A cloud server aging prediction method based on SVR and Gaussian function

Abstract

The present invention discloses a cloud server aging prediction method based on SVR and Gaussian function, which includes collecting resource and performance data of cloud servers; obtaining resource and performance sequence data of cloud servers; preprocessing the obtained sequence data, and extracting, evaluating and fusing features of the sequence data based on a sliding window with a length of l; obtaining SVs based on the feature data obtained after SVR fitting processing, and fitting the obtained SVs with a Gaussian function; predicting the future performance change trend of the cloud server with the obtained Gaussian function, and calculating the remaining life of the cloud server; the method of the present invention enables the prediction model to predict the aging state of the cloud server in a long period of time in the future in real time and accurately, and estimate the remaining available life.

Description

Technical Field

[0001] The present invention belongs to the technical field of prediction methods for the usable life of complex systems, and in particular relates to a cloud server aging prediction method based on SVR and Gaussian function. Background Art

[0002] With the development of cloud computing, cloud servers are becoming more and more common. Cloud servers have the characteristics of long-term operation, high complexity and frequent resource exchange, which increases the risk of resource exhaustion and software system anomalies and failures. As failures and resource consumption accumulate, cloud server systems will slowly degrade in performance, increase in failure rates and even crash. This phenomenon is called "software aging". Software aging is a complex process affected by many factors, such as software anomalies, resource utilization, workload, etc. It is a common phenomenon that cloud server systems experience software aging. For example, with the long-term operation of cloud service systems, system performance gradually decreases and server response time gradually increases. When the server response time is higher than a certain threshold, the cloud server system enters an out-of-service state, also known as a "pseudo-failure". Unexpected interruptions caused by aging in enterprise systems (especially cloud platforms) can result in considerable revenue losses. A survey shows that IT downtime can reach an average of 14 hours per year, resulting in a loss of 26.5 billion US dollars. Therefore, predicting cloud server system resource and performance data is the key to dealing with software aging.

[0003] An effective and commonly used method to deal with software failure is the "software regeneration" technique. It proactively recovers the system by cleaning up the failure before a complete or partial new state failure occurs. The benefits of software regeneration technology depend largely on the time of anti-aging. The downtime or overhead caused by such operations is not negligible, and frequent software regeneration may have a negative impact on system availability. On the contrary, the ideal anti-aging strategy is to recover the system when it is close to failure.

[0004] Therefore, predicting the availability of cloud servers and calculating the remaining life of cloud servers can provide a theoretical basis for online pre-maintenance of cloud systems. Most of the existing prediction methods for the aging trend of cloud server system software are time series analysis or intelligent algorithms. For example, models such as ARMA, particle filtering and neural networks are used for trend prediction, but they can only make predictions in a short period of time, and cannot predict the changing trend of the system for a long period of time in the future, and cannot estimate the remaining life of the cloud server. In addition, the traditional prediction model must re-optimize parameters and train after the aging trend of the cloud server changes. In order to overcome the shortcomings of the above traditional prediction models, this paper proposes a prediction model based on SVR and Gaussian function, which can predict the aging indicators of cloud services for a long period of time in the future, and can adaptively change the model parameters when the aging trend of the cloud server changes, avoiding the secondary training of the model. Summary of the invention

[0005] The purpose of the present invention is to provide a cloud server aging prediction method based on SVR and Gaussian function, which overcomes the inherent defects of the above-mentioned metric-based software aging prediction method.

[0006] The technical solution adopted by the present invention is a cloud server aging prediction method based on SVR and Gaussian function, which is specifically implemented in the following steps:

[0007] Step 1: Collect resource and performance data of the cloud server;

[0008] Step 2: Obtain cloud server resource and performance sequence data;

[0009] Step 3, preprocessing the sequence data obtained in step 2;

[0010] Step 4, feature extraction, feature evaluation and feature fusion are performed on the sequence data based on a sliding window of length l;

[0011] Step 5, obtaining SVs based on the feature data obtained after processing in step 4 of SVR fitting;

[0012] Step 6, fitting the SVs obtained in step 5 with a Gaussian function;

[0013] Step 7: Based on the Gaussian function obtained in step 6, the future performance change trend of the cloud server is predicted, and the remaining life of the cloud server is calculated.

[0014] The present invention is also characterized in that:

[0015] The cloud server resource and performance sequence data in step 2 include CPU idle rate, available memory, average load and response time;

[0016] The sequence data preprocessing in step 3 is specifically to use a normalization processing method to map the original data of the cloud server to the interval [-1,1];

[0017] The pretreatment is specifically implemented in the following steps:

[0018] Step 3.1, find the average value of the sequence data and record it as x mean , the maximum value is recorded as x max ;

[0019] Step 3.2, subtract x from the sequence data mean ;

[0020] Step 3.3, use the sequence data obtained in step 3.2 and divide it by the maximum value x max ;

[0021] Step 4 is specifically implemented as follows:

[0022] Step 4.1, use a sliding window of length l to segment the sequence data;

[0023] Step 4.2, based on the segmentation result of step 4.1, calculate the feature sequence of the sequence data, and calculate the single dispatch of each feature sequence according to the feature data single dispatch algorithm, and select the feature sequence whose single dispatch is greater than the monotonic threshold σ=0.1 as the feature set;

[0024] Step 4.3, based on the feature selection results of step 4.2, use PCA to perform feature fusion;

[0025] Wherein step 5 is specifically implemented as follows:

[0026] Step 5.1, suppose the given training sample w, b are the model parameters to be determined, and the SVR model is shown in formula (1):

[0027] f(x)=w T x+b (1)

[0028] In step 5.2, let the deviation threshold be ε and the regularization constant be C, that is, the loss is calculated when and only when the absolute value of the deviation between f(x) and y is greater than ε. Therefore, the SVR regression problem can be formalized as formula (2):

[0029]

[0030] Step 5.3, introduce the slack variable ξ i , Formula (2) can be rewritten as:

[0031]

[0032] Step 5.4, introduce the Lagrange multiplier μ i , α i , The Lagrange function of formula (4) is obtained by the Lagrange multiplier method. Let the formula (4) constructed For w,b,ξ i , The partial derivative of is 0, and we can get formula (5):

[0033]

[0034]

[0035] Step 5.5, substitute formula (5) into formula (4) to obtain the dual problem of SVR:

[0036]

[0037] Where 0≤α i , And meet the conditions:

[0038]

[0039] Step 5.6, if and only if f(x i )-y i -ε-ξ i =0 when α i ≠0, if and only if hour That is, only when the sample (x i ,y i ) does not fall into the ε-interval band, the corresponding α i , Substituting formula (4) into formula (1), the solution of SVR is as follows:

[0040]

[0041] Wherein step 6 is specifically implemented as follows:

[0042] Step 6.1, set the data set It can be described by the following Gaussian function:

[0043]

[0044] Step 6.2, convert formula (8) into natural logarithm:

[0045]

[0046] Step 6.3, substitute formula (10) into formula (9), then the matrix form of formula (9) is (11):

[0047]

[0048]

[0049] Step 6.4, formula (11) can be written as: Z = XB. According to the least squares method, the generalized least squares solution of the matrix B is: B = (X T X) -1 X T Z, and then use formula (10) to solve α, β, σ to get the final form of formula (8);

[0050] Wherein step 7 is specifically implemented according to the following steps:

[0051] The Gaussian function obtained in step 6 is used to predict the future performance change trend of the cloud server. According to formula (12), the predicted time point T i The calculated RTTF is recorded as RTTF i , the prediction error of RTTF is calculated according to formula (13) and recorded as RTTF i E ; and calculate the remaining life of the cloud server;

[0052] RTTF k = inf{r:f(r+T k )≥δ} (12)

[0053] In the formula, RTTF k is the mixed model at time point T k It is estimated that the aging index of the cloud server reaches f(r+T for the first time k ) of RTTF = r, f(r+T k ) is the mixed model at time point r+T k The predicted value of k Practical Then the RTTF estimation error The calculation formula is as follows:

[0054]

[0055] The beneficial effects of the present invention are:

[0056] The cloud server aging prediction method based on SVR and Gaussian function of the present invention is based on the changing law characteristics of cloud server performance indicator data. The cloud server aging prediction method based on SVR and Gaussian function selects characteristic data with higher single scheduling, abandons the current deep learning gradient descent algorithm, and enables the prediction model to predict the aging status of the cloud server in a long period of time in the future in real time and accurately, and estimate the remaining usable life. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a cloud server response time variation diagram in a cloud server aging prediction method based on SVR and Gaussian function of the present invention;

[0058] Figure 2 It is a feature fusion result diagram in a cloud server aging prediction method based on SVR and Gaussian function of the present invention;

[0059] Figure 3 It is an overall prediction result diagram of a cloud server aging prediction method based on SVR and Gaussian function of the present invention;

[0060] Figure 4 It is a graph of prediction results at different moments in a cloud server aging prediction method based on SVR and Gaussian function of the present invention. DETAILED DESCRIPTION

[0061] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0062] Based on the relationship between the aging state and response time of a cloud server, the present invention proposes a hybrid method combining support vector regression (SVR) and Gaussian function (GF) to predict the software aging trend of the cloud server and estimate the remaining available time (Remain Time To Failure, RTTF); in the hybrid model, SVR based on a linear kernel function is used to fit the feature data extracted from the aging data of the cloud server and obtain support vectors (SVs); then GF is used to fit the change trend of the aging index of the cloud server and estimate the RTTF of the cloud server; by combining aging data, a machine learning algorithm (SVR) and a priori model (GF) to predict and estimate the aging trend of the cloud server RTTF, not only the advantages of the model-based software aging prediction method and the metric-based software aging prediction method are integrated, but also the shortcomings of the two methods are improved; on the one hand, SVR is used to sparsely represent the feature data of the aging index to reduce the cost of model training; on the other hand, the use of Gaussian function to fit SVs reduces the model's dependence on the historical data of the aging index, and can more accurately predict the change trend of the aging index. Especially in the initial stage of cloud server startup, when there are fewer aging indicator datasets.

[0063] The present invention provides a cloud server aging prediction method based on SVR and Gaussian function, and takes the cloud system server database query response time as an example for detailed description. Figure 1 As shown, the specific implementation steps are as follows:

[0064] Step 1: Collect resource and performance data of the cloud server;

[0065] Step 2: Obtain cloud server resource and performance sequence data, including CPU idle rate, available memory, average load, response time, etc.

[0066] Step 3: preprocess the sequence data obtained in step 2. Before predicting the aging of the cloud server, the data needs to be preprocessed. Otherwise, the convergence of the model prediction process is poor, which increases the difficulty and time of data training, and finally leads to a large prediction error. The normalization method is used to map the original data of the cloud server to the interval [-1,1], so that the prediction model is stable and the prediction convergence speed is fast. Specifically, the following steps are included:

[0067] Step 3.1, find the average value of the sequence data and record it as x mean , the maximum value is recorded as x max ;

[0068] Step 3.2, subtract x from the sequence data mean ;

[0069] Step 3.3, use the sequence data obtained in step 3.2 and divide it by the maximum value x max ;

[0070] Step 4: Extract features, evaluate features and fuse features of the sequence data based on a sliding window of length l: extract mathematical statistical features of the feature data, including but not limited to m features such as mean, standard deviation, skewness, kurtosis and waveform factor, and obtain a feature data set of the sequence data.

[0071] Step 4.1, use a sliding window of length l to segment the sequence data;

[0072] Step 4.2, based on the segmentation result of step 4.1, calculate the feature sequence of the sequence data, and calculate the single dispatch of each feature sequence according to the feature data single dispatch algorithm. The results are shown in Table 1. The feature sequence whose single dispatch is greater than the monotonic threshold σ=0.1 is selected as the feature set;

[0073] Table 1 Characteristics and single scheduling

[0074]

[0075]

[0076]

[0077] Step 4.3, based on the feature selection results of step 4.2, use PCA to perform feature fusion. The results are as follows Figure 2 As shown;

[0078] Step 5, based on the feature data obtained after SVR fitting step 4, obtain SVs:

[0079] Step 5.1, suppose the given training sample w, b are the model parameters to be determined, and the SVR model is shown in formula (1):

[0080] f(x)=w T x+b (1)

[0081] In step 5.2, let the deviation threshold be ε and the regularization constant be C, that is, the loss is calculated when and only when the absolute value of the deviation between f(x) and y is greater than ε. Therefore, the SVR regression problem can be formalized as formula (2):

[0082]

[0083] Step 5.3, introduce the slack variable ξ i, Formula (2) can be rewritten as:

[0084]

[0085] Step 5.4, introduce the Lagrange multiplier μ i , α i , The Lagrange function of formula (4) is obtained by the Lagrange multiplier method. Let the formula (4) constructed For w,b,ξ i , The partial derivative of is 0, and we can get formula (5):

[0086]

[0087]

[0088] Step 5.5, substitute formula (5) into formula (4) to obtain the dual problem of SVR:

[0089]

[0090] Where 0≤α i , And meet the conditions:

[0091]

[0092] Step 5.6, if and only if f(x i )-y i -ε-ξ i =0 when α i ≠0, if and only if hour That is, only when the sample (x i ,y i ) does not fall into the ε-interval band, the corresponding α i , Substituting formula (4) into formula (1), the solution of SVR is as follows:

[0093]

[0094] Step 6, fit the SVs obtained in step 5 with a Gaussian function:

[0095] Step 6.1, set the data set It can be described by the following Gaussian function:

[0096]

[0097] Step 6.2, convert formula (8) into natural logarithm:

[0098]

[0099] Step 6.3, substitute formula (10) into formula (9), then the matrix form of formula (9) is (11):

[0100]

[0101]

[0102] Step 6.4, formula (11) can be written as: Z = XB. According to the least squares method, the generalized least squares solution of the matrix B is: B = (X T X) -1 X T Z, and then use formula (10) to solve α, β, σ to get the final form of formula (8);

[0103] Step 7: Based on the Gaussian function obtained in step 6, predict the future performance change trend of the cloud server and calculate the remaining life of the cloud server:

[0104] The Gaussian function obtained in step 6 is used to predict the future performance change trend of the cloud server. According to formula (12), the predicted time point T i The calculated RTTF is recorded as RTTF i , the prediction error of RTTF is calculated according to formula (13) and recorded as RTTF i E ; and calculate the remaining life of the cloud server; the result is as follows Figure 3 As shown, the prediction comparison of i and j at different times is Figure 4 As shown, under the same conditions, the time consumption comparison between the method proposed by the present invention and the traditional model is shown in Table 3;

[0105] RTTF k = inf{r:f(r+T k )≥δ} (12)

[0106] Table 3 Time consumption

[0107]

[0108] In the formula, RTTF k is the mixed model at time point T k It is estimated that the aging index of the cloud server reaches f(r+T for the first time k ) of RTTF = r, f(r+T k ) is the mixed model at time point r+T k The predicted value of k Practical Then the RTTF estimation error The calculation formula is as follows, and the prediction results are shown in Table 2:

[0109]

[0110] Table 2 RTTF prediction results

[0111]

Claims

1. A cloud server aging prediction method based on SVR and Gaussian function, characterized in that: Please follow the steps below to implement: Step 1: Collect resource and performance data of the cloud server; Step 2: Obtain cloud server resource and performance sequence data; The cloud server resource and performance sequence data in step 2 include CPU idle rate, available memory, average load and response time; Step 3, preprocessing the sequence data obtained in step 2; The sequence data preprocessing in step 3 is specifically to use a normalization processing method to map the original data of the cloud server to the interval [-1,1]; The pretreatment is specifically carried out according to the following steps: Step 3.1, find the average value of the sequence data and record it as x mean , the maximum value is recorded as x max ; Step 3.2, subtract x from the sequence data mean ; Step 3.3, use the sequence data obtained in step 3.2 and divide it by the maximum value x max ; Step 4, feature extraction, feature evaluation and feature fusion are performed on the sequence data based on a sliding window of length l; Please follow the steps below to implement: Step 4.1, use a sliding window of length l to segment the sequence data; Step 4.2, based on the segmentation result of step 4.1, calculate the feature sequence of the sequence data, and calculate the single dispatch of each feature sequence according to the feature data single dispatch algorithm, and select the feature sequence whose single dispatch is greater than the monotonic threshold σ=0.1 as the feature set; Step 4.3, based on the feature selection results of step 4.2, use PCA to perform feature fusion; Step 5, obtaining SVs based on the feature data obtained after processing in step 4 of SVR fitting; The step 5 is specifically implemented according to the following steps: Step 5.1, suppose the given training sample w, b are the model parameters to be determined, and the SVR model is shown in formula (1): f(x)=w T x+b (1) In step 5.2, let the deviation threshold be ε and the regularization constant be C, that is, the loss is calculated when and only when the absolute value of the deviation between f(x) and y is greater than ε. Therefore, the SVR regression problem can be formalized as formula (2): Step 5.3, introduce the slack variable ξ i , Formula (2) can be rewritten as: Step 5.4, introduce the Lagrange multiplier μ i , α i , The Lagrangian function of formula (4) is obtained by the Lagrangian multiplier method. Let the formula (4) constructed For w,b,ξ i , The partial derivative of is 0, and we can get formula (5): Step 5.5, substitute formula (5) into formula (4) to obtain the dual problem of SVR: in And meet the conditions: Step 5.6, if and only if f(x i )-y i -ε-ξ i =0 when α i ≠0, if and only if hour That is, only when the sample (x i ,y i ) does not fall into the ε-interval band, the corresponding α i , Substituting formula (4) into formula (1), the solution of SVR is as follows: Step 6, fitting the SVs obtained in step 5 with a Gaussian function; The step 6 is specifically implemented according to the following steps: Step 6.1, set the data set It can be described by the following Gaussian function, where α, β, σ are real numbers and 0<α: Step 6.2, convert formula (8) into natural logarithm: Step 6.3, substitute formula (10) into formula (9), then the matrix form of formula (9) is (11): Step 6.4, formula (11) can be written as: Z = XB. According to the least squares method, the generalized least squares solution of the matrix B is: B = (X T X) -1 X T Z, and then use formula (10) to solve α, β, σ to get the final form of formula (8); Step 7: predict the future performance change trend of the cloud server based on the Gaussian function obtained in step 6, and calculate the remaining life of the cloud server; The step 7 is specifically implemented according to the following steps: The Gaussian function obtained in step 6 is used to predict the future performance change trend of the cloud server. According to formula (12), the predicted time point T is calculated. i The calculated RTTF is recorded as RTTF i , the prediction error of RTTF is calculated according to formula (13) and recorded as RTTF i E ; And calculate the remaining life of the cloud server; RTTF k =inf{r:f(r+T k )≥δ} (12) In the formula, RTTF k is the mixed model at time point T k It is estimated that the aging index of the cloud server reaches f(r+T for the first time k ) of RTTF = r, f(r+T k ) is the mixed model at time point r+T k The predicted value of k Practical Then the RTTF estimation error The calculation formula is as follows:

Images