Method, device and electronic equipment for dynamically adjusting TCP full connection queue in container

Abstract

The application provides a method and device for dynamically adjusting a TCP full connection queue in a container and an electronic device. The device state parameters of a server device are obtained, the device state parameters are input into a target full connection queue prediction model trained in advance, and the full connection queue value of the server device is adjusted according to the output result of the model. According to the embodiment of the application, the correlation between the device state parameters and the full connection queue value adjustment amount is learned by means of a training machine learning algorithm model until the full connection queue prediction model trained can accurately output the corresponding full connection queue value adjustment amount according to the input device state parameters. In actual application, the full connection queue value is dynamically adjusted according to the current device state of the server device, thereby reducing the problem that the TCP connection request is rejected due to the fixed full connection queue value threshold, and further ensuring the communication and data transmission efficiency between Pods.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet operation and maintenance, and particularly relates to a method and device for dynamically adjusting a TCP full connection queue in a container and an electronic device. BACKGROUND

[0002] Kubernetes, K8s for short, is a container orchestration engine, and is specifically used for managing containerized applications on a plurality of hosts of a cloud platform. A container is an independent running environment that can isolate an application program and its association, and K8s deploys containers in a Pod. A Pod is the smallest unit managed by K8s, and a Pod can include a plurality of containers. The Pods communicate and transmit data without obstacles through internal TCP full connections.

[0003] In the process of building a TCP full connection, after receiving an ACK (Acknowledge character) packet of a third handshake of a client, a kernel removes the connection from a half-connection queue, then creates a new full connection, and puts the connection into a full-connection queue, waiting for an application program to call an accept function to take away the connection.

[0004] The existing containers in a Pod of K8s take a default value as a threshold value of an internal full-connection queue, and when the threshold value is reached, a new TCP full connection is refused to be established. However, the default value is usually small, and for some scenarios of high-performance TCP full connection services, the default small full-connection queue value will affect the communication and data transmission efficiency between Pods. SUMMARY

[0005] Therefore, the embodiments of the present application provide a method and device for dynamically adjusting a TCP full connection queue in a container and an electronic device, so as to dynamically adjust the TCP full-connection queue value in the container, and further guarantee the communication and data transmission efficiency between Pods.

[0006] In a first aspect, the embodiments of the present application provide a method for dynamically adjusting a TCP full connection queue in a container, and the method is applied to a server device, and the method comprises the following steps.

[0007] obtaining a device state parameter of the server device, wherein the device state parameter comprises a CPU usage, a memory usage, and a disk I / O waiting time;

[0008] input the device state parameter into a target full connection queue prediction model trained in advance, and adjust the full connection queue value of the service end device according to the result output by the target full connection queue prediction model;

[0009] The target full connection queue prediction model is trained in advance by the following method:

[0010] input the preprocessed sample data set into an initial full connection queue prediction model; the preprocessed sample data set includes a plurality of training sample data, each of which includes device state parameters of each service end device and an actual full connection queue value adjustment amount corresponding to the service end device, and the full connection queue prediction model is a model constructed based on a machine learning algorithm;

[0011] obtain a predicted full connection queue value adjustment amount output by the full connection queue prediction model based on the device state parameters in each of the training sample data according to a target correlation relationship, wherein the target correlation relationship is learned based on each of the training sample data, and is used to represent the correlation between the device state parameters and the full connection queue value adjustment amount;

[0012] adjust the model parameters of the full connection queue prediction model according to a target data difference between the predicted full connection queue value adjustment amount and the actual full connection queue value adjustment amount until the target data difference converges;

[0013] determine the full connection queue prediction model corresponding to the time when the target data difference converges as the target full connection queue prediction model.

[0014] In some possible embodiments, the full connection queue prediction model is a comprehensive model obtained by integrating a plurality of basic learning models and secondary learning models based on a stacking ensemble learning algorithm;

[0015] The basic learning model includes a first basic learning model, a second basic learning model, and a third basic learning model. The first basic learning model is a model constructed based on a linear regression algorithm, the second basic learning model is a model constructed based on a support vector machine algorithm, and the third basic learning model is a model constructed based on a decision tree algorithm.

[0016] The secondary learning model is used to integrate the basic full connection queue value adjustment amounts output by each of the basic learning models to generate the predicted full connection queue value adjustment amount.

[0017] In some possible embodiments, the sample data set comprises: a training data set, and the obtaining of the predicted full connection queue value adjustment amount by the full connection queue prediction model comprises:

[0018] dividing the training data set into different training data subsets, and inputting each training data subset into the first base learning model, the second base learning model, and the third base learning model respectively, and outputting a respective base full connection queue value adjustment amount by each base learning model based on the input training data subset;

[0019] constructing new sample data according to each base full connection queue value adjustment amount and the device state parameter corresponding to each base full connection queue value adjustment amount, and inputting the new sample data into the secondary learning model, and outputting the predicted full connection queue value adjustment amount by the secondary learning model based on the input sample data.

[0020] In some possible embodiments, the adjusting of the model parameters of the full connection queue prediction model according to the target data difference between the predicted full connection queue value adjustment amount and the actual full connection queue value adjustment amount comprises:

[0021] constructing a loss function according to the predicted full connection queue value adjustment amount and the actual full connection queue value adjustment amount;

[0022] updating the model parameters of the secondary learning model based on the function value of the loss function by back propagation until the function value of the loss function converges.

[0023] In some possible embodiments, the first base learning model is trained in the following manner:

[0024] inputting a first sample data set into an initial first base learning model, and outputting a first base full connection queue value adjustment amount by the first base learning model based on the following formula:

[0025] first base full connection queue value adjustment amount = w0 + w1 x CPU usage rate + w2 x memory usage rate + w3 x disk I / O latency;

[0026] wherein w0 is an intercept, w1 is a weight of the CPU usage rate, w2 is a weight of the memory usage rate, and w3 is a weight of the disk I / O latency;

[0027] The mean square error between the first basic full connection queue value adjustment amount and an actual full connection queue value adjustment amount corresponding to the first basic full connection queue value adjustment amount is calculated, and each weight is adjusted according to the mean square error until the mean square error converges.

[0028] In some possible embodiments, the second basic learning model is trained in the following manner:

[0029] The second sample data set is input into an initial second basic learning model, and input sample data is calculated by the second basic learning model based on a support vector machine (SVM) objective function until optimal w and optimal b are determined:

[0030]

[0031] wherein i = 1, 2, 3, …, N, N is the number of sample data in the second sample data set, x i is a feature vector of the ith sample data in the second sample data set, y i is a label corresponding to an actual full connection queue value adjustment amount of the ith sample data, ω is a normal vector of a hyperplane of the SVM, and b is an intercept of the hyperplane of the SVM.

[0032] In some possible embodiments, the third basic learning model is constructed in the following manner in advance:

[0033] According to each sample data in a third sample data set, and according to different device state parameters, a best split point corresponding to each device state parameter and a sample data subset corresponding to each best split point are determined;

[0034] Each sample data subset is used as each leaf node in a decision tree, and the third basic learning model is constructed.

[0035] The third basic learning model is used to recursively calculate a full connection queue value adjustment amount mean value corresponding to each leaf node, and the full connection queue value adjustment amount mean value corresponding to the leaf node is determined as a third basic full connection queue value adjustment amount of the leaf node.

[0036] In a second aspect, an apparatus for dynamically adjusting a TCP full connection queue in a container is provided, and the apparatus is applied to a server device. The apparatus comprises:

[0037] An obtaining module is configured to obtain a device state parameter of the server device, wherein the device state parameter comprises a CPU usage, a memory usage, and a disk I / O waiting time.

[0038] The determining module is configured to input the device state parameter into a target full connection queue prediction model trained in advance, and adjust a full connection queue value of the server device according to a result output by the target full connection queue prediction model.

[0039] The target full connection queue prediction model is trained in advance in the following manner:

[0040] The preprocessed training sample data set is input into an initial full connection queue prediction model, wherein the preprocessed training sample data set includes a plurality of training sample data, each of the training sample data includes device state parameters of each server device and an actual full connection queue value adjustment amount corresponding to the server device, and the full connection queue prediction model is a neural network model constructed based on a machine learning algorithm.

[0041] A predicted full connection queue value adjustment amount output by the full connection queue prediction model based on the device state parameters in each of the training sample data according to a target correlation relationship is obtained, wherein the target correlation relationship is learned based on each of the training sample data, and the target correlation relationship is used to represent a correlation between the device state parameters and the full connection queue value adjustment amount.

[0042] Model parameters of the full connection queue prediction model are adjusted according to a target data difference between the predicted full connection queue value adjustment amount and the actual full connection queue value adjustment amount until the target data difference converges.

[0043] The full connection queue prediction model corresponding to the target data difference when the target data difference converges is determined as the target full connection queue prediction model.

[0044] In a third aspect, an electronic device is provided, and the electronic device includes:

[0045] a processor; and

[0046] a memory storing a program,

[0047] The program includes instructions that, when executed by the processor, cause the processor to perform the method for dynamically adjusting a TCP full connection queue in a container according to the first aspect.

[0048] In a fourth aspect, a non-transitory computer-readable storage medium storing computer instructions is provided, and the computer instructions are used to cause a computer to perform the method for dynamically adjusting a TCP full connection queue in a container according to the first aspect.

[0049] The application has the following beneficial effects:

[0050] The application provides a method, device and electronic equipment for dynamically adjusting a TCP full connection queue in a container. The method is applied to a server device. A device state parameter of the server device is obtained, the device state parameter is input into a target full connection queue prediction model, a queue value is output by the target full connection queue prediction model based on the input device state parameter, and the full connection queue value of the server device is adjusted according to the output result of the model. The target full connection queue prediction model is a machine learning algorithm model trained by using the device state parameter and the actual full connection queue value adjustment amount as training sample data. The full connection queue prediction model learns the correlation between the device state parameter and the full connection queue value adjustment amount through continuous training until it can accurately output the corresponding predicted full connection queue value adjustment amount according to the device state parameter.

[0051] According to the embodiments of the application, the correlation between the device state parameter and the full connection queue value adjustment amount is learned by training the machine learning algorithm model until the target full connection queue prediction model can accurately output the corresponding full connection queue value adjustment amount according to the input device state parameter. In actual application, the full connection queue value can be quickly and accurately adjusted according to the current device state of the server device, thereby reducing the problem that the TCP connection request is rejected due to the fixed full connection queue value threshold, and further ensuring the communication and data transmission efficiency between Pods. BRIEF DESCRIPTION OF DRAWINGS

[0052] In the following description of exemplary embodiments in conjunction with the accompanying drawings, more details, features and advantages of the application are disclosed, in which:

[0053] Figure 1 A flowchart of a method for dynamically adjusting a TCP full connection queue in a container is shown;

[0054] Figure 2 A flowchart of a training method of a target full connection queue prediction model is shown;

[0055] Figure 3 A model architecture diagram of a target full connection queue prediction model is shown;

[0056] Figure 4 A logical architecture diagram of a device for dynamically adjusting a TCP full connection queue in a container is shown;

[0057] Figure 5 A structural block diagram of an exemplary electronic device capable of implementing the embodiments of the application is shown. Detailed Implementation

[0058] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0059] It should be understood that the steps described in the method embodiments of this application may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this application is not limited in this respect.

[0060] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc., mentioned in this application are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0061] It should be noted that the terms "a" and "a plurality of" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0062] Firstly, this application provides a method for dynamically adjusting the TCP full-connection queue within a container. This method is applicable to any electronic device capable of dynamically adjusting the TCP full-connection queue within a container, including but not limited to personal mobile terminals, computers, or servers. As a preferred embodiment, this method for dynamically adjusting the TCP full-connection queue within a container is primarily applied to server-side devices in TCP connection-based devices. It can also be a virtual device acting as a server, such as a Pod acting as a server in a Kubernetes Pod. Figure 1 As shown, the method includes the following steps:

[0063] S11. Obtain the device status parameters of the server device, including: CPU utilization, memory utilization, and disk I / O wait time.

[0064] S12, input the device state parameter into a target full connection queue prediction model trained in advance, and adjust the full connection queue value of the server device according to the result output by the target full connection queue prediction model.

[0065] As an embodiment, the target full connection queue prediction model can be trained in the following steps: Figure 2 The target full connection queue prediction model is trained in the following steps:

[0066] S21, input the preprocessed sample data set into an initial full connection queue prediction model; wherein the preprocessed sample data set includes a plurality of training sample data, each of which includes device state parameters of each server device and an actual full connection queue value adjustment amount corresponding to the server device, and the full connection queue prediction model is a model constructed based on a machine learning algorithm;

[0067] S22, obtain a predicted full connection queue value adjustment amount output by the full connection queue prediction model based on the device state parameters in each of the training sample data according to a target correlation; wherein the target correlation is a target correlation learned based on each of the training sample data, and the target correlation is used to represent the correlation between the device state parameters and the full connection queue value adjustment amount;

[0068] S23, adjust the model parameters of the full connection queue prediction model according to a target data difference between the predicted full connection queue value adjustment amount and the actual full connection queue value adjustment amount until the target data difference converges;

[0069] S24, determine the full connection queue prediction model corresponding to the converged target data difference as the target full connection queue prediction model.

[0070] The method is applied to a server device, and the device state parameters of the server device are obtained. The device state parameters are input into a target full connection queue prediction model trained in advance. A queue value is output by the target full connection queue prediction model based on the input device state parameters. The full connection queue value of the server device is adjusted according to the result output by the model. Since the target full connection queue prediction model is a pre-trained artificial intelligence model obtained by training a model based on a machine learning algorithm with device state parameters and actual full connection queue value adjustment amounts as training sample data, the full connection queue prediction model continuously learns the correlation between the device state parameters and the full connection queue value adjustment amount through continuous training until it can accurately output the corresponding predicted full connection queue value adjustment amount according to the device state parameters.

[0071] With the embodiments of the present application, the correlation between the device state parameters and the full connection queue value adjustment amount is learned by means of training the machine learning algorithm model until the full connection queue prediction model obtained by training can accurately output the corresponding full connection queue value adjustment amount according to the input device state parameters. In actual application, the full connection queue value can be quickly, accurately and dynamically adjusted according to the current device state of the server device, thereby reducing the problem that the TCP connection request is rejected due to the fixed full connection queue value threshold, and further ensuring the communication and data transmission efficiency between Pods.

[0072] The above steps S11, S12, S21-S24 will be described in detail below in combination with specific examples:

[0073] From the perspective of the training and application of the model, the above steps S11 and S12 belong to the application stage of the full connection queue prediction model, and steps S21-S24 belong to the training stage of the full connection queue prediction model. The model training stage and the model application stage can occur on the same device or on different devices. That is, the execution subject device of the above steps S21-S24 and the execution subject of steps S11-S12 can be the same device, or steps S21-S24 can be executed in device A, and after the training of the full connection queue prediction model is completed in device A, the target full connection queue prediction model obtained by training is deployed to device B, and device B runs the target full connection queue prediction model to determine the appropriate full connection queue value adjustment amount based on its own device state parameters to adjust its own full connection queue value.

[0074] In the embodiments of the present application, when step S11 is executed, the device state parameters of the server device are obtained, which can be obtained by reading the monitoring results of the device state monitoring function module. As an implementation manner, the device state monitoring function module can be any type of existing device state parameter monitoring software tool. For example, the device state parameters of the server device can be obtained by real-time monitoring of the CPU usage, memory usage, disk I / O (Input / Output) latency of the server device through the Prometheus (a monitoring and alarm system) system deployed in K8s.

[0075] Then, by inputting the device state parameters as model input data into the target full connection queue prediction model trained by steps S21-S24, step S12 is executed, the output result of the target full connection queue prediction model is obtained, and the current full connection queue value is adjusted based on the output result of the target full connection queue prediction model, thereby achieving the effect of dynamically adjusting the full connection queue value threshold of the server device and reducing the rejection of TCP full connection requests sent by the client device.

[0076] The output result of the target full connection queue prediction model depends on the output result type specified during the training phase. As an implementation, the output result of the target full connection queue prediction model can be a direct output of a full connection queue value, which is the new full connection queue value threshold. As another implementation, the output result of the target full connection queue prediction model can be a full connection queue value adjustment amount. When step S12 is performed, the full connection queue value adjustment amount can be added or subtracted from the existing full connection queue value to achieve the effect of dynamically adjusting the full connection queue value.

[0077] In the embodiments of the present application, the target full connection queue prediction model is the key to dynamically adjusting the full connection queue value, and how to train the target full connection queue prediction model is the core of achieving the dynamic adjustment of the full connection queue value. The training process of the target full connection queue prediction model will be described in detail below in conjunction with specific examples, i.e., steps S21-S24 will be described in detail:

[0078] In the embodiments of the present application, when step S11 or step S21 is performed, as an implementation, the Prometheus system can be used to obtain the original data of each Pod running in K8s. The Pod original data is the device state data of the Pod in history and the full connection queue value adjustment amount of the Pod. In the embodiments of the present application, the device state data can include CPU usage, memory usage, disk I / O latency, process quantity, high concurrency quantity, and other data of each Pod. The device state parameters can be flexibly selected according to the actual scene, and other device state parameters such as process quantity and high concurrency quantity can be added based on CPU usage, memory usage, and disk I / O latency. The full connection queue value adjustment amount, which is equivalent to Δ full connection queue value, is an increment that is added to the existing full connection queue value to obtain a new full connection queue value. Hereinafter, the full connection queue value is referred to as S, and the full connection queue value adjustment amount is referred to as ΔS.

[0079] That is, before step S21 is performed, the historical CPU usage, memory usage, disk I / O latency, and other device state parameters of each Pod, as well as the historical ΔS of each Pod, can be obtained by Prometheus to construct sample data for training or testing the full connection queue prediction model. As an implementation, the historical CPU usage, memory usage, disk I / O latency, and other device state parameters of each Pod, as well as the historical ΔS of each Pod, can be added to the sample data for training or testing the full connection queue prediction model to add label encoding, feature selection, and other operations to generate more accurate sample data.

[0080] Wherein, although △S itself is a numerical characteristic, it does not need to be labeled and encoded, but assuming that there are other types of characteristics in the data set, such as the model of the service device Server Model, it can be converted into a numerical characteristic using label encoding. Specifically, assuming that the sample data set obtained is as shown in Table 1:

[0081] Table 1. A sample data set table

[0082]

[0083] By converting the Server Model type of characteristic into a numerical characteristic, such as converting it into the corresponding sequential number in alphabetical order, i.e., Model A is converted into the number 0, Model B is converted into the number 1, and so on, the following Table 2 shows a sample table:

[0084] Table 2. A sample data set table after label encoding processing

[0085]

[0086] Based on the numerical conversion obtained sample data set, the correlation between the device state parameters and △S can be determined by correlation analysis method, and the influence degree (or importance) of each device state parameter on △S. Among them, as an implementation, the linear correlation degree between variables can be determined by the following formula:

[0087]

[0088] Wherein, x i and y i are the i-th observation values of two variables, and are the means of two variables, and n is the number of observations.

[0089] If you want to evaluate the correlation between CPU usage, memory usage, disk I / O latency and full connection queue value adjustment △S, you can replace CPU usage, memory usage, disk I / O latency in Table 2 above as x i and △S (as y i ) into the above linear correlation degree calculation formula, and calculate the correlation coefficient of CPU usage, memory usage, disk I / O latency. Specifically, as shown in Table 3:

[0090] Table 3. A correlation coefficient table

[0091] Feature name (device status parameter name) Correlation coefficient between device status parameter and ΔS CPU usage rate (%) 0.3 Memory usage rate (%) 0.4 Disk I / O latency (ms) 0.7

[0092] Based on the calculated correlation coefficients, the correlation between each device status parameter and the full-connection queue value adjustment can be determined. The larger the correlation coefficient, the stronger the correlation between the corresponding device status parameter and the full-connection queue value adjustment. It is evident that the correlation between memory utilization, disk I / O latency, and the full-connection queue value adjustment is stronger than that between CPU utilization. Therefore, during the training of the full-connection queue prediction model, disk I / O latency and memory utilization can be preferentially selected as the input sample feature set for training the full-connection queue prediction model.

[0093] After labeling and feature selection of the aforementioned sample dataset, a sample feature set is obtained, which includes several sample features. As one implementation method, this sample feature set can be divided into a training dataset and a test dataset. Specifically, the first 80% of the sample features in the sample feature set can be designated as the training dataset, and the remaining 20% ​​as the test dataset. The fully connected queue prediction model is first trained using the training dataset, and then the functionality of the trained fully connected queue prediction model is tested using the test dataset to determine whether the trained fully connected queue prediction model can accurately output the corresponding fully connected queue value adjustment.

[0094] In some possible embodiments, the fully connected queue prediction model is a comprehensive model obtained by integrating several basic learning models and secondary learning models based on the Stacking ensemble learning algorithm; specifically, it can be as follows: Figure 3 As shown, the basic learning models in this fully connected queue prediction model include: a first basic learning model, a second basic learning model, and a third basic learning model. A secondary learning model is appended after the basic learning models using a stacking ensemble learning algorithm. The secondary learning model integrates the basic fully connected queue value adjustments output by each of the basic learning models to generate the predicted fully connected queue value adjustment. Specifically, the first basic learning model is constructed using a linear regression algorithm, the second basic learning model is constructed using a support vector machine algorithm, and the third basic learning model is constructed using a decision tree algorithm.

[0095] Based on this, in some possible embodiments, the training sample data set can be divided into different training data subsets, and then each training data subset is input into the first basic learning model, the second basic learning model and the third basic learning model respectively, and the first basic learning model, the second basic learning model and the third basic learning model are trained through different training data subsets. For example, the training data set can be divided into subset A, subset B and subset C, then subset A is input into the first basic learning model to train the first basic learning model, subset B is input into the second basic learning model to train the second basic learning model, and so on.

[0096] The output of the first basic learning model constructed by the linear regression algorithm, the output of the second basic learning model constructed according to the support vector machine algorithm, and the output of the third basic learning model constructed according to the decision tree algorithm are connected with the input of the secondary learning model, and the respective basic full connection queue value adjustment amount is output by each basic learning model based on the input training data subset. In this way, multiple machine learning algorithms such as linear regression, support vector machine and decision tree can be integrated into one comprehensive model, so that the advantages of different algorithms can be fully utilized, and the prediction accuracy and application range of the entire comprehensive model can be improved. It is worth noting that in the embodiments of the present application, the basic learning model can not only include the first basic learning model, the second basic learning model and the third basic learning model, but also can include other basic learning models constructed based on other types of machine learning algorithms. For example, it can also include a fourth basic learning model constructed by using a random forest algorithm. Similarly, the output of the fourth basic learning model can be connected to the secondary learning model, and the secondary learning model can perform secondary prediction based on the input basic full connection queue value adjustment amount (equivalent to a primary prediction result), so as to output a precise prediction result of the full connection queue value adjustment amount.

[0097] In the embodiments of the present application, new sample data can be constructed according to each basic full connection queue value adjustment amount and the device state parameter corresponding to each basic full connection queue value adjustment amount, and the new sample data is input into the secondary learning model, and the secondary learning model outputs the predicted full connection queue value adjustment amount based on the input sample data. In this way, the amount of input sample data of the secondary learning model can be increased, the prediction accuracy of the secondary learning model can be further improved, and the output accuracy of the entire full connection queue prediction model can be improved.

[0098] In some possible embodiments, the first basic learning model is a model constructed based on a linear regression algorithm, which is trained in advance by the following method:

[0099] The first sample data set is input into an initial first basic learning model, and the first basic learning model outputs a first basic full connection queue value adjustment amount based on the following formula:

[0100] The first basic full connection queue value adjustment amount = w0 + w1 x CPU usage rate + w2 x memory usage rate + w3 x disk I / O latency.

[0101] wherein w0 is an intercept, w1 is a weight of the CPU usage rate, w2 is a weight of the memory usage rate, and w3 is a weight of the disk I / O latency.

[0102] The mean square error between the first basic full connection queue value adjustment amount and an actual full connection queue value adjustment amount corresponding to the first basic full connection queue value adjustment amount is calculated, and each weight is adjusted according to the mean square error until the mean square error converges.

[0103] Specifically, the least square method or other optimization algorithm can be used to estimate the weight values of the above-mentioned device state parameters and the intercept, and w0-w3 can be calculated through the following mean square error formula:

[0104]

[0105] wherein J(w) is the mean square error, y i is an actual full connection queue value adjustment amount corresponding to the i-th sample data input into the first basic learning model, is a predicted full connection queue value adjustment amount output by the first basic learning model for the i-th sample data input, and n is the number of sample data participating in model training in the first sample data set. In this way, the first basic learning model of the linear regression algorithm model can learn how to combine the predicted result with the original device state parameter features to give an accurate predicted full connection queue value adjustment amount.

[0106] In some possible embodiments, the second basic learning model is a model constructed based on a support vector machine (SVM) algorithm, which is trained in advance by the following method:

[0107] The second sample data set is input into an initial second basic learning model, and the second basic learning model calculates the input sample data based on the following support vector machine (SVM) objective function until the optimal w and the optimal b are determined:

[0108]

[0109] wherein i = 1, 2, 3, ···, N, N is the number of sample data in the second sample data set, x ia feature vector of the i-th sample data in the second sample data set, y i a label corresponding to the actual full connection queue value adjustment of the i-th sample data, ω is a normal vector of a hyperplane of the support vector machine SVM, and b is an intercept of the hyperplane of the support vector machine SVM.

[0110] The basic idea of the support vector machine SVM is to find a hyperplane, so that the projection interval between the feature vector corresponding to the input device state parameter and the actual full connection queue value adjustment corresponding to the feature vector on the hyperplane is maximized, and this hyperplane is called a maximum interval hyperplane. In the embodiment of the present application, y i a label corresponding to the actual full connection queue value adjustment of the i-th sample data, which can be generated based on the actual full connection queue value adjustment of the sample data by using some existing label generation algorithm. In the embodiment of the present application, for some linearly inseparable cases, slack variables and penalty coefficients can be introduced to handle abnormal values generated in the process, and specific slack variables and penalty coefficients can refer to existing technical disclosure documents, which will not be described here.

[0111] The feature vector x i The corresponding CPU usage, memory usage, and disk I / O usage time are input into the formula of the support vector machine SVM, and the SVM library such as SVR in scilkit-learn is used to train the second basic learning model. Then, the feature vector and the label are substituted into the model for training, until the optimal hyperplane parameters w and b are found, and the objective function is minimized and the above training constraints are satisfied. In the embodiment of the present application, as the model is trained and the application process is deepened, new sample data will inevitably be generated in the process and added to the model training process. For the new sample data, the trained model can be used for prediction, and the predicted value The new predicted value can be calculated by the formula f(x)=sign(ωx+b) for classification problems or similar regression formulas (for regression problems).

[0112] In some possible embodiments, the third basic learning model is a model constructed based on a decision tree algorithm, which is trained in advance by the following method:

[0113] According to each sample data in the third sample data set, according to different device state parameters, the best split point corresponding to each device state parameter and the sample data subset corresponding to each best split point are determined.

[0114] Each sample data subset is used as each leaf node in the decision tree, and the third basic learning model is constructed.

[0115] and the third basic learning model recursively calculates a full connection queue value adjustment amount mean value corresponding to each of the leaf nodes based on each of the leaf nodes, and determines the full connection queue value adjustment amount mean value corresponding to the leaf node as a third basic full connection queue value adjustment amount of the leaf node.

[0116] In the embodiments of the present application, an initial decision tree is first constructed to obtain an initial third basic learning model. Then, different device state parameters are determined to correspond to different split features, i.e., split features are selected for the initial third basic learning model, wherein one split feature corresponds to one device state parameter. For example, the CPU usage rate can be regarded as a split feature, and the memory usage rate is another split feature. Then, the corresponding optimal split point is determined for each split feature.

[0117] The specific way of determining the optimal split point is to traverse all possible values of each device state parameter and calculate the information gain when each possible value is used as a split point. For example, the CPU usage rate is used as a split feature, and the CPU usage rate = 75% is used as a possible value. Then, the sample data set is divided into two subsets: subset 1: CPU usage rate <= 75%, and subset 2: CPU usage rate > 75%.

[0118] Taking the sample data set shown in Table 2 above as an example, the following can be obtained:

[0119] Subset 1 (CPU usage rate <= 75%): sample 1, sample 3, sample 4 (assuming that the CPU utilization rates of sample 1, sample 3, and sample 4 are less than 75%).

[0120] Subset 2 (CPU usage rate > 75%): sample 2, sample 5 (assuming that the CPU utilization rates of sample 2 and sample 5 are greater than 75%).

[0121] For each subset, the above process is recursively calculated to divide the subset until a recursive stop condition is met. The recursive stop condition can be that the size of the subset is less than a certain threshold. Assuming that the average value of the full connection queue value adjustment amount of subset 1 and subset 2 is used as the prediction value of the leaf node, the following can be obtained:

[0122] The mean value of the full connection queue value adjustment amount of subset 1 = (50+-30+60) / 3 = 26.67 (rounded to the unit of 27).

[0123] The mean value of the full connection queue value adjustment amount of subset 2 = (80+90) / 2 = 85.

[0124] In this way, the CPU usage rate = 75% as a split point can obtain two branches:

[0125] If the CPU usage rate ≤ 75%, output the predicted full connection queue value adjustment amount = 27;

[0126] Otherwise, output the predicted full connection queue value adjustment amount = 85.

[0127] Based on this, information gain calculation can be performed. Specifically, by calculating the variance of the data set before and after splitting, the information gain of the data set is calculated. For example, taking data set D as an example, assuming that data set D contains n sample data, and each sample data has a target full connection queue value adjustment amount y, the variance of data set D before and after splitting can be calculated, and the variance change amount is calculated. The variance change amount represents the information gain of the data set D. Since the average value is used as the predicted value of the leaf node in the embodiment of the present application, the explicit calculation of the information gain can be skipped, and the average value can be directly used to split the data set. For each leaf node, the average value of the actual full connection queue value adjustment amount is calculated as the predicted full connection queue value adjustment amount.

[0128] In the above manner, the first, second, and third basic learning models can be trained, and then the test data set can be used to test the trained first, second, and third basic learning models. The first, second, and third basic learning models respectively output corresponding basic full connection queue value adjustment amounts for the input sample data. Then, the basic full connection queue value adjustment amount is used as the full connection queue value adjustment amount in the new sample data, and is added to the original sample data set to form a new sample data set. For example, assuming that for the training sample data set 1, the first, second, and third basic learning models respectively calculate the basic full connection queue value adjustment amounts of 48, 52, and 50, a new sample data set 2 is constructed, which includes the following features: concurrent connection number, CPU usage rate, memory usage rate, disk I / O latency, 48, 52, and 50. Then, the new sample data set is input into the secondary learning model, and the secondary learning model is used for secondary conversion and utilization, to provide more information and perspectives for the secondary learning model, and to output more accurate predicted full connection queue value adjustment amounts.

[0129] Specifically, as an implementation manner, the second sample data set input into the secondary learning model can be preprocessed, and the second sample data set can be divided into a support data set and a query data set according to the task related to the adjustment amount of the full connection queue value. The support data set is used to train the secondary learning model, and the query data set is used to evaluate the model performance of the secondary learning model.

[0130] For task 1, the support data set can be constructed as shown in Table 4 below, and the query data set can be constructed as shown in Table 5 below:

[0131] Table 4. Illustration of support data set for task 1

[0132]

[0133] Table 5. Illustration of query data set for task 1

[0134]

[0135] For task 2, the support data set can be constructed as shown in Table 6 below, and the query data set can be constructed as shown in Table 7 below:

[0136] Table 6. Illustration of support data set for task 2

[0137]

[0138] Table 7. Illustration of query data set for task 2

[0139]

[0140] By randomly sampling a small number of data points from the support data set described above, for example, using only 2 sample data as the support data set for each task, and then training the secondary learning model using the support data set, the secondary learning model learns how to generate the predicted full connection queue value adjustment amount according to the support data set. Then in this process, the model parameters of the secondary learning model are constantly adjusted according to the target data difference between the predicted full connection queue value adjustment amount and the actual full connection queue value adjustment amount, until the target data difference is less than a data difference threshold, or converges.

[0141] As an implementation, a loss function can be constructed according to the predicted full connection queue value adjustment amount and the actual full connection queue value adjustment amount, and then the model parameters of the secondary learning model are updated based on the function value of the loss function by backpropagation until the function value of the loss function converges. Specifically, the query data set can be input into the secondary learning model, the secondary learning model can output the predicted full connection queue value adjustment amount based on the query data set, and then a mean square error loss function or other suitable loss function for regression problems can be constructed, the model parameters of the secondary learning model are updated using the backpropagation algorithm according to the function value of the loss function until the secondary learning model converges to a stable performance.

[0142] The above takes task 1 as an example, assuming that a neural network is selected as a secondary learning model, and the secondary learning model parameters are initialized. In the first iteration, sample 1 and sample 2 can be randomly sampled from the support data set of task 1 as support data, and used to train the secondary learning model, which can be a linear regression model. After training, the query data set (sample 4 and sample 5) is used to calculate the loss, and the model parameters of the secondary learning model are updated according to the loss.

[0143] In a second aspect, the present application provides a device for dynamically adjusting a TCP full connection queue in a container, which is applied to a server device, wherein, as shown in the figure, Figure 4 The device 40 comprises:

[0144] The acquisition module 401 is configured to acquire device state parameters of the server device, wherein the device state parameters comprise CPU usage, memory usage, disk I / O latency, etc.

[0145] The determination module 402 is configured to input the device state parameters into a target full connection queue prediction model trained in advance, and adjust the full connection queue value of the server device according to the result output by the target full connection queue prediction model.

[0146] In some embodiments, the device 40 further comprises a model training module 403 configured to:

[0147] input a preprocessed training sample data set into an initial full connection queue prediction model, wherein the preprocessed training sample data set comprises a plurality of training sample data, each of the training sample data comprises device state parameters of each server device and an actual full connection queue value adjustment amount corresponding to the server device, and the full connection queue prediction model is a neural network model constructed based on a machine learning algorithm;

[0148] acquire a predicted full connection queue value adjustment amount output by the full connection queue prediction model based on the device state parameters in each of the training sample data according to a target correlation, wherein the target correlation is learned based on each of the training sample data, and the target correlation is used to represent the correlation between the device state parameters and the full connection queue value adjustment amount;

[0149] adjust model parameters of the full connection queue prediction model according to a target data difference between the predicted full connection queue value adjustment amount and the actual full connection queue value adjustment amount until the target data difference converges;

[0150] determine the full connection queue prediction model corresponding to the time when the target data difference converges as the target full connection queue prediction model.

[0151] The names of the messages or information exchanged between the plurality of devices in the embodiments of the present application are used only for illustrative purposes, and are not intended to limit the scope of the messages or information.

[0152] In a third aspect, an example embodiment of the present application further provides an electronic device, comprising: at least one processor; and a memory connected with the at least one processor in communication. The memory stores a computer program capable of being executed by the at least one processor, and the computer program, when executed by the at least one processor, is configured to cause the electronic device to perform the method according to the embodiments of the present application.

[0153] An example embodiment of the present application further provides a non-transitory computer readable storage medium storing a computer program, wherein the computer program, when executed by a processor of a computer, is configured to cause the computer to perform the method according to the embodiments of the present application.

[0154] An example embodiment of the present application further provides a computer program product comprising a computer program, wherein the computer program, when executed by a processor of a computer, is configured to cause the computer to perform the method according to the embodiments of the present application.

[0155] Reference Figure 5 A block diagram of the structure of an electronic device 500 that can be a server or a client of the present application will now be described, which is an example of a hardware device that can be applied to various aspects of the present application. The electronic device is intended to represent a variety of forms of digital electronic computing devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent a variety of forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections, and relationships, and their functions, are meant only as examples, and are not meant to limit implementations of the present application described and / or claimed herein.

[0156] As Figure 5 shown, the electronic device 500 includes a computing unit 501 that can perform various appropriate actions and processes in accordance with a computer program stored in a read-only memory (ROM) 502 or a computer program loaded into a random access memory (RAM) 503 from a storage unit 508. In the RAM 503, various programs and data required for the operation of the electronic device 500 can also be stored. The computing unit 501, the ROM 502, and the RAM 503 are connected to each other through a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0157] A plurality of components in the electronic device 500 are connected to the I / O interface 505, including: an input unit 506, an output unit 507, a storage unit 508, and a communication unit 509. The input unit 506 can be any type of device that can input information to the electronic device 500, and can receive inputted digital or character information, and generate key signal inputs related to user settings and / or function controls of the electronic device. The output unit 507 can be any type of device that can present information, and can include, but is not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 508 can include, but is not limited to, a magnetic disk, an optical disk. The communication unit 509 allows the electronic device 500 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and can include, but is not limited to, a modem, a network card, an infrared communication device, a wireless communication transceiver, and / or a chipset, such as a Bluetooth™ device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0158] The computing unit 501 can be various general and / or special purpose processing components having processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 501 performs various methods and processes described above. For example, in some embodiments, the aforementioned method of dynamically adjusting the TCP full-connection queue within a container can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 508. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device 500 via the ROM 502 and / or the communication unit 509. In some embodiments, the computing unit 501 can be configured to perform the aforementioned method of dynamically adjusting the TCP full-connection queue within a container by any other appropriate means, such as by means of firmware.

[0159] Program code for carrying out the methods of the present application can be written in any combination of one or more programming languages. The program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, produces the functions / operations specified in the flowcharts and / or the block diagrams. The program code can be embodied in whole or in part within a machine, executed partially on the machine, partially on a remote machine, or entirely on a remote machine or server.

[0160] In the context of this application, a machine-readable medium can be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine- readable medium can include but is not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium will include one or more lines of electrical connections, portable computer disks, hard disk drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fibers, portable compact disc read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0161] As used in this application, the terms "machine-readable medium" and "computer- readable medium" refer to any computer program product, apparatus and / or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal that can be used to provide machine instructions and / or data to a programmable processor.

[0162] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0163] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0164] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

Claims

1. A method for dynamically adjusting the TCP full connection queue within a container, characterized in that, The method is applied to a server-side device, and the method includes: Obtain the device status parameters of the server device, including: CPU utilization, memory utilization, and disk I / O wait time; The device status parameters are input into a pre-trained target fully connected queue prediction model, and the full connected queue value of the server device is adjusted according to the output of the target fully connected queue prediction model. The target fully connected queue prediction model is pre-trained in the following manner: The preprocessed sample dataset is input into the initial fully connected queue prediction model; wherein, the preprocessed sample dataset includes several training sample data, each of which includes the device status parameters of each server device and the actual fully connected queue value adjustment amount corresponding to the server device; the fully connected queue prediction model is a model built based on machine learning algorithms. Obtain the predicted fully connected queue value adjustment amount, wherein the predicted fully connected queue value adjustment amount is obtained by the fully connected queue prediction model based on the device state parameters in each of the training sample data and output according to the target correlation relationship, wherein the target correlation relationship is the correlation relationship learned based on each of the training sample data, and the target correlation relationship is used to characterize the correlation between the device state parameters and the fully connected queue value adjustment amount; Based on the target data difference between the predicted fully connected queue value adjustment amount and the actual fully connected queue value adjustment amount, the model parameters of the fully connected queue prediction model are adjusted until the target data difference converges. The fully connected queue prediction model corresponding to the convergence of the target data differences is determined as the target fully connected queue prediction model.

2. The method according to claim 1, characterized in that, The fully connected queue prediction model is a comprehensive model obtained by integrating several basic learning models and secondary learning models based on the Stacking ensemble learning algorithm. The basic learning model includes: a first basic learning model, a second basic learning model, and a third basic learning model; the first basic learning model is a model constructed based on the linear regression algorithm, the second basic learning model is a model constructed based on the support vector machine algorithm, and the third basic learning model is a model constructed based on the decision tree algorithm. The secondary learning model is used to synthesize the basic fully connected queue value adjustment amounts output by each of the basic learning models to generate the predicted fully connected queue value adjustment amounts.

3. The method according to claim 2, characterized in that, The sample dataset includes: a training data set; the step of obtaining the adjustment amount of the predicted fully connected queue value includes: The training data set is divided into different training data subsets, and each training data subset is input into the first basic learning model, the second basic learning model, and the third basic learning model respectively. Each basic learning model outputs its own basic fully connected queue value adjustment based on the input training data subset. New sample data is constructed based on the adjustment amount of each basic fully connected queue value and the device state parameters corresponding to each basic fully connected queue value adjustment amount. The new sample data is then input into the secondary learning model, which outputs the predicted adjustment amount of the fully connected queue value based on the input sample data.

4. The method according to claim 2 or 3, characterized in that, The step of adjusting the model parameters of the fully connected queue prediction model based on the target data difference between the predicted fully connected queue value adjustment amount and the actual fully connected queue value adjustment amount, until the target data difference converges, includes: A loss function is constructed based on the predicted adjustment amount of the fully connected queue value and the actual adjustment amount of the fully connected queue value. Based on the value of the loss function, the model parameters of the secondary learning model are updated through backpropagation until the value of the loss function converges.

5. The method according to claim 2, characterized in that, The first basic learning model was trained in the following way: The first sample dataset is input into the initial first basic learning model, and the first basic learning model outputs the adjustment amount of the first basic fully connected queue value based on the following formula: First basic fully connected queue value adjustment amount = w 0+ w 1×CPU utilization+ w 2×memory utilization+ w 3× disk I / O wait time; in, w 0 is the intercept. w 1 represents the weight of CPU utilization. w 2 represents the weight of memory usage. w 3 represents the weight of disk I / O latency; Calculate the mean square error between the adjustment amount of the first basic fully connected queue value and the actual adjustment amount of the fully connected queue value corresponding to the adjustment amount of the first basic fully connected queue value, and adjust each weight according to the mean square error until the mean square error converges.

6. The method according to claim 2, characterized in that, The second basic learning model is trained in the following way: The second sample dataset is input into the initial second basic learning model. The second basic learning model then performs calculations on the input sample data based on the following Support Vector Machine (SVM) objective function until the optimal model is determined. Optimal b : in, i =1,2,3,...,N, where N is the number of samples in the second sample dataset. Let i be the feature vector of the i-th sample data in the second sample dataset. The label corresponding to the actual fully connected queue value adjustment for the i-th sample data. Let be the normal vector of the hyperplane of the Support Vector Machine (SVM). b Let be the intercept of the hyperplane of the Support Vector Machine (SVM).

7. The method according to claim 2, characterized in that, The third basic learning model is pre-constructed in the following manner: Based on the sample data in the third sample dataset, the optimal split point corresponding to different device state parameters and the sample data subset corresponding to each optimal split point are determined according to different device state parameters. The third basic learning model is constructed by using each of the aforementioned sample data subsets as leaf nodes in the decision tree. The third basic learning model recursively calculates the average adjustment amount of the fully connected queue value corresponding to each leaf node based on each leaf node, and determines the average adjustment amount of the fully connected queue value corresponding to each leaf node as the third basic fully connected queue value adjustment amount of the leaf node.

8. A device for dynamically adjusting the TCP full connection queue within a container, characterized in that, The device is used in a server-side device, and the device includes: The acquisition module is used to acquire the device status parameters of the server device, including: CPU utilization, memory utilization, and disk I / O wait time. The determination module is used to input the device status parameters into a pre-trained target fully connected queue prediction model, and adjust the full connection queue value of the server device according to the output of the target fully connected queue prediction model; The target fully connected queue prediction model is pre-trained in the following manner: The preprocessed training sample dataset is input into the initial fully connected queue prediction model; wherein, the preprocessed training sample dataset includes several training sample data, each training sample data includes the device status parameters of each server device and the actual fully connected queue value adjustment amount corresponding to the server device, and the fully connected queue prediction model is a neural network model built based on machine learning algorithms; Obtain the predicted fully connected queue value adjustment amount, wherein the predicted fully connected queue value adjustment amount is obtained by the fully connected queue prediction model based on the device state parameters in each of the training sample data and output according to the target correlation relationship, wherein the target correlation relationship is the correlation relationship learned based on each of the training sample data, and the target correlation relationship is used to characterize the correlation between the device state parameters and the fully connected queue value adjustment amount; Based on the target data difference between the predicted fully connected queue value adjustment amount and the actual fully connected queue value adjustment amount, the model parameters of the fully connected queue prediction model are adjusted until the target data difference converges. The fully connected queue prediction model corresponding to the convergence of the target data differences is determined as the target fully connected queue prediction model.

9. An electronic device, characterized in that, The electronic device includes: Processor; and Stored program memory, The program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7.

Images