A method for predicting cloud server software aging based on EGGM model
Through the EGGM model, the cloud server performance indicator data is subjected to EEMD decomposition and GA-GRU-MLR model prediction, which solves the problem of low accuracy caused by data randomness and non-periodicity in the aging prediction of cloud server software, and achieves high-precision prediction effect .
Patent Information
- Application Number
- CN202110959442.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-08-20
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2041-08-20
AI Technical Summary
The existing cloud server software aging prediction method has too low prediction accuracy due to the randomness and non-periodic characteristics of the data. Especially in the interval where data fluctuations are severe, it is difficult for traditional machine learning and mathematical statistical models to accurately predict.
Using the prediction method based on the EGGM model, the cloud server performance index data is decomposed into high and low frequency IMF components and Residual components through EEMD, combined with GA optimization GRU model and MLR model, and multi-step iterative prediction of each component is carried out, and the final prediction result is finally reconstructed. .
It improves the accuracy of aging prediction of cloud server software, can better fit the changing trends of performance indicator data, overcomes the problems of local optimal solutions and low prediction accuracy of traditional methods, and achieves high-precision prediction and analysis.
Smart Images

Figure CN113791960B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of time series prediction and relates to a method for predicting cloud server software aging based on an EGGM model. Background Art
[0002] With the rapid development of the Internet, many small and medium-sized Internet service providers, such as online game companies, Internet business companies, and financial securities companies, usually adopt cloud server solutions due to insufficient technical strength or to save money and facilitate server operation and maintenance, which has led to the booming development of cloud computing technology and cloud server business. The main features of cloud servers are long-term operation, high scalability, and frequent resource exchange. With the rapid increase in the complexity and concurrency of cloud server software, the service pressure on cloud servers is increasing. During the long-term high-load operation of cloud servers, as faults and resource consumption accumulate, the performance of cloud servers will slowly decline, the failure rate will also increase, and severe cases may even cause cloud servers to crash, resulting in service interruptions. This phenomenon is called "software aging."
[0003] In the process of predicting software aging of cloud servers, the most important thing is to calculate the software aging threshold of cloud servers, which provides an important theoretical basis for online maintenance. Although the performance index data of cloud servers usually has certain periodic characteristics, when the number of users accessing Internet products running on cloud servers increases during holidays, the performance index data of cloud servers will have certain high load, randomness and non-periodic components, which makes short-term prediction difficult.
[0004] At present, the prediction technology of cloud server aging can be divided into three categories. The first category is the prediction method based on traditional mathematical statistical models, such as multivariate linear regression, time series method, etc. The multivariate linear regression model is simple, and has high prediction accuracy and fast prediction speed for low-frequency data, but for complex nonlinear data or high-frequency data, multivariate linear regression is difficult to obtain high prediction accuracy. The time series method can make accurate predictions for data with high stability and strong periodicity, but when the data has non-periodic and non-stationary features, the time series method is difficult to obtain effective prediction results. The second category is the prediction method based on machine learning, such as recurrent neural network (RNN), gated recurrent unit (GRU), support vector machine (SVM), etc. Among the machine learning methods, neural networks are widely used in nonlinear system prediction with their powerful multivariate mapping capabilities and have achieved high prediction accuracy. The RNN neural network realizes the feature mining of time series data by inputting the neuron state value of the previous moment into the current neuron. GRU neural network is based on RNN neural network. By adding gate structure to control the influence of previous moments, it solves the defects of gradient explosion and vanishing of RNN, making GRU neural network better than RNN in processing and mining time series. Support vector machine is a small sample statistical learning theory based on the principle of structural risk minimization and the concept of dimension. It avoids the problems of network structure selection, over-learning, under-learning and local minima of artificial neural network and other methods, but the prediction results of fluctuation peak are not very ideal. Although machine learning algorithms have many advantages over traditional algorithms, their prediction results depend to a large extent on the quality and quantity of data. High-quality and massive data often make it easier to obtain accurate prediction results. For time series data with high dimensions and small data volume, due to the influence of noise and local features, it is difficult for machine learning algorithms to find effective rules from the data to establish an accurate prediction model. In addition, machine learning algorithms generally have the disadvantages of difficult hyperparameter determination, high resource consumption and slow computing speed. Due to the characteristics of nonlinearity, random distribution and periodicity of the resource and performance data of cloud server software aging, the prediction accuracy of the above two methods is not high, especially in the interval where the software aging data fluctuates violently, the prediction accuracy of these models will be lower. The third type is the prediction method using a combined model. First, the time series data is feature decomposed, usually using wavelet transform (WT), empirical mode decomposition (EMD), ensemble empirical mode decomposition, etc., and then the decomposed data is predicted using traditional mathematical statistical model methods or machine learning methods, and then reconstructed to obtain the final prediction value.When wavelet transform processes non-stationary series, different basis functions and orders have different effects, which makes the method have a certain a priori nature and increases the complexity of the prediction model in actual use. EMD is very suitable for processing nonlinear and non-stationary time series data. It is essentially a stationary processing of time series data, decomposing complex time series data into a linear combination of a finite number of intrinsic mode functions (IMF) with high to low frequencies. The decomposed IMF components contain the local characteristic signals of the original data at different time scales. Therefore, by performing EMD decomposition on time series data, it is possible to achieve a separate prediction of local features, achieve a non-interference effect, and reduce the difficulty of prediction. However, the data decomposed by EMD has some modal aliasing, which will have a certain impact on the prediction effect. In order to better solve the problem of modal aliasing, an improved ensemble empirical mode decomposition method (EEMD) has emerged. The EEMD method performs auxiliary processing on time series data by adding white noise, and utilizes the uniform distribution characteristics of the white noise spectrum. When the data is added to the white noise background with consistent distribution throughout the entire time-frequency space, the data of different time scales will be automatically distributed to the appropriate reference scale, and due to the characteristics of zero-mean noise, after multiple averaging, the noise will cancel each other out, and the result of the integrated mean can be used as the final result. The EEMD method achieves the purpose of significantly improving the decomposition effect of the EMD method by significantly reducing the modal aliasing phenomenon. The present invention decomposes the cloud server performance index data into several IMF components and residual components with different frequencies by the EEMD method. Each IMF component represents a local feature on a certain time scale. While fully mining the information contained in the time series data, it avoids mutual interference between local features. Subsequently, the GA-GRU model and the multiple linear regression (MLR) model after optimizing the GRU hyperparameters by the genetic algorithm (GA) are used to perform multi-step iterative predictions on the high and low frequency IMF components and residual components, respectively, and the obtained results are reconstructed and restored to obtain the final prediction value. The present invention builds a cloud server platform and collects cloud server performance indicator data. When the data contains noise, it achieves better prediction effect and can better fit the changing trend of cloud server performance indicator data. Summary of the invention
[0005] The purpose of the present invention is to provide a method for predicting cloud server software aging based on the EGGM model, which solves the problem that traditional machine learning prediction methods (such as SVM, GRU, etc.), prediction methods based on traditional mathematical statistical models (such as MLR, time series method, etc.) and traditional time series decomposition methods (such as EMD decomposition, wavelet decomposition, etc.) have low prediction accuracy when predicting cloud server software aging due to the randomness and non-periodic components of cloud server performance indicator data.
[0006] The technical solution adopted by the present invention is a method for predicting cloud server software aging based on the EGGM model, which specifically includes the following steps:
[0007] Step 1, extracting cloud server performance indicator series data;
[0008] Step 2: Perform EEMD decomposition on the cloud server performance indicator data extracted in step 1 to obtain multiple IMF components with frequencies from high to low and one Residual component;
[0009] Step 3, normalize the high-frequency IMF component, low-frequency IMF component and residual component obtained in step 2, and map the data sequence to the [0,1] interval;
[0010] Step 4, constructing a GRU model using the high-frequency IMF components normalized in step 3, and obtaining the prediction results of the GRU model for the data obtained in step 3;
[0011] Step 5, the GRU model obtained in step 4 is optimized through GA to obtain a GA-GRU model;
[0012] Step 6, predicting the high-frequency IMF component through the GA-GRU model obtained in step 5 to obtain the prediction result of the GA-GRU model;
[0013] Step 7, predicting the low-frequency IMF component and residual component obtained in step 4 by using the MLR model to obtain the MLR model prediction result;
[0014] Step 8, reconstructing the GA-GRU model prediction result obtained in step 6 and the MLR model prediction result obtained in step 7 to obtain the final prediction result;
[0015] Step 9: Compare the final prediction result obtained in step 8 with the actual data.
[0016] The present invention is also characterized in that:
[0017] Step 1 specifically includes: extracting cloud server performance indicator series data by writing an automated script on the cloud server, where the extracted performance indicator series data includes CPU idle rate, available memory, average load, and response time;
[0018] Wherein step 2 is specifically implemented as follows:
[0019] Step 2.1, set the number of decompositions m;
[0020] Step 2.2, add Gaussian white noise to the original sequence x(t), the standard deviation of the added white noise is 0.2 times the standard deviation of the original sequence, and the mean is 0, and the sequence after adding white noise is obtained; as shown in the following formula (1):
[0021] x′(t)=x(t)+εn(t) (1)
[0022] In the formula, ε n (t)(n=1,2,...,m) is random Gaussian white noise, m is the length of the original sequence;
[0023] Step 2.3, get all the i is the IMF order;
[0024] Step 2.4, repeat steps 2.2 and 2.3 m times to obtain all IMFs;
[0025] Step 2.5, calculate the average value of m IMFs obtained by equation (2):
[0026]
[0027] Step 3 is specifically implemented as follows:
[0028] Step 3.1, record the sequence data vector as x, and then record the maximum and minimum values as X max and X min ;
[0029] Step 3.2: Subtract the minimum value X from each data in the sequence data. min ;
[0030] Step 3.3: Divide the sequence data obtained in step 3.2 by the difference between the maximum value and the minimum value (X max -X min ); the normalization formula is shown in formula (3):
[0031]
[0032] Step 4 is specifically implemented as follows:
[0033] Using the normalized high-frequency IMF components obtained in step 3, divide the data set into training set and test set for each component, and then construct the three functional modules of the GRU model, including: input layer, hidden layer and output layer. The input layer accepts the initial parameters of the neural network; the hidden layers are connected to each other through GRU neurons to build a multi-layer network and iteratively calculate; the output layer obtains the model prediction results and then merges the outputs;
[0034] Step 5 is specifically implemented as follows:
[0035] The specific method of using GA to optimize the hyperparameters of the GRU model is: constructing a hyperparameter search space for the GRU prediction model, using GA to globally search the hyperparameter space, screening the optimal hyperparameters, obtaining the optimal hyperparameter combination, and establishing a performance-optimized GRU prediction model;
[0036] The specific steps of using GA to optimize the GRU model in step 5 are as follows:
[0037] Step 6.1, population coding initialization; the coding method used is to convert the parameters to be optimized into gene strings and form chromosomes; for the hyperparameters of the GRU neural network, they are encoded in binary with different bits;
[0038] Step 6.2, calculate the individual fitness value; the fitness value quantifies the adaptability of the individual in the evolution process. The higher the fitness value, the smaller the model prediction error formed by the hyperparameter combination represented by the individual. The fitness function is as follows (4):
[0039]
[0040] Where: y(k) is the actual observed value of the sample time series, is the predicted value, T represents the amount of data in the sample parameter time series;
[0041] Step 6.3, use genetic algorithms to perform selection, crossover and mutation operations among individuals to achieve changes in individual information. When the information of an individual changes, a new individual is generated in the hyperparameter search space. By continuously generating new individuals, the optimal hyperparameter combination is explored. When the new individual meets the training error or reaches the number of iterations, the search ends and the optimal GRU prediction model is output.
[0042] The beneficial effects of the present invention are:
[0043] The present invention uses ensemble empirical mode decomposition EEMD, genetic algorithm GA, gated recurrent unit GRU neural network and multivariate linear regression MLR. First, cloud server performance index data is decomposed by EEMD to obtain several groups of high- and low-frequency IMF components and residual components. Then, the high-frequency IMF components are predicted by a GA-GRU model after GA optimizes GRU hyperparameters. The low-frequency IMF components and residual components are predicted by an MLR model. Finally, the predicted values of each component are reconstructed to obtain the final predicted value. The EEMD-GA-GRU-MLR prediction method overcomes the problem that traditional methods are prone to fall into local optimal solutions and have too low prediction accuracy during the prediction process. Compared with traditional methods, the present method can not only accurately extract the data fluctuation change characteristics of cloud server software aging, but also accurately predict local features with strong randomness, and finally realize high-precision prediction and analysis of performance index parameters of cloud server software aging. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 This is the data flow of EEMD decomposition of cloud server database query response time series data in the method for predicting cloud server software aging based on the EGGM model of the present invention;
[0045] Figure 2 It is the EEMD decomposition data result of the cloud server database query response time series data in the method for predicting cloud server software aging based on the EGGM model of the present invention;
[0046] Figure 3 It is the GA chromosome encoding result in the method for predicting cloud server software aging based on the EGGM model of the present invention;
[0047] Figure 4 The specific process of the EEMD-GA-GRU-MLR prediction method in the method for predicting cloud server software aging based on the EGGM model of the present invention;
[0048] Figure 5 The GRU model structure in the method for predicting cloud server software aging based on the EGGM model of the present invention;
[0049] Figure 6 A cloud server database query response time series data graph in the method for predicting cloud server software aging based on the EGGM model of the present invention;
[0050] Figure 7 This is a comparison chart of the prediction accuracy of the EEMD-GAGRU-MLR model, the GAGRU model, the SVM model, and the MLR model in the method for predicting cloud server software aging based on the EGGM model of the present invention;
[0051] Figure 8 This is an absolute error comparison chart of the EEMD-GAGRU-MLR model, the GAGRU model, the SVM model, and the MLR model in the method for predicting cloud server software aging based on the EGGM model of the present invention. DETAILED DESCRIPTION
[0052] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0053] The present invention provides a method for predicting cloud server software aging based on the EGGM model, which specifically includes the following steps:
[0054] Step 1: extract the cloud server performance indicator series data by writing an automated script on the cloud server. The extracted performance indicator series data include CPU idle rate, available memory, average load and response time.
[0055] Step 2, such as Figure 1 and Figure 2 As shown, the cloud server performance indicator data extracted in step 1 is decomposed by EEMD to obtain multiple IMF components with high to low frequencies and a Residual component:
[0056] Step 2.1, set the number of decompositions m;
[0057] Step 2.2, add Gaussian white noise to the original sequence x(t), the standard deviation of the added white noise is 0.2 times the standard deviation of the original sequence, and the mean is 0, and the sequence after adding white noise is obtained; as shown in the following formula (1):
[0058] x′(t)=x(t)+εn(t) (1)
[0059] In the formula, ε n (t)(n=1,2,...,m) is random Gaussian white noise, m is the length of the original sequence;
[0060] Step 2.3, get all the i is the IMF order;
[0061] Step 2.4, repeat steps 2.2 and 2.3 m times to obtain all IMFs;
[0062] Step 2.5, calculate the average value of m IMFs obtained by equation (2):
[0063]
[0064] Step 3: Normalize and preprocess the high-frequency IMF component, low-frequency IMF component, and Residual component decomposed in step 2, and map the data sequence to the [0,1] interval:
[0065] Step 3.1, record the sequence data vector as x, and then record the maximum and minimum values as X max and X min ;
[0066] Step 3.2: Subtract the minimum value X from each data in the sequence data. min ;
[0067] Step 3.3: Divide the sequence data obtained in step 3.2 by the difference between the maximum value and the minimum value (X max -X min ); the normalization formula is shown in formula (3):
[0068]
[0069] Step 4: Using the normalized high-frequency IMF components obtained in step 3, divide the data set into a training set and a test set for each component, and then construct three functional modules of the GRU model, including: input layer, hidden layer and output layer. The input layer accepts the initial parameters of the neural network; the hidden layers are connected to each other through GRU neurons to construct a multi-layer network and iteratively calculate; the output layer obtains the model prediction results and then merges the outputs;
[0070] Step 5, the specific method of using GA to optimize the hyperparameters of the GRU model is: constructing a GRU prediction model hyperparameter search space, using GA to globally search the hyperparameter space, screening the optimal hyperparameters, obtaining the optimal hyperparameter combination, and establishing a performance-optimized GRU prediction model; the hyperparameter search range space set in this application is shown in Table 1; in Table 1, the hyperparameter value range is represented by [a:b:c], where a is the starting point, b is the end point, and c is the numerical interval;
[0071] Table 1. Hyperparameter search space of GRU prediction model
[0072] Tab.1 Hyperparameter search space of GRU prediction model
[0073]
[0074] The specific steps of using GA to optimize the GRU model are as follows:
[0075] Step 5.1, population coding initialization; the coding method used is to convert the parameters to be optimized into the form of gene strings and form chromosomes; for the hyperparameters of the GRU neural network, they are encoded in binary with different bits; the coding process is as follows Figure 3 As shown;
[0076] Step 5.2, calculate the individual fitness value; the fitness value quantifies the adaptability of the individual in the evolution process. The higher the fitness value, the smaller the model prediction error formed by the hyperparameter combination represented by the individual. The fitness function is as follows (4):
[0077]
[0078] Where: y(k) is the actual observed value of the sample time series, is the predicted value, T represents the amount of data in the sample parameter time series;
[0079] Step 5.3, use genetic algorithm to perform selection, crossover and mutation operations between individuals to achieve the change of individual information. When the information of a certain individual changes, a new individual is generated in the hyperparameter search space. By continuously generating new individuals, the optimal hyperparameter combination is explored. When the new individual meets the training error or reaches the number of iterations, the search ends and the optimal GRU prediction model is output;
[0080] Step 6, predicting the high-frequency IMF component through the GA-GRU model obtained in step 5; the GA-GRU model is a GRU model after optimizing the network hyperparameters, and the high-frequency IMF component data set is put into the network training to obtain the optimal result.
[0081] Step 7, predict the low-frequency IMF component and Residual component obtained in step 4 through the MLR model; MLR sets the time window to 3 through grid search parameter adjustment, and then predicts the linear time series data such as the low-frequency IMF component and the Residual component to obtain the prediction result;
[0082] Step 8, reconstruct the GA-GRU model prediction results obtained in step 6 and the MLR model prediction results obtained in step 7;
[0083] Step 9, comparing the final prediction result obtained in step 8 with the real data, it can be seen that this method basically fits the changing trend of the cloud server performance indicator data.
[0084] Example
[0085] The present application embodiment uses the cloud server database query response time series data as an example. The EEMD decomposition data flow is as follows: Figure 1 As shown, the EEMD decomposition data results are as follows Figure 2 As shown, the GA chromosome encodes Figure 3 As shown, the specific implementation process of this method is as follows Figure 4 As shown, the structure of a single GRU hidden layer is as follows Figure 5 As shown, the cloud server database query response time series is as follows Figure 6As shown in Figure 2, the prediction accuracy of the EEMD-GAGRU-MLR model is compared with that of the GAGRU model, SVM model, and MLR model. Figure 7 As shown, the absolute error comparison of EEMD-GAGRU-MLR model, GAGRU model, SVM model, and MLR model is Figure 8 The error comparison of different models is shown in Table 2. The root mean square error RMSE, mean absolute error MAE and mean absolute percentage error MAPE are used as evaluation indicators, respectively, as shown in the following formulas (5), (6) and (7), where RMSE is the standard deviation, N is the number of data samples, and y predictive is the predicted value, y true is the actual value.
[0086]
[0087]
[0088]
[0089] Table 2 Comparison of prediction errors of different models
[0090]
[0091] As can be seen from the above table, the RMSE, MAE and MAPE of the present invention are lower than those of other methods. Therefore, the prediction method of decomposing the cloud server performance index data by EEMD, then using GA-GRU to predict the high-frequency IMF component, using MLR to predict the low-frequency IMF component and the residual component, and finally superimposing and reconstructing the prediction results of each IMF component to obtain a complete prediction result is feasible and highly accurate. This method solves the problem of low prediction accuracy in traditional methods due to the randomness and non-periodic components of cloud server performance index data, and can accurately fit cloud server performance index data.
Claims
1. A method for predicting cloud server software aging based on the EGGM model, characterized in that: The specific steps include: Step 1, extracting cloud server performance indicator series data; Step 2: Perform EEMD decomposition on the cloud server performance indicator data extracted in step 1 to obtain multiple IMF components with frequencies from high to low and one Residual component; Step 3, normalize the high-frequency IMF component, low-frequency IMF component and residual component obtained in step 2, and map the data sequence to the [0,1] interval; Step 4, constructing a GRU model using the high-frequency IMF components normalized in step 3, and obtaining the prediction results of the GRU model for the data obtained in step 3; Step 5, the GRU model obtained in step 4 is optimized through GA to obtain a GA-GRU model; The specific method of using GA to optimize the hyperparameters of the GRU model is: constructing a hyperparameter search space for the GRU prediction model, using GA to globally search the hyperparameter space, screening the optimal hyperparameters, obtaining the optimal hyperparameter combination, and establishing a performance-optimized GRU prediction model; Step 5.1, population coding initialization; the coding method used is to convert the parameters to be optimized into gene strings and form chromosomes; for the hyperparameters of the GRU neural network, they are encoded in binary with different bits; Step 5.2, calculate individual fitness value; The fitness value quantifies the adaptability of an individual during the evolution process. The higher the fitness value, the smaller the model prediction error formed by the hyperparameter combination represented by the individual. The fitness function is as follows (4): (4) Where: is the actual observed value of the sample time series, is the predicted value, T represents the amount of data in the sample parameter time series; Step 5.3, use genetic algorithm to perform selection, crossover and mutation operations between individuals to achieve the change of individual information. When the information of a certain individual changes, a new individual is generated in the hyperparameter search space. By continuously generating new individuals, the optimal hyperparameter combination is explored. When the new individual meets the training error or reaches the number of iterations, the search ends and the optimal GRU prediction model is output; Step 6, predicting the high-frequency IMF component through the GA-GRU model obtained in step 5 to obtain the prediction result of the GA-GRU model; Step 7, predicting the low-frequency IMF component and the residual component obtained in step 3 by using the MLR model to obtain the MLR model prediction result; Step 8, reconstructing the GA-GRU model prediction result obtained in step 6 and the MLR model prediction result obtained in step 7 to obtain the final prediction result; Step 9: Compare the final prediction result obtained in step 8 with the actual data.
2. The method for predicting cloud server software aging based on the EGGM model according to claim 1, characterized in that: The step 1 specifically includes: extracting cloud server performance indicator sequence data by writing an automated script on the cloud server, and the extracted performance indicator sequence data includes CPU idle rate, available memory, average load and response time.
3. The method for predicting cloud server software aging based on the EGGM model according to claim 1, characterized in that: The step 2 is specifically implemented according to the following steps: Step 2.1, set the number of decompositions m ; Step 2.2: Add Gaussian white noise to the original sequence x(t) In the equation (1), the standard deviation of the added white noise is 0.2 times the standard deviation of the original sequence, and the mean is 0. The sequence after adding white noise is obtained as follows: (1) In the formula, (t) is the nth additional random Gaussian white noise, where n =1,2,..., m ; Step 2.3, get all the ( t ), n =1,2,… m , i is the IMF order; Step 2.4, repeat m Next, in steps 2.2 and 2.3, all IMFs are obtained; Step 2.5, through formula (2) we get m Find the average value of the IMF: (2)。 4. The method for predicting cloud server software aging based on the EGGM model according to claim 1, characterized in that: The step 3 is specifically implemented according to the following steps: Step 3.1, record the sequence data vector as x , and then record the maximum and minimum values as X max and X min ; Step 3.2: Subtract the minimum value from each data in the sequence data X min ; Step 3.3: Divide the sequence data obtained in step 3.2 by the difference between the maximum value and the minimum value. (X max -X min ) ; The normalization formula is shown in formula (3): (3)。 5. The method for predicting cloud server software aging based on the EGGM model according to claim 1, characterized in that: The step 4 is specifically implemented according to the following steps: Using the normalized high-frequency IMF components obtained in step 3, divide the data set into training set and test set for each component, and then construct the three functional modules of the GRU model, including: input layer, hidden layer and output layer. The input layer accepts the initial parameters of the neural network; the hidden layers are connected to each other through GRU neurons to build a multi-layer network and iterative calculation; the output layer obtains the model prediction results and then merges the outputs.
