Real-time prediction method for server equipment load

Abstract

The invention belongs to the technical field of cloud computing, and particularly relates to a server equipment load real-time prediction method. The method comprises the following steps: collecting an equipment parameter sequence of a server in a past period of time, and dividing the equipment parameter sequence into a sample sequence and a label sequence; the sample sequence and the tag sequence are preprocessed; for the preprocessed sample sequence, clustering each coverage according to the service information, so that the similar coverage is clustered into one class, and introducing a timestamp to obtain a final sample sequence; establishing a random forest regression prediction model by adopting a random forest regression algorithm; inputting the final sample sequence and the tag sequence into the random forest regression prediction model for training to obtain a trained real-time prediction model; and inputting the equipment parameter sequence of the server in the current time period into the trained real-time prediction model to obtain the current load parameter of the server. According to the method, the load prediction accuracy is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of cloud computing, and in particular relates to a real-time prediction method for server equipment load. Background Art

[0002] Content Delivery Network (CDN) is a distributed network architecture used to optimize Internet content transmission. CDN distributes content to the nodes closest to users by deploying multiple cache servers (cache groups) around the world, thereby improving content access speed and reliability. Each cache group contains a group of devices with similar performance, which achieves load balancing through domain name hashing to improve cache hit rate.

[0003] In actual applications, cache groups usually run a variety of services in a mixed manner. Each service has different characteristics, mainly reflected in the file size and file request frequency. It is known that the operation of a single machine is closely related to the type of service. For example, the bandwidth of a single machine running short video services can reach 3G, while the bandwidth of running image services can only reach 1.5G or even lower. Exceeding the single machine operation limit will lead to a decrease in service indicators (download speed, card frame rate) and an increase in load.

[0004] Currently, the basic unit of CDN scheduling is the cache group, so before scheduling, it is necessary to evaluate the impact of the scheduling action on the load.

[0005] The traditional load assessment method is mainly calculated by the following formula:

[0006] The new cache group load factor = current load factor + incoming service bandwidth * single machine running load factor, and check whether it exceeds 1.0. However, this method has shortcomings for the following reasons:

[0007] (1) The current load factor assessment is inaccurate;

[0008] (2) The single-machine load factor is simply equivalent to the cache group factor, which is inaccurate;

[0009] Based on the above problems, it is expected to train a machine learning model that can learn iteratively and use the model's learning of historical data to predict more accurate load assessments. Summary of the invention

[0010] In order to solve the above technical problems, the present invention provides a real-time prediction method for server equipment load, which collects the trend changes of historical data series of server load, clusters cache-like services using business information, and introduces timestamp features to capture the periodic laws of the sequence, and establishes a random forest regression prediction model to optimize the prediction of the maximum load capacity of the equipment.

[0011] The technical solution adopted by the present invention to solve the technical problem includes the following steps:

[0012] A real-time prediction method for server device load, the method comprising:

[0013] (1) collecting a device parameter sequence of the server in the past period of time, and dividing the device parameter sequence into a sample sequence and a label sequence; and preprocessing the sample sequence and the label sequence;

[0014] (2) For the preprocessed sample sequence, cluster each coverage according to the business information, so that similar coverages are clustered into one category, and introduce timestamps to obtain the final sample sequence;

[0015] (3) establishing a random forest regression algorithm to establish a random forest regression prediction model; inputting the final sample sequence and the label sequence into the random forest regression prediction model for training to obtain a trained real-time prediction model;

[0016] (4) Inputting the device parameter sequence of the server in the current time period into the trained real-time prediction model to obtain the current load parameters of the server.

[0017] By collecting and preprocessing server device parameter sequences and using the random forest regression algorithm to establish a prediction model, the server load parameters can be predicted in real time and accurately. Compared with traditional methods, this solution improves the accuracy and real-time performance of load assessment, solves the problem of inaccurate assessment of current load factors and single-machine load factors, thereby optimizing CDN resource allocation and improving content access speed and system reliability.

[0018] Furthermore, the server device load refers to the maximum carrying capacity of the server CPU or the maximum carrying capacity of memory resources.

[0019] Furthermore, in step (1), the sample sequence is input in units of cache groups; the label sequence is output in units of single devices; a cache group contains multiple different devices, and the output of the cache group is the average of the outputs of the devices in the group; therefore, according to the device-cache group mapping table, the input and output need to be unified in order to perform modeling;

[0020] The sample sequence and the label sequence are preprocessed, specifically:

[0021] The sample sequence and the tag sequence are inner-connected according to the device-cache group mapping table, and aligned according to the time interval relationship, and the missing interval data in the sample sequence and the tag sequence are interpolated according to the average of the upper and lower moments;

[0022] The data with abnormal label sequences are directly deleted, and the sample sequences are normalized.

[0023] Furthermore, step (2) is specifically as follows: for the preprocessed sample sequence, clustering is performed according to the unit CPU bandwidth running volume, unit I / O bandwidth running volume, and unit SSD I / O bandwidth running volume corresponding to each coverage in the business information table, so that similar coverages in the cache are clustered into one category; so that the input feature dimension is increased; and at the same time, the timestamp feature is introduced to obtain the final sample sequence.

[0024] Furthermore, in step (3), a random forest regression algorithm is used to establish a random forest regression prediction model, specifically:

[0025] Use bootstrap resampling to collect training data several times with replacement from the training data set to generate a sample subset;

[0026] According to the sample subset, several decision trees are constructed, and out-of-bag sample data generated by each decision tree is obtained; according to the out-of-bag sample data generated by each decision tree, several corresponding decision regression trees are constructed respectively;

[0027] The actual value of the out-of-bag sample data is obtained according to the target variable, and the out-of-bag sample data is input into each decision tree for prediction to generate several predicted values; the absolute value of the difference between the predicted value and the actual value is taken as the error value of each predicted value, and the average value of the error value of each predicted value is taken as the out-of-bag error value; the quantity parameter and depth parameter of each decision tree are adjusted until the difference between the out-of-bag error value generated after adjustment and the out-of-bag error value before adjustment is less than a preset threshold, thereby generating a random forest prediction regression model.

[0028] Furthermore, according to the out-of-bag sample data generated by each decision tree, several corresponding decision regression trees are constructed respectively, specifically:

[0029] A binary tree is constructed according to a sample set of a decision tree. Starting from the root node of the binary tree, the first feature in the feature set of the decision tree is used as a split point of the binary tree, each sample in the sample set is divided into a left set and a right set, and the mean square error of the left set and the mean square error of the right set are calculated; the split point with the smallest sum of the mean square error of the left set and the mean square error of the right set is selected as the optimal split point of the first feature; the optimal split point under each feature in the feature set is calculated in turn, the mean square error of each feature under the optimal split is compared, and the feature with the smallest mean square error is selected as the optimal feature; two child nodes are generated in the current node, the optimal split point of the optimal feature is matched to the two child nodes, the sets in the two child nodes are split again for the next time, and the optimal split point and the optimal feature in the next split are found until the depth of the tree reaches the maximum depth, and the corresponding decision regression tree is generated.

[0030] Furthermore, in step (3), the final sample sequence and the label sequence are input into the random forest regression prediction model for training to obtain a trained real-time prediction model; specifically:

[0031] 1) Dividing the final sample sequence and the label sequence into a training set and a test set;

[0032] 2) inputting the training set into the random forest regression prediction model for training;

[0033] 3) Calculate the loss function between the model prediction result and the label sequence, use the gradient descent method to propagate and update the model parameters, and continuously iterate and update until the loss function converges to the minimum or reaches the maximum number of iterations, that is, the training is completed, and a trained real-time prediction model is obtained.

[0034] Furthermore, the loss function is expressed as follows:

[0035]

[0036] in: Represents Y target and The loss function between target is the tag sequence, is the model prediction result, y i represents the load output of sample sequence i, represents the load output predicted by the model for sample sequence i; n represents the total number of samples.

[0037] The beneficial technical effects of the technical solution of the present invention are:

[0038] A real-time prediction method for server device load provided by the present invention clusters each coverage according to business information, so that similar coverages are clustered into one category, thereby making the input features more fine-grained and increasing the input feature dimension; introducing timestamps to capture periodicity makes it easier to capture the periodic variation of data; combining business information clustering and periodic information to improve the accuracy of machine learning prediction; showing superiority over traditional methods in load prediction, compared with traditional methods, it can effectively improve the accuracy of load prediction, provide cloud server providers with more accurate scheduling and decision-making basis, thereby reducing the operating costs of cloud server providers. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a CDN system structure diagram;

[0040] Figure 2 The present invention is a flowchart of a method for real-time prediction of server device load in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0042] like Figure 1 As shown in the figure, the CDN operation architecture is that there are multiple equipment clusters (cache groups) in the computer room. The performance of the equipment in the cache group is similar. The domain name hash is used to do load balancing within the cache group to improve the hit rate. The cache group generally runs a variety of services. Each business has different characteristics, which are mainly reflected in the file size and file request frequency. The basic unit of current scheduling is the cache group. Therefore, the load brought by the scheduling action is evaluated before scheduling, so as to better achieve cluster load balancing.

[0043] exist Figure 1 In the context of the CDN operation architecture shown in FIG. , the present invention provides an embodiment of a method for real-time prediction of the maximum carrying capacity of a device based on iterative learning, such as Figure 2 As shown, the method comprises the following steps:

[0044] Step 1: Collect the device parameter sequence of the server in the past period of time, and divide the device parameter sequence into a sample sequence and a label sequence; pre-process the sample sequence and the label sequence;

[0045] Specifically, the device parameter sequence includes an input part and an output part. The input part includes on-demand hot, on-demand cold, download hot, download cold, pictures, number of devices, total CPU frequency, average CPU main frequency, memory size, number of SSDs, and number of hard disk blocks; the output part includes download speed, freeze rate, average load, network card utilization, I / O utilization, and CPU utilization;

[0046] The input part corresponds to the sample sequence, and the output part corresponds to the label sequence; the sample sequence is the input in units of cache groups; the label sequence is the output in units of single devices; a cache group contains multiple different devices, and the output of the cache group is the average of the outputs of the devices in the group; therefore, according to the device-cache group mapping table, the input and output need to be unified in order to perform modeling;

[0047] The sample sequence and the label sequence are preprocessed, specifically:

[0048] The sample sequence and the tag sequence are inner-connected according to the device-cache group mapping table, and aligned according to the time interval relationship, and the missing interval data in the sample sequence and the tag sequence are interpolated according to the average of the upper and lower moments;

[0049] The data with abnormal label sequences are directly deleted, and the sample sequences are normalized, thereby improving the consistency, integrity and quality of the data, providing a reliable data basis for subsequent model training. Specifically, the judgment criteria for data with abnormal label sequences are: data <= 0 and the load is greater than 300.

[0050] In this embodiment, the collected sample sequence and label sequence are recorded as:

[0051]

[0052] Sequence X represents the sample sequence, and Y represents the label sequence, where X is the set of x, Y is the set of y, n represents the characteristic dimension of the sample sequence, m is the length of the sample sequence, and k is the characteristic dimension of the label sequence y;

[0053] Step 2: For the preprocessed sample sequence, cluster each coverage according to the business information, so that similar coverages are clustered into one category, and introduce timestamps to obtain the final sample sequence;

[0054] Specifically, k-means clustering is adopted to perform clustering according to the three values ​​of unit CPU bandwidth running volume, unit I / O bandwidth running volume and unit SSD I / O bandwidth running volume in the business table. Similar business refers to the proximity of the above three indicators. In addition, because the traffic has periodic rules, such as weekly and monthly regularity, the timestamp is introduced in the present invention to capture the periodicity.

[0055] Specifically, after K-Means clustering according to business information, the input features become more fine-grained, so that the input feature dimension increases. In the experiment, the categories are set to 10. (Taking the total CPU frequency as an example, initially it is just the total CPU frequency. After clustering, it is subdivided into the CPU frequency of cluster 1 and the CPU frequency of cluster 2 according to business coverage. There are 10 types of CPU frequencies in cluster 10, which makes the granularity finer. Therefore, the sample sequence after clustering is:

[0056]

[0057] Each input feature of the sample sequence is expanded to 10 dimensions, making the input features more fine-grained while keeping the label sequence unchanged. The input is further segmented according to business coverage. For example, short video traffic requires high bandwidth, while listening to music requires less traffic. Better segmentation makes it easier to capture data patterns.

[0058] At the same time, the timestamp feature x is introduced t , which helps the model capture the periodic changes in the data; taking the week as an example, the value of the timestamp on this Monday is the same as the value of the timestamp on next Monday.

[0059] Among them, the timestamp feature is introduced. The timestamp feature is specifically expressed in days, weeks, months, or morning and evening peaks, thereby increasing the dimension and time dependence of the input features, and improving the expressiveness and prediction accuracy of the model.

[0060] Therefore, the final sample sequence is:

[0061]

[0062] Step 3: Using a random forest regression algorithm to establish a random forest regression prediction model; inputting the final sample sequence and the label sequence into the random forest regression prediction model for training to obtain a trained real-time prediction model;

[0063] Specifically, the random forest regression algorithm is used to establish a random forest regression prediction model, which is as follows:

[0064] Using bootstrap resampling method, several training data are collected with replacement from the training data set to generate a sample subset; according to the sample subset, several decision trees are constructed, and the out-of-bag sample data generated by each decision tree is obtained; according to the out-of-bag sample data generated by each decision tree, several corresponding decision regression trees are constructed respectively;

[0065] Obtain the actual value of the out-of-bag sample data according to the target variable: The process of building a decision tree according to the sample data is usually performed recursively. The specific steps are as follows:

[0066] Select a feature as the partitioning criterion for the current node, and use indicators such as information gain to evaluate the importance of the feature. Divide the sample data into different subsets based on the selected features, and each subset corresponds to a branch. For each subset, repeat the above steps and select the best feature for partitioning until the termination condition is met, such as reaching the maximum depth or the samples on the node belong to the same category. At the leaf node, determine the prediction result based on the category distribution or average value of the sample.

[0067] Out-of-bag sample data is sample data that is not used for training when building a random forest. Each time a decision tree is built, a sample subset is generated using bootstrap resampling, and some samples are not selected. These unselected samples are out-of-bag sample data. Out-of-bag sample data is used to evaluate the performance and generalization ability of decision trees, because these samples are not involved in building decision trees and can be used to verify the accuracy of the model.

[0068] The effects of this step are as follows: Generating multiple sample subsets through bootstrap resampling can increase sample diversity and reduce the risk of overfitting. Building multiple decision trees and evaluating them using out-of-bag sample data can provide an estimate of the generalization ability of the model. Predicting out-of-bag sample data through decision regression trees can provide an estimate of the performance of the model on unseen samples. The prediction results of random forests are usually the average or voting results of multiple decision regression trees, which can improve the stability and accuracy of the model.

[0069] Input the out-of-bag sample data into each decision tree for prediction, and generate several predicted values; take the absolute value of the difference between the predicted value and the actual value as the error value of each predicted value, and take the average value of the error value of each predicted value as the out-of-bag error value; adjust the quantity parameter and depth parameter of each decision tree until the difference between the out-of-bag error value generated after adjustment and the out-of-bag error value before adjustment is less than the preset threshold, and generate a random forest prediction regression model. This not only improves the prediction accuracy and stability of the model, enhances the robustness of the model to data noise, but also effectively prevents the occurrence of overfitting.

[0070] Specifically, according to the out-of-bag sample data generated by each decision tree, several corresponding decision regression trees are constructed respectively, specifically:

[0071] A binary tree is constructed according to the sample set of the decision tree, starting from the root node of the binary tree, the first feature in the feature set of the decision tree is used as the split point of the binary tree, each sample in the sample set is divided into a left set and a right set, and the mean square error of the left set and the mean square error of the right set are calculated; the split point with the smallest sum of the mean square error of the left set and the mean square error of the right set is selected as the optimal split point of the first feature; the optimal split point under each feature in the feature set is calculated in turn, the mean square error of each feature under the optimal split is compared, and the feature with the smallest mean square error is selected as the optimal feature; two child nodes are generated in the current node, the optimal split point of the optimal feature is matched to the two child nodes, the sets in the two child nodes are split again for the next time, and the optimal split point and the optimal feature in the next split are found until the depth of the tree reaches the maximum depth, and the corresponding decision regression tree is generated. The embodiment of the present invention improves the model's ability to capture complex data patterns, enhances the generalization ability of the model, and improves the accuracy and reliability of the overall prediction.

[0072] Specifically, in the process of establishing a multi-input and multi-output real-time prediction model, a random forest prediction regression model is constructed based on the time series characteristics and target variables in the training data set. The number of trees in the random forest and the maximum depth of the tree are set, and cross-validation is performed to select the optimal parameters, specifically:

[0073] Initialize the number of decision trees T and the maximum depth D of the tree in the random forest, for example, initialize T = 100, D = 10; use the bootstrap resampling method to construct 100 decision trees from the reconstructed training data set with replacement N times, generate the sample subset corresponding to each decision tree and obtain OOB data (i.e., the out-of-bag sample data generated by each decision tree) at the same time.

[0074] For each decision tree sample data, 100 CART decision regression trees are constructed. The steps are as follows:

[0075] Assume that the sample subset data corresponding to one of the decision trees is as follows:

[0076] Sample set: Samples = {s1, s2, ...s N}, a total of N samples;

[0077] Feature set: Features = {f1,f2,...f m}, each sample in the sample set corresponds to m features, M refers to the M feature data collected for each sample during data collection, but it is randomly selected when constructing each decision tree in the random forest. Features are used to construct the tree.

[0078] Starting from the root node, a binary tree is constructed for the sample set Samples, and each feature value fm of the feature set Features is traversed as a split point to divide the samples in the sample set Samples into two sets. The left set R1(j,f m )={x|χ j ≤f m}(j represents the jth eigenvalue, sample eigenvalue x j <=f m ), right set R2(j,f m )={x|χ j >f m}(j represents the jth eigenvalue, sample eigenvalue x j >f m ), each set is a tree node, and the mean square error (MSE) of the two sets is calculated when the split point is:

[0079]

[0080] in, y i is the true label sequence, c1 is the predicted value of set R1, c2 is the predicted value of set R2, N m is the number of samples in the set.

[0081] Right now, represents the mean of the target variable y;

[0082] choose The smallest f m As the optimal split point f under the corresponding feature m ;

[0083] Calculate the optimal split point for each feature, and compare the mean square error of each feature under the optimal split. Select the feature with the smallest mean square error as the optimal feature for this split, that is:

[0084]

[0085] After determining the optimal feature and the optimal split point fm′, two child nodes are generated from the current node, and the sample set is matched to the two child nodes according to the features to obtain two sets, R1(j,f m' )={x|x j ≤f m'} and R2(j,f m' )={x|x j >f m'};

[0086] The sets in the two child nodes are split until the depth of the tree reaches the maximum depth D, and the construction of the CART regression tree is completed. The OOB error is calculated using the OOB out-of-bag data and the generated regression decision tree. The out-of-bag data is predicted using the 100 generated decision trees. The absolute value of the difference between the predicted value and the actual value is calculated and the average of all absolute values ​​is taken as the OOB error value. The two parameters of the number of decision trees T and the maximum depth D of the tree in the random forest are adjusted until the OOB error value does not decrease significantly, and a trained random forest prediction regression model is generated.

[0087] The final sample sequence and the label sequence are input into the random forest regression prediction model for training to obtain a trained real-time prediction model; specifically:

[0088] 1) Dividing the final sample sequence and the label sequence into a training set and a test set;

[0089] 2) inputting the training set into the random forest regression prediction model for training;

[0090] 3) Calculate the loss function between the model prediction result and the label sequence, use the gradient descent method to propagate and update the model parameters, and continuously iterate and update until the loss function converges to the minimum or reaches the maximum number of iterations, that is, the training is completed, and a trained real-time prediction model is obtained.

[0091] The expression of the loss function is as follows:

[0092]

[0093] in: Represents Y target and The loss function between target is the tag sequence, is the model prediction result, y i represents the load output of sample sequence i, represents the load output predicted by the model for sample sequence i; n represents the total number of samples. This improves the real-time prediction capability and accuracy of the model and ensures the efficiency and reliability of the model in practical applications.

[0094] Step 4: Load prediction: Input the device parameter sequence of the server in the current time period into the trained real-time prediction model to predict the current load parameters of the server in real time.

[0095] In this embodiment, after the random forest regression prediction model is constructed, the output to be predicted can be obtained by averaging the target variable values ​​output by all decision trees of the forest model according to the input feature data.

[0096] The present invention provides a real-time prediction method for server device load, which combines business information clustering and periodic information to improve the accuracy of machine learning prediction. As shown in Table 1 and Table 2, the method provided by the present invention shows superiority over traditional methods in load prediction, and can effectively improve the accuracy of load prediction; it provides cloud server providers with more accurate scheduling and decision-making basis, thereby reducing the operating costs of cloud server providers.

[0097] Table 1 Evaluation index data of traditional methods

[0098]

[0099]

[0100] Table 2 Evaluation index data optimized by the real-time prediction method provided by the present invention

[0101]

[0102] Matters not covered in the present invention are known technologies. The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A real-time prediction method for server equipment load, characterized in that: The method comprises: (1) collecting a device parameter sequence of the server in the past period of time, and dividing the device parameter sequence into a sample sequence and a label sequence; and preprocessing the sample sequence and the label sequence; (2) For the preprocessed sample sequence, cluster each coverage according to the business information, so that similar coverages are clustered into one category, and introduce timestamps to obtain the final sample sequence; (3) using a random forest regression algorithm to establish a random forest regression prediction model; inputting the final sample sequence and the label sequence into the random forest regression prediction model for training to obtain a trained real-time prediction model; (4) Inputting the device parameter sequence of the server in the current time period into the trained real-time prediction model to obtain the current load parameters of the server.

2. A real-time prediction method for server device load according to claim 1, characterized in that: Server device load refers to the maximum load capacity of the server CPU or the maximum load capacity of memory resources.

3. A real-time prediction method for server device load according to claim 1, characterized in that: In step (1), the sample sequence is input in units of cache groups; the label sequence is output in units of single devices; The sample sequence and the label sequence are preprocessed, specifically: The sample sequence and the tag sequence are inner-connected according to the device-cache group mapping table, and aligned according to the time interval relationship, and the missing interval data in the sample sequence and the tag sequence are interpolated according to the average of the upper and lower moments; The data with abnormal label sequences are directly deleted, and the sample sequences are normalized.

4. A real-time prediction method for server device load according to claim 1, characterized in that: Step (2) is specifically as follows: for the preprocessed sample sequence, clustering is performed according to the unit CPU bandwidth running volume, unit I / O bandwidth running volume, and unit SSDI / O bandwidth running volume corresponding to each coverage in the business information table, so that similar coverages in the cache are clustered into one category; the input feature dimension is increased; and the timestamp feature is introduced at the same time to obtain the final sample sequence.

5. A real-time prediction method for server device load according to claim 1, characterized in that: In step (3), the final sample sequence and the label sequence are input into the random forest regression prediction model for training to obtain a trained real-time prediction model; specifically: 1) Dividing the final sample sequence and the label sequence into a training set and a test set; 2) inputting the training set into the random forest regression prediction model for training; 3) Calculate the loss function between the model prediction result and the label sequence, use the gradient descent method to propagate and update the model parameters, and continuously iterate and update until the loss function converges to the minimum or reaches the maximum number of iterations, that is, the training is completed, and a trained real-time prediction model is obtained.

6. A real-time prediction method for server device load according to claim 6, characterized in that: The expression of the loss function is as follows: in: Represents Y target and The loss function between target is the tag sequence, is the model prediction result, y i represents the load output of sample sequence i, represents the load output predicted by the model for sample sequence i; n represents the total number of samples.

Images