Edge cloud server utilization prediction method, prediction device and storage medium based on boosting algorithm

Abstract

This invention provides a method, device, and storage medium for predicting edge cloud server utilization based on a boosting algorithm. The method includes: S1 obtaining predicted bandwidth utilization data based on a boosting algorithm model; S2 establishing a recent data focus mechanism, including: calculating the time difference between the actual bandwidth utilization data and the predicted bandwidth utilization data of each server; generating a random number between the minimum and maximum time difference for each actual bandwidth utilization data of a server; retaining bandwidth utilization data with time differences less than the random number to generate server focus feature data; S3 further training the boosting algorithm model based on the server focus feature data to generate an R-A boosting algorithm model to output predicted bandwidth utilization data for each server, thereby effectively filtering the data scale on the input side of the boosting model to improve model training speed and prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence applications, and in particular to an edge cloud server utilization prediction method, prediction device and storage medium based on a boosting algorithm. Background Art

[0002] With the development of edge computing network architecture, the potential of edge cloud services has been further explored, and edge cloud platforms for serving infrastructure equipment suppliers, content providers (CPs) and network users have shown great commercial value. The advantages of edge cloud service platforms are: 1. It can integrate various heterogeneous edge server resources to maximize the value of hardware; 2. It can provide flexible deployment of edge network services for various content providers; 3. It can take advantage of edge computing to provide low-latency, high-reliability edge services for top-level users.

[0003] Although the edge cloud service platform has all the advantages mentioned above, it faces important issues such as unified management of a large number of heterogeneous devices, precise deployment of task instances on servers, and efficient real-time scheduling between user needs and servers. Among them, the bandwidth utilization attribute of the edge cloud server is the main feature that the platform must refer to when deploying tasks on the server.

[0004] In actual production, the bandwidth utilization of the server has a strong positive correlation with the interests of downstream resource providers, platforms, and upstream content providers. That is, the higher the comprehensive bandwidth utilization of the server, the corresponding increase in the benefits of the three parties. In addition, a high server bandwidth utilization indicates that the task is well adapted to the corresponding server, which ensures that the server's hardware resources are not wasted and also ensures a good user experience for the users served by the server.

[0005] Therefore, how to correctly deploy tasks on the server to ensure maximum bandwidth utilization of the server is the core strategy for task deployment on the edge cloud service platform. The formulation of this strategy is inseparable from the accurate estimation of the bandwidth utilization of heterogeneous servers under different tasks and at different time nodes.

[0006] In the past, these estimation tasks were usually solved by human analysis and decision-making. However, with the development of artificial intelligence, machine learning methods have been widely used in academia and industry. The inventors began to try to predict the utilization of edge cloud servers by training machine learning models. For example, the inventors have tried to use clustering regression (Clustering) algorithm, logistic regression (Logistic Regression, LR) algorithm, multilayer perceptron (MultilayerPerceptron, MLP), XGBoosting algorithm, etc. to predict the utilization of edge cloud servers, but the results were not ideal.

[0007] Therefore, the inventor believes that the prediction accuracy of existing algorithms of this type needs to be improved, and for this reason, the inventor hopes to further improve the prediction accuracy through this case. Summary of the invention

[0008] To this end, the main purpose of the present invention is to provide an edge cloud server utilization prediction method, prediction device and storage medium based on the Boosting algorithm, so as to integrate the recent data attention screening mechanism (Recent Attention) on the basis of the Boosting algorithm, thereby effectively screening the data scale on the input side of the Boosting model to improve the model training speed and prediction accuracy.

[0009] In order to achieve the above-mentioned purpose, according to one aspect of the present invention, a method for predicting edge cloud server utilization based on a boosting algorithm is provided, and the steps include: S1, based on a boosting algorithm model, obtaining predicted bandwidth utilization data; S2, establishing a recent data focus mechanism, including: calculating the time difference between the actual bandwidth utilization data of each server and the predicted bandwidth utilization data; generating a random number between a minimum time difference and a maximum time difference for each piece of actual bandwidth utilization data of the server; retaining bandwidth utilization data with a time difference less than the random number, and generating server focus feature data; S3, based on the server focus feature data, further training the boosting algorithm model to generate an R-Aboosting algorithm model to output the predicted bandwidth utilization data of each server.

[0010] In a possible preferred implementation, the R-Aboosting algorithm model generation step in step S3 includes: S31 defining the boosting base model as a base prediction model based on a regression decision tree; S32 adjusting the complexity of the objective function and the decision tree to generate a single regression decision tree base model by splitting the decision tree nodes through an approximate algorithm; S33 using residual learning based on server-focused feature data to train and integrate all single regression decision tree base models to generate an R-Aboosting algorithm model.

[0011] In a possible preferred implementation, it also includes: step S34, setting the server utilization rate to a day scale, dividing the server focus feature data by day, and limiting the R-Aboosting algorithm model training to only provide prediction results for the day after the latest date of the currently accumulated data.

[0012] In a possible preferred implementation, it also includes: step S44, setting the server utilization rate to natural weeks as the scale, dividing the server focus feature data into natural weeks, the steps include: using a window jump data acquisition method, placing the latter group of weekly data adjacent to the former group of weekly data but without intersection, and after data division, averaging the server characteristics and bandwidth utilization rate of a group of weekly data according to the server ID as a new weekly server focus feature data for R-Aboosting algorithm model training, and outputting the average bandwidth utilization rate in a natural week starting from the next day after the latest date of the currently accumulated data.

[0013] In a possible preferred implementation, in step S32, the decision tree nodes are split by an approximate algorithm, and candidate split points are newly proposed before each split, so as to generate a single regression decision tree base model.

[0014] In a possible preferred embodiment, the steps of determining the complexity of the objective function and the decision tree in step S32 include:

[0015] Establishing the loss function

[0016]

[0017] in To predict bandwidth utilization, is the actual bandwidth utilization, N is the number of servers;

[0018] Add a regular term to the loss function to form the objective function:

[0019]

[0020] in, is the complexity of all t decision trees The sum is added to the objective function as a regular term to prevent the model from overfitting. The Taylor formula is expanded to transform the objective function into the following formula:

[0021]

[0022] Where g is the first-order derivative of the loss function, and h is the second-order derivative of the loss function.

[0023] In a possible preferred embodiment, the regularization term of the objective function is a vector consisting of the number of leaf nodes T of all decision trees generated and the weights of all nodes: of The paradigm is jointly determined, and its formula is:

[0024] .

[0025] In order to achieve the above-mentioned purpose, according to two aspects of the present invention, a method for predicting the utilization of edge cloud servers based on a boosting algorithm is also provided, and the steps include: S1 obtaining the characteristics of each server including: server performance characteristics, corresponding deployment task categories, and predicted time characteristics, and modeling the edge cloud server bandwidth utilization prediction as a regression task; S2 cleaning the server historical data set, scaling it proportionally so that the data falls into a preset specific interval, and converting it into a dimensionless pure numerical value for the boosting algorithm to learn, and outputting the predicted bandwidth utilization data; S3 establishing a recent data focus mechanism, including: calculating the time difference between the actual bandwidth utilization data of each server and the predicted bandwidth utilization data; generating a random number between the minimum and maximum time difference for each server actual bandwidth utilization data; retaining the bandwidth utilization data with a time difference less than the random number, and generating server focus feature data; S4 further training the boosting algorithm model based on the server focus feature data to generate an R-Aboosting algorithm model to output the predicted bandwidth utilization data of each server.

[0026] In a possible preferred implementation, the R-Aboosting algorithm model generation step in step S4 includes: S41 defining the boosting base model as a base prediction model based on a regression decision tree; S42 adjusting the complexity of the objective function and the decision tree to generate a single regression decision tree base model by splitting the decision tree nodes through an approximate algorithm; S43 using residual learning based on server-focused feature data to train and integrate all single regression decision tree base models to generate an R-Aboosting algorithm model.

[0027] In a possible preferred implementation, it also includes: step S44, setting the server utilization rate to a day scale, dividing the server focus feature data by day, and limiting the R-Aboosting algorithm model training to only provide prediction results for the day after the latest date of the currently accumulated data.

[0028] In a possible preferred implementation, it also includes: step S44, setting the server utilization rate to natural weeks as the scale, dividing the server focus feature data into natural weeks, the steps include: using a window jump data acquisition method, placing the latter group of weekly data adjacent to the former group of weekly data but without intersection, and after data division, averaging the server characteristics and bandwidth utilization rate of a group of weekly data according to the server ID as a new weekly server focus feature data for R-Aboosting algorithm model training, and outputting the average bandwidth utilization rate in a natural week starting from the next day after the latest date of the currently accumulated data.

[0029] In a possible preferred implementation, in step S42, the decision tree nodes are split by an approximate algorithm, and candidate split points are newly proposed before each split, so as to generate a single regression decision tree base model.

[0030] In a possible preferred embodiment, the steps of determining the complexity of the objective function and the decision tree in step S42 include:

[0031] Establishing the loss function

[0032]

[0033] in To predict bandwidth utilization, is the actual bandwidth utilization, N is the number of servers;

[0034] Add a regular term to the loss function to form the objective function:

[0035]

[0036] in, is the complexity of all t decision trees The sum is added to the objective function as a regular term to prevent the model from overfitting. The Taylor formula is expanded to transform the objective function into the following formula:

[0037]

[0038] Where g is the first-order derivative of the loss function, and h is the second-order derivative of the loss function.

[0039] In a possible preferred embodiment, the regularization term of the objective function is a vector consisting of the number of leaf nodes T of all decision trees generated and the weights of all nodes. of The paradigm is jointly determined, and its formula is:

[0040] .

[0041] In order to achieve the above-mentioned purpose, according to three aspects of the present invention, there is also provided an edge cloud server utilization prediction device, which includes: a data processing unit, a data collection unit, a storage unit and a data output unit, wherein the data collection unit collects log data including server focus feature data from each edge cloud server and sends it to the data processing unit, the storage unit stores a step execution program of the edge cloud server utilization prediction method based on the boosting algorithm as described above, and the data processing unit calls the execution program of the storage unit to process the log data, so as to output the predicted bandwidth utilization data results of each server through the data output unit.

[0042] In order to achieve the above objectives, according to the four aspects of the present invention, a computer-readable storage medium is also provided, on which a computer program is stored, wherein when the computer program is executed by a computer processor, the steps of the edge cloud server utilization prediction method based on the boosting algorithm as described in any of the above items are implemented.

[0043] The edge cloud server utilization prediction method, prediction device and storage medium based on the Boosting algorithm provided by the present invention can effectively screen the data scale on the input side of the Boosting model by establishing a recent data focus mechanism, thereby improving the model training speed and the prediction accuracy. Other beneficial effects will be further described later. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0045] Figure 1 This is a schematic diagram of periodic changes in server bandwidth according to the first embodiment of the present invention;

[0046] Figure 2 is a schematic diagram of one-hot encoding of the ISP feature of the first embodiment of the present invention;

[0047] Figure 3 is a schematic diagram of a schematic diagram of a server bandwidth variation diagram of the first embodiment of the present invention;

[0048] Figure 4 is a schematic diagram of a recent data focus (RA) mechanism according to a first embodiment of the present invention;

[0049] Figure 5 This is a data diagram showing the effect improvement after adding the data focus mechanism in the first embodiment of the present invention;

[0050] Figure 6is a schematic diagram of the structure of a decision tree according to a first embodiment of the present invention;

[0051] Figure 7 is a schematic diagram of the structure of a decision tree according to a first embodiment of the present invention;

[0052] Figure 8 is a data graph showing the error performance of the RA Boosting algorithm of the first embodiment of the present invention under long-term prediction;

[0053] Fig. 9 It is a schematic diagram of single-day model prediction and training logic of the RA Boosting algorithm of the first embodiment of the present invention;

[0054] Fig.10 It is a logical diagram of natural cycle model prediction and training of the RA Boosting algorithm according to the first embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the technical solution of the present invention, the specific technical solution of the present invention will be clearly and completely described in conjunction with the embodiments below to help those skilled in the art further understand the present invention. Obviously, the embodiments described in this case are only embodiments of a part of the present invention, not all of the embodiments. It should be pointed out that for those of ordinary skill in the art, the embodiments in this application and the features in the embodiments can be combined with each other without departing from the concept of the present invention and without conflicting with each other. Based on the embodiments in the present invention, all other embodiments obtained without creative work by those of ordinary skill in the art should belong to the disclosure and protection scope of the present invention.

[0056] In addition, the terms "first", "second", "S1", "S2", etc. in the specification, claims and drawings of the present invention are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those described herein. At the same time, the terms "including" and "having" and any variations thereof in the present invention are intended to cover non-exclusive inclusions. Unless otherwise clearly specified and limited, the terms "set", "layout", "install", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be a connection between the two elements. For those of ordinary skill in the art, the specific meanings of the above terms in this case can be understood based on the specific circumstances and in combination with the prior art.

[0057] Boosting is an algorithm that can promote weak learners to strong learners. The present invention is conceived to model edge cloud servers and analyze the influence relationship between the server performance characteristics, spatiotemporal characteristics, task category characteristics and bandwidth utilization indicators of edge cloud servers. Then, the machine learning model based on the Boosting algorithm is used to capture features to achieve prediction of bandwidth utilization changes of heterogeneous servers in the edge cloud system at different time scales.

[0058] Therefore, the ultimate effect achieved by the present invention is to provide, based on the server performance characteristics of any server (such as bandwidth, CPU, memory, etc.), the task category characteristics planned to run on the server and the prediction time characteristics, the model outputs the predicted bandwidth utilization of the server in this state.

[0059] To this end, we first need to model the edge cloud server bandwidth utilization prediction as a regression task. The hardware characteristics of the , ,in is the number of server features, which is determined based on subsequent feature correlation and importance analysis. In addition, we use Representation Server The task categories deployed on is the total number of existing tasks.

[0060] Since the same server will show completely different bandwidth utilization performance after deploying different tasks, the present invention encodes the task category into the characteristics of the server.

[0061] Finally, the present invention also predicts the time at which the server is located As the features considered by the model (with a granularity of days), in addition to the time feature dt, the periodicity of server bandwidth utilization is taken into account, such as Figure 1 As shown, there are periodic changes in bandwidth utilization of servers on the same date, week, and month.

[0062] Therefore, the time feature Further split into the week of the forecast date ( ), the week of the month ( ), the month (mon) of the year and other attributes are used as the time features of the prediction server.

[0063] According to the above definition, the bandwidth utilization rate that an edge cloud server can achieve by deploying a task on a certain date is , let the prediction model be , then the problem can be formally defined as:

[0064]

[0065] in is the number of servers, and the goal of this case is to minimize the predicted server bandwidth utilization The actual bandwidth utilization The difference between.

[0066] Specifically, in the problem, the error or loss function used in the present invention is the Mean Absolute Error (MAE), which is defined as:

[0067]

[0068] Therefore, based on the general steps of the regression prediction task and the complex heterogeneous characteristics of edge cloud services, this patent proposes the following steps to establish and train the prediction model and verify the effect, as follows:

[0069] 1. Data Collection and Preprocessing

[0070] The present invention utilizes the real large-scale historical model data set accumulated by the edge cloud service platform and adopts the mean substitution method to fill the continuous feature data. The steps include: cleaning missing values, cleaning format content, and cleaning logical errors.

[0071] In addition, it is preferred to use the box plot method to remove outliers in the data. The box plot method determines whether the feature value in a record is abnormal based on the upper and lower quartiles of the data (a certain feature).

[0072] For example: The box plot method first calculates the value of a certain feature and .in is the one-quarter quantile, is the third quartile, is the difference between the two. or When the value is judged as an outlier, the record is deleted as an outlier. After testing, it is preferred to delete only extreme outliers. For this reason, in this embodiment, it is preferred to It is set to 5, but those skilled in the art can adjust it according to actual conditions.

[0073] After cleaning the data and removing outliers, the data needs to be standardized because the data in each field has a large span and is subject to data unit restrictions. The data needs to be scaled so that it falls into a very small specific interval and converted into a dimensionless pure value, so that indicators of different units can be weighted.

[0074] In this embodiment, min-max standardization is preferably used to adjust the continuous characteristics of the server. Transform to obtain a dimensionless new feature sequence ,in .

[0075] 2. Discrete Feature Encoding

[0076] In an optional implementation, to facilitate staff understanding, the value of the discrete feature is usually in string format, or a pre-set code (such as bz for company B's business, zjtd for company Z's business). This type of feature cannot be directly learned by the model, so the discrete features need to be pre-processed.

[0077] The example of the present invention uses a one-hot encoding method to encode discrete features such as tasks, server locations, bandwidth types, and operator types. The one-hot encoding uses an N-bit state register to encode N states. Each state has its own independent register bit, and at any time, only one of them is valid.

[0078] For example, Figure 2 As shown in the figure, the ISP types of all servers are mainly China Telecom, China Mobile and China Unicom. One-hot encoding of the ISP of any server is obtained in the following form: first, the ISP feature is split into three types: isp_China Telecom, isp_China Mobile and isp_China Unicom. If the ISP type of the server is China Telecom, the corresponding code is 100.

[0079] 3. Establishing a Recent-Attention Data Focus Mechanism for the Boosting Algorithm

[0080] Based on the observation and verification of historical data, the inventor found that the closer the server bandwidth utilization change data is to the predicted date, the greater the contribution to the prediction of the future bandwidth utilization of the predicted server. This is because the actual running bandwidth of the edge cloud server is based on the business bandwidth of the most recent date. In particular, when the edge cloud server has recently experienced task switching or network fluctuations, the characteristics and bandwidth utilization of premature historical data records will become dirty data and bring errors to the model.

[0081] like Figure 3 As shown in the figure, the edge cloud server underwent a system adjustment on December 7, and the bandwidth increased and stabilized at 1.4GB. However, a large amount of previous data showed that the server's bandwidth utilization did not exceed 1G. If the full amount of data was input into the model for learning, the model would tend to predict that the server's bandwidth utilization would be low. However, in reality, the server's utilization for this business has reached 1.4 times the original.

[0082] For this reason Figure 4As shown in the figure, in order to increase the model's ability to fit recent data of the regular server, this case illustrates the design of a recent data focus (RA) mechanism for the Boosting algorithm. The main steps of this mechanism are as follows:

[0083] 1. First, calculate the time difference between each server bandwidth utilization data and the predicted data;

[0084] 2. Generate a random number between the minimum and maximum time difference for each piece of bandwidth utilization data;

[0085] 3. Use the random number as the retention threshold for the data. If the time difference of the data is greater than the threshold, delete the data; otherwise, keep the data.

[0086] like Figure 5 As shown in the figure, after adding the Recent-Attention (RA) mechanism, the prediction error of the R-Aboosting algorithm model was significantly reduced. The MAE of single-day prediction was reduced by 0.028 (28.87%), and the prediction accuracy increased by 6.1%. For natural week prediction, MAE was reduced by 0.065 (42.76%), and the prediction accuracy increased by 13%.

[0087] 4. Define the Boosting base model

[0088] The working principle of the Boosting algorithm is to combine a series of base prediction models with weaker capabilities to obtain a strong prediction model to improve the generalization ability of the model. Therefore, in the present invention, it is preferred to establish a base prediction model based on a regression decision tree, in order to determine the predicted value of the server bandwidth utilization through the joint decision of many decision trees.

[0089] The decision tree is as follows Figure 6 As shown in the figure, it is a predictor composed of a root node (containing the full set of samples), internal nodes (corresponding to feature attribute tests), and leaf nodes (representing the results of decisions). If the target variable of the leaf node uses a continuous value (usually a real number), the decision tree is called a regression tree.

[0090] In the boosting algorithm, it is preferred to define the basic model of a regression decision tree by defining the following two parts:

[0091] 1. Weight vector of leaf nodes ;

[0092] 2. The mapping relationship q from instance (sample) to leaf node;

[0093]

[0094] in is a one-dimensional vector of length T, representing the weight of each leaf node of tree q, and its function is to map the input to a leaf node. After generating enough base decision tree models, the training samples are randomly sampled with replacement by using the Bootstrap method, and repeated You will get it later The training samples together form the training data set A subset of training data , each training data subset trains a decision tree as a sample of the root node of the decision tree.

[0095] 5. Designing the complexity of the objective function and the tree

[0096] The objective function and tree complexity are key attributes that need to be referred to when splitting decision tree nodes. The present invention uses the mean absolute error to represent the predicted bandwidth utilization. The actual bandwidth utilization The difference between:

[0097]

[0098] In addition, we add a regularization term on the basis of the loss function to form the objective function:

[0099]

[0100] in, is the complexity of all t decision trees The sum is added to the objective function as a regular term to prevent the model from overfitting. The Taylor formula is expanded to transform the objective function into the following formula:

[0101]

[0102] Where g is the first-order derivative of the loss function, and h is the second-order derivative of the loss function.

[0103] In addition to the loss function, the complexity of the boosting base model decision tree defined in this embodiment is The number of leaves The fewer leaf nodes, the simpler the model. In addition, leaf nodes should not have too high weights. (Similar to the weight of each variable in LR), so the regularization term of the objective function is the vector consisting of the number of leaf nodes of all decision trees generated and the weights of all nodes The paradigm is jointly determined, and its formula is:

[0104]

[0105] 6. Splitting decision tree nodes through approximate algorithms:

[0106] The traditional decision tree node splitting algorithm generally adopts a greedy algorithm, that is, traversing all features at each node, traversing all samples for each feature to find the best splitting point as the splitting position, and splitting two new nodes on the left and right at the node.

[0107] The split benefit of each node is calculated by the difference between the objective function before and after the split, such as:

[0108]

[0109] in is the cumulative sum of the first-order partial derivatives of the samples contained in the left leaf node, which is a constant. It is the accumulated sum of the n-th order partial derivatives of the samples contained in the left leaf node, which is also a constant.

[0110] However, when the amount of data is too large, the greedy algorithm cannot read it into memory for calculation. For this reason, this case uses an approximate algorithm, which first proposes candidate partition points based on the quantiles of the feature distribution, then maps the continuous features to the buckets divided by these candidate points, and then aggregates the statistical information to find the best split points for all intervals. In order to improve the accuracy, this case prefers to re-propose candidate split points before each column division (local strategy). Although it increases the calculation steps, it is closer to the effect that the greedy algorithm can achieve. Combining the approximate algorithm with the local strategy, the algorithm description is as follows:

[0111]

[0112] The first for loop: find the candidate set of cutting mats for feature k according to the quantile of the feature distribution The purpose of this is to propose some of the split points without traversing all the split points. The method of obtaining candidate split points for a feature k is called proposal (strategy). In this model, the local strategy is preferably adopted, that is, the candidate point is selected for each split.

[0113] The second for loop: maps the value of each feature to the bucket interval divided by the candidate point set corresponding to the feature, that is, The sample statistics G and H in each bucket interval are accumulated, and finally the best split point is found based on these accumulated statistics. The purpose of this is to obtain the G and H values ​​of the candidate split points for each feature.

[0114] 7. Ensemble a single regression decision tree based on residual learning

[0115] Refer to GBDT gradient boosting decision tree, such as Figure 7 As shown, the present invention uses residual learning to train and integrate all individual regression decision tree-based models. The steps of residual learning are as follows:

[0116] 1. Use all samples to train the first decision tree. After obtaining the prediction results, calculate the difference between each sample and the label as the new label value of the sample.

[0117] 2. Retrain decision tree 2 using the new samples generated by the above process.

[0118] 3. Repeat this process until the overall objective function of the model is less than a certain value or the maximum number of decision trees is reached.

[0119] 8. Generate RA Boosting algorithm model

[0120] After the above training process is completed, the final output of the model is an additive model composed of k base models, and its final output result is:

[0121]

[0122] IX. Dual-Time-Scale Training and Prediction

[0123] Generally speaking, the utilization logs of edge cloud servers are usually based on a daily time scale, that is, each piece of data represents the characteristics and bandwidth conditions of a server on a single day. If the prediction and training logic of the model is based on data logs of this scale, the model provides a prediction of bandwidth utilization for a certain day in the future. This result can be used to discover faulty machines and recommend task switching on special dates (such as competition momentum). However, since the model only outputs the server utilization for a single day, it is not very practical for businesses that consider utilization changes over a longer period of time, such as tasks on new servers, faulty machines, and task switching on low-utilization machines.

[0124] like Figure 8 As shown, after experimental verification, the inventors found that the training error of the model for future dates will increase as the date is extended. When the error (MAE) is increased to above 0.10, the prediction accuracy of the model will be less than 80%. At the same time, the average prediction error (MAE) of the first seven days reached 0.189, which does not meet the requirements of the actual industry.

[0125] The main reason for this phenomenon also comes from the autocorrelation of edge server bandwidth changes. As the predicted date gets longer, the gap between it and the most recent data date learned by the model will increase, and the model's difficulty in predicting bandwidth utilization on that date will also increase. This has been verified in many inventions and studies.

[0126] To solve this problem, the present invention proposes two different prediction and training logics: single-day prediction and natural week prediction model.

[0127] Specifically, in order to ensure the accuracy of single-day prediction, the present invention limits the time of single-day prediction to one day in the future, that is, the model only provides the prediction results for the day after the latest date of the currently accumulated data, and its training logic remains unchanged, such as Fig. 9 As shown, learning and fitting are performed on a single day’s data log.

[0128] In addition, Fig.10 As shown, the present invention also designs a prediction model for the average bandwidth utilization of the server in a natural week (7 days), and its prediction logic is to predict the average bandwidth utilization of the server in a natural week starting from the next day.

[0129] In order to train the natural week prediction model, the data must first be divided into natural weeks and averaged. The division method adopts the window jumping data acquisition method, that is, the last group of data is adjacent to the previous group of data but there is no overlap. For example: all the data from the 1st to the 7th is taken as the front group, then the last group of data in the sliding window is from the 2nd to the 8th, and there is overlap between the data, but the inventor has tried it and found that the effect is not good, so it is preferred to use the jumping window method, and the last group of data is from the 8th to the 15th, thereby achieving no overlap with the previous group of data, thereby extracting data by group according to data time.

[0130] After the data is divided, the server characteristics and bandwidth utilization of a group of weekly data are averaged according to the server ID as a new weekly server focus feature data for R-Aboosting algorithm model training, and the average bandwidth utilization of a natural week starting from the next day of the latest date of the current accumulated data is output.

[0131] It should be noted that if the task feature in the server feature has not changed within a week, it is preferably retained. If there is a change in the middle, it can be merged into two data, each corresponding to a task, and the time feature is preferably taken as the date of the first day of the natural week, such as: the time feature of the data entry after the merger of 20220304-20220310 is 20220304, which is used as the time feature of the merged data to facilitate subsequent recommendations.

[0132] After experimental verification, the model prediction error obtained by the proposed natural week training logic is 0.087, and the accuracy is 84%, while the average error of single-day prediction for 7 consecutive days is 0.189, and the accuracy is 67%. This shows that the natural week prediction and training logic can effectively improve the model's effect on long-term prediction. The higher prediction accuracy enables the present invention to better support various businesses of the edge cloud service platform.

[0133] Based on the above method example, the present invention also provides an edge cloud server utilization prediction device, which includes: a data processing unit, a data collection unit, a storage unit and a data output unit, wherein the data collection unit collects log data including server focus feature data from each edge cloud server and sends it to the data processing unit, wherein the storage unit stores a step execution program of the edge cloud server utilization prediction method based on the boosting algorithm as described above, and the data processing unit calls the execution program of the storage unit to process the log data, so as to output the predicted bandwidth utilization data results of each server through the data output unit.

[0134] Based on the above method examples, the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a computer processor, the steps of the edge cloud server utilization prediction method based on the boosting algorithm as described in any of the above items are implemented to obtain server predicted bandwidth utilization data.

[0135] Test example

[0136] In order to verify the prediction accuracy of the server bandwidth utilization prediction of this embodiment, this case uses the mean absolute error (MAE) indicator and the mean error rate (MER) to evaluate the bandwidth utilization prediction ability of the model based on the evaluation indicators of general regression tasks. The calculation method is as follows:

[0137]

[0138]

[0139] In order to verify the advantages of the present invention, this case compared a variety of prediction methods, including:

[0140] 1. Clustering

[0141] 2. Logistic Regression (LR) Algorithm

[0142] 3. Multilayer Perceptron (MLP)

[0143] 4. XGBoosting algorithm.

[0144]

[0145] Table 1 Comparison of experimental results

[0146] As shown in Table 1, the experimental results show that compared with the existing model methods, the RA Boosting algorithm achieves the lowest prediction error in the MAE and MER indicators of single-day prediction and natural week prediction, indicating that the RA Boosting algorithm model in this case has a higher prediction accuracy.

[0147] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is only limited by the claims and their full scope and equivalents. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0148] Those skilled in the art can understand that, in addition to implementing the system, device and its various modules provided by the present invention in a purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps. Therefore, the system, device and its various modules provided by the present invention can be considered as a hardware component, and the modules included therein for implementing various programs can also be regarded as structures within the hardware component; the modules for implementing various functions can also be regarded as both software programs for implementing the method and structures within the hardware component.

[0149] In addition, all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including a number of instructions to enable a single-chip microcomputer, a chip or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), disk or optical disk and other media that can store program codes.

[0150] In addition, various implementation modes of the embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the embodiments of the present invention, they should also be regarded as the contents disclosed by the embodiments of the present invention.

Claims

1. The edge cloud server utilization prediction method based on the boosting algorithm includes the following steps: S1 obtains the predicted bandwidth utilization data based on the boosting algorithm model; S2 establishes a recent data focus mechanism, including: calculating the time difference between the actual bandwidth utilization data of each server and the predicted bandwidth utilization data; generating a random number between the minimum and maximum time difference for each piece of actual bandwidth utilization data of the server; retaining the bandwidth utilization data with a time difference less than the random number, and generating server focus feature data; S3 further trains the boosting algorithm model based on the server focus feature data to generate an R-Aboosting algorithm model to output the predicted bandwidth utilization data of each server, the steps comprising: Define the boosting base model as a base prediction model based on a regression decision tree; Adjust the complexity of the objective function and decision tree to generate a single regression decision tree base model by splitting the decision tree nodes through an approximate algorithm; Based on the server-focused feature data, residual learning is used to train and integrate all individual regression decision tree base models to generate the R-Aboosting algorithm model.

2. The edge cloud server utilization prediction method based on the boosting algorithm includes the following steps: S1 obtains the characteristics of each server including: server performance characteristics, corresponding deployment task categories, and predicted time characteristics, and models the edge cloud server bandwidth utilization prediction as a regression task; After S2 cleans the server historical data set, it scales it so that the data falls into a preset specific interval and converts it into a dimensionless pure value for the boosting algorithm to learn and output the predicted bandwidth utilization data; S3 establishes a recent data focus mechanism, including: calculating the time difference between the actual bandwidth utilization data of each server and the predicted bandwidth utilization data; generating a random number between the minimum and maximum time difference for each server actual bandwidth utilization data; retaining the bandwidth utilization data with a time difference less than the random number, and generating server focus feature data; S4 further trains the boosting algorithm model based on the server focus feature data to generate an R-Aboosting algorithm model to output the predicted bandwidth utilization data of each server, the steps including: S41 defines the boosting base model as a base prediction model based on a regression decision tree; S42 adjusts the complexity of the objective function and the decision tree to generate a single regression decision tree base model by splitting the decision tree nodes through an approximate algorithm; S43 uses residual learning based on server-focused feature data to train and integrate all individual regression decision tree base models to generate an R-Aboosting algorithm model.

3. The edge cloud server utilization prediction method based on the boosting algorithm according to claim 2, further comprising: Step S44, setting the server utilization rate to a day scale, dividing the server focus feature data by day, and limiting the R-Aboosting algorithm model training to only provide prediction results for the day after the latest date of the currently accumulated data.

4. The edge cloud server utilization prediction method based on boosting algorithm according to claim 2, further comprising: Step S44, setting the server utilization rate to natural weeks as the scale, and dividing the server focus feature data into natural weeks, the steps include: adopting the method of window jumping to obtain data, placing the latter group of weekly data adjacent to the former group of weekly data but without crossing, and after the data is divided, averaging the server characteristics and bandwidth utilization rate of a group of weekly data according to the server ID as a new weekly server focus feature data for R-Aboosting algorithm model training, and outputting the average bandwidth utilization rate of a natural week starting from the next day of the latest date of the currently accumulated data.

5. According to the edge cloud server utilization prediction method based on boosting algorithm in claim 2, wherein in step S42, the decision tree nodes are split by an approximate algorithm, and candidate split points are re-proposed before each split to generate a single regression decision tree base model.

6. The edge cloud server utilization prediction method based on boosting algorithm according to claim 2, wherein the steps of calculating the complexity of the objective function and the decision tree in step S42 include: Establishing the loss function in To predict bandwidth utilization, U i is the actual bandwidth utilization, N is the number of servers; Add a regular term to the loss function to form the objective function: in, The complexity Ω of all t decision trees is summed up and added to the objective function as a regular term to prevent the model from overfitting. The objective function is expanded by Taylor's formula to transform it into the following formula: Where g is the first-order derivative of the loss function, and h is the second-order derivative of the loss function.

7. According to the edge cloud server utilization prediction method based on boosting algorithm in claim 6, the regular term of the objective function is jointly determined by the number of leaf nodes T of all decision trees generated and the L2 norm of the vector ω composed of all node weights, and the formula is:

8. An edge cloud server utilization prediction device, comprising: A data processing unit, a data collection unit, a storage unit and a data output unit, wherein the data collection unit collects log data including server focus feature data from each edge cloud server and sends it to the data processing unit, the storage unit stores a step execution program of the edge cloud server utilization prediction method based on the boosting algorithm as described in any one of claims 1 to 7, and the data processing unit calls the execution program of the storage unit to process the log data, so as to output the predicted bandwidth utilization data results of each server through the data output unit.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a computer processor, the steps of the edge cloud server utilization prediction method based on the boosting algorithm as described in any one of claims 1 to 7 are implemented.

Images