Intelligent control method and system of switch

Abstract

The invention relates to the technical field of network communication, and discloses an intelligent control method and system of a switch. The method comprises the following steps: receiving a heterogeneous data stream in a target network domain; performing data processing by using a preset flow feature aggregation model, and generating a flow dynamic pressure field distribution diagram and a protocol priority mapping matrix; constructing an adaptive scheduling threshold model, and generating a multi-target flow control instruction set; a cascade effect is simulated based on a network congestion chained conduction model, and collaborative routing parameters are optimized; and iteratively optimizing through a mixed integer programming framework, and outputting a network scheduling action sequence to a switch control engine. The system correspondingly comprises a data receiving module, a data processing module, an instruction generation module, a simulation optimization module and an instruction output module. According to the method, heterogeneous data can be effectively processed, congestion can be accurately predicted, flow control can be optimized, the overall performance of the network can be improved, and communication requirements in a complex network environment can be met.

Description

Technical Field

[0001] The present invention relates to the technical field of network communications, and in particular to an intelligent control method and system for a switch. Background Art

[0002] In today's digital age, the scale and complexity of network communications are growing exponentially. With the continuous emergence of various emerging network applications, such as high-definition video streaming, large-scale online gaming, and cloud services, the types and characteristics of network traffic are becoming increasingly complex and diverse, posing unprecedented challenges to the traffic control and management capabilities of switches.

[0003] Traditional switch control methods have numerous limitations when dealing with complex network traffic. For one thing, they mostly rely on static rules or simple threshold settings for traffic scheduling, failing to perceive dynamic changes in network traffic in real time. For example, in the face of sudden traffic spikes, the inability to adjust scheduling strategies in a timely manner can easily lead to network congestion, which in turn increases packet loss and network latency, severely impacting the user experience. For example, in online video conferencing, when multiple participants simultaneously start streaming high-definition video, network traffic surges instantly, and traditional switches may be unable to quickly allocate sufficient bandwidth resources, resulting in video freezes, audio interruptions, and other issues.

[0004] On the other hand, traditional methods struggle to effectively handle heterogeneous data flows. In modern networks, data formats and characteristics generated by different types of network devices, protocols, and applications vary significantly. Traditional switches lack the ability to uniformly analyze and process this heterogeneous data. For example, in a network environment that includes IoT devices, mobile terminals, and traditional PCs, small data packet traffic generated by IoT devices, real-time communication traffic from mobile terminals, and large file transfer traffic from PCs coexist. Traditional switches are unable to fully exploit the underlying information in this data, making it difficult to properly schedule and manage traffic based on its characteristics.

[0005] Furthermore, traditional network congestion control relies primarily on local feedback, lacking a global understanding of the overall network status. When congestion occurs at a node in the network, this localized congestion information may not be promptly transmitted to other relevant nodes, causing congestion to spread throughout the network, creating a chain reaction and further deteriorating network performance. Furthermore, traditional methods fail to fully account for the dynamic changes in network topology and the differing priorities of different protocols when optimizing network routing, resulting in inefficient utilization of network resources.

[0006] As networks continue to expand and network applications become increasingly complex, the intelligent control capabilities of traditional switches are no longer sufficient to meet the demands of modern networks for efficient, stable, and reliable communication. Consequently, there is an urgent need for a switch control method and system that can process heterogeneous data streams in real time, accurately predict network congestion, optimize flow control instruction sets, and intelligently schedule traffic based on dynamic network changes. Summary of the Invention

[0007] The object of the present invention is to provide an intelligent control method and system for a switch to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent control of a switch, the method comprising: Receiving heterogeneous data streams within a target network domain, the data streams including port traffic timing sequences, protocol type distribution matrices, and device status telemetry information; Based on a preset traffic feature aggregation model, multi-dimensional tensor alignment and feature fusion are performed on the heterogeneous data to generate a traffic dynamic pressure field distribution map and a protocol priority mapping matrix; According to the pressure field distribution map, an adaptive scheduling threshold model is constructed to generate a multi-objective flow control instruction set, wherein the instruction set includes a load balancing topology strategy and a queue cache dynamic allocation scheme; Based on a preset network congestion chain conduction model, simulating the cascading effect of packet loss events within a preset future time window, and optimizing the collaborative routing parameters in the control instruction set; The collaborative routing parameters are iteratively optimized through a mixed integer programming framework, and a network scheduling action sequence is output to a switch control engine.

[0009] Preferably, the steps of constructing the traffic feature aggregation model include: Collect network traffic case libraries across time periods and construct a multi-dimensional training set that includes traffic burst patterns, link oscillation data, and congestion triggering conditions; Performing latent variable inference on the multi-dimensional training set through a Bayesian probabilistic graphical model to extract independent distributions of traffic dominant features and background noise; Combined with the network fluid dynamics equations, partial differential constraints for the dynamic correlation between features are constructed; The partial differential constraint condition is embedded in the spatiotemporal graph convolutional network to generate the traffic feature aggregation model that supports online updating.

[0010] Preferably, the adaptive scheduling threshold model includes: Dynamically classifying traffic congestion density levels according to the Poisson gradient field of the pressure field distribution map; Calculate the link overload risk score based on the device processing capability matrix and protocol priority weights; The risk score and the congestion density level are nonlinearly mapped using a Sigmoid function to generate a port-specific scheduling threshold; An activation condition of the multi-stage traffic shaping protocol is triggered according to the threshold.

[0011] Preferably, the steps of constructing the network congestion chain conduction model include: Collect data related to packet retransmission in historical congestion events and construct a congestion causal graph dataset; The conduction delay and packet loss rate parameters between events are extracted through Granger causality test; Combining the small-world network theory, a congestion propagation directed weighted graph is constructed to quantify the buffer dependency strength between nodes; The dependency strength and the real-time traffic disturbance factor are input into a gated recurrent unit network to generate the network congestion chain conduction model.

[0012] Preferably, the method further comprises: Identify key cascade bottleneck nodes based on simulation results of the network congestion chain conduction model; configuring a pre-cache diffusion strategy for the node in the control instruction set; Based on the diffusion strategy, a cross-device collaborative forwarding solution is automatically generated, including a priority inversion suppression instruction and a buffer pre-allocation solution.

[0013] Preferably, the calculation of the link overload risk score includes: Obtain real-time topology connectivity tensors and protocol processing latency heatmaps to build a network resilience assessment hypercube. Calculate the low-rank approximate weights between multi-dimensional features through tensor decomposition algorithm; Performing a Hadamard product operation on the approximate weight and the evaluation hypercube to obtain a comprehensive risk score; The calculation formula for the comprehensive risk score is: ; Where, Indicates the comprehensive risk score value, Indicates the The processing delay coefficient of the class protocol, Indicates the The priority weight of the class protocol, Indicates the network basic processing capability constant, represents element-wise product, Indicates the total number of protocol categories.

[0014] Preferably, the embedding of the partial differential constraint condition includes: Performing Lie group symmetry analysis on the latent variable inference results to screen out association patterns that conform to network conservation laws; Generate feature evolution paths that meet dynamic constraints through Markov chain Monte Carlo sampling; Path data is used to regularize the node embedding of the spatiotemporal graph convolutional network to ensure that the model output conforms to the physical laws of traffic propagation.

[0015] Preferably, the execution of the mixed integer programming framework includes: Define a multi-objective optimization loss function, including the dual metrics of end-to-end delay and bandwidth utilization; Solve the integer programming problem of each port scheduling strategy using branch and bound algorithm; In each round of optimization, the slack variable threshold is dynamically adjusted according to the degree of constraint conflict; Outputting the network scheduling action sequence that meets the Pareto optimality condition; Wherein, the loss function is: ; Where, represents the loss value, represents the delay penalty term, represents the bandwidth utilization gain term, is the dynamic weight factor.

[0016] Preferably, the method further comprises: After configuring the pre-cache diffusion strategy, monitor the queue occupancy rate change gradient of the bottleneck node in real time; If the change gradient exceeds the preset critical value, the simulated annealing optimization mechanism is triggered to replan the spatiotemporal collaborative path for cross-device forwarding.

[0017] Preferably, the present invention further includes an intelligent control system for a switch, the system comprising: Data receiving module: used to receive heterogeneous data streams within the target network domain, the data streams including port traffic timing sequence, protocol type distribution matrix and device status telemetry information; Data processing module: Based on the preset traffic feature aggregation model, it performs multi-dimensional tensor alignment and feature fusion on the heterogeneous data to generate a traffic dynamic pressure field distribution map and a protocol priority mapping matrix; Instruction generation module: constructing an adaptive scheduling threshold model based on the pressure field distribution map and generating a multi-objective flow control instruction set, the instruction set including a load balancing topology strategy and a queue cache dynamic allocation scheme; Simulation optimization module: based on a preset network congestion chain conduction model, simulates the cascading effect of packet loss events within a preset time window in the future, and optimizes the collaborative routing parameters in the control instruction set; Instruction output module: Iteratively optimizes the collaborative routing parameters through a mixed integer programming framework and outputs the network scheduling action sequence to the switch control engine.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The intelligent control method and system of the switch proposed in the present invention have a significant positive impact on improving network performance. First of all, the invention can efficiently process heterogeneous data streams within the target network domain, covering port traffic timing sequences, protocol type distribution matrices and device status telemetry information. Through the preset traffic feature aggregation model, these complex data are subjected to multi-dimensional tensor alignment and feature fusion to generate a traffic dynamic pressure field distribution map and a protocol priority mapping matrix. This process not only fully taps the potential value in the data, but also provides a comprehensive and accurate basis for subsequent traffic scheduling. For example, in a complex network environment containing multiple network applications, the model can clearly display the traffic pressure conditions in different areas and the priority of each protocol, enabling the switch to allocate resources in a targeted manner, giving priority to ensuring the traffic transmission of critical services, and effectively avoiding network jams caused by unreasonable resource allocation.

[0019] An adaptive scheduling threshold model, built based on a dynamic traffic pressure field distribution map, dynamically categorizes traffic congestion density based on real-time network status, calculates link overload risk scores, and generates port-specific scheduling thresholds. This refined threshold setting triggers the activation of a multi-level traffic shaping protocol, which in turn generates a multi-objective flow control instruction set, including a load balancing topology strategy and a dynamic queue buffer allocation scheme. The load balancing topology strategy dynamically adjusts the network topology to achieve more even traffic distribution, preventing overload on certain links or nodes due to excessive traffic. The dynamic queue buffer allocation scheme flexibly adjusts the switch's internal queue buffer resources based on real-time traffic changes, ensuring timely data packet processing and significantly reducing packet loss. In practical application scenarios, such as large data center networks, when a large number of users simultaneously access server resources, the adaptive scheduling threshold model can quickly respond, rationally allocate resources, and ensure efficient data transmission.

[0020] Utilizing a pre-set chain-congestion model, the system simulates the cascading effects of packet loss events within a preset future time window and optimizes collaborative routing parameters. This mechanism can predict network congestion trends in advance, identify key cascading bottleneck nodes, configure pre-caching and diffusion strategies for them, and automatically generate cross-device collaborative forwarding solutions. These measures effectively mitigate the chain reaction of network congestion and improve the network's fault tolerance. When a key node in the network shows signs of congestion, the system can take timely action to redirect data packets to alternative paths for transmission. It also pre-allocates cache resources to potentially affected nodes to ensure normal network operation.

[0021] A mixed integer programming framework iteratively optimizes collaborative routing parameters and outputs a Pareto-optimal network scheduling action sequence to the switch control engine. This process comprehensively considers both end-to-end latency and bandwidth utilization, dynamically adjusting their weighting based on different network requirements and application scenarios to achieve overall network performance optimization. In network applications with high real-time requirements, such as live streaming, the system prioritizes reducing end-to-end latency to ensure smooth video playback. In file transfer scenarios with high bandwidth utilization requirements, the system focuses on improving bandwidth utilization to accelerate file downloads.

[0022] The switch intelligent control method and system of the present invention can effectively improve the overall performance of the network, achieve efficient, stable and reliable communication in a complex and changeable network environment, meet the diverse needs of different users and application scenarios, and have broad application prospects and important practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a working principle diagram of the switch intelligent control method according to the present invention; Figure 2 This is a flowchart of node processing and path optimization based on the congestion model; Figure 3 Flowchart for link overload risk score calculation and decision making; Figure 4 Flowchart generated for optimization and scheduling sequences in a mixed integer programming framework. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] See also Figure 1-4The present invention provides an intelligent switch control method and system, aiming to improve the traffic control and scheduling capabilities of network switches in complex network environments, ensuring efficient and stable network operation. The specific implementation steps are as follows: The switch receives heterogeneous data streams within the target network domain through its built-in data receiving interface. These data streams include port traffic time series, which record the traffic volume of each port at different time points, reflecting the changes in network traffic over time; a protocol type distribution matrix, which displays the proportion and distribution of different network protocols in network traffic; and device status telemetry information, which covers the switch's hardware status, such as CPU utilization and memory usage, as well as link status, such as bandwidth utilization and latency. By obtaining this data, the switch can fully understand the current network operation status.

[0026] The received heterogeneous data is processed using a preset traffic feature aggregation model. This model performs tensor alignment on multidimensional data, aligning different types of data across time and space, and then performs feature fusion. After processing, a dynamic traffic pressure field distribution map is generated. This map intuitively displays traffic pressure in different areas of the network, helping to identify potential areas of network congestion. A protocol priority mapping matrix is also generated to clarify the priority order of different network protocols in network transmission, providing a basis for subsequent traffic scheduling.

[0027] Based on the generated pressure field distribution map, an adaptive scheduling threshold model is constructed. This model generates a multi-objective flow control instruction set based on the relevant characteristics of the pressure field. The load balancing topology strategy in the instruction set adjusts the network topology to ensure a more balanced distribution of traffic within the network and avoid overloading certain links or nodes. The dynamic queue cache allocation scheme rationally allocates queue cache resources within the switch based on real-time changes in network traffic, ensuring that packets are properly processed and minimizing packet loss.

[0028] Using a pre-defined chain-congestion model, we simulate the cascading effects of packet loss events within a pre-defined time window. This simulation predicts the potential chain reactions caused by network congestion and helps optimize the co-routing parameters within the control instruction set. These parameters determine the transmission paths of packets within the network. Optimized parameters can improve the network's fault tolerance and reduce the impact of congestion on network performance.

[0029] Using a mixed-integer programming framework, collaborative routing parameters are iteratively optimized. This optimization process considers multiple factors, such as network latency and bandwidth utilization, ultimately outputting a sequence of network scheduling actions that meet the requirements. This sequence is then sent to the switch control engine, which adjusts the switch's operation in real time based on these instructions, enabling intelligent control of network traffic and ensuring stable and efficient network operation.

[0030] The implementation of the present invention will be further described below with reference to Examples 1 to 6.

[0031] Example 1: In practical applications, the construction of a traffic feature aggregation model plays a key role in achieving intelligent switch control. First, a cross-period network traffic case library is collected. This case library data comes from a wide range of sources, covering network traffic data in different time periods and under different network load conditions. From this data, a multi-dimensional training set is constructed, including traffic burst patterns, link oscillation data, and congestion trigger conditions. Traffic burst patterns record sudden and significant increases in traffic, including their occurrence time and duration; link oscillation data reflects unstable fluctuations in link bandwidth; and congestion trigger conditions define the traffic flow, link status, and other conditions that will trigger network congestion.

[0032] The Bayesian probabilistic graphical model is used to perform latent variable inference on a multi-dimensional training set. This model can uncover hidden relationships and patterns in the data. In this process, it extracts the independent distributions of dominant traffic characteristics and background noise. Dominant traffic characteristics are key factors influencing network traffic fluctuations, such as the large amounts of data transmitted by specific applications. Background noise, on the other hand, is a random factor that has a smaller impact on network traffic.

[0033] To more accurately describe the relationships between features, we combined the network fluid dynamics equations to construct partial differential constraints for the dynamic associations between features. The network fluid dynamics equations describe the flow patterns of network traffic within a network. The partial differential constraints constructed based on these equations constrain the relationships between traffic features at a physical level.

[0034] When embedding partial differential constraints into the spatiotemporal graph convolutional network, the latent variable inference results are first subjected to Lie group symmetry analysis. This Lie group symmetry analysis can screen out correlation patterns that conform to network conservation laws, ensuring that the model adheres to the underlying physical laws of the network when processing data. Next, Markov chain Monte Carlo sampling is used to generate feature evolution paths that conform to dynamic constraints. Markov chain Monte Carlo sampling is a random sampling method that efficiently samples in complex spaces, generating qualified feature evolution paths. Finally, these path data are used to perform regularized training on the node embeddings of the spatiotemporal graph convolutional network. Through regularized training, the model output conforms to the physical laws of traffic propagation, resulting in a traffic feature aggregation model that supports online updates. This allows the model to update in real time even in a constantly changing network environment, accurately processing heterogeneous data and providing reliable support for subsequent traffic control.

[0035] In this embodiment, the traffic feature aggregation model extracts key information from massive amounts of network traffic data through a complex multi-step process. By combining physical laws with mathematical models for optimization, it effectively addresses complex and ever-changing network traffic conditions, laying a solid foundation for intelligent switch control. Its online update feature also enables it to adapt to ever-changing network environments, maintaining accurate analysis and processing of network traffic.

[0036] Example 2: The adaptive scheduling threshold model dynamically categorizes traffic congestion density levels based on the Poisson gradient field of the dynamic traffic pressure field distribution map. The Poisson gradient field reflects the changing trend and speed of the pressure field. By analyzing this field, network congestion can be classified into different levels, such as mild, moderate, and severe congestion. This provides a more intuitive understanding of the severity of network congestion and provides a basis for subsequent scheduling decisions.

[0037] When calculating the link overload risk score, we need to obtain the real-time topology connectivity tensor and protocol processing delay heatmap to construct a network resilience assessment hypercube. The real-time topology connectivity tensor reflects the connectivity and stability of the network topology, while the protocol processing delay heatmap shows the delay of different protocols during processing. Combining these two factors to construct the network resilience assessment hypercube allows for a comprehensive assessment of network performance from multiple dimensions.

[0038] The tensor decomposition algorithm calculates low-rank approximate weights between multidimensional features. The tensor decomposition algorithm decomposes high-dimensional tensors into low-dimensional tensor combinations. This method generates low-rank approximate weights between multidimensional features. These weights reflect the importance of different features in assessing link overload risk.

[0039] Perform a Hadamard product operation on the approximate weight and the evaluation hypercube to obtain the comprehensive risk score. The calculation formula for the comprehensive risk score is: ; Where, Indicates the comprehensive risk score value, which comprehensively reflects the overload risk level of the link; Indicates the The processing delay coefficient of the protocol class, which reflects the The degree to which delays in the processing of similar protocols affect the overload risk; Indicates the The priority weight of the protocol class reflects the importance of different protocols in network transmission; Indicates the network basic processing capability constant, which represents the processing capability level of the network itself; It represents the element-wise product, which multiplies the delay coefficients and priority weights of different protocols by their corresponding elements and then sums them. Indicates the total number of protocol categories, covering all different types of protocols in the network.

[0040] After obtaining the comprehensive risk score, it is nonlinearly mapped to the congestion density level using a Sigmoid function to generate a port-specific scheduling threshold. The Sigmoid function maps input values to a range between 0 and 1. Through this nonlinear mapping, an appropriate scheduling threshold is generated for each port based on the congestion situation and risk score. When the port's traffic reaches or exceeds this threshold, the activation conditions of the multi-level traffic shaping protocol are triggered, which shapes and schedules the traffic to ensure stable network operation.

[0041] In this way, the adaptive scheduling threshold model dynamically adjusts scheduling thresholds based on the real-time network status and protocol characteristics, achieving refined control over network traffic and improving overall network performance and reliability. It comprehensively considers factors such as network topology, protocol processing latency, protocol priority, and basic network processing capabilities, making the generated scheduling thresholds more reasonable and effective, and better able to cope with complex and changing network traffic conditions.

[0042] Example 3: The network congestion chain transmission model is used to predict the cascading effects of network congestion. In practice, data related to packet retransmissions from historical congestion events is collected. This data details the retransmissions of packets during congestion, including the number of retransmissions, the retransmission interval, and the links through which the retransmitted packets traveled. Based on this data, a congestion causal graph dataset is constructed. This dataset graphically illustrates the causal relationships between congestion events, providing an intuitive data structure for subsequent analysis.

[0043] Granger causality testing is used to extract the transmission delay and packet loss rate parameters between events. Granger causality testing is a commonly used causal relationship testing method. In network congestion analysis, it can determine whether the occurrence of one congestion event leads to the occurrence of another, as well as the time delay and packet loss rate between them. This method can accurately obtain the dynamic relationship parameters between congestion events.

[0044] A congestion propagation directed weighted graph is constructed based on small-world network theory. Small-world network theory suggests that in some complex networks, nodes are connected by short paths, even in large networks. In a congestion propagation directed weighted graph, nodes represent devices or links in the network, edges represent the direction of congestion propagation, and weights quantify the buffer dependency strength between nodes. Buffer dependency strength reflects the degree to which the buffer status of one node affects another. For example, when a node's buffer is full, it may cause congestion in connected nodes. The extent of this impact is represented by the buffer dependency strength.

[0045] The dependency strength and real-time traffic perturbation factor are input into a gated recurrent unit network to generate a chain-like congestion model. A gated recurrent unit network is a specialized recurrent neural network that effectively processes time series data and captures long-term dependencies within the data. This model uses dependency strength and real-time traffic perturbation factors as inputs and processes them through the gated recurrent unit network to generate a chain-like congestion model that accurately predicts the cascading effects of packet loss events within a preset future time window. This model provides an important basis for optimizing co-routing parameters in control instruction sets, enabling proactive measures to avoid or mitigate the impact of network congestion.

[0046] By building this network congestion chain transmission model, we can deeply analyze the propagation mechanism of network congestion, accurately predict congestion trends, and provide strong support for intelligent switch control. It fully utilizes historical data and advanced mathematical theories and models to adapt to congestion prediction needs in different network environments and improve network reliability and stability.

[0047] Example 4: In the process of intelligent control based on the network congestion chain transmission model, identifying key cascading bottleneck nodes and implementing corresponding strategies are important means of improving network performance. After simulating the cascading effects of packet loss events within a preset future time window, the network congestion chain transmission model can accurately identify key cascading bottleneck nodes by analyzing the simulation results. These nodes are key locations in the network that are prone to triggering chain reactions of congestion. Once they become congested, they are likely to cause congestion in surrounding nodes as well, thus affecting the performance of the entire network.

[0048] For identified critical cascade bottleneck nodes, a pre-cache diffusion strategy is configured in the control instruction set. This strategy aims to allocate cache resources to bottleneck nodes before they become congested, and to diffuse some packets into the caches of other nodes, alleviating pressure on the bottlenecks. For example, lower-priority packets can be temporarily stored in other nodes with free cache space, and then transmitted when network conditions improve.

[0049] Based on the pre-cache diffusion strategy, a cross-device collaborative forwarding solution is automatically generated, which includes priority inversion suppression instructions and buffer pre-allocation solutions. Priority inversion suppression instructions are used to adjust the transmission priority of data packets to avoid increased congestion caused by unreasonable priorities. For example, in some cases, data packets with lower priority may block higher priority data packets due to inappropriate transmission timing. This situation can be corrected by priority inversion suppression instructions. The buffer pre-allocation solution allocates buffer resources in advance to nodes that may be affected by congestion based on the real-time status of the network and the load of the nodes, ensuring that data packets can be processed in a timely manner.

[0050] After configuring the pre-cache diffusion strategy, it is necessary to monitor the queue occupancy gradient of the bottleneck node in real time. The queue occupancy gradient reflects the rate of change of the number of data packets in the bottleneck node queue. If the gradient exceeds the preset critical value, it means that the congestion of the bottleneck node may be rapidly deteriorating. At this time, the simulated annealing optimization mechanism is triggered to re-plan the spatiotemporal collaborative path for cross-device forwarding. The simulated annealing optimization mechanism is a probability-based optimization algorithm that can search for the optimal solution within a certain range. When re-planning the path, the mechanism will comprehensively consider factors such as the network topology, node load, and data packet priority to find an optimal cross-device forwarding path to alleviate the congestion of the bottleneck node and ensure the normal operation of the network.

[0051] This series of operations effectively addresses network congestion, improving fault tolerance and transmission efficiency. Identifying key nodes, configuring pre-caching and flooding strategies, and real-time monitoring and path optimization all work together to form a complete network congestion response system, providing a more reliable foundation for intelligent switch control.

[0052] Example 5: A mixed integer programming framework is used to optimize the collaborative routing parameters and output the network scheduling action sequence. In the actual implementation process, the multi-objective optimization loss function is first defined. This function includes the dual indicators of end-to-end delay and bandwidth utilization. The loss function is: ; in, Represents the loss value, which comprehensively reflects the network performance in terms of end-to-end delay and bandwidth utilization; Delay penalty is a term used to measure the impact of the delay experienced by data packets from the source node to the destination node on network performance. The greater the delay, the higher the penalty value. Indicates the bandwidth utilization gain item, which reflects the utilization efficiency of network bandwidth. The higher the bandwidth utilization, the greater the gain item value. is a dynamic weight factor that can adjust the relative importance of end-to-end delay and bandwidth utilization in the loss function according to actual network requirements and application scenarios. For example, for applications with high real-time requirements, such as video conferencing, the The value of makes the network more focused on reducing latency; and for applications with large data transmission volume, such as file downloads, it can be reduced The value of , pay more attention to bandwidth utilization.

[0053] After defining the loss function, the branch-and-bound algorithm is used to solve the integer programming problem for each port's scheduling policy. This algorithm is a commonly used method for solving integer programming problems. It continuously decomposes the problem into subproblems and estimates the bounds of each subproblem, gradually narrowing the search range and ultimately finding the optimal solution. During the solution process, the scheduling policy for each port is optimized to determine the transmission order and transmission time of packets on each port.

[0054] During each optimization round, the slack variable threshold is dynamically adjusted based on the degree of constraint conflict. The degree of constraint conflict reflects the degree of conflict between different constraints during the optimization process. When constraint conflicts are severe, appropriately adjusting the slack variable threshold can relax some constraints, allowing the optimization algorithm to find feasible solutions within a wider range. When constraint conflicts are minor, tightening the slack variable threshold can improve solution accuracy.

[0055] Through continuous iterative optimization, a network scheduling action sequence that satisfies the Pareto optimality condition is ultimately output. This condition means that no single objective can be further optimized without sacrificing other objectives. A network scheduling action sequence that meets this condition achieves an optimal balance between end-to-end latency and bandwidth utilization, effectively improving overall network performance and providing a scientific and rational decision-making basis for intelligent switch control.

[0056] The mixed-integer programming framework uses rigorous mathematical models and algorithms to comprehensively consider multiple network factors, optimizing collaborative routing parameters and generating network scheduling action sequences. It can flexibly adjust optimization objectives and parameters based on different network requirements and actual conditions, providing strong support for efficient and stable network operation.

[0057] Example 6: This embodiment is used to describe an intelligent control system for a switch according to the present invention. The intelligent control system is composed of multiple modules with different functions but closely cooperating with each other. The modules work together to realize intelligent control of the switch to ensure efficient and stable operation of the network.

[0058] Data Receiving Module: This module serves as the data interface between the intelligent control system and the target network domain, responsible for acquiring critical information. Its hardware architecture includes high-performance network interface cards (NICs) with high-speed data acquisition capabilities, enabling real-time capture of various data signals within the target network domain. At the software level, a dedicated data capture program accurately identifies and extracts port traffic timing sequences, a protocol type distribution matrix, and device status telemetry information. The port traffic timing sequence records port traffic changes over time, providing a basis for analyzing network traffic trends. The protocol type distribution matrix displays the traffic share of different network protocols, helping to understand the network application composition. Device status telemetry information covers the switch's hardware status (such as CPU utilization and memory usage) and link status (such as bandwidth utilization and latency), reflecting the switch's real-time operational status. The data receiving module encapsulates this heterogeneous data in a specific format and rapidly transmits it to the data processing module via an internal data bus, ensuring data timeliness and integrity.

[0059] Data Processing Module: This module is one of the core processing units of the intelligent control system and features a built-in preset traffic feature aggregation model. Upon receiving heterogeneous data from the data receiving module, the data processing module rapidly activates the model. During processing, multidimensional tensor alignment is first performed to match and calibrate data of different dimensions and formats in both time and space, enabling data analysis within a unified framework. Subsequently, complex algorithms are used to fuse features, exploring the inherent connections between the data and generating a dynamic traffic pressure field distribution map and a protocol priority mapping matrix. For example, by fusion analysis of port traffic timing sequences and the protocol type distribution matrix, the flow pressure magnitude and distribution range in different regions are determined, and a dynamic traffic pressure field distribution map is plotted. Based on the transmission requirements and network policies of different protocols, a protocol priority mapping matrix is generated, providing a priority basis for subsequent traffic scheduling. The data processing module utilizes parallel computing technology and optimized data processing algorithms to effectively improve processing speed and accuracy, ensuring that complex data processing tasks are completed in a short period of time.

[0060] Instruction Generation Module: Based on the pressure field distribution map generated by the data processing module, the instruction generation module constructs an adaptive scheduling threshold model. This module comprehensively considers factors such as network topology, device processing capability matrix, and protocol priority weights to calculate a link overload risk score and dynamically classify traffic congestion density levels based on the score. A sigmoid function is used to nonlinearly map the risk score to the congestion density level to generate port-specific scheduling thresholds. When port traffic reaches or exceeds these thresholds, the activation conditions of the multi-level traffic shaping protocol are triggered, which in turn generates a multi-objective traffic control instruction set. The load balancing topology strategy within the instruction set adjusts the network topology to guide traffic distribution and avoid link and node overload. The dynamic queue buffer allocation scheme flexibly allocates internal queue buffer resources within the switch based on real-time traffic changes, ensuring proper packet processing and minimizing packet loss. The instruction generation module can rapidly adjust scheduling strategies based on real-time network changes and generate precise and effective control instructions.

[0061] Simulation and Optimization Module: This module uses a pre-defined network congestion chain transmission model to simulate the cascading effects of packet loss events within a preset future time window. This module collects packet retransmission correlation data from historical congestion events to construct a congestion causal graph dataset. Using Granger causality tests, it extracts the transmission delay and packet loss rate parameters between events. Incorporating small-world network theory, it constructs a directed weighted graph of congestion propagation, quantifying the buffer dependency strength between nodes. This dependency strength and real-time traffic perturbation factors are input into a gated recurrent unit network to generate a network congestion chain transmission model. This model predicts congestion trends, identifies critical cascading bottleneck nodes, and configures pre-caching diffusion strategies for these nodes in the control instruction set. It also automatically generates cross-device coordinated forwarding solutions, including priority inversion suppression instructions and buffer pre-allocation schemes. Based on the simulation results, the module also optimizes the coordinated routing parameters in the control instruction set to improve network fault tolerance and transmission efficiency.

[0062] Instruction Output Module: The instruction output module is responsible for outputting the network scheduling action sequence optimized by the simulation optimization module to the switch control engine. This module iteratively optimizes the collaborative routing parameters using a mixed integer programming framework. During the optimization process, a multi-objective optimization loss function is defined that includes the dual indicators of end-to-end delay and bandwidth utilization. The integer programming problem of each port scheduling strategy is solved using a branch-and-bound algorithm, and the slack variable threshold is dynamically adjusted based on the degree of constraint conflict. Ultimately, a network scheduling action sequence that meets the Pareto optimality condition is output. The instruction output module uses a high-speed data transmission interface to ensure that the optimized instructions can be accurately and promptly transmitted to the switch control engine, allowing the switch to control network traffic in real time based on these instructions, achieving efficient and stable network operation.

[0063] Through the collaborative work of the above modules, the switch's intelligent control system can fully and accurately perceive the network status, quickly generate and execute effective control instructions, realize intelligent control of the switch, meet the diverse needs under different network environments, and provide strong guarantees for the stable operation of the network.

[0064] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0065] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent control of a switch, characterized in that: include: Receiving heterogeneous data streams within a target network domain, the data streams including port traffic timing sequences, protocol type distribution matrices, and device status telemetry information; Based on a preset traffic feature aggregation model, multi-dimensional tensor alignment and feature fusion are performed on the heterogeneous data to generate a traffic dynamic pressure field distribution map and a protocol priority mapping matrix; According to the pressure field distribution map, an adaptive scheduling threshold model is constructed to generate a multi-objective flow control instruction set, wherein the instruction set includes a load balancing topology strategy and a queue cache dynamic allocation scheme; Based on a preset network congestion chain conduction model, simulating the cascading effect of packet loss events within a preset future time window, and optimizing the collaborative routing parameters in the control instruction set; The collaborative routing parameters are iteratively optimized through a mixed integer programming framework, and a network scheduling action sequence is output to a switch control engine.

2. The intelligent control method according to claim 1, characterized in that: The steps of constructing the traffic feature aggregation model include: Collect network traffic case libraries across time periods and construct a multi-dimensional training set that includes traffic burst patterns, link oscillation data, and congestion triggering conditions; Performing latent variable inference on the multi-dimensional training set through a Bayesian probabilistic graphical model to extract independent distributions of traffic dominant features and background noise; Combined with the network fluid dynamics equations, partial differential constraints for the dynamic correlation between features are constructed; The partial differential constraint condition is embedded in the spatiotemporal graph convolutional network to generate the traffic feature aggregation model that supports online updating.

3. The intelligent control method according to claim 1, characterized in that: The adaptive scheduling threshold model includes: Dynamically classifying traffic congestion density levels according to the Poisson gradient field of the pressure field distribution map; Calculate the link overload risk score based on the device processing capability matrix and protocol priority weights; The risk score and the congestion density level are nonlinearly mapped using a Sigmoid function to generate a port-specific scheduling threshold; An activation condition of the multi-stage traffic shaping protocol is triggered according to the threshold.

4. The intelligent control method according to claim 1, characterized in that: The steps of constructing the network congestion chain conduction model include: Collect data related to packet retransmission in historical congestion events and construct a congestion causal graph dataset; The conduction delay and packet loss rate parameters between events are extracted through Granger causality test; Combining the small-world network theory, a congestion propagation directed weighted graph is constructed to quantify the buffer dependency strength between nodes; The dependency strength and the real-time traffic disturbance factor are input into a gated recurrent unit network to generate the network congestion chain conduction model.

5. The intelligent control method according to claim 4, characterized in that: Also includes: Identify key cascade bottleneck nodes based on simulation results of the network congestion chain conduction model; configuring a pre-cache diffusion strategy for the node in the control instruction set; Based on the diffusion strategy, a cross-device collaborative forwarding solution is automatically generated, including a priority inversion suppression instruction and a buffer pre-allocation solution.

6. The intelligent control method according to claim 3, characterized in that: The calculation of the link overload risk score includes: Obtain real-time topology connectivity tensors and protocol processing latency heatmaps to build a network resilience assessment hypercube. Calculate the low-rank approximate weights between multi-dimensional features through tensor decomposition algorithm; Performing a Hadamard product operation on the approximate weight and the evaluation hypercube to obtain a comprehensive risk score; The calculation formula for the comprehensive risk score is: ； Where, Indicates the comprehensive risk score value, Indicates the The processing delay coefficient of the class protocol, Indicates the The priority weight of the class protocol, Indicates the network basic processing capability constant, represents element-wise product, Indicates the total number of protocol categories.

7. The intelligent control method according to claim 2, characterized in that: The embedding of the partial differential constraint condition includes: Performing Lie group symmetry analysis on the latent variable inference results to screen out association patterns that conform to network conservation laws; Generate feature evolution paths that meet dynamic constraints through Markov chain Monte Carlo sampling; Path data is used to regularize the node embedding of the spatiotemporal graph convolutional network to ensure that the model output conforms to the physical laws of traffic propagation.

8. The intelligent control method according to claim 1, characterized in that: The implementation of the mixed integer programming framework includes: Define a multi-objective optimization loss function, including the dual metrics of end-to-end delay and bandwidth utilization; Solve the integer programming problem of each port scheduling strategy using branch and bound algorithm; In each round of optimization, the slack variable threshold is dynamically adjusted according to the degree of constraint conflict; Outputting the network scheduling action sequence that meets the Pareto optimality condition; Wherein, the loss function is: ； Where, represents the loss value, represents the delay penalty term, represents the bandwidth utilization gain term, is the dynamic weight factor.

9. The intelligent control method according to claim 5, characterized in that: Also includes: After configuring the pre-cache diffusion strategy, monitor the queue occupancy rate change gradient of the bottleneck node in real time; If the change gradient exceeds the preset critical value, the simulated annealing optimization mechanism is triggered to replan the spatiotemporal collaborative path for cross-device forwarding.

10. An intelligent control system for a switch, characterized in that: include: Data receiving module: used to receive heterogeneous data streams within the target network domain, the data streams including port traffic timing sequence, protocol type distribution matrix and device status telemetry information; Data processing module: Based on the preset traffic feature aggregation model, it performs multi-dimensional tensor alignment and feature fusion on the heterogeneous data to generate a traffic dynamic pressure field distribution map and a protocol priority mapping matrix; Instruction generation module: constructing an adaptive scheduling threshold model based on the pressure field distribution map and generating a multi-objective flow control instruction set, the instruction set including a load balancing topology strategy and a queue cache dynamic allocation scheme; Simulation optimization module: based on a preset network congestion chain conduction model, simulates the cascading effect of packet loss events within a preset time window in the future, and optimizes the collaborative routing parameters in the control instruction set; Instruction output module: Iteratively optimizes the collaborative routing parameters through a mixed integer programming framework and outputs the network scheduling action sequence to the switch control engine.

Images