Distributed server load balancing system and method based on multi-modal data fusion

Abstract

The invention discloses a distributed server load balancing system based on multi-modal data fusion. The distributed server load balancing system comprises a multi-modal sensing layer, a fusion analysis layer, an intelligent decision-making layer and an elastic execution layer, the multi-modal sensing layer collects multi-source heterogeneous data; the fusion analysis layer is used for mining an association relationship between data of different dimensions; meanwhile, converting multi-source data into a unified feature vector, and outputting a dynamic load balancing strategy; the intelligent decision-making layer generates an executable load balancing strategy according to the fusion analysis result; and the elastic execution layer converts the strategy generated by the intelligent decision-making layer into an actual action, and deals with load change through an elastic mechanism. According to the method, server performance, network topology and request semantic data are fused, a unified feature tensor is generated by using a cross-modal attention mechanism, and load association is accurately quantified; on the basis of cooperation of a global mixed integer programming model and a local lightweight DQN agent, load migration is completed in a very short time, and the response speed is increased.

Description

Technical Field

[0001] This invention relates to the field of server technology, and more specifically to a distributed server load balancing system and method based on multimodal data fusion. Background Technology

[0002] Traditional distributed server load balancing systems primarily rely on single-dimensional monitoring data, such as CPU / memory metrics, to trigger load migration or resource scaling through static thresholds. However, this existing approach has drawbacks: data fragmentation, with multi-source heterogeneous data such as server performance, network status, and request semantics processed independently, failing to capture cross-dimensional correlations, such as the coupling effect between high-concurrency requests and network congestion. Summary of the Invention

[0003] To address these issues, this invention provides a distributed server load balancing system and method based on multimodal data fusion, thereby resolving the problems in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A distributed server load balancing system based on multimodal data fusion includes a multimodal perception layer, a fusion analysis layer, an intelligent decision-making layer, an elastic execution layer, and a server cluster.

[0006] The multimodal perception layer collects heterogeneous data from multiple sources across three dimensions: server, network, and user.

[0007] The fusion analysis layer is used to uncover the correlation between data from different dimensions; at the same time, it transforms multi-source data into a unified feature vector and outputs a dynamic load balancing strategy.

[0008] Based on the results of the fusion analysis, the intelligent decision-making layer generates an executable load balancing strategy and optimizes the graph neural network module and temporal convolutional network module built into the fusion analysis layer through feedback iteration.

[0009] The elastic execution layer transforms the strategies generated by the intelligent decision-making layer into actual actions and responds to load changes through elastic mechanisms; the elastic execution layer collects the strategy execution effect in real time and feeds it back to the multimodal perception layer.

[0010] Server clusters distribute massive requests across multiple server nodes through clustered deployment; they can dynamically expand or shrink based on traffic fluctuations while isolating traffic from different services; server clusters can provide diverse data sources for the fusion analytics layer.

[0011] Furthermore, the data types collected by the multimodal perception layer include those from the server side, the network side, and the user side; the multimodal perception layer deploys three types of heterogeneous sensor networks, namely a performance probe cluster, a network topology scanner, and a semantic parsing engine.

[0012] Furthermore: the performance probe cluster can collect dynamic indicators of server nodes in real time, including traditional resource indicators, hardware status indicators, and application-level characteristic indicators.

[0013] The network topology scanner can construct a dynamic network state map and continuously monitor key parameters between nodes, including communication latency, link packet loss rate, and cross-rack bandwidth cost, updating the topology weight matrix every 1-20 seconds.

[0014] The semantic parsing engine can analyze the semantic features of the request content through natural language processing technology, extract key business attributes, and construct a request profile vector.

[0015] Furthermore, the multimodal perception layer can standardize the data, including normalizing data from different modalities; and align multi-source data through timestamps to ensure consistency in subsequent fusion analysis.

[0016] Furthermore: the fusion analysis layer employs a heterogeneous neural network fusion architecture to process multimodal data; the fusion analysis layer includes a graph neural network module, a temporal convolutional network module, and a cross-modal attention mechanism;

[0017] The graph neural network module can process the topological relationship data of server clusters and build a node influence model; it learns the communication patterns between servers through a multi-layer graph convolutional network, outputs multi-dimensional node embedding vectors, and quantitatively evaluates the structural importance of each node in the cluster.

[0018] The Temporal Convolutional Network (TCN) module can analyze the spatiotemporal characteristics of performance metrics, using dilated causal convolution to capture long-period dependencies and predict load trends for the next 20-40 seconds.

[0019] A cross-modal attention mechanism is used to align and fuse heterogeneous data. By calculating the correlation weights of performance data, network state, and request semantics through a multi-head attention layer, a unified feature tensor is generated.

[0020] Furthermore: the intelligent decision-making layer adopts a two-stage collaborative decision-making mechanism to achieve a balance between global optimization and local adjustment, that is, it adopts a global optimizer and a local optimizer;

[0021] The global optimizer constructs a multi-objective optimization model based on mixed-integer programming. The objective function is to minimize {α × maximum server load + β × network cost + γ × SLA default rate}. Here, maximum server load is the load value of the server node with the highest load in the cluster; network cost is the network overhead incurred by cross-node communication; and the SLA default rate is the proportion of service quality agreement (SLA) defaults. The dynamic weight coefficients α, β, and γ are automatically adjusted every n minutes through deep reinforcement learning. When the system load is unbalanced, the value of α is automatically increased; when network congestion is severe, the value of β is automatically increased; and when service quality deteriorates, the value of γ is automatically increased.

[0022] The local adjuster deploys lightweight DQN agents on each server node to respond to local state changes in real time. The agent evaluates the long-term benefits of migration, scaling, and unloading operations through an action value function, achieving local optimality while satisfying global constraints. When a load deviation >15% is detected, the agent autonomously performs load migration or elastic resource allocation within 50-200ms.

[0023] Furthermore: the global optimizer and local adjuster decisions are dynamically coordinated through a two-layer feedback control loop, the specific implementation process of which is as follows:

[0024] Uplink synchronization: The local adjuster agent uploads a state summary to the global optimizer every 5 seconds. The state summary includes key indicators such as load gradient, decision distribution, and constraint satisfaction rate.

[0025] Downlink correction: The global optimizer analyzes the cluster state and sends policy correction vectors to the local adjuster agents to adjust their reward function parameters;

[0026] When a node load deviation > 25% or an SLA default rate > 5% is detected, the global optimizer layer directly takes over control and implements a forced load migration.

[0027] Furthermore: the policy distribution of the elastic execution layer is achieved by pushing policy parameters to the load balancer via API; and by calling the cloud platform API to realize elastic scaling of resources.

[0028] Furthermore: the elastic execution layer achieves fine-grained control through a programmable data plane; the elastic execution layer includes a heterogeneous resource abstraction engine, a dynamic weight adjuster, and a feedback control loop; the heterogeneous resource abstraction engine unifies heterogeneous computing resources into a computing unit matrix, supporting task-oriented distribution based on hardware characteristics; the dynamic weight adjuster modifies the load balancing weight in real time according to the decision layer's instructions, with a response latency of ≤10ms; the feedback control loop continuously optimizes decisions by monitoring execution results, using a PID controller to dynamically adjust resource allocation parameters.

[0029] To achieve the above objectives, the present invention also provides a distributed server load balancing method based on multimodal data fusion, comprising the following steps:

[0030] Step 1: System initialization and configuration;

[0031] The multimodal perception layer is deployed by deploying a network topology scanner, configuring a semantic parsing engine, and deploying a performance probe cluster. The fusion analysis layer parameters are initialized by initializing node features and edge weights of the graph neural network and constructing a graph structure based on the initial topology matrix. The kernel size and time step of the temporal convolutional network are configured to adapt to the temporal characteristics of system performance data. Weight parameters for the cross-modal attention mechanism are set, and the fusion priority of different modal data is defined. The intelligent decision layer loads a pre-trained load balancing strategy model and initializes the strategy correction vector. Rules for generating state summaries are defined. The elastic execution layer connects to the feedback loop, configuring the trigger threshold for the local adjuster. The scheduling blueprint for the global optimizer is defined. The heterogeneous resource abstraction engine is connected to unify the resource representation across different server architectures.

[0032] Step 2: Multimodal data acquisition and preprocessing; acquire data from the server layer, network layer, and application layer; perform two-level parsing of client requests using a semantic parsing engine; first-level parsing extracts the request type and target service; second-level parsing matches key parameters using regular expressions;

[0033] Step 3: Multimodal data fusion; The graph neural network model processes topological data, inputting the topological weight matrix into the graph neural network model to generate node embeddings through graph convolution operations; The temporal convolutional network model processes temporal data, inputting the temporal data collected by the performance probe into the temporal convolutional network model to extract short-term and long-term temporal features; Cross-modal attention fusion, inputting the node embeddings generated by the graph neural network model, the temporal features extracted by the temporal convolutional network model, and the semantic parsing results into the cross-modal attention layer; Calculate the attention weights for each modality; Generate a unified feature tensor;

[0034] Step 4: Intelligent Decision Making and Strategy Generation; State Summary Generation: Input the unified feature tensor into the state summary generator and compress it into a key indicator vector; Strategy Correction Vector Calculation: Based on the current state summary and historical strategies, generate correction vectors through the strategy network; Combine the correction vectors with preset rules to generate a load balancing strategy; Local adjustment strategy is to migrate non-core tasks of high-load servers; Global optimization strategy is to dynamically adjust the cluster resource pool allocation ratio.

[0035] Step 5: Flexible execution and dynamic adjustment;

[0036] When the local adjuster executes, it triggers task migration when the CPU utilization of a certain server exceeds 85%.

[0037] Filter out transferable tasks; encapsulate the tasks into a unified format using a heterogeneous resource abstraction engine and send them to a low-load server;

[0038] The global optimizer executes and schedules operations every n seconds based on the cluster resource pool status.

[0039] The feedback control loop is adjusted, and the dynamic weight adjuster updates the topology weight matrix based on the execution results.

[0040] The present invention has the following advantages: The present invention integrates server performance, network topology, and request semantic data, and uses a cross-modal attention mechanism to generate a unified feature tensor to accurately quantify load correlation; based on the collaboration of a global mixed integer programming model and a local lightweight DQN proxy, load migration is completed in a very short time, and the response speed is improved.

[0041] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0042] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0043] Figure 1 This is a flowchart illustrating the implementation of a distributed server load balancing system based on multimodal data fusion, as provided in one embodiment of this application. Detailed Implementation

[0044] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Technical engineers in the field can make some non-essential improvements and adjustments to the present invention based on the above-described content. 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.

[0045] Please see Figure 1A distributed server load balancing system based on multimodal data fusion includes a multimodal perception layer, a fusion analysis layer, an intelligent decision-making layer, an elastic execution layer, and a server cluster.

[0046] The multimodal perception layer collects heterogeneous data from multiple sources across three dimensions: servers, networks, and users, providing a foundation for subsequent analysis.

[0047] The data types collected by the multimodal perception layer include those from the server side, network side, and user side. Server-side data includes hardware resource utilization such as CPU, memory, and disk I / O; network-side data includes transmission quality indicators such as inter-node bandwidth, latency, and packet loss rate; and user-side data includes behavioral data such as request type (API / page), concurrency, and response time.

[0048] The multimodal perception layer is the data acquisition front end of the load balancing system of this invention. The multimodal perception layer deploys three types of heterogeneous sensor networks, namely performance probe clusters, network topology scanners, and semantic parsing engines.

[0049] The performance probe cluster can collect more than 200 dynamic metrics of server nodes in real time, including traditional resource metrics (CPU utilization, memory pressure, I / O latency, network bandwidth, etc.), hardware status metrics (GPU memory utilization, NPU inference latency, disk health status, etc.), and application-level feature metrics (request processing latency, microservice call chain topology, database connection pool saturation, etc.).

[0050] The network topology scanner can construct a dynamic network state map and continuously monitor key parameters such as inter-node communication latency (RTT≤1ms), link packet loss rate (≤0.01%), and cross-rack bandwidth cost (10Gbps / 1Gbps mixed environment), updating the topology weight matrix every 1-20 seconds.

[0051] The semantic parsing engine can analyze the semantic features of the request content through natural language processing technology, extract key business attributes (such as the resolution requirements of video transcoding requests, the model type of AI inference tasks, and the SLA level of real-time transactions), and construct a request profile vector (128-dimensional feature encoding).

[0052] The multimodal perception layer can standardize data, including normalizing data from different modalities (e.g., mapping CPU utilization to the 0-1 range) to solve the problem of dimensional differences; and aligning multi-source data with timestamps (e.g., 1-second granularity) to ensure consistency in subsequent fusion analysis.

[0053] The fusion analysis layer is used to uncover the correlation between data from different dimensions (such as the correlation between high concurrency requests and server load); at the same time, it transforms multi-source data into a unified feature vector and outputs a dynamic load balancing strategy.

[0054] The fusion analysis layer uses a heterogeneous neural network fusion architecture to process multimodal data; the fusion analysis layer includes a graph neural network module, a temporal convolutional network module, and a cross-modal attention mechanism.

[0055] Graph Neural Network (GNN) module: processes topological relationship data of server clusters and builds a node influence model; learns the communication patterns between servers through a 3-layer Graph Convolutional Network (GCN) and outputs a 32-dimensional node embedding vector to quantitatively evaluate the structural importance (Centrality Score) of each node in the cluster.

[0056] Temporal Convolutional Network (TCN) module: Analyzes the spatiotemporal characteristics of performance metrics, uses dilated causal convolution to capture long-period dependencies, and predicts load trends for the next 20-40 seconds. In this embodiment, Dilation = 8.

[0057] Cross-modal attention mechanism: to achieve alignment and fusion of heterogeneous data, and to generate a unified feature tensor by calculating the correlation weights of performance data, network state and request semantics through a multi-head attention layer.

[0058] Based on the results of the fusion analysis, the intelligent decision-making layer generates an executable load balancing strategy and optimizes the model through feedback iteration.

[0059] The intelligent decision-making layer adopts a two-stage collaborative decision-making mechanism to achieve a balance between global optimization and local adjustment, namely, using a global optimizer and a local optimizer.

[0060] Global Optimizer: A multi-objective optimization model is constructed based on Mixed Integer Programming (MIP). The objective function is to minimize {α × maximum server load + β × network cost + γ × SLA default rate}. Where, maximum server load is the load value of the server node with the highest load in the cluster; network cost is the network overhead generated by cross-node communication; SLA default rate is the proportion of service quality agreement defaults; the dynamic weight coefficients (α, β, γ) are automatically adjusted every 5 minutes through deep reinforcement learning (DRL); when the system load is unbalanced, the value of α is automatically increased; when network congestion is severe, the value of β is automatically increased; when service quality degrades, the value of γ is automatically increased.

[0061] Local adjuster: Deploy a lightweight DQN agent (model size ≤ 2MB) on each server node to respond to local state changes in real time; the agent evaluates the long-term benefits of operations such as migration, scaling, and unloading through the action value function (Q-value) to achieve local optima while satisfying global constraints; when a load deviation > 15% is detected, it autonomously performs load migration or elastic resource allocation within 100ms.

[0062] The global optimizer and the local adjuster make decisions through a two-layer feedback control loop, and the specific implementation process is as follows:

[0063] Uplink synchronization: The local adjuster agent uploads a state summary (compression rate of 80%) to the global optimizer every 5 seconds, including key indicators such as load gradient, decision distribution, and constraint satisfaction rate.

[0064] Downlink correction: The global optimizer analyzes the cluster state and sends a policy correction vector (16-dimensional) to the local adjuster agent to adjust its reward function parameters.

[0065] Emergency Intervention: When a node load deviation > 25% or an SLA default rate > 5% is detected, the global optimizer layer directly takes over control and implements forced load migration.

[0066] The elastic execution layer translates the strategies generated by the intelligent decision-making layer into actual actions and responds to load changes through elastic mechanisms.

[0067] Policy delivery is implemented by pushing policy parameters (such as weights and routing rules) to the load balancer (such as dynamic Nginx configuration) via API; and by calling cloud platform APIs to achieve elastic scaling of resources (such as second-level startup / stop of Kubernetes containers).

[0068] The elastic execution layer collects the strategy execution results (such as throughput and error rate) in real time and feeds them back to the multimodal perception layer.

[0069] The elasticity mechanism of the elastic execution layer includes horizontal scaling and vertical scaling.

[0070] Horizontal scaling automatically increases / decreases server instances based on load (e.g., scaling up a database cluster during a promotional period).

[0071] Vertical adjustment involves dynamically allocating resources (e.g., prioritizing high-priority requests by scheduling them to low-load nodes).

[0072] For example, after receiving the "expand database" instruction, the execution layer starts two new instances via Kubernetes and offloads 20% of the requests to the cache server.

[0073] The elastic execution layer achieves fine-grained control through a programmable data plane; the elastic execution layer includes a heterogeneous resource abstraction engine, a dynamic weight adjuster, and a feedback control loop.

[0074] The heterogeneous resource abstraction engine unifies heterogeneous computing resources such as CPU, GPU, and FPGA into a computing unit matrix, supporting task-oriented distribution based on hardware characteristics.

[0075] The dynamic weight adjuster modifies the load balancing weights in real time according to the instructions of the decision-making layer, with a response latency of ≤10ms; it supports 7 routing strategies (including the improved WRR-Resource-Aware weighted round-robin).

[0076] The feedback control loop continuously optimizes decisions by monitoring execution results and uses a PID controller to dynamically adjust resource allocation parameters.

[0077] Server clusters distribute massive requests across multiple server nodes through clustered deployment, avoiding single-point overload and improving overall throughput; they can dynamically expand or shrink based on traffic fluctuations while isolating traffic from different services.

[0078] In addition, server clusters can provide diverse data sources for the fusion analytics layer.

[0079] Server-side data consists of performance metrics such as CPU and memory directly reported by cluster nodes; network-side data consists of inter-cluster network traffic captured by sFlow / NetFlow probes, used to analyze bandwidth bottlenecks; user-side data consists of user behavior logs (such as API call frequency) recorded when the cluster processes requests, providing data for the intelligent decision-making layer to optimize strategies.

[0080] Cluster nodes act as data acquisition units, reporting performance data in real time through agents (such as Prometheus Exporter); server clusters achieve elastic resource scheduling through container orchestration (such as Kubernetes).

[0081] The distributed server load balancing method based on multimodal data fusion includes the following steps:

[0082] Step 1: System Initialization and Configuration

[0083] (1) Deploy the multimodal perception layer, deploy a network topology scanner to scan and generate an initial network topology weight matrix; configure a semantic parsing engine to define semantic parsing rules for client requests (such as HTTP header analysis and API path identification); deploy a performance probe cluster, install probes on each server, and collect metrics such as CPU, memory, and network bandwidth.

[0084] (2) Initialization of parameters for fusion analysis layer

[0085] Initialize the node features and edge weights of the graph neural network (GNN) and construct the graph structure based on the initial topology matrix; configure the kernel size and time step of the temporal convolutional network (TCN) to adapt to the temporal characteristics of the system performance data; set the weight parameters of the cross-modal attention mechanism and define the fusion priority of different modal data (topology, semantics, performance).

[0086] (3) Intelligent decision-making layer strategy loading

[0087] Load the pre-trained load balancing policy model (e.g., a policy network based on reinforcement learning) and initialize the policy correction vector. Define the rules for generating the state summary (e.g., compress the feature tensor into key metrics: server load rate, request queue length).

[0088] (4) Interconnection between the flexible execution layer and the feedback loop

[0089] Configure the trigger threshold for local adjusters (e.g., initiate millisecond-level task migration when CPU utilization > 80%); define the scheduling blueprint for global optimizers (e.g., Kubernetes-based cluster resource pool partitioning rules); connect heterogeneous resource abstraction engines to unify resource representation across different server architectures (x86 / ARM).

[0090] Step 2: Multimodal data acquisition and preprocessing

[0091] (1) Dynamic network topology scanning; The network topology scanner performs a BGP routing table fetch every n seconds and updates the topology weight matrix (edge ​​weight = link delay × bandwidth utilization).

[0092] (2) Request semantic parsing; The semantic parsing engine performs two levels of parsing on the client request; Level 1 parsing extracts the request type (GET / POST) and target service (such as user service, order service); Level 2 parsing matches key parameters (such as user ID, product SKU) through regular expressions.

[0093] (3) Performance data collection; The performance probe cluster collects the following indicators at 100ms intervals: Server layer: CPU utilization, memory remaining, disk I / O wait time; Network layer: Inbound / outbound traffic rate, TCP retransmission rate; Application layer: Request processing latency, error code distribution (such as the proportion of 500 errors).

[0094] Step 3: Multimodal data fusion

[0095] (1) Graph Neural Network (GNN) processes topology data. The topology weight matrix is ​​input into the GNN and node embeddings (server node state representations) are generated through graph convolution operations.

[0096] (2) Temporal convolutional network (TCN) processes time series data. The time series data collected by the performance probe (such as CPU utilization sequence) is input into the TCN to extract short-term and long-term time series features.

[0097] (3) Cross-modal attention fusion: The node embedding generated by GNN, the temporal features extracted by TCN, and the semantic parsing results are input into the cross-modal attention layer; the attention weights of each modality are calculated; and a unified feature tensor is generated (dimension: number of servers × number of fused features).

[0098] Step 4: Intelligent Decision Making and Strategy Generation

[0099] (1) State summary generation: Input the unified feature tensor into the state summary generator and compress it into a key indicator vector:

[0100] (2) Policy correction vector calculation: Based on the current state summary and historical policies, a correction vector is generated through a policy network (such as LSTM);

[0101] (3) The final decision output combines the correction vector with preset rules (such as "prioritize core services") to generate a load balancing strategy; the local adjustment strategy is to migrate non-core tasks of high-load servers; the global optimization strategy is to dynamically adjust the allocation ratio of cluster resource pool.

[0102] Step 5: Flexible Execution and Dynamic Adjustment

[0103] (1) Local adjuster execution: When the CPU utilization of a certain server is >85%, task migration is triggered;

[0104] Filter out transferable tasks (such as stateless image processing tasks); encapsulate the tasks into a unified format through a heterogeneous resource abstraction engine and send them to a low-load server.

[0105] (2) The global optimizer executes Kubernetes scheduling every n seconds based on the cluster resource pool status; when scaling up, when the request queue length is >1000, a backup Pod is started; when scaling down, when the resource utilization is <30% for 5 minutes, the redundant Pod is terminated.

[0106] (3) Feedback control loop adjustment: The dynamic weight adjuster updates the topology weight matrix based on the execution result.

[0107] For example, an e-commerce platform faces the following challenges during the "Double 11" shopping festival:

[0108] Traffic peaked 10 times higher than usual (QPS surged from 50,000 / second to 500,000 / second); core services (order processing, payment) required to maintain low latency (<200ms); the server cluster consisted of 200 x86 servers and 50 ARM servers, with a heterogeneous architecture.

[0109] The system implementation steps are as follows:

[0110] (1) Initialization phase: Deploy a performance probe cluster to focus on monitoring CPU and network metrics of order service and payment service; Configure a semantic parsing engine to prioritize parsing requests containing the / order / create and payment paths.

[0111] (2) Data collection and fusion; The network topology scanner found that the latency of the core link (order service → database) rose to 15ms and updated the topology weight; The performance probe detected that the CPU utilization of the payment service cluster reached 90%, and the time series data was input into the TCN to predict that the load would exceed the threshold in the next 10 seconds.

[0112] (3) Intelligent decision generation; the status summary shows that the payment service cluster load rate is 95%, the request queue length is 800, and the network latency is 25ms; the policy correction vector adjusts the resource allocation priority and decides to migrate the non-core sub-tasks of the payment service (such as log recording) to the ARM server.

[0113] (4) Elastic execution and feedback; the local adjuster immediately migrates log tasks to the ARM server, releasing x86 server resources; the global optimizer starts a backup payment service Pod (an additional 10 x86 servers) to evenly distribute core tasks; the feedback control loop adjusts the topology weights based on the execution results (CPU utilization drops to 75%) to strengthen the priority of core links.

[0114] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A distributed server load balancing system based on multimodal data fusion, characterized in that, It includes a multimodal perception layer, a fusion analysis layer, an intelligent decision-making layer, a flexible execution layer, and a server cluster; The multimodal perception layer collects heterogeneous data from multiple sources across three dimensions: server, network, and user. The fusion analysis layer is used to uncover the correlation between data from different dimensions; at the same time, it transforms multi-source data into a unified feature vector and outputs a dynamic load balancing strategy. Based on the results of the fusion analysis, the intelligent decision-making layer generates an executable load balancing strategy and optimizes the graph neural network module and temporal convolutional network module built into the fusion analysis layer through feedback iteration. The elastic execution layer transforms the strategies generated by the intelligent decision-making layer into actual actions and responds to load changes through elastic mechanisms; the elastic execution layer collects the strategy execution effect in real time and feeds it back to the multimodal perception layer. Server clusters distribute requests across multiple server nodes through clustered deployment. Dynamically expand or shrink capacity based on traffic fluctuations, while isolating traffic from different services; Server clusters can provide diverse data sources for the fusion analytics layer.

2. The distributed server load balancing system based on multimodal data fusion according to claim 1, characterized in that, The data types collected by the multimodal perception layer include those from the server side, the network side, and the user side; the multimodal perception layer deploys three types of heterogeneous sensor networks, namely performance probe clusters, network topology scanners, and semantic parsing engines.

3. The distributed server load balancing system based on multimodal data fusion according to claim 2, characterized in that, The performance probe cluster can collect dynamic indicators of server nodes in real time, including traditional resource indicators, hardware status indicators, and application-level characteristic indicators. The network topology scanner can construct a dynamic network state map and continuously monitor key parameters between nodes, including communication latency, link packet loss rate, and cross-rack bandwidth cost, updating the topology weight matrix every 1-20 seconds. The semantic parsing engine can analyze the semantic features of the request content through natural language processing technology, extract key business attributes, and construct a request profile vector.

4. The distributed server load balancing system based on multimodal data fusion according to claim 1, characterized in that, The multimodal perception layer can perform data standardization processing, including normalizing data from different modalities; Aligning multi-source data with timestamps ensures consistency in subsequent fusion analysis.

5. The distributed server load balancing system based on multimodal data fusion according to claim 1, characterized in that, The fusion analysis layer employs a heterogeneous neural network fusion architecture to process multimodal data; the fusion analysis layer includes a graph neural network module, a temporal convolutional network module, and a cross-modal attention mechanism; The graph neural network module can process the topological relationship data of server clusters and build a node influence model; it learns the communication patterns between servers through a multi-layer graph convolutional network, outputs multi-dimensional node embedding vectors, and quantitatively evaluates the structural importance of each node in the cluster. The Temporal Convolutional Network (TCN) module can analyze the spatiotemporal characteristics of performance metrics, using dilated causal convolution to capture long-period dependencies and predict load trends for the next 20-40 seconds. A cross-modal attention mechanism is used to align and fuse heterogeneous data. By calculating the correlation weights of performance data, network state, and request semantics through a multi-head attention layer, a unified feature tensor is generated.

6. The distributed server load balancing system based on multimodal data fusion according to claim 1, characterized in that, The intelligent decision-making layer adopts a two-stage collaborative decision-making mechanism to achieve a balance between global optimization and local adjustment, namely, using a global optimizer and a local optimizer; The global optimizer constructs a multi-objective optimization model based on mixed-integer programming. The objective function is to minimize {α × maximum server load + β × network cost + γ × SLA default rate}. Here, maximum server load is the load value of the server node with the highest load in the cluster; network cost is the network overhead incurred by cross-node communication; and the SLA default rate is the proportion of service quality agreement (SLA) defaults. The dynamic weight coefficients α, β, and γ are automatically adjusted every n minutes through deep reinforcement learning. When the system load is unbalanced, the value of α is automatically increased; when network congestion is severe, the value of β is automatically increased; and when service quality deteriorates, the value of γ is automatically increased. The local adjuster deploys lightweight DQN agents on each server node to respond to local state changes in real time. The agent evaluates the long-term benefits of migration, scaling, and unloading operations through an action value function, achieving local optimality while satisfying global constraints. When a load deviation >15% is detected, the agent autonomously performs load migration or elastic resource allocation within 50-200ms.

7. The distributed server load balancing system based on multimodal data fusion according to claim 6, characterized in that, The global optimizer and local adjuster decisions are dynamically coordinated through a two-layer feedback control loop, and the specific implementation process is as follows: Uplink synchronization: The local adjuster agent uploads a state summary to the global optimizer every 5 seconds. The state summary includes key indicators such as load gradient, decision distribution, and constraint satisfaction rate. Downlink correction: The global optimizer analyzes the cluster state and sends policy correction vectors to the local adjuster agents to adjust their reward function parameters; When a node load deviation > 25% or an SLA default rate > 5% is detected, the global optimizer layer directly takes over control and implements a forced load migration.

8. The distributed server load balancing system based on multimodal data fusion according to claim 1, characterized in that, The policy delivery of the elastic execution layer is achieved by pushing policy parameters to the load balancer via API; Use cloud platform APIs to achieve elastic scaling of resources.

9. The distributed server load balancing system based on multimodal data fusion according to claim 1, characterized in that, The elastic execution layer achieves fine-grained control through a programmable data plane; the elastic execution layer includes a heterogeneous resource abstraction engine, a dynamic weight adjuster, and a feedback control loop; The heterogeneous resource abstraction engine unifies heterogeneous computing resources into a computing unit matrix, supporting task-oriented distribution based on hardware characteristics; The dynamic weight adjuster modifies the load balancing weights in real time according to the instructions of the decision-making level, with a response delay of ≤10ms; the feedback control loop continuously optimizes decisions by monitoring the execution results and uses a PID controller to dynamically adjust resource allocation parameters.

10. A distributed server load balancing method based on multimodal data fusion, characterized in that, Includes the following steps: Step 1: System initialization and configuration; Multimodal perception layer deployment includes deploying a network topology scanner, configuring a semantic parsing engine, and deploying a performance probe cluster. The process involves several key steps: initializing parameters for the fusion analysis layer, initializing node features and edge weights for the graph neural network, and constructing the graph structure based on the initial topology matrix; configuring the kernel size and time step of the temporal convolutional network to adapt to the temporal characteristics of system performance data; setting weight parameters for the cross-modal attention mechanism and defining the fusion priority of different modal data; loading the intelligent decision-making layer strategy, loading the pre-trained load balancing strategy model, and initializing the strategy correction vector; defining the rules for generating state summaries; connecting the elastic execution layer to the feedback loop and configuring the trigger threshold for the local adjuster; defining the scheduling blueprint for the global optimizer; and connecting to the heterogeneous resource abstraction engine to unify the resource representation of different server architectures. Step 2: Multimodal data acquisition and preprocessing; acquiring data from the server layer, network layer, and application layer; The semantic parsing engine performs two levels of parsing on client requests: the first level of parsing extracts the request type and target service; Second-level parsing uses regular expressions to match key parameters; Step 3: Multimodal data fusion; The graph neural network model processes the topology data, inputting the topology weight matrix into the graph neural network model and generating node embeddings through graph convolution operations; The temporal convolutional network model processes the time series data, inputting the time series data collected by the performance probe into the temporal convolutional network model to extract short-term and long-term time series features; Cross-modal attention fusion inputs node embeddings generated by graph neural network models, temporal features extracted by temporal convolutional network models, and semantic parsing results into the cross-modal attention layer; Calculate the attention weights for each modality; Generate a unified feature tensor; Step 4: Intelligent Decision Making and Strategy Generation; State Summary Generation: Input the unified feature tensor into the state summary generator and compress it into a key indicator vector; Strategy Correction Vector Calculation: Based on the current state summary and historical strategies, generate correction vectors through the strategy network; Combine the correction vectors with preset rules to generate a load balancing strategy; Local adjustment strategy is to migrate non-core tasks of high-load servers; Global optimization strategy is to dynamically adjust the cluster resource pool allocation ratio. Step 5: Flexible execution and dynamic adjustment; When the local adjuster executes, it triggers task migration when the CPU utilization of a certain server exceeds 85%. Filter for transferable tasks; The heterogeneous resource abstraction engine encapsulates tasks into a unified format and sends them to low-load servers. The global optimizer executes and schedules operations every n seconds based on the cluster resource pool status. The feedback control loop is adjusted, and the dynamic weight adjuster updates the topology weight matrix based on the execution results.

Images