Method for optimizing application elastic capacity expansion and contraction strategy in multi-cloud environment

Abstract

The invention relates to the technical field of cloud computing resource management, and discloses an application elastic capacity expansion and contraction strategy optimization method in a multi-cloud environment, and the method comprises the steps: obtaining a multi-cloud environment data set composed of public cloud platform monitoring data, edge node load data and a historical capacity expansion and contraction strategy library; abnormal value cleaning, template matching and multi-resolution filtering are carried out on public cloud data through a multi-cloud load feature extraction module, and load features such as a CPU utilization rate fluctuation curve are extracted; performing time alignment and sliding window statistics on the edge node data by using a multi-dimensional resource fusion algorithm, and generating resource associated parameters such as computing resource redundancy; and performing space-time correlation analysis on the historical strategy library to generate an evolution graph. And inputting the load features, the resource correlation parameters and the evolution atlas into an elastic capacity expansion and contraction optimization model composed of a feature enhancement module and a collaborative reasoning module, and generating a strategy feature vector. According to the method, intelligent optimization of application of an elastic capacity expansion and contraction strategy in a multi-cloud environment is realized.

Description

Technical Field

[0001] This invention relates to the field of cloud computing resource management technology, specifically to an optimization method for application elastic scaling strategies in a multi-cloud environment. Background Technology

[0002] With the rapid development of cloud computing technology, multi-cloud environments are gradually becoming an important choice for enterprise application deployment due to their significant advantages in resource scheduling flexibility, cost control, and business continuity. In multi-cloud environments, application workloads often exhibit complex and variable characteristics, which places extremely high demands on the elastic scaling capabilities of applications. However, current elastic scaling strategies face many unresolved issues in practical applications.

[0003] From a data processing perspective, traditional methods exhibit significant shortcomings when dealing with massive amounts of data in multi-cloud environments. Their methods for processing public cloud platform monitoring data are relatively limited, making it difficult to comprehensively and accurately extract application load characteristics. For example, they cannot accurately capture key information such as CPU utilization fluctuation curves, memory allocation patterns, and request latency distribution characteristics, resulting in a lack of reliable data support for subsequent scaling decisions. Furthermore, their processing of edge node load data is particularly flawed. Edge node load data includes multi-dimensional information such as computing resource utilization, storage I / O throughput, and network traffic peaks. Traditional methods often fail to effectively integrate and analyze this data, making it difficult to generate accurate resource correlation parameters, such as computing resource redundancy, storage bottleneck coefficients, and network latency thresholds, thus affecting the accurate assessment of overall resource status.

[0004] Traditional methods also have significant limitations in strategy optimization. They typically fail to fully utilize information from historical scaling strategy libraries, lack spatiotemporal correlation analysis of historical strategies, and cannot generate scaling evolution maps. This results in strategy optimization lacking guidance from historical experience, making it difficult to adapt to complex and ever-changing business scenarios. Furthermore, existing elastic scaling models often fail to achieve effective feature enhancement and collaborative reasoning when processing multi-dimensional data, leading to inaccurate generated strategy feature vectors. This, in turn, affects the construction of multi-level scaling strategy networks, making it difficult for the output scaling decisions and confidence levels to meet the needs of actual business operations.

[0005] In practical applications, problems such as incomplete handling of outliers, inaccurate load time-series segmentation, and untimely dynamic adjustment of resource baselines still exist. These problems are intertwined, leading to low resource utilization, high response latency, and poor cost control in traditional elastic scaling strategies in multi-cloud environments, severely restricting the full realization of the advantages of multi-cloud environments. Therefore, there is an urgent need for an elastic scaling strategy optimization method that can comprehensively process multi-cloud environment data, accurately extract load characteristics, effectively integrate resource parameters, and fully utilize historical strategy information to improve the elasticity and reliability of applications in multi-cloud environments and reduce operating costs. Summary of the Invention

[0006] The purpose of this invention is to provide an optimization method for application elastic scaling strategies in a multi-cloud environment, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an optimization method for application elastic scaling strategies in a multi-cloud environment, the method comprising: Acquire a multi-cloud environment data set; the multi-cloud environment data includes public cloud platform monitoring data, edge node load data, and a historical scaling-up / scaling strategy library; the edge node load data includes computing resource utilization, storage I / O throughput, and network traffic peaks; Based on the monitoring data of the public cloud platform, application load characteristics are extracted through the multi-cloud load feature extraction module. The load characteristics include CPU utilization fluctuation curve, memory allocation pattern and request latency distribution characteristics. Based on the edge node load data, resource association parameters are generated through a multi-dimensional resource fusion algorithm. The resource association parameters include computational resource redundancy, storage bottleneck coefficient, and network latency threshold. Spatiotemporal correlation analysis is performed on the historical expansion and contraction strategy library to generate an expansion and contraction evolution map; The application load characteristics, resource association parameters, and scaling evolution map are input into the elastic scaling optimization model to generate a strategy feature vector. Based on the policy feature vector, a multi-level scaling-down policy network is constructed through a dynamic optimization algorithm, and the scaling-down decisions and confidence levels are output. The hierarchy of the multi-level scaling-down policy network represents the scaling-down trigger condition branches, and the nodes represent resource scheduling weights and policy priorities.

[0008] Preferably, the step of extracting application load features through the multi-cloud load feature extraction module includes: The public cloud platform monitoring data is cleaned of outliers to generate a standardized load dataset. Based on a predefined application template library, a pattern matching algorithm is used to locate key business load areas and segment the load time series. A multi-resolution filtering algorithm is used to extract the trend features of the load time series, and candidate expansion / shrinkage regions are generated by combining the resource utilization fluctuation curve. The candidate scaling regions are compared with the baseline load parameters in the application template library to generate request latency distribution characteristics and confidence scores.

[0009] Preferably, the step of generating resource association parameters through a multi-dimensional resource fusion algorithm includes: The edge node load data is processed to align the acquisition time and eliminate cross-node data deviation. The mean and variance of computing resource utilization, storage I / O throughput and network traffic peak are calculated based on the sliding window statistical method to generate a dynamic resource baseline. Based on the difference between the dynamic resource baseline and the resource threshold, calculate the resource redundancy and storage bottleneck coefficient. The spectral characteristics of peak network traffic are compared with a preset network latency threshold to generate a network anomaly index, which is then encoded as a resource association parameter.

[0010] Preferably, the elastic scaling optimization model includes a feature enhancement module and a collaborative inference module, wherein the feature enhancement module includes: The CPU utilization fluctuation curve in the application load characteristics is normalized to obtain the first enhancement vector; The storage bottleneck coefficient in the resource association parameters is discretized and encoded to generate a second enhancement vector; Spatiotemporal convolution operations are performed on the expansion and contraction evolution map to extract policy propagation features and obtain the third enhancement vector; The first enhancement vector, the second enhancement vector, and the third enhancement vector are merged into a high-dimensional fusion feature through a feature concatenation layer.

[0011] Preferably, the collaborative reasoning module includes: Channel attention weighting is applied to the high-dimensional fusion features to generate a feature importance matrix; Policy association features are extracted using a cross-level feature interaction algorithm to generate a policy association matrix; The feature importance matrix and the policy association matrix are multiplied element-wise to generate collaborative reasoning features; By superimposing collaborative reasoning features with the original high-dimensional fused features through residual connections, a policy feature vector is output.

[0012] Preferably, the step of constructing a multi-level scaling strategy network through a dynamic optimization algorithm includes: Initialize the hierarchical attributes based on the scaling trigger condition branch, and generate a weight allocation matrix based on the policy priority; The policy feature vector is used as the input feature, and the weight allocation matrix consists of resource scheduling weights and policy priorities. The weight parameters of each level are iteratively updated using the gradient backpropagation algorithm, and the weight allocation matrix is ​​adjusted accordingly. Generate an optimal decision sequence that covers all triggering conditions based on the adjusted weight allocation matrix.

[0013] Preferably, the method for constructing the application template library includes: Collect standard load samples of various typical cloud-native applications and extract baseline resource utilization and request latency characteristics; The baseline resource utilization rate is transformed to generate application templates for multiple business scenarios. Templates are categorized according to application type and associated with a baseline load parameter database; Store the categorized templates as an application template library.

[0014] Preferably, the parameter optimization method for the dynamic resource baseline includes: Calculate the initial sliding window length and statistical interval based on the historical load data distribution; The parameter combinations are traversed using a grid search algorithm to select the parameters that provide the best balance between baseline stability and anomaly detection sensitivity.

[0015] Preferably, the parameter update method of the gradient backpropagation algorithm includes: The loss function between levels is defined as the sum of the decision error and the regularization term of the resource scheduling weight; The weight parameters of each level are initialized to random values, and the input features are normalized vectors; The gradient update amount of each level based on the previous level is calculated using the chain rule, and the optimal weight path is recorded. The weight allocation matrix is ​​updated by backpropagation based on the optimal weight path, and the final decision result is generated.

[0016] Preferably, the method for collecting edge node load data includes: Deploy lightweight monitoring agents at edge nodes to collect real-time data on computing resource utilization, storage I / O throughput, and network traffic peaks. The collected data is timestamped to eliminate cross-node clock skew. The raw data is initially aggregated by edge computing nodes to generate minute-level statistical values; The aggregated data is encrypted, transmitted, and stored in the edge database to form edge node load data.

[0017] Compared with the prior art, the beneficial effects of the present invention are: The proposed optimization method for application elastic scaling strategies in a multi-cloud environment improves the accuracy and efficiency of scaling decisions through multi-dimensional data fusion and intelligent analysis. This method acquires a multi-cloud environment data set including public cloud platform monitoring data, edge node load data, and a historical scaling strategy library. The edge node load data covers various aspects such as computing resource utilization, storage I / O throughput, and network traffic peaks, laying the foundation for comprehensive system status analysis.

[0018] Based on public cloud platform monitoring data, the multi-cloud load feature extraction module is used to clean and standardize the data for outliers. Combined with a predefined application template library, key business load areas are located. Trend features are extracted and candidate scaling areas are generated through multi-resolution filtering algorithms. Finally, request latency distribution features and confidence scores are generated, ensuring the comprehensiveness and accuracy of load feature extraction and avoiding decision bias caused by incomplete feature extraction in traditional methods.

[0019] For edge node load data, a multi-dimensional resource fusion algorithm is used to align the collection time and eliminate cross-node data deviation. A dynamic resource baseline is generated by using a sliding window statistical method, and then parameters such as resource redundancy and bottleneck coefficient are calculated. The peak spectrum characteristics of network traffic are compared with preset thresholds to generate an anomaly index, which realizes an accurate characterization of the resource status of edge nodes and provides a scientific basis for resource scheduling.

[0020] A spatiotemporal correlation analysis is performed on the historical scaling strategy library to generate an evolutionary map, which is then input into the elastic scaling optimization model along with load characteristics and resource correlation parameters. The model's feature enhancement module normalizes the CPU utilization fluctuation curve, discretizes and encodes the storage bottleneck coefficient, and performs spatiotemporal convolution operations on the evolutionary map to generate high-dimensional fused features. The collaborative inference module generates strategy feature vectors through operations such as channel attention weighting and cross-level feature interaction, fully exploring the potential correlation between historical strategies and current data, and improving the rationality and adaptability of strategy generation.

[0021] Based on policy feature vectors, a multi-level scaling policy network is constructed through a dynamic optimization algorithm. The hierarchical attributes are initialized according to the triggering conditions, and the weight parameters are iteratively updated using the gradient backpropagation algorithm to generate the optimal decision sequence covering all triggering conditions. This achieves dynamic optimization and precise triggering of scaling decisions, improving the system's response speed to load changes and resource scheduling efficiency.

[0022] The construction of the application template library provides a benchmark for load feature comparison by collecting typical cloud-native application load samples, performing scenario transformations, and classifying and storing them. Parameter optimization of the dynamic resource baseline utilizes a grid search algorithm to select optimal parameters, ensuring the stability of the resource baseline and the sensitivity of anomaly detection. The gradient backpropagation algorithm, combined with a loss function and regularization term, ensures the accuracy of weight parameter updates. Edge node data collection ensures the real-time performance, accuracy, and security of data through the deployment of lightweight monitoring agents, timestamp standardization, and encrypted data aggregation and transmission. Attached Figure Description

[0023] Figure 1 This is a schematic diagram illustrating the working principle of the application elastic scaling strategy optimization method in a multi-cloud environment described in this invention. Figure 2 Design diagram of the feature enhancement module for the elastic scaling optimization model; Figure 3 A diagram illustrating the construction of a multi-level scaling strategy network; Figure 4 Design diagram for edge node load data acquisition system. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figures 1-4 The present invention relates to an optimization method for application elastic scaling strategies in a multi-cloud environment, the specific implementation steps of which are as follows: Acquire a multi-cloud environment data set, which includes public cloud platform monitoring data, edge node load data, and a historical scaling-up / scaling strategy library. The edge node load data includes computing resource utilization, storage I / O throughput, and peak network traffic.

[0026] Based on public cloud platform monitoring data, application load characteristics are extracted through a multi-cloud load feature extraction module. These load characteristics include CPU utilization fluctuation curves, memory allocation patterns, and request latency distribution characteristics.

[0027] Based on edge node load data, resource association parameters are generated through a multi-dimensional resource fusion algorithm. These resource association parameters include computational resource redundancy, storage bottleneck coefficient, and network latency threshold.

[0028] Spatiotemporal correlation analysis is performed on the historical expansion and contraction strategy library to generate an expansion and contraction evolution map.

[0029] The application load characteristics, resource association parameters, and scaling evolution map are input into the elastic scaling optimization model to generate a strategy feature vector.

[0030] Based on policy feature vectors, a multi-level scaling policy network is constructed using a dynamic optimization algorithm, outputting scaling decisions and confidence levels. The hierarchy of the multi-level scaling policy network represents the branches of scaling trigger conditions, and the nodes represent resource scheduling weights and policy priorities.

[0031] Example 1: After acquiring the multi-cloud environment dataset, the public cloud platform monitoring data needs to be processed to extract application load characteristics. Outlier cleaning is performed on the public cloud platform monitoring data. During the collection, transmission, and storage of public cloud platform monitoring data, various factors such as network fluctuations, hardware failures, and software anomalies may interfere, resulting in outlier data points. If these outlier data points are not processed, they will adversely affect subsequent load characteristic extraction, leading to inaccurate results. Therefore, it is necessary to set a reasonable threshold range or use statistical methods, such as the mean plus or minus the standard deviation, to identify and remove these outlier data points, thereby generating a standardized load dataset.

[0032] Based on a predefined application template library, a pattern matching algorithm is used to analyze the business load in a standardized load dataset. The application template library stores standard load samples and related characteristics of various typical cloud-native applications. The pattern matching algorithm can match and identify the current load data according to these templates, thereby locating critical business load areas. Critical business load areas refer to the load portions that are crucial to application performance and business operation. After locating the critical business load areas, the load data needs to be segmented according to the time dimension to form a load time series. This divides continuous load data into multiple time segments, facilitating subsequent analysis and processing of load characteristics across different time periods.

[0033] A multi-resolution filtering algorithm is employed to process the load time series. This algorithm analyzes the load time series across different time scales and frequency ranges, extracting trend features. It filters out short-term fluctuations and noise, preserving long-term trends and key change patterns. Simultaneously, by combining this with resource utilization fluctuation curves, candidate scaling-up / scaling regions are identified. These curves visually reflect changes in computing and memory resource usage over time. Analyzing these fluctuations identifies periods with high or low resource utilization, thus determining candidate regions that may require scaling-up / scaling operations.

[0034] The candidate scaling regions are compared with baseline load parameters in the application template library. These baseline load parameters are typical characteristics of applications under normal operating conditions, obtained through extensive testing and verification, including resource utilization and request latency. By comparing these parameters, the distribution of request latency within the candidate scaling regions can be analyzed, generating request latency distribution characteristics and corresponding confidence scores. The request latency distribution characteristics reflect the application's response time under different load conditions, while the confidence score indicates the reliability and accuracy of these characteristics.

[0035] Building the application template library first requires collecting standard load samples of various typical cloud-native applications. These samples should cover different types of cloud-native applications, such as microservice applications and containerized applications, to ensure the diversity and representativeness of the template library. When collecting samples, it is necessary to extract baseline resource utilization and request latency characteristics, as these characteristics form the basis for building the application templates.

[0036] Scenario transformation is applied to the baseline resource utilization rate. Scenario transformation can simulate different business scenarios and load conditions, such as peak periods, off-peak periods, and sudden traffic surges, thereby generating application templates suitable for multiple business scenarios. Through scenario transformation, application templates become more flexible and versatile, adapting to different business needs.

[0037] Templates are categorized according to application type. Different types of applications have different business characteristics and resource requirements, so categorization makes the template library more organized and facilitates subsequent use and management. Simultaneously, a benchmark load parameter database is linked, associating each template with its corresponding benchmark load parameters to quickly obtain the necessary information for subsequent comparisons and analyses.

[0038] Store the categorized templates as an application template library. When storing, consider data security and accessibility to ensure the template library can be quickly retrieved and used when needed.

[0039] Example 2: When processing edge node load data to generate resource association parameters, a multi-dimensional resource fusion algorithm needs to be executed according to a specific process. The edge node load data needs to be aligned to the time of acquisition. Since edge nodes are distributed in different physical locations, their local clocks may deviate, and the start time and frequency of data acquisition may also differ due to device variations. This makes it impossible to directly compare cross-node data in the time dimension. Therefore, a unified timestamp calibration mechanism is needed to map the data acquired by each node to the same time coordinate system. Specifically, the Network Time Protocol (NTP) can be used to synchronize the clocks of edge nodes, ensuring that the timestamps of all data are consistent. For data with inconsistent acquisition frequencies, interpolation or resampling methods can be used to unify them to the same time interval, eliminating data deviations caused by time asynchrony, thereby generating a time-aligned standardized dataset.

[0040] The time-aligned load data is processed using a sliding window statistical method. Appropriate sliding windows are set for three key metrics: compute resource utilization, storage I / O throughput, and peak network traffic. The size of the sliding window needs to be determined based on the fluctuation characteristics of historical load data. For example, for compute resource utilization, if historical data shows significant fluctuations on a minute-by-minute scale, the window length can be set to 5-10 minutes. Within each window, the mean and variance of each metric are calculated. The mean reflects the average resource usage level within that time period, while the variance reflects the dispersion of the data and can be used to measure the stability of resource usage. By continuously moving the sliding window, a dynamically changing resource baseline is generated, which reflects the real-time trend of edge node resource load changes.

[0041] Parameter optimization of the dynamic resource baseline is a crucial step. In the initial stage, the initial length and statistical interval of the sliding window are calculated based on the distribution characteristics of historical load data. For example, by analyzing the autocorrelation function of historical data, the time range within which the data exhibits correlation is determined, serving as a reference for the initial window length. Then, a grid search algorithm is used to iterate through different parameter combinations, such as window lengths of 5, 10, and 15 minutes, and statistical intervals of 1 and 2 minutes, evaluating the baseline for each parameter combination. Evaluation metrics include baseline stability and anomaly detection sensitivity. Baseline stability is measured by calculating the fluctuation amplitude of the baseline during normal load periods; smaller fluctuation amplitudes indicate higher stability. Anomaly detection sensitivity is observed by simulating abnormal load scenarios to assess the baseline's ability to identify anomalies. The parameter combination that performs best in both metrics is selected as the final dynamic resource baseline parameters, ensuring that the baseline accurately reflects the normal state of resource load.

[0042] After generating a dynamic resource baseline, resource correlation parameters need to be calculated based on the difference between the baseline and preset resource thresholds. For computing resource utilization, the preset thresholds are typically determined based on the maximum processing capacity of the hardware and business requirements, such as setting the CPU utilization threshold to 80% and the memory utilization threshold to 70%. When calculating resource redundancy, the threshold is subtracted from the mean of the dynamic baseline, and then divided by the threshold; the resulting ratio is the computing resource redundancy, reflecting the remaining space of resources relative to the threshold. For storage I / O throughput, the preset storage bottleneck threshold is determined based on the maximum throughput of the storage device and business I / O requirements. The storage bottleneck coefficient is calculated by comparing the mean of the dynamic baseline with the storage bottleneck threshold. If the mean is close to or exceeds the threshold, the bottleneck coefficient is large, indicating a potential storage bottleneck; conversely, it is small. This coefficient quantifies the potential bottleneck risk of the storage system.

[0043] For peak network traffic, its spectral characteristics need to be compared with a preset network latency threshold. First, spectral analysis is performed on the peak network traffic data. A Fast Fourier Transform (FFT) can be used to convert the time-domain signal to a frequency-domain signal, extracting the main frequency components and energy distribution characteristics. The preset network latency threshold is determined based on the service's tolerance for network latency; for example, critical services require network latency to be no more than 50ms. The energy of frequency components related to network latency (such as high-frequency noise components) in the spectral characteristics is compared with the threshold. If the energy exceeds the threshold, it indicates a potential risk of abnormal network latency, generating a corresponding network anomaly index. This index is encoded into resource-related parameters using specific coding rules (such as normalization to the range of 0-100) to characterize the degree of network anomaly.

[0044] The data collection process for edge node load data must ensure the real-time nature and accuracy of the data. A lightweight monitoring agent is deployed on the edge nodes. This agent employs a low-power, high-reliability design to avoid excessive burdening the normal business operations of the edge nodes. The monitoring agent collects data in real time, including computing resource utilization (such as CPU utilization and memory utilization), storage I / O throughput (such as read / write operations per second and read / write bytes per second), and peak network traffic (such as maximum data packets per second and bytes per second). After data collection, the data is immediately timestamped and standardized. By calibrating with the local clock of the edge node (which has been synchronized to the NTP server), a precise timestamp is added to each data entry, eliminating cross-node clock skew.

[0045] The collected data, after being timestamped and standardized, undergoes initial aggregation processing by edge computing nodes. These nodes possess local computing capabilities, enabling them to aggregate the raw data at a minute-level granularity, such as calculating the average CPU utilization per minute, and the peak and average storage I / O throughput. This aggregation process reduces data transmission volume while preserving key statistical characteristics. The aggregated data is then transmitted to the edge database via an encrypted transmission channel (such as SSL / TLS protocol) to ensure data security during transmission. Upon arrival at the edge database, the data is stored according to a specific data structure, forming complete edge node load data and providing reliable data support for subsequent resource association parameter generation.

[0046] Example 3: The elastic scaling optimization model includes a feature enhancement module and a collaborative inference module, which work together to generate accurate policy feature vectors. The feature enhancement module begins by normalizing the CPU utilization fluctuation curve in the application load characteristics. Since the CPU utilization value range may vary significantly across different time periods or application scenarios, directly using the raw data can lead to an imbalance in the weights between features. Therefore, it is necessary to map the CPU utilization fluctuation curve values ​​to a unified interval, such as [0,1]. Specifically, the maximum and minimum values ​​in the curve are first determined, and then a linear transformation is used to convert the value of each data point into a normalized value, resulting in the first enhancement vector. This vector can eliminate the influence of differences in the numerical range, ensuring that the fluctuation characteristics of CPU utilization are reasonably reflected in subsequent processing.

[0047] Next, the storage bottleneck coefficient in the resource association parameters is processed and discretized. The storage bottleneck coefficient is typically a continuous value, and directly inputting it into the model may not effectively capture its inherent semantic information. Discretization encoding divides the storage bottleneck coefficient into several intervals based on its numerical range, with each interval corresponding to a discrete encoded value. For example, if the storage bottleneck coefficient ranges from 0 to 1, it can be divided into four intervals: 0-0.2, 0.2-0.5, 0.5-0.8, and 0.8-1, encoded as 0, 1, 2, and 3 respectively. In this way, the continuous storage bottleneck coefficient is converted into a discrete second augmentation vector, enabling the model to more clearly identify different levels of storage bottleneck states.

[0048] For scaling evolution maps, spatiotemporal convolution operations are required to extract policy propagation features. A scaling evolution map is a visual representation of the distribution and evolution of historical scaling policies in time and space, containing information such as policy triggering conditions, execution effects, and mutual influences. Spatiotemporal convolution operations can simultaneously process the temporal series features and spatial correlation features in the map. Specifically, the map is scanned using convolutional layers of a convolutional neural network (CNN), utilizing convolutional kernels of different sizes to capture spatiotemporal features at different scales. For example, a 3×3 convolutional kernel is used to extract local spatiotemporal correlation features, while a larger kernel is used to capture broader policy propagation trends. After multiple rounds of convolution operations, the temporal evolution patterns and spatial propagation modes of the policies are extracted, resulting in a third enhancement vector.

[0049] After obtaining the first, second, and third augmentation vectors, they are merged into a high-dimensional fusion feature through a feature concatenation layer. The feature concatenation layer sequentially joins the elements of the three vectors to form a higher-dimensional feature vector. For example, if the first augmentation vector has a dimension of 10, the second augmentation vector has a dimension of 5, and the third augmentation vector has a dimension of 15, then the concatenated high-dimensional fusion feature will have a dimension of 30. This concatenation operation can integrate features from different sources and of different types into the same feature space, providing rich input information for subsequent collaborative inference.

[0050] The collaborative inference module applies channel attention weighting to high-dimensional fused features. Different channels (i.e., feature elements at different positions) of the high-dimensional fused features contribute differently to the final policy feature vector, and the channel attention mechanism automatically learns the importance of each channel. Specifically, the high-dimensional fused features are first subjected to global average pooling and global max pooling to obtain the mean and maximum values ​​for each channel dimension, respectively. Then, these two sets of information are input into a multilayer perceptron (MLP), processed by an activation function to generate channel attention weights. Finally, the feature value of each channel is multiplied by its corresponding attention weight to generate a feature importance matrix. This matrix highlights the feature channels that are more critical to policy optimization and suppresses less important channels, thereby improving the expressive power of the features.

[0051] A cross-level feature interaction algorithm is used to extract policy association features and generate a policy association matrix. Complex relationships exist between features at different levels within high-dimensional fused features, and the cross-level feature interaction algorithm can uncover these relationships. Specifically, the idea of ​​a graph neural network (GNN) can be adopted, treating high-dimensional fused features as nodes in a graph and the relationships between features as edges. Through graph convolution operations, each node (feature) can aggregate information from its neighboring nodes (related features), thereby capturing long-distance dependencies and complex associations between features. After multiple rounds of graph convolution operations, a policy association matrix is ​​generated, which represents the strength and pattern of associations between different features.

[0052] Element-wise multiplication of the feature importance matrix and the policy association matrix generates collaborative inference features. This element-wise multiplication combines the importance weights of channels with the relationships between features, further highlighting the connections between important features. For example, if a feature channel has a high weight in the feature importance matrix and that feature has a strong association with other features in the policy association matrix, their product will be larger, thus strengthening the collaborative inference features. This method generates more discriminative features, providing a more accurate basis for policy optimization.

[0053] The policy feature vector is output by superimposing collaborative inference features with the original high-dimensional fused features through residual connections. The introduction of residual connections is to avoid losing original information during feature transformation. Specifically, the elements at corresponding positions of the collaborative inference features and the original high-dimensional fused features are added to obtain the final policy feature vector. This approach utilizes effective features extracted during collaborative inference while retaining fundamental information from the original features, improving feature integrity and reliability. The output policy feature vector integrates information from application load features, resource association parameters, and scaling evolution maps, accurately representing the application status and scaling requirements in the current multi-cloud environment, providing high-quality input for subsequent construction of multi-level scaling policy networks.

[0054] In the feature enhancement module, normalization, discretization, and spatiotemporal convolution operations process different types of features, transforming them into forms suitable for model processing; the feature concatenation layer achieves the fusion of multi-source features. In the collaborative inference module, operations such as channel attention weighting, cross-level feature interaction, element-wise multiplication, and residual connections progressively optimize and enhance the fused features, ultimately generating feature vectors containing rich policy information. These steps work closely together to form a complete feature processing flow, ensuring the extraction of the most valuable information from the original data and providing a solid feature foundation for optimizing application elastic scaling strategies in multi-cloud environments.

[0055] Example 4: When constructing a multi-level scaling-up / scaling strategy network using a dynamic optimization algorithm, each step must be completed according to a specific logical flow. The hierarchical attributes must be initialized based on the scaling-up / scaling trigger condition branches, and a weight allocation matrix must be generated based on the policy priority. The scaling-up / scaling trigger condition branches are typically composed of combinations of resource thresholds under various business scenarios. For example, scaling up is triggered when CPU utilization exceeds 80% and memory utilization exceeds 70%, or scaling down is triggered when CPU utilization is below 30% for a certain period. Each trigger condition branch corresponds to a level in the multi-level scaling-up / scaling strategy network. During initialization, the trigger condition parameters, level depth, and logical relationships with other levels for each level must be clearly defined.

[0056] The determination of strategy priorities is based on the historical performance of scaling-up and scaling-down strategies and the importance of the business. Scaling-up and scaling-down strategies for critical business scenarios are assigned higher priority, while strategies for non-critical scenarios have lower priority. The weight allocation matrix consists of resource scheduling weights and strategy priorities. Resource scheduling weights reflect the importance of different resource types (such as computing resources, storage resources, and network resources) during the scheduling process, while strategy priorities reflect the execution order of different strategies. For example, in the weight allocation matrix, computing resource scheduling weights might account for 40%, storage resources for 30%, and network resources for 30%, while the priority of scaling-up strategies for critical business scenarios is set to 1, and the priority of scaling-down strategies for non-critical business scenarios is set to 0.5.

[0057] The policy feature vector is used as input features and fed into the constructed multi-level scaling policy network. The policy feature vector is a high-dimensional vector generated by the elastic scaling optimization model, containing comprehensive information about application load characteristics, resource association parameters, and scaling evolution patterns. Before being input into the network, the policy feature vector needs to be normalized to map its numerical range to a uniform interval, avoiding network training bias caused by large differences in feature values. The specific normalization method is as follows: calculate the maximum and minimum values ​​of each element in the feature vector, and transform the value of each element into a value within the interval [0,1] through a linear transformation.

[0058] The gradient backpropagation algorithm iteratively updates the weight parameters of each layer, adjusting the weight allocation matrix. The parameter update method of the gradient backpropagation algorithm begins by defining a loss function between layers, which is the sum of the decision error and the regularization term of the resource scheduling weights. The decision error measures the difference between the current network output scaling decision and the expected decision, and can be calculated using methods such as mean squared error. The regularization term of the resource scheduling weights is used to prevent overfitting and enhance the model's generalization ability, typically employing L1 or L2 regularization methods.

[0059] The weights of each layer are initialized to random values, typically within a small range such as [-0.1, 0.1], to avoid unstable network training due to excessively large initial weights. The input features are normalized policy feature vectors, which are fed into the first layer of the network. After the weights of this layer are calculated, the output features are obtained and passed to the next layer. This process is repeated layer by layer until the final scaling decision output is generated.

[0060] The chain rule is used to calculate the gradient update of each layer based on the previous layer. The chain rule, a derivative rule in calculus used for finding the derivative of composite functions, enables backpropagation of gradients in neural networks. Specifically, starting from the output layer of the network, the gradient of the output error with respect to the weights of the last layer is calculated, and then the gradients of the weights of each layer are calculated layer by layer backward. During the calculation, the gradient update of each layer is recorded, along with the weight path that minimizes the loss function during training, i.e., the optimal weight path.

[0061] The weight allocation matrix is ​​updated using backpropagation based on the optimal weight path. During iterative training, after each round of forward and backward propagation, the weight allocation matrix is ​​adjusted based on the calculated gradient update. The adjustment is made by subtracting the product of the learning rate and the gradient from the weight parameters. The learning rate is a preset hyperparameter that controls the step size of the weight updates. Through multiple iterations of training, the weight allocation matrix is ​​continuously optimized, causing the loss function to gradually decrease until it converges to a small value. The weight allocation matrix at this point is the optimized result.

[0062] Based on the adjusted weight allocation matrix, an optimal decision sequence covering all triggering conditions is generated, outputting scaling decisions and confidence levels. The optimized weight allocation matrix accurately reflects the importance of each triggering condition branch, the weight of resource scheduling, and the priority of the strategy. Based on this matrix, all possible triggering conditions are analyzed, and the optimal scaling strategy for each triggering condition is calculated, forming a complete decision sequence. This decision sequence covers scaling decisions under various business scenarios and resource states, providing comprehensive elastic scaling guidance for applications in multi-cloud environments.

[0063] The output of scaling decisions includes the decision type (e.g., scaling up, scaling down, maintaining the status quo), the amount of resource adjustment (e.g., adding 2 compute nodes, removing 1 storage node), and the timing of the execution strategy. Confidence represents the reliability of the decision, calculated by comparing the model's performance during training with the degree of matching between the model and the current input features. For example, when the input features highly match the features of a certain scenario in the training set, the confidence of the output decision will be high; conversely, if the feature matching is low, the confidence will be low.

[0064] In constructing a multi-level scaling strategy network, initializing hierarchical attributes and generating weight allocation matrices lay the structural foundation for network construction; the input of policy feature vectors provides the basis for decision-making; iterative optimization using the gradient backpropagation algorithm ensures the accuracy of network parameters; and the final generated optimal decision sequence and confidence output provide specific and executable decision-making schemes for elastic scaling of applications in multi-cloud environments. Each stage is interconnected and mutually influential, forming a complete dynamic optimization process that enables the constructed multi-level scaling strategy network to adapt to complex and ever-changing multi-cloud environments, achieving precise scheduling and elastic management of application resources.

[0065] Example 5: During the collection and processing of load data at edge nodes, multiple steps are required to ensure the real-time performance, accuracy, and security of the data. First, a lightweight monitoring agent must be deployed at the edge nodes. This agent employs a low-resource-consumption architecture, typically running as a container or lightweight process to avoid significantly impacting the business processing performance of the edge nodes. The monitoring agent integrates a data acquisition module that can obtain real-time computing resource utilization, specifically including CPU utilization, the time ratio between user mode and system mode, and the used, free, and cached memory usage. It also collects storage I / O throughput, covering metrics such as the number of read / write operations per second, the number of bytes read / written, and the I / O queue length. Furthermore, it collects network traffic peaks, including inbound and outbound byte rates, the number of data packets, and the time of peak occurrence.

[0066] After data collection, the data needs to be timestamped. Since edge nodes may be located in different physical locations, their local clocks may deviate due to hardware differences or network latency. Therefore, synchronization with a clock server via Network Time Protocol (NTP) is necessary to ensure that all nodes' clocks are consistent. For each piece of collected data, the local clock of the edge node must be calibrated before adding a timestamp to eliminate cross-node clock deviations. In practice, a time synchronization thread can be set up in the data acquisition module to periodically communicate with the NTP server to obtain the standard time and adjust the local clock. Then, the calibrated timestamp is added immediately upon data generation to ensure the accuracy and consistency of the data timestamps.

[0067] Data, after being timestamped and standardized, is initially aggregated by edge computing nodes. Edge computing nodes typically possess local computing capabilities, enabling real-time processing of the raw data. The aggregation process is executed at a minute-level granularity. For example, for CPU utilization data collected every minute, the average, maximum, and minimum values ​​are calculated; for storage I / O throughput, the total number of read / write bytes and average I / O response time per minute are recorded; for network traffic peaks, the maximum inbound and outbound traffic per minute is recorded. This aggregation process reduces data volume while retaining key statistical characteristics, facilitating subsequent transmission and storage. The aggregation algorithm employs a sliding window mechanism, continuously updating the statistical results within the window over time to ensure that the aggregated data reflects the load status of the edge nodes in real time.

[0068] The aggregated data needs to be sent to the edge database via an encrypted transmission channel. To ensure data transmission security, SSL / TLS encryption protocols are used to encapsulate the data, preventing it from being stolen or tampered with during transmission. Establishing the transmission channel requires authentication and key negotiation. The edge computing node and the edge database server verify each other's identities using digital certificates and negotiate to generate a symmetric encryption key for data encryption and decryption. During data transmission, a streaming transmission method is used, sending the aggregated minute-level statistical data in batches to avoid transmission delays due to large data volumes.

[0069] Once data arrives at the edge database, it is stored according to a pre-defined data structure. The edge database employs a distributed storage architecture, supporting high availability and scalability, and can dynamically adjust storage nodes based on the number of edge nodes and the amount of data. The database table structure design must include time fields, node identifier fields, various resource indicator fields, and status fields. The time field records the data collection time, the node identifier field distinguishes different edge nodes, the resource indicator fields store data such as computing resource utilization, storage I / O throughput, and peak network traffic, and the status field records the data collection status and integrity flags. During storage, an index is created on the time field for quick querying and retrieval of historical data.

[0070] In the parameter optimization process of dynamic resource baseline, the initial sliding window length and statistical interval are determined based on the distribution characteristics of historical load data. First, the time series of historical load data is analyzed, the autocorrelation function of the data is calculated, and the time range in which the data shows correlation is determined, serving as a reference for the initial sliding window length. For example, if historical data shows that CPU utilization has a strong correlation within 10 minutes, and the correlation decreases significantly after 10 minutes, then the initial sliding window length is set to 10 minutes. The initial value of the statistical interval is usually set to 1 minute to capture short-term fluctuations in the data.

[0071] When using a grid search algorithm to traverse different parameter combinations, the search range for each parameter needs to be defined. The search range for the sliding window length can be set from 5 to 30 minutes, increasing in 5-minute increments; the search range for the statistical interval can be set from 1 to 5 minutes, increasing in 1-minute increments. For each parameter combination, tests are performed on historical load data to calculate baseline stability and anomaly detection sensitivity. Baseline stability is measured by calculating the standard deviation of the baseline during normal load periods; a smaller standard deviation indicates higher stability. Anomaly detection sensitivity is achieved by injecting simulated outliers into historical data and observing the baseline's ability to identify anomalies; a higher detection rate indicates higher sensitivity.

[0072] During the grid search process, the stability and sensitivity indices for each parameter combination are recorded, and a parameter-indicator relationship graph is plotted to identify the optimal parameter combination for both indices. For example, when the sliding window length is 15 minutes and the statistical interval is 2 minutes, the baseline standard deviation is minimized and the anomaly detection rate is highest; therefore, this parameter combination is selected as the optimization parameter for the dynamic resource baseline. After parameter optimization, it is applied to the sliding window statistical process to ensure that the generated dynamic resource baseline accurately reflects the normal resource load status of edge nodes, providing a reliable basis for subsequent calculations of parameters such as resource redundancy and storage bottleneck coefficients.

[0073] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing application elastic scaling strategies in a multi-cloud environment, characterized in that, include: Acquire a multi-cloud environment data set; the multi-cloud environment data includes public cloud platform monitoring data, edge node load data, and a historical scaling-up / scaling strategy library; the edge node load data includes computing resource utilization, storage I / O throughput, and network traffic peaks; Based on the monitoring data of the public cloud platform, application load characteristics are extracted through the multi-cloud load feature extraction module. The load characteristics include CPU utilization fluctuation curve, memory allocation pattern and request latency distribution characteristics. Based on the edge node load data, resource association parameters are generated through a multi-dimensional resource fusion algorithm. The resource association parameters include computational resource redundancy, storage bottleneck coefficient, and network latency threshold. Spatiotemporal correlation analysis is performed on the historical expansion and contraction strategy library to generate an expansion and contraction evolution map; The application load characteristics, resource association parameters, and scaling evolution map are input into the elastic scaling optimization model to generate a strategy feature vector. Based on the policy feature vector, a multi-level scaling-down policy network is constructed through a dynamic optimization algorithm, and the scaling-down decisions and confidence levels are output. The hierarchy of the multi-level scaling-down policy network represents the scaling-down trigger condition branches, and the nodes represent resource scheduling weights and policy priorities.

2. The method for optimizing application elastic scaling strategies in a multi-cloud environment according to claim 1, characterized in that, The step of extracting application load features through the multi-cloud load feature extraction module includes: The public cloud platform monitoring data is cleaned of outliers to generate a standardized load dataset. Based on a predefined application template library, a pattern matching algorithm is used to locate key business load areas and segment the load time series. A multi-resolution filtering algorithm is used to extract the trend features of the load time series, and candidate expansion / shrinkage regions are generated by combining the resource utilization fluctuation curve. The candidate scaling regions are compared with the baseline load parameters in the application template library to generate request latency distribution characteristics and confidence scores.

3. The method for optimizing application elastic scaling strategies in a multi-cloud environment according to claim 1, characterized in that, The generation of resource association parameters through a multi-dimensional resource fusion algorithm includes: The edge node load data is processed to align the acquisition time and eliminate cross-node data deviation. The mean and variance of computing resource utilization, storage I / O throughput and network traffic peak are calculated based on the sliding window statistical method to generate a dynamic resource baseline. Based on the difference between the dynamic resource baseline and the resource threshold, calculate the resource redundancy and storage bottleneck coefficient. The spectral characteristics of peak network traffic are compared with a preset network latency threshold to generate a network anomaly index, which is then encoded as a resource association parameter.

4. The method for optimizing application elastic scaling strategies in a multi-cloud environment according to claim 1, characterized in that, The elastic scaling optimization model includes a feature enhancement module and a collaborative inference module. The feature enhancement module includes: The CPU utilization fluctuation curve in the application load characteristics is normalized to obtain the first enhancement vector; The storage bottleneck coefficient in the resource association parameters is discretized and encoded to generate a second enhancement vector; Spatiotemporal convolution operations are performed on the expansion and contraction evolution map to extract policy propagation features and obtain the third enhancement vector; The first enhancement vector, the second enhancement vector, and the third enhancement vector are merged into a high-dimensional fusion feature through a feature concatenation layer.

5. The method for optimizing application elastic scaling strategy in a multi-cloud environment according to claim 4, characterized in that, The collaborative reasoning module includes: Channel attention weighting is applied to the high-dimensional fusion features to generate a feature importance matrix; Policy association features are extracted using a cross-level feature interaction algorithm to generate a policy association matrix; The feature importance matrix and the policy association matrix are multiplied element-wise to generate collaborative reasoning features; By superimposing collaborative reasoning features with the original high-dimensional fused features through residual connections, a policy feature vector is output.

6. The method for optimizing application elastic scaling strategies in a multi-cloud environment according to claim 1, characterized in that, The construction of a multi-level scaling strategy network through a dynamic optimization algorithm includes: Initialize the hierarchical attributes based on the scaling trigger condition branch, and generate a weight allocation matrix based on the policy priority; The policy feature vector is used as the input feature, and the weight allocation matrix consists of resource scheduling weights and policy priorities. The weight parameters of each level are iteratively updated using the gradient backpropagation algorithm, and the weight allocation matrix is ​​adjusted accordingly. Generate an optimal decision sequence that covers all triggering conditions based on the adjusted weight allocation matrix.

7. The method for optimizing application elastic scaling strategies in a multi-cloud environment according to claim 2, characterized in that, The method for constructing the application template library includes: Collect standard load samples of various typical cloud-native applications and extract baseline resource utilization and request latency characteristics; The baseline resource utilization rate is transformed to generate application templates for multiple business scenarios. Templates are categorized according to application type and associated with a baseline load parameter database; Store the categorized templates as an application template library.

8. The method for optimizing application elastic scaling strategy in a multi-cloud environment according to claim 3, characterized in that, The parameter optimization method for the dynamic resource baseline includes: Calculate the initial sliding window length and statistical interval based on the historical load data distribution; The parameter combinations are traversed using a grid search algorithm to select the parameters that provide the best balance between baseline stability and anomaly detection sensitivity.

9. The method for optimizing application elastic scaling strategies in a multi-cloud environment according to claim 6, characterized in that, The parameter update method of the gradient backpropagation algorithm includes: The loss function between levels is defined as the sum of the decision error and the regularization term of the resource scheduling weight; The weight parameters of each level are initialized to random values, and the input features are normalized vectors; The gradient update amount of each level based on the previous level is calculated using the chain rule, and the optimal weight path is recorded. The weight allocation matrix is ​​updated by backpropagation based on the optimal weight path, and the final decision result is generated.

10. The method for optimizing application elastic scaling strategies in a multi-cloud environment according to claim 1, characterized in that, The method for collecting edge node load data includes: Deploy lightweight monitoring agents at edge nodes to collect real-time data on computing resource utilization, storage I / O throughput, and network traffic peaks. The collected data is timestamped to eliminate cross-node clock skew. The raw data is initially aggregated by edge computing nodes to generate minute-level statistical values; The aggregated data is encrypted, transmitted, and stored in the edge database to form edge node load data.

Images