Communication networking method and device of computing power server, equipment and storage medium

Abstract

This application discloses a communication networking method, apparatus, device, and storage medium for a computing server, relating to the field of communication technology. The method constructs a standardized node feature set through multi-dimensional parameter collection and principal component analysis, quantifies performance evaluation, and reduces dimensionality to lay the foundation for the topology. An improved Prim algorithm generates an initial topology with composite weights and connection number constraints, optimizing performance and preventing single-point overload. A graph neural network aggregates neighbor information, predicts link efficiency and global vulnerability, and intelligently evaluates the network. Real-time load and prediction are fused to establish adaptive link priority, responding to load changes. A deep Q-network undergoes multi-action iterative optimization to balance efficiency and robustness, outputting a self-optimizing dynamic topology. Multi-path configuration selects independent paths without shared intermediate nodes, improving fault tolerance and reliability. Overall, the method uses data-driven approaches and intelligent algorithms to achieve network self-awareness, self-optimization, and self-recovery, providing high-throughput, low-latency, and fault-tolerant communication infrastructure for high-computing clusters.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a communication networking method, apparatus, device and storage medium for a computing server. Background Technology

[0002] With the explosive growth in computing power demand from scenarios such as artificial intelligence training, high-performance computing, and edge intelligence, computing server clusters are showing a trend towards large-scale, distributed, and heterogeneous development. Servers within a cluster are often distributed in different geographical locations, such as racks within a data center or cross-regional edge nodes. The computing power, network bandwidth, and real-time load of nodes vary significantly, and computing tasks have high requirements for communication latency, bandwidth stability, and fault tolerance.

[0003] As a key channel for computing power collaboration, the efficiency of communication networking directly determines the overall performance of the cluster. Inefficient networking can lead to idle computing power (such as high-performance nodes being unable to participate in collaboration due to communication bottlenecks), a surge in task response latency (such as cross-node data synchronization blocking), and wasted resources (such as redundant links excessively occupying bandwidth).

[0004] However, the core challenge in networking communication for computing servers lies in the fact that unified modeling of heterogeneous node features requires the integration of multi-dimensional heterogeneous data such as geographic coordinates, computing power configuration, network bandwidth, real-time load, and historical latency to construct a unified feature model. However, dimensionality reduction and physical meaning mapping of high-dimensional data can easily lose key information, affecting the accuracy of subsequent decisions. At the same time, under dynamic load, traditional static topology algorithms build networks based on fixed weights, which cannot adjust link priorities according to task fluctuations. Reconstruction also requires balancing communication efficiency and energy consumption, and the lack of quantitative evaluation can lead to over- or under-construction. In addition, topology vulnerability assessment often ignores the proportion of connectivity reduction after the failure of key nodes, while multi-path redundancy configuration to enhance fault tolerance faces problems such as path combination explosion, complex detection of intermediate node sharing, uneven traffic distribution, and high failure switching latency, making it difficult to balance reliability and resource efficiency. Particularly noteworthy is that multi-path deployment requires traversing the global reachability matrix to screen for paths without shared intermediate nodes, facing challenges such as exponential growth in the number of paths and difficulties in dynamic adaptation and coordination of traffic shaping and forward error correction coding.

[0005] These challenges collectively point to the technical problem of low network communication efficiency in traditional static topologies and experience-based configurations. Summary of the Invention

[0006] The main objective of this application is to provide a communication networking method, apparatus, device, and storage medium for computing servers, so as to solve the problem of low networking communication efficiency in the prior art based on traditional static topology and empirical configuration.

[0007] To achieve the above objectives, this application provides the following technical solution:

[0008] A communication networking method for computing power servers, wherein the communication networking method applies a cluster of computing power servers to be networked, the cluster of computing power servers to be networked includes a plurality of computing power servers, and the communication networking method includes:

[0009] Step S1: Collect the geographical location parameters, core calculation parameters, and data transmission parameters of all computing power servers to be networked, and construct a standardized node feature set including computing power intensity, bandwidth potential, and load sensitivity through principal component analysis and standardization processing.

[0010] Step S2: Define each computing server as an undirected graph node, using the product of the computing power intensity and the bandwidth potential as the basic weight and the reciprocal of the geographical location parameter as the auxiliary weight. Use the Prim improved algorithm to find an initial topology that covers all undirected graph nodes with the goal of limiting the number of direct connections of a single node to no more than a preset threshold and maximizing the sum of the basic weight and the auxiliary weight.

[0011] Step S3: Using a graph neural network model, the feature vector of each undirected graph node in the standardized node feature set is used as a node attribute, and each topological adjacency relationship of the initial topology is used as an edge attribute. After aggregating the neighbor information through the network layer, the expected communication efficiency of each link and the overall topological vulnerability index are obtained.

[0012] Step S4: Obtain the real-time load data of each computing server, and calculate the link priority ranking of each undirected graph node by combining the expected communication efficiency and the overall topology vulnerability index.

[0013] Step S5: Define the action space and reward function based on the pre-trained deep Q network decision model, input the link priority sorting into the deep Q network decision model, and iterate the link priority sorting through all action spaces until the reward function reaches its maximum value to obtain the optimized dynamic topology;

[0014] Step S6: Based on the optimized dynamic topology, configure at least two independent physical paths without shared intermediate nodes for each pair of nodes using the multipath transmission control protocol to obtain a multipath mapping table networking scheme.

[0015] Beneficial effects:

[0016] Steps S1 to S6 systematically achieve efficient communication networking of the computing server cluster. Specifically, Step S1 uses multi-dimensional parameter collection and principal component analysis to construct a standardized set of node features, enabling quantitative evaluation of server performance and feature dimensionality reduction, providing a data foundation for subsequent topology construction. Secondly, Step S2's improved Prim algorithm introduces composite weights and connection number constraints to traditional graph theory methods, generating an initial topology that optimizes network performance while avoiding single-point overload. Next, Step S3 employs a graph neural network for intelligent topology analysis, accurately predicting link communication efficiency and global vulnerability index through neighbor information aggregation, achieving intelligent evaluation of network performance. Step S4 then establishes an adaptive link priority mechanism through dynamic fusion of real-time load data and prediction results, enabling the network to respond to real-time workload changes. Step S5 uses a deep Q-network decision model for iterative optimization in a multi-action space, achieving a balance between communication efficiency and topology robustness, resulting in a self-optimizing dynamic topology. Finally, Step S6, through a multi-path configuration scheme and independent path design without shared intermediate nodes, significantly improves the network's fault tolerance and transmission reliability. The overall solution combines data-driven approaches with intelligent algorithms to achieve self-awareness, self-optimization, and self-recovery of high-performance computing networks, providing computing clusters with high-throughput, low-latency, and fault-tolerant communication infrastructure.

[0017] As a further improvement to this application, in step S1, the geographical location parameters, core calculation parameters, and data transmission parameters of all computing power servers to be networked are collected. Principal component analysis and standardization processing are used to construct a standardized set of node features, including computing power intensity, bandwidth potential, and load sensitivity, comprising:

[0018] Step S11: Collect the geographical location parameters, core calculation parameters, and data transmission parameters of all computing power servers to be networked, and classify and store all parameters according to the node number of each computing power server to be networked to form the original parameter dataset.

[0019] Step S12: Perform data interpolation and preprocessing on the original parameter dataset to obtain a normalized parameter matrix;

[0020] Step S13: Input the normalized parameter matrix into the pre-trained principal component analysis model to calculate the covariance matrix;

[0021] Step S14: Sort the covariance matrix from largest to smallest according to its eigenvalues ​​and retain the first few principal components whose cumulative contribution rate exceeds a preset percentage threshold to construct a dimension-reduced projection matrix.

[0022] Step S15: Perform a linear transformation on the normalized parameter matrix using the dimensionality reduction projection matrix to obtain the principal component score vector for each node;

[0023] Step S16: Based on the physical meaning of the principal components, the first principal component is defined as the computing power intensity, the second principal component is defined as the bandwidth potential, and the third principal component is defined as the load sensitivity.

[0024] Step S17: Standardize the computing power intensity, the bandwidth potential, and the load sensitivity respectively, and integrate them to obtain the standardized node feature set.

[0025] Beneficial effects:

[0026] Steps S11 to S17, through systematic data acquisition and feature engineering, achieve accurate quantitative characterization and effective dimensionality reduction of the performance of computing server cluster nodes, providing standardized feature inputs with high signal-to-noise ratio for subsequent communication networking. Specifically, step S11 constructs the original dataset through multi-dimensional parameter acquisition, providing a comprehensive data foundation for feature extraction; step S12's normalization process eliminates differences in parameter dimensions, making data from different sources comparable; steps S13 to S15's principal component analysis achieves intelligent dimensionality reduction of the feature space, significantly reducing computational complexity while retaining key information; step S16's feature definition based on physical meaning ensures the interpretability of the dimensionality reduction results, making the three core features—computing power intensity, bandwidth potential, and load sensitivity—directly correspond to the server's computing, transmission, and scheduling capabilities; and step S17's standardization process further eliminates differences in feature scale, making performance comparisons between different servers more objective. These steps together form a complete feature processing pipeline: after cleaning and transformation of the original data, principal component analysis extracts the most representative feature dimensions, which are then given physical meaning by combining domain knowledge, and finally, the feature scale is standardized and unified. This pipeline not only solves the challenge of structurally representing heterogeneous server parameters, but also effectively suppresses the interference of redundant features on network construction through dimensionality reduction, enabling subsequent topology optimization to focus on the key factors that truly affect communication performance. The construction of a standardized node feature set essentially transforms complex server hardware characteristics into quantifiable network construction parameters, providing a unified performance evaluation benchmark for computing clusters. Improving the adaptability and robustness of the networking method effectively addresses the networking needs of server clusters of different sizes and configurations, laying a data foundation for subsequent graph theory-based network optimization.

[0027] As a further improvement to this application, step S2 defines each computing server as an undirected graph node, using the product of the computing power intensity and the bandwidth potential as the basic weight and the reciprocal of the geographical location parameter as the auxiliary weight. Using the Prim improved algorithm, with the goal of limiting the number of direct connections of a single node to no more than a preset threshold and maximizing the sum of the basic weight and the auxiliary weight, an initial topology covering all undirected graph nodes is found, including:

[0028] Step S21: Define each computing server in the standardized node feature set as an undirected graph node, and construct a node set including all nodes;

[0029] Step S22: Extract the computing power intensity and bandwidth potential of each node and calculate their product to obtain the basic weight of each node.

[0030] Step S23: Extract the geographic location parameters of each node and take their reciprocals to obtain the auxiliary weights of each node;

[0031] Step S24: Initialize an empty edge set as the edge set of the initial topology, create a set of visited nodes and randomly select a node to add to the set of visited nodes;

[0032] Step S25: Create a candidate edge set, which includes all edges connecting the visited node set and the unvisited node set. The weight of the current edge is the sum of the basic weight and auxiliary weight of the two nodes connected to the current edge.

[0033] Step S26: Select the two endpoints of the edge with the largest current weight in the candidate edge set to further select the nodes with fewer connected edges among the two endpoints, and restrict the current direct connection number of the selected nodes to be less than the preset threshold.

[0034] Step S27: Add the edge with the largest current weight to the edge set, and add the unvisited node connected by the edge with the largest current weight to the visited node set;

[0035] Step S28: Remove all edges connecting newly added nodes to other nodes in the visited node set from the candidate edge set, and update the current direct connection count record for all nodes in the visited node set.

[0036] Step S29: Repeat steps S26 to S28 until the visited node set contains all nodes, and output the initial topology structure consisting of all selected edges.

[0037] Beneficial effects:

[0038] Steps S21 to S29 intelligently construct the initial topology of the computing server cluster using an improved Prim algorithm. Steps S21 to S23 define graph nodes and calculate weights, providing structured input for subsequent operations. Steps S24 to S28 construct a complete iterative optimization process, where a dynamic update mechanism for the visited node set ensures topology connectivity, real-time maintenance of the candidate edge set improves search efficiency, and continuous updating of connection count records ensures strict enforcement of constraints. The synergistic effect of these steps enables the algorithm to efficiently traverse all possible connection combinations, ultimately outputting the optimal topology covering all nodes in step S29. Firstly, a composite weight design achieves a balance between network performance and geographical constraints. The basic weights are the product of computing power and bandwidth potential, directly reflecting the computational and transmission coordination capabilities of nodes. The auxiliary weights use the reciprocal of the geographical location parameter, effectively suppressing latency issues caused by long-distance connections. This weight design ensures that the generated topology significantly reduces the impact of physical distance on communication quality while maintaining computational efficiency. Secondly, the introduced single-node connection limit mechanism overcomes the limitations of the traditional Prim algorithm. By dynamically verifying the number of connected edges at both ends of a candidate edge, it ensures that the number of direct connections for each node does not exceed a preset threshold. This constraint not only avoids over-connected hub nodes in the network but also evenly distributes data transmission pressure, giving the topology better load balancing characteristics. Finally, by prioritizing the selection of nodes with the largest weights and fewer connections, the goal of maximizing the sum of basic and auxiliary weights is achieved. This optimization strategy ensures that the generated initial topology achieves optimal overall performance, providing a high-quality input foundation for subsequent graph neural network analysis.

[0039] As a further improvement to this application, step S3 involves using a graph neural network model to take the feature vector of each undirected graph node in the standardized node feature set as a node attribute and each topological adjacency relationship of the initial topology as an edge attribute. After aggregating neighbor information through the network layer, the expected communication efficiency of each link and the overall topology vulnerability index are obtained, including:

[0040] Step S31: Construct a graph neural network model architecture with three graph convolutional layers, each graph convolutional layer being configured with an independent weight matrix;

[0041] Step S32: The feature vector of each undirected graph node in the standardized node feature set is used as a node attribute of the graph neural network model, and all topological adjacency relationships in the initial topology are used as edge attributes;

[0042] Step S33: In the first graph convolutional layer, neighbor information aggregation is performed on each undirected graph node. The feature vectors of adjacent nodes connected by edge attributes are linearly transformed by the first layer weight matrix and then summed with the feature vectors of the corresponding center nodes of the adjacent nodes. The first layer node embedding representation is obtained by processing through the activation function.

[0043] Step S34: Input the first layer node embedding representation into the second graph convolutional layer, perform neighbor information aggregation again and extend the aggregation range to two-hop neighbor nodes, and process the aggregated feature vector through the second layer weight matrix linear transformation and the activation function to obtain the second layer node embedding representation;

[0044] Step S35: Input the second layer node embedding representation into the third graph convolutional layer, concatenate the embedding representations of each node and its corresponding direct neighbors, and map the concatenated vector to a scalar output through a fully connected layer. The scalar output is the evaluation value of the expected communication efficiency.

[0045] Step S36: Calculate the global graph-level representation vector based on the second-layer node embedding representation of all nodes;

[0046] Step S37: Input the global graph-level representation vector into a multilayer perceptron with two output nodes. One of the output nodes is processed by the Sigmoid function to obtain the overall topological fragility index.

[0047] Beneficial effects:

[0048] Steps S31 to S37 construct a graph neural network model to achieve intelligent evaluation of the communication network performance of computing servers. The three-layer graph convolutional architecture in step S31 provides sufficient depth for feature learning; the node and edge attribute definitions in step S32 clearly define the structured representation of the input data; the stepwise information aggregation in steps S33 to S35 achieves feature extraction from local to global perspectives; the global graph-level representation computation in step S36 provides the foundation for vulnerability assessment; and the dual-output design in step S37 enables the model to simultaneously optimize communication efficiency and network robustness. The synergistic effect of these steps allows the model to accurately capture complex dependencies in the network, and the evaluation results reflect both the intrinsic performance of nodes and the propagation characteristics of the network topology.

[0049] As a further improvement to this application, step S4 involves acquiring the real-time load data of each computing server, and calculating the link priority ranking of each undirected graph node based on the expected communication efficiency and the overall topology vulnerability index, including:

[0050] Step S41: Obtain the real-time load data of each computing server through an external cluster monitoring system, and combine all the real-time load data of a computing server into a real-time load data vector.

[0051] Step S42: Normalize the real-time load data vector to obtain a standardized real-time load vector;

[0052] Step S43: Take the arithmetic mean of the standardized real-time load vectors of the nodes at both ends of the current link as the load factor of the current link, and perform weighted fusion with the expected communication efficiency to obtain the comprehensive link evaluation vector.

[0053] Step S44: The overall topology vulnerability index is introduced as an adjustment factor into the comprehensive link evaluation vector, and the comprehensive link evaluation vector is corrected by the exponential decay function to obtain the corrected comprehensive link evaluation vector.

[0054] Step S45: Calculate the difference between the corrected integrated link evaluation vector of the current node and the standardized real-time load vector of the neighboring nodes, and use it as the link value score of the current node and all its neighboring nodes.

[0055] Step S46: Sort all neighboring nodes of the current node in descending order according to the link value score to obtain the link priority sort of the current node.

[0056] Beneficial effects:

[0057] Steps S41 to S46 achieve intelligent prioritization of computing server cluster links by dynamically integrating real-time load status and network performance indicators. Step S41 uses load data collected in real-time by the cluster monitoring system, which complements the communication efficiency indicators predicted by the graph neural network introduced in step S43. This ensures that the priority ranking reflects both the current intensity of computing resource usage and the potential capacity of network transmission. This dynamic fusion mechanism enables the network scheme to quickly respond to server load fluctuations and avoid communication bottlenecks caused by local overload. Secondly, the overall topology vulnerability index introduced in step S44 serves as an adjustment factor. The comprehensive evaluation results are corrected through an exponential decay function, allowing the priority ranking to optimize communication efficiency while proactively avoiding vulnerable links in the network. This design significantly improves the robustness of the network scheme. When some links experience performance degradation, the system can automatically adjust the data transmission path, reducing the impact of single-point failures on overall communication. Finally, the link value scoring and priority ranking mechanism constructed in steps S45 and S46, by quantifying the differences in communication value between nodes, provides a precise basis for action selection for the deep Q-network decision model. This allows subsequent topology optimization to focus on adjusting critical links, improving algorithm iteration efficiency. The synergistic effect of these steps gives the networking scheme adaptive characteristics, enabling it to dynamically adjust communication strategies according to real-time load changes, ensuring efficient execution of computing tasks while maintaining network stability. This process not only improves the reliability of data transmission but also lays a solid foundation for subsequent topology optimization, enabling the computing server cluster to maintain efficient communication under complex workloads.

[0058] As a further improvement to this application, step S5 defines an action space and a reward function based on a pre-trained deep Q-network decision model, inputs the link priority ranking into the deep Q-network decision model, and iterates the link priority ranking through all action spaces until the reward function reaches its maximum value, thereby obtaining the optimized dynamic topology, including:

[0059] Step S51: Define the deep Q-network decision model as having three types of action spaces, namely, a subset of link addition / deletion actions, a subset of node grouping / merging actions, and a subset of cross-group direct connection actions;

[0060] Step S52, define the reward function of the deep Q network decision model as the sum of the communication efficiency improvement value and the reconstruction energy consumption reduction value minus the overall topology vulnerability index;

[0061] Step S53: Convert the link priority sorting into the initial state representation of the deep Q network decision model, and input the initial state representation into the deep Q network decision model;

[0062] Step S54: In each decision cycle of the deep Q-network decision model, calculate the Q value of each sub-action in the three action spaces according to the state representation of the current decision cycle, and select the sub-action with the largest Q value as the selected action of the current decision cycle.

[0063] Step S55: After executing the selected action, add or delete links, group or merge nodes, or directly connect across groups according to the action type of the selected action to obtain an intermediate topology.

[0064] Step S56: Recalculate the reward values ​​of the two reward functions according to the intermediate topology, and sum the two reward values ​​as the immediate reward for the current decision cycle, while obtaining the updated state representation;

[0065] Step S57: Select the actions, immediate rewards, and updated state representations of the same decision cycle into an experience tuple, and randomly select an experience tuple to update the deep Q-network decision model using gradient descent.

[0066] Step S58: Repeat steps S54 to S57 to iteratively optimize the link priority ranking through the deep Q network decision model until the output values ​​of the two reward functions reach their respective historical maximum values ​​and remain stable.

[0067] Step S59: Output the network topology corresponding to the fact that the output values ​​of the two reward functions have reached their respective historical maximum values ​​and remain stable as the optimized dynamic topology.

[0068] Beneficial effects:

[0069] Steps S51 to S59 achieve dynamic and intelligent optimization of the computing server cluster communication network scheme through iterative optimization of the deep Q-network decision model. Firstly, comprehensive network topology optimization is achieved through the synergistic effect of multiple action spaces. The link addition / deletion action space directly adjusts physical connection relationships; the node grouping / merging action space optimizes data routing paths through logical grouping; and the cross-group direct connection action space breaks down traditional grouping communication barriers. The combination of these three action spaces enables the model to intervene at multiple levels, from local connection adjustments to global structure optimization, improving the adaptability and flexibility of the network topology. This allows the network scheme to automatically adjust the communication structure according to real-time load changes, maintaining efficient and stable data transmission performance. Secondly, the designed dual-reward function mechanism achieves a balanced optimization of communication efficiency and network robustness. The sum of the improvement in communication efficiency and the reduction in reconstruction energy consumption serves as a positive reward, guiding the model to prioritize topology adjustment strategies that improve data transmission efficiency and reduce energy consumption. Meanwhile, the negative value of the overall topology vulnerability index acts as a negative constraint, prompting the model to actively avoid fragile connections that could lead to network outages. This combined positive and negative reward design allows the optimization process to simultaneously consider performance improvement and reliability assurance, avoiding the problem of traditional methods that overemphasize a single metric while neglecting network stability. Finally, the model achieves adaptive optimization in complex network environments through a continuous learning mechanism of experience replay and gradient descent. The selected action, immediate reward, and updated state representation in each decision cycle are stored as experience tuples. Model parameters are updated through random sampling and gradient descent. This mechanism allows the model to continuously accumulate optimization experience and gradually converge to the optimal solution. The final optimized dynamic topology not only meets the communication requirements of the current workload but also possesses robustness against future load changes.

[0070] As a further improvement to this application, step S6, based on the optimized dynamic topology, configures at least two independent physical paths without shared intermediate nodes for each pair of nodes using a multipath transmission control protocol, to obtain a multipath mapping table networking scheme, including:

[0071] Step S61: Extract the direct and indirect connection relationships between all node pairs from the optimized dynamic topology and construct a global reachability matrix;

[0072] Step S62: For each pair of nodes, traverse all transportable paths in the global reachability matrix and constrain the transportable paths to consist of continuous physical links and not contain duplicate nodes.

[0073] Step S63: Perform intermediate node sharing detection on all transmittable paths of each node pair, and filter out path combinations that do not have shared intermediate nodes.

[0074] Step S64: From the path combinations that do not share intermediate nodes, select at least two independent physical paths for each pair of nodes based on the combination with the fewest path hops.

[0075] Step S65: Configure the selected independent physical path for each group of node pairs through the multipath transmission control protocol, and establish the mapping relationship between path identifiers and physical links;

[0076] Step S66: Store the mapping relationships of all node pairs in groups according to the source node to obtain a multi-path mapping table networking scheme that includes path priority and link status.

[0077] Beneficial effects:

[0078] Steps S61 to S66 achieve high reliability and fault tolerance in the communication network of the computing server cluster through a multi-path transmission control protocol. Step S61 extracts direct and indirect connections from the optimized dynamic topology, ensuring the comprehensiveness of the path search, while step S62's constraint on the traversal of continuous physical links guarantees the physical feasibility of the paths. Furthermore, a rigorous path selection mechanism ensures the independence of multiple paths. Step S63's intermediate node sharing detection technically eliminates the risk of resource contention between paths. Step S64's optimization strategy, based on minimizing path hops, further optimizes transmission efficiency while satisfying the constraint of no shared nodes. This dual selection mechanism ensures that the configured independent physical paths possess both fault isolation capabilities and maintain the lowest possible transmission latency. Finally, steps S65 to S66, through protocol implementation and mapping table generation, transform the theoretical path configuration into an executable networking scheme. The specific implementation of the multipath transmission control protocol establishes a reliable correspondence between path identifiers and physical links, while the systematic storage of the multipath mapping table enables centralized management of network status and provides a structured basis for dynamic routing adjustments. This complete closed loop from topology analysis to protocol implementation enables the computing server cluster to automatically switch faulty paths without interrupting service, significantly improving the continuous operation capability of critical computing tasks.

[0079] To achieve the above objectives, this application also provides the following technical solutions:

[0080] A communication networking device for a computing server, the communication networking device being applied to the communication networking method described above, the communication networking device comprising:

[0081] The standardized node feature set construction module is used to collect the geographical location parameters, core computing parameters, and data transmission parameters of all computing power servers to be networked. Through principal component analysis and standardization processing, it constructs a standardized node feature set including computing power intensity, bandwidth potential, and load sensitivity.

[0082] The initial topology search module is used to define each computing server as an undirected graph node, using the product of the computing power intensity and the bandwidth potential as the basic weight and the reciprocal of the geographical location parameter as the auxiliary weight. The Prim improved algorithm is used to search for an initial topology covering all undirected graph nodes with the goal of limiting the number of direct connections of a single node to no more than a preset threshold and maximizing the sum of the basic weight and the auxiliary weight.

[0083] The topology link parameter calculation module is used to take the feature vector of each undirected graph node in the standardized node feature set as node attribute and each topological adjacency relationship of the initial topology structure as edge attribute through the graph neural network model, and obtain the expected communication efficiency of each link and the overall topology vulnerability index after aggregating the neighbor information through the network layer.

[0084] The link priority ranking calculation module is used to obtain the real-time load data of each computing server, and calculate the link priority ranking of each undirected graph node by combining the expected communication efficiency and the overall topology vulnerability index.

[0085] The link priority ranking iteration module is used to define the action space and reward function based on the pre-trained deep Q network decision model, input the link priority ranking into the deep Q network decision model and iterate the link priority ranking through all action spaces until the reward function reaches its maximum value, thereby obtaining the optimized dynamic topology.

[0086] The independent physical path allocation module is used to configure at least two independent physical paths without shared intermediate nodes for each pair of nodes based on the optimized dynamic topology through the multipath transmission control protocol, thereby obtaining a multipath mapping table networking scheme.

[0087] To achieve the above objectives, this application also provides the following technical solutions:

[0088] An electronic device includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the communication networking method described above.

[0089] To achieve the above objectives, this application also provides the following technical solutions:

[0090] A computer-readable storage medium storing program instructions, which, when executed by a processor, enable the communication networking method described above. Attached Figure Description

[0091] Figure 1This is a flowchart illustrating the steps of one embodiment of a communication networking method for a computing server according to this application.

[0092] Figure 2 This is a functional module diagram of one embodiment of a communication networking device for a computing server according to this application;

[0093] Figure 3 This is a schematic diagram of the structure of an embodiment of the electronic device of this application;

[0094] Figure 4 This is a schematic diagram of the structure of one embodiment of the storage medium of this application. Detailed Implementation

[0095] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0096] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0097] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0098] like Figure 1 As shown, this embodiment provides an example of a communication networking method for computing power servers. In this embodiment, the communication networking method is applied to a cluster of computing power servers to be networked, which includes several computing power servers.

[0099] Specifically, the communication networking method includes the following steps:

[0100] Step S1: Collect the geographical location parameters, core calculation parameters, and data transmission parameters of all computing power servers to be networked. Construct a standardized node feature set including computing power intensity, bandwidth potential, and load sensitivity through principal component analysis and standardization processing.

[0101] Further, in step S1, the geographical location parameters, core calculation parameters, and data transmission parameters of all computing power servers to be networked are collected. A standardized node feature set, including computing power intensity, bandwidth potential, and load sensitivity, is constructed through principal component analysis and standardization. Specifically, this includes the following steps:

[0102] Step S11: Collect the geographical location parameters, core calculation parameters, and data transmission parameters of all computing power servers to be networked, and classify and store all parameters according to the node number of each computing power server to be networked to form the original parameter dataset.

[0103] Preferably, taking three computing servers as an example, this embodiment uniformly defines one computing server as one node. Then, the geographical location parameter collects the Euclidean distance from each node to the cluster center. For example, if the cluster center is (0,0) by default, then the distance d1 of the first node Server_1 (1,2) is √(1 / 2)Euclidean distance. 2 +2 2 =√5 is approximately 2.236; calculate the core parameters by collecting the number of cores × clock speed, for example, Server_1 is 32 cores × 2.5GHz, and get the original computing power value of 80; collect the data transmission parameters by collecting the network card bandwidth × historical average throughput, for example, Server_1 = 100Gbps × 80Gbps, and get the original bandwidth value of 8000.

[0104] Preferably, the original parameter dataset is categorized and stored according to node number, for example, {Server_1: {d:2.236, computing power:80, bandwidth:8000, load fluctuation:0.2}, ...}), with ellipses representing subsequent data collection from other dimensions.

[0105] Step S12: Perform data interpolation and preprocessing on the original parameter dataset to obtain the normalized parameter matrix.

[0106] Preferably, missing data in the original parameter dataset can be filled using linear interpolation. For example, if the load fluctuation of Server_2 is missing, it can be filled using the average value of Server_1 and Server_3.

[0107] Step S13: Input the normalized parameter matrix into the pre-trained principal component analysis model to calculate the covariance matrix.

[0108] Step S14: Sort the covariance matrix by eigenvalues ​​from largest to smallest and retain the top principal components whose cumulative contribution rate exceeds a preset percentage threshold to construct a dimension-reduced projection matrix.

[0109] Step S15: Perform a linear transformation on the normalized parameter matrix using the dimension reduction projection matrix to obtain the principal component score vector for each node.

[0110] Step S16: Based on the physical meaning of the principal components, the first principal component is defined as computing power intensity, the second principal component is defined as bandwidth potential, and the third principal component is defined as load sensitivity.

[0111] Step S17: Standardize the computing power intensity, bandwidth potential, and load sensitivity respectively, and integrate them to obtain a standardized node feature set.

[0112] Preferably, the covariance matrix of the normalized matrix can reflect the correlation of features. After eigenvalue decomposition, principal components with a cumulative contribution rate of ≥90% can be retained. For example, if the cumulative contribution rate of three eigenvalues ​​λ1=2.5, λ2=1.2, and λ3=0.3 is 92.5%, then all three can be selected.

[0113] The physical meaning of the principal components is as follows:

[0114] The first principal component is computing power intensity, which is the sum of cores and clock speed.

[0115] The second principal component is bandwidth potential, which is the sum of network interface card (NIC) and throughput.

[0116] The third principal component is the load sensitivity, which is the standard deviation of the load fluctuation.

[0117] The Z-scores of the three principal components are standardized and integrated to obtain a standardized set of node features, for example, Server_1={computing power intensity:1.07, bandwidth potential:0.92, load sensitivity:-0.85}.

[0118] Beneficial effects:

[0119] Steps S11 to S17, through systematic data acquisition and feature engineering, achieve accurate quantitative characterization and effective dimensionality reduction of the performance of computing server cluster nodes, providing standardized feature inputs with high signal-to-noise ratio for subsequent communication networking. Specifically, step S11 constructs the original dataset through multi-dimensional parameter acquisition, providing a comprehensive data foundation for feature extraction; step S12's normalization process eliminates differences in parameter dimensions, making data from different sources comparable; steps S13 to S15's principal component analysis achieves intelligent dimensionality reduction of the feature space, significantly reducing computational complexity while retaining key information; step S16's feature definition based on physical meaning ensures the interpretability of the dimensionality reduction results, making the three core features—computing power intensity, bandwidth potential, and load sensitivity—directly correspond to the server's computing, transmission, and scheduling capabilities; and step S17's standardization process further eliminates differences in feature scale, making performance comparisons between different servers more objective. These steps together form a complete feature processing pipeline: after cleaning and transformation of the original data, principal component analysis extracts the most representative feature dimensions, which are then given physical meaning by combining domain knowledge, and finally, the feature scale is standardized and unified. This pipeline not only solves the challenge of structurally representing heterogeneous server parameters, but also effectively suppresses the interference of redundant features on network construction through dimensionality reduction, enabling subsequent topology optimization to focus on the key factors that truly affect communication performance. The construction of a standardized node feature set essentially transforms complex server hardware characteristics into quantifiable network construction parameters, providing a unified performance evaluation benchmark for computing clusters. Improving the adaptability and robustness of the networking method effectively addresses the networking needs of server clusters of different sizes and configurations, laying a data foundation for subsequent graph theory-based network optimization.

[0120] Step S2: Define each computing server as an undirected graph node, using the product of computing power and bandwidth potential as the basic weight and the reciprocal of the geographical location parameter as the auxiliary weight. Use the Prim improved algorithm to find an initial topology that covers all undirected graph nodes with the goal of limiting the number of direct connections of a single node to no more than a preset threshold and maximizing the sum of the basic and auxiliary weights.

[0121] Further, in step S2, each computing server is defined as an undirected graph node, with the product of computing power intensity and bandwidth potential as the basic weight and the reciprocal of the geographical location parameter as the auxiliary weight. Using the Prim improved algorithm, with the goal of limiting the number of direct connections of a single node to no more than a preset threshold and maximizing the sum of the basic and auxiliary weights, an initial topology covering all undirected graph nodes is found. This specifically includes the following steps:

[0122] Step S21: Define each computing server in the standardized node feature set as an undirected graph node, and construct a node set including all nodes.

[0123] Step S22: Extract the computing power intensity and bandwidth potential of each node and calculate their product to obtain the basic weight of each node.

[0124] Step S23: Extract the geographic location parameters of each node and take their reciprocals to obtain the auxiliary weights of each node.

[0125] Preferably, the node base weight is w_base(u) = computing power intensity × bandwidth potential, for example, Server_1 = 1.07 × 0.92 ≈ 0.98; the node auxiliary weight is w_aux(u) = 1 / geographical distance, for example, Server_1 = 1 / 2.236 ≈ 0.447; the edge weight is w(e = uv) = (w_base(u) + w_base(v)) + (w_aux(u) + w_aux(v)), for example, the edge weight of Server_1-2 = (0.98 + 0.85) + (0.447 + 0.5).

[0126] =2.777).

[0127] Step S24: Initialize an empty edge set as the edge set of the initial topology, create a set of visited nodes and randomly select a node to add to the set of visited nodes.

[0128] Step S25: Create a candidate edge set, which includes all edges connecting the visited node set and the unvisited node set. The weight of the current edge is the sum of the basic weight and auxiliary weight of the two nodes connected to the current edge.

[0129] Step S26: Select the two endpoints of the edge with the largest current weight in the candidate edge set to further select the nodes with fewer connected edges among the two endpoints, and restrict the current direct connection number of the selected nodes to be less than a preset threshold.

[0130] Preferably, the preset threshold can be set to 3.

[0131] Step S27: Add the edge with the largest current weight to the edge set, and add the unvisited node connected by the edge with the largest current weight to the visited node set.

[0132] Step S28: Remove all edges connecting newly added nodes to other nodes in the visited node set from the candidate edge set, and update the current direct connection count record for all nodes in the visited node set.

[0133] Step S29: Repeat steps S26 to S28 until the set of visited nodes contains all nodes, and output the initial topology structure consisting of all selected edges.

[0134] Preferably, steps S24 to S29 constitute the calculation flow of the Prim improved algorithm. Based on the examples above, the calculation flow of the Prim improved algorithm is mainly as follows:

[0135] ① Initialization: Define an empty edge set E, visited node U = {randomly selected Server_1}, and candidate edge set C = {edges between Server_1 and other nodes};

[0136] ② Iterative edge selection (limiting the number of connections per node to k=3):

[0137] Select the edge with the largest weight from C (e.g., Server_1-2, weight 2.777).

[0138] Check if the current number of connections (0) on the target node (Server_2) is less than 3 → Allow.

[0139] Add the edge to E, and add Server_2 to U.

[0140] Update C (remove the edge between Server_2 and the node inside U, and add the edge between Server_2 and the unvisited node).

[0141] ③ Termination: When U contains all nodes (e.g., Server_1→2→3→4), output the initial topology formed by the edge set E (e.g., E={(1-2), (1-3), (2-4)}).

[0142] Preferably, the computation flow of the Prim improved algorithm is shown in the following simplified pseudocode:

[0143] def prim_improved(nodes, edges, k=3):

[0144] U = {np.random.choice(nodes)} # Random starting nodes

[0145] E = [] # Initial topological edge set

[0146] degree = {n:0 for n in nodes} # Number of node connections

[0147] while len(U) < len(nodes):

[0148] # Candidate edges: Edges between U and unvisited nodes

[0149] candidates = [(edges[(u,v)], u, v) for u in U for v in nodesif v not in U and (u,v) in edges]

[0150] If not candidates: break

[0151] # The edge with the largest selection weight

[0152] w_max, u, v = max(candidates, key=lambda x: x[0])

[0153] if degree[v] < k: # Check connection limit

[0154] E.append((u, v))

[0155] U.add(v)

[0156] degree[u] += 1

[0157] degree[v] += 1

[0158] return E

[0159] Beneficial effects:

[0160] Steps S21 to S29 intelligently construct the initial topology of the computing server cluster using an improved Prim algorithm. Steps S21 to S23 define graph nodes and calculate weights, providing structured input for subsequent operations. Steps S24 to S28 construct a complete iterative optimization process, where a dynamic update mechanism for the visited node set ensures topology connectivity, real-time maintenance of the candidate edge set improves search efficiency, and continuous updating of connection count records ensures strict enforcement of constraints. The synergistic effect of these steps enables the algorithm to efficiently traverse all possible connection combinations, ultimately outputting the optimal topology covering all nodes in step S29. Firstly, a composite weight design achieves a balance between network performance and geographical constraints. The basic weights are the product of computing power and bandwidth potential, directly reflecting the computational and transmission coordination capabilities of nodes. The auxiliary weights use the reciprocal of the geographical location parameter, effectively suppressing latency issues caused by long-distance connections. This weight design ensures that the generated topology significantly reduces the impact of physical distance on communication quality while maintaining computational efficiency. Secondly, the introduced single-node connection limit mechanism overcomes the limitations of the traditional Prim algorithm. By dynamically verifying the number of connected edges at both ends of a candidate edge, it ensures that the number of direct connections for each node does not exceed a preset threshold. This constraint not only avoids over-connected hub nodes in the network but also evenly distributes data transmission pressure, giving the topology better load balancing characteristics. Finally, by prioritizing the selection of nodes with the largest weights and fewer connections, the goal of maximizing the sum of basic and auxiliary weights is achieved. This optimization strategy ensures that the generated initial topology achieves optimal overall performance, providing a high-quality input foundation for subsequent graph neural network analysis.

[0161] Step S3: Using a graph neural network model, the feature vector of each undirected graph node in the standardized node feature set is used as a node attribute, and each topological adjacency relationship in the initial topology is used as an edge attribute. After aggregating the neighbor information through the network layer, the expected communication efficiency of each link and the overall topology vulnerability index are obtained.

[0162] Further, in step S3, the feature vector of each undirected graph node in the standardized node feature set is used as a node attribute, and each topological adjacency relationship in the initial topology is used as an edge attribute through a graph neural network model. After aggregating the neighbor information through the network layer, the expected communication efficiency of each link and the overall topology vulnerability index are obtained. Specifically, this includes the following steps:

[0163] Step S31: Construct a graph neural network model architecture with three graph convolutional layers, each graph convolutional layer having its own independent weight matrix.

[0164] Preferably, the graph neural network model architecture (GNN) has three graph convolutional layers (GCN), in which the weight matrices W1, W2, and W3 of the three GCN layers are independent, the activation function of the hidden layer is the ReLU activation function, and the output layer is the Sigmoid function.

[0165] Step S32: Use the feature vector of each undirected graph node in the standardized node feature set as the node attribute of the graph neural network model, and use all topological adjacency relationships in the initial topology as edge attributes.

[0166] Step S33: In the first graph convolutional layer, neighbor information aggregation is performed on each undirected graph node. The feature vectors of adjacent nodes connected by edge attributes are linearly transformed by the first layer weight matrix and then summed with the feature vectors of the corresponding center nodes of the adjacent nodes. The first layer node embedding representation is obtained by processing through the activation function.

[0167] Step S34: Input the first layer node embedding representation into the second graph convolutional layer, perform neighbor information aggregation again and extend the aggregation range to two-hop neighbor nodes, and process the aggregated feature vector through the second layer weight matrix linear transformation and activation function to obtain the second layer node embedding representation.

[0168] Step S35: Input the second layer node embedding representation into the third graph convolutional layer, concatenate the embedding representations of each node and its corresponding direct neighbors, and map the concatenated vector to a scalar output through a fully connected layer. The scalar output is the evaluation value of the expected communication efficiency.

[0169] Step S36: Calculate the global graph-level representation vector based on the second-layer node embedding representation of all nodes.

[0170] Step S37: Input the global graph-level representation vector into a multilayer perceptron with two output nodes. One of the output nodes is processed by the Sigmoid function to obtain the overall topological fragility index.

[0171] Preferably, steps S32 to S37 specifically constitute the forward propagation process of the GNN model, wherein:

[0172] The node attributes of the input data are standardized feature sets, and the edge attributes are initial topological adjacency relationships.

[0173] In the first-layer GCN (one-hop aggregation), the feature of node u is the weighted sum of its own feature and the features of its neighbors: u_h1 = ReLU[W1·(u_h0+Σv_h0)], and node v is a neighbor of u.

[0174] In the second-layer GCN (two-hop aggregation), the features of node u are extended to its two-hop neighbors: u_h2=ReLU[W2·(u_h1+Σv_h1)].

[0175] The third layer outputs the link efficiency by concatenating the h2 values ​​of the two nodes at both ends of the edge: h_concat=concat(u_h2,v_h2), and mapping it to a scalar eff(uv)=sigmoid(W3·h_concat) through the fully connected layer, which is the expected communication efficiency.

[0176] The topological fragility index is calculated by the global average of all nodes h2: inputting an MLP (2 layers of ReLU + Sigmoid) yields the topological fragility index F∈[0,1], and the larger the value, the more fragile the topology.

[0177] Preferably, the simplified Python code for steps S32 to S37 is as follows:

[0178] import torch

[0179] from torch_geometric.nn import GCNConv

[0180] class GNN(torch.nn.Module):

[0181] def __init__(self, in_dim=3, hid_dim=16):

[0182] super().__init__()

[0183] self.conv1 = GCNConv(in_dim, hid_dim) # First layer GCN

[0184] self.conv2 = GCNConv(hid_dim, hid_dim) # Second layer GCN

[0185] self.fc_eff = torch.nn.Linear(2*hid_dim, 1) # Link efficiency layer

[0186] self.mlp_frag = torch.nn.Sequential( # Vulnerability MLP)

[0187] torch.nn.Linear(hid_dim, hid_dim / / 2), torch.nn.ReLU(),

[0188] torch.nn.Linear(hid_dim / / 2, 1), torch.nn.Sigmoid() )

[0190] def forward(self, x, edge_index):

[0191] h1 = torch.relu(self.conv1(x, edge_index)) # First layer output

[0192] h2 = torch.relu(self.conv2(h1, edge_index)) # Second layer output

[0193] # Calculate link efficiency (all edges)

[0194] src, dst = edge_index

[0195] h_concat = torch.cat([h2[src], h2[dst]], dim=1)

[0196] eff = torch.sigmoid(self.fc_eff(h_concat)).squeeze()

[0197] # Calculate the vulnerability index (global average)

[0198] frag = self.mlp_frag(h2.mean(dim=0)).squeeze()

[0199] return eff, frag

[0200] Beneficial effects:

[0201] Steps S31 to S37 construct a graph neural network model to achieve intelligent evaluation of the communication network performance of computing servers. The three-layer graph convolutional architecture in step S31 provides sufficient depth for feature learning; the node and edge attribute definitions in step S32 clearly define the structured representation of the input data; the stepwise information aggregation in steps S33 to S35 achieves feature extraction from local to global perspectives; the global graph-level representation computation in step S36 provides the foundation for vulnerability assessment; and the dual-output design in step S37 enables the model to simultaneously optimize communication efficiency and network robustness. The synergistic effect of these steps allows the model to accurately capture complex dependencies in the network, and the evaluation results reflect both the intrinsic performance of nodes and the propagation characteristics of the network topology.

[0202] Step S4: Obtain the real-time load data of each computing server, and calculate the link priority ranking of each undirected graph node by combining the expected communication efficiency and the overall topology vulnerability index.

[0203] Further, in step S4, the real-time load data of each computing server is obtained, and the link priority ranking of each undirected graph node is calculated by combining the expected communication efficiency and the overall topology vulnerability index. This specifically includes the following steps:

[0204] Step S41: Obtain real-time load data of each computing server through an external cluster monitoring system, and combine all real-time load data of a computing server into a real-time load data vector.

[0205] Step S42: Normalize the real-time load data vector to obtain a standardized real-time load vector.

[0206] Preferably, real-time load processing can be achieved by collecting the real-time load of the node, for example, Server_1's CPU=70%, memory=60%, IO=50Gbps, to obtain the normalized vector L_1=[0.7,0.6,0.5].

[0207] Step S43: Take the arithmetic mean of the standardized real-time load vectors of the nodes at both ends of the current link as the load factor of the current link, and weight and fuse it with the expected communication efficiency to obtain the comprehensive link evaluation vector.

[0208] Step S44: The overall topology vulnerability index is introduced as an adjustment factor into the comprehensive link evaluation vector, and the comprehensive link evaluation vector is corrected by the exponential decay function to obtain the corrected comprehensive link evaluation vector.

[0209] Preferably, the weighted fusion can be calculated by S_i=A×eff_i+ (1-A)×L_i, where A is the priority efficiency, A=0.6, for example eff_1=0.8, L_1=0.7, resulting in S_i=0.6×0.8+0.4×0.7=0.76; the vulnerability correction can be calculated by S_i'=S_i×exp(-B×F), where B is the vulnerability penalty, B=0.5, for example the overall topological vulnerability index F=0.3, resulting in exp(-0.15)=0.861, resulting in S_1'=0.76×0.861≈0.654.

[0210] Step S45: Calculate the difference between the corrected integrated link evaluation vector of the current node and the standardized real-time load vector of the neighboring nodes, and use it as the link value score of the current node and all its neighboring nodes.

[0211] Step S46: Sort all neighboring nodes of the current node in descending order according to the link value score to obtain the link priority sort of the current node.

[0212] Preferably, the priority sorting can be calculated by the link value score between node u and its neighbor v: Score(uv) = S_u' - L_v'; the link priority of u is obtained by sorting the neighbors in descending order of the score. For example, if neighbor v_1 has a score of 0.1, v_2 has a score of 0.05, and v_3 has a score of -0.02, then the sorting is v_1 > v_2 > v_3.

[0213] Beneficial effects:

[0214] Steps S41 to S46 achieve intelligent prioritization of computing server cluster links by dynamically integrating real-time load status and network performance indicators. Step S41 uses load data collected in real-time by the cluster monitoring system, which complements the communication efficiency indicators predicted by the graph neural network introduced in step S43. This ensures that the priority ranking reflects both the current intensity of computing resource usage and the potential capacity of network transmission. This dynamic fusion mechanism enables the network scheme to quickly respond to server load fluctuations and avoid communication bottlenecks caused by local overload. Secondly, the overall topology vulnerability index introduced in step S44 serves as an adjustment factor. The comprehensive evaluation results are corrected through an exponential decay function, allowing the priority ranking to optimize communication efficiency while proactively avoiding vulnerable links in the network. This design significantly improves the robustness of the network scheme. When some links experience performance degradation, the system can automatically adjust the data transmission path, reducing the impact of single-point failures on overall communication. Finally, the link value scoring and priority ranking mechanism constructed in steps S45 and S46, by quantifying the differences in communication value between nodes, provides a precise basis for action selection for the deep Q-network decision model. This allows subsequent topology optimization to focus on adjusting critical links, improving algorithm iteration efficiency. The synergistic effect of these steps gives the networking scheme adaptive characteristics, enabling it to dynamically adjust communication strategies according to real-time load changes, ensuring efficient execution of computing tasks while maintaining network stability. This process not only improves the reliability of data transmission but also lays a solid foundation for subsequent topology optimization, enabling the computing server cluster to maintain efficient communication under complex workloads.

[0215] Step S5: Define the action space and reward function based on the pre-trained deep Q-network decision model, input the link priority sorting into the deep Q-network decision model, and iterate the link priority sorting through all action spaces until the reward function reaches its maximum value, thus obtaining the optimized dynamic topology.

[0216] Further, in step S5, the action space and reward function are defined based on the pre-trained deep Q-network decision model. The link priority ranking is input into the deep Q-network decision model and the link priority ranking is iterated through all action spaces until the reward function reaches its maximum value, thus obtaining the optimized dynamic topology. This specifically includes the following steps:

[0217] Step S51 defines the deep Q network decision model as having three types of action spaces: a subset of link addition and deletion actions, a subset of node grouping and merging actions, and a subset of cross-group direct connection actions.

[0218] Step S52, define the reward function of the deep Q network decision model as the sum of the communication efficiency improvement value and the reconstruction energy consumption reduction value minus the overall topology vulnerability index.

[0219] Preferably, the action space is ① link addition / deletion (adding / deleting edges), ② node grouping and merging (merging two nodes into a group), ③ cross-group direct connection (adding edges between groups); the reward function R can be set as R=(ΔEff+ΔEnergy)-F, where ΔEff=communication efficiency improvement value, ΔEnergy=reconstruction energy consumption reduction value, and F=vulnerability index.

[0220] Preferably, the communication efficiency improvement value ΔEff can be defined as a relative increment, that is, the ratio of the improvement of the new topology to the old topology ΔEff=(Eff_new−Eff_old) / Eff_old, where Eff_old≠0. If the extreme case of Eff_old=0 occurs, then ΔEff=Eff_new.

[0221] Before executing the DQN selected action (in the old topology state), Eff_old calls the graph neural network model in step S3, inputs the node feature set and adjacency relationships of the old topology, outputs the expected communication efficiency of all links, and calculates the arithmetic mean:

[0222] ,in, This represents the total number of links in the old topology. Let be the expected communication efficiency of the i-th link. The output is obtained from step S35.

[0223] In this process, after Eff_new performs the selected action (link addition / deletion / node merging / cross-group direct connection), an intermediate topology (new topology) is obtained. The graph neural network model in step S3 is called again, and the node features and adjacency relationships of the new topology are input. Similarly, the average value of the expected communication efficiency of all links is calculated.

[0224] ,in, This represents the total number of links in the old topology. Let be the expected communication efficiency of the j-th new link.

[0225] Preferably, ΔEnergy measures the relative reduction in energy consumption of the new topology compared to the old topology. Its core is to quantify the energy consumption changes (including link activation, node scheduling, and merging overhead) during topology maintenance and reconstruction. Therefore, the energy consumption model can consist of three parts, superimposed in a linear relationship:

[0226] ,in, To activate the total number of links, the energy consumption coefficient per unit time for each link is α, for example, α = 0.1W / Gbps for an optical fiber link; The number of independent nodes is given by the node management energy consumption coefficient β, for example, β = 5 watts for a single node; The number of node mergings is γ, and the energy consumption coefficient for each merging scheduling is γ, for example, γ=20 watts, including the overhead of computing resource coordination. The specific accurate data can be obtained through actual collection. This embodiment is only for illustrative purposes.

[0227] Similarly, the relative energy reduction (ΔEnergy) is defined as the proportion of energy consumed by the old topology compared to the new topology. ΔEnergy = (Energy old − Energy new) / Energy old, where Energy old ≠ 0. If Energy old = 0, then ΔEnergy = −Energy new.

[0228] Among them, the old topology energy consumption (Energy_old) is based on the old topology structure before the action is executed, and the number of old links is counted. Number of old nodes Number of historical mergers Substitute into the energy consumption model:

[0229] .

[0230] Among them, the new topology energy consumption (Energy_new) is calculated by counting the number of new links in the new topology after the selected action is executed. Number of new nodes Number of new mergers Substitute into the energy consumption model:

[0231] .

[0232] In summary, the sign conventions for ΔEff and ΔEnergy are as follows: a positive ΔEff indicates improved efficiency, and a positive ΔEnergy indicates reduced energy consumption. Together, they constitute the "benefit term" of the reward function. The energy consumption model parameters (α, β, γ) need to be fitted using historical operation and maintenance data (such as link power consumption testing and node management energy consumption monitoring) to ensure that the calculation closely reflects the real-world scenario. The reward function R = ΔEff + ΔEnergy − F uses a negative sign to transform "vulnerability deterioration" into a "penalty term," achieving a multi-objective balance between efficiency, energy consumption, and reliability.

[0233] Step S53: Convert the link priority sorting into the initial state representation of the deep Q network decision model, and input the initial state representation into the deep Q network decision model.

[0234] Step S54: In each decision cycle of the deep Q-network decision model, calculate the Q value of each sub-action in the three action spaces according to the state representation of the current decision cycle, and select the sub-action with the largest Q value as the selected action for the current decision cycle.

[0235] Step S55: After executing the selected action, add or delete links, group or merge nodes, or directly connect across groups according to the action type of the selected action to obtain the intermediate topology.

[0236] Step S56: Recalculate the reward values ​​of the two reward functions based on the intermediate topology, and sum the two reward values ​​as the immediate reward for the current decision cycle, while obtaining the updated state representation.

[0237] Step S57: Select the actions, immediate rewards, and updated state representations for the same decision cycle into an experience tuple, and randomly select an experience tuple to update the deep Q-network decision model using gradient descent.

[0238] Step S58: Repeat steps S54 to S57 to iteratively optimize the link priority ranking through the deep Q-network decision model until the output values ​​of the two reward functions reach their respective historical maximum values ​​and remain stable.

[0239] Step S59: Output the network topology corresponding to the fact that the output values ​​of the two reward functions have reached their respective historical maximum values ​​and remain stable, as the optimized dynamic topology.

[0240] Preferably, steps S53 to S59 constitute the DQN decision-making process. Based on the example above, the specific DQN decision-making process is as follows:

[0241] ① Initialization: Encode the link priority sorting into a state vector (e.g., embedding in 128 dimensions) and input it into the pre-trained DQN.

[0242] ② Iterative optimization:

[0243] The Q-values ​​for three actions are calculated in each round, such as adding / deleting links (Q=0.8), merging (Q=0.6), and cross-group (Q=0.7).

[0244] Choose the action with the highest Q value to execute, such as adding or deleting links by adding an edge between Server_1 and Server_4.

[0245] Calculate the immediate reward R and store the experience tuple (state, action, R, new state) into the experience pool.

[0246] Randomly sample batches of experience and update DQN using gradient descent.

[0247] ③ Termination: When the reward reaches its historical maximum value for 10 consecutive rounds, the optimized dynamic topology is obtained.

[0248] Beneficial effects:

[0249] Steps S51 to S59 achieve dynamic and intelligent optimization of the computing server cluster communication network scheme through iterative optimization of the deep Q-network decision model. Firstly, comprehensive network topology optimization is achieved through the synergistic effect of multiple action spaces. The link addition / deletion action space directly adjusts physical connection relationships; the node grouping / merging action space optimizes data routing paths through logical grouping; and the cross-group direct connection action space breaks down traditional grouping communication barriers. The combination of these three action spaces enables the model to intervene at multiple levels, from local connection adjustments to global structure optimization, improving the adaptability and flexibility of the network topology. This allows the network scheme to automatically adjust the communication structure according to real-time load changes, maintaining efficient and stable data transmission performance. Secondly, the designed dual-reward function mechanism achieves a balanced optimization of communication efficiency and network robustness. The sum of the improvement in communication efficiency and the reduction in reconstruction energy consumption serves as a positive reward, guiding the model to prioritize topology adjustment strategies that improve data transmission efficiency and reduce energy consumption. Meanwhile, the negative value of the overall topology vulnerability index acts as a negative constraint, prompting the model to actively avoid fragile connections that could lead to network outages. This combined positive and negative reward design allows the optimization process to simultaneously consider performance improvement and reliability assurance, avoiding the problem of traditional methods that overemphasize a single metric while neglecting network stability. Finally, the model achieves adaptive optimization in complex network environments through a continuous learning mechanism of experience replay and gradient descent. The selected action, immediate reward, and updated state representation in each decision cycle are stored as experience tuples. Model parameters are updated through random sampling and gradient descent. This mechanism allows the model to continuously accumulate optimization experience and gradually converge to the optimal solution. The final optimized dynamic topology not only meets the communication requirements of the current workload but also possesses robustness against future load changes.

[0250] Step S6: Based on the optimized dynamic topology, configure at least two independent physical paths without shared intermediate nodes for each pair of nodes using the multipath transmission control protocol to obtain the multipath mapping table networking scheme.

[0251] Further, in step S6, based on the optimized dynamic topology, at least two independent physical paths without shared intermediate nodes are configured for each pair of nodes using the multipath transmission control protocol to obtain a multipath mapping table networking scheme, which specifically includes the following steps:

[0252] Step S61: Extract the direct and indirect connections between all node pairs from the optimized dynamic topology and construct a global reachability matrix.

[0253] Step S62: For each pair of nodes, traverse all transportable paths in the global reachability matrix and constrain the transportable paths to consist of continuous physical links and not contain duplicate nodes.

[0254] Step S63: Perform intermediate node sharing detection on all transmittable paths of each node pair, and filter out path combinations that do not have shared intermediate nodes.

[0255] Preferably, path traversal and filtering can be performed by extracting the global reachability matrix from the optimized topology, for example, the paths from Server_1 to 4 are 1-2-4 and 1-3-4; traversing all simple paths (without duplicate nodes), filtering combinations with no shared intermediate nodes, for example, the intermediate node of 1-2-4 is 2, and the intermediate node of 1-3-4 is 3, which are unshared, so they are retained.

[0256] Step S64: From the path combinations that do not share intermediate nodes, select at least two independent physical paths for each node pair based on the combination with the fewest path hops.

[0257] Step S65: Configure the selected independent physical path for each group of node pairs through the multipath transmission control protocol and establish the mapping relationship between path identifiers and physical links.

[0258] Step S66: Store the mapping relationships of all node pairs in groups according to the source node to obtain a multi-path mapping table networking scheme that includes path priority and link status.

[0259] Preferably, the path selection and configuration selects the path combination with the fewest hops. For example, if 1-2-4 and 1-3-4 both have 2 hops, then these two are selected. Multipath TCP (MPTCP) is used to configure the path, and a path ID-physical link mapping is established. For example, Path_1-4,1=1→2→4 and Path_1-4,2=1→3→4. The mapping relationship (including path priority and link status) is stored in groups according to the source node to obtain the multipath mapping table. For example, the entry for destination node 4 of source node 1 contains 2 paths.

[0260] Beneficial effects:

[0261] Steps S61 to S66 achieve high reliability and fault tolerance in the communication network of the computing server cluster through a multi-path transmission control protocol. Step S61 extracts direct and indirect connections from the optimized dynamic topology, ensuring the comprehensiveness of the path search, while step S62's constraint on the traversal of continuous physical links guarantees the physical feasibility of the paths. Furthermore, a rigorous path selection mechanism ensures the independence of multiple paths. Step S63's intermediate node sharing detection technically eliminates the risk of resource contention between paths. Step S64's optimization strategy, based on minimizing path hops, further optimizes transmission efficiency while satisfying the constraint of no shared nodes. This dual selection mechanism ensures that the configured independent physical paths possess both fault isolation capabilities and maintain the lowest possible transmission latency. Finally, steps S65 to S66, through protocol implementation and mapping table generation, transform the theoretical path configuration into an executable networking scheme. The specific implementation of the multipath transmission control protocol establishes a reliable correspondence between path identifiers and physical links, while the systematic storage of the multipath mapping table enables centralized management of network status and provides a structured basis for dynamic routing adjustments. This complete closed loop from topology analysis to protocol implementation enables the computing server cluster to automatically switch faulty paths without interrupting service, significantly improving the continuous operation capability of critical computing tasks.

[0262] It is worth noting that the data examples in this embodiment are only for illustrating the principle, and the actual data needs to be obtained through data collection.

[0263] Beneficial effects of steps S1 to S6:

[0264] Steps S1 to S6 systematically achieve efficient communication networking of the computing server cluster. Specifically, Step S1 uses multi-dimensional parameter collection and principal component analysis to construct a standardized set of node features, enabling quantitative evaluation of server performance and feature dimensionality reduction, providing a data foundation for subsequent topology construction. Secondly, Step S2's improved Prim algorithm introduces composite weights and connection number constraints to traditional graph theory methods, generating an initial topology that optimizes network performance while avoiding single-point overload. Next, Step S3 employs a graph neural network for intelligent topology analysis, accurately predicting link communication efficiency and global vulnerability index through neighbor information aggregation, achieving intelligent evaluation of network performance. Step S4 then establishes an adaptive link priority mechanism through dynamic fusion of real-time load data and prediction results, enabling the network to respond to real-time workload changes. Step S5 uses a deep Q-network decision model for iterative optimization in a multi-action space, achieving a balance between communication efficiency and topology robustness, resulting in a self-optimizing dynamic topology. Finally, Step S6, through a multi-path configuration scheme and independent path design without shared intermediate nodes, significantly improves the network's fault tolerance and transmission reliability. The overall solution combines data-driven approaches with intelligent algorithms to achieve self-awareness, self-optimization, and self-recovery of high-performance computing networks, providing computing clusters with high-throughput, low-latency, and fault-tolerant communication infrastructure.

[0265] like Figure 2 As shown, this embodiment provides an example of a communication networking device for a computing server. In this embodiment, the communication networking device is applied to the communication networking method as described in the above embodiment.

[0266] Specifically, the communication networking device includes a standardized node feature set construction module 1, an initial topology search module 2, a topology link parameter calculation module 3, a link priority ranking calculation module 4, a link priority ranking iteration module 5, and an independent physical path allocation module 6.

[0267] The standardized node feature set construction module 1 collects the geographical location parameters, core calculation parameters, and data transmission parameters of all computing power servers to be networked. Through principal component analysis and standardization, it constructs a standardized node feature set including computing power intensity, bandwidth potential, and load sensitivity. The initial topology search module 2 defines each computing power server as an undirected graph node, using the product of computing power intensity and bandwidth potential as the basic weight and the reciprocal of the geographical location parameter as the auxiliary weight. Using the Prim improved algorithm, it aims to limit the number of direct connections of a single node to no more than a preset threshold and maximize the sum of the basic and auxiliary weights to find an initial topology covering all undirected graph nodes. The topology link parameter calculation module 3 uses a graph neural network model to calculate the feature vector of each undirected graph node in the standardized node feature set as node attributes and the topology link parameter of each node in the initial topology. Adjacency relationships are used as edge attributes. After aggregating neighbor information at the network layer, the expected communication efficiency of each link and the overall topology vulnerability index are obtained. The link priority ranking calculation module 4 is used to obtain the real-time load data of each computing server and calculate the link priority ranking of each undirected graph node by combining the expected communication efficiency and the overall topology vulnerability index. The link priority ranking iteration module 5 is used to define the action space and reward function based on the pre-trained deep Q network decision model, input the link priority ranking into the deep Q network decision model, and iterate the link priority ranking through all action spaces until the reward function reaches its maximum value, thus obtaining the optimized dynamic topology. The independent physical path allocation module 6 is used to configure at least two shared intermediate node independent physical paths for each pair of nodes based on the optimized dynamic topology through the multipath transmission control protocol, thus obtaining the multipath mapping table networking scheme.

[0268] Furthermore, the standardized node feature set construction module 1 specifically includes a first standardized node feature set construction unit, a second standardized node feature set construction unit, a third standardized node feature set construction unit, a fourth standardized node feature set construction unit, a fifth standardized node feature set construction unit, a sixth standardized node feature set construction unit, and a seventh standardized node feature set construction unit that are electrically or signal-connected in sequence; the seventh standardized node feature set construction unit is electrically or signal-connected to the initial topology search module 2.

[0269] The system comprises four main components: a first standardized node feature set construction unit, a second standardized node feature set construction unit, and a third standardized node feature set construction unit. The first standardized node feature set construction unit collects the geographical location parameters, core calculation parameters, and data transmission parameters of all computing power servers to be networked, and categorizes and stores all parameters according to the node number of each server to form the original parameter dataset. The second standardized node feature set construction unit performs data interpolation and preprocessing on the original parameter dataset to obtain a normalized parameter matrix. The third standardized node feature set construction unit inputs the normalized parameter matrix into a pre-trained principal component analysis model to calculate the covariance matrix. The fourth standardized node feature set construction unit sorts the covariance matrix by its eigenvalues ​​from largest to smallest. The first unit selects and retains the top principal components whose cumulative contribution rate exceeds a preset percentage threshold to construct a dimension-reduced projection matrix; the fifth standardized node feature set construction unit is used to perform a linear transformation on the normalized parameter matrix through the dimension-reduced projection matrix to obtain the principal component score vector of each node; the sixth standardized node feature set construction unit is used to define the first principal component as computing power intensity, the second principal component as bandwidth potential, and the third principal component as load sensitivity according to the physical meaning of the principal components; the seventh standardized node feature set construction unit is used to standardize computing power intensity, bandwidth potential, and load sensitivity respectively, and integrate them to obtain a standardized node feature set.

[0270] Furthermore, the initial topology search module 2 specifically includes a first initial topology search unit, a second initial topology search unit, a third initial topology search unit, a fourth initial topology search unit, a fifth initial topology search unit, a sixth initial topology search unit, a seventh initial topology search unit, an eighth initial topology search unit, and a ninth initial topology search unit that are electrically or signal-connected in sequence; the first initial topology search unit is electrically or signal-connected to the seventh standardized node feature set construction unit, and the ninth initial topology search unit is electrically or signal-connected to the topology link parameter calculation module 3.

[0271] The initial topology search unit is used to define each computing server in the standardized node feature set as an undirected graph node, constructing a node set including all nodes. The second initial topology search unit extracts the computing power intensity and bandwidth potential of each node and calculates their product to obtain the basic weight of each node. The third initial topology search unit extracts the geographical location parameters of each node and takes their reciprocals to obtain the auxiliary weight of each node. The fourth initial topology search unit initializes an empty edge set as the edge set of the initial topology, creates a visited node set, and randomly selects a node to add to the visited node set. The fifth initial topology search unit creates a candidate edge set, which includes all edges connecting the visited node set and the unvisited node set, with the weight of the current edge being the base weight of the two nodes connected by the current edge. The sum of the weight and auxiliary weights; the sixth initial topology search unit is used to select the two endpoints of the edge with the largest current weight in the candidate edge set to further select the nodes with fewer connected edges among the two endpoint nodes, and to limit the current direct connection number of the selected nodes to less than a preset threshold; the seventh initial topology search unit is used to add the edge with the largest current weight to the edge set, and add the unvisited nodes connected by the edge with the largest current weight to the visited node set; the eighth initial topology search unit is used to remove all edges connecting the newly added nodes to other nodes in the visited node set from the candidate edge set, and update the current direct connection number record of all nodes in the visited node set; the ninth initial topology search unit is used to repeat the execution of the sixth to the eighth initial topology search units until the visited node set contains all nodes, and output the initial topology structure composed of all selected edges.

[0272] Furthermore, the topology link parameter calculation module 3 specifically includes a first topology link parameter calculation unit, a second topology link parameter calculation unit, a third topology link parameter calculation unit, a fourth topology link parameter calculation unit, a fifth topology link parameter calculation unit, a sixth topology link parameter calculation unit, and a seventh topology link parameter calculation unit that are electrically or signal-connected in sequence; the first topology link parameter calculation unit is electrically or signal-connected to the ninth initial topology structure finding unit, and the seventh topology link parameter calculation unit is electrically or signal-connected to the link priority sorting calculation module 4.

[0273] The system comprises the following components: a first topology link parameter calculation unit, which constructs a graph neural network model architecture with three graph convolutional layers, each configured with an independent weight matrix; a second topology link parameter calculation unit, which uses the feature vector of each undirected graph node in the standardized node feature set as a node attribute of the graph neural network model and uses all topological adjacency relationships in the initial topology as edge attributes; a third topology link parameter calculation unit, which performs neighbor information aggregation on each undirected graph node in the first graph convolutional layer, and after linear transformation of the feature vectors of adjacent nodes connected by edge attributes through the first layer weight matrix, sums them with the feature vectors of the corresponding center nodes of the adjacent nodes, and obtains the first layer node embedding representation through activation function processing; and a fourth topology link parameter calculation unit, which inputs the first layer node embedding representation into the second graph convolutional layer. The neighbor information aggregation is performed again, and the aggregation range is extended to two-hop neighbor nodes. The aggregated feature vector is processed by the second-layer weight matrix linear transformation and activation function to obtain the second-layer node embedding representation. The fifth topology link parameter calculation unit is used to input the second-layer node embedding representation into the third graph convolutional layer, and concatenate the embedding representation of each node and its corresponding direct neighbor. The concatenated vector is mapped to a scalar output through a fully connected layer. The scalar output is the evaluation value of the expected communication efficiency. The sixth topology link parameter calculation unit is used to calculate the global graph-level representation vector based on the second-layer node embedding representation of all nodes. The seventh topology link parameter calculation unit is used to input the global graph-level representation vector into a multilayer perceptron with two output nodes. One of the output nodes is processed by the Sigmoid function to obtain the overall topology vulnerability index.

[0274] Furthermore, the link priority ranking calculation module 4 specifically includes a first link priority ranking calculation unit, a second link priority ranking calculation unit, a third link priority ranking calculation unit, a fourth link priority ranking calculation unit, a fifth link priority ranking calculation unit, and a sixth link priority ranking calculation unit that are electrically or signal-connected in sequence; the first link priority ranking calculation unit is electrically or signal-connected to the seventh topology link parameter calculation unit, and the sixth link priority ranking calculation unit is electrically or signal-connected to the link priority ranking iteration module 5.

[0275] The system comprises the following components: a first link priority ranking calculation unit, which acquires real-time load data from each computing server through an external cluster monitoring system and combines all real-time load data from a single computing server into a real-time load data vector; a second link priority ranking calculation unit, which normalizes the real-time load data vector to obtain a standardized real-time load vector; a third link priority ranking calculation unit, which takes the arithmetic mean of the standardized real-time load vectors of the nodes at both ends of the current link as the load factor of the current link and weights it with the expected communication efficiency to obtain a comprehensive link evaluation vector; a fourth link priority ranking calculation unit, which introduces the overall topology vulnerability index as an adjustment factor into the comprehensive link evaluation vector and corrects the comprehensive link evaluation vector through an exponential decay function to obtain a corrected comprehensive link evaluation vector; a fifth link priority ranking calculation unit, which calculates the difference between the corrected comprehensive link evaluation vector of the current node and the standardized real-time load vectors of its neighboring nodes, as the link value score of the current node and all its neighboring nodes; and a sixth link priority ranking calculation unit, which sorts all its neighboring nodes in descending order according to their link value scores to obtain the link priority ranking of the current node.

[0276] Furthermore, the link priority ranking iteration module 5 specifically includes a first link priority ranking iteration unit, a second link priority ranking iteration unit, a third link priority ranking iteration unit, a fourth link priority ranking iteration unit, a fifth link priority ranking iteration unit, a sixth link priority ranking iteration unit, a seventh link priority ranking iteration unit, an eighth link priority ranking iteration unit, and a ninth link priority ranking iteration unit that are electrically or signal-connected in sequence; the first link priority ranking iteration unit is electrically or signal-connected to the sixth link priority ranking calculation unit, and the ninth link priority ranking iteration unit is electrically or signal-connected to the independent physical path allocation module 6.

[0277] The deep Q-network decision model comprises the following components: The first link priority ranking iteration unit defines three action spaces: a subset of link addition / deletion actions, a subset of node grouping / merging actions, and a subset of cross-group direct connection actions. The second link priority ranking iteration unit defines the reward function of the deep Q-network decision model as the sum of the communication efficiency improvement and the reconstruction energy reduction, minus the overall topology vulnerability index. The third link priority ranking iteration unit converts the link priority ranking into the initial state representation of the deep Q-network decision model and inputs this initial state representation into the model. The fourth link priority ranking iteration unit calculates the Q-value of each sub-action in the three action spaces based on the current decision cycle's state representation in each decision cycle of the deep Q-network decision model, and selects the sub-action with the largest Q-value as the selected action for the current decision cycle. The fifth link priority ranking iteration unit performs link priority ranking after executing the selected action, based on the action type of the selected action. The process involves several steps: adding or deleting paths, grouping or merging nodes, or directly connecting across groups to obtain an intermediate topology; the sixth link priority ranking iteration unit recalculates the reward values ​​of the two reward functions based on the intermediate topology, sums the two reward values ​​as the immediate reward for the current decision period, and obtains the updated state representation; the seventh link priority ranking iteration unit groups the selected actions, immediate rewards, and updated state representations for the same decision period into an empirical tuple, and randomly selects one empirical tuple to update the deep Q-network decision model using gradient descent; the eighth link priority ranking iteration unit repeats steps S54 to S57 to iteratively optimize the link priority ranking using the deep Q-network decision model until the output values ​​of both reward functions reach their respective historical maximum values ​​and remain stable; the ninth link priority ranking iteration unit outputs the network topology corresponding to the point where the output values ​​of both reward functions reach their respective historical maximum values ​​and remain stable as the optimized dynamic topology.

[0278] Furthermore, the independent physical path allocation module 6 specifically includes a first independent physical path allocation unit, a second independent physical path allocation unit, a third independent physical path allocation unit, a fourth independent physical path allocation unit, a fifth independent physical path allocation unit, and a sixth independent physical path allocation unit that are electrically or signal-connected in sequence; the first independent physical path allocation unit is electrically or signal-connected to the ninth link priority sorting iteration unit.

[0279] The system comprises the following components: a first independent physical path allocation unit extracts the direct and indirect connections between all node pairs from the optimized dynamic topology to construct a global reachability matrix; a second independent physical path allocation unit traverses all transmittable paths for each node pair in the global reachability matrix, constraining transmittable paths to consist of continuous physical links without duplicate nodes; a third independent physical path allocation unit performs intermediate node sharing detection on all transmittable paths for each node pair, filtering out path combinations without shared intermediate nodes; a fourth independent physical path allocation unit selects at least two independent physical paths for each node pair from path combinations without shared intermediate nodes, using the path hop count as the criterion; a fifth independent physical path allocation unit configures the selected independent physical paths for each node pair using a multipath transmission control protocol, establishing a mapping relationship between path identifiers and physical links; and a sixth independent physical path allocation unit stores the mapping relationships of all node pairs in groups according to the source node, obtaining a multipath mapping table network scheme including path priority and link status.

[0280] It should be noted that this embodiment is a functional module embodiment based on the above method embodiment. For the preferred, extended, limited, exemplified, and principle-explained parts of this embodiment, please refer to the above embodiments. This embodiment will not repeat them here.

[0281] Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Figure 3 As shown, the electronic device 7 includes a processor 71 and a memory 72 coupled to the processor 71.

[0282] The memory 72 stores program instructions for implementing the federated learning-based collaborative energy-saving method for government data clusters in any of the above embodiments.

[0283] The processor 71 is used to execute program instructions stored in the memory 72 for collaborative energy saving of government data clusters based on federated learning.

[0284] The processor 71 can also be referred to as a CPU (Central Processing Unit). The processor 71 may be an integrated circuit chip with signal processing capabilities. The processor 71 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0285] Furthermore, Figure 4 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. See also: Figure 4 In this embodiment of the application, the storage medium 8 stores program instructions 81 capable of implementing all the above methods. These program instructions 81 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0286] In the several embodiments provided in this application, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only 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. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, signal, or other forms.

[0287] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A communication networking method for a computing server, the communication networking method being applied to a cluster of computing servers to be networked, the cluster of computing servers to be networked comprising a plurality of computing servers, characterized in that, The communication networking method includes: Step S1: Collect the geographical location parameters, core calculation parameters, and data transmission parameters of all computing power servers to be networked, and construct a standardized node feature set including computing power intensity, bandwidth potential, and load sensitivity through principal component analysis and standardization processing. Step S2: Define each computing server as an undirected graph node, using the product of the computing power intensity and the bandwidth potential as the basic weight and the reciprocal of the geographical location parameter as the auxiliary weight. Use the Prim improved algorithm to find an initial topology that covers all undirected graph nodes with the goal of limiting the number of direct connections of a single node to no more than a preset threshold and maximizing the sum of the basic weight and the auxiliary weight. Step S3: Using a graph neural network model, the feature vector of each undirected graph node in the standardized node feature set is used as a node attribute, and each topological adjacency relationship of the initial topology is used as an edge attribute. After aggregating the neighbor information through the network layer, the expected communication efficiency of each link and the overall topology vulnerability index are obtained. Step S4: Obtain the real-time load data of each computing server, and calculate the link priority ranking of each undirected graph node by combining the expected communication efficiency and the overall topology vulnerability index. Step S5: Define the action space and reward function based on the pre-trained deep Q network decision model, input the link priority sorting into the deep Q network decision model, and iterate the link priority sorting through all action spaces until the reward function reaches its maximum value to obtain the optimized dynamic topology; Step S6: Based on the optimized dynamic topology, configure at least two independent physical paths without shared intermediate nodes for each pair of nodes using the multipath transmission control protocol to obtain a multipath mapping table networking scheme.

2. The method of claim 1, wherein, Step S1: Collect the geographical location parameters, core calculation parameters, and data transmission parameters of all computing servers to be networked. Construct a standardized node feature set, including computing power intensity, bandwidth potential, and load sensitivity, through principal component analysis and standardization. Step S11: Collect the geographical location parameters, core calculation parameters, and data transmission parameters of all computing power servers to be networked, and classify and store all parameters according to the node number of each computing power server to be networked to form the original parameter dataset. Step S12: Perform data interpolation and preprocessing on the original parameter dataset to obtain a normalized parameter matrix; Step S13: Input the normalized parameter matrix into the pre-trained principal component analysis model to calculate the covariance matrix; Step S14: Sort the covariance matrix from largest to smallest according to its eigenvalues ​​and retain the first few principal components whose cumulative contribution rate exceeds a preset percentage threshold to construct a dimension-reduced projection matrix. Step S15: Perform a linear transformation on the normalized parameter matrix using the dimensionality reduction projection matrix to obtain the principal component score vector for each node; Step S16: Based on the physical meaning of the principal components, the first principal component is defined as the computing power intensity, the second principal component is defined as the bandwidth potential, and the third principal component is defined as the load sensitivity. Step S17: Standardize the computing power intensity, the bandwidth potential, and the load sensitivity respectively, and integrate them to obtain the standardized node feature set.

3. The method of claim 1, wherein, Step S2: Define each computing server as an undirected graph node. Use the product of the computing power intensity and the bandwidth potential as the basic weight, and the reciprocal of the geographical location parameter as the auxiliary weight. Using the Prim improved algorithm, with the goal of limiting the number of direct connections per node to no more than a preset threshold and maximizing the sum of the basic and auxiliary weights, find an initial topology covering all undirected graph nodes, including: Step S21: Define each computing server in the standardized node feature set as an undirected graph node, and construct a node set including all nodes; Step S22: Extract the computing power intensity and bandwidth potential of each node and calculate their product to obtain the basic weight of each node. Step S23: Extract the geographic location parameters of each node and take their reciprocals to obtain the auxiliary weights of each node; Step S24: Initialize an empty edge set as the edge set of the initial topology, create a set of visited nodes and randomly select a node to add to the set of visited nodes; Step S25: Create a candidate edge set, which includes all edges connecting the visited node set and the unvisited node set. The weight of the current edge is the sum of the basic weight and auxiliary weight of the two nodes connected to the current edge. Step S26: Select the two endpoints of the edge with the largest current weight in the candidate edge set to further select the nodes with fewer connected edges among the two endpoints, and restrict the current direct connection number of the selected nodes to be less than the preset threshold. Step S27: Add the edge with the largest current weight to the edge set, and add the unvisited node connected by the edge with the largest current weight to the visited node set; Step S28: Remove all edges connecting newly added nodes to other nodes in the visited node set from the candidate edge set, and update the current direct connection count record for all nodes in the visited node set. Step S29: Repeat steps S26 to S28 until the visited node set contains all nodes, and output the initial topology structure consisting of all selected edges.

4. The method of claim 1, wherein, Step S3: Using a graph neural network model, the feature vector of each undirected graph node in the standardized node feature set is used as a node attribute, and each topological adjacency relationship in the initial topology is used as an edge attribute. After aggregating neighbor information through the network layer, the expected communication efficiency of each link and the overall topology vulnerability index are obtained, including: Step S31: Construct a graph neural network model architecture with three graph convolutional layers, each graph convolutional layer being configured with an independent weight matrix; Step S32: The feature vector of each undirected graph node in the standardized node feature set is used as a node attribute of the graph neural network model, and all topological adjacency relationships in the initial topology are used as edge attributes; Step S33: In the first graph convolutional layer, neighbor information aggregation is performed on each undirected graph node. The feature vectors of adjacent nodes connected by edge attributes are linearly transformed by the first layer weight matrix and then summed with the feature vectors of the corresponding center nodes of the adjacent nodes. The first layer node embedding representation is obtained by processing through the activation function. Step S34: Input the first layer node embedding representation into the second graph convolutional layer, perform neighbor information aggregation again and extend the aggregation range to two-hop neighbor nodes, and process the aggregated feature vector through the second layer weight matrix linear transformation and the activation function to obtain the second layer node embedding representation; Step S35: Input the second layer node embedding representation into the third graph convolutional layer, concatenate the embedding representations of each node and its corresponding direct neighbors, and map the concatenated vector to a scalar output through a fully connected layer. The scalar output is the evaluation value of the expected communication efficiency. Step S36: Calculate the global graph-level representation vector based on the second-layer node embedding representation of all nodes; Step S37: Input the global graph-level representation vector into a multilayer perceptron with two output nodes. One of the output nodes is processed by the Sigmoid function to obtain the overall topological fragility index.

5. The method of claim 1, wherein, Step S4: Obtain the real-time load data of each computing server, and calculate the link priority ranking of each undirected graph node based on the expected communication efficiency and the overall topology vulnerability index, including: Step S41: Obtain the real-time load data of each computing server through an external cluster monitoring system, and combine all the real-time load data of a computing server into a real-time load data vector. Step S42: Normalize the real-time load data vector to obtain a standardized real-time load vector; Step S43: Take the arithmetic mean of the standardized real-time load vectors of the nodes at both ends of the current link as the load factor of the current link, and perform weighted fusion with the expected communication efficiency to obtain the comprehensive link evaluation vector. Step S44: The overall topology vulnerability index is introduced as an adjustment factor into the comprehensive link evaluation vector, and the comprehensive link evaluation vector is corrected by the exponential decay function to obtain the corrected comprehensive link evaluation vector. Step S45: Calculate the difference between the corrected integrated link evaluation vector of the current node and the standardized real-time load vector of the neighboring nodes, and use it as the link value score of the current node and all its neighboring nodes. Step S46: Sort all neighboring nodes of the current node in descending order according to the link value score to obtain the link priority sort of the current node.

6. The method of claim 1, wherein, Step S5: Define the action space and reward function based on the pre-trained deep Q-network decision model. Input the link priority ranking into the deep Q-network decision model and iterate the link priority ranking through all action spaces until the reward function reaches its maximum value, obtaining the optimized dynamic topology, including: Step S51: Define the deep Q-network decision model as having three types of action spaces, namely, a subset of link addition / deletion actions, a subset of node grouping / merging actions, and a subset of cross-group direct connection actions; Step S52, define the reward function of the deep Q network decision model as the sum of the communication efficiency improvement value and the reconstruction energy consumption reduction value minus the overall topology vulnerability index; Step S53: Convert the link priority sorting into the initial state representation of the deep Q network decision model, and input the initial state representation into the deep Q network decision model; Step S54: In each decision cycle of the deep Q-network decision model, calculate the Q value of each sub-action in the three action spaces according to the state representation of the current decision cycle, and select the sub-action with the largest Q value as the selected action of the current decision cycle. Step S55: After executing the selected action, add or delete links, group or merge nodes, or directly connect across groups according to the action type of the selected action to obtain an intermediate topology. Step S56: Recalculate the reward values ​​of the two reward functions according to the intermediate topology, and sum the two reward values ​​as the immediate reward for the current decision cycle, while obtaining the updated state representation; Step S57: Select the actions, immediate rewards, and updated state representations of the same decision cycle into an experience tuple, and randomly select an experience tuple to update the deep Q-network decision model using gradient descent. Step S58: Repeat steps S54 to S57 to iteratively optimize the link priority ranking through the deep Q network decision model until the output values ​​of the two reward functions reach their respective historical maximum values ​​and remain stable. Step S59: Output the network topology corresponding to the fact that the output values ​​of the two reward functions have reached their respective historical maximum values ​​and remain stable as the optimized dynamic topology.

7. The method of claim 1, wherein, Step S6: Based on the optimized dynamic topology, configure at least two independent physical paths without shared intermediate nodes for each pair of nodes using the multipath transmission control protocol to obtain a multipath mapping table networking scheme, including: Step S61: Extract the direct and indirect connection relationships between all node pairs from the optimized dynamic topology and construct a global reachability matrix; Step S62: For each pair of nodes, traverse all transportable paths in the global reachability matrix and constrain the transportable paths to consist of continuous physical links and not contain duplicate nodes. Step S63: Perform intermediate node sharing detection on all transmittable paths of each node pair, and filter out path combinations that do not have shared intermediate nodes. Step S64: From the path combinations that do not share intermediate nodes, select at least two independent physical paths for each pair of nodes based on the combination with the fewest path hops. Step S65: Configure the selected independent physical path for each group of node pairs through the multipath transmission control protocol, and establish the mapping relationship between path identifiers and physical links; Step S66: Store the mapping relationships of all node pairs in groups according to the source node to obtain a multi-path mapping table networking scheme that includes path priority and link status. 8.A computing power server communication networking device, applied to the communication networking method in any one of claims 1 to 7, characterized in that, The communication networking device includes: The standardized node feature set construction module is used to collect the geographical location parameters, core computing parameters, and data transmission parameters of all computing power servers to be networked. Through principal component analysis and standardization processing, it constructs a standardized node feature set including computing power intensity, bandwidth potential, and load sensitivity. The initial topology search module is used to define each computing server as an undirected graph node, using the product of the computing power intensity and the bandwidth potential as the basic weight and the reciprocal of the geographical location parameter as the auxiliary weight. The Prim improved algorithm is used to search for an initial topology covering all undirected graph nodes with the goal of limiting the number of direct connections of a single node to no more than a preset threshold and maximizing the sum of the basic weight and the auxiliary weight. The topology link parameter calculation module is used to take the feature vector of each undirected graph node in the standardized node feature set as node attribute and each topological adjacency relationship of the initial topology structure as edge attribute through the graph neural network model, and obtain the expected communication efficiency of each link and the overall topology vulnerability index after aggregating the neighbor information through the network layer. The link priority ranking calculation module is used to obtain the real-time load data of each computing server, and calculate the link priority ranking of each undirected graph node by combining the expected communication efficiency and the overall topology vulnerability index. The link priority ranking iteration module is used to define the action space and reward function based on the pre-trained deep Q network decision model, input the link priority ranking into the deep Q network decision model and iterate the link priority ranking through all action spaces until the reward function reaches its maximum value, thereby obtaining the optimized dynamic topology. The independent physical path allocation module is used to configure at least two independent physical paths without shared intermediate nodes for each pair of nodes based on the optimized dynamic topology through the multipath transmission control protocol, thereby obtaining a multipath mapping table networking scheme.

9. An electronic device, comprising: The method includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the communication networking method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed by a processor, enable the communication networking method as described in any one of claims 1 to 7.

Images