Server performance prediction method, electronic device, storage medium and program product

Abstract

The present application discloses a server performance prediction method, electronic device, storage medium and program product, which relate to the technical field of server performance prediction. In particular, during data collection, not only the time series data of the server in a preset monitoring period is collected, but also the data change sequence of the server determined according to a preset time interval in the preset monitoring period is collected, thereby realizing automatic adjustment of the sampling frequency according to real-time data changes and adaptive dynamic sampling. Through multi-scale decomposition, feature information at different frequencies can be effectively extracted, which helps to mine the deep data structure of the data and provides high-quality input for subsequent cross-modal information fusion. Through cross-modal information fusion, key information can be extracted, the influence of irrelevant noise can be reduced, the quality of feature representation can be improved, and accurate prediction of server performance can be achieved. This solves the technical problem of inaccurate server performance prediction results and achieves the technical effect of improving the accuracy of performance prediction results.

Description

Technical Field

[0001] The present application relates to the technical field of server performance prediction, and in particular to a server performance prediction method, electronic device, storage medium, and program product. Background Art

[0002] Server performance prediction technology aims to predict various server performance indicators over the next period of time, such as central processing unit (CPU) utilization, memory usage, and disk input / output (I / O), through historical data analysis and model training. This technology is crucial for server resource management, fault prevention, and load balancing. Accurate performance prediction can proactively identify potential issues, optimize resource allocation, and improve system stability and efficiency.

[0003] Therefore, it is very necessary to implement a solution that can accurately predict server performance. Summary of the Invention

[0004] The present application provides a server performance prediction method, electronic device, storage medium and program product to at least solve the problem of inaccurate server performance prediction results in related technologies.

[0005] This application provides a server performance prediction method, including:

[0006] Obtaining time series data of at least one server in the server cluster during a preset monitoring period and a data variation sequence determined at preset time intervals, where the end time of the preset monitoring period is the current time;

[0007] Determining, based on the time series data and the data variation sequence, a dual-coded feature vector corresponding to the at least one server, the dual-coded feature vector including a time domain feature obtained by time domain coding and a time domain feature obtained by frequency domain coding and conversion;

[0008] Performing multi-scale decomposition on the double-coded feature vector to obtain a time-frequency domain joint feature;

[0009] Performing cross-modal information fusion on the time-frequency domain joint features to obtain a spatiotemporal feature matrix;

[0010] Performing server topology modeling based on the spatiotemporal feature matrix to obtain a node-level state representation of the at least one server;

[0011] A timing prediction is performed based on the node-level state representation to generate a performance prediction result corresponding to the at least one server.

[0012] The present application also provides a server performance prediction device, comprising:

[0013] A data acquisition module, configured to acquire time series data of at least one server in a server cluster during a preset monitoring period and a data variation sequence determined at preset time intervals, wherein the end time of the preset monitoring period is the current time;

[0014] an encoding module, configured to determine, based on the time series data and the data variation sequence, a dual-encoded feature vector corresponding to the at least one server, the dual-encoded feature vector comprising a time domain feature obtained by time domain encoding and a time domain feature obtained by frequency domain encoding and conversion;

[0015] A decomposition module, configured to perform multi-scale decomposition on the double-coded feature vector to obtain a joint feature in the time-frequency domain;

[0016] A feature extraction module is used to perform cross-modal information fusion on the time-frequency domain joint features to obtain a spatiotemporal feature matrix;

[0017] a modeling module, configured to perform server topology modeling based on the spatiotemporal feature matrix to obtain a node-level state representation of the at least one server;

[0018] A prediction module is used to perform time series prediction based on the node-level state representation to generate a performance prediction result corresponding to the at least one server.

[0019] The present application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for implementing the steps of any of the above-mentioned server performance prediction methods when executing the computer program.

[0020] The present application also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned server performance prediction methods are implemented.

[0021] The present application also provides a computer program product, including a computer program, which implements the steps of any of the above-mentioned server performance prediction methods when executed by a processor.

[0022] Through the present application, during data collection, not only the time series data of the server in the preset monitoring period is collected, but also the data change sequence determined by the server in the preset monitoring period according to the preset time interval is collected. The data change sequence reflects the data fluctuation of the server in the preset monitoring period, realizes automatic adjustment of the sampling frequency according to real-time data changes, and realizes adaptive dynamic sampling; by encoding the collected time series data and data change sequence in the time domain and frequency domain, a dual-encoded feature vector is obtained, which combines the information of the data in the time domain and frequency domain, enhances the comprehensiveness and accuracy of the feature representation, and can effectively extract feature information at different frequencies through multi-scale decomposition, which helps to explore the deep data structure of the data and provide high-quality input for subsequent cross-modal information fusion. Through cross-modal information fusion, key information can be extracted, the influence of irrelevant noise is reduced, and the quality of feature representation is improved. By performing server topology modeling based on a high-quality spatiotemporal feature matrix, the node-level state representation of the server is obtained, and then the performance of the server is predicted based on the node-level state representation, thereby realizing accurate prediction of the server performance. Therefore, the technical problem of inaccurate server performance prediction results can be solved, and the technical effect of improving the accuracy of performance prediction results can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0024] Figure 1 A flow chart of a server performance prediction method provided in Example 1 of the present application;

[0025] Figure 2 A flow chart of a server performance prediction method provided in Example 2 of the present application;

[0026] Figure 3 A flow chart of a server performance prediction method provided in Example 3 of the present application;

[0027] Figure 4 A flow chart of a server performance prediction method provided in Example 4 of the present application;

[0028] Figure 5 A flowchart of a server performance prediction method provided in Example 5 of the present application;

[0029] Figure 6 A flow chart of a server performance prediction method provided in Example 6 of the present application;

[0030] Figure 7 A schematic diagram of the structure of a server performance prediction device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0031] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0032] It should be noted that, in the description of this application, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. The terms "first," "second," etc., in this application are used to distinguish similar objects, and are not used to describe a particular order or sequence.

[0033] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0034] Server performance prediction technology aims to predict various server performance indicators over the future by analyzing historical data and training models. Currently, related technologies in the field of server performance prediction typically use a fixed sampling frequency when collecting data. This can easily lead to insufficient sampling frequency when the data fluctuates drastically, and oversampling when the data is relatively stable. Furthermore, traditional performance prediction methods often focus only on surface features and ignore the deeper information structure within the data, resulting in inaccurate server performance prediction results.

[0035] In response to the above problems, the present disclosure provides a server performance prediction solution, which realizes adaptive dynamic data collection by collecting a sequence of data changes determined at preset time intervals within a preset monitoring period, and can automatically adjust the sampling frequency according to real-time data changes. This not only improves the flexibility and efficiency of data collection, but also reduces unnecessary resource consumption while ensuring data accuracy. It is particularly suitable for scenarios where data change rates are different in server clusters, ensuring that more refined data is obtained when data fluctuates violently, and reducing the sampling frequency during data stability to save resources; through dual-stream time coding, the collected time-aligned mixed frequency indicator data is subjected to spatiotemporal alignment processing, thereby realizing the ability to extract subtle change features as well as long-term trends and periodic components from the original data; the enhanced time domain coding function captures short-term changes, while the frequency domain coding reveals the long-term patterns in the data; by using a trainable wavelet neural network to perform multi-scale decomposition of the eigenvectors of the unified time base, a fine analysis of different frequency components is realized. The application of wavelet not only enhances the ability to capture signal details, but also further improves flexibility and adaptability by introducing a trainable mechanism to optimize parameter settings. By adopting a gated cross-attention mechanism to perform cross-modal fusion of frequency-domain and time-domain joint features, effective integration of information from different modalities is achieved. This mechanism allocates attention weights by calculating the similarity between queries, keys, and values, and combines gated units to regulate information flow, thereby enhancing the focus on key information and reducing the impact of irrelevant noise, thereby significantly improving the quality of feature representation. The high-quality features finally obtained are used for modeling and predicting server performance, ensuring that the prediction results are both highly accurate and capable of handling complex practical application scenarios.

[0036] The embodiments of the present application provide a server performance prediction method, and the method is described in detail in conjunction with the execution process of the server performance prediction method.

[0037] Figure 1 This is a flow chart of a server performance prediction method provided in Example 1 of the present application. The method can be executed by the server performance prediction device provided in the embodiment of the present disclosure. The device can be implemented in software and / or hardware and can be integrated into an electronic device.

[0038] like Figure 1 As shown, the server performance prediction method includes the following steps:

[0039] Step 101: Obtain time series data of at least one server in a server cluster in a preset monitoring period and a data variation sequence determined according to a preset time interval, where the end time of the preset monitoring period is the current time.

[0040] The preset monitoring period is greater than the preset time interval. The preset monitoring period can be set based on actual needs, for example, the preset monitoring period can be set to one week, one day, or three hours. The preset time interval can also be set based on actual needs, and can be combined with the server's data collection frequency to set the preset time interval based on actual needs and resource availability. For example, if the server's data collection frequency is once per second, the preset time interval can be set to 3 seconds, 5 seconds, or the like. This disclosure does not limit the values ​​of the preset monitoring period and the preset time interval.

[0041] In this embodiment, the end time of the preset monitoring period is the current time. In other words, the acquired data is the time series data from a certain historical moment to the current moment for at least one server, as well as the data change sequence determined at preset time intervals within the preset monitoring period. The time interval between a certain historical moment and the current moment is the preset monitoring period. For example, if the preset monitoring period is one hour, the acquired data is the time series data from the previous hour to the current moment, as well as the data change sequence determined at preset time intervals within that hour.

[0042] Among them, each server in the server cluster collects data according to its own data collection frequency, and obtains status data at each collection moment (different time points), such as CPU utilization, memory usage, etc. In this embodiment, the status data of at least one server in the server cluster is monitored at different time points within a preset monitoring period to obtain time series data, and the data change amount at each time point within the preset monitoring period is determined according to the preset time interval to form a data change amount sequence. i , for each time point within the preset monitoring period (denoted as t ), the corresponding data change can be expressed by the following formula (1):

[0043] (1)

[0044] in, Indicates the i Servers at time t The amount of data change, Indicates the i Servers at time t Status data, Indicates the i Servers at time point ( ) status data, Indicates the preset time interval.

[0045] Through the above-mentioned data collection method of determining the data change sequence according to the preset time interval, it is possible to collect multi-source heterogeneous data of the server cluster, obtain time-aligned mixed-frequency indicator data, and realize adaptive dynamic sampling. This sampling method can automatically adjust the sampling frequency according to real-time data changes, which not only improves the flexibility and efficiency of data collection, but also reduces unnecessary resource consumption while ensuring data accuracy. It is particularly suitable for situations where the data change rate in the server cluster is different, ensuring that more detailed data is obtained when the data fluctuates violently, and reducing the sampling frequency to save resources during data stability.

[0046] Step 102: Determine a dual-coded feature vector corresponding to at least one server based on the time series data and the data variation sequence. The dual-coded feature vector includes a time domain feature obtained by time domain coding and a time domain feature obtained by frequency domain coding and conversion.

[0047] In this embodiment, after obtaining the time series data and data change sequence of at least one server, for each server, the time series data and data change sequence corresponding to the server can be encoded in two different ways, namely, in the time domain and in the frequency domain, to obtain a double-encoded feature vector.

[0048] Exemplarily, a dual-stream time coding method can be used to perform spatiotemporal alignment processing on the time-aligned mixed frequency indicator data obtained in the previous step to obtain a dual-coded feature vector with a unified time base. The dual-stream time coding includes time domain coding and frequency domain coding, and the resulting dual-coded feature vector includes the time domain features obtained by time domain coding and the time domain features obtained by frequency domain coding and conversion.

[0049] Step 103: Perform multi-scale decomposition on the double-coded feature vector to obtain a joint feature in the time and frequency domains.

[0050] In this embodiment, after obtaining the dual-coded feature vector corresponding to each server, multi-scale decomposition can be further performed on each server's dual-coded feature vector to obtain a joint time-frequency domain feature. Because the dual-coded feature vector is obtained through both time-domain coding and frequency-domain coding, and includes features in both the time and frequency domains, the features obtained after decomposition still include features in both the time and frequency domains, and are therefore called joint time-frequency domain features.

[0051] The dual-encoded feature vectors of each server can be decomposed using a commonly used multi-scale decomposition algorithm. This disclosure does not limit the specific multi-scale decomposition algorithm used. For example, a wavelet transform can be used to perform multi-scale decomposition on the dual-encoded feature vectors of each server to separate short-term fluctuations and long-term trends. Alternatively, a pre-trained time series decomposition model can be used to perform multi-scale decomposition on the dual-encoded feature vectors of each server.

[0052] Step 104: perform cross-modal information fusion on the joint time-frequency domain features to obtain a spatiotemporal feature matrix.

[0053] In this embodiment, cross-modal information fusion can be performed on the time-frequency domain joint features of each server obtained by decomposition to obtain an enhanced spatiotemporal feature matrix.

[0054] For example, based on the attention mechanism, cross-modal information fusion can be performed on the joint time-frequency domain features of each server to obtain the spatiotemporal feature matrix corresponding to each server.

[0055] Step 105: Perform server topology modeling based on the spatiotemporal feature matrix to obtain a node-level state representation of at least one server.

[0056] In this embodiment, after obtaining the spatiotemporal feature matrix for each server, server topology modeling can be performed based on these spatiotemporal feature matrices to obtain node-level state representations of these servers. The node-level state representation reflects the topological relationship between server nodes. It aggregates the feature information of neighboring nodes and updates the features of the current node, making the feature representation of each server node more accurate. This provides strong support for subsequent performance prediction and ensures the reliability and accuracy of server performance predictions.

[0057] For example, the spatiotemporal feature matrix of each server can be input into a pre-trained graph convolutional network for modeling, and the graph convolutional network outputs the corresponding node-level state representation.

[0058] Step 106: Perform time series prediction based on the node-level state representation to generate a performance prediction result corresponding to at least one server.

[0059] In this embodiment, after obtaining the node-level state representation of the server, time series prediction can be performed based on the node-level state representation to obtain the performance prediction result corresponding to each server.

[0060] Exemplarily, a multi-scale time series Transformer model can be used to perform time series prediction on the node-level state representation. After the encoder of the multi-scale time series Transformer model extracts the input node-level state representation, the decoder of the multi-scale time series Transformer model makes predictions based on the input features to generate server performance indicators of each server in a certain time period in the future as performance prediction results.

[0061] The server performance prediction method of the disclosed embodiment not only collects the time series data of the server in a preset monitoring period, but also collects the data change sequence of the server determined at preset time intervals within the preset monitoring period. The data change sequence reflects the data fluctuation of the server in the preset monitoring period, realizes automatic adjustment of the sampling frequency according to real-time data changes, and realizes adaptive dynamic sampling. The collected time series data and data change sequence are encoded in the time domain and frequency domain to obtain a dual-encoded feature vector, which combines the information of the data in the time domain and frequency domain, enhances the comprehensiveness and accuracy of the feature representation, and can effectively extract feature information at different frequencies through multi-scale decomposition, which helps to explore the deep data structure of the data and provide high-quality input for subsequent cross-modal information fusion. Through cross-modal information fusion, key information can be extracted, the influence of irrelevant noise can be reduced, and the quality of feature representation can be improved. By performing server topology modeling based on a high-quality spatiotemporal feature matrix, a node-level state representation of the server is obtained, and the server performance is predicted based on the node-level state representation, thereby realizing accurate prediction of server performance. Therefore, the technical problem of inaccurate server performance prediction results can be solved, and the technical effect of improving the accuracy of performance prediction results can be achieved.

[0062] In an optional embodiment of the present disclosure, Figure 2 As shown, based on the above embodiment, step 102 may include the following sub-steps:

[0063] Step 201: Perform time domain coding processing on the time series data to obtain a first time domain feature.

[0064] In this embodiment, for each server, for the time series data of the server, an enhanced time domain coding function can be used to capture the subtle change characteristics in the original time series data to obtain a time domain feature (for ease of description and distinction, referred to as the first time domain feature).

[0065] Among them, the time domain coding function can be expressed as shown in the following formula (2):

[0066] (2)

[0067] in, x tRepresents a time point in time series data t Corresponding data points; Express x t The eigenvalues ​​obtained by time domain coding; α It is a preset amplitude parameter used to control the amplitude of the sine waveform; is the preset frequency parameter, is the preset phase offset.

[0068] Step 202: Perform frequency domain encoding on the data variation sequence and then convert it into the time domain to obtain a second time domain feature.

[0069] In this embodiment, for each server, a frequency domain coding function may be introduced for the data variation sequence of the server to extract the long-term trend and periodic components of the data in the data variation sequence.

[0070] Assumptions y ( ω ) is obtained by Fast Fourier Transform (FFT) from The frequency domain representation obtained is then frequency domain coding The expression can be expressed as the following formula (3):

[0071] (3)

[0072] in, y ( ω ) represents the frequency domain representation of the data, j represents the imaginary unit, ω represents the angular frequency variable.

[0073] In this embodiment, by performing frequency domain coding on the data variation sequence, different frequency components in the data can be better separated and processed.

[0074] Next, the obtained frequency domain coding features can be converted to the time domain by calculating the inverse transform from the frequency domain to the time domain to obtain the time domain features (for ease of description and distinction, referred to as the second time domain features). The conversion formula from the frequency domain to the time domain can be expressed as shown in the following formula (4):

[0075] (4)

[0076] in, F -1 represents the inverse Fourier transform operation, Indicates the converted i Servers at time t The second time domain eigenvalue of .

[0077] Step 203: Determine a dual-coding feature vector based on the first time domain feature and the second time domain feature.

[0078] In this embodiment, after the first time domain feature and the second time domain feature of each server are determined, the dual-encoded feature vector of the server can be determined based on the first time domain feature and the second time domain feature of the same server.

[0079] Assume that i Servers at time t The double-coded feature vector value of V i ( t ),but:

[0080] .

[0081] in, Indicates the i Servers at time t The enhanced data is obtained after the data is time-domain encoded. Indicates the i Servers at time t The feature representation is obtained after the data change is encoded in the frequency domain and converted to the time domain. It can be seen that the double-encoded feature vector is the concatenation of the first time domain feature and the second time domain feature at the same time point.

[0082] In an optional embodiment of the present disclosure, for the obtained dual-encoded feature vector, all feature vectors in the dual-encoded feature vector can be adjusted to the same time base through the currently commonly used interpolation method to obtain a dual-encoded feature vector with a unified time base.

[0083] The server performance prediction method of the embodiment of the present disclosure adopts an enhanced time domain coding function to capture subtle change characteristics in the original time series data, and introduces a frequency domain coding function to extract the long-term trend and periodic components of the data in the data change sequence, thereby constructing a dual-coded feature vector for each server. The dual-stream time coding enhances the comprehensiveness and accuracy of the feature representation by combining time domain and frequency domain information. The enhanced time domain coding captures subtle changes in the data, while the frequency domain coding reveals long-term trends and periodic components. The dual coding strategy can not only improve the model's ability to understand complex patterns, but also enable subsequent analysis to be based on richer feature information, thereby improving the robustness and reliability of server performance prediction.

[0084] In an optional embodiment of the present disclosure, Figure 3 As shown, based on the above embodiment, step 103 may include the following sub-steps:

[0085] Step 301: Determine the decision variable corresponding to the current moment based on the data variation sequence.

[0086] In this embodiment, for the data variation sequence of each server, the decision variable corresponding to each server at the current moment can be determined.

[0087] In an optional embodiment of the present disclosure, for each server, when calculating the decision variable, the data changes in the data change sequence can be traversed, and the currently traversed data change can be converted to obtain corresponding conversion parameters. For example, each data change in the data change sequence of each server can be input into the adaptive adjustment basis function shown in the following formula (5) to determine the conversion parameter:

[0088] (5)

[0089] in, x That is, it represents a data change in the data change sequence of the aforementioned embodiment. k and x 0 represents a preset parameter. It can be understood that for each data variation in the data variation sequence, the corresponding conversion parameter is obtained by converting it using the above formula (5), forming a conversion parameter sequence.

[0090] Next, the decision variables can be determined based on the conversion parameters. For example, the obtained conversion parameters can be integrated to obtain the decision variables. The calculation formula of the decision variables can be expressed as shown in the following formula (6):

[0091] (6)

[0092] in, D i ( t ) indicates the i Servers at time t The decision variables, Indicates the preset time interval, τ Represents a moment within a preset time interval. F () function is the above-mentioned adaptive adjustment basic function, Indicates the i Servers at time τ The amount of data change.

[0093] As an example, for each server, within a preset time interval (corresponding to a time window) closest to the current moment, the above formula (6) can be used to determine the decision variable corresponding to the current moment based on the data changes at all moments in the time window. The t in the above formula (6) represents the current moment.

[0094] As another example, for each server, the above formula (6) can be used to calculate the decision variables corresponding to each preset time interval within the entire preset monitoring period, and the decision variable at the current moment can be determined based on the calculated decision variables. For example, the mean of the decision variables can be calculated as the decision variable at the current moment.

[0095] Step 302: Determine the transformation scale parameter based on the decision variables.

[0096] In this embodiment, after the decision variables of each server at the current moment are determined, the transformation scale parameters may be determined based on the decision variables of each server.

[0097] As an example, the transformation scale parameter can be determined by the following formula (7):

[0098] (7)

[0099] in, Represents the decision variable of the server at the current moment, represents the transformation scale parameter, Indicates the preset initial value of the scale parameter, γ Indicates the preset adjustment coefficient and sets γ The value of is less than 0.

[0100] Through the above formula (7), we can achieve The value of dynamically adjusts the transformation scale parameter. When it is larger, a smaller scale parameter is used to capture more details; when When is smaller, a larger scale parameter is used to capture long-term trends.

[0101] Step 303: Perform continuous wavelet transform on the dual-coded feature vector based on the transform scale parameter to obtain wavelet coefficients at different scales.

[0102] In this embodiment, for each server, a continuous wavelet transform may be performed on the double-coded feature vector of the server using corresponding transformation scale parameters to obtain wavelet coefficients at different scales.

[0103] As an example, a trainable wavelet neural network can be used to perform multi-scale decomposition on the feature vector of the unified time base to obtain wavelet coefficients at different scales. Among them, the Morlet wavelet can be selected as the basis function. For the dual-encoded feature vector of each server, the continuous wavelet transform (CWT) is used to perform multi-scale decomposition on it. The expression of CWT is shown in formula (8):

[0104] (8)

[0105] in, It represents the transformation scale parameter that determines the scaling degree of the wavelet; b represents the translation parameter; Indicated on scale and location b The wavelet coefficients on ; represents the complex conjugate of the wavelet basis function; Indicates the i The double-encoded feature vectors of the servers.

[0106] Step 304: Generate time-frequency domain joint features based on wavelet coefficients at different scales.

[0107] In this embodiment, a trainable mechanism is introduced to optimize the parameters in the wavelet basis function and the transformation scale parameters. After multi-scale decomposition through the above-mentioned continuous wavelet transform, wavelet coefficients at different scales are obtained, and then the wavelet coefficients at different scales are reorganized into a new feature matrix to obtain frequency domain-time domain joint features, that is, time-frequency domain joint features.

[0108] The server performance prediction method of the embodiment of the present disclosure determines the decision variable corresponding to the current moment based on the data change sequence, and then determines the transformation scale parameter based on the decision variable, thereby realizing dynamic adjustment of the scale parameter of the wavelet transform according to the decision variable. By selecting Morlet wavelet as the basic function, a trainable wavelet neural network is used to perform continuous wavelet transform to realize multi-scale decomposition. Based on the parameters and scale parameters in the dynamically optimized wavelet basis function, the wavelet coefficients at different scales are reorganized to form a new feature matrix, and the time-frequency domain joint features are obtained. The feature information at different frequencies can be effectively extracted, and the expressiveness of the features can be further improved by optimizing the parameter settings. The application of Morlet wavelet can not only capture the detailed features of the signal, but also adapt to different frequency requirements by adjusting the scale parameters, which helps to explore deep data structures, provides high-quality input for subsequent cross-modal fusion, and significantly improves the accuracy and detail of performance prediction.

[0109] In an optional embodiment of the present disclosure, Figure 4 As shown, based on the above embodiment, step 104 may include the following sub-steps:

[0110] In step 401, the time-frequency domain joint features are input into the attention module of the gated cross attention network to obtain the attention feature matrix output by each attention branch of the attention module.

[0111] In this embodiment, for each server, the obtained time-frequency domain joint features are input into the attention module of the gated cross attention network. The concepts of query (Query, denoted as Q), key (Key, denoted as K) and value (Value, denoted as V) are introduced into the attention module to construct an attention mechanism. For the input time-frequency domain joint features, the attention mechanism of the attention module maps the input time-frequency domain joint features to three vectors of query Q, key K and value V. The attention score matrix (denoted as A) is calculated based on Q and K. For example, the dot product of Q and K can be used as the similarity measure between the two to obtain the attention score matrix, that is:

[0112] .

[0113] in, d k Indicates the dimension size of the key vector K, which is used to scale the dot product result to stabilize the gradient, K T Represents the transposed matrix of the key vector K.

[0114] Next, the obtained attention score matrix is ​​normalized by the Softmax function, and the attention score is converted into a probability distribution to obtain the attention weight matrix, where the attention weight matrix can be determined by the following formula (9):

[0115] (9)

[0116] in, Represents the first in the determined attention weight matrix i OK j The attention weights of the columns, Indicates the first i OK j The attention score of the column.

[0117] Next, the obtained attention weight matrix is ​​multiplied by the value vector V to achieve the fusion of cross-modal information and obtain the attention feature matrix.

[0118] Through the above method, each attention branch of the attention module can determine the attention feature matrix corresponding to each branch and output it.

[0119] In step 402, the attention feature matrix output by each attention branch is spliced ​​in feature dimension to obtain a feature splicing matrix.

[0120] In this embodiment, for the attention feature matrices output by each attention branch, these attention feature matrices can be spliced ​​along the feature dimension to obtain a feature splicing matrix spliced ​​along the feature dimension.

[0121] In step 403, the feature concatenation matrix is ​​input into the gate unit of the gated cross attention network to obtain the spatiotemporal feature matrix output by the gate unit.

[0122] In this embodiment, the gating unit of the gated cross-attention network includes a fully connected layer and a Sigmoid activation function. The feature splicing matrix is ​​input into the gating unit. The trainable weight parameters in the fully connected layer and the Sigmoid activation function constitute a dynamic weight matrix, which is then used to perform weighted summation on the input feature splicing matrix to obtain a spatiotemporal feature matrix.

[0123] Understandably, when server load fluctuates periodically, the gating unit's weighting parameters automatically increase the weight of the frequency domain branch, while decreasing it accordingly in bursty traffic scenarios, achieving dynamic feature optimization. This entire process is continuous, and as more data is input, the gating unit's weighting parameters become more optimized.

[0124] The server performance prediction method of the embodiment of the present disclosure calculates the attention feature matrix of the joint features in the time and frequency domains through the attention mechanism, and adjusts the information flow in combination with the gating mechanism to obtain an enhanced spatiotemporal feature matrix. The gated cross-attention mechanism effectively integrates information from different modalities, thereby enhancing the network's attention to key information and reducing the impact of irrelevant noise, enabling the network to flexibly select important information for processing, thereby improving not only the quality of feature representation, but also the network's ability to cope with complex environments, ultimately achieving high robustness and stability of the prediction results.

[0125] In an optional embodiment of the present disclosure, Figure 5 As shown, based on the above embodiment, step 105 may include the following sub-steps:

[0126] Step 501: Obtain an adjacency matrix corresponding to the at least one server.

[0127] The adjacency matrix describes the connection relationship between at least one server. The initialization adjacency matrix can be determined based on the actual topological structure of at least one server, and the initialization adjacency matrix is ​​updated. The initialization adjacency matrix is ​​added to the self-loop matrix (unit matrix) to obtain the adjacency matrix.

[0128] Step 502: Determine a degree matrix based on the adjacency matrix.

[0129] In this embodiment, based on the adjacency matrix of at least one server, a corresponding degree matrix can be determined. The degree matrix is ​​a diagonal matrix, and the elements on its diagonal are the degrees of the corresponding nodes (ie, servers), that is, the number of connections with other nodes.

[0130] The degree matrix is ​​calculated as follows:

[0131] .

[0132] in, Indicates the adjacency matrix i OK j Elements of the column, Indicates the degree matrix i OK i Elements of the column, j Represents the adjacency matrix Middle and Node i The indices of all other connected nodes are obtained by mapping the adjacency matrix No. i Sum all elements in the row to get the node i degree.

[0133] Step 503: Input the spatiotemporal feature matrix, degree matrix, and adjacency matrix into the graph convolutional network for processing to obtain the node-level state representation output by the graph convolutional network.

[0134] In this embodiment, the graph convolutional network can include at least one layer of graph convolution. When constructing a multi-layer graph convolutional network, the output of each layer serves as the input of the next layer. Graph convolution operations are used to propagate node information and learn node-level state representations (node ​​feature matrices), gradually refining more abstract and high-level node representations. Before each graph convolution, the adjacency matrix and corresponding degree matrix are re-evaluated and updated based on the current server topology. The graph convolution layer is defined as:

[0135] .

[0136] in, H (m+1) Indicates the first m +1) layer node feature matrix, H (m) Indicates the m The node feature matrix of the layer, the input layer The node feature matrix is ​​the spatiotemporal feature matrix of at least one server node; represents the adjacency matrix with self-loops added, express The degree matrix of W (m) Indicates the m The weight matrix of the layer, σ Represents an activation function (such as the ReLU function).

[0137] The output after the last layer of graph convolution That is the node-level state representation required, where L is the number of layers in the graph convolutional network.

[0138] The server performance prediction method of the embodiment of the present disclosure determines the adjacency matrix and the degree matrix, applies graph convolution operations to propagate node information and learn node-level state representation, refines high-level node representation, and obtains node-level state representation, so that the network can adapt to the dynamic changes in the relationship between servers, thereby more accurately reflecting the actual situation. As a result, it not only improves the accuracy of node representation, but also provides strong support for subsequent time series prediction, ensuring the reliability and accuracy of future server performance predictions.

[0139] In an optional embodiment of the present disclosure, Figure 6 As shown, based on the above embodiment, step 106 may include the following sub-steps:

[0140] Step 601: Perform multi-scale feature extraction on the node-level state representation to obtain multi-scale fusion features.

[0141] In this embodiment, the obtained node-level state representation is a feature matrix, and multi-scale feature extraction can be performed on the node-level state representation to obtain multi-scale fusion features.

[0142] As an example, a multi-scale temporal Transformer network can be used to extract multi-scale features from the node-level state representation to obtain multi-scale fusion features. The multi-scale temporal Transformer network includes a multi-scale feature extraction module and a Transformer architecture. The multi-scale feature extraction module is used to capture the pattern changes at different time scales in the input node-level state representation. The Transformer architecture is used to encode the features at each time scale extracted by the multi-scale feature extraction module after multi-scale feature extraction to obtain multi-scale fusion features. Among them, the multi-scale feature extraction module contains multiple parallel branches, each branch is responsible for processing the data of the node-level state representation at a specific time scale to obtain features at different time scales. The Transformer architecture first performs position encoding on the features output by each branch of the multi-scale feature extraction module, adding position information to each element in the time series; then, a multi-head self-attention mechanism is used to capture the internal dependencies of the time series. The input of the multi-head self-attention mechanism is the output of the position encoding. The output of the multi-head self-attention mechanism is passed through a feedforward neural network to further refine the feature representation. Finally, the feature representations refined for the features output by each branch are fused through weighted summation to form a unified feature representation, obtaining multi-scale fused features.

[0143] By designing a multi-scale feature extraction module to capture pattern changes at different time scales, and using the Transformer architecture to encode features at each time scale, and then performing feature fusion through weighted summation, the multi-scale temporal Transformer network captures pattern changes at different time scales by designing a multi-scale feature extraction module, enabling the network to fully understand the temporal characteristics of the data, and using the Transformer architecture to encode features at each time scale, further enhancing the network's ability to capture long-term and short-term dependencies.

[0144] Step 602: Based on the multi-scale fusion features, a regression model is used to predict performance indicator data of at least one server in the next monitoring period.

[0145] In this embodiment, after obtaining the multi-scale fusion features, the regression model can be used to predict the performance indicator data of at least one server in a certain time period in the future (eg, the next monitoring cycle).

[0146] It should be noted that currently commonly used regression models can be used for performance prediction, and the present disclosure does not limit the specific form of the regression model.

[0147] Step 603: Based on the node-level state representation, performance pulse data of at least one server is generated using a traffic pulse adversarial generation network.

[0148] In this embodiment, the obtained node-level state representation may be further used to generate performance pulse data of at least one server using a Traffic Pulse Generative Adversarial Network (TP-GAN).

[0149] The TP-GAN network consists of two parts: a generator and a discriminator. The generator's task is to generate data that simulates real server traffic pulses, while the discriminator is used to distinguish between real data and generated data. During the TP-GAN network optimization process, the generator attempts to deceive the discriminator into mistaking the generated data for real data, while the discriminator strives to improve its recognition ability to accurately distinguish between real and fake data. The network parameters of the generator and discriminator are updated with the goal of minimizing their respective losses.

[0150] In this embodiment, the node-level state representation is input into the generator of the TP-GAN network. The generator synthesizes performance pulse data of at least one server based on its own network parameters and the node-level state representation.

[0151] Step 604: Generate a performance prediction curve corresponding to at least one server based on the performance indicator data and the performance pulse data.

[0152] In this embodiment, for the same server, the performance indicator data and performance pulse data of the server can be combined to generate a performance prediction curve of the server. The performance indicator data is adjusted by the performance pulse data to reflect the sudden traffic situation, and then the performance prediction curve of the server is drawn based on the adjusted data.

[0153] The server performance prediction method of the embodiment of the present disclosure obtains multi-scale fusion features by performing multi-scale feature extraction on the node-level state representation, capturing the pattern changes at different time scales to facilitate a more comprehensive understanding of the temporal characteristics and long-term and short-term dependencies of the data. The performance prediction curve is generated by combining the traffic pulse adversarial generation network, which can not only simulate the actual burst traffic situation, but also adjust the prediction curve to better reflect the abnormal situations that may occur in the future, thereby improving the authenticity and practicality of the prediction curve, and ensuring that the prediction results have both high accuracy and can cope with complex practical application scenarios.

[0154] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus the necessary general hardware platform, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method.

[0155] An embodiment of the present application also provides a server performance prediction device. Figure 7 A schematic diagram of the structure of a server performance prediction device provided in an embodiment of the present application is shown in FIG. Figure 7 As shown, the server performance prediction device 70 includes: a data acquisition module 710, an encoding module 720, a decomposition module 730, a feature extraction module 740, a modeling module 750 and a prediction module 760.

[0156] The data acquisition module 710 is used to acquire the time series data of at least one server in the server cluster during a preset monitoring period and the data change sequence determined according to a preset time interval, where the end time of the preset monitoring period is the current time;

[0157] An encoding module 720 is configured to determine a dual-encoded feature vector corresponding to at least one server based on the time series data and the data variation sequence, where the dual-encoded feature vector includes a time domain feature obtained by time domain encoding and a time domain feature obtained by frequency domain encoding and conversion.

[0158] A decomposition module 730 is used to perform multi-scale decomposition on the double-coded feature vector to obtain a joint feature in the time and frequency domains;

[0159] The feature extraction module 740 is used to perform cross-modal information fusion on the joint time-frequency domain features to obtain a spatiotemporal feature matrix;

[0160] A modeling module 750 is configured to perform server topology modeling based on the spatiotemporal feature matrix to obtain a node-level state representation of at least one server;

[0161] The prediction module 760 is configured to perform time series prediction based on the node-level state representation and generate a performance prediction result corresponding to at least one server.

[0162] Optionally, the encoding module 720 is further configured to:

[0163] Perform time domain coding processing on the time series data to obtain the first time domain feature;

[0164] The data variation sequence is encoded in the frequency domain and then converted to the time domain to obtain the second time domain feature;

[0165] A dual-encoded feature vector is determined based on the first time-domain feature and the second time-domain feature.

[0166] Optionally, the decomposition module 730 includes:

[0167] A first determining unit is used to determine a decision variable corresponding to a current moment based on a data variation sequence;

[0168] a second determining unit, configured to determine a transformation scale parameter based on the decision variable;

[0169] A transform unit, configured to perform a continuous wavelet transform on the double-coded feature vector based on a transform scale parameter to obtain wavelet coefficients at different scales;

[0170] The generation unit is used to generate joint features in the time and frequency domains based on wavelet coefficients at different scales.

[0171] Optionally, the first determining unit is further configured to:

[0172] Traverse the data variation in the data variation sequence, convert the current traversed data variation, and obtain corresponding conversion parameters;

[0173] Determine the decision variables based on the transformation parameters.

[0174] Optionally, the feature extraction module 740 is further configured to:

[0175] Input the joint time-frequency domain features into the attention module of the gated cross-attention network to obtain the attention feature matrix output by each attention branch of the attention module;

[0176] The attention feature matrix output by each attention branch is spliced ​​in feature dimension to obtain a feature splicing matrix;

[0177] The feature splicing matrix is ​​input into the gate unit of the gated cross attention network to obtain the spatiotemporal feature matrix output by the gate unit.

[0178] Optionally, the modeling module 750 is further configured to:

[0179] Obtain an adjacency matrix corresponding to at least one server;

[0180] Determine the degree matrix based on the adjacency matrix;

[0181] The spatiotemporal feature matrix, degree matrix and adjacency matrix are input into the graph convolutional network for processing to obtain the node-level state representation of the graph convolutional network output.

[0182] Optionally, the prediction module 760 is further configured to:

[0183] Perform multi-scale feature extraction on node-level state representation to obtain multi-scale fusion features;

[0184] Based on the multi-scale fusion features, a regression model is used to predict the performance indicator data of at least one server in the next monitoring cycle;

[0185] Based on the node-level state representation, a traffic pulse adversarial generation network is used to generate performance pulse data of at least one server;

[0186] Based on the performance indicator data and the performance pulse data, a performance prediction curve corresponding to at least one server is generated.

[0187] For the description of the features in the embodiment corresponding to the server performance prediction device, please refer to the relevant description of the embodiment corresponding to the server performance prediction method, which will not be repeated here.

[0188] An embodiment of the present application further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any of the above-mentioned server performance prediction method embodiments.

[0189] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above-mentioned server performance prediction method embodiments when running.

[0190] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0191] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps of any of the above-mentioned server performance prediction method embodiments are implemented.

[0192] An embodiment of the present application also provides another computer program product, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in any of the above-mentioned server performance prediction method embodiments are implemented.

[0193] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0194] The above is a detailed introduction to a server performance prediction method, electronic device, storage medium, and program product provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core ideas. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A server performance prediction method, characterized in that: include: Obtaining time series data of at least one server in the server cluster during a preset monitoring period and a data variation sequence determined at preset time intervals, where the end time of the preset monitoring period is the current time; Determining, based on the time series data and the data variation sequence, a dual-coded feature vector corresponding to the at least one server, the dual-coded feature vector including a time domain feature obtained by time domain coding and a time domain feature obtained by frequency domain coding and conversion; Performing multi-scale decomposition on the double-coded feature vector to obtain a time-frequency domain joint feature; Performing cross-modal information fusion on the time-frequency domain joint features to obtain a spatiotemporal feature matrix; Performing server topology modeling based on the spatiotemporal feature matrix to obtain a node-level state representation of the at least one server; Performing time series prediction based on the node-level state representation to generate a performance prediction result corresponding to the at least one server; The performing of server topology modeling based on the spatiotemporal feature matrix to obtain a node-level state representation of the at least one server includes: Obtaining an adjacency matrix corresponding to the at least one server; determining a degree matrix based on the adjacency matrix; The spatiotemporal feature matrix, the degree matrix, and the adjacency matrix are input into a graph convolutional network for processing to obtain the node-level state representation output by the graph convolutional network.

2. The server performance prediction method according to claim 1, characterized in that: The determining, based on the time series data and the data variation sequence, a double-encoded feature vector corresponding to the at least one server includes: Performing time domain coding processing on the time series data to obtain a first time domain feature; Performing frequency domain encoding on the data variation sequence and then converting it into time domain to obtain a second time domain feature; The double-coded feature vector is determined based on the first time-domain feature and the second time-domain feature.

3. The server performance prediction method according to claim 1, wherein: The multi-scale decomposition of the double-coded feature vector to obtain a time-frequency domain joint feature includes: Determining the decision variable corresponding to the current moment based on the data variation sequence; determining a transformation scale parameter based on the decision variables; Performing a continuous wavelet transform on the dual-coded feature vector based on the transform scale parameter to obtain wavelet coefficients at different scales; The time-frequency domain joint features are generated based on the wavelet coefficients at different scales.

4. The server performance prediction method according to claim 3, wherein: The determining of the decision variable corresponding to the current moment based on the data variation sequence includes: Traversing the data variation in the data variation sequence, converting the currently traversed data variation to obtain corresponding conversion parameters; The decision variables are determined based on the transformation parameters.

5. The server performance prediction method according to claim 1, wherein: The cross-modal information fusion of the time-frequency domain joint features to obtain a spatiotemporal feature matrix includes: Inputting the time-frequency domain joint feature into the attention unit of the gated cross attention network to obtain the attention feature matrix output by each attention branch of the attention unit; Perform feature dimension splicing on the attention feature matrix output by each attention branch to obtain a feature splicing matrix; The feature splicing matrix is ​​input into the gating unit of the gated cross attention network to obtain the spatiotemporal feature matrix output by the gating unit.

6. The server performance prediction method according to claim 1, wherein: The performing time series prediction based on the node-level state representation to generate a performance prediction result corresponding to the at least one server includes: Performing multi-scale feature extraction on the node-level state representation to obtain multi-scale fusion features; Based on the multi-scale fusion features, using a regression model to predict performance indicator data of the at least one server in the next monitoring cycle; generating performance pulse data of the at least one server using a traffic pulse adversarial generation network based on the node-level state representation; Based on the performance indicator data and the performance pulse data, a performance prediction curve corresponding to the at least one server is generated.

7. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the server performance prediction method according to any one of claims 1 to 6 when executing the computer program.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, the steps of the server performance prediction method according to any one of claims 1 to 6 are implemented.

9. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the server performance prediction method according to any one of claims 1 to 6 are implemented.

Images