Dynamic network load balancing method and system

Abstract

The invention discloses a dynamic network load balancing method and system. According to the method, multi-dimensional operation data is collected through a node cluster deployed at the edge of a network, an original data set comprehensively reflecting the operation state of the network is constructed, then the original data is deeply analyzed, features of key information are screened out and integrated into structured feature vectors, and the feature vectors of the key information are analyzed. And inputting the feature vector into a pre-constructed AI prediction model to generate a prediction result of the network load, generating a dynamic load scheduling strategy according to the prediction result, and issuing a strategy instruction to an edge node cluster and a region control center, so that the edge node cluster and the region control center jointly realize dynamic balance of the network load. According to the method, by predicting the network load change in advance, resource capacity expansion and path optimization can be completed before the traffic peak comes; meanwhile, due to the cooperation mechanism of the edge nodes and the regional control center, low-delay processing of local flow is guaranteed, overall planning and allocation of global resources are achieved, and the stability and reliability of the system are enhanced.

Description

Technical Field

[0001] The present disclosure generally relates to the field of computer network technology, and in particular to a dynamic network load balancing method and system. Background Art

[0002] Network Load Balancing (NLB) is a technology that evenly distributes network traffic to multiple servers or computing resources, aiming to optimize resource utilization, improve system availability and response speed.

[0003] Traditional load balancing strategies, such as round-robin and weighted methods, usually make traffic distribution decisions based on historical information such as server load data and traffic patterns collected over a period of time. However, when faced with sudden traffic bursts such as live streaming and online gaming peaks, server utilization may be uneven. Some servers may be overloaded due to sudden traffic bursts, while other servers may not be fully utilized, making it impossible to achieve real-time and reasonable traffic distribution.

[0004] Therefore, how to achieve dynamic network resource regulation is a problem that those skilled in the art urgently need to solve. Summary of the Invention

[0005] In view of the above-mentioned defects or deficiencies in the prior art, it is desired to provide a dynamic network load balancing method and system that can realize the regulation of dynamic network resources.

[0006] In a first aspect, an embodiment of the present application provides a dynamic network load balancing method, comprising:

[0007] Collect multi-dimensional operation data through edge node clusters;

[0008] Extracting features from the multidimensional operating data, and combining the extracted features to obtain a feature vector;

[0009] Input the feature vector into the AI prediction model to predict the network load in the future time period and generate a prediction result;

[0010] A dynamic load scheduling strategy is generated based on the prediction results and sent to the edge node cluster and the regional control center to collaboratively perform resource optimization and traffic scheduling of the service instance.

[0011] In one embodiment, generating a dynamic load scheduling strategy according to the prediction result includes:

[0012] When the prediction error rate exceeds the error rate threshold, emergency response processing is performed;

[0013] When the predicted traffic growth rate exceeds the growth rate threshold, pre-capacity expansion processing is performed;

[0014] When all the prediction indicators are within the normal range, automatic balancing processing is performed;

[0015] The emergency response process, the pre-capacity expansion process, and the automatic balancing process respectively perform resource allocation optimization operations that match the prediction index.

[0016] In one embodiment, the emergency response processing includes: performing a combination of node expansion, traffic redirection, and SSL offloading;

[0017] The pre-capacity expansion process includes: performing an elastic resource allocation operation based on a predicted growth rate.

[0018] In one embodiment, the deployment method of the AI prediction model includes:

[0019] Deploy a global model on the cloud, where the global model is built based on an LSTM network structure to generate initial model parameters;

[0020] A local copy of the global model is deployed on each edge node in the edge node cluster as a local model, and the local model is used to perform local load prediction.

[0021] In one embodiment, the training method of the AI prediction model includes:

[0022] Each edge node independently trains its local model based on local data to generate gradient parameters;

[0023] Receive and aggregate the gradient parameters of each edge node to generate a global gradient update;

[0024] Adjusting the parameters of the global model according to the global gradient update amount, and extracting the adjusted global model parameters;

[0025] The global model parameters are sent to each of the edge nodes so as to synchronize the parameters of each of the local models according to the global model parameters.

[0026] In one embodiment, before receiving and aggregating the gradient parameters of each edge node, the method further includes:

[0027] Encrypting the corresponding gradient parameter by each edge node to obtain an encrypted gradient parameter;

[0028] Then, the gradient parameters of each edge node are received and aggregated, including: based on a federated learning framework, receiving and aggregating each encrypted gradient parameter through a secure communication protocol.

[0029] In one embodiment, performing feature extraction on the multidimensional operation data includes:

[0030] Mapping the multi-dimensional operation data to a four-dimensional coordinate system of time-space-service-network-business to generate coordinate data in four-dimensional space;

[0031] A feature association relationship is extracted based on the coordinate data through a spatiotemporal encoder.

[0032] In one embodiment, generating a dynamic load scheduling strategy according to the prediction result includes:

[0033] Construct a multi-dimensional routing decision matrix;

[0034] Mapping the state parameter of each edge node in the edge node cluster to a corresponding dimension of the routing decision matrix to generate a dimension index;

[0035] Dynamically adjusting the weight coefficients of each dimension in the routing decision matrix according to the prediction results;

[0036] Calculate a comprehensive score for each candidate path based on the dimensional indicators and the weight coefficients;

[0037] The candidate path with the highest comprehensive score is selected as the optimal route, and a dynamic load scheduling strategy including the optimal path is generated.

[0038] In a second aspect, an embodiment of the present application provides a dynamic network load balancing system, comprising: an edge node cluster, a regional control center, and a cloud collaboration center;

[0039] The edge node cluster includes several edge nodes, each of which is used to collect multi-dimensional operation data of its own node;

[0040] The regional control center is used to collect the multi-dimensional operation data of each edge node cluster, perform feature extraction on the multi-dimensional operation data, combine the extracted features to obtain a feature vector; and input the feature vector into the AI prediction module;

[0041] The cloud-based collaborative center is deployed with an AI prediction module and a dynamic scheduling engine; the AI prediction module is used to predict the network load in the future time period based on the feature vector and generate a prediction result; the dynamic scheduling engine is used to obtain the prediction result, generate a dynamic load scheduling strategy based on the prediction result, and send it to the edge node cluster and the regional control center to collaboratively perform resource optimization and traffic scheduling of the service instance.

[0042] In one embodiment, each of the edge nodes includes: a smart probe, a local scheduler, and an edge AI engine;

[0043] Wherein, the intelligent probe is used to collect multi-dimensional operation data;

[0044] The edge AI engine is used to generate a lightweight feature vector based on the multi-dimensional operation data, and input the lightweight feature vector into the local model of the AI prediction model to predict the local load and generate a local prediction result;

[0045] The local scheduler is used to combine the dynamic load scheduling strategy generated in the cloud with the local prediction result to perform container expansion and contraction and / or dynamic routing adjustment.

[0046] The dynamic network load balancing method provided by this application collects multi-dimensional operating data through a node cluster deployed at the edge of the network to construct an original data set that comprehensively reflects the network operation status. The original data is then deeply analyzed to screen out the features of key information and integrate them into a structured feature vector. The feature vector is input into a pre-built AI prediction model to make a forward-looking prediction of the network load in the future time period, generate a prediction result, and then generate a targeted dynamic load scheduling strategy based on the prediction result. The strategy instructions are sent to the edge node cluster and the regional control center. The two work together. The edge node is responsible for the real-time scheduling of local traffic and the elastic allocation of resources, while the regional control center coordinates cross-node resource migration from a global perspective to jointly achieve dynamic balancing of network load. By predicting network load changes in advance, this method can complete resource expansion and path optimization before the traffic peak arrives, thereby avoiding resource waste or service performance degradation caused by uneven server load. At the same time, the collaborative mechanism between the edge node and the regional control center not only ensures low-latency processing of local traffic, but also realizes the coordinated allocation of global resources, enhancing the stability and reliability of the system.

[0047] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0049] Figure 1 A flow chart of a dynamic network load balancing method provided in an embodiment of the present application is shown;

[0050] Figure 2 An exemplary structural block diagram of a dynamic network load balancing system provided in an embodiment of the present application is shown;

[0051] Figure 3 A schematic diagram of core components under a distributed architecture provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0052] The present application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0053] It should be noted that, in the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments. Although the embodiments of the present application provide the method operation instruction steps shown in the following embodiments or drawings, more or fewer operation instruction steps may be included in the method based on routine or no creative labor. In steps where there is no necessary causal relationship logically, the execution order of these steps is not limited to the execution order provided in the embodiments of the present application. During the actual processing process or when the device is executed, the method may be executed in the order of the methods shown in the embodiments or drawings or in parallel.

[0054] Example 1:

[0055] This application proposes a dynamic network load balancing method, please refer to Figure 1 , Figure 1 FIG. 1 shows a flow chart of a dynamic network load balancing method provided by an embodiment of the present application. Figure 1 As shown, the method includes:

[0056] S101. Collect multi-dimensional operation data through edge node clusters;

[0057] Utilizing a cluster of multiple nodes distributed at the edge of the network, data covering multiple dimensions reflecting network operation status and business conditions is collected. These multiple dimensions include one or more of spatiotemporal characteristics, service characteristics, network characteristics, and business characteristics, comprehensively presenting network operation status and providing a comprehensive and accurate data foundation for subsequent load forecasting and scheduling strategy formulation. Spatiotemporal characteristics include user latitude and longitude, time zone information, and traffic patterns based on sliding window analysis. Network characteristics include bandwidth utilization, packet loss rate, and the frequency of BGP (Border Gateway Protocol) routing table changes.

[0058] After collecting data, components on each edge node perform preliminary local collation and aggregation, integrating data from different dimensions into a structured data format for subsequent transmission and processing. This collated data communicates with the central management node via the gRPC protocol, synchronizing node status (CPU / memory utilization, network latency, and other information) and passing the collected data on to subsequent processing.

[0059] S102, extracting features from the multi-dimensional operating data, and combining the extracted features to obtain a feature vector;

[0060] The raw data collected by edge nodes contains a lot of noise (such as instantaneous network fluctuations) and redundant information (such as duplicate device IDs). At the same time, the dimensions and distribution of data of different dimensions (such as geographic coordinates and business conversion rates) vary greatly.

[0061] To this end, this step filters out the key information with the most predictive value, such as traffic trends, service dependencies, etc., from the multi-dimensional original data of time, space, service, network, and business collected by the edge nodes. The specific settings can be made according to the needs of the actual application scenario, and this is not limited in this embodiment.

[0062] The extracted features are integrated into a structured numerical vector as input to the AI prediction model.

[0063] S103: Input the feature vector into the AI prediction model to predict the network load in the future time period and generate a prediction result;

[0064] Traditional static load balancing strategies (such as polling and weighting methods) rely on historical data or current status and cannot cope with sudden traffic (such as live streaming and game launches), resulting in scheduling delays and server load imbalance. In order to meet the real-time requirements of high-concurrency scenarios and complete resource allocation before the load peak arrives, this embodiment uses the already constructed AI prediction model and uses the obtained feature vector as input data. The model uses the learned patterns and rules to estimate the network load in a specific time period in the future, such as traffic growth rate, server load rate, error rate and other indicators, and outputs the corresponding prediction results to provide a forward-looking basis for network resource scheduling.

[0065] By accurately predicting future network load, the system can sense potential load pressure in advance, allocate resources and schedule traffic before the network load becomes too high, avoid server failures due to overload, improve the stability and reliability of the entire network system, and ensure the continued normal operation of network services.

[0066] S104: Generate a dynamic load scheduling strategy based on the prediction results and send it to the edge node cluster and regional control center to collaboratively perform resource optimization and traffic scheduling of the service instance.

[0067] The AI prediction model outputs a forecast of future network load, such as a prediction of a significant increase in network traffic in a certain area over the next 10 minutes. Based on this forecast data, the system conducts in-depth analysis, taking into account various factors, such as predicted traffic growth rate, error rate, and other key indicators, as well as the demand and priority of different services for network resources, to generate a real-time dynamic load scheduling strategy.

[0068] The generated dynamic load scheduling policies are sent to the edge node cluster and regional control center. These policies contain detailed scheduling instructions, such as which edge nodes to scale, which service instances to migrate, and how to adjust the traffic distribution ratio between regions. Reliable communication mechanisms, such as the Google Remote Procedure Call (gRPC) protocol, ensure that the policies are accurately and promptly delivered to the edge node cluster and regional control center.

[0069] Among them, edge nodes are responsible for local traffic scheduling and resource adjustment, such as forwarding traffic to low-load nodes in real time, processing local traffic nearby through edge nodes, and reducing transmission delays; the regional control center coordinates cross-node resources from a global perspective, such as migrating hotspot services to neighboring nodes, coordinating resources within the region, and ensuring load balance within the region.

[0070] Based on the above introduction, the dynamic network load balancing method provided in this embodiment collects multi-dimensional operational data from a node cluster deployed at the network edge, constructing a raw data set that comprehensively reflects the network's operational status. This raw data is then deeply analyzed to identify key information features and integrate them into a structured feature vector. This feature vector is then fed into a pre-built AI prediction model to proactively estimate network load conditions for future time periods, generating a prediction result. Based on the prediction result, a targeted dynamic load scheduling strategy is then generated and issued to the edge node cluster and regional control center. The two work together, with the edge node responsible for real-time scheduling of local traffic and flexible resource allocation, while the regional control center coordinates cross-node resource migration from a global perspective, jointly achieving dynamic network load balancing. By predicting network load changes in advance, this method can complete resource expansion and path optimization before traffic peaks arrive, thereby avoiding resource waste or service performance degradation caused by uneven server load. Furthermore, the collaborative mechanism between the edge node and the regional control center ensures low-latency processing of local traffic while enabling coordinated allocation of global resources, enhancing system stability and reliability.

[0071] Example 2:

[0072] The above embodiment does not limit how to perform dynamic load scheduling based on different prediction results. In this embodiment, a specific method for generating a load scheduling strategy is proposed. Specifically, generating a dynamic load scheduling strategy based on the prediction results in step S104 may include the following sub-steps:

[0073] Step S41: When the predicted error rate exceeds the error rate threshold, perform emergency response processing;

[0074] Step S42: When the predicted traffic growth rate exceeds the growth rate threshold, pre-capacity expansion processing is performed;

[0075] Step S43: When all the prediction indicators are within the normal range, perform automatic balancing processing;

[0076] When the prediction error rate exceeds the error rate threshold, an emergency response is initiated. A high prediction error rate may mean that the current server is overloaded and cannot handle the requests. At this time, the emergency response module will perform resource allocation optimization operations, such as expanding the capacity of more nodes (node-345 and node-678, etc.) to allow more servers to handle requests; redirecting traffic to other areas to avoid excessive local load; enabling SSL (Secure Sockets Layer) termination function to reduce the burden of encrypted data on the server, allowing the server to use more resources for operations such as processing business requests, so that resource allocation can better meet current needs, reduce error rates, and improve system availability. Optionally, emergency response processing includes: performing a combination of node expansion, traffic redirection, and SSL offloading operations.

[0077] If the predicted traffic growth rate exceeds the growth rate threshold, pre-scaling is triggered. For example, if a significant traffic increase is predicted, this triggers pre-allocation of resources. Based on the predicted traffic growth, resources are added to the relevant nodes or more node replicas are prepared. This ensures that the system has sufficient resources to cope with traffic peaks, avoiding server lags or crashes caused by insufficient resources and optimizing resource allocation in advance to accommodate the upcoming high load. Optionally, pre-scaling includes performing elastic resource allocation based on the predicted growth rate.

[0078] When all predicted indicators are within normal ranges, the system starts optimizing resource allocation by comprehensively considering various factors. For example, it dynamically adjusts the load balancing algorithm based on different time periods (time and space), service requirements (service), network conditions (network), and business characteristics (business). It selects the most appropriate algorithm, such as round-robin, weighted, or adaptive scheduling, to ensure a more even and reasonable distribution of resources across nodes, regions, and services, thereby improving overall resource utilization.

[0079] It should be noted that emergency response processing, pre-capacity expansion processing, and automatic balancing processing will adopt different methods to perform resource allocation optimization operations that match the corresponding prediction indicators based on their corresponding prediction indicators, making resource utilization more efficient and reasonable. This embodiment does not limit the specific methods of emergency response processing, pre-capacity expansion processing, and automatic balancing processing; this embodiment only uses the above operations as examples.

[0080] The method for generating dynamic load scheduling strategies provided in this embodiment divides strategies into three types: emergency response, pre-capacity expansion, and automatic balancing. On the one hand, this method implements differentiated strategies for different risk levels to avoid the problem of resource waste or insufficient response caused by a "one-size-fits-all" approach, and standardizes strategy execution to reduce the complexity of decision-making. On the other hand, this method allocates resources on demand. For example, emergency response only triggers high-cost operations during crises, and pre-capacity expansion is based on predicted precise traffic demand. This can reduce static resource redundancy, reduce operating costs, and improve resource utilization. It converts the "foresight" of AI predictions into "operational hierarchical response capabilities," enabling the network load balancing system to respond to extreme emergencies while efficiently handling daily traffic fluctuations, ultimately improving the stability and competitiveness of the overall system.

[0081] Example 3:

[0082] Currently, AI prediction models are deployed locally or in the cloud. Taking cloud deployment as an example, if all data is uploaded to the cloud for processing, if it is applied to a large-scale distributed network, it will generate huge bandwidth pressure and transmission delays.

[0083] In order to adapt to the load balancing requirements of large-scale distributed networks, improve response speed, and optimize resource efficiency, this embodiment proposes a deployment method for an AI prediction model, which adopts a cloud-edge collaborative deployment architecture and realizes efficient and accurate network load prediction through the division of labor and collaboration between the cloud global model and the edge local model.

[0084] Specifically, a global model is deployed in the cloud. This model, built on an LSTM (Long Short-Term Memory Network) architecture, is used to generate initial model parameters. This model is trained using full historical data, such as network-wide traffic and multi-regional business models, to learn macroscopic patterns in network load fluctuations and generate initial model parameters (such as the neural network weight matrix). These parameters serve as a "knowledge template" to provide basic predictive capabilities for edge nodes.

[0085] At the same time, a local copy of the global model is deployed on each edge node in the edge node cluster. A lightweight copy of the global model is deployed as the local model. The local model is used to perform local load forecasting. The local model is initialized with initial parameters sent from the cloud and combines local real-time data (such as user behavior and device status in a specific area) to perform localized load forecasting. For example, it can predict short-term traffic fluctuations in a city's business district.

[0086] This cloud-edge collaborative AI prediction architecture, through the "global knowledge sharing + local intelligent execution" model, achieves triple optimization of prediction accuracy, response speed and resource efficiency while ensuring data security.

[0087] In order to protect data privacy and improve the adaptability of the model, based on the deployment architecture of the above-mentioned AI prediction model, this embodiment further proposes a model training method, which fully utilizes the computing resources and local data value of edge nodes while ensuring data privacy, thereby achieving efficient iteration and global optimization of the AI model.

[0088] Specifically, the training process of the AI prediction model can be performed as follows:

[0089] 1. Each edge node independently trains its local model based on local data to generate gradient parameters;

[0090] Each edge node independently trains its local model based on locally collected data and calculates the gradient of the model parameters, which is the direction of model optimization. For example, a node in City A discovers that local traffic has a special fluctuation pattern at 8 pm.

[0091] 2. Receive and aggregate the gradient parameters of each edge node to generate a global gradient update;

[0092] Edge nodes upload the calculated gradient parameters to the cloud, rather than the raw data. The cloud server then performs a weighted average of the gradients from all edge nodes to generate a global gradient update representing the global optimization direction.

[0093] 3. Adjust the parameters of the global model according to the global gradient update amount, and extract the adjusted global model parameters;

[0094] The cloud adjusts the parameters of the global model based on the global gradient update, for example, strengthening the ability to identify traffic patterns at 8 pm. The adjusted model parameters become the new "knowledge template."

[0095] 4. Send the global model parameters to each edge node so that the parameters of each local model can be synchronized according to the global model parameters.

[0096] The cloud sends the updated global model parameters to each edge node, which then replaces the corresponding parts of its local model with these parameters, achieving model synchronization. At this point, the models of all edge nodes have "learned" the global traffic patterns while retaining the unique characteristics of local data.

[0097] Based on the above training method, the original data (such as user location and business content) always remains locally on the edge node, and the cloud only obtains gradient information, thus achieving data privacy protection for the edge node; at the same time, local training of the edge node can capture regional characteristics and improve the adaptability of the model. Moreover, the amount of data of the gradient parameter is much smaller than the original data, which can significantly reduce communication costs.

[0098] To further enhance privacy protection, before receiving and aggregating the gradient parameters from each edge node in step 2, each edge node can further encrypt the corresponding gradient parameters to obtain encrypted gradient parameters. After generating the gradient parameters, each edge node first uses an encryption algorithm to mask the gradients and convert them into ciphertext. The cloud receives only this ciphertext data and directly performs aggregation calculations in ciphertext form, without decryption. Finally, the cloud updates the global model parameters based on the aggregated encrypted results, ensuring that the gradient data remains confidential during transmission and processing. Receiving and aggregating the gradient parameters from each edge node specifically involves receiving and aggregating the encrypted gradient parameters via a secure communication protocol within the federated learning framework. This encrypted gradient aggregation technique provides the dual guarantees of a "mathematical lock" and a "secure channel," completely addressing data privacy issues in federated learning without sacrificing model performance. Of course, depending on the needs of different application scenarios, gradient parameters can also be left unencrypted, which is not a limitation here.

[0099] Example 4:

[0100] The above embodiments do not limit the specific algorithm used to extract features from the multi-dimensional operational data collected by edge node clusters to improve the degree to which features retain key information. This embodiment proposes a feature extraction method that transforms complex multi-dimensional operational data into feature representations with spatiotemporal associations and business semantics by constructing a four-dimensional coordinate system and a spatiotemporal encoder. Specifically, the multi-dimensional operational data can be mapped to a four-dimensional coordinate system of spatiotemporal-service-network-business to generate coordinate data in four-dimensional space; the spatiotemporal encoder then extracts feature associations based on the coordinate data.

[0101] The collected multi-dimensional operational data is mapped into a four-dimensional coordinate system: space-time-service-network-business. The space-time dimension covers user latitude and longitude, time zone information, and traffic patterns derived from a 5-minute sliding window analysis. The service dimension includes the dependencies of the microservice call chain and the database connection pool status. The network dimension involves the bandwidth utilization of the eth0 (Ethernet interface 0), network packet loss rate, and BGP routing table change frequency collected via the SNMP protocol (Simple Network Management Protocol). The business dimension includes user device type, user session history, and purchase conversion rate. This allows data to be sorted and categorized from different perspectives, providing a clear framework for feature extraction.

[0102] The data mapped to the four-dimensional coordinate system is processed using a spatiotemporal encoder. The spatiotemporal encoder consists of a one-dimensional convolutional layer, an LSTM layer, and a Transformer layer. The one-dimensional convolutional layer maps the 15-dimensional input features to a 64-dimensional feature space, capturing local patterns using a convolution kernel of size 5. The LSTM layer processes the convolved sequence data, extracts temporal dependencies, and outputs a 128-dimensional hidden state. The Transformer layer uses a multi-head self-attention mechanism to capture long-range dependencies in the sequence, with a model dimension of 256. Through this series of operations, the spatiotemporal encoder can simultaneously capture local features, short-term temporal dependencies, and long-term dependencies in the data, thereby extracting feature correlations between the data.

[0103] The features extracted by the spatiotemporal encoder are combined according to a specific order and rules. The feature values obtained after each processing step are arranged in an orderly manner to form a fixed-length vector, the feature vector. This feature vector integrates key information from the multidimensional operational data and serves as input to the subsequent AI prediction model, providing data support for predicting network load in future time periods and enabling the model to perform more accurate analysis and predictions based on these features.

[0104] The feature extraction method provided in this embodiment, through the combination of a four-dimensional coordinate system + a spatiotemporal encoder, enables the feature vector to contain richer semantic information and improves the feature expression capability; at the same time, by explicitly modeling spatiotemporal associations, the system can more accurately identify abnormal patterns, thereby ensuring the accuracy of recognition.

[0105] Embodiment 5:

[0106] In edge computing scenarios, the network environment is complex, nodes are widely distributed and highly heterogeneous (such as edge servers, fog devices, and terminal nodes), and business needs are diversified. In order to improve the accuracy and adaptability of scheduling strategies, this embodiment proposes a method for generating dynamic load scheduling strategies based on prediction results to achieve data-driven intelligent decision-making.

[0107] Specifically, you can generate a dynamic load scheduling strategy by following these steps:

[0108] 1. Construct a multi-dimensional routing decision matrix;

[0109] Specifically, a five-level routing decision matrix can be constructed, encompassing four dimensions: geographic proximity, link quality, service load, and security level. Geographic proximity is used to measure the physical distance between edge nodes and users / data sources (e.g., longitude and latitude differences). The closer the distance, the lower the latency. Specifically, the Haversine distance calculation method can be used. Link quality is used to monitor network link bandwidth utilization, packet loss rate, round-trip time (RTT), and other indicators in real time. End-to-end latency measurements can be performed based on ICMP / BFD. Service load reflects the current working status of edge nodes, including CPU / memory utilization, task queue length, and request response time. Security levels can be determined based on the node's security certification status, vulnerability risk score, and data encryption level (e.g., levels A / B / C).

[0110] A routing decision matrix includes:

[0111] 1. Geographic proximity: Haversine distance calculation (accuracy ≤ 10m);

[0112] 2. Link quality: end-to-end delay measurement based on ICMP / BFD;

[0113] 3. Node status: comprehensive CPU / memory / disk IO utilization (weight ratio 4:3:2:1);

[0114] 4. Service affinity: Session persistence (session stickiness, timeout = 30 minutes);

[0115] 5. Cost optimization: Prioritize scheduling to nodes with purchased GPU (Graphics Processing Unit) resources.

[0116] In this embodiment, only the above routing decision matrix is taken as an example. Other matrix forms constructed according to the requirements of actual application scenarios can refer to the introduction of this embodiment and will not be repeated here.

[0117] 2. Map the state parameters of each edge node in the edge node cluster to the corresponding dimension of the routing decision matrix to generate dimension indicators;

[0118] The edge node agent collects status parameters such as geographic location, network probe data, service monitoring data, and security scan results in real time, normalizes the raw data of different dimensions, and ensures the comparability of indicators in each dimension.

[0119] 3. Dynamically adjust the weight coefficients of each dimension in the routing decision matrix based on the prediction results;

[0120] Among them, the dynamic adjustment logic of each weight needs to be set accordingly according to the weight type. For example, when the prediction results show that the risk of network attack is increased, the security level weight is increased from 20% to 40%, and high-security level nodes are given priority; when the traffic is predicted to grow explosively, the link quality weight is increased from 30% to 50% to ensure data transmission stability; in daily load balancing scenarios, the default weights (geographic proximity 30%, link quality 30%, service load 25%, security level 15%) can be used.

[0121] 4. Calculate a comprehensive score for each candidate path based on the indicators and weight coefficients of each dimension;

[0122] Based on the network topology of edge nodes, all reachable paths are generated, and a comprehensive score is calculated for each candidate path. The weighted sum of each dimension is calculated, or other scoring methods are adopted to obtain the score of each path.

[0123] 5. Select the candidate path with the highest comprehensive score as the optimal route and generate a dynamic load scheduling strategy including the optimal path.

[0124] Based on the above introduction, the method for generating a dynamic load scheduling strategy provided in this embodiment can flexibly switch between multiple objectives such as "low latency", "high reliability" and "strong security" through dynamic adjustment of weights to achieve multi-objective optimization. It can also flexibly adjust priorities according to prediction results to achieve data-driven intelligent decision-making.

[0125] Example 6:

[0126] Further references Figure 2 , which shows an exemplary structural block diagram of a dynamic network load balancing system according to an embodiment of the present application, which mainly includes: an edge node cluster, a regional control center, and a cloud collaboration center;

[0127] Among them, the edge node cluster is deployed at the edge of the network close to the user or data source (such as urban base stations, enterprise branches), and includes several edge nodes. Each edge node is used to collect multi-dimensional operation data of its own node, such as user location, device status, network traffic, service response time, etc.

[0128] The regional control center aggregates data from edge nodes within a region, such as all edge nodes in a province or city, performs feature engineering, and converts raw data into feature vectors, alleviating computing pressure on the cloud. Specifically, the regional control center collects multi-dimensional operational data from the edge node cluster, extracts features from this multi-dimensional operational data, combines the extracted features to generate feature vectors, and inputs these feature vectors into the AI prediction module.

[0129] The cloud collaboration center is deployed with an AI prediction module and a dynamic scheduling engine. The AI prediction module uses a global historical data training model to predict the network load in the future time period based on the feature vector and generate prediction results. The dynamic scheduling engine is used to obtain the prediction results, generate a dynamic load scheduling strategy based on the prediction results, and send it to the edge node cluster and regional control center to coordinate the resource optimization and traffic scheduling of service instances.

[0130] The overall data flow path is: edge node (data collection) → regional control center (feature extraction) → cloud (prediction + strategy generation) → region / edge (execution scheduling).

[0131] The system provided in this embodiment adopts a three-level distributed architecture of "cloud-edge-end". The edge layer can process data locally, reduce transmission delays, and support offline autonomy; the cloud integrates multi-region information based on a global data perspective to improve prediction accuracy.

[0132] Embodiment seven:

[0133] Based on the system architecture provided by the above embodiment, in order to meet the real-time requirements under dynamic network changes, smart probes, local schedulers and edge AI engines can be further built in the edge nodes as core components to form a closed loop of "data collection-local intelligence-dynamic execution".

[0134] Among them, the intelligent probe serves as the data entry point and is used to collect multi-dimensional operating data of edge nodes in real time;

[0135] The edge AI engine performs lightweight processing on the original data, mainly used to generate lightweight feature vectors based on multi-dimensional operation data, and input the lightweight feature vectors into the local model of the AI prediction model to predict the local load and generate local prediction results.

[0136] The local scheduler serves as the execution hub, integrating the cloud's global scheduling strategies (such as cross-regional resource migration) with local prediction results. It is mainly used to combine the dynamic load scheduling strategies generated by the cloud with local prediction results to perform container scaling and / or dynamic routing adjustments, dynamically adjust the resource allocation of edge nodes (such as Kubernetes container scaling) and traffic paths (such as switching routes through the Envoy proxy), and achieve "global optimization + local emergency" collaborative scheduling.

[0137] The edge node structure provided in this embodiment not only retains the cloud's global optimization capability for network load, but also gives the edge node the autonomous response speed to deal with sudden changes. The cloud strategy is responsible for cross-regional resource balancing (such as migrating hotspot services to adjacent edge nodes), and local scheduling focuses on real-time micro-adjustments. Through global-local collaborative optimization, resource utilization can be significantly improved and response speed can be improved.

[0138] Embodiment 8:

[0139] In order to deepen the understanding of the architecture deployment and functions of the dynamic network load balancing system, this embodiment introduces the core components under a distributed architecture, such as Figure 3 shown.

[0140] The edge computing layer deploys over 300 edge nodes, equipped with NVIDIA T4 GPUs (16GB of video memory) and Intel Snow Lake CPUs (40 cores). Each node runs the lightweight EdgeOS operating system (containerized deployment, image size <80MB). The core components of the edge computing layer include intelligent probes, a local scheduler, and an edge AI engine.

[0141] Smart Probe: Implements kernel-level traffic sampling (programmable network interface) based on eBPF, collects four-dimensional data on time, space, service, network, and business (such as user location, microservice call chain, and bandwidth utilization), and senses edge environment status in real time, providing raw data for prediction.

[0142] Local Scheduler: Integrates the Envoy proxy and Kubernetes controller, primarily used to perform autonomous edge scheduling (e.g., responding to traffic bursts within 50ms) and coordinate resource allocation with cloud-based policies. The Envoy proxy is used for dynamic traffic control, and the Kubernetes controller is used for container orchestration.

[0143] Edge AI Engine: Deploys a TensorRT-accelerated inference model (FP16 precision), mainly used for local prediction of future loads and support emergency scheduling decisions.

[0144] The regional control layer has three regional control centers (Beijing, Shanghai, and Guangzhou), equipped with high-performance computing clusters (AWS EC2p5.4xlarge instances), a model training platform (PyTorchDistributed training framework), and a policy configuration center (Consul service discovery + Etcd distributed lock). The high-performance computing cluster is used to process data from multiple edge nodes and perform deep feature extraction; the model training platform is used to train regional prediction models and optimize edge node parameters; and the policy configuration center is used to manage the status of edge nodes within the region and coordinate the issuance of global scheduling policies.

[0145] The cloud collaboration layer deploys a global AI training cluster (with Kilobyte V100 GPU resources), a model repository (supporting version rollbacks and A / B testing), and a monitoring dashboard (Grafana + Prometheus real-time visualization). The global AI training cluster trains global models based on full data; the model repository stores and distributes optimized model parameters to edge nodes; and the monitoring dashboard displays real-time network load status, prediction accuracy, and resource utilization.

[0146] This architecture achieves a balance between low-latency response and high-accuracy prediction through a three-level system of edge real-time perception, regional collaborative optimization, and cloud-based global decision-making, supporting intelligent scheduling of millions of concurrent requests.

[0147] It should be understood that the units described in the above system and the reference Figure 1 The steps in the method described above correspond to each other. Therefore, the operations and features described above for the method are also applicable to the device and the units contained therein, and will not be repeated here. The system can be pre-implemented in the browser or other security application of the server, or loaded into the browser or its security application of the server by downloading or other means. The corresponding units in the system can cooperate with the units in the server to implement the solutions of the embodiments of the present application.

[0148] The division of the several units mentioned in the detailed description above is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided into multiple units to be embodied.

[0149] It should be noted that for details not disclosed in the dynamic network load balancing system of the embodiments of the present application, please refer to the details disclosed in the above embodiments of the present application, and no further details will be given here.

[0150] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operating instructions of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the aforementioned module, program segment, or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than the order marked in the accompanying drawings. For example, the boxes represented by two connections can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of the boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operating instruction, or can be implemented using a combination of dedicated hardware and computer instructions.

[0151] The units or modules involved in the embodiments described in this application may be implemented in software or hardware. The units or modules described may also be provided in a processor. The names of these units or modules do not, in certain circumstances, constitute limitations on the units or modules themselves.

[0152] The above description is merely a preferred embodiment of the present application and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the aforementioned disclosed concepts. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A dynamic network load balancing method, characterized in that: include: Collect multi-dimensional operation data through edge node clusters; Extracting features from the multidimensional operating data, and combining the extracted features to obtain a feature vector; Input the feature vector into the AI prediction model to predict the network load in the future time period and generate a prediction result; A dynamic load scheduling strategy is generated based on the prediction results and sent to the edge node cluster and the regional control center to collaboratively perform resource optimization and traffic scheduling of the service instance.

2. The method according to claim 1, wherein Generating a dynamic load scheduling strategy according to the prediction result includes: When the prediction error rate exceeds the error rate threshold, emergency response processing is performed; When the predicted traffic growth rate exceeds the growth rate threshold, pre-capacity expansion processing is performed; When all the prediction indicators are within the normal range, automatic balancing processing is performed; The emergency response process, the pre-capacity expansion process, and the automatic balancing process respectively perform resource allocation optimization operations that match the prediction index.

3. The method according to claim 2, wherein The emergency response process includes: performing a combination of node expansion, traffic redirection, and SSL offloading; The pre-capacity expansion process includes: performing an elastic resource allocation operation based on a predicted growth rate.

4. The method according to claim 1, wherein The deployment method of the AI prediction model includes: Deploy a global model on the cloud, where the global model is built based on an LSTM network structure to generate initial model parameters; A local copy of the global model is deployed on each edge node in the edge node cluster as a local model, and the local model is used to perform local load prediction.

5. The method according to claim 4, wherein The training method of the AI prediction model includes: Each edge node independently trains its local model based on local data to generate gradient parameters; Receive and aggregate the gradient parameters of each edge node to generate a global gradient update; Adjusting the parameters of the global model according to the global gradient update amount, and extracting the adjusted global model parameters; The global model parameters are sent to each of the edge nodes so as to synchronize the parameters of each of the local models according to the global model parameters.

6. The method according to claim 5, wherein Before receiving and aggregating the gradient parameters of the edge nodes, the method further includes: Encrypting the corresponding gradient parameter by each edge node to obtain an encrypted gradient parameter; Then, the gradient parameters of each edge node are received and aggregated, including: based on a federated learning framework, receiving and aggregating each encrypted gradient parameter through a secure communication protocol.

7. The method according to claim 1, wherein Extracting features from the multidimensional operating data includes: Mapping the multi-dimensional operation data to a four-dimensional coordinate system of time-space-service-network-business to generate coordinate data in four-dimensional space; A feature association relationship is extracted based on the coordinate data through a spatiotemporal encoder.

8. The method according to claim 1, wherein Generating a dynamic load scheduling strategy according to the prediction result, including: Construct a multi-dimensional routing decision matrix; Mapping the state parameter of each edge node in the edge node cluster to a corresponding dimension of the routing decision matrix to generate a dimension index; Dynamically adjusting the weight coefficients of each dimension in the routing decision matrix according to the prediction results; Calculate a comprehensive score for each candidate path based on the dimensional indicators and the weight coefficients; The candidate path with the highest comprehensive score is selected as the optimal route, and a dynamic load scheduling strategy including the optimal path is generated.

9. A dynamic network load balancing system, characterized in that: include: Edge node clusters, regional control centers, and cloud collaboration centers; The edge node cluster includes several edge nodes, each of which is used to collect multi-dimensional operation data of its own node; The regional control center is used to collect the multi-dimensional operation data of each edge node cluster, perform feature extraction on the multi-dimensional operation data, combine the extracted features to obtain a feature vector; and input the feature vector into the AI prediction module; The cloud-based collaborative center is deployed with an AI prediction module and a dynamic scheduling engine; the AI prediction module is used to predict the network load in the future time period based on the feature vector and generate a prediction result; the dynamic scheduling engine is used to obtain the prediction result, generate a dynamic load scheduling strategy based on the prediction result, and send it to the edge node cluster and the regional control center to collaboratively perform resource optimization and traffic scheduling of the service instance.

10. The system according to claim 9, wherein Each of the edge nodes includes: an intelligent probe, a local scheduler, and an edge AI engine; Wherein, the intelligent probe is used to collect multi-dimensional operation data; The edge AI engine is used to generate a lightweight feature vector based on the multi-dimensional operation data, and input the lightweight feature vector into the local model of the AI prediction model to predict the local load and generate a local prediction result; The local scheduler is used to combine the dynamic load scheduling strategy generated in the cloud with the local prediction result to perform container expansion and contraction and / or dynamic routing adjustment.

Images