback window function, and the normalized feedback value of node timing are integrated to generate a feedback package.

[0030] As a preferred embodiment of the full-stack resource big data analysis and dynamic optimization method for computing servers described in this invention, the steps for performing runtime sequence normalization processing on the feedback packets to obtain the window feedback dataset are as follows:

[0031] Deconstruct all feedback packets to obtain the globally aligned feedback matrix;

[0032] Based on the time length of the dynamic feedback window function, time-weighted smoothing calculation is performed on the global aligned feedback matrix to obtain the weighted smoothed feedback value. Global statistical features are obtained by calculating the mean and standard deviation of the weighted smoothed feedback across all nodes.

[0033] Based on global statistical features, the runtime normalized feedback value is calculated to obtain the window feedback dataset.

[0034] As a preferred embodiment of the full-stack resource big data analysis and dynamic optimization method for computing servers described in this invention, the steps of calculating the correlation between resource standard indicators and performance indicators in the window feedback dataset, dynamically adjusting weight factors by analyzing the influence of different resource dimensions, obtaining resource contribution vectors, and generating an executable optimization strategy package are as follows.

[0035] By calculating the Pearson correlation coefficient matrix between the resource standard indicators and the window feedback dataset, the correlation matrix is ​​obtained. The average absolute correlation is calculated for each resource dimension to obtain the set of influence strength of the resource dimension.

[0036] By combining the set of influence intensity of resource dimensions and the set of historical weight factors, the weight factors are corrected through a time decay dynamic update formula to obtain a dynamic set of weight factors;

[0037] Based on a dynamic set of weighting factors and resource standard indicators, a resource contribution vector is calculated and combined with a comprehensive indicator of computing power and data correlation to generate an executable optimization strategy package.

[0038] As a preferred embodiment of the full-stack resource big data analysis and dynamic optimization method for computing servers described in this invention, the steps of distributing the execution parameters in the optimization strategy package to each node after applying security limits and executing the optimization strategy are as follows:

[0039] Based on resource standard indicators and historical stable parameters, calculate the upper and lower safety limits of each resource dimension and generate a safety tuning strategy package by combining it with an executable tuning strategy package.

[0040] Through a unified scheduling interface, security optimization strategy packages are distributed to each computing server node. The nodes then adjust their resources according to the parameters in the security optimization strategy packages and obtain updated node resource status data.

[0041] As a preferred embodiment of the full-stack resource big data analysis and dynamic optimization method for computing servers described in this invention, the steps of real-time verification of resource status, calculation of comprehensive performance improvement value, and adjustment of full-stack multi-dimensional resource elements based on the comprehensive performance improvement value are as follows:

[0042] Based on the updated node resource status data, a set of resource status is collected, a set of performance improvement differences is calculated, and a comprehensive performance improvement value is generated by combining the resource contribution vector and the set of dynamic weight factors.

[0043] Performance verification is performed based on the comprehensive performance improvement value, a performance verification status identifier is generated, and the full-stack multi-dimensional resource elements are adjusted according to the comprehensive performance improvement value to obtain a new set of full-stack multi-dimensional resource elements.

[0044] As a preferred embodiment of the full-stack resource big data analysis and dynamic optimization method for computing servers described in this invention, the steps of collecting execution feedback information and writing it back to the control plane to generate a receipt message are as follows:

[0045] Collect node operation logs and resource status data of each computing server node after executing the security tuning strategy package, and obtain execution feedback information;

[0046] The execution feedback information is written to the control plane feedback database, generating an acknowledgment message.

[0047] The beneficial effects of this invention are as follows: by constructing a dynamic feedback window function and performing runtime sequence normalization on the feedback package, real-time synchronization and timing alignment of the full-stack resource status are achieved; by dynamically updating the execution time decay and dynamically correcting the weight factor by analyzing the influence of resource dimensions, multi-dimensional resource adaptive optimization is achieved, improving the system resource coordination efficiency and ensuring the accuracy of the tuning strategy. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of a dynamic optimization method for full-stack resource big data analysis for computing servers.

[0050] Figure 2 This is a flowchart for the collection and standardization of multi-dimensional resource elements across the entire stack.

[0051] Figure 3 A flowchart for constructing a comprehensive index and dynamic feedback window function that correlates computing power and data.

[0052] Figure 4 A flowchart for optimizing strategy execution and verification. Detailed Implementation

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0054] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0055] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0056] Reference Figures 1-4 This is one embodiment of the present invention, which provides a dynamic optimization method for full-stack resource big data analysis of computing servers, including the following steps:

[0057] S1. Collect the full-stack multi-dimensional resource elements of the computing server and integrate them into a structured data vector of full-stack multi-dimensional resource elements.

[0058] The full-stack multi-dimensional resource elements include CPU utilization, GPU computing power utilization, system memory utilization, network bandwidth utilization, storage access rate, input / output latency, power consumption, device temperature, energy efficiency ratio, and thermal margin.

[0059] Collect full-stack multi-dimensional resource elements, perform time synchronization processing on the full-stack multi-dimensional resource elements, and perform interpolation correction on the full-stack multi-dimensional resource elements with missing time points to obtain continuous full-stack multi-dimensional resource elements.

[0060] Furthermore, at the operation layer of each computing server node, the utilization rate of each core is calculated by monitoring the processor instruction execution time and idle time, and the CPU utilization rate is recorded in real time with a CPU utilization sampling period. At the GPU driver layer, the runtime utilization information of each GPU core's computing unit is obtained, and the ratio of the number of executing threads to the number of idle threads is calculated to obtain the GPU computing power utilization rate. By recording memory allocation and the number of idle page frames at the operation memory management interface layer, the system memory utilization rate is calculated. At the node network interface layer, the uplink and downlink data transmission volume per unit time is monitored and compared with the physical link bandwidth limit to obtain the network bandwidth utilization rate. At the storage control layer, the number of read and write operations per second and the amount of data transmitted are counted to obtain the storage access rate. Rate; Record the time interval from initiation to completion of input / output requests at the input / output scheduling layer, and obtain the input / output latency by calculating the average latency value; Read the output values ​​of power sensors or power meters in real time at the power management interface layer, record the power consumption data of each component, and obtain the power consumption; Periodically read the temperature data of the CPU, GPU, and key components of the motherboard through the temperature sensor interface embedded in the motherboard to form a temperature sampling sequence and obtain the device temperature; Based on the power consumption acquisition and performance monitoring results, calculate the energy efficiency ratio by the ratio of the computational load per unit of task to the power consumption; Calculate the temperature rise difference of the heat dissipation component by reading the temperature of the heat dissipation component, the fan speed, and the ambient temperature, and compare the temperature rise difference of the heat dissipation component with the maximum allowable operating temperature to obtain the heat dissipation margin.

[0061] It should be noted that the CPU utilization sampling period is set based on the processor clock frequency and task scheduling interval. The CPU utilization sampling interval is determined by statistically analyzing the average task execution time within a single scheduling cycle. For example, setting the CPU utilization sampling period to 1 second can cover the entire task scheduling cycle.

[0062] Furthermore, based on the sampling timestamps of all computing server nodes, a globally unified timeline is constructed. Clock correction is performed on each computing server node according to the master control time signal source to generate a unified time reference sequence. Based on the unified time reference sequence and the full-stack multi-dimensional resource elements collected by each node, the full-stack multi-dimensional resource elements of different nodes in adjacent sampling periods are aligned according to the timeline. When some elements are missing at a given time point, a gap marker is retained on the timeline, generating a time-aligned set of full-stack multi-dimensional resource elements. A linear interpolation compensation algorithm is used to interpolate and correct the missing time points in each dimension of the time-aligned full-stack multi-dimensional resource elements, obtaining the interpolated and corrected elements. The interpolated and corrected elements are then reconstructed in chronological order to reconstruct a complete time series, generating continuous full-stack multi-dimensional resource elements.

[0063] Based on continuous full-stack multi-dimensional resource elements, an energy response balance acquisition algorithm is used to calculate the energy response balance coefficient and generate a structured data vector of full-stack multi-dimensional resource elements.

[0064] Furthermore, based on continuous full-stack multi-dimensional resource elements, the average service demand time of CPU, GPU, and input / output under standard load is obtained through offline calibration. Based on the queuing theory relationship between input / output latency and average service demand time, input / output utilization is obtained through reverse calculation, and the range of input / output utilization values ​​is narrowed to... Based on CPU utilization, GPU computing power utilization, and input / output utilization, the throughput of the CPU side, GPU side, and input / output side are obtained respectively through the service demand law. According to the bottleneck law, the minimum value among the CPU side throughput, GPU side throughput, and input / output side throughput is taken as the final throughput. The equivalent power is obtained through the power equivalent mapping formula, and the energy response balance coefficient is calculated through the energy response balance acquisition algorithm.

[0065] It should be noted that the equivalent power is obtained through the power equivalent mapping formula, expressed as:

[0066] ;

[0067] in, Indicates equivalent power. This represents the computing power equivalence coefficient. Indicates the final throughput.

[0068] It should be noted that the computing power equivalence coefficient is calculated per 1 GFLOP (10 9 The average energy consumption corresponding to the next floating-point operation is obtained.

[0069] It should be noted that the energy response balance acquisition algorithm is expressed as:

[0070] ;

[0071] in, Represents the energy response balance coefficient. This represents the electrical power of each node in the computing server. Indicates the length of the energy observation window. Represents the integral variable. Represents the zero-point protection constant. Indicates network bandwidth utilization. This represents the historical average bandwidth.

[0072] It should be noted that, The electrical power of each node in the computing server is obtained by collecting the instantaneous power conversion rate of each node. The zero-point protection constant is set offline by calibrating the power integral of the computing server nodes during idle operation. Specifically, when the computing server is in a no-task or low-load state, the power integral within the energy observation window is continuously calculated with the energy observation window length as the sampling period. The power integral set within the energy observation window is obtained, and the minimum stable value of the mean of the non-zero power integral in the power integral set within the energy observation window is used as the value of the zero-point protection constant. The value range of the zero-point protection constant is... .

[0073] The length of the energy observation window is determined by the task latency and resource fluctuation characteristics of the computing server node. Specifically, within the continuous operation cycle of the computing server node, the time series of task latency and full-stack multidimensional resource elements are collected. The average value of the task latency is calculated to obtain the average task latency. By performing autocorrelation analysis on the time series of full-stack multidimensional resource elements, the main period of resource fluctuation is extracted. The geometric mean of the average task latency and the main period of resource fluctuation is used as the length of the energy observation window.

[0074] Furthermore, based on the energy response balance coefficient, the energy efficiency status of computing power and energy consumption is determined, the sampling ratio of computing power and energy is balanced, weighted full-stack multi-dimensional resource elements are obtained, and the sampled values ​​of all weighted full-stack multi-dimensional resource elements are output in a structured manner to generate a structured data vector of full-stack multi-dimensional resource elements.

[0075] It should be noted that the energy efficiency status of computing power and energy consumption is determined based on the energy response balance coefficient. Specifically, when... When computing power and energy consumption are determined to be in a high-energy-efficiency state, the energy consumption sampling frequency is reduced and the priority of bandwidth and input / output sampling frequency is increased. When computing power and energy consumption are in a low-energy-efficiency state, the sampling frequency weights of power consumption and temperature are increased.

[0076] S2. Standardize and aggregate the structured data vectors of full-stack multi-dimensional resource elements to obtain resource standard indicators. By calculating the coupling relationship between computing power, data, and energy efficiency, obtain the comprehensive indicator of the correlation between computing power and data.

[0077] For each resource element in the structured data vector of full-stack multidimensional resource elements, perform linear standardization processing according to a unified interval to obtain standardized data of full-stack multidimensional resource elements.

[0078] Furthermore, based on historical observation data, the upper and lower limits of the linear standardization interval of each resource element in the full-stack multi-dimensional resource element structured data vector are determined. Using a linear interval mapping method, interval normalization and mean correction are performed on each resource element in the full-stack multi-dimensional resource element structured data vector. All resource elements after interval normalization and mean correction are then recombined into resource standard indicators.

[0079] The sliding window mean correction method is used to smooth and correct the standardized data of full-stack multidimensional resource elements, generating corrected full-stack multidimensional resource element data.

[0080] Furthermore, a sliding window mean correction method is adopted. Based on the sampling period and data fluctuation characteristics, the sliding window length is set for the standardized data of full-stack multidimensional resource elements, and the standardized data of full-stack multidimensional resource elements is smoothed and corrected to generate corrected full-stack multidimensional resource element data.

[0081] It should be noted that the corrected full-stack multidimensional resource element data includes corrected CPU utilization, corrected GPU computing power utilization, corrected system memory utilization, corrected network bandwidth utilization, corrected storage access rate, corrected input / output latency, corrected power consumption, corrected device temperature, corrected energy efficiency ratio, and corrected heat dissipation margin.

[0082] The corrected full-stack multidimensional resource element data is divided into a set of computing power elements, a set of communication elements, and a set of energy efficiency elements. A three-dimensional coupled aggregation calculation of computing power, data, and energy efficiency is performed to obtain resource standard indicators.

[0083] Furthermore, the corrected CPU utilization, corrected GPU computing power utilization, and corrected system memory utilization are included in the computing power element set; the corrected network bandwidth utilization, corrected storage access rate, and corrected input / output latency are included in the communication element set; and the corrected power consumption, corrected device temperature, corrected energy efficiency ratio, and corrected heat dissipation margin are included in the energy efficiency element set. Indexes consistent with timestamps are established for each of the computing power element set, communication element set, and energy efficiency element set, and directional consistency processing is performed on the entries in each of these sets. Through a three-dimensional coupled aggregation calculation algorithm of computing power, data, and energy efficiency, the coupling relationship between the computing power element set, communication element set, and energy efficiency element set is constructed, generating resource aggregation values.

[0084] It should be noted that the three-dimensional coupled and aggregated calculation algorithm for computing power, data, and energy efficiency is expressed as follows:

[0085] ;

[0086] in, Represents the aggregated value of resources. Represents a set of computing power elements. Represents a collection of communication elements, Represents the set of energy efficiency factors. This represents the corrected full-stack multidimensional resource element data corresponding to the entries in the computing power element set. This represents the corrected full-stack multidimensional resource element data corresponding to the entries in the communication element set. This represents the corrected full-stack multidimensional resource element data corresponding to the entries in the energy efficiency element set. This represents the weight coefficient corresponding to the item in the computing power element set. This represents the weight coefficient corresponding to the item in the communication element set. This represents the weight coefficient corresponding to the item in the energy efficiency factor set.

[0087] It should be noted that, The specific method for setting up the sampling layer weights is as follows: First, based on the energy response balance coefficient, the energy efficiency status of computing power and energy consumption is determined. After balancing the sampling weights of computing power and energy, the priority weights of various resource samples in the weighted multi-dimensional resource elements of the entire stack are statistically analyzed and divided into initial weight coefficient sets for computing power elements, communication elements, and energy efficiency elements, thus obtaining a sampling layer weight set. Second, based on the stability indicators of various resources in the weighted multi-dimensional resource elements of the entire stack during the current time period, the sampling-confidence weight correction algorithm is used to correct the confidence level of the sampling layer weight set, obtaining the intermediate weights after confidence level correction. The intermediate weights after confidence level correction are then statistically analyzed and divided to obtain an intermediate weight set. Third, using a proportional normalization method, the intermediate weights after confidence level correction in the intermediate weight set are normalized and aggregated to obtain the aggregation layer weight coefficients. Finally, the aggregation layer weight coefficients are divided into weight coefficients corresponding to entries in the computing power element set, the communication element set, and the energy efficiency element set.

[0088] Specifically, by calculating the inverse ratio of the variance of each type of resource in the corresponding time series, the stability index of each type of resource in the weighted full-stack multidimensional resource elements is obtained in the current time period.

[0089] It should be noted that the sampling-confidence weight correction algorithm is as follows:

[0090] ;

[0091] in, Indicates category index, This represents the median weight after confidence level adjustment. This represents the sampling layer weights in the sampling layer weight set. This represents the volatility coefficient.

[0092] It should be noted that, Only one of the initial weight coefficient sets for computing power, communication, and energy efficiency is used; the volatility coefficient is set by calculating the standard deviation of the category index within the most recent observation window and normalizing it, with a value range of [value missing]. .

[0093] Furthermore, all resource aggregate values ​​are stored with timestamps to obtain standard resource metrics.

[0094] Based on resource standard indicators, and combined with sets of computing power elements, communication elements, and energy efficiency elements, a coupled input matrix of computing power, data, and energy efficiency is constructed.

[0095] Furthermore, using the timestamps of resource standard indicators as the primary index, after aligning the timestamps and filling gaps in the sets of computing power elements, communication elements, and energy efficiency elements, an initial input matrix for computing power, data, and energy efficiency is constructed. The numerical domain and unit consistency of the input matrix for computing power, data, and energy efficiency are verified, and abnormal rows in the initial input matrix for computing power, data, and energy efficiency are removed to generate the input matrix for computing power, data, and energy efficiency.

[0096] It should be noted that the gap filling process is as follows: when there is a single missing point in the current timestamp, the most recent valid sample is used to fill the gap to maintain the continuity of the sequence. When the number of consecutive missing points is greater than three, the line is marked as unusable and removed in the final step.

[0097] Based on the coupling input matrix of computing power, data, and energy efficiency, the coupling coefficient is obtained by calculating the coordinated change rate.

[0098] Furthermore, based on the coupling input matrix of computing power, data, and energy efficiency, after aligning the time window with the index, the coupling coefficient is obtained by calculating the cooperative rate of change.

[0099] It should be noted that the calculated rate of co-variance is expressed as:

[0100] ;

[0101] in, This represents the coupling coefficient.

[0102] Based on the coupling coefficient, a comprehensive index of the correlation between computing power and data is obtained by calculating the standard deviation of the set of energy efficiency factors.

[0103] Furthermore, the standard deviation of the energy efficiency factor set is calculated, and a comprehensive index relating computing power and data is obtained through a standard deviation constraint coupling correction algorithm.

[0104] It should be noted that the standard deviation constraint coupling correction algorithm is expressed as:

[0105] ;

[0106] in, This represents a comprehensive indicator of the correlation between computing power and data. This represents the energy efficiency fluctuation suppression coefficient. It represents the standard deviation of the set of energy efficiency factors.

[0107] It should be noted that the energy efficiency fluctuation suppression coefficient is set by fitting the sensitivity relationship between energy efficiency fluctuation and performance loss using historical operating data. Specifically, in a long-running computing server, the standard deviation sequence of the energy efficiency factor set and the corresponding performance loss rate sequence are collected over historical periods. By calculating the regression relationship between the performance loss rate and the change in the standard deviation of the energy efficiency factor set, a sensitivity function between energy efficiency fluctuation and performance loss is constructed. The slope of the linear fit of the sensitivity function is defined as the sensitivity coefficient. Based on the sensitivity coefficient, the least squares method is used to fit the historical sample data to determine the optimal value of the energy efficiency fluctuation suppression coefficient, satisfying the balance between the comprehensive index of computing power and data correlation and the response to energy efficiency fluctuation. The range of the energy efficiency fluctuation suppression coefficient is as follows: .

[0108] S3. Based on the comprehensive index of computing power and data correlation, construct a dynamic feedback window function, collect the performance feedback data of each node, obtain the feedback package, perform runtime sequence normalization processing on the feedback package, and obtain the window feedback dataset.

[0109] A dynamic feedback window function is constructed by calculating the relative rate of change of the comprehensive index of computing power and data correlation.

[0110] It should be noted that the dynamic feedback window function is expressed as:

[0111] ;

[0112] in, This indicates the duration of the dynamic feedback window function. This indicates the initial time length of the dynamic feedback window function. Indicates the window adaptation coefficient. This represents the relative rate of change of a comprehensive indicator relating computing power and data.

[0113] It should be noted that the initial time length of the dynamic feedback window function is determined by statistically analyzing the task execution latency, resource utilization fluctuation cycle, and historical feedback stability data of the computing power server under different business scenarios. The center value of the average fluctuation cycle stability interval of the comprehensive index of computing power and data correlation is taken as the initial time length of the dynamic feedback window function. The window adaptation coefficient is set by a collaborative response calculation method based on the rate of change of the comprehensive index of computing power and data correlation and the fluctuation amplitude of node performance feedback data. Specifically, during the operation of the dynamic feedback window function, the time series rate of change of the comprehensive index of computing power and data correlation is continuously collected, and the mean square error of the node performance feedback data within the same time interval is calculated simultaneously. After normalizing the rate of change of the comprehensive index of computing power and data correlation and the mean square error of the node performance feedback data, the ratio of the normalized rate of change of the comprehensive index of computing power and data correlation to the node performance feedback data is calculated to obtain the instantaneous response ratio. A moving average operation is performed on the instantaneous response ratio within the time length of the dynamic feedback window function, and the average response value obtained after the moving average operation is taken as the value of the window adaptation coefficient. The range of the window adaptation coefficient is as follows. .

[0114] Based on the duration of the dynamic feedback window function, performance feedback data is collected on each computing server node to obtain node performance feedback data.

[0115] Furthermore, using the duration of the dynamic feedback window as a unified sampling period, the local sampling clock of each computing server node is synchronized. Within each computing server node, within the sampling period corresponding to the duration of the dynamic feedback window, multi-dimensional resource elements across the entire stack are continuously collected, performance feedback data is collected, and performance feedback data is cached. When the duration of the dynamic feedback window reaches its termination time, node performance feedback data is generated by calculating the average, maximum, minimum, and standard deviation of the performance feedback data.

[0116] Perform time normalization processing on the node performance feedback data to obtain the node time-series normalized feedback value.

[0117] Furthermore, based on the time length of the dynamic feedback window, the node performance feedback data is divided into equally spaced time series according to the sampling timestamp. The maximum and minimum values ​​of each performance index in the node performance feedback data within the equally spaced time series are used as the normalization boundary. Linear normalization calculation is performed on the time series of each performance index to obtain the node time-series normalized feedback value.

[0118] The comprehensive index of computing power and data association, the time length of the dynamic feedback window function, and the normalized feedback value of node timing are integrated to generate a feedback package.

[0119] Furthermore, using the comprehensive index of computing power and data association as the core control parameter for feedback information integration, the weight allocation logic of the feedback package is determined. The time length of the dynamic feedback window function is used as a unified benchmark for the time dimension. The node time-series normalized feedback values ​​are time-aligned and indexed. The node time-series normalized feedback values ​​are weighted and aggregated according to the weight ratio of the comprehensive index of computing power and data association to generate a node-level comprehensive feedback vector. The node-level comprehensive feedback vector and the corresponding dynamic feedback window timestamp are encapsulated together to generate a feedback package.

[0120] Deconstruct all feedback packets to obtain the globally aligned feedback matrix.

[0121] Furthermore, using the comprehensive index of computing power and data association, the time length of the dynamic feedback window function, and the node time-series normalized feedback value as input, each feedback package is deconstructed at the field level to extract node identifiers, timestamps, and performance feedback parameters. Using the time length of the dynamic feedback window function as a unified time benchmark, time interpolation and synchronization are performed on all node time-series normalized feedback values ​​to fully align the performance feedback data of different nodes in the time dimension. Using the comprehensive index of computing power and data association as a synchronization constraint, cross-node smoothing is performed on the time-aligned node time-series normalized feedback values ​​to maintain the coordinated response trend of computing power, data, and energy efficiency characteristics in the time series. The smoothed node time-series normalized feedback values ​​are arranged according to node identifiers and unified time indices to construct a global aligned feedback matrix.

[0122] Based on the time length of the dynamic feedback window function, time-weighted smoothing is performed on the global aligned feedback matrix to obtain the weighted smoothed feedback value. Global statistical features are obtained by calculating the mean and standard deviation of the weighted smoothed feedback across all nodes.

[0123] Furthermore, using the time length of the dynamic feedback window function as the time weighting parameter, the node time-series normalized feedback value of each time segment in the global aligned feedback matrix is ​​weighted; the weighted feedback values ​​of all nodes are smoothed along the time dimension to obtain the weighted smoothed feedback value; based on the weighted smoothed feedback values ​​of all nodes, the weighted smoothed feedback mean and weighted smoothed feedback standard deviation of all nodes are calculated, and the weighted smoothed feedback mean and weighted smoothed feedback standard deviation of all nodes are statistically analyzed as global statistical features.

[0124] Based on global statistical features, the runtime normalized feedback value is calculated to obtain the window feedback dataset.

[0125] Furthermore, based on global statistical features, a normalization transformation is performed on the node time-series data in the global alignment feedback matrix to obtain runtime time-series normalized feedback values. Using the comprehensive index of computing power and data association as the time series alignment benchmark, the runtime time-series normalized feedback values ​​are mapped to time intervals. The magnitude of the runtime time-series normalized feedback values ​​after time interval mapping is adjusted according to the global statistical features. The adjusted runtime time-series normalized feedback values ​​and the comprehensive index of computing power and data association are structurally integrated according to the time length of the dynamic feedback window function to generate a window feedback dataset.

[0126] It should be noted that the runtime sequence normalization feedback value after mapping the time interval is adjusted based on global statistical features. Specifically, the runtime sequence normalization feedback value is scaled proportionally using the weighted smoothed feedback mean of all nodes as the central benchmark, and the standard deviation of the weighted smoothed feedback of all nodes is used as a dynamic adjustment factor to adjust the response amplitude of different nodes in the process of computing power changes, data throughput and energy efficiency fluctuations.

[0127] S4. Calculate the correlation between the resource standard indicators and the performance indicators in the window feedback dataset. By analyzing the degree of influence of different resource dimensions, dynamically adjust the weight factors, obtain the resource contribution vector, and generate an executable tuning strategy package.

[0128] By calculating the Pearson correlation coefficient matrix between the resource standard indicators and the window feedback dataset, a correlation matrix is ​​obtained. The average absolute correlation is calculated for each resource dimension to obtain the set of influence strengths of the resource dimensions.

[0129] Furthermore, when calculating the Pearson correlation coefficient matrix between the resource standard indicators and the window feedback dataset, the parameters of each resource dimension in the resource standard indicators are used as row vectors, and the performance indicators in the window feedback dataset are used as column vectors. The linear correlation is calculated element by element to form a correlation matrix. The absolute value of the correlation coefficient of each resource dimension on all nodes in the correlation matrix is ​​taken and the average value is calculated to obtain the average absolute correlation of all nodes. The average absolute correlation of all nodes of all resource dimensions is arranged in order of resource type to generate a set of resource dimension influence intensity.

[0130] By combining the set of resource dimension influence intensity and the set of historical weight factors, the weight factors are corrected through a time decay dynamic update formula to obtain a dynamic set of weight factors.

[0131] Furthermore, using the influence intensity values ​​of each resource dimension in the resource dimension influence intensity set as the reference for weight adjustment at the current moment, time series matching is performed on the weight values ​​of the corresponding resource dimensions in the historical weight factor set, and the results of time series matching are weighted and calculated using a time decay dynamic update formula.

[0132] It should be noted that the dynamic update formula for time decay is expressed as:

[0133] ;

[0134] in, Represents dynamic weighting factors. This represents the set of influence strengths of resource dimensions. Represents historical weighting factors. This represents the current weight adjustment coefficient. Indicates the time decay coefficient. This indicates the time interval for updating the weighting factors. This represents the index of a specific dimension within the corresponding set of resource standard indicators.

[0135] It should be noted that the current weight adjustment coefficient is set by calculating the normalized ratio of the standard deviation of the resource dimension influence intensity set to the variance of the historical weight factor set. Specifically, the standard deviation of the resource dimension influence intensity set within the current time window is calculated, and the variance of the historical weight factor set within the previous time window is calculated. The ratio obtained after interval normalization of the standard deviation of the resource dimension influence intensity set within the current time window to the variance of the historical weight factor set within the previous time window is used as the value of the current weight adjustment coefficient. The range of the current weight adjustment coefficient is as follows: The time decay coefficient is set using historical load fluctuation cycles and resource feedback stability analysis methods. Specifically, time-series analysis is performed on the changes in resource dimension weights of the computing server over multiple dynamic feedback window functions. By calculating the autocorrelation function of the weight changes, the average stabilization period of the resource weights is determined. The time decay coefficient is then determined based on the average stabilization period of the resource weights, ensuring that the resource weights decay to their initial values ​​after one complete stabilization period. The range of values ​​for the horizontal and time decay coefficients is as follows: .

[0136] Based on a dynamic set of weighting factors and resource standard indicators, a resource contribution vector is calculated and combined with a comprehensive indicator of computing power and data correlation to generate an executable optimization strategy package.

[0137] Furthermore, based on the dynamic weight factor set and resource standard indicators, the resource standard indicators are multiplied dimension-wise with the corresponding weights in the dynamic weight factor set to obtain the comprehensive contribution value of different resource dimensions under the current computing power, data, and energy efficiency synergy state. All comprehensive contribution values ​​are normalized to generate a resource contribution vector. Using the comprehensive index of computing power and data correlation as a global adjustment factor, the resource contribution vector is scaled proportionally. The scaled resource contribution vector is then encapsulated into a parameterized structure according to the resource dimension order to generate an executable optimization strategy package.

[0138] S5. After applying safety limits to the execution parameters in the optimization strategy package, distribute the optimization strategy to each node, execute the optimization strategy and verify the resource status in real time, calculate the comprehensive performance improvement value, and adjust the full-stack multi-dimensional resource elements according to the comprehensive performance improvement value.

[0139] Based on resource standard indicators and historical stable parameters, the upper and lower safety limits of each resource dimension are calculated and combined with executable tuning strategy packages to generate a safety tuning strategy package.

[0140] Furthermore, based on resource standard indicators and historical stable parameters, the upper and lower safety limits of each resource dimension are calculated. The execution parameters of each resource dimension in the executable tuning strategy package are compared with the corresponding upper and lower safety limits. When the execution parameters exceed the corresponding upper and lower safety limits, a limit correction is performed according to the boundary values ​​to generate a set of limit-adjusted safe execution parameters. Using the comprehensive index of computing power and data association as a global constraint factor, the limit-adjusted safe execution parameters are uniformly weighted and adjusted. The weighted and adjusted limit-adjusted safe execution parameters are integrated and encoded with the comprehensive index of computing power and data association to generate a safety tuning strategy package.

[0141] It should be noted that historical stability parameters include: the historical mean and variance of resource standard indicators over multiple periods, used to measure the degree of long-term resource fluctuation; the mean and variance of operating deviations in performance feedback data, used to reflect execution stability; the fluctuation range of energy efficiency ratio and power consumption ratio in multiple operating periods, used to assess energy consumption stability; and the periodic stability indicators of network bandwidth utilization, input / output latency, and storage access rate, used to quantify the long-term controllability of communication and storage dimensions.

[0142] It should be noted that the upper and lower security boundaries of the resource dimension include the upper security boundary and the lower security boundary of the resource dimension.

[0143] It should be noted that the safety upper and lower bounds for each resource dimension are calculated as follows:

[0144] ;

[0145] in, This represents the lower bound of security in the resource dimension. Indicates the upper limit boundary of the resource dimension. Indicates resource standard indicators, This represents the safety fluctuation coefficient.

[0146] It should be noted that the safety fluctuation coefficient is set through variance analysis and normalization mapping algorithms of historical stable parameters to ensure the stability of each resource dimension during multi-period operation. Specifically, based on historical stable parameters, historical performance sequences of each resource dimension during multi-period operation are extracted to construct a set of historical stable parameter time series, and the variance of the historical stable parameter time series for each resource dimension is calculated to obtain a set of historical stable variances. The max-min normalization method is used to normalize the set of historical stable variances to obtain a normalized set of historical stable variances. A monotonically increasing smoothing mapping function is used to perform normalization mapping operations, mapping the normalized variance value of each resource dimension to the corresponding safety fluctuation coefficient value, generating a set of safety fluctuation coefficients. The set of safety fluctuation coefficients is smoothed and corrected according to the sliding window length within each time period to obtain a smoothed set of safety fluctuation coefficients. The safety fluctuation coefficient is set according to the statistical characteristics of historical stable parameters during multi-period operation to ensure that the upper and lower safety limits of the resource dimensions can dynamically match the long-term operating characteristics of the computing server and prevent abnormal resource expansion. The value range of the safety fluctuation coefficient is [insert range here]. .

[0147] Through a unified scheduling interface, security optimization strategy packages are distributed to each computing server node. The nodes then adjust their resources according to the parameters in the security optimization strategy packages and obtain updated node resource status data.

[0148] Furthermore, the security execution parameters of each resource dimension, the comprehensive indicators of computing power and data association, and the corresponding timestamp information in the security tuning strategy package are indexed and matched according to the node identifier to generate a node allocation instruction set. According to the time synchronization mechanism, the node allocation instruction set is distributed to each computing power server node through a unified scheduling interface. After receiving the security tuning strategy package, each computing power server node performs real-time resource adjustment according to the corresponding resource dimension security execution parameters. After executing the security tuning strategy package, each computing power server node obtains the updated node resource status data by collecting the updated full-stack multi-dimensional resource elements within the time length of the dynamic feedback window function.

[0149] It should be noted that real-time resource adjustments are performed based on the corresponding resource dimension security execution parameters. For example, based on the CPU security execution parameters in the security execution parameter set, the CPU core frequency and thread allocation are adjusted; based on the GPU security execution parameters in the security execution parameter set, the GPU task ratio and video memory allocation rate are adjusted; based on the system memory security execution parameters in the security execution parameter set, memory page replacement is performed; based on the network bandwidth security execution parameters in the security execution parameter set, the bandwidth allocation weight is adjusted; and based on the power consumption security execution parameters in the security execution parameter set, power consumption limits and power saving strategies are implemented.

[0150] Based on the updated node resource status data, a set of resource statuses is collected, a set of performance improvement differences is calculated, and a comprehensive performance improvement value is generated by combining the resource contribution vector and the set of dynamic weight factors.

[0151] Furthermore, based on the updated node resource status data, the monitoring time series of all computing server nodes are time-aligned according to the time length of the dynamic feedback window function. The updated node resource status data is indexed, matched, and standardized by timestamp. Using the node identifier as the index key and the timestamp as the sorting basis, the synchronized data of all resource dimensions are combined to form a multi-dimensional time series matrix, generating a resource status set. The resource status set is then differentially calculated with the resource standard indicator set before the previous round of optimization to obtain a performance improvement difference set. Using the resource contribution vector as the influence weight and the dynamic weight factor set as the time decay correction term, the performance improvement difference set is weighted and summed to generate a comprehensive performance improvement value.

[0152] Performance verification is performed based on the comprehensive performance improvement value, a performance verification status identifier is generated, and the full-stack multi-dimensional resource elements are adjusted according to the comprehensive performance improvement value to obtain a new set of full-stack multi-dimensional resource elements.

[0153] Furthermore, a performance verification threshold is set, and the performance verification threshold and the comprehensive performance improvement value are used to determine the performance. A performance verification status identifier is generated, and the multi-dimensional resource elements of the whole stack are adjusted differently according to the performance verification status identifier to obtain a new set of multi-dimensional resource elements of the whole stack.

[0154] It should be noted that the performance verification threshold is set as follows: a baseline distribution is constructed using samples of comprehensive performance improvement values ​​within the time frame of the dynamic feedback window function before tuning, and the upper quantile is taken as the upper bound of the noise; the sum of the noise upper bound and the minimum improvement step size given by the business SLA (Service Level Agreement) is taken as the lower bound of the performance verification threshold; a hysteresis margin is set based on the short-term fluctuation variance of performance indicators in the full-stack multi-dimensional resource elements, and the sum of the lower bound of the performance verification threshold and the hysteresis margin is taken as the upper bound of the performance verification threshold. The range of the lower bound of the performance verification threshold is as follows: The upper limit of the performance verification threshold can be set to the following values: .

[0155] It should be noted that the performance verification threshold and the comprehensive performance improvement value are used to determine the performance verification. Specifically, when the comprehensive performance improvement value is greater than or equal to the upper limit of the performance verification threshold, the performance verification status is generated as optimized and meets the standard; when the comprehensive performance improvement value is lower than the lower limit of the performance verification threshold, the performance verification status is generated as needing adjustment.

[0156] It should be noted that the multi-dimensional resource elements of the whole stack are adjusted differently according to the performance verification status. Specifically, when the performance verification status indicates that the optimization has been achieved, the resource allocation strategy is maintained and the resource weights are gradually converged in the next tuning cycle. When the performance verification status indicates that adjustment is needed, the resource dimensions in the resource standard indicators are reversed according to the deviation direction of the comprehensive performance improvement value. A new set of multi-dimensional resource elements of the whole stack is obtained by updating and collecting the multi-dimensional resource elements of the whole stack.

[0157] S6. Collect execution feedback information and write it back to the control plane, generating a receipt message.

[0158] Collect node operation logs and resource status data of each computing server node after executing the security tuning strategy package, and obtain execution feedback information.

[0159] Furthermore, a node operation monitoring process is initiated on each computing server node to collect real-time multi-dimensional resource elements across the entire stack after the execution of the security tuning strategy package. The node operation monitoring process generates time-synchronized node operation logs based on the execution parameters in the security tuning strategy package. Based on the time synchronization identifier of each computing server node, the timestamp field in the node operation log is matched one-to-one with the collection time field in the node resource status data. Using the node identifier, task number, and sampling time in the security tuning strategy package as the joint primary key, the operation event records in the node operation log are mapped at the field level to the performance index values ​​in the node resource status data. The associated results are then aggregated in intervals according to the time length of the dynamic feedback window function to form execution feedback information.

[0160] The execution feedback information is written to the control plane feedback database, generating an acknowledgment message.

[0161] Furthermore, a unique index key is constructed based on the node identifier, task number, and sampling time generated by each computing server node when executing the security tuning strategy package. The execution feedback information is written into the feedback information table in the control plane feedback database one by one through the control plane data interface. During the writing process, data consistency verification and time synchronization verification are performed. After the writing is completed, the control plane feedback database generates a write-back status identifier, and the control plane generates a receipt message based on the write-back status identifier.

[0162] In summary, this invention achieves real-time synchronization and timing alignment of the entire stack resource state by constructing a dynamic feedback window function and performing runtime timing normalization on the feedback package; it also achieves multi-dimensional resource adaptive optimization by dynamically updating the execution time decay and dynamically adjusting the weight factor by analyzing the influence of resource dimensions, thereby improving the system's resource coordination efficiency and ensuring the accuracy of the tuning strategy.

[0163] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A dynamic optimization method for full-stack resource big data analysis of computing servers, characterized in that: include, Collect full-stack multi-dimensional resource elements of computing power servers and integrate them into a structured data vector of full-stack multi-dimensional resource elements; Standardize and aggregate the structured data vectors of full-stack multi-dimensional resource elements to obtain resource standard indicators. By calculating the coupling relationship between computing power, data, and energy efficiency, obtain a comprehensive indicator of the correlation between computing power and data. Based on the comprehensive index of computing power and data correlation, a dynamic feedback window function is constructed to collect the performance feedback data of each node, obtain the feedback package, perform runtime sequence normalization processing on the feedback package, and obtain the window feedback dataset. The correlation between resource standard indicators and performance indicators in the window feedback dataset is calculated. By analyzing the degree of influence of different resource dimensions, the weight factors are dynamically adjusted to obtain the resource contribution vector and generate an executable tuning strategy package. After the execution parameters in the tuning strategy package are subject to safety limits, they are sent to each node. The tuning strategy is executed and the resource status is verified in real time. The comprehensive performance improvement value is calculated, and the multi-dimensional resource elements of the whole stack are adjusted according to the comprehensive performance improvement value. Collect execution feedback information and write it back to the control plane, generating a receipt message; The standardization and aggregation of the structured data vectors of full-stack multi-dimensional resource elements to obtain resource standard indicators are performed as follows: Perform linear standardization on each resource element in the full-stack multi-dimensional resource element structured data vector according to a unified interval to obtain full-stack multi-dimensional resource element standardized data. A sliding window mean correction method is used to smooth and correct the standardized data of full-stack multidimensional resource elements, generating corrected full-stack multidimensional resource element data. The corrected full-stack multidimensional resource element data is divided into computing power element set, communication element set and energy efficiency element set. Three-dimensional coupled aggregation calculation of computing power, data and energy efficiency is performed to obtain resource standard indicators. The process of calculating the coupling relationship between computing power, data, and energy efficiency to obtain a comprehensive index linking computing power and data involves the following steps. Based on resource standard indicators, and combined with sets of computing power elements, communication elements, and energy efficiency elements, a coupled input matrix of computing power, data, and energy efficiency is constructed. Based on the coupling input matrix of computing power, data, and energy efficiency, the coupling coefficient is obtained by calculating the coordinated change rate; Based on the coupling coefficient, a comprehensive index of the correlation between computing power and data is obtained by calculating the standard deviation of the set of energy efficiency factors.

2. The method for dynamic optimization of full-stack resource big data analysis for computing servers as described in claim 1, characterized in that: The steps for collecting full-stack multi-dimensional resource elements from the computing power server and integrating them into a structured data vector of full-stack multi-dimensional resource elements are as follows. Collect full-stack multi-dimensional resource elements, perform time synchronization processing on the full-stack multi-dimensional resource elements, and perform interpolation correction on the full-stack multi-dimensional resource elements with missing time points to obtain continuous full-stack multi-dimensional resource elements. Based on continuous full-stack multi-dimensional resource elements, an energy response balance acquisition algorithm is used to calculate the energy response balance coefficient and generate a structured data vector of full-stack multi-dimensional resource elements.

3. The method for dynamic optimization of full-stack resource big data analysis for computing servers as described in claim 2, characterized in that: The process involves constructing a dynamic feedback window function based on a comprehensive index linking computing power and data, collecting performance feedback data from each node, and obtaining feedback packets. The steps are as follows: A dynamic feedback window function is constructed by calculating the relative rate of change of the comprehensive index of computing power and data correlation. Based on the time length of the dynamic feedback window function, performance feedback data is collected on each computing server node to obtain node performance feedback data. Perform time normalization processing on the node performance feedback data to obtain the node time-series normalized feedback value; The comprehensive index of computing power and data association, the time length of the dynamic feedback window function, and the normalized feedback value of node timing are integrated to generate a feedback package.

4. The method for dynamic optimization of full-stack resource big data analysis for computing servers as described in claim 3, characterized in that: The steps for performing runtime normalization on the feedback packets to obtain the window feedback dataset are as follows. Deconstruct all feedback packets to obtain the globally aligned feedback matrix; Based on the time length of the dynamic feedback window function, time-weighted smoothing calculation is performed on the global aligned feedback matrix to obtain the weighted smoothed feedback value. Global statistical features are obtained by calculating the mean and standard deviation of the weighted smoothed feedback across all nodes. Based on global statistical features, the runtime normalized feedback value is calculated to obtain the window feedback dataset.

5. The method for dynamic optimization of full-stack resource big data analysis for computing servers as described in claim 4, characterized in that: The steps for calculating the correlation between resource standard indicators and performance indicators in the window feedback dataset, analyzing the influence of different resource dimensions, dynamically adjusting weight factors, obtaining resource contribution vectors, and generating an executable tuning strategy package are as follows. By calculating the Pearson correlation coefficient matrix between the resource standard indicators and the window feedback dataset, the correlation matrix is ​​obtained. The average absolute correlation is calculated for each resource dimension to obtain the set of influence strength of the resource dimension. By combining the set of influence intensity of resource dimensions and the set of historical weight factors, the weight factors are corrected through a time decay dynamic update formula to obtain a dynamic set of weight factors; Based on a dynamic set of weighting factors and resource standard indicators, a resource contribution vector is calculated and combined with a comprehensive indicator of computing power and data correlation to generate an executable optimization strategy package.

6. The method for dynamic optimization of full-stack resource big data analysis for computing servers as described in claim 5, characterized in that: The steps for implementing the optimization strategy package, including applying safety limits to the execution parameters, are as follows: The package is then distributed to each node for execution of the optimization strategy. Based on resource standard indicators and historical stable parameters, calculate the upper and lower safety limits of each resource dimension and generate a safety tuning strategy package by combining it with an executable tuning strategy package. Through a unified scheduling interface, security optimization strategy packages are distributed to each computing server node. The nodes then adjust their resources according to the parameters in the security optimization strategy packages and obtain updated node resource status data.

7. The method for dynamic optimization of full-stack resource big data analysis for computing servers as described in claim 6, characterized in that: The steps for real-time verification of resource status, calculation of comprehensive performance improvement value, and adjustment of full-stack multi-dimensional resource elements based on the comprehensive performance improvement value are as follows: Based on the updated node resource status data, a set of resource status is collected, a set of performance improvement differences is calculated, and a comprehensive performance improvement value is generated by combining the resource contribution vector and the set of dynamic weight factors. Performance verification is performed based on the comprehensive performance improvement value, a performance verification status identifier is generated, and the full-stack multi-dimensional resource elements are adjusted according to the comprehensive performance improvement value to obtain a new set of full-stack multi-dimensional resource elements.

8. The method for dynamic optimization of full-stack resource big data analysis for computing servers as described in claim 7, characterized in that: The steps for collecting execution feedback information, writing it back to the control plane, and generating a receipt message are as follows: Collect node operation logs and resource status data of each computing server node after executing the security tuning strategy package, and obtain execution feedback information; The execution feedback information is written to the control plane feedback database, generating an acknowledgment message.

Images