Switch configuration method and system based on AI adaptive congestion control

Abstract

The invention relates to the technical field of computer network communication and intelligent control, and discloses a switch configuration method and system based on AI adaptive congestion control, and the method comprises the steps: collecting the multi-dimensional feature data of a network in real time through a data plane programmable pipeline of a switch, and transmitting the multi-dimensional feature data to an AI intelligent decision engine through an encryption channel; based on the multi-dimensional feature data, executing congestion prediction, strategy generation and interpretability analysis through an AI intelligent decision engine to obtain a congestion control regulation and control strategy; converting the congestion control regulation and control strategy into an executable switch configuration command, and issuing the executable switch configuration command to switch hardware to dynamically adjust congestion control parameters of the switch; and acquiring the network performance index after configuration execution, calculating the profit value of the congestion control strategy, and updating the model weight of the AI intelligent decision engine based on the profit value. According to the invention, the adaptability of the network to dynamic loads and complex scenes is effectively improved, delay and packet loss are reduced, and the operation and maintenance complexity is reduced.

Description

Technical Field

[0001] This application relates to the fields of computer network communication and intelligent control technology, such as a switch configuration method and system based on AI adaptive congestion control. Background Technology

[0002] With the widespread deployment of artificial intelligence training, large-scale model inference, cloud computing, and high-performance computing (HPC) clusters, the scale of global data interaction is growing exponentially. The demand for high-density, high-bandwidth, and low-latency network transmission has become the core foundation for supporting the efficient operation of these computing scenarios. The transmission performance of the network system directly determines the utilization rate of overall computing resources and the efficiency of task execution. Therefore, as the core hub, the congestion control mechanism of the switch is crucial to ensuring network performance.

[0003] In current network communication technologies, traditional switches generally employ statically configured congestion control mechanisms, primarily including three typical methods: First, achieving lossless link transmission through priority-based pause frames, which temporarily stops the transmission of corresponding traffic by sending pause frames to traffic with specified priorities, thus ensuring lossless link transmission. Second, adjusting the source-end rate based on explicit notifications, where the switch sends explicit congestion notifications to the traffic source, triggering the source to actively reduce its transmission rate to alleviate congestion. Third, employing static queue scheduling and buffer allocation strategies, which pre-configure fixed queue scheduling rules and buffer resource allocation ratios to allocate dedicated scheduling priorities and buffer spaces for different types of traffic. These static congestion control mechanisms typically rely on preset fixed thresholds and cannot dynamically adjust according to real-time changes in service load, resulting in a delayed response to network status changes. Furthermore, static strategies react slowly to sudden traffic spikes or hotspots, easily causing delays and packet loss, impacting critical service performance. Moreover, parameter configuration largely depends on manual experience, making it difficult to adapt to complex and ever-changing network topologies and diverse application scenarios, leading to high operational complexity. Therefore, a congestion control solution is urgently needed to address these issues.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0006] This disclosure provides a switch configuration method and system based on AI adaptive congestion control to solve the technical problem that existing static congestion control mechanisms lack dynamic adaptability and scenario adaptation flexibility.

[0007] In some embodiments, a switch configuration method based on AI adaptive congestion control includes: Multidimensional feature data of the network is collected in real time through the programmable pipeline of the data plane of the switch, and transmitted to the trained AI intelligent decision engine through an encrypted channel. The AI ​​intelligent decision engine runs on a domestic AI chip platform. Based on multidimensional feature data, congestion prediction, strategy generation, and interpretability analysis are performed through an AI intelligent decision engine to obtain congestion control and regulation strategies. The congestion control policy is converted into executable switch configuration commands and sent to the switch hardware to dynamically adjust the switch's congestion control parameters. Collect network performance metrics after configuration execution, calculate the benefit value of congestion control and regulation strategies, and update the model weights of the AI ​​intelligent decision engine based on the benefit value.

[0008] In some embodiments, the multidimensional feature data includes: Traffic layer metrics, including port bandwidth, rate changes, and burst traffic detection results; Transport layer metrics, including TCP window changes, RTT distribution, and packet loss rate; Application layer features, including application classification results for AI training stream, storage stream, and control stream; Topology layer information includes the load of adjacent switches, link latency, and congestion distribution.

[0009] In some embodiments, the AI ​​intelligent decision engine performs congestion prediction through a collaborative model of time-series neural networks and graph neural networks, including: The temporal dynamic features of a single link are extracted using a time series neural network and used as the initial features of the nodes to be input into a graph neural network. By using graph neural networks to aggregate the topological dependencies between nodes and links in the network, topology enhancement features for congestion prediction are formed. Based on topology enhancement features, the congestion probability and severity level within a preset time window are predicted. The preset time window is dynamically determined according to the real-time requirements of the business.

[0010] In some embodiments, the AI ​​intelligent decision engine generates execution policies based on a reinforcement learning framework, such as PPO or DQN. The state space of the reinforcement learning framework includes link utilization, queue dynamic features, RTT statistics, and ECN labeling trends. The action space includes ECN threshold, PFC triggering conditions, queue scheduling weights, and buffer resource configuration. The reward function adopts a multi-objective weighted approach to comprehensively evaluate the degree of latency improvement, bandwidth utilization improvement, and packet loss rate reduction. Congestion control strategies include: dynamically adjusting ECN / PFC thresholds, dynamically adjusting queue scheduling weights and buffer allocations, reallocating traffic priorities and paths, providing ultra-high priority guarantees for critical task flows, and performing predictive congestion mitigation.

[0011] In some embodiments, the AI ​​intelligent decision engine performs interpretability analysis through a feature attribution mechanism, which associates congestion prediction results or congestion control strategies with model input features to obtain the contribution of each feature to the decision and generates visual explanation results. The method also includes logging relevant information about the interpretability analysis process to meet audit requirements.

[0012] In some embodiments, the method further includes: Deploy distributed agents on multiple switches, where an AI intelligent decision engine runs within the distributed agents; The model parameters of each distributed agent are synchronized through a control bus or a centralized controller, and the network state characteristics obtained through local processing are shared. The network state characteristics include congestion distribution and bottleneck link information. A federated learning mechanism is adopted, in which each agent trains locally and uploads the model update parameters. The central node aggregates and generates a global model and then distributes it to achieve global congestion optimization and traffic balancing among multiple switches. The central node is either a centralized controller or a designated distributed agent.

[0013] In some embodiments, the method further includes: Based on network performance metrics or the internal state of the AI ​​intelligent decision engine, detect abnormal network device operating status or model failure in the AI ​​intelligent decision engine, and automatically roll back the switch configuration to a historical stable state. When a network loop or broadcast storm is detected, the faulty node is identified and the forwarding path is updated by verifying connectivity between multiple nodes and comparing the consistency of status reporting. Continuously monitor the distribution of input data and the accuracy of prediction output of the model in the AI ​​intelligent decision engine. When model drift or data deviation exceeds the preset range, trigger model retraining or parameter correction. Actively limit the rate of identified abnormal traffic patterns to defend against network attacks.

[0014] In some embodiments, switch configuration commands are sent to the switch hardware via the domestic switch SDK interface, including: In an independent logical storage area established inside the switch, the consistency and validity of the set of configuration parameters to be issued in the switch configuration command are verified. After successful verification, the set of configuration parameters to be distributed is switched to the running configuration of the switch through atomic operations to achieve uninterrupted service updates. When a network performance degradation or device malfunction is detected due to configuration changes, a transactional rollback mechanism is triggered to atomically restore the switch configuration to its historical stable state before the switchover.

[0015] In some embodiments, the method is implemented in a domestic environment, including: The AI ​​intelligent decision engine model is adapted and optimized for inference based on the instruction set structure of domestic AI chips, including operator adaptation, memory access optimization and parallelism adjustment. The encrypted channel uses the national cryptographic SM2 / SM4 algorithm to encrypt the transmitted control communication data; This method is compatible with domestic operating systems, and the issuance of configuration commands and logging are both signed and audited to achieve end-to-end independent control.

[0016] In some embodiments, a switch configuration system based on AI adaptive congestion control is used to perform the method described in any of the preceding embodiments, the system comprising: The multi-dimensional traffic awareness module is deployed on the switch data plane to collect multi-dimensional feature data of the network in real time through a programmable pipeline and output it through an encrypted channel. The AI ​​intelligent decision engine is used to receive multi-dimensional feature data, perform congestion prediction, strategy generation and interpretability analysis, and output congestion control and regulation strategies; wherein, the AI ​​intelligent decision engine runs on a domestic AI chip platform; The adaptive configuration module is used to convert control strategies into executable switch configuration commands and send them to the switch hardware through the domestic switch SDK interface to dynamically adjust congestion control parameters. The feedback learning module is used to collect network performance metrics after configuration execution, calculate policy reward values, and update the model weights of the AI ​​intelligent decision engine based on the reward values. The collaborative control module is used to deploy distributed agents among multiple switches, synchronize information through a centralized controller or control bus, and achieve global congestion optimization and traffic balancing of multiple switches based on a federated learning mechanism. The self-healing and self-defense module is used to automatically detect network anomalies and model failures, roll back configurations, switch fault paths, monitor model health, and defend against abnormal traffic. The switch configuration method and system based on AI adaptive congestion control provided in this disclosure can achieve the following technical effects: This application utilizes an AI intelligent decision-making engine deployed on a domestically developed AI chip platform. Firstly, it employs a programmable pipeline on the switch's data plane to collect and encrypt multi-dimensional feature data in real time, providing the AI ​​intelligent decision-making engine with a dynamic data foundation reflecting real-time network load and service status. This overcomes the shortcomings of traditional solutions that rely on fixed thresholds and cannot perceive dynamic service conditions. Secondly, the engine simultaneously performs congestion prediction, policy generation, and interpretability analysis, dynamically generating precise control policies based on real-time data. This enables proactive intervention in sudden traffic surges and complex scenarios, resolving issues of delayed response and rigid manual configuration. Next, this application automatically translates the policies into configuration commands and dynamically distributes them to the switch hardware, achieving real-time and precise adjustment of network parameters. This replaces cumbersome and delayed manual configuration, enabling the system to adapt to complex network topologies and diverse application scenarios. Finally, by collecting performance metrics after policy execution and updating model weights, a closed-loop self-learning system is formed, allowing the congestion control mechanism to continuously optimize and adapt to dynamic changes in service load. This application effectively improves the network's adaptability to dynamic loads and complex scenarios, reduces latency and packet loss, and decreases operational complexity.

[0017] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0018] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein: Figure 1 This is a flowchart illustrating a switch configuration method based on AI adaptive congestion control provided in an embodiment of this disclosure; Figure 2 This is a schematic diagram of the workflow of an AI intelligent decision engine provided in an embodiment of this disclosure; Figure 3 This is a schematic diagram of an adaptive configuration and closed-loop learning process provided in an embodiment of this disclosure; Figure 4 This is a diagram of a multi-switch collaborative deployment architecture provided in an embodiment of this disclosure; Figure 5 This is a system architecture diagram of a switch configuration based on AI adaptive congestion control provided in an embodiment of this disclosure. Detailed Implementation

[0019] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0020] The terms "first," "second," etc., used in the embodiments of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0021] Unless otherwise stated, the term "multiple" means two or more.

[0022] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0023] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0024] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0025] In the current field of network communication technology, congestion control in traditional switches largely relies on statically configured congestion control mechanisms. These include methods such as using priority-based pause frames to ensure link integrity, implementing source-end rate adjustments based on explicit notifications, and employing static queue scheduling and buffer allocation strategies. However, these static congestion control mechanisms have significant shortcomings: congestion assessment depends on fixed thresholds and cannot be dynamically adjusted based on service load; responses to sudden traffic spikes and hotspots are delayed, easily leading to transmission delays and data loss; related parameter configurations require manual intervention, making it difficult to adapt to complex and ever-changing network topologies and application scenarios; and existing solutions are often deeply coupled with foreign algorithms and chips, making them unsuitable for direct porting to domestic switch chip platforms.

[0026] To address the aforementioned problems, this disclosure provides a switch configuration method and system based on AI adaptive congestion control. The switch configuration method based on AI adaptive congestion control provided in this disclosure will be described below with reference to the accompanying drawings.

[0027] Figure 1This is a flowchart illustrating a switch configuration method based on AI adaptive congestion control provided in an embodiment of this disclosure. Figure 1 As shown, the method includes the following steps: S101: Multi-dimensional feature data of the network is collected in real time through the programmable pipeline of the data plane of the switch, and transmitted to the trained AI intelligent decision engine through an encrypted channel. The AI ​​intelligent decision engine runs on a domestic AI chip platform.

[0028] In some embodiments, this application is based on a fully domestically produced hardware and software architecture and data processing mechanism. The data acquisition process relies on the programmable pipeline (such as P4 pipeline, NP pipeline, etc.) built into the domestically produced switch chips (such as H3C, Unisplendour, Phytium network chips). This pipeline has high parallel processing capability and, combined with intelligent sampling algorithms, achieves millisecond-level network status updates without affecting data forwarding performance, ensuring the real-time nature of data acquisition and the high efficiency of system operation.

[0029] In some embodiments, the collected multidimensional feature data comprehensively covers all key layers of network transmission, including traffic layer metrics, transport layer metrics, application layer features, and topology layer information. Traffic layer metrics focus on real-time link dynamics, including port bandwidth, rate changes, and burst traffic detection results; transport layer metrics revolve around core data transmission parameters, including TCP window changes, RTT distribution, and packet loss rate; application layer features include application classification results for AI training streams, storage streams, and control streams; and topology layer information relates to the global network state, including adjacent switch load, link latency, and congestion distribution. Of course, multidimensional feature data includes, but is not limited to, the data listed above.

[0030] In some embodiments, the collected raw multidimensional feature data must first undergo a local preprocessing process, including: eliminating differences in magnitude between different indicators through data normalization, reducing transmission bandwidth consumption through feature compression, and filtering out invalid data caused by hardware errors or interference through anomaly filtering, ensuring the data quality transmitted to the AI ​​intelligent decision engine. The preprocessed data is transmitted through an encrypted channel, which uses the national cryptographic SM2 / SM4 algorithm for end-to-end encryption, effectively preventing theft and tampering risks during data transmission and meeting domestic security compliance requirements.

[0031] Furthermore, the AI ​​intelligent decision-making engine can be deployed on domestic AI chip platforms such as Ascend, Tianshu, and Suiyuan. To achieve deep compatibility between the engine and domestic chips, the model inference process needs to be specifically optimized according to the chip's instruction set structure and operator characteristics. This includes the reconstruction and adaptation of operator adaptation (such as LSTM time-series computation operators and GNN topology aggregation operators), access optimization combined with the chip's memory architecture, and adjustment of task parallelism based on the number of chip cores and parallel computing capabilities. If domestic AI chips require customized drivers or dedicated plugins, the system reserves standardized interfaces for compatibility and integration. Simultaneously, the engine is compatible with domestic operating systems such as Kylin and Tongxin UOS. Configuration commands and log recordings are signed and audited to ensure that the entire process from data collection and transmission to AI decision-making meets the end-to-end self-reliant and controllable domestic requirements.

[0032] S102: Based on the multidimensional feature data, the congestion prediction, strategy generation, and interpretability analysis are performed by the AI ​​intelligent decision engine to obtain the congestion control and regulation strategy.

[0033] Figure 2 This is a schematic diagram of the workflow of an AI intelligent decision-making engine provided in an embodiment of this disclosure. The following is in conjunction with... Figure 2 This application describes the process by which the AI ​​intelligent decision engine performs congestion prediction, strategy generation, and interpretability analysis.

[0034] In some embodiments, the AI ​​intelligent decision engine performs congestion prediction through a collaborative model of time-series neural networks (such as LSTM) and graph neural networks (GNNs). The time-series neural network extracts the temporal dynamic features of a single link over a period of time, specifically based on the historical performance sequences of links or ports in multi-dimensional feature data (including bandwidth utilization, queue depth, latency statistics, ECN labeling ratio, and other indicators reflecting link dynamic trends). It accurately captures temporal patterns such as rate fluctuations and load changes. Its output temporal representation serves as the initial feature input to the graph neural network, which focuses on characterizing the topological dependencies between multiple nodes and links in the network. By aggregating topological information such as adjacent switch load, link latency, and congestion distribution, it correlates the temporal features of a single link with the global topology, forming a topology-enhanced feature that combines temporal dynamism and topological correlation. Based on this topology-enhanced feature, the model can predict the probability of congestion and its corresponding severity level within a preset time window at both local (port-level) and global (network-level) granularities. Here, the congestion probability is calculated through the normalized output layer of the model to quantify the likelihood of congestion occurring. The severity is divided into multiple levels according to probability ranges or trend characteristics to guide subsequent parameter adjustments. The value of the preset time window is dynamically determined according to the real-time requirements of the business. For example, a shorter window is used for latency-sensitive businesses to ensure response speed, while a longer window is used for batch processing businesses to improve prediction accuracy. Ultimately, a balance is achieved between prediction accuracy and response speed.

[0035] In some embodiments, the AI ​​intelligent decision engine generates execution strategies based on a reinforcement learning framework, such as PPO or DQN. The state space of this framework comprehensively covers key operational indicators of the current network, including core data such as link utilization, queue dynamics, RTT statistics, and ECN labeling trends. The action space includes adjustable congestion management-related parameters such as ECN thresholds, PFC triggering conditions, queue scheduling weights, and buffer resource configurations. The reward function employs a multi-objective weighted approach, without fixing specific weight values, but comprehensively evaluating the degree of latency improvement, bandwidth utilization enhancement, and packet loss rate reduction to ensure that the strategy optimization direction aligns with network performance requirements.

[0036] In some embodiments, the congestion control and regulation strategies generated based on the reinforcement learning framework have multi-dimensional adaptability: when dynamically adjusting ECN / PFC thresholds, the thresholds are converged when the congestion trend intensifies and appropriately relaxed when the trend eases, with the adjustment step size related to the congestion level and the threshold range limited to a reasonable range supported by the device. When dynamically adjusting queue scheduling weights and buffer allocations, resources are dynamically reallocated for different services based on factors such as the proportion of service types, historical packet loss, and long-term queue occupancy, ensuring sufficient support for critical services. When real-time reallocation of traffic priorities and paths, data transmission paths are optimized in conjunction with ECMP or multi-path forwarding mechanisms. When providing ultra-high priority guarantees for critical task flows, this is achieved by increasing scheduling priority, reserving dedicated buffers, and assigning higher weights to the scheduling algorithm. In this application, critical tasks refer to data flows in business scenarios within domestically produced data centers, intelligent computing clusters, and high-performance network systems that have stringent requirements for network transmission performance (latency, packet loss rate, bandwidth stability) and directly impact the operational efficiency of core businesses, the continuity of task execution, or the overall computing power output. The criteria for determining critical tasks can be based on a combination of business attributes, transmission requirements, and system operational priorities. When performing predictive congestion mitigation, before detecting future congestion risks, alternative paths with relatively low loads are selected in advance within the topology for traffic diversion. The diversion ratio is dynamically set based on link load relationships, thus preventing congestion from occurring at the source.

[0037] In some embodiments, the AI ​​intelligent decision engine performs interpretability analysis through feature attribution mechanisms such as SHAP or LIME. Specifically, it deeply correlates congestion prediction results or generated congestion control strategies with model input features (such as traffic intensity, port rate, queue trends, topology dependencies, etc.) to accurately calculate the contribution of each input feature to the decision result and clarify the core influencing factors in decision formation. Simultaneously, the system generates various forms of visual explanation results, such as feature importance ranking charts, decision path diagrams, and contribution visualizations, to help administrators intuitively understand the basis of AI decision logic and significantly improve policy transparency. Furthermore, to meet data center security and compliance requirements, this application will also record detailed logs of the interpretability analysis process, including analysis time, participating features, administrator access behavior, and other key information, ensuring that AI parameter tuning is based on evidence and fully adapts to data center audit requirements.

[0038] Through the coordinated operation of congestion prediction, strategy generation, and interpretability analysis, the AI ​​intelligent decision engine can output accurate, interpretable, and adaptable congestion control strategies for domestically produced network environments, laying the foundation for subsequent configuration and execution.

[0039] S103: The congestion control and regulation strategy is converted into an executable switch configuration command and sent to the switch hardware to dynamically adjust the switch's congestion control parameters.

[0040] In some embodiments, the congestion control and regulation strategies output by the AI ​​intelligent decision engine (including dynamic adjustment of ECN / PFC thresholds, queue scheduling weights and buffer allocations, scheduling algorithm switching, etc.) are converted by the adaptive configuration module into standardized configuration commands compatible with domestic switch hardware. This conversion process fully adapts to the characteristics of domestic switch SDK interfaces and supports parameter distribution and status feedback through standardized interfaces (such as API, CLI or configuration interfaces) provided by domestic switch chip SDKs such as H3C, Unisplendour, and Phytium. No specific protocol or vendor implementation method is limited here, ensuring deep compatibility with the all-domestic hardware system.

[0041] In some embodiments, configuration commands must undergo a strict consistency and legality verification process before being issued. Figure 3 This is a schematic diagram of an adaptive configuration and closed-loop learning process provided in an embodiment of this disclosure, combined with... Figure 3 The system pre-establishes an independent logical storage area within the switch to store the set of configuration parameters to be issued (including queue weights, rate limits, buffer sizes, ECN / PFC control parameters, scheduling algorithm types, etc.) in this independent space, avoiding interference with the currently running configuration. During the verification process, the system will focus on verifying whether the parameters comply with the hardware specifications of domestic switches (such as buffer resource allocation not exceeding the total buffer capacity of the device, and threshold settings within a reasonable range supported by the chip), whether there are logical conflicts between parameters (such as compliance of the total queue scheduling weights, and matching of path configuration with the network topology), and whether they are compatible with other currently running service configurations. This ensures that the configuration to be issued is executable and secure, and avoids equipment anomalies or service interruptions caused by improper configuration from the source.

[0042] In some embodiments, after successful verification, the system instantly switches the set of configuration parameters to be distributed to the switch's operating configuration through atomic operations. This atomic operation ensures that there is no intermediate transition state during the configuration switch process, avoiding potential service transmission interruptions, data loss, or network instability during configuration updates, truly achieving uninterrupted "hot configuration" updates. After the switch is completed, the configuration command takes effect immediately, dynamically adjusting the switch's congestion control parameters, including but not limited to real-time optimization of queue weights, dynamic adjustment of rate limits, flexible allocation of buffer resources, instant refresh of ECN / PFC control parameters, adaptive switching of scheduling algorithms (such as WRR, DRR, SP), and intelligent updating of hardware flow table entries (effectively avoiding flow table oscillations), ensuring that the congestion control and regulation strategies generated by AI decisions are quickly implemented and respond promptly to changes in network congestion.

[0043] In some embodiments, to ensure the security and stability of configuration changes, the network operating status and device status can be monitored in real time after the configuration takes effect. If the network performance indicators deteriorate due to the configuration change (such as a significant increase in latency, excessive packet loss rate, or abnormal decrease in bandwidth utilization), or if the device malfunctions (such as port offline or configuration execution error), a transactional rollback mechanism is immediately triggered. This rollback mechanism strictly adheres to the atomicity principle, instantly and completely restoring the switch configuration to its historical stable state before the configuration switchover. The entire rollback process does not affect the continuous transmission of current services, ensuring the consistency and security of the configuration state, effectively avoiding network failures caused by configuration changes, and ensuring the stable and reliable operation of domestically produced switch systems.

[0044] S104: Collect network performance metrics after configuration execution, calculate the benefit value of the congestion control and regulation strategy, and update the model weights of the AI ​​intelligent decision engine based on the benefit value.

[0045] In some embodiments, to ensure that AI decision-making capabilities are dynamically optimized according to network conditions, network performance indicators after the configuration policy takes effect are collected through distributed collection nodes and embedded collection mechanisms in switches. These indicators include end-to-end latency, latency jitter, link bandwidth utilization and fluctuation, instantaneous and cumulative packet loss rate, data retransmission rate, and other core indicators. At the same time, parameter changes and network state differences before and after configuration execution are recorded synchronously to ensure that the collected data can accurately map the actual impact of policy execution on network performance and provide comprehensive and reliable basic data for subsequent evaluation.

[0046] In some embodiments, a reward signal is generated based on the collected network performance indicators. The reward value is used to evaluate the actual effectiveness of the congestion control strategy. The calculation of the reward value is based on the multi-dimensional performance improvement trend after strategy implementation. It does not rely on a fixed calculation formula but comprehensively considers the achievement of key objectives such as latency reduction, bandwidth utilization improvement, and packet loss rate reduction, while also taking into account implicit requirements such as network stability and service transmission continuity, ultimately generating a quantified reward signal. This reward value directly serves as the core standard for judging the suitability of the current congestion control strategy. If the reward value shows a positive trend, it indicates that the strategy matches the current network scenario and can effectively improve network performance. If the reward value remains below a preset threshold, it indicates that the strategy's effect has not met expectations, and the model optimization process needs to be initiated.

[0047] In some embodiments, based on the calculated reward value, the model weights of the AI ​​intelligent decision engine are dynamically updated using a reinforcement learning algorithm. The update process balances model stability and scenario adaptability. Specifically, when the reward value is consistently positive and the model prediction accuracy remains stable, a local layer update method is used to fine-tune the model parameters, avoiding decision fluctuations caused by over-updates. When the reward value is low, the model prediction accuracy decreases, or there are significant changes in scenarios such as network topology or service type, a full model update is initiated to ensure that the engine can quickly adapt to the new network operating state. Simultaneously, during the model update process, the parameter configurations corresponding to historically successful strategies can be associated and stored with scenario features to generate an evolutionary strategy library. When encountering scenarios with similar traffic patterns, topologies, or service requirements, reference strategies can be directly selected from the strategy library, and key parameters can be dynamically adjusted based on the current real-time network status to achieve rapid strategy migration and adaptation. The entire process requires no manual intervention and automatically performs feature self-reconstruction based on the characteristics of different network scenarios such as high-performance computing, storage networks, and large model training, ensuring that the AI ​​intelligent decision engine can continuously iterate and optimize, maintaining the optimal congestion control decision level in the long term.

[0048] In some embodiments, in order to achieve global congestion optimization and traffic balancing in large-scale network scenarios, this application also achieves joint control of multiple devices through distributed intelligent agent deployment and federated learning. Figure 4 This is a diagram of a multi-switch collaborative deployment architecture provided in this embodiment of the disclosure, combined with... Figure 4 Distributed intelligent agents are deployed on multiple switches, each running an AI-powered intelligent decision-making engine. These agents, deployed as independent functional units on the switch control plane, can operate independently or be integrated with control plane components without affecting data plane forwarding performance, ensuring service continuity. Communication links are established between the distributed agents via a control bus or centralized controller. This enables real-time synchronization of model parameters, ensuring consistent decision-making logic across switches. Furthermore, it allows for the sharing of locally processed key network status characteristics, including network-wide congestion distribution maps, bottleneck link locations, and load conditions, providing comprehensive data support for global decision-making. Building upon this foundation, a federated learning mechanism is employed to enable collaborative training of the global model. Each distributed agent performs small-scale local training based on locally collected traffic data and operational status, generates model update parameters, and uploads them to the central node (which can be a centralized controller or a designated distributed agent). The central node then aggregates all uploaded update parameters (e.g., through weighted integration) to generate a globally optimized model that considers the entire network scenario. This model is then distributed to each distributed agent, enabling each switch to execute congestion control based on a globally optimal strategy. This effectively avoids the performance degradation of the entire network caused by local optima and adapts to the deployment requirements of large-scale AI computing clusters or supercomputing networks.

[0049] In some embodiments, regarding anomaly detection and configuration self-healing, the system continuously monitors network performance indicators (such as latency fluctuations, sudden changes in packet loss rate, and abnormal bandwidth utilization) and the internal operating status of the AI ​​intelligent decision engine (such as prediction errors and parameter rationality). When network device malfunctions, parameter configuration failures, or model decision failures are detected (such as persistent queue anomalies, latency fluctuations exceeding set ranges, and significantly increased model prediction errors), a configuration rollback mechanism is automatically triggered to atomically restore the switch configuration to its historical optimal stable state. This optimal state is determined based on the comprehensive benefits of historical strategies, ensuring a balance between latency, throughput, and stability. Regarding network fault handling, when abnormal fluctuations such as network loops or sudden broadcast storms are detected, faulty nodes are accurately identified through methods such as bidirectional connectivity verification between multiple nodes and consistency comparison of status reporting. Subsequently, the forwarding path is automatically adjusted to avoid faulty areas, ensuring the continuity of network transmission. Regarding model health maintenance, the system continuously monitors the input data distribution and prediction output accuracy of the AI ​​intelligent decision engine. By comparing the current data distribution with historical stable distributions, it determines whether model drift or data deviation has occurred. When the deviation exceeds a preset range, it automatically triggers model retraining or parameter correction processes to ensure the long-term reliability of AI decisions. Furthermore, in terms of security defense, by analyzing abnormal patterns such as sudden traffic spikes and continuous growth trends, it identifies security threats such as DDoS attacks and protocol attacks, proactively limiting and controlling related abnormal traffic to comprehensively enhance the network system's security protection capabilities.

[0050] Based on the same inventive concept as the above-described AI-based adaptive congestion control-based switch configuration method, this application also discloses a system for configuring switches based on AI adaptive congestion control in some embodiments, which is used to execute the switch configuration method disclosed in the above embodiments. Figure 5 This is a system architecture diagram of a switch configuration based on AI adaptive congestion control provided in an embodiment of this disclosure, combined with... Figure 5The system includes: a multi-dimensional traffic perception module, deployed on the switch data plane, used to collect multi-dimensional feature data of the network in real time through a programmable pipeline and output it through an encrypted channel; an AI intelligent decision engine, used to receive the multi-dimensional feature data, perform congestion prediction, policy generation and interpretability analysis, and output congestion control and regulation policies; wherein the AI ​​intelligent decision engine runs on a domestic AI chip platform; an adaptive configuration module, used to convert the regulation policies into executable switch configuration commands and send them to the switch hardware through the domestic switch SDK interface to dynamically adjust congestion control parameters; a feedback learning module, used to collect network performance indicators after configuration execution, calculate policy benefit values, and update the model weights of the AI ​​intelligent decision engine based on the benefit values; a collaborative control module, used to deploy distributed agents among multiple switches, synchronize information through a centralized controller or control bus, and realize global congestion optimization and traffic balancing of multiple switches based on a federated learning mechanism; and a self-healing and self-defense module, used to realize automatic detection of network anomalies and model failures, configuration rollback, fault path switching, model health monitoring, and abnormal traffic defense.

[0051] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions executed by a computer device (which may be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, and other media capable of storing program code; it can also be a transient storage medium.

[0052] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for descriptive purposes only and is not intended to limit the scope of protection. As used in the description herein, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used herein means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0053] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0054] The methods and products (including but not limited to devices and equipment) disclosed in the embodiments herein can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection shown or discussed between each other may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

Claims

1. A method for configuring a switch based on AI adaptive congestion control, characterized in that, The method comprises: Real-time collection of multi-dimensional feature data of the network through the data plane programmable pipeline of the switch, transmission to the trained AI intelligent decision engine through an encrypted channel, wherein the AI intelligent decision engine runs on a domestic AI chip platform; Based on the multi-dimensional feature data, the AI intelligent decision engine performs congestion prediction, strategy generation and explainability analysis to obtain a congestion control regulation strategy; The congestion control regulation strategy is converted into executable switch configuration commands and is issued to the switch hardware to dynamically adjust the congestion control parameters of the switch; After the network performance indicators are collected after the configuration is executed, the benefit value of the congestion control regulation strategy is calculated, and the model weight of the AI intelligent decision engine is updated based on the benefit value.

2. The AI adaptive congestion control based switch configuration method of claim 1, wherein, The multi-dimensional feature data comprises: Traffic layer indicators, including port bandwidth, rate change and burst traffic detection results; Transmission layer indicators, including TCP window change, RTT distribution and packet loss rate; Application layer features, including application classification results of AI training flow, storage flow and control flow; Topology layer information, including adjacent switch load, link delay and congestion distribution. 3.The AI adaptive congestion control based switch configuration method of claim 1, wherein, The AI intelligent decision engine performs congestion prediction through a collaborative model of time series neural network and graph neural network, comprising: Using the time series neural network to extract the time series dynamic features of a single link as the initial feature input of the graph neural network; Using the graph neural network to aggregate the topological dependency of each node and link in the network to form topological enhanced features for congestion prediction; Based on the topological enhanced features, the congestion probability and severity level in the future preset time window are predicted, wherein the preset time window is dynamically determined according to the real-time requirement of the business.

4. The AI adaptive congestion control based switch configuration method of claim 1, wherein, The AI intelligent decision engine performs strategy generation based on a reinforcement learning framework, and the reinforcement learning framework is PPO or DQN, wherein The state space of the reinforcement learning framework includes link utilization rate, queue dynamic features, RTT statistics, ECN marking trend, the action space includes ECN threshold, PFC trigger condition, queue scheduling weight and buffer resource configuration, and the reward function adopts a multi-objective weighted manner to comprehensively evaluate the delay improvement, bandwidth utilization rate improvement and packet loss rate reduction degree; The congestion control regulation strategy includes: dynamically adjusting the ECN / PFC threshold, dynamically adjusting the queue scheduling weight and buffer allocation, redistributing the priority and path of the traffic, providing super high priority protection for critical task flow, and performing predictive congestion relief.

5. The AI adaptive congestion control based switch configuration method of claim 1, wherein, The AI intelligent decision engine performs explainability analysis through a feature attribution mechanism, associates the congestion prediction result or congestion control regulation strategy with the model input features, obtains the contribution of each feature to the decision, and generates a visual explanation result; The method further comprises recording relevant logs of the explainability analysis process to meet the audit requirements.

6. The AI adaptive congestion control based switch configuration method of claim 1, wherein, The method further comprises: Deploying a distributed agent on multiple switches, wherein the AI intelligent decision engine runs in the distributed agent; Model parameters of each distributed intelligent agent are synchronized through a bus or a centralized controller, and network state features including congestion distribution and bottleneck link information obtained through local processing are shared; A federal learning mechanism is adopted for local training by each intelligent agent and uploading of model update parameters, and a global model is generated by a central node for distribution to achieve global congestion optimization and flow balancing among multiple switches, wherein the central node is the centralized controller or a designated distributed intelligent agent.

7. The AI adaptive congestion control based switch configuration method of claim 1, wherein, The method further comprises: Based on the network performance indicators or the internal state of the AI intelligent decision engine, the running state of the network equipment is detected for abnormality or the model in the AI intelligent decision engine is detected for failure, and the switch configuration is automatically rolled back to a historical stable state; When a network loop or broadcast storm is detected, the connectivity between multiple nodes is verified and a state consistency comparison is performed to identify a faulty node and update a forwarding path; The input data distribution of the model in the AI intelligent decision engine and the prediction output accuracy are continuously monitored, and when model drift or data deviation is detected to exceed a preset range, retraining or parameter correction of the model is triggered; The identified abnormal traffic pattern is actively rate-limited to defend against network attacks. 8.The AI adaptive congestion control based switch configuration method of claim 1, wherein, The switch configuration command is issued to the switch hardware through a domestic switch SDK interface, including: In a separate logical storage area established in the switch, consistency and legality verification is performed on a set of to-be-issued configuration parameters in the switch configuration command; After verification, the set of to-be-issued configuration parameters is switched to the running configuration of the switch through an atomic operation to achieve uninterrupted update of services; When network performance indicators are monitored to be degraded or equipment is abnormal due to configuration changes, a transactional rollback mechanism is triggered to atomically restore the configuration of the switch to a historical stable state before switching. 9.The AI adaptive congestion control based switch configuration method of claim 1, wherein, The method is implemented in a domestic environment, including: The model of the AI intelligent decision engine is adapted to operators and optimized for reasoning based on the instruction set architecture of a domestic AI chip, including operator adaptation, memory access optimization, and parallelism adjustment; The encryption channel uses the national cryptographic SM2 / SM4 algorithm to encrypt the transmitted control communication data; The method is compatible with a domestic operating system, and the issuance and log recording of configuration commands are signed and audited to achieve end-to-end autonomous controllability.

10. An AI adaptive congestion control based switch configuration system, characterized in that, The system for performing the method of any one of claims 1 to 9 comprises: A multi-dimensional flow perception module deployed in the data plane of the switch for real-time collection of multi-dimensional feature data of the network through a programmable pipeline and output through an encrypted channel; An AI intelligent decision engine for receiving the multi-dimensional feature data and performing congestion prediction, policy generation, and explainability analysis to output congestion control regulation policies; wherein the AI intelligent decision engine runs on a domestic AI chip platform; An adaptive configuration module for converting the regulation policies into executable switch configuration commands and issuing them to the switch hardware through a domestic switch SDK interface to dynamically adjust congestion control parameters; a feedback learning module configured to collect network performance indicators after configuration execution, calculate a policy benefit value, and update model weights of the AI intelligent decision engine based on the benefit value; a collaborative control module configured to deploy distributed agents among multiple switches, synchronize information through a centralized controller or a control bus, and achieve global congestion optimization and traffic balancing of multiple switches based on a federated learning mechanism; a self-healing and self-defense module configured to achieve automatic detection of network anomalies and model failure, configuration rollback, fault path switching, model health monitoring, and abnormal traffic defense.

Images