Cloud server performance degradation prediction method based on time series segmentation

Abstract

The application discloses a cloud server performance degradation prediction method based on time sequence segmentation, first extracts performance resource time sequence data on a cloud server, decomposes the obtained time sequence data by adopting a DTW-BU time sequence segmentation algorithm, respectively constructs LSTM models for segmented subsequences, and predicts cloud server resource time sequence data, verifies model precision by using root mean square error and average absolute percentage error, predicts system performance degradation trend according to time sequence prediction data of the LSTM model, tests data fitting degree, and determines a software regeneration time node according to prediction data threshold value; the application can improve the precision of cloud server performance degradation prediction results, avoid overfitting phenomenon in the prediction process, and solves the problem of how to perform software regeneration at an optimal time point for cloud server performance degradation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of time series prediction, and particularly relates to a cloud server performance degradation prediction method based on time series segmentation. BACKGROUND

[0002] With the development of cloud computing technology, cloud servers are increasingly used. Cloud servers have the characteristics of long-term operation, high complexity and frequent resource exchange, which increases the risk of resource depletion and abnormal and failure of software systems. With the accumulation of resource consumption, the cloud server system will have a slow performance decline, an increased failure rate and even a system crash. The main reasons for system performance degradation include consumption of operating system resources, data destruction and accumulation of errors. In important systems such as military defense, telecommunications systems, financial systems, securities systems and commercial systems, software performance degradation has occurred, affecting the normal operation of the entire business system and causing immeasurable economic losses to enterprises and institutions.

[0003] A common means to cope with system performance degradation is software rejuvenation technology. This technology actively restores the system before a fault occurs by cleaning up internal system errors. Software rejuvenation technology largely depends on the time of software rejuvenation. The downtime or overhead caused by such operations cannot be ignored, and frequent software rejuvenation can reduce system availability. Therefore, accurate prediction of software performance degradation trends, calculation of performance degradation thresholds and determination of rejuvenation time according to the thresholds can provide a theoretical basis for system preventive maintenance.

[0004] When predicting the system performance degradation trend, the change characteristics of the system performance index time series data need to be analyzed. It is usually necessary to combine the time series similarity measurement method to decompose the time series, reduce the dimension of the original time series, and then predict the time series trend in order to discover and predict the system performance degradation pattern in real time. Typical time series segmentation algorithms include sliding window algorithm, top-down (TD) algorithm and bottom-up (BU) algorithm. Time series segmentation methods are often used in combination with time series similarity measurement methods, such as slope deviation distance measurement and dynamic time warping algorithm (DTW). The sliding window method cannot capture the change information of the original time series with large fluctuations. The top-down TD algorithm has a relatively high time complexity and is relatively sensitive to noise data. In contrast, the bottom-up BU algorithm not only has a low time complexity, but also has linear scalability to data sets, and the segmentation effect is relatively good.

[0005] In addition, the existing cloud server system performance degradation trend prediction method is mostly time series analysis or intelligent algorithm. The time series analysis method adopts a recurrent neural network, particle filtering and other models to predict the trend, the model is simple, but the required data amount is large and the prediction accuracy is low for data with large fluctuations. The intelligent algorithm includes neural network, support vector machine and the like, the prediction accuracy of the algorithm is also not high when predicting time series data. The cloud server resource and performance data have the characteristics of nonlinearity, randomness and burstiness, the prediction accuracy of the above prediction methods for cloud server performance degradation is not high, and overfitting phenomenon is prone to occur. SUMMARY

[0006] The purpose of the present application is to provide a time series segmentation DTW-BU algorithm based on the combination of dynamic time warping (DTW) algorithm and bottom-up (BU) algorithm, and to predict the cloud server performance degradation trend by using a deep learning model Long Short-Term Memory (LSTM), and finally to determine the software regeneration time according to the predicted time series data; the present application can improve the accuracy of the cloud server performance degradation prediction result, and avoid overfitting phenomenon in the prediction process, thereby providing a theoretical basis for system regeneration maintenance.

[0007] The technical solution adopted by the present application is a cloud server performance degradation prediction method based on time series segmentation, which is implemented according to the following steps:

[0008] Step 1, extracting performance resource time series data on the cloud server;

[0009] Step 2, performing time series segmentation on the time series data obtained in step 1 by using the DTW-BU algorithm;

[0010] Step 3, constructing an LSTM model for each sub-sequence segmented in step 2, training the LSTM model, and predicting the cloud server resource time series data;

[0011] Step 4, merging the sub-sequences trained in step 3;

[0012] Step 5, verifying the model accuracy by using the root mean square error and the mean absolute percentage error;

[0013] Step 6, predicting the system performance degradation trend according to the time series prediction data of the LSTM model, testing the fitting degree of the data, and determining the time node of software regeneration according to the prediction data threshold.

[0014] The present application also has the following characteristics:

[0015] The performance resource time series data on the cloud server in step 1 includes CPU idle rate data and system available memory data.

[0016] where the segmentation process in step 2 is specifically as follows:

[0017] Step 2.1, cloud server performance degradation original time series X ( x 1, x 2,… x T ), length T , the time series X Data segments are connected two by two, and divided into T / 2 segments;

[0018] Step 2.2, calculate the similarity measure Seg i and Seg j of two segments DTW Value, the calculation formula is as follows:

[0019] (1)

[0020] Where ω t Indicates the weight value corresponding to the t Variable, and , , And Respectively represent the two segments Seg i And Seg j The difference in trend, mean and time span on the t Variable, ε , λ And γ Respectively represent the weight value of the difference of the three, and + + =1. Here DTW The smaller the value, the more similar the two time series segments are;

[0021] Step 2.3, select the smallest DTW Value from it in a loop, if the minimum value is less than the set segmentation threshold δ , then merge the corresponding two segments Seg i And Seg j

[0022] Step 2.4, repeat step 2.1 to step 2.3 until the DTW Value of all segments is not less than the segmentation threshold δ , the segmentation ends;​

[0023] Step 2.5, obtaining the sequence X The segmented feature similar d Subsequences Seg 1 , Seg 2, … , Seg d}.

[0024] Wherein step 3 is implemented according to the following steps:

[0025] Step 3.1, original data conversion, normalizing cloud resource time series data;

[0026] Step 3.2, generating model training data set, determining the window length of the training set;

[0027] Step 3.3, determining and adjusting the structure of the LSTM network model;

[0028] Step 3.4, model training and result prediction;

[0029] Wherein the verification process in step 5 is: fitting the prediction results in step 3 with the actual time series to verify the prediction accuracy:

[0030] The root mean square error (RMSE) and the mean absolute percentage error (MAPE) of the actual monitoring data and the model prediction data of the cloud server resource are compared, and the calculation formula is as follows:

[0031] (2)

[0032] (3)

[0033] Wherein, actual ( t ) represents the actual time series data; forecast ( t ) represents the time series data predicted by the LSTM model; RMSE E represents the expected value of the square of the error; MAPE The smaller the value of E, the better the accuracy of the prediction model;

[0034] The performance resource time series data of the cloud server is randomly divided into training set and validation set according to the proportion of 4:1; then the time series data is imported into the model as parameters to obtain the prediction data, and compared with the actual cloud resource time series to obtain the fitting curve;

[0035] The cloud server resource time sequence prediction data threshold in step 6 is determined according to an LSTM model, so as to determine a time node for software regeneration, and the cloud server is in a state close to downtime, and the software regeneration operation is selected at this time to keep the system in a good running state.

[0036] The present application has the following advantages:

[0037] The present application adopts a prediction method combining a DTW-BU time sequence segmentation algorithm and an LSTM model to predict the degradation trend of the cloud server, avoids overfitting in the model prediction process, overcomes the limitation of low prediction accuracy of traditional prediction methods for non-stationary time sequence data and time sequence data with large fluctuations, and thus can more accurately predict the performance degradation trend of the cloud server, and provide an accurate time node for the software regeneration operation. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is the overall flowchart of the cloud server performance degradation prediction method based on time sequence segmentation of the present application;

[0039] Figure 2 is the cloud server CPU idle resource time sequence diagram in the cloud server performance degradation prediction method based on time sequence segmentation of the present application;

[0040] Figure 3 is the cloud server system available memory resource time sequence diagram in the cloud server performance degradation prediction method based on time sequence segmentation of the present application;

[0041] Figure 4 is the DTW-BU time sequence segmentation algorithm flowchart in the cloud server performance degradation prediction method based on time sequence segmentation of the present application;

[0042] Figure 5 is the LSTM model structure diagram used in the cloud server performance degradation prediction method based on time sequence segmentation of the present application. DETAILED DESCRIPTION

[0043] The present application will be described in detail below in combination with the drawings and specific embodiments.

[0044] The present application provides a cloud server performance degradation prediction method based on time sequence segmentation, as shown in Figure 1 The specific implementation is as follows:

[0045] Step 1: Extract the performance resource time sequence data on the cloud server, including CPU idle rate data and system available memory data, and the time sequence data is specifically as shown in Figure 2 and Figure 3 .

[0046] The segmentation process in step 2 specifically involves: firstly, preprocessing the time series data; then dividing the data into segments of length [missing information]. T The time series data points are connected pairwise and divided into non-overlapping segments. T Two initial segments are formed, and the similarity DTW value of adjacent segments is calculated and merged. Then, the minimum DTW value is selected from these segments in a loop. If the minimum value is less than a set segmentation threshold, the segmentation is terminated. δ If the two corresponding segments are merged, the DTW value of the new segment and its preceding and following segments is calculated. This process is repeated until the DTW value between all segments is not less than the segmentation threshold, at which point the segmentation ends. The given sequence is decomposed into feature-similar subsequences from the bottom up, with each segment considered as a cloud server health state with different degrees of decay. Figure 4 As shown:

[0047] Step 2.1, Original time series of cloud server performance degradation X =( x 1, x 2,… x T ), length is T time series X Data segments are concatenated in pairs and divided into T / 2 segments;

[0048] Step 2.2, calculate pairwise segments Seg i and Seg j Similarity measurement DTW The value is calculated using the following formula:

[0049] (1)

[0050] in ω t Indicates the first t The weight values ​​corresponding to each variable, and , , and These represent two segments respectively. Seg i and Seg j In the t Differences in trend, mean, and time span across individual variables. ε , λ and γ These represent the weight values ​​for the differences among the three, and also have... + + =1; here DTW The smaller the value, the more similar the segments of the two time series are;

[0051] Step 2.3, the minimum value is selected from the cycle DTW If the minimum value is less than the set segmentation threshold δ , then the corresponding two segments are merged Seg i And Seg j ;

[0052] Step 2.4, repeat steps 2.1 to 2.3 until all segments DTW The value is not less than the segmentation threshold δ , the segmentation is finished;

[0053] Step 2.5, get the sequence X Segmented feature similar d Subsequence{ Seg 1 , Seg 2, … , Seg d};

[0054] Wherein step 3 is implemented according to the following steps:

[0055] Step 3.1, original data conversion, normalizing the cloud resource time series data;

[0056] Step 3.2, generate model training data set, determine the window length of the training set;

[0057] Step 3.3, determine and adjust the structure of the LSTM network model;

[0058] Step 3.4, model training and result prediction;

[0059] Wherein the verification process in step 5 is: fitting the prediction results in step 3 with the actual time series to verify the prediction accuracy:

[0060] The root mean square error (RMSE) and the mean absolute percentage error (MAPE) of the actual monitoring data of the cloud server resources and the model prediction data are compared, and the calculation formula is as follows:

[0061] (2)

[0062] (3)

[0063] In the formula, actual ( t ) represents the actual time series data; forecast ( t ) represents the time series data predicted by the LSTM model; RMSE E represents the expected value of the square error.MAPE The smaller the value is, the better the prediction model has the accuracy;

[0064] The performance resource time series data of the cloud server is randomly divided into a training set and a validation set at a ratio of 4:1; then the time series data is imported into the model as a parameter to obtain prediction data, and a fitting curve is obtained by comparing the actual cloud resource time series;

[0065] Among them, the cloud server resource time series prediction data threshold in step 6 according to the LSTM model is as shown in the following formula: Figure 5 Thus, the time node for software regeneration is determined, and the cloud server is in a state close to downtime. At this time, the software regeneration operation is selected to keep the system in a good running state.

Claims

1. A cloud server performance degradation prediction method based on time series segmentation, characterized in that, The specific steps are as follows: Step 1: Extract time-series data of performance resources from the cloud server; Step 2: The time series data obtained in Step 1 is divided into new subsequences using the DTW-BU algorithm; Step 3: Construct LSTM models for the multiple subsequences segmented in Step 2, train the LSTM models, and predict the time series data of cloud server resources. Step 4: Merge the subsequences predicted and trained in Step 3; Step 5: Verify the model accuracy using root mean square error and mean absolute percentage error; Step 6: Based on the time series prediction data of the LSTM model, predict the system performance degradation trend, check the goodness of fit of the data, and determine the time node for software regeneration based on the predicted data threshold. The segmentation process in step 2 is specifically as follows: Step 2.1, Original time series of cloud server performance degradation X =( x 1, x 2,… x T ), length is T time series X Data segments are concatenated in pairs and divided into T / 2 segments; Step 2.2, calculate pairwise segments Seg i and Seg j Similarity measurement DTW The value is calculated using the following formula: （1） in ω t Indicates the first t The weight values ​​corresponding to each variable, and , , and These represent two segments respectively. Seg i and Seg j In the t Differences in trend, mean, and time span across individual variables. ε , λ and γ These represent the weight values ​​for the differences among the three, and also have... + + =1; here DTW The smaller the value, the more similar the segments of the two time series are; Step 2.3: Iteratively select the smallest from among them. DTW The value, if the minimum value is less than the set segmentation threshold. δ Then merge the two corresponding segments. Seg i and Seg j ; Step 2.4, repeat steps 2.1 to 2.3 until all segments are complete. DTW The values ​​are all not less than the segmentation threshold. δ The segmentation is complete; Step 2.5, obtain the sequence X Similar features after segmentation d Subsequences { Seg 1 Seg 2, … , Seg d } 2. The cloud server performance degradation prediction method based on time series segmentation according to claim 1, characterized in that, The time-series data of performance resources on the cloud server in step 1 includes CPU idle rate data and system available memory data.

3. The cloud server performance degradation prediction method based on time series segmentation according to claim 1, characterized in that, Step 3 is implemented in the following steps: Step 3.1, raw data transformation, normalize the cloud resource time series data; Step 3.2: Generate the model training dataset and determine the window length of the training set; Step 3.3: Determine and adjust the network model structure; Step 3.4: Model training and result prediction.

4. The cloud server performance degradation prediction method based on time series segmentation according to claim 1, characterized in that, The verification process in step 5 specifically involves: fitting the prediction results from step 3 with the actual time series to verify the prediction accuracy. The root mean square error (RMSE) and mean absolute percentage error (MAPE) of the actual monitoring data and model prediction data of cloud server resources are compared. The calculation formula is as follows: （2） （3） In the formula, actual ( t () represents the actual time series data; forecast ( t () represents the time series data predicted by the LSTM model; RMSE This represents the expected value of the square of the error; MAPE The smaller the value, the better the accuracy of the prediction model; The time series data of cloud server performance resources are randomly divided into training and validation sets in a 4:1 ratio. Then, the time series data is imported into the model as parameters to obtain prediction data, and the fitted curve is obtained by comparing it with the actual cloud resource time series.

5. The cloud server performance degradation prediction method based on time series segmentation according to claim 1, characterized in that, In step 6, the time node for software regeneration is determined based on the cloud server resource time series prediction data threshold of the LSTM model. When the cloud server is in a state close to crashing, the software regeneration operation is selected to maintain the system in a good operating state.

Images