Edge cloud computing load balancing method

Abstract

The invention relates to the technical field of edge cloud computing, discloses an edge cloud computing load balancing method, and aims to solve the load balancing problem caused by wide node distribution, limited resources and dynamic change in an edge computing environment. A traditional load balancing algorithm is based on a static data allocation request, cannot accurately predict future load change, and easily causes node overload or resource idling. According to the method, a neural network algorithm, especially a long-short term memory (LSTM) network model, is adopted to construct and train a load prediction model, and the future load level of the edge computing node is predicted according to the historical use condition of the edge computing node. A monitoring and data acquisition system is deployed, factor data influencing node loads are collected in real time and input into a prediction model, and a resource scheduling strategy is dynamically formulated. According to the method, the dynamic and efficient utilization of the edge computing nodes is realized, the response speed and the operation efficiency of the system are improved, and the user experience is optimized.

Description

Technical Field

[0001] The present invention relates to the field of edge cloud computing technology, and in particular to an edge cloud computing load balancing method. Background Art

[0002] With the rapid development of the Internet of Things, big data and artificial intelligence technologies, edge computing, as an emerging computing model, has gradually been widely used in various industries because it can reduce data transmission delays, improve data processing efficiency and protect user privacy. Edge computing deploys data processing capabilities at the edge of the network, that is, near the device or terminal, to achieve instant data processing and feedback, greatly improving the system's response speed and operating efficiency.

[0003] In edge computing environments, how to achieve efficient load balancing has become an urgent problem to be solved due to the wide distribution of nodes, limited resources and dynamic changes. Traditional load balancing algorithms mostly allocate requests based on current or historical static data, and cannot accurately predict the changing trend of future loads. Therefore, when facing burst traffic or periodic load peaks, it is easy to cause some nodes to be overloaded while other node resources are idle, and it is impossible to effectively achieve dynamic optimization allocation of resources. The edge computing environment is complex and changeable, and factors such as node performance, network conditions, and user behavior may affect the load situation. Traditional load balancing methods are difficult to adapt to these dynamic changes in real time, and cannot achieve accurate load prediction and resource scheduling, which affects the overall performance of the system and user experience. To this end, we propose a load balancing method for edge cloud computing. Summary of the invention

[0004] The purpose of the present invention is to provide an edge cloud computing load balancing method to solve the problems raised in the above background technology.

[0005] The working principle of this method is based on neural network algorithms, especially the long short-term memory network (LSTM) model, to achieve load balancing in edge cloud computing environments. It first collects historical load data of edge computing nodes and uses feature selection algorithms to filter out features that are useful for the prediction model to build a training data set. Then, the LSTM model is used to iteratively train the training data set to obtain a load prediction model that can accurately predict the future load level of the node. In practical applications, this method deploys a monitoring and data acquisition system to collect data on factors that affect node load in real time, and inputs this data into the trained load prediction model, thereby dynamically predicting the future load level of the node. Finally, based on the prediction results, a resource scheduling strategy is dynamically formulated and implemented in the load balancer to achieve dynamic and efficient utilization of edge computing nodes.

[0006] To achieve the above object, the present invention provides the following technical solution: an edge cloud computing load balancing method, the method comprising the steps of:

[0007] Step 1: Use a neural network algorithm to build and train a load prediction model. The load prediction model is used to predict the load level of the node at a certain point in the future based on the usage of the edge computing node. The specific steps of building and training the load prediction model include:

[0008] Step 1.1, collect historical load data of edge computing nodes, where the historical load data is data of all factors that affect the node load at different load levels;

[0009] Step 1.2: Clean the collected data to handle missing values ​​and outliers;

[0010] Step 1.3, standardize the cleaned feature data;

[0011] Step 1.4: Use feature selection algorithm to select features that are useful for the prediction model and form a feature data set;

[0012] Step 1.5: Mark the load level of each data and form a training data set together with the corresponding feature data. The training data set is further divided into a training set and a validation set.

[0013] Step 1.6: Select the long short-term memory network LSTM model to build a sampling prediction model and initialize the parameters of the model;

[0014] Step 1.7: Use the training set data to iteratively train the LSTM model. During the training process, regularly use the validation set to evaluate the model performance until the model performance meets the requirements.

[0015] Step 1.8: After training is completed, save the weights and biases of the LSTM model and deploy the model for subsequent sampling position prediction;

[0016] Step 2: Deploy a monitoring and data acquisition system to collect data on all factors that affect the edge computing node load in real time;

[0017] Step 3: Clean the data collected in step 2 to deal with missing values ​​and outliers;

[0018] Step 4: Input the cleaned data in step 3 into the load prediction model trained in step 1 to obtain the load level of the edge computing node at a certain point in the future;

[0019] Step 5: According to the load level obtained in step 4, dynamically formulate a resource scheduling strategy, wherein the scheduling strategy includes the number of edge computing nodes required to bear the load;

[0020] Step 6: Implement the scheduling strategy in the load balancer to achieve dynamic and efficient utilization of edge computing nodes.

[0021] Preferably, the historical load data collected for training the load prediction model include resource usage indicators, network status indicators, and user behavior data of nodes at different load levels; the resource usage indicators include CPU usage, memory usage, disk I / O, storage space usage, number of processes and number of threads; the network status indicators include network bandwidth usage, latency, packet loss rate, number of connections and network traffic; the user behavior data includes user request volume, user access pattern, user session duration, user geographic location distribution and user device type.

[0022] Preferably, in step 1.2, for missing values, a linear interpolation algorithm is used to fill them; for outliers, a threshold judgment method is used to regard data points that exceed a preset threshold range as outliers and replace them with the mean of their neighboring points.

[0023] Preferably, in step 1.3, the Z-score standardization method is used to subtract the mean of each eigenvalue and divide it by its standard deviation so that the processed data conforms to the standard normal distribution. The formula for standardization is: Among them, X is the original data, μ is the mean of the original data, σ is the standard deviation of the original data, and Z is the standardized data.

[0024] Preferably, in step 1.4, a filtering method is used for feature selection, which evaluates the importance of the feature by calculating the correlation or statistical index between each feature and the target variable; the linear correlation between each feature and the target variable is measured using the Pearson correlation coefficient, and the calculation formula of the Pearson correlation coefficient is:

[0025]

[0026] Among them, r represents the correlation coefficient between the feature and the target variable, n represents the number of samples, x represents the feature value, and y represents the target variable value;

[0027] According to the calculated correlation coefficient, the features whose absolute values ​​are greater than the set threshold are selected as useful features to form a feature data set for the subsequent training of the load prediction model.

[0028] Preferably, the specific steps of initializing the LSTM model include:

[0029] i) Initialize the structural parameters of the LSTM model, including the feature dimension of the input layer, the number of hidden layers and the number of hidden units in each layer, and the number of nodes in the output layer, where the feature dimension of the input layer matches the number of features extracted that affect the load of the edge computing nodes, and the number of nodes in the output layer corresponds to the dimension of the predicted load level;

[0030] ii) Initialize the model weights and biases using small random numbers;

[0031] iii) Setting the activation function of the LSTM model, including setting the Sigmoid function for the gating mechanism and setting the Tanh function for the calculation of candidate memory cells and hidden states;

[0032] iv) Configure the optimization algorithm and set the learning rate parameters;

[0033] Use the training set data to iteratively train the LSTM model. The specific method is as follows:

[0034] In each iteration, a batch of input data is fed into the LSTM model, and the memory cells and hidden states are updated by calculating the values ​​of the input gate, forget gate, and output gate, and finally the predicted output result is obtained; the loss function value is calculated based on the predicted output and the true label; the gradient of the loss function with respect to the model parameters is calculated through the back propagation algorithm; the model weights and biases are updated according to the gradient using the optimization algorithm; during the training process, the validation set is regularly used to evaluate the model performance, and the learning rate is adjusted or the training is stopped early as needed to prevent overfitting.

[0035] Preferably, in step 5, a threshold-based scheduling algorithm is adopted, and multiple load thresholds are preset, each threshold corresponds to a different resource scheduling strategy; the load level predicted in step 5 is compared with the preset load threshold to determine the threshold range to which the current load level belongs; according to the threshold range, the corresponding resource scheduling strategy is selected;

[0036] The resource scheduling strategy includes the number of edge computing nodes required to bear the load, and the required number of nodes is calculated using a linear mapping function. The calculation formula is:

[0037]

[0038] Among them, round() represents a rounding function, the predicted load level represents the load level obtained in step 5, the minimum threshold and the maximum threshold represent the preset load threshold ranges, and the maximum number of nodes and the minimum number of nodes represent the maximum and minimum number of edge computing nodes available, respectively.

[0039] Preferably, in step 2, for each key factor affecting the load of the edge computing node, a corresponding data acquisition module is configured, and the data of the key factor is captured and recorded in real time using a preset sampling frequency and time interval; and a data transmission interface between the data acquisition module and the central data storage system is formulated so that the collected data can be transmitted to the central data storage system in real time and accurately for subsequent processing and analysis;

[0040] Preferably, the data acquisition module adopts a lightweight data acquisition protocol.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] Through the neural network algorithm, the present invention can accurately predict the future load level of edge computing nodes, thereby realizing dynamic resource scheduling based on prediction. This method ensures that resources are allocated when needed, avoids idleness and waste of resources, and significantly improves resource utilization efficiency. As resources are more effectively utilized and the system response speed is accelerated, users will experience less delay and higher service quality when using edge computing services.

[0043] The edge computing environment is complex and changeable. Through real-time monitoring and data collection, the present invention can dynamically adapt to changes in factors such as node performance, network conditions, and user behavior, so that the load balancing strategy can more accurately respond to actual load conditions, thereby improving the stability and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A diagram showing the processing steps of the method of the present invention;

[0045] Figure 2 Flowchart for initializing, training, and validating optimization for load-like prediction models;

[0046] Figure 3 Diagram of the steps for dynamically formulating resource scheduling strategies. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] See also Figure 1-3 The present invention provides a technical solution: an edge cloud computing load balancing method, the steps of the method comprising:

[0049] Step 1: Build and train a load prediction model. The load prediction model is used to predict the load level of the node at a certain point in the future based on the usage of the edge computing node. The specific steps of building and training the load prediction model include:

[0050] Step 1.1, collect historical load data: for edge computing nodes, collect their historical load data at different load levels. These data include all factors that affect the node load, including the resource usage indicators, network status indicators, and user behavior data of the nodes at different load levels; the resource usage indicators include CPU usage, memory usage, disk I / O, storage space usage, number of processes and number of threads; the network status indicators include network bandwidth usage, latency, packet loss rate, number of connections and network traffic; the user behavior data includes user request volume, user access pattern, user session duration, user geographic location distribution and user device type.

[0051] Step 1.2, data cleaning: Clean the collected historical load data to handle missing values ​​and outliers, including filling missing values, smoothing outliers, etc., to ensure the integrity and consistency of the data.

[0052] For outliers, the threshold judgment method is used to regard data points that exceed the preset threshold range as outliers and replace them with the mean of their neighboring points.

[0053] Step 1.3, data standardization: Standardize the cleaned feature data to eliminate the dimensional differences between different features and improve the training effect of the model. Use the Z-score standardization method to subtract the mean of each feature value and divide it by its standard deviation so that the processed data conforms to the standard normal distribution. The formula for standardization is: Among them, X is the original data, μ is the mean of the original data, σ is the standard deviation of the original data, and Z is the standardized data.

[0054] Step 1.4, feature selection: Use feature selection algorithm to screen features that are useful for the prediction model and form a feature data set, aiming to reduce data dimension and improve model training efficiency and prediction accuracy.

[0055] Step 1.5, build training data set: mark the load level of each data, and together with the corresponding feature data, form a training data set. The training data set is further divided into a training set and a validation set to evaluate the performance of the model during the training process.

[0056] Step 1.6, Select and initialize the LSTM model: Select the long short-term memory network (LSTM) model as the load forecasting model and initialize the model parameters. The LSTM model is suitable for load forecasting because it is good at processing time series data.

[0057] Step 1.7, model training: Use the training set data to iteratively train the LSTM model. During the training process, the validation set is regularly used to evaluate the model performance until the model performance meets the requirements, aiming to ensure that the model has good generalization ability and prediction accuracy.

[0058] Step 1.8, Model deployment: After training is completed, save the weights and biases of the LSTM model and deploy the model. The deployed model can be used to predict the load level of the edge computing node in the future.

[0059] Step 2: Deploy monitoring and data acquisition system

[0060] Collect data on all factors that affect the load of edge computing nodes in real time, and provide real-time data support for subsequent load prediction and resource scheduling.

[0061] Step 3: Data cleaning

[0062] The real-time data collected in step 2 is cleaned and missing values ​​and outliers are processed to ensure the accuracy and reliability of the data, which helps to improve the accuracy of subsequent load forecasting.

[0063] Step 4: Load Forecasting

[0064] Input the real-time data cleaned in step 3 into the load prediction model trained in step 1 to obtain the load level of the edge computing node at a certain point in the future. Through accurate load prediction, support is provided for subsequent resource scheduling.

[0065] Step 5: Develop a resource scheduling strategy

[0066] According to the load level obtained in step 4, a resource scheduling strategy is dynamically formulated. The scheduling strategy includes the number of edge computing nodes required to bear the load and the specific allocation plan of the nodes, aiming to ensure that while meeting the load requirements, the maximum utilization of resources and the optimal operation of the system are achieved.

[0067] Step 6: Implement scheduling strategy

[0068] Implement the scheduling strategy formulated in step 5 in the load balancer to achieve dynamic and efficient utilization of edge computing nodes. The last step is the final link of resource scheduling. By implementing the scheduling strategy, real-time, dynamic, and efficient resource allocation and management of edge computing nodes can be achieved.

[0069] The present invention will be further described below in conjunction with Examples 1 to 3:

[0070] Embodiment 1:

[0071] In step 1.4, the filtering method is used for feature selection. The filtering method is a feature selection technique that is independent of any subsequent machine learning algorithm. In the implementation of the present invention, the filtering method is used to preliminarily screen those features that may be correlated with the target variable (i.e., the load level). This method is based on the statistical relationship or correlation between the features and the target variable, and can quickly remove those features that are irrelevant to the target variable or have a weak correlation, thereby reducing the dimension of the feature space and improving the efficiency of subsequent model training.

[0072] In order to determine the strength of the correlation between each feature and the target variable, the present invention uses the Pearson Correlation Coefficient as a measurement indicator. The Pearson correlation coefficient is a widely used statistic that can quantify the linear relationship between two variables. In the context of the present invention, a feature (such as CPU usage, memory usage, etc.) is regarded as a variable, and the target variable is the load level of the edge computing node. By calculating the Pearson correlation coefficient between each feature and the target variable, it is possible to evaluate which features play an important role in predicting the load level. The calculation formula of the Pearson correlation coefficient is:

[0073]

[0074] Among them, r represents the correlation coefficient between the feature and the target variable, n represents the number of samples, x represents the feature value, and y represents the target variable value.

[0075] This formula calculates the covariance between the feature and the target variable and divides it by the standard deviation of the two variables, resulting in a value between -1 and 1 that represents the strength and direction of the linear correlation between the feature and the target variable. Values ​​close to 1 or -1 indicate strong correlation, while values ​​close to 0 indicate weak or no correlation. In the implementation of the present invention, this formula is used to calculate the Pearson correlation coefficient between each feature and the target variable, and features useful for the prediction model are selected based on these coefficients.

[0076] After calculating the Pearson correlation coefficient between each feature and the target variable, the next step is to select useful features based on these correlation coefficients. Specifically, a correlation coefficient threshold is set, and then those features whose absolute values ​​are greater than this threshold are selected. The purpose of this step is to screen out features with strong linear correlation with the target variable, which are considered to contribute significantly to the accuracy of the prediction model. By removing those features with weak correlation, the complexity of the feature space can be reduced, and the training efficiency and generalization ability of the model can be improved. Ultimately, the selected useful features will constitute a feature subset for subsequent load prediction model training.

[0077] After determining the useful feature subsets, the next step is to use these features to construct a feature dataset. This feature dataset will be used as input data to train the load forecasting model. When constructing the feature dataset, it is necessary to ensure the integrity and consistency of the data in order to provide a high-quality data foundation for model training. The feature dataset should contain the values ​​of all selected features, and these values ​​should correspond to the values ​​of the target variable. By constructing the feature dataset using the selected feature subset, an optimized, low-dimensional feature space can be provided for subsequent model training, thereby improving the accuracy and efficiency of the load forecasting model.

[0078] Embodiment 2:

[0079] As a variant of recurrent neural network (RNN), the LSTM model is particularly good at processing time series data. In load forecasting, historical load data usually exhibits significant time series characteristics. LSTM can capture long-term dependencies and periodic patterns in these time series, thereby improving the accuracy of forecasts. The load situation in edge computing environments is often affected by a variety of dynamic factors, such as node performance, network conditions, user behavior, etc. The LSTM model can update its internal state in real time and adjust the forecast results according to the latest input data to adapt to these dynamic changes and improve the real-time and accuracy of the forecast.

[0080] In load forecasting, in addition to historical load data, there may also be a variety of external factors involved, such as weather, holidays, economic indicators, etc. The LSTM model can integrate these multi-source heterogeneous data into the model by introducing an additional input layer, and work together on the forecast results, thereby improving the robustness and comprehensiveness of the forecast.

[0081] The load prediction model is used to predict the load level of the node at a certain point in the future based on the usage of the edge computing node. For example, all factor data that affect the load of the edge computing node are collected in real time, and these real-time data are input into the deployed load prediction model to obtain the load level prediction result of the node at a certain point in the future (such as 5 minutes, 10 minutes or 30 minutes later).

[0082] The specific steps to initialize the LSTM model include:

[0083] i) Initialize the structural parameters of the LSTM model. The resulting parameters include the feature dimension of the input layer, which matches the number of features extracted that affect the load of the edge computing node. The number of features is determined based on the collected historical load data, which includes resource usage indicators (such as CPU usage, memory usage, etc.), network status indicators (such as network bandwidth usage, latency, etc.), and user behavior data (such as user request volume, user access pattern, etc.). The number of hidden layers and the number of hidden units in each layer determine the complexity and learning ability of the model; the number of nodes in the output layer corresponds to the dimension of the predicted load level.

[0084] ii) Initialize the model weights and biases. In order to avoid the problem of gradient vanishing or exploding in the early stage of training, small random numbers are used to initialize the weights and biases.

[0085] iii) Set the activation function of the LSTM model. In order to achieve nonlinear transformation of the gating mechanism, select the Sigmoid function; and in order to calculate the candidate memory units and hidden states, select the Tanh function.

[0086] iv) Configure the optimization algorithm and set an appropriate learning rate parameter to control the update step size of the model during training.

[0087] In each iteration, the following steps are performed:

[0088] i) A batch of input data is fed into the LSTM model. The data contains feature information extracted from historical load data, such as resource usage indicators, network status indicators, and user behavior data.

[0089] ii) By calculating the values ​​of the input gate, forget gate, and output gate, the memory unit and hidden state are updated to finally obtain the predicted output result.

[0090] iii) Calculate the loss function value based on the predicted output and the true label to evaluate the performance of the model on the current batch of data.

[0091] iv) The gradient of the loss function with respect to the model parameters is calculated through the back-propagation algorithm. These gradients indicate how to adjust the model parameters to reduce the loss.

[0092] v) Use an optimization algorithm to update the model weights and biases based on the gradients, thereby improving the model’s predictive power.

[0093] During the training process, the model performance is evaluated regularly using the validation set (which also contains comprehensive historical load data of resource usage indicators, network status indicators, and user behavior data). If the model performance on the validation set is no longer improving or there are signs of overfitting, the learning rate is adjusted or training is stopped early as needed.

[0094] Through the implementation of the above steps, the present invention can effectively initialize, train and optimize the LSTM model to perform load pre-calculation of edge cloud computing.

[0095] Embodiment 3:

[0096] The specific method of dynamically formulating resource scheduling strategies is as follows:

[0097] ①Use threshold-based scheduling algorithm:

[0098] In order to achieve dynamic resource scheduling, the present invention presets multiple load thresholds, which are set based on experience or historical data, and divides the load level into different ranges. Each range corresponds to a specific resource scheduling strategy to ensure that edge computing nodes can be effectively allocated and used under different load levels.

[0099] ② Compare the predicted load level with the preset load threshold:

[0100] The present invention predicts the future load level through the LSTM model and compares the predicted value with the preset load threshold. The purpose of the comparison is to determine within which preset threshold the current predicted load level falls. This step is the key to selecting an appropriate resource scheduling strategy because it determines how resources should be allocated under a specific load level.

[0101] ③Select the corresponding resource scheduling strategy:

[0102] After determining the threshold range to which the current load level belongs, we further select resource scheduling strategies corresponding to this range. These strategies are based on historical data and experience and are designed to achieve optimal resource allocation under different load levels. Strategies may include increasing or decreasing the number of edge computing nodes, adjusting the processing capacity of nodes, etc., to ensure load balancing and improve overall service quality.

[0103] ④Calculate the number of edge computing nodes required:

[0104] The resource scheduling strategy includes the number of edge computing nodes required to bear the load, and the required number of nodes is calculated using a linear mapping function. The calculation formula is:

[0105]

[0106] Among them, round() represents a rounding function, the predicted load level represents the load level obtained in step 5, the minimum threshold and the maximum threshold represent the preset load threshold ranges, and the maximum number of nodes and the minimum number of nodes represent the maximum and minimum number of edge computing nodes available, respectively.

[0107] Assume that the preset minimum load threshold is 100, the maximum load threshold is 500, the minimum number of available edge computing nodes is 5, and the maximum number of nodes is 20. When the load level predicted by the LSTM model is 300, the required number of nodes is calculated according to the linear mapping function as follows:

[0108]

[0109] Therefore, when the predicted load level is 300, resource allocation needs to be dynamically adjusted to ensure that 13 edge computing nodes are available to achieve load balancing.

[0110] ⑤Dynamically adjust the resource allocation of edge computing nodes:

[0111] Based on the calculated number of nodes, the resource allocation of edge computing nodes is dynamically adjusted, which means that when the load is high, more nodes will be added to process requests, and when the load is low, nodes will be reduced to save resources.

[0112] In step 2, in order to achieve accurate prediction of edge computing node load and dynamic resource scheduling, the present invention configures a corresponding data acquisition module for each key factor affecting the edge computing node load.

[0113] For each key factor that affects the edge computing node load, such as network bandwidth, processor usage, memory usage, storage I / O, etc., the present invention is equipped with a dedicated data acquisition module, which uses a preset sampling frequency and time interval to capture and record the data of the key factors in real time. The setting of the sampling frequency and time interval is based on the analysis and prediction requirements of historical data to ensure that the collected data can accurately reflect the real-time status of the key factors.

[0114] In order to ensure that the collected data can be transmitted to the central data storage system in real time and accurately for subsequent processing and analysis, the present invention has developed a data transmission interface between the data acquisition module and the central data storage system. The interface is based on an efficient and stable data transmission protocol to ensure reliable data transmission. At the same time, the interface also takes into account the security and integrity of the data, and adopts encryption and verification mechanisms to prevent data from being tampered with or lost during transmission.

[0115] Select or design a lightweight data acquisition protocol, which should have a concise data format and an efficient transmission mechanism to reduce the network load and computing overhead during the data acquisition process. You can consider using existing lightweight communication protocols, such as MQTT (Message Queuing Telemetry Transport) or CoAP (Constrained Application Protocol), or customize a more streamlined protocol according to actual needs. For the selected lightweight protocol, define the data format and transmission rules between the data acquisition module and the central data storage system. The data format should be compact and easy to parse to reduce bandwidth consumption during data transmission. The transmission rules should specify the data transmission frequency, time interval, retransmission mechanism, etc. to ensure the reliability and integrity of the data. Taking the MQTT protocol as an example, the present invention can implement the MQTT client in the data acquisition module, publish the collected key factor data as an MQTT message to the specified topic (Topic), and the central data storage system acts as an MQTT broker (Broker), subscribes to these topics and receives messages. The MQTT protocol is very suitable for lightweight data acquisition and transmission in edge computing environments due to its concise data format and efficient transmission mechanism.

[0116] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0117] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for edge cloud computing load balancing, characterized in that: The steps of the method include: Step 1: Use a neural network algorithm to build and train a load prediction model. The load prediction model is used to predict the load level of the node at a certain point in the future based on the usage of the edge computing node. The specific steps of building and training the load prediction model include: Step 1.1, collect historical load data of edge computing nodes, where the historical load data is data of all factors that affect the node load at different load levels; Step 1.2: Clean the collected data to handle missing values ​​and outliers; Step 1.3, standardize the cleaned feature data; Step 1.4: Use feature selection algorithm to select features that are useful for the prediction model and form a feature data set; Step 1.5: Mark the load level of each data and form a training data set together with the corresponding feature data. The training data set is further divided into a training set and a validation set. Step 1.6: Select the long short-term memory network LSTM model to build a sampling prediction model and initialize the parameters of the model; Step 1.7: Use the training set data to iteratively train the LSTM model. During the training process, regularly use the validation set to evaluate the model performance until the model performance meets the requirements. Step 1.8: After training is completed, save the weights and biases of the LSTM model and deploy the model for subsequent sampling position prediction; Step 2: Deploy a monitoring and data acquisition system to collect data on all factors that affect the edge computing node load in real time; Step 3: Clean the data collected in step 2 to deal with missing values ​​and outliers; Step 4: Input the cleaned data in step 3 into the load prediction model trained in step 1 to obtain the load level of the edge computing node at a certain point in the future; Step 5: According to the load level obtained in step 4, dynamically formulate a resource scheduling strategy, wherein the scheduling strategy includes the number of edge computing nodes required to bear the load; Step 6: Implement the scheduling strategy in the load balancer to achieve dynamic and efficient utilization of edge computing nodes.

2. The edge cloud computing load balancing method according to claim 1, characterized in that: The historical load data collected for training the load prediction model includes resource usage indicators, network status indicators, and user behavior data of nodes at different load levels; the resource usage indicators include CPU usage, memory usage, disk I / O, storage space usage, number of processes and number of threads; the network status indicators include network bandwidth usage, latency, packet loss rate, number of connections and network traffic; the user behavior data includes user request volume, user access pattern, user session duration, user geographic location distribution and user device type.

3. The edge cloud computing load balancing method according to claim 2, characterized in that: In step 1.2, for missing values, linear interpolation algorithm is used to fill them; for outliers, threshold judgment method is used to regard data points beyond the preset threshold range as outliers and replace them with the mean of their neighboring points.

4. The edge cloud computing load balancing method according to claim 3, characterized in that: In step 1.3, the Z-score standardization method is used to subtract the mean of each feature value and divide it by its standard deviation so that the processed data conforms to the standard normal distribution. The formula for standardization is: Among them, X is the original data, μ is the mean of the original data, σ is the standard deviation of the original data, and Z is the standardized data.

5. The edge cloud computing load balancing method according to claim 4, characterized in that: In step 1.4, the filtering method is used for feature selection. This method evaluates the importance of features by calculating the correlation or statistical index between each feature and the target variable; the Pearson correlation coefficient is used to measure the linear correlation between each feature and the target variable. The calculation formula of the Pearson correlation coefficient is: Among them, r represents the correlation coefficient between the feature and the target variable, n represents the number of samples, x represents the feature value, and y represents the target variable value; According to the calculated correlation coefficient, the features whose absolute values ​​are greater than the set threshold are selected as useful features to form a feature data set for the subsequent training of the load prediction model.

6. The edge cloud computing load balancing method according to claim 5, characterized in that: The specific steps to initialize the LSTM model include: i) Initialize the structural parameters of the LSTM model, including the feature dimension of the input layer, the number of hidden layers and the number of hidden units in each layer, and the number of nodes in the output layer, where the feature dimension of the input layer matches the number of features extracted that affect the load of edge computing nodes, and the number of nodes in the output layer corresponds to the dimension of the predicted load level; ii) Initialize the model weights and biases using small random numbers; iii) Setting the activation function of the LSTM model, including setting the Sigmoid function for the gating mechanism and setting the Tanh function for the calculation of candidate memory cells and hidden states; iv) Configure the optimization algorithm and set the learning rate parameters; Use the training set data to iteratively train the LSTM model. The specific method is as follows: In each iteration, a batch of input data is fed into the LSTM model, and the memory cells and hidden states are updated by calculating the values ​​of the input gate, forget gate, and output gate, and finally the predicted output result is obtained; the loss function value is calculated based on the predicted output and the true label; the gradient of the loss function with respect to the model parameters is calculated through the back propagation algorithm; the model weights and biases are updated according to the gradient using the optimization algorithm; during the training process, the validation set is regularly used to evaluate the model performance, and the learning rate is adjusted or the training is stopped early as needed to prevent overfitting.

7. The edge cloud computing load balancing method according to claim 1, characterized in that: In step 5, a threshold-based scheduling algorithm is adopted, and multiple load thresholds are preset, each threshold corresponds to a different resource scheduling strategy; the load level predicted in step 5 is compared with the preset load threshold to determine the threshold range to which the current load level belongs; according to the threshold range, the corresponding resource scheduling strategy is selected; The resource scheduling strategy includes the number of edge computing nodes required to bear the load, and the required number of nodes is calculated using a linear mapping function. The calculation formula is: Among them, round() represents a rounding function, the predicted load level represents the load level obtained in step 5, the minimum threshold and the maximum threshold represent the preset load threshold ranges, and the maximum number of nodes and the minimum number of nodes represent the maximum and minimum number of edge computing nodes available, respectively.

8. The edge cloud computing load balancing method according to claim 1, characterized in that: In step 2, for each key factor affecting the edge computing node load, a corresponding data acquisition module is configured, and the data of the key factor is captured and recorded in real time using a preset sampling frequency and time interval; A data transmission interface between the data acquisition module and the central data storage system is also developed so that the collected data can be transmitted to the central data storage system in real time and accurately for subsequent processing and analysis.

9. The edge cloud computing load balancing method according to claim 8, characterized in that: The data acquisition module adopts a lightweight data acquisition protocol.

Images